doi
stringlengths 28
28
| title
stringlengths 19
311
| abstract
stringlengths 217
5.08k
| plain language summary
stringlengths 115
4.83k
| article
stringlengths 3.87k
161k
|
|---|---|---|---|---|
10.1371/journal.pcbi.1002484 | Theoretical Analysis of Competing Conformational Transitions in Superhelical DNA | We develop a statistical mechanical model to analyze the competitive behavior of transitions to multiple alternate conformations in a negatively supercoiled DNA molecule of kilobase length and specified base sequence. Since DNA superhelicity topologically couples together the transition behaviors of all base pairs, a unified model is required to analyze all the transitions to which the DNA sequence is susceptible. Here we present a first model of this type. Our numerical approach generalizes the strategy of previously developed algorithms, which studied superhelical transitions to a single alternate conformation. We apply our multi-state model to study the competition between strand separation and B-Z transitions in superhelical DNA. We show this competition to be highly sensitive to temperature and to the imposed level of supercoiling. Comparison of our results with experimental data shows that, when the energetics appropriate to the experimental conditions are used, the competition between these two transitions is accurately captured by our algorithm. We analyze the superhelical competition between B-Z transitions and denaturation around the c-myc oncogene, where both transitions are known to occur when this gene is transcribing. We apply our model to explore the correlation between stress-induced transitions and transcriptional activity in various organisms. In higher eukaryotes we find a strong enhancement of Z-forming regions immediately 5′ to their transcription start sites (TSS), and a depletion of strand separating sites in a broad region around the TSS. The opposite patterns occur around transcript end locations. We also show that susceptibility to each type of transition is different in eukaryotes and prokaryotes. By analyzing a set of untranscribed pseudogenes we show that the Z-susceptibility just downstream of the TSS is not preserved, suggesting it may be under selection pressure.
| The stresses imposed on DNA within organisms can drive the molecule from its standard B-form double-helical structure into other conformations at susceptible sites within the sequence. We present a theoretical method to calculate this transition behavior due to stresses induced by supercoiling. We also develop a numerical algorithm that calculates the transformation probability of each base pair in a user-specified DNA sequence under stress. We apply this method to analyze the competition between transitions to strand separated and left-handed Z-form structures. We find that these two conformations are both competitive under physiological environmental conditions, and that this competition is especially sensitive to temperature. By comparing its results to experimental data we also show that the algorithm properly describes the competition between melting and Z-DNA formation. Analysis of large gene sets from various organisms shows a correlation between sites of stress-induced transitions and locations that are involved in regulating gene expression.
| DNA structure has been known to be polymorphic since the earliest days of its investigation. Rosalind Franklin in her initial fiber diffraction studies found two distinct DNA structures, which she called the A-form and the B-form [1]. Transitions between these forms could be induced by changes of hydration state, with B-DNA being the hydrated form and hence presumably the biologically relevant structure. Yet the first determination of DNA structure at atomic resolution found that the sequence crystallizes into a left-handed helix, called Z-DNA [2].
Although the standard right-handed B-form helix is known to be the prevalent structure in vivo, DNA can assume many other conformations. Some, such as the A-form and the strand separated state, can occur in any DNA sequence, although the latter prefers A+T-rich sequences. Other conformations have either specific sequence requirements or strong preferences for certain sequence types. These include the Z-form, which prefers alternating purine-pyrimidine sequences, the cruciform, which requires a high degree of inverted repeat symmetry, the triple stranded H-form, which needs long, mirror symmetric homopurine or homopyrimidine runs, and the four stranded G-quadriplex structure, which requires four runs of G's in close proximity.
Transitions from B-form to alternate DNA structures can be induced in susceptible sequences in a variety of ways, including changes of temperature, ionic conditions, hydration, or superhelical state. The first three of these conditions are approximately constant in vivo; only the level of imposed DNA superhelicity is subject to physiological changes that can affect the propensity of the molecule to transform to alternate conformations. Substantial levels of negative superhelicity are imposed on DNA in vivo by gyrase enzymes in prokaryotes, and by transcriptional activity in all organisms [3]. Although the superhelicity imposed by transcription in eukaryotes is transient, it is known to travel over kilobase distances and to persist long enough to drive DNA structural transitions [4].
Negative superhelicity imposes undertwisting torsional stresses on the DNA, which can induce transitions to alternate conformations that are less twisted in the right-handed sense than is B-DNA. Transitions to such states decrease the local helical twist, and thereby relieve some of the imposed superhelical stress. A transition will become favored at equilibrium when the amount of stress energy it relieves exceeds its energy cost. In vitro experiments have demonstrated superhelical transitions from the B-form to each of several types of alternate structures, including Z-DNA [5], [6], H-DNA [7], locally strand separated DNA [8], [9], and cruciforms [10]–[12]. A structural transition also has been observed to occur in a superhelical plasmid that contains a quadriplex-susceptible region, although it was not verified that the alternate structure involved is the quadriplex [13]. It has been suggested that this region instead prefers to form H-DNA, to which it also is susceptible [14].
Since genomic DNA often has numerous sites whose sequences are susceptible to forming these alternate structures, in principle there are many different combinations of transitions that can occur in response to imposed negative superhelicity. Moreover, the transition behavior of each susceptible site is coupled to the behaviors of all other sites that experience the same superhelical stress. This coupling occurs because each transition relieves some of the imposed stress, which alters the probability of transition at all other sites throughout the region involved. In this way imposed superhelicity induces a global competition among all the sites that are susceptible to any type of transition. Non-linear and highly complex correlations occur among the transition behaviors of susceptible regions throughout the domain. To analyze this competition in its full complexity, it is necessary to develop methods that can treat the simultaneous occurrence of multiple competing transitions of different types. This paper presents the first computational method to analyze competing superhelical transitions in this way.
Several theoretical models have been developed previously to analyze superhelical transitions in DNA. The earliest models were mechanical in nature, treating the transition as an “on-off” mechanism at a single susceptible site within a sequence that was otherwise unable to transform [15]–[18]. Subsequently, more detailed models were developed that used statistical mechanics to analyze transitions at a single susceptible site in a transition-resistant background [19]–[21]. This strategy is the basis for the Z-Hunt algorithm, which searches for individual Z-susceptible regions within a sequence by assessing the ability of each to undergo transition when placed alone in an otherwise non-transforming plasmid [22]. Competitions between two susceptible sites within a non-transforming background were also treated in this way. In some cases these involved two sites susceptible to the same type of transition [18], [23], [24]. In others idealized competitions between different types of transitions were examined, such as cruciform extrusion vs B-Z transitions [18], [25] and denaturation vs B-Z transitions [16], [26]. Although these models illuminated basic properties of superhelical transitions, they did not include the full competition among multiple sites that can occur in genomic sequences.
It was soon recognized that a complete statistical mechanical treatment was required to accurately simulate the competitive behavior of conformational transitions in superhelical DNA sequences of kilobase lengths, which commonly contain numerous sites whose sequences render them imperfectly susceptible to transitions of several types. Several algorithms have been developed to analyze superhelical transitions to a single type of alternate structure in DNA sequences where every base pair is regarded as being able to assume that structure [27]–[31]. A conformational state is determined by specifying which base pairs are in the B-form state and which are in the alternate structure. These states are weighted according to the Boltzmann distribution, from which equilibrium properties of the system are determined under given environmental conditions and levels of supercoiling. This approach has been applied individually to each of several types of transitions, including strand separation and B-Z transitions, and a modified version has been used to treat cruciform extrusion [17], [30]–[32]. Some of the techniques that have been developed are formally exact but computationally very slow [30], while others are approximate. Among the approximate methods, the SIDD (stress-induced duplex destabilization) and the SIBZ (stress-induced B-Z transition) algorithms for treating strand separation and B-Z transitions, respectively, are based on a similar algorithmic strategy, which has proven to be both highly accurate and computationally efficient [31], [32].
In order to develop quantitatively accurate statistical mechanical methods, it is necessary to have detailed knowledge of the alternate conformations being analyzed. One must know the geometry and flexibility of each alternate conformation, the energies of junctions between that structure and others (most importantly the B-form), and the sequence-specific energetics of the transition from B-DNA to each conformation. Since strand separation and B-Z transitions have been implicated in biological functions, they have been widely studied. So the information regarding the energetics of these two transitions is available that enables their quantitative analysis. For this reason in this paper we focus on applying our multi-state approach to the competition between denaturation and B-Z transitions in superhelical molecules.
Local separation of the two DNA strands at the correct times and locations is necessary for the initiation of transcription and replication, two key functions of DNA. Superhelical strand separation was the first DNA transition to be rigorously modeled in a way that enabled the analysis of sequences having arbitrary lengths [28], [29], [32], [33]. The SIDD algorithm that was developed for this purpose has been applied to analyze a wide variety of DNA sequences, including complete genomes. Its results agree closely with experimental observations of the level of supercoiling required to drive stand separation and the locations of the melted regions within a sequence in all cases where experiments have been performed [29], [32], [34]–[36]. Since it costs less energy to melt an AT base pair than a GC base pair, local strand separation tends to occur in the A+T-rich regions of a sequence. Stress-induced duplex destabilization has been implicated in a variety of important biological processes, including the initiation of transcription from specific promoters, the functioning of replication origins in yeast and viruses, and scaffold attachment in eukaryotes [34]–[42].
Shortly after the discovery of Z-DNA it was theoretically predicted and experimentally verified that transitions to this structure could be driven by physiologically attainable levels of negative superhelicity [2], [16], [19], [20]. Z-DNA has been experimentally detected at inserted Z-susceptible regions in torsionally stressed bacterial DNA, both in vitro and in vivo [21], [43]–[47]. There is strong indirect evidence suggesting that Z-DNA also may occur in eukaryotic genomes in vivo [48]–[51]. At present, specific biological activities of Z-DNA have not been fully elucidated, although there is substantial indirect evidence that it may serve regulatory functions in several processes [48]. The repeat unit of Z-DNA is a dinucleotide, with one base pair in the anti and the other in the syn conformation. Although Z-DNA is known to prefer alternating purine-pyrimidine sequences, specifically or runs, it can occur in other base sequences at a higher energy cost [22]–[24], [43], [45], [52]–[54]. The junctions energies and the free energies of the B-Z transition have been determined for all ten dinucleotides, including their dependence on their anti/syn character [21], [22], [25], [52]–[54].
The first theories developed to study B-Z transitions treated highly simplified cases in which a transition could only occur at a single uniformly Z-susceptible site [16], [18]–[20], [25]. An extension of this approach has been developed, which uses a thermodynamic model to calculate the propensity of an individual segment, extracted from a genomic sequence, to form Z-DNA when placed in a Z-resistant background [22], [55]. However, a base composition-dependent statistical mechanical model is required to calculate the competitive B-Z transition behavior of kilobase length DNA sequences. We have recently implemented the first algorithm, called SIBZ, that performs this type of analysis [31]. The SIBZ algorithm uses the same basic computational strategy as SIDD, but substantial modifications were needed to treat the B-Z transition. The results of SIBZ agree well with experimental measurements of the onset of transition as a function of superhelicity [5], [52], as well as experimental determinations of the locations where the superhelical B-Z transition occurs within genomic DNA sequences [49]–[51].
In this paper we develop the first algorithm that evaluates the statistical mechanical equilibrium behavior of a negatively supercoiled DNA molecule of kilobase length and specified sequence that is susceptible to multiple types of conformational transitions. Our method calculates separate transition profiles (i.e. the probability of transition of each base pair in the sequence) for each type of competing transformation. It also can calculate ensemble averages of other important parameters, including the number of transformed base pairs of each type, the number of regions experiencing each type of transition, the overall probability of transition to each type of secondary structure, and the probabilities of different types of transitions occurring simultaneously. The algorithm we develop to handle multiple competing transitions is based on and generalizes the numerical strategy used in SIDD for approximating the exact partition function. Although it necessarily makes some approximations due to the computational limitations of exactly evaluating the partition function, the SIDD-based approach has been demonstrated to provide accurate results in reasonable computational times.
In principle the method we present can be used to analyze competitions among any number of different types of transitions. In practice, however, to make quantitative predictions one must know both the geometry of the relevant secondary structures and the base pair-specific energetics of transitions to those conformations. Here we implement the model for a three state system in which each base pair can occur in the B-form, which is regarded as the ground state, or in either of two alternate conformations. We explicitly analyze the competition between superhelical strand separation and B-Z transitions since their energy parameters are known at comparable temperatures and ionic conditions. Although the energetics governing these two transitions have the same orders of magnitude in the physiological temperature range, denaturation is more temperature dependent while B-Z transition causes greater relaxation. Our analysis shows that, because of these properties, the competition between these two types of transitions is quite complex, involving the interplay between base composition effects, imposed superhelical density, and environmental conditions. We call the new algorithm BDZtrans, for B-form to the Denatured and/or Z-form transitions.
Constraining a DNA molecule either into a closed circle or a topological loop fixes its linking number , the number of times either strand links through the closed circle formed by the other strand. When the resulting topological domain is relaxed under physiological conditions, the molecule is entirely B-form with average twist rate of turns per base pair (bp). The linking number of the relaxed state of a domain containing base pairs is given by .
A molecule that is negatively superhelical has linking number that is smaller than its relaxed value, so its linking difference (also called its superhelicity) is . The superhelix density normalizes the level of superhelicity relative to the length of the DNA experiencing it.
Superhelicity imposes torsional stresses on DNA. These stresses can be relieved by various deformations, including local conformational transitions to secondary structures having unstressed helical twist that is smaller than that of the B-form, . This includes a wide variety of DNA structures, such as strand separated DNA, Z-form, H-form, cruciform, and quadriplex DNA.
A statistical mechanical model is required in order to analyze the equilibrium transition behavior of a superhelical domain having a specified base sequence. At equilibrium the available conformational states are occupied according to their free energies. If a state has energy , then at equilibrium it is occupied at a frequency proportional to , where . So the equilibrium probability of that state is , where the normalization factor is the partition function of the system, given by . Once the resulting equilibrium distribution (also called the Boltzmann distribution) has been obtained, it is straightforward to compute ensemble averages of any parameter of interest [56]. If a parameter has value in a state , then its ensemble average value is(1)
Analysis of the superhelical transition behavior of DNA requires knowledge of the structure of each alternate conformation, and of the free energies associated with transitions from the B-form to that conformation. The energy required to transform each base pair must be known, including its dependence on base sequence, under the conditions of temperature and ionic strength assumed in the calculations. In addition, there is a substantial nucleation energy for transitions to alternate DNA structures, which may be regarded as the energy needed to form junctions between that structure and its neighboring structures. Although this may vary according to the neighboring structure that it abuts, in practice this is usually B-form DNA. In cases where the alternate structure is flexible, such as strand separated DNA, significant twist can be absorbed by additional winding of that structure. This requires free energy that is modeled by an elastic Hooke's law as a quadratic function of the twist density. This term is not required for structures that are approximately as stiff as B-DNA, such as Z-DNA, or for structures that cannot undergo interstrand twisting, such as cruciforms. Finally, there is a free energy associated with the residual superhelicity , which is the superhelicity not absorbed by the changes of twist consequent on transition.
Suppose that in a state there are regions of alternate structure. Different regions may be in different alternate conformations. Let the -th such region consist of base pairs in an alternate conformation that has unstressed helicity . Its formation relaxes an amount of helicity given by turns/bp. If this alternate structure is flexible, let each base pair be twisted by (radians/bp) away from its unstressed structure. If it is not flexible then . Also, suppose that the twist (in turns) of each junction at the -th region is . Then the residual superhelicity of the domain, which is the amount of superhelicity remaining to stress the molecule after all changes of twist due to transformations to the alternate conformations have occurred, is given by(2)The first term in this equation is the linking difference imposed on the molecule. The free energy associated with superhelicity has been measured to be quadratic [57].
Consider a topological domain consisting of base pairs that is susceptible to types of conformational transitions, . Suppose that a state of this domain contains specific base pairs in conformation . The total free energy of the state is given by(3)The first term in this expression is the total nucleation energy associated to this state. Suppose that there are junctions between the B-form and the -th alternate conformation, each of which has energy . Also assume that there are junctions between the alternate conformations and , , each with energy . Then the total nucleation energy associated with this arrangement is(4)
The second term in Eq. (3) sums the transition energies of the -th base pair to be transformed to conformation over the regions of transition. This transition energy varies with the type of transition the base pair experiences, the identity of the base pair (and sometimes also the identities of its neighbors), temperature, and ionic conditions. This term is summed over the number of base pairs involved in each transition.
The third term in Eq. (3) is the Hooke's law torsional energy associated to the twisting of alternate conformation . The parameter is the torsional stiffness coefficient of conformation , and is its helical twist rate away from its relaxed conformation, measured in rad/bp. This term is required for strand separation, where the necessary parameter values have been determined [29]. However, it is not needed for transitions to other alternate conformations, such as the B-Z transition, whose twisting (along with that of the B-form DNA) is regarded as being incorporated into the residual superhelicity . The free energy for conformations such as cruciforms, in which the two strands are physically separated in space and do not twist around each other, also does not contain this term.
More generally the helical twist could be considered to fluctuate independently for each transformed base pair. This has been done in a formally exact analysis of superhelical strand separation [30]. However, no significant difference was seen between the results of analyses that allowed independent twists, and those where all denatured base pairs were assumed to have the same twist , as in Eq. (3). Therefore, in this analysis we choose the latter strategy.
The last term in Eq. (3) is the quadratic energy associated with the residual superhelicity , as defined in Eq. (2). The values of the constant and the other energy parameters are discussed below for the specific transitions modeled there.
The expressions presented in this section can be applied to analyze molecules in which any number of superhelically driven transitions compete, provided the helical twist rates and transition energies of the alternate structures are known.
Several computational strategies have been developed to analyze superhelically driven transitions in genomic DNA. Historically, the first fully developed algorithm focused on strand separation, since this is the only transition known to be required for essential biological processes such as the initiation of transcription and of replication. Although an exact theoretical method has been implemented that is capable of computing transition probabilities of individual base pairs in kilobase-scale genomic sequences, it proved too computationally cumbersome for widespread use [30]. However, an alternate strategy has been developed, called SIDD, that performs accurate and efficient approximate calculations. The initial SIDD algorithm focused on the superhelical strand separation transition [32]. This approach subsequently was modified into the SIBZ algorithm in order to treat superhelical B-Z transitions [31].
In this paper we further develop this computational strategy to enable efficient calculations of the equilibrium properties of superhelical molecules that are susceptible to multiple types of competing transitions. Although we focus specifically on the competition between denaturation and B-Z transitions in kilobase length superhelically constrained DNA sequences, in principle this approach can be applied to any number of different transitions.
The basic strategy of the algorithm is first to determine the lowest energy state of the system . Then a threshold is set, and all states having energies less than are found and included in the analysis [28], [32]. The number of states that are included, and hence the execution time, increases with , whose value must be chosen to suit the conditions assumed in the calculation and the level of accuracy desired.
Extensive calculations using both SIDD and SIBZ have shown that this approach has an attractive combination of efficiency and accuracy. Comparisons of the SIDD results with those from an exact method show that this approximate approach has an accuracy of at least four significant digits in all calculated parameters at physiologically attained superhelicities when a threshold is chosen between 10 and 12 kcal/mol [32]. This accuracy is more than sufficient for comparison with experimental data. Although states having energies above the threshold are not explicitly included, a density of states technique has been developed that can approximately correct some calculated parameters for the cumulative effect of these high energy states [28]. Such corrections are beyond the fourth decimal of accuracy and hence are rarely needed in practice.
Calculations on 5 kb segments under standard conditions take on average about 10 seconds on one Opteron processor, although some segments can require up to 5 minutes or even longer. The difference in execution times strongly depends on the base composition of the sequence. For example, when there is a dominant transition region with a low energy cost, transformed states that do not include this region have a much higher energy. Therefore, relatively few states are found in the energy range determined by the threshold, resulting in a quick execution time. However, if no dominant transition regions are present, many states will have comparable energies, requiring a longer run time.
The free energy associated to each state of our competitive system, shown in Eq. (3), is comprised of two parts. The first two terms in this equation assign nucleation and base-dependent transition energies to the set of transformed base pairs in each state. These energies only depend on which base pairs are transformed and the alternate structures that they assume. As their domains are discrete sets, they are collectively referred to as the discrete part of the energy expression. Once the secondary structures of all base pairs have been specified, it remains to partition the balance of the superhelical deformation between residual superhelicity and twisting of the denatured regions. Since this partitioning can be done in a continuous manner, this term is described as the continuous part of the energy. It contains the energy terms arising from the superhelical constraint and from the twist. This separation of the energy expression into discrete and continuous parts facilitates finding the minimum energy state, as well as the states that satisfy the threshold condition, in a computationally efficient way [32].
We first calculate the energy associated to the discrete states, the nucleation energy given in Eq. (4) and the base-dependent transition energies . We consider segments of length along the molecule. These segments are regarded as being susceptible to any type of transition, as long as they meet the sequence requirements for that alternate conformation. This is done for values of up to a limit . The value of is chosen so that all states with longer runs of transition of any type will have energies higher than the threshold at physically reasonable values of the superhelix density . In addition, some transition types may have a lower limit on the segment length .
Consider a circular molecule base pairs long. (We discuss linear molecules below.) In this molecule there are different segments of each length , , one starting at each base location. For simplicity, each segment is assumed to border B-form DNA on both sides. The total transition energy of each segment is(5)The free energies found this way are sorted according to increasing energy into separate arrays for each transition type. The rows of these arrays are indexed by the length of the transformed segment. Each array has columns, equal to the length of the sequence being analyzed. In these arrays the first position of each transformed segment is stored along with its energy, as this information is required later in the calculation. Since the discrete components of the state energy are additive for multiple run states, these sorted arrays are also used to determine the discrete energies of states in which more than one run of transition is present.
To consider the one-run states of the system, we add the appropriate quadratic free energy associated with the residual superhelicity to each entry in the -th row of all the arrays containing the discrete energies. This is done for each type of transition. (The manner in which the torsional deformation energy is treated in the strand separation transition is described in the next section.) The resulting energy values remain sorted within their rows, which enables an efficient search to be conducted for the lowest energy state among the untransformed or 1-run states. This lowest energy is taken as the initial value of . We next find all states whose energies satisfy , as described below. If in this process a multirun state is discovered whose energy is less than the current value of the minimum energy, then is assigned this lower value, which is used in the subsequent calculation. In practice this reassignment only occurs for a small fraction of sequences analyzed, and only when analyzed at extreme negative superhelicities. However, when it occurs more states are included than the final threshold cutoff condition requires, giving a correspondingly (very slightly) more accurate approximation.
For multiple run states, the procedure followed is similar to that described above for one run states. For each number of runs, the algorithm considers all transition types, and the total number of runs , where is the maximum number of runs considered. In general, each transition type includes a high initiation cost for each additional run. When the number of runs becomes large enough, all such states will have energies that exceed the threshold, and hence will not be included in the analysis. For this reason it is appropriate to impose a limit on the total number of runs that are considered. This is done by calculating whether any state with a given number of runs could satisfy the energy threshold condition by assuming that all transforming base pairs have the lowest possible transition energies, so the transition becomes isoenergetic. If it is found that such a state could satisfy the threshold condition, then a search of states with that number of runs can be instituted.
The Boltzmann factors associated with each state are accumulated into arrays for each transition type that are indexed by the lengths of their participating segments and their positions within the sequence. The contributions to the partition function for each type of conformation are collected separately in order to calculate their individual probability profiles. For multiple run states in which more than one type of transition occurs, the information for each run is placed in the appropriate array according to its length, position, and transition type. Details of these procedures may be found elsewhere [32].
A variety of equilibrium properties of the transition may be calculated from the information that is collected in these arrays. This includes the probability that each base pair in the sequence is in a particular alternate conformation, the expected number of runs of each transition type, the probability of the state with no transition, and other attributes of interest.
We focus henceforth on analyzing the competition between denaturation and B-Z transitions in a superhelical plasmid base pairs long and having any specified sequence. We first examine the residual superhelicity associated with this competition, given in Eq. (2), and then consider the state energy described in Eq. (3).
Strand separated DNA, being untwisted when unstressed, has turns/bp. It follows that the transition of base pairs from B-form to the unstressed strand separated state involves a twist decrease of turns. Since single-stranded DNA is highly flexible, the two separated strands in a melted region are able interwind in order to further relieve supercoiling stresses. The amount of helical interwinding that occurs is denoted by in radians/bp.
The helicity of Z-DNA is turns/bp, the minus sign indicating that it is twisted in the left-handed sense. So the decrease of helicity for each base pair experiencing this transition is , which is approximately of untwisting per transformed base pair. Since the Z-form is torsionally stiff we do not consider its twist fluctuations separately, but rather regard them, together with those of the B-form regions, as part of the residual superhelicity. In addition, each B-Z junction requires an untwisting of turns [21].
Since the Z-form is favored in G+C-rich regions and strand separation in A+T-rich regions, in practice they are unlikely to both be competitive at the same regions. In particular, junctions where strand separated DNA directly abuts Z-form DNA are unlikely to occur in low energy states under the conditions assumed below. In this case the nucleation energy of Eq. (4) can be written as(6)where there are runs of conformation (). (A run is defined as a segment in which all base pairs are in the same alternate structure.) Here the nucleation energy of a single run of type is , the cost of producing two junctions between B-DNA and that conformation.
We consider a state in which there are denatured base pairs in runs, and Z-form base pairs in runs. Because the unit cell of Z-DNA is a dinucleotide, is an even number. Then the residual superhelicity whose general form is given in Eq. (2) becomes(7)
The total free energy associated to this state is given by(8)The values of the various energy parameters found in this equation are discussed in the section below.
In describing a particular state one first specifies the conformation of each base pair in the sequence being analyzed. Here they may be either B-form, Z-form, or melted. This determines the numbers and of transformed base pairs, and the numbers and of runs for each transition. This fixes all the factors in Eq. (7) except for the residual superhelicity and the twist of the denatured regions. There is a continuum of ways to partition the balance of the topological constraint between and of the single stranded regions. In previous papers we have developed and evaluated a number alternative ways of treating this partitioning [28], [30]. We found that high accuracy can be achieved by minimizing the total free energy associated with these two quantities, which are the two terms on the right in Eq. (8), subject to the condition that the sum remains constant. This minimum occurs when(9)Combining previously described terms and using this minimization condition in Eq. (8), we obtain the following expression for the free energy of a state when denaturation and B-Z transitions compete:(10)
In the present implementation we assume that strand separation is governed by copolymeric transition energies. That is, every or base pair is assigned the same separation free energy , while every or base pair is given separation free energy . Nearest neighbor energetics have been measured for strand separation under various environmental conditions [58]–[60], and their use has been implemented as an option in the SIDD algorithm [61]. Although these can easily be incorporated into the present analysis, we choose to use the computationally slightly faster copolymeric energetics, since little practical difference has been seen between the results found using these two approaches.
The free energy of strand separation depends on temperature according to the relationship(11)where for A or T bases and for C or G bases, respectively. The enthalpy of this transition has been measured to be kcal/mol and kcal/mol [29]. The entropy term in this equation is related to the transition temperature , which is the temperature at which the transition energy . In turn, varies with ionic strength according to(12)where is the salt concentration in molar units, the temperature is in degrees Kelvin, , and .
The remaining energy parameters found in Eq. (10) that involve strand separation have been well evaluated at salt concentration 0.01 M and temperature 310 K, where the most sensitive experiments to analyze superhelical strand separation were conducted [9]. Analysis of these experimental results determined the torsional stiffness coefficient to be , and the nucleation energy to be kcal/mol [62]. We assume no temperature dependence for these parameters. The superhelical energy parameter is , where is the gas constant, is the temperature, and is the number of base pairs in the sequence being analyzed [63], [64].
The unit cell of Z-DNA is a dinucleotide (i.e. two neighboring base pairs), with one in the anti and the other in the syn configuration. Therefore, we henceforth regard the B-Z transition energy as referring to the energy of forming a unit cell of Z-DNA, hence associated to two base pairs. In addition, there is an energy cost required when two neighboring dinucleotide repeat units break the anti-syn alternation, as happens for example in the (AS)(SA) arrangement. This Z-Z junction energy is denoted by and is added to the total Z-forming free energy of states in which these junctions occur [31].
B-Z transition energetics have been determined for each of the ten possible dinucleotides. Most of these free energies were experimentally measured at room temperature and 0.1 M sodium concentration [21], [52]–[54], although some were estimated [22], [25]. The B-Z transition energies of all ten dinucleotide pairs and the corresponding Z-Z junction energies at these environmental conditions are given in [22], [31].
When the transition properties of uniformly Z-susceptible inserts within a pBR322-derived plasmid were determined from two-dimensional gel electrophoresis experiments, it was found that the B-Z transition behaved the same at C as at C [26]. This suggests that there is no significant temperature dependence of the B-Z transition, at least for the inserted sequence that was used. For this reason we assume that the dinucleotide, Z-Z junction, and the B-Z junction energies are all independent of temperature. If future measurements find differently, this assumption can easily be modified.
There is an intricate interplay between the superhelical B-Z and local denaturation transitions that arises from the different energy dependencies of these reactions. Transition to Z-form involves a greater change of twist than does strand separation, so it relieves more superhelical stress. This suggests that the B-Z transition will occur at less extreme superhelicities than strand separation at physiological temperatures, other influences remaining fixed. However, the transition free energy of strand separation is highly temperature dependent, whereas the B-Z transition appears to be approximately independent of temperature. In consequence, as temperature rises one expects a change of behavior in a superhelical molecule that is susceptible to both types of transitions. At relatively low temperatures one expects B-Z transitions to dominate, with strand separation occurring only at more extreme superhelicities after the low energy Z-susceptible regions have transformed. However, as the temperature rises the free energy cost of strand separation diminishes, so one expects this transition to become more competitive. The calculations we report below suggest that these energy effects can render the competition between these two types of superhelical transitions quite complex in practice.
In previous sections we summarized the general algorithmic strategy for treating multiple transition types. Here we describe how this algorithm is tailored specifically to model the competition between superhelical strand separation and Z-DNA formation.
The state energy has been separated into the discrete ( and ) and continuous () parts as described in Eq. (10). We first consider the energy associated to the discrete states. We regard each segment of length along the molecule to be susceptible both to strand separation and, when is even, to Z-form. In a circular molecule of length and longest segment this produces a matrix of denaturation energies of dimension , and a matrix of B-Z transition energies of dimension . (Here the square brackets denote the greatest integer function.) We limit the minimum number of Z-forming dinucleotides in a single run to four, since shorter runs of Z-DNA have not been found experimentally [21]. We subtract three from the number of lengths considered because we do not include Z-runs comprised of 2, 4, or 6 dinucleotides.
As described in the previous section, we assign copolymeric energies to denatured regions according to their base sequences. It follows that the discrete energy associated with the denaturation of specific base pairs is(13)where is the number of denatured or base pairs, so is the number of denatured or pairs, and is the number of denatured regions present.
Next, we determine the energetics of transition to Z-form of runs that together contain transformed dinucleotides, hence base pairs. First, we find the most energetically favorable anti/syn conformation according to its base sequence, and then we calculate the total energy by summing the energies of each Z-DNA dinucleotide and all the occurring Z-Z junctions. Details of this procedure are provided elsewhere [31]. The discrete B-Z transition free energy is given by(14)where is the number of Z-Z junctions, and and are the dinucleotide and junction energies, respectively.
The discrete energies are calculated from the above equations for single runs of transition of any length in the range . The free energies found this way are sorted according to increasing energy into arrays for denaturation and for the B-Z transition as described above. The rows of these arrays are indexed by the length of the transformed region, which is base pairs for denaturation and dinucleotides for the B-Z transition.
In each execution of the algorithm one initially fixes the imposed linking difference and the temperature . This determines the parameters associated with the continuous component of the state free energy. This is the quadratic last term in Eq. (10), which varies with the numbers of melted bases and of Z-form bases, and the number of Z-runs : . We note that does not depend upon the positions of the runs of transition within the sequence. If there are runs of transition in a state and transformed base pairs in the -th run, then(15)where {0, 1} denote the denatured and the Z-form states, respectively.
For one-run states of the system, we add to each entry in the -th row of the array containing discrete energies for strand separation, and to the corresponding entries of the array. Since the nucleation energy is significant for both denaturation and B-Z transitions ( kcal/mol in both cases), a limit on the total number of runs considered may be imposed. In SIDD analyses with threshold kcal/mol it was found that states with more than three runs were never found under reasonable conditions [32]. For the SIBZ algorithm it was determined that states with more than four runs would generally not occur [31]. Therefore we impose a limit of four runs in the present algorithm.
For each number of runs, the algorithm is iterated for , where when a maximum of four simultaneous runs is allowed. This arrangement assures that there are three types of two-run states (two melted regions, two Z-regions, or one melted and one Z-region). Similarly, there are four types of three-run states, and five types of four-run states. The energy associated with a state is calculated by adding the appropriate energies from the discrete arrays to the total superhelical energy for each number of runs of the appropriate types. For example, the energy of a two-run state where bases are strand separated and bases are in Z-form is , where and depend on the base sequence of the runs involved. Whenever the total energy associated to a state satisfies , its Boltzmann factor is calculated and added to the appropriate arrays as described above.
We analyze linear molecules by connecting their ends with an inserted sequence, and then treating them as circular. To fully isolate one end from the other, the insert must be chosen so it is energetically highly disfavored to undergo any transition. For the B-Z and denaturation transitions this is achieved by inserting a segment between the ends. Since this run of G's is unlikely to either melt or form Z-DNA, its insertion prevents any artificial correlation between the two ends from arising, hence correctly simulates a linear sequence. The linking difference that corresponds to the specified superhelix density is imposed on the resulting molecule. The algorithm only reports the results for the actual sequence, and disregards those from the insert.
In order to assess the computational accuracy of the BDZtrans algorithm we compare its results to those from an exact analysis of a simplified situation in which competition is limited to two homopolymeric sites. Specifically, we consider a 5 kb plasmid in which there is one uniformly Z-susceptible region and a second region, at a distance from the first, that is uniformly susceptible to denaturation. The A+T-rich, easily melted segment has length , while the Z-susceptible site is a dinucleotide repeat containing base pairs. By uniform susceptibility we mean that both transitions are homopolymeric: all dinucleotides in the Z-susceptible segment have the same transition energy , and all base pairs in the denaturation-susceptible region have the same transition energy . All other parts of the plasmid are regarded as being unable to undergo any form of transition, so the competition is exclusively between these two sites. In our simulations this is achieved by giving high transition energy values to all base pairs that are not in these regions. This example approximately corresponds to an experimental situation where a highly Z-susceptible sequence, such as , is inserted into a plasmid that contains a dominating SIDD site, such as the -lactamase terminator in the pBR322 plasmid.
This case is analytically solvable by standard procedures that have been presented and applied elsewhere [16], [26]. Here we also solve it using the BDZtrans algorithm, and compare the results. In the BDZtrans analysis we use an energy threshold of kcal/mol and consider only states with four or fewer runs of transition. Calculations were performed using both methods over a range of superhelical densities and temperatures for various combinations of segment lengths and .
The analytic calculation allows the A+T-rich segment only to melt, and the segment only to assume the Z-form. However, the BDZtrans algorithm allows both regions to undergo either type of transition. In all situations where these two segments either are untransformed or experience their expected transitions, we find that the results from the two methods agree exactly up to the accuracy of double precision. (Data not shown.) Having established the high computational accuracy achieved by the BDZtrans algorithm, we now can use it to analyze other situations, where exact calculations are not possible.
The BDZtrans algorithm has been applied to analyze several situations, the results of which are reported here. First, we consider a simplified case where a plasmid contains two regions that are uniformly susceptible to superhelical transitions, one to strand separation and the other to the B-Z transition. Second, we perform an analysis of the only experimental data presently available regarding the competition between B-Z transitions and strand separation in superhelical plasmids. We show that the predictions of BDZtrans agree closely with this experimental data when energetics that are appropriate to the buffer conditions of the experiment are used. Third, we analyze the competition between strand separation and B-Z transitions in the pBR322 plasmid. Specifically, we demonstrate that there is a strong temperature dependence to this competitive behavior. Fourth, we analyze the superhelical competition between B-Z transitions and denaturation in the control regions of the c-myc oncogene. Both transitions have been shown to occur when this gene is transcribing, and have been proposed to regulate its expression. However, to date little is known regarding the competitive interactions between these transitions. Lastly, we apply BDZtrans to study transition behavior around genomic sites that regulate transcription. We compare the patterns of transition found in eukaryotic genomes around transcription start sites (TSS) with those that occur around the sites where the transcript terminates. We also compare the patterns around the TSS that are found in eukaryotes (human and mouse) with those found in a prokaryote (E. coli), and with those that occur in a class of pseudogenes that are not transcribed.
We first analyze the competition between two regions in an otherwise transition-resistant background. The specific problems addressed here were chosen to eluciate the complexities that can arise in superhelical competitions between strand separation and B-Z transitions, even in simplified situations. The intricacies of these interactions result primarily from three factors. First, as shown in Eq. (11) and Eq. (12), strand separation is strongly temperature dependent, with the transition temperatures of specific regions depending both on base composition and on ionic strength. In contrast, the B-Z transition appears to be essentially independent of temperature [26]. One anticipates that this will cause significant variations with temperature of the competition between these two types of superhelical transitions. Second, the B-Z transition relaxes substantially more superhelical stress per transforming base pair than does strand separation. Lastly, the relative lengths of susceptible regions also can strongly influence their competitions [18].
We first consider the case where the Z-susceptible insert is a sequence placed at position 1000 in a 5 kb plasmid, and the denaturation-susceptible region is at location 3000. The temperature is set at 300 K, and the superhelix density is allowed to vary. We only consider negative superhelix densities , although the results are presented in graphs as a function of . Since 300 K is substantially lower than the transition temperature for poly-A (see Eq. (12)), these conditions should favor the B-Z transition. However, as the A-rich region is long, its transition will produce more relaxation than will the short Z-forming region. We use BDZtrans to calculate the probability of transition of each base pair in each susceptible region. We then average these values over the lengths of the regions involved to find the average probability of melting of base pairs in the A-segment, and of Z-formation in the CG-region.
The results of these calculations are shown in Fig. 1(a), which graphs the average probability of each type of transition in the corresponding segment as a function of superhelicity. When there is not enough superhelical stress to drive any transition. As the superhelix density becomes progressively more extreme, B-Z transition in the segment occurs first and dominates in the range . However, since the Z-susceptible region contains ten base pairs, it can only relax less than two superhelical turns. Although denaturation relaxes less stress per transformed base pair, in this sequence the A-rich segment is much longer than the Z-susceptible segment, and hence can relax substantially more superhelicity. In the range of there is a coordinated reversion of the Z-forming region back to B-form, coupled to denaturation of the A-rich segment. In this range it is energetically too costly for both transitions to occur, so transformation of the longer meltable region becomes favored because it provides more relaxation. When the region is essentially fully melted, and additional stress induces the B-Z transition. Around both segments are essentially completely transformed. Similar coupled transition-reversion events have been noted for other competitions, including those between two Z-susceptible regions [24], [31], between two cruciform extrusions, and between cruciform extrusion at one site and B-Z transitions at another [18].
In Fig. 1(b) we show the results for the same system obtained when it is analyzed using the two-state algorithms, SIDD for denaturation and SIBZ for the B-Z transition. One sees that disregarding the competition between different types of transitions can result in an entirely incorrect representation of the transition behavior of the plasmid. First, the onset of the melting transition occurs at a lower value of in Fig. 1(b), since in SIDD denaturation is not competing with the B-Z transition, which is first to transform in the competitive situation, as shown in Fig. 1(a). Further, the reversion of the B-Z transition apparent in Fig. 1(a) is not captured here, because the competition between that transition and melting is not considered. Using the individual algorithms separately wrongly predicts a B-Z transition with probability close to one for −0.055, whereas including the competition with denaturation shows this probability actually to be below 0.2.
Next, we analyze the reverse situation, in which a short melting region competes with a longer Z-susceptible region. These calculations were performed at a temperature of = 340 K, which is higher than the transition temperature of = 322 K for A+T-rich DNA at 0.01 M salt concentration. The average transition probabilities calculated for this situation are plotted as functions of in Fig. 2. Although the high temperature should favor denaturation, the high nucleation energy of denaturation keeps the region in the B-form state when the molecule is relaxed at this temperature. The onset of transition under these conditions occurs around . Since the Z-susceptible site is much longer and hence can relax more superhelicity, B-Z transition is the first to occur. The short AT-region can only relax about 1.5 turns of superhelicity while B-Z transition, although energetically more expensive, can relieve much more stress. This greater stress relief favors the latter transition, even at this high temperature. At the B-Z transition becomes essentially complete, and denaturation starts to occur as a second transition. In the range both segments are completely transformed. However, at extreme superhelicities of , the melting probability of the segment is seen to fall gradually back to zero. At this level of supercoiling BDZtrans finds that this segment transitions from the denatured state to the Z-form. This behavior can be understood by comparing the energies required for the AT-insert either to melt or to form Z-DNA, assuming that the entire segment is already in Z-form. By equating these two free energies, it is straightforward to obtain the critical superhelix density at which the insert begins to favor the B-Z transition over denaturation. For this sequence at T = 340 K we find . The value for agrees exactly with the result obtained by BDZtrans for where the two probabilities of the segment are equal, which occurs at the intersection of their two curves in Fig. 2. This behavior also could not be predicted from separate analyses of each type of transition.
These examples illustrate some of the complexities that can occur in multi-state superhelical transitions, even in artificially simple situations. In particular, the susceptibility of a region within a sequence to undergo a certain transition is not always a simple function of its base composition. Although at high temperatures it generally it takes less energy to melt an A+T-rich region than to transform it to the Z-helix, the results presented in Fig. 2 shows that under certain circumstances the opposite behavior can happen at equilibrium, even well above the melting transition temperature where one might imagine that denaturation would dominate. These examples also show the importance of competing together all transition-susceptible sites in the sequence, rather than simply analyzing the propensity of each individual region to transform independent of the rest of the domain.
To date only one experimental investigation has been performed of the competition between melting and the B-Z transition in supercoiled DNA [26]. The pAT153 plasmid used in those experiments is a derivative of pBR322 that contains an A+T-rich, easily meltable 105 bp region at its -lactamase gene terminator. Two other plasmids were constructed by inserting a Z-susceptible sequence into pAT153 in order to observed the competition between strand separation and the B-Z transition in a situation where the regions involved do not abut. The pCG8/vec plasmid was constructed by inserting the highly Z-susceptible sequence, and the pTG12/vec plasmid was constructed by inserting . This insert is also susceptible to B-Z transition, although it requires approximately twice the energy per dinucleotide to transform as does .
Each plasmid was subjected to two-dimensional gel electrophoresis to determine its transition behavior over a wide range of linking differences. The amount of residual superhelicity present at each linking difference can be measured directly from the 2-D gel data, as described elsewhere [65]. From this information the extent of transition-induced relaxation can be found as . We analyzed the transition behavior of each of these three plasmids directly from the original gel images, which were kindly provided by the experimental investigators [26].
To compare these experimental results with theory we used the BDZtrans algorithm to calculate the equilibrium transition behavior of each complete plasmid over the experimental range of linking differences. We consider superhelical competition between all base pairs in each plasmid, rather than isolating the competition between the transition susceptible sites. By inserting the condition from Eq. (9) into Eq. (7), we obtain the following expression for the ensemble average value of the residual superhelicity when strand separation and B-Z transitions are competing:(16)Here the expressions in angled brackets are ensemble averages of the bracketed parameters, which are calculated using our numerical method. The relaxation experienced by a topoisomer due to transitions is given by . This quantity can be compared to the extent of relaxation experienced by the experimental plasmids, which is determined from the 2-D gels.
A single transition was observed experimentally in the pAT153 plasmid, which contains the A+T-rich region found in pBR322, but no Z-susceptible insert. This transition was reported to be highly temperature dependent, indicating that it is strand separation [26]. In contrast, two transitions were observed as negative superhelicity was increased in the pCG8/vec plasmid at = 305 K. The first transition, at the less extreme superhelicity, was suggested to be a B-Z transition both by chemical probing, and by the degree of unwinding it exhibited. This transition was found to be essentially independent of temperature, behaving the same at 281 K as at 305 K. This further confirmed its B-Z character, as strand separation is well known to be highly temperature dependent. The nature of the second transition in this plasmid was not determined experimentally. However, it was assumed to be denaturation, primarily because it behaved in a qualitatively similar manner to the transition observed in pAT153. The pTG12/vec plasmid also showed two transitions, which were delayed in linking difference relative to those seen in pCG8/vec.
The 2-D gel experiments were performed in TBE/2 buffer, which contained 45 mM tris borate and 0.5 mM EDTA at pH 8.3. Unfortunately, the energetics of denaturation and of the B-Z transition have not previously been determined under these conditions. In particular, the energetics for strand separation described previously were determined at pH 7.0 [9], and no correction for a higher pH is known. Since this difference in pH level constitutes a 20-fold decrease in the counterion concentration, it is reasonable to suppose that it could affect the energetics of melting, which are known to vary with the concentrations of larger monovalent counterions.
When we ran the BDZtrans algorithm on the pAT153, CG8/vec, and TG12/vec plasmids using the energy values described in the “Energy Parameters” section, we found that their qualitative experimental behaviors were correctly depicted by the numerical model. However, it was apparent that the transition energy parameter values used were too large, since the transition behavior predicted by BDZtrans was consistently shifted to more extreme superhelicities than were observed experimentally. Therefore, our first task in comparing numerical calculations with experiments was to determine the transition energetics appropriate to the experimental conditions.
We did this in the following manner. First, we used the SIDD algorithm, which considers only strand separation, to fit the experimental data on the melting transition in pAT153. To determine the relaxation as described above, we set the parameters and to zero in Eq. (16), since the B-Z transforming insert is not present in this plasmid. When performing these fits to the data, we chose to vary only the transition energetics per base pair and to keep all other parameters fixed. The best fit with experiment was achieved when the sequence averaged transition energetics of the easily melted region are 0.45 kcal/mole per bp. This is about 0.3 kcal/mole/bp smaller than the value derived from the information in the “Energy Parameters” section, which pertain under other experimental conditions. It is well known that the melting energy of DNA decreases as salt concentration is lowered, as is shown in Eq. (12). The present analysis suggests that a similar decrease may also occur when the counterion is .
Next, we used the SIBZ algorithm, which considers only the B-Z transition, to analyze the first transition in the pCG8/vec plasmid. In this case we set in Eq. (16), since there is no denaturation in this regime. We find that the best fitting B-Z transition energetics for the CG dinucleotide is approximately 0.1 kcal/mole/bp lower than the values found in [22], [31]. The same analysis of the first transition in the pTG12/vec plasmid gives a similar result for the TG dinucleotide. We note that these results are in qualitative agreement with those found by the authors of the experimental paper [26], who used a different and rather simpler method of analysis.
Finally, we used the BDZtrans algorithm to analyze the full competition between strand separation and B-Z transitions in the pCG8/vec and pTG12/vec plasmid sequences. This was done at T = 305 K using the fitted energy parameter values found above for both transitions. The BDZtrans results are plotted as solid lines in Fig. 3(a) and (b), while the dots with error bars represent experimental data. The horizontal axis shows the imposed superhelicity , and the vertical axis plots the relaxation . We find close agreement in both cases between the results of BDZtrans and the experimental data. This accord shows that, given the correct energy parameters, the BDZtrans algorithm captures the competition between two alternate structural transitions in a quantitatively accurate manner.
Both experiments and SIDD analysis have shown that the superhelical pBR322 plasmid contains a dominant melting region that is about 105 bp long and coincides with the -lactamase gene terminator [9], [30]. Antibody binding experiments and SIBZ analysis of this plasmid has shown that it also contains several short Z-forming segments, the longest of which consists of 14 bp [31], [66]. Here we use the BDZtrans algorithm to analyze how these two transitions compete under various conditions. Specifically, we calculate the probability of each transition occurring anywhere in the plasmid. This probability is defined as , where = , , and is the sum of all the Boltzmann factors for all states in which at least one region is in conformation . The probability that both types of transition occur in the same molecule also is calculated. We analyze the transition behavior of this plasmid at base pair resolution by also calculating the probabilities of each base being in either alternate conformation at equilibrium. The transition profiles show the graphs of these probabilities as functions of position.
First, we used BDZtrans to analyze the pBR322 plasmid sequence at superhelix density and various temperatures. The results are shown in Fig. 4(a), which plots the probabilities of each transition as a function of temperature. It is apparent that the B-Z transition dominates at low temperatures, while strand separation prevails at high temperatures. The probability of both transitions occurring simultaneously also is shown. Although at the value of never exceeds 0.5 at any temperature, at more extreme superhelicities both transitions will occur simultaneously with high probability. We define the phenomenological competitive transition temperature to be the temperature at which both transitions are equally probable. At this superhelix density K.
Figs. 4(b) and 4(c) plot the average numbers of transformed base pairs and runs of transition, respectively, for the pBR322 plasmid as functions of temperature at . Separate curves are shown for each transition type, and the total number of runs for both transitions is also given. The transition behavior seen in these graphs arises from the fact that the Z-susceptible regions in the pBR322 plasmid are several but short, while the region most susceptible to strand separation is about 105 bp long. At low temperatures and the B-Z transition is seen to dominate over strand separation. In this regime, although multiple sites are in Z-form, they together comprise only 25 to 30 base pairs on average. The difference in the numbers of base pairs undergoing each type of transition that is seen in Fig. 4(b) is also a consequence of the fact that the B-Z transition relieves substantially more superhelical stress per base pair than does strand separation.
As the temperature increases beyond 308 K strand separation comes to dominate, and the propensity to form Z-DNA falls back to zero. In this regime the number of denatured base pairs increases to large values. This behavior is a consequence of the strong temperature dependence of denaturation. With increasing temperature the energy required to denature a region goes down, so more of the imposed superhelicity is partitioned to this transition at equilibrium. Since the dominant destabilized site is long, all this melting can be accomodated within that site until the temperature reaches around 315 K, where the average number of runs start to exceed one, as shown in Fig. 4(c). At this point the first site is fully melted and a second site located near the promoter region of the -lactamase gene also starts to melt.
Next, we compare the transition profiles calculated by BDZtrans for each type of transition with those calculated for denaturation alone by SIDD and for B-Z transitions alone by SIBZ. These three profiles are calculated at superhelix density and K, and are shown in Fig. 5. Comparison of these profiles shows that the sites that transform when the two transitions are competing coincide with the dominant sites that are predicted when each transition is treated alone. However, the probabilities of transition at these sites are significantly smaller when the two types of transition are allowed to compete. The probability of the dominant melting region drops from when calculated by SIDD to when calculated by BDZtrans. Similarly, the transition probability of the dominant Z-forming region changes from when only the B-Z transition is allowed, to when both conformations compete. One sees that calculations in which only one alternate conformation is considered tend to overstate the transition probabilities relative to a more realistic analysis in which multiple transition types compete together.
Table 1 shows sample numerical results and technical information regarding these computations. The execution times reported here are for calculations performed on a MacBook Pro with dual Intel processors. The SIBZ algorithm is slowest, as the B-Z transition occurs at multiple runs () under these conditions. In consequence, SIBZ also includes the largest number of states. SIDD is fastest because melting occurs predominantly in single run states, so fewer states satisfy the threshold condition for this transition. When the full competition is analyzed using BDZtrans, the average number of runs of denaturation and of Z-formation are both smaller, so there are an average of 1.8 runs in this case. The total number of transformed base pairs also decreases for each transition. The execution time of BDZtrans is intermediate between those of SIDD and of SIBZ, as is the number of states it includes.
These calculations show that the results from BDZtrans are qualitatively consistent with those from SIDD and SIBZ in that the competing transitions are largely limited to sites that dominate when each transition is considered alone. However, the probabilities of transition found by BDZtrans are not related in any simple way to those found by the other two algorithms. In particular, they are not weighted averages of the probabilities found by SIDD and SIBZ, and there is no direct way in which one could estimate the competitive behavior of these two transitions from the profiles found using the independent analyses. Rather, this behavior is determined by complex, globally coupled, non-linear interactions, and can only be assessed by a full analysis that simultaneously considers both types of competing transitions.
Regulation of the c-myc oncogene has been intensively studied, in part because mutations involving this gene have been implicated in various cancers. Substantial evidence has been found suggesting that superhelical DNA structural transitions play roles in regulating c-myc. Its 5′ flank contains a SIDD site called FUSE, located 2 kb upstream from the promoters, three AluI fragments containing Z-forming sites, and a potentially either G-quadriplex forming or H-forming site called the CT element around 1 kb upstream from the promoters [14]. Experiments have shown that each of these sites can be driven into their alternate structure by the superhelicity that is induced by transcription [13], [50], [51], [67], [68]. Much is known about the mechanism by which superhelical denaturation of the FUSE element regulates transcription through binding of single strand-specific regulatory proteins. Less is known about the roles of the other two alternate structures, and nothing is known about the competition among them or how it modulates transcription [67].
Here we use BDZtrans to investigate how the strand separation and B-Z transitions compete in this region. Specifically, we analyze the transition behavior of a 5 kb region around the c-myc gene that includes the FUSE element and the three Z-susceptible sites. Fig. 6(a) shows the transition profile of this region calculated using BDZtrans at T = 310 K and = −0.06. The upper panel in the figure marks the locations of the FUSE element and the three AluI Z-sites, labeled Z1, Z2, and Z3. The locations of the promoters also are shown. Under these conditions one sees a clear melting peak at the FUSE element. For Z-DNA there are two small peaks at Z2 and Z3, and only an insignificant transition probability at Z1.
Fig. 6(b) shows the overall probabilities and of each type of transition as a function of the superhelical density at T = 310 K. At low levels of negative superhelicity only the B-Z transition is present. As becomes more negative the probability of melting also increases, out-competing Z-DNA in some ranges. At −0.06, both transitions have high probabilities of occurrence. However, as shown in Fig. 6(a) denaturation is substantially confined to the FUSE element, while the propensity to form Z-DNA is distributed among several regions, each of which has only a low transition probability. One sees that Z-formation is predicted to occur at lower superhelical densities than does FUSE melting. This suggests that the presence of the Z-forming regions will delay the onset of FUSE melting to more extreme superhelicities than would be required in their absence. Moreover, when FUSE melting occurs, it is facilitated by the partial reversion of the Z-DNA back to B-form. In this way the B-Z transitions are predicted to have regulatory effects through their modulation of FUSE melting.
Transcription in eukaryotes has been shown to produce enough negative superhelicity to drive structural transitions in regions upstream from active genes [4]. This suggests that superhelically driven transitions to alternate DNA structures could occur there in vivo, where they might serve transcriptional regulatory functions. SIBZ analysis has found an enrichment of regions with Z-forming potential around transcription start sites (TSSs) [31]. Other less rigorous methods, such as Z-Catcher and Z-Hunt, have found qualitatively similar patterns of Z-DNA enrichment around TSSs [31], [55], [69], [70], although they find different numbers of sites than we do. Here we examine how superhelical B-Z transitions compete with denaturation at these locations, as well as in the regions where transcription terminates. We compare the transition properties of the TSS regions in eukaryotes with those in a prokaryote, and with those from a class of pseudogenes that are not transcribed.
This paper develops the first computational method to analyze the statistical mechanical equilibrium behavior of a negatively supercoiled DNA with any base sequence that is subject to multiple, competing secondary structural transitions. This method calculates the probability of each base pair transforming into any of its available alternative secondary structures, as well as the ensemble average values of other parameters of interest. The analysis of multi-state competitions is required when sites susceptible to each type of transition are present in the same topological domain, as occurs in virtually all domains of kilobase length.
The first implementation of this method is the BDZtrans algorithm, which analyzes the competition between superhelical strand separation and B-Z transitions. This competition was chosen for two reasons. First, complete information is available regarding the energetics of both of these transitions under comparable environmental conditions. Second, every base pair in DNA is susceptible to forming either alternate structure, at least in principle. So these two transitions compete in every DNA sequence. Using this algorithm we show that the competition between these transitions in superhelical DNA is highly intricate, depending in complex ways on base sequence, superhelix density, and temperature. Due to the temperature dependence of the energetics of strand separation, B-Z transitions dominate at low temperatures and denaturation becomes increasingly competitive as temperature increases. In the physiologically important temperature range 300–315 K, both types of transitions are reasonably competitive. Their interactions also depend in complex ways on the sequences and lengths of the transforming regions, and on the superhelix density. In an illustrative sample calculation we documented conditions in which B-Z transitions are preferred over denaturation at high superhelix densities, even when the temperature is above the melting temperature of A+T-rich DNA.
To determine how strand separation and B-Z transitions interact in practice in superhelical domains, we used BDZtrans to analyze 12,841 mouse gene sequences at = 305 K and superhelix density = −0.06. For each sequence in this set we assessed its equilibrium distribution, then determined the fraction of conformations in that distribution that had specific properties of interest. First, for every sequence in this set the probability of having no transition was essentially zero; virtually every conformation in the equilibrium distribution of every sequence was found to undergo some sort of transition under these conditions. Next, for each sequence we determined the frequency in its equilibrium distribution of conformations in which both denatured and Z-form sites were simultaneously present. We found that approximately half of these sequences have equilibrium distributions in which more than 10% of the molecules have coexisting Z-form and denatured regions. In 30% of the sequences these states dominate the equilibrium distribution. That is, more than half the molecules in the equilibrium distribution contain both Z-form and denatured regions. This shows the prevalence of states involving all three conformations in superhelically stressed genomic sequences, and indicates the importance of using computational methods that analyze their interactions.
We have shown that one cannot develop an accurate analysis of multistate transitions by amalgamating results from two-state techniques. To this end we compared the results from BDZtrans with those from SIDD and SIBZ, two-state algorithms that treat strand separation and B-Z transitions, respectively. Although the dominant transition regions are often correctly identified by the individual algorithms, they substantially overestimate both the number of such regions and their relative propensities to experience transition. This happens because each transition type in fact competes with the other, transitions to which decrease the effective level of supercoiling. A variety of examples have shown that sequences susceptible to both types of transition can exhibit particularly complex behaviors that cannot be captured by combining the results from the two-state SIDD or SIBZ analyses. In essence, this is because one cannot get an accurate depiction of an equilibrium distribution that contains many conformations in which denatured and Z-form sites coexist by mixing one distribution in which only denatured states occur with a second distribution in which only Z-forming states are present. This is why a full multi-state analysis is required to accurately depict competitions involving multiple alternate conformations in superhelical DNA.
Comparisons of the BDZtrans results with those from experiments investigating the superhelical competition between strand separation and B-Z transitions shows that, when the correct energetics are used, the BDZtrans algorithm accurately depicts the competitive transition behavior that was observed in these experiments [26].
We performed the first theoretical analysis of the competition between superhelical denaturation and B-Z transitions in the control regions of the c-myc oncogene, where both transitions are known to occur in vivo, and have been posited to serve regulatory functions [67]. Our results suggest that B-Z transitions near the c-myc promoters could modulate the known regulatory effects of strand separation at the upstream FUSE element. When the energetics of formation of the quadriplex that also can occur in this region become available, we will model the full three-way competitive interactions that can occur among these transitions in a quantitatively precise manner. We anticipate that this approach will illuminate the competitions among these three transitions, and thereby assist scientists to design experiments that assess their regulatory interactions.
We used BDZtrans to analyze the competitive transition behaviors of collections of mouse and human gene sequences. Each sequence was 5000 bp long, aligned and centered on their annotated transcription start site (TSS). We found a sharp increase of Z-forming sites that peaks just before the TSS, then continues a short distance into the transcribed region. This apparent enrichment suggests that B-Z transitions might be involved in the transcriptional regulation of some genes. Interestingly, the BDZtrans analysis of these mammalian gene sets also found that sites susceptible to superhelical denaturation are highly depleted over a broad region extending approximately one kilobase on either side of the TSS. The similarities of the patterns found for both transitions in human and mouse sequences suggests that these may be universal properties of mammalian genomes, and may also occur in other eukaryotes. This question will be investigated in future work.
The depletion of stress-denaturable sites around mammalian TSSs may seem surprising, as strand separation is an essential step in the initiation of transcription from every gene. However, this process is stringently regulated by interactions between the DNA and a large number of other molecules. It is possible that the occurrence of superhelically denatured sites in 5′ gene flanks could disrupt this regulation in some manner. It has been shown that transcription can be initiated by the presence of single stranded regions of DNA alone, without requiring any other regulatory factors [76]. So if a site susceptible to superhelical denaturation occurred within the first kilobase 5′ of a gene, where its transcription would produce enough negative superhelicity to drive denaturation, the resulting open region could initiate unintended additional rounds of transcription. If this were a deleterious event, sites susceptible to superhelical strand opening would be expected to be disfavored near TSSs. We note, however, that superhelical destabilizations at more remote positions are known to serve specific regulatory functions. The FUSE element that is located 2 kb upstream from the major c-myc promoters regulates transcription of this gene in humans by processes involving superhelical destabilization [36].
The situation may be expected to be rather different in prokaryotes. These organisms are highly gene dense, so the intergenic transcriptional regulatory machinery must be positioned close to the genes or operons they control. Also, superhelicity is not transient in prokaryotes, but is maintained within domains by enzymatic as well as by transcriptional processes. The superhelix density in E. coli changes with environmental conditions and growth state, and is coupled directly to the expression levels of genes that are differentially expressed under these conditions. Our results suggest that superhelical denaturation would be highly competitive with B-Z transitions at temperatures characteristic of growth phase in a host, while B-Z transitions would dominate at the lower temperatures that occur outside of a host. However, the level of negative superhelicity imposed on the genome by gyrase also is higher during growth phase than in stationary phase. So a careful analysis of this situation requires that both effects be included. This matter also will be investigated in future work.
BDZtrans analysis of a prokayotic gene set finds the opposite transition behavior in their 5′ flanks than was found for eukaryotes. A clear enrichment of denatured sites just upstream of the TSS (i.e. the +1 position in prokaryotic genes) has been found at = 310 K in E. coli, as well as in many other prokaryotic genomes that have been analyzed previously with the SIDD algorithm. Interestingly, at the temperature where the probabilities of strand separation and of B-Z transition are comparable in eukaryotes, in E. coli one finds that Z-DNA dominates away from the +1 gene positions. This result suggests that fundamental differences may exist in the process of transcription as it occurs in prokaryotes and in eukaryotes.
When we compared the competitive transition behaviors around transcription start (TSS) and gene stop (TES) sites in humans, we found that opposite patterns prevail in the two regions. At the TSS the number of Z-susceptible sites is increased and the number of denaturation-susceptible sites decreased, relative to more distant regions. The opposite pattern occurs in the 3′ regions proximal to TESs. In these locations there is a clear and substantial enrichment of denaturing states, and a slight diminishing of Z-susceptible sites. This suggests that denatured DNA might play some role in processes occurring near gene 3′ flanks.
The transition properties of transcribed mouse genes have been compared to those of a set of processed pseudogenes that do not transcribe. The results obtained by BDZtrans show no apparent pattern for either transition upstream of the pseudogene start sites. However, there is a substantial decrease in the number of sites susceptible to either type of transition just downstream of the pseudogene “start” positions. This result suggests that the maintenance of Z-susceptible sites just 3′ of the TSS in transcribed mouse genes may be under selection pressure, disappearing when that pressure is removed.
To illustrate the practical utility of the BDZtrans algorithm, suppose that superhelical transition at a specific site is hypothesized to serves some regulatory function. To establish this hypothesis one must first show that the superhelical transition actually occurs at the site, and then prove that it exerts a regulatory effect. These questions are frequently investigated by inserting a segment containing the putatively regulatory susceptible site into a plasmid, perhaps along with a reporter gene. However, if superhelical transition at the test site is found not to occur in the plasmid it could be either because the hypothesis is false, or because in the plasmid that site competes with different alternatives than it does in its genomic context. Conversely, just because the transition occurs in the plasmid does not automatically mean that it also will occur in its genomic context. One can only draw inferences from the plasmid behavior regarding the genomic activity if the behaviors of the test site in the two contexts are comparable. The theoretical methods developed in this paper enable investigator to assess how the transition behavior of a site would be expected differ when it is placed in different contexts. Use of these methods will enable experimenters to design plasmids that most accurately address their questions.
After transition at the test site has been shown to occur in the plasmid, it remains to establish that it is the superhelical transition itself serves the regulatory function, and not some other attribute of the site. To do this one must alter the transition properties of the test site without changing its other attributes - in particular the local base sequence of the region involved. One can insert at a remote position on the plasmid a different susceptible site that is designed to outcompete the transition at the test site. If transition at the insert site happens first, it will delay the transition at the test site to more extreme superhelicities. If the regulatory effect is delayed to the same degree, this is strong evidence that it is indeed the superhelical transition that exerts the regulatory effect. To design experiments of this sort one needs a way to assess how various inserts would compete with a given test site within a given plasmid. The multistate analytical methods developed here will enable experimenters to make these assessments. Since B-Z transitions relax the most superhelicity per base pair, under most conditions they transform at less extreme superhelicities than do other types of transitions. So the natural choice for a competitive insert would be a Z-susceptible site. If the transition whose putative regulatory properties are being tested is either denaturation or another BZ transition, then use of the BDZtrans method presented in this paper will enable experimenters to design the correct systems to rigorously test their hypotheses.
These examples show how the techniques presented in this paper can be of immediate use to experimenters. Our precise quantitative method has the potential to enable the design of much more accurate and rigorous experiments than would otherwise be possible.
The multistate methods developed here are capable of treating competitions involving all the possible secondary structures that can be driven by supercoiling. In addition to the Z-form and denatured conformations, this could include G-quadriplexes, H-form DNA, cruciforms, and possibly others. However, in order to make quantitative predictions of the superhelical competitive behavior of sequences containing sites that can form these structures, their transition energetics must be known under the assumed environmental conditions. This limits the present applicability of our method to treating competitions involving B-Z transitions and denaturation, as was done here. Information is available regarding the energetics of superhelical cruciform extrusion at perfect inverted repeat sequences, and the energy costs of some types of imperfections are known [65]. So analyses that include extrusion of cruciforms are being developed. Unfortunately, sufficiently complete information regarding the energetics of forming general G-quadriplexes and H-form triplexes is not available at present. The approach presented here will become applicable to more situations as our understanding of transition energetics improves. In particular, information regarding the energetics of formation of the quadriplex at the CT site in the c-myc 5′ flank is expected to be available soon.
A website is available (http://benham.genomecenter.ucdavis.edu) where members of the scientific community may submit sequences of interest to them for analysis by the BDZtrans algorithm. The sequence must be either in FASTA format or in a file that contains sequence characters exclusively. Sequences of any length up to 10 kb may be submitted, although sequences of length around 5 kb are preferred. This site may also be used for SIDD and/or SIBZ analyses.
|
10.1371/journal.pgen.1002139 | Maps of Open Chromatin Guide the Functional Follow-Up of Genome-Wide Association Signals: Application to Hematological Traits | Turning genetic discoveries identified in genome-wide association (GWA) studies into biological mechanisms is an important challenge in human genetics. Many GWA signals map outside exons, suggesting that the associated variants may lie within regulatory regions. We applied the formaldehyde-assisted isolation of regulatory elements (FAIRE) method in a megakaryocytic and an erythroblastoid cell line to map active regulatory elements at known loci associated with hematological quantitative traits, coronary artery disease, and myocardial infarction. We showed that the two cell types exhibit distinct patterns of open chromatin and that cell-specific open chromatin can guide the finding of functional variants. We identified an open chromatin region at chromosome 7q22.3 in megakaryocytes but not erythroblasts, which harbors the common non-coding sequence variant rs342293 known to be associated with platelet volume and function. Resequencing of this open chromatin region in 643 individuals provided strong evidence that rs342293 is the only putative causative variant in this region. We demonstrated that the C- and G-alleles differentially bind the transcription factor EVI1 affecting PIK3CG gene expression in platelets and macrophages. A protein–protein interaction network including up- and down-regulated genes in Pik3cg knockout mice indicated that PIK3CG is associated with gene pathways with an established role in platelet membrane biogenesis and thrombus formation. Thus, rs342293 is the functional common variant at this locus; to the best of our knowledge this is the first such variant to be elucidated among the known platelet quantitative trait loci (QTLs). Our data suggested a molecular mechanism by which a non-coding GWA index SNP modulates platelet phenotype.
| Genome-wide scans have revealed multiple genetic regions underlying complex traits. However, the transition from an initial association signal to identifying the functional DNA change(s) has proved challenging. Many of the DNA changes discovered are located outside protein-coding regions and may exert their effects through gene regulation. We screened genetic regions associated with hematological traits in erythroblasts (red blood cells) and megakaryocytes (platelet-producing cells) and mapped sites of open chromatin, which harbor active gene regulatory elements. We investigated a DNA sequence change located within a site of open chromatin at chromosome 7 in megakaryocytes, but not erythroblasts, known to be associated with platelet volume. We showed that this DNA change is functional due to alteration of the binding site of a transcription factor, which regulates the expression of a gene that affects platelet characteristics. Mice lacking this gene revealed significant differences in expression of several important platelet genes compared to wild-type mice. The approach described here can be applied in different cell types to functionally follow-up association signals with many other biological traits by identification of the causative base change and how it affects gene function, thus paving the way to clinical benefit.
| In recent years, genome-wide association (GWA) studies have driven the discovery of genetic loci associated with a multitude of complex traits and diseases. However, due to linkage disequilibrium (LD), the single-nucleotide polymorphisms (SNPs) assayed in large-scale GWA studies typically yield a proxy for the actual causative variant(s) and often fail to even pinpoint the underlying gene [1]–[3]. Therefore, despite great success in biological discovery, GWA studies have shed little light on which are the functional variants, common and/or low-frequency, and how they relate to biological mechanisms.
Several strategies have been proposed to identify causative sequence variants underlying GWA signals, including fine-mapping (i.e. targeted resequencing followed by additional genotyping of large cohorts) and functional annotation using biochemical assays (e.g. gene expression studies) [1], [3]. Since many initial association signals are localized at intronic and intergenic regions, it has been postulated that they may correspond to gene regulatory variants [4]. Therefore, functional variants may exert their effects through regulation of gene expression, which can vary substantially among tissues and cell types [5], [6].
Integrated studies have illustrated how particular trait-associated alleles are mechanistically relevant to cellular functions and pathways [7]–[12]. In a recent study, Gaulton et al. used the formaldehyde-assisted isolation of regulatory elements (FAIRE) technique to identify pancreatic islet-specific open chromatin regions, some of which embed known diabetes risk variants [13]. FAIRE is an elegant approach to isolating nucleosome-depleted regions (NDRs) that encompass active regulatory elements, and thus, to accessing regulatory variants [14], [15].
Heritable hematological traits are of particular interest in studying the architecture of complex traits, since relevant cell types for functional assays are easily accessible. In addition, genetic loci associated with platelet counts, SH2B3–ATXN2 and PTPN11, have been reported to be associated with coronary artery disease (CAD) and myocardial infarction (MI) suggesting a possible role for platelets as an intermediate phenotype [16].
Here we describe maps of open chromatin generated by FAIRE in a megakaryocytic and an erythroblastoid cell line at known GWA loci associated with hematological quantitative traits, CAD and MI. We report substantial differences in chromatin architecture among the two cell types and assess the phenotypic causality of variants that fall into NDRs specific to either cell type. We demonstrate the procedure with the locus at chromosome 7q22.3, which is associated with mean platelet volume and function, where we identify a site of open chromatin in megakaryocytes but not erythroblasts and elucidate the molecular mechanism that contributes to the platelet phenotype.
We profiled chromatin at 62 non-redundant genetic loci representing all known associations, as of November 2009, with 11 cardiovascular traits (Table 1) in a megakaryocytic (MK cells) and an erythroblastoid cell line (EB cells). Maps of open chromatin were created with FAIRE, applying two different formaldehyde cross-linking times and subsequent hybridization to a custom 385,000-oligonucleotide array spanning the selected loci (Materials and Methods). Open chromatin regions showed high concordance across cross-linking conditions, with 95.3% and 93.1% overlapping regions in MK and EB cells, respectively. To achieve higher stringency, we retained only concordant open chromatin regions in each cell type for further analysis (Figure S1). At the interrogated loci, we identified 254 and 251 NDRs in MK and EB cells, respectively (Table S1A), of which 147 (57.9% and 58.6%, respectively) were common to both cell types. FAIRE peak density (per megabase) in our data set is consistent with that in foreskin fibroblast cells reported by the ENCODE Project [17] (Table S2).
We then analyzed the 254 and 251 NDRs in MK and EB cells, respectively, in context of their genomic location (intergenic, intronic, overlap 5′-untranslated region (UTR), overlap 3′-UTR or exonic; Table S3). We note that our observations are based on a selected set of loci and therefore cannot be extrapolated to the whole genome. NDRs were most frequently located at non-coding segments (98.2% and 92.5% of peaks found only in MK and EB cells, respectively, and 98.1% of peaks common to both cell types). Promoter/5′-UTR regions were enriched in NDRs common to both cell types (28.4%) compared to open chromatin specific to either cell type, namely 4.6% (6.2-fold) and 5.6% (5.1-fold) for MK and EB cells, respectively.
At the interrogated loci, NDRs clustered around transcription start sites (TSS). In MK and EB cells, 70.5% and 76.1% of all FAIRE peaks were located within 20 kb of a TSS, respectively (Figure S2). However, accessible chromatin regions as far as 264 kb upstream of a TSS were also detected (TBX3 gene locus). Open chromatin observed in MK but not EB cells was located on average 2.80 kb upstream of a TSS. NDRs found in EB but not MK cells were located on average 1.77 kb, whereas NDRs common to both cell types were located on average 0.98 kb upstream of a TSS.
To assess the cell type specificity of NDRs marked by FAIRE in our experiment, we determined the number of peaks in lineage-specific genes for MK and EB cells present on the array (Figure S3; Table S1B). A significant enrichment of FAIRE peaks at MK lineage-specific genes was observed in MK cells, when compared to the number of peaks in EB cells (Wilcoxon rank-sum test, P=0.0225). A similar trend of enrichment was observed in EB lineage-specific genes in EB cells (P=0.0781). This result highlights the importance of studying chromatin architecture and gene regulatory circuits in a cell type-dependent manner.
At seven of the 62 tested loci, we found SNPs in strong LD with the corresponding GWA index SNP (r2≥0.8, Phase II HapMap, CEU population) located within a NDR (Table 2). Five out of the seven loci are associated with platelet-related quantitative trait loci (QTLs). We compared the position of sites of open chromatin and sites were either found only in MK but not EB cells (‘MK-specific’, n=2), in EB but not MK cells (‘EB-specific’, n=1) or in both cell types (n=4). The two MK-specific NDRs harboring SNPs associated with mean platelet volume (MPV) are located at an intergenic region of the FLJ36031–PIK3CG gene locus (Figure S4A) and an intronic region of DNM3 (Figure S4B). Both genes, PIK3CG and DNM3, are upregulated in megakaryocytes as compared to erythroblasts (2.08- and 6.42-fold, respectively, according to the HaemAtlas, a systematic analysis of expression profiles in differentiated human blood cells) [18]. The EB-specific NDR is located at an intergenic region of the HBS1L–MYB gene cluster (Figure S4C). Sequence variants at this locus are known to be associated with mean corpuscular volume (MCV) of erythrocytes, mean corpuscular hemoglobin (MCH) and red blood cell counts (RBC). HBS1L and MYB are upregulated in EB cells (1.30- and 2.40-fold, respectively, according to the HaemAtlas). At the four NDRs common to both cell types, we found variants associated with: platelet signaling (PLS) located at the promoter regions of PEAR1 (Figure S4D) and RAF1 (Figure S4E); MPV found at an intronic region of TMCC2 (Figure S4F); and systolic blood pressure (SBP) at an intronic region of C10orf32 (CYP17A1 gene cluster; Figure S4G). Expression profiles of these genes, based on the HaemAtlas data, confirmed transcription in both MK and EB cells (Figure S5A).
Cell-specific open chromatin regions are likely to play a regulatory role in modulating gene expression in a cell-specific manner. Sequence variants in such NDRs have the potential to impact cell-specific traits, for example MPV. As a proof of principle, we investigated the 65-kb locus associated with mean platelet volume and function at chromosome 7q22.3 (Figure 1A), which harbors a MK-specific NDR as described above. The recombination interval exhibited a total of six distinct NDRs (Figure 1B), two of which are common to both cell types located at an evolutionary conserved element 10 kb upstream of the GWA index SNP rs342293 and at the promoter region of FLJ36031, three of which are specific to EB cells and one of which is specific to MK cells. The latter NDR contained SNPs associated with MPV: rs342293 (MAF=0.48, Phase II+III HapMap, CEU) and its best proxy rs342294 (r2=1.0, MAF=0.48, Phase II+III HapMap, CEU). Data from the 1000 Genomes Project [19] (Pilot 1, CEU) revealed 34 SNPs in LD (r2≥0.8) with rs342293, and confirmed that only rs342293 and rs342294 fall into the MK-specific NDR. This NDR was absent in both FAIRE-chip data sets (8 and 12 min formaldehyde cross-linking conditions) in EB cells (Figure 1C). Irrespective of LD to rs342293, there are no sequence variants reported by the 1000 Genomes Project (Pilot 1, CEU) within sites of open chromatin at the recombination interval in MK cells.
Next, we performed an in silico analysis of transcription factor binding sites at the 7q22.3 locus (Table S4). Of the 34 SNPs in LD with rs342293, four (rs342240, rs342247, rs342292 and rs342293) disrupt an in silico predicted transcription factor binding site. However, only rs342293 is located within an experimentally verified site of open chromatin in MK cells. Among these four SNPs, rs342293 is the most strongly associated with mean platelet volume [16].
The SNP rs342293 is located within overlapping predicted binding sites for the transcription factors BARX2, EVI1, GATA1, HHEX, HOXC8, HOXC9 and LBX2 (Figure 1D). Based on the HaemAtlas data, of these seven transcription factors only EVI1, GATA1 and HHEX are expressed in megakaryocytes (Figure S5B). The C>G substitution leads to disruption of the predicted binding sites of EVI1 and GATA1 (Figure 2A). It is worth noting that in silico analysis predicted a RUNX1 transcription factor binding site only 5 bp apart from the EVI1-like binding site.
To obtain the full spectrum of sequence variation at this region, we sequenced the NDR (chr7:106,159,393–106,159,887; 494 bp) in 643 healthy individuals. No other common or low-frequency variants were detected, although a rare, possibly private, SNP was detected in one individual as a heterozygous position (chr7:106,159,601; A>C) located in a polyA-region (Table S5). Based on the above evidence and the absence of any additional common or low-frequency sequence variant at the MK-specific NDR, rs342293 remains as the only likely putative functional candidate underlying the MPV association signal at the 7q22.3 locus. However, it cannot be excluded that additional functional variants may exist outside NDRs and exert their function through, for example disruption of a yet to be identified transcription factor binding site or modulation of DNA methylation.
We then performed electrophoretic mobility shift assays (EMSA) in nuclear extracts from the megakaryocytic cell line CHRF-288-11. We observed a band shift with probes surrounding rs342293 for both the ancestral allele rs342293-C and the alternative allele rs342293-G (Figure 2B). However, the bands were unequally shifted suggesting differential protein binding properties at this position depending on the allele of rs342293. The specific unlabeled competitors supported specificity of the retarded bands. Further, our results suggested superior protein binding to the probe containing the C-allele. With supershift experiments using an EVI1 antibody, we confirmed binding of EVI1 to probes containing the ancestral C-allele, but not to those harboring the alternative G-allele. Supershift experiments with a GATA1 antibody did not support in vitro binding of GATA1 transcription factors to this site. We also confirmed RUNX1 transcription factor binding in vitro by demonstrating a supershift for both probes with a RUNX1 antibody (Figure S6). We validated these findings by performing chromatin immunoprecipitation combined with next-generation DNA sequencing (ChIP-seq) with GATA1 and RUNX1 antibodies in primary human megakaryocytes. We observed no significant GATA1 but weak RUNX1 binding at this locus further corroborating the EMSA results (Figure S7).
Soranzo et al. investigated the association of rs342293 with transcript levels of all known genes within 1 Mb of the GWA index SNP (Figure S5C) in platelets and reported a weak expression QTL (eQTL) association with PIK3CG transcript levels (permutation P=0.047) [20]. We replicated this finding in an independent sample cohort of 24 healthy individuals showing the same genotypic effect (P=0.0542; Figure 3A). In both sample cohorts, none of the other genes within the 1-Mb interval, PRKAR2B, HBP1, PBEF1 and COG5, had a statistically significant eQTL with rs342293 (P>0.1; data not shown). Next, we assessed the PIK3CG-eQTL in three different types of white blood cells, macrophages, monocytes and B cells (lymphoblastoid cell line, LCLs), as well as in different tissues, fat and skin (Table S6). We observed a genotypic effect on PIK3CG transcript abundance in macrophages (P=0.0018; Figure 3B), but not in monocytes and LCLs. Further, we did not observe an association in fat and skin tissues. Our data strengthened the evidence of rs342293 modulating PIK3CG transcript levels in platelets and showed that this eQTL is also present in macrophages.
To further our understanding of the role of PIK3CG in platelets, we performed whole-genome gene expression profiling in whole blood of Pik3cg knockout mice. We identified 220 differentially expressed genes between knockout (n=3) and wild type mice (n=3) with a fold-change of at least ±1.5 (Table S7). Functional ontology classification of these genes (Table S8) revealed enrichment for ‘regulation of biological characteristic’, e.g. cell size and volume (GO term: 0065008; P=2.97×10−14), and ‘blood coagulation’ (GO term: 0007596; P=7.87×10−12). Notably, this gene list includes Gp1bb (fold-change of −2.203 in Pik3cg−/− compared to wild type mice), Gp5 (−3.211), Gp9 (−0.996), Gp6 (−0.711) and Vwf (−1.381). All five genes are transcribed in the megakaryocytic lineage in humans, based on the HaemAtlas data. The platelet glycoproteins GP1BB, GP5 and GP9, together with GP1BA constitute the platelet membrane receptor for the plasma protein Von Willebrand Factor (VWF), which is encoded by VWF [21].
To explore the signaling pathways of PIK3CG in humans, we analyzed 191 human orthologs of the 220 differentially expressed genes between Pik3cg−/− and wild type mice. Canonical pathway enrichment analysis based on the curated gene sets of the Molecular Signatures Database (MSigDB) v3.6 [22] revealed that the top six enriched gene sets were related to platelets: ‘platelet degranulation’ (P=3.44×10−8), ‘platelet activation’ (P=5.05×10−8), ‘formation of platelet plug’ (P=2.85×10−7), ‘hemostasis’ (P=7.48×10−6), ‘formation of fibrin clot and clotting cascade’ (P=1.39×10−5) and ‘platelet adhesion to exposed collagen’ (P=1.86×10−5).
We constructed a protein-protein interaction network centered on the proteins encoded by the 191 transcripts described above. First-order interactors of these ‘core’ proteins were obtained from Reactome, an open-source manually curated database of human biological pathways [23], [24]. We filtered interactors on their expression levels in MK cells (Materials and Methods). The resulting network incorporated 45 core proteins centered on PIK3CG consisting of 642 nodes and 1067 edges (Figure 4).
The intersection of maps of open chromatin with variants identified through GWA studies can facilitate the search for underlying functional variant(s). We applied the FAIRE assay to generate a catalog of NDRs in a megakaryocytic and an erythroblastoid cell line at 62 selected genetic loci associated with hematological and cardiovascular traits. We provided evidence that open chromatin profiles exhibit distinct patterns among different cell types and that cell-specific NDRs can be useful in prioritizing regions for further functional analysis. As proof, we elucidated the molecular basis of the association with mean platelet volume and function at chromosome 7q22.3. We identified a NDR in MK but not EB cells containing the index SNP rs342293 for this association and demonstrated that the alleles of rs342293 differentially bind the transcription factor EVI1. Thus, to the best of our knowledge this is the first functional variant to be elucidated among the known platelet QTLs.
Expression QTL data in platelets and macrophages provided statistical support that rs342293 affects PIK3CG gene expression levels. PIK3CG is transcribed in megakaryocytes but only weakly expressed in erythroblasts [18], which is in agreement with the MK-specific properties of the identified NDR. However, additional work is required to scrutinize all possible targets of this regulatory module. The closest gene FLJ36031, which spans only 2 kb, has no reported protein product and function and requires further characterization. PIK3CG is located 134 kb downstream of rs342293 and encodes the phosphoinositide-3-kinase γ-catalytic subunit. The lipid kinase PIK3CG (PI3Kγ) is a member of the class I PI3Ks and catalyzes the conversion of phosphatidylinositol-4,5-bisphosphate (PtdIns(4,5)P2; PIP2) to phosphatidylinositol-3,4,5-trisphosphate (PtdIns(3,4,5)P3; PIP3) downstream of cell surface receptor activation [25]–[27]. In megakaryocytes and platelets, PIP3 is crucial in the collagen-induced regulation of phospholipase C and initiation of megakaryopoiesis and proplatelet formation [28]. Functional studies with Pik3cg knockout mice have indicated a role in wound healing [29], ADP-induced platelet aggregation and thrombosis [30], [31]. It is noteworthy that PIK3CG has also a prominent role in macrophage activation [27].
Whole-genome gene expression profiling in Pik3cg−/− mice showed differential expression of genes involved in important platelet-related pathways compared to wild type mice, most notably Vwf and its platelet membrane receptor components. Our recent meta-analysis of GWA studies in 68,000 Northern Europeans showed that common sequence variation at the VWF and GP1BA loci exert an effect on platelet volume and counts, respectively (NS, unpublished data). This recent finding together with the established knowledge that Mendelian mutations in the genes encoding the platelet VWF (Von Willebrand disease, type 2; OMIM: 613554) and its receptor (Bernard-Soulier syndrome; OMIM: 231200) are causative of giant platelets, give biological credence to the observed effects of Pik3cg knockout on the transcription of these platelet genes. The protein-protein interaction network centered on PIK3CG highlighted additional proteins implicated in severe platelet disorders, including TUBB1 (macrothrombocytopenia, autosomal dominant, TUBB1-related; OMIM: 613112), F5 (factor V deficiency; OMIM: 612309) and P2RY12 (bleeding disorder due to P2RY12 defect; OMIM: 609821). Therefore, it is plausible to assume that differences in the PIK3CG transcript levels in humans, based on the different alleles of rs342293, may lead to changes in the abundance of platelet membrane proteins that are key regulators of platelet formation.
We previously showed an association of rs342293-G with decreased platelet reactivity in humans, assessed as the proportion of binding to annexin V and fibrinogen, as well as P-selectin expression, after activation of platelets with collagen-related peptide (CRP-XL) [20]. We also observed that rs342293 is likely to modify events downstream of signaling via the collagen signaling receptor glycoprotein VI, which is encoded by GP6 [32]. A recent GWA study showed the same locus at chromosome 7q22.3 to be associated also with epinephrine-induced platelet aggregation [33]. The index SNP in that study, rs342286, and rs342293 are in high LD (r2=0.87, Phase II HapMap, CEU) making rs342293 the putative causative variant underlying both functional associations. The observation that platelet functional events triggered via an immunoreceptor tyrosine-based activation motif on the Fc receptor γ-chain (ITAM-FcR-γ) or G-protein-coupled receptor, for collagen and epinephrine, respectively, are both modified by differences in the PIK3CG transcript level is compatible with the notion that both signaling cascades require PIP3 and that silencing of Pik3cg in mice reduces the expression of the Gp6 gene.
Based on the above data, we propose a model (Figure 5) in which the DNA sequence containing the C- but not the G-allele of rs342293 binds the transcription factor EVI1, which acts as transcriptional repressor of PIK3CG in megakaryocytes and ultimately affects platelet phenotype.
EVI1 (ecotropic viral integration site-1) encodes a protein with two zinc-finger domains (ZF1 and ZF2), which feature distinct DNA-binding specificities [34]. EVI1 mainly promotes hematopoietic differentiation into the megakaryocytic lineage. EVI1 was frequently reported to be a repressor of transcription that has the potential to recruit diverse regulatory proteins. For example, EVI1 antagonizes the growth-inhibitory effect of transforming growth factor-β (TGF-β), a potent regulator of megakaryopoiesis [35], by interacting with SMAD3 via ZF1, and inhibiting SMAD3 from binding to DNA [36], [37]. EVI1 contains domains that interact with RUNX1 (Runt-related transcription factor 1), which is the α-subunit of the transcription factor CBF (core binding factor). The interaction of EVI1 with the DNA-binding domain Runt of RUNX1 leads to the destabilization of the DNA-RUNX1 complex and subsequent loss of RUNX1 function [38].
By converting the read-out system from microarrays to high-throughput next-generation sequencing, genome-wide open chromatin profiles can be interrogated. This would scale up analysis to the whole genome generating a catalog of open chromatin profiles to annotate association loci identified in past and future GWA studies. Furthermore, resources such as the completed 1000 Genomes Project will make the requirement for resequencing any identified NDR redundant. However, access to cell types relevant to the studied trait is not always feasible and can be a limitation for functional studies. For instance, we did not observe any intersection of FAIRE peaks with variants associated with CAD/MI or hypertension indicating that cell types other than the megakaryocytic and erythroblastoid cell lines analyzed here may be more suitable.
In addition to rs342293, we identified six putative functional variants associated with hematological and cardiovascular traits that fall into open chromatin (Table 2). Follow-up functional studies of these variants will further test the applicability of the approach we have proposed here. For example, the MK-specific NDR containing rs2038479 was found to mark an alternative promoter of a dynamin 3 (DNM3) transcript with expression restricted to the megakaryocytic lineage (WHO, unpublished data). Correlation of signatures of open chromatin with experimentally determined transcription factor binding sites studied in different cell types can systematically and rapidly translate GWA signals into functional components and biological mechanisms, thus paving the way to clinical benefit.
All human subjects were recruited with appropriate informed consent in Cambridgeshire and enrolled in the Cambridge BioResource (http://www.cambridgebioresource.org.uk/). The study has ethical approval from the NHS Cambridgeshire Research Ethics Committee. The care and use of all mice in this study was carried out in accordance with the UK Home Office regulations under the Animals (Scientific Procedures) Act 1986.
We designed a 385,000-oligonucleotide tiling array (Roche NimbleGen) using 72 known genetic loci associated with the traits listed in Table 1 (National Human Genome Research Institute catalog of published GWA studies, http://www.genome.gov/gwastudies/). We considered only genetic loci that reached genome-wide significance with the threshold of P<5×10−8 (or as otherwise indicated in Table S9A) in a GWA study conducted with individuals of Northern and Western European ancestry (CEU population). In addition, we selected genetic loci based on biological evidence, where there was suggestive evidence of association. For each locus, we included the entire genetic region of the index SNP as defined by recombination hotspots based on Phase II HapMap [39]. If a recombination interval exceeded 500 kb, we included the closest target gene ±10 kb. In order to assess patterns of cell type specificity, we included eight lineage-specific reference genes for each of megakaryocytes, erythroblasts and monocytes on the array (Table S9B). We selected these transcripts on the basis of their expression profiles (Figure S5D) [18]. The oligonucleotide (50–75-mer probes) tiling array provided a mean probe span of 23 bp and harbored only probes unique to the human genome (build: hg18, coverage: 79%). In summary, a total of 62 unique complex trait loci and 24 reference gene loci representing 9.59 Mb and 1.77 Mb of genomic DNA, respectively, were selected for the array design.
We performed FAIRE in the human cell lines CHRF-288-11 and K562 on tiling arrays. Both cell lines exhibit markers characteristic for megakaryocytes/platelets [40], [41] and erythroblasts [42], [43], respectively. CHRF-288-11 cells were maintained in RPMI-1640 medium, supplemented with 20% horse serum (heat-inactivated) and 1% L-glutamine-penicillin-streptomycin solution. K562 cells were maintained in RPMI-1640 medium, supplemented with 10% fetal bovine serum (non-heat-inactivated), 2 mM GlutaMAX-I and 1% L-glutamine-penicillin-streptomycin solution. Cells were grown at 37°C, 5% CO2 and 100% humidified atmosphere. FAIRE experiments were carried out as previously described [44]. Briefly, for each cell type, 20×106 cells were cross-linked with 1% formaldehyde for 8 or 12 min. Chromatin was extracted and subjected to 12 sonication cycles (30 sec of high pulse, 30 sec of rest) using the Bioruptor UCD-200 (Diagenode). DNA fragments depleted of proteins (‘open chromatin’) were phenol-chloroform extracted and ethanol precipitated. The samples were treated with RNase A and cleaned-up using the MinElute PCR Purification Kit (QIAGEN). A DNA amplification step was not applied.
DNA extracted from cross-linked cells (FAIRE, input sample) and uncross-linked cells (reference sample) were labeled with Cy5 and Cy3 dye, respectively, and subsequently co-hybridized to the tiling array. Dual-color sample labeling, array hybridization, washing and scanning were performed using the dual-channel microarray platform by Roche NimbleGen according to the manufacturer's protocol (NimbleGen Arrays User's Guide, ChIP-chip Analysis v4.1).
Experimental data were analyzed using the software NimbleScan v2.5 (Roche NimbleGen). The two-channel raw signal intensities were scaled between channels by subtracting the Tukey bi-weight mean for the log2-ratio values for all features from each log2-ratio value. In the scaled log2-ratio data, peaks were identified using a sliding window approach [45]. For each chromosome, the log2-ratio cut-off value was calculated as the percentage of a hypothetical maximum (Pmax=arithmetic mean+6× standard deviation). The peak finding process was repeated using a series of log2-ratio cut-off values from Pstart to Pend. The following settings for the peak-finding analysis were applied: sliding window: 300 bp; min. probes>cut-off in peak=4; all probes in peak>cut-off=2; Pstart=90%, Pend=15%, Pstep=0.5, number of steps: 100. Log2-ratio and peak data sets were displayed as UCSC Genome Browser (http://genome.ucsc.edu/) custom tracks. For the genomic characterization of FAIRE peaks, we extracted all annotation from the Ensembl database v54 (build: hg18). Proxy-SNPs to reported GWA index SNPs were gathered using the Genome-wide Linkage Disequilibrium Repository and Search Engine (GLIDERS) [46], with the following settings: Phase II HapMap v23 (CEU); MAF limit≥0.05; r2 limit≥0.8; no distance limits. The FAIRE microarray data sets are available online in the Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo/) under accession number GSE25716.
DNA samples from a total of 643 individuals of Northern European ancestry were subjected to capillary DNA sequencing of the targeted locus at chromosome 7q22.3. Details of the sequencing protocol are described online (http://www.sanger.ac.uk/resources/downloads/human/exoseq.html). DNA was amplified by PCR and applied to bi-directional sequencing using Big Dye chemistry on 3730 DNA sequencers (Applied Biosystems). We used the following primer pairs (for two sequence-tagged sites): 5′-TGG AAA ATT ACA AAA GTC CCA AA, 5′-GAG AAA GGA TCA TGA GGG AGA A; and 5′-ACA AAA GTC CCA AAA TTT CAC A, 5′-GAG AAA GGA TCA TGA GGG AGA. The resulting products map to the following genomic location: chr7:106,159,328–106,159,998 (671 bp) and chr7:106,159,337–106,159,998 (662 bp). Pre-processed sequence traces were analyzed using the semi-automated analysis software ExoTrace, developed at the Wellcome Trust Sanger Institute (http://www.sanger.ac.uk/resources/downloads/human/exoseq.html). Potential SNP positions were indicated by the software and then reviewed manually.
In silico transcription factor binding sites were predicted using the software MatInspector v8.01 [47], with the following parameters: matrix group: vertebrates; core=1.00; matrix=optimized+0.02; tissue: hematopoietic system.
Nuclear extract was prepared from CHRF-288-11 cells with the NE-PER Nuclear and Cytoplasmic Extraction Reagents (Thermo Fisher Scientific). Oligonucleotides were designed based on the genomic sequence surrounding rs342293; the SNP position is shown in bold: 5′-biotin-AGC CCT GTG GTT TTA ATT ATC/G TTG AGG TTC AGG CTC A. Competitor probes were prepared without biotin tags. The labeled strands were annealed with the unlabeled complementary strands using a standard protocol. All oligonucleotides were provided by Sigma-Aldrich. We performed gel mobility shift assays with the LightShift Chemiluminescent EMSA Kit (Thermo Fisher Scientific) according to the manufacturer's protocol. Each 20-µl binding reaction contained 1× binding buffer, 75 ng/µl poly(dI/dC), 2.5% glycerol, 0.05% NP-40, 87.5 mM KCl and 6.25 mM MgCl2. For competition assays, we used 200-fold molar excess of the unlabeled probe. Reactions were incubated for 2 hr at room temperature. Supershift experiments were performed with EVI1 (sc-8707 X, Santa Cruz Biotechnology), GATA1 (ab28839, Abcam) and RUNX1 (sc-28679 X, Santa Cruz Biotechnology) antibodies.
Gene expression profiling and genotypic data sets were obtained from different sources depending on the studied cell type. RNA from all studied cell types was isolated using a TRIZOL standard protocol. Total RNA was quantified using the NanoDrop (Labtech International) and quality-checked using the 2100 Bioanalyzer (Agilent Technologies). Standard protocols were applied to generate biotinylated cRNA and hybridize to Illumina BeadChips (platelets: HumanWG-6 v2; macrophages/monocytes: HumanRef-8 v3; LCLs/fat/skin: HumanHT-12 v3). Then, arrays were washed and scanned. All experimental procedures were carried out according to the manufacturer's protocol. Even though different versions of Illumina platforms were used, the probe for PIK3CG is the same across all chips (probe-ID: ILMN_1770433). Genotyping was performed on DNA extracted from whole blood using TaqMan (Applied Biosystems) SNP genotyping assays (platelets study) and Illumina SNP arrays (macrophages/monocytes study: Human 1.2M-Custom and Human 670-Quad-Custom; LCLs/fat/skin study: Human 1M-Duo) following the manufacturer's instructions. Expression QTL analysis with rs342293 or its proxy-SNP was performed with the software Genevar (Gene Expression Variation) using a window of ±1 Mb centered on the SNP [48]. The strength of the relationship between alleles and gene expression intensities was estimated using Spearman's rank correlation and reported as nominal P-values.
Pik3cg knockout mice were obtained from sources previously described [49], backcrossed onto the C57BL/6J Jax genetic background for eight generations (B6J;129-Pik3cgtm1Pngr) and then maintained as a closed colony by intercrossing from within the colony (C57BL/6J Jax contribution: 99.6%). We performed PCR genotyping using a standard protocol with the following primer pairs: 5′-TCA GGC TCG GAG ATT AGG TA, 5′-GCC CAA TCG GTG GTA GAA CT (wild type); 5′-GGA CAC GGC TTT GAT TAC AAT C, 5′-GGG GTG GGA TTA GAT AAA TG (mutant) [49]. Whole blood from three Pik3cg−/− and three C57BL/6J Jax wild type mice (all females, age: 13–16 weeks, Mouse Breeders Diet (Lab Diets 5021-3)) was collected from terminally anesthetized mice via the retro-orbital sinus. We extracted total RNA using the Mouse RiboPure-Blood RNA Isolation Kit (Ambion). Total RNA was quantified using the NanoDrop and quality-checked using the 2100 Bioanalyzer. Standard protocols were used to generate biotinylated cRNA and hybridize to MouseWG-6 v2 Expression BeadChips. Arrays were washed, detected and scanned. All experimental procedures were carried out according to the manufacturer's protocol. On the raw expression data, we performed background subtraction, variance-stabilizing transformation and quantile normalization across all samples with the R package lumi (http://bioconductor.org/packages/release/bioc/html/lumi.html). Technical replicates were averaged and the differentially expressed transcripts between wild type and knockout mice were identified by calculating the log2-fold changes of the averaged expression values. P-values were calculated by 1-way analysis of variance (ANOVA). We performed gene ontology term enrichment analysis using AmiGO v1.7 [50], with the following parameters: gene expression fold-change cut-off: ±1.5; background: MGI; P-value cut-off: 1×10−5; minimum number of gene products: 10. The whole-genome gene expression data sets of Pik3cg−/− and wild type mice are available online in the GEO database under accession number GSE26111.
Of the 220 differentially expressed genes between Pik3cg−/− and wild type mice, we retrieved 191 orthologous human genes using BioMart (http://www.ensembl.org/biomart/martview/) and their respective proteins using UniProt (http://www.uniprot.org/). These ‘core’ proteins were used as primary seeds to develop the protein-protein interaction network. We determined first-order interactors of core proteins using Reactome v36. Only clustered non-redundant first-level interactions between human proteins that were connected to the largest connected component were considered. Based on the HaemAtlas data, we neglected interactors that are not expressed in MK cells (P>0.01). The described network is available for download in Cytoscape v2.8.0 format (Dataset S1).
|
10.1371/journal.pgen.1000980 | Cushing's Syndrome and Fetal Features Resurgence in Adrenal Cortex–Specific Prkar1a Knockout Mice | Carney complex (CNC) is an inherited neoplasia syndrome with endocrine overactivity. Its most frequent endocrine manifestation is primary pigmented nodular adrenocortical disease (PPNAD), a bilateral adrenocortical hyperplasia causing pituitary-independent Cushing's syndrome. Inactivating mutations in PRKAR1A, a gene encoding the type 1 α-regulatory subunit (R1α) of the cAMP–dependent protein kinase (PKA) have been found in 80% of CNC patients with Cushing's syndrome. To demonstrate the implication of R1α loss in the initiation and development of PPNAD, we generated mice lacking Prkar1a specifically in the adrenal cortex (AdKO). AdKO mice develop pituitary-independent Cushing's syndrome with increased PKA activity. This leads to autonomous steroidogenic genes expression and deregulated adreno-cortical cells differentiation, increased proliferation and resistance to apoptosis. Unexpectedly, R1α loss results in improper maintenance and centrifugal expansion of cortisol-producing fetal adrenocortical cells with concomitant regression of adult cortex. Our data provide the first in vivo evidence that loss of R1α is sufficient to induce autonomous adrenal hyper-activity and bilateral hyperplasia, both observed in human PPNAD. Furthermore, this model demonstrates that deregulated PKA activity favors the emergence of a new cell population potentially arising from the fetal adrenal, giving new insight into the mechanisms leading to PPNAD.
| Carney complex is a rare familial disease characterized by a predisposition to develop multiple endocrine tumors and highly morbid syndromes due to endocrine overactivities. Its most frequent endocrine manifestation, hypersecretion of glucocorticoids i.e. Cushing's syndrome, is caused by micronodular adrenal gland hyperplasia, an unusual neoplasia which combines both hyperplastic and atrophic areas. Inactivating mutations of the gene encoding the regulatory subunit 1α (R1α) of the cAMP–dependent protein kinase were frequently found in these patients, but the causal link between loss of R1α and onset of this adrenal disorder had not yet been established. Here, we describe the first mouse model mimicking this disease and provide mechanistic insights into endocrine overactivity and neoplastic transformation. Indeed, we show that lack of R1α induces autonomous expression of genes involved in steroid biosynthesis and resurgence of hyperplastic fetal-like cells with concomitant defects in cell renewal of the adult cortex. Our data therefore represent a substantial conceptual advance on the cellular dynamics involved in adrenal gland homeostasis. They suggest that regression of fetal structures may be important to establish normal endocrine functions and to allow cell renewal in the definitive cortex. Failure to clear out cells of fetal features in R1α-deficient adrenals leads to morbid hyperplasia.
| Primary pigmented nodular adrenocortical disease (PPNAD) is a rare form of bilateral micronodular adrenocortical hyperplasia leading to high morbidity due to ACTH (adreno corticotropic hormone)-independent Cushing's syndrome. PPNAD may be either sporadic or regarded as the most frequent endocrine manifestation of Carney complex (CNC), an autosomal dominant multiple neoplasia syndrome characterized by cardiac myxomas, spotty skin pigmentation and endocrine overactivity [1]. Cushing's syndrome in PPNAD is most diagnosed in children and young adults. Both isolated PPNAD and CNC have been associated with inactivating mutations in PRKAR1A, the gene encoding the type 1α regulatory subunit (R1α) of the cAMP-dependent protein kinase (PKA) [2], [3]. Among CNC patients with Cushing's syndrome, the frequency of PRKAR1A mutations is about 80%. Tumour-specific loss of heterozygosity within the chromosomal region harboring PRKAR1A is observed in tumours from CNC patients and isolated PPNAD, suggesting that PRKAR1A is a potential tumour suppressor gene [4]. Because general homozygous loss of Prkar1a is lethal in early mouse embryos, various haploinsufficiency and tissue-specific knock-out models have been engineered to demonstrate its tumour suppressor activity [5], [6]. General down-regulation of R1α levels has been achieved either in mouse lines heterozygous for a null allele of Prkar1a [6], [7] or in a transgenic line carrying an inducible antisense-construct [8]. Both approaches indicate that haploinsufficiency for Prkar1a predisposes to tumour formation in a spectrum of endocrine and non-endocrine tissues that are cAMP-responsive; the mouse phenotype partially overlaps with the human one. However haploinsufficiency in mouse models does not appear to be sufficient to promote tumour formation in a subset of tissues known for their propensity to develop neoplasms in CNC patients. Thus, complete loss of Prkar1a using heart-, Schwann cell- or pituitary-specific knockouts was required to induce tumours in these tissues [9]–[11]. To date, although PPNAD is the most frequent endocrine disorder observed in CNC patients [12], little is known on its pathophysiology. No clear adrenal lesions nor Cushing's syndrome were observed in mouse models of haploinsufficiency, suggesting that complete loss of Prkar1a might be required to phenocopy human phenotype. To address directly this question and obtain a possible mouse model for PPNAD, we produced mice with targeted Prkar1a gene inactivation in adreno-cortical cells by mating Prkar1a floxed mice with Akr1b7-Cre mouse line, a Cre expressing line allowing specific gene ablation in the steroidogenic lineage of the adrenals [13]. Adrenal cortex-specific Prkar1a knockout mice (AdKO) develop pituitary-independent Cushing's syndrome and evident signs of deregulated adreno-cortical cells differentiation and hyperplasia. These defects lead to improper maintenance and expansion of foetal adrenal cells in adult adrenals and establishment of tumoural conditions. Deregulation of the inhibin-activin signalling pathway seems to be implicated in this improper maintenance in AdKO mice model and in the human pathology. Our data provide the first in vivo evidence that the absence of R1α subunit of PKA is sufficient to induce the autonomous adrenal hyper-activity and bilateral hyperplasia observed in PPNAD. They also strongly suggest that deregulated PKA activity positively affects the maintenance of foetal characteristics in adult adrenal glands.
To assess the impact of complete loss of Prkar1a on adreno-cortical function and initiation of PPNAD, we crossed the Akr1b7:Cre line [13] with mice carrying the conditional null allele Prkar1aloxP [6] to produce adrenal cortex-specific KO mice of the Akr1b7:Cre;Prkar1aloxP/loxP genotype (Figure 1A). In this study, these mice were referred to as AdKO mice, and wild-type mice (WT) were of the Prkar1aloxP/loxP genotype. The Prkar1aΔ2 allele (KO allele) was detected by PCR in the DNA extracted from AdKO adrenals but absent from gonads and WT tissues (Figure 1B). As expected the intact conditional allele was still detected in the adrenals since Cre-mediated recombination was not supposed to occur in the whole organ but only in the cortex. Western blot and RT-QPCR analyses confirmed that Prkar1a gene expression was impaired in the adrenal glands of AdKO mice at both the mRNA (50% decrease) and protein levels (60% decrease) (Figure S1A and Figure 1C) when compared to WT. The 60% loss of R1α protein in adrenal tissue lysate of AdKO mice was accompanied by a significant increase in accumulation of R2β and C PKA subunits (Figure 1C), a phenomenon that is also observed in PPNAD [8]. By contrast, no significant changes were observed at the mRNA levels, indicating that upregulation of R2β and C subunits, involved a post-trancriptional mechanism (Figure S1B). We performed mRNA in situ hybridization and immunostaining to confirm that the decrease of Prkar1a gene expression was due to Cre-mediated gene ablation within the cortical compartment. As shown in Figure 1D, R1α mRNA signal was unaffected in medulla but was lost in the vast majority of cortical cells. These observations were confirmed at the protein level by R1α immunostaining (Figure 1D).
AdKO mice were born at expected Mendelian frequency and no difference in viability, weight or blood glucose values was observed up to 18 months of age when compared to WT mice (data not shown). Adrenal endocrine function and histological differentiation were explored in groups of mice of both sexes at 5, 10 and 18 months of age.
Visual examination of AdKO females from the age of 10 months onwards revealed neck humps formed of large accumulations of adipose tissue. This “buffalo hump” phenotype was never observed in WT females (Figure 2A) nor in males of both genotypes (not shown). Alteration of the repartition of fat depots is one of the features of “classic” Cushing's syndrome and is observed in PPNAD patients with PRKAR1A inactivation. In agreement with these observations, basal corticosterone levels in plasma were at least 2-fold higher in 10- and 18-month-old AdKO females than in age-matched WT, while no difference could be detected at 5 months (Figure 2B). Basal corticosterone levels were not affected in AdKO males of 5 months (8.9±6.3 ng/mL in WT vs 7.8±5.0 ng/mL in AdKO), 10 months (8.5±6.5 ng/mL in WT vs 7.8±2.1 ng/mL in AdKO) or 18 months of age (7.0±4.6 ng/mL in WT vs 8.2±4.2 ng/mL in AdKO). ACTH levels were measured in plasma of 10 months females. Importantly, ACTH levels measured in AdKO females were at least unchanged or tended to decrease (21±7 pg/mL in WT vs 14±7 pg/mL in AdKO), indicating that their basal hypercorticosteronaemia was independent of pituitary and likely resulted from primary adrenal overactivity. To explore the mechanism of hypercorticosteronaemia, AdKO mice that had not declared frank Cushing's syndrome, i.e. 5 months females and 10 months males, were injected with dexamethasone to induce a complete blockade of the hypothalomo-pituitary-adrenal (HPA) axis and subsequent suppression of ACTH production (Figure 2C–2E). Dexamethasone suppression test led to the expected decrease of adrenal weight (measured in females) as well as cortical atrophy in WT mice but had no effect on AdKO adrenals (Figure 2C, 2D and S2). Moreover, corticosterone levels were undetectable in plasma of WT mice after dexamethasone treatment whereas they remained unaltered in AdKO mice (Figure 2E). Finally, ACTH replacement in dexamethasone-treated mice restored corticosterone levels in WT and led to a further increase in AdKO mice, indicating that lack of R1α did not impair ACTH inducibility of steroidogenesis (Figure S3). Altogether, these data demonstrated that the adrenal glands of AdKO mice acquired the ability to secrete corticosterone in an autonomous manner leading to frank (in females) or subclinical (in males) ACTH-independent Cushing's syndrome.
Glucocorticoid biosynthesis depends on the continuous ACTH stimulation of adrenal steroidogenic and detoxification genes, through the cAMP/PKA signalling pathway. We thus studied the expression level of ACTH-dependent (Star, Akr1b7, Cyp11a1, Cyp11b1) and -independent (Cyp11b2) genes in WT and AdKO adrenal glands (Figure S4). RT-QPCR showed that basal hypercorticosteronaemia found in 10 months AdKO females correlated with a significant increase in Star mRNA levels (Figure S4B). A corresponding rise of StAR protein accumulation was confirmed by western blot (Figure S4B, inset). Consistent with their milder phenotype, males did not show any significant change in the expression of steroidogenic genes. By contrast, when AdKO mice with subclinical Cushing's syndrome (5-month-old females and 10-month-old males) were submitted to dexamethasone suppression test, most of the ACTH-responsive genes remained upregulated when compared to WT (Figure S4C). As expected, Cyp11b2 gene expression remained unchanged, showing that this response of mutant mice depended on ACTH signalling. We then checked whether Prkar1a ablation in AdKO mice led to the expected increase in PKA signalling, by measuring the PKA kinase activity and CREB phosphorylation on ser133 residue. Kinase assays demonstrated that basal PKA activity (in the absence of cAMP) was increased in mutant adrenals while total activity (in the presence of cAMP) remained unchanged (Figure 3A). In agreement with an increase in basal PKA activity, the amount of P-CREB in AdKO adrenals was doubled when compared to WT (Figure 3B). All these converging data demonstrated that the adrenal cortex-specific ablation of the Prkar1a gene led to primary pituitary-independent hypercorticosteronaemia through enhancement of PKA signalling.
Histological abnormalities consisting of large eosinophilic fœtal-like cells emerging from the innermost part of the adrenal cortex were detected in 5-month-old AdKO mice (Figure 4A and 4D). Eosinophilic cells appeared clearly hypertrophic when compared to WT spongiocytes (229±42 µm2 vs 110±15, p<0.01) (Figure 4H and 4K, insets). This hypertrophic cell population expanded centrifugally to represent more than 50% of the cortex at 10 months (Figure 4B and 4E) and most of the cortex by 18 months (Figure 4C and 4F). Simultaneously, neighbouring zona fasciculata cells, still arranged in tighly packed cords in 5-month-old mutant mice, became gradually disorganized in 10-month-old adrenals and appeared completely atrophic at 18 months (Figure 4G–4L). At this stage, the zona glomerulosa no longer appeared as a continuous layer of cells but as groups of glomeruli, isolated from each other by small hyperplastic spindle-shaped basophilic cells, arising from the subcapsular region (Figure 4L). This region contains adrenal stem/progenitor cells ensuring the continuous renewal of the adult cortex [14]–[16]. We thus assessed possible changes in the contents of adrenocortical progenitors in AdKO females of 10 and 18 months of age by quantifying expression of the progenitor-specific marker Shh and Gli1 [16] and the potential stem cell marker Pod1 [17]. RT-QPCR analyses showed that expression of these genes were not affected by the genotype nor the age (Figure S5), suggesting that the number of adrenocortical progenitors may not be affected in AdKO mice. The different cell populations in this area were characterized by double immunostaining for Sf1, a marker of steroidogenic lineage, and for β-catenin, a marker of subcapsular steroidogenic lineage mostly represented by zona glomerulosa cells [18]. Double immunostaining of WT adrenals from 18-month-old mice confirmed that both markers (Sf1 staining in the nucleus and β-catenin mostly at the cell periphery) colocalized in the cells of zona glomerulosa that formed a continuous layer in the outermost cortex (Figure 4M). By contrast, in age-matched AdKO adrenals, the general disorganization of the innercortex (Sf1-positive, β-catenin-negative cells) and the discontinuous aspect of the subcapsular/glomerulosa zone (Sf1/β-catenin-positive cells) were obvious (Figure 4N). A detailed view of this area showed hyperplastic spindle-shaped cells, both Sf1- and β-catenin-negative, were surrounding the Sf1- and β-catenin-positive glomeruli (Figure 4O). In previous mouse models for adrenocortical tumours, the first signs of neoplastic transformation seemed to coincide with the emergence of Gata-4 positive cells growing centripetally from the subcapsular region [15]. Thus, we examined expression of this transcription factor by immunostaining on 18-month-old adrenal sections. As shown in Figure 4P-4Q, numerous cells with Gata-4 nuclear staining could be detected within the subcapsular hyperplastic region of AdKO adrenals while, as expected, very rare positive cells were observed in WT.
To determine the mechanisms leading to early hyperplasia of the large eosinophilic cells and late hyperplasia of the small spindle-shaped cells, cellular proliferation within adrenal glands of 5–18 months mice was assessed by immunodetection of Ki67 (Figure 5A). At 5 or 10 months, WT and AdKO adrenals did not show any difference in the number of Ki67 positive cells (not shown). By contrast at 18 months, the number of proliferative cells was more than doubled in mutant adrenal cortex. All Ki67-positive cells were also and thus probably corresponded to the large eosinophilic cells but not to the Sf1-negative small spindle-shaped cells (Figure 5B). The increased cell proliferation was only evident in late stage of the AdKO phenotype and then could not be sufficient to explain the hyperplasia of innercortex observed in 5-10-month-old mice. We thus tested the sensitivity of adrenocortical cells to apoptosis induced by dexamethasone injections in 5-month-old mice [19]. Although apoptotic cells, assessed by positive staining for cleaved caspase 3, were detected within the cortex of both genotypes upon dexamethasone treatment, adrenal sections from AdKO mice showed 62% less apoptotic cells than WT (Figure 5C). These data support the view that early hyperplasia observed in 5-10-month-old AdKO mice could be, at least in part, the result of decreased sensitivity to apoptosis. TGFβ superfamily members inhibin and activin play a critical role in the growth dynamics of transient zones in the developing adrenal of both human and mouse [20], [21]. The expression of genes encoding the activin subunits (Inhβa and Inhβb), the inhibin subunit (Inhα) and the activin-binding protein follistatin (Fst) were compared in WT and AdKO adrenals (Figure 5D). The mRNA levels of inhibin subunit and follistatin were 2-fold higher in AdKO adrenals whereas expression of activin subunits remained unchanged. To assess the relevance of this observation in the human adrenal, we realised an immunostaining against INHIBIN-α on sections from normal (3 patients) and PPNAD-affected adrenocortical tissues (5 patients) (Figure 5E and Figure S6). Hypertrophic cells that form the nodules were strongly stained. By contrast, no significant signal could be detected in the surrounding tissue corresponding to zona fasciculata nor in sections of normal adrenal tissues. Hence, in both mice and humans, R1αdepletion in adrenal cortex led to increased expression of TGFβ members known for their antagonistic effects on apoptotic action of activins [22].
Adrenocortical-specific ablation of the Prkar1a gene led to expansion of hypertrophic eosinophilic cells emerging from the innermost part of the cortex, adjacent to adrenal medulla (Figure 4). We hypothesized that this cell population could originate from the X-zone, a transient zone of fœtal origin that regresses during the first pregnancy in female and at puberty in male mice [23], [24]. Hence, we compared adrenal zonation in WT and mutant mice by using Akr1b7 and 20α-HSD immunostaining to delineate zona fasciculata and X-zone, respectively [25], [26] (Figure 6 and Figure S7). Adrenal cortex from WT virgin females showed canonical concentric organization consisting of three adjacent zones: X-zone (20α-HSD-positive and Akr1b7-negative), zona fasciculata (20α-HSD-negative and Akr1b7-positive) and zona glomerulosa (20α-HSD-negative and Akr1b7-negative) (Figure 6A). By contrast, in the adrenal gland of 5-month-old AdKO virgin females although the 20α-HSD-expressing cells remained adjacent to the medulla, the X-zone and zona fasciculata were now overlaping in the innermost part of the cortex, and some isolated cells co-expressed both 20α-HSD and Akr1b7 markers (Figure 6B). As expected, in 10-month-old parous WT females, X-zone had completely regressed and 20α-HSD-expressing cells were no longer detected (Figure 6C). By contrast, adrenal cortex from age-matched parous AdKO females showed a persistent large X-like-zone that has clearly expanded in a centrifugal direction (Figure 6D). At this stage, most 20α-HSD positive cells of the X-like-zone also expressed Akr1b7 and the typical packed cords organization of zona fasciculata was no longer observed in the Akr1b7-positive/20α-HSD-negative remaining cortex. In 18-month-old females, the X-like-zone had further expanded and now represented most of the cortex. Akr1b7-positive/20α-HSD-negative cells were repelled to adrenal periphery (Figure 6E). Interestingly, examination of the proliferative potential of the X-like-zone using double immunostaining showed that Ki67-positive/20α-HSD-positive cells could be found in both 10 and 18-month-old AdKO females (Figure S8).
Many 20α-HSD positive cells were also detected in 18-month-old AdKO males although the observed phenotype was milder than in females (Figure S9). Indeed, most 20α-HSD-expressing cells were not Akr1b7-positive and the overlap of X-like-zone with zona fasciculata was limited to the innermost cortex. In mouse, natural X-zone regression at puberty can be suppressed by castration of pre-pubescent males. To explore the possible reasons for the less pronounced phenotype in AdKO males, we tested whether gonadectomy occuring at 3-weeks of age (before natural X-zone regression) could accelerate the onset of the adrenal defects in 3-month-old adult males (Figure S10). 20α-HSD/Akr1b7 co-immunostaining showed that castration allowed the maintenance of a classical X-zone in WT, and of a X-like-zone overlaping fasciculata in AdKO adult males. When compared to shame-operated AdKO males of the same age, gonadectomized AdKO males showed a high number of Akr1b7-positive/20α-HSD-positive cells, these cells being never observed in gonadectomized WT males (Figure S10A). Consistently, corticosterone levels were more elevated in gonadectomized than in shame-operated AdKO males (Figure S10B). By contrast, gonadectomy had no impact on corticosterone levels in WT males. These data suggested that gonadectomy was able to amplify the phenotype in male AdKO mice.
The persistence of X-zone marker suggested that foetal characteristics were maintained throughout adult life of AdKO mice. Cyp17 is a steroidogenic enzyme involved in the biosynthesis of precursors of both sex steroids and cortisol. In rodents, however, Cyp17 is only transiently expressed in the foetal adrenal and is therefore considered a foetal marker [27]. RT-QPCR analyses showed that, as opposed to WT, most adult AdKO adrenal glands expressed high levels of Cyp17 transcripts (Figure 7A). In addition, Cyp17 positive immunostaining was detected within the innercortex of AdKO adrenal glands (Figure 7B). More importantly, this expression led to the production of a functional Cyp17 enzyme since AdKO mice produced detectable levels of cortisol (Figure 7C). Because both corticosterone and cortisol production required the continuous expression of genes encoding Cyp21 and Cyp11b1 biosynthetic enzymes, we examined whether their expressions were maintained throughout the progression of the AdKO phenotype (Figure S11). RT-QPCR analyses demonstrated that Cyp21 and Cyp11b1 expression levels were unchanged with age in both WT or AdKO females.
Altogether, these data demonstrate that adrenal-specific ablation of Prkar1a altered the differentiation program of the adult cortex by promoting the improper maintenance and centrifugal expansion of steroidogenic competent foetal-like cells, that had the capacity to proliferate and to produce glucocorticoids (cortisol and corticosterone).
Here, we shown that the adrenal-specific ablation of Prkar1a, the Carney Complex gene 1 (CNC1), in mouse reproduced the essential features of PPNAD observed in humans carrying PRKAR1A mutations. AdKO mice developed ACTH-independent Cushing's syndrome and cortical hyperplasia combined with atrophic areas that are typical hallmarks of PPNAD [1]. This mouse model definitively proves the central role of PRKAR1A gene defects in the etiology of PPNAD. Furthermore, the discovery of an unexpected role of Prkar1a in the repression of foetal features in adrenal cortex provides novel mechanistic insight into the cellular dynamics leading to definitive adrenal tissue or, when disturbed, to morbid hyperplasia.
AdKO mice phenocopied most of the features of adrenal overactivity seen in patients. From a clinical point of view, PPNAD is difficult to diagnose because Cushing's syndrome usually develops slowly. Hypercortisolism may be mild or even periodic, with no clear decrease in plasma ACTH levels [28]–[30]. Adrenal-specific disruption of Prkar1a triggered subclinical hypercorticosteronaemia revealed upon blockade of pituitary ACTH, in 5-month-old mice. Around one year of age, it evolved into frank Cushing's syndrome with low, but still detectable levels of plasma ACTH. Contrasting with PPNAD, we did not detect any paradoxical rise in corticosterone levels after dexamethasone injection in AdKO mice, but only a resistance to ACTH blockade. Paradoxical response would rely on increased expression of the glucocorticoid receptor (GR) that was shown to activate PKA in PPNAD nodules [31], [32]. We did not find any AdKO-dependent increase in GR expression neither by measuring mRNA levels nor by immunohistochemical analyses (not shown). This paradoxical response could likely be a feature of human cells since adrenal cultures from Prkar1a haploinsufficient mice did not show paradoxical dexamethasone response in perifusion experiments [32]. Another discrepancy between AdKO adrenals and PPNAD was the absence of pigmentation in the mice glands. Hyper-pigmentation in PPNAD nodules relies on the accumulation of lipofuscin and is a consequence of autophagic deficiency [33], [34]. This decreased autophagy was thought to originate from the R1α loss and consecutive activation of mTOR signalling [35]. Lipofuscin is made of aldehyde-linked protein residues that make it non-degradable and that form under chronic mild oxidative stress conditions, at a rate inversely related to the average lifespan of species [36]. We thus speculate that more efficient enzymatic defenses against reactive aldehydes forming aducts [25] and/or shorter lifespan might preserve mice from adrenal lipofuscin accumulation under R1α depletion. The cytomegalic aspect of eosinophilic cells arising from the innercortex is a hallmark of hyperplasia seen in PPNAD patients or in AdKO mice, and could be linked to unbuffered mTOR activity [35], [37] that is a prerequisite to increased cell size [38]. Works are in progress to explore the contribution of this pathway in the pathophysiology of the AdKO model.
Although they were attributed to the lack of R1α regulatory subunit of PKA, until now, the mechanisms leading to adrenal overactivity in PPNAD were not clear. The PKA heterotetrameric holoenzyme is composed of a dimer of regulatory subunits combined with two catalytic subunits. When the regulatory subunits bind cAMP, they dissociate from the catalytic subunits, which in turn, exhibit their kinase activity [39]. In previous knockout mouse models with general loss of R1α, basal PKA activity (linked to free catalytic subunits only) measured in embryos was found increased whereas total PKA activity (cAMP-stimulated) was decreased [5]. When measured in mouse embryonic fibroblasts both activities were increased [40]. The net impacts of R1α depletion on PKA activity could therefore depend on the cell type or tissue context. Consistent with these studies, specific depletion of R1α in adrenals mainly triggered a rise in basal PKA activity, attested both by increased catalytic activity and CREB phosphorylation. This resulted in a net gain of Star gene expression and therefore increased basal steroidogenesis. In agreement with our findings, StAR gene expression was found upregulated in a serial analysis of gene expression (SAGE) of PPNAD tissues [41].
The most intriguing phenotype observed during the follow-up of AdKO mice was an atypical hyperplasia of foetal-like cortex emerging at the corticomedullary junction which, over time, extended to the periphery. However, concomitant atrophy of the adult cortex resulted in net adrenal size equivalent to WT. These results were reminiscent of early histopathological studies of PPNAD, showing that micronodules seemed to arise from the medulla-cortex boundary. These were composed of eosinophilic giant cells that were surrounded by mostly atrophic cortex resulting in an otherwise normal-sized gland [1], [42], [43]. In mice, a transient cell layer, termed the X-zone, is adjacent to the medulla and regresses at puberty in males and at the first pregnancy in females [23]. Pioneer work from Morohashi‘s laboratory provided unequivocal genetic proofs that the X-zone was a remnant of foetal cortex forming before the definitive cortex and that distinct pools of precursor cells within the foetal cortex contributed to either the definitive cortex or the transient X-zone [24], [44]. One pool of precursors activated transiently the foetal adrenal-specific enhancer of Ad4BP/Sf1 (FAdE) and contributed to the definitive cortex while the second pool maintained FAdE activated and contributed to the formation of the transient X-zone [24]. We showed previously that the developmental pattern of Akr1b7-Cre mediated recombination was reminiscent of that observed with the FAdE construct and that it occurred in both foetal and definitive adrenocortical cells [13]. Here, we provided evidence that loss of R1α during adrenal development resulted in two major abnormalities: unbuffered PKA activity leading to endocrine overactivity, and persistence of foetal-like cells that expanded across the adult cortex. In human, INHIBIN-α is more expressed in foetal than in adult adrenals [45]. In PPNAD, besides the hypertrophic aspect of the cells, the overexpression of INHIBIN-α specifically in the nodules could be interpreted as another sign of foetal origin of these cells. In mice, the foetal character of these hyperplastic and hypertrophic cells was attested both by persistent expression of 20α-HSD, an X-zone marker, and by re-expression of Cyp17, an enzyme otherwise restricted to the embryonic period in rodent adrenals [26], [27]. However, in contrast to natural X-zone cells that had no reported steroidogenic potential, the foetal-like hyperplastic cells of AdKO adrenals had acquired full steroidogenic competence of zona fasciculata cells and produced both corticosterone and cortisol. It is tempting to speculate that this (hyper)cortisolism, hitherto never described in mouse, could participate to the Cushing's syndrome of AdKO mice.
Our data demonstrated that AdKO adrenals were less sensitive to apoptosis than WT. Apoptosis mediated by TGFβ family members largely contributed to the regression of both human foetal zone and mice X-zone in which both inhibin and follistatin opposed to the apoptotic signal triggered by activins [20], [21]. This is in good agreement with our observation that inhibin-follistatin/activin transcripts ratio was augmented in AdKO adrenal glands and could therefore contribute to strengthen anti-apoptotic paracrine signals. In addition, R1α depletion could render foetal cells less sensitive to TGFβsignalling. This mechanism likely occured in human cells since we recently demonstrated that R1α knockdown in NCI-H295R adrenocortical cells enhanced their resistance to TGFβ-stimulated apoptosis [46]. Consistent with all these observations, here we showed an increased immunostaining for INHIBIN-α specifically in the nodules of PPNAD samples. Converging data in the literature highlight the importance of inhibin/activin system as a paracrine mediator of cAMP/ACTH signalling in both foetal and adult tissues [20], [47]–[49]. Accordingly, maintaining derepression of PKA activity in the adrenal glands of AdKO mice from the foetal period to adulthood would favour the maintenance of high levels of inhibin. Interestingly, in a murine cell line postulated to originate from the X-zone [50], inhibin was shown to counteract the repressive effect of activin on the Cyp17 gene [51]. This could provide a plausible but yet-non-demonstrated mechanism for re-expression of Cyp17 in the adult mutant gland of AdKO mice.
Most adrenocortical tumours and Cushing's syndrome are more frequent in females than in males [52]. PPNAD does not escape this rule [12]. By the age of 40 years, more than 70% of the female carriers of PRKAR1A defects had clinical evidence of this disease, whereas only 45% of the male carriers were concerned. In addition, PPNAD was diagnosed at a younger age in females than in males. These clinical outcomes were strikingly reminiscent of the phenotype of AdKO mice that was earlier and more severe (Cushing's syndrome and hyperplasia) in females than in males. Although the influence of sex-specific hormones cannot be ruled out as suggested by the (permissive) aggravating effect of gonadectomy in AdKO males, some gender specificities of foetal cortex cells could account for these differences in mouse. First, foetal expression of Cyp17 was nearly completely down-regulated at E14.5 in males, whereas down-regulation occurred only at birth in females [53]. Second, post-natal foetal cortex, the X-zone, regressed at puberty in males but only during first pregnancy in females. Thus, foetal cells (and among them, foetal/X-zone precursor cells) remained for a longer time in the female cortex. Since our mouse model showed that R1α loss contributed to foetal-like cortex persistence and expansion, it is reasonable to assume that enrichment in foetal cells (or foetal precursor cells) predisposes AdKO females to manifest more severe PPNAD. To our knowledge, possible gender differences in foetal cortex dynamic changes have never been addressed in human adrenals.
According to the centripetal model, cell renewal in the adrenal cortex depends on a common pool of stem/progenitor cells located in the periphery (within the fibroblastic capsule and/or the glomerulosa) which migrate centripetally from this zone, differenciate to successively adopt all the cortical fates and finally enter into apoptosis in the innermost cortex (Figure 8). According to cell lineage-tracing studies, two populations of progenitors contributed to the formation of the adult cortex, one located in the capsule expressed Gli1 and the second in a subcapsular position expressed Shh [16]. Although the role of capsule/subcapsule in the centripetal renewal of the adult definitive cortex is now fully established, there are also genetic evidences that in developing adrenal, the definitive cortex and the transient X-zone originate from different foetal precursors [24], [44].
According to the centripetal model, dividing cells are essentially present at the periphery while apoptotic cells preferentially concentrate at the inner cortex ([54], reviewed in [17]). The balance between these two opposing gradients could be essential for homeostatic maintenance of the adrenal cortex i.e. the establishment of a centripetal differentiation. In AdKO mice, the centrifugal expansion of foetal-like cells with proliferative potential emerging from the inner cortex and the progressive atrophy of zona fasciculata could indicate that this balance is perturbed. A possible model may be proposed to illustrate these observations (Figure 8). Indeed, loss of R1α allowed the maintenance of cells of foetal features, which otherwise were transient. This maintenance could result from both an improvement of their proliferative capacity and from their decreased sensitivity to apoptotis and/or alteration of the apoptotic gradient (as suggested by the increased expression of Inhα and Follistatin genes known to antagonise activins signalling). In addition, loss of R1α induced a progressive atrophy of the zona fasciculata in AdKO adrenals that was reminiscent to defects in cell renewal. Indeed, whereas the foetal-like cells undergo a continuous centrifugal expansion across the cortex, no gain in adrenal size was detected and no increase in cell apoptosis accompanied the concomitant atrophy of zona fasciculata. On the other hand, hyperplasia of non-steroidogenic subcapsular spindle-shaped cells was observed in elder mice (18 months) and their accumulation eventually affected the integrity of zona glomerulosa. At least two non-exclusive mechanisms could account for defective cell renewal in the definitive cortex of AdKO mice: depletion of progenitor cells or impaired capacity of these progenitors to undergo centripetal differentiation and clonal replenishment of the cortex. The latter mechanism seems more likely in our model. Indeed, the expression levels of markers for stem/progenitor cells (Figure S5) were unaltered in AdKO mice suggesting that progressive atrophy of the definitive cortex was not due to their depletion. Similar to observations made in most mouse adrenal tumour models, with age, adrenal glands of AdKO mice accumulated subcapsular Gata-4-positive spindle-shaped cells that are supposed to descend from multipotent progenitors capable to engage toward adrenal or gonadal fates [15], [55], [56]. In AdKO mice, late accumulation of Gata-4-positive cells could reflect an incapacity for the progenitor cells to properly differenciate into adrenocortical cells. This would prevent efficient renewal of the definitive cortex, which as a result, would become atrophic over time. An other hypothesis emerges from a recent report of Hammer and colleagues showing that inhibin-α prevented aberrant proliferation and differentiation of subcapsular adrenocortical progenitor cells [57]. Indeed, as opposed to Inhα−/− mice, adrenal glands of AdKO mice have increased inhibin-α expression. This is consistent with PPNAD samples and could therefore decrease proliferation/differentiation of progenitors. The possible dual role for inhibin-α in enhancing survival of foetal cells and impeding renewal of definitive cortex will have to be demonstrated in AdKO mice in an Inhα−/− context. In a symmetric point of view, the slow but continuous centrifugal expansion of hypertrophic foetal-like cells, would imply that a reservoir of foetal cortex precursor cells could lay in the juxtamedullary region and that differentiating steroidogenic foetal cells could emerge and replenish the cortex (Figure 8). Lineage tracing experiments will be required to confirm this hypothesis.
By developing a mouse model of PPNAD, we established for the first time that Prkar1a, the Carney Complex gene 1, not only controls adrenocortical endocrine activity but also prevents the maintenance of foetal remnants. The loss of R1α acts, at least, by increasing PKA activity and possibly by PKA independent effects mediated through alteration of protein interactions that remain to be deciphered [35], [58]. The data existing in the literature and our present results strongly suggest a role for the inhibin-activin signalling pathway in the progression of the disease. Adrenal hyperplasia observed in PPNAD is classified as a neoplastic lesion. Although we showed that R1α loss induced tumoural conditions in adrenal glands (resistance to apoptosis, cell hypertrophy, mild proliferation), profound alterations in zonal differentiation and cell renewal suggest that PPNAD should also be considered as a developmental disease.
Informed signed consent for the analysis of adrenal tissue and for genetic diagnosis was obtained from the patients and the study was approved by an institutional review board (Comité Consultatif de Protection des Personnes dans la Recherche Biomédicale, Cochin Hospital, Paris). PPNAD paraffin sections were performed from adrenal samples of patients with isolated PPNAD or PPNAD with Carney complex who underwent bilateral adrenalectomy for ACTH-independent Cushing's syndrome. All patients carried germline inactivating mutations of the PRKAR1A gene.
Animal studies were done in agreement with standards described by the NIH Guide for Care and Use of Laboratory Animals as well as with the local laws and regulations applicable to animal manipulations in France. For all analyses, groups of 17–20 mice of each genotype (WT and AdKO) and each age (5, 10 and 18 months) were constituted. Adult mice were injected s.c. with vehicle (sesame oil), dexamethasone acetate for 4 days (75 µg twice daily; Sigma-Aldrich, L'Isle d'Abeau Chesnes, France) and injected i.m. with long-acting ACTH (1.2 U, Synacthene, Novartis Pharma S.A., Rueil-Malmaison, France) the day before and in the morning of the experiment.
Mouse genomic DNA (from tail, adrenal or gonad) was extracted and analyzed by PCR. Genotyping for the 0.5 Akr1b7-Cre transgene was carried out using the following conditions: 94°C, 45 s; 55°C, 45 s; 72°C, 45 s for a total of 40 cycles (primers: 5′-CCTGGAAAATGCTTCTGTCCG-3′; 5′-CAGGGTGTTATAAGCAATCCC-3′) and Prkar1aloxP/loxP intact or knockout allele were genotyped using the following conditions: 94°C, 90 s; 58°C, 90 s; 72°C, 90 s for a total of 35 cycles (primer a: 5′- CACTGCAGGGGCCTATTTTA -3′; primer b: 5′-TGTCTAGCTTGGGGTGGACT-3′, primer c: 5′-CATCCATCTCCTATCCCCTTT-3′).
Mice were sacrified by decapitation at 8–9 am with minimum handling (within 1 min), trunk blood was collected in eppendorf tubes containing 5 µL EDTA 0.5 M and placed immediately at 4°C. Samples were spun down at 4000 g for 5 min at 4°C and the resultant plasma was stored at −20°C for corticosterone or cortisol analysis, or at −80°C for ACTH analysis.
Corticosterone concentrations in plasma were determined by radioimmunoassay (RIA) using a commercially available kit (ICN Biomedicals, Orsay, France). ACTH dosage in plasma were performed by solid-phase, two-site sequentiel chemiluminescent immunometric assay (Siemens Healthcare Diagnostic SAS, Saint-Denis, France) using an Immulite 2000 analyzer. Cortisol concentrations were determined by electrochemiluminescence immunoassay (Roche Diagnostic, Meylan, France) using a Modular Analytics E170 analyzer.
PKA activity was quantified in freshly dissected adrenals using the following commercial kit: PepTag assay for non-radioactive detection of cAMP-dependant protein kinase (Promega Corp., Charbonnière, France).
Total RNA and DNA (for genotype confirmation) were isolated from tissue with the Qiagen DNA/RNA Mini kit (Qiagen, Courtaboeuf, France). Total RNAs (1 µg) were reverse-transcribed by Moloney murine leukaemia virus reverse transcriptase (Promega Corp., Charbonnière, France) according to the manufacturer's instructions. Quantitative real-time PCR was performed using the iCycler BioRad system and BioRad IQ5 optical system software (BioRad, Marnes-la-Coquette, France) under standard conditions (40 cycles of 95°C for 15 seconds and 60°C for 60 seconds). All primer/probe sets were obtained from Applied Biosystems: Prkar1a, Prkar1b, Prkar2a, Prkar2b, Prkaca, Star, Akr1b7, Cyp11a1, Cyp11b1, Cyp11b2, Cyp21, Shh, Gli-1, Pod-1, Ppib, Cyp17, Inhα, Inhβa, Inhβb, Fst (Applied Biosystems, Courtaboeuf, France). For quantification of transcripts, all PCR were performed in triplicate and the DCt method was used to calculate mRNA levels relative to a Peptidylprolyl isomerase B (Ppib) standard.
A 442 bp 3′untranslated part of the Prkar1a cDNAs was amplified using the following primers: 5′-GGGCGTTGGAATTACTGAGA-3′; 5′-CTCCCAAATAGAACCCGACA -3′; and subcloned in pGEM-T easy vector (Promega Corp., Charbonnière, France). Antisense riboprobes were synthesized and labelled with digoxigenin (Boehringer Mannheim, Mannheim, Germany). Adrenals were fixed in 4% paraformaldehyde overnight, embedded in paraffin and sectioned. Sections were treated for 15 min with proteinase K (3 µg/ml) at room temperature and washed with glycine (2 mg/ml) and then with PBS. They were fixed with 4% paraformaldehyde for 5 min and washed with PBS. Samples were incubated in hybridization mix (50% formamide; 4x SSC; 10% Dextran sulphate; 1x Denhart's; Salmon sperm DNA 250 µg/ml; tRNA 250 µg/ml) for 1 h at 42°C. Digoxygenin labelled probe was added to the hybridization mix and incubated overnight at 42°C. Slides were then treated to a series of washes in 2x SSC and 1x SSC at 42°C and 0.2x SSC at room temperature. Sections were washed in buffer 1 (150 mM NaCl; 100 mM Tris, pH 7.5), blocked by Boehringer blocking reagent in buffer 1 then incubated 1 h at room temperature with peroxidase-conjugated anti-digoxygenin antibody. After several washes in buffer 2 (150 mM NaCl; 100 mM Tris, pH 9.5; 5 mM MgCl2), peroxidase activity was detected by incubation with 0.18 mg/ml BCIP and 0.34 mg/ml NBT in buffer 2. In situ hybridization slides were observed and photographed on an Axiophot microscope (Carl Zeiss, Zurich, Switzerland).
Adrenals were fixed overnight in 4% PFA and embedded in paraffin. Sections were then cut and deparaffinized in Histoclear. For general morphology, sections were stained with haematoxylin and eosin.
For mouse-anti-human-INHIBIN-α immunodectection, unmasking solution was sodium citrate buffer 10 mM pH 6, Tween 0.05%. For co-localisation experiments of Akr1b7/Ki67 with 20α-HSD, the following protocol of limit detection was used: deparaffinised sections were incubated for 20 min at 95°C with Unmasking Solution (Vector Laboratories, Peterborough England). For the first detection, rabbit-anti-Akr1b7 antibody [59] (1/1000) or rabbit-anti-Ki67 (1/500, Thermo Fischer Scientific, Elancourt, France) was revealed using a secondary biotinylated goat anti-rabbit antibody, Vectastain ABC amplification kit (Vector Laboratories, Peterborough England) and TSA fluorescein HRP substrate (Perkin Elmer, Courtaboeuf, France). For the second detection, slides were incubated with a rabbit-anti-20α-HSD antibody at 1/2000 (kind gift from Dr Y. Weinstein, Ben-Gurion University, Israël) revealed by goat anti-rabbit Alexa 555 at 1/1000 (Molecular probes, Cergy Pointoise, France). Sections were then incubated 5 min with Hoechst at 1 µg/ml (Sigma-Aldrich, L'Isle d'Abeau Chesnes, France), rinsed, mounted in PBS-glycerol, and photographed on an Axiophot microscope (Carl Zeiss, Zurich, Switzerland).
The following antibodies: Mouse-anti-R1α (1/50, BD Biosciences, Le pont de Claix, France); rabbit-anti-Sf1 (1/1000, kind gift from Dr K. Morohashi, Kyushu University, Japan), rabbit-anti-Cyp17 (1/5000, kind gift from Dr A. Conley, University of California, USA); rabbit-anti-cleaved-Caspase-3 (1/400, Cell signalling, Saint-Quentin-en-Yvelines, France); goat-anti-GATA-4 (1/100, Tebu-Santa Cruz, Le Perray en Yvelines, France), mouse-anti-β-catenin (1/500, BD Biosciences, Le pont de Claix, France), mouse-anti-human-INHIBIN-α (1/75, AbD Serotec, Oxford, UK) were detected using the same protocol as Akr1b7. The secondary biotinylated antibodies were donkey anti-goat to detect Gata-4 and sheep anti-mouse to detect R1α and β-catenin. Gata-4 detection was performed using the Novared Kit (Abcys, Paris, France).
For the double staining β-catenin/Sf1, 20 min in 0.02% HCl are necessary to abolish the rest of the peroxidase activity after the first immuno-reaction. Detection of Cyp17 was done without incubation in unmasking solution. An InSitu Pro VSi (Intavis AG) automated processor was used for immunodetection.
Adrenal samples and western blotting were done as described previously [60]. The Primary antibodies were used at the following dilutions: rabbit-anti-StAR (1/5000, kind gift from Dr Stocco, Texas Tech University Health Sciences Center, USA); mouse-anti-R1α (1/500); mouse-anti-R2α (1/1000); mouse-anti-R2β (1/1000); mouse-anti-Cαβ (1/1000, BD Biosciences, Le Pont de Claix, France), rabbit-anti-CREB (1/1000), rabbit-anti-P-CREB (1∶1000, Cell signalling, Saint-Quentin-en-Yvelines, France); rabbit-anti-βTubulin (1/1000, Sigma-Aldrich, L'Isle d'Abeau Chesnes, France). Quantification of western blot signals was performed using the Quantity One software (Biorad, Marnes la Coquette, France).
For statistical analysis, a Student t test was performed to determine whether there were differences between the two groups. Mann-Whitney test was used in Figure 7A. A P value of 0.05 was considered significant.
|
10.1371/journal.ppat.1002966 | Immunity to Intracellular Salmonella Depends on Surface-associated Antigens | Invasive Salmonella infection is an important health problem that is worsening because of rising antimicrobial resistance and changing Salmonella serovar spectrum. Novel vaccines with broad serovar coverage are needed, but suitable protective antigens remain largely unknown. Here, we tested 37 broadly conserved Salmonella antigens in a mouse typhoid fever model, and identified antigen candidates that conferred partial protection against lethal disease. Antigen properties such as high in vivo abundance or immunodominance in convalescent individuals were not required for protectivity, but all promising antigen candidates were associated with the Salmonella surface. Surprisingly, this was not due to superior immunogenicity of surface antigens compared to internal antigens as had been suggested by previous studies and novel findings for CD4 T cell responses to model antigens. Confocal microscopy of infected tissues revealed that many live Salmonella resided alone in infected host macrophages with no damaged Salmonella releasing internal antigens in their vicinity. In the absence of accessible internal antigens, detection of these infected cells might require CD4 T cell recognition of Salmonella surface-associated antigens that could be processed and presented even from intact Salmonella. In conclusion, our findings might pave the way for development of an efficacious Salmonella vaccine with broad serovar coverage, and suggest a similar crucial role of surface antigens for immunity to both extracellular and intracellular pathogens.
| Salmonella infections cause extensive morbidity and mortality worldwide. A vaccine that prevents systemic Salmonella infections is urgently needed but suitable antigens remain largely unknown. In this study we identified several antigen candidates that mediated protective immunity to Salmonella in a mouse typhoid fever model. Interestingly, all these antigens were associated with the Salmonella surface. This suggested that similar antigen properties might be relevant for CD4 T cell dependent immunity to intracellular pathogens like Salmonella, as for antibody-dependent immunity to extracellular pathogens. Detailed analysis revealed that Salmonella surface antigens were not generally more immunogenic compared to internal antigens. However, internal antigens were inaccessible for CD4 T cell recognition of a substantial number of infected host cells that contained exclusively live intact Salmonella. Together, these results might pave the way for development of an efficacious Salmonella vaccine, and provide a basis to facilitate antigen identification for Salmonella and possibly other intracellular pathogens.
| Enteric fever caused by systemic Salmonella infection causes tremendous morbidity and mortality worldwide [1]. Current control strategies become increasingly inefficient as a result of increasing antimicrobial resistance [2], [3] and emergence of Salmonella serovars that are not covered by currently available safe vaccines [4], [5]. This situation generates an urgent medical need for novel Salmonella vaccines with broad serovar coverage.
Early killed whole-cell vaccines containing mixtures of different serovars provide broad protection, but cause unacceptable adverse reactions [1]. As an alternative to whole-cell vaccines, subunit vaccines containing a few defined Salmonella components could minimize adverse reactions. Indeed, vaccines containing the capsular polysaccharide Vi antigen provide moderate protection and excellent safety [1]. On the other hand, serovars Paratyphi A and non-typhoidal Salmonella (NTS) that cause an increasing number of invasive salmonelloses [6], lack the Vi antigen and are therefore not covered by Vi vaccines [5]. Apart from Vi, few Salmonella antigens have been identified, and all of these provide at best moderate levels of protection against challenge infection with virulent Salmonella strains in the commonly used mouse typhoid fever model. Moreover, antigens such as flagellin [7] and OmpD [8] are poorly conserved among relevant serovars.
For extracellular pathogens with antibody-mediated immunity, protective antigens must be surface-exposed [9], and this enables an effective strategy for priorization of antigen candidates [9]. Humoral response to surface antigens can also contribute to immunity to intracellular pathogens such as invasive Salmonella [10]. Indeed, Vi which induces protective antibody responses in human vaccinees, forms an extracellular capsule around Salmonella Typhi [11]. Two additional antigens that confer partial immunity in the mouse typhoid fever model, flagellin [7] and SseB [12], are also part of Salmonella surface structures (flagella, translocon complex of a type III secretion system). Furthermore, outer membrane preparations (but not the outer membrane component lipopolysaccharide) have been suggested to mediate protective humoral immune responses against extracellular Salmonella bacteremia [13] and attenuated Salmonella strains in the mouse model [8], [14]. A number of porins such as OmpC, OmpD, and OmpF are highly abundant in such outer membrane preparations suggesting that they might represent the actual protective antigens [8], [14], [15].
However, immunity to Salmonella critically depends also on CD4 T cells [10]. Unfortunately, protective T cell antigens seem to be rare, and priorization of candidates is difficult since relevant antigen properties for CD4 T cell responses remain unclear [9], [16], [17]. One key precondition for protective responses is expression of the respective Salmonella antigen during infection [18], and some data suggest that highly abundant antigens might be particularly well recognized by CD4 T cells [12], [19]. Antigen in vivo expression can be deduced from various complementary approaches including screening of promoter trap libraries [20], [21], proteomics [22], serum antibody response [23]–[26], as well as mutant virulence phenotypes.
In addition to antigen expression, antigen immunogenicity could play a major role. Antigen detection by cognate CD4 T cells requires antigen processing and presentation of the resulting small peptides by major histocompatibility (MHC) class II molecules. Peptide sequence properties that are characteristic for well recognized epitopes, can be used for genome-wide prediction of promising antigens [27]. However, a large number of non-protective antigens contain putative high-score epitopes [16], [18], [28] which could compromise the discriminatory power of this approach.
Experimental detection of immune responses to an antigen in convalescent individuals that have survived infection, demonstrates that this antigen was expressed in vivo and could be recognized by the immune system [23]–[25]. Indeed, this approach has been recently shown to facilitate identification of protective Chlamydia antigens [29]. On the other hand, many immunodominant antigens in convalescent individuals lack protective efficacy, while a number of protective antigens may induce immune responses below the detection threshold during natural infection [17].
Another antigen property that can affect CD4 T cell responses is antigen localization. In particular, secreted or surface-associated antigens might induce particularly strong cellular immune responses because of superior processing, kinetic advantages compared to internal antigens, and/or physical association with pathogen-associated molecular patterns (PAMP) such as lipopolysaccharide that provide potent stimuli for innate and adaptive immunity [14], [30]–[36]. Indeed, secretion/surface localization has been widely used to prioritize candidates for antigen identification. However, antigens with likely internal localization can also induce specific CD4 T cell responses that mediate protection against various intracellular pathogens [37], [38].
Taken together, relevant antigen properties for CD4 T cell mediated immunity to intracellular pathogens remain poorly characterized, and this impairs antigen priorization for vaccine development. To address this issue, we compared here 37 diverse Salmonella antigens in a mouse model that closely mimics human typhoid fever [39]. The results suggested that recognition of surface-associated antigens might be necessary to detect and combat live intracellular Salmonella, whereas recognition of internal antigens would mediate futile non-protective attack of already dead Salmonella. In conclusion, we propose a similar crucial role of surface-associated antigens for immunity to both extracellular and intracellular pathogens.
To determine immune responses to Salmonella antigens, we selected 21 broadly conserved Salmonella proteins. We selected several subunits of the SPI-2 type III secretion system since the putative translocon subunit SseB of this system showed promising protectivity in previous studies [12], [26]. We also included several porins since a previous study had shown that OmpD conferred protection against an attenuated Salmonella mutant [8]. To explore the role of antigen localization we selected additional proteins localized in Salmonella cytosol, inner membrane, periplasm, and outer membrane/surface. To explore the role of antigen abundance, we determined absolute quantities of more than 1100 Salmonella in infected mouse spleen. Specifically, we purified Salmonella from infected mouse spleen using flow cytometry as described [22]. We determined absolute protein quantities in these ex vivo purified Salmonella using shot-gun proteomics with 30 isotope labeled reference peptides and the iBAQ quantification method [40] (for detailed description see Materials and Methods section). From these data, we selected additional antigens with a large range of abundances (Tab. 1).
To determine potential cross-protection between different serovars, we cloned the corresponding genes from Salmonella enterica serovar Typhi (except for OmpD which was obtained from serovar Typhimurium since it is absent in serovar Typhi). We expressed the proteins as C-terminal His6-fusions in E. coli followed by Ni-affinity chromatography purification. We purified the control antigen GFP-His6 using the same protocol.
We determined immune responses to these antigens in genetically resistant, convalescent mice that had survived infection with virulent Salmonella enterica serovar Typhimurium. We detected antigen-specific CD4 T cells in spleen using a sensitive CD154 assay [41] and measured serum IgG antibody responses using ELISA. All tested antigens were recognized by CD4 T cells (Fig. 1A; Tab. 1), many of which secreted IFNγ or IL-17 upon stimulation. Both cytokines play crucial roles in immunity to Salmonella [10]. Frequencies of responsive CD4 T cells were in the same range as for flagellin, which has been considered an immunodominant antigen [42]. These data suggested that Salmonella infection elicited a broad cellular immune response against a large number of in vivo expressed antigens from all Salmonella compartments in agreement with data observed for S. Typhi infected human patients [43]. There was no correlation between in vivo antigen abundance as determined by proteome analysis of ex vivo purified Salmonella, and CD4 T cell frequency or cytokine profile (Tab. 1).
Serum antibody responses revealed similar broad recognition of antigens from several Salmonella compartments (Fig. 1B) in agreement with previous data for human typhoid fever patients [24]–[26], [44]. Interestingly, the three immunodominant humoral antigens T2461, PhoN, and PcgL were all highly expressed in vivo (Tab. 1) suggesting a potential impact of antigen dose on antibody responses to Salmonella, although responses to minor antigens did not correlate with antigen abundance. PhoN has been previously recognized as an immunodominant antigen [26].
Many of the tested Salmonella antigens were capable to induce cellular and humoral immune responses. To test if these responses could confer protective immunity, we tested the 21 recombinant Salmonella antigens in immunization/challenge infection experiments in genetically susceptible BALB/c mice. Based on the results, we selected 16 additional Salmonella antigens primarily from the outer membrane, and tested them using the same experimental immunization/challenge approach (however, we did not measure their immunogenicity in convalescent mice). For simplicity, we discuss results for both antigen sets together. Out of 37 tested antigens, only few antigens enabled prolonged survival after oral challenge infection with virulent Salmonella compared to control immunization with the unrelated antigen GFP (Fig. 2; Tab. 1; poor survival of PhoN-vaccinated animals confirmed recently published data [26]). In fact, only two antigens (T0937 and T2672) mediated protective immune responses with P-values below 0.05 in our small experimental groups of only five mice per antigen. Replicate experiments with larger group sizes might yield statistical significant results for additional candidates such as SseB that has already been shown to be protective in two independent previous studies. Such experiments will be required to select individual antigens for vaccine development in future studies. On the other hand, the primary focus of this study was to identify antigen properties that correlate with protectivity. For this purpose, the somewhat noisy survival times detected with small animal groups were still helpful.
As an example, survival times did not correlate with CD4 T cell responses (Fig. 3A) or serum antibody levels (Fig. 3B) during natural infection of resistant mice. This could partially reflect differences in MHC class II haplotypes (H2d in BALB/c vs. H2b in 129/Sv), courses of infection, and potential differences in Salmonella biology in susceptible vs. resistant mice. However, a recent large-scale study reports comprehensive immunogenicity data for BALB/c mice and other mouse strains that had been immunized with attenuated Salmonella, as well as for human patients [26]. Several antigens that prolonged survival of immunized BALB/c mice after Salmonella challenge infection in our experiments, elicit detectable antibody responses in various mouse strains including BALB/c. However, none of these antigens was found to be immunodominant [26] and antibodies to antigens IroN and CirA with the longest survival times were not detected in this and previous studies. This could reflect differential antigen expression in virulent vs. attenuated Salmonella, different routes of administration, and/or differential expression at various stages of disease progression. Together, these data provide no evidence for immunodominance in convalescent or immune individuals as a prerequisite for protectivity.
Interestingly, in vivo expression levels also did not correlate with survival times (Fig. 4A). In fact, the two antigens that enabled the longest survival, IroN and CirA, had in vivo expression levels that were below our detection threshold. By comparison, antigens T2461 and PhoN were highly expressed in vivo and induced potent CD4 T cell and humoral responses in convalescent individuals, yet failed to prolong survival (in agreement with previous observations [26]).
In contrast to immunogenicity and in vivo abundance, antigen localization seemed to be crucial (Fig. 4B). In fact, antigens enabling prolonged survival times were exclusively associated with the Salmonella surface, either as experimentally validated outer membrane-associated lipoproteins [45], as outer membrane proteins, or as the translocon complex of the type III secretion system encoded by Salmonella pathogenicity island two (SPI-2) (Tab. 1). These data suggested distinct immune responses to Salmonella outer membrane/surface antigens that fundamentally differ from those to internal antigens.
On the other hand, surface localization alone was not sufficient for protectivity. As examples, membrane proteins PgtE, PagC, and Tsx were highly expressed in vivo and PgtE and PagC elicited potent CD4 T cell responses in convalescent individuals (Fig. 1A). PagC is also well recognized by antibodies and CD4 T cells of human typhoid fever patients [24]. However, PagC, PgtE, and Tsx failed to prolong survival. Interestingly, structural models revealed that these proteins were largely buried in the outer membrane bilayer (Fig. 5), and their extracellular loops contained at most one predicted CD4 T cell epitope each, and only up to two linear antibody epitopes, respectively. Importantly, key amino acids in exposed T cell epitopes differed among Salmonella serovars which might have impaired cross-protectivity of serovar Typhi antigens against serovar Typhimurium challenge infection. Similar observations were also made for non-protective TolC, OmpC, OmpD, and OmpF. By contrast, antigens IroN, CirA, and FepA that enabled extended survival after challenge infection, had extracellular loops with several highly conserved T and B cell epitopes (Fig. 5). Further studies with larger data sets will be required to validate the relevance of these structural properties for protectivity.
The strong bias for surface-associated Salmonella antigens might have been expected based on previous data for model antigens suggesting superior immunogenicity of surface antigens compared to internal antigens [30], [46]–[50]. However, these model antigen data were in striking contrast to results from us and others demonstrating comparable immune responses to autologous Salmonella antigens from all Salmonella compartments (Fig. 1A,B). Furthermore, there was no obvious correlation between immunogenicity and survival times (Fig. 3A,B).
To better understand these discrepancies between model antigens and autologous Salmonella antigens, we re-visited the impact of antigen localization using a well-characterized, sensitive model system in which a MHC II-restricted T cell epitope from ovalbumin comprising amino acids 319 to 343 (OVA) is recognized by adoptively transferred cognate T cell receptor transgenic CD4 T cells [51], [52]. We targeted the OVA epitope to different Salmonella compartments by fusing it to various proteins with known localization: GFP_OVA (cytosol [53]), OVA_MglB (periplasm [54]), Lpp_OVA (inner leaflet of the outer membrane [55]), and OVA_AIDA (outer leaflet of the outer membrane [56]) (Fig. 6A). To modulate expression levels, we used ribosome binding sites with differential translation initiation efficiency [12]. We expressed these fusion proteins from an in vivo inducible promoter [57] in an attenuated Salmonella enterica serovar Typhimurium aroA strain [58]. Antigen expression and localization was validated in in vitro cultures using cell fractionation followed by western blotting, trypsin treatment, and antibody binding (Fig. S2). Interestingly, small fractions of both outer membrane antigens LPP_OVA and partially processed OVA_AIDA were released to the extracellular surroundings when expressed at high levels (Fig. S2C) in agreement with previous findings for similar proteins [59]–[61].
We infected BALB/c mice with Salmonella strains by intragastric gavage of 1010 CFU. All Salmonella strains colonized intestinal Peyer's patches with peak tissue loads of 3×104 to 1.5×105 CFU at day seven post infection as observed before for attenuated Salmonella aroA [62]. All constructs stably maintained their respective ovalbumin-expression plasmids (>80% at 7 days post infection). To determine antigen-specific CD4 T cell induction, we adoptively transferred OVA-specific TCR-transgenic CD4 T cells one day prior to Salmonella infection. OVA-specific T cells upregulated the early activation marker CD69 and formed blasts in mice infected with Salmonella expressing ovalbumin model antigens, but not in mice infected with control Salmonella (Fig. 6B) as observed previously [53]. CD4 T cell induction kinetics were similar for all constructs and consistent with our previous observations [53] suggesting a response to Salmonella in situ antigen expression, but not to the inoculum [57], [63].
To compare T cell responses against the various Salmonella constructs, we measured T cell blast formation at peak Salmonella colonization at day seven post infection. Salmonella tissue loads varied somewhat between individual mice but for each construct, there was a linear relationship between the number of ovalbumin-specific DO11.10 blasts and Salmonella loads (Fig. 6C) in agreement with our earlier observations [57]. To determine the specific immunogenicity of each Salmonella strain, we calculated the average ratio of DO11.10 CD4 T cell blasts per viable Salmonella (i.e., the slopes in Fig. 6C) [57]. The data revealed comparable immunogenicity of model antigens GFP_OVA and OVA_MglB (Fig. 6D). In contrast, high-level expression of surface-associated LPP_OVA and OVA_AIDA induced superior responses that clearly surpassed responses even to saturating amounts [12] of internal GFP_OVA.
The OVA_AIDA fusion protein contained a fragment of the virulence factor AIDA from enteropathogenic E. coli and a cystein-deficient variant of the cholera toxin B subunit from Vibrio cholerae [64]. Both components might have stimulatory effects [65], [66] that could potentiate ovalbumin immunogenicity. To test this potentially confounding factor, we compared Salmonella expressing a suboptimal level of cytosolic GFP_OVA [12] (some 54.000 copies per Salmonella cell) to Salmonella expressing the same amount of GFP_OVA together with AIDA and cholera toxin B. Both strains induced DO11.10 T cell blasts with similar efficacy (Fig. 6D) suggesting that AIDA and cholera toxin B expression had no impact on the immunogenicity of Salmonella-encoded OVA.
Taken together, these findings suggested that antigens from all Salmonella compartments could induce specific CD4 T cell responses, but highly expressed outer membrane-associated antigens were clearly superior in agreement with previous observations in other model systems. However, these data were in striking contrast to responses to autologous Salmonella antigens (see discussion).
The fundamentally superior protectivity of surface-associated Salmonella antigens might reflect their unique accessibility to antigen processing and presentation in infected host cells in contrast to internal Salmonella antigens that are shielded by the Salmonella envelope, and thus remain invisible for the host immune system until Salmonella is damaged and the bacterial cell breaks open. To detect intact and damaged Salmonella in infected tissues, we used cytosolic GFP as a marker for internal antigens.
Salmonella expressing GFP from the chromosomal in vivo induced locus sifB were readily detected in infected tissue homogenates using flow cytometry [12] (Fig. 7A). Flow cytometric counts for GFP+ Salmonella closely correlated with viable counts as determined by plating (Fig. 7A, inset) suggesting that detectable GFP levels were present in all live Salmonella.
Confocal microscopy of infected spleen and liver sections revealed many particles that were stained by a polyclonal antibody to Salmonella lipopolysaccharide, had typical Salmonella size and morphology, and contained GFP (Fig. 7B) as previously observed [53] suggesting that these particles represented live intact Salmonella. In addition, we also detected numerous lipopolysaccharide-positive particles with distorted shapes that lacked detectable GFP (Fig. 7C), and likely represented killed and partially degraded Salmonella. Such particles were absent in non-infected control sections. Some Salmonella killing during acute infections had previously been proposed [67]–[69]. We observed some infected cells containing both intact and damaged Salmonella, but a large number of live Salmonella resided alone (or together with other live Salmonella) in infected cells with no detectable dead Salmonella. In such infected cells, internal Salmonella antigens were thus shielded and inaccessible for immune recognition.
There is an urgent medical need for an efficacious Salmonella vaccine with broad coverage of invasive serovars. One important bottleneck in the development of such a vaccine is the identification of suitable protective antigens. In this study, we identified broadly conserved S. Typhi antigen candidates that prolonged survival after S. Typhimurium challenge infection in the mouse typhoid fever model. The protectivity of some of these candidates should be confirmed with larger experimental groups to select the best antigen candidates for vaccine development in future studies.
Two siderophore receptors (IroN, CirA) enabled the longest survival (Tab. 1) consistent with previous studies that revealed siderophore receptors including IroN as promising vaccine antigens in other models [70]–[72]. Interestingly, siderophore receptors are induced by iron starvation and/or activation of the PhoPQ two component sensory system [73]. IroN and CirA induction could thus contribute to increased protective efficacy of membrane preparations from iron-starved Salmonella [74], or live attenuated Salmonella phoQ24 with constitutive hyperactivation of the PhoP response regulator [75].
On the other hand, all identified antigens still provided at most partial protection against challenge infection with virulent Salmonella suggesting a need for additional antigens. Unfortunately, protective Salmonella antigens might be rather rare as even among the 37 tested in vivo expressed antigens that were all highly immunogenic during infection, only a small minority enabled prolonged survival. OmpC, OmpD, and OmpF were previously proposed as potential protective antigens based on data obtained for enriched Salmonella membrane preparations. However, all three antigens failed to protect in our model. This could reflect higher stringency of our model (challenge infection with virulent Salmonella vs. highly attenuated mutant Salmonella), denatured three-dimensional structures of our recombinant antigen preparations vs. native antigens, and/or presence of undetected minor protective antigens (such as IroN and CirA) besides OmpC, OmpD, and OmpF in the previously used outer membrane antigen preparations.
Additional protective Salmonella antigens could be identified by comprehensive immunization/challenge experiments, but this would require extensive animal experimentation. Antigen priorization using relevant antigen properties could help to narrow down the number of antigen candidates to more practical numbers. Unfortunately, some previously proposed antigen properties seemed to have limited relevance for protectivity in our model. This included Salmonella in vivo expression levels, sequence-based antigenicity predictions, and immunodominance in convalescent individuals. Poor correlation of antigen immunodominance with protective efficacy has also been observed in tuberculosis [17]. On the other hand, immune recognition in convalescent individuals can still provide valuable information about antigen expression during at least some stage of infection that might be difficult to obtain otherwise [23], [29]. Such data thus could greatly help to prioritize antigen candidates [26].
In contrast to antigen abundance and immunodominance, surface-association appeared to be an essential prerequisite. Surprisingly, some surface-associated proteins that enabled prolonged survival also included lipoproteins which were likely to reside in the inner leaflet of the outer membrane facing the internal periplasmic space with no exposure to the outside. It is possible that some lipoproteins might flip across the outer membrane as observed for other Gram-negative bacteria [76]. Moreover, some lipoprotein fraction is constantly released to the outside through outer membrane vesicle shedding [59], [60].
Several mechanisms could contribute to the striking superiority of surface-associated antigens. Antibody responses are important for full protection against virulent Salmonella [10], and protective antibody responses must be directed against surface antigens [9]. On the other hand, CD4 T cells are even more important for immunity to Salmonella at least in the mouse typhoid fever model [10], and it is unclear why CD4 T cells should respond to surface-associated antigens in a fundamentally different way compared to the much larger number of internal antigens.
In fact, early cell culture experiments suggested no impact of Salmonella antigen localization on CD4 T cell recognition of infected cells [77]. However, in this study a large amount of antigen was already present in the inoculum, and rapid killing of the majority of phagocytosed Salmonella [78] would have released this antigen from all Salmonella compartments. Several subsequent in vivo studies suggested that surface-associated model antigens might have intrinsically higher immunogenicity compared to internal model antigens [30], [46]–[50]. However, the various model antigen targeting constructs could have differed in antigen in vivo expression levels, antigen stability, and epitope processing. Fusion partners could also have direct immunomodulatory effects. We therefore re-visited this issue and tried to control some of these factors. Our results clearly supported the previous finding of superior immunogenicity of highly expressed surface-associated model antigens in Salmonella.
In surprising contrast to these data from model antigens, however, humoral and cellular immune responses in Salmonella-infected convalescent mice did not show any bias for surface-associated autologous Salmonella antigens in this as well as in a recent large-scale study [26]. Broad recognition of antigens from all pathogens compartments has also been observed in Salmonella Typhi-infected or Chlamydia-infected human patients [24], [25], [29], [43], [44]. Model antigens and autologous antigens were also discordant with respect to the impact of antigen abundance. Specifically, our data for ovalbumin model antigens in this and a previous study [12], as well as similar findings for Mycobacterium bovis BCG overexpressing Ag85b [19], suggested that high in vivo expression levels enhance antigen immunogenicity. However, for autologous Salmonella antigens in vivo expression levels did not correlate with protectivity. Striking discrepancies between results for model antigens vs. autologous antigens have also been observed in other pathogens [38].
Some of the discrepancies could reflect technical issues. In particular, strong expression of foreign surface model antigens might induce subtle alteration in Salmonella in vivo properties such as increasing outer membrane vesicle shedding or alterations in protein secretion that could affect antigen presentation and immune recognition. Furthermore, model antigens might not be representative of autologous antigens that may have been shaped by host/pathogen co-evolution selecting for weak immunogenicity. Regardless of the actual causes of these discrepancies, our data indicated that in contrast to evidence from model antigens, protective Salmonella surface-associated antigens were not more immunogenic compared to internal antigens.
As an alternative explanation, surface-associated antigens might become more rapidly available for immune recognition compared to internal antigens that are only released after some pathogen damage. This could be relevant since early immune responses might facilitate infection control [32]. In the mouse typhoid fever model, however, a detectable fraction of Salmonella is rapidly killed early during infection as observed in this and previous studies [67], [69] similar to events during Mycobacterium infection [79]. Consistent with these observations, CD4 T cell induction kinetics in the ovalbumin model system were similar for Salmonella strains with internal or surface-associated OVA-expression.
Instead, we propose an alternative explanation based on the observation that many live Salmonella resided alone, or together with other live Salmonella, in infected host cells with no dead Salmonella releasing their internal antigens. As a consequence, Salmonella internal antigens remained inaccessible for antigen processing and presentation in these cells. In contrast, surface-exposed Salmonella antigens, or antigens released by outer membrane vesicle shedding, could be accessible for processing and presentation to cognate CD4 T cells for initiation of protective anti-Salmonella effector mechanisms (Fig. 8). In comparison, CD4 T cells recognizing internal Salmonella antigens would have limited impact on infection control because they miss many cells containing live Salmonella and instead direct their responses to host cells containing already dead Salmonella. According to this model, surface-associated antigens thus differ fundamentally from internal antigens because they are uniquely accessible in host cells containing only live Salmonella.
Surface-associated/secreted antigens have been shown to be crucial for CD8 T cell-dependent immunity to Listeria infection [31], [33]. Our data suggested that such antigens might also be crucial for CD4 T cell mediated immunity to Salmonella and potentially other intracellular pathogens. Interestingly, some internal antigens have been shown to confer partial protection in infectious diseases caused by intracellular pathogens such as Leishmania [38] and Mycobacterium [37]. In these infections live and dead pathogens often co-occur in the same host microenvironments [80], [81] suggesting that both internal and surface-associated antigens might be available for T cell recognition and initiation of antimicrobial immune effector mechanisms targeting both live and already dead pathogens [82]. We speculate that full protection might still require immune detection of all live pathogens including those that reside in microenvironments with yet no accessible internal antigens from dead pathogens. Further studies are required to test this hypothesis.
This study suggested novel Salmonella antigens that conferred partial protection against virulent Salmonella in a stringent typhoid fever model. High sequence conservation among relevant Salmonella serovars and cross-protection of serovar Typhi antigens against serovar Typhimurium challenge infection, suggested that some of these antigens might help to pave the way for a broadly protective vaccine against systemic Salmonella infection. In addition, our findings suggested that surface-associated antigens might represent particular promising antigens for both humoral and cellular immunity to Salmonella, since recognition of surface antigens uniquely enables detection and destruction of live Salmonella in relevant host microenvironments. This crucial importance of antigen localization could facilitate discovery of additional protective antigens for Salmonella and potentially other intracellular pathogens.
All animal experiments were approved (license 2239, Kantonales Veterinäramt Basel-Stadt) and performed according to local guidelines (TschV, Basel) and the Swiss animal protection law (TschG).
Antigens were PCR-amplified from Salmonella enterica serovar Typhi Ty2 (or Salmonella enterica serovar Typhimurium SL1344 [58] for ompD), cloned as His6-fusions by conventional ligation into pET22b, or by Enzyme Free Cloning into plasmid pLICHIS [83], and overexpressed in E. coli BL21. GFP_His6 was cloned as control antigen. Antigens were purified from washed inclusion bodies using immobilized metal ion affinity chromatography (Protino Ni TED 1000, Macherey Nagel) followed by ion exchange chromatography (Ion exchange spin columns, Pierce Thermo Scientific, cationic or anionic resins depending on antigen isoelectric point).
Salmonella expressing the green fluorescent protein (GFP) were sorted infected using flow cytometry as described [22]. Preparation of tryptic peptides and analysis by LC-MS/MS was done essentially as described [84] with some modifications. Given the limited sample material Protein LoBind tubes and pipette tips (Axygen) were used throughout the procedure. Frozen FACS sorted Salmonella pellets were resuspended in 15 µl lysis buffer (100 mM ammonium bicarbonate, 8 M urea, 0.1% RapiGest) and sonicated for 2×30 seconds. The released proteins were reduced and alkylated, and first digested for 4 hrs with sequencing grade LysC peptidase (10 ng/µl; Promega) before overnight trypsin digestion. The detergent was cleaved by adding 2M HCL and 5% TFA to final concentrations of 50 mM and 0.5% respectively, and incubating for 45 min at 37°C. Prior to centrifugation to remove the cleaved detergent (14,000×g, 10 min, 4°C) a mixture containing 32 heavy labeled reference peptides were added to the samples (5*10−5 fmoles per Salmonella for expected “high” abundance proteins, 5*10−6 fmoles per Salmonella for expected “low” abundance proteins; Tab. S1). The recovered peptides were desalted on C18 reverse-phase spin columns (Macrospin columns, Harvard apparatus), dried under vacuum and subjected to LC-MS/MS using an LTQ-Orbitrap-Velos instrument (Thermo-Fischer Scientific). The amount of material analyzed in a single shot in the MS depended on the infection load, and corresponded to peptides derived from between 5*105 and 2*106 sorted Salmonella, plus contaminating mouse material which escaped detection in the cell sorter [22]. We analyzed samples from seven independently infected mice. In order to increase the number of Salmonella protein identifications, MS-sequencing was focused on previously identified peptides from Salmonella using the recently developed inclusion list driven workflow [84]. Each sample was analyzed twice in succession in the MS to verify technical reproducibility. Peptides and proteins were database searched against a decoy database consisting of the SL1344 genome sequence (ftp://ftp.sanger.ac.uk/pub/pathogens/Salmonella/), GFP_OVA, 204 frequently observed contaminants, all mouse entries from SwissProt (Version 57.12), and all sequences in reversed order (total 42502 entries) using the Mascot search algorithm. The search criteria were set as follows: full tryptic specificity was required (cleavage after lysine or arginine residues); 2 missed cleavages were allowed; carbamidomethylation (C) was set as fixed modification; oxidation (M) as variable modification. The mass tolerance was set to 10 ppm for precursor ions and 0.5 Da for fragment ions. The false discovery rate was set to 1% for protein and peptide identifications. In addition to Salmonella proteins a substantial number of mouse proteins were identified in the samples as previously noted [22]. Absolute quantities were determined for those 18–20 “anchor” proteins that were detected along with a corresponding labeled AQUA peptide using the Trans-Proteomic Pipeline (TPP,V4.4.0). We then used the iBAQ method [40] to establish absolute quantities of all remaining protein identifications, with a linear model error of between 47 and 60%.
Translational fusions of the ovalbumin peptide containing amino acids 319 to 343 to various proteins with differential targeting in the Salmonella cell were constructed by PCR cloning. All fusion genes were cloned into a pBR322-derived plasmid backbone [53] downstream of a Salmonella genome fragment containing the in vivo inducible pagC promoter [57] and ribosomal binding site 1 (AAGAA) or 2 (AGCAG) for low or high translation initiation efficiencies [12]. To generate ova_aida, coding sequence for the ovalbumin peptide (ova) was inserted between the signal peptide derived from cholera toxin B and the HA tag in plasmid pLAT260 [85]. A control plasmid coding for CTB_AIDA and GFP_OVA was also constructed. To generate lpp_ova, lpp without the C-terminal lysine codon that can cross-link to peptidoglycan [86], was amplified from E. coli DH5α and fused with ova and a C-terminal HA tag. To generate ova_mglB, mglB gene without the signal peptide sequence was amplified from E. coli DH5α and fused with a ctB signal sequence followed by ova and the HA tag. The construction of gfp_ova has been described [87]. The various plasmids were transformed into attenuated Salmonella enterica serovar Typhimurium aroA SL3261 [58].
Ovalbumin expression was assessed by western blotting with a polyclonal antibody to ovalbumin (Sigma) that recognizes the OVA peptide comprising amino acids 319 to 343 [53]. Salmonella outer membranes were prepared by extraction with L-lauryl sarcosinate as described [85]. Periplasm was prepared by chloroform extraction as described [88]. Culture supernatants were sterile filtered (0,2 µm pore size) and subjected to TCA precipitation [89]. To assess ovalbumin surface accessibility, intact or lysed Salmonella cells were treated with 50 µg ml−1 trypsin at 37°C for 10 min. In addition, Salmonella were stained with an antibody to the HA tag, and examined by fluorescence microscopy.
Female 8 to 12 weeks old 129/Sv mice were obtained from Charles River. Mice were orally infected with 109 CFU Salmonella enterica serovar Typhimurium SL1344 [58] from late log cultures using a round-tip stainless steel needle. Control mice were sham-infected. Mice were sacrificed 6 months after infection. Splenocytes were isolated and tested for antigen-specific CD4 T cell responses as described [41]. Unstimulated T cells from convalescent mice as well as antigen-stimulated T cells from naïve control mice showed only weak background responses (Fig. S1). Some antigens gave also weak responses for T cells from convalescent mice (depending on the individual mouse). Together, these data suggested that antigen-nonspecific background responses to E. coli contaminants that might have been present in trace amounts in our antigen preparations did not result in unspecific T cell responses in our assay. Plasma was tested for antigen-specific IgG responses using ELISA with an IgG calibration curve for absolute quantification.
Female, 8 to 12 weeks old BALB/c mice were obtained from Charles River. Groups of 5 mice were immunized subcutaneously with 10 µg antigen emulsified in complete Freund's adjuvant followed by a second immunization with incomplete Freund's adjuvant four weeks later. After additional four weeks, mice were orally infected with 6×105 CFU Salmonella enterica serovar Typhimurium SL1344 [58] from late log cultures using a round-tip stainless steel needle. Infected BALB/c were monitored twice daily and sacrificed when moribund.
BALB/c and DO11.10 mice [51] were bred in the Bundesamt für gesundheitlichen Verbraucherschutz und Veterinärmedizin (Berlin, Germany) under specific-pathogen free conditions. Adoptive transfer of 4×106 DO11.10 T cells into syngenic age- and sex-matched BALB/c mice was performed one day before infection as described [53]. For infection, attenuated Salmonella strains carrying expression cassettes for various ovalbumin fusion proteins were grown to late log phase and harvested. Bacteria were washed twice and resuspended in LB containing 3% sodium bicarbonate. Doses containing ca. 1010 cfu in 100 µl were administered intragastrically to chimeric mice with a round-tip stainless steel needle. At various time points post infection, mice were anesthetized and sacrificed. DO11.10 T cell blast formation was determined by flow cytometry as described [53]. Aliquots of the same Peyer's patch preparations were treated with 0.1% Triton x-100 to release intracellular Salmonella for CFU determination by plating, and for quantitation of GFP_OVA in vivo expression levels by two-color flow cytometry as described [87]. Many TCR tg models show substantial clonal expansion upon antigen stimulation. However, in our Salmonella model we observe only weak and variable accumulation of tg CD4 T cells in infected tissues which might reflect the fact that even at peak Salmonella loads only about 1 ng antigen is present [87]. Instead, blastogenesis as measured by CD69 upregulation and increased forward scatter provides a sensitive antigen-specific readout.
BALB/c mice with Salmonella loads of 106 to 107 in spleen and liver were sacrificed. 10 µm cryosections were stained with polyclonal rabbit antibodies to Salmonella lipopolysaccharide (SIFIN) and anti-CD68 (abcam) followed by Alexa 546-conjugated goat anti-rabbit and Alexa 647-conjugated goat anti-rat antibodies (Invitrogen). Sections were examined by confocal microscopy (Leica, SP5).
Structural models for selected Salmonella outer membrane antigens based on solved structures of homologues were obtained from SWISS-MODEL [90] available at http://swissmodel.expasy.org. Linear B-cell epitopes were predicted using FBCPred [91] available at http://ailab.cs.iastate.edu/bcpreds/predict.html using an epitope length of 14 and 90% specificity. Peptides that bind to MHC II I-Ad and/or I-Ed were predicted using RANKPEP [92] available at http://imed.med.ucm.es/Tools/rankpep.html with a binding threshold yielding 85% sensitivity for detection of well-defined epitopes in MHCII haplotype databases (the default setting of RANKPEP).
|
10.1371/journal.pntd.0006107 | Seroprevalence of rickettsial infections and Q fever in Bhutan | With few studies conducted to date, very little is known about the epidemiology of rickettsioses in Bhutan. Due to two previous outbreaks and increasing clinical cases, scrub typhus is better recognized than other rickettsial infections and Q fever.
A descriptive cross-sectional serosurvey was conducted from January to March 2015 in eight districts of Bhutan. Participants were 864 healthy individuals from an urban (30%) and a rural (70%) sampling unit in each of the eight districts. Serum samples were tested by microimmunofluorescence assay for rickettsial antibodies at the Australian Rickettsial Reference Laboratory.
Of the 864 participants, 345 (39.9%) were males and the mean age of participants was 41.1 (range 13–98) years. An overall seroprevalence of 49% against rickettsioses was detected. Seroprevalence was highest against scrub typhus group (STG) (22.6%) followed by spotted fever group (SFG) rickettsia (15.7%), Q fever (QF) (6.9%) and typhus group (TG) rickettsia (3.5%). Evidence of exposure to multiple agents was also noted; the commonest being dual exposure to STG and SFG at 5%. A person’s likelihood of exposure to STG and SFG rickettsia significantly increased with age and farmers were twice as likely to have evidence of STG exposure as other occupations. Trongsa district appeared to be a hotspot for STG exposure while Punakha district had the lowest STG exposure risk. Zhemgang had the lowest exposure risk to SFG rickettsia compared to other districts. People living at altitudes above 2000 meters were relatively protected from STG infections but this was not observed for SFG, TG or QF exposure.
This seroprevalence study highlights the endemicity of STG and SFG rickettsia in Bhutan. The high seroprevalence warrants appropriate public health interventions, such as diagnostic improvements and clinical treatment guidelines. Future studies should focus on vector profiles, geospatial, bio-social and environmental risk assessment and preventive and control strategies.
| Rickettsial infections including scrub typhus and Q fever are not widely recognised in Bhutan although the country is situated in an endemic Asian region. With two recorded outbreaks, scrub typhus is known to occur but other rickettsial infections and Q fever have not been recorded. In this first seroprevalence study of rickettsial infections, an overall seroprevalence of 49% was detected against rickettsioses amongst 864 participants. Seroprevalence was highest against scrub typhus group (STG) (22.6%) followed by spotted fever group (SFG) rickettsia (15.7%), Q fever (QF) (6.9%) and typhus group (TG) rickettsia (3.5%). Evidence of exposure to multiple agents were also noted; the commonest being dual exposure to STG and SFG at 5%. A person’s likelihood of exposure to STG and SFG significantly increased with age and farmers were twice as likely to have evidence of STG exposure as other occupations. Trongsa district in central Bhutan appeared to be a hotspot for STG exposure. People living at altitudes above 2000 meters were relatively protected from STG infections but this was not observed for SFG, TG and QF exposure. The findings from this study may direct future research on rickettsioses in Bhutan. These neglected tropical diseases warrant specific public health interventions in Bhutan.
| Rickettsial infections including scrub typhus and Q fever are usually referred to as rickettsiosis [1]. Rickettsioses are zoonotic infections transmitted to humans through bites of infected ticks, fleas, lice and mites or through aerosols generated during exposure to infected placentas and birth fluids of mammals in the case of QF [2]. The family Rickettsiaceae includes two genera, Rickettsia and Orientia, which include many human pathogens some of which cause lethal infections with up to 30% mortality without treatment [3, 4]. The genus Rickettsia has more than 20 species making up several groups among which the spotted fever group (SFG) and typhus group (TG) are established human pathogens [4, 5]. The SFG rickettsia includes the etiologic agents of Rocky Mountain spotted fever (R. rickettsii) and Mediterranean spotted fever (R. conorii) and many others. The TG rickettsia include agents of epidemic (R. prowazekii) and endemic (R. typhi) typhus [4]. Orientia has two species; O. tsutsugamushi and O. chuto [6], together forming the scrub typhus group (STG). Coxiella burnetii is the causal agent of Q fever (QF). Of all the methods to detect rickettsial infections, antibody detection by serology is the most commonly used, microimmunofluorescence assay (IFA) being the currently accepted gold standard [7]. After an infection, IgM can be detectable for months and IgG for years [7, 8].
SFG and TG rickettsia occur worldwide and are a significant cause of morbidity in south-east Asia [9]. STG was originally thought to be confined to the Asia-Pacific region but now has been reported from the Middle East [6], Africa [10, 11] and South America [12]. Q fever has a worldwide distribution [13] except New Zealand [14] although fears of its introduction have been raised [15]. Rickettsioses are both emerging and re-emerging infections [16, 17]. Despite being endemic in Asia and causing significant burden to public health, true prevalence studies of these infections are limited. In India, rickettsial diseases including scrub typhus have been documented in several states from all parts of the country [1]. A seroepidemiology study in northeast India, in areas bordering Bhutan reported a sero-prevalence of 30.8%, 13.8% and 4.2% against STG, SFG and TG respectively [18]. In Darjeeling, another Indian district near Bhutan, a 2005 study reported an overall incidence of STG at 34 cases/100,000 population/pa, varying from 2 cases/100,000 population in July to 20/100,000 population in September and decreasing to zero in December [19]. Q fever has been under-reported from India and recent data are lacking [20]. A Chinse study reports an overall Q fever prevalence of 10% and highlights it as an under-reported and underdiagnosed illness [21].
Although situated in the endemic Asia Pacific region, Bhutan has reported scrub typhus cases only since 2009 [22] and SFG, TG and QF have not been reported to date. Rickettsial diseases (excluding Q fever) have been included in the national notifiable diseases since 2010 with increasing reports, mostly scrub typhus, from 118 cases in 2011 to 605 cases in 2015. Despite the increasing notifications and improving awareness, there are currently no clinical guidelines on management of rickettsial infections in Bhutan, and awareness needs improving. There are no reports of Q fever in Bhutan owing to the lack of diagnostic facility both in the human and animal sector at present. Therefore, a serological investigation was undertaken to determine the seroprevalence of rickettsial infections including QF in Bhutan.
Bhutan is composed of 20 districts and 205 subdistricts with an estimated population of 770,000 in 2016 [23]. The Bhutan national census in 2005 reported on 1044 rural villages/chiwogs and 311 urban towns as primary sampling units (PSUs). Population density in different districts vary between 9–64 people/km2 [23]. For this study, the 20 districts were stratified into four regions; eastern (5 districts), central (4 districts), western (5 districts) and southern (6 districts) as defined by the Bhutan National Statistical Bureau (NSB) [23] for their national surveys. From each region, two districts were selected with a probability proportionate to size (PPS) method, selecting eight of twenty districts for the study (Fig 1). A rural and an urban PSU were selected from each district by the same PPS method. To assess the influence of altitude on exposure, altitude of places were arbitrarily grouped into low (<1000 meters), medium (1000–2000 meters) and high altitude (>2000 meters).
This descriptive cross-sectional sero-survey was carried out from January to March 2015, during the dry winter and early spring season. The sample size was calculated using a multi-stage cluster sampling method. Persons <13 years were excluded due to the possible risk of complications during blood sampling in remote areas. The sample size needed to estimate the number of households to be surveyed with a 95% confidence interval and other assumptions (50% prevalence rate, 0.05 margin of error, a design effect of 2 and an expected rate of participation of 90%) was calculated to be 864. Based on Bhutan’s urban-rural population proportion of 30:70 [23], 30% of the participants were taken from urban and the remaining from rural settings; therefore, of the 864 households selected, 256 were from urban and 608 from rural settings. Each of the eight selected districts contributed 108 households (76 rural and 32 urban households). The households were taken from the household list with unique identification numbers developed during the previous national surveys (National Health Survey 2012 and NCD STEPS Survey 2014). When a selected PSU had a lesser number of households than required, a nearby PSU was added. After selection of the household, all eligible members (≥13 years) present in the house were listed and one member was selected for the study through a lottery system. After selection, written consent was obtained; demographic details and environmental exposure history were taken by trained laboratory personnel through a face-to-face interview and blood samples were collected. Serum was extracted and stored at -70°C until shipment to Australia.
Serum samples were shipped at room temperature to the Australian Rickettsial Reference Laboratory (ARRL) [24], a nationally accredited laboratory for rickettsial testing, where serological testing was carried out by indirect microimmunofluorescence assay (IFA) [25]. Antibodies against SFG rickettsia were individually tested using R. australis, R. honei, R. conorii, R. africae, R. rickettsii and R. felis antigens; TG rickettsia using R. prowazekii and R. typhi antigens; STG using O. tsutsugamushi (Gilliam, Karp and Kato strains) and O. chuto antigens, and QF using C. burnetii phase I and phase II antigens. Samples were initially screened at low dilutions and titrated to end-point (titre) when positive. With slight modification from the usual ARRL interpretation criteria [24, 25], antibody titres of ≥1:256 for IgG and/or ≥1:1024 for IgM against any of the SFG, TG and STG antigens were considered positive for the rickettsial group agents. Similarly an antibody titre of ≥1:50 for IgG or IgA and ≥1:100 against IgM against C. burnetii phase I or II or both were considered positive for Q fever. Positive and negative control wells were included in each slide during testing.
Data were entered into an Excel spreadsheet and analysed using STATA software version 14. Chi-squared or Fischer’s exact test was used to explore the association between seropositivity and study variables considering p values of ≤0.05 significant. Univariate logistic regression was used to determine crude odds ratio (COR) and p values. All variables with p values 0.2 or less in the univariate analysis were taken for multivariate logistic regression to determine adjusted OR (AOR) and corresponding p values of <0.05 considered significant.
The study was approved by the Bhutan Research Ethics Board of Health (REBH) (Ref: REBH/Approval/2014/019) and the Human Research Ethics Committee (HREC), University of Newcastle, Australia (Ref: H-2016-0085). All individuals or parent/guardian provided written consent before participation.
A total of 864 participants were enrolled from the eight districts and all selected candidates consented to the study. There were 345 (39.9%) males and the mean age of participants was 41.1 (range 13–98) years. Most participants belonged to the age group of 26–40 years. Farmers 414 (47.9%) were the highest group by occupation (Table 1).
In seropositive participants, most had IgG antibody titres of 1:256 or 1:512 against STG, SFG and TG rickettsia and titres of 1:100 or 1:200 against Coxiella phase II IgG, IgA or IgM. A very small number of high antibody titres of up to 1: 2048 in STG (≈3%) and SFG (≈ 0.1%) were seen in participants. Overall, the most prevalent rickettsial infection was STG (22.6%) and the least prevalent was TG rickettsia (3.5%). Evidence of past exposure to multiple agents was also seen; the commonest being dual exposure to SFG and STG (5%) (Fig 2).
Seropositivity rates were not significantly different between males and female for all four infectious agents. The prevalence of each infection, especially STG and SFG, appeared to increase with age and farmers exhibited the highest seropositivity rates.
Thirty percent (256) of the participants were from urban areas. Among all participants, 550 (63.7%) reported having animal contact and almost half (426) had pets at home. In addition, 620 (71.8%) reported contact with vegetation and forest during their daily activities, 205 (24.3) recollected suffering from febrile illness in the past, 337 (40.8%) had past tick bites, 153 (18.0%) had an eschar in the past and 202 (23.6%) had past flea bites.
Many of the demographic and environmental variables showed significant baseline correlation with seropositivity against each infection in Chi-squared or Fisher’s exact test (Table 2) and a few were statistically significant in the logistic regression analysis.
The comparative seropositivity in the urban and rural sampling units of the eight districts and the overall national prevalence of all four infections are shown in Fig 3 and the estimated proportion of each infection at different sampling units (urban and rural) of the eight districts presented in Fig 4. Significant epidemiological factors and seropositivity are shown for STG (Table 3) and SFG (Table 4). No factors showed association with QF or TG rickettsia seropositivity, likely due to the small number of seropositives.
The prevalence of STG seropositivity increased with age. The odds of exposure to STG infection was significantly higher in farmers compared to other occupations. Punakha district had the lowest risk of exposure to STG infections while people living in Trongsa district were three times more likely to be infected as those in other districts. Contact with domestic animals more than doubled the odds of exposure to STG. People residing at high altitude had 89% lower odds of being exposed to STG compared to those residing at lower altitude (AOR 0.11, p 0.002, 95% CI 0.03, 0.44), (Table 3).
SFG rickettsial seropositivity prevalence also increased with age. A person over 55 years of age was three times more likely to have been exposed to SFG than the younger age groups. Compared to other districts, Zhemgang district had a significantly lower odds of exposure to SFG rickettsia (AOR 0.34, p value = 0.010, 95% CI 0.15, 0.77). Altitude did not affect the prevalence of SFG (Table 4).
This study revealed an overall seroprevalence of 48.7% against rickettsioses in Bhutan with the highest prevalence to scrub typhus (22.6%) followed by SFG rickettsia (15.7%), Q fever (6.9%) and TG rickettsia (3.5%). Evidence of past exposure to two or more rickettsial agents was seen in 11.1% of the participants depicting probable dual or multiple infections in an endemic setting or possibly cross-reacting antibodies. This is the first seroprevalence study on rickettsioses in Bhutan and may be used as baseline data for subsequent studies in this country although it is recognised that prevalence estimates may vary when measured at different times of the year. The limitations of the findings from the exclusion of children (<13 years old) should be borne in mind. This was required to avoid risks of complications during blood sampling especially in remote areas where medical assistance is hard to obtain. Unintended but unavoidable exclusion of potential participants could have also occurred due to a member of the household being away from home during sampling. Information on past fevers, tick bites and eschar might have had drawbacks due to participants failing to comprehend technical terms. In addition, inter-district, urban versus rural as well as high versus low altitudes comparisons was not precise due to the highly variable landscape within and between districts. Unavailability of adequate local data on climatic conditions, environmental and geospatial information at primary sampling unit (urban and rural) level made it impossible to explain inter-district differences of exposure to the infections.
The findings in this study were similar to a seroprevalence study in north-east India, that reported the highest seroprevalence against STG (30.8%) followed by SFG (13.8%) and TG (4.2%) [18]. The similarity is noteworthy due to the proximity of these areas to Bhutan. Similar occurrences of these infections in neighbouring countries may benefit from coordinated cross-border prevention and control activities.
The odds of exposure sequentially increased with increasing age of participants in case of STG and SFG rickettsiae. This mirrors the situation in endemic areas where increasing number of people would be exposed to the infections as they advance through life leading to an accumulation of older seropositive people in the community. In south-east Asia, murine typhus was reported more in urban dwellers, while STG and SFG were more prevalent in rural dwellers [9]. However, in Bhutan, this study did not find any significant differences in any of the four infections between urban and rural residents probably reflecting similar environmental conditions between the two populations. This is supported by finding no significant differences between occupational groups for all infections, with the exception of STG where farmers had higher seropositivity rates compared to other occupations. There were differences between a few districts for STG and SFG infections, highlighting hot spots for these two infections. Trongsa district in central Bhutan appeared to be a hotspot for STG infection and Punakha district exhibited significantly low odds of exposure. STG exposures were significantly low amongst participants in high altitude areas. This may be explained by cold weather at high altitude areas not favouring mite survival. Of all the districts, Zhemgang in south-central Bhutan had the lowest odds of exposure to SFG rickettsia. Unlike STG infections, altitude had no effect on SFG, TG and QF exposure. This could be due to different tick species at different altitudes transmitting different infections. Expanding primary and secondary clusters of STG infections were also reported in China [26]. Such clusters or hotspots would benefit from focused public health interventions especially where resources are limited as in the case of Bhutan. Targeting prevention and control activities in hotspot areas could be effective and cost saving.
Antibody titres of 1:256 or 1:512 were the commonest observed antibody levels amongst the participants. A small number of participants with higher antibody titres of 1:1024 or 1:2048 may have been due to recent infections (symptomatic or asymptomatic) or due to recurrent subacute exposures stimulating antibody production. Cross-reactions between antibodies within the rickettsial species, especially between SFG and TG rickettsia, are known to occur. Therefore, persons with mixed antibodies may not necessarily have had multiple infections but may be due to cross-reacting antibodies resulting from the one infection. This could also explain some of the observed multi-species exposures. Background antibody levels in endemic situations are known to interfere with serological diagnosis of acute infections especially with rapid point-of-care diagnostics [17]. This is worthy of note in the Bhutanese setting where only rapid point-of-care diagnostics are available currently. There is urgent need to improve diagnostic facilities in Bhutan to provide more specific assays such as the microimmunofluorescence assay and molecular diagnostics, especially in the main centres. Point-of-care diagnostics could still be useful in the smaller districts and remote health centres for ease of use.
Rickettsioses have been associated with poor maternal and neonatal outcomes including stillbirth and low birth weights [27, 28] in endemic situations. The role of rickettsioses in the high maternal and neonatal mortality and morbidities in Bhutan deserves to be studied. Scrub typhus has also been known to involve the central nervous system manifesting as meningoencephalitis [29–31] and STG has been recently reported as a significant cause of encephalitis in northeast India [32]. This is important in the Bhutanese context in light of establishing the causes of acute encephalitis syndromes including meningococcal infection, Japanese encephalitis and other viral meningitis syndromes which are poorly understood at present. Documented deaths in Bhutan due to proven scrub typhus had resulted from meningoencephalitis and gastrointestinal perforation [22], which are known to be severe complications of scrub typhus. Understanding these occurrences in endemic areas could be helpful in averting preventable deaths from such complications.
Rickettsioses are important causes of illnesses in international travellers [33]. In a study on the spectrum of illness amongst ill returned travellers from six GeoSentinel sites, rickettsial infections were significant causes with 17 and 32 patients/1000 cases returning from south central and south-east Asia respectively [34]. Bhutan is an emerging destination for international travellers involving in activities like camping, trekking, cultural and rural home-stays. The high prevalence of rickettsioses could potentially expose travellers to these infections. Therefore travellers should be aware of the risk and become educated on preventive measures. In addition, educating travellers would keep them vigilant for any febrile illnesses during travel or upon returning to their home countries, enabling them to provide a detailed travel history and for their treating doctor to include rickettsial infections in their differential diagnoses.
There are limited prevalence studies on rickettsioses in south-east Asia [9]. Studies are even scantier in south and central Asia including Bhutan where most published studies were focused on clinical cases and acute febrile patients. Therefore, a prevalence study of these neglected but re-emerging infections in these endemic areas should be carried out with active regional collaborations and participations. This first seroprevalence study in Bhutan highlighted the endemicity of rickettsioses especially STG and SFG rickettsia. Findings on TG rickettsia and Q fever should be interpreted with caution due to the detection of fewer positive cases. This high rickettsial seroprevalence needs attention from the Bhutan Ministry of Health such as appropriate public health interventions, diagnostic improvement and clear clinical treatment guidelines. Future studies should focus on vector profiles, geospatial, bio-social and environmental risk assessment and preventive and control strategies formulation.
|
10.1371/journal.ppat.1002105 | Low CCR7-Mediated Migration of Human Monocyte Derived Dendritic Cells in Response to Human Respiratory Syncytial Virus and Human Metapneumovirus | Human respiratory syncytial virus (HRSV) and, to a lesser extent, human metapneumovirus (HMPV) and human parainfluenza virus type 3 (HPIV3), can re-infect symptomatically throughout life without significant antigenic change, suggestive of incomplete or short-lived immunity. In contrast, re-infection by influenza A virus (IAV) largely depends on antigenic change, suggestive of more complete immunity. Antigen presentation by dendritic cells (DC) is critical in initiating the adaptive immune response. Antigen uptake by DC induces maturational changes that include decreased expression of the chemokine receptors CCR1, CCR2, and CCR5 that maintain DC residence in peripheral tissues, and increased expression of CCR7 that mediates the migration of antigen-bearing DC to lymphatic tissue. We stimulated human monocyte-derived DC (MDDC) with virus and found that, in contrast to HPIV3 and IAV, HMPV and HRSV did not efficiently decrease CCR1, 2, and 5 expression, and did not efficiently increase CCR7 expression. Consistent with the differences in CCR7 mRNA and protein expression, MDDC stimulated with HRSV or HMPV migrated less efficiently to the CCR7 ligand CCL19 than did IAV-stimulated MDDC. Using GFP-expressing recombinant virus, we showed that the subpopulation of MDDC that was robustly infected with HRSV was particularly inefficient in chemokine receptor modulation. HMPV- or HRSV-stimulated MDDC responded to secondary stimulation with bacterial lipopolysaccharide or with a cocktail of proinflammatory cytokines by increasing CCR7 and decreasing CCR1, 2 and 5 expression, and by more efficient migration to CCL19, suggesting that HMPV and HRSV suboptimally stimulate rather than irreversibly inhibit MDDC migration. This also suggests that the low concentration of proinflammatory cytokines released from HRSV- and HMPV-stimulated MDDC is partly responsible for the low CCR7-mediated migration. We propose that inefficient migration of HRSV- and HMPV-stimulated DC to lymphatic tissue contributes to reduced adaptive responses to these viruses.
| The respiratory viruses human respiratory syncytial virus (HRSV) and, to a lesser extent, human metapneumovirus virus (HMPV) and human parainfluenza virus (HPIV3), can re-infect humans throughout life without significant antigenic change, suggesting that immunity to these viruses is incomplete. In contrast, re-infection by influenza A virus (IAV) depends on antigenic change, suggestive of more complete immunity. Dendritic cells (DC) take up virus antigen at the site of infection, undergo maturation, and migrate to the lymphatic tissue to present antigen to T lymphocytes, orchestrating the immune response. In response to antigen uptake, DC switch chemokine receptors on their surface, decreasing expression of receptors for inflammatory chemokines in the infected tissue, and increasing expression of CCR7, the sole chemokine receptor that directs DC to migrate to lymphatic tissue. By stimulating human DC in vitro, we found that, in contrast to HPIV3 and IAV, HMPV and HRSV did not efficiently induce the switching of these surface receptors. In cell migration assays, we showed that, compared with IAV-treated DC, HRSV- or HMPV-treated DC migrated less efficiently to CCL19, a chemokine that directs T cell migration to lymphatic tissue. Thus, during infection with HRSV and HMPV, inefficient migration of DC to the lymphatics could reduce the adaptive immune response to these viruses.
| The paramyxoviruses human respiratory syncytial virus (HRSV), human metapneumovirus (HMPV) and human parainfluenza virus type 3 (HPIV3) are common respiratory pathogens. HRSV is the most important viral agent of severe pediatric respiratory tract disease worldwide [1], [2], followed by HPIV3 [3], [4] and HMPV [5], [6], [7], [8]. The orthomyxovirus influenza virus type A (IAV) infects and causes respiratory tract disease in all age groups [9], [10], [11].
These human respiratory viruses share a tropism for the respiratory epithelium and have overlapping spectra of disease, ranging from rhinitis to bronchiolitis and pneumonia [12], [13]. IAV usually induces long-term immunity following infection, such that re-infection depends on significant antigenic change [14], [15]. In contrast, HMPV, HRSV and HPIV3 are able to symptomatically re-infect humans throughout life without significant antigenic change. This is particularly common with HRSV. Glezen and colleagues followed children from birth, and found that more than two-thirds were infected with HRSV during the first year of life, and almost half of these individuals were re-infected during each of the next two years [16]. In experimental infections of adults, typically 50–80% of subjects are re-infected with HRSV, and the majority has acute illness [17]. In another study, adults were challenged at intervals of 2–6 months over a period of 26 months with the same HRSV isolate, with the result that 73% were infected at least twice and 43% at least three times, and more than half of these infections were symptomatic [18]. These observations have been widely interpreted to indicate that HRSV in particular blunts or skews the immune response, resulting in suboptimal protection.
Antigen-presenting dendritic cells (DC) are critical for a functional adaptive immune response. During a lower respiratory tract infection, the number of DC in the bronchi and lung increases by chemotactic influx of precursors that originate primarily from circulating monocytes [19], [20], [21], [22]. Migration to non-lymphoid peripheral tissues such as the lung is mediated by so called “inflammatory” chemokine receptor-ligand pairs, including CCR1-CCL3/MIP-1α, CCR2-CCL2/MCP-1 or CCR5-CCL5/RANTES. Exposure of DC to antigen in peripheral tissue initiates DC maturation. During maturation, DC increase the surface expression of co-stimulatory molecules such as CD38, CD40, CD80 CD86, and CD83 [23], [24]. DC also change their expression of cell surface chemokine receptors: expression of CCR1, CCR2, and CCR5 is reduced, reducing responsiveness to inflammatory chemokines, and expression of CCR7 is increased [25], [26]. CCR7 has two specific ligands, CCL19 and CCL21, which are expressed by endothelial cells in lymphatic venules, in high endothelial venules (HEV) in lymph nodes, and in the T cell zone of lymphoid organs [27], [28], [29]. CCL19 and CCL21 direct migration of maturing, CCR7-expressing DC through the afferent lymphatics to the draining lymph nodes, and control DC positioning within defined functional lymphoid compartments [25], [26], [30], [31] for efficient interaction with naïve and/or antigen-specific memory T lymphocytes. DC have a key role in determining the magnitude and quality of the adaptive immune response.
We previously reported that HMPV, HRSV, and HPIV3 induce low-to moderate levels of human monocyte-derived dendritic cell (MDDC) maturation, cytokine/chemokine expression, and CD4 T cell proliferation, with the magnitude increasing slightly in the order HRSV, HMPV, and HPIV3 [32], [33]. MDDC generated in vitro from primary human monocytes by treatment with IL-4 and GM-CSF represent an appropriate model for lung DC because monocytes give rise to myeloid DC in the resting lung [34] and mucosa [35], and are phenotypically and functionally similar to DC located at sites of inflammation in vivo [36]. In the present study, we expanded our previous findings by screening MDDC for expression of genes related to maturation in response to HMPV, HPIV3, HRSV and, for comparison, IAV. We found that CCR7 mRNA and protein expression were substantially increased in response to HPIV3 and IAV, but minimally increased in response to HMPV and HRSV. These differences detected by qRT-PCR and flow cytometry were functionally relevant, since MDDC stimulated with HMPV or HRSV were less efficient in their migration along a CCR7 concentration gradient than IAV- and HPIV3 stimulated MDDC. Secondary stimulation of HRSV- or HMPV-exposed MDDC with the strong DC activator LPS enhanced CCR7 expression and in vitro migration, suggesting that suboptimal stimulation, rather than inhibition, is responsible for this poor-migration phenotype. Finally, we provide evidence that low CCR7 expression by MDDC in response to HRSV and HMPV is at least partly due to the low level of expression of pro-inflammatory cytokines (TNF-α, IL-1α and IL-6).
Elutriated monocytes were obtained from healthy donors at the Department of Transfusion Medicine of the National Institutes of Health, under a protocol (99-CC-0168) approved by the IRB of the Clinical Center, NIH. Written informed consent was obtained from all donors.
Recombinant (r) HMPV (strain CAN97-83), rHRSV (strain A2) and rHPIV3 (strain JS) with or without the GFP gene were described previously [12], [37], [38]. The present study employed a genetically “stabilized” version of rHMPV, in which the SH gene was modified to silently remove tracts of A and T residues that had been sites of spontaneous mutations during passage in vitro [39]. Human Influenza/A/Udorn/72, a wildtype virus of subtype H3N2, was used as control.
All viruses were grown on Vero cells and purified by centrifugation through sucrose step gradients as described previously [32]. Sucrose purified viruses were pelleted by centrifugation to remove sucrose. Virus pellets were resuspended in Advanced RMPI 1640 (Invitrogen, Carlsbad, CA) supplemented with 2 mM L-glutamine (aRPMI), and aliquots were snap frozen and stored at −80°C until use. Virus titers were determined by immuno-plaque assay on Vero cells under methylcellulose overlay (containing trypsin for titration of rHMPV and IAV) as described previously [37]. In some experiments, UV-inactivated viruses were included as controls which were prepared using a Stratalinker UV cross-linker (Agilent, Santa Clara, CA) at 0.5 J/cm2. Complete inactivation was monitored by plaque assay (limit of detection: 5 plaque forming units per mL).
Elutriated monocytes were obtained from healthy donors at the Department of Transfusion Medicine of the National Institutes of Health, under a protocol (99-CC-0168) approved by the IRB of the Clinical Center, NIH. As previously described [32], monocytes were subjected to CD14+ sorting on an Automacs separator (Miltenyi Biotec, Auburn CA), and cultured in presence of recombinant human IL-4 (R&D Systems, Minneapolis, MN) and recombinant human granulocyte-macrophage colony-stimulating factor (GM-CSF, Bayer Healthcare, Wayne, NJ) for 7 days to generate immature MDDC. These were confirmed by flow cytometry to be CD14−, CD38−, CD80low, CD86low, CD40low, CD54low.
Immature MDDC were seeded in 12-well plates at 6×105 cells per well and were infected with live virus at an MOI of 3 PFU/cell, or with an equivalent amount of UV-inactivated virus, stimulated with 1 µg/ml of the superantigen Staphylococcus enterotoxin B (SEB; Sigma, St Louis, MO) or with 1 µg/ml of lipopolysaccharide (LPS) from Escherichia Coli O55:B5 (Sigma). The infectivity of rgHMPV, rgHRSV and rgHPIV3 for MDDC was similar (approximately 3–5% of GFP+ MDDC at 24 or 48 h post infection, no significant differences at the p≤0.05 confidence level for any of the data sets of this study). In some experiments, immature MDDC infected with rgHMPV, rgHRSV, rgHPIV3 at an input MOI of 3 PFU/cell were further stimulated 4 to 6 h later with 1 µg/ml of LPS or 150 IU of Interferon (IFN)-β (PBL Interferon source, Piscataway, NJ) or with a cocktail of pro-inflammatory cytokines of 6 ng/ml TNF-α, 10 ng/ml IL-6 and 0.36 ng/ml IL-1α (R&D systems). All inoculations or stimulations were performed in advanced RMPI 1640 (Invitrogen) supplemented with 10% heat-inactivated FBS (Hyclone, Logan, UT), 2 mM L-glutamine (Invitrogen), 200 U/ml penicillin, and 200 µg/ml streptomycin (Invitrogen) at 37°C in 5% CO2.
Cell-associated RNA was isolated using the RNeasy mini kit (Qiagen) as recommended by the manufacturer and treated with DNAse I to remove residual genomic DNA. Analysis was done in two ways. The first involved a custom made low-density Taqman gene array containing 62 genes. Here, 1 µg of isolated RNA was reverse transcribed using SuperScript II (Invitrogen) in a 50 µl mix using random primers, and each cDNA mix was loaded onto an array in triplicate. The second method involved individual RT-PCR reactions. Here, 600 ng of isolated RNA was reverse transcribed using superscript II (Invitrogen) in a 25 µl mix using random primers. The reverse transcription product was diluted three-fold, and two µl of the diluted cDNA mix were used in each quantitative TaqMan PCR (Applied Biosystems, CA) for quantification of the targets of interest, namely CCR1 (Hs00174298_m1), CCR5 (Hs00152917_m1) and CCR7 (Hs00171054_m1). qPCR results were analyzed using the comparative threshold cycle (ΔΔCT) method, normalized to 18S rRNA and expressed as fold change over mock.
To determine the surface expression level of chemokine receptors, cells were stained with allophycocyanin (APC)-conjugated anti-human mAbs [anti-CCR1 (CD191, clone 53504), anti-CCR2 (CD192, clone 48607), anti-CCR5 (CD195, clone 2D7), anti-CCR7 (CD197, clone 2H4) (BD Biosciences, San Jose, CA)]. Isotype-matched mAbs were included as controls. Propidium iodide staining was used to exclude dead cells from further analysis. At 48 h post infection, the median viability of MDDC from six independent experiments was 85% for HMPV-, 86% for HRSV-, and 82% for HPIV3-exposed MDDC, reflecting the anti-apoptotic effects of virus-induced DC maturation [32]. In order to avoid interference, CCR1, 2, 5 and 7 expression was analyzed individually. At least 20,000 events were acquired using a FACSCalibur flow cytometer (BD Biosciences) and analyzed using FlowJo version 8.8.6 software (Tree Star, Inc., Ashland, OR).
After 48 h stimulation, migration of the virus-stimulated MDDC in response to a CCL19 concentration gradient was evaluated using polycarbonate 5-µm diameter pore size transwells (Corning, Lowell, MA). 1×105 live MDDC were seeded in the upper chamber, and incubated in presence or absence of CCL19 (1 µg/ml, (R&D Systems, Minneapolis, MN) in the lower chamber. Duplicate wells were used for each condition. After 3 h incubation, MDDC from the lower chamber were harvested, and the cell density of live cells was determined using a FACS Calibur flow cytometer (BD Biosciences). For each sample, data acquisition was performed for 1 min at constant flow using 200 µl final volume. Forward scatter, side scatter, live/dead staining, and GFP expression were analyzed using FlowJo version 8.8.6 software (Tree Star, Inc., Ashland, OR). The average number of MDDC specifically migrating in response to CCL19 was calculated as follows: (Average number of stimulated MDDC migrated in the presence of CCL19) – (Average number of stimulated MDDC migrated in the absence of CCL19).
Data sets were assessed for significance using parametric one-way repeated measures ANOVA with the Tukey post hoc tests for normally distributed data sets or the non-parametric Friedman test with Dunns post hoc test. A log10 transformation was applied to data sets when necessary to obtain equal standard deviation among groups, a necessary requirement of both tests. Statistics were performed using Prism 5 (GraphPad Software, Inc, San Diego, CA). Data were only considered significant at P<0.05. Analysis of CCR5 and CCR7 expression: To account for the smaller data set of the IAV control (n = 8 donors, except for IAV, n = 6 donors), data were analyzed using an unbalanced repeated measures ANOVA (JMP version 8.0.2; SAS, Cary, NC).
We used RT-qPCR to survey maturation-related gene expression in MDDCs from 3 donors 24 h after exposure to either the superantigen SEB or to purified live or UV-inactivated rHRSV, rHMPV, rHPIV3, or IAV (Fig. S1A and B). In general, all four live viruses induced the up-regulation of the same array of genes but differed in the intensity of up-regulation increasing in the order rHRSV<rHMPV<IAV<rHPIV3. The donors also had substantial responses to UV-inactivated IAV, but weak responses to UV-inactivated rHMPV, rHRSV or rHPIV3. Donors 1 and 2 were refractory to stimulation by rHMPV and rHRSV, respectively.
Among the genes surveyed, expression of CCR7 mRNA was substantially increased in response to IAV and rHPIV3, but not in response to rHMPV and rHRSV (Fig. S1A). Based on these preliminary results, we analyzed CCR7 mRNA expression by qPCR in additional donors (total n = 9, Fig. 1), and found that, while IAV and HPIV3 induced a strong increase of CCR7 mRNA (median increases of 23-fold and 7.2 fold, respectively), HRSV and HMPV only induced a 2.2- and 2.5-fold increase compared to mock treated cells. The effects of HMPV and HRSV on CCR7 expression were significantly smaller compared to HPIV3 and IAV (Fig. 1). By contrast, expression of CCR1 and CCR5 mRNA was increased in response to all viruses, but without any statistical difference between the viruses, except that the CCR5 mRNA expression was significantly different between rHPIV3 and IAV (Fig. 1). Because CCR7 has a unique role in DC migration towards lymph nodes and the subsequent adaptive response [26], we explored the effect of these viruses on MDDC chemokine receptor expression and migration.
We next used flow cytometry to measure surface expression of CCR1, 2, 5, and 7 on MDDC 48 h after exposure to the different viruses (Fig. 2). We included CCR2 in this analysis since, like CCR1 and CCR5, it directs monocytes and DC to inflamed tissue and is down-regulated during DC maturation. LPS was used as positive control because it strongly activates DC [40], [41]. Fig. 2A shows primary data for a representative donor, and Fig. 2B–C show the compiled results for six to eight donors. In this and all subsequent experiments, we used versions of rHMPV, rHRSV, and rHPIV3 that express GFP from an added gene (rgHMPV, rgHRSV, and rgHPIV3, respectively).
In mock-treated MDDC, substantial subpopulations of cells expressed CCR1 (median 91% of total), CCR2 (34%), and CCR5 (75%), and were CCR7-negative or low (Fig. 2A, B, C). High CCR1/2/5 and low CCR7 values would be typical for immature DC residing in peripheral tissue. As expected, LPS treatment induced a significant down-regulation of CCR1 (32%), CCR2 (9%), and CCR5 (24%), and up-regulation of CCR7 (48% positive cells) (Fig. 2A, B, C). Stimulation of MDDC with rgHPIV3 or IAV also induced a significant decrease in frequency of cells expressing the inflammatory chemokine receptors CCR1, 2, and 5 compared to mock-treated cells (Fig. 2B). However, only IAV significantly decreased all median fluorescence intensities (MFIs) (Fig. 2C).
In contrast, stimulation with rgHMPV or rgHRSV had only moderate effects on chemokine receptor surface expression. Cell surface expression of CCR1 and 2 decreased after stimulation with rgHRSV and rgHMPV, but the difference compared to mock-treated MDDC was not significant (Fig. 2B and C), except for the MFI of CCR1 (Fig. 2C). Stimulation with rgHMPV or rgHRSV reduced the percentage of CCR5+ MDDC significantly compared to mock-treated MDDC, but treatment with IAV reduced CCR5 expression significantly more than rgHMPV or rgHRSV treatment (Fig. 2B). CCR5 expression of rgHPIV3 stimulated MDDC was intermediate between HMPV and HRSV on the one hand, and IAV on the other hand, with no significant differences to any of the viruses (Fig. 2C).
The limited down-regulation of CCR1, 2, and 5 in response to rgHMPV and rgHRSV was coupled with a weak increase of CCR7 expression occurring on only a small subpopulation of cells (Fig. 2B, median 7% CCR7+ cells for for rgHMPV, and 6% for rgHRSV, with no statistical difference to mock). Stimulation with rgHPIV3 or IAV was associated with a significantly stronger up-regulation of CCR7 than mock, rgHMPV or rgHRSV stimulation, resulting in 13% and 37% CCR7+ cells, respectively.
Taken together, these results showed that compared to LPS, IAV, and rgHPIV3, stimulation with rgHMPV and rgHRSV induced a smaller down-regulation of surface expression of CCR1, CCR2, and CCR5, and a smaller up-regulation of CCR7 surface expression occurring on a smaller fraction of cells.
We used flow cytometry to compare chemokine receptor surface expression on virus-exposed cells that were GFP-positive versus GFP-negative (Fig. 4). We previously showed that, following infection with rgHMPV, rgHRSV, or rgHPIV3 at an MOI of 3, approximately 3–5 % of MDDC were GFP+ at 24 or 48 h post-infection [32]. This was indicative of robust viral gene expression, which was confirmed by RT-qPCR. In the GFP– population, we detected a low level of viral gene expression, suggestive of abortive virus replication [32]. Thus, comparing host gene expression in GFP+ and GFP– cells provides an indication of the effects of a robust versus abortive infection. Fig. 4A shows primary data for GFP expression and CCR7 surface expression for a single donor, and Fig. 4B summarizes data for the expression of CCR1, 2, 5, and 7 for six donors.
After treatment of MDDC with rgHMPV or rgHPIV3, the extent of down-regulation of CCR1, 2, and 5 was similar between the GFP+ and GFP− MDDC (Fig. 4). In contrast, after rgHRSV treatment, these receptors were decreased only in the GFP− cells; indeed, in the GFP+ cells, CCR2 and CCR5 expression was slightly increased compared to mock treated cells. Thus, robust rgHRSV gene expression did not induce the down-regulation of the inflammatory chemokine receptors CCR1, 2, and 5 that normally occurs as part of DC maturation.
CCR7 was expressed at higher levels in the GFP+ cells than in the GFP− cells after treatment with rgHMPV or rgHPIV3, indicating that robust infection by these viruses stimulated rather than inhibited expression (Fig. 4A and B). In contrast, CCR7 expression was not increased in either the GFP+ or the GFP− subpopulations of cells treated with rgHRSV.
One possible explanation for the weak chemokine receptor modulation and migration by rgHMPV- and rgHRSV-treated MDDC was direct virus-mediated inhibition. Alternatively, it was possible that these viruses were insufficiently stimulatory, perhaps due to the low production of cytokines by virus-treated MDDC as described in our previous study [32]. We therefore investigated whether exposure of virus-stimulated MDDC to secondary stimulation with LPS or to higher concentrations of cytokines would result in more efficient chemokine receptor modulation and migration. We tested possible cytokine and IFN candidates based on the gene expression analysis described above (Fig. S1) and previously published data by ourselves and others [32], [42], [43], [44], [45], [46]. The individual additions of IFN-β, IL-28, IL-29, TNF-α, IL-1α, IL-6 and prostaglandin E2 to virus-treated MDDC had little or no effect on CCR7 mRNA levels or on the ability of MDDC to migrate to a CCL19 concentration gradient (data not shown). These preliminary results confirmed previously published data showing that CCR7 is not an IFN-regulated gene in human or mouse DC [47], [48], [49]. Thus, the poor up-regulation of CCR7 by rgHMPV and rgHRSV is unlikely to be the result of a more stringent IFN antagonism by these viruses.
We next tested the effect of a cocktail of pro-inflammatory cytokines containing TNF-α, IL-1α and IL-6 on chemokine receptor expression, with each cytokine in concentrations similar to those induced by LPS under our experimental conditions [32]. MDDC (n = 4 donors) were treated with rgHMPV or rgHRSV, and, 4–6 h later, received a secondary stimulation with the cocktail of pro-inflammatory cytokines or with LPS. The expression levels of CCR7 mRNA were quantified 24 h post-infection (Fig. 5A). The secondary treatment with LPS induced a significant (p<0.05) increase of CCR7 mRNA expression in rgHMPV- and rgHRSV-stimulated MDDC. Thus, the relatively low level of expression of CCR7 in MDDC exposed to rgHMPV or rgHRSV was not due to an irreversible block. Following treatment with the cocktail of pro-inflammatory cytokines, there was an increase of CCR7 mRNA in mock-, rgHMPV- or rgHRSV-stimulated MDDC, although there was substantial individual variation and this increase did not reach statistical significance. This suggests that the low level of expression of CCR7 mRNA in MDDC stimulated with rgHMPV or rgHRSV might be partly a consequence of the low levels of TNF-α, IL-1α and IL-6 produced after exposure to rgHMPV or rgHRSV.
To measure cell surface protein expression, MDDC that were treated with rgHMPV or rgHRSV and given a secondary stimulation with the pro-inflammatory cytokine cocktail or LPS, as described above, were analyzed by flow cytometry at 48 h post-infection. Consistent with the results at the mRNA level, stimulation with the proinflammatory cytokine cocktail induced a partial decrease in CCR1, 2 and 5 as well as a partial increase in CCR7 surface expression (Fig. 5B and C; B: 1 representative donor, and C: n = 6 donors). Secondary stimulation with LPS had stronger effects in all cases.
We also evaluated replicate samples to investigate if the profile of CCR7 mRNA and protein expression correlated with the ability of MDDC to migrate to a CCL19 concentration gradient, measured 48 h post-infection (Fig. 5D, n = 5 donors). Indeed, secondary stimulation with LPS induced a strong and significant (p≤0.05) increase of migration of rgHMPV- and rgHRSV-stimulated MDDC as compared to virus-treated cells given a mock secondary treatment. Following secondary stimulation of virus-treated cells with the cocktail of pro-inflammatory cytokines, there was an increase in migration of mock-, rgHMPV- and rgHRSV-stimulated MDDC, although this did not reach statistical significance, and did not reach the level of increase induced by LPS. Taken together, these results suggest that the low concentration of TNF-α, IL-1α and IL-6 induced by rgHMPV and rgHRSV is partly responsible for the low CCR7 mediated migration.
We next investigated possible effects of robust viral infection (indicated by intracellular GFP expression) on chemokine receptor expression following treatment with the pro-inflammatory cytokine cocktail or LPS. This was done by infecting MDDC (n = 6 donors) with rgHMPV, rgHRSV, or rgHPIV3, subjecting them to a secondary stimulation with the pro-inflammatory cytokine cocktail or LPS at 4 h post-infection, and using flow cytometry to analyze the cell surface expression of CCR1, 2, 5, and 7 in the GFP-positive versus the GFP-negative populations at 48 h post-infection (Fig. 6).
Secondary stimulation of rgHMPV-, rgHRSV-, or rgHPIV3-stimulated MDDC with LPS decreased cell surface expression of CCR1, 2, and 5 on both GFP+ and GFP− cells. Secondary stimulation with the cocktail of pro-inflammatory cytokines also induced a decrease in surface expression of CCR1, 2, and 5. However, the magnitude of the effect usually was less than that observed with LPS.
Secondary stimulation of rgHMPV-, rHRSV-, or rgHPIV3-exposed MDDC with LPS induced an equally strong increase of CCR7 surface expression on GFP+ and GFP− cells, compared to cells that did not receive the secondary treatment (Fig. 6). Secondary stimulation of virus-infected cells with the pro-inflammatory cocktail also induced increases in CCR7 expression on both GFP− and GFP+ cells, although only in the case of rgHRSV GFP−+ and GFP− cells and rgHPIV3 GFP− cells was this difference statistically significant compared to cells receiving a mock secondary treatment. This provided further evidence that the poor expression of CCR7 in MDDC exposed to rgHRSV or rgHMPV could be overcome by secondary stimulation with LPS, and substantially overcome by secondary stimulation with the cocktail of pro-inflammatory cytokines. These increases were observed both in GFP+ and GFP− cells, indicating that robust viral infection did not irreversibly block CCR7 expression.
Compared to HPIV3 or IAV, stimulation of human MDDC with HRSV or HMPV in vitro resulted in inefficient maturational changes in chemokine receptor usage – namely down-regulation of CCR1, CCR2, and CCR5 and up-regulation of CCR7 – that are necessary for DC migration in vivo following antigen uptake. MDDC stimulated with HRSV or HMPV did not migrate efficiently towards a CCL19 gradient in an in vitro assay, compared to HPIV3 or IAV, confirming that the poor surface expression of CCR7 had functional consequences. The weak chemokine receptor modulation and migration by MDDC exposed to HMPV and HRSV, viruses that are thought to induce incomplete immunity, was particularly evident compared to MDDC exposed to IAV, a virus that induces effective immunity.
In vivo, maturing, antigen-bearing DC migrate from peripheral tissue to secondary lymphatic tissue and localize in defined lymphoid compartments, where they present antigens to CD4+ and CD8+ T lymphocytes, initiating and polarizing the T cell response [26], [50]. DC migration to and positioning within lymphatic tissue are critical towards mounting an effective adaptive immune response [50]. While there are multiple chemokine receptors that direct immature DCs towards peripheral sites, CCR7 is the only receptor that mediates migration toward and positioning within lymphatic compartments for interaction with T lymphocytes [30], [51], [52], [53]. Thus, differential effects of pathogens on CCR7 expression in particular could be functionally relevant for differences in the immune response to these pathogens. Accordingly, the reduced migration observed in our in vitro assay for HMPV- and HRSV-treated MDDC following stimulation with HRSV and HMPV suggests that, during an HMPV or HRSV infection in vivo, maturing DC migrate with reduced efficiency from the infected mucosa towards secondary lymphatic tissues. This might lead to reduced adaptive immune responses that could explain the greater ability of HMPV and HRSV to reinfect humans throughout life without need for significant antigenic change.
The present study was done with primary human cells from multiple donors. While the use of cells from an outbred population provides data with substantial individualistic differences and reduced statistical significance compared to convenient, uniform hosts like inbred mice, it is important to note that the natural host of the viruses in the present study is the human and not the mouse. Direct in vivo studies of virus-specific effects on DC migration during respiratory infections of humans are difficult, especially in children. Gill et al [54] noted that DC persisted in the lungs of children hospitalized for HRSV infection for as long as 8 weeks following the resolution of infection [55]. Resorting to data from mice, sustained increases in pulmonary DC have also been observed following HRSV infection [56]. Lucken et al [57] tracked the migration of mouse DC following HRSV infection and showed that the increase in DC numbers in the mouse mediastinal lymph node was slower compared to IAV or Sendai virus infection [58], [59], [60]. These observations would be consistent with inefficient migration from the lung to lymphoid tissue. Our in vitro studies now provide a mechanism for these previous in vivo observations. In addition, we provided data that MDDC maturation also was reduced with HMPV compared to HPIV3 and IAV.
We previously provided data indicating that the level of MDDC maturation in response to exposure to HMPV and HRSV is lower compared to HPIV3 [32] and IAV (not shown). In vivo, the combination of these two factors, namely reduced overall maturation and inefficient CCR7-CCL19 driven migration, might result in additive net effects that could affect both the magnitude and the quality of the adaptive immune response. Compared to infection with IAV, HRSV and HMPV infections may yield lower overall numbers of virus-stimulated mature DC in the afferent lymphatics. Reduced expression of co-stimulatory surface molecules and reduced cytokine expression could affect the quality of the response as well as its magnitude. In addition, the inefficient migration of maturing DCs may also play a role in viral pathogenesis: specifically, the sustained presence of mature DC in the mouse lung has been suggested to contribute to airway inflammation [56].
Another paramyxovirus, measles virus (MeV), was recently shown to inhibit CCR7-driven DC migration. Interference with DC maturation and function is considered to be central to MeV-induced immunosuppression. Compared to LPS, MeV infection failed to promote the switch from CCR5 to CCR7 expression, and MeV-matured DC exhibited chemotactic responses to CCL3 rather than to CCL19 [61]. Inhibition of CCR7-driven migration was also described for vaccinia virus and for herpes simplex virus type 1 [45], [62], [63]. However, the effects of reduced DC maturation and migration on long-term protection might be particularly significant for respiratory viruses such as HMPV and HRSV. Both of these viruses are restricted in tropism to the superficial cell layer of the respiratory tract, and protection against re-infection has reduced effectiveness (compared to viremic viruses, for example) due to the short-lived nature of local IgA antibodies, the inefficiency with which serum antibodies access the respiratory lumen, and the down-regulation of virus-specific CD8+ T cell functionality in the respiratory tract [64]. Thus, even modest decreases in the magnitude of the adaptive response could result in decreases in viral clearance and protection against re-infection.
We used recombinant GFP-expressing viruses to distinguish between effects in robustly infected (GFP-positive) and uninfected/abortively-infected (GFP-negative cells) MDDC. This revealed additional differences between the viruses. For MDDC infected with HMPV or HPIV3, the GFP-positive population expressed significantly more surface CCR7 than the GFP-negative population. In contrast, for MDDC infected with HRSV, the GFP-positive subpopulation resembled the GFP-negative population in having very low CCR7 surface expression. Thus, whereas robust infection with HMPV and HPIV3 stimulated expression of CCR7, robust infection with HRSV did not. Furthermore, GFP-positive cells infected with HRSV showed no down-regulation of CCR1, 2, and 5 surface expression. Thus, compared to HMPV or HPIV3, even the subpopulation of DC that is robustly infected with HRSV and contains abundant intracellular antigen would not be mobilized for migration. This would impede the delivery of HRSV antigen from the periphery to lymphoid tissue. Furthermore, DC that are robustly infected with a virus can readily process newly synthesized viral antigens for display on MHC class I molecules and presentation to CD8+ T cells. Reduced migration of DC that are robustly infected with HRSV to lymphoid tissue would reduce this activity. This would make activation of CD8+ T cells more dependent on cross-presentation by non-infected DC, and could reduce the efficiency of CD8+ T cell activation during HRSV infection, reducing viral clearance and the disease-sparing regulatory effects of HRSV-specific CD8+ T cells [65].
Secondary stimulation of HRSV- or HMPV-stimulated MDDC with LPS, a strong DC activator, resulted in up-regulation of CCR7 expression on both GFP-negative and GFP-positive cells and increased in vitro migration. In contrast, with vaccinia virus or human cytomegalovirus, a secondary stimulation of the infected DC with LPS failed to up-regulate the CCR7 chemokine receptor [45], [62]. LPS is a strong NFκ-B and AP-1 dependent DC activator [66], [67]. Secondary stimulation of HRSV- and HMPV-infected MDDC with the NFκ-B/AP-1-dependent pro-inflammatory cytokines TNF-α, IL-1α and IL-6, at concentrations comparable to those induced by LPS treatment, up-regulated CCR7 expression and was pro-migratory. This suggests that, in contrast to MeV, vaccinia virus, or herpes simplex virus, suboptimal stimulation, rather than inhibition, is responsible for the poor-migration phenotype of pneumovirus-exposed MDDC.
In summary, compared to HPIV3 and, in particular, IAV, the pneumoviruses HMPV and HRSV were inefficient in inducing the maturation-related changes in cell surface chemokine receptor expression in MDDC that are necessary in vivo to re-direct DC from the periphery to lymphoid tissue. Consistent with this, both HRSV and HMPV were poor inducers of MDDC maturation and migration in vitro. These effects could be contributing factors in the incomplete nature of protection induced by HRSV infection in humans.
|
10.1371/journal.pgen.1002023 | Systematic Detection of Polygenic cis-Regulatory
Evolution | The idea that most morphological adaptations can be attributed to changes in the
cis-regulation of gene expression levels has been gaining
increasing acceptance, despite the fact that only a handful of such cases have
so far been demonstrated. Moreover, because each of these cases involves only
one gene, we lack any understanding of how natural selection may act on
cis-regulation across entire pathways or networks. Here we
apply a genome-wide test for selection on cis-regulation to two
subspecies of the mouse Mus musculus. We find evidence for
lineage-specific selection at over 100 genes involved in diverse processes such
as growth, locomotion, and memory. These gene sets implicate candidate genes
that are supported by both quantitative trait loci and a validated
causality-testing framework, and they predict a number of phenotypic
differences, which we confirm in all four cases tested. Our results suggest that
gene expression adaptation is widespread and that these adaptations can be
highly polygenic, involving cis-regulatory changes at numerous
functionally related genes. These coordinated adaptations may contribute to
divergence in a wide range of morphological, physiological, and behavioral
phenotypes.
| Evolution can involve changes that are advantageous—known as
adaptations—as well as changes that are neutral or slightly deleterious,
which are established through a process of random drift. Discerning what
specific differences between any two lineages are adaptive is a major goal of
evolutionary biology. For gene expression differences, this has traditionally
proven to be a challenging question, and previous studies of gene expression
adaptation in metazoans have been restricted to the single-gene level. Here we
present a genome-wide analysis of gene expression evolution in two subspecies of
the mouse Mus musculus. We find several groups of genes that
have likely been subject to selection for up-regulation in a specific lineage.
These groups include genes related to mitochondria, growth, locomotion, and
memory. Analysis of the phenotypes of these mice indicates that these
adaptations may have had a significant impact on a wide range of phenotypes.
| To what extent the evolution of gene expression cis-regulation
drives the evolutionary innovations of life is an important unresolved question.
While some contend that changes in cis-regulation are responsible
for the majority of morphological adaptations [1], others point out that only a
few such cases have been demonstrated [2], [3] (we distinguish here between
cis-regulatory changes that have been shown to affect
phenotypes, of which there are a moderate number [4], [5], and those that have further been
shown to be adaptive, of which there have been far fewer [2], [3]; adaptive changes are those that
are subject to positive selection as a result of increasing fitness).
This long-standing paucity of examples of adaptive cis-regulatory
divergence was due in large part to the fact that historically it has not been
possible to formally demonstrate the presence of cis-regulatory
adaptation from genome-wide data [3]. Sequence-based approaches have often been used to scan
the genome for accelerated divergence in non-coding regions [6]–[9], but what fraction of these
represent positive selection on cis-regulation remains unknown;
other possible explanations include changes in local mutation rate or biased gene
conversion rate [10], or selection on non-coding RNAs, recombination control
elements, DNA replication origins, or any other non-coding feature of genomes (e.g.
[6]).
Moreover, even when accelerated evolution does reflect
cis-regulatory adaptation, the target genes often cannot be
identified, since transcriptional enhancers can act on distant genes in many
species.
Alternatively, many studies have attempted to detect genes under positive selection
from genome-wide gene expression data, but have been unable to demonstrate the
presence of positive selection due to the lack of a null model of neutrality [3], [11]. For example, the
finding that gene expression divergence among three populations of
Fundulus fish species correlates better with the species'
environment than with their phylogeny [12] is consistent either with
widespread adaptation to the environment, or with a neutral mutation affecting many
gene expression levels being shared between two populations by chance; these cannot
be distinguished without a null model of neutral change. Similarly, studies that
rank genes by their ratio of gene expression divergence between species to diversity
within species [13]–[14] can identify promising candidates for follow-up studies,
but cannot distinguish between neutral and adaptive evolution without knowing how
the expression of a “neutral gene” would evolve [3].
Several studies have succeeded in developing accurate neutral models of gene
expression change by quantifying expression divergence when selection is
artificially weakened in the lab [15]–[17]. In these studies positive selection on a gene's
expression would be indicated by a greater divergence between species than expected
from the neutral model; less divergence than expected would reflect negative
selection. Although these studies have had the potential to discover positive
selection, they have only uncovered negative selection—i.e. all genes have
shown less divergence between species than expected under neutrality. However since
these studies can only measure “average” selection pressures (much like
the dN/dS metric for coding regions), genes even with fairly
frequent episodes of positive selection on expression would go undetected if they
are most often subject to negative selection [3]. Therefore the lack of any
positive selection on gene expression identified in these studies is not evidence
against the existence of such positive selection.
This landscape has changed with the recent publication of two studies of selection on
genome-scale gene expression data in Saccharomyces yeast [3], . In one of
these [18], we
used the directionality of gene expression quantitative trait loci (eQTL; reviewed
in [20]) to
demonstrate that at least 242 gene expression levels (and likely many more) have
been subject to lineage-specific selection (i.e. different selective regimes between
two lineages) since the divergence of two strains of S. cerevisiae,
and then employed population-genetic analyses to show that most of these represent
positive selection, as opposed to relaxed negative selection. Although this work
expanded the number of known cases of gene expression adaptation (across all
species) by over 10-fold, it revealed little insight into the higher-level traits
being selected. In another important recent study, Bullard et al. [19] examined the
allele-specific expression (ASE) levels of gene sets (e.g. pathways, co-expressed
gene clusters, etc.) in a hybrid between S. cerevisiae and another
yeast, S. bayanus. ASE implies the presence of a
cis-acting polymorphism affecting expression, and consistent
directionality of ASE within a gene set implies lineage-specific selection (see
below for further explanation). This method has great promise for identifying the
biological processes affected by gene expression adaptation, though it remains
unknown if the gene sets implicated in this work have been subject to positive (as
opposed to relaxed negative) selection [19]. Interestingly, parallel
analysis of the genomic sequences of these same gene sets revealed no cases of
either promoters or protein-coding regions under positive selection [19].
Here we apply a gene set-based test of selection on gene expression to M.
musculus. Although mouse is a heavily studied model organism, both in
the lab and in the wild, no cases of gene expression adaptation have been
demonstrated in this species (one example, the Agouti gene, has
been found in Peromyscus deer mice [21]). Our results show that both
“traditional” eQTL mapping in an F2 population as well ASE analysis in
an F1 hybrid can be used to detect lineage-specific selection on gene sets, and that
data from additional strains can be used to polarize the changes and infer the
probable action of positive selection. Moreover, we expand the known extent of gene
expression adaptation in M. musculus from zero genes to over 100,
and find that a great deal of such adaptation may occur in parallel on many genes of
small effect, in contrast to all previously known cases of gene expression
adaptation [1],
[2] aside
from our work in yeast [18]. Finally, our results suggest that gene expression
adaptations can affect behavioral and physiological phenotypes, in addition to their
more well-established role in morphological evolution [1].
The test of lineage-specific selection we use is based upon an idea first
formalized by Orr [22] in an elegant test of selection on quantitative
traits: under neutrality, QTLs for any given trait are expected to be unbiased
with respect to their directionality. In other words, given two parents (A and
B) of a genetic cross, A alleles at any QTL would be expected to be equally
likely as B alleles to increase the trait value. If a significant bias is
seen—e.g., among 20 QTLs for a trait, the A allele increases the trait
value at all of them—neutrality may be rejected in favor of
lineage-specific selection (in the absence of ascertainment bias [see Text
S1]). At present, no gene expression levels have been mapped to a
sufficient number of eQTLs to reject neutrality for any single gene. However, if
the expression levels of an entire group of genes is treated as a single trait,
and each eQTL used in the test is independent (i.e. caused by a distinct
polymorphism), then lineage-specific selection can be detected as a bias in the
directionality of eQTLs for the gene set being tested [3], [19] (This approach will have
the greatest power for gene sets containing genes that predominantly have the
same direction of effect on a trait under selection; for gene sets with a
significant fraction of genes that act in opposition, selection in one direction
could result in upregulation of some, and downregulation of others.).
The independence of eQTLs for different genes is critical for this test, since a
single eQTL that affected many genes could lead to a strong bias in the
directionality of effect even in the absence of lineage-specific selection
(Figure 1, strain A
versus B). To ensure that each eQTL is independent, we considered only local
eQTLs—that is, eQTLs located at genetic markers that are close in the
genome to the gene whose expression they control. These local eQTLs have been
shown to be primarily cis-acting [23] (so we refer to these as
cis-eQTL for brevity), though we note that our test of
selection is equally valid for local trans-acting eQTLs. Since
a single cis-eQTL could conceivably control multiple nearby
genes, and thus violate the requirement for independence, we also discard genes
that are located close to others in the same gene set (see Methods).
At any eQTL, either the allele from parent A up-regulates expression (and thus
parent B's down-regulates), or the allele from parent A down-regulates
expression (and thus parent B's up-regulates). In our test we include an
equal number of each type (arbitrarily termed “+” and
“–”), so that any gene set that is not under lineage-specific
selection should have close to the same number of genes in each eQTL direction
(Figure 1, strain A
versus C). This null expectation requires no assumptions about gene sets or
eQTLs or the complex biological networks involved, but follows simply from the
fact that we constrain the total number of + and – eQTLs to be equal
(relaxing this constraint to allow different numbers of + and – eQTLs
is straightforward, and requires only adjusting the null expectation; e.g. if we
adjust our cutoffs so that 60% of all eQTLs are +, then any random
or non-lineage-specific-selected gene set is expected to have ∼60%
+ eQTLs). A hypergeometric p-value, testing whether the
observed data deviate from this expectation by having an excess of either +
or – eQTLs (Figure 1,
strain A versus D), constitutes the test. Although this method will have greater
power for gene sets with many cis-eQTLs, any variation in the
total number of cis-eQTLs per gene set (whether due to real
biological differences, or experimental design, e.g. gene sets not
well-represented on the expression array) will not lead to false-positive
results, since these will affect + and – eQTLs equally. Further, the
test is sensitive to both positive selection and relaxed negative selection
acting on a gene set, as long as that selection is present in only one of the
two lineages; thus it is a test of lineage-specific selection, although positive
selection can be inferred with additional data (see below). In this sense, it is
similar to the McDonald-Kreitman test [24], which also cannot
distinguish between positive and relaxed negative selection [25]. However
unlike the McDonald-Kreitman test, as well as nearly all other previous tests of
selection (on both gene expression levels and DNA/protein sequences), this is
not dependent on any assumptions about either demographic histories or a subset
of neutral sites (see Text S1).
We applied our test of selection to eQTL data from a cross between two diverged
inbred mouse strains, C57BL/6J (B6) and CAST/EiJ (CAST). B6, like most commonly
used lab strains, is a mosaic of several lineages [26], but primarily Mus
musculus domesticus. CAST represents Mus musculus
castaneus, a subspecies thought to have diverged from the primary
B6 progenitor strains ∼500,000 years ago [27]. This divergence, as well
as recent selection during inbreeding in the lab, provides ample opportunity for
adaptive changes to have accumulated in each lineage.
To map cis-eQTLs, we produced 442 F2 animals, either with B6 as
the F0 paternal strain (referred to here as CxB F2 animals) or maternal strain
(referred to as BxC F2 animals). Each mouse was genotyped at 1,438 informative
genetic markers, and genome-wide gene expression was measured in adult brain,
liver, and skeletal muscle (see Methods).
Cis-eQTLs were found by linear regression of gene
expression levels on genotypes separately in each of four cohorts: CxB females,
CxB males, BxC females, and BxC males. A total of 5,000
cis-eQTLs in each cohort—the strongest 2,500 in each
direction (corresponding to a false discovery rate [FDR]
<10% in each cohort)—were retained for analysis. Using the same
number of + and – eQTLs allows us to apply our simple yet robust null
expectation of neutrality to any gene set: regardless of what complex biological
networks and population histories underlie the eQTLs, any gene set not subject
to lineage-specific selection (including random gene sets) will show an
approximately equal number of + and – eQTLs, following the binomial
distribution. Throughout this work we report gene sets significant at either a
high-confidence (<2% FDR) or medium-confidence (<15% FDR)
cutoff, with FDRs estimated by testing randomly generated gene sets matched in
size to the real ones (see Methods).
We began by testing gene sets from the Gene Ontology (GO) Consortium [28], since these
have been shown to be useful in a wide range of applications (while any
particular gene's GO classification and expression data may be imperfect,
the sheer number of genes and expression measurements being used make this a
potentially powerful test; any inaccuracies in the gene set assignments may lead
to false negatives, but are unlikely to result in false positives). Applying the
hypergeometric test to 531 GO gene sets (each with at least 50 members)
separately in each tissue, we found one high-confidence set (FDR
= 1.5%, meaning that there is approximately a
1.5% probability that this enrichment is due to chance, given the number
of gene sets tested, and the overlap in content between gene sets): genes in the
“mitochondria” set were biased towards increased expression from B6
cis-eQTL alleles (“B6-upregulation”) in liver
(Table 1; see Table S1
for gene lists). These results were consistent across all four cohorts (Figure 2a), not only at the
gene-set level, but also for specific genes within those sets (see Text S1),
underscoring their robustness. SNPs that could disrupt microarray probe
hybridization are unlikely to explain the results, since these did not show any
enrichment in the B6-upregulated mitochondria-related genes (see Text S1).
The number of genes affected by selection can be estimated as the difference
between the numbers of cis-eQTLs in each direction (see Text S1);
in mitochondria, this is estimated separately in each cohort as 32-35 genes in
females and 47-48 genes in males (Figure 2a, green numbers). We note this will be conservative if any
of the CAST-upregulated cis-eQTLs were fixed by positive
selection as well. No additional gene sets were observed with medium
confidence.
To increase our statistical power, we combined results across tissues, since many
cis-eQTLs in our data were not tissue-specific. Seven
additional gene sets were found: one at high-confidence and six at
medium-confidence (Table
1; see Table S2 for results from all 531 gene sets). Two of the seven sets
were related to mitochondria at different levels of the GO hierarchy
(“mitochondrial inner membrane” and “intracellular
organelle”), while the other five represented a diverse collection of
functions. As an example, locomotory genes—which are biased towards
CAST-upregulation in all three tissues—are shown in Figure 2b. Similar to the mitochondria gene
set, the specific genes implicated in each cohort overlapped extensively (see
Text
S1). In sum, these results suggest that lineage-specific selection
involving these subspecies can be inferred for several functional
categories.
We also applied our method to other types of gene sets. Testing 41 modules of
genes co-expressed in each F2 population (see Methods), we did not find any significant enrichments for biased
directionality of cis-eQTLs. However testing 75 pathways from
the KEGG database [29], we found one at medium confidence (FDR
= 4.5%): the JAK/STAT pathway was biased towards
cis-upregulation in CAST brain (Table 1).
To complement the microarray-based approach described above, we turned to
sequencing RNA isolated from F1 mice to directly identify allele-specific
expression (ASE). While this approach does not offer the richness in terms of
understanding genetically regulated networks and their interactions that can be
achieved in a large F2 cross, it does address two drawbacks of the microarray
approach described above: 1) our microarrays cannot provide direct evidence of
cis-regulation (since local eQTLs can occasionally be
trans-acting [23]), so we cannot be confident that our results truly
reflect selection solely on cis-acting elements; and 2) there
is considerable time and expense associated with rearing, genotyping, and
expression profiling of hundreds of F2 mice.
We and others have shown that high-throughput mRNA sequencing (RNA-seq) in F1
hybrid mice is an effective approach to studying ASE [30]–[32]. mRNA levels can be
accurately estimated by simply counting the density of reads from each
transcript. Since heterozygous SNPs are present at a 1∶1 ratio in the
genome, any significant deviation from this ratio in the number of sequence
reads that can be mapped to each individual allele (as a result of containing a
heterozygous SNP) indicates ASE. When the allele-specificity associates in
reciprocal crosses with SNP genotype—as opposed to parent-of-origin, as
seen for imprinted loci [30]–[31]—this implies the presence of a
cis-acting eQTL. These cis-eQTL target
genes can then be used as input for our selection test, in exactly the same
fashion as those found using microarrays in an F2 population.
We searched for ASE in a set of ∼78 million sequence reads from F1 hybrid BxC
and CxB embryos we generated previously [30]. Because this is not only a
different technology, but also a different developmental stage (embryonic day
9.5) and tissue (whole embryos), we were encouraged to see several of our
strongest hits replicate. For example, mitochondrial genes show a bias towards
higher expression of B6 alleles, whereas locomotory-related genes show the
opposite (Figure 3a). Gene
sets that were biased in adults but not in F1 embryos might be tissue and/or
stage-specific, or may be missing due to lower power of our RNA-seq data for
weakly expressed genes (this is not an inherent limitation of RNA-seq, since
power is limited only by the number of reads). In addition, genes lacking any
B6/CAST sequence polymorphisms are not assayable by allele-specific RNA-seq.
In addition to replicating some hits from adult mice, the F1 embryo data revealed
new significant gene sets as well. Two gene sets reached high-confidence:
“calmodulin binding” and “memory” (Table 1 and Figure 3b), both showing a
bias towards B6-upregulation. Although unannotated SNPs overlapping RNA-seq
reads can cause a marginal alignment bias resulting in an apparent up-regulation
of the B6 reference genome alleles, our analysis indicates this is unlikely to
underlie the significance of these gene sets (see Text S1).
Consistent with previous work in yeast [19], we conclude that RNA-seq
is a cost-effective alternative for measuring selection on
cis-regulation, particularly between lineages with a high
density of exonic sequence differences.
An important question is whether the lineage-specific selection we detected has
had any detectable impact on organismal phenotypes. Examination of the gene sets
in Table 1 reveals that
specific predictions can be made for the gene sets belonging to the GO
“biological process” and “cellular component” ontologies
(Table 2). For
example, cis-eQTLs leading to higher expression of growth or
locomotory genes may be (at least naively) expected to increase growth or
locomotion, since these gene annotations were typically identified by observing
a reduction of growth or locomotion in a gene knockout/knockdown model; genes
leading to increased growth or locomotion when inactivated are far less common
(for example, among genes annotated as growth regulators [28], 40 have a mutant phenotype of
decreased body size, whereas only two are associated with increased body size
[33]).
These effects could either be strong, like all previous examples of adaptive
cis-regulatory adaptation in metazoans [1], [2]; or
subtle, considering that many loci are being selected in parallel and thus may
only exert major phenotypic effects in aggregate. We were unable to make any
phenotypic predictions for the GO molecular function terms (“calmodulin
binding”, “G-protein coupled receptor activity”,
“receptor activity”, or “enzyme inhibitor activity”), or
the JAK-STAT pathway.
If the loci we identified have major phenotypic effects, they should be
detectable by QTL mapping in our F2 mice. One phenotype we predicted to be
affected was measured for every F2 individual in our cross: naso-anal length,
which approximately reflects the sum of growth over the lifetime of the mice. In
females, we found two significant QTLs for length, on chromosomes 2 and 15,
while in males the strongest QTL was on chromosome 5 (Figure 4, red lines). In all three cases, the
B6 alleles were associated with greater length, as expected since B6 alleles
tend to increase expression of growth-related genes (whose knockout/knockdown
phenotype is typically a reduction in growth). Strikingly, the strongest QTL
from each gender overlapped almost perfectly with two of the strongest (genotype
versus expression level r2>0.5)
cis-eQTLs in the growth-related gene set (Figure 4, blue lines), and the
weaker female length QTL coincided with a weaker
(r2>0.2) but still highly significant
growth-related gene cis-eQTL (Figure 4a, green line). This overlap is
unlikely to occur by chance, considering that only ∼0.5% of
cis-eQTLs are as close to the length QTLs as each of these
are (probability of overlap by chance, p<0.001; see Methods). The three genes are Dcaf13 (also
known as WDSOF1), an rRNA processing factor; Ept1, a CDP-alcohol
phosphatidyltransferase (orthologous to human SELI); and Sp3, a transcription
factor. All three are well-conserved, and have been implicated in positive
regulation of growth either by mouse knockout [34], or RNAi experiments
involving their orthologs in Caenorhabditis elegans
[35]. This
highly significant overlap suggests that these genes may be responsible for the
length QTLs.
To further test the hypothesis that the cis-eQTLs for these
three genes affect mouse length, we applied a statistical approach for inferring
causality of eQTLs for other traits [36]–[37] that has been
extensively tested and validated using transgenic mice [38]. For all three genes,
causality for length was strongly (p<0.001) supported in at
least one tissue. This provides further support for a role of these eQTLs in the
length phenotype.
An alternative method to assess the phenotypic importance of these gene sets is
to compare the predictions to phenotypes of B6 and CAST mice. Although QTL
mapping cannot be performed with only two strains (typical mapping populations
consist of hundreds of F2 individuals or recombinant inbred lines)—and
thus causal loci cannot be implicated—concordance of predictions with
observed phenotypes can at least serve as evidence that the selection on
cis-regulation of these gene sets is phenotypically
relevant. To this end, we searched the literature for studies where phenotypes
we predicted to be affected by selection (Table 2) were measured in B6 and CAST. For
three of our four predictions, we found multiple studies testing the relevant
phenotypes. From the growth regulator gene set, we predicted larger size of B6
mice (measured by length, as above, or by total body mass), and indeed they are
known to have nearly twice the mass of CAST mice, from an early age through
adulthood [39], [40]. From the adult locomotory-related gene set showing
CAST-upregulation (found in both the microarray and RNA-seq data, Figure 1 strain B, and Figure 2a), we predicted
higher locomotor activity in CAST, which has indeed been observed [40], [41]. In fact,
one study [41]
found that daytime activity of CAST was over six times higher than that of B6.
The B6-upregulation of the memory-related gene set (Figure 2b) predicted increased memory in B6
(since knockout of most memory-related genes results in reduced, not increased,
memory). In two studies employing the Morris Water Maze (MWM) to measure
learning and memory, B6 significantly outperformed CAST [40], [42]. In fact, CAST showed no
capacity at all for memory in this context (see Text S1).
In sum, all three of our predictions that have been addressed in previous
publications were confirmed by multiple independent studies. We did not find any
studies contradicting these predictions.
Our fourth prediction—that mitochondria would be more abundant in B6, as a
result of the B6-upregulation of many mitochondrial genes (most notably genes
related to the inner membrane, but also mitochondrial small ribosomal subunits
[combined-tissue
p = 4.5×10−8],
among others) observed in both the microarray and RNA-seq data—has not, to
our knowledge, been tested by previous studies. Therefore we isolated nuclear
and mitochondrial genomic DNA from livers (the tissue with the strongest
B6-upregulation of mitochondrial genes) of B6 and CAST adult mice, and measured
the ratio of their mitochondrial to nuclear genome copy number by qPCR (see
Methods). Consistent with our
prediction, we found a small but highly significant
(p<0.001) difference between B6 and CAST, with B6 showing a
12.9% increase in abundance. Therefore, all four of our predictions have
been confirmed—three retrospectively and one
prospectively—underscoring the ability of our selection test to predict
phenotypic differences, and suggesting that these differences may have been
shaped by lineage-specific selection on cis-regulation (though
we note that other traits could also have been affected by, or been the primary
targets of, the lineage-specific selection in these gene sets).
To better understand the selection that has acted on these phenotypes, we sought
to determine on which lineage the majority of changes in each trait had
occurred. This can be achieved by including an outgroup species in the analysis:
for example, if a trait value in B6 is much further from the outgroup than is
the CAST trait value, then the most parsimonious explanation is that the
majority of divergence occurred on the B6 lineage. As with all parsimony-based
methods, this indicates the most likely evolutionary scenario (i.e. that
requiring the fewest changes), but cannot formally rule out any less
parsimonious explanation.
To perform this analysis we searched for measurements of the four traits in Table 2 from additional
mouse strains. Mus spretus (SPRET) is an ideal outgroup, being
the species most closely related to Mus musculus. We found
published measurements from SPRET for two of the traits, growth and memory. For
growth, the adult mass of SPRET was found to be statistically indistinguishable
from CAST [39]—and about half of that of B6—indicating
that the change in growth likely took place along the B6 lineage. Similarly for
memory, SPRET showed no evidence of recall in the MWM [42], similar to CAST but in
stark contrast to B6—again implicating the B6 lineage as the probable
source of divergence. In fact, B6 showed significantly greater recall than all
of the 12 other strains tested [42].
Although locomotory behavior has not been measured in an outgroup (to our
knowledge), it was measured in nine strains in addition to B6 and CAST [41], including
seven wild-derived strains that are more closely related to CAST than is B6 or
other lab strains [43]. Since CAST had over twice the daytime locomotory
activity of any other strain tested [41]—including the closely
related wild strains—the majority of divergence can be inferred to have
likely taken place on the CAST lineage, after its divergence from the other wild
strains (in this case, B6 is the outgroup). The much lower daytime activity
level of B6 was similar to most of the wild strains, as well as another lab
strain [43].
In sum, the phenotypic changes can be polarized for three of the traits. These
results rest on the logic of parsimony: that a phenotypic change in one lineage
is more likely than independent changes in the same trait—of the same
direction and magnitude—in two lineages. Under the assumption that the
phenotypic divergence was driven by (and thus occurred on the same branch as)
the expression divergence, all three cases can be inferred to have likely been
caused by cis-upregulation of the relevant gene sets.
As mentioned above, our test of lineage-specific selection cannot by itself
distinguish between positive selection and relaxed negative selection (analogous
to the McDonald-Kreitman test [24], [25]). However recent evidence from saturation mutagenesis
studies showing that the vast majority of random cis-regulatory
mutations cause downregulation (see Text S1) suggests that relaxed negative
selection would likewise be biased towards downregulation. If this is indeed the
case for the gene sets we have implicated, then relaxed negative selection is
unlikely to explain the upregulation of these three traits/gene sets, leading to
the conclusion that their divergence was most likely due to the action of
positive selection for upregulation. However given the qualitative nature of
this argument, we cannot yet quantify the precise probability that positive
selection has been acting upon the cis-regulation of these gene
sets.
Using a systematic genome-scale approach to inferring lineage-specific selection
acting on cis-regulation, we found that over 100 genes belonging to
several gene sets have undergone lineage-specific selection in mouse, which may have
impacted diverse morphological and behavioral phenotypes. This work reports the
first cases of adaptive cis-regulatory evolution in M.
musculus, and expands the classes of traits (in any species) known to
be affected by gene expression adaptation, which previously did not include any
behavioral phenotypes. Methodologically, we augment previous work [19] by showing
that adding information from an outgroup can suggest the likely action of positive
selection (as opposed to relaxed negative selection) when that selection was for
cis-acting upregulation. Two interesting questions for future
work are how much of this selection occurred since the introduction of these strains
to the lab, and for selection that occurred on the wild B6 ancestors, how much
occurred in Mus musculus domesticus (the primary ancestor of B6
[26]) as
opposed to Mus musculus musculus. Interestingly, wild M. m.
domesticus tend to be larger than wild M. m. castaneus
when reared in a common laboratory environment (C. Pfeifle, personal communication),
suggesting that this adaptation was likely to have occurred in the wild. Another
question raised by these findings is what are the relevant “units of
selection” [44] for these polygenic adaptations; though regardless of the
answer, our conclusions regarding the extent of selection on
cis-regulation will not be affected.
Because the RNA-seq version of this approach can be applied rapidly and inexpensively
to hybrids between any two diverged lineages (including outbred lineages), we expect
it will find use in a wide range of taxa. In fact, it can be applied to any ASE data
from a hybrid between diverged lineages. Published ASE data sets from a variety of
species (e.g. [45],
[46]) can now
be similarly re-analyzed for cis-regulatory selection. This
approach can also be applied to any of the numerous published eQTL data sets
involving crosses between diverged parental lines.
Our approach is quite different from all previous studies of metazoan
cis-regulatory adaptation [1]–[4], which have identified single
genes with extremely strong effects on phenotypes such as pigmentation (e.g. [21], [47], [48]) or skeletal
structure (e.g. [49]). Our results reveal several important insights that
could not have been found at this single-gene level. For example, the only
previously known case of pathway-level gene expression adaptation was from our work
on the ergosterol biosynthesis pathway in S. cerevisiae, where six
genes clustered in the pathway have undergone selection for down-regulation [18]. Our present
results extend this considerably, demonstrating that polygenic
cis-regulatory adaptation can operate in parallel on dozens of
genes within a single functional group or pathway, and that this has occurred in
multiple gene sets during recent mouse evolution. Although each gene under such
coordinate selection may be expected to have a less extreme phenotypic effect than
those previously reported [1], [2], [21], [47]–[49], the sum of their effects could be quite strong. One
important question that can now start to be addressed is how often
cis-regulatory adaptation proceeds via dramatic changes in
single genes, as opposed to more subtle changes distributed across an entire gene
set [3]. Much of
the answer may ultimately depend on factors such as the strength/duration of
selection (with intense/short-term selection pressure likely favoring extreme
single-locus changes) and the genetic architecture of the trait in question.
A second open question is how often cis-regulatory adaptation occurs
by upregulation versus downregulation of genes; our results suggest that the
majority of the adaptation we discovered was due to upregulation, in contrast to
most previous (single-locus) studies, which have predominantly identified cases of
trait loss via downregulation [2]. Interestingly, we previously observed a preponderance of
upregulation in a genome-wide study of gene expression adaptation in S.
cerevisiae
[18], suggesting
that this pattern may be widespread. Again, which of these is more common in a
particular species may depend on the nature of the selective pressure and the
underlying genetic architecture.
Third, it has been proposed that gene expression adaptation may be responsible for
most morphological adaptations in part because it offers a solution to the issue of
pleiotropy. For a gene expressed in many tissues or stages of development, an amino
acid change (in a constitutive exon) will affect the protein produced in all of
these different contexts. Even if this change is adaptive in one or two of them, it
has been argued that it would be highly unlikely to be advantageous in all of them
[1]. In
contrast, the modular nature of cis-regulation allows for a change
in expression in just one tissue or stage, without affecting any other; thus
pleiotropic constraints should not be as severe, and adaptation should be able to
proceed [1].
Predictions from this are that genes expressed more broadly will be more likely to
adapt via cis-regulation, and that these adaptations will only
affect a small part of the genes' expression patterns. Two recent studies
attempted to test this idea. In one of these [50], genes near noncoding elements
with accelerated evolution in the human lineage were proposed to have undergone
human-specific selection on cis-regulation (though the authors
acknowledged that such acceleration need not indicate positive selection); however
no enrichment was found for these genes to be expressed in more tissues than
average. In the other [51], genes were classified as either
“morphogenes” or “physiogenes” based on their mouse knockout
phenotypes; morphogenes (which tend to be expressed in fewer tissues) had higher
dN/dS (an indicator of selection on protein-coding regions), while physiogenes had a
higher magnitude of expression change between human and mouse, consistent with the
prediction of greater adaptive expression change in broadly expressed genes. However
this study did not distinguish between adaptive versus non-adaptive
change, or cis versus trans regulation, or tissue-specific
versus non-specific expression changes, so the relevance to
theories of tissue-specific adaptive cis-regulatory evolution is
not clear. Our results suggest that although most of the genes in our most
significant gene sets are broadly expressed (not shown), their expression in all
three tissues was affected by the recent selection on
cis-regulation we detected (Table S2; all gene sets from Table 1 were significant in all
three tissues, except for the JAK/STAT pathway); thus these adaptations were not
tissue-specific, so do not support pleiotropy-based arguments for the expected
prevalence of tissue-specific gene expression adaptation (we note that while the
adaptations did not result in tissue-specific expression changes, the selection may
have acted to change expression in just one tissue, with the rest changing as a
side-effect). Of course, since we have only examined three tissues in two mouse
strains, much more work is required to determine how general this conclusion is.
Finally, because of its genome-scale perspective, our approach may eventually help to
address many other fundamental questions that cannot be addressed by single-locus
studies [3], such
as what fraction of gene expression divergence is adaptive, and what fraction of
evolutionary adaptation occurs at the level of cis-regulation.
Ethics statement: All mouse work was conducted according to Institutional Animal
Care and Use Committee regulations.
C57BL/6J (B6) mice were intercrossed with M. m. castaneus
(CAST/EiJ) mice to generate 442 F2 progeny (276 females, 166 males). All mice
were maintained on a 12 h light–12 h dark cycle and fed ad libitum. Mice
were fed Purina Chow until 10 wk of age, and then fed western diet (Teklad
88137, Harlan Teklad) for the subsequent 8 wk. Mice were fasted overnight before
they were killed. Their tissues were collected, flash frozen in liquid nitrogen,
and stored in −80°C prior to RNA isolation.
RNA preparation and array hybridizations were performed at Rosetta Inpharmatics.
The custom ink-jet microarrays used were manufactured by Agilent Technologies.
The array used consisted of 2,186 control probes and 23,574 non-control
oligonucleotides extracted from mouse Unigene clusters and combined with RefSeq
sequences and RIKEN full-length cDNA clones.
Mouse tissues were homogenized, and total RNA extracted using Trizol reagent
(Invitrogen) according to manufacturer's protocol. Three micrograms of
total RNA was reverse transcribed and labeled with either Cy3 or Cy5
fluorochrome. Labeled complementary RNA (cRNA) from each F2 animal was
hybridized against a cross-specific pool of labeled cRNAs constructed from equal
aliquots of RNA from 150 F2 animals and parental mouse strains for each of the
three tissues. The hybridizations were performed to single arrays (individuals
F2 samples labeled with Cy5 and reference pools labeled with Cy3 fluorochromes)
for 24 h in a hybridization chamber, washed, and scanned using a confocal laser
scanner. Arrays were quantified on the basis of spot intensity relative to
background, adjusted for experimental variation between arrays using average
intensity over multiple channels, and fitted to a previously described error
model to determine significance (type I error) [52]. All microarray data are
available at NCBI GEO (GSE16227).
Genomic DNA was isolated from tail sections using standard methods and genotyping
was performed by Affymetrix (Santa Clara, CA) using the Affymetrix GeneChip
Mouse Mapping 5K Panel.
The RNA-seq data were described previously [30]. All data are available at
the NCBI SRA (accession SRA008621.10).
eQTL scans were performed by linear regression of expression log ratios against
genotypes (coded as 0, 1, and 2), separately in each tissue for each of the four
cohorts (CxB females, CxB males, BxC females, and BxC males). eQTL were
designated as “local” (and likely cis-acting) if
the regression between the expression level of a gene and a genetic marker
within 1 megabase of the transcription start site was significant (where
significance was defined as the cutoff resulting in 2,500 eQTLs in each
direction; see below). Testing for dominance (comparing the average heterozygote
value to the average of the two average homozygote values) revealed evidence for
non-additivity at only a small fraction of local eQTLs (as expected for
cis-eQTLs, which typically act additively), so dominance
effects were not included in our eQTL mapping.
We implemented the following strategy to isolate local eQTL effects in the
presence of unlinked marker correlations. First the strongest local eQTL was
identified, and expression of the target gene was then corrected for its effects
by taking the residuals of expression when regressed against the eQTL genotype.
The corrected expression level was then subjected to a whole-genome eQTL scan to
identify the strongest trans-eQTL. Once this
trans-eQTL was identified, its effects were regressed out
of the original expression levels for the gene. These
trans-corrected expression levels were then regressed
against all local genetic markers once again, to identify the strength and
direction of effect for the cis-eQTL. This process allows us to
achieve a more accurate estimate of local eQTL effect sizes, even in the
presence of unlinked trans-eQTLs or correlations between
unlinked genetic markers (we note that removing trans effects
is not necessary for our test, though we have found it to improve our ability to
estimate cis effects). More generally, our focus on local eQTLs
allows us to isolate the effect of the local polymorphism(s) on gene expression,
regardless of other effects (e.g. environmental effects,
trans-eQTL not captured in our regression approach, epistatic
interactions, feedback, etc.); of course such effects are widespread, but they
will only weaken the correlation between a genetic marker's genotype and a
nearby gene expression level, potentially causing us to miss some local eQTLs,
but not resulting in false-positive results.
A total of 5,000 genes with the strongest cis-eQTLs (2,500 in
each direction) in each tissue/cohort combination were analyzed. The decision to
use an equal number of eQTLs in each direction does not reflect any biological
aspects or assumptions, but instead is merely an arbitrary choice. Whether the
total “true” numbers of cis-eQTLs in each direction are actually
equal is not addressed here (nor is it directly relevant for interpreting our
test's results). Altering the proportion of eQTLs in each direction by up
to 10% (a 60/40 ratio) in either direction did not have any impact on our
results (i.e. the gene sets in Table 1 were not affected, although FDRs were changed slightly).
FDRs for each tissue/cohort combination were estimated by randomization. We first
shuffled genotype labels so that one individual's entire set of genotypes
was paired with another individual's expression levels. Then the entire
eQTL detection procedure was carried out, and the number of
cis-eQTLs above the cutoffs associated with the top 5,000
eQTLs in the real data were counted. Randomizations were repeated at least 1,000
times. The estimated FDR equals the average number of significant eQTLs in the
randomized data divided by 5,000 (the number in the real data). This procedure
yielded a maximum FDR of 9.7% in the smaller cohorts (BxC), and an FDR of
<2% in the larger (CxB) ones. An equal number of eQTLs were used in
each cohort so that results between cohorts would be directly comparable. We
note that 5,000 eQTLs represents an average of ∼3.5 eQTLs per genetic
marker, which is not surprising given that linkage disequilibrium extends for
many megabases in a mouse F2 cross, so a single marker captures many
polymorphisms.
Gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG)
classifications were tabulated for each gene on the microarray. Only the 531 GO
gene sets (from all levels of the GO hierarchy and all three GO branches:
Biological Process, Molecular Function, and Cellular Component) and 75 KEGG gene
sets containing at least 50 genes on our microarray were tested, since small
gene sets have little statistical power in our test. If multiple genes from a
gene set had cis-eQTLs and were located within 2 mb of each
other in the genome, all but one in the cluster were discarded from the
analysis, to ensure that the eQTLs being tested are all independent (the 2 mb
cutoff was chosen since the most distant known cis-regulation
is an enhancer ∼1 mb from its target gene; so allowing 1 mb from a
cis-regulatory mutation in each direction yields 2 mb). All
of the cases of clustered eQTLs within a gene set showed the same direction of
effect (either up-regulated by the B6 allele, or by the CAST allele, but not a
mixture of both), so the choice of which gene(s) to exclude had no effect on the
test's results. Relaxing our distance cutoff results in a small increase in
the sample size and gene set enrichment significance. Likewise, increasing the
distance cutoff excludes a small fraction of genes, marginally decreasing the
enrichment significance.
The effect directions for the cis-eQTLs of a gene set were then
tested for departure from the expected 1∶1 ratio of +/– alleles
by comparing to the hypergeometric expectation. The results are similar to
testing using the binomial expectation, but the hypergeometric takes into
account the fact that if many + alleles have already been observed in a
gene set, further genes in that set are actually slightly less likely to have
+ alleles by chance (since the total number of + and – alleles
included in our list is equal). Coexpression modules were constructed for each
tissue as previously described [53]. A total of 41 modules containing at least 50 genes
were tested (10 in brain, 14 in liver, and 17 in muscle).
Hypergeometric p-values for each gene set in each tissue/cohort were then
combined across cohorts by Fisher's method, to yield the single-tissue
p-values for each gene set. The FDR was estimated in two ways. In the first
approach, genotype labels were permuted as described above, and the entire eQTL
detection and directionality test procedure was carried out. This yielded zero
false positives even over many thousands of randomizations. However this
randomization strategy does not account for the fact that a gene with a
B6-upregulating cis-eQTL in one cohort is likely to have
B6-upregulating alleles in other cohorts as well. In order to capture this
effect in our permutations, we carried out a second randomization procedure. We
used the cis-eQTL results from the real data, but randomly
shuffled the gene set assignments for each gene. In this test, the consistency
of eQTL directions across tissues and cohorts is perfectly preserved, and only
the effect of the gene set assignments is randomized. With this procedure, false
positives were found at all cutoffs tested; FDRs were estimated at several
cutoffs, and are shown in Table
1. We note that although the data from different tissues are not
entirely independent, since they come from the same mice, this does not present
a problem for estimating FDRs because we combined the p-values in the same way
for both real and permuted data. In addition, the non-independence of gene sets
is not a problem, since this overlap is perfectly captured by our randomization
procedure.
For the multi-tissue analysis, the three single-tissue p-values for each gene set
were combined by Fisher's method, both for real and randomized data. This
was expected to increase power because it decreases false-positive eQTLs, though
it is also possible that the non-tissue-specific eQTLs this procedure enriches
are more likely to be the result of recent selection. FDRs were estimated as
described above for the single-tissue analysis. We also tested combining only
results from mice of each gender, but did not find any sex-specific gene set
enrichments.
The RNA-seq data were analyzed as follows. Sequence reads overlapping
heterozygous SNPs were assigned to alleles as described [30]. All reads from each allele
of each RefSeq gene were then summed to generate the total number of reads from
each allele. Distinct transcripts from the same gene cannot be discerned with
this approach (as with the vast majority of microarrays), so each gene was
treated as if it produced a single transcript (we note that since GO annotations
are typically for genes, and not individual transcripts, having
transcript-specific data would not substantially affect our results). SNPs with
no reads from one allele were discarded, since these are likely to reflect SNP
annotation errors. Binomial p-values were calculated for each gene, using the
expected 1∶1 ratio of reads from each allele. The most extreme 25%
of genes with allele-specific information (2,037 genes) in each direction were
retained for GO analysis. The GO analysis was carried out with the
hypergeometric test as described above, except that no p-values were combined
because only a single tissue/cohort was used. Randomizations were performed by
replacing the cis-eQTL target genes with randomly chosen genes,
and repeating the hypergeometric test.
The probability of QTLs for naso-anal length overlapping with eQTLs for the
growth regulator gene set was calculated as follows. The peaks for all three
eQTLs shown in Figure 4 were
within the 0.5 LOD support interval of the top three length QTLs (one in males,
two in females). Across all 3,834 eQTLs at this strength
(r2>0.2), only 0.6% were within this
interval of the male length QTL and 0.3% for each female length QTL.
Since these are independent, and 27 eQTLs from the growth gene set reached this
cutoff, the chance of all three overlaps is
27×0.006×26×0.003×25×0.003 = 0.00095.
Interestingly, the eQTL overlapping the strongest length QTL in each gender were
both in the top 12 strongest growth eQTL (r
2>0.5), so even just the overlap of those two is significant at
p = 0.002. Testing the overlap with the three length QTLs
in random groups of 27 eQTLs supported these calculations. In males there is one
length QTL where the CAST allele is associated with greater length, but this was
not included in our overlap analysis because we only posit that the alleles
increasing B6 growth have been under positive selection and are present in the
list of growth genes with B6-upregulating cis-eQTL. eQTL scans
shown in Figure 4 were
performed using CxB brain; brain was chosen because it is the tissue with the
strongest growth gene eQTL direction bias, and CxB was chosen because it is the
larger of the two cohorts. Expression levels were from CxB female brains in
Figure 4a, and CxB male
brains in Figure 4b, to
match genders with the length QTL shown.
We performed quantitative PCR with SYBR green, amplifying both nuclear and
mitochondrial DNA from B6 and CAST liver tissue. The ratio of
mitochondrial/nuclear DNA gives an estimate of the mitochondrial abundance in
each strain, and the ratio of these ratios indicates their relative levels. The
following primer sequences were used: nuclear, CCTTGGACATTAGCACATGG and AACTGTCTCCCCTGACCAAC; mitochondrial,
ACAATGTTAGGGCCTTTTCG and
GTTCCCAGAGGTTCAAATCC. No
off-target effects were observed for either primer pair. Each reaction was
repeated 48 times to ensure consistency. The 99% confidence interval for
the B6:CAST ratio of mitochondrial/genomic DNA (a ratio of ratios) was 1.06
– 1.20, and the 99.9% confidence interval was 1.04 –
1.23.
|
10.1371/journal.pcbi.1004402 | Dynamic Integration of Value Information into a Common Probability Currency as a Theory for Flexible Decision Making | Decisions involve two fundamental problems, selecting goals and generating actions to pursue those goals. While simple decisions involve choosing a goal and pursuing it, humans evolved to survive in hostile dynamic environments where goal availability and value can change with time and previous actions, entangling goal decisions with action selection. Recent studies suggest the brain generates concurrent action-plans for competing goals, using online information to bias the competition until a single goal is pursued. This creates a challenging problem of integrating information across diverse types, including both the dynamic value of the goal and the costs of action. We model the computations underlying dynamic decision-making with disparate value types, using the probability of getting the highest pay-off with the least effort as a common currency that supports goal competition. This framework predicts many aspects of decision behavior that have eluded a common explanation.
| Choosing between alternative options requires assigning and integrating values along a multitude of dimensions. For instance, when buying a car, different cars may vary for their price, quality, fuel economy and more. Solving this problem requires finding a common currency to allow integration of disparate value dimensions. In dynamic decisions, in which the environment changes continuously, this multi-dimensional integration must be updated over time. Despite many years of research, it is still unclear how the brain integrates value information and makes decisions in the presence of competing alternatives. In the current study, we propose a probabilistic theory that allows dynamically integrating value information into a common currency. It builds on successful models in motor control and decision-making. It is comprised of a series of control schemes with each of them attached to an individual goal, generating an optimal action-plan to achieve that goal starting from the current state. The key novelty is the relative desirability computation that integrates good- and action- values to a single dynamic variable that weighs the individual action-plans as a function of state and time. By dynamically integrating value information, our theory models many key results in movement decisions that have previously eluded a common explanation.
| A soccer player moves the ball down the field, looking for an open teammate or a chance to score a goal. Abstractly, the soccer player faces a ubiquitous but challenging decision problem. He/she must select between many competing goals while acting, whose costs and benefits can change dynamically during ongoing actions. In this game scenario, the attacker has options to pass the ball to one of his/her teammates. An undefended player is preferred, but this opportunity will soon be lost if the ball is not quickly passed. If all teammates are marked by opposing players, other alternatives like holding the ball and delaying the decision may be better. Critically, the best option is not immediately evident before acting. To decide which strategy to follow at a given moment requires dynamically integrating value information from disparate sources. This information is diverse relating to both the dynamic value of the goal (i.e., relative reward of the goal, probability that reward is available for that goal) and the dynamic action cost (i.e., cost of actions to pursue that goal, precision required), creating a challenging problem in integrating information across these diverse types in real time. Despite intense research in decision neuroscience, dynamic value integration into a common currency remains poorly understood.
Previous explanations fall into two categories. The goods-based theory [1–7] proposes that all the decision factors associated with an option are integrated into a subjective economic value independently computed for each alternative. This view is consistent with evidence suggesting convergence of value information in the prefrontal cortex [3–5, 7]. Critically, action planning starts only after a decision is made. While this view is sufficient for decisions like buying or renting a house, modifications are needed for decisions while acting. Alternatively, an action-based theory proposes that options have associated action-plans. According to this theory, when the brain is faced with multiple potential goals, it generates concurrent action-plans that compete for selection and uses value information to bias this competition until a single option is selected [8–14]. This theory has been received apparent support from neurophysiological [8–10, 15–19] and behavioral [11, 12, 20–23] studies. Although the action-based theory explains competition, it leaves mysterious how action cost is integrated with good value (also referred as stimulus value in some decision-making studies [13]) that have different currencies and how goods-based decisions that do not involve action competition are made. To solve complex decision problems, the brain must dynamically integrate all the factors that influence the desirability of engaging in an action-plan directed towards a goal.
We propose a theory of dynamic value integration that subsumes both goods-based and action-based theories. We provide a simple, computationally feasible way to integrate online information about the cost of actions and the value of goods into an evolving assessment of the desirability of each goal. By integrating value information into a common currency, our approach models many key results in decision tasks with competing goals that have eluded a common explanation, including trajectory averaging in rapid reaching tasks with multiple potential goals, a common explanation for errors due to competition including the global-effect paradigm in express saccadic movements [24], and a unified explanation for the pattern of errors due to competition in sequential decisions [25].
This section describes analytically the computational theory developed in this study to model decisions in tasks with competing goals. We used a reaching task as a paradigm. Full details of the architecture and stochastic optimal control methodology that underlies the control schemes of our theory is in S1 Text for reaching and S2 Text for saccade models.
Stochastic optimal control has proven a powerful tool at modeling goal-directed movements, such as reaching [26], grasping [27] and walking [28] (for review see [29]). It involves solving for a policy π that maps states into actions ut = π(xt) by minimizing a cost function penalizing actions and deviations from a goal. Despite the growing popularity of optimal control models, most of them are limited to tasks with single goals, because policies are easily defined towards a single goal. On the other hand, it is unclear how to define policies in the presence of multiple goals, each of which may provide different reward and may require different effort. The core difficulty is to develop a single policy that selects actions that pursue many targets but ultimately arrives at only one.
One of the simplest solutions is to carefully construct a composite cost function that incorporates all targets. However, naive applications of this approach can produce quite poor results. For instance, an additive mixture of quadratic cost functions is a new cost function with a minimum that does not lie at any of the competing targets. The difficulty is that quadratic cost functions do not capture the winner-take-all implicit reward structure, since mixtures of quadratics reward best for terminal positions in between targets. Even when such a cost function can be constructed, it can be very difficult to solve the policy, since these types of decision problems are P-SPACE complete—a class of problems more intractable than NP-complete. Any dynamic change in targets configuration requires a full re-computation, which makes the approach difficult to implement as a real-time control strategy [30].
To preserve simplicity, we propose to decompose the problem into policy solutions for the individual targets. The overall solution should involve following the best policy at each moment, given incoming information. We can construct a simple cost function that has this property using indicator variables ν(xt). The indicator variables encode the policy that has the lowest future expected value from each state—in other words, it categorizes the state space into regions where following one of the policies to a goal i is the best option. In essence, a goal i “owns” these regions of the state space. We can write the cost function that describes this problem as a ν-weighted mixture of individual cost functions J j ′ s:
J = ∑ j = 1 N ν j ( x t ) J j ( x t , π j ) J = ∑ j = 1 N ν j ( x t ) ( ( x T j - S p j ) T Q T j ( x T j - S p j ) + ∑ t = 1 T j π j ( x t ) T R π j ( x t ) ︸ J j ( x t , π j ) ) (1)
where N is the total number of targets and νj is the indicator variable associated with the target j. The cost function Jj(xt, πj) describes the individual goal for reaching the target j starting from the current state xt and following the policy πj for time instances t = [t1, ⋯, tTj]. Tj is the time-to-contact that target j and S is a matrix that picks out the hand and target positions from the state vector. The first term of the cost Jj is the accuracy cost that penalizes actions that drive the end-point of the reaching trajectory away from the target position pj. The second term is the motor command cost that penalizes the effort required to reach the target. Both the accuracy cost and the motor command cost characterize the “action cost” Vπj(xt) for implementing the policy πj at the state xt. Matrices QTj and R define the precision- and the control- dependent costs, respectively (see S1 Text for more details).
When there is no uncertainty as to which policy to implement at a given time and state (e.g., actual target location is known), the ν-weighted cost function in Eq (1) is equivalent to the classical optimal control problem. The best policy is given by the minimization of the cost function in Eq (1) with νj = 1 for the actual target j and νi ≠ j = 0 for the rest of the non-targets. However, when there is more than one competing target in the field, there is uncertainty about which policy to follow at each time and state. In this case, the best policy is given by minimizing the expected cost function with expectation across the probability distribution of the indicator variable ν. This minimization can be approximated by the weighted average of the minimization of the expected individual cost functions, Eq (2).
π m i x ( x t ) = ∑ j = 1 N ⟨ ν j ( x t ) ⟩ ν arg min π j J j ( x t , π j ) = ∑ j = 1 N ⟨ ν j ( x t ) ⟩ ν π j * ( x t ) (2)
where ⟨.⟩ν is the expected value across the probability distribution of the indicator variable ν, and π j * ( x t ) is the optimal policy to reach goal j starting from the current state xt. For notational simplicity, we omit the * sign from the policy π, and from now on πj(xt) will indicate the optimal policy to achieve the goal j at state xt.
The first problem is to compute the weighting factor ⟨νj(xt)⟩ν, which determines the contribution of each individual policy πj(xt) to the weighted average πmix(xt). Let’s consider for now that all the alternative targets have the same good values and hence the behavior is determined solely by the action costs. Recall that Vπj(xt) represents the value function—i.e., cost that is expected to accumulate from the current state xt to target j including the accuracy penalty at the end of the movement, under the policy πj(xt). This cost partially characterizes the probability of achieving at least Vπj(xt) starting from state x(t) at time t and adopting the policy πj(xt) to reach the target j, Eq (3):
P ( V π j ( x t ) | π j ( x t ) , x t , Δ t ) = λ e - 1 λ V π j ( x t ) (3)
where λ is the free “inverse temperature” parameter (S3 Text). This assumption can be taken as is, or justified from the path integral approach in [31] and [32]. The probability that the value function of the policy πj at the current state xt is lower than the rest of the alternatives P(Vπj(xt) < Vπi ≠ j(xt)) can be approximated by the softmax-type equation in Eq (4), which gives an estimate of the probability of νj at xt:
P ( V π j ( x t ) < V π i ≠ j ( x t ) ) ≈ P ( ν j | x t ) = λ e - 1 λ V π j ( x t ) ∑ i = 1 N λ e - 1 λ V π i ( x t ) (4)
where N is the total number of targets (i.e., and total number of policies) that are available at the current state.
Given that all targets have the same good values, the probability P(νj|xt) characterizes the “relative desirability” rD(πj(xt)) of the policy πj to pursue the goal j at a given state xt. It reflects how desirable is to follow the policy πj at that state with respect to the alternatives. Therefore, we can write that:
r D ( π j ( x t ) ) = P ( V π j ( x t ) < V π i ≠ j ( x t ) ) (5)
However, in a natural environment the alternative goals are usually attached with different values that we should take into account before making a decision. We integrate the good values into the relative desirability by computing the probability that pursing the goal j will result in overall higher pay-off rj than the alternatives, P(rj > ri ≠ j):
r D ( π j ( x t ) ) = P ( V π j ( x t ) < V π i ≠ j ( x t ) ) P ( r j > r i ≠ j ) (6)
To integrate the goods-related component on the relative desirability, we consider two cases:
The reward magnitude is fixed and equal for all targets, but the receipt of reward is probabilistic. In this case, the probability that the value of the target j is higher than the rest of the alternatives is given by the reward probability of this target P(target = j|xt) = pj:
P ( r j > r i ≠ j ) = p j (7)
The target provides a reward with probability pj, but the reward magnitude is not fixed. Instead, we assume that it follows a distribution r j ∼ ( 1 - p j ) δ ( r j ) + p j N ( μ j , σ j 2 ), where δ(rj) is the Delta dirac function, and μj and σj are the mean and the standard deviation of the reward attached to the target j. For simplicity reasons, we focus on the case with two potential targets, in which the goal is to achieve the highest pay-off after N trials. In this case, the goods-related component of the desirability function is P ( r ¯ 1 > r ¯ 2 ), where r ¯ j = 1 N ∑ k = 1 N r j ( k ) , j = 1 , 2 is the net reward attached to the target j—i.e., the average reward received from the target j across N trials.
To compute the probability P ( r ¯ 1 > r ¯ 2 ), we need the probability distribution of P ( r ¯ j ) , j = 1 , 2. Given p(n) = Binomial(n, pj, N) is the probability of receiving n-times reward after N trials, the probability distribution of r ¯ j is:
P ( r ¯ j ) = ∑ n = 0 N p ( n ) ( 1 N ∑ k = 1 n r j ( i ) ) (8)
We can show that a mean based on n samples has a Normal distribution N ( n N μ j , σ j 2 n ). Therefore, the distribution of r ¯ j can be written as:
P ( r ¯ j ) = ∑ n = 0 N p ( n ) N ( r ¯ j ; n N μ j , σ j 2 n ) (9)
For a large number of trials N > > 0, p(n) is concentrated around n = pj N and r ¯ j ∼ N ( p j μ j , σ j 2 p j N ) , j = 1 , 2. To compute P ( r ¯ 1 > r ¯ 2 ) = P ( r ¯ 1 - r ¯ 2 > 0 ), we define a new random variable, Z = r ¯ 1 - r ¯ 2, which has Normal distribution with mean p1 μ1 − p2 μ2 and variance σ 1 2 p 1 N + σ 2 2 p 2 N. We can show that P ( r ¯ 1 > r ¯ 2 ) = P ( Z > 0 ) is given as:
P ( r ¯ 1 > r ¯ 2 ) = 1 2 e r f c ( p 2 μ 2 - p 1 μ 1 2 ( σ 1 2 p 1 N + σ 2 2 p 2 N ) ) (10)
where erfc is the complementary error function. Using that erfc(x) = 1 − erf(x), where erf is the error function, we can write that:
P ( r ¯ 1 > r ¯ 2 ) = 1 2 + 1 2 e r f ( p 1 μ 1 - p 2 μ 2 2 ( σ 1 2 p 1 N + σ 2 2 p 2 N ) ) = C u m N o r m ( Z ; p 1 μ 1 - p 2 μ 2 , σ 1 2 p 1 N + σ 2 2 p 2 N ) (11)
This result is consistent with the common practice of modeling choice probabilities as a softmax function between options. For example, the cumulative normal distribution can be approximated by the following logistic function [33]:
P ( r 1 > r 2 ) = l ( r ¯ 1 - r ¯ 2 ; p 2 μ 2 - p 1 μ 1 , S 1 . 6 ) (12)
where S = σ 1 2 p 1 N + σ 2 2 p 2 N.
In the preceding sections we developed a theory for the case that targets are presented simultaneously and the expected reward depends only on successfully reaching the target—i.e. reward availability is not state- and time- dependent. However, decisions are not limited only to this case but often involve goals with time-dependent values. In this section, we extend our approach to model visuomotor tasks with sequential goals, focusing on a pentagon copying task.
The theory precedes as before, with a set of control schemes that instantiate policies πj(xt)—where (j = 1, ⋯ 5)—that drive the hand from the current state to the vertex j. However, to draw the shape in a proper spatial order, we cannot use the same policy mixing as with simultaneously presented goals. Instead, we have to take into account the sequential constraints that induce a temporal order across the vertices. We can conceive the vertices as potential goals that provide the same amount of reward, but with different probabilities (i.e., similar to scenario 2 in the reaching task) with the exception that we design the target probability to be time- and state- dependent, so that it encodes the order of policies for copying the pentagon. The target probability P(vertex = j|xt) describes the probability that the vertex j is the current goal of the task at the state xt after departing from the vertex j − 1, or in other words, it describes the probability that we copy the segment defined by the two successive vertices j − 1 and j.
We define an indicator function ej that is 1 if we arrive at vertex j and 0 otherwise.
P ( v e r t e x = j | x t ) = P ( e j = 0 , e j - 1 = 1 | x t ) = P ( e j = 0 | x t ) p ( e j = 1 | x t ) = (13)
= ( 1 - P ( τ a r r i v e j < t ) ) P ( τ a r r i v e j - 1 < t ) (14)
where τ a r r i v e j is the time to arrive at vertex j -i.e, time to complete drawing the segment defined by the vertices j − 1 and j.
Let’s assume that we are copying the shape counterclockwise starting from the purple vertex (see right inset in Fig 1A), at the initial state at time t = 0. The probability distribution of time to arrive at vertex j, τ a r r i v e j, is given by Eq (15).
P ( τ a r r i v e j ) = ∑ k = 1 j P ( τ a r r i v e k | τ a r r i v e k - 1 ) P ( τ a r r i v e k - 1 ) (15)
where P ( τ a r r i v e k | τ a r r i v e k - 1 ) is the probability distribution of time to arrive at vertex k given that we started from vertex k − 1. We generated 100 trajectories between two successive vertices and found that P ( τ a r r i v e k | τ a r r i v e k - 1 ) can be approximated by a Normal distribution N ( μ τ a r r i v e , σ τ a r r i v e 2 ). Using Eq (15), we show that P ( τ a r r i v e j ) is also Gaussian distribution, but with j times the mean and variance—N ( j μ τ a r r i v e , j σ τ a r r i v e 2 ) as shown in Fig 1A. Considering that, we estimate that target probability P(vertex = j|xt), Fig 1B. Each time that we arrive at a vertex, we condition on completion, and P(vertex = j|xt) is re-evaluated for the next vertices.
The basic architecture of the model is a set of control schemes, associated with individual goals, Fig 2. Each scheme is a stochastic optimal control system that generates both a goal-specific policy πj, which is a mapping between states and best-actions, and an action-cost function that computes the expected control costs to achieve the goal j from any state (see S1 Text for more details). It is important to note that a policy is not particular a sequence of actions—rather it is a controller that tells you what action-plan uj (i.e., sequence of actions ui) to take from a state xt to the goal (i.e., πj(xt) = uj = [ut, ut+1, ⋯ utend]). In addition, the action-cost function is a map cost(j) = Vπj(xt) that gives the expected action cost from each state to the goal.
Let’s reconsider the soccer game scenario and assume a situation in which the player has 3 alternative options to pass the ball (i.e., 3 unmarked teammates) at different distances from the current state xt. In such a situation, the control schemes related to these options become active and suggest 3 action-plans (u1 = π1(xt), u2 = π2(xt) and u3 = π3(xt)) to pursue the individual options. Each of the alternative action-plans is assigned with value related to the option itself (e.g., teammates’ performance, distance of the teammates to the goalie) and with cost required to implement this plan (e.g., effort). For instance, it requires less effort to pass the ball to the nearby teammate No.1, but the distant teammate No.2 is considered a better option, because he/she is closer to the opponent goalie. While the game progresses, the cost of the action-plans and the estimates of the values of the alternative options change continuously. To make a correct choice, the player should integrate the incoming information online and while acting. However, the value of the options and the cost of the actions have different “currencies”, making the value integration a challenging procedure. The proposed theory uses a probabilistic approach to dynamically integrate value information from disparate sources into a common currency that we call the relative desirability function w(xt). While common currency usually refers to integration in value space, relative desirability combines in the space of policy weights. Using relative desirability, integration of disparate values is accomplished by combining each different type of value in its own space, then computing the relative impact of that value on the set of available policies.
The crux of our approach is that to make a decision, we only need to know what is the current best option and whether we can achieve it. This changes the complex problem of converting action costs to good values into a simple problem of maximizing the chances of getting the best of the alternatives that are currently available. To integrate value information with different “currencies”, we compute the probability of achieving the most rewarding option from a given time and state. This probability has both action-related and goods-related components with an intuitive interpretation: the probability of getting the highest reward with the least effort. We call this value relative desirability (rD) because it quantifies the attractiveness of the policy π for each goal i from state xt relative to the alternative options:
r D ( π i ( x t ) ) = P ( c o s t ( i ) < c o s t ( j ≠ i ) | x t ) P ( r e w a r d ( i ) > r e w a r d ( j ≠ i ) | x t ) (16)
The first term is the “action-related” component of the relative desirability and describes the probability that pursuing the goal i has lowest cost relative to alternatives, at the given state xt. The second term refers to the “goods-related” component and describes the probability that selecting the goal i will result in highest reward compared to the alternatives, at the current state xt. Note that the relative desirability values of the alternative options are normalized so that they all sum to 1.
To illustrate the relative desirability function, consider a reaching task with two potential targets presented in left (target L) and right (target R) visual fields (gray circles in Fig 3A). For any state xt where the policy to the right target is more “desirable” than to the left target, we have the following inequality:
r D ( π R ( x t ) ) > r D ( π L ( x t ) ) (17)
This inequality predicts two extreme reaching behaviors—a direct movement to the target R (i.e., winner-take-all) when rD(πR(xt)) > > rD (πL(xt)), and a spatial averaging movement towards an intermediate position between the two targets when rD(πR(xt)) ≈ rD(πL(xt)). Rearranging this equation, using P(reward(L) > reward(R)) = 1 − P(reward(R) > reward(L)) we see that the relative desirability to pursue the target R increases with the odds that target L has more reward and lower cost:
r D ( π R ( x t ) ) > r D ( π L | x t ) ) → ( P ( r e w a r d ( R ) > r e w a r d ( L ) ) 1 - P ( r e w a r d ( R ) > r e w a r d ( L ) ) ) ( P ( c o s t ( R ) < c o s t ( L ) | x t ) 1 - P ( c o s t ( R ) < c o s t ( L ) | x t ) ) > 1 (18)
To gain more insight on how action cost and good value influence the reaching behavior, we visualize the relative desirability to reach the right target in 3 scenarios (the desirability related to the left target is a mirror image of the right one):
For this case,
P ( r e w a r d ( R ) > r e w a r d ( L ) ) = 1 - P ( r e w a r d ( R ) > r e w a r d ( L ) )
which means target R is more desirable when
P ( c o s t ( R ) > c o s t ( L ) | x t ) < 0 . 5
Now the action cost (and hence relative desirability) is a function of the hand-state, making them difficult to illustrate. For a point-mass hand in 2D, the hand state is captured by the 4D position-velocity. To visualize this 4D relative desirability map in two dimensions, we “slice” through the 4D position-velocity space by making velocity a function of position in the following way. All trajectories are constrained to start at position (0, 0) with zero velocity. We then allow the trajectory to arrive at one of a set of spatial positions (100 total) around a circle of radius 85% of the distance between the start point and the midpoint (black star) of the two potential targets. For each of these points, we constrain the hand velocity to have direction (red arrows in Fig 3A) in line with the start point (gray square in Fig 3A) and the hand position on the circle (black dots in Fig 3A). We set the magnitude of the velocities to match the speed of the optimal reaching movement at 85% of completion (blue trace for left target in Fig 3A). From each position-velocity pair on the circle, we sample 100 optimal movements to each of the two targets (solid and discontinuous traces are illustrated examples for reaching the left and the right target, respectively). We discretize the space and compute the action cost to reach the targets from each state—the expected cost from each state to the goal following the policy for that goal, including an accuracy penalty at the end of the movement. Fig 3B and 3C depict these action costs, where blue indicates low cost and red indicates high cost, respectively. Fig 3D illustrates the action costs converted into relative desirability values to reach the right target (indicted by a solid gray circle), where blue and red regions correspond to states with low and high desirability, respectively. Notice desirability increases rapidly as the reach approaches a target, resulting in winner-take-all selection of an action-plan once moving definitely towards a target. However, when the hand position is about the same distance from both targets (greenish areas) there is no dominant policy, leading to strong competition and spatial averaging of the competing policies.
In this case, desirability also depends on the probability of reward. Since both targets provide the same amount of reward, but with different probabilities, the goods-related term simplifies:
P ( r e w a r d ( R ) > r e w a r d ( L ) ) = P ( t a r g e t = R ) = p R
where pR describes the probability of earning reward by pursuing the right target. Hence, the target R is more desirable in a state xt when
P ( c o s t ( R ) > c o s t ( L ) | x t ) < p R
The relative desirability function for the right target is illustrated in Fig 3E, when pR is 4 times higher than the probability of the left target (pR = 0.8, pL = 0.2). The right target is more desirable for most states (reddish areas), unless the hand position is already nearby the left target (blue areas), predicting frequent winner-take-all behavior -i.e., direct reaches to the right target.
More generally, the reward magnitude attached to each target is not fixed, but both the reward magnitude and reward probability vary. We assume that target j provides a reward with probability pj, and that the magnitude follows a Normal distribution with mean μj and standard deviation σj. Hence, the distribution of the rewards attached to the left target (L) and right target (R) is a mixture of distributions:
r e w a r d ( L ) ∼ ( 1 - p L ) δ ( r e w a r d ( L ) ) + p L N ( μ L , σ L 2 ) (19)
r e w a r d ( R ) ∼ ( 1 - p R ) δ ( r e w a r d ( R ) ) + p R N ( μ R , σ R 2 ) (20)
where δ is the Dirac function.
In visuomotor decision tasks, the ultimate goal is usually to achieve the highest reward after N trials. In this case, the probability that the right target provides overall higher reward than the left one over N trials can be approximated by a logistic function l with argument pR μR − pL μL (see Materials and Methods section for more details). When the reward values are precisely encoded, this simplifies to:
P ( r e w a r d ( R ) > r e w a r d ( L ) ) ≈ l ( p R μ R - p L μ L ) (21)
Hence, pursuing the target R is more desirable in a state xt when
P ( c o s t ( R ) > c o s t ( L ) | x t ) < l ( p R μ R - p L μ L ) (22)
Fig 3F illustrates the heat map of the relative desirability values at different states of the policy to reach the right target (solid gray circle), when both targets have the same reward probability pL = pR = 0.5, but μR = 4μL, i.e. reward(R) ∼ 0.5δ(reward(R)) + 0.5N(2, 1) and reward(L) ∼ 0.5δ(reward(L)) + 0.5N(0.5, 1). Similar to the previous scenario, reaching behavior is dominated mostly by the goods-related component and consequently reaching the right target is more desirable than reaching the left target for most states (reddish areas), leading frequently to “winner-take-all” behavior.
Several studies have shown that reaching decisions made while acting follow a “delay-and-mix” policy, with the mixing affected by target configuration and task properties [11, 12, 22, 23]. Subjects were trained to perform rapid reaching movements either to a single target or to two equidistant, equiprobable targets (i.e., actual target location is unknown prior to movement onset in two-target trials). Black and green traces in Fig 4A show single-target trials, characterized by trajectories straight to the target location. Red and blue traces show the delay-and-mix policy for reaches in two-target trials—an initial reaching movement towards an intermediate position between the two stimuli followed by corrective movements after the target was revealed. Relative desirability predicts this behavior (Fig 3D), for equiprobable reward (scenario 1). In this case, the relative desirability is determined solely by the distance from the current hand position to the targets. Since targets are equidistant, the reaching costs are comparable and hence the two competing policies have about the same desirability values for states between the origin and the target locations (see the greenish areas in Fig 3D). Hence, the weighted mixture of policies produces spatial averaging trajectories (red and blue traces in Fig 4E). Note that each controller i, which is associated with the potential target i, generates an optimal policy πi(xt) to reach that target starting from the current state xt. On single-target trials, the actual location of the target is known prior to movement onset and hence the desirability is 1 for the cued target. Consequently the simulated reaches are made directly to the actual target location (green and black traces in Fig 4E).
The competition between policies is also modulated by spatial location of the targets [12]. When one of the targets was shifted, reaching trajectories shifted towards a new intermediate position Fig 4B. This behavior is also captured by our framework—perturbing the spatial distribution of the potential targets, the weighted policy is also perturbed in the same direction Fig 4F. This finding is somehow counterintuitive, since the targets are no longer equidistant from the origin and it would be expected that the simulated reach responses would be biased towards the closer target. However, the magnitude of the perturbation is too small to change the action costs enough to significantly bias the competition. More significant are the action costs required to change direction once the target is revealed, and these costs are symmetric between targets.
Reaching behavior is also influenced by goods-related decision variables, like target probability. When subjects were informed that the potential targets were not equiprobable, the reach responses were biased towards the target with the highest reward probability [11]. This finding is consistent with relative desirability predictions in scenario 2—targets with higher reward probabilities are more desirable than the alternative options for most of the states. Reward probabilities learned via feedback can also be modeled in the same framework. Instead of informing subjects directly about target probabilities, the experimenters generated a block of trials in which one of the targets was consecutively cued for action [22]. Subjects showed a bias towards the cued target that accumulated across trials (Fig 4C) consistent with probability learning. We modeled this paradigm by updating the reward probability using a simple reinforcement learning algorithm (see S4 Text for more details). In line with the experimental findings, the simulated reach responses were increasingly biased to the target location that was consecutively cued for action on the past trials, Fig 4G.
Unlike most value computation methods, our approach can make strong predictions for what happens when additional targets are introduced. A previous study showed that by varying the number of potential targets, reaching movements were biased towards the side of space that contains more targets [12], Fig 4D. Our approach predicts this effect due to normalization across policies. When there are more targets in one hemifield than the other, there are more alternative reaching policies towards this space biasing the competition to that side, Fig 4H. Overall, these findings show that weighting individual policies with the relative desirability values can explain many aspects of human behavior in reaching decisions with competing goals.
A good theory should predict not only successful decisions, but also decisions that result in errors in behavior. Experimental studies provide fairly clear evidence that humans and animals follow a “delay-and-mix” behavior even when it appears pathological. A typical example is the “global effect” paradigm that occurs frequently in oculomotor decisions with competing goals. When two equally rewarded targets are placed in close proximity—less than 30° angular distance—and the subject is free to choose between them, saccade trajectories usually end on intermediate locations between targets [24, 34, 35]. To test whether our theory can capture this phenomenon, we modeled the saccadic movements to individual targets using optimal control theory (see S2 Text for more details) and ran a series of simulated oculomotor decision tasks. Consistent with the experimental findings, the simulated eye movements land primarily in a position between the two targets for 30° target separation (gray traces in Fig 5A), whereas they aim directly to one of them for 90° target separation (black traces in Fig 5A).
We visualize the relative desirability of the left target (i.e., desirability to saccade to the left target) at different states, both for 30° and 90° target separation. We followed a similar procedure as for the reaching case but used an ellipse. Particularly, individual saccadic movements are constrained to start at (0, 0) and arrive at one of the sequence (100 total) of spatial positions with zero velocity around an ellipse with center intermediate between the two targets (black star), with minor axis twice the distance between the origin and the center of the ellipse, and major axis double the length of the minor axis (Fig 5B). For each position on the ellipse, we generate 100 optimal saccadic movements and evaluate the relative desirability to saccade to the left target (solid gray circle) at different states. Fig 5C depicts the heat-map of the relative desirability for 30° target separation. The black traces represent the average trajectories for direct saccadic movements, when only a single target is presented. Notice that regions defined by the starting position (0, 0) (gray square) and the locations of the targets is characterized by states with strong competition between the two saccadic policies (greenish areas). Consequently the weighted mixture of policies results frequently in spatial averaging movements that land between the two targets. On the other hand, when the targets are placed in distance, such as the 90° case presented in Fig 5D, the targets are located in areas in which one of the policies clearly dominates the other, and therefore the competition is easily resolved.
Fig 5E shows examples of saccadic movements (left column) with the corresponding time course of relative desirability values to saccade to the left and the right target (right column). The first two rows show trials from the 30° target separation task, where the competition between the two saccadic policies results in global effect (upper panels) and saccadic movement to the right target (middle panels). The two policies have about the same relative desirability values at different states resulting in a strong competition. Because saccades are ballistic with little opportunity for correction during the trajectory, competition produces the global effect paradigm. However, if the competition is resolved shortly after saccade onset, the trajectory ends up to one of the targets. On the other hand, when the two targets are placed in distance, the competition is easily resolved and the mixture of the policies generates direct movements to one of the targets (lower panel).
These findings suggest that the competition between alternative policies depends on the geometrical configuration of the targets. We quantified the effects of the targets’ spatial distribution to eye movements by computing the percentage of averaging saccades against the target separation. The results presented in Fig 5F (red line) indicate that averaging saccades were more frequent for 30° target separation and fell off gradually as the distance between the targets increases (see the Discussion section for more details on how competition leads to errors in behavior). This finding is also in line with experimental results from an oculomotor decision study with express saccadic movements in non-human primates (green, blue and cyan lines in Fig 5F describe the performance of 3 monkeys [24]).
In previous sections we considered decisions between multiple competing goals. However, ecological decisions are not limited only to simultaneous goals, but often involve choices between goals with time-dependent values. Time-dependent values mean that some of the goals may spoil or have limited period of worth such that they must be reached within a time window or temporal order. A characteristic example is sequential decision tasks that require a chain of decisions between successive goals. Substantial evidence suggests that the production of sequential movements involves concurrent representation of individual policies associated with the sequential goals that are internally activated before the order is imposed upon them [25, 36–39]. To model these tasks using our approach, the critical issue is how to mix the individual control policies. State-dependent policy mixing as described previously will dramatically fail, since the desirability values do not take into account the temporal constraints. However, it is relatively easy to incorporate the sequential constraints and time-dependence into the goods-related component of the relative desirability function. We illustrate how sequential decision tasks can be modeled using a simulated copying task used in neurophysiological [25, 40] and brain imaging studies [41, 42].
Copying geometrical shapes can be conceived as sequential decisions with goal-directed movements from one vertex (i.e., target) of the shape to another in a proper spatial order. To model this, each controller j provides a policy πj to reach the vertex j starting from the current state. We encode the order of the policies using a time-dependent target reward probability p(vertex = j|xt) that describes the probability that vertex j is the current goal of the task at state xt (see Materials and Methods section for more details). In fact, it describes the probability to copy the segment defined by the successive vertices j − 1 and j at a given state xt.
We evaluated the theory in a simulated copying task with 3 geometrical shapes (i.e., equilateral triangle, square and pentagon). Examples of movement trajectories from the pentagon task is shown in Fig 6A. Fig 6B depicts the time course of the relative desirability values of the segments from a successful trial. The desirability of each segment peaks once the model starts copying that segment and falls down gradually, whereas the desirability of the following segment starts rising while copying the current segment. Notice that the competition is stronger for middle segments than the first or the last segment in the sequence. Consequently, errors, such as rounding of corners and transposition errors (i.e., copying other segments than the current one in the sequence) are more frequent when copying the middle segments of the shape, than during the execution of the early or late segments. These simulation results are congruent with studies showing that human/animal accuracy in serial order tasks is better during early or late elements in the sequence [25, 43]. A characteristic example is illustrated in Fig 6C, in which the competition between copying the “blue” and the “green” segments resulted in an error trial.
Notice also that the temporal pattern of desirability values is congruent with populations of neural activity in prefrontal cortex during the copying task that encode each of the segments [25]. The strength of the neuronal population corresponding to a segment predicted the serial position of the segment in the motor sequence, providing a neural basis for Lashley’s hypothesis. Interestingly, the temporal evolution of the population activities resembles the temporal evolution of the relative desirabilities of policies in our theory. This finding provides a direct neural correlate of relative desirability suggesting that the computations in our model are biologically plausible. Finally, Fig 6D and 6E illustrate examples of movement trajectories for copying an equilateral triangle and a square.
How the brain dynamically selects between alternatives challenges a widely used model of decisions that posit comparisons of abstract representations of “goods” [1–7]. According to this model, the brain integrates all the decision variables of an option into a subjective economic value and makes a decision by comparing the values of the alternative options. Most importantly, the comparison is taking place within the space of goods, independent of the sensorimotor contingencies of choice [5]. While abstract representation of values have been found in brain areas like orbitrofrontal cortex (OFC) and ventromedial prefrontal cortex (vmPFC) [4, 44], these representations do not necessarily exclude the involvement of sensorimotor areas in decisions between actions. Recent studies provide evidence for an “action-based” theory involving competition between concurrent prepared actions associated with alternative goals [9, 10, 12–14, 17]. The main line of evidence of this theory is recent findings from neurophysiological studies [15, 16, 45, 46] and studies that involve reversible inactivation of sensorimotor regions [47, 48]. According to these studies, sensorimotor structures, such as the lateral intraparietal area (LIP) [48], the dorsal premotor cortex (dPM) [16], the superior colliculus (SC) [47] and the parietal reach region (PRR) [45, 46] are causally involved in decisions.
Despite the attractiveness of the “action-based” theory to model decisions between actions, what has been missing is a computational theory that can combine good values (e.g., money, juice reward) with action costs (e.g., amount of effort) into an integrated theory of dynamic decision-making. Previous studies have used principles from Statistical Decision Theory (SDT) to model human behavior in visuomotor decisions [49]. According to these studies, action-selection can be modeled as a decision problem that maximizes the desirableness of outcomes, where desirableness can be captured by an expected gain function. Despite the significant contribution of these studies to the understanding of the mechanisms of visuomotor decisions, they have focused mostly on static environments, in which the availability and the value of an option do not change with time and previous actions. Additionally, the expected gain functions usually involve the integration of decision values that have the same currency, such as expected monetary gains and losses—e.g., humans perform rapid reaching movements towards displays with regions that, if touched within a boundary lead to monetary reward, otherwise to monetary penalty [50, 51]. In the current study, we propose a probabilistic model that shows how value information from disparate sources with different “currencies” can be integrated in a manner that is both online and can be updated during action execution. The model is based on stochastic optimal control theory and is consistent with the view that decision and action are merged in a parallel rather than serial order. It is comprised of a series of control schemes that each of them is attached to an individual goal and generates a policy to achieve that goal starting from the current state. The key to our model is the relative desirability value that integrates the action costs and good values to a single variable that weighs the individual control policies as a function of state and time. It has intuitive meaning of the probability of getting the highest pay-off with the least cost following a specific policy at a given time and state. Because the desirability is state- and time- dependent, the weighted mixture of policies produces a range of behavior automatically, from “winner-take-all” to “weighted averaging”. By dynamically integrating in terms of probabilities across policies, relative-desirability varies with decision context. Relative desirability’s effective exchange rate changes whenever action costs increase or decrease, the set of options change, or the value of goods increase. Moreover, relative desirability is dynamic and state-dependent, allowing for dynamic changes in the effective exchange rate between action costs and the good values. We believe these properties are critical for maintaining adaptability in a changing environment. Throughout our evolutionary history, new opportunities and dangers constantly present themselves, making a fixed exchange rate between action costs and good value maladaptive.
The proposed computational framework can be conceived as analogous to classical value-comparison models in decision making, such as the drift diffusion model (DDM) [52] and the leaky competing accumulator (LCA) model [53], but for decisions that require continuous evaluation of in-flowing value information during ongoing actions. In the standard version of these models, choosing between two options is described by accumulator-to-threshold mechanisms. Sensory evidence associated with each alternative is accumulated, until the integrated evidence for one of them reaches a decision threshold. Despite the success of these frameworks to model a variety of decision tasks, they are difficult to extend beyond binary choices, require a pre-defined decision threshold and are mainly applied in perceptual decisions, in which decision precedes action. Unlike these models, the proposed computational theory can model decisions between multiple alternatives that either are presented simultaneously or sequentially, does not require any pre-defined decision threshold and can handle tasks in which subjects cannot wait to accumulate evidence before making a choice. The relative desirability integrates dynamically both sensory and motor evidence associated with a particular policy and reflects the degree to which this policy is best to follow at any given time and state with respect to the alternatives.
We tested our theory in a series of visuomotor decision tasks that involve reaching and saccadic movements and found that it captures many aspects of human and animal behavior observed in recent decision studies with multiple potential targets [11, 12, 21–23]. In line with these studies, the theory predicts the “delay-and-mix” behavior, when the competing goals have about the same good values and action costs and the “pre-selection” behavior, when one of the alternative goals is clearly the best option.
The present computational theory bears some similarities with the Hierarchical Reinforcement Learning (HRL) models used extensively in decision-making studies [54]. According to HRL theory, decision-making takes place at different level of abstractions, where the higher levels select the best current goal and the lower levels generate the optimal policy to implement the choice. Although, the HRL implements the dynamic aspects of decision-making by re-evaluating the alternative options and selecting the best one at a given time and state, there are two fundamental differences with the present theory. First, HRL always selects the best policy and typically pursues it until all the actions in the sequence have been performed. On the other hand, our computational theory generates a weighted average of the alternative policies and executes only part of it before re-evaluating the alternative option (i.e., see S5 Text about the “receding horizon control” theory). Second, the HRL uses a softmax transformation to evaluate the alternative options, whereas the proposed computational theory uses both the expected reward and the effort cost associated with each alternative. Additionally, other similar modular frameworks consisting of multiple control systems, such as the MOSAIC model [55] and the Q-decomposition framework [56], have been previously proposed to model tasks with multiple goals. However, these frameworks do not incorporate the idea of integrating both the good values and action costs into the action selection process. Hence, they fail to make predictions on how value information from disparate sources influences the motor competition and how this competition can lead to erroneous behavior.
We developed our model for cases where the competing options are similar. These are also cases where the relative effort and reward desirabilities are similar. For two options, it means the relative desirabilities would be far from zero or one. Here we consider extreme situations where one option requires much more effort or supplies much less reward. For extreme cases, the relative desirability calculation appears to break down and produces an “indeterminate” form for each alternative option. Here we explain why that happens, and how the indeterminacy is avoided by adding even a tiny amount of noise in implementing the calculation.
To illustrate the indeterminacy, consider selecting between an “extremely hard but very rewarding” and an “extremely easy but unrewarding” option. The hard option offers significantly higher reward than the easy option reward(Hard) > > reward(Easy), but it requires significantly higher effort to get it than the easy one cost(Hard) > > cost(Easy). According to the definition of the relative desirability, the reward-related component of the desirability will approach 1 for the hard option and 0 for the easy option, since P(reward(Hard) > reward(Easy)) = 1. On the other hand, the effort-related component of the desirability will be 0 for the hard option and 1 for the easy option, since P(cost(Hard) > cost(Easy)) = 1. The relative reliability multiplies these values and renormalizes, leading to the indeterminate form rD(option(1)) = 0*1/(0*1+1*0) = 0/0 and rD(option(2)) = 1*0/(0*1+1*0) = 0/0. In this case the model apparently fails to make a coherent choice. As long as the probability formula for reward and effort are continuous mappings, this indeterminacy will only be experienced in the limit that one option is infinitely harder to get (inaccessible) while the accessible option is comparably worthless.
However, the indeterminacy is an extreme example of an important class of problems where effort and reward values for the two options are in conflict with each other. Because there is a trade-off associated with reward vs effort neither option is clearly better than the other. While none of the decisions modeled here have extreme conflict, we nevertheless believe that the indeterminacy described above will never occur in a biological decision-making system due to the effects of even tiny amounts of noise on the relative desirability computation. If we assume that desirability values are the brain’s estimate of how “desirable” one option is with respect to alternatives in terms of expected outcome and effort cost, then it is reasonable to assume these estimates are not always precise. In other words, biological estimates of desirability should manifest stochastic errors, which we model by including noise in the estimates. In the S6 Text we show the effect of this noise is profound. For the extreme scenario in which P(reward(Hard) > reward(Easy)) = 1 and P(cost(Hard) > cost(Easy)) = 1, in the presence of noise the relative desirability of each option is 0.5. Thus, indeterminacy produces a lack of preference—since the “easy” option dominates the “hard” option in terms of effort, but the “hard” option is better than the “easy” option in terms of reward. In general, cases with extreme conflict will produce lack of preference, but these cases are also unstable—small changes in factors affecting the valuation such as the internal states of the subject (e.g., hunger level, fatigue level) can produce large shifts in preference. In the S6 Text, we further discuss the effects of noise in decisions with multiple options.
One of the key assumptions in our study is that the brain continuously evaluates the relative desirability—i.e., the probability that a given policy will result in the highest pay-off with the least effort—in decisions with competing options. Although this idea is novel, experimental studies provide evidence that the brain maintains an explicit representation of “probability of choice” when selecting among competing options (for a review see [9]). For binary perceptual decisions, this probability describes the likelihood of one or another operant response, whereas for value-based decisions it describes the probability that selecting a particular option will result in the highest reward. Classic experimental studies reported a smooth relationship between stimulus parameters and the probability of choice suggesting that the brain translates value information to probabilities when making decisions [57, 58]. Additionally, neurophysiological recordings in non-human primates revealed activity related to the probability of choice in the lateral intraparietal area (LIP) both in “two-alternative force-choice eye movement decisions” and in “value-based oculomotor decisions”. In the first case, the animals performed the random-dot motion (RDM) direction discrimination task while neuronal activity was recorded from the LIP [59]. The activity of the LIP neurons reflects a general decision variable that is monotonically related to the logarithm of the likelihood ratio that the animals will select one direction of motion versus the other. In classic value-based decisions, the animals had to select between two targets presented simultaneously in both hemifields [15]. The activity of the LIP neurons is modulated by a number of decision-related variables including the expected reward and the outcome probability. These experimental findings have inspired previous computational theories to model perceptual- and value-based decisions [9]. According to these studies, when the brain is faced with competing alternatives, it implements a series of computations to transform sensory and value information into a probability of choice. The proposed idea of the relative desirability value can be conceived as an extension of these theories taking into account both the expected reward and the expected effort related to a choice.
One of the novelties of this theory is that it predicts not only successful decisions, but decisions that result in poor or incorrect actions. A typical example is the “global effect” paradigm that occurs frequently in short latency saccadic movements. When the goal elements are located in close proximity and subjects are free to choose between them, erroneous eye movements usually land at intermediate locations between the goals [24, 35]. Although the neural mechanisms underlying the global effect paradigm have not been understood fully yet, the prevailing view suggests that it occurs due to unresolved competition between the populations of neurons that encode the movements towards the two targets. Any target in the field is represented by a population of neurons that encodes the movement direction towards its location as a vector. The strength of the population is proportional to the saliency (e.g., size, luminance) and the expected pay-off of the target. When two similar targets are placed in close proximity, the populations corresponding to them will be combined to one mean population with the direction of the vector towards an intermediate location. If one of the targets is more salient or provide more reward than the other, the vector is biased to this target location. Since subjects have to perform saccadic movements to one of the targets, the competition between the two populations has to be resolved in time by inhibiting one of them. The time to suppress the neuronal activity that encodes one of the alternatives may be insufficient for short latency saccades resulting in averaging eye movements. Our findings are consistent with this theory. The strength of the neuronal population is consistent with relative desirability of the policy that drives the effector directed to the target. When the two equally rewarded targets are placed in close proximity, the two policies generate similar actions. Given that both targets are attached with the same goods-related values, the relative desirability of the two policies are about the same at different states, resulting in a strong competition. Because saccades are ballistic with little opportunity for correction during movement, the competition produces averaging saccades. On the other hand, placing the two targets in distance, the two saccadic policies generate dissimilar actions and consequently the competition is easier to be resolved in time.
Competition between policies in closely aligned goals can also explain errors in sequential decision tasks that involve serial order movements as described by Lashley [36]. The key idea in Lashley’s pioneer work (1951) is that the generation of serial order behavior involves the parallel activation of sequence of actions that are internally activated before each of the actions are executed. The main line of evidence of this hypothesis was the errors that occur frequently in serial order tasks, such as speech [37], typing [38], reaching [39] and copying of geometrical shapes [25]. For instance, a common error in typing and speaking is to swap or transpose nearby letters, even words. Lashley suggested that errors in sequential tasks would be most likely to occur when executing nearby elements within a sequence. Recent neurophysiological studies provide the neural basis of the Lashley’s hypothesis showing that the serial characteristics of a sequence of movements are represented in an orderly fashion in the prefrontal cortex, in time before the start of drawing [25, 40]. Training monkeys to copy geometrical shapes and recording the activity of individual neurons in the prefrontal cortex, the experimenters were able to identify populations of neurons that encode each of the segments [25]. The strength of the neuronal population corresponding to a segment predicted the serial position of the segment in the motor sequence. Interestingly, the temporal evolution of the strength of the segment representation during the execution of the trajectories for copying the shapes resembles the temporal evolution of the relative desirabilities of policies in our theory. This finding suggests that the strength of the neuronal population of a particular segment may encode the relative desirability (or components of the desirability) of copying that segment at a given time with respect to the alternatives. This hypothesis is also supported by error analysis in the serial order tasks, which showed that errors more frequently occurred when executing elements with nearly equal strength of representation. In a similar manner, our theory predicts that when two policies have about equal relative desirabilities over extended periods of the movement, the competition between them may lead to errors in behavior.
Finally, our theory provides a conceptual alternative in understanding important aspects of neurological disorders that cause deficits in choice behavior, such as the spatial extinction syndrome. This syndrome is a subtle form of hemispatial neglect that occurs frequently after brain injury. It is characterized by the inability to respond to stimuli in the contralesional hemifield, but only when a simultaneous ipsilesional stimulus is also presented [60]. Recent studies reported contralesional bias that reminiscent the extinction syndrome, in oculomotor decision tasks after reversible pharmacological inactivation of the LIP [48] and the Pulvinar [61] in monkeys. According to our theory, this effect could be related to a deficit in value integration after inactivation, rather than simply sensory attention deficit.
In sum, decisions require integrating both good values and action costs, which are often time and state dependent such that simple approaches pre-selection of goals or fixed weighted mixture of policies cannot account for the complexities of natural behavior. By focusing on a fundamental probabilistic computation, we provide a principled way to dynamically integrate these values that can merge work on decision making with motor control.
|
10.1371/journal.ppat.1004901 | Varicella Viruses Inhibit Interferon-Stimulated JAK-STAT Signaling through Multiple Mechanisms | Varicella zoster virus (VZV) causes chickenpox in humans and, subsequently, establishes latency in the sensory ganglia from where it reactivates to cause herpes zoster. Infection of rhesus macaques with simian varicella virus (SVV) recapitulates VZV pathogenesis in humans thus representing a suitable animal model for VZV infection. While the type I interferon (IFN) response has been shown to affect VZV replication, the virus employs counter mechanisms to prevent the induction of anti-viral IFN stimulated genes (ISG). Here, we demonstrate that SVV inhibits type I IFN-activated signal transduction via the JAK-STAT pathway. SVV-infected rhesus fibroblasts were refractory to IFN stimulation displaying reduced protein levels of IRF9 and lacking STAT2 phosphorylation. Since previous work implicated involvement of the VZV immediate early gene product ORF63 in preventing ISG-induction we studied the role of SVV ORF63 in generating resistance to IFN treatment. Interestingly, SVV ORF63 did not affect STAT2 phosphorylation but caused IRF9 degradation in a proteasome-dependent manner, suggesting that SVV employs multiple mechanisms to counteract the effect of IFN. Control of SVV ORF63 protein levels via fusion to a dihydrofolate reductase (DHFR)-degradation domain additionally confirmed its requirement for viral replication. Our results also show a prominent reduction of IRF9 and inhibition of STAT2 phosphorylation in VZV-infected cells. In addition, cells expressing VZV ORF63 blocked IFN-stimulation and displayed reduced levels of the IRF9 protein. Taken together, our data suggest that varicella ORF63 prevents ISG-induction both directly via IRF9 degradation and indirectly via transcriptional control of viral proteins that interfere with STAT2 phosphorylation. SVV and VZV thus encode multiple viral gene products that tightly control IFN-induced anti-viral responses.
| In this manuscript we demonstrate that the immediate early protein ORF63 encoded by varicella zoster virus (VZV) and simian varicella virus (SVV) interferes with interferon type I-mediated activation of JAK-STAT signaling and thereby inhibits the expression of interferon stimulated genes. ORF63 blocks this pathway by degrading IRF9, which plays a central role in JAK-STAT signaling. In addition, both viruses code for immune evasion mechanisms affecting the JAK-STAT pathway upstream of IRF9, which results in the inhibition of STAT2 phosphorylation. By fusing a degradation domain derived from dihydrofolate reductase (DHFR) to ORF63 we further demonstrate that this protein is essential for SVV growth and gene expression, indicating that ORF63 also affects IFN-signaling indirectly by regulating the expression of other immune evasion genes.
| The alphaherpesvirus varicella zoster virus (VZV) is the causative agent of chickenpox. After primary infection, VZV establishes latency in sensory ganglia. Reactivation from latency, which typically occurs in elderly individuals, can cause shingles or herpes zoster that is associated with a number of debilitating complications, including postherpetic neuralgia [1]. In vivo research on VZV is limited because the virus does not produce varicella or zoster in animals [2, 3]. Simian varicella virus (SVV) is closely related to VZV sharing about 75% DNA homology and exhibiting a highly similar genome organization [4]. Inoculation of nonhuman primates, including African green monkeys and Cynomolgus macaques, results in a persistent viremia [4]. In contrast, infection of rhesus macaques (RM) with SVV results in a primary infection followed by latency that is similar to VZV infection in humans. SVV-induced skin lesions are resolved by 21 days post infection which correlates with the absence of virus DNA in blood. Latent SVV can be detected in ganglia of infected RM [5]. Infection of RM with SVV thus represents a robust animal model that recapitulates most hallmarks of a primary human VZV infection.
The innate host immune response to viral infection is dominated by interferons (IFNs) that are subdivided in three families, namely types I, II and III. In particular, several subtypes of IFNα and IFNβ that represent type I IFNs are key players in the anti-viral innate immune response [6]. Transcription of IFN is initiated by pattern recognition receptors (PRRs) engaging pathogen associated molecular patterns (PAMPs) such as double-stranded RNA, lipopolysaccharide and cytosolic DNA. Downstream signaling pathways lead to the activation of transcription factors such as IFN regulatory factor (IRF) 3 and nuclear factor κB (NFκB) that induce the transcription of IFNβ. Secreted IFNβ can signal in an autocrine and paracrine fashion by interacting with the type I IFN receptor complex (consisting of IFNAR1 and IFNAR2) both on infected and neighboring uninfected cells [7]. Receptor binding activates the JAK-STAT signaling pathway, which results in the expression of hundreds of IFN-stimulated genes (ISG) and corresponding proteins that inhibit virus growth by counteracting multiple molecular steps of the replication cycle and by signaling to innate immune cells including natural killer cells [8, 9]. In addition, type I IFNs have been shown to be involved in dendritic cell maturation and antigen presentation thereby stimulating the development of virus-specific adaptive immune responses [10, 11].
The ability of VZV and SVV to spread and establish latency in the presence of these immediate immune responses implies that both viruses display evasion strategies that circumvent or counteract the induction or function of IFNs and ISGs. VZV infection of human skin xenografts in severe combined immunodeficiency (SCIDhu) mice showed that VZV-infected cells do not express type I IFN, while uninfected bystander cells stained positive for the cytokine [12]. Several reports have shown that the effect of IFN is counteracted by at least four different VZV-encoded proteins: both IE62 and ORF47 alter the phosphorylation of IRF3 and prevent gene activation [13, 14], whereas ORF61 inhibits pathogen-induced cytokine expression by degrading IRF3 [15] and by blocking the activation of NFκB via inhibiting IκBα [16]. In addition, Cohen et al. showed the deletion of ORF63 severely attenuates the growth of the virus in the presence of IFNα but not IFNγ, suggesting a possible involvement of ORF63 in regulating JAK-STAT signaling [17].
Expression of IFNα and IFNβ and subsequent binding to their receptor triggers the dimerization of IFNAR1 and IFNAR2, which activates the Janus kinases JAK1 and TYK2 that are constitutively associated with the receptor. The kinases phosphorylate the receptor creating a docking site for the transcription factors STAT1 and STAT2. Subsequent phosphorylation of the STAT proteins by the JAKs leads to conformational changes within the molecules that allow the formation of stable STAT1, STAT2 and IFN regulatory factor 9 (IRF9) complex termed ISGF3. ISGF3 then shuttles to the nucleus where it activates the transcription of type I IFNs and ISGs [18]. VZV-infected cells in skin xenografts in the SCIDhu mice did not show nuclear localization or phosphorylation of STAT1, in contrast to uninfected bystander cells [12]. Since IFNα produced by bystander cells can induce JAK-STAT signaling in VZV-infected cells, the absence of pSTAT1 in infected cells suggests that VZV interferes not only with IRF3-mediated activation of IFN transcription but also with JAK/STAT signaling.
The VZV ORF63 protein is a 30 kDa immediate early protein that is phosphorylated by host and viral kinases [19, 20]. The protein is abundantly present during lytic infection and its expression has also been observed in latently infected ganglia [21, 22]. A duplicate of the ORF63 gene, designated ORF70, is found in the terminal repeat region [19]. VZV ORF63/70 regulates viral gene expression and impairs the expression of certain cellular genes [23–26]. In addition, VZV ORF63/70 inhibits apoptosis in cultured primary human neurons [27]. SVV encodes both ORF63/70 orthologous proteins that share 52% amino acid identity with their VZV homologs [28]. These gene products are required for replication of SVV in cell culture [29]. In vivo studies using SVV-infected RM and African green monkeys confirmed expression of the ORF63 protein in ganglia during latent infection [5].
In this report, we show that both SVV and VZV interfere with type I IFN signaling. SVV inhibits type I IFN-induced ISG expression by downregulating the expression of IRF9 and prevents phosphorylation of STAT2 upon IFN stimulation. We also observed a minor decrease in STAT2 levels in SVV-infected cells. SVV ORF63 was found to be responsible for the reduction in IRF9 expression, but was not directly involved in downregulation of STAT2 expression or inhibition of STAT2 phosphorylation. These data suggest that multiple SVV proteins counteract type I IFN signaling. Similarly, we observed reduced levels of STAT2 and IRF9 proteins in VZV-infected cells and the cells were refractory to IFN-induced STAT2 phosphorylation. Ectopic expression of VZV ORF63 affected steady state levels of IRF9, but did not block STAT2 expression or phosphorylation. Thus, the multi-level inhibition of JAK-STAT signaling seen in SVV-infected cells is conserved in VZV-infected cells.
Telomerized rhesus fibroblasts (TRFs), TRF-ISRE cells [30], the African green monkey kidney epithelial cell line Vero, Vero-CRE (kindly provided by Dr. Linda van Dyk, University of Colorado, Denver) [31], telomerized human fibroblasts (THF)-ISRE [32], human embryonic kidney (HEK) 293T cells (ATCC), and the human fibroblast cell line MRC-5 (ATCC) were maintained in DMEM supplemented with 10% heat-inactivated fetal bovine serum (FBS), 140 IU of penicillin/ml and 140 μg of streptomycin/ml.
TRFs and Vero cells were infected with a recombinant SVV Delta strain, in which eGFP was inserted between US2 and US3 through homologous recombination [33]. A confluent monolayer of TRF or Vero cells was infected by cocultivation of SVV.eGFP-infected cells with uninfected cells at the indicated ratios. Complete infection was confirmed by visualizing eGFP using fluorescence microscopy. Infected cells were maintained in DMEM supplemented with 2% FBS and harvested at 48 hours p.i. For VZV infections, we used the recombinant VZV Oka strain, in which eGFP was fused to the N-terminus of ORF66 (generously provided by P.R. Kinchington, University of Pittsburgh, Pennsylvania) [34]. VZV.eGFP-infected MRC-5 cells were cocultivated with uninfected MRC-5 cells at a 1:5 ratio.
Rhesus IFNα2 was obtained from R&D systems. Human IFNα and universal type I IFN (uIFN) were obtained from PBL Assay Science. MG132 (Fisher Scientific) was dissolved in DMSO and used at the indicated concentrations for 16 hours. Control wells were treated with same concentration of DMSO without MG132. Trimethoprim (TMP; Sigma-Aldrich) was dissolved in DMSO and used 10 μM or less where indicated. Viral cultures were supplemented with fresh TMP every 24 hours. The following antibodies were used for detection of endogenous and viral proteins in western blot: anti-ISG15 F-9 (Santa Cruz), anti-ISG54/IFIT2 (Abcam), anti-Mx-1 (GeneTex), anti-STAT1 M22 (Santa Cruz), anti-phosphorylated STAT1 Tyr701 (Santa Cruz), anti-STAT2 C20 (Santa Cruz), anti-phosphorylated STAT2 Tyr690 (Cell Signaling Technology), anti-IRF9/ISGF3γ clone 6 (BD Biosciences), anti-GAPDH 6C5 (Santa Cruz), anti-p84 5E10 (GeneTex), anti-IRF1 H-205 (Santa Cruz), anti-IRF3 (Santa Cruz) and anti-FLAG M2 (Sigma-Aldrich). The monoclonal antibodies specific for SVV and VZV ORF63 (clone 63_6), ORF62 (clone 62_6) and ORF31 (clone 31C_8) have been previously described [35]. STAT2 C20 (Santa Cruz) was also used for immunofluorescence microscopy.
TRFs were transduced with a replication-defective lentivirus encoding firefly luciferase downstream of an IFN-stimulated response element (ISRE) and a lentivirus constitutively expressing renilla luciferase driven by a CMV promotor (Qiagen). Transduced TRFs (TRF-ISRE cells) were selected by culturing in the presence of 4 μg/ml puromycin. TRF-ISRE cells were infected with SVV.eGFP as described above and 24 hours p.i. the cells were seeded in a black 96 well plate (Corning Incorporated). At 42 hours p.i., the cells were stimulated with rhesus or human IFNα to induce expression of ISRE-driven firefly luciferase. After 6 hours, expression of firefly and renilla luciferase was measured using the Dual-Glo luciferase assay system (Promega). Luminescence was measured on a Veritas microplate luminometer (Promega). Data are presented as the ratio between firefly luciferase expression and renilla luciferase expression. For the experiment described in Fig 1B the cells were sorted for high GFP expression at 40 hours p.i. using a FACS Aria II cytometer. We seeded 2000 of the sorted and mock-infected cells cells in black 96 well plate and stimulated with uIFN for 6 hours, after which luciferase expression was measured using the ONE-Glo luciferase assay system (Promega).
HEK 293T cells were co-transfected with an ISRE-firefly luciferase reporter plasmid (pGL3-ISRE-Luc) and the pcDNA3.1 expression vectors (described below). At 24 and 42 hours post transfection, the cells were treated with 5000 U/ml uIFN for 6 hours. Cells were transferred to a black 96 wells plate before or right after treatment with uIFN treatment. Expression of firefly luciferase was measured using the ONE-Glo luciferase assay system (Promega).
To generate nuclear and cytoplasmic fractions, cells were resuspended in dounce buffer (100 mM KCl, 20 mM Hepes [pH 7.4], 0.1 mM EDTA, 3% sucrose) supplemented with HALT protease and phosphatase inhibitors (Thermo Scientific). After 15 minutes 10% Nonidet P-40 was added to the lysate in a 1:20 ratio and lysates were vortexed for 10 seconds. The cytoplasmic fraction was removed immediately and nuclei were washed twice with D-PBS and lysed in RIPA buffer (10 mM Tris-HCl [pH 7.4], 150 mM NaCl, 1 mM EDTA, 1% Triton X-100, 1% Sodium Deoxycholate, 0.1% SDS) supplemented with protease and phosphatase inhibitors. For all other experiments, cells were lysed directly in Laemmli sample buffer (100 mM Tris-HCL [pH 8.0], 4% SDS, 20% glycerol, 10% 2-mercaptoethanol, Bromophenol blue). Proteins were separated by SDS-page and transferred to polyvinylidene difluoride membranes (Thermo Scientific). Membranes were first incubated with the indicated antibodies, which was followed by incubation with horseradish peroxidase (HRP)-conjugated secondary antibodies specific for mouse (Santa Cruz) or rabbit (Thermo Scientific) IgG. Binding of secondary antibodies to the membranes was visualized by using Pierce ECL 2 (Thermo Scientific).
Mock- or SVV-infected TRFs were grown on cover slips, washed twice with PBS and fixed with 3.7% formaldehyde (Fisher Scientific) at room temperature (RT) for 40 minutes. After washing with PBS, residual formaldehyde was quenched with 50 mM Ammonium Chloride for 10 minutes and the cells were permeabilized with 0.1% Triton for 5 to 7 minutes. Non-specific protein binding sites were blocked with 2% bovine serum albumin (BSA) (Fisher Scientific) and cells were incubated with STAT2-specific antibody in 2% BSA for 1 hour at 37°C. Cells were washed with 2% BSA and incubated with the secondary antibody Alexa Fluor 594 Goat anti-Rabbit (Life Technologies). Cells were then washed with 2% BSA, followed by one PBS wash. Cover slips were mounted on glass slides using Prolong Gold Anti-fade reagent (Cell Signaling). Staining was visualized on a Zeiss Axioskop 2 Plus fluorescence microscope, and images were taken using AxioVision v4.6 software (Zeiss).
The recombinant adenoviruses expressing SVV ORF63 (AdORF63) and VZV AdORF63 were produced as previously described [36]. The vector contains a tetracycline-responsive promoter and requires the addition of a tetracycline-regulated transactivator (tTA) [37], which was provided by co-infecting with AdtTA. TRFs cultured in six-well clusters were co-transduced with the purified ORF63 adenoviruses and AdtTA at the indicated MOI in 0.5 ml of serum-free DMEM. After 1.5 hour of incubation at 37°C, 1.5 ml of DMEM supplemented with 10% FBS was added and incubation continued for a total of 48 hours. Where indicated, doxycycline was added to the infections to regulate tTA-dependent gene expression.
Total RNA was extractedand treated with DNase using the NucleoSpin RNA isolation kit (Machery Nagel) according to the manufacturer's protocol. The concentration of the RNA samples was measured using the NanoDrop 1000 Spectrophotometer (Thermo Scientific). Single-stranded cDNA was made from total RNA using random hexamers (TaKaRa) to prime first-strand synthesis by Maxima Reverse Transcriptase (Thermo Scientific) as recommended by the manufacturer. The induction of ISG54 and Mx-1 mRNA expression upon uIFN stimulation was determined using SYBR green-based semiquantitative real-time RT-PCR (qPCR) using the following primers: ISG54 Fw: 5’-gttactggaactaataggacac-3’, ISG54 Rev: 5’-tggcaagaatggaaca-3’, Mx-1 Fw: 5’-atgatcgtcaagtgccg-3’, Mx-1 Rev: 5’-gccttgccttcctcca-3’. SVV ORF63 expression was confirmed using the primers ORF63 Fw: 5’-CAGCGTCCTACAGTGAC-3’ and ORF63 Rev: 5’-GTTGCTGGTAGCATCATC-3’. Levels of IRF9 mRNA were determined using the primers IRF9 Fw: 5’- TACCCGAAAACTCCGGAAC-3’ and IRF9 Rev: 5’-AAGAAGGCAGCATCCTGG-3’. Levels of STAT2 mRNA were determined using the primers STAT2 Fw: 5’- ATGGCGCAGTGGGAAATG-3’ and STAT2 Rev: 5’-ctgccagttctggtcttc-3’. Reactions were performed using SYBR green PCR core reagents and Platinum Taq DNA Polymerase (Invitrogen). Relative expression of ISG54, Mx-1, and ORF63 was calculated using the method described by Livak and Schmittgen [38]. GAPDH was used as a housekeeping gene to establish a baseline against which target genes were compared (Fw: 5’-GCACCACCAACTGCTTAGCAC-3’, Rev: 5’- TCTTCTGGGTGGCAGTGATG-3’). For IRF9 and STAT2 mRNA expression we calculated delta cycle threshold (ΔCt) by subtracting background Ct (GAPDH) from the Ct value for IRF9/STAT2.
Purified DNA from TRFs infected with SVV.eGFP was used as a template for PCR amplification of SVV ORF63. PCR was performed using AccuPrime Taq DNA polymerase High Fidelity (Life Technologies) using the primers 5’-AATAAAGGATCCGCCACCATGCAGGCGCCCCGAG-3’ (Fw) and 5’-AATAAAGAATTCTTATGTATTGTGTACAGACTCTCGTAACTCCGTG-3’ (Rev) to amplify the coding sequence of the SVV ORF63 gene. The PCR-generated product was inserted into pcDNA3.1-IRES-nlsGFP using BamHI/EcoRI sites, creating pcDNA3.1 ORF63. To create pcDNA3.1 FLAG-ORF32 we amplified ORF32 from the same template with the primers 5'-(AATAAAGGATCCGCCACCATGGCATCATCTAATACTTGCGAAGAACAAAATAATTCTA)-3' (Fw) and 5'-(AATAAAGAATTCTTActtatcgtcgtcatccttgtagtcATCCGTTTCGCTCTCGCTAGATGAAGGTTG)-3' (Rev) using the Expand High Fidelity PCR system (Roche). The PCR-generated product was inserted into the pcDNA3.1-IRES-nlsGFP vector using BamHI/EcoRI sites. VZV ORF63 was amplified from DNA extracted from VZV.eGFP-infected MRC-5 cells using the Expand Expand High Fidelity PCR system (Roche) and the following primers: 5'-(AATAAAGGATCCGCCACCATGTTTTGCACCTCACCGGC)-3' (Fw) and 5'-(AATAAAGAATTCCTACACGCCATGGGGGGGCGGTATATC)-3' (Rev). The resulting insert was cloned into pcDNA3.1-IRES-nlsGFP using BamHI/EcoRI sites. Rhesus IRF9 was synthesized and codon-optimized for expression in rhesus cell lines by GenScript. The insert was cloned from the pUC57 plasmid into pcDNA3.1-IRES-nlsGFP using BamHI/EcoRI restriction sites, creating pcDNA3.1 IRF9. All ligations were performed using the Rapid DNA Dephos and Ligation kit (Roche). The DNA sequences of all expression plasmids were verified.
HEK 293T cells were transfected with the indicated plasmids and pGL3-ISRE-Luc (a kind gift from Dr. John Hiscott, Vaccine and Gene Therapy Institute, Florida) using the Lipofectamine 2000 reagent (Life Technologies) using the manufacturers protocol. For the IRF9 overexpression experiment we used 1 μg pGL3-ISRE-Luc and the indicated amounts pcDNA3.1 (p)ORF63 and pcDNA3.1 (p)IRF9. Control pcDNA3.1 or pRetro-E2 expressing GFP was used to equalize all transfection samples to a total of 6 μg DNA. Transfection efficiency was confirmed measuring GFP expression in all samples using Synergy HTX Multi-Mode Reader (Bio-Tek) or by staining for specific proteins in western blot. To test whether VZV ORF63 inhibits IFN-signaling, we used 1 μg of pGL3-ISRE-Luc and 3 μg of the other indicated plasmids. 48 hours post transfection IFN-signaling was assessed using a luciferase assay, described below.
GIPZ lentivirus constructs expressing shRNA specific for human IRF9 were obtained from Open Biosystems/GE Healthcare. The constructs used are V3LHS-322329 (shRNA-1), V3LHS-322332 (shRNA-2), and V2LHS-69847 (shRNA-3). Replication deficient lentiviruses were produced by transfecting the shRNA vectors into HEK 293T cells and providing the vesicular stomatitis virus G (pMD2.G VSV-G, Addgene) protein and the packaging plasmid psPAX2 (Addgene) in trans. The plasmids were transfected using Lipofectamine LTX (Life Technologies). 48 hours post transfection the supernatant containing lentivirus was harvested and transferred to target cells, which were transduced in the presence of 5 μg/ml Polybrene (Hexadimethrine bromide; Sigma-Aldrich). After 24 hours, this process was repeated. The resulting cell lines were grown in the presence of 3 μg/ml Puromycin to select for shRNA expressing cells.
P-values were determined using unpaired Student’s t-test.
To prepare the SVV ORF63/70 mutant, we used an SVV BAC containing the complete SVV genome and eGFP driven by the CMV immediate-early promoter [39]. To introduce mutations into SVV ORF63/70, we used the two-step red-mediated mutagenesis protocol [40]. Mutagenesis of SVV BAC using this protocol has been previously described [29]. Briefly, we used a recombinant plasmid encoding red fluorescent protein (RFP) interrupted by the kanamycin gene (kindly provided by Dr. Benedikt Kaufer, Freie Universität Berlin, Germany). Using oligonucleotide primers specific for regions flanking SVV ORF63/70 at the 5’-end and RFP-specific sequences at the 3’-end, we amplified a 1748 bp DNA fragment containing RFP/kanamycin (ORF63 mRFP Fw: TACCATCTGAATGTTACGTACATAAATAAAACGCTTCTCAATGGCCTCCTCCGAGGACG, ORF63 mRFP Rev: GACAGGGGTAACATGTTAGCGGCTCCCTATTGGGTAAGGGACTACAAGGCGCCGGTGGAG). The DNA fragment was used to transform E. coli GS1783 containing wild-type SVV BAC. We selected kanamycin-resistant colonies and extracted recombinant BAC DNA and confirmed recombination using Hind III digestion and agarose gel electrophoresis. We identified the recombinant BAC clones that contained RFP/kanamycin in place of SVV ORF63 and eliminated the kanamycin cassette. Complete replacement of SVV ORF 63 sequences by RFP was confirmed by sequence analysis.
DHFR domains were introduced at the C-terminus of ORF63/70 using a plasmid containing the destabilization domain dihydrofolate reductase (DHFR) derived from E. coli (kindly provided by Dr. Thomas Wandless, Stanford University, California). We introduced the kanamycin-cassette at a unique restriction site (PmeI) within the sequences encoding DHFR. We amplified DHFR/kanamycin with primers specific for SVV ORF63/70 (ORF70 DHFR Fw: CCATCTGAATGTTACGTACATAAATAAAACGCTTCTCAATGATCAGTCTGATTGCGGCGTTAGCGGT, ORF70 DHFR Rev: CCATCTGAATGTTACGTACATAAATAAAACGCTTCTCAATGATCAGTCTGATTGCGGCGTTAGCGGT) by PCR and transformed of E. coli GS1783 containing mutant SVV BAC in which SVV ORF63 was replaced with RFP. After elimination of the kanamycin cassette, mutant SVV BAC in which DHFR was fused at the amino terminus of SVV ORF70 was identified by HindIII digestion and gel electrophoresis. Proper fusion of DHFR to SVV ORF70 was confirmed by sequence analysis. The recombinant BAC was purified and used to transfect Vero cells. Infected cells were grown in the presence of 10 μM trimethoprim (TMP) to stabilize the ORF70-DHFR fusion protein. SVV plaques expressing eGFP and RFP were identified and isolated using a fluorescent microscope. Sequentially, mutant SVV was passaged four to five times in Vero-CRE cells. Passaging the virus allowed recombination of ORF70-DHFR to ORF63 location, which lead to the loss of RFP. In addition, BAC vector and eGFP sequences within the virus are flanked by loxP sites, thus passing the virus in Vero cells stably expressing cre recombinase resulted in the elimination of these non-viral sequences [39]. SVV plaques that were negative for both eGFP and RFP were purified and transferred from Vero-CRE cells to TRFs via serial passage. DNA extracted from Vero cells infected with SVV mutant was used for sequence analysis to confirm proper fusion of DHFR to the C-termini of both ORF63 and ORF70.
In vitro growth curves for wild type and mutant SVV were generated as described [41]. Briefly, a monolayer of uninfected Vero cells in 25 cm2 tissue culture flasks were infected with approximately 5X102 Vero cells previously infected with either wild type or ORF63-DHFR SVV. At 3, 24, 48, 72, 96, 120, 144 and 168 hours p.i. cells were trypsinized, diluted and seeded on triplicate dishes containing uninfected Vero cells. After approximately one week, infected cells were stained with crystal violet and infectious plaques were counted.
SVV Delta, ORF63/70-DHFR SVV, SVV ORF63 gene, and SVV ORF32 gene: GenBank NC_002686. VZV pOka and VZV ORF63 gene: GenBank AB097933.
To determine if SVV interferes with type I IFN-mediated responses, we studied IFN-stimulated response element (ISRE)-dependent transcription in SVV-infected luciferase reporter cells. We used telomerized rhesus fibroblasts (TRFs) stably expressing firefly luciferase under the control of the ISRE as well as constitutively expressing renilla luciferase to control for differences in cell viability between the samples. TRF-ISRE cells were infected with SVV.eGFP at a ratio of 5:1 (uninfected to SVV-infected cells) and, after 42 hours, incubated with rhesus or human IFNα for 6 hours. Productive virus infection was confirmed by visualizing eGFP expression using immunofluorescence microscopy. The firefly and renilla signal was measured and the ratio of these values reflected ISRE activity. In mock-infected cells, incubation with increasing concentrations of rhesus IFNα corresponded with increased ISRE activity. However, only a minimal response to IFNα was observed in SVV-infected cells (Fig 1A, left panel). The rhesus reporter cells were also activated by human IFNα and a comparable reduction in ISRE activity was observed in the SVV-infected cells (Fig 1A, right panel). A dose-dependent increase of luciferase activity in SVV-infected cells was not due to the presence of uninfected cells since this was also observed when the infected cells were sorted for high GFP expression by flow cytometry before IFN-treatment (Fig 1B), suggesting that high concentrations of IFN can partially overcome the inhibition by SVV. The ISRE element drives the expression of interferon stimulated genes (ISG). To study if SVV inhibits IFNα-induced ISG-expression, we infected TRFs with SVV.eGFP for 40 hours and incubated with increasing concentrations of recombinant universal type I IFN (uIFN) for 8 hours. Productive SVV infection was confirmed by the detection of SVV ORF63 expression (Fig 1C). Expression of ISG15, ISG54 and Mx-1 was observed in all mock-infected IFNα-stimulated samples, but was absent in SVV-infected cells (Fig 1C). These data show that SVV inhibits IFNα-mediated activation of ISRE-dependent reporter gene expression and ISG protein expression.
The engagement of IFN with the IFN-receptor results in the activation of the JAK-STAT signal transduction pathway. The resulting phosphorylation of STAT1 and STAT2 allows their heterodimerization and association with IRF9, forming the ISGF3 complex that subsequently shuttles to the nucleus to initiate ISRE-dependent transcription [42]. The nuclear translocation of the ISGF3 complex is thus essential for ISRE activation. We analyzed IFN-induced nuclear localization of STAT in SVV-infected cells. TRFs were infected at a 10:1 ratio with SVV.eGFP and incubated with uIFN for 40 minutes at 48 hours post infection (p.i.). In uninfected cells, STAT2 was found predominantly in the cytosol in the absence of IFN and in the nucleus upon IFN-treatment (Fig 2A). In contrast, STAT2 was not translocated to the nucleus in SVV-infected cells (green/eGFP) upon IFN-treatment (Fig 2A). In addition, we isolated cytoplasmic and nuclear fractions of SVV.eGFP-infected Vero cells (ratio 5:1) and determined the cellular localization of STAT2 by western blot. Separation of cytosol and nuclei was confirmed using GAPDH and the nuclear matrix protein p84 (Fig 2B). In uninfected cells, STAT2 was found in the cytosolic fraction in the absence of IFN-treatment, whereas stimulation with uIFN led to the redistribution of STAT2 to both cytoplasmic and nuclear fractions. In contrast, STAT2 remained predominantly cytosolic in SVV-infected cells even upon IFN-treatment (Fig 2B). The SVV ORF62 protein was found in both cytoplasmic and nuclear fractions. This distribution is consistent with reports for the homologous VZV protein that, while primarily nuclear during early times of infection, localizes to the cytoplasm at later times of infections as a results of phosphorylation by ORF66 [43, 44]. Thus, in an asynchronous infection one would expect both nuclear and cytoplasmic expression of ORF62. Taken together, these data suggest that SVV inhibits IFN-dependent ISG-induction by abrogating the IFN-associated translocation of STAT2.
Next we examined whether the inhibition of STAT2 nuclear translocation correlated with a SVV-mediated reduction in steady state levels of ISGF3 members or impaired STAT1/STAT2 phosphorylation. TRFs infected with SVV.eGFP for 48 hours were stimulated with IFN for 20 minutes. Steady state levels and IFN-induced phosphorylation of STAT1 were comparable between SVV- and mock-infected cells (Fig 3A). In contrast, IFN-induced STAT2 phosphorylation was absent in SVV-infected cells and steady state levels of the protein also appeared to be reduced (Fig 3A). Densitometric analysis of STAT2 protein using four independent experiments confirmed an approximately 25% decrease in STAT2 levels (Fig 3B). However, this decrease was not statistically significant. In contrast, we observed a significant decrease of more than 50% in IRF9 levels by SVV (Fig 3A and 3B). Interestingly, the reductions in STAT2 and IRF9 protein levels were observed regardless of IFN stimulation (Fig 3B). Since IRF9 drives the nuclear translocation and retention of phosphorylated STAT1 and STAT2 [42], we studied the localization of residual IRF9 in IFN-stimulated SVV-infected Vero cells by analyzing isolated cytoplasmic and nuclear fractions in western blots. In control cells, IFN-stimulation triggered the increased translocation of IRF9 from the cytosol to the nucleus. However, this increased nuclear translocation of IRF9 was not observed in SVV-infected cells (Fig 3C). Furthermore, this experiment confirmed the SVV-mediated reduction in IRF9 expression levels. These data suggest that SVV abrogates JAK-STAT signaling by both preventing the phosphorylation of STAT2 and reducing STAT2 and IRF9 protein levels.
Ambagala et al. showed that wild type VZV can replicate in cells preincubated with IFNα, but an ORF63 deletion mutant could not [17]. This observation suggested that VZV ORF63 might be involved in the ability of VZV to evade IFN responses. VZV and SVV ORF63 share 52% overall amino acid homology [4, 45]. To determine if SVV ORF63 plays a role in the inhibition of IFN-stimulated responses observed in SVV-infected cells, we constructed a recombinant adenovirus expressing SVV ORF63 under the control of a tetracycline-responsive promotor (AdORF63). ORF63 expression is induced by co-infection with a recombinant adenovirus expressing the tetracycline-regulated transactivator (AdTA) [37]. These adenoviruses lack the E1-region [36], and therefore unable to interfere with IFN-signaling [46]. TRFs were co-transduced with a multiplicity of infection (MOI) of 10 of AdTA and with increasing MOIs of AdORF63. The expression of ORF63 was monitored at 48 hours p.i. (S1A Fig). Increasing MOI correlated with increasing ORF63 expression levels as expected. However, higher ORF63 levels also resulted in decreased GAPDH expression, suggesting that high expression levels of ORF63 may be cytotoxic. The transactivator expressed by AdTA can be inactivated by the tetracycline derivative doxycycline (Dox). To fine-tune ORF63 expression levels, we transduced TRFs with AdORF63 (MOI 20) and AdTA (MOI 10) in the presence of decreasing amounts of Dox. We detected robust ORF63 expression in cells that were incubated with 1 ng/ml Dox (S1B Fig) and these expression levels were comparable to ORF63 expression in cells infected with SVV (S1C Fig). At 1 ng Dox, GAPDH levels were not affected whereas lower Dox concentrations resulted in decreased GAPDH levels reflecting reduced cell viability (S1B Fig). To examine whether ORF63 inhibits IFN signaling we treated TRFs with IFN for up to 16 hours in the presence of 1ng Dox and studied ISG expression using qPCR (Fig 4A). In addition we used 1000 ng/ml Dox to inhibit ORF63 expression and under these conditions we observed increased expression of Mx-1 and ISG54 mRNA, reaching peak expression at 8 and 4 hours of IFN stimulation, respectively. In ORF63 expressing cells, however, Mx-1 and ISG54 mRNA levels were severely reduced at all time points (Fig 4A). ORF63 expression was confirmed by qPCR (Fig 4A, lower right panel). We also examined ISG protein expression by western bloting in the absence or presence of ORF63: after 8 hours of stimulation with IFN, high levels of ISG15, ISG54 and Mx-1 were detected in mock-transduced cells and in AdORF63-transduced cells treated with 1000 ng of Dox, whereas expression was absent or barely detectable in AdORF63-transduced cells treated with 1 ng Dox (Fig 4B). Taken together, these data indicate that SVV ORF63 inhibits type I IFN-induced gene expression.
The inhibition of IFN-induced ISG expression by ORF63 correlated with our observations in SVV-infected cells. Since reduced STAT2 phosphorylation as well as decreased amounts of STAT2 and IRF9 proteins were observed in SVV-infected cells (Fig 3A and 3B), we examined whether expression of ORF63 leads to the inhibition of IFN-induced STAT2 phosphorylation and reduced steady state levels of STAT2 and IRF9. Expression and phosphorylation status of members of the JAK-STAT pathway were examined in AdORF63/AdTA-transduced TRFs stimulated with uIFN for 20 minutes or 8 hours. Despite an inhibition of IFN-induced ISG expression in ORF63-expressing cells (Fig 5A, lower panel) the expression levels and phosphorylation status of STAT1 and STAT2 were unchanged (Fig 5A, upper panel). However, we did observe a reduction in steady state levels of IRF9 when ORF63 was present (Fig 5A). To confirm that ORF63 affects IRF9 expression, we transduced TRFs with AdORF63/AdTA with decreasing concentrations of Dox to obtain increasing ORF63 expression levels. Western blot analyses revealed that increasing levels of ORF63 inversely correlated with decreasing IRF9 levels (Fig 5B and 5C). Interestingly, reduced IRF9 expression was observed in both unstimulated and IFN-stimulated samples (Fig 5B), suggesting that ORF63 reduces IRF9 regardless of IFN signaling. To confirm that ORF63 does not affect STAT2 levels we transduced TRFs with AdORF63/AdTA in the presence of decreasing amounts of Dox. We observed a reduction in STAT2 levels in cells expressing high levels of ORF63 (0.1 and 0 ng/ul Dox) (Fig 5D). However, STAT1 and GAPDH expression were also affected in these samples, but when we normalized STAT1 and STAT2 expression to GAPDH expression the reduction in STAT1 or STAT2 levels was not significant when ORF63 was expressed (Fig 5E). To determine whether ORF63 affects the transcription of IRF9 we studied IRF9 mRNA levels in TRFs transduced with AdORF63/AdTA in the presence of decreasing amounts of Dox. However, we did not observe a reduction of IRF9 mRNA levels upon increasing transcription of ORF63 suggesting that ORF63 does not affect IRF9 transcription (Fig 5F).
To analyze if ORF63 promotes IRF9 degradation via the proteasome we incubated ORF63-expressing cells with increasing concentrations of the proteasome inhibitor MG132 for 16 hours prior to lysing the cells. Treating ORF63-expressing cells (+) with increasing concentrations of MG132 reversed IRF9 degradation, but we also observed a slight increase in IRF9 levels in control cells (-) (Fig 6A). To determine if the rescue of IRF9 expression in ORF63-expressing cells was due to reduced turnover of residual IRF9 or due to actively blocking ORF63-mediated IRF9 degradation, we averaged the ratio of IRF9 and GAPDH expression in MG132-treated (10 μM) control or ORF63-expressing cells in four independent experiments (Fig 6B). While MG132-treatment increased IRF9 expression in control cells, the difference was not statistically significant. In contrast, MG132-treatment of ORF63-expressing cells resulted in a fivefold increase in IRF9 expression (Fig 5B, p = 0.007). From these data we conclude that ORF63 promotes IRF9 degradation in a proteasome-dependent manner. In addition to IRF9, related IRF proteins play an important role in the regulation of the expression of IFN and ISGs [47–49]. To determine whether SVV ORF63 affects the expression of IRFs other than IRF9, we monitored the steady state expression of IRF1 and IRF3 in TRFs transduced with AdORF63/AdTA in the presence of decreasing amounts of Dox. Increasing levels of ORF63 reduced IRF9 expression levels, but not that of IRF1 or IRF3 (Fig 6C). While we cannot formally rule out that SVV ORF63 might affect other IRFs it seems likely that SVV ORF63 specifically induces the degradation of IRF9 and thus preventing IFN-mediated ISG induction.
To determine whether the reduction of IRF9 was sufficient to prevent ISG-induction, we attempted to recapitulate the ORF63-effect by reducing IRF9 levels using small hairpin RNA (shRNA). Telomerized human fibroblasts (THF) stably expressing ISRE-luciferase (THF-ISRE) were transduced with lentivectors expressing IRF9-specific shRNA. Western blots of IRF9 showed that IRF9 expression levels were reduced in cells expressing shRNA-1 and completely absent in cells expressing shRNA-2 and -3 (Fig 7A). To compare steady state levels of IRF9 upon translational inhibition by shRNA to that of post-translational degradation by ORF63, we transduced THF-ISRE cells with AdORF63/AdTA in the presence of decreasing amounts of Dox. In these human cells, optimal ORF63 levels were observed at 0.1 ng/ml Dox or in the absence of Dox without reduction of GAPDH levels whereas only partial induction of ORF63 was observed at 1 ng Dox. A reduction of IRF9 levels consistent with ORF63 expression was observed. Reduced steady state levels of IRF9 were comparable to the partial reduction of IRF9 observed in shRNA-1 expressing cells while IRF9 was completely absent from shRNA-2 and 3 expressing cells (Fig 7B). However, when the THF-ISRE cells expressing the three shRNAs were incubated with uIFN for 4 or 8 hours, ISG54 mRNA expression was only partially reduced in the presence of shRNA-1, whereas shRNA-2 and shRNA-3 largely prevented ISG induction (Fig 7C). Analysis of IFN-induced ISG54-expression by western blotting confirmed these results (Fig 7D). These data demonstrate that removal of IRF9 is sufficient to inhibit ISG-expression and they are consistent with ORF63 affecting IRF9 protein turnover rather than IRF9 transcription or translation.
To further determine whether IRF9 is the primary JAK/STAT-associated target of ORF63, we took advantage of the fact that HEK 293T cells do not respond efficiently to type I IFN unless IRF9 is overexpressed [50]. Co-transfection of a plasmid encoding an ISRE-luciferase reporter with increasing amounts of rhesus IRF9-expressing plasmid resulted in an IRF9-dependent increase of luciferase expression upon treatment with IFN at 24 hours post transfection (Fig 8A; black lined graph). Interestingly, optimal luciferase stimulation was observed with 25–100 ng of the IRF9 plasmid whereas higher IRF9 concentrations resulted in decreased luciferase activity. This decrease might be due to the fact that higher IRF9 levels activate ISRE activity even in the absence of IFN-treatment (Fig 8B), and a prolonged stimulation might induce negative regulators of IFN signaling [51]. Co-transfection of 1 μg ORF63-expressing plasmid resulted in the inhibition of IRF9-induced luciferase expression both in the presence or absence of IFN (Fig 8A and 8B; gray lined graphs) suggesting that ORF63 inhibition cannot be overcome by increasing IRF9 levels. However, ORF63 needs to be in excess of IRF9 for complete inhibition since complete inhibition of IRF9-dependent ISG-induction was only observed when at least 250 ng of ORF63 plasmid was co-transfected with 50ng IRF9-expressing plasmid (Fig 8C). Together with the finding that ORF63 inhibited IRF9-dependent ISRE-transcription even in the absence of IFN-stimulation (Fig 8B) these data further support the conclusion that degradation of IRF9 is the major mechanism by which ORF63 inhibits JAK/STAT signaling.
To study the role of ORF63 in the context of virus-induced inhibition of JAK-STAT signaling we constructed a conditionally ORF63/ORF70-expressing mutant using a recombinant bacterial artificial chromosome containing the complete SVV genome (SVV BAC) [39]. Recently, using the SVV BAC an ORF63/ORF70 SVV mutant was constructed by introducing stop codons in both genes. These mutations severely affected replication of the virus in vitro consistent with ORF63 being essential for viral growth [29]. To generate a mutant virus in which expression levels of ORF63 could be conditionally regulated, we fused the destabilizing domain (DD) of dihydrofolate reductase (DHFR) to the C-termini of ORF63 and ORF70 using two-step red-mediated mutagenesis (Fig 9A) [40]. The addition of DD-DHFR to any protein results in rapid proteasomal degradation of the fusion protein unless DHFR is stabilized with trimethoprim (TMP) [52]. We were able to recover DD-DHFR-tagged SVV in the presence of TMP and growth curves confirmed that viral growth was reduced by about 50% during the first 96 hours of infection compared to that of unmodified BAC-derived SVV when grown in 10 μM TMP (Fig 9B). A more severe reduction in viral growth was observed beyond that time point. To study the effect of TMP-removal on ORF63/70-DHFR SVV replication, we infected Vero cells in the presence of 10 μM TMP until viral plaques were detected (Fig 9C, left panel). Removal of TMP and passaging ORF63/70-DHFR SVV-infected Vero cells resulted in very few plaques (Fig 9C, middle panel). However, virus in these cultures could be rescued when 10 μM of TMP was added back to the culture (Fig 9C, right panel). These observations confirmed that the virus expressing DD-DHFR-tagged ORF63/ORF70 was able to replicate in tissue culture in the presence of TMP and viral growth can be regulated by removal of TMP. To determine whether DHFR-fusion to ORF63/ORF70 affected viral inhibition of the JAK-STAT pathway, we infected TRFs with wild type or ORF63/70-DHFR SVV in the presence of 10 μM TMP and studied expression of STAT2 and IRF9 as well as phosphorylation of STAT2 after 20 minutes of stimulation with IFN (Fig 9D). We observed a reduction in STAT2 and IRF9 expression levels as well as a complete inhibition of STAT2 phosphorylation in both wild type- and ORF63/70-DHFR-infected cells, indicating that the evasion of JAK-STAT signaling was not affected by the presence of the DHFR domains (Fig 9D). Since the removal of TMP from the culture media triggers degradation of ORF63-DHFR and ORF70-DHFR, we infected TRF with ORF63/70-DHFR SVV and cultured the cells in varying concentrations of TMP. We observed a dose-dependent decrease in ORF63 expression and ORF63 was no longer detectable when the virus was grown in 0.02 μM or less TMP (Fig 9E). This reduction in ORF63 levels led to increased levels of IRF9, indicating that expression of these proteins was inversely correlated (Fig 9E). However, the expression of ORF31 (glycoprotein B) was also affected by reducing TMP concentrations, similar to ORF63 expression (Fig 9E, western blot and graph). This result revealed that ORF63/70 is required for ORF31 expression and most likely for the expression of other SVV genes as well [24]. Restoration of IFN-sensitivity could therefore not be unequivocally assigned to the absence of ORF63 since SVV ORF31 was absent as well.
In the experiments described above, we established that SVV inhibits JAK-STAT signaling by interfering with IFN-induced phosphorylation of STAT2 and by modulating the degradation of STAT2 and IRF9. Previous reports have shown that VZV interferes with IFN-mediated signaling [12, 17], however, the evasion mechanisms involved in this inhibition are largely unknown. To determine if VZV inhibits JAK-STAT signaling similar to SVV, we infected human fibroblasts, MRC-5 cells, with VZV.eGFP for 48 hours and activated JAK-STAT signaling by incubating the cells with uIFN for 20 minutes prior to harvesting the cells. Similar to SVV, expression levels of IRF9 and STAT2 were reduced in VZV-infected cells (Fig 10A, upper panel). Moreover, complete inhibition of IFN-induced STAT2 phosphorylation was observed, indicating that VZV employs mechanisms of JAK-STAT evasion that are comparable to those of SVV (Fig 10A, upper panel). However, unlike SVV (Fig 5A), IFN-induced STAT1 phosphorylation is blocked by VZV (Fig 10A, lower panel). To determine whether VZV ORF63 prevents ISG-induction, we co-transfected HEK 293T cells with the ISRE-luciferase reporter plasmid and plasmids encoding VZV ORF63 or, as a control, FLAG-tagged SVV ORF32. The cells were stimulated with uIFN for 6 hours at 42 hours post transfection, after which ISRE-luciferase expression was measured. In contrast to untransfected controls or SVV ORF32, expression of VZV ORF63 caused a reduction in IFN-induced luciferase expression (Fig 10B). Expression of the viral proteins was confirmed by western blot (Fig 10C). To assess if VZV ORF63 plays a role in the reduction of IRF9 expression observed in VZV-infected cells (Fig 10A), we expressed the viral protein using the tetracycline-inducible adenovirus system described in Fig 4. MRC5 cells were co-transduced with VZV AdORF63 and AdTA, incubated with the indicated concentrations of Dox and after 48 hours steady state levels of IRF9 were analyzed using western blot. We observed a prominent decrease in IRF9 expression levels in cells that expressed VZV ORF63 (Fig 10D, left panel). To determine if the reduction in STAT2 expression levels and phosphorylation observed in VZV-infected cells was due to VZV ORF63, we stimulated the transduced MRC5 with uIFN for 20 minutes to activate the pathway and studied STAT2 by western blot. STAT2 steady state levels and phosphorylation were unaffected by the presence of VZV ORF63 (Fig 10D, left panel). The same experiment was performed in TRFs and we observed a VZV ORF63-induced reduction of IRF9, but not of STAT2, in those cells as well (Fig 10D, right panel). Taken together these data show that SVV and VZV employ similar mechanisms to interfere with JAK-STAT signaling and that the ORF63 proteins of both viruses contribute to this inhibition by mediating the degradation of IRF9. VZV ORF63 induced IRF9 degradation in both MRC5 and TRFs (Fig 10D), indicating that the pathway’s inhibitory target is conserved between human and rhesus cells.
The data presented here show that both SVV and VZV inhibit IFN-mediated ISG induction and reduce the expression of IRF9. In addition, we observed a marginal decrease in STAT2 levels and a complete inhibition of IFN-dependent phosphorylation of STAT2 in both SVV- and VZV-infected cells. Since IRF9 is essential for JAK-STAT signaling, the degradation of IRF9 by ORF63 is thus part of a multi-pronged inhibition of IFN-mediated antiviral gene induction. These observations are consistent with previous reports showing that VZV prevents induction of Mx-1 in brain fibroblasts and reduces STAT2 levels in infected brain fibroblasts [53]. We further demonstrate that SVV ORF63 promotes the degradation of IRF9 in a proteasome-dependent manner, but does not affect STAT2 protein levels or its phosphorylation. SVV ORF63 and VZV ORF63 share 52% amino acid identity and there is evidence that these proteins are expressed in latently infected ganglia [5, 21, 22, 28, 45, 54–56]. We observed that inhibition of IFN-signaling by both proteins correlated with degradation of IRF9 whereas neither protein affected STAT2 expression or phosphorylation. Together with the previous report by Ambagala et al. demonstrating IFN-sensitivity of ORF63-deficient VZV [17] our data thus strongly suggest that ORF63 plays a central role in IFN-signaling inhibition by both SVV and VZV. In addition, our observations reveal that both viruses likely encode additional ORFs responsible for the inhibition of STAT2 phosphorylation and the reduction of STAT2 expression.
The reduction in IRF9 levels observed in SVV/VZV-infected cells and ORF63-expressing cells did not result from reduced transcription (Fig 5F and S2A Fig). Rather, ORF63-mediated the proteasomal degradation of IRF9 independently of IFN-signaling. Although this is the first description of IFN-evasion by IRF9 targeting for a herpesvirus, IRF9 (p48) targeting has been demonstrated for several other viruses. The non-structural protein (NSP) 1 of the simian rotavirus strain SA11-4F was shown to induce proteasomal degradation of all IRF proteins that contain an IRF association domain (IAD), which include IRF3, IRF5, IRF7 and IRF9 [57]. In contrast, our data suggest that ORF63 targets IRF9, but not IRF3 which in VZV is degraded by ORF61 [15]. Adenovirus E1A protein blocks IFN-induced protein expression by reducing IRF9 expression levels, yet the mechanism of this immune evasion is unknown [46, 58]. Incidentally, the presence of E1A in HEK293 cells [59, 60] could be responsible for the requirement of exogenous IRF9 for IFN-dependent ISG-induction. In our co-transfection experiments ORF63 thus seemed to be able to eliminate IRF9 once the E1A-dependent IRF9-inhibition was breached. Additionally, the reovirus type 1 Lang μ2 protein causes nuclear accumulation of IRF9, independent of IFN-stimulation, which results in severely impaired JAK-STAT signaling [61]. Finally, the human papillomavirus (HPV) 16 E7 oncoprotein was shown to interact with IRF9 thereby preventing ISGF3 formation [62]. We did not observe an interaction between IRF9 and ORF63 in immunoprecipitations studies performed in SVV-infected and ORF63-expressing TRFs. Therefore, the exact mechanism by which ORF63 elicits the degradation of IRF9 still needs to be elucidated.
IRF9 is a key player in the JAK-STAT signaling pathway activated by type I IFN: unphosphorylated STAT2 is complexed with IRF9 and the pair continuously shuttles between the cytoplasm and the nucleus, which is driven by the nuclear localization signal of IRF9 and the nuclear export signal (NES) of STAT2 [63]. Upon activation of JAK1 and TYK2, phosphorylated STAT1 and STAT2 dimerize, which results in the loss of the NES of STAT2 [63]. The requirement of IRF9 for anti-viral immune responses was demonstrated by Kimura et al. using IRF9 knock out murine cells. Replication of herpes simplex virus type 1 and vesicular stomatitis virus (a rhabdovirus) was greatly enhanced in IFNα-treated cultures of infected IRF9-/- cells, while the cytokine limited viral replication in wild type cells [64]. In addition, Maiwald et al. created a mathematical model based on experimental data that shows that IRF9 determines the peak time and intensity of type I IFN-induced responses [65]. Our data using IRF9-specific shRNA are consistent with these studies and demonstrate the requirement of IRF9 for efficient ISG induction in human cells. Whereas shRNA-expressing cells seemed to express lower levels of IRF9 than ORF63-expressing cells it needs to be considered that ORF63 needs protein expression to act on IRF9 whereas shRNAs prevent protein expression itself. Thus, we concluded that reduced IRF9 levels are responsible for inhibition of IFN-induced ISG expression in both shRNA and ORF63-expressing cells. We therefore propose that that ORF63 blocks the JAK-STAT pathway by reducing IRF9 levels. This conclusion is further supported by the ability of ORF63 to counteract IRF9-mediated ISG-induction in HEK 293T cells both in the presence and absence of IFN (Fig 8).
Since IRF9-degradation was also observed in VZV-infected and VZV ORF63-transduced cells (Fig 10) it seems highly likely that the restoration of IRF9-levels was responsible for the previously described hyper-sensitivity of ORF63-deficient VZV [17]. The VZV deletion mutant used in this study contained a truncation of ORF63 in which only the first 24 amino acids of ORF63 was expressed. This mutant did not replicate in the presence of IFNα, but was able to replicate, albeit at reduced levels, in some cell types such as the osteosarcoma cell line U2OS [17]. To determine whether IRF9-depletion would restore the ability of this VZV ORF63 deletion virus to replicate in fibroblasts we infected THF-ISRE and THF-ISRE shRNA-3 cells with the VZV-deletion mutant (kindly provided by Jeff Cohen). However, we did no observe an increased growth of this virus upon IRF9-depletion. This indicates that additional functions of ORF63, either directly affecting host pathways or indirectly via other viral proteins, contribute to the reduced growth of the deletion virus.
Introducing stop codons in ORF63 and ORF70 severely impaired the growth of SVV in IFN-deficient Vero cells [29]. Similarly, when we fused a DHFR domain to the C-terminus of both ORF63 and ORF70 to create an inducible knock out for both proteins, removal of TMP resulted not only in the degradation of ORF63 and ORF70 but also prevented expression of other SVV genes that depend on ORF63 function. Because of the requirement of ORF63 for viral replication we were unable to directly address the biological significance of IRF9-degradation for evasion of JAK-STAT signaling by these varicelloviruses. Cohen et al. found that cells infected with the VZV ORF63 deletion virus were highly susceptible to IFN treatment [17], suggesting that ORF63-induced degradation of IRF9 plays a prominent role in evasion of this pathway. However, since ORF63 is required for viral early gene expression and viral replication [23–25], deletion of protein could also affect expression of the as yet unknown inhibitor of STAT2-phorphorylation as well. Thus, ORF63 likely impacts IFN-resistance both directly, by reducing IRF9 expression, and indirectly by regulating the expression of other IFN-inhibitory genes. Our data are thus consistent with a multipronged inhibition of JAK/STAT signaling by SVV and VZV to ensure efficient evasion of this innate immune pathway. It is not uncommon for a virus to target a signaling pathway at multiple levels. For example, VZV codes for at least three independent strategies devoted to inhibiting IRF3-driven expression of IFNs, which including ORF61 [15], ORF47 [14] and IE62 [13]. In addition, it is conceivable that IRF9-inhibition by the immediate early gene ORF63 precedes STAT2 inhibition due to sequential expression of the respective inhibitory proteins during viral infection. The relative contribution of each inhibitory pathway during viral infection thus still needs to be elucidated.
We observed that both SVV and VZV reduce levels and phosphorylation of STAT2, whereas STAT1 levels were not affected. In contrast, reduced phosphorylation of STAT1 was only observed in VZV-infected cells. The diminished STAT2 levels did not result from reduced transcription (S2B Fig). Previous reports demonstrated diminished levels of STAT1 in VZV-infected human fibroblasts treated with IFNγ [66] and reduced STAT1 phosphorylation in VZV-infected skin xenografts in the SCIDhu model [12]. A recent report confirmed VZV-mediated downregulation of STAT2, but downregulation of STAT1 was inconsistent between the experiments and inhibition of STAT1 phosphorylation was not observed [53]. Since STAT1 is shared between type I and type II IFN signal transduction pathways, it represents an attractive target for viral innate immune evasion. However, observations concerning STAT1 expression and phosphorylation in VZV-infected cells are inconsistent (our data, [12, 53, 66]), possibly due to the use of type II IFN and different cell lines. Therefore, we cannot rule out that SVV targets the phosphorylation of STAT1 in cell types other than fibroblasts. Since STAT1 phosphorylation was not reduced in SVV-infected TRFs it is unlikely that SVV interferes upstream of STAT1/STAT2 since the binding of IFN to its receptor triggers the activation of the tyrosine kinases JAK1 and TYK2 which in turn phosphorylate STAT1 and STAT2 resulting in the formation of the ISGF3 complex [67, 68]. If SVV would target either JAK1 or TYK2 one would expect a reduction in STAT1 phosphorylation. For VZV however, this possibility cannot be ruled out since STAT1 phosphorylation is inhibited. Although SVV and VZV-infected cells displayed lower expression levels of STAT2 the remaining STAT2 was not phosphorylated. Therefore, SVV and VZV most likely affect STAT2 phosphorylation directly.
Primary VZV infection starts with respiratory mucosal inoculation [69]. Recently, type III IFNs (IFNλ1–3) have been implicated in playing an important role in limiting (herpes)viral replication in mucosal tissues [70–72]. These cytokines bind to the IL28Rα/IL10Rβ receptor complex [73, 74], which is predominantly expressed on epithelial cells [75]. Engagement of the receptor results in the activation of the JAK-STAT pathway and induction ISG expression [74]. The Kaposi’s sarcoma-associated herpesvirus protein vIRF2 was shown to inhibit both IFNα and IFNλ-mediated ISG expression, by reducing the levels of STAT1 and IRF9 [76, 77]. Similarly, the inhibition of JAK-STAT signaling by VZV is likely to block IFNλ-mediated signaling in epithelial cells, thereby immediately limiting the host anti-viral responses to allow further spreading.
In addition to inhibiting IFN-dependent signal transduction, VZV inhibits activation of the IFN-gene itself and IFN-independent ISG-induction. The VZV proteins IE62, ORF47, and ORF61 all target IRF3-mediated induction of IFNβ and ISG genes, such as ISG15, ISG54 and IS56 [14, 15, 22]. VZV ORF61 also blocks the TNFα-mediated activation of NFκB by inhibiting IκBα [16]. Taken together with our observations, it appears that multiple VZV and SVV ORFs are devoted to interfering at sequential steps along the induction of IFN and anti-viral ISG.
Despite efficiently counteracting IFN activation and IFN-dependent signaling, several reports have indicated the importance of type I IFNs in limiting VZV replication and spread in vivo. Children that were treated for leukemia and were suffering from varicella showed significantly reduced dissemination of the virus in response to the administration of intra-muscular IFNα [78]. In addition, experiments in severe combined immunodeficiency mice engrafted with VZV-infected human skin (SCIDhu mice) showed that preventing IFN signaling with antibodies specific for the interferon receptor resulted in larger cutaneous lesions compared to mice that were untreated [12]. These in vivo responses are most likely explained by the fact that the establishment of an anti-viral state prior to infection is more difficult to overcome than IFN-responses in infected cells. IFNα-treatment of melanoma cells prior to VZV infection led to a reduction in plaque formation [17]. Thus, IFN likely prevents viral spread during the acute phase of infection by inducing an antiviral state in target cells. However, further sensitizing VZV to IFN by therapeutically blocking viral IFN-evasion mechanisms could improve the ability of IFN to prevent spread and/or reactivation of VZV in vivo. This concept could be experimentally tested in non-human primates using the SVV model. Since ORF63 expression have been demonstrated in latently-infected ganglia of SVV-infected rhesus macaques [5], it is also conceivable that IFN evasion is essential for maintaining viral latency. Dissecting the role of ORF63 in limiting IFN signaling could thus lead to a better understanding of the role of IFN in viral latency and reactivation.
|
10.1371/journal.pbio.1000442 | Rac1-Dependent Collective Cell Migration Is Required for Specification of the Anterior-Posterior Body Axis of the Mouse | Cell migration and cell rearrangements are critical for establishment of the body plan of vertebrate embryos. The first step in organization of the body plan of the mouse embryo, specification of the anterior-posterior body axis, depends on migration of the anterior visceral endoderm from the distal tip of the embryo to a more proximal region overlying the future head. The anterior visceral endoderm (AVE) is a cluster of extra-embryonic cells that secretes inhibitors of the Wnt and Nodal pathways to inhibit posterior development. Because Rac proteins are crucial regulators of cell migration and mouse Rac1 mutants die early in development, we tested whether Rac1 plays a role in AVE migration. Here we show that Rac1 mutant embryos fail to specify an anterior-posterior axis and, instead, express posterior markers in a ring around the embryonic circumference. Cells that express the molecular markers of the AVE are properly specified in Rac1 mutants but remain at the distal tip of the embryo at the time when migration should take place. Using tissue specific deletions, we show that Rac1 acts autonomously within the visceral endoderm to promote cell migration. High-resolution imaging shows that the leading wild-type AVE cells extend long lamellar protrusions that span several cell diameters and are polarized in the direction of cell movement. These projections are tipped by filopodia-like structures that appear to sample the environment. Wild-type AVE cells display hallmarks of collective cell migration: they retain tight and adherens junctions as they migrate and exchange neighbors within the plane of the visceral endoderm epithelium. Analysis of mutant embryos shows that Rac1 is not required for intercellular signaling, survival, proliferation, or adhesion in the visceral endoderm but is necessary for the ability of visceral endoderm cells to extend projections, change shape, and exchange neighbors. The data show that Rac1-mediated epithelial migration of the AVE is a crucial step in the establishment of the mammalian body plan and suggest that Rac1 is essential for collective migration in mammalian tissues.
| The specification of the anterior-posterior body axis of the mouse embryo depends on migration of the anterior visceral endoderm (AVE) to a position that overlies the future head. By high-resolution imaging of intact embryos we show that movement of the AVE is a form of collective cell migration, as the migrating cells retain tight and adherens junctions while they migrate and exchange neighbors within the plane of the visceral endoderm epithelium. Using conditional knockouts, we find that the small GTPase Rac1 is absolutely required for specification of the anterior-posterior axis and acts cell-autonomously within the AVE to allow cells to extend long, dynamic lamellar projections that are required for movement. Rac1-mediated epithelial migration of the AVE is a crucial step in the establishment of the mammalian body plan, and Rac1 may be important for collective migration in general in mammalian tissues, including invading tumor cells.
| Between the time of implantation and gastrulation, the pluripotent cells of the mammalian epiblast become restricted to specific lineages in a series of inductive interactions that depend on both intercellular signals and highly orchestrated cell rearrangements. One day after implantation (e5.5), the embryonic region that will give rise to the three germ layers of the mouse is a single-layered cup-shaped columnar epithelium (the epiblast) that is surrounded by the squamous visceral endoderm (VE) epithelium. At this stage, the mouse embryo is elongated in its proximal-distal axis, where the site of connection to the uterine tissue defines the proximal pole. Proximal-distal differences in the pattern of gene expression first become apparent at e5.5, when a group of VE cells at the distal tip of the embryo (the distal visceral endoderm (DVE)) expresses a distinctive set of molecular markers, including the transcription factor Hex. Between e5.5 and e6.0, this population of cells migrates proximally and comes to lie on the presumptive anterior side of the embryo, adjacent to the embryonic/extra-embryonic boundary [1],[2], where the cells are known as the anterior visceral endoderm (AVE). The cells of the AVE secrete localized Nodal and Wnt inhibitors that confine Wnt and Nodal signals to the opposite side of the embryo, where the primitive streak is then specified. Thus migration of DVE/AVE cells converts the early proximal-distal asymmetry into the definitive anterior-posterior (AP) axis of the animal.
Although migration of mammalian cells has been studied extensively in culture, little is known about the dynamics of cell migration in intact mouse embryos. As the AVE lies on the surface of the embryo and AVE migration is completed in about 5 h, it has been possible to image migration of AVE cells in vivo [2]. The AVE is therefore ideal for studies of the cell behaviors and the genetic control of mammalian cell migration in vivo.
The mechanisms of AVE migration are the subject of considerable debate. It has been observed that migrating AVE cells extend filopodia in the direction of movement, which suggests that they may migrate towards an unidentified chemoattractant [2] or away from a chemorepellant [3]. Alternatively, it has been proposed that the cells might migrate in response to instructive cues from the extracellular matrix [4],[5]. The non-canonical Wnt pathway proteins Celsr1 and Testin are expressed in the AVE [6], which suggests that AVE cells might move through planar polarity-dependent cellular rearrangements analogous to those that take place in the extending germ band of Drosophila [7],[8]. In mouse embryos that lack Prickle1, a non-canonical Wnt pathway protein, the AVE fails to move; however, this defect is accompanied by a disruption of apical-basal polarity, so it is not clear whether Prickle1 regulates AVE migration directly [9].
We showed previously that the actin regulator Nap1, a component of the WAVE complex, is important for AVE migration [10]. The WAVE complex is crucial for migration of many cell types; it promotes the formation of branched actin networks at the leading edge of migrating cells and thereby pushes the cell membrane forward [11]–[13]. The ENU-induced Nap1khlo allele was identified in a genetic screen for mutations that affect embryonic patterning [10]. The most dramatic phenotype of Nap1khlo mutants is the duplication of the AP body axis, seen in about 25% of the mutant embryos, and correlated with partial migration of the AVE. These studies indicated that WAVE-mediated migration was important for AVE migration and axis specification, but the low penetrance of the axis duplication phenotype left open the possibility that other, WAVE-independent mechanisms might be crucial for AVE migration. The WAVE complex can act downstream of Nck or Rac [14]; activation by Rac requires simultaneous interaction with acidic phospholipids [15]. Because the AP body axis is specified normally in mouse mutants that lack Nck [16], we hypothesized that the small GTPase Rac1 might act upstream of the WAVE complex to direct AVE migration.
The functions of mammalian Rac1 have been studied in a variety of processes. Conditional inactivation of Rac1 in mouse fibroblasts has confirmed that Rac plays roles in lamellipodia formation, cell-matrix adhesion, and cell survival [17]. Tissue-specific gene inactivation experiments have also implicated mouse Rac1 in a variety of processes, including EGF-induced cell proliferation of the neural crest [18], canonical Wnt signaling during limb outgrowth [19], actin rearrangements required for myoblast fusion [20], regulation of p38 MAP kinase activity and of Indian Hedgehog expression in chondrocytes [21], and maintenance of stem cells in the skin through regulation of c-Myc expression [22]. Nevertheless, no genetic studies have addressed the roles of Rac proteins in vertebrate morphogenesis.
Although the mouse genome encodes three forms of Rac, only Rac1 is expressed in the early embryo [23],[24]. Because the tissue organization of the early mouse embryo is relatively simple, we set out to define precise functions of Rac in tissue migration in vivo in the intact early embryo. It was previously shown that Rac1 null embryos die at the time of gastrulation [25]. Here we show that Rac1 is essential for the establishment of the AP axis of the mouse embryo. We use high-resolution imaging to show that wild-type AVE cells move in a coordinated fashion and extend long projections that can span nearly the entire embryonic region. In embryos that lack Rac1, AVE cells lack all projections and fail to move from their original location at the distal tip of the embryo. These findings demonstrate that Rac1 is a critical link that connects morphogenetic signals to AVE migration, allowing the establishment of the AP axis of the mammalian body plan.
It was previously described that Rac1 null mutant embryos are delayed in growth as early as e5.75 and arrest between e6.0 and e7.5 [25]. We generated a null allele of Rac1 by germ line deletion of a conditional allele [26]. Most (∼60%) of the Rac1 null embryos arrested before e7.5, and all had arrested by e7.5; thus, this allele recapitulated the early lethality previously seen in Rac1 null embryos.
The Rac1 null embryos that survived to e7.5 often showed a constriction at the boundary between embryonic and extra-embryonic regions (Figure 1A, arrow). This phenotype has been observed in mutants in which specification of the AP body axis is disrupted (e.g. FoxA2 [27]; Axin [28]; Nodal [29]; Otx2 [30]; Lim1 [31]). We therefore examined expression of markers of early AP patterning in Rac1 null embryos.
Brachyury (T), an early marker of the primitive streak, is expressed exclusively on the posterior side of wild-type e6.5–e7.5 embryos. T was expressed ectopically in e7.5 Rac1 null embryos in a ring at the embryonic/extra-embryonic border (Figure 1A). One day earlier, at the onset of gastrulation (Figure 1B; e6.5), T was expressed in a spot or ring at the embryonic/extra-embryonic border in the mutants. Wnt3 is the earliest marker of the primitive streak [32]. At e6.5, Wnt3 was expressed on the posterior side of wild-type embryos and in a ring at the level of the embryonic/extra-embryonic boundary of Rac1 mutant embryos (Figure 1C). Nodal, which is required for primitive streak formation, was expressed in e6.5 Rac1 mutants (Figure 1F). Despite the ectopic expression of T and Wnt3, gastrulation did take place in the mutants, as shown by the formation of a mesodermal layer (Figure S1; [25]). In a minority of e7.5 Rac1 embryos (∼30%), streak markers, such as T or the Wnt reporter BAT-gal (Figure 1D), did not encircle the circumference of the embryo and instead were restricted to one side of the embryo. The primitive streak never extended distally but remained as a spot at the embryonic/extra-embryonic boundary. Thus the processes that trigger primitive streak formation are ectopically initiated at all positions around the circumference of most Rac1 null embryos.
The position of the primitive streak depends on the earlier migration of the AVE cells from the distal tip of the embryo to the embryonic/extra-embryonic boundary. We therefore examined the expression of Cer1, a marker of the AVE. In wild-type embryos, Cer1 expressing cells had moved from the distal tip to the embryonic/extra-embryonic boundary at e6.5. In contrast, in Rac1 null embryos, Cer1 expression was restricted to the distal tip (Figure 1E).
Migration of AVE cells can be followed at cellular resolution using the Hex-GFP transgene, which is expressed specifically in the AVE when the cells are at the distal tip of the embryo and continues to be expressed during their migration [2]. Hex-GFP expression in Rac1 null embryos was indistinguishable from that in wild-type embryos at e5.5, before AVE migration (not shown). While Hex-GFP-expressing cells in wild-type embryos had completed migration to the extra-embryonic border at e6.5, Hex-GFP-positive cells remained in their original position at the distal tip of the embryo in Rac1 null embryos (Figure 2). The Hex-GFP expressing cells at the distal tip of the Rac1 null embryos gradually formed a grape-like cluster (Figure 2, Figure S2B), as has been seen in other mutants in which AVE cells fail to migrate (FoxA2 (Figure S2C, [27]), Otx2 [30], Nap1 (Figure S2D), and β-catenin [33]). Thus migration of the AVE cells fails in the absence of Rac1; the failure of AVE migration is sufficient to explain the failure to specify the primitive streak at a single position during gastrulation.
A number of genes, including Cripto and Nodal, are required in the epiblast for AVE migration [34]–[36]. To test whether Rac1 acts in the epiblast to promote AVE migration, we crossed animals carrying the Rac1 conditional allele (Rac1cond) with mice that carry the Sox2-Cre [37] to remove Rac1 from the epiblast but not the VE (we refer to these embryos as Rac1 epiblast-deleted). In contrast to the early lethality of the null embryos, Rac1 epiblast-deleted embryos survived until e8.5. The Rac1 epiblast-deleted embryos had a single AP axis, as marked by expression of Brachyury (Figure 3A), although they subsequently showed a variety of later morphogenetic defects (Migeotte and Anderson, in preparation). The AVE migrated normally in the Rac1 epiblast-deleted embryos, as shown by location of the Cer1-expressing cells in the proximal embryonic region at e6.5 (Figure 3B). Thus Rac1 is not required in the epiblast for axis specification or AVE migration.
To test whether Rac1 is required within the cells of the VE for their migration, we carried out the reciprocal experiment in which we deleted Rac1 in the VE using a Cre driven by the transthyretin promoter (Ttr-Cre) [38]. Ttr-Cre is expressed in all VE cells by e5.5 [38], although a Cre reporter line showed that the expression was not uniform among cells at e5.5 and became stronger by e6.0. Rac1cond; Ttr-Cre embryos (Rac1 VE-deleted) survived longer than Rac1 null embryos, to as late as e8.25. At e7.5, they displayed a constriction between embryonic and extra-embryonic regions and the headfolds had abnormal morphology (Figure 3C). Marker analysis showed that the initial specification of the AP body axis was defective in the majority of Rac1 VE-deleted embryos: T (Figure 3E) and Wnt3 (Figure 3F) were expressed at e6.5 in a ring at the embryonic/extra-embryonic border, as in the null embryos.
AVE migration was disrupted in the Rac1 VE-deleted embryos. At e6.5, Cer1 was expressed close to the distal tip of the embryo in ∼50% of Rac1 VE-deleted embryos (Figure 3D), indicating defects in AVE migration. At cellular resolution, the distribution of Hex-GFP expressing cells showed that AVE migration was abnormal in most Rac1 VE-deleted embryos, when assayed at e6.5 or e7.5 (Figure 4A,B). However, most AVE cells had moved away from the distal tip of the embryo in the majority of e7.5 Rac1 VE-deleted embryos (Figure 4). The milder defects in AVE migration and axis specification seen in the Rac1 VE-deleted embryos compared to the nulls were probably due to the perdurance of Rac1 protein after the relatively late expression of the Ttr-Cre transgene. In contrast to the abnormal expression of streak markers at e6.5, the partial migration of the AVE in VE-deleted embryos correlated with the specification of an AP axis at e7.5, as shown by staining for Brachyury (Figure 3C); a similar correction of earlier defects has been seen in other mutants that disrupt AVE migration (Otx2 [30]; Nap1 [10]). Nevertheless, the dependence of AVE migration on the presence of Rac1 in the VE, but not in the epiblast, demonstrates that Rac1 acts autonomously within the VE to promote migration of the AVE.
We confirmed that wild-type AVE cells retain E-cadherin-positive adherens junctions during the time of migration (Figure 5 and Figure S3) [39]. Tight junctions, visualized by ZO1 expression, were present between wild-type AVE cells and between AVE cells and other cells in the VE during (Figure S4A) and after migration (Figure S4B and C). Apical-basal polarity was also retained during migration, as wild-type migrating AVE cells contacted the basement membrane at all times and expressed ZO1 exclusively at their apical surfaces (Figure S4A).
Because Rac-dependent modulation of adherens junctions is thought to play a role in cell migration and rearrangements [40], we examined the organization of the VE epithelium in Rac1 mutants. Adherens junctions (marked by E-cadherin, Figure 5) and tight junctions (marked by ZO1, Figure S4) were present in the mutant embryos at e5.5, and expression of these junctional markers was indistinguishable from wild-type. Apical-basal polarity was also normal in the mutant VE at e5.5 (not shown). Between e5.5 and e5.75, some mutant AVE cells lost contact with the basement membrane (Figure S4A), prefiguring the distal AVE cell clustering we observed in most e6.5 Rac1 null embryos (Figure S2).
Other aspects of VE organization appeared normal in Rac1 mutant embryos: phospho-histone H3-positive mitotic cells were present in the Rac1 mutant VE, and little apoptosis was observed in the VE layer, although there was large-scale apoptosis in the epiblast of Rac1 null embryos at e6.5 (not shown). The VE is required to transport nutrients to and waste from the early embryo [41], but Rac1 VE-deleted embryos were comparable in size to wild-type littermates and showed normal epithelial organization at e6.5 and e7.5 (Figure 4). Thus global aspects of epithelial organization of the VE do not depend on Rac1 and are unlikely to be responsible for defects in AVE migration.
E-cadherin staining showed that VE cells throughout the embryonic region of wild-type embryos had a range of cell shapes (Figure 5A,B; Figure S3). In the extra-embryonic portion of e5.5–e6.0 embryos, most VE cells were regularly packed in a hexagonal array; in contrast, the VE cells in the embryonic region had irregular cell shapes, frequently with only 3–4 sides (Figure S5) and some multicellular rosette-like structures with vertices of 6 or more cells were observed (Figure S3A, Figure 6A). High-resolution static images revealed that some AVE cells showed protrusions characteristic of migrating cells (Figures 5B and 6A). In contrast, Rac1 mutant cells were more uniform in shape (Figure 2, Figure 5A and B, Figure 6A).
While there was strong E-cadherin expression between neighbors in most of the wild-type VE, we observed more diffuse E-cadherin expression on some faces of migrating Hex-positive cells (Figure 5D). Careful examination of Z-sections in embryos where the direction of migration could be predicted showed that the apparently weaker staining corresponded to cell boundaries that were tilted along the apical-basal axis, such that the proximal side of the cell (the presumptive leading edge) was basally located and partially covered by the cell in front (Figure 5D and E). In contrast, expression of E-cadherin appeared uniform on Hex-positive Rac1 mutant cells (Figure 5B and Figure S3B), consistent with their failure to migrate. Thus Rac1 mutant AVE cells fail to undergo the cell-shape changes that accompany migration.
The organization of F-actin revealed by phalloidin staining, like the E-cadherin staining, showed differences in cell shapes between wild-type and mutant VE cells. 3D reconstructions of wild-type embryos stained for F-actin highlighted the variety of cell shapes among embryonic VE cells (Figure S5), as well as the presence of multicellular rosettes (Figure 6A). In contrast, Rac1 mutant cells were uniform and rounded (Figure 6A).
Because our studies were motivated by the hypothesis that Rac1 acts upstream of the WAVE complex to reorganize F-actin during AVE cell migration, we examined the organization of F-actin in migrating VE cells. Individual sections from Z-stacks showed that most wild-type VE cells, and pre-migration AVE cells, had strong cortical actin. In contrast, some migrating AVE cells had weaker apical F-actin, probably reflecting dynamic F-actin rearrangements in those cells (Figure 6B, arrow). Both Hex-GFP-positive and -negative Rac1 mutant VE cells showed strong cortical actin (Figure 6B, Figure 4), consistent with a requirement for Rac1 in the dynamic rearrangements of the actin cytoskeleton required for migration.
We used time-lapse video microscopy to compare the behaviors of migrating AVE cells in cultured wild-type and mutant embryos (Figure S6). As described previously [2], wild-type Hex-GFP cells in the early phase of migration extended transient protrusions directed towards the proximal region of the embryo (Figure 7A and B, Figure 8A; Videos S1, S2, and S3).
Live imaging revealed dramatic dynamic cellular behaviors that were not captured by static images. In two time-lapse experiments in which we could follow AVE migration from beginning to end, we saw that 5–7 leading Hex-positive cells extended parallel long lamellar protrusions toward the proximal region of the embryo (Videos S1 and S2). Long projections were present for about 3 h (Figure 7B, Video S1). The vectors of movement of individual cell bodies were highly oriented towards the proximal region of the embryo (Figure S7).
In most embryos there were non-green cells between Hex-GFP+ cells. The GFP-negative cells may have corresponded to AVE cells that did not express the transgene [2] or AVE cells may have moved in smaller groups to navigate through the VE epithelium. Cell tracking showed that AVE cells exchanged neighbors during migration (Figure 7B, Video S1). Although all Hex-GFP cells belonged to the single-layered VE epithelium, some with small apical surfaces were mostly hidden by their more spread neighbors and became visible at the apical surface of the epithelium during the observation period (marked by * in Figure 7B, Video S1), highlighting the shape changes that occur during migration. We observed no more than two mitoses in AVE cells per embryo during the migration period, which makes it unlikely that cell division contributed to cell intercalation. Approximately 5 h after the appearance of long forward projections, the leading Hex-GFP cells arrived at the embryonic/extra-embryonic border. They ceased forward movement, sent shorter lateral projections, and spread laterally on the anterior side of the embryo (Figure 7A and B, Figure 8B, Videos S1 and S2). After migration, AVE cells became more dispersed on the anterior surface, probably due to the growth of the embryo.
Analysis of the Z-planes for individual cells (Figure 8A and B), 3D reconstruction (Video S4), and analysis of cells captured in a profile view (Figure 8C) showed that the long projections were located on the basal side of cells, giving cells the appearance of a snail (Figure 8D). The tight packing of trailing GFP-positive cells that did not have an edge with non-GFP-cells made it difficult to assess the cell contours with certainty. However, the trailing cells that we could observe in detail displayed shorter projections that were also on the basal side (Figure S8 and Video S5).
Morphometric analysis of protrusion dynamics suggested that the long lamellar protrusions on the leading Hex-expressing cells might probe the path that the cells would follow. Most long projections extended, then partially retracted, and extended again at a slightly different angle, keeping the same global direction (Figures 9A and S9; Videos S1, S2, S3, S6, and S7). Individual lamellar protrusions lasted for 1–2.5 h and their maximal length was 2.5 times as long as the diameter of the cell body (Figures 9C and S10, Video S8). The lamellar projections ended in two smaller lateral filopodia-like protrusions that extended and collapsed more dynamically than the lamella as a whole (Figures 9A and S9; Videos S3, S6, and S7). While the long projections were extended, the cell bodies displayed pulsing contractions with little net movement, after which the cell body moved rapidly following the direction of the projection (Video S3).
Live imaging of a Rac1 null embryo from e5.75 to e6 made it possible to observe the formation of the grape-like cluster at the embryo tip. Some mutant AVE cells located at the distal tip acquired a pear shape, with a smaller basal surface, and lost contact with the basal membrane, but did not shed off the embryo, presumably due to the acquisition of new cell-cell contacts (Figures S2 and S11, Videos S9 and S10).
Live imaging of Rac1 VE-deleted embryos confirmed that Rac1 is required for the cell shape changes and the formation of the protrusions that accompany cell migration (Figure 7C, Video S11). No active movement of Hex-GFP-positive cells was observed. Although the growth of the embryo appeared normal, very little reorganization of cells within the epithelium occurred: we observed no neighbor exchange or cell shape changes. Mutant cells were round (Figure S10, Video S12), and protrusions observed in a low proportion of Hex-GFP-positive cells were short and transient. In the two Rac1 VE-deleted mutants we imaged for a long period, we found two AVE cells per embryo that extended short protrusions (less than half the cell body size) that lasted only 10–20 min (Figure 9B and D, Video S13). Thus Rac1 is required for both cellular events that accompany the coordinated epithelial migration of AVE cells: the extension of long processes and the reorganization of the epithelium.
Embryological and genetic studies have shown that migration of the AVE cells is crucial for the establishment of the AP axis of the mouse embryo [4]. Although genetic studies in the mouse have defined requirements for the Wnt and Nodal pathways in establishing the competence of the AVE cells to migrate, remarkably little is known about the signals or cellular mechanisms that direct migration of these cells. Here we have demonstrated that Rac1, a key regulator of cell movement, is absolutely essential for the establishment of the AP axis of the mouse embryo.
In the absence of Rac1, all cells in the proximal epiblast acquire a posterior identity and no AP axis is defined. This profound disruption of AP patterning suggested that there might be an earlier defect in AVE migration. Although Rac1 is required for survival and adhesion of the mesoderm ([25]; Migeotte and Anderson in preparation), we find that Rac1 is not required for epithelial organization, cell survival, or proliferation in the VE. Wnt signaling is required for specification of the AVE [42]. It has been argued that Rac1 is required for canonical Wnt signaling [19]; in contrast, the AVE was specified normally in Rac1 null embryos, as marked by the expression of Cer1 and Hex. We find, however, that Rac1 is specifically and absolutely required for migration of AVE cells.
We showed previously that mutations that disrupt Nap1, a component of the WAVE regulatory complex, cause duplications of the AP axis with ∼25% penetrance, associated with the failure of normal migration of AVE cells [10]. Our findings support our prediction that Rac1 acts upstream of Nap1 to control directed cell migration of the AVE. Our previous studies did not test whether Nap1 was required in the migrating cells themselves. Here we were able to use a conditional allele to show that Rac1 is required in the VE, and not in the epiblast, for migration of AVE cells.
Recently we have characterized a stronger allele of Nap1 caused by a gene trap insertion, Nap1RRQ. The Nap1RRQ phenotype is intermediate between the Nap1khlo and Rac1 null phenotypes. AVE migration is disrupted in half of Nap1khlo embryos [10] but is disrupted in all Nap1RRQ, which leads to defects in AP axis specification in all mutants (Figure S12). The Nap1RRQ allele is not a null allele, as some wild-type transcript is present in mutant embryos due to splicing around the gene trap insertion (not shown). Nevertheless, the strong phenotype of the Nap1RRQ allele highlights the importance of the WAVE complex in AVE migration and argues strongly that one of the principle activities of Rac1 in AVE migration is to control activation of WAVE.
Our analysis of cellular behavior during AVE migration is also consistent with the central role of the WAVE complex. During wild-type AVE migration, the leading cells extend very long, lamellipodia-like projections (the type of structures that depend on WAVE) that are completely dependent on Rac activity. Rac1 VE-deleted embryos, which have decreased Rac1 activity in the VE, show a preferential loss of lamellipodia, although they can extend a few short projections. Our findings suggest that AVE migration is driven by an extracellular signal that activates Rac1 in the AVE cells to direct them to extend proximal, WAVE-dependent lamellar projections.
In contrast to the variable AVE migration defects seen in Nap1 mutants, 100% of Rac1 null mutants display no migration at all. Because the characterized alleles of Nap1 are not null, it is not known whether the complete loss of Nap1 would cause a stronger defect in AVE migration. However, the stronger phenotype caused by loss of Rac1 raises the possibility that Rac1 has roles in the AVE in addition to the regulation of the WAVE complex.
Rac1 could for example influence movement of AVE cells through regulation of the dynamics of adherens junctions within the VE [43]–[45]. Although wild-type migrating AVE cells retain E-cadherin between all neighbors during migration, the neighbor exchanges we observe indicate that junctions must be dynamically remodeled during migration. There is no global loss of junctions in the mutants, as Rac1 VE cells express adherens and tight junction proteins at comparable levels to wild-type cells. Nevertheless, Rac1 mutant VE cells at the onset of migration appear to be rounder than wild-type cells: they lack the tight packing seen in wild-type and their apical surfaces are dome-shaped. During migration, the shapes of both Hex-GFP positive and Hex-GFP negative cells within the embryonic VE become irregular. In contrast, cell shapes in the VE of Rac1 null mutants remain regular. These findings suggest that Rac1 could contribute to the mechanical forces within the VE epithelium that favor migration in the wild-type embryo.
There has been controversy in the field concerning a possible role for asymmetries in the blastocyst in the establishment of the AP body axis [42],[46]–[48]. We found that although most Rac1 mutants expressed markers of the primitive streak radially around the embryonic circumference, 30% of null embryos examined showed some asymmetry in the expression of streak markers. In these embryos, the cluster of distally located DVE/AVE cells was slightly skewed away from the side of streak marker expression, suggesting that some aspect of AP asymmetry might be independent of AVE migration. If these skewed DVE cells secrete Wnt and Nodal inhibitors, that might account for the residual asymmetry in Rac1 mutants. However, in the majority of Rac1 null embryos, this shift is either too small to be detected or nonexistent and does not lead to posterior restriction of streak markers. We therefore conclude that there may be a small AP bias derived from the blastocyst embryo, but that it is not detectable in most embryos and that any early bias must be reinforced through cell migration.
Collective cell migration, in which cells migrate while retaining epithelial organization, is common in both invertebrates and vertebrates [49]–[51]. One common feature of collective migration is that cells at the leading edge display some characteristics typical of mesenchymal cells, such as the presence of highly dynamic actin-rich cellular protrusions, while retaining junctions with their neighbors. This cohesive mode of migration allows cell guidance as well as mechanical coupling and shaping of tissues.
The live imaging experiments made it possible to define the cellular behaviors of the wild-type AVE during migration, which, together with our static analysis, suggested that AVE movement has the properties of collective epithelial migration. Although the Hex-expressing cells do not form a single coherent group, clusters of AVE cells insert basal projections between neighboring cells and pull themselves forward through the epithelium, while retaining junctions amongst themselves and with adjacent cells, as well as with the basal membrane. The leading AVE cells change shape from columnar to cuboidal and send long extensions that can extend several cell diameters ahead of the cell body. The long protrusions on the leading cells are unusual in structure: they have two smaller protrusions at their tips. The long protrusions are stable, lasting 1–2 h, similar to the long cellular extensions found in many cells that move in a directional manner during development [52], such as Drosophila border cells [53] and tracheal cells [54]. These stable projections are very different from short-lived lamellipodia found in cells that follow rapidly changing gradients such as neutrophils. We suggest that these unusual properties may facilitate the reception of signals that guide the direction of AVE migration.
All the features of AVE migration were lost in Rac1 mutant embryos: null mutant cells failed to extend any projections, and no modulation of junctions or no neighbor exchanges were observed. The behavior of Rac1 mutant AVE cells is very different from what has been observed in Rac1 null fibroblasts, which migrate as rapidly as wild-type cells in response to PDGF by extending pseudopodia-like extensions and do not depend on lamellipodia or membrane localization of Arp2/3 for movement [55]. The difference between the two cell types suggests that migration within an epithelium may be more dependent on Rac1 than is the migration of mesenchymal cells. Rac has previously been shown to be important for collective migration in Drosophila in the border cells of the ovary [56],[57] and during tracheal branching [40]. Collective cell migration is likely to be a primary mode of tissue movement in mammalian branching morphogenesis [58] and tumor invasion [51]; our findings on the AVE raise the possibility that Rac1 is also a crucial regulator of these types of mammalian collective migration.
A Rac1 conditional allele [26] was crossed to mice expressing CAG-Cre [59] to generate a null allele. The Rac1 conditional allele was analyzed in a C3H background and the null allele was analyzed on a mixed C3H/CD1 background. Sox2-Cre [37] and Ttr-Cre [38] were used to delete the gene in the epiblast and VE, respectively. Epiblast-deleted embryos were of the genotype Rac1 floxed/null; Sox2-Cre/+, and VE-deleted embryos were of the genotype Rac1 floxed/null; Ttr-Cre/+. An ES cell line (RRQ139) carrying a gene trap insertion downstream of exon 6 of the Nap1 gene was made by BayGenomics (http://www.mmrrc.org/). This gene trap should produce a fusion of the first 297 amino acids of Nap1 and β-geo. Mice derived from the ES cell clone were kept on a C3H background and are referred to as Nap1RRQ. The Foxa2 allele [60] was analyzed on a mixed CD1/C3H background. The Hex-GFP line was previously described [2]. The BAT-gal line was characterized in [61].
Embryos at e6.5 and older were dissected in PBS 0.4% BSA; e5.5 embryos were dissected in PB1 [62]. In situ hybridization and X-gal staining were carried out as described [63]. Whole-mount embryos were imaged using a Zeiss Axiocam HRC digital camera on a Leica MZFLIII microscope. For immunofluorescence, embryos were fixed in 4% paraformaldehyde (PFA) in PBS for 1 to 3 h on ice and washed in PBS. Fixed embryos were embedded in OCT and cryosectioned at 8 µm thickness. Staining was performed in PBS, 0.1% Triton, 1% heat-inactivated goat serum for e6.5 and e7.5 embryos, and as described in [39] for e5.5 embryos. Whole-mount embryos were imaged on a Leica DM1RE2 inverted confocal. Confocal datasets were analyzed using the Volocity software package (Improvision Inc.).
Antibodies were: rat anti-E-cadherin, 1∶500 (Sigma); rabbit anti-GFP, 1∶500 (Invitrogen); chick anti-GFP, 1∶500 (Abcam); rabbit anti-ZO1, 1∶200 (Zymed). Rac1 antibodies showed background staining, which made it impossible to directly measure the level of residual Rac1 protein in the VE of e5.5 embryos. F-actin was visualized using 10 U/ml TRITC-Phalloidin (Molecular Probes), and nuclei using DAPI (Sigma). Secondary antibodies were from Invitrogen.
Embryos were cultured in 50% DMEM-F12 with HEPES and L-glutamine without phenol red (Gibco), 50% rat serum (Taconic or Harlan) in a dome covered with mineral oil. Imaging was performed using a spinning disk Perkin Elmer UltraVIEW VoX Confocal Imaging System. Nikon Eclipse Ti microscope was equipped with Nikon Plan Fluor 20×, 0.5NA or Plan-NEOFLUAR 40×/1.3 NA, with Hamamatsu EM-CCD C91100-13 (high frame rate format of 512×512 pixels), Hamamatsu C9100-50 (Japan), or Andor iXonEM+EMCCD Camera (Belfast, Northern Ireland) and LiveCell TM temperature-controlled stage (Live Cell Pathology Devices Inc.). Acquisition was controlled by Volocity 5.0.3 Build 4 software or MetaMorph (Molecular Devices). 488 nm laser power was set to 2%. Exposure time was in a range of 400–800 ms. Acquisition frequency was set for 6 time points per hour.
Epifluorescence time-lapse video microscopy was performed on Axiovert 200 M equipped with LD Plan-NEOFLUAR 0.4NA 20× lens, Hamamatsu C4742-80 Orca-ER, and temperature-controlled stage. Acquisition was controlled by Axiovision software.
Images from the live imaging experiments were calibrated, aligned, and adjusted for digital contrast using MetaMorph (Molecular Devices). All measurements were performed using MetaMorph and all data were transferred to Excel (Microsoft) for analysis and representation. Manual alignment of images was performed using Align Stack function (MetaMorph). To analyze migration of AVE cells, we used the distal border of the embryo as a reference point to align images. To monitor protrusive activity of individual cells, alignment was performed against the distal border of the embryo or the center of the cell body. To monitor behavior of individual AVE cells, we manually highlighted cells and the embryo distal border in shades of different colors (50% opacity) using Adobe Photoshop. Maximal projections or images of individual optical sections were used to determine the outline of cells. Towards the end of experiments, when the signal became saturated because the expression of the transgene increased, we used the cell shape of previous time frame assuming that the cell shape did not change significantly. Relative protrusion length was determined by ratio of the protrusion length over the diameter of the cell body. The diameter of cell body was defined as double the distance from the center of the nucleus to the posterior cell border. For the determination of the cell shape factor, we used high resolution Z-stacks of live embryos (step size 0.2–0.5 µm) and chose candidate cells with obvious protrusions. We used multiple optical sections to outline the cell perimeter (Region tool, Metamorph). Thresholded images were used to calculate the shape factor (shape factor = 4πA/P2), which is calculated from the perimeter (P) and the area (A) of the object (the cell) using the Integrated Morphometry tool (Metamorph). 1 corresponds to a perfect circle and values <1 represent progressively more elongated or irregular shapes.
|
10.1371/journal.pcbi.1000529 | Modeling Structure-Function Relationships in Synthetic DNA Sequences using Attribute Grammars | Recognizing that certain biological functions can be associated with specific DNA sequences has led various fields of biology to adopt the notion of the genetic part. This concept provides a finer level of granularity than the traditional notion of the gene. However, a method of formally relating how a set of parts relates to a function has not yet emerged. Synthetic biology both demands such a formalism and provides an ideal setting for testing hypotheses about relationships between DNA sequences and phenotypes beyond the gene-centric methods used in genetics. Attribute grammars are used in computer science to translate the text of a program source code into the computational operations it represents. By associating attributes with parts, modifying the value of these attributes using rules that describe the structure of DNA sequences, and using a multi-pass compilation process, it is possible to translate DNA sequences into molecular interaction network models. These capabilities are illustrated by simple example grammars expressing how gene expression rates are dependent upon single or multiple parts. The translation process is validated by systematically generating, translating, and simulating the phenotype of all the sequences in the design space generated by a small library of genetic parts. Attribute grammars represent a flexible framework connecting parts with models of biological function. They will be instrumental for building mathematical models of libraries of genetic constructs synthesized to characterize the function of genetic parts. This formalism is also expected to provide a solid foundation for the development of computer assisted design applications for synthetic biology.
| Deciphering the genetic code has been one of the major milestones in our understanding of how genetic information is stored in DNA sequences. However, only part of the genetic information is captured by the simple rules describing the correspondence between gene and proteins. The molecular mechanisms of gene expression are now understood well enough to recognize that DNA sequences are rich in functional blocks that do not code for proteins. It has proved difficult to express the function of these genetic parts in a computer readable format that could be used to predict the emerging behavior of DNA sequences combining multiple interacting parts. We are showing that methods used by computer scientists to develop programming languages can be applied to DNA sequences. They provide a framework to: 1) express the biological functions of genetic parts, 2) how these functions depend on the context in which the parts are placed, and 3) translate DNA sequences composed of multiple parts into a model predicting how the DNA sequence will behave in vivo. Our approach provides a formal representation of how the biological function of genetic parts can be used to assist in the engineering of synthetic DNA sequences by automatically generating models of the design for analysis.
| “How much can a bear bear?” This riddle uses two homonyms of the word “bear”. The first instance of the word is a noun referring to an animal, and the second is a verb meaning “endure”. Although the word “bear” has over 50 different meanings in English, its meaning in any given sentence is rarely ambiguous. In a simple case like this riddle, the meaning of each word can be deciphered by looking at other words in the same sentence. In other cases, it is necessary to take into account a broader context to properly interpret the word. For instance, it may be necessary to read several sentences to decide if “bear claw” refers to a body part or a pastry. A reader will progressively derive the meaning of a text by recognizing structures consistent with the language grammar. It is often difficult to understand the meaning of a text by relying exclusively on a dictionary.
It is interesting to compare this bottom-up emergence of meaning with the top-down approach that made genetics so successful. The discipline was built upon a quest to define hereditary units that could be associated with observable traits well before the physical support of heredity was discovered [1],[2]. The one-to-one relationship between genes and traits was later refined by Beadle and Tatum's hypothesis that the gene action was mediated by enzymes [3],[4]. Cracking the genetic code has been one of the major milestones in understanding the information content of nucleic acids sequences. By demonstrating the colinearity of DNA, RNA, and protein sequences, the genetic code was instrumental in the identification of specific DNA sequences as genes. The influence of this legacy on contemporary biology cannot be underestimated. Models used in quantitative genetics predict phenotypes from unstructured lists of alleles at different loci [5],[6]. Similarly, genome annotations remain very gene-centric. Most bioinformatics databases have been designed to collect information relative to coding regions or candidate genes. Few, if any, annotations of non-coding regions or higher order structures are being systematically recorded even for model organisms like yeast [7],[8].
Yet, despite its success, the notion of gene appears insufficient to express the complexity of the relation between an organism genome and its phenotype [1],[9] The elucidation of the molecular mechanisms controlling gene expression has revealed a web of molecular interactions that have been modeled mathematically to show that important phenotypic traits are the emerging properties of a complex system [10]–[15]. The development of this more integrated understanding of the cell physiology leads to a progressive adoption of the more neutral notion of genetic part as a replacement for the notion of genes associated with specific traits. Making sense of the list of parts generated in genomics, proteomics, and metabolomics has been a major challenge for the systems biology community [16]–[21].
It is becoming apparent that the genetic code captures only a small fraction of the information content of DNA molecules [22],[23]. Yet, if there is a general agreement that the cell dynamics is somehow coded in genetic sequences, no formal relationship between DNA sequences and dynamical models of gene expression has been proposed so far. In particular, the formalization of the biological functions of genetic parts has remained elusive. As a result, building models of gene networks encoded in DNA sequences remains a labor-intensive process. This limitation has hampered the development of large families of models needed to analyze phenotypic data generated by libraries of related genetic constructs [24]–[28].
Synthetic biology is likely to be instrumental in refining our understanding of the design of natural biological systems [29]. Just like the genetic code was partly elucidated through the de novo chemical synthesis of DNA molecules [30],[31], the redesign of genomic sequences will shed a new light on the relations between structure and function in genetic sequences [32]–[34]. By considering biological parts as the building blocks of artificial DNA sequences [35], designing new parts that do not exist in nature [26]–[28], and making parts physically available to the community [36], synthetic biology calls for a systematic functional characterization of genetic parts [37]. These efforts are still limited by the difficulty in expressing how the function of biological parts may be influenced by the structure of the DNA sequence in which they are used. It has been shown that a partial redesign of the genomic sequences of two viruses had a significant effect on the virus fitness even though the redesigns preserved the protein sequences [33],[38]. Just as the context of the expression “bear claw” helps understand its meaning, it is necessary to consider the entire structure of the DNA molecule coding for particular genes to appreciate how those genes contribute to the phenotype.
One possible approach to this problem is to extend the linguistic metaphor used to formulate the central dogma. The notions of genetic code, transcription, and translation are derived from a linguistic representation of biological sequences. Several authors have modeled the structure of various types of biological sequences using syntactic models [39]–[46]. However, these structural models have not yet been complemented by formal semantic models expressing the sequence function. An interesting attempt to use grammars to model the dynamics of gene expression did not rely on a description of the DNA sequence structure. Instead, this grammar described how various inducible or repressible promoters can transition between different states under the control of environmental parameters [47]. The simple semantic model stored in a knowledge base established a correspondence between the strings generated by the syntax and the physiological state of the cell. The Sequence Ontology [48] and the Gene Regulation Ontology [49] represent other attempts to associate semantic values with biological sequences. Their controlled vocabularies can be used by software applications to manage knowledge. However, the semantics derived from these ontologies is a semantics of the sequence annotation, not of the sequences themselves.
We recently described a fairly simple syntactic model of synthetic DNA sequences [50] capable of generating a large number of previously published synthetic genetic constructs [24],[25],[51]. We have now enhanced this initial syntactic model with a formal semantic model capable of expressing the dynamics of the molecular mechanisms coded by the DNA sequences. Specialized terms like syntax, semantics, and others are defined in Table 1. Our approach uses attribute grammars [52], a theoretical framework developed in the 60s to establish a formal correspondence between the text of a computer program and the series of microprocessor operations it codes for [53],[54]. Even though other types of semantic models have been developed since then [55],[56], attribute grammars still represent a good compromise between simplicity and expressivity, an important characteristic to ensure that the framework can be used by non-computer scientists. Attribute grammars make it possible to use well characterized compilation algorithms to translate a DNA sequence into a mathematical model of the molecular interactions it codes for. As the static source code of a program directs the dynamic series of operations carried out by the microprocessor based on user inputs, the compilation process translates the static information of cells coded by DNA sequences into a dynamical model of the development of a phenotype in response to environmental influences [57].
The translation of a gene network model from a genetic sequence is very similar to the compilation of the source code of a computer program into an object code that can be executed by a microprocessor (Figure 1). The first step consists in breaking down the DNA sequence into a series of genetic parts by a program called the lexer or scanner. Since the sequence of a part may be contained in the sequence of another part, the lexer is capable of backtracking to generate all the possible interpretations of the input DNA sequences as a series of parts. All possible combinations of parts generated by the lexer are sent to a second program called the parser to analyze if they are structurally consistent with the language syntax. The structure of a valid series of parts is represented by a parse tree [50] (Figure 2). The semantic evaluation takes advantage of the parse tree to translate the DNA sequence into a different representation such as a chemical reaction network. The translation process requires attributes and semantic actions. Attributes are properties of individual genetic parts or combinations of parts. Semantic actions are associated with the grammar production rules. They specify how attributes are computed. Specifically, the translation process relies on the semantic actions associated with parse tree nodes to synthesize the attributes of the construct from the attributes of its child nodes, or to inherit the attributes from its parental node. In our implementation, the product of the translation is a mass action model of the network of molecular interactions encoded in the DNA sequence. By using the standardized format of Systems Biology Markup Language (SBML), the model can be analyzed using existing simulation engines [58]–[60].
We have developed a simple grammar compact enough to be presented extensively, yet sufficiently complex to represent basic epistatic interactions. The grammar generates constructs composed of one or more gene expression cassettes. The gene expression cassettes are themselves composed of a promoter, cistron, and transcription terminator. Finally, a cistron is composed of a Ribosome Binding Site (RBS) and a coding sequence (gene). The syntax is composed of 12 production rules (P1 to P12) displayed in bold characters in Figure 3 where each entry is composed of a rewriting rule (bold), and semantic actions (curly brackets). The symbol ε refers to an empty string, [ , ] to a list, [] to an empty list, and the ‘+’ sign indicates the concatenation operation on two lists. This syntax is comparable to the one described previously [50] except that we introduced the extra non-terminal restConstructs to allow the generation of constructs with multiple cassettes without introducing parsing problems due to direct left recursions [61].
The attributes of a part include the kinetic rates related to this part and the interaction information. For example, the attributes of a promoter include a transcription rate along with a list of proteins repressing it and the kinetic parameters of the protein-DNA interactions. For non-terminal variables corresponding to combinations of parts such as cistrons, the attributes include a list of proteins, a list of promoters, and a list of chemical equations. The equation list is used to store the model of the system behavior, while the lists of promoters and proteins are recorded for computing the molecular interactions resulting from the DNA sequence. The complete set of attributes used in this simple grammar is listed in Table 2.
If many attributes can be computed locally by only considering a small fragment of the DNA sequence, other attributes are global properties of the system. For instance, the computation of protein-DNA interactions requires access to a global list of proteins expressed by the constructs. However, this list is not available until all of the different cassettes have been parsed. The problem is overcome by using a multiple-pass compilation method. In the first pass, the compiler does not do any structural validation but builds the list of proteins in the system and passes the list as an inherited attribute to the second pass. In the second pass, the promoter-protein interactions can be calculated locally at the level of each cassette. Rules P1 to P5 define the structure of a design, while rules P6 to P12 cover the selection of a specific part for each category. In the semantic action, the relation between an attribute and its variable is indicated by a dot and constants are enclosed by brackets. For instance, gene.mRNA_degration_rate = [k6] indicates that the value of the attribute mRNA_degration_rate of a gene is a constant k6. The attribute repressor_list used in P6 and P7 includes the name of the repressor, the stoichiometry, and the kinetic constants of the forward and reverse reactions of the protein-DNA interaction. Table S1 details the parsing steps and computational dependence of each step. Finally, the equation writing operations are handled by functions typed in italics in Figure 3 and defined in Figure 4.
The translation of the DNA sequence into a mathematical model is available as the equation_list attribute of constructs. The model outputs are generated by equations generators, which are purposely decoupled from the semantic actions. The decoupling enables the flexibility of using different equation formats to describe a biological process. The translation of the construct composed of the parts pro_u rbsA gene_v t1 pro_v rbsB gene_u t1 generates the equations displayed in the [Reactions] section of Figure 5. Each line is composed of a reaction index (R1 to R12), the chemical equation itself, and one or two reaction parameters depending on the reaction reversibility. The initial values have been computed by assigning 1 to variables representing DNA sequences and prompting the user to set the initial condition of proteins. The scripts and data used in this report are available in Dataset S1.
The semantic model presented in the previous section is completely modular since the parameters of the model describing the construct behavior are attributes of individual parts, not of higher order structures. For instance, in the previous model (Figures 3 and 4), translational efficiency is primarily determined by the RBS sequence [62],[63]. This association between RBS and translation rate was successfully used to design one of the first artificial gene networks [24] and is still used by many synthetic biology software applications [64]–[67]. Yet, it is also well known that translation initiation can be attenuated by stable mRNA secondary structures [68]–[70]. This leads to a situation where a translational rate can no longer be considered the attribute of an individual part but needs to be considered as the attribute of a specific combination of parts. This type of context-dependency can naturally be expressed using attribute grammars since the translation reaction is computed at the cistron level, not at the level of individual parts. Rule P5 of Figure 3 can be modified by introducing a new function to retrieve the translation rate for specific combination of gene and RBS.
The get_translation_rate function checks for specific cases of interactions between an RBS and coding sequence first. If none is found, then the default RBS translation rate is used.
This approach is illustrated in Table 3 using previously published data demonstrating the interference between the RBS and coding sequence [68]. Specifically, this report provides the relation expression observed in 23 different constructs generated by combining different variants of the RBS and MS2 coat protein gene. This data set has been reorganized in Table 3 by sorting the constructs according to the RBS and gene variants they used. Three of the constructs using the WT RBS sequence resulted in a maximum level of expression while the expression of the gene variants ORF4, ORF5, and ORF6 were expressed at a much lower level due to the greater stability of the mRNA secondary structure. A similar pattern is observed for other RBS variants (RBS1, RBS2, RBS3, RBS7). For all of these RBS variants, it is possible to define the translation_rate function by associating the default translation rate with the maximum expression rate. Specific translation rates associated with particular pairs of RBS and gene variants are recorded separately.
The semantic model in Figures 3 and 4 is a compact proof of concept example, but it does not capture a number of features commonly found in actual genetic constructs. In order to demonstrate that our approach is capable of modeling more realistic DNA sequences, we have extended this semantic model (Supplementary Materials) to translate the DNA sequences of previously published DNA plasmids that include polycistronic cassettes in different orientations [24]. This plasmid library was generated by 32 different genetic parts (three promoters: pLtetO-1, pLs1con, ptrc-2; eight RBS: rbsA to rbsH; and four genes: tetR, cIts, lacI, and gfp and one terminator, all in both orientations). The syntax generates 72 different single gene expression constructs in each orientation. By combining two genes repressing each other in a construct, it is possible to make bistable artificial gene networks that are represented in Figure 6. These bistable networks can be used as a genetic switch.
To demonstrate the potential use of a semantic model to search for a desirable behavior in a large genetic design space, we have generated the DNA sequences of all 41,472 possible sequences (722×8 RBS for the reporter gene) having the same structure as previously described switches. All sequences were translated into separate model files and a script was developed to perform a bistability analysis of each model. Parameters of the semantic model were obtained by qualitatively matching the experimental results of the six previously published switches [24] and are summarized in Table S2. Most of the automatically generated sequences led to inherently non-bistable networks because the necessary repressor/promoter pairs did not match. Since this specific example is particularly well understood, we could have generated a limited number of targeted constructs. Yet, we chose to generate all possible sequences to demonstrate the generality of our approach. In particular, it was important to evaluate the computational cost of generating and translating DNA sequences to ensure that it would not prevent a systematic exploration of more complex design spaces. It takes only minutes to generate 41,472 sequences and translate them into SBML files. Hence, the computational cost of this step is negligible compared to the time required by the simulation of the SBML files.
Bistability was tested numerically by integrating the differential equations until they converged to a steady state starting from two different initial conditions. The two initial conditions started with one protein level very high and the other very low and vice versa. We characterized the bistability by computing the ratio of reporter concentration for the two steady state values. In order to globally verify the behavior of this large population of models, we focused on the 3,072 constructs potentially capable of bistability, 1,408 of which were found to be bistable. We further reduced the number of constructs used to verify the translation process from 3,072 to 384 by assuming that two constructs differing only in the RBS in 5′ of the reporter gene would produce the same ratio of steady state values. Figure 6 visualizes the behavior of these 384 constructs. Constructs that are not bistable have a ratio of 1. This ratio gives insight into how the construct is expected to be experimentally detectable. Since most experimental methods cannot give an exact value of protein concentration, a high ratio is desired to rise above experimental noise. Each of the 6 windows is analogous to the previously described two-parameter bifurcation diagram for that pair of repressors [24]. This gives confidence that both the semantic model of DNA sequences and the compiler used to translate automatically generated DNA sequences give results consistent with manually developed models of this family of gene networks. In the long term, the advantage to our approach over a traditional two-parameter bifurcation is the association of discrete parameter values with specific parts. This will prove particularly valuable when the context-dependencies of parameter values are better documented experimentally.
This example demonstrates the benefit of building a semantic model of synthetic DNA sequences. Even a small library of genetic parts can generate large numbers of artificial gene networks having no more than a few interacting genes. A syntactic model describing how parts can be combined into constructs is a compact representation of the genetic design space generated from the parts library. While it is possible to manually build mathematical models capturing the dynamics of some of these artificial gene networks individually, it becomes desirable to automate the process to ensure the model consistency when building large families of related models derived from the same parts library. By considering genetic parts as the terminal symbols of an attribute grammar, it becomes possible to automatically generate models of numerous artificial gene networks derived from this parts library and quickly identify the optimal designs [71].
The parameter values used in the previous example were selected to match an extremely small set of six experimental data points. Although the under-determination of the model does not make it possible to precisely estimate the value of these parameters, the example illustrates how the framework could provide valuable guidance in selecting specific parts for a design. Considering that the exact value of parameters for parts is still a far off perspective, the automatic exploration of the design space presented here will provide useful guidance in construct design. For example, robust constructs from the cusp interior of the tetR/cI and lacI/cI pairings could be built and tested while less robust switches based on the lacI/tetR pairing would be avoided. As more is learned about these parts including the specific rates in different genetic contexts, the predictive ability of such maps will increase. Other motifs could be explored in a similar manner. For example, oscillators [11] could be explored by permuting parts and calculating the model-predicted existence of oscillations as well as their period or amplitude.
The approach presented in this report will be implemented into GenoCAD [72], the web-based tool we have developed to give biologists access to our syntactic design framework. Through GenoCAD, users will benefit from the syntactic and semantic models of various parts sources (GenoCAD provided library, MIT Registry of Standard Biological Parts, or user created parts library). Initially, users will be able to translate their designs into SBML files that could be imported in SBML-compliant simulation tools (www.sbml.org/SBML_Software_Guide) for further analysis. At a later stage, simulation results and more advanced numerical analyses will be seamlessly integrated in GenoCAD's workflow. One of the major obstacles toward the implementation of such semantic models in GenoCAD is the development of a data model allowing users to understand and possibly edit the functional model of the parts they use.
A function description language called Genetic Engineering of living Cells (GEC) was recently introduced to specify the properties of a design [67]. GEC is capable of finding a DNA sequence that implements the desirable phenotypic functions. Several other software applications have been recently released to design biological systems from standardized genetic parts. ASMPART [65], SynBioSS [66], a specialized ProMot package [64] and TinkerCell (www.tinkercell.com) illustrate this trend. These tools are still exploratory. One of their limitations is the requirement to define parts in a specialized format, such as SBML or Modeling Description Language (MDL). Furthermore, instead of defining parts interactions in the underlying parts data models, these tools rely on the user to manually define them textually [66] or graphically [64]. As a result of this specific limitation, several of these tools do not appear suitable for the automatic exploration of a design space. Moreover, they tend to rely on a loosely defined relationship between the structure of the genetic constructs and their behavior. They allow parts to be assembled in any order without regard for biological viability.
Still, the scripts developed to generate our results are of lesser importance than the application of the theory of semantics-based translation using attribute grammars to the translation of DNA sequences into dynamical models representing the molecular interactions they encode. Since this approach is used to develop the compilers of many computer languages [56],[73], a wealth of existing theoretical results and software tools can find new applications in the life sciences. For instance, we have implemented semantic models of DNA sequences into two widely used but very different programming environments, Prolog [74] and ANTLR [75]. Future research efforts will need to investigate the pros and cons of different compiler generators and different parsing algorithms for analyzing even genome-scale DNA sequences and how they impact the ability of grammars to express various features of DNA sequences. Also, the type of attributes associated with parts is flexible. Here we primarily use mass action kinetic rates as attributes, but we could just as easily have used an emerging synthetic biology unit like polymerase per second (PoPS) [37],[76].
Ultimately, tools capable of automatically generating models of the behavior of synthetic DNA sequences will be important for the advancement of synthetic biology [71]. However, these tools will need to be able to express that the contribution of a genetic part to the phenotype of an organism depends largely on the local and global context in which it is placed. The interference between RBS and coding sequence is just one example of the biological complexity that computer assisted design applications will have to properly consider.
Before it will be used to build synthetic genetic systems meeting user-defined specifications, the semantic model of DNA sequences presented in this report will be instrumental in the quantitative characterization of structure-function relationships in synthetic DNA sequences. The vision of applying quantitative engineering methods to biological problems has been recognized as a promising avenue to biological discovery [29]. The critical role of artificial gene networks in the characterization of molecular noise affecting the dynamics of gene networks [77] illustrates the potential of synthetic biology as a route to refine the understanding of basic biological processes.
Ongoing efforts aim to carefully define how parts should fit together syntactically and what attributes are needed to characterize their function. For example, the sequence between the RBS and the start codon has been shown to play an important role in translation rate [63]. The question arises whether the RBS should be defined to include the spacing, or if there should be a separate parts category for the spacer. The rapid development of gene synthesis techniques [78] will make it possible to investigate these questions with a base-level resolution. Beyond libraries of parts for designing expression vectors, similar curation efforts could lead to the identification of parts in genomic sequences, whereby the hypothetical function of these parts as they are expressed in attribute grammars could be tested by genome refactoring [33].
|
10.1371/journal.pcbi.1005811 | Filament turnover tunes both force generation and dissipation to control long-range flows in a model actomyosin cortex | Actomyosin-based cortical flow is a fundamental engine for cellular morphogenesis. Cortical flows are generated by cross-linked networks of actin filaments and myosin motors, in which active stress produced by motor activity is opposed by passive resistance to network deformation. Continuous flow requires local remodeling through crosslink unbinding and and/or filament disassembly. But how local remodeling tunes stress production and dissipation, and how this in turn shapes long range flow, remains poorly understood. Here, we study a computational model for a cross-linked network with active motors based on minimal requirements for production and dissipation of contractile stress: Asymmetric filament compliance, spatial heterogeneity of motor activity, reversible cross-links and filament turnover. We characterize how the production and dissipation of network stress depend, individually, on cross-link dynamics and filament turnover, and how these dependencies combine to determine overall rates of cortical flow. Our analysis predicts that filament turnover is required to maintain active stress against external resistance and steady state flow in response to external stress. Steady state stress increases with filament lifetime up to a characteristic time τm, then decreases with lifetime above τm. Effective viscosity increases with filament lifetime up to a characteristic time τc, and then becomes independent of filament lifetime and sharply dependent on crosslink dynamics. These individual dependencies of active stress and effective viscosity define multiple regimes of steady state flow. In particular our model predicts that when filament lifetimes are shorter than both τc and τm, the dependencies of effective viscosity and steady state stress on filament turnover cancel one another, such that flow speed is insensitive to filament turnover, and shows a simple dependence on motor activity and crosslink dynamics. These results provide a framework for understanding how animal cells tune cortical flow through local control of network remodeling.
| In this paper, we develop and analyze a minimal model for a 2D network of cross-linked actin filaments and myosin motors, representing the cortical cytoskeleton of eukaryotic cells. We implement coarse-grained representations of force production by myosin motors and stress dissipation through an effective cross-link friction and filament turnover. We use this model to characterize how the sustained production of active stress, and the steady dissipation of elastic stress, depend individually on motor activity, effective cross-link friction and filament turnover. Then we combine these results to gain insights into how microscopic network parameters control steady state flow produced by asymmetric distributions of motor activity. Our results provide a framework for understanding how local modulation of microscopic interactions within contractile networks control macroscopic quantities like active stress and effective viscosity to control cortical deformation and flow at cellular scales.
| Cortical flow is a fundamental and ubiquitous form of cellular deformation that underlies cell polarization, cell division, cell crawling and multicellular tissue morphogenesis [1–6]. Cortical flows originate within a thin layer of cross-linked actin filaments and myosin motors, called the actomyosin cortex, that lies just beneath the plasma membrane [7]. Local forces produced by bipolar myosin filaments are integrated within cross-linked networks to build macroscopic contractile stress [8–10]. At the same time, cross-linked networks resist deformation and this resistance must be dissipated by network remodeling to allow macroscopic deformation and flow. How force production and dissipation depend on motor activity and network remodeling remains poorly understood.
One successful approach to modeling cortical flow has relied on coarse-grained phenomenological descriptions of actomyosin networks as active fluids, whose motions are driven by gradients of active contractile stress and opposed by an effectively viscous resistance [11]. In these models, spatial variation in active stress is typically assumed to reflect spatial variation in motor activity and force transmission [12], while effective viscosity is assumed to reflect the internal dissipation of elastic resistance due to local remodeling of filaments and/or cross-links [7, 13]. Models combining an active fluid description with simple kinetics for network assembly and disassembly, can successfully reproduce the spatiotemporal dynamics of cortical flow observed during polarization [11], cell division [14, 15], cell motility [16, 17] and tissue morphogenesis [18]. However, it remains a challenge to connect this coarse-grained description of cortical flow to the microscopic origins of force generation and dissipation within cross-linked actomyosin networks.
Studies in living cells reveal fluid-like stress relaxation on timescales of 10-100s [1, 2, 11, 19–21], which is thought to arise through a combination of cross link unbinding and actin filament turnover [7, 13, 22]. Theoretical [23, 24] and computational [25–27] studies reveal that cross-link unbinding can endow actin networks with complex time-dependent viscoelasticity. However, while cross-link unbinding is sufficient for viscous relaxation (creep) on very long timescales in vitro, it is unlikely to account for the rapid cortical deformation and flow observed in living cells [26, 28–31]. Experimental studies in living cells reveal rapid turnover of cortical actin filaments on timescales comparable to stress relaxation (10-100s) [32–35]. Perturbing turnover can lead to changes in cortical mechanics and in the rates and patterns of cortical flow [33, 36]. However, the specific contributions of actin turnover to stress relaxation and how these depend on network architecture remain unclear.
Recent work has also begun to reveal mechanisms for active stress generation in disordered actomyosin networks. Theoretical studies suggest that spatial heterogeneity in motor activity along individual filaments, and asymmetrical filament compliance (stiffer in extension than in compression), are sufficient for macroscopic contraction [37, 38], although other routes to contractility may also exist [38]. Local interactions among actin filaments and myosin motors are sufficient to drive macroscopic contraction of disordered networks in vitro [39], and the kinematics of contraction observed in these studies support a mechanism based on asymmetrical filament compliance and filament buckling. However, in these studies, the filaments were preassembled and network contraction was transient, because of irreversible network collapse [40], or buildup of elastic resistance [41], or because network rearrangements (polarity sorting) dissipate the potential to generate contractile force [42–45]. This suggests that network turnover may play essential role(s) in allowing sustained production of contractile force. Recent theoretical and modeling studies have begun to explore how this might work [46–48], and to explore dynamic behaviors that can emerge when contractile material undergoes turnover [15, 49]. However, it remains a challenge to understand how force production and dissipation depend individually on the local interplay of network architecture, motor activity and filament turnover, and how these dependencies combine to mediate tunable control of long range cortical flow.
Here, we construct and analyze a simple computational model that bridges between the microscopic description of cross-linked actomyosin networks and the coarse grained description of an active fluid. We represent actin filaments as simple springs with asymmetric compliance; we represent dynamic binding/unbinding of elastic cross-links as molecular friction [50–52] at filament crossover points; we represent motor activity as force coupling on a subset of filament cross-over points with a simple linear force/velocity relationship [53]. Finally, we model filament turnover by allowing entire filaments to appear with a fixed probability per unit area and disappear with fixed probabilities per unit time. We use this model to characterize: first the passive response of a cross-linked network to externally applied stress, then the buildup and maintenance of active stress against an external resistance, and finally the steady state flows produced by an asymmetric distribution of active motors in which active stress and passive resistance are dynamically balanced across the network. Our results reveal how network remodeling can tune cortical flow through simultaneous effects on active force generation and passive resistance to network deformation.
Our goal is to construct a minimal model that is sufficiently detailed to capture essential microscopic features of cross-linked actomyosin networks (actin filaments with asymmetric compliance, dynamic cross-links, active motors and and continuous filament turnover), but simple enough to explore, systematically, how these microscopic features control macroscopic deformation and flow. We focus on 2D networks because they capture a reasonable approximation of the quasi-2D cortical actomyosin networks that govern flow and deformation in many eukaryotic cells [11, 54], or the quasi-2D networks studied recently in vitro [39, 55].
Fig 1 provides a schematic overview of our model’s assumptions. We model each filament as an oriented elastic spring with rest length L. If index i enumerates over all filaments, then the state of a filament i is defined by the positions of its endpoints bi and pi, marking its barbed (+) and pointed (-) ends respectively. We define u i ^ to be the unit vector oriented along filament i towards its barbed end.
We assume (Fig 1A) that local deformation of filament i gives rise to equal and opposite elastic restoring forces on its endpoints:
F p , i elas = μ γ i u i ^ , F b , i elas = - F p , i elas (1)
where γi = (|bi − pi| − L)/L is the strain on filament i, and μ is a normalized spring constant. To model asymmetric filament compliance, we use a piecewise linear approximation to the non-linear entropic force extension curve for semi-flexible polymers with lengths less than the persistence length [56, 57]. We set μ = μe if the strain is positive (extension), and μ = μc if the strain is negative (compression).
Previous models have represented cross-linkers as elastic connections between pairs of points on neighboring filaments that bind and unbind with either fixed or force-dependent probabilities [25, 58]. Here, we introduce a coarse-grained representation of crosslink dynamics by introducing an effective drag force that couples every pair of overlapping filaments, and which represents a molecular friction arising from the time-averaged contributions of many individual transient crosslinks (Fig 1B). This is a reasonable approximation if all pairs of overlapping filaments have equal access to a non-limiting pool of cross links, and if the rate at which filaments move past one another is slow relative to the unbinding rate of individual crosslinks [59]. This coarse-grained approach has been used to model frictional forces arising from ionic cross-linking of actin filaments in vitro [60, 61], and simple force-velocity relationships for systems of cytoskeletal filaments and cross-linking motors [53, 62–64].
To implement coupling through effective drag, for any pair of overlapping filaments i and j, we write the drag force on filament i as:
F i , j ξ = - ξ ( v i - v j ) (2)
where ξ is the drag coefficient and vi, vj are the average centroid velocities of filaments i and j. We apportion this drag force to the two endpoints pi and bi of filament i as follows: If xi,j is the position of the filament overlap, then we define λi,j = |xi,j − pi|/|bi − pi| to be the fractional position of the overlap point along filament i, and we assign ( 1 - λ i , j ) F i , j ξ to endpoint pi and λ i , j F i , j ξ to endpoint bi.
The total crosslink coupling forces on endpoints pi and bi, due to overlaps along filament i, can then be written:
F p , i xl = ∑ j ( 1 - λ i , j ) F i , j ξ F b , i xl = ∑ j λ i , j F i , j ξ (3)
where the sums are taken over all filaments j that overlap with filament i.
This model assumes a linear relation between the drag force and the velocity difference between attached filaments. Although non-linearities can arise through force dependent detachment kinetics and/or non-linear force extension of cross-links, we assume here that these non-linear effects are of second or higher order.
We add motor activity at the point of overlap between two filaments i and j as follows: For each filament in the pair, we impose an additional force of magnitude υ, directed towards its pointed (-) end (Fig 1C):
F i υ = - υ u i ^ (4)
We impose an equal and opposite force on its overlapping partner. We distribute these forces to filament endpoints as described above for crosslink coupling forces. Thus, the total force on endpoints i and i+1 due to motor activity on overlap points between filaments i and j can be written as:
F p , i motor = ∑ j ( 1 - λ i , j ) ( F i υ - F j υ ) q i , j = υ ∑ j ( 1 - λ i , j ) ( u j ^ - u i ^ ) q i , j F b , i motor = υ ∑ j ( λ i , j ) ( u j ^ - u i ^ ) q i , j (5)
where j enumerates over all filaments j that overlap with filament i, and qi,j equals 0 or 1 depending on whether there is an “active” motor at this location. To model dispersion of motor activity, we set qi,j = 1 on a randomly selected subset of filament overlaps, such that q ¯ = ϕ, where q ¯ indicates the mean of q (Fig 1C).
To write the full equation of motion for a network of actively and passively coupled elastic filaments, we assume the low Reynold’s number limit in which inertial forces can be neglected, and we equate the sum of all forces acting on all filament endpoints to zero to obtain:
0 = - ζ v i p L / 2 - F p , i xl + F p , i elas + F p , i motor 0 = - ζ v i b L / 2 - F b , i xl + F b , i elas + F b , i motor (6)
where the first terms in each equation represent the hydrodynamic drag on the half-filaments adjoining endpoints pi or bi with respect to motion at velocities v i p or v i b against the surrounding fluid, and ζ is the drag coefficient.
We used a mikado model approach [65] to initialize a minimal network of overlapping unstressed linear filaments in a rectangular 2D domain. We generate individual filaments by laying down straight lines, of length L, with random position and orientation. We define the density using the average distance between cross-links along a filament, lc. A simple geometrical argument can then be used to derive the number of filaments filling a domain as a function of L and lc [66]. Here, we use the approximation that the number of filaments needed to tile a rectangular domain of size Dx × Dy is 2DxDy/Llc, and that the length density is therefore simply, 2/lc.
Although we do not model thermal forces explicitly, the contribution of thermal fluctuations to filament elasticity are embedded in our coarse-grained representation of asymmetrical filament compliance. In principle, thermally-driven transverse fluctuations of filament segments between crosslink points could influence crosslink binding kinetics. However, for the network mesh sizes considered here, lc <= 0.5μm, the root mean square amplitude of these fluctuations is predicted to be < 5nm (see e.g [56]), suggesting that these effects will be minor. Hence, we have chosen to ignore them here.
In living cells, actin filament assembly is governed by multiple factors that control filament nucleation, branching and elongation. Likewise filament disassembly is governed by multiple factors that promote filament severing and monomer dissociation at filament ends. Here, we implement a very simple model for filament turnover in which entire filaments appear with a fixed rate per unit area, kapp and disappear at a rate kdissρ, where ρ is the filament density (Fig 1D). With this assumption, in the absence of network deformation, the density of filaments will equilibrate to a steady state density, kapp/kdiss, with time constant τr = 1/kdiss. In deforming networks, filament density will also decrease under extensional strain and increase under compressional strain. Thus filament density will be set by a dynamic interplay of deformation and density equilibration via turnover (see below and (S1 Appendix A.3)). To implement this model, at fixed time intervals τs < 0.01 ⋅ τr (i.e. 1% of the equilibration time), we selected a fraction, τs/τr, of existing filaments (i.e. less than 1% of the total filaments) for degradation. We then generated a fixed number of new unstrained filaments kappτsDxDy at random positions and orientations within the original domain. We refer to kdiss = 1/τr as the turnover rate, and to τr as the turnover time.
Further details regarding our simulation approach and references to our code can be found in the Supplementary Information (S1 Appendix A.1). Briefly, eqs 1–6 define a coupled system of ordinary differential equations that can be written in the form:
A · x ˙ = f ( x ) (7)
where x is a vector of filament endpoint positions, x ˙ the endpoint velocities, A is a matrix with constant coefficients that represent crosslink coupling forces between overlapping filaments, and f(x) represents the active (motor) and elastic forces on filament endpoints. We smoothed all filament interactions, force fields, and constraints linearly over small regions such that the equations contained no sharp discontinuities. We used a fourth-order Runga-Kutta method to numerically integrate this system of equations to find the time evolution of the positions of all filament endpoints. We generate a network of filaments with random positions and orientations as described above within a domain of size Dx by Dy. For all simulations, we imposed periodic boundaries in the y-dimension. To impose an extensional force per unit length (2D stress) on the network, we constrained all filament endpoints within a fixed distance 0.05 ⋅ Dx from the left edge of the domain to be non-moving, then we imposed a rightwards force on all endpoints within a distance 0.05 ⋅ Dx from the right edge of the domain, such that the force per unit length of boundary equals the desired stress value. To simulate free contraction, we removed all constraints at domain boundaries; to assess buildup and maintenance of contractile stress under isometric conditions, we used periodic boundary conditions in both x and y dimensions.
In our 2D model, we measure stress as a force per unit length. We measured the internal network stress at each axial position by summing the axial (x) component of the tensions on all filaments intersecting that position, and dividing by network height Dy. We quantified two different forms of strain: the average filament strain, which measures the deformations of individual filaments, and the cumulative network strain, a normalized measure of network deformation defined as the change in axial length of a patch of network divided by its original length. These two measures can differ because filaments can slide relative to one another during deformation and because strained filaments are replaced by unstrained filaments during network turnover. We measured the strain on individual filaments as defined above from γi = (|bi − pi| − L)/L. Then we averaged this measurement over all filaments in a network to obtain an average filament strain. To measure the average network strain rate, we first measured the mean velocity v(X) at position X (relative to the network boundary at x = 0) to be the average velocities of all filaments intersecting that position. In the cases where we measure network strain or strain rates, we observed an approximately linear dependence of v(X) on X; hence the strain rate is approximately uniform across the network (Fig 2B and S6 Fig). Accordingly, for each filament, we took 1 X d X d t to be an estimate of the strain rate on the network between x = 0 and x = X. We averaged this estimate over all filaments in a domain to get an average strain rate. Finally, to estimate the cumulative network strain at a given time T in the simulation, we integrated the strain rate with respect to time for t = 0 to T.
We assigned biological plausible reference values for all parameters (See Table 1). For individual analyses, we sampled the ranges of parameter values around these reference values shown in S1 Table.
The goal of this study is to understand how cortical flow is shaped by the simultaneous dependencies of active stress and effective viscosity on filament turnover, crosslink drag and on “network parameters” that control filament density, elasticity and motor activity. We approach this in three steps: First, we analyze the passive deformation of a cross-linked network in response to an externally applied stress; we identify regimes in which the network response is effectively viscous and characterize the dependence of effective viscosity on network parameters and filament turnover. Second, we analyze the buildup and dissipation of active stress in cross-linked networks with active motors, as they contract against an external resistance; we identify conditions under which the network can produce sustained stress at steady state, and characterize how steady state stress depends on network parameters and filament turnover. Finally, we confirm that the dependencies of active stress and effective viscosity on network parameters and filament turnover are sufficient to predict the dynamics of networks undergoing steady state flow in response to spatial gradients of motor activity.
Thus far, we have considered independently how network remodeling controls the passive response to an external stress, or the steady state stress produced by active contraction against an external resistance. We now consider how these two forms of dependence can combine to shape steady state flow produced by spatial gradients of motor activity. To this end, we model a simple scenario in which motor activity is continuously patterned such that the right half network has uniformly high levels of motor activity (controlled by υ, with ϕ = 0.5), while the left half network has none (ϕ = 0). For simplicity, we imposed periodic boundary conditions at left and right boundaries. Under these conditions, with filament turnover, we expect the right half network to contract continuously against a passive resistance from the left half network. Given the highly asymmetric filament compliance, the internal resistance of the right half network to active compression should be negligible compared to the external resistance of the left half network to extension. Thus the steady state flow should be well-described by:
γ ˙ = σ s s η (11)
where σss is the active stress generated by the right half-network (less the internal resistance to filament compression), η is the effective viscosity of the left half network and strain rate γ ˙ is measured in the left half-network. Note that strain rate can be related to the steady state flow velocity v at the boundary between right and left halves through v = γ ˙ D x. Therefore, we can understand the dependence of flow speed on filament turnover and other parameters using the approximate relationships summarized by eqs 9 and 10 for η and σss. As shown in Fig 8, there are two qualitatively distinct possibilities for the dependence of strain rate on τr, depending on the relative magnitudes of τm and τc. In both cases, for fast enough turnover (τr < min(τm, τc)), we expect weak dependence of strain rate on τr (γ ˙ ∼ τ r 1 / 4). For all parameter values that we sampled in this study (which were chosen to lie in a physiological range), τm > τc. Therefore we predict the dependence of steady state strain rate on τr shown in Fig 8A.
To test this prediction, we simulated the simple scenario described above for different values of τr, with all other parameter values initially fixed at their reference values (Fig 9A). As expected, for all values of τr, the asymmetric pattern of active contraction gave rise to steady state flow, characterized by continuous contraction of the right half-network and expansion of the left half-network, with peak velocity at the boundaries between right (contracting) and left (expanding) domains (Fig 9B). At long times, the average strain on individual filaments reached a plateau (see S6 Fig), but the cumulative network strain increased linearly with time (Fig 9C), indicating steady state flow with a constant strain rate and peak velocity. Plotting steady state strain rate vs filament lifetime τr confirmed the predicted dependence: Steady state strain rates approached zero with increasing τr; however, for decreasing τr, steady state strain rates increased steadily, before reaching an approximate plateau on which strain rate varied by less than 15% over more than two decades of variation in τr (Fig 9D).
We repeated these simulations for a wider range of parameter values, and saw the same qualitative dependence of γ ˙ on τr in all cases. Using eq 9 with τr < τc and eq 10 with τr < τm, and the theoretical or empirical scaling relationships found above for ηc, τc, σm and τm, we predict a simple scaling relationship for γ ˙ (for small τr, see Fig 9D):
γ˙=υξL(τr)1/4 (12)
Indeed, when we plot the steady state measurements of γ ˙, normalized by υ/ξL, for all parameter values, the data collapse onto a single curve for small τr. Thus. our simulations identify a flow regime, characterized by sufficiently fast filament turnover, in which the steady state flow speed is buffered against variation in turnover, and has a relatively simple dependence on other network parameters.
Cortical flows arise through a dynamic interplay of force production and dissipation within cross-linked actomyosin networks. Here we combined computer simulations with simple theoretical analysis to explore how this interplay depends on motor activity, crosslink dynamics, network architecture and filament turnover. Our results reveal two essential requirements for filament turnover during cortical flow: (a) to allow the continuous relaxation of elastic resistance without catastrophic loss of network connectivity and (b) to prevent the dissipation of active stress through local network rearrangements. We find that biphasic dependencies of active stress and passive relaxation on filament lifetime define multiple modes of steady state flow with distinct dependencies on network parameters and filament turnover.
We identify two distinct modes of passive response to uniaxial stress: a low turnover mode in which filaments strain to an elastic limit before turning over, and effective viscosity depends on crosslink density and effective crosslink friction, and a high turnover mode in which filaments turn over before reaching an elastic limit and effective viscosity is proportional to elastic resistance and approximately proportional to filament lifetime. We note that the weakly sub-linear dependence of effective viscosity on filament lifetime that we observe in the high turnover regime may simply reflect a failure to capture very local modes of filament deformation, since a previous study [68] in which filaments were represented as connected chains of smaller segments predicted linear dependence of effective viscosity on filament lifetime. While previous studies have emphasized individual roles for cross-link unbinding or filament turnover in stress relaxation [7, 13, 22], here we have capture their distinct contributions within a single self-consistent modeling framework.
Our simulations confirm the theoretical prediction [37, 39, 69] that spatial heterogeneity of motor activity and asymmetric filament compliance are sufficient to support macroscopic contraction of unconstrained networks. However, under isometric conditions, and without filament turnover, our simulations predict that active stress cannot be sustained. On short timescales, motor forces drive local buildup of extensional stress, but on longer timescales, active local filament rearrangements lead, invariably, to a decay in active stress. These rearrangements can lead to macroscopic network tearing and fragmentation, as previously described [40, 47]. However, stress decay can also occur when the distribution of network filaments remains more uniform and the network remains globally connected in the sense that every filament overlaps, and can exchange frictional crosslink forces, with many others. Under these conditions, network rearrangements involve a slower rebalancing of extensile and compressive forces on network filaments. Our results suggest that when filaments can slide relative to one another, the motor forces that produce active stress will drive changes either in connectivity, or in the distributions of forces along individual filaments, or both, that inevitably lead to a decrease in active stress. Thus for contractile networks to maintain isometric tension on long timescales, they must either form stable crosslinks to prevent filament rearrangements, or they must continuously turnover network filaments (or active motors) to renew the local potential for production of active stress.
Indeed, our simulations predict that filament turnover is sufficient for maintenance of active stress. As in the passive case, they predict biphasic dependence of steady state stress on filament turnover: For short-lived filaments (τr < τm), steady state stress increases linearly with filament lifetime because filaments have more time to build towards peak extensional stress before turning over. For longer-lived filaments (τr > τm), steady state stress decreases monotonically with filament lifetime because local rearrangements decrease the mean contributions of longer lived filaments. These findings imply that for cortical networks that sustain contractile stress under approximately isometric conditions, tuning filament turnover can control the level of active stress, and there will be an optimal turnover rate that maximizes the stress, all other things equal. This may be important, for example in early development, where contractile forces produced by cortical actomyosin networks maintain, or drive slow changes in, cell shape and tissue geometry [7, 70].
For cortical networks that undergo steady state flows driven by spatial gradients of motor activity, our simulations predict that the biphasic dependencies of steady state stress and effective viscosity on filament lifetime define multiple regimes of steady state flow, characterized by different dependencies on filament turnover (and other network parameters). In particular, the approximately linear dependencies of steady state stress and effective viscosity on filament lifetime for short-lived filaments define a fast turnover regime in which steady state flow speeds are buffered against variations in filament lifetime, and are predicted to depend in a simple way on motor activity and crosslink resistance. Measurements of F-actin turnover times in cells that undergo cortical flow [32, 71–75] suggests that they may indeed operate in this fast turnover regime. For reference values of model parameters, the steady state strain rates predicted for the high turnover regime (≈ 2x10−4/sec) are approximately ten-fold lower than those measured in polarized C. elegans zygotes (≈ 1 − 2x10−3/sec) [6, 11, 76]. This is reasonable agreement, given uncertainties about these reference values. For example, our reference value for effective crosslink friction ξ is 10-100-fold higher than friction coefficients measured for single crosslinkers and molecular motors in vitro [77, 78]. A ten-fold lower value for this parameter would yield a ten-fold increase in the predicted steady state strain rate (12). Interestingly, recent studies in C. elegans embryos suggests that cortical flow speeds are surprisingly insensitive to depletion of factors (ADF/Cofilin) that govern filament turnover [11], again consistent with our model’s predictions. Stronger tests of our model’s predictions will require more systematic analyses of how flow speeds vary with filament and crosslink densities, motor activities, and filament lifetimes.
|
10.1371/journal.ppat.1003122 | Adaptive Immunity Alters Distinct Host Feeding Pathways during Nematode Induced Inflammation, a Novel Mechanism in Parasite Expulsion | Gastrointestinal infection is often associated with hypophagia and weight loss; however, the precise mechanisms governing these responses remain poorly defined. Furthermore, the possibility that alterations in feeding during infection may be beneficial to the host requires further study. We used the nematode Trichinella spiralis, which transiently inhabits the small intestine before migrating to skeletal muscle, as a biphasic model of infection to determine the cellular and molecular pathways controlling feeding during enteric and peripheral inflammation. Through the infection of genetically modified mice lacking cholecystokinin, Tumor necrosis factor α receptors and T and B-cells, we observed a biphasic hypophagic response to infection resulting from two separate immune-driven mechanisms. The enteroendocrine I-cell derived hormone cholecystokinin is an essential mediator of initial hypophagia and is induced by CD4+ T-cells during enteritis. In contrast, the second hypophagic response is extra-intestinal and due to the anorectic effects of TNFα during peripheral infection of the muscle. Moreover, via maintaining naive levels of the adipose secreted hormone leptin throughout infection we demonstrate a novel feedback loop in the immunoendocrine axis. Immune driven I-cell hyperplasia and resultant weight loss leads to a reduction in the inflammatory adipokine leptin, which in turn heightens protective immunity during infection. These results characterize specific immune mediated mechanisms which reduce feeding during intestinal or peripheral inflammation. Importantly, the molecular mediators of each phase are entirely separate. The data also introduce the first evidence that I-cell hyperplasia is an adaptively driven immune response that directly impinges on the outcome to infection.
| Infection with intestinal parasites often results in a period of reduced appetite which can result in weight loss; however the factors which control these feeding alterations and the reason why they occur is unknown. We used the nematode parasite Trichinella spiralis, which during its life cycle causes intestinal and muscular inflammation, as a mouse infection model to study the factors which alter feeding during infection. We found that the mouse immune response to the parasite was driving two periods of reduced feeding by two distinct immune mediators during the intestinal and muscular periods of infection. Interestingly, the immune system was utilizing a hormone which usually terminates feeding during our daily meals to cause a reduction in weight and fat deposits. Furthermore, we found that a reduction in these fat deposits and their associated hormones actually helped the mouse expel the parasite from the intestine. Hence the immune driven weight loss was actually beneficial to the mouse's ability to resolve an infection. Our study provides novel insights into how the immune system interacts with feeding pathways during intestinal inflammation and may help us design new strategies for helping people with parasitic infections of the gut.
| Intestinal inflammation is commonly associated with reduced feeding (hypophagia) and weight loss [1], [2], yet the mechanisms and underlying principles of these responses is unknown. Infection with the intestinal parasites Ascaris suum and Trichostrongylus colubriformis results in hypophagia that is coupled with an increase in cholecystokinin (CCK) released from I-cells [3], [4]; a subset of intestinal epithelial enteroendocrine cells (EECs). Despite only comprising 1% of the epithelium, EECs collectively form the largest mammalian endocrine system. Regulatory peptides and amines are released from EECs in response to luminal nutrients [5] and these peptides signal via vagal afferent fibers to feeding control centers in the brain. These EEC signals in concert with leptin, produced from adipose tissues indicating levels of fat deposits, ultimately control our daily short-term feeding patterns. However, the true biological function and molecular mechanisms that orchestrate the pathways driving hypophagia and weight loss during inflammation have not been addressed.
The nematode Trichinella spiralis produces a well characterized CD4+ T-cell, Th2 driven transient inflammation in the small intestine culminating in worm expulsion via a mast cell dependent process [6]. Recently we have observed a hypophagic response during the Th2 driven enteritis induced by T. spiralis infection [7]. However, the full mechanisms controlling hypophagia during enteritis and the precise effects reduced feeding have on immunity to intestinal infection require further elucidation. T. spiralis is experimentally highly attractive since the enteritis fully resolves, but is closely followed by a peripheral inflammatory phase characterized by skeletal muscle invasion and myositis as part of the parasite's life cycle.
Here, we demonstrate that this two step inflammatory process following T. spiralis infection is mirrored by a biphasic hypophagic response, and mediated by two separate adaptive immune driven mechanisms. We have characterized these two phases using genetically modified mice lacking functional CCK or adaptive immunity and demonstrated that CD4+ T-cells drive I-cell hyperplasia and the resulting CCK is an essential mediator of the initial hypophagia observed during enteritis. Conversely the second phase of hypophagia during skeletal peripheral myositis is CCK-independent but mediated by the anorectic actions of TNFα signaling. Furthermore, we demonstrate for the first time that this immune-EEC driven alteration in feeding also contributes a protective role during gastrointestinal infection. The hypophagia and resulting weight loss causes a reduction in fat secreted leptin, and the reduction in this hormone, which also acts as an inflammatory adipokine, augments the protective Th2 immune response aiding parasite expulsion. These results highlight the importance of the immunoendocrine axis in the gut during infection induced immunity and provide a biological function and associated mechanism for commonly associated infection induced weight loss. These data have wide-acting implications for the biology of gut infection and inflammation, and may inform new leptin-derived therapeutic strategies. Furthermore, the T. spiralis infected mouse presents a novel preclinical platform to study the biological mechanisms affecting food intake in inflammatory disorders, and has the unique potential to experimentally dissociate gastrointestinal from peripheral signals in an individual model.
Proximal enteritis induced by T. spiralis has been associated with a period of hypophagia and an increase in CCK and serotonin secreting EECs [7]. Here, mice were examined for alterations in feeding during T. spiralis induced inflammation. Interestingly, a biphasic response in feeding was observed following infection (Fig. 1A). Mice became hypophagic from days 6–10 post infection (p.i.), during the transient period of T. spiralis induced enteritis, and we observed a significant increase in CCK positive I-cells in wild-types (Fig. 1C and E) mirroring hypophagia at days 6 and 9 p.i. Feeding then returned to baseline levels until undergoing a second period of hypophagia from day 18–19 p.i. The secondary period of hypophagia occurs during the period of muscle invasion and peripheral myositis, caused when larvae form the “nurse cell” in which the parasite resides. To further investigate the biological mediators of the hypophagic responses the two phases were mechanistically explored using a panel of genetically modified mouse strains.
CCKlacZ mice, which do not express or secrete CCK peptide due to a knock in of a LacZ cassette [8], were infected with T. spiralis and their food intake monitored. Strikingly, the initial period of hypophagia was completely absent in infected CCKlacZ mice (Fig. 1B). LacZ positive I-cell hyperplasia was indistinguishable from that of the “natural” I-cell response in wild-type mice (Fig. 1D and E). However, the absence of CCK during this I-cell hyperplasia resulted in the complete absence of initial hypophagia in CCKlacZ mice, despite comparable enteritis (Fig. S1). Even with the absence of CCK and initial hypophagia, CCKlacZ mice still exhibited secondary hypophagia from day 18–19 p.i. and this was comparable to infected wild-type mice (Fig. 1B). The second period of hypophagia occurred in wild-type and CCKlacZ mice despite the resolution of enteritis (Fig. S1) and transpires during the period of larvae encysting within skeletal muscle, representing an extraintestinal inflammatory response to the same biological agent. The lack of I-cell hyperplasia in wild-type mice at day 20 p.i. (Fig. 1E) and the presence of the second phase of hypophagia in CCKlacZ mice (Fig. 1B) confirm this extraintestinal period of hypophagia during myositis as CCK independent.
Taken together, these data demonstrate a biphasic hypophagia correlating to T. spiralis induced enteritis and peripheral myositis, respectively. Furthermore, increased I-cell function, through the release of CCK, is essential for the initial hypophagia during enteritis, but not the secondary episode during peripheral myositis.
As EEC hyperplasia during inflammation has been previously linked to T-lymphocytes [7], [9], the biphasic hypophagia generated by T. spiralis was examined in severe combined immunodeficient (SCID) mice, which lack B and T-cells. SCID mice demonstrated a complete absence of initial hypophagia during T. spiralis induced enteritis and secondary hypophagia during peripheral myositis (Fig. 2A and B). The lack of hypophagia was mirrored by a complete lack of I-cell hyperplasia in parasitized SCID mice (Fig. 2C). This absence of hyperplasia was not seen in all epithelial secretory cells as, concurrent with previous findings [10], indistinguishable goblet cell hyperplasia occurred in infected SCID, adoptively transferred SCID and wild-type animals (Fig. 2D).
CD4+ T-cells play a key role in the resolution of T. spiralis infection [11], so to assess if CD4+ T-cells could restore I-cell hyperplasia and hypophagia in SCID mice, CD4+ T-cells (>90% purity; Fig. S2A), were adoptively transferred into SCID recipients before infection. Successful reconstitution was evident from CD4+ splenocytes present post-transfer and via successful worm expulsion kinetics (Fig. S2B and C).The adoptive transfer of CD4+ T-cells into SCID mice restored I-cell hyperplasia (Fig. 2C) and initial hypophagia during T. spiralis induced enteritis. Recipient mice began to eat less from day 4 p.i., with significant hypophagia at days 6 and 7 (Fig. 2E). This hypophagia was not a direct result of cell transfer alone, as uninfected reconstituted mice displayed no hypophagia (Fig. 2E). Interestingly, the adoptive transfer did not restore the secondary period of hypophagia during the peripheral inflammation induced by T. spiralis (Fig. 2A).
Collectively, these data confirm that the biphasic alterations in feeding behavior during T. spiralis induced gastrointestinal and peripheral inflammation is mediated by the adaptive immune system. Furthermore, CD4+ T-cells are identified as the key initiator in I-cell hyperplasia and resulting CCK driven hypophagia during T. spiralis induced enteritis. However, the adoptive transfer of functional CD4+ T-cells did not restore the second phase of hypophagia occurring during nurse cell formation-induced myositis. Therefore CD4+ T-cells are not sufficient for this secondary hypophagic period, during T. spiralis induced myositis.
We next sought to investigate which factors of the adaptive immune response were responsible for the second phase of hypophagia seen during peripheral inflammation induced during the period of nurse cell formation. Both CD4+ and CD8+ T-cells are present during parasite encystation [12] and many apoptotic factors, including TNFα, are detected during nurse cell formation [13]. Interestingly, TNFα is associated with cachexia in parasite infections [14], [15]. Consequently, we examined serum cytokine levels throughout T. spiralis infection. Indeed, TNFα was significantly increased in the serum of infected mice at the time of secondary hypophagia (Fig. 3A), comparable to levels known to directly cause cachexia in mouse infection models [16]. To test the function of increased TNFα during myositis we infected p55/p75−/− mice, which lack functional TNFα receptors, and assessed if TNFα was responsible for hypophagia during T. spiralis infection. Although initial I-cell hyperplasia and hypophagia during enteritis were present in infected p55/p75−/− mice (Fig. 3B, C and D), remarkably, infected p55/p75−/− mice displayed no period of secondary hypophagia (Fig. 3B and C). Therefore, although the initial CD4+ T-cell and CCK driven hypophagia during enteritis is independent of TNFα, a peripheral peak in TNFα during myositis is functionally responsible for the second phase of hypophagia, via the receptors p55 and/or p75, during T. spiralis induced peripheral inflammation.
Secretory cell hyperplasia during intestinal infection is known to be advantageous during infection. Various goblet and Paneth cell products have been show to have anti-parasitic affects [17], [18]. We therefore tested whether I-cell hyperplasia and hypophagia are simply by-products of a parallel switch towards this secretory lineage, or whether I-cell hyperplasia is in itself advantageous during infection. A link between hypophagia, weight loss and immunity is the adipokine leptin; produced mainly by adipose tissue it is a peripheral signal to the body of fat mass deposits but also acts as a pro-inflammatory Th1 cytokine [19]. Therefore reductions in leptin would be anticipated to occur following CCK induced hypophagia and consequent weight loss in this experimental model. Loss of leptin may consequently enhance Th2 immune responses which are protective during nematode infection.
CD4+ T-cell mediated I-cell driven hypophagia during enteritis was seen to result in significant weight loss at days 8 and 12 p.i., accompanied by a visible reduction in abdominal fat pads, whereas the brief TNFα driven secondary hypophagia produced no significant alteration in weight at day 20 p.i. (Fig. 4a). This weight loss was correlated with a reduction in serum leptin levels from day 6 p.i. (Fig. 4B). To determine whether alterations in leptin could influence a protective Th2 driven intestinal I immune response, mesenteric lymph node (mLN) cells were polarized towards a Th2 phenotype in the presence or absence of leptin. The addition of leptin resulted in a significant increase in the amount of intracellular pro-inflammatory IFN-y detectable in CD4+ T-cells, as well as a significant reduction in the protective Th2 cytokine IL-4 (Fig. 4C). To assess if the reduction in leptin during T. spiralis induced enteritis enhances immunity to infection, leptin levels were maintained at basal levels during hypophagia via recombinant leptin injection (Fig. 5A). Strikingly, the restoration of basal leptin levels resulted in delayed expulsion of adult worms and a corresponding increase in nurse cell encystation (Fig. 5B). Although no increase in IFN-y levels was seen in re-stimulated mLNs of leptin treated mice, a significant decrease in both Th2 cytokines IL-4 and IL-13 was seen at day 8 p.i (Fig. 5C). A key Th2 driven expulsion mechanism of T. spiralis is mastocytosis and this was seen to be significantly reduced upon the restoration of basal leptin levels analogous to delayed adult worm expulsion (Fig. 5D). Taken together this suggests that CD4+ T-cells drive a cascade in which I-cell hyperplasia produces hypophagia and weight loss, lowering pro-inflammatory leptin levels which feed back to influence the protective Th2 immune response, augmenting mastocytosis and allowing parasite expulsion.
During enteritis food intake is often significantly reduced, perhaps serving to inhibit consumption of contaminated food or to prevent further gut injury. EECs are implicated in this response: elevated I-cell produced CCK and hypophagia have been demonstrated during Ascaris suum and Trichostrongylus colubriformis infection, leading to the hypothesis that CCK was responsible for inflammation induced alterations in feeding [3], [4] However, the true biological function and molecular mechanisms that orchestrate the pathways driving hypophagia and weight loss during inflammation have not been addressed. The life cycle of T. spiralis has uniquely allowed us to investigate these questions during both intestinal and peripheral inflammation. Our data demonstrate that these separate inflammatory episodes are mirrored by a biphasic hypophagia driven by two independent immune mediated mechanisms. I-cell hyperplasia and CCK are essential for the initial hypophagia during enteritis, which is orchestrated by CD4+ T-cells. Importantly, we have also identified the second phase of hypophagia during the period of nurse cell formation to be mediated by a separate myositis-induced immune mechanism, fully dependent on the actions of TNFα.
Furthermore we show for the first time that immune driven weight loss during enteritis results in reduced levels of the Th1 adipokine leptin augmenting a protective Th2 response during infection. Intestinal inflammation is often associated with hypophagia and weight loss [1], [2] and we have now determined an important immune driven mechanism to explain why this is biologically functional. We have therefore identified a novel molecular pathway and can include I-cell hyperplasia and weight loss as an adaptively driven immune response which, through alterations in leptin, is beneficial during intestinal infection.
Our finding that T. spiralis infected mice lacking CCK (CCKlacZ) do not undergo initial hypophagia correlates with our own and other previous findings that a single treatment with the CCK1 receptor antagonist loxiglumide partially restores food intake in parasitized animals [7], [20]. This partial restoration now seems likely a result of the short half-life of loxiglumide as opposed to alternative satiety factors playing an essential role. The previous characterization of CCKlacZ mice demonstrated normal food intake, fat absorption and mass compared to wild-types [21], [22]. This coupled with our own observations rules out any underlying defect in feeding of CCKlacZ mice being responsible for the absence of hypophagia during infection. An alternative possibility is that the loss of CCK in brain neurons, rather than gut EECs, underpins the complete absence of hypophagia. However, the persistence of the second phase of hypophagia in CCKlacZ mice and our previous findings involving loxiglumide [7] which does not cross the blood brain barrier, make I-cell hyperplasia and CCK-vagal interactions most likely to mediate gut-induced hypophagia.
T. spiralis infected SCID mice were seen to display I-cell hyperplasia and hypophagia only upon reconstitution of CD4+ T-cells. These data clearly indicate that CCK induced satiety, via vagal afferent fibers signaling to feeding control centres in the brain [5], is an inherent pathway that is utilized by the adaptive immune system to bring about hypophagia and weight loss. This finding corresponds to recent studies showing that CD4+ T-cells restore 5-HT cell hyperplasia in Trichuris muris infected SCID mice [23]. The precise mechanism by which CD4+ T-cells cause EEC hyperplasia during infection remains to be elucidated. EECs have been shown to possess functional TLRs [24] and IL-13 receptors are present on 5-HT cells [9]. However, it has previously been established that during T. spiralis infection in SCID mice NK cells produce ample levels of IL-13 to induce goblet cell hyperplasia [10] yet this IL-13 appears not sufficient to cause EEC hyperplasia. We also detected mRNA for both TNFα receptors p55 and p75 on EECs, yet mice genetically deficient in TNF receptor signaling demonstrated initial hypophagia and enteritis (Fig. S3A–C) arose independently of its actions.
As all gut epithelial subtypes are derived from pluripotent crypt stem cells [25], immune cell mediators may alter the transcription factors at the stem cell level leading to altered EEC hyperplasia during infection. Alternatively, analogous alterations in the post-stem cell, neurogenin+, EEC specific progenitor cells could alter EEC proliferation. BrdU labeling studies in experimentally-induced inflammation demonstrated EEC hyperplasia does occur at the level of the stem/progenitor cell rather than fully differentiated epithelial cells [26]. Indeed, the uncoupling of goblet and I cell hyperplasia seen here in infected SCID mice supports the hypothesis that neurogenin+ EEC precursors, rather than the stem cell itself, is targeted by an unknown CD4+ dependent mechanism during T. spiralis driven I-cell hyperplasia. Identifying which factors drive this CD4+ T-cell-stem/progenitor cell interaction is an exciting area for further study.
We have identified a novel and specific role for TNFα in hypophagia during T. spiralis induced peripheral inflammation. The absence of the second phase of hypophagia from p55/p75 −/− mice suggests that the drop in food intake during this extra-enteric inflammatory period is due to the anorexic effects of systemic TNFα or downstream targets. The significant peak of serum TNFα seen in mice correlated with the second period of hypophagia, occurring during nurse cell development and myositis, and strongly supports this notion. Indeed, we observed serum levels comparable to levels known to directly cause cachexia in mouse infection models [16]. Furthermore, TNFα is associated with cachexia in Trypanosoma cruzi infection [14] and during schistosomiasis [15]. There are numerous modes of action by which TNFα could cause anorexia [27]. Although central TNFα levels were not directly monitored, the 20 pg/ml serum TNFα measured during secondary hypophagia is below the levels required to induce anorexia by central administration [28]. It is therefore most likely that TNFα is acting on peripheral afferent nerves, as low level localized cytokine production can trigger afferent nerves without causing an increase in circulating cytokine levels [29]. It is also possible that myalgia and malaise may have contributed to reduced food intake: appetite per se cannot be measured in mice.
The observed systemic peak in TNFα occurs during the period of encystation of T. spiralis new born larvae in cells of the striated muscle. Encystations are likely to arise from day 4–10 post-infection with rapid growth of the parasite occurring over the following 20 days as terminally differentiated muscle cells re-enter the cell cycle and establish a niche for the parasite [12]. Early non-significant increases in systemic TNFα were seen as early as day 8 post-infection; day 18–21 post-infection may therefore indicate a “tipping point” in peripheral TNFα levels, where significant myositis breaches the threshold required to produce anorexia. TNFα has been shown to be involved in nurse cell formation [13], yet we observed no alteration in nurse cell development in p55/p75−/− mice that could alternatively explain the observations seen (Fig. S3C–E). The cellular source of TNFα remains to be elucidated. CD4+ and CD8+ T-cells are reported to be present during parasite encystation [12], as are macrophages interestingly peaking during hypophagia [30]. However, as we illustrate here, given the absence of secondary hypophagia in SCID mice reconstituted with CD4+ T-cells, where macrophages are present, the likelihood is that CD8+ T-cells may be the source of cachectic TNFα. Further studies are therefore required to ascertain the cellular source of TNFα which drives the hypophagia during T. spiralis induced myositis.
Immune mediated secretory cell hyperplasia during intestinal infection is advantageous as goblet and Paneth cell products have been show to have anti-parasitic affects [17], [18]. We therefore postulated whether I-cell hyperplasia and hypophagia are simply by-products of a parallel switch towards these beneficial secretory lineages or whether I-cell hyperplasia is in itself advantageous in nematode expulsion. Stimulation of the vagus nerve via nutritional release of CCK has also been shown to protect against hemorrhagic shock [31]. Therefore I-cell hyperplasia during nematode infection may represent a previously unidentified anti-inflammatory response. We therefore hypothesized that a reduction in weight as a result of I-cell induced hypophagia may alter the levels of the Th1 adipokine leptin [19]. A reduction in leptin could enhance the protective Th2 immune response to nematode infection. Indeed significant weight loss and reduced leptin levels did occur during T. spiralis induced hypophagia. Recent data on splenocytes demonstrated that leptin alters polarized CD4+ T-cells towards a Th1 phenotype via alterations in proliferation in vitro [32] and we demonstrated parallel results in mLN cells for the first time. Unfortunately CCKlacZ mice have overall reduced basal levels of leptin [33] and were hence unsuitable to study the affect of reduced leptin on intestinal inflammation. We therefore maintained basal leptin levels in infected hypophagic mice and strikingly saw a significant reduction in Th2 cytokines and mastocytosis culminating in delayed worm expulsion. Interestingly mastocytosis was similar in both leptin reconstituted and wild-type mice at day 8 p.i. demonstrating that initially mastocytosis can establish, but without the I-cell driven reduction in Th1 polarizing leptin it is blunted later in infection. These results complement other recent studies in identifying the adipokine leptin as a molecule which can greatly influence the response to infection. Mice lacking the leptin receptor are highly susceptible to infection from protozoa [34], pneumonia [35] and Listeria [36] demonstrating how malnutrition can compromise Th1 driven immunity. However, our data demonstrate that brief alterations in leptin can benefit immunity in terms of Th2 driven resistance to infection. Indeed, a recent study has demonstrated that leptin receptor deficient mice are resistant to experimentally induced Th2-mediated colitis [32]. The precise action of leptin in our studies may be as a direct result of effects on CD4+ T-cell IL-4 production altering mast cell differentiation, proliferation and migration [37] or due to direct effects on mast cells which have recently been shown to express leptin receptors [38]. Leptin may also directly act on Th2 cytokine production itself as opposed to indirect alterations on Th1 cytokine production [32]. Further study is therefore required to address the leptin-mast cell axis which alters parasite expulsion in our model.
In conclusion, we have identified two separate immune mediated mechanisms of hypophagia during infection induced gastrointestinal and peripheral inflammation, which act via the distinct pathways of I-cell hyperplasia and TNFα cachexia. Furthermore, we demonstrate for the first time an immunoendocrine feedback loop, in which CD4+ T-cell driven weight loss via CCK reduces leptin levels which impinge on CD4+ T-cell driven effector mechanisms for gastrointestinal infection resolution. Our data elucidate inflammation and weight loss, not just as commonly associated phenomena, but highlights them as a novel immune driven mechanism in parasite expulsion. These data offer potential specific treatment targets to modulate feeding and immune function during inflammatory diseases of the intestine.
Mice were housed in specific pathogen free conditions and experiments were carried out in accordance with the United Kingdom Home Office Scientific Procedures Act (1986) under Department for Environment, Food and Rural Affairs license.
Male C57BL/6 and BALB/c mice were obtained from Harlan-Olac Ltd. CCKlacZ mice have a LacZ cassette knocked into the CCK locus on a C57BL/6 background, so homozygote animals are CCK null but faithfully express LacZ in the I cell population [8]. TNFα receptor null p55/p75 −/− (C57BL/6 background) and severe combined immunodeficient mice (SCID, BALB/c background) were generated as previously described [8], [39].
The maintenance, infection and recovery of T. spiralis were carried out as previously described [40].
Mice were individually weighed on a daily basis. Food intake per mouse was derived by weighing the chow (B and K, Hull, UK) daily.
Proximal small intestine was fixed and stained and I-cells were enumerated using, CCK specific, L421 anti-proCCK as previously described [7]. For CCKlacZ detection, transverse 12 µm sections of tissue were cut and fixed in 0.2% glutaraldehyde and stained with X-gal as previously described [8]. Mast or goblet cell sections were stained in toludine blue or Schiff's reagent, respectively. After mounting, positive cells were enumerated in 20 randomly selected villus crypt units (VCU) and results presented as mean number of positive cells/20 VCU (± s.e.).
Mesenteric lymph node (mLN) cells were prepared from day 7 p.i. BALB/c mice, in RPMI-1640, supplemented with 10% fetal calf serum, 100 µg/ml penicillin/streptomycin and 1 mM L-glutamine (complete media). CD4+ T-cells were isolated via negative selection using an isolation kit (Miltenyi Biotec). Evaluation of CD4+ purity was via flow cytometry. SCID mice received 4×106 cells in P.B.S. via intraperitoneal (i.p.) injection 2 days before infection.
mLN cells at 5×106 cells/ml in complete media received 50 µg/ml of T. spiralis antigen (Ag). Supernatants were collected after 24 hrs and cytokines measured using a cytometric bead array kit (BD).
Serum was obtained from blood at the time of sacrifice via centrifugation at 15000×g and cytokines measured using a cytometric bead array kit (BD).
Mouse leptin ELISA (Linco) was used to detect mouse serum leptin according to manufacturer's instructions.
During the period of significant hypophagia, mice were treated at 10 a.m. and 6 p.m. via an i.p. injection of recombinant leptin (R and D) at 0.5 µg/g of initial body weight or control vehicle PBS [41].
2×106/ml mLN cells were stimulated via 5 µg/ml αCD28, 3 µg/ml αCD3 (BD) and polarized via 50 ng/ml IL-4 (Peprotech), 50 µg/ml anti-IFN-γ with/without 500 ng/ml recombinant leptin. At 120 hrs 1 µg/ml Brefeldin A/1 µg/ml monensin (Sigma-Aldrich) for IFN-y/IL-4 staining was added for 4 hrs before blocking with anti-FcγR (BD). Cells were stained for CD4 (BD) for 30 mins at 4°C before fixing in FACS fix buffer (1% formaldehyde, 0.1% BSA and 0.05% NaN3 in PBS). Cells were permeabilised in 0.1% saponin (Sigma-Aldrich) and stained with biotinylated anti-IFN-γ/anti-IL-4 (BD) for 25 mins at RT. Controls were stained with isotype controls (BD). Biotinylated antibodies were detected by streptavidin APC conjugate (Caltag) at 1/200 in saponin for 25 minutes RT. Cells were analyzed on a FACScalibur using Flowjo.
Two experimental groups were compared using Student's t-test. Three or more groups were compared using the Kruskal-Wallis test, Dunn's multiple comparison post-test. A p value of ≤0.05 was considered statistically significant. *, P<.05; **, P<.01; or ***, P<.005 for indicated comparisons, error bars represent SE of means.
|
10.1371/journal.ppat.1004645 | Disruption of an M. tuberculosis Membrane Protein Causes a Magnesium-dependent Cell Division Defect and Failure to Persist in Mice | The identification of Mycobacterium tuberculosis genes necessary for persistence in vivo provides insight into bacterial biology as well as host defense strategies. We show that disruption of M. tuberculosis membrane protein PerM (Rv0955) resulted in an IFN-γ-dependent persistence defect in chronic mouse infection despite the mutant’s near normal growth during acute infection. The perM mutant required increased magnesium for replication and survival; incubation in low magnesium media resulted in cell elongation and lysis. Transcriptome analysis of the perM mutant grown in reduced magnesium revealed upregulation of cell division and cell wall biosynthesis genes, and live cell imaging showed PerM accumulation at the division septa in M. smegmatis. The mutant was acutely sensitive to β-lactam antibiotics, including specific inhibitors of cell division-associated peptidoglycan transpeptidase FtsI. Together, these data implicate PerM as a novel player in mycobacterial cell division and pathogenesis, and are consistent with the hypothesis that immune activation deprives M. tuberculosis of magnesium.
| The success of Mycobacterium tuberculosis (Mtb) as a human pathogen is due to ability to persist in chronic infection, despite a robust adaptive immune response by the host. The mechanisms by which Mtb achieves this are, however, poorly understood. Here we show that a novel integral membrane protein, Rv0955/PerM, is essential for Mtb persistence during chronic mouse infection. The perM mutant required increased magnesium compared to wild type Mtb for replication and survival in culture and elongated in media with reduced magnesium concentration. Transcriptomic, electron microscopy and live cell imaging approaches provided evidence that PerM is involved in cell division. The survival defects of the perM mutant in reduced magnesium and during chronic mouse infection are consistent with the hypothesis that magnesium deprivation constitutes an IFN-γ dependent host defense strategy. This work also has potential clinical implications, as disruption of PerM renders Mtb susceptible to β-lactam antibiotics, which are commonly used to treat non-mycobacterial infections.
| With an estimated one-third of the world’s population latently infected with Mycobacterium tuberculosis (Mtb), the question remains: how is this pathogen able to persist in vivo? In the mouse model, Mtb infection is characterized by an acute phase of logarithmic bacterial growth lasting approximately three weeks, followed by a plateau in bacterial burden, persisting as a chronic infection. The transition from acute to chronic infection—from logarithmic bacterial growth to stable bacterial counts—results from the onset of the adaptive immune response and activation of host macrophages by CD4+ T cell-derived IFN-γ [1,2]. IFN-γ enhances the antimicrobial capacity of macrophages by numerous mechanisms including promotion of phagosome maturation and acidification via induction of the GTPase Irgm1 and production of reactive nitrogen and oxygen species mediated by nitric oxide synthase and phagocyte oxidase [3–6]. However, IFN-γ induces hundreds of genes in macrophages [7] and the array of environmental modifications occurring within these macrophages and leading to control of Mtb growth is not entirely understood.
Mtb persistence mutants (per mutants) are a unique class of strains that are competent for replication during acute infection, but attenuated during chronic infection [8]. Several previously identified per mutants provide information about the processes required for survival in the activated macrophage following the onset of adaptive immunity. For example, a per phenotype was observed for an Mtb mutant lacking isocitrate lyase-1, an enzyme involved in the glyoxylate shunt and methylcitrate cycle, as well as a mutant lacking the cholesterol transporter Mce4, indicating that cholesterol and fatty acids are carbon sources required by Mtb to survive during chronic infection [9,10].
Macrophage activation promotes phagosomal maturation and intraphagosomal acidification [6,11,12]. In a screen for Mtb transposon mutants hypersusceptible to acid stress, we previously identified 21 genes whose interruption lead to reduced viability in low pH [13]. The majority of these genes are annotated to have functions related to cell wall processes. These included two independent transposon mutants of the previously uncharacterized Mtb gene rv0955, a 1,368 base pair open reading frame, which is annotated to encode an integral membrane protein with a predicted topology of ten transmembrane helices (S1 Fig.) [14–16]. Rv0955 is highly conserved among mycobacteria and actinobacteria, but has no known homologues in other species, and no conserved sequence motifs to predict its function. It is included among the 219 mycobacterial “core” genes noteworthy for their conservation among mycobacterial species, including Mtb and M. leprae [17]. These core genes lack homologues in other bacteria, suggesting that their function may be unique to mycobacteria, and making them potential targets for mycobacteria-specific drugs.
Here, we investigated the function of the previously uncharacterized Mtb Rv0955 protein. Disruption of rv0955 resulted in a striking persistence defect in chronic mouse infection with a 300-fold decline in bacterial burden in the lungs. We therefore named this gene perM, encoding a persistence-associated integral membrane protein. As Vandal et al. noted, the acid susceptibility of the perM mutant—similar to many of the mutants identified in the screen—was detergent-dependent, observed only when the bacteria were exposed to a combination of low pH and Tween-80 detergent [13]. We thus sought to investigate mechanisms beyond protection from acid, which might account for the strong attenuation of the mutant in vivo.
We found that the perM mutant required increased magnesium (Mg2+) compared to wild type (wt) Mtb for replication and survival in culture. Mg2+ is among the most abundant divalent cations in both prokaryotic and eukaryotic cells, and is essential for bacterial growth. In bacteria, Mg2+ serves a wide range of roles: it functions as a cofactor with ATP in numerous enzymatic reactions, enables the formation of tRNA and ribosomal tertiary structure, and regulates stability of the cell wall and membrane [18–20]. Mg2+ also impacts virulence in Salmonella enterica by regulating the PhoP/PhoQ two-component system [21].
In Mtb, two Mg2+-dependent mutants have been identified: Mtb∆phoP and Mtb∆mgtC [22,23]. PhoP shows high similarity to the PhoP response regulator of Salmonella enterica and is required in Mtb for the synthesis of several complex cell wall lipids as well as replication in macrophages and mice [22,24,25]. MgtC is required for virulence of both Mtb and Salmonella enterica and inhibits the bacterial F1F0 ATP synthase to maintain physiological ATP levels and intrabacterial pH [23,26].
Mg2+ restriction remains a plausible but unconfirmed antimycobacterial mechanism employed by the host. In media with low Mg2+ concentrations, the perM mutant elongated and upregulated expression of cell division and cell wall biosynthesis genes. Furthermore, Mtb PerM accumulated at the putative division septa in the closely related M. smegmatis. Disruption of perM resulted in pronounced hypersusceptibility to beta-lactam antibiotics, including cephalexin and piperacillin, which are specific inhibitors of the cell division-associated peptidoglycan synthesis protein FtsI. This work characterizes a novel mycobacterial protein necessary for persistence in vivo and implicated in cell division, and is consistent with the hypothesis that Mtb has reduced access to Mg2+ during chronic infection.
PerM was previously identified in a screen for Mtb genes required for acid resistance [13]. To examine the role of PerM in vivo, we monitored replication and survival of a perM transposon mutant, perM::tn, in wild type mice. PerM::tn established infection and replicated during the acute phase, with only a 5-fold reduction in peak bacterial burden, measured by colony forming units (CFU), compared to wt (P = 0.032) at 21 days (Fig. 1A). However, perM::tn exhibited a severe persistence defect in chronic infection, with a 300-fold reduction in CFU in the lungs at fourteen weeks post-infection.
In agreement with these growth patterns, histological analysis revealed markedly fewer and smaller lesions in perM::tn-infected lung tissue compared to wt Mtb-infected mice, a difference observed at the 148 day post-infection time point, but not at the end of acute infection (Fig. 1B). Genetic complementation of the perM mutant with a wild-type copy of the gene expressed chromosomally under control of the hsp60 promoter restored persistence and increased granulomatous inflammation, indicating that attenuation in vivo was due to disruption of perM.
The adaptive immune response to Mtb is characterized by IFN-γ mediated activation of host macrophages. To examine whether death of perM::tn in vivo was dependent on host IFN-γ, we infected IFN-γ knockout mice, which are unable to control replication of wt Mtb [1,2]. PerM::tn replicated in IFN-γ knockout mice, but at a slower rate than wt Mtb (Fig. 1C). IFN-γ knockout mice infected with wt Mtb had to be sacrificed at day 50, because they were moribund, in contrast to IFN-γ knockout mice infected with perM::tn, which remained healthy through the end of the experiment (day 106). These results indicate that killing of perM::tn in wt mice requires host IFN-γ, while the mutant also exhibits an IFN-γ-independent replication defect.
Since later-occurring persistence defects like that of perM::tn often depend on the adaptive immune response of the host, we hypothesized that PerM might cause a more robust immune response than wt Mtb. To examine this possibility, we infected bone marrow derived mouse macrophages with equal numbers of wt, perM::tn and complemented mutant and measured cytokine concentrations in macrophage culture supernatants 24 hours later. Supernatants of macrophages infected with perM::tn contained elevated levels of proinflammatory cytokines, including TNF-α, IL-6, IL-12 p70, the anti-inflammatory cytokine IL-10, and the chemokine KC (Fig. 2A).
To assess the immune response to perM::tn during mouse infection we measured cytokine transcripts in mouse lungs by quantitative real-time polymerase chain reaction (qRTPCR), focusing on pro-inflammatory cytokines required to attenuate Mtb growth in vivo [1,27]. We did not observe significant differences in IFN-γ or TNF-α mRNA levels at 2 weeks post-infection, when bacterial titers of perM::tn were 3-fold lower than those of wt (S2A–S2B Fig.). In an independent experiment, IL-12p40 protein levels in lung homogenates from mice infected with wt or perM::tn were similar at 1 or 2 weeks post-infection and increased in wt compared to mutant infected lungs or 3 weeks post-infection (S2C–S2D Fig.). Differences in CFU confounded interpretation of these data, as bacterial counts of perM::tn were 3-, 5- and 6-fold reduced compared to wt at weeks 1, 2 and 3, respectively; however, these results suggested that attenuation of the mutant in vivo was not exclusively due to a more robust immune response that preceded in perM::tn-infected mice.
We sought to better understand the properties of perM::tn leading to the increased innate immune response by macrophages infected ex vivo. In a mixed-strain Mtb infection of macrophages, TNF-α production was similar to that induced by infection with perM::tn alone at the same total multiplicity of infection (MOI) (Fig. 2B). The dominance of the mutant suggested that the difference in response to these strains was due to an immunostimulatory effect of the mutant, as opposed to a suppressive effect of intact PerM protein produced by wt Mtb. The stimulatory effect of perM::tn was reproduced by exposure of macrophages to formalin-killed Mtb (Fig. 2C) and to cell-free Mtb-conditioned culture media (Fig. 2D), indicating that the stimulatory component(s) were shed or secreted by live perM::tn, but did not require viable bacteria for production or release during macrophage infection. In the absence of NOD and TLR2 signaling, perM::tn still elicited higher levels of TNF-α than wt (Fig. 2E). NOD and TLR2 are required for the macrophage response to bacterial peptidoglycan and triacylated lipoproteins, respectively, suggesting that the hyperinflammatory phenotype of perM::tn is not tied specifically to one of these cell wall components. TNF-α production was, however, significantly lower in cultures from knockout macrophages compared to wt macrophages, indicating that these receptors are important for TNF-α production following infection with both strains. Together, these data suggest that a combination of cellular components, both released into the medium during growth and expressed on the surface of killed perM::tn, function to stimulate increased inflammatory signaling in macrophages.
In liquid media, perM::tn replicated at a near-normal rate (Fig. 3A), but formed a loose aggregate during growth (Fig. 3B). Unlike previously described mycobacterial biofilms [28], these aggregates formed on the bottom of standing cultures, rather than at the liquid-air interface, and could be readily dispersed by shaking or pipetting. These aggregates suggested a perturbation of the perM::tn cell envelope. Since extracellular magnesium (Mg2+) has been shown to overcome phenotypes of mutants with cell envelope defects [22], we asked whether reduction of Mg2+ would affect growth or survival of perM::tn. Strains were cultured in nominally Mg2+-free Sauton’s minimal media, and supplemented with Mg2+ at a range of concentrations up to 2000 μM, the normal concentration in Sauton’s media (Fig. 3C). Wt Mtb died in nominally Mg2+-free media, but survived and replicated at Mg2+ concentrations of 25 μM and higher. In contrast, perM::tn exhibited death, observed by decreasing CFU counts, and lysis, observed by decreasing absorbance, at Mg2+ concentrations 100 μM and below. At 250 and 500 μM Mg2+, perM::tn replicated, but at a slower rate than wt and the complemented mutant.
The requirement for additional Mg2+ was specific, as other cations, including Mn2+, Ca2+, Zn2+, and Fe3+, could not restore growth of perM::tn in reduced (100 or 250 μM) Mg2+ media (S3 Fig.). For further experiments, Mtb was grown in modified Sauton’s media containing 250 or 500 μM Mg2+ (“reduced” Mg2+), concentrations at which perM::tn displayed a growth defect without apparent death or lysis, or 2000 μM (“high”) Mg2+.
Within the IFN-γ-activated macrophage, Mtb is subject to numerous stresses, including low pH, reactive nitrogen intermediates, reactive oxygen species, and nutrient limitation [29,30], and it has been postulated that Mg2+ restriction may be an additional stress encountered by intraphagosomal pathogens including Mtb [23,31,32]. The inability of perM::tn to replicate and survive in low Mg2+ raised the possibility that the persistence defect in vivo might follow depletion of intraphagosomal Mg2+ in activated macrophages. We infected resting and IFN-γ-activated macrophages with wt, perM::tn and the complemented mutant following growth in high (2 mM) and reduced (250 μM) Mg2+. The mutant displayed a growth defect in resting macrophages, which was larger when it was pre-cultured in reduced magnesium. Survival of perM::tn in IFN-γ-activated macrophages was impaired in comparison to wt and complemented mutant, but only following pre-culture in reduced magnesium (S4 Fig.). We examined whether perM::tn was more susceptible than wt Mtb to a range of stresses in vitro, including exposure to hydrogen peroxide, lysozyme, detergent, acidified sodium nitrite, free fatty acid, zinc, and copper, as well as carbon starvation, iron depletion, and a multi-stress assay combining a fatty acid carbon source, reduced pH, hypoxia, and sodium nitrite (S5 Fig.). The mutant survived at wt levels under all conditions, indicating that perM::tn does not have a general viability defect, but rather, appears to be specifically vulnerable to reduced Mg2+.
The increased Mg2+ requirement of perM::tn suggested a possible role for PerM in Mg2+ transport. However, analysis of total Mg2+ content by inductively coupled mass spectrometry (ICP-QQQ) showed no significant differences in strains grown in either high (2000 μM) or reduced (250 and 25 μM) Mg2+ (S6 Fig.), suggesting that PerM is not required for Mg2+ acquisition. Given our data, along with work in Salmonella suggesting a redundancy of Mg2+ transporters that ensures significant Mg2+ uptake [33], the persistent defect of perM::tn is unlikely the result of impaired Mg2+ transport.
In the context of the host response, infection of mouse macrophages with Mtb pre-grown in reduced (500 μM) Mg2+ media resulted in a 2-fold increase in TNF-α production by macrophages infected with perM::tn, but not wt or complemented strains (Fig. 3D). This suggests that either the immunostimulatory component(s) of perM::tn are more highly produced, secreted, or shed in reduced Mg2+; or that Mtb growth in reduced Mg2+ leads to increased exposure of these components to macrophage pattern recognition receptors and induction of a proinflammatory response.
To gain insight into the function of PerM, we compared the transcriptomes of wt and perM::tn Mtb grown in high (2000 μM) and reduced (250 μM) Mg2+ in three independent experiments, using a p-value of 0.05 and 2-fold cutoff to identify differentially regulated genes. In reduced Mg2+, 41 genes were differentially expressed between strains, all of which except one were upregulated in the mutant (Table 1). Sixteen of these genes are annotated with predicted or possible roles in cell division and/or cell wall biosynthesis. Upregulation of a subset of these genes was confirmed by qRTPCR analysis (Fig. 4A).
Genes listed were differentially expressed at least 2-fold in perM::tn compared to wt grown for 5 days in media supplemented with 250 μM Mg2+. Fold change values are averages of three independent experiments, P<0.05. Annotations adapted from TB Database (tbdb.org), TubercuList (tuberculist.epfl.ch) and PATRIC (patricbrc.org). FC, fold change in perM::tn compared to wt. Genes also regulated greater than 2-fold between strains in 2000 μM Mg2+ are marked with *.
Cell division genes more highly expressed in the mutant compared to wt under reduced Mg2+ included ftsK and xerC, involved in chromosome segregation; ftsI, necessary for peptidoglycan crosslinking during division; and ftsW, whose product likely translocates peptidoglycan precursors across the cell membrane [34] and interacts with both FtsI as well as cell division initiator FtsZ in mycobacteria [35]. Also upregulated in the mutant were genes encoding four putative penicillin binding proteins (FtsI, DacB1, Rv2864c, and Rv1433), enzymes which carry out the transpeptidation necessary for crosslinking of cell wall peptidoglycan strands; Rv3717, a possible peptidoglycan amidase with a role in cell wall remodeling; and Rv0519c, a possible mycolyltransferase involved in mycolic acid processing [36]. Secreted fibronectin-binding protein C (FbpC), a possible trehalose mycolyltransferase thought to have both antigenic and cell wall biosynthesis roles, also showed increased expression in the mutant. Notably, expression of ftsZ, encoding the cytosolic, tubulin-like initiator of cell division, was not increased in the mutant at either Mg2+ concentration, nor was expression of genes in the cell wall biosynthetic gene cluster (rv3779-rv3809c) contributing to mycolic acid, arabinogalactan, and LAM synthesis [37], pointing towards a specific response rather than a global induction of all cell division and cell wall biosynthesis genes in the mutant.
Seven genes were upregulated in the mutant compared to wt in both high and reduced Mg2+, with more pronounced differences in expression between strains in reduced Mg2+ (Tables 1, S1), suggesting that Mg2+ reduction exacerbates differential transcriptional responses that are also present in high Mg2+.
Comparison of gene expression in wt Mtb in reduced versus high Mg2+ revealed only two genes meeting the 2-fold cutoff: pe20 was upregulated in reduced Mg2+ and fadD5 was downregulated (S2 Table). This transcriptional response was far less pronounced than that previously identified consisting of 24 genes differentially regulated in wt Mtb grown in media with or without Mg2+ [22], suggesting that the transcriptional response to Mg2+ starvation in wt Mtb was not triggered at 250 μM Mg2+, used in our experiment. The gene expression pattern of perM::tn in 250 μM Mg2+ (S2 Table) did not resemble Mg2+-starved wt Mtb [22], contrary to what might be expected if Mg2+ uptake were impaired in the mutant.
The increased expression of cell division and cell wall biosynthesis genes in the mutant suggested a possible defect in these processes. To examine the impact of perM disruption on cell morphology, Mtb was grown in a range of Mg2+ concentrations, fixed, and imaged by scanning electron microscopy (SEM). PerM::tn exhibited Mg2+-dependent defects in morphology and division. Median cell length increased as the concentration of Mg2+ decreased, and some mutant bacilli exhibited bulging at the poles in reduced Mg2+ (Fig. 4B,C).
To examine localization of PerM, GFP-tagged Mtb PerM protein was expressed in wt Mycobacterium smegmatis, a non-pathogenic species closely related to Mtb and itself containing an PerM homolog with 73% identity. Mtb requires containment within a biosafety level 3 facility, which prevented us from performing live cell imaging experiments in Mtb. Live cell imaging of recombinant M. smegmatis revealed that PerMMtb localized to the membrane and it accumulated at the mid-cell division site (Fig. 5), similar to mycobacterial cell division proteins, such as FtsI and FtsZ, as well as peptidoglycan synthesis enzymes, such as penicillin binding protein 1 [38–40].
We next compared sensitivity of perM::tn and wt Mtb to a range of compounds targeting cell wall biosynthesis, as well as drugs with other established targets. The majority of compounds assayed exhibited a similar minimum inhibitory concentration (MIC) in wt and perM::tn (Table 2 and S7 Fig.), with a shift of 2-fold or less considered insignificant. However, perM::tn was acutely sensitive to growth inhibition by β-lactam antibiotics, which target penicillin binding proteins that carry out the transpeptidation reaction resulting in crosslinking of cell wall peptidoglycan, a final step in peptidoglycan synthesis. The shift in MIC was most pronounced for cephalexin and piperacillin, β-lactams that specifically inhibit FtsI, the transpeptidase required for peptidoglycan crosslinking during bacterial cell division [41–43]. β-lactamase activities in wt and perM::tn were not significantly different (S3 Table) excluding the possibility that impaired β-lactamase activity caused the mutant’s increased susceptibility to β-lactams. Notably, the MICs of vancomycin and D-cycloserine, which inhibit earlier steps in peptidoglycan synthesis than do β-lactams, were similar for wt and perM::tn. Furthermore, there was little to no shift in MIC of isoniazid and ethambutol, which inhibit production of other cell wall components (mycolic acids and arabinogalactan, respectively), indicating that the perM::tn is not broadly hypersusceptible to interference with cell wall biosynthesis.
Minimum inhibitory concentration (MIC) of various drugs against wt and perM::tn Mtb. MIC90 values in μg/mL, determined by minimum concentration at which OD580 was less than 10% that of untreated control. FC, fold change reduction of perM::tn MIC compared to wt MIC.
This work implicates a novel mycobacterial membrane protein in cell division and demonstrates its requirement for Mtb persistence in vivo. The persistence defect of the PerM mutant is one of the most dramatic per phenotypes observed to date, and to our knowledge the first noted in a mutant of an Mtb membrane protein. Global gene expression profiling revealed increased expression of cell division and cell wall biosynthesis genes in the mutant, and these increases exacerbated during growth in reduced Mg2+. Several additional observations support the hypothesis that PerM plays a role in cell division. First, the mutant elongated in reduced Mg2+, with additional morphological changes at very low Mg2+. Second, the mutant exhibited hypersusceptibility to β-lactam antibiotics, which inhibit the enzymes necessary for crosslinking of cell wall peptidoglycan. In particular, the mutant was hypersusceptible to piperacillin and cephalexin, β-lactams that specifically target the cell division-associated peptidoglycan transpeptidase, FtsI [42–44]. Third, PerM localized to the mid-cell region in M. smegmatis, similarly to previously studied mycobacterial proteins involved in cell division and peptidoglycan biosynthesis [38–40]. The mutant hyperstimulated mouse macrophages ex vivo, a phenotype exacerbated after culture in reduced Mg2+, which may be related to shedding of cell wall components during a compromised cell division process.
The inability of perM::tn to replicate and survive at low Mg2+ suggested that PerM may play a role in Mg2+ acquisition, could be necessary for the adaptive response of Mtb to low Mg2+, or that Mg2+ might serve a compensatory function to mask physiological defects caused by the absence of PerM. We examined the first possibility by ICP-QQQ analysis, which revealed perM::tn and wt Mtb to contain the same total Mg2+, even when grown in reduced Mg2+ media. Furthermore, gene expression data from the mutant showed a regulation pattern distinct from that of Mg2+-starved wt Mtb [22]. The second possibility, that PerM is a component of the bacterial response to low Mg2+, is similarly not supported by the gene expression profile of wt Mtb grown in low Mg2+ [22]. However, it is possible that PerM, constitutively expressed, is required for a successful adaptive response to Mg2+ starvation, perhaps through interaction with Mg2+ response proteins. Future protein interaction studies may shed light on this question.
Our work supports the third possibility, that Mg2+ serves a compensatory function in the mutant through stabilization of a weakened cell envelope; in particular, our data suggest that the mutant cell envelope may be especially vulnerable during cell division. While the role of Mg2+ in cell wall stability is widely acknowledged, the mechanism by which this occurs is not entirely clear. In Salmonella, outer membrane permeability decreased in high Mg2+, and a phoP Salmonella mutant with lipopolysaccharide alterations displayed increased permeability and susceptibility to numerous antibiotics in low Mg2+, but behaved like wt Salmonella when Mg2+ was high [45]. This suggests a role for Mg2+ in stabilizing the outer membrane, perhaps through interaction with negatively-charged lipopolysaccharide. In the Gram-positive B. subtilis, which lacks an outer membrane, high Mg2+ partially suppressed the growth defect of a mutant lacking teichoic acid suggesting that Mg2+ might be able to compensate for loss of teichoic acid in the cell wall [46]. On the other hand, high concentrations of Mg2+ may serve to stabilize an otherwise vulnerable peptidoglycan sacculus. B. subtilis mutants lacking MreB, RodB and PonA—proteins thought to be involved in peptidoglycan synthesis—display morphological and growth defects that were rescued by high Mg2+ [47–49]. Of note, peptidoglycan synthesis decreased and peptidoglycan precursors accumulated in Mg2+-deprived B. subtilis [50], and in Salmonella, lipid A acylation increased in response to Mg2+ deprivation [20], suggesting that the influence of Mg2+ on peptidoglycan integrity may occur by several mechanisms, both structural and regulatory. It has also been proposed that Mg2+ might affect the degree of peptidoglycan crosslinking that occurs; stabilize or regulate important cell-wall synthases or hydrolases; or serve to stiffen the cell envelope [45,51].
The upregulation of cell division genes in perM::tn, combined with the hypersensitivity of the mutant to specific inhibitors of FtsI, suggests a role for PerM in peptidoglycan synthesis or remodeling during cell division. The perM mutant was not hypersusceptible to all peptidoglycan synthesis inhibitors: the MICs of vancomycin and cycloserine were similar for mutant and wt Mtb. Cycloserine, an analog of D-alanine, blocks synthesis of cytoplasmic peptidoglycan precursors [52], while vancomycin prevents both the early transglycosylation step necessary for incorporation of peptidoglycan monomer into the sacculus, as well as the final crosslinking of peptidoglycan by transpeptidases [53]. The specific vulnerability of perM::tn to β-lactams, which target the transpeptidation step, suggests that PerM may play a role in late peptidoglycan biosynthesis during cell division. Interestingly, a conditional mutant of ripA, which encodes an essential mycobacterial peptidoglycan hydrolase, was similarly hypersusceptible to a β-lactam, carbenicillin, but not to cycloserine following ripA depletion [54].
While its specific mechanism of action remains to be determined, it is plausible that PerM, as an integral membrane protein with 10 transmembrane helices, could serve a structural role, recruiting or anchoring key cell division proteins, such as peptidoglycan transpeptidases or hydrolases, to the division site. It may serve to bridge cytoplasmic proteins, such as FtsZ or early peptidoglycan synthesis machinery, with cell division proteins in the periplasm, or it could be involved in transport of cell envelope components.
The perM mutant withstood numerous stresses in vitro, including reactive oxygen and nitrogen species, cell wall-perturbing detergent, and carbon starvation, showing a specific vulnerability to a low-Mg2+ environment. Surprisingly, the perM mutant was not more susceptible than wt to exposure to arachidonic acid at pH 5.5, despite its sensitivity to Tween-80 at pH 4.5, which suggested that free oleic acid might be toxic to the mutant at low pH. It is plausible that oleic acid released from Tween-80 and arachidonic acid cause toxicity by different mechanisms. In addition, the lower pH of the Tween-80 containing medium may have contributed to the enhanced killing of the mutant.
Survival of perM:tn in IFN-γ activated macrophages was impaired, when the bacteria were pre-grown in reduced Mg2+. IFN-γ activated, Mtb infected macrophages have a limited lifespan ex vivo, which prevented extending the time course of the ex vivo infection to better mimic the mouse infection. It is possible that pre-growth in reduced Mg2+ has the same impact as replication in the acute phase of mouse infection, but interpretation of the ex vivo macrophage infection data is difficult and does not allow direct conclusions about the intraphagosomal availability of Mg2+.
Previous work revealed that macrophage activation by IFN-γ results in changes in the intraphagosomal concentrations of several metals, but Mg2+ was not measured [55]. The Salmonella-containing phagosome was estimated to contain 10 to 50 μM Mg2+, based on strong induction of the Mg2+-regulated mgtB gene in Salmonella in both low Mg2+ media and upon uptake by mammalian cells [56,57]; however, measurement of intraphagosomal Mg2+ using nanosensor particles showed the concentration to be approximately 1 mM in the first two hours of infection [58]. The intraphagosomal concentration of Mg2+ in vivo, after days or weeks of Mtb infection, remains a topic of speculation. Our work lends support to the hypothesis of an Mg2+-depleted environment in the Mtb-containing activated macrophage in vivo. Unfortunately, measuring intraphagosomal Mg2+ concentrations is extremely challenging. Purification of Mtb infected phagosomes is difficult and it is unknown if the purification process alters phagosomal ion concentrations. Fluorescent Mg2+ reporters exhibit much higher affinity for Ca2+ and also bind Zn2+, while PEPPLE (probe encapsulated by biologically localized embedding) technology suffers from low magnesium affinity [59]. Future development of novel and better sensors for magnesium is required to overcome these obstacles.
The PerM mutant stimulated a hyperinflammatory cytokine response in infected macrophages ex vivo. While we did not detect elevated cytokine levels in lungs of mice infected with the perM mutant compared to wt-infected mice, we cannot rule out the possible contribution of a hyperinflammatory response to the persistence defect. The cytokine measurements might have been confounded by differences in bacterial loads, and even a small, perhaps difficult to quantify, difference in the host immune response could synergize with Mg2+ restriction to result in killing of perM::tn; or the response may be localized, with lesion-centric inflammation contributing to killing perM::tn, but little impact on total cytokine levels in the lungs.
Growth of Mtb in reduced Mg2+ prior to macrophage infection resulted in an augmented response to the mutant, but no difference in response to wt Mtb, suggesting that cell wall instability of the mutant may contribute to the hyperinflammatory phenotype. In light of other evidence linking PerM to cell division, it is plausible that the mutant sheds multiple components of the cell wall during a stalled or otherwise compromised division process, resulting in increased stimulation of macrophage response pathways.
The bacterial cell wall is the target of many drugs in current use. The remarkable sensitivity of perM::tn to cephalexin and piperacillin, antibiotics routinely and safely used in clinical practice, suggests an exciting possibility of PerM as a co-target. An inhibitor of PerM could potentially be used to sensitize Mtb to β-lactam antibiotics, extending their use to mycobacterial infections.
Mouse studies were performed following National Institutes of Health guidelines for housing and care of laboratory animals and performed in accordance with institutional regulations after protocol review and approval by the Institutional Animal Care and Use Committee of Weill Cornell Medical College (protocol # 2008–0006, pH homeostasis in Mycobacterium tuberculosis).
PerM::tn, the Mtb H37Rv transposon mutant of gene perM (rv0955) containing a ΦMycoMarT7 transposon insertion at nucleotide 701, was isolated in a screen for acid-sensitive mutants described previously [13]. Mtb strains were grown in a humidified incubator at 37°C with 5% CO2 in Sauton’s media with 0.05% Tween 80 or 0.05% tyloxapol; Middlebrook 7H9 medium (Difco) containing 0.2% glycerol, 0.5% bovine serum albumin, 0.2% dextrose, 0.085% NaCl, and 0.05% Tween 80; or Middlebrook 7H11 agar (Difco) containing 10% OADC supplement (Becton Dickinson) and 0.5% glycerol. Nominally magnesium-free Sauton’s media was prepared with 0.8 mM citric acid, 9 mM sodium citrate, 3 mM potassium phosphate, 30 mM L-asparagine, and 6% glycerol, chelated overnight with 20 g/L Chelex 100 resin (Bio-Rad), filtered to remove Chelex, supplemented with 0.2 mM ferric ammonium citrate and 5 μM zinc sulfate, and adjusted to pH 7.4. Before use, 0.05% Tween 80 and 2 mM MgCl2 were added unless otherwise indicated. We call this medium “nominally magnesium-free” as trace residual magnesium is likely present. Hygromycin B (50 μg/ml), kanamycin (15 μg/ml) and streptomycin (20 μg/ml) were included when required for selection.
Rv0955 was PCR amplified from H37Rv genomic DNA and cloned behind the hsp60 promoter into a plasmid that integrates into the chromosomal phage integration attL5 site. For localization studies, GFP was fused to the C-terminus of Rv0955 and expressed from the hsp60 promoter on an integrative plasmid.
Female C57BL/6, or IFN-γ-/- mice (Jackson Laboratory) were infected using an inhalation exposure system (Glas-Col) with early-log-phase Mtb to deliver approximately 100 bacilli per mouse. Bacterial numbers were enumerated by plating serial dilutions of lung or spleen homogenates on 7H11 agar plates for CFU. Upper left lung lobes were fixed in 10% buffered formalin, embedded in paraffin and stained with hematoxylin and eosin.
Bone marrow derived macrophages were harvested and differentiated as previously described [13] and seeded at 4x106 cells/mL, with or without 50 ng/mL murine IFN-γ (R&D Systems). Approximately sixteen hours later, macrophages were infected at a multiplicity of infection (MOI) of 0.1 with a single cell suspensions of log-phase Mtb grown for 6 days in 250 or 2000 μM MgCl2. Monolayers were washed with PBS 4 hours post-infection to remove extracellular bacteria. After 4 hours, 3 days, or 6 days, macrophages were lysed with 0.5% Triton X-100 and bacteria were enumerated by plating serial dilutions on 7H11 agar plates. Half of the media in each well was replaced with fresh media after 3 days.
Bone marrow derived macrophages from C56BL/6 or TLR2-/- mice (Jackson Laboratories) were harvested and differentiated as previously described [13]. Immortalized macrophage cell lines from wild type, and Nod1/2-/- mice [60] were a gift from M. A. Kelliher at the University of Massachusetts. Macrophages were seeded at 4x105 cells/ml (wt macrophages) or 6x105 cells/ml (knockout macrophages). After 16 hours, they were infected at the indicated MOI with a single cell suspension of log-phase Mtb. For experiments using dead bacteria, Mtb was fixed in 10% formalin for 16 hours, washed twice in PBS, and added to Mtb at an MOI of 20. For exposure of macrophages to Mtb-conditioned culture media, Mtb was grown for 8 days in detergent-free Sauton’s media containing 2 mM MgCl2, then culture supernatant was passed through a 0.2 μm filter, concentrated approximately 10-fold in Amicon Ultra-15 Centrifugal Filter Units (Millipore), and added to macrophages at a volume equivalent to 10 μg protein. Supernatants were collected after 24 hours, passed through a 0.2 μm filter, and stored at -80°C. Cytokine levels were quantified using BD OptEIA ELISA kits for mouse TNF or IL-12p40 (BD Biosciences), or a multiplex ELISA Mouse ProInflammatory 7-Plex Tissue Culture Kit (Meso Scale Discovery). Tissue processing, RNA isolation, and real-time PCR were performed as previously described [61].
Mtb was grown to log phase in Sauton’s media containing 2000 μM MgCl2 prior to each experiment. Single cell suspensions were prepared in assay medium by centrifugation at 800 rpm for 12 minutes, then diluted to OD 0.02–0.05 and incubated in the following conditions: 3 days in pH 4.5 in media containing 0.05% Tween 80 or tyloxapol; 5 weeks in PBS with 0.05% tyloxapol; 24 hours in 7H9 media with 0.05% Tween 80 and 2.5 mg/ml lysozyme; 5 hours in 7H9 media with 0.05% Tween 80 and 0.1% SDS; 3 days in 7H9 media at pH 5.5 with 0.05% tyloxapol and 5 mM NaNO2; 3 days in 7H9 media at pH 5.5 with 0.05% tyloxapol and 50 μM arachidonic acid; 3 hours in 7H9 media containing 10 mM H2O2. For the multi-stress survival assay, Mtb was incubated in 1% oxygen for 14 days in modified Sauton’s media at pH 5.5 containing 0.05% tyloxapol, 0.05% butyrate, 0.5 mM sodium nitrite, 2000 μM MgCl2, and without glycerol. The exposure times to different stress conditions were selected so that viability of wt Mtb was reduced by approximately 5- to 10-fold. For conditions in which wt Mtb survived without significant or very slow death (carbon starvation, multi-stress model) extended incubation times were chosen. To determine viability, serial dilutions of cultures were plated on 7H11 plates.
Mtb was grown to log phase, washed twice in assay medium, and diluted to OD580 0.02 in plates containing two-fold serial dilutions of MgCl2, MgSO4, ZnCl2, MnCl2, CaCl2, CuCl2, or ferric ammonium citrate. For experiments testing various cations as substitutes for Mg2+, a basal level of either 100 or 250 μM MgCl2 was added as indicated. For experiments involving ZnCl2 or ferric ammonium citrate, modified Sauton’s medium was prepared without the respective cation.
Mtb was washed twice in nominally Mg2+-free Sauton’s media, diluted to OD 0.1 in Sauton’s media containing 250 or 2000 μM added MgCl2. After 5 days, cultures were washed twice in PBS with 0.05% Tween 80. To determine the impact of very low Mg2+, cultures were grown in Sauton’s media containing 2000 μM added MgCl2 until mid-log phase, then washed twice in nominally Mg2+-free Sauton’s media and incubated in Sauton’s media containing 25 μM added MgCl2. Pellets were collected at 0 hour, 3 hours and 10 hours post inoculation. After normalizing for biomass, pellets were heated at 80°C for 1 hour to kill Mtb, then resuspended in 200 μL 70% nitric acid, trace element grade (Fisher) and heated at 80°C for 2h before ICP-QQQ processing and analysis. Samples were analyzed on an Agilent 8800 ICP-QQQ running in MS/MS mode. Instrument daily performance qualification and method specific tuning was achieved by the expert AutoTune function of the MassHunter software (B.01.02). Typical sample introduction parameters for direct injection were used; RF Power 1550W, sample depth 8 mm, carrier gas 0.95 L/min, and dilution gas was set at 0.15 L/min. These parameters resulted in an oxide ratio of 0.8% (CeO/Ce). Prior to analysis, samples were diluted to a final volume of 2 ml and analyzed against multi-element external calibration standards (Agilent, Wilmington, DE). NIST 1643e was used as a standard reference material for calibration verification and monitor any possible drift during the analytical run.
Mtb was grown standing for 5 days in Sauton’s media containing 250 or 2000 μM MgCl2 and 0.05% Tween 80. Flasks were shaken for 5 hours prior to harvest. Cultures were mixed with an equal volume of GTC buffer containing guanidinium thiocyanate (4 M), sodium lauryl sulfate (0.5%), trisodium citrate (25 mM), and 2-mercaptoethanol (0.1 M) and pelleted by centrifugation. Bacterial RNA was isolated as previously described [62]. For microarray experiments, RNA was labeled using a Low Input Quick Amp Labeling Kit (Agilent). Microarrays were custom-designed (Genotypic Technology, Bangalore, India). Analysis was performed using Agilent GeneSpring software. The complete Microarray data sets have been submitted to the Gene Expression Omnibus (GEO) database.
For gene expression analysis by quantitative real-time PCR, cDNA was generated using MuLV Reverse Transcriptase (Invitrogen) and quantified using Roche Light Cycler 480 Real-Time PCR System with primers and TaqMan probes designed using Primer3 (http://bioinfo.ut.ee/primer3-0.4.0). Primer and probe sequences are available upon request.
Mtb was grown for 5 days in Sauton’s media containing 25, 250, 500, or 2000 μM MgCl2 and 0.05% Tween 80 before fixation, processing, and imagining by scanning electron microscopy as previously described [13]. Cell lengths were measured using Adobe Photoshop software.
Mtb was grown to early log phase and diluted to an optical density of 0.02 in Sauton’s medium containing 2 mM MgCl2 and 0.05% Tween 80. Bacteria were then exposed to twofold dilutions of piperacillin, cephalexin, ampicillin, meropenem, rifampicin, polymyxin B, DCCD, vancomycin, D-cycloserine, streptomycin, chloramphenicol, isoniazid, and ethambutol (Sigma-Aldrich). For assays of meropenem, DCCD, isoniazid, chloramphenicol, and rifampicin, all wells contained 0.5% DMSO. For the assay of cephalexin, all wells contained 4 mM NH4OH. The MIC was recorded as the minimum concentration at which growth, measured by optical density (OD580), was inhibited by at least 90%, as compared to a control containing no antibiotic, after approximately 2 weeks.
M. smegmatis expressing PerM-GFP was sealed in B04A microfluidic flow chamber plates (Cell Asic, part of EMD Millipore) and perfused with Middlebrook 7H9 broth at 37°C. Cells were visualized by fluorescence microscopy using an inverted Olympus IX-70 microscope equipped with a GFP filter set, a Photometrics CoolSnap QE cooled CCD camera, and an Insight SSI 7 color solid state illumination system. Snapshots were captured every 15 minutes.
The chromogenic cephalosporin nitrocefin (Fisher) was used to assay ß-lactamase activity as previously described [63] in whole-cell lysates of Mtb saturated cultures grown in Sauton’s media with 2 mM MgCl2.
|
10.1371/journal.ppat.1003625 | Calcineurin Plays Key Roles in the Dimorphic Transition and Virulence of the Human Pathogenic Zygomycete Mucor circinelloides | Many pathogenic fungi are dimorphic and switch between yeast and filamentous states. This switch alters host-microbe interactions and is critical for pathogenicity. However, in zygomycetes, whether dimorphism contributes to virulence is a central unanswered question. The pathogenic zygomycete Mucor circinelloides exhibits hyphal growth in aerobic conditions but switches to multi-budded yeast growth under anaerobic/high CO2 conditions. We found that in the presence of the calcineurin inhibitor FK506, Mucor exhibits exclusively multi-budded yeast growth. We also found that M. circinelloides encodes three calcineurin catalytic A subunits (CnaA, CnaB, and CnaC) and one calcineurin regulatory B subunit (CnbR). Mutations in the latch region of CnbR and in the FKBP12-FK506 binding domain of CnaA result in hyphal growth of Mucor in the presence of FK506. Disruption of the cnbR gene encoding the sole calcineurin B subunit necessary for calcineurin activity yielded mutants locked in permanent yeast phase growth. These findings reveal that the calcineurin pathway plays key roles in the dimorphic transition from yeast to hyphae. The cnbR yeast-locked mutants are less virulent than the wild-type strain in a heterologous host system, providing evidence that hyphae or the yeast-hyphal transition are linked to virulence. Protein kinase A activity (PKA) is elevated during yeast growth under anaerobic conditions, in the presence of FK506, or in the yeast-locked cnbR mutants, suggesting a novel connection between PKA and calcineurin. cnaA mutants lacking the CnaA catalytic subunit are hypersensitive to calcineurin inhibitors, display a hyphal polarity defect, and produce a mixture of yeast and hyphae in aerobic culture. The cnaA mutants also produce spores that are larger than wild-type, and spore size is correlated with virulence potential. Our results demonstrate that the calcineurin pathway orchestrates the yeast-hyphal and spore size dimorphic transitions that contribute to virulence of this common zygomycete fungal pathogen.
| Calcineurin is a Ca2+/calmodulin-dependent, serine/threonine-specific protein phosphatase. In pathogenic fungi, calcineurin is involved in morphogenesis and virulence. Therefore, calcineurin is an attractive antifungal drug target. The roles of calcineurin in virulence have been established in both major human pathogenic fungi (Candida species, Cryptococcus neoformans/gattii, Aspergillus fumigatus) and in plant pathogenic fungi (Magnaporthe oryzae, Ustilago maydis/hordei). However, the role of calcineurin is currently unknown in pathogenic zygomycetes. We found that the calcineurin inhibitors FK506 and cyclosporine A inhibit the growth of a prevalent zygomycete pathogen, Mucor. This fungus grows as multi-budded yeast under anaerobic conditions, and we have found that even in aerated culture (which without FK506 would result in abundant hyphal growth), Mucor exhibits yeast growth when exposed to FK506. Mucor cnbR mutants that lack the calcineurin regulatory subunit essential for calcineurin activity, are locked in perpetual yeast phase growth, indicating that calcineurin is required for hyphal growth. We further demonstrated that these yeast-locked mutants are attenuated for virulence, illustrating that hyphae or the yeast-hyphal transition are linked to virulence. These findings indicate that: 1) calcineurin governs the yeast/hyphae morphogenic transition; 2) a link exists between respiration and the calcineurin pathway; and 3) calcineurin inhibitors are attractive anti-mucormycosis drug candidates.
| Mucor circinelloides is one causal agent of mucormycosis, an uncommon but frequently lethal fungal infection of humans. Several other species belonging to the Mucorales order also cause mucormycosis, including some Rhizopus species, Absidia, Cunninghamella, Rhizomucor, Apophysomyces trapeziformis, and others [1]–[3]. Mucormycosis is an emerging fungal infection often afflicting immunocompromised and vulnerable populations including patients with diabetes, AIDS, hematologic malignancies, or trauma [3]–[8]. Susceptible hosts also include patients with solid organ transplants and those with high serum iron levels [4], [8], [9]. Mucormycosis is the second most common fungal infection among patients with hematological malignancies and transplants. This infection is associated with high mortality rates [4], [8], [10]: ∼50% for all mucormycosis infections and >90% for disseminated infections [4]–[6], [9], [11]. Recently, a cluster of mucormycosis cases was reported among the tornado victims in Joplin, MO in the spring of 2011; 13 cases were reported with five fatalities. Based on 28S ribosomal DNA sequence the responsible species has been identified as A. trapeziformis [3]. Despite the increasing incidence of disease, high mortality rate, and unmet clinical needs for therapy, zygomycetes are understudied compared to other pathogenic fungi.
Mucor species are dimorphic fungi and exhibit either hyphal or yeast growth depending upon the conditions (reviewed in [12]). Hyphal growth of Mucor was first thought to originate from species of Saccharomyces by transmutation [13] until Louis Pasteur discovered that Mucor grows as a multi-budded yeast under anaerobic/high CO2 growth conditions [14]. Later, Bartnicki-Garcia and Nickerson rediscovered the induction of Mucor yeast growth by CO2 [15]–[17]. Although dimorphic Mucor species vary in their responses to the environment during morphogenic changes, common critical factors that induce yeast growth of Mucor species include oxygen concentration, CO2 concentration, and carbon source (reviewed in [12]). In addition, several chemicals that inhibit mitochondrial function [including potassium cyanide and antimycin A (which block electron transport) or oligomycin and phenyl alcohol (which inhibit oxidative phosphorylation)] induce yeast growth in Mucor spp., even in aerobic conditions [18]–[20]. Inhibition of the synthesis of cytochrome b and other mitochondrial components by chloramphenicol also results in Mucor yeast growth [21], [22]. Thus, respiration, anaerobic conditions, and high CO2 environments all contribute to the yeast-hyphal transition. Cerulenin, a lipid metabolism inhibitor, and cycloleucine, an S-adenosylmethionine (SAM) synthetase inhibitor, both block the yeast to hyphal growth transition under aerobic conditions [23], [24]. Notably, adding cyclic AMP (cAMP) to the culture medium induces yeast growth of Mucor spp. (reviewed in [12] and references therein). cAMP activates cAMP-dependent protein kinase A (PKA), implicating a role for protein kinase A in the Mucor dimorphic transition, and this is supported by a series of recent studies, including genetic analyses in which genes encoding PKA regulatory subunits were disrupted by homologous recombination [25]–[31]. PKA also plays a key role in various morphogenesis processes including germination, branching, and polarized growth in Mucor [26], [32]–[37]
Dimorphism has evolved in multiple lineages in the fungal kingdom, including ascomycetes, basidiomycetes, and zygomycetes [38]–[40]. In many pathogenic fungi, the morphogenic transition is closely related to pathogenicity. A group of so-called “dimorphic fungi” grow as molds at lower environmental temperatures; however, in association with the host these fungi undergo a thermal dimorphic switch, resulting in yeast growth as virulent forms at 37°C (reviewed in [41], [42]). The dimorphic fungal pathogens include Histoplasma capsulatum, Paracoccidioides brasiliensis, Coccidioides immitis/posadasii, Blastomyces dermatitidis, and Penicillium marneffei, among others. The most prevalent fungal commensal and pathogen, Candida albicans, exhibits yeast and hyphal growth, and the ability of this fungus to grow as both forms is necessary for virulence; yeast-locked or hyphal-locked mutants are both avirulent [39], [40], [43], [44]. The basidiomycete human pathogen Cryptococcus neoformans changes its morphology; it grows mainly as yeast but during the sexual cycle the fungus forms hyphae and produces sexual spores that are readily disseminated as infectious propagules [45]–[47]. C. neoformans also exhibits pseudohyphal growth, which is advantageous for escape from amoebae, one of its natural predators [48], [49]. In the plant pathogenic basidiomycete Ustilago maydis, the yeast form is not virulent; however, the hyphae produced by mating cause disease on plant hosts and transform infected plant tissue into galls [50]. Therefore, morphogenic transitions play major roles during host infection by pathogenic fungi. Dimorphism in zygomycetes, especially in Mucor spp. was studied decades ago; however, our knowledge of its contributions to pathogenesis is limited and the underlying genetics, beyond the involvement of anaerobic/high CO2, respiration, and PKA is not well established.
Our study reveals that calcineurin plays a crucial role during the yeast-hyphal growth transition of M. circinelloides. The calcineurin pathway is conserved throughout eukaryotes. Calcineurin is a Ca2+/calmodulin-dependent serine/threonine specific protein phosphatase that consists of two subunits: the catalytic A subunit, which has phosphatase activity, and the regulatory B subunit, which binds calcium and the A subunit to activate the enzyme complex. Both subunits are required for enzyme activity. The drugs tacrolimus (FK506) and cyclosporine A (CsA) form complexes with the immunophilins FKBP12 and cyclophilin A, respectively. The protein-drug complexes bind to the hydrophobic interface between the calcineurin A and B subunits to inhibit phosphatase activity [51], [52]. We found that FK506 inhibits hyphal growth and instead induces yeast growth in the pathogenic zygomycete Mucor. Gain of function mutations in the calcineurin regulatory B and catalytic A subunit genes were identified that confer resistance to FK506. Furthermore, disruption of the calcineurin regulatory B subunit required for calcineurin activity results in a yeast-locked phenotype, supporting our conclusion that calcineurin regulates the dimorphic transition. Notably, the cnbRΔ yeast phase locked mutants were less virulent, suggesting that either hyphae are more virulent than yeast or the morphogenic switch is central to pathogenicity of this fungus. In addition, cAMP-dependent protein kinase A activity was found to be elevated during yeast compared to hyphal growth, and also when calcineurin activity was inhibited by FK506 or mutation. We also show that calcineurin is involved in hyphal polarity and spore size dimorphism, which is linked to virulence [53].
Calcineurin is involved in the morphogenesis and virulence of multiple pathogenic fungi (reviewed in [54], [55]): in C. neoformans, calcineurin is required for growth at 37°C [56]; in Candida spp., calcineurin functions in antifungal drug resistance/tolerance, survival in serum, and virulence [57]–[60]; calcineurin plays a role in morphogenesis in P. brasiliensis [61]; and in Aspergillus fumigatus, calcineurin regulates morphogenesis and thereby pathogenesis [62]. Therefore, calcineurin is a promising antifungal drug target and this study provides a novel foundation to develop approaches to control the emerging fungal infection mucormycosis.
We found that in Mucor the calcineurin inhibitor FK506 inhibits hyphal growth and drives multi-budded yeast growth (Figure 1 and Videos S1 and S2). Eight different M. circinelloides isolates (103 spores) were each inoculated aerobically on YPD agar media or YPD agar media containing FK506 (1 µg/mL). At 4 days post-inoculation, the colonies grown on YPD with FK506 were significantly smaller than those on YPD alone (Figure 1A). The smaller, compact colonies consisted entirely of yeast form Mucor (data not shown). In liquid culture, with vigorous shaking for aeration, the Mucor isolate CBS277.49 grew exclusively as a mold (Figure 1B); however, this isolate exhibited only yeast growth when cultured in the same fashion but in the presence of FK506 (1 µg/mL). The multi-budded yeast phenotype imposed by FK506 phenocopies Mucor yeast growth under anaerobic/high CO2 growth conditions (Figure 1B). This observation indicates that there is a link between respiration/CO2 sensing and the calcineurin pathway. On the other hand, another calcineurin inhibitor cyclosporine A (CsA) did not induce multi-budded yeast growth, and instead resulted in abnormal, stunted hyphal growth (Figure S1). Higher concentrations of CsA did not produce yeast growth (data not shown). This result suggests that FK506-specific calcineurin inhibition plays a role in the hyphal-yeast switch. We note that although CsA does stunt hyphal growth, the efficacy may not be sufficient to enforce yeast growth compared to FK506 if levels of the cyclophilin A-CsA complex are insufficient to inhibit all calcineurin activity. This hypothesis is supported by findings that 1) in a sensitized background in which calcineurin activity has been reduced with a sub-inhibitory concentration of FK506, CsA is now fully able to block the dimorphic transition and impose yeast growth (Figures S2 and 2) in regulatory B subunit mutant CNBR-1 (N125Y, note that the all gain-of-function mutations are designated with capital letters) mutants isolated as being resistant to FK506 (described below), CsA imposes a multi-budded yeast phenotype (See supplemental figure S5).
We identified calcineurin pathway components in the M. circinelloides genome. Unusually high numbers of calmodulin and calmodulin (CAM) kinase orthologs (nine calmodulins) were identified in the genomes with six CAM kinases Figure S3). Interestingly, the Mucor genome encodes three calcineurin catalytic A subunits along with a single calcineurin regulatory B subunit, FKBP12, and cyclophilin A.
We further examined the three catalytic A subunit (cna) genes and found that the Cna ortholog proteins contain an N-terminal phosphatase domain and a C-terminal regulatory domain containing the calcineurin B subunit binding domain, cyclophilin A-cyclosporine A complex binding domain, FKBP12-FK506 complex binding domain (the domains were determined based on [63]), calmodulin binding domain, and auto-inhibitory domain (Figure 2), further demonstrating that the identified Cna proteins are bona fide calcineurin catalytic A subunits. The sequences of the Cna proteins are distinct, and there is 73% identity shared between CnaA and CnaB; 64% identity between CnaA and CnaC; and 62% identity between CnaB and CnaC. It is interesting to consider how multiple paralogs of the cna genes evolved in the zygomycetes (Figure S3). In phylogenetic analyses, the three zygomycete species including M. circinelloides, Rhizopus delemar, and Phycomyces blakesleeanus, are conserved on a common branch, including McCnaA, PbCnaA, RdCnaA, and RdCnaB. This observation suggests that the calcineurin A subunit gene in this group may be the more ancestral (Figure S3).
FKBP12 is an immunophilin family protein with cis-trans peptide prolyl isomerase activity that serves as the cellular receptor for FK506 and rapamycin [64]. When bound to FK506, FKBP12 binds to the interface between the calcineurin catalytic A and regulatory B subunits, inhibiting phosphatase activity by occluding substrate access to the active site [51]. FKBP12 also binds to rapamycin to inhibit the Tor pathway. Disruption of the gene encoding FKBP12 confers resistance to FK506 and rapamycin (in the absence of FKBP12, FK506 fails to bind calcineurin) [65]. Amino acid substitutions in the calcineurin regulatory B and catalytic A subunit surface that interact with the FKBP12-FK506 complex result in resistance to FK506 [63], [66]. Calcineurin FK506 resistant mutants expressing FKBP12 remain rapamycin sensitive. Sensitivity is defined as FK506 enforced yeast growth or rapamycin limiting hyphal growth of the fungus; resistance to FK506 is defined as strains forming hyphae instead of yeast in the presence of FK506, and resistance to rapamycin as strains growing as in non-drug media.
To test whether these molecular principles of FK506 sensitivity and resistance apply in M. circinelloides, we grew wild-type strains (103 spores) in solid YPD media containing 1 µg/ml FK506 until resistant mycelial growth emerged out of yeast colonies as previously described [65]. Through this approach six FK506 resistant spontaneous mutants, MSL11, MSL12, MSL13, MSL14, MSL15, and MSL16 were generated (Table 1 and Figure 3). These FK506 resistant mutants were as rapamycin sensitive as the wild-type parental strain and in contrast to an fkbAΔ strain lacking FKBP12 that is FK506 and rapamycin resistant. Another mutant strain SM4 has a base substitution in the fkbA gene resulting in a leucine to proline substitution (L91P) in FKBP12 that confers resistance to FK506 but not to rapamycin [65]. Five of six spontaneous mutants remained sensitive to CsA, whereas the MSL16 mutant (CNBR-3) was found to be cross-resistant to CsA (see Figure S6). These observations led us to examine the genes involved in the mode of inhibition by FK506, and we sequenced the cnbR, cnaA, cnaB, and cnaC calcineurin genes as well as the fkbA gene to test if any mutations were present in any of these genes in the six FK506 resistant isolates.
Four strains had mutations in only the cnbR gene (Figure 3 and S4). The MSL11 and MSL15 mutants (CNBR-1) contained a point mutation, A496T, which results in an amino acid alteration, N125Y. It is interesting that two independent mutant strains contained an identical mutation in the cnbR gene and these strains displayed yeast growth in the presence of CsA (CsA hypersensitive) (Figure S5). This mutation may both impair FKBP12-FK506 binding and compromise the stability of the calcineurin A–B complexes to result in CsA hypersensitivity. The MSL12 strain (CNBR-2) contained an insertion of three nucleotides ‘CCA’ at the 513th bp of the ORF (513_514insCCA), resulting in an insertion of histidine (H) after 129th aspragine, Asn129_Gln130insHis. The MSL16 mutant (CNBR-3) contained a point mutation, G487T, which results in the amino acid substitutions, V122F. Interestingly, all of these amino acid alterations are present in the latch region of the calcineurin B regulatory subunit that is known to interact with the FKBP12-FK506 complex and to be involved in phosphatase activity [52], [66]. It is likely that the N125Y, N129_Q130insH, and V122F changes alter the interaction between FKBP12-FK506 and the calcineurin B-calcineurin A complex, resulting in resistance to FK506. Interestingly, the V122F substitution may also alter the interaction between cyclophilin-CsA and the calcineurin B-calcineurin A complex as the MSL16 mutant (CNBR-3, V122F) was found to be cross-resistant to CsA (Figure S6).
The two other mutant strains had mutations in only the cnaA gene (Figure 3 and S7). The MSL13 mutant (CNAA-1) has a point mutation, G1514C in the ORF, resulting in an amino acid change S378T, which lies in the amino acid adjacent to a previously known FK506 resistant mutant in calcineurin A in S. serevisiae (W430C) [63]. The MSL14 mutant (CNAA-2) has a point mutation, A1489G in the ORF, resulting in the amino acid substitution, N370D. These mutations occurred in the binding domain for calcineurin B and for the FKBP12-FK506 complex and likely reduce the binding affinity of the FKBP12-FK506 complex to the interface between the calcineurin A and B subunits, resulting in a less efficient inhibition of calcineurin activity by FK506. The fact that two FK506 resistant alleles were isolated in one of the three calcineurin A catalytic subunit (CnaA) suggests that it alone is sufficient to promote hyphal growth when cells are exposed to FK506, as the remaining two calcineurin A catalytic subunits (CnaB and CnaC) in complex with the CnbR regulatory subunit should be inhibited by FKBP12-FK506 under these conditions. The isolation and analysis of these FK506 resistant mutants in calcineurin A and B subunits further support the conclusions that the calcineurin pathway is: 1) conserved, 2) responsible for the dimorphic transition, and 3) the target of FK506 action in M. circinelloides.
FK506 treatment results in yeast growth in Mucor; therefore, we hypothesized that calcineurin activity is essential to maintain hyphal growth. To test this hypothesis we disrupted the cnbR gene encoding the calcineurin B regulatory subunit as described in the Materials and Methods. The cnbR gene in the wild-type strain MU402 (pyrG−, leuA−) was replaced with a cnbRΔ::pyrG allele, and the gene deletion was confirmed by 5′ and 3′ junction PCR as well as ORF spanning PCR, and further confirmed by Southern blot analysis to show precise homologous recombination and the absence of any ectopic integration events (Figures S8 and S9). We generated two independent cnbRΔ deletion mutants from separate transformations.
The wild-type and cnbRΔ mutants (103 cells) were inoculated in the center of solid YPD medium and the strains were grown for three days at 30°C under normal aerobic conditions. The wild-type formed a large colony consisting of complex mycelia, whereas the cnbRΔ mutants displayed a compact yeast colony (Figure 4A). In liquid YPD media, wild-type Mucor grew as a filamentous mold; however, the cnbRΔ mutants displayed only yeast growth, even in aerobic conditions (Figure 4B). No impact of FK506 was observed on the cnbRΔ mutants consistent with a complete absence of the drug target (Figure 4C). This phenotype of the cnbRΔ mutants is essentially identical to that of wild-type Mucor grown in the presence of FK506 in aerobic conditions (Videos S2 and S3). The observations that mutation or inhibition of calcineurin enforces yeast growth are the key findings that allow us to conclude that the calcineurin pathway orchestrates the dimorphic switch in Mucor.
To assess the virulence of yeast vs. hyphae and to test the role of the calcineurin pathway in virulence, the cnbRΔ yeast-locked mutants were compared to both wild-type spores and wild-type yeast cells in animal virulence models (Figure 5). The R7B (leuA−) strain served as wild-type to exclude the possibility that leucine auxotrophy could generate a biased result, as the cnbRΔ mutants were generated in the MU402 (pyrG− leuA−) strain background (pyrG− leuA− cnbR::pyrG). Wild-type spores, wild-type yeast, and cnbRΔ mutant yeast (two independent mutant isolates) were quantified and cohorts of 10 wax moth larvae (a heterologous host system) per strain were infected with an inoculum of 20,000 infectious units. Wild-type spores and wild-type yeast both showed 100% mortality by four days post-infection (Figure 5A). Wild-type yeast are able to switch into growth as hyphae; thus, it was anticipated that wild-type yeast would be as virulent as wild-type spores. On the other hand, the cnbRΔ yeast-locked mutants did not kill the larvae after 8 days and were indistinguishable from the PBS control. When 40,000 cells were inoculated, the yeast-locked cnbRΔ mutants showed a higher level of virulence (60 to 70% mortality by day 4 post-infection), but were still significantly less virulent than wild-type spores (WT vs. cnbRΔ1, p = 0.00059) (Figure 5B). These results indicate that: 1) morphogenesis contributes to virulence in this fungus; 2) the calcineurin pathway is involved in virulence, and 3) calcineurin inhibitors have promise as antifungal drugs for zygomycete infections.
That hyphae are more virulent than yeast, or alternatively that ability to transition between yeast to hyphae is a key virulence factor as observed in Candida albicans [40], is further supported by observations of tissues from infected animal host models (Figure S10). In a wax moth mucormycosis system, wild-type mainly presented as the hyphal form, whereas the yeast-locked mutants were exclusively as yeast. In a murine mucormycosis system, we found that the majority of an M. circinelloides fungal burden was present in the brain as hyphae.
Yeast growth can be induced by elevated levels of CO2 [15], [16]. Carbon dioxide is converted into bicarbonate ions (HCO3−) by carbonic anhydrase (reviewed in [67] and references therein). Subsequent activation of adenylyl cyclase by CO2, HCO3−, or both occurs, resulting in the production of cAMP. cAMP then binds to the regulatory subunit of protein kinase A (PKAR) and the protein kinase A (PKA) catalytic subunits are released to result in activation. Previous studies suggested that PKA may play a major role during yeast growth, given that the PKA genes are over-expressed during yeast growth, and the regulatory subunit gene (pkaR) that inhibits PKA activity is expressed at a higher level during the yeast to hyphal transition [25], [27], [28]. We further confirmed that Mucor grows as yeast in the presence of high concentrations of bicarbonate in the media. At 50 mM bicarbonate, Mucor exhibited both hyphal and yeast growth; at 75 mM, Mucor grew only as a yeast (Figure 6A).
We adapted a PKA activity assay used in other studies of Mucor morphogenesis to test how PKA activity is differentially regulated during yeast and hyphal growth [27], [28], [68]. Due to the nature of this assay performed in the presence of cAMP, the maximum possible PKA activity was measured. Crude cell extracts (0.5 µg total protein) were prepared and PKA activity was measured by using the PKA specific substrate kemptide as previously described [27], [68]. Compared to hyphal growth, PKA activity was elevated approximately five-fold during yeast growth in high CO2 growth conditions (Figure 6B). During yeast growth enforced by FK506, PKA activity was higher. In the two independent cnbRΔ yeast-locked mutants, PKA activity was also higher than in wild-type (Figure 6B). These results demonstrate that 1) higher activity of PKA is associated with, and may be necessary for, yeast growth; and 2) calcineurin and PKA may play antagonistic roles during the yeast-hyphal dimorphic transition. It is noteworthy that the M. circinelloides genome encodes four PKA regulatory subunits and their subcellular localization may generate compartmentalized PKA activity [26]–[28]. In the non-dimorphic zygomycete R. delemar, PKA activity was also elevated by FK506 treatment, and the same result was observed in the basidiomycete pathogen C. neoformans (Figure S11).
We tested whether the three catalytic A subunit genes (cnaA, cnaB, and cnaC) are differentially regulated during the dimorphic transition. Initial northern blots showed that the probes for each cna gene were cross-reactive due to sequence similarity (data not shown); therefore, we performed quantitative RT-PCR with gene specific primers to specifically examine expression of each individual gene. In RT-PCR analyses, we found that all three cna genes were expressed, indicating that none were pseudogenes (data not shown). The PCR products from cDNAs obtained for each gene were sequenced to re-annotate the genes. Based on the ORF sequences, we designed a pair of specific RT-PCR primers across two exons of each cna gene. The actin gene was used for normalization.
The expression levels of the cna genes were evaluated using ct values from RT-PCR with hyphal growth conditions as a control. We found that cnaC is expressed at higher levels during yeast growth under microaerobic conditions but at lower levels during hyphal growth. Interestingly, during yeast growth driven by FK506, cnaC is also overexpressed (Figure 7). The expression levels of cnaA and cnaB were moderately reduced during the dimorphic switch. These results suggest that CnaC plays a specific role during the yeast growth phase and could therefore be less sensitive to inhibition by FK506, possibly as the result of overexpression in molar excess above FKBP12 levels. In parallel studies, Drs. Praveen Juvvadi and William Steinbach at Duke University found that M. circinelloides cnaC partially complements the radial growth defect of the A. fumigatus calcineurin catalytic A subunit (cnaA) mutant and also that MuCnaC-GFP localizes to septa, as does AfCnaA (Juvvadi P. R. et al., manuscript in preparation). M. circinelloides hyphae are coenocytic (aseptate), but during yeast growth or pseudohyphal growth, septa are formed; thus, CnaC function may be associated with septum formation.
Of the three cna genes found in Mucor, homologs for the cnaA gene were found in all three of the zygomycete genomes analyzed (Figure S3), indicating that the cnaA gene may be the more ancestral calcineurin catalytic A subunit gene. Therefore, we chose to disrupt the cnaA gene to further test the roles of calcineurin in Mucor. A disruption allele consisting of the pyrG gene flanked by ∼1 kb of the 5′ upstream and 3′ downstream untranslated regions of the cnaA gene were introduced as described in the Materials and Methods. Gene replacement by recombination was confirmed by 5′ and 3′ junction PCR, ORF spanning PCR, and Southern blot (Figures S12 and S13). Two independent cnaAΔ::pyrG mutants were obtained from separate transformations.
Interestingly, the cnaAΔ mutants exhibited much higher sensitivity to FK506 and CsA compared to wild-type (Figure 8A). Sensitivity was assessed based on the ability of the drugs to inhibit growth of the strains tested; for example, on YPD medium supplemented with FK506 (0.025 to 1 µg/ml), the two cnaAΔ mutants displayed significantly reduced growth compared to wild-type. Both cnaAΔ mutants also exhibited hypersensitivity to CsA. In liquid YPD with vigorous aeration, more abundant yeast growth was observed in the cnaA mutants compared to wild-type. Notably, the cnaAΔ mutants exhibited only yeast growth in the presence of 0.1 µg/ml FK506 in the media, which is a concentration insufficient to inhibit hyphal growth and induce yeast growth in wild-type (Figure 8B). We explain these observations as resulting from the absence of one catalytic A subunit out of the three expressed (the mutants still have intact cnaB and cnaC genes), leading to a relative decrease in calcineurin, similar to haplo-insufficiency in S. cerevisiae [69]. The cnaAΔ mutants were also more sensitive to SDS, indicating that calcineurin may be involved in cell wall integrity (Figure 8C) as observed in Candida species [57], [70]. Mucor was largely resistant to Congo red and calcofluor white and there was no significant difference between wild-type and cnaA mutants with either (data not shown).
We found that the cnaAΔ mutants produce larger spores than the wild-type (WT: 12.00±2.94 µm; cnaAΔ1: 16.08±4.04 µm; cnaAΔ2: 16.02±3.70 µm) (Figure 9A and B). The differences in spore size between the wild-type and the two cnaAΔ mutants are significant based on a two sample independent t-test (p<0.0001 in both WT vs. cnaAΔ1 and WT vs. cnaAΔ2, N = 80). The cnaA mutant spores were multinucleate as are wild-type spores. This larger spore size is an intriguing phenotype because spore size is a virulence factor as shown in our previous study, wherein larger spores were more virulent than smaller spores [53].
To test if the enlarged spore phenotype conferred by mutations in the cnaA gene might contribute to virulence, we infected wax moth larvae with 5,000 spores of wild-type or the cnaAΔ mutants. Ten larvae were used for each strain. At the given inoculum, wild-type showed a moderate level of virulence; however, the two cnaA mutants were significantly more virulent compared to the wild-type (Figure 9C). The cnaAΔ mutants also formed hyphae in hosts similar to wild-type. (Figure S14). The experiments were performed independently three times with similar results. As observed in our previous study [53], in the interaction with macrophages, the larger cnaAΔ mutant spores produced a germ tube inside of macrophages, whereas wild-type remained as spores without germ tube emergence at 3.5 hours of co-culture (Figure S15), indicating that the mutants possibly avoided the innate host immune system by sending out germ tubes earlier and more efficiently than the wild-type. This suggests that calcineurin is involved in spore size control as a negative regulator and may therefore also contribute to virulence, although other calcineurin-CnaA-dependent functions may also contribute.
The cnaAΔ mutations also confer abnormal hyphal polarity. To carefully examine hyphal growth of the cnaAΔ mutants, we applied a time course imaging technique with a microscope equipped with an automatic shutter system. Wild-type and cnaAΔ mutant spores were placed on YPD agar media, and images were captured every 30 seconds. When the cnaAΔ mutants germinated, the germ tubes exhibited a tip-splitting phenotype and abnormal branching, whereas wild-type germ tubes elongated in a single direction with no tip-splitting (Figure 10 and Videos S4 and S5). This phenotype of the cnaAΔ mutants indicates that CnaA is necessary to maintain a single hyphal polarity in Mucor. A role for calcineurin in hyphal polarity is conserved in other filamentous fungi, including Neurospora crassa and A. fumigatus [62], [71], indicating that this is a general role of calcineurin.
Morphogenic transitions (for example, yeast to hyphae and vice versa) are common in fungi and evolved in multiple fungal lineages. Many well known fungal pathogens change their morphology in response to their environments and morphogenesis is linked to their virulence. Human fungal pathogens encompass ascomycetes, basidiomycetes, and zygomycetes [72]. In ascomycetes and basidiomycetes, the relationships between dimorphism and pathogenicity are relatively well established. However, whether dimorphism is associated with virulence in zygomycetes was not known. Our study has revealed how the dimorphic transition is regulated and linked to virulence and how the calcineurin pathway directs this process in the human pathogenic zygomycete M. circinelloides.
Several zygomycete species, especially those belonging to the order Mucorales, are known to cause the lethal fungal infection mucormycosis. These species include Mucor spp., Rhizopus spp., Rhizomucor spp., Cunninghamella spp., and A. trapeziformis [1]–[3]. Unlike ascomycetous and basidiomycetous fungal pathogens, the current status of research into zygomycete pathogens is in its infancy. A limited ability to conduct genetics studies is one of the barriers to advancement in our understanding of zygomycete pathogens. M. circinelloides is one of the best developed systems for genetics and molecular biology among zygomycete species. Therefore, Mucor provides a foundation from which to understand mucormycosis.
Mucor is a dimorphic fungus and the morphogenic transition is known to occur in response to environmental conditions, especially low levels of oxygen and high levels of carbon dioxide. Several other studies have implicated protein kinase A in this response [25]–[31]. Other factors known to be involved predominantly alter mitochondrial functions. Our study revealed that the calcineurin pathway is a key regulator of this developmental process.
FK506 treatment imposes yeast growth in Mucor (Figure 1). Previously, we found that FKBP12 mutants are resistant to FK506 (resistance is defined as strains forming hyphae instead of yeast in the presence of FK506) [65]. FKBP12 is a member of the immunophilin family of proteins, which have cis-trans peptidyl- prolyl isomerase activity [64]. When bound to FK506, FKBP12 binds to the interface between the calcineurin catalytic A and regulatory B subunits, inhibiting phosphatase activity by occluding substrate access to the active site [51]. FKBP12 also binds to rapamycin to inhibit the Tor pathway. In other fungi, disruption of the gene encoding FKBP12 confers resistance to FK506 and rapamycin (in the absence of FKBP12, FK506 fails to bind calcineurin). Similar results are observed in Mucor; a splice site mutant (adenine into guanine at the 316th nucleotide), Leu91Pro substitution mutant, and FKBP12Δ null mutants are all resistant to FK506 [65].
In this study, we found that the N125Y, N129_Q130insH, and V122F alterations in the latch region of the calcineurin B regulatory protein confer resistance to FK506 (Figure 3 and S4). A previous site-directed mutagenesis study documented that the latch region is essential for immunophilin-immunosuppressant complex docking [52]. These mutations in the CnbR subunit of Mucor may prevent or limit docking of FKBP12-FK506 onto the calcineurin complex to result in less inhibition of calcineurin. Similar results has been observed in C. neoformans where a two amino acid insertion in the latch area of calcineurin B similarly confers resistance to FK506 [66]. In addition, the N370D and S378T mutations in the calcineurin B and FKBP12-FK506 binding domain of the CnaA catalytic subunit (Figure 3 and S7) confer resistance to FK506 and lead to the conclusion that Mucor has a conserved calcineurin pathway. Furthermore, the calcineurin B regulatory subunit gene disruption mutants only exhibited yeast growth, even during vigorous aeration conditions and there was no impact of FK506 on these strains that lack calcineurin (Figure 4). These observations all support the conclusion that calcineurin is a key factor in the dimorphic yeast to hyphae transition in Mucor.
Calcineurin has been suggested as a candidate antifungal drug target in many other pathogenic fungi, including Candida spp., C. neoformans, A. fumigatus, Magnaporthe oryzae, and U. maydis (reviewed in [55]). The observation that calcineurin inhibitors block hyphal growth in zygomycete pathogens parallels findings in other fungal pathogens. That transplant patients receiving FK506 treatment have a lower incidence of mucormycosis contributes to validate calcineurin as a potential antifungal drug target [73]. FK506 displays synergistic effects when combined with other antifungal agents, including azoles and echinocandins in C. albicans (reviewed in [74]). In addition, calcineurin inhibitors exhibit an in vitro synergistic inhibition on zygomycete growth [75]–[77]. A recent study suggests that combination therapy with posaconazole and FK506 is efficacious in both a Drosophila heterologous host model and a cutaneous mucormycosis murine model system [78]. We found that yeast-locked cnbRΔ mutants are not as virulent as wild-type in a heterologous wax moth model system (Figure 5). One possibility is that hyphae are a more virulent form of Mucor compared to yeast. However, although the wild-type existed as hyphae and the yeast-locked mutants existed as yeast inside hosts (Figure S10), we cannot rule out the possibility that the ability to transition between yeast and hyphae is a major factor in virulence during host infections as observed in C. albicans [40]. Currently there are no known hyphal-locked mutants in Mucor that could be used to test these hypotheses further. This observation overall supports two important conclusions: first, dimorphism is linked to virulence in this fungal species and second, calcineurin inhibitors may be an effective way to control mucormycosis.
However, because the calcineurin pathway is conserved throughout eukaryotes, including humans, calcineurin inhibitors have their own inherent risks as antifungal drugs, including for example host immune suppression. In humans, calcineurin dephosphorylates the cytoplasmic subunit of the transcription factor nuclear factor of activated T-cells (NFATc) and promotes its nuclear localization to induce interleukin 2 (IL-2) and stimulate T cell proliferation [79]. Thus, calcineurin inhibitors are immunosuppressive drugs. It is possible that the effects of calcineurin inhibitors as immunosuppressants will overpower their antifungal activities. This situation was also observed in our murine host experiments, in which FK506 treatment detrimentally affected mice infected with Mucor (Figure S16). Therefore ideal calcineurin inhibitors would be FK506 analogs that are less immunosuppressive to the host but which remain active against fungal calcineurins.
Mucor encodes three calcineurin catalytic A subunits (Figures 2 and S3). The presence of the triplicated cna genes might have been derived from individual gene duplication events, which is supported by phylogenetic analyses and the presence of repetitive sequences linked to the cna genes (Figures S2 and S17). However, in Mucor, synteny around the three cna genes was not conserved (Figure S17), indicating that the duplication of the cna genes may have resulted from individual gene duplication events or segmental gene duplications that occurred earlier and that flanking synteny has decayed with evolutionary time. The presence of the same repetitive elements up- and downstream of the cna genes supports the hypothesis that individual gene duplication events may have produced the three versions of the cna genes in Mucor. However, in R. delemar, the triplicated copies of the cna genes arose via a different trajectory, involving a recent whole genome duplication and then a segmental gene duplication. Interestingly, the non-pathogenic Mucorales species Phycomyces has only a single cna gene (Figure S3). These observations indicate that different evolutionary trajectories gave rise to genes linked to virulence during speciation in the Mucorales order.
The cnaA gene is conserved in the three Mucorales species, and this gene might therefore have a conserved ancestral role in this fungal lineage. The Mucor cnaAΔ mutants display several interesting phenotypes: an abnormal polarity during hyphal growth, larger spore production, hypersensitivity to calcineurin inhibitors, and cell wall defects (Figures 8, 9, and 10, and Videos S3 and S4). The abnormal polarity phenotypes include tip-splitting and atypical branching patterns. In several other fungal species, calcineurin is involved in hyphal polarity in accord with our observations in Mucor [62], [71], [80]. A remarkable phenotype of the cnaAΔ mutants is the production of larger spores (Figure 9). In a previous study, we showed that larger spore size is linked to higher virulence in Mucor. Accordingly, the larger spores of the cnaAΔ mutants were also found to be more virulent in the heterologous host model and macrophage model (Figures 9 and S15), further supporting the hypothesis that fungal cell gigantism is involved in virulence as observed in titan/giant cells of C. neoformans and multinucleate spherules of Coccidioides immitis/posadasii [53], [81]–[84]. This observation is at one level paradoxical because we suggest calcineurin as a target of antifungal drugs. However, we found that the growth of the cnaAΔ mutants was more sensitive to FK506 and CsA, and that the cnaAΔ mutants also exhibit cell wall defects (Figure 8). Taken together with the finding that cnbRΔ mutants that should lack all calcineurin activity are attenuated in virulence, these results indicate that calcineurin inhibitors remain attractive antifungal drug candidates.
Calcineurin regulates many cellular processes in eukaryotes, including T-cell activation, polarized growth of neurites, long-term memory transitions, signaling in cardiac hypertrophy in humans as well as morphogenesis and virulence in fungi [55], [79], [85]–[87]; however, to our knowledge, a connection between respiration and calcineurin has not been previously described. Our finding that calcineurin negatively regulates protein kinase A (PKA) activity is also novel (Figures 6 and S11). An antagonistic relationship between calcineurin and PKA has been observed in U. maydis and S. cerevisiae [88], [89]. Our study verifies that calcineurin activity is required for hyphal growth and PKA activity is correlated with yeast growth. Further, calcineurin activity might attenuate PKA activity in Mucor. Similar results were observed in two other pathogenic fungi, C. neoformans and R. delemar; thus, this antagonistic interconnection between calcineurin and PKA may be a conserved functional circuit in several fungal species. It is not clear if calcineurin directly acts to dephosphorylate PKA to control its activity or indirectly regulates PKA activity, for example, by regulation at the transcriptional level or via other proteins. Alternatively, calcineurin and PKA may share common targets (e.g. Crz1) for dephosphorylation and phosphorylation, respectively (Figure 11). Future investigation of this signaling network will be a promising venture.
In conclusion, our finding that calcineurin orchestrates the dimorphic transition in Mucor is a significant advance since the first finding that CO2/low oxygen was the key factor [15]–[17]. Another key major advance of this study is that dimorphism is linked to virulence in the zygomycete pathogen Mucor. This study on morphogenesis and pathogenicity in the emerging zygomycete fungal pathogens will foster development of anti-mucormycosis targets for novel treatment options.
The animal studies at the Duke University Medical Center were in full compliance with all of the guidelines of the Duke University Medical Center Institutional Animal Care and Use Committee (IACUC) and in full compliance with the United States Animal Welfare Act (Public Law 98–198). The Duke University Medical Center IACUC approved all of the vertebrate studies under protocol number A061-12-03. The studies were conducted in the Division of Laboratory Animal Resources (DLAR) facilities that are accredited by the Association for Assessment and Accreditation of Laboratory Animal Care (AAALAC).
Strains and plasmids used in this study are listed in Table 1. M. circinelloides strains were grown on yeast and dextrose agar (YPD, 10 g/L yeast extract, 20 g/L peptone, 20 g/L dextrose, 2% agar) or yeast peptone glucose (YPG, 3 g/L yeast extract, 10 g/L peptone, 20 g/L glucose, 2% agar, pH = 4.5) media for spore production at room temperature (24°C) or at 30°C in the light. YPD was used for liquid culture with vigorous shaking for aeration at 30°C. For anaerobic/high CO2 conditions, 500 mL flasks were completely filled with liquid YPD media that had been sonicated to eliminate residual O2 gas. The spores of experimental strains (107 spores) were then inoculated on the bottom of the flasks. The flasks were sealed with parafilm and placed in an incubator at 30°C without shaking. Plasmids used in this study were maintained in the Escherichia coli DH5α-T1R strain (Invitrogen, Carlsbad, CA) and manipulated as previously described [90]. All microbial strains were handled under appropriate Biosafety Level 2 conditions (BSL2). All chemicals for media, buffer, and supplements were purchased from Difco Laboratories (Sparks, MD) or Sigma-Aldrich (Saint Louis, MO) unless otherwise stated.
Liquid YPD or YPD agar plates were supplemented with FK506 (1 µg/mL) or cyclosporine A (100 µg/mL). Various concentrations of FK506 were also used (0.025 to 1 µg/mL). FK506 and CsA were obtained from Astellas Pharma Inc. (North Brook, IL, USA) and LC Laboratories (Woburn, MA), respectively.
Three available zygomycete genomes were examined, including M. circinelloides (http://genome.jgi-psf.org/Mucci1/Mucci1.home.html), R. delemar (http://www.broadinstitute.org/annotation/genome/rhizopus_oryzae/MultiHome.html), and P. blakesleeanus (http://genome.jgi.doe.gov/Phybl2/Phybl2.home.html). To assess the calcineurin components, including the calcineurin catalytic A and regulatory B subunits, FK506 binding protein 12 (FKBP12), cyclophilin A, calmodulin, and calmodulin-dependent protein kinase (CAM kinase), we applied a BLAST search by using S. cerevisiae and C. neoformans calcineurin pathway proteins. Proteins in the zygomycete genomes identified via a BLAST with S. cerevisiae and C. neoformans proteins were re-BLASTED against the S. cerevisiae and C. neoformans genomes. When the proteins in the zygomycete genomes returned the S. cerevisiae and C. neoformans counterpart proteins as the best hits, we defined them as the calcineurin components in the zygomycete genomes.
Fungal mass was lyophilized, and total RNA was extracted from the lyophilized fungal mass using the Ambion RiboPure-Yeast Kit (Invitrogen, Carlsbad, CA, USA). cDNA was synthesized from the extracted total RNA using the AffinityScript qPCR cDNA Synthesis Kit (Agilent Technologies, Santa Clara, CA, USA). The ORFs of the three cna genes were each amplified with a pair of primers: JOHE23748 and JOHE23750 for cnaA; JOHE23751 and JOHE23753 for cnaB; JOHE23754 and JOHE23756 for cnaC. Primers used in this study are listed in Table S1. PCR products obtained were directly sequenced with the primers above and additionally with JOHE20751 for cnaA, JOHE23765 and JOHE23766 for cnaB, and JOHE23767 and JOHE23768 for cnaC.
Deduced amino acid sequences of the cna genes were aligned with CLUSTALW, along with the calcineurin catalytic A subunit protein sequences from R. delemar, P. blakesleeanus, S. cerevisiae, Aspergillus nidulans, C. neoformans, and Homo sapiens. Phylogenies were constructed using a PhyML 3.0 software [91], which allowed phylogenies to be inferred and levels of support ascertained. Phylogenetic trees were drawn with the Dendroscope program [92] with aligned sequences.
Calcineurin mutants MSL11 (CNBR-1), MSL12 (CNBR-2), MSL13 (CNAA-1), MSL14 (CNAA-2), MSL15 (CNBR-1), and MSL16 (CNBR-3) were isolated by placing a drop with M. circinelloides spores (R7B, MU406, MU407, MU420, MU416, and MU402 strains, respectively) (Table 1) in the center of YPD solid medium containing FK506 (1 µg/mL), as described before [65]. Between 5 days and 2 or 3 weeks of growth, some of the yeast colonies showed resistant mycelial outgrowths, from which spores were collected and passaged on FK506-containing solid medium until homokaryotic mycelial growth was observed. Spore suspensions were collected and tested for rapamycin resistance on YPD solid medium containing rapamycin (100 µg/mL). Total genomic DNA was isolated from the mutants and the genes encoding fkbA, cnbR, cnaA, cnaB, and cnaC were sequenced.
Wild-type and mutant allele sequences obtained for cnbR, cnaA, cnaB, and cnaC were deposited in the GenBank under the following accession numbers, KC460400 (cnbR), KC460401 (cnaA), KC460402 (cnaB), KC460403 (cnaC), KC503248 (CNBR-1), KC503249 (CNBR-2), KC512815 (CNBR-3), KC503250 (CNAA-1), and KC503251 (CNAA-2).
To disrupt the cnbR gene, we constructed a disruption allele containing the pyrG gene flanked by ∼1 kb of the 5′ and 3′ untranslated regions of the cnbR gene via overlap PCR. The 5′ region was amplified with primers JOHE22226 and JOHE22227, and the 3′ region was amplified with primers JOHE22230 and JOHE22231 from the genomic DNA of wild-type M. circinelloides strain CBS277.49 (the genome sequence reference strain). The pyrG gene was also amplified with primers JOHE22228 and JOHE22229 from CBS277.49 genomic DNA. The three fragments were then subjected to an overlap PCR with nested primers JOHE22236 and JOHE22237 to isolate a disruption allele as described [93]. The cnbRΔ::pyrG cassette was purified and cloned into plasmid pCR2.1-TOPO following the manufacturer's instructions (Invitrogen, Carlsbad, CA). The resulting plasmid, pCnbR-KO, was used to generate a greater amount of disruption construct DNA via PCR.
Strain MU402 (leuA−, pyrG−) was transformed with the obtained disruption cassette to disrupt the target cnbR gene, and transformation was carried out as described previously [94]. In brief, protoplasts (confirmed by lysis in water) were obtained from 2.5×108 germinated spores of strain MU402 (pyrG−, leuA−) [95] by incubation with 0.03 unit/mL chitosanase RD (US Biological, Marblehead, MA) and 1 mg/mL lysing enzymes (L-1412; Sigma-Aldrich, Saint Louis, MO) at 30°C for 90 min. Protoplasts were incubated with 5 µg of DNA and electroporation was performed in 0.2 cm cuvettes (pulse at 0.8 kV, 25 µF, and 400 Ohms). pyrG+ transformants were selected on MMC medium (1% casamino acids, 0.05% yeast nitrogen base without amino acids and ammonium sulfate, 2% glucose) pH 3.2, supplemented with 0.5 M sorbitol [95]. Mucor is coenocytic (aseptate), and therefore, an intensive selection procedure is required to obtain progeny with a homozygous karyotype. Thirty-nine transformants were obtained from eight independent transformations. The transformants were then grown in MMC selective medium for several vegetative cycles to increase the proportion of transformed nuclei. Spores of the transformants were spread onto MMC medium with uridine (200 mg/L) and MMC medium without uridine to determine the ratio between pyrG+ and pyrG− progeny. Once the ratio reached 1∶1, genomic DNA from the transformants was extracted and used as a template for 5′ and 3′ junction PCRs to identify transformants in which homologous replacement of the cnbR locus with the pyrG allele had occurred. We found two independent transformants that contained the cnbRΔ::pyrG allele. Then, the two transformants underwent intensive vegetative selection cycles until all progeny were pyrG+, and 5′ and 3′ junction PCRs then allowed us to confirm the disruption of the cnbR gene (Figure S8). JOHE22226 and JOHE37644 were used to amplify the 5′ junction and JOHE22231 and JOHE37645 to amplify the 3′ junction. Primers JOHE22226 and JOHE22231 are specific to the 5′- and 3′ regions outside of and upstream and downstream of the disruption cassette, respectively and were used for ORF spanning PCR. Primers JOHE37644 and JOHE37645 are specific to the marker gene pyrG. Primers For Southern blot analysis, primers JOHE22237 and JOHE39531 were used to amplify the probe (Figure S9).
Disruption of the cnaA gene was performed with the primers listed in Table S1. Thirty-seven transformants were obtained from eight independent transformations, and then screened for homologous recombination as described above. The cnaA gene disruption was confirmed by the same procedure described above. Primers JOHE26840 and JOHE37644 were used to amplify the 5′ junction and primers JOHE26845 and JOHE37645 to amplify the 3′ junction (Figure S11). Primers JOHE26840 and JOHE26845 were also used for ORF spanning PCR. The plasmid containing the cnaA disruption cassette in pCR2.1-TOPO was designated pAL1-1 (Table 1). Primers JOHE26847 and JOHE39530 were used to amplify the probe for Southern blot analysis (Figure S12).
Wild-type and the cnaAΔ mutant spores were obtained from cultures on solid YPG media that were incubated for 4 days in the light. Sterile deionized water was then added to the media and spore suspensions were collected by scraping with a spreader. For the wild-type yeast, spores were inoculated at the bottom of a 500 mL flask completely filled with liquid YPD media and the flask was incubated without shaking to render the growth conditions anaerobic. The cnbRΔ mutant yeast were grown on YPD agar (pH = 6.5) for one day at 30°C and yeast colonies were scraped and resuspended in sterile deionized water. All inocula were washed twice with sterile PBS and quantified with a hemocytometer. Inocula in PBS were injected into wax moth larvae, and the survival rate of the host was monitored. The significance of the mortality rate data was evaluated by using Kaplan–Meier survival curves with the PRISM statistics software (GraphRad Software, Inc.).
Each strain of Mucor, R. delemar RA99-880, and C. neoformans H99 was grown in appropriate conditions. Crude protein extracts were obtained with radioimmunoprecipitation (RIPA) lysis buffer (Santa Cruz Biotechnology, Inc., Santa Cruz, CA) containing PMSF, sodium orthovanadate (a phosphatase inhibitor), and protease inhibitor cocktail solutions as recommended by the manufacturer. In brief, total fungal cells were briefly washed with PBS and soaked in the RIPA buffer with microbeads (426 to 600 µm, Sigma-Aldrich, Saint Louis, MO). Each sample was vigorously shaken in a bead beater 5 times for 1 minute each, rested for 5 minutes at 4°C, and centrifuged for 15 minutes at 12,000 rpm. Supernatants containing crude protein extracts were transferred into a new tube. Total protein was quantified with Bradford solution (BioRad, Hercules, CA) using a standard curve with bovine serum albumin. Protein kinase A activity of the crude protein extracts was measured with the SignaTECT cAMP-Dependent Protein Kinase (PKA) Assay System (Promega, Madison, WI) following the manufacturer's instructions. Three independent experiments were performed with consistent results.
To determine the level of expression of the cnaA, cnaB, and cnaC genes during hyphal and yeast growth, quantitative RT-PCR assays were performed. Based on the ORF sequences, we designed pairs of specific RT-PCR primers across two exons of each cna gene: JOHE24019 and JOHE24020 for cnaA; JOHE24021 and JOHE24022 for cnaB; JOHE24023 and JOHE24024 for cnaC and JOHE24075 and JOHE24076 for actin. qRT-PCR was performed using the Brilliant SYBR Green QPCR Master Mix (Agilent Technologies), and the expression of the actin gene was used for normalization. The expression levels of the cna genes were evaluated by ct values from RT-PCR. Multiple independent experiments were performed with concordant results.
Spores and hyphae were observed with a Zeiss Axioskop 2 Plus equipped with an AxioCam MRm camera (Carl Zeiss Inc., Thornwood, NY). To analyze nuclei, spores were fixed with 3.7% formaldehyde in 50 mM potassium phosphate buffer, pH 7.0, containing 0.2% Triton X-100 (v/v). The specimens were then mounted on a glass slide with ProLong Gold antifade reagent with DAPI (Invitrogen, Carlsbad, CA).
Time lapse analyses for the germination of wild-type and cnaA mutants were performed by using a Zeiss Axio Observer Z1 microscope system (Carl Zeiss Inc., Thornwood, NY) with an Opto-electronically motorized XY stage, Pecon XL S1 incubator, and Coolsnap ES2 high-resolution CCD camera, which are housed in the Duke University Light Microscopy Core Facility (LMCF). The images were sequentially obtained every 30 seconds and combined as a movie (30 frames per second) by using MetaMorph 7.6.5 (Molecular devices Inc., Sunnyvale, CA).
|
10.1371/journal.pcbi.1004802 | Dynamical Allocation of Cellular Resources as an Optimal Control Problem: Novel Insights into Microbial Growth Strategies | Microbial physiology exhibits growth laws that relate the macromolecular composition of the cell to the growth rate. Recent work has shown that these empirical regularities can be derived from coarse-grained models of resource allocation. While these studies focus on steady-state growth, such conditions are rarely found in natural habitats, where microorganisms are continually challenged by environmental fluctuations. The aim of this paper is to extend the study of microbial growth strategies to dynamical environments, using a self-replicator model. We formulate dynamical growth maximization as an optimal control problem that can be solved using Pontryagin’s Maximum Principle. We compare this theoretical gold standard with different possible implementations of growth control in bacterial cells. We find that simple control strategies enabling growth-rate maximization at steady state are suboptimal for transitions from one growth regime to another, for example when shifting bacterial cells to a medium supporting a higher growth rate. A near-optimal control strategy in dynamical conditions is shown to require information on several, rather than a single physiological variable. Interestingly, this strategy has structural analogies with the regulation of ribosomal protein synthesis by ppGpp in the enterobacterium Escherichia coli. It involves sensing a mismatch between precursor and ribosome concentrations, as well as the adjustment of ribosome synthesis in a switch-like manner. Our results show how the capability of regulatory systems to integrate information about several physiological variables is critical for optimizing growth in a changing environment.
| Microbial growth is the process by which cells sustain and reproduce themselves from available matter and energy. Strategies enabling microorganisms to optimize their growth rate have been extensively studied, but mostly in stable environments. Here, we build a coarse-grained model of microbial growth and use methods from optimal control theory to determine a resource allocation scheme that would lead to maximal biomass accumulation when the cells are dynamically shifted from one growth medium to another. We compare this optimal solution with several cellular implementations of growth control, based on the capacity of the cell to sense different physiological variables. We find that strategies maximizing growth in steady-state conditions perform quite differently in dynamical conditions. Moreover, the control strategy with performance close to the theoretical maximum exploits information of more than one physiological variable, suggesting that optimization of microbial growth in dynamical rather than steady environments requires broader sensory capacities. Interestingly, the ppGpp alarmone system in the enterobacterium Escherichia coli, known to play an important role in growth control, has structural similarities with the control strategy approaching the theoretical maximum. It senses a discrepancy between the concentrations of precursors and ribosomes, and adjusts ribosome synthesis in an on-off fashion. This suggests that E. coli is adapted for environments with intermittent, rapid changes in nutrient availability.
| Microorganisms adapt their physiology to changes in nutrient availability in the environment. This involves changes in the expression of a large number of genes, encoding proteins with a variety of cellular functions, such as transporters for the uptake of nutrients, enzymes for the conversion of nutrients to energy and building blocks for macromolecules, the components of the transcriptional and translational machinery, and transcription factors to preferentially direct RNA polymerase to specific promoters [1, 2]. Fundamentally, the reorganization of gene expression in response to changes in environmental conditions is a resource allocation problem. It poses the question how microorganisms redistribute their protein synthesis capacity over different cellular functions when constrained by the changing environment.
The mechanisms responsible for resource allocation in microbial cells are usually assumed to have been optimized through evolution, so as to maximize the offspring of cells in their natural environment. How this general principle manifests itself on the level of cellular physiology is not straightforward though. Many studies have reasoned that growth-rate maximization provides a selective advantage to microorganisms, because it allows competitors to be outgrown when resources are scarce. Others have shown, however, that appropriate optimization criteria will depend on the structure of the environment and the ecosystem, as well as on the molecular properties of metabolic pathways [3–7]. For instance, in environments without competition for a shared resource, maximization of growth yield rather than growth rate is expected to provide a selective advantage. Although what counts as optimal is thus context-dependent, growth and evolution experiments in Escherichia coli have shown that in certain conditions bacterial metabolism is indeed geared towards growth-rate maximization [8–10].
For this reason, growth-rate maximization is a central hypothesis in a number of recent theoretical studies of resource allocation using coarse-grained models of the cell [11–13]. The models deliberately reduce the molecular complexity of regulatory networks so as to focus on generic explanatory principles [14]. Along these lines, Molenaar et al. developed a series of simple models of the microbial cell, taking into account that growth requires the synthesis of proteins playing a role in metabolism (transporters, enzymes) and gene expression (ribosomes), in varying proportions. Allocation parameters that maximize the growth rate were shown to account, at least in a qualitative way, for the variation of the amount of ribosomal protein as a fraction of total protein in different growth media, and for the occurrence of overflow metabolism above certain growth rates [11]. Using another coarse-grained model of the cell, centered on amino acid supply (metabolism) and demand (protein synthesis), Scott et al. derived empirical growth laws with linear relations between the ribosomal protein fraction and the growth rate, in conditions where the nutrient supply or demand are altered [12, 13]. In their model, maximization of growth rate requires maximization of amino acid flux and is achieved for a specific, unique value of the ribosomal protein fraction. Based on a structurally similar model, Maitra and Dill related optimal resource allocation to the basic constants of the metabolic and gene expression machinery, in particular energy efficiency [15].
The assumption of growth-rate maximization may lead to correct predictions in some situations, but ignores the regulatory mechanisms achieving resource allocation and therefore cannot provide a causal explanation of cellular behavior [16]. Several studies have used coarse-grained models to understand which control strategies microorganisms employ to achieve (optimal) resource allocation [13, 17, 18]. Scott et al. have shown that a robust feedforward control strategy, based on the sensing of the amino acid pool size and the corresponding adjustment of the fraction of ribosomes producing ribosomal proteins, allows the ribosomal protein fraction to be maintained close to its optimal value under a variety of growth conditions [13]. The authors suggest that this control strategy involves the signalling molecule ppGpp, in agreement with conclusions drawn from a recent kinetic model of the regulatory mechanisms achieving optimal adjustment of the ribosomal protein fraction [17]. Weiße et al. also developed a coarse-grained model of microbial growth based on resource allocation trade-offs [18]. Without including specific regulatory interactions, the model accounts for the above-mentioned bacterial growth laws, predicts host-circuit interactions in synthetic biology, and relates gene regulation to the nutrient composition of the medium.
The above studies consider resource allocation at steady state, where all intensive variables describing the growing microbial culture, in particular the concentrations of its molecular components, are constant (see [19] for a precise definition of steady-state growth and the closely related notions of balanced and exponential growth). This requires an environment to be stable over a long period of time. Such conditions can be achieved in the laboratory [20], but many microorganisms naturally experience frequently-changing conditions. For example, E. coli can cycle between two distinct habitats, the mammalian intestine and the earth’s surface (water, sediment, soil) [21]. The bacteria transit through different microenvironments in the intestinal system, where they encounter different mixes of sugars [22]. They are even more challenged in the open environment outside the host, with a greatly fluctuating availability of carbon and energy sources and a large variability in temperature, osmolarity, oxygen, and microbial communities [23, 24].
This situation motivates a dynamical perspective on microbial growth and resource allocation [25–28]. However, fundamental results like the growth laws uncovered for steady-state conditions are still lacking. In particular, extending the results reviewed above to dynamical conditions raises the following questions: Are control strategies that maximize steady-state growth also optimal in dynamical environments? If this is not the case, then which alternative strategies would be optimal for such conditions? And finally, how do these strategies compare with the regulatory mechanisms that have actually evolved in microorganisms?
The aim of this study is to address the above fundamental questions in a specific dynamical growth scenario, namely a transition between two steady states following an environmental perturbation. In particular, we consider the upshift of a microbial culture from a medium supporting growth at a low rate to a medium supporting growth at a high rate [28]. We develop a coarse-grained model of the cell, inspired by the self-replicator model of Molenaar et al. [11], and reformulate our questions in the context of optimal control theory [29] to identify control schemes maximizing biomass production over an interval of time, the dynamical equivalent of growth-rate maximization.
We show that Pontryagin’s Maximum Principle suggests that optimal resource allocation after a growth transition is achieved by a bang-bang-singular control law [29], a conjecture confirmed by direct numerical optimization. This optimal solution provides a gold standard against which possible control strategies of the cell can be compared. We consider simple strategies that drive the system to the steady state enabling growth at the maximal rate in the new medium, after the upshift. In a dynamical growth scenario, the strategy sensing the concentration of precursor metabolites emerges as the best candidate, consistent with the analysis of Scott et al. that feedforward activation of the rate of synthesis of ribosomal proteins, involving ppGpp-mediated sensing of the amino acid pool [30–32], is the key regulatory mechanism for growth control. It is possible, however, to define a strategy approaching the theoretical optimum even more closely by exploiting information on both the precursor concentration and the abundance of the gene expression machinery. Interestingly, a thorough analysis of the functioning of the ppGpp system, as described by a kinetic model of the synthesis and degradation of this signalling molecule, suggests similarities between our two-variable control strategy and the regulation of the transcription of ribosomal RNA by ppGpp [17].
The results presented here generalize the analysis of control strategies enabling optimal growth of microorganisms from steady-state to dynamical scenarios. The control strategies are formulated in the context of a coarse-grained model of resource allocation, based on minimal assumptions, that accounts for empirical growth laws at steady state. The analysis shows that during growth transitions, control strategies based on information of a single variable are outperformed by systems measuring several variables. This conclusion agrees with the intuition that, in dynamical environments, there may be an evolutionary pressure towards more elaborate sensory systems. From a methodological point of view, our study illustrates how optimal control theory can provide novel insights into complex biological phenomena [33].
Resource allocation in bacteria involves the distribution of cellular resources (precursor metabolites and energy) over processes supporting maintenance and growth [1]. A simple modelling tool for analyzing resource allocation questions in a precise way are so-called self-replicator models. These models have a long history in various domains of chemistry, biology, physics, and computer science [34], and were recently put to use as an analytical tool in systems biology [11] (see also [35]). We will show that despite their simplicity, which make them tractable for mathematical analysis, self-replicator models are sufficiently expressive to account for empirical observations and make testable predictions.
Bearing in mind that the major constituents of the cell are macromolecules (DNA, RNA, proteins), produced from precursor metabolites, a fundamental resource allocation question is the following: How much of the cellular resources are invested in the making of new macromolecules (gene expression machinery) and how much in performing other functions, in particular producing metabolic enzymes involved in the uptake of nutrients and their conversion to precursor metabolites (metabolic machinery)? In order to address this question, we consider a self-replicating system composed of the gene expression machinery (R) and the metabolic machinery (M). The system, shown schematically in Fig 1, is thus defined by two macroreactions which are conveniently written as:
S ⟶ V M P , P ⟶ V R α R + ( 1 - α ) M . (1)
The first reaction, catalyzed by M, converts external substrates (S) into precursor metabolites (P). The second reaction, catalyzed by R, converts precursors into macromolecules (R and M). The resource allocation parameter α ∈ [0, 1] defines the proportion of precursor mass used for making gene expression machinery as compared to metabolic machinery. We will interchangeably use the symbols M, R, S, and P for the components of the replicators themselves and their total mass [g]. We will denote the rates at which the macroreactions occur by VR and VM [g h-1].
The self-replicator system in Fig 1 is based on a number of simplifying assumptions. First, cell division is not explicitly modelled and replication should therefore be interpreted as the growth of (the mass of) a cell population. This amounts to the assumption that individual cells in a growing populations have the same macromolecular composition. Second, degradation of the macromolecules is ignored. In other words, we assume that macromolecules are stable and that their degradation rates are negligible with respect to the rates of other reactions in the system. Third, we consider only two classes of macromolecules (R and M). In particular, we do not assume that an irreducible mass fraction of the precursors is dedicated to cell maintenance [12]. The system could be easily extended to relax the above assumptions, but this would complicate the analysis of the model and obscure the points we want to make.
In what follows, it will be more convenient to describe the quantities in the system as intracellular concentrations rather than as the total mass in the cell population. To this end, we define the volume Vol [L] of the cell population as follows:
Vol = β ( M + R ) , (2)
with β a conversion constant [L g-1] equal to the inverse of the cytoplasmic density. Dividing each variable M, R, and P by Vol yields the concentrations m, r, and p of metabolic enzymes, ribosomes and other components of the gene expression machinery, and precursor metabolites, respectively [g L-1]. Henceforth, these variables as well as Vol and α will be considered functions of time t [h].
The dynamics of the self-replicator in Fig 1 can be described by the following system of ordinary differential equations (see S1 Text for the derivation):
d p d t = v M ( s , r ) - v R ( p , r ) ( 1 + β p ) , (3) d r d t = v R ( p , r ) ( α ( t ) - β r ) , (4)
where s [g L-1] denote the (extracellular) concentration of substrate. vM(s, r) [g L-1 h-1] and vR(p, r) [g L-1 h-1] denote the precursor synthesis rate and the macromolecule synthesis rate, respectively. The growth rate μ [h-1] of the replicator system is defined as the relative increase of the volume, and can be rewritten with Eqs 3 and 4 as proportional to the macromolecule synthesis rate (S1 Text):
μ = 1 Vol d Vol d t = 1 M + R d ( M + R ) d t = β v R ( p , r ) . (5)
The precursor concentration changes through the joint effect of the precursor synthesis rate vM(⋅), the macromolecule synthesis rate vR(⋅), and the rate of growth dilution (β vR(⋅)p). The change in concentration of ribosomes and other components of the gene expression machinery is the net effect of the ribosome synthesis rate (α(⋅)vR(⋅)) and the rate of growth dilution (β vR(⋅)r). Remark that it is not necessary to add an equation for m because it follows from Eq 2 that r + m = 1/β, and therefore dm/dt = −dr/dt.
We use Michaelis-Menten kinetics to define the synthesis rate of each reaction:
v M ( s , r ) = k M m s K M + s = k M ( 1 / β - r ) s K M + s , (6) v R ( p , r ) = k R r p K R + p , (7)
with rate constants kM, kR [h-1] and half-saturation constants KM, KR [g L-1]. Note that the rate of precursor synthesis is proportional to the concentration of the components of the metabolic machinery, while the macromolecule synthesis rate is proportional to the concentration of the components of the gene expression machinery. These catalytic effects correspond to the dashed arrows in Fig 1. The rate constant kM depends both on the quality of the nutrients in the medium (higher kM for a richer medium) and on the metabolic efficiency of the macroreaction converting the substrate into precursors (higher kM for a more efficient reaction). For convenience, we henceforth assume that the environmental conditions do not change over the time-interval considered, either because s is constant or because s ≫ KM, corresponding to a situation in which the substrate is available in excess. In both cases, eM(s) = kM s/(KM + s) is approximately constant, so that we can write
v M ( r ) = e M ( 1 / β - r ) . (8)
The rate constant kR characterizes the efficiency of the gene expression machinery, depending on the elongation rate of ribosomes, among other things. The ratio p/KR is an indicator of the saturation of the gene expression machinery by precursors.
The system of Eqs 3 and 4 thus has four parameters (eM, kR, KR, β), one of which characterizes the input from the environment (eM). The order of magnitude of the parameters can be inferred from data in the literature, as explained in S2 Text. Below we use the following values for the parameters eM = 3.6 h-1, kR = 3.6 h-1, KR = 1 g L-1, β = 0.003 L g-1 (S1 Table). However, it should be emphasized that the conclusions of this paper do not depend on the exact quantitative values of these parameters.
An interesting property of the model is that it is built on minimal assumptions, basically the two macroreactions and the definition of the volume as proportional to the total mass of macromolecules. Like in [11, 13, 25], these assumptions directly lead to the expression of the growth rate in Eq 5, without additional assumptions.
The nullcline for r is given by r = 0, r = α/β, and p = 0, while the nullcline for p is defined by
r = e M β e M + k R p K R + p ( 1 + β p ) .
The nullclines define a single stable steady state (p*, r*) (Fig 2A and Methods). At this steady state, the growth rate is constant and denoted by μ*. The nullcline for p is defined by the environment eM. The nullcline for r, and thus the location of the steady state with the associated growth rate, are given by α. Fig 2B shows the dependency of the steady-state growth rate μ* on the resource allocation parameter α. As can be seen, μ* is maximal for a specific, unique value of α, which we denote α o p t *. That is, the model predicts that there is a single optimal way to divide the precursor flux over the synthesis of the gene expression machinery and the metabolic machinery. The same result, using a similar model, was obtained by Scott et al. [13]. The self-replicator model is simple enough to derive an algebraic expression for computing α o p t * and the corresponding maximal growth rate μ o p t * (Methods and S1 Text), which will simplify analysis of the system in later sections.
In order to validate the model, we verified that it can account for data on the macromolecular composition of E. coli at steady state [12]. When optimizing α for different values of eM (assuming cells attain maximal growth), the model predicts a relation between α o p t * and μ o p t * (colored dots and black dashed line in Fig 3A) that is quasi-linear for high growth rates. We compared this prediction with the results of experiments where the relation between the growth rate and the mass ratio of total RNA and protein was determined in different growth media (Fig 3B). In the framework of our model, different media correspond to different values of eM, and different total RNA/protein mass ratios to different values of α (up to a conversion factor), allowing a direct comparison of the model predictions in Fig 3A with the data in Fig 3B (see Methods). As can be seen, the model is able to account for the observed quasi-linear relation between the growth rate and the total mass ratio of RNA and protein. Moreover, for realistic values of kR and eM, a good quantitative fit is obtained (Methods and S1 Table).
The data from Scott et al. also reveal a second apparently linear relation between the growth rate and the total RNA/protein mass ratio. This relation is obtained when varying, in the same growth medium, the efficiency of protein synthesis by adding different doses of an inhibitor of translation (chloramphenicol) [12]. Using the model, we computed α o p t * and μ o p t *, for constant environment eM and different values of the efficiency of protein synthesis kR (dashed colored lines in Fig 3A). As can be seen in Fig 3B, the model also captures the second linear relation in the data.
We conclude that the self-replicator model is able to reproduce known observations of resource allocation in bacteria, so-called growth laws [12]. The model is similar to a model recently proposed by Scott et al. [13]. Contrary to the latter model, the translation rate is not assumed to be constant in the self-replicator model, but rather depends on precursor abundance, as proposed by the same authors in [36].
The above analysis of bacterial growth has two major limitations. First, the predictions of optimal resource allocation (the value of α leading to the maximal growth rate) hold at steady state, for a constant environment, whereas most bacteria are not expected to encounter such conditions outside the laboratory. An allocation of resources that is optimal for steady-state growth and constant over time may not be optimal in dynamical growth conditions. Second, while it predicts which value of α is optimal at steady state, the model says nothing about the strategies that could be used to control resource allocation and set α to its optimal value. In other words, how could bacterial cells use sensors of changes in their internal state and the environment to optimally adjust α? In what follows, we will address the above two questions, after having given a precise statement of the problem of optimal resource allocation in a dynamical environment in the next section.
A self-replicator at steady state accumulates biomass according to Vol(0) eμ* t, t ∈ [0, τ], when μ* is the growth rate at steady state. The accumulation of biomass is obviously maximal when the growth rate is maximal (μ * = μ o p t *). In dynamical conditions, the growth rate is not constant and biomass accumulation is described more generally by:
d Vol d t = μ ( t ) Vol .
In other words, when integrating over the time interval [0, τ]:
ln Vol ( τ ) Vol ( 0 ) = ∫ 0 τ μ ( t ) d t . (9)
Since the logarithm is an increasing function, maximizing the biomass produced over [0, τ] requires maximization of the right-hand side of the equation.
In a changing environment, maximization of the integral in Eq 9 will generally require the optimal value of α to be a function of time instead of a specific constant value. This dynamical resource allocation problem can be formulated in a more precise way using concepts from optimal control theory [29]. Let J be the objective function
J ( α ) = ∫ 0 τ μ ( t ) d t = ∫ 0 τ β v R ( p , r ) d t ,
where α : ℝ+ → [0, 1] is a time-dependent function. The time evolution of p and r is determined by the self-replicator model of Eqs 3 and 4, and p and r thus depend on eM and α. Moreover, let U = { α : R + → [ 0 , 1 ] } be the set of admissible controls. The optimal dynamical control problem then consists in finding the time-varying function αopt(t) that maximizes J(α) over the time-interval [0, τ]:
α o p t = arg max α ∈ U J ( α ) . (10)
In what follows, we will simplify the above problem by considering that the environment changes in a step-wise fashion at t = 0, but remains constant over the time-interval [0, τ], that is, eM(t) = eM. More specifically, we focus on the case of a nutrient upshift, corresponding to a step-wise increase of eM. This upshift scenario corresponds to classical experiments in bacterial physiology [37–39], reviewed in [28], and is frequently encountered in the life cycle of a microorganism [1]. Notice that more complex environments can be approximated by a sequence of step-wise nutrient upshifts and downshifts.
Optimal dynamical control problems for two-dimensional nonlinear dynamical systems, like the problem of Eq 10, are generally difficult to solve. However, we will show that the class of functions to which αopt belongs can be identified, and we will use numerical optimization to identify a particular αopt maximizing J.
As a preliminary step, in order to simplify the analysis, the variables in the self-replicator model of Eqs 3 and 4 are made nondimensional, by defining t ^ = k R t, p ^ = β p, and r ^ = β r. This leads to the following ODE system:
d p ^ d t ^ = ( 1 - r ^ ) E M - ( 1 + p ^ ) r ^ p ^ K + p ^ , (11) d r ^ d t ^ = r ^ p ^ K + p ^ ( α ( t ^ ) - r ^ ) , (12)
where K = β KR and EM = eM/kR. The nondimensional growth rate is given by:
μ ^ = μ k R = p ^ K + p ^ r ^ . (13)
Notice that the nondimensionalized system depends on a single parameter K, in addition to the constant environment EM, which functions as an input to the system.
Analysis of the nondimensionalized system allows a number of properties of the solution of the optimal control problem of Eq 10 to be derived (Methods and S3 Text). First, by applying a version of the well-known Pontryagin Maximum Principle [40], we can prove that the optimal solution is obtained for an alternating sequence of α(⋅) = 0 and α(⋅) = 1, possibly ending with an intermediate value of α(⋅), corresponding to the optimal steady state ( p ^ ( t ) , r ^ ( t ) ) = ( p ^ o p t * , r ^ o p t * ), that is, the steady state leading to the optimal growth rate μ ^ o p t * in the post-upshift environment EM. Second, if the optimal solution reaches the optimal steady state for the new environment, then it does so after an infinite number of switches of α(⋅) between 0 and 1. Third, this switching behavior is characterized by a so-called switching curve r ^ = φ ( p ^ ) in the ( p ^ , r ^ )-plane, which passes through ( p ^ o p t * , r ^ o p t * ). The switching curve divides the phase plane into two regions, such that α(⋅) switches to 0 when the system is in the region above φ and to 1 when the system is below φ (black dashed curve in Fig 4A).
In line with these results, we conjecture that the optimal solution consists in a switching transient towards the optimal steady state for the new environment, and remains at this steady state until the next environmental change. Such a solution is known as a bang-bang-singular solution in the control theory literature [29]. Formally, the solution of Eq 10 can be described as
α o p t ( t ^ ) = 0 , if r ^ ( t ^ ) > φ ( p ^ ( t ^ ) ) , 1 , if r ^ ( t ^ ) < φ ( p ^ ( t ^ ) ) , α o p t * , if ( p ^ ( t ^ ) , r ^ ( t ^ ) ) = ( p ^ o p t * , r ^ o p t * ) . (14)
Notice that the optimal solution involves dynamical feedback from the state of the system to the control variable α(⋅), and is therefore an instance of closed-loop optimization [29].
The optimal control problem of Eq 10 was also solved numerically by a direct method using the bocop software [41] (see Methods for details). A time discretization allows the problem to be transformed into a nonlinear optimization problem solved here by interior point techniques. The optimal trajectories obtained numerically confirm our conjecture that the optimal control is bang-bang-singular. An example solution, obtained by numerical optimization is shown in Fig 4. At time t ^ = 0, EM jumps from a low to a high value, corresponding to a nutrient upshift (dashed black line in Fig 4B). The optimal solution αopt consists of a sequence of switches between α = 1, corresponding to maximal synthesis of the gene expression machinery, and α = 0, corresponding to maximal synthesis of the metabolic machinery, until ( p ^ o p t * , r ^ o p t * ) is reached. α is then set to α o p t *, the value leading to the maximum growth rate in the new medium (here 0.5, for EM = 1). The sequence of switches of α in Fig 4B corresponds to successive crossings of the switching curve in Fig 4A. In particular, the switch just after t ^ = 2 corresponds to the first crossing of the switching curve; the subsequent switches accumulate around the steady state and are therefore difficult to identify in the plot.
What is the biological relevance of the bang-bang-singular solution maximizing growth of the bacterial self-replicator? In order to answer this question, we will investigate in the next two sections the different ways in which microorganisms could implement or have been shown to implement feedback growth control by sensing the environment and cellular physiology. Although the idealized solution proposed by optimal control theory will obviously not be found in nature, actual control strategies may produce solutions that are close. The optimal solution can thus be used as a gold standard, a benchmark for comparing actual control strategies.
The control strategies that microbial cells have evolved to bring resource allocation in line with changes in the environment involve a variety of molecular mechanisms [42]. These mechanisms are responsible for sensing the environment and the physiological state of the cell, as well as for adjusting the expression of genes that encode components of the transcriptional and translational machinery, enzymes, transporters, and proteins with other metabolic functions.
In the framework of the self-replicator model of bacterial growth, control strategies take the form of feedback control laws mapping the value of system variables to a value of the control variable α(⋅). In this section, we explore two such strategies, the first exploiting information on the quality and quantity of substrate present in the environment, as reflected in the value of EM, and the second using information on the precursor concentration p ^. The feedback control strategies are graphically displayed in Fig 5, as an extension of the self-replicator of Fig 1. We pose a number of mathematical constraints on the feedback control strategies considered below. First, we require the control laws to be functions of the variables of the self-replicator but not involve derivatives or integrals of these variables. Second, for a constant environment EM, the control strategies must drive the system to a unique stable and non-trivial steady state, enabling a non-zero growth rate. Third, this steady state must equal the optimal steady state for that environment, given by ( p ^ o p t * , r ^ o p t * ).
The first control strategy is defined by the function f : ℝ+ → [0, 1], mapping EM to α:
α = f ( E M ) . (15)
Notice that α is constant because EM is fixed to the value defining the new environment after the upshift. What would be an appropriate choice for f? An advantage of the self-replicator model is that the optimal allocation at steady state can be explicitly formulated as a function of EM (Eq 22 in Methods, with derivation in S1 Text). This function is the unique function satisfying all of the above criteria (S1 Text). S1 Fig plots f and shows that it is conveniently approximated by a Michaelis-Menten function, i.e.,
α ( · ) = E M E M + K m E , (16)
with the dimensionless half-saturation constant KmE. The interest of the approximation is that it demonstrates that the control strategy can be described by a simple and ubiquitous response curve in biochemical kinetics.
As an example of a regulatory system resembling the above control strategy consider the phosphotransferase system responsible for the uptake of glucose, the preferred substrate of E. coli [43]. In the presence of glucose, the EIIAGlc component of the phosphotransferase system is mostly unphosphorylated, since the phosphate groups are used for the conversion of extracellular glucose to intracellular glucose-6-phosphate. When glucose disappears from the medium, however, the glucose uptake rate decreases and, correspondingly, the phosphorylated fraction of EIIAGlc increases. The phosphorylation state of EIIAGlc thus provides an indirect read-out of glucose availability. In response to this signal, a variety of metabolic processes are upregulated or downregulated, notably involving the signalling molecule cAMP which activates the pleiotropic transcription factor Crp [43, 44].
How does the control strategy of Eq 15, which we call a nutrient-only strategy, perform in comparison with the optimal solution derived in the previous section? That is, how much biomass does this strategy produce compared with the maximal amount of biomass that can theoretically be obtained after a nutrient upshift? In order to answer these questions, we simulated the response to a sudden upshift of the self-replicator of Eqs 11 and 12 controlled by the nutrient-only strategy of Eq 15. The results are shown in Fig 6. Panel A shows the trajectory of the controlled self-replicator system and panel D plots the evolution of the amount of biomass as a fraction of the amount of biomass produced by the optimal strategy. While the system does reach the steady state that is optimal for EM, the nutrient-only strategy has poor performance in the transient phase immediately following the nutrient upshift. As can be seen from the solution trajectory in Fig 6A, fixing α to the value that enables optimal growth at steady state leads to a huge transient overshoot of the precursor concentration. The overshoot reveals that resource allocation is initially suboptimal, with too many resources invested in the metabolic machinery at the expense of the gene expression machinery. This causes a transiently suboptimal growth rate, leading to lower biomass accumulation (Eq 9).
One way to avoid the transient precursor imbalance observed in Fig 6A would be to exploit information on the precursor concentration in the control strategy. The second strategy considered here, which we label a precursor-only strategy, does exactly this: it involves a feedback control law g : ℝ+ → [0, 1] mapping p ^ to α:
α = g ( p ^ ) . (17)
Since p ^ will vary during the upshift experiment, α is not constant, contrary to the nutrient-only strategy above. In the Methods section, we present a function g satisfying the requirements listed in the beginning of this section, in particular that the system converge to a stable steady state ensuring maximal growth in the new environment. Moreover, we show that any other choice for g leads to lower biomass production. The function is plotted in S1 Fig, and as shown in the same panel, is conveniently approximated by a Hill function with cooperativity coefficient 2:
α ( · ) = p ^ 2 p ^ 2 + K m p 2 , (18)
where Kmp is a dimensionless half-saturation constant.
While converging to the same steady state, this second strategy, which we will refer to as the precursor-only strategy, performs much better than the nutrient-only strategy after an upshift, as shown in Fig 6. We simulated the response to a nutrient upshift of the self-replicator of Eqs 11 and 12 with the precursor-only strategy of Eq 17. The relative biomass increases by 51% and reaches 94% of the biomass produced by the optimal control strategy (the theoretical maximum). The precursor-only strategy notably avoids the inefficient transient accumulation of precursors directly after the nutrient upshift, by alternatingly investing more resources in gene expression (consumption of precursors) and metabolism (production of precursors). In this respect, the oscillatory time profile of α (Fig 6C) is somewhat reminiscent of the bang-bang-singular control in the solution of the optimal control problem (Fig 4B).
Both strategies, nutrient-only and precursor-only, drive the self-replicator towards the same steady state. Whereas the two strategies are thus indistinguishable when the analysis is restricted to steady state, the precursor-only strategy is shown to perform much better in a dynamical upshift scenario, in the sense that the biomass produced is much closer to that produced by the optimal strategy. Several authors have concluded that control strategies based on precursor sensing are key for maintaining optimal growth at steady state. Scott et al. argue that a strategy similar to the precursor-only approach above allows robust control of amino acid supply and demand, resulting in optimal steady-state growth over a range of nutrient conditions [13]. They associate this strategy with ppGpp-mediated control of the synthesis of ribosomal proteins [30–32]. The signalling molecule ppGpp accumulates in response to an increase in the level of uncharged tRNA, when amino acid concentrations in the cell drop. This causes ribosomes to “stall” and leads to RelA-mediated conversion of GTP to ppGpp, the molecular details of which are still subject of debate [32, 45]. Since ppGpp inhibits the transcription of ribosomal RNAs [46], the concentration of the latter decreases, leading to more inactive ribosomal proteins and, through a well-characterized post-transcriptional autoregulatory mechanism, a lower synthesis rate of ribosomal proteins [31, 47]. Our analysis adds to the above study a novel insight: measuring precursors does not only enable resource allocation control to achieve maximal growth at steady state, but is also a good strategy in a dynamical context.
While the precursor-only strategy is thus seen to lead to good results, Fig 6D shows that there remains room for improvement. It seems reasonable to expect that control strategies exploiting information of not just a single variable, but several variables simultaneously, could further improve the performance of the self-replicator during a growth transition.
In the quest for further improvements, a natural starting-point would be to consider the curve defining the optimal steady states ( p ^ o p t * , r ^ o p t * ) for different environments EM. This curve is defined by a function mapping p ^ * to r ^ *, which is actually the same as the function g introduced in the precursor-only strategy (Methods and S1 Fig), given that at steady state r ^ = α (Eq 12). The curve can be seen as representing an optimal balance between precursors and the gene expression machinery, in the sense that the maximal growth rate attainable for a given precursor concentration p ^ requires a concentration r ^ of ribosomes and other components of the gene expression machinery equal to g ( p ^ ). If either r ^ > g ( p ^ ) or r ^ < g ( p ^ ), the growth rate is suboptimal.
These considerations suggest an intuitive control strategy, namely to avoid an imbalance between p ^ and r ^ at all times, and remain as close as possible to the curve defined by g. In particular, when the gene expression machinery is more abundant than what is optimal given the available precursors (r ^ > g ( p ^ )), its synthesis is switched off (α = 0). Conversely, when r ^ < g ( p ^ ), synthesis of the gene expression machinery is switched on. This strategy thus tries to restore “as quickly as possible” the optimal balance between precursors p ^ and the gene expression machinery r ^, giving rise to a so-called on-off control strategy:
α = h ( p ^ , r ^ ) = 0 , if r ^ > g ( p ^ ) , 1 , if r ^ < g ( p ^ ) , α o p t * if ( p ^ , r ^ ) = ( p ^ o p t * , r ^ o p t * ) . (19)
As shown in the Methods section, the on-off strategy drives the system to a stable steady state ensuring growth at the maximal rate. Notice that, contrary to the strategies discussed in the previous section, the value of α selected by the on-off strategy depends on both p ^ and r ^ (Fig 5). It thus uses more information on the state of the system than the nutrient-only and precursor-only strategies.
Fig 7 shows the performance of the on-off strategy after a nutrient upshift, as compared to the precursor-only strategy. The transition is seen to be nearly perfect, in the sense that 98% of the optimal biomass is produced by the strategy. The time course of α in panel D is very similar to the optimal time course obtained by numerical optimization, shown in Fig 4B, and clearly brings out the bang-bang-singular nature of the solution. These results show that a strategy exploiting complete information on the internal state of the self-replicator can lead to near-optimal performance, outcompeting a strategy that uses partial information on the internal state (precursor abundance only).
Are microbial cells equipped with mechanisms implementing a strategy similar to the on-off strategy? A possible candidate would again be the ppGpp system. A kinetic model of ppGpp metabolism and the regulation of the synthesis of ribosomal proteins was recently presented by Bosdriesz et al. [17]. The model proved capable of accounting for a range of experimental data, including the steady-state concentration of ppGpp as a function of the growth rate [48] and the dynamical response of ppGpp to a nutrient upshift or downshift [49]. A major conclusion of the model is that the steady-state concentration of ppGpp exhibits a strongly ultrasensitive response to deviations of the ribosomal protein fraction from the optimal ribosomal protein fraction at a given growth rate. These deviations from optimality, in turn, lead to a switch-like response of the synthesis rate of ribosomal proteins (Fig. 4 in Bosdriesz et al. [17]).
How does this mechanistic model of ppGpp regulation relate to the on-off strategy presented above? In order to answer this question, we first need to find a correspondence between the variables p and r of our coarse-grained model and the concentrations of molecular species in the kinetic model of Bosdriesz et al. This is rather straightforward to achieve, by equating p to the total amino acid concentration and r to the ribosome concentration. Second, S4 Text shows that by making two simplifying assumptions, ppGpp can be expressed as a function of the total amino acid concentration and the ribosome concentration. In particular, we assume that concentrations of all individual amino acids are equal, and that the concentrations of charged tRNAs and ppGpp evolve fast relative to the dynamics of the amino acid and ribosome concentrations. The third step consists in positing an explicit relation between ppGpp and α, based on the regulatory action of ppGpp on the transcription of ribosomal RNA [46]:
α ( · ) = K I K I + ppGpp ( · ) , (20)
with KI a Michaelis-Menten inhibition constant [μmol L-1] and ppGpp the (time-varying) intracellular concentration of ppGpp [μmol L-1].
The response function for ppGpp thus obtained and evaluated for a range of amino acid and ribosome concentrations is represented in Fig 8, and visually compared with the on-off strategy. As can be seen, the two response surfaces are very similar. In other words, the ultrasensitive response of the synthesis rate of ribosomal proteins to the suboptimal allocation of cellular resources, derived from a model of the molecular mechanisms involved in the synthesis, degradation, and regulatory action of ppGpp [17], implements a control strategy that is close to the optimal predicted by a control-theoretical analysis of the self-replicator. While the role of ppGpp in maintaining optimal resource allocation was already pointed out by Scott et al. and Bosdriesz et al., the latter studies were restricted to optimizing steady-state growth. A major insight from the analysis in this section is that this conclusion seems to carry over to dynamical scenarios as well. Fundamentally, the analysis suggests that the ppGpp system is a likely candidate to fulfill this role because it integrates information on the imbalance between precursor concentration and abundance of the gene expression machinery.
Quantitative growth laws are empirical regularities pointing at fundamental properties of microbial life [50]. Recent work has led to the precise theoretical formulation of growth laws and has shown that they can be derived from basic assumptions on the molecular processes responsible for the assimilation of nutrients and their conversion to biomass [11, 13, 15, 17, 18]. The growth laws are uniquely defined under the hypothesis that microorganisms allocate resources in such a way as to maximize their growth rate. Several of the above-mentioned studies have analyzed feedback control strategies on the molecular level enabling cells to achieve optimal resource allocation in a robust manner. The control strategies exploit information on the physiological state of the cell to adjust the (relative) rate of synthesis of different classes of proteins (ribsomes, metabolic enzymes, …). Whereas the growth laws describe microbial growth at steady state, most microorganisms live in complex, continuously changing environments. Despite some precursory work [25, 26], questions about the dynamics of microbial growth remain largely unanswered: Which resource allocation schemes are optimal in changing environments? Which dynamical control strategies lead to (near-)optimal resource allocation? How do these strategies compare with those actually implemented by microorganisms?
We have addressed the above questions by means of a self-replicator model of microbial growth, which, like other coarse-grained models of bacterial growth [11, 13, 15], is capable of reproducing the growth laws at steady state (Fig 3). A first major contribution of our work is to show that, in the case of a dynamical upshift scenario, optimal production of biomass requires a bang-bang-singular resource allocation scheme (Fig 4). That is, the optimal self-replicator should iteratively allocate all of its resources to the gene expression machinery (bang control input) and the metabolic machinery (another bang control input), until the steady state enabling maximal growth in the post-upshift environment is reached, corresponding to a trade-off in the allocation of resources to the two processes (singular control input).
Bang-bang phenomena are widespread in a variety of life processes. Applications of optimal control theory to reproductive strategies in insects [51], the development of intestinal crypts [52], and the activation of metabolic pathways [53, 54] have led to bang-bang or bang-bang-singular strategies. In optimal control problems, such a solution arises with systems where the differential equations are linear in the control variable (in our case, α(⋅)). Examples of applications that are close to the problem considered here are the control of gene expression for adaptation to environmental changes [25, 55], and the allocation of resources between nutrient uptake and growth in microorganisms [26, 56]. Whereas the former applications focus on minimization of response times, the latter also optimize biomass during a growth transition, using a different model, not derived from first principles as in this study. However, the optimal solution of the corresponding optimal control problem is also bang-bang-singular, thus showing that our conclusions are robust to model variations.
Our second major contribution is the assessment of how different feedback control strategies perform with respect to each other and to the gold standard determined from optimal control theory. We show that the precursor-only and nutrient-only strategies, both of which drive the self-replicator to the steady state with maximal growth rate in a static environment, perform quite differently in a dynamical upshift scenario (Fig 6). While the precursor-only strategy is better than the nutrient-only strategy in a dynamical environment, it is in turn outperformed by a so-called on-off strategy, which achieves a near-perfect growth transition by exploiting information on the imbalance between the precursor concentration and the abundance of the gene expression machinery (Fig 7). The superior performance of the on-off strategy can be intuitively explained by the fact that during a growth transition the two variables are not fully correlated, which means that sensing both instead of either one provides additional information in a dynamical context.
Interestingly, the on-off strategy is based on a feedback control law that very much resembles the response function for ppGpp-mediated regulation of the synthesis of ribosomal RNAs in E. coli [17]. The role of ppGpp in controlling microbial growth has been amply documented [30–32]. For example, Potrykus et al. observed that in cells without ppGpp (ppGpp0 mutants) the RNA/protein mass ratio, a proxy for our resource allocation variable α, does not change with the growth rate, which has led these authors to conclude that ppGpp is the major source of growth-rate control in E. coli [57]. The central importance of ppGpp in the reallocation of gene expression resources in E. coli following changes in nutrient availability has also been mapped with higher resolution, using genome-wide transcriptome studies [58, 59]. In nearly all bacterial species examined so far, ppGpp is known to accumulate in response to an increase in the level of uncharged tRNA [60], although the molecular details of ppGpp metabolism and the range of other functions of the alarmone may greatly vary across species [32, 60, 61]. While it has thus been well-established that regulation by ppGpp is an evolutionary conserved mechanism of growth control in the bacterial cell, our analysis provides a new perspective by suggesting that ppGpp enables optimal reallocation of resources after a growth transition, dynamically maximizing the accumulation of biomass.
The model on which the above results are based is built from first principles by distinguishing two fundamental cellular processes: metabolism (converting nutrients to precursors) and gene expression (converting precursors to the proteins that make up biomass) (Fig 1). Despite its simplicity, our self-replicator model is capable of reproducing the empirical growth laws and of making testable predictions on the time-course profile of the resource allocation variable α and on the concentrations p and r of components of the gene expression machinery and metabolic machinery, respectively (see Fig 8 and below). The model can be easily extended with more details on protein synthesis, central carbon and energy metabolism, stress systems, or cell membranes, but this would make the mathematical analysis of the model dynamics and the optimal control problem more complicated. Notice, however, that the direct numerical approach for solving the optimal control problem remains applicable, even for more fine-grained models (Fig 4, see also [27]).
The comparison of different control strategies during a classical growth transition should be interpreted carefully, in a qualitative rather than quantitative manner. Whereas the differences in performance based on the biomass ratio Vol/Volopt of the control strategies are robust, the absolute numbers for the biomass ratio will depend on details of the growth experiment chosen and the exact parameter values. Another implicit assumption in the analysis of the control strategies is that the costs of their molecular implementation can be neglected. This is not true in general, since every control strategy requires resources to be diverted towards the synthesis of sensory systems and regulatory proteins, with possibly detrimental effects on growth. In other words, a control strategy entails a trade-off between the growth burden of regulation and the growth benefit of the improved capability to adapt to changes in the environment [62, 63]. The analysis of control strategies could be refined by adding a reaction to the self-replicator that models the loss of resources incurred by regulatory strategies. While in general the growth burden of a control strategy requiring information on several aspects of cellular physiology is expected to be higher, notice that a single regulatory system may be capable of sensing more than one variable. For example, we show that ppGpp levels in the cell carry information on both the metabolic and the gene expression state (Fig 8), thus integrating several signals in a cost-efficient manner.
The model predictions for the dynamical adaptation of resource allocation after a nutrient upshift suggest several interesting experimental tests. In particular, the switching profile of the resource allocation variable α is a promising candidate for experimental validation. The most straightforward option would be direct measurement of the synthesis rate of ribosomal proteins, using a translational fusion of a fluorescent reporter with a ribosomal protein [45, 64]. However, a more indirect approach based on the quantification of ppGpp concentrations in the cell or the activity of the ribosomal RNA (rRNA) promoters would also be a possibility. Interestingly, some data are already available in the literature. For instance, Gausing has reviewed data on the synthesis of ribosomal proteins after a nutrient upshift, showing that the synthesis rate goes through “a series of rapid changes” resembling oscillations [65]. Later work attributed this pattern to regulation on the transcriptional level [66]. Friesen et al. observed oscillatory patterns in ppGpp concentrations after a nutrient upshift, with an initial response resembling bang control for an upshift to a particularly rich medium [67]. Murray et al. also present data on the ppGpp concentration after a nutrient upshift [49], but with a lower temporal resolution and no clear oscillatory pattern. All of the above measurements were carried out on the population level, which means that switching patterns may be obscured by desynchronisation of the individual cells. More sophisticated experimental set-ups are necessary for the decisive validation of the model predictions, allowing gene expression in single cells to be followed over time in tightly regulated growth conditions [68, 69]. In addition, the model could be validated on other dynamical scenarios, for example nutrient downshifts [49, 70].
Apart from its interest for fundamental science, resource allocation is also a critical question in biotechnology, where there exists an inherent trade-off between the maximization of yield and productivity [71]. High yield means that most of the substrate is converted to a metabolite, peptide or recombinant protein of interest, but this leads to low productivity if the remaining nutrient influx is insufficient to sustain population growth. Engineered control of resource allocation may help in establishing the right trade-off, the most profitable balance between yield and productivity, in a biotechnological process. Such a trade-off could be attained either in steady-state conditions (the incoming nutrient flux is optimally distributed over growth and production) or in dynamical conditions (alternating utilization of the incoming nutrient flux for growth or production) [72–74]. When extended with heterologous metabolic pathways, the self-replicator models used in this study would provide an adequate in-silico test bed for the rapid screening and comparison of alternative control strategies in bioprocess engineering.
The nondimensional version of the model, given by Eqs 11 and 12, was used for a steady-state analysis of the self-replicator. Eqs 11 and 12 were derived from the original model of Eqs 3 and 4 by means of the following rescalings:
p ^ = β p , r ^ = β r , t ^ = k R t , E M = e M / k R , K = β K R .
As shown in S1 Text, for a constant environment EM and constant resource allocation α, the system has two steady states: a trivial unstable steady state ( p ^ * , r ^ * ) = ( 0 , 1 ), allowing no growth in the absence of precursors, and a steady state with a positive growth rate given by
( p ^ * , r ^ * ) = ( 1 - α ) E M - α + [ ( 1 - α ) E M - α ] 2 + 4 α ( 1 - α ) E M K 2 α , α . (21)
The two eigenvalues of the Jacobian matrix evaluated at ( p ^ * , r ^ * ) are negative (S1 Text), so that this steady state is stable.
The growth rate at steady state, as a function of p ^ * and r ^ *, is given by Eq 13, which we repeat here for clarity:
μ ^ * = p ^ * K + p ^ * r ^ * .
Evaluating d p ^ / d t = 0 at ( p ^ * , r ^ * ) allows r ^ *, and therefore μ ^ *, to be written as a function of p ^ * (S1 Text). Accordingly, we can compute ∂ μ ^ * / ∂ p ^ * and, when setting this partial derivative to 0, determine the maximum growth rate at steady state μ o p t * and the optimal resource allocation α o p t * bringing about this maximal growth rate. As shown in S1 Text, μ o p t * and α o p t * can be written as explicit functions of either the environment EM:
α o p t * = E M + K E M E M + 2 K E M + 1 , μ ^ o p t * = E M E M + 2 K E M + 1 , (22)
or the precursor abundance p ^ o p t *:
α o p t * = p ^ o p t * p ^ o p t * + K K + p ^ o p t * ( 1 + p ^ o p t * ) , μ ^ o p t * = p ^ o p t * 2 p ^ o p t * 2 + 2 K p ^ o p t * + K . (23)
The above equations were used for the derivation of the control strategies (see below).
As can be seen by comparing Fig 3A and 3B, growth-rate maximization in the self-replicator model leads to a good qualitative correspondence with the growth laws. In order to determine if a good quantitative fit of the model with the data from Scott et al. [12] can be obtained, for reasonable parameter values, we estimated eM and kR in Eqs 3 and 4 from the measured RNA/protein mass ratios. At steady state, the RNA/protein mass ratio can be interpreted as proportional to r ^ * (and thus α o p t *), with an unknown (dimensionless) proportionality constant γ (see [12] for details on the use of the RNA/protein mass ratio as a proxy for the ribosomal protein mass fraction):
r ^ * = α o p t * = γ RNA mass protein mass . (24)
Reformulating Eq 22 in terms of the original parameters eM and kR, which have physical dimensions facilitating the biological interpretation of their values, we obtain a straighforward relation between eM, kR, K, α o p t * and μ o p t *:
α o p t * = e M + K e M k R e M + 2 K e M k R + k R , μ o p t * = e M k R e M + 2 K e M k R + k R . (25)
Eqs 24 and 25 were used to estimate values of kR and γ, as well as eM for each of the six growth conditions, from the measurements of the growth rate and the RNA/protein mass ratio. The value K was not estimated from the experimental data, but set to a value inferred from the literature (S1 Text). The optimization process was carried out by means of the differential evolution algorithm of Storn and Price [75]. The results are shown in Fig 3B, while the estimated parameter values are summarized in S1 Table. The parameter values are in very good agreement with order-of-magnitude values determined from the literature (S2 Text and S1 Table).
The optimal control problem of Eq 10 consists in identifying the function αopt(t) that maximizes the integral of the growth rate μ ^ over an interval [0, τ]. In order to solve this problem, we first redefined it over an infinite horizon (i.e., τ → ∞) in order to avoid boundary effects occurring over finite time intervals, in particular the depletion of precursors just before reaching τ. With U = { α : R + → [ 0 , 1 ] } the set of admissible controls, the full optimization problem for the nondimensionalized system is given by
max α ∈ U J ( α ) ≡ ∫ 0 ∞ r ^ ( t ^ ) p ^ ( t ^ ) K + p ^ ( t ^ ) d t ^ . (26)
Since J(α) diverges, we actually consider overtaking optimality: A solution is overtaking optimal if its performance index catches up with the performance index of any other solution ([40], see S3 Text for details).
Necessary conditions on optimal trajectories can be obtained by the Infinite Horizon Maximum Principle [40], an extension of the well-known Pontryagin Maximum Principle. Analysis of the Hamiltonian of the system of Eqs 11 and 12 and the associated adjoint system shows that the optimal trajectory is a concatenation of bang arcs (α(⋅) = 0 or α(⋅) = 1) and possibly a singular arc corresponding to the optimal steady state ( p ^ ( t ) , r ^ ( t ) ) = ( p ^ o p t * , r ^ o p t * ), that is, the steady state leading to the optimal growth rate μ ^ o p t * in the new environment after the upshift (S3 Text). Moreover, from the Kelley condition [76], we can show that if the optimal trajectory has a singular arc, then it must enter this singular arc via a chattering arc, i.e., with an infinite number of switches of α(⋅) between 0 and 1 (S3 Text). The chattering arc is characterized by a switching curve r ^ = φ ( p ^ ) in the ( p ^ , r ^ )-plane, which passes through ( p ^ o p t * , r ^ o p t * ). The switching curve divides the phase plane into two regions, such that α(t) switches to 0 when the system is in the region above φ and to 1 when the system is below φ (S3 Text and Fig 4).
The above results have led to the conjectured optimal solution of Eq 14. In parallel, we numerically solved the problem of Eq 26 by a direct method using the bocop software [41]. A time discretization allows the optimal control problem to be transformed into a nonlinear optimization problem, solved here by interior point techniques. A discretization by a Lobatto IIIC formula (6th order) was used with 4000 time steps, and the relative tolerance for the NLP solver was set to 10−14. The optimal trajectories thus obtained are composed of a chattering arc followed by a steady state corresponding to the singular arc (Fig 4). The switching curve φ ( p ^ ) was computed from numerical simulations with different initial conditions.
As described in the Results section, we are interested in control strategies satisfying the following conditions:
It can be directly verified from the functions f, g, and h defining the nutrient-only, precursor-only, and on-off control strategies (Eqs 15, 17 and 19) that they are indeed static functions of the system variables (or the system input, in the case of the nutrient-only strategy). Here we show that the other two conditions are also satisfied for all three strategies.
Following Eq 15, the nutrient-only strategy is defined by α = f(EM), so that α is constant after the upshift. As shown above and in S1 Text, this means that the system controlled by the nutrient-only strategy has a single nontrivial stable steady state (Condition C2). In addition, in this case the optimal steady state is attained for α o p t * defined as in Eq 22, and the following function f therefore guarantees Condition C3:
f ( E M ) = E M + K E M E M + 2 K E M + 1 . (27)
In S1 Text, it is shown that Eq 27 is the only definition of f satisfying all conditions. S1 Fig shows a plot of f(EM) together with a biologically plausible Michaelis-Menten approximation (Eq 16).
The full specification of the precursor-only strategy demands an expression for the function g in Eq 17. Recall that Eq 23 defines α o p t * in terms of the precursor concentration p ^ o p t *, which leads us to propose the following function g:
g ( p ^ ) = p ^ p ^ + K K + p ^ ( 1 + p ^ ) . (28)
As shown in S1 Text by computing the Jacobian, the system given by Eqs 11, 12 and 28 has a single nontrivial stable steady state for any environment EM (Condition C2). Moreover, Eq 28 guarantees this steady state to be optimal (Condition C3). This can be seen by noting that at steady state, d r ^ / d t = 0 implies r ^ * = g ( p ^ * ) (Eq 12). In order for the self-replicator to attain a maximal growth rate at steady rate, Eq 23 needs to be satified, which is the case for the above choice of the function g. Like for f, Eq 28 is the only choice for g satisfying C1–C3. S1 Fig shows a plot of g ( p ^ ) together with a biologically plausible Hill approximation (Eq 18).
The on-off control strategy is defined in Eq 19 and repeated below:
h ( p ^ , r ^ ) = 0 , if r ^ > g ( p ^ ) , 1 , if r ^ < g ( p ^ ) , α o p t * , if ( p ^ , r ^ ) = ( p ^ o p t * , r ^ o p t * ) . (29)
This strategy drives the system to a single steady state, because the p ^-nullcline crosses the function g ( p ^ ) only once, as shown graphically in Fig 7A. In S1 Text we argue that this steady state is stable, by taking into account so-called sliding modes on the switching curve [77] (Condition C2). Moreover, the steady state coincides with the optimal steady state ( p ^ o p t * , r ^ o p t * ) by construction, so that Condition C3 is satisfied as well. Fig 8A shows a plot of h ( p ^ , r ^ ).
Note that since h(⋅) is discontinuous, numerical instabilities occur during simulations. We therefore used the following continuous approximation of this function:
g ( p ^ ) 100 g ( p ^ ) 100 + r ^ 100 , if r ^ ≠ g ( p ^ ) . (30)
The approximation causes α to take intermediate values (instead of 0 or 1) just before reaching the optimal steady state in Fig 7C. For numerical simulations of the ODE system, we used the CVODE solver [78] from SUNDIALS 2.6.2 [79].
|
10.1371/journal.pgen.1004579 | Disruption of SUMO-Specific Protease 2 Induces Mitochondria Mediated Neurodegeneration | Post-translational modification of proteins by small ubiquitin-related modifier (SUMO) is reversible and highly evolutionarily conserved from yeasts to humans. Unlike ubiquitination with a well-established role in protein degradation, sumoylation may alter protein function, activity, stability and subcellular localization. Members of SUMO-specific protease (SENP) family, capable of SUMO removal, are involved in the reversed conjugation process. Although SUMO-specific proteases are known to reverse sumoylation in many well-defined systems, their importance in mammalian development and pathogenesis remains largely elusive. In patients with neurodegenerative diseases, aberrant accumulation of SUMO-conjugated proteins has been widely described. Several aggregation-prone proteins modulated by SUMO have been implicated in neurodegeneration, but there is no evidence supporting a direct involvement of SUMO modification enzymes in human diseases. Here we show that mice with neural-specific disruption of SENP2 develop movement difficulties which ultimately results in paralysis. The disruption induces neurodegeneration where mitochondrial dynamics is dysregulated. SENP2 regulates Drp1 sumoylation and stability critical for mitochondrial morphogenesis in an isoform-specific manner. Although dispensable for development of neural cell types, this regulatory mechanism is necessary for their survival. Our findings provide a causal link of SUMO modification enzymes to apoptosis of neural cells, suggesting a new pathogenic mechanism for neurodegeneration. Exploring the protective effect of SENP2 on neuronal cell death may uncover important preventive and therapeutic strategies for neurodegenerative diseases.
| Protein modification by SUMO is a reversible and evolutionarily conserved process. Members of the SUMO-specific protease (SENP) family are known to reverse SUMO-conjugation in many defined systems, but their importance in mammalian development and pathogenesis remains largely elusive. Although SUMO-conjugated proteins have been shown to aberrantly accumulate in patients with neurodegeneration, there is no evidence supporting a direct involvement of SUMO modification enzymes in human diseases. This study reveals that disruption of SENP2 causes neurodegeneration through modulation of mitochondrial morphogenesis. Our findings provide a causal link of SUMO modification enzymes to cell survival, suggesting a new pathogenic mechanism for neurodegeneration. Exploring the protective effect of SENP2 on neuronal cell death may uncover important preventive and therapeutic strategies for neurodegenerative diseases.
| Emerging evidence suggests the importance of protein modification by Small Ubiquitin-related Modifier (SUMO) in neural development and function [1]–[3]. Abnormal SUMO modification has been found in several neurodegenerative diseases, characterized by progressive loss or dysfunction of neurons [4]–[6]. Unlike ubiquitin with a well-established role in protein degradation, SUMO is involved in protein trafficking, cell proliferation and survival, as well as ubiquitin-mediated proteolysis [7]–[11]. Covalent conjugation of SUMO to protein substrates, also known as sumoylation, is a reversible process catalyzed by SUMO ligases [12],[13]. The removal of SUMO, also known as desumoylation, is mediated by SUMO proteases [14], [15]. Although these proteins have been shown to reverse sumoylation in various physiological systems, their roles in mammalian development and disease remain largely unknown. SUMO-specific protease 2 (SENP2) is found in three alternatively spliced forms exhibiting differential subcellular localizations [16]. Genetic inactivation of Senp2 reveals its requirement in development of trophoblast stem cell niches and lineages during development of the placenta [17]. Although SENP2 mutants display embryonic defects including brain and heart abnormalities, they are likely associated with placental insufficiency which requires further investigation [17], [18].
Enhanced sumoylation and accumulation of SUMO-conjugated proteins have been widely observed in patients with various neurodegenerative disorders [19]–[22]. Among the most notable ones are polyglutamine disorders, including Huntington's disease (HD) caused by a trinucleotide expansion, and neuronal intranuclear inclusion disease (NIID). The encoded CAG expansions result in production of toxic proteins carrying extended glutamine repeats. In HD, SUMO1 conjugation of the disease protein Huntingtin (Htt) contributes to the disease pathology possibly by stabilizing the toxic Htt [20]. SUMO-modified targets/substrates also accumulate in the nuclear aggregates of NIID, a multisystem neurodegenerative disease characterized by large intranuclear inclusions in neurons of the central and peripheral nervous systems [21]. In autosomal recessive juvenile parkinsonism, the SUMO pathway might affect protein degradation mediated by the disease protein Parkin, an E3-ubiquitin ligase [23]. Targeting the SUMO pathway may offer new strategies for disease prevention and therapy. However, there is no evidence indicating a direct involvement of SUMO modification regulators/enzymes in neurodegenerative disease. Information providing a causal link of SUMO dysregulation to neural cell survival is also very limited.
We previously created a mouse strain carrying a null allele of SENP2 [17]. The knockout of SENP2 led to severe developmental abnormalities in trophoblast stem cell niches and lineages during placentation [17]. Although brain and heart deformities were also detected in the SENP2-null embryos (Figure S1, Maruyama et al., unpublished, and [18]), we speculated these are secondary defects due to placental insufficiencies [17]. To analyze the involvement of SENP2 and the importance of SUMO modification in neural development and disease, we first examined its expression pattern. In situ hybridization detected the presence of SENP2 mRNA in the developing mouse brain at embryonic day 14.5 (E14.5) and postnatal day 0 (P0), P7 and P14 (Figure 1A). SENP2 was expressed in subventricular neural progenitors and differentiated cells of the cerebral cortex (Figure 1A). To definitively assess our speculation on the contribution of placental deficiencies to the embryonic deformities, we took a genetic approach by creating a mouse model deficient for SENP2 during neural development. A new mouse strain carrying a SENP2ΔSUMOFx allele, permitting removal of the protease core domain using the Cre-loxP system, was generated (Figure S2). The presence of Cre caused an in-frame deletion, resulting in production of a SENP2 mutant deficient for the SUMO protease activity. Using EIIa-Cre to remove the protease core domain, we generated a mouse strain carrying SENP2ΔSUMO mutant allele expressing the truncated SENP2 (Figure S3). The SENP2ΔSUMOΔ/Δ embryos were significantly smaller or underdeveloped compared to their SENP2ΔSUMO+/+ and SENP2ΔSUMO+/Δ littermates at E10.5 (Figure S3A–B). Development of all three trophoblast layers was severely impaired in the homozygous mutants (Figure S3C–J). These extraembryonic and embryonic defects are highly reminiscent to the SENP2 nulls [17], suggesting that the protease core domain deletion results in a loss of function mutation. We also were able to obtain mice heterozygous for the deleted allele without any noticeable abnormality, further suggesting that there is no dominant phenotype associated with the mutation.
Next, we generated a SENP2ΔSUMO-Nes model, in which SENP2 is ablated in the neural progenitor cells by Nestin-Cre (Figure S4). At newborn, no obvious defects associated with the deletion could be detected, including neuronal differentiation (Fu and Hsu, unpublished), indicating that SENP2 is not essential for embryonic neural development. The embryonic deformities observed in the SENP2 nulls were attributed to placental insufficiency. However, the SENP2ΔSUMO-Nes mice displayed movement difficulties at P10. They developed paralysis around P16 (Figure 1B and Supplementary Video S1; 100% penetrance, n = 20 SENP2ΔSUMO-Nes mutants collected from 10 litters), and died at the age of 3 weeks. The size of the mutant brains was slightly smaller comparable to the control at P0, but later on exhibited a gradual reduction (Figure 1B). At P14, the mutant brain looked transparent, and was much smaller than the control (Figure 1B; *, p<0.05; **, p<0.01, n = 3). Histology revealed no obvious defects at P0 but severe brain abnormalities at P7 and P14 associated with the SENP2 deficiency (Figure 1C). The cerebral cortex of SENP2ΔSUMO-Nes became significantly thinner and malformed. Other CNS regions, e.g. midbrain, cerebellum, hippocampus and spinal cord were also affected by the mutation although the phenotypes were less severe (Figure S5). The results suggested an essential role of SENP2 in neural development at postnatal, but not prenatal, stages.
The neurodegenerative phenotype of SENP2ΔSUMO-Nes prompted us to examine programmed cell death affected by the mutation. Immunostaining of active Caspase 3 and TUNEL analysis revealed that abnormal apoptosis is, not detectable at P0, but increased at P4 (Caspase 3: 0.46±0.12% in control vs. 1.52±0.33% in mutant) and highly enhanced at P7 (Caspase 3: 0.82±0.08% in control vs. 9.4±0.59% in mutant; TUNEL: 0.91±0.17% in control vs. 12.09±0.87% in mutant) (Figure 2A, p<0.01, ∼700 cells were counted in each of 3 independent experiments, mean ± SEM). The apoptotic abnormality, albeit less severe at this stage, was also observed in other CNS regions (Figure S6). To further elucidate the mechanism underlying the neural cell death of SENP2ΔSUMO-Nes, we examined expression of the activated form of Bak, a proapoptotic effector which promotes programmed cell death through modulation of mitochondrial morphogenesis [24], . In the SENP2ΔSUMO-Nes cerebral cortices, Bak activation is stimulated at P0 (1.39±0.41% in control vs. 3.03±0.17% in mutant) and P7 (3.4±0.36% in control vs. 8.21±0.59% in mutant), suggesting an association of mitochondrial dysfunction with the SENP2 mutation (Figure 2B, p<0.01, ∼700 cells were counted in each of 3 independent experiments, mean ± SEM). Neurons derived from the cerebral cortices of mouse embryonic brains were then cultured in vitro for examination of mitochondrial dynamics. Fluorescent labeling of the mitochondria revealed a more than 2.5-fold increase of neurons containing fragmented, but not tubular/rod-like, mitochondria in the cell body and neurite caused by the mutation (20.8±4.4% in control vs. 55.3±7.8% in mutant) (Figure 2C, p<0.002, ∼200 neurons were counted in each of 3 independent experiments, mean ± SEM). Electron microscopy analysis of the P7 brain sections further identified fragmentation of the mitochondria in the cerebral cortical neurons of SENP2ΔSUMO-Nes (Figure 2D). The mitochondrial cisternae are generally intact although few of them show alterations on the inner membrane. The results thus suggested a protective effect of SENP2 on neuronal cell survival. SENP2 plays an essential role in the regulation of mitochondrial dynamics during postnatal development of CNS.
We then examined whether the SENP2 deficiency causes imbalances of sumoylation, resulting in accumulations of SUMO-conjugated proteins. Immunostaining of SUMO1 showed increased levels of the sumoylated proteins (26.1±1.5% in control vs. 39.4±4.5% in mutant), indicating that SENP2 deficiency induces hyper-sumoylation (Figure 3A, p<0.01, ∼700 cells were counted in each of 3 independent experiments, mean ± SEM). Although SENP2 was shown to regulate the Mdm2-p53 pathway [16], [17], the expression and the activity of p53 and Mdm2 were not altered in these mutants (Fu and Hsu, unpublished). The neural defects caused by the SENP2 deletion most likely were not associated with p53-induced apoptosis, which is a mitochondrial independent event. Examination of protein extracts isolated from the P7 cerebral cortices revealed an elevation of SUMO1 association in the mutants (Figure 3B). The loss of SENP2 activated Bak (Figure 2B), which has been shown to promote sumoylation of Dynamin regulated protein 1 (Drp1) and its association with mitochondria during programmed cell death [24], [25]. Therefore, we tested if Drp1 is affected in the SENP2ΔSUMO-Nes mutants. Not only the stability (1.9-fold), but also SUMO1 association with Drp1 (2.7-fold), was enhanced by the mutation while RanGAP1, a known substrate of SENP2, did not appear to be affected (Figure 3B). We then examined the mitochondrial association of Drp1 in primary neurons derived from the cerebral cortices of mouse embryonic brains. The mutation apparently promoted Drp1 association with the mitochondria (Fig. S7). The results implied that dysregulation of Drp1 may cause mitochondrial defects, leading to the development of neurodegeneration in the SENP2ΔSUMO-Nes mutants.
Drp1 has been implicated in neural degenerative diseases with disruption of mitochondrial dynamics [26], [27]. To test if Drp1 plays a role in this pathogenic process, we investigated its regulation by SENP2. Our previous report showed that three gene products of SENP2 (SENP2, SENP2M and SENP2S), generated by alternative splicing, leading to the use of distinct translation initiation sites, exhibit distinct subcellular localizations and functions [16]. The SENP2, SENP2M and SENP2S isoforms are predominately located to the nucleus, cytoplasmic vesicles and perinuclear region, and cytoplasm, respectively [16]. First, we examined which of these isoforms might be involved in the regulation of Drp1 using a parental cell line and its stably transformed variants, which express high levels of different isoforms [16]. Whole cells or mitochondria only prepared from these cell lines were used to isolate extracts, followed by protein analysis. Overexpression of a HA tagged SUMO1 led to hyper-sumoylation of total as well as the mitochondrial proteins in the parental cells which occurs less effectively in all SENP2 variants (Figure 4A). SUMO1 also promotes total cell, cytoplasmic and mitochondrial accumulation of Drp1, suggesting that its stability is modulated by sumoylation. However, this regulatory process, not affected by SUMO2 and SUMO3, is apparently a SUMO1-specific regulation (Figure S8A, B). Moreover, high levels of SENP2S, but not SENP2 and SENP2M, prevented the SUMO1-induced accumulation of Drp1 to the mitochondria (Figure 4A). SENP2S also decreased the SUMO1-induced accumulation of Drp1 in the cytoplasm. Thus suggests that the Drp1 reduction mediated by SENP2S is caused by protein degradation rather than decreased targeting to the mitochondria (Figure 4A). Immunoprecipitation-immunoblot analysis further showed that the SUMO1-association of endogenous Drp1 is eliminated by SENP2S, but not other isoforms (Figure 4B). Although certain levels of reduction were detected in the SENP2 and SENP2M analyses, these might be attributed to the disruption of cellular compartmentalization in vitro. To further examine the ability of SENP2 to remove SUMO1 from Drp1, we used in vitro reconstitution analysis (Figure 4C). Recombinant enzymes, including Ubc9 and SAE1/2, were first utilized to perform the SUMO1 conjugation of Drp1. The addition of purified SENP2 efficiently was able to reverse this sumoylation process (a ∼3.8-fold decrease), suggesting Drp1 as a direct substrate of SENP2 (Figure 4C). Because of differential subcellular distributions of the SENP2 isoforms (SENP2 in nucleus; SENP2M in Golgi; SENP2S in cytoplasm) [16], their co-localizations with Drp1 were then investigated. Double labeling analysis indicated an extensive co-localization between Drp1 and SENP2S (Figure 4D). Using a proximity ligation assay examining protein-protein association within the cells, we found that SENP2S exhibited an isoform-specific interaction with Drp1 (Figure 4E). The interaction apparently took place in the mitochondria and cytoplasm (Figure 4F). Furthermore, using siRNA specifically knocking down SENP2 to an expression level at ∼17% (Figure 4H), we found that its reduction promotes Drp1 association with the mitochondria (Figure 4G), resulting in a 2.2-fold increase compared to the control (Figure 4H). A mitochondrial protein with higher molecular mass, which is probably the SUMO1-associated Drp1, was also increased in the SENP2 siRNA treated cells. Consistent with our analysis in the primary neuron (Figure 2C), the knockdown of SENP2 also enhanced mitochondrial fragmentation in the cell line (Figure 4I, p<0.01, ∼200 cells were counted in each of 3 independent experiments, mean ± SEM). These results imply an isoform-specific effect of SENP2 on Drp1 stabilization and mitochondrial accumulation through modulation of SUMO1-specific conjugation.
The isoform-specific regulation of Drp1 by SENP2S suggests its potential involvement in modulating mitochondrial dynamics. Using DsRed2-Mito labeling, mitochondrial morphology was examined in HCT116 and HCT116-SENP2S cells. Similar to previous findings [28], overexpression of Drp1 and SUMO1 caused fragmentation of the mitochondria in these cells (Figure 5A–B, A′–B′, I). However, the SUMO1-induced mitochondrial fission was prohibited by high levels of SENP2 (Figure 5C, C′, G, G′, I, p<0.01, ∼100 cells were counted in each of 3 independent experiments, mean ± SEM). This might be attributed to the regulatory effects of SENP2 on Drp1 sumoylation and stability. Therefore, we examined if Drp1 is involved in the SENP2-mediated protection of mitochondrial fragmentation. High levels of Drp1 were able to overcome the protective effect of SENP2 on the SUMO1-induced mitochondrial fission (Figure 5D, D′, H, H′, I, p<0.01, ∼300 cells counted, n = 3, mean ± SEM). In contrast, high levels of SENP2S did not seem to affect the Drp1-induced mitochondrial fission, suggesting that Drp1 acts downstream of SENP2S in the regulatory pathway. These results not only indicated a role of SENP2 in controlling mitochondrial dynamics but also suggested that SENP2 exerts its effects through modulation of Drp1.
This study demonstrates that SENP2 controls the SUMO1-mediated modification of Drp1 essential for the regulation of mitochondrial dynamics. Targeted disruption of SENP2 induces neurodegeneration through promotion of Drp1 sumoylation and mitochondrial fragmentation. Impaired desumoylation results in neural cell death suggesting a new pathogenic mechanism for neurodegenerative diseases. Dysregulation of several aggregation-prone proteins which are sumoylation substrates have been implicated in neurodegeneration [19], [20], [22], [29], [30]. However, there is no evidence showing a direct involvement of SUMO modification enzymes in human diseases. Our findings suggest that enhanced sumoylation may also be attributed to mutations in the SUMO regulators in addition to the substrates. A balanced sumoylation is pivotal for neuronal cell survival. Hyper-sumoylation resulting from stimulation of SUMO ligases or disruption of SUMO proteases can lead to neural cell death. Our findings imply that targeting the SUMO protease may correct an imbalance of sumoylation and desumoylation. The SENP2ΔSUMO-Nes mice might have potential in modeling human diseases associated with the SUMO pathway.
An association of the SUMO pathway with the regulation of mitochondrial dynamics has also been demonstrated in this study. Mitochondrial dysfunction has a strong association with neurodegenerative diseases [31]–[33]. Mitochondria possess a highly dynamic nature, undergoing frequent fusion and fission [34]. Due to large energy demands and long extended processes of the neurons, they are particularly sensitive and vulnerable to mitochondrial abnormalities. Enhanced mitochondrial fission induces apoptosis during neurodegeneration [31]–[33]. Mitochondrial dynamics is regulated by the GTPase dynamin-related protein Drp1, whose function is modulated by SUMO modification. In cells, overexpression of SUMO1 prevents Drp1 degradation, resulting in its stabilization and activation [35]. The SUMO1-induced Drp1 promotes mitochondrial fission which can be altered by manipulating the SENP activity [36], [37]. Data presented in this study strongly suggest that SENP2 is the physiological enzyme essential for this regulation. SENP2 controls mitochondrial dynamics through modulation of Drp1 in neural development and disease. Furthermore, Drp1 regulation by the SUMO pathway is causally linked to neural degeneration.
The SENP2 deficiency causes cell survival issues through increases in mitochondrial fission, leading to the development of neurodegeneration. As Drp1 appears to be a direct substrate of SENP2, dysregulation of mitochondrial dynamics is likely the primary cause of defects induced by the SENP2 disruption. Further study of mice with aberrant expression of Drp1 in the neural cells promises new insight into this regulatory mechanism. It remains possible that the aberrant mitochondrial phenotype is one of the main causes, which acts parallel with another cellular abnormality or is a consequence of other cellular abnormalities, e.g. failure of neural connection. Therefore, it is interesting to test if prevention of mitochondrial apoptosis can alleviate the defects caused by SENP2 deficiency. Further examination on the role of mitochondria dynamics promises new insight into the SENP2-mediated neuronal cell death.
The involvement of SENP2 in neural development and degeneration opens new opportunities to develop therapeutic targets in the SUMO pathway. As sumoylation has been shown to counter against ubiquitination, manipulation of the SUMO pathway may also alter the ubiquitination-mediated degradation for the prevention and treatment of neurological disorders. Although SENP2 may have a general effect on the neurons, it remains possible that a specific subtype is more sensitive to the loss of SENP2. In the SENP2 mutants, we identify different degrees of neurodegeneration in the cerebral cortex, hippocampus, cerebellum and spinal cord with the cerebral cortex being most severely affected. A disruption of SENP2 in a specific neuronal subtype may further divulge its role in neurodegenerative diseases. Testing the protective role of SENP2 in neural cell survival in disease conditions is also likely to gain a knowledge base of neurodegenerative diseases, leading to new therapeutic strategies.
The pCS2-SENP2, pHASUMO1, pGEX-4T-SAE1/2, pGEX-2T-Ubc9, pCS2-SENP2, pCS2-SENP2M and pCS2-SENP2S DNA plasmids were described previously [16]. The pGEX-2T-Drp1 clone was generated by inserting a DNA fragment encoding Drp1 into the pGEX-2T vector (GE HealthCare). The SRa-HA-SUMO2, pcDNA3-HA-SUMO3 and pDsRed2-Mito clones were from Addgene or Clontech Laboratories. C3H10T1/2 and HCT116 cells and their derivatives were cultured in DMEM media with 10% fetal bovine serum and antibiotics [16], [38]. Isolation, culture and differentiation of primary neural progenitor cells were performed as described [38], [39].
The SENP2ΔSUMOFx ES cell lines were generated by electroporation of a targeting vector, containing the insertion of an orphan loxP site in intron 15 and another loxP site and a pgk-neo cassette flanked by two FRT sites in intron 16, into CSL3 ES cells [17], [40], [41]. Twenty mouse ES cell clones heterozygous for the targeted allele were obtained by homologous recombination (targeting efficiency: 23/112). Two independent clones were injected into blastocysts to generate chimeras which were bred to obtain mice carrying the targeted allele. These mice were then crossed with the R26Flp mice to remove the pgk-neo cassette to obtain the SENP2ΔSUMOFx mouse strain. Mice were genotyped by PCR analysis using primers (5′-TCTCACTTGAAACCGTAGGGACC-3′ and 5′-GAAGGAAGGACTGGAGGAGAGAAG-3′) to identify the 5′ loxP locus, primers (5′-TTGTCAGAAGCAGTGTCCTGCG-3′ and 5′-GACTGGGAAGATATGAACTCGGC-3′) to identify the 3′ loxP locus. The deleted allele was identified using primers (5′-TCTCACTTGAAACCGTAGGGACC-3′ and 5′-GACTGGGAAGATATGAACTCGGC-3′). The PCR was performed by denaturation at 95°C for 5 min and 35 cycles of amplification (95°C for 30 s, 67°C for 30 s, and 72°C for 60 s), followed by a 7-min extention at 72°C. The SENP2lacZ and Nestin-Cre mouse strains and genotyping methods were reported previously [17], [40]. To generate the SENP2ΔSUMO mouse strain expressing a deficient protein, mice carrying the SENP2ΔSUMOFx allele were crossed with EIIa-Cre transgenic mice to delete the protease core domain in the germ cells [41]. To examine the production of SENP2 transcript, the reverse transcription products were subject to PCR amplifications using primers 5′-CAGTCTCTACAATGCTGCC-3′ and 5′-CTGTCACTCTGATCTTTGG-3′ (exons 3–5), primers 5′-GTGAGCTCATGAGTTCTGG-3′ and 5′-GTCGCTCCAATAACTTTCG-3′ (exons 5–7), primers 5′-GGAGGAGCAGAATCATGG-3′ and 5′-CTCAAAATCTCATCTGGTGG-3′(exons 8–11) and primers 5′-AGGTACATTGGAGCCTGGTG-3′ and 5′-AGCAACTGCTGGTGAAGGAT-3′ (exons 13–17). The PCR reaction was performed by denaturation at 94°C for 5 min and 30 cycles of amplification (94°C for 30 s, 53°C for 30 s, and 72°C for 45 s), followed by a 7-min extension at 72°C. Care and use of experimental animals described in this work were approved by and comply with guidelines and policies of the University of Committee on Animal Resources at the University of Rochester.
Samples were fixed, paraffin embedded, sectioned and stained with hematoxylin/eosin for histological evaluation [17],[42]. The in situ hybridization was performed as described [17], [38], [43], [44]. In brief, sections were incubated with the digoxygenin labeled RNA probes generated by in vitro transcription [17], [43], followed by recognition with an alkaline phosphatase conjugated anti-digoxygenin antibody and visualization with BM-purple [38], [44]. TUNEL staining was performed with ApopTag (Millipore) as described [45], [46]. For electron microscopy, mice were fixed by perfusion with fixative (2% paraformaldehyde, 2.5% Glutaraldehyde, 0.1M sodium cacodylate, 6.8% sucrose). The dissected tissues were then fixed with 1% osmium tetroxide, embedded in EPON/Araldite resin and cut in seventy nm sections, followed by staining with aqueous uranyl acetate and lead citrate and examined using Hitachi 7650 transmission electron microscope.
Proximity ligation assay (PLA) was performed using Duolink In Situ reagents (Duolink Bioscience). Briefly, cells were fixed and incubated with rabbit anti-Drp1 and mouse anti-myc tag antibodies. Two oligonucleotide-labeled anti-rabbit and anti-mouse PLA probes, which bind to each other when they are in close proximity, were then used to generate fluorescent signals. Immunostaining of cells [47] and tissue sections [48]–[50] were performed by incubation with primary antibodies, followed by detection with fluorescence-conjugated or horseradish peroxidase-conjugated secondary antibodies. Images were taken using Zeiss Axio Observer microscope equipped with deconvolution analysis. To determine the mitochondrial morphology, cells were either stained by MitoTracker or transient expression of DsRed2-Mito. For statistical analysis, cells containing MitoTracker or DsRed2-Mito positive mitochondria were counted and scored for tubular/rod-like or fragmented mitochondria. Immunoblot was performed by isolation of protein extracts from mitochondria, cells or tissues using M-PER (Pierce) in the presence of protease inhibitor cocktail, followed by electrophoresis as described [16], [17], [38], [48]. Isolation of mitochondria was performed using a Mitochondria Isolation Kit (Thermo Fisher) according to the manufacture's description. Immunoprecipitation was performed using Pierce Classic IP Kit. Briefly, cells were lysed in buffer containing 0.025M Tris, 0.15M NaCl, 0.001M EDTA, 1% NP-40, 5% glycerol. Approximately 500 µg of protein lysates were mixed with 2 µg of antibodies overnight, followed by incubation with Protein A/G agarose for 1 hour at 4°C. The antibody-bound complex was then incubated with elution buffer for 5 min at 100°C, and collected by centrifugation for SDS-PAGE analysis. Mouse monoclonal antibodies, Actin (Thermo Fisher), HA (Cell Signaling), and myc tag (Santa Cruz); rabbit polyclonal antibodies Bak (Novus Biologicals), caspase-3 (BD Biosciences), Cox IV (Cell Signaling), Drp1 (Novus Biologicals), and SUMO1 (Cell Signaling), were used in these analyses.
Recombinant GST-SAE1/SAE2, GST-Ubc9, GST-Drp1, HA-SUMO1 and myc tagged (MT)-SENP2 proteins expressed in Escherichia coli were affinity purified. The 20-µl reaction buffer containing 50 mM Tris-HCl (pH 7.6), 5 mM magnesium chloride, 10 mM ATP, 1 µg of GST-ASE1/2, 2 µg of GST-Ubc9, 10 µg of GST-HA-SUMO1 and 200 ng of GST-Drp1 with the presence of protease inhibitor cocktail was incubated for 3 h at 37°C. The desumoylation reaction was then carried out in 10 µl of the above sumoylation mixture with the addition of purified MT-SENP2 for overnight at 37°C. The samples were then analyzed by SDS-PAGE and immunoblot analysis of Drp1 and SUMO1.
|
10.1371/journal.pgen.1006214 | Human Management of a Wild Plant Modulates the Evolutionary Dynamics of a Gene Determining Recessive Resistance to Virus Infection | This work analyses the genetic variation and evolutionary patterns of recessive resistance loci involved in matching-allele (MA) host-pathogen interactions, focusing on the pvr2 resistance gene to potyviruses of the wild pepper Capsicum annuum glabriusculum (chiltepin). Chiltepin grows in a variety of wild habitats in Mexico, and its cultivation in home gardens started about 25 years ago. Potyvirus infection of Capsicum plants requires the physical interaction of the viral VPg with the pvr2 product, the translation initiation factor eIF4E1. Mutations impairing this interaction result in resistance, according to the MA model. The diversity of pvr2/eIF4E1 in wild and cultivated chiltepin populations from six biogeographical provinces in Mexico was analysed in 109 full-length coding sequences from 97 plants. Eleven alleles were found, and their interaction with potyvirus VPg in yeast-two-hybrid assays, plus infection assays of plants, identified six resistance alleles. Mapping resistance mutations on a pvr2/eIF4E1 model structure showed that most were around the cap-binding pocket and strongly altered its surface electrostatic potential, suggesting resistance-associated costs due to functional constraints. The pvr2/eIF4E1 phylogeny established that susceptibility was ancestral and resistance was derived. The spatial structure of pvr2/eIF4E1 diversity differed from that of neutral markers, but no evidence of selection for resistance was found in wild populations. In contrast, the resistance alleles were much more frequent, and positive selection stronger, in cultivated chiltepin populations, where diversification of pvr2/eIF4E1 was higher. This analysis of the genetic variation of a recessive resistance gene involved in MA host-pathogen interactions in populations of a wild plant show that evolutionary patterns differ according to the plant habitat, wild or cultivated. It also demonstrates that human management of the plant population has profound effects on the diversity and the evolution of the resistance gene, resulting in the selection of resistance alleles.
| Viruses cause plant diseases, whose severity is considered to increase under plant cultivation. Hence, it is highly relevant to understand the genetics of plant virus resistance, and its variation in wild and cultivated plants. Analyses of plant pathogen resistance have focused on R proteins, which recognize pathogen molecules triggering defenses according to a gene-for-gene interaction. Alternatively, infection may require the interaction of plant and pathogen molecules, mutations impairing this interaction resulting in recessive resistance according to a matching-alleles model. We analyse here the variation of a recessive resistance gene in wild and cultivated populations of a plant, focusing on chiltepin, a wild pepper currently undergoing incipient cultivation in Mexico. The pvr2 gene encodes the translation initiation factor eIF4E1, which must interact with the viral VPg for potyvirus infection. A high genetic variation was found for pvr2/eIF4E1 but, at odds with reports for R genes, there was no evidence for selection of resistance in wild chiltepin populations. However, data supported selection for resistance in cultivated populations, in spite of no phenotypic differences between wild and cultivated plants, and similar potyvirus incidences. Results demonstrate that cultivation has profound effects on the diversity and evolution of resistance.
| Host-parasite interactions often show a high degree of genetic specificity, in that only a subset of parasite genotypes can infect and multiply in each host genotype [1–6]. The outcome (infection vs. resistance) of the host genotype-by-parasite genotype interaction can be integrated into coevolutionary models that differ in the underlying infection matrices [5]. The different proposed models stem from two general ones, the gene-for-gene (GFG) and the matching-alleles (MA) models, which were initially proposed to explain plant-parasite and invertebrate-parasite interactions, respectively [1,7], although evidence indicates that they are not taxonomically restricted [5]. These two models differ widely in their conceptual framework. In the GFG model, there is a hierarchy of resistance alleles in the host and infectivity alleles in the parasite, so that some host resistance alleles are intrinsically better than others, conferring resistance to a larger set of parasite genotypes, and similarly, some parasite alleles determining infectivity are intrinsically better than others, allowing infection of a larger set of host genotypes. In the MA model, there is no hierarchy of resistance (infectivity) alleles, and a particular host genotype is better at resisting a subset of parasite genotypes, and worse at resisting the rest of parasite genotypes, and a parasite genotype is better at infecting a subset of host genotypes, and worse at infecting the rest [1]. Both models also differ in the mechanisms determining host-parasite interactions. In the GFG model, infection occurs when the host genotype does not recognize the parasite genotype, i.e., matches between host and parasite molecules do not occur, while in the MA model successful infection requires molecular matches between host and parasite [5,7]. Hence, the evolution of resistance (infectivity) loci will differ if host-parasite interactions correspond to GFG or MA models. Notably, models predict that costs associated with resistance (infectivity) are required to maintain polymorphisms at resistance (infectivity) loci in the host (parasite) population in GFG interactions, but not in MA ones [1,7–9]. Accordingly, evidence of resistance costs has been reported for GFG interactions [10–12] but, to our knowledge, costs of resistance have not been analysed in MA interactions.
In the last 20 years a big progress has been made in understanding the molecular genetics of plant-parasite, including plant-virus, interactions. Resistance determined by single dominant genes (R genes) is based on host recognition of genotype-specific parasite molecules, being thus compatible with a GFG model, while recessive resistance prevents the matching of the specific host and parasite molecules required for infection, according to a MA model [13–17]. Molecular analyses of the genetic variation of resistance loci in host populations refer almost entirely to R genes determining resistance to cellular pathogens. R genes are considered to have evolved in response to the negative effects of parasite infection on the host fitness [13,18,19], that is, to virulence sensu [20]. Data from different systems show that R genes are hypermutagenic, and suggest that they are frequently under balancing selection [21]. In contrast with the effort devoted to understand the evolution of R genes, the molecular evolution of recessive resistance genes (in fact, susceptibility genes) has been seldom analysed. This gap is especially important in the case of plant-virus interactions, as a large fraction of monogenic resistance of plants to viruses is recessive [15,22]. Thus, the few published reports refer to plant-virus interactions [23–25], and focus on analyses of germplasm collections of crops, rather than on wild plant populations. Human-driven and natural selection on plant genomes can be very different in both cultivated and wild plant populations [26–29]. Thus, a full understanding of the evolutionary dynamics of MA-like plant-parasite interactions requires analyses in wild plant populations, as well as of comparisons between wild and cultivated ones.
Within this scenario, the aim of this work is to analyse the evolutionary patterns of plant recessive resistance loci involved in MA-like interactions, and how these patterns are affected by human management of the host populations. For this, we studied a wild plant that is currently undergoing incipient domestication, the wild pepper or chiltepin, Capsicum annuum var. glabriusculum (Dunal) [30]. Chiltepin is considered as the ancestor of the domesticated pepper C. annuum var. annuum L. [31], an economically important crop that was domesticated in Mesoamerica [32,33]. Chiltepin is a 5–10 year-lived perennial bush distributed from northern Colombia to south western United States. In Mexico, it grows in a variety of environments from the evergreen tropical forests of the Yucatan peninsula and the Gulf of Mexico to the dry deciduous forests of central and western Mexico and to the Sonoran desert [33–35]. Chiltepin plants grow and reproduce during the rainy season and their pungent fruits are consumed by birds, which disperse the seeds [34]. In some regions, fruits are harvested from wild populations for human usage [36] and their high value has led to its very recent cultivation. In the last 25 years, chiltepin cultivation has progressed from home gardens to monocultures in small traditional fields, where they are managed as an annual crop [35]. However, cultivated chiltepin does not show obvious phenotypic differences with wild populations and does not present any of the major traits of pepper domestication syndrome, such as larger, pendulous, non-deciduous fruits of different colours and pungency, flower morphology favoring selfing, and synchronized high germination rates [37]. Genetic variation is high in wild populations and shows a strong spatial structure associated with the biogeographical province of origin, and cultivation results in a significant loss of both genetic diversity and spatial genetic structure [35].
Wild and cultivated chiltepin populations are infected by potyviruses, reaching incidences of up to 42% according to population and year [38]. Thus, this work focuses on the recessive resistance gene pvr2, which has alleles in pepper (Capsicum spp.) conferring recessive resistance to virus species in the genus Potyvirus [39]. Potyviruses are a numerous group of economically important plant viruses with tubular particles encapsidating a single-stranded messenger-sense RNA genome of about 10000 nucleotides (nt), with a virus-encoded protein covalently linked to its 5’ end (VPg) and a polyadenylated tail at its 3’ end [40]. As for most characterized recessive resistance genes to viruses in plants [15,41,42], pvr2 encodes an eukaryotic translation initiation factor, specifically, factor eIF4E1 [39]. Recessive resistance is expressed as immunity (no infection) or decreased virus multiplication [15,43,44], and the various pvr2 resistance alleles reported differ from the susceptibility wild type allele in a small and mainly non-conservative number of amino acid changes [22,23,39,45]. It has been shown that the potyviral VPg interacts directly with pvr2/eIF4E1 in yeast two-hybrid and in vitro binding assays, and the physical interaction between pvr2/eIF4E1 and the virus VPg is required for virus infection [46–49], although the exact role in the potyvirus life cycle of eIF4E-VPg interaction remains a matter of discussion [15,50]. Mutations at pvr2/eIF4E1 that prevent its interaction with the VPg lead to resistance [22,23,51] and mutations at the VPg central domain that restore the pvr2/eIF4E1-VPg interaction allow infection [23]. Thus, the pvr2/eIF4E1-VPg-determined pepper-potyvirus interaction corresponds mechanistically to a MA model.
The pvr2/eIF4E1 allelic diversity has been extensively screened in accessions of C.annuum var. annuum (domestic bell and chili pepper) and, to a lesser extent, in its relatives in the Capsicum genus, reporting one of the largest allelic series of eIF4E, including different susceptibility and resistance alleles to potyviruses [22,23,39,45,52,53]. Genetic variation and functional analyses have provided evidence of selection at pvr2/eIF4E1 for potyvirus resistance [23]. However, these analyses were based for the largest part on accessions of domestic Capsicum species, and included few accessions of wild relatives, so that selection for potyvirus resistance could be associated with selection pressures (including potyvirus infection) specific of, or modulated by, the agroecosystem environment.
The reported incidence of potyviruses infection in chiltepin, together with the high genetic diversity of wild chiltepin populations in a variety of habitats in Mexico, and its incipient domestication, makes the chiltepin-potyvirus interaction a unique system to analyse the genetic variation and the evolutionary patterns of a recessive resistance gene (pvr2/eIF4E1), as well as the potential effects of human management of a host plant and its habitat on the diversity and the evolution of resistance, the two goals of this study. To attain these goals we (i) obtained the nucleotide sequence of pvr2/eIF4E1 in plants collected from wild and cultivated chiltepin populations in different biogeographical provinces of Mexico; (ii) analysed the genetic diversity and structure of pvr2/eIF4E1 according the region of origin and the level of human management; (iii) identified and characterized functionally the different pvr2/eIF4E1 alleles present in chiltepin populations; (iv) analysed the effect of these mutations on pvr2/eIF4E1 structure, (v) evaluated the frequency of potyvirus resistance in the populations and (vi) assessed the incidence of potyvirus infection in chiltepin populations. Our results suggest that resistance probably has associated costs due to functional constraints on pvr2/eIF4E1. Also, in wild chiltepin populations pvr2/eIF4E1 accumulated synonymous changes, and the frequency of resistance alleles was low, while in cultivated populations pvr2/eIF4E1 accumulated non-synonymous changes and the frequency of resistance alleles was significantly higher than in wild populations. These results are evidence of stronger selection for resistance under cultivation, and indicate a role of human management on the evolution of pvr2/eIF4E1.
The coding sequence of the pvr2/eIF4E1 gene has a length of 687nt and encodes a predicted protein of 228 amino acids. The variability of the pvr2/eIF4E1 coding sequence was evaluated in 97 chiltepin plants, 70 from wild and 27 from cultivated populations. These plants were randomly selected from 16 wild and 9 cultivated populations (2–4 plants per population) to represent the diversity of the species in six biogeographical provinces of Mexico (S1 Table). Note that neither the total number of sampled populations nor the ratio of wild to cultivated ones is evenly distributed across biogeographical provinces (S1 Table), which reflects the abundance of chiltepin and the intensity of cultivation [35]. A total of 12.4% of plants were identified as heterozygous at the pvr2/eIF4E1 locus (S2 Table). The proportion of heterozygous plants was similar between wild and cultivated populations (χ2 = 1.3; P = 0.253), the same result being obtained when the plants from cultivated populations were compared with three random subsets of wild plants of the same size (χ2<30; P>0.083). For wild populations, the proportion of heterozygous plants significantly varied between biogeographical provinces (χ2 = 17.9; P = 0.003), which was due to the higher frequency of heterozygotes in AZP: when populations from this province were not included in the analysis, heterozygosity no longer depended on province (χ2 = 2.42; P = 0.659). From these 97 plants, a total of 109 coding sequences of the pvr2/eIF4E1 gene were obtained, 77 from wild and 32 from cultivated populations, and 17 haplotypes were identified at the nucleotide sequence level (Table 1, S1 Table). No significant difference in haplotype richness was observed between wild and cultivated populations over all biogeographical provinces (χ2 = 2.4; P = 0.169) a result that, again, held regardless of sample size (χ2<1.5; P = 0.903).
The genetic diversity of the coding sequence was of 0.00359 ± 0.00115 nucleotide substitutions per site for the whole set of 109 pvr2/eIF4E1 sequences and of 0.00655± 0.00130 for the concatenated sequenced introns (Table 2, see S3 Table for detailed intron diversity). Coding sequence diversity was highest in YUC and SMO, and lowest in SON and CPS (Table 2). Plants grown from seeds of fruits purchased at local markets were also analysed, named as local market populations. People selling the fruits claimed that they had been collected from local wild chiltepin populations, which was confirmed on the basis of the polymorphisms of nine microsatellite markers [35]. To further check if local market populations were derived from fruits harvested from wild populations and, thus, represented their genetic diversity, the genetic differentiation of the pvr2/eIF4E1 coding sequences between wild and local market populations was analysed. The value of the fixation index FST between these two groups of populations was very low and not significantly different from zero (FST(W/LM)<0.001, P = 0.388), showing no genetic differentiation between these two types of populations that, hence, can be clumped into a single class (wild populations). When the genetic diversity was analysed according to habitat, it was found to be 1.4 times higher in the cultivated than in the wild populations (0.00400 vs. 0.00292, Table 2) and the FST value between wild and cultivated populations (FST(habitat) = 0.208, P<0.001) indicated that pvr2/eIF4E1 was genetically structured according to habitat, a result that held when the comparison was between sequences from cultivated plants and random subsets of sequences from wild plants of the same size (χ2>0.107; P<0.001).
The diversity of the pvr2/eIF4E1 coding sequences also showed a strong spatial structure, both at the population level (FST = 0.625, P<10−4 and FST = 0.643, P<10−4, for all or only wild populations, respectively) and at the level of the biogeographical province (FST = 0.522, P<10−4 and FST = 0.584, P<10−4, for all or only wild populations, respectively). More specifically, the chiltepin populations of each biogeographical province were genetically differentiated for the pvr2/eIF4E1 coding sequences, except between CPS/SON, CPS/CPA and CPS/YUC regions (S4 Table). To analyse if this spatial structure followed a model of isolation by distance, a Mantel test was performed between the matrices of genetic and geographical distances among chiltepin wild populations. Data showed that the distribution of the genetic variation of pvr2/eIF4E1 was not correlated with the geographic distance (r = 0.220, P>0.065; S1 Fig).
Table 2 also shows the nucleotide diversity of the pvr2/eIF4E1 coding sequence at synonymous and non-synonymous positions and the dN/dS ratio indicates that pvr2/eIF4E1 is globally under mild negative selection (dN/dS = 0.899). When sequences from wild and cultivated populations were analysed separately, dN/dS values were significantly different. Evidence for negative selection on pvr2/eIF4E1 was stronger in wild populations (dN/dS = 0.605), while it appeared to be under positive selection in cultivated populations (dN/dS = 1.784). However, no site under positive selection was consistently identified by the different methods applied (see Material and Methods), either when all sequences were analysed together or according to habitat, wild or cultivated. Only codon 205 was identified as under positive selection by the REL method. Tajima’s D (DT) showed negative values for pvr2/eIF4E1 (-0.691; -0.868 and -0.519 for all, wild and cultivated populations, respectively) which did not depart from the null hypothesis of neutrality. However, a sliding window analysis of DT across the entire pvr2/eIF4E1 coding sequence revealed regions with strongly positive DT values, around codon 105 for wild populations and between codons 67 and 77 in cultivated populations (Fig 1). Positions 67–77 include those determining potyvirus resistance (see below) and position 105 has a polymorphism exclusive to AZP province.
At the amino acid sequence level, a total of eleven allelic variants were identified based on 10 polymorphic sites, 7 of which were localized in exon 1 (Fig 2). Eight of these alleles had been reported previously within the Capsicum genus [22,23,45], three of them conferring susceptibility to potyviruses (pvr2+, pvr1+ and pvr217) and five conferring resistance (pvr21, pvr22, pvr24, pvr27, pvr29). The eight previously reported alleles represented 87 out of the 109 pvr2/eIF4E1 sequences (i.e. 79.8%) obtained in this study (Fig 2). The 3 new alleles (named pvr223 to pvr225) were characterized by single (pvr223 and pvr224) or double (pvr225) mutations relative to the reference allele pvr2+ (Fig 2). Interestingly, two of the three amino acid changes identified in these new alleles involved new polymorphic sites in comparison with previously reported alleles (codons 40 and 105, Fig 2). The three new alleles were identified in wild populations, allele pvr223 was identified in CPA represented by only one sequence, and alleles pvr224 and pvr225 were identified in AZP, representing 21 out of the 24 sequences (87.5%) from this biogeographical province (Fig 2).
A minimum spanning network (MSN) connecting all pvr2/eIF4E1 alleles in the chiltepin population (Fig 3) showed that the tomato orthologous pot-1+/eIF4E used as outgroup was connected to the pvr1+ allele, which is the root of the network. The MSN also shows that most pvr2/eIF4E1 alleles were connected by steps of just one amino acid substitution. Interestingly, the new allele pvr223 corresponds to one of the most parsimonious putative intermediates described in Moury et al [44] to connect pvr2+ to pvr29. However, one intermediate (labelled “1” in the network), needed to connect pvr223 to pvr29 is still missing, and sequence comparison of all previously described pvr2/eIF4E1 alleles [22,23,45] did not reveal any sequence corresponding to this intermediate. MSN analysis demonstrated that the mutation D205G occurred at least twice in the evolution of pvr2/eIF4E1 in chiltepin.
To test if the new pvr2/eIF4E1 alleles identified in the chiltepin population were not impaired in the essential eIF4E1 function in mRNA translation, we analysed their ability to complement the eIF4E knockout yeast strain JO55 as in Charron et al [23]. Assays showed no growth difference in the selective medium between the yeasts complemented with the fully functional pvr2/eIF4E1 susceptibility allele pvr2+ and the newly described ones (S2 Fig), strongly suggesting that alleles pvr223, pvr224 and pvr225 are functional in translation.
Next, for all the pvr2/eIF4E1 alleles identified in chiltepin populations we analysed the interaction between eIF4E1 and viral VPg, as in the interaction of pepper with Tobacco etch virus (TEV) and Potato virus Y (PVY) there is strong correlation between absence of interaction and resistance. The physical interaction between the 11 pvr2/eIF4E1 proteins encoded and the VPg of the avirulent PVY-LYE84 isolate was analysed using yeast two-hybrid (Y2H) system. Differences of growth on selective medium were observed for yeast transformed with the constructs containing the different pvr2/eIF4E1 proteins and PVY-LEY84 VPg (Fig 4, S3 Fig), which confirmed the interaction pattern reported for the previously characterized alleles, i.e. interactions between the pvr2/eIF4E1-VPg for pvr2+ and pvr1+ susceptibility alleles, and no interaction for the resistance alleles pvr21 to pvr29. The proteins encoded by the pvr217, pvr224 and pvr225-alleles interacted with the PVY-LYE84 VPg, suggesting that they are susceptibility alleles. In contrast, the eIF4E1 encoded by pvr223 did not, suggesting it is a resistance allele toward PVY-LYE84 (Fig 4, S3 Fig). A detailed analysis of the effects of the mutations present in these alleles relative to pvr2+ (Fig 2), which has been taken as reference for susceptibility [23,44,45], showed that the single mutation V67E (characterising pvr24) is sufficient to abolish the pvr2/eIF4E1-VPg interaction (S3 Fig). Similarly, the mutation A68E defining pvr223 and also present in pvr29, is sufficient to disrupt the pvr2/eIF4E1-VPg interaction (S3 Fig). Conversely, the single mutations A15V, D40G, K71R and V105I did not impair that interaction (S3 Fig). When these results were compared with a phylogeny of the pvr2/eIF4E1 haplotypes, it was apparent that the interaction between pvr2/eIF4E1 and PVY-LYE84 VPg was more stable for the alleles corresponding to the most ancestral haplotypes (pvr1+, pvr2+, pvr217, pvr224 and pvr225) than for the more derived pvr2 alleles (pvr21, pvr22, pvr24, pvr27, pvr29 and pvr223 (Fig 4, see also Fig 3). Interaction assays were also performed between the pvr2/eIF4E1 alleles identified in chiltepin populations and the VPg of TEV-HAT isolate, and demonstrated that the pvr2/eIF4E1-VPg interaction was efficient except for the pvr22 allele as previously reported [23].
Finally, chiltepin plants were inoculated with isolates PVY-LYE84 and TEV-HAT (see Material and Methods) in order to confirm the susceptibility/resistance phenotypes of the new alleles deduced from the Y2H assays. Since the pvr217 and pvr223 alleles are infrequent in chiltepin populations (Fig 2), the CPA populations where they were found were not included in this analysis. However, as alleles pvr224 and pvr225 are prevalent in AZP (Fig 2), 40 plants from seeds of the BER-W population were inoculated with each virus, and all of them showed symptoms 21 days after inoculation and high viral accumulation as detected by ELISA. The pvr2/eIF4E1 coding sequences were obtained from 10 randomly chosen plants among those inoculated with PVY-LYE84: 8 plants were homozygous for pvr224, 1 plant was homozygous for pvr225 and 1 plant was a pvr224/pvr225 heterozygote, which confirmed that the pvr224 and pvr225 alleles confer susceptibility to PVY-LYE84 and TEV-HAT.
Altogether, the described assays indicated that 6 out of 11 pvr2/eIF4E1 alleles found in chiltepin populations confer resistance to PVY-LYE84 infection. However, most pvr2/eIF4E1 sequences obtained in this study (83 out of 109, i.e. 76.1%) correspond to susceptibility alleles. When the distribution of resistance alleles in the sampled plants was analysed, it was found that 20.6% of plants would be resistant to PVY-LYE84 (Table 3). Resistance frequency significantly differed among biogeographical provinces (for all populations: χ2 = 58.2, P<10−4; for wild populations: χ2 = 29.5, P<10−4; for cultivated populations: χ2 = 20.2, P = 10−4), being highest in populations from SMO and YUC (for overall population: 84.2% and 25.0%, respectively; for wild populations: 66.7% and 10.0%, respectively; for cultivated populations: 92.3% and 100.0%, respectively; Table 3). Interestingly, the frequency of resistant plants was significantly higher in cultivated populations than in wild ones (55.6% and 8.6%, respectively, χ2 = 25.4; P<10−4; S5 Table, Table 3).
Most previously reported mutations in the pvr2/eIF4E1 protein of Capsicum spp. resulting in potyvirus resistance were predicted to be in the cap binding pocket [23,54]. None of the amino acid substitutions detected in pvr2/eIF4E1 of chiltepin relative to pvr2+, except D109N, were located at the sites interacting with the mRNA m7GTP cap or the eIF4G factor (S4 Fig). Since no experimental structure is available for the eIF4E1 protein of Capsicum, a three-dimensional model was built in order to locate and to predict the structural effects in pvr2/eIF4E1 of the mutations identified in chiltepin. First, the amino acid sequence of the C. annuum var. annuum pvr2+ reference allele was aligned with those of eIF4E proteins with known crystal structure (from Homo sapiens, Mus musculus, Triticum aestivum, and Pisum sativum). A phylogeny of these five eIF4E was reconstructed (S5A Fig), and their secondary structures were compared (S5B Fig), which showed a very high conservation except for the N-terminal domain which is longer in human, wheat and pepper (S5B Fig). The non-conserved N-terminal domain was demonstrated to be flexible in yeast [55], and our analysis confirmed that this domain is predicted to be disordered in human, wheat and in Capsicum (S5B and S5C Fig).
The 3D-models generated independently for the 11 pvr2/eIF4E1 alleles in Fig 2 confirmed, first, that the N-terminal domain of pvr2/eIF4E1 is flexible, and second, that the structural core of pvr2/eIF4E1 protein is not significantly altered by any amino acid substitution identified in chiltepin populations (S6 Fig). With the single exception of residue 109, which is placed in the β strand spanning amino acid positions 107–115, all the analyzed mutations involve residues located at loops (S6 Fig). Loops connecting secondary structure elements exhibit a great conformational flexibility and are usually exposed to the aqueous environment. Correspondingly, all mutations in pvr2/eIF4E1 alleles locate at the protein surface and, interestingly, they are close to the domain involved in the m7GTP cap recognition and far distant from the interface associated with eIF4G recruitment (Fig 5). It must be also noticed that being part of the disordered N-terminal region, the mutation A15V and to a lesser extent, the mutation D40G, should not alter significantly the essential functions of the pvr2/eIF4E1 protein.
In addition to being localized at the surface of the protein (Fig 5), most amino acid substitutions (6 out of 10) involved steric changes associated to side chain volumes (except for A15V, K71R, V105I and D109N mutations) as well as noticeable local variations of the electrostatic potential in the protein surface (Fig 6). For the new alleles pvr223, pvr224 and pvr225, only the mutation A68E in pvr223 introduced a large change in electrostatic potential relative to pvr2+, from a strong positive to a clearly negative potential in the external surface of the protein (Fig 6). It is interesting to note that there is a perfect correlation between all significant changes of electrostatic potential in pvr2/eIF4E1 and the disruption of its interaction with PVY VPg (Fig 6). Our results reveal that drastic changes in the local electrostatic potential of surface regions caused by some mutations (e.g. neutral to negative in V67E or neutral to positive in L79R) have a great impact in terms of disrupting the interaction with PVY-LYE84 VPg. Finally, as the N-terminal tails are disordered in the 3D models of all 11 pvr2/eIF4E1 alleles, variations among alleles in the electrostatic potential of those disordered regions are in part translated to nearby regions of the structural core. This is why the electrostatic potential of the structurally conserved core is not exactly the same in all alleles, which could indirectly alter the function of the pvr2/eIF4E1 protein.
To estimate the incidence of potyvirus infection in chiltepin populations, we analysed by ELISA leaf samples of 955 plants collected in wild and cultivated populations between 2007 and 2010. A total of 147 samples were ELISA positive, indicating a global Potyvirus incidence of 15.4% (Table 4). Potyvirus incidence varied significantly according to biogeographical province (χ2 = 50.2, P<10−4), being highest in SON and AZP (23.8% and 24.1%, respectively), where pvr2/eIF4E1 resistance alleles were not identified. Potyvirus incidence varied significantly according to year (from 8.5% in 2008 to 22.2% in 2010; χ2 = 15.0, P = 0.002). This temporal variation was solely due to wild populations, in which incidence varied according to year (χ2 = 24.1, P<10−4; Table 4), which was not the case for the cultivated ones (χ2 = 1.5, P = 0.676; Table 4), indicating a more constant challenge of virus infection in human-managed populations. Habitat, wild or cultivated, was not a factor on Potyvirus incidence (χ2 = 0.3, P = 0.597; Table 4), however, the percentage of infected plants that showed disease symptoms (mosaic, leaf distortion) was significantly higher in cultivated than in wild populations (45.5% and 9.8%, respectively; χ2 = 24.6, P<10−4) whereas it did not differ according to biogeographical province (χ2 = 7.3, P = 0.202) (Table 5).
To identify which Potyvirus species infected chiltepin populations in Mexico, we amplified a highly conserved region of NIb gene from the most ELISA positive samples. Amplification was successful from 8 samples, 4 from AZP, collected in 2008 and 2009, 3 from SON, 2007, and 1 from CPA, 2009, yielding two groups of sequences: those from SON and CPA were 99% identical to Pepper mottle virus (PepMoV), and those from AZP were 83% identical to Tobacco etch virus (TEV) (S7 Fig). Amplification and sequence determination of the genes encoding the VPg and CP in these samples confirmed the results based on the NIb fragment (VPg: 97% and 77% of identity with PepMoV and TEV, respectively; CP: 98% and 83% of identity with PepMoV and TEV, respectively). The TEV-like potyvirus differed in 30 out of the 188 VPg amino acid positions from TEV but none of them included a site reported to be involved in pvr2 resistance-breaking (S8 Fig).
In this study, the genetic diversity of the recessive resistance gene pvr2/eIF4E1 to potyviruses was analysed in the wild ancestor of domesticated pepper, Capsicum annuum var. glabriusculum (chiltepin), with the aim of inferring the evolutionary pattern of a resistance locus involved in matching-allele (MA)-like interactions, and of evaluating the impact of incipient domestication on that pattern. For that, we compared the diversity of pvr2/eIF4E1 for wild and cultivated chiltepin populations in six biogeographic provinces within its distribution range in Mexico, and we determined the phenotype of susceptibility or resistance of pvr2/eIF4E1 alleles by the analysis of the interaction between pvr2/eIF4E1 and PVY-LYE84 VPg in a yeast two hybrid (Y2H) assay, and by the response of plants to viral inoculations. Infection requires the physical interaction between pvr2/eIF4E1 and the potyviral VPg, and it has been shown that there is a perfect correlation between pvr2/eIF4E1-VPg interaction-no interaction in Y2H and susceptibility-resistance in plants [22,23,51]. Also, the lack of physical interaction between pvr2/eIF4E1 and PVY-LYE84 VPg has been shown to be an efficient way of identifying resistance to potyviruses in Capsicum spp. However, interactions of particular pvr2/eIF4E1 resistance alleles with the VPg of other potyviruses may be more stable, resulting in susceptibility. Indeed, among the 25 previously described pvr2/eIF4E1 alleles, 23 confer resistance to PVY-LYE84 and only one to TEV-HAT [22,23,44,56].
In 109 pvr2/eIF4E1 full-length coding sequences obtained from 97 chiltepin plants, 17 haplotypes were identified at the nucleotide sequence level, which largely differed in frequency. The most frequent one, haplotype D, accounted for 28% of total sequences, and the other four haplotypes encoding the susceptibility allele pvr1+, which according to the minimum spanning network (MSN) and phylogenetic analyses represents the basal state of pvr2/eIF4E1 in chiltepin (Figs 2 and 3), accounted for 44% of total sequences (Fig 2). Allele frequency also varied according to biogeographical province, so that the genetic diversity of pvr2/eIF4E1 coding sequence was 2.5–5 times higher in YUC and SMO than in the other four biogeographical provinces (Table 2). Also, the most basal pvr2/eIF4E1 haplotype (G, Fig 2) was only identified in YUC. These results are consistent with the higher genetic diversity of chiltepin in YUC and SMO estimated from nuclear microsatellite makers (SSRs) [35] and with reports that identify the Yucatan peninsula and the areas around the Gulf of Mexico as centres of diversity and domestication of C. annuum [33,57]. Analyses of nuclear SSRs have shown a strong spatial structure of chiltepin genetic diversity according to biogeographical province [35], which was also the case for pvr2/eIF4E1, both when the coding sequence or the introns (S3 Table) were analysed. However, at odds with results from SSRs, which showed evidence of isolation by distance, the genetic distance among chiltepin populations at pvr2/eIF4E1 poorly correlated with geographical distance. The discrepancy between the spatial structure of the variation of putatively neutral genetic markers and of pvr2/eIF4E1 suggests that this gene is under selection associated with environment-specific factors. Although other factors may certainly be involved, selection on pvr2/eIF4E1 could be associated with resistance to potyviruses, as potyvirus incidence differs according to biogeographical province (Table 4).
In agreement with the hypothesis that there is selection on pvr2/eIF4E1 for resistance, MSN and phylogenetic analyses indicate that pvr2/eIF4E1 has evolved to confer potyvirus resistance. Most pvr2/eIF4E1 alleles can be connected by just one amino acid substitution, and the allelic diversity found in chiltepin allowed to identify alleles, as pvr223, which were predicted as most parsimonious intermediates in pvr2/eIF4E1 evolution by Moury et al [44] (Fig 3). Analyses showed that the susceptibility allele pvr1+ is at the base of pvr2/eIF4E1 phylogeny. From that state, evolution has proceeded towards decreasing the stability of the interaction between pvr2/eIF4E1 and PVY-LYE84 VPg, i.e., towards resistance, as judged by yeast growth in a selective medium complemented by a Y2H assay interaction (Fig 4). The most supported node in pvr2/eIF4E1 phylogeny splits haplotypes encoding susceptibility alleles pvr1+ and pvr217, from a cluster built of two less strongly supported subclusters, one including haplotypes corresponding to susceptibility alleles pvr224 and pvr225, and the other including haplotypes corresponding to susceptibility alleles pvr2+, from which all other haplotypes, encoding resistance alleles, derive (Fig 4). The pattern of evolution into this last cluster including both susceptibility and resistance alleles is compatible with a hypothesis of selection on pvr2/eIF4E1 resulting in the evolution of a variety of resistance alleles, as was concluded from the analysis of a set of 25 accessions of Capsicum annuum [23]. Interestingly, when the phylogeny of all reported pvr2/eIF4E1 alleles was reconstructed, resistance also appeared as a derived state, and evolution to resistance occurred in different phylogenetic clusters (S9 Fig). Although support for the internal nodes of the phylogeny was not strong, the topology was consistent regardless of the method of phylogenetic reconstruction, or when the phylogeny was based on only first and second codon positions (S9 and S10 Figs). Phylogenies derived from third codon positions (S11 Fig) did not present an informative pattern, supporting the significance of the main clusters in the other phylogenies. However, at odds with previous analyses [23], when the alleles in our chiltepin data set are considered, evidence of selection for resistance is weaker: most (10/17) haplotypes encoded susceptibility alleles and a large number of pvr2/eIF4E1 polymorphisms in the chiltepin population were due to synonymous nucleotide substitutions, so that 7/17 haplotypes encoded the susceptibility alleles pvr1+ (5 haplotypes) and pvr2+ (2 haplotypes). In contrast, only non-synonymous mutations were found in the data set analysed by Charron et al [23]. Accordingly, no site, including those that determine potyvirus resistance, was identified in our data set as being under positive selection, with the possible exception of codon 205, in which the mutation D205G confers potyvirus resistance and occurred at least twice during pvr2/eIF4E1 evolution in chiltepin (Fig 3). Positive selection on codons involved in potyvirus resistance was only detected in a data set including a wide range of plant species [44].
In the chiltepin population the frequency of potyvirus resistance was moderate, as 21.6% of plants were predicted to be resistant to PVY-LYE84, and 26.0% of pvr2/eIF4E1 sequences corresponded to resistance alleles (Table 3). Most resistance alleles were identified in SMO populations, and among resistance alleles only pvr23 and pvr24 were found in more than one biogeographical province (Fig 2). Interestingly, 55.6% of plants, and 62.5% of pvr2/eIF4E1 sequences were resistant to PVY-LYE84 in cultivated populations, as compared with 8.4% of plants and 7.8% of sequences in wild ones, and the higher proportion of resistance in cultivated populations held for the three biogeographical provinces in which resistance alleles/plants were found (YUC, SMO and CPA, Table 3). Four out of seven nucleotide sequence haplotypes encoding resistance alleles were found in cultivated populations. Heterozygosity at the pvr2/eIF4E1 locus was not different in wild or cultivated populations (Table 1, S5 Table), while for SSRs heterozygosity was higher in wild than in cultivated populations, and values were higher than for pvr2/eIF4E1 [35]. Nucleotide diversity at pvr2/eIF4E1 was higher in cultivated than in wild populations, whereas a significant decrease in genetic variation at neutral markers in cultivated populations was previously demonstrated in chiltepin [35] as it is commonly observed during plant domestication [26–28]. Also, there was a higher fraction of non-synonymous substitutions in cultivated populations than in wild ones, resulting in dN/dS ratios indicative of positive selection, as opposed with data from wild populations (Table 2). Last, DT values were positive for the region between codons 67 and 77, which includes most determinants of potyvirus resistance (Region I in Fig 2), in cultivated but not in wild populations (Fig 1). Thus, all data taken together indicate that selection for potyvirus resistance is stronger in cultivated than in wild chiltepin populations, and results in higher diversification of the pvr2/eIF4E1 gene. It is noteworthy that both a ~55% frequency of potyvirus resistance and evidence of diversifying selection was found by Charron et al [23] in 25 accessions of C. annuum, mostly cultivated. High frequency of eIF4E-mediated resistance to the bymoviruses (in family Potyviridae) Barley yellow mosaic virus and Barley mild mosaic virus has also been found in accessions from domesticated barley varieties, with evidence of diversifying selection for resistance [24]. The eIF4E alleles conferring resistance to the potyvirus Pea seed borne mosaic virus were only found in domestic pea accessions, in spite of high variability of the locus in wild accessions [25]. So, these reports of other host-virus systems agree with a hypothesis of cultivation-associated selection for resistance at eIF4E.
Although the ecological changes associated with cultivation are considered to favor the incidence of plant pathogens [58,59], which is certainly the case for begomoviruses and other viruses infecting chiltepin in Mexico [38,60], potyvirus incidence in chiltepin did not differ according to habitat (Table 4). However, potyvirus incidence varied less among years in cultivated than in wild populations (Table 4), indicating a more constant challenge of virus infection. Interestingly, in chiltepin populations localized in anthropic environments and tolerated but not cultivated by humans, i.e. “let-standing” populations [35], potyvirus incidence varied temporally as in wild populations (χ2 = 9.1, P = 0.028) strongly suggesting that cultural practices favor a more constant potyvirus prevalence. More significantly, infection in cultivated populations was much more virulent, as 5 times more infected plants showed disease symptoms in cultivated than in wild populations (Table 5), and disease expression can be a good proxy of virulence in plant virus interactions [61–63]. Differences in selection for potyvirus resistance in the wild and under cultivation can be due to human-driven directional selection, as a response to strong symptom expression in cultivated populations, or to natural selection caused by cultivation conditions favoring a more constant and stronger effect of potyvirus infection. The role of natural selection during plant domestication is often overlooked and has been recently emphasized [29]. Also, the shorter generation time in cultivated populations, where chiltepin is managed as an annual crop, as compared with the 4–6 year perennial life span in the wild, could favor a higher selection rate per generation for resistance in the cultivated populations. We cannot at present evaluate the relative role of these contrasting factors on the evolution of potyvirus resistance in chiltepin wild and cultivated populations.
The core structure of the pvr2/eIF4E1 protein would not be affected significantly by the amino acid substitutions found in chiltepin. However, substitutions that uncoupled the pvr2/eIF4E1-VPg interaction, resulting in resistance, were around the cap-binding pocket and strongly affected the electrostatic surface potential at this region, which is reasonable to expect would affect the binding of eIF4E to the cap of cellular mRNAs and, hence its efficiency in translation initiation. Thus, potyvirus resistance would have a cost even if the resistance alleles are fully functional for translation in yeast complementation assays. The location of amino acid substitutions on the protein structure, the low dN/dS values and the low frequency of resistance alleles in wild chiltepin populations, altogether support a hypothesis of functional constraints translating into costs limiting the evolution of pvr2/eIF4E1 towards potyvirus resistance. Capsicum plants carrying an eIF4E1 loss-of-function allele, which could provide evidence on eIF4E1 involvement in development/plant fitness and thus of mutation costs, are not available. A TILLING eIF4E1 knock out allele in cultivated tomato was not associated with obvious developmental defaults under greenhouse conditions [64], although it might be detrimental under more stressful wild conditions. Costs of resistance have been often reported in GFG-like plant-pathogen interactions [10–12,65], but are not a feature of the evolution of pure MA interactions. However, it is considered that real-world host-parasite interactions that mechanistically correspond to a MA model would fall within a continuum between pure MA and GFG models, in which partial infection with less successful parasite multiplication occurs, with correspondingly partial costs of resistance and infectivity [5,7]. This seems indeed to be the case of the pvr2/eIF4E1-mediated interaction between Capsicum and potyviruses, as infections largely differ in efficiency and costs of infectivity have been reported [66–68]. Our present results suggest that resistance costs could also determine the evolutionary dynamics of the Capsicum-Potyvirus interaction.
The evolution of dominant resistance genes (R genes) of plants to cellular pathogens, which are involved in GFG-like interactions, has been analysed extensively. Data indicate that R genes are hypermutagenic and often under balancing selection [21,69–72]. The present work focuses on the analysis of the evolution of a recessive resistance gene involved in a MA-like interaction in populations of a wild plant. It also compares evolutionary dynamics between plant populations under different levels of human management. Notably, results show a quite different pattern depending on the level of human management of the habitat. While there is no evidence of high genetic variation or of selection on pvr2/eIF4E1 in wild chiltepin populations, as often reported for R genes [21,69–72], there is evidence of selection on pvr2/eIF4E1 for potyvirus resistance in the cultivated populations, which is compatible with a hypothesis of balancing selection maintaining pvr2/eIF4E1 resistance diversity. These major results are perhaps unexpected as cultivation of chiltepin is recent and has not yet resulted in domestication or in obvious phenotypic changes, and the cultivated populations here analysed are not genetically differentiated from sympatric wild ones according to the variation of nuclear SSRs markers [35]. It is widely accepted that human management of plant habitats heavily influence the epidemiology of plant pathogens, including plant viruses [59,73], as has been shown for viruses infecting chiltepin [38,60]. This study shows that human management of the habitat may also have a deep impact on the evolution of plant-pathogen interactions, an underexplored topic in need of more research.
Chiltepin plants were sampled during the summers of 2007–2010 at different sites over the species distribution in Mexico [35]. Plant samples were collected from chiltepin populations growing in a variety of habitats under different levels of human management [35]. For analyses of the pvr2/eIF4E1 gene we focused on those from the most extreme levels of human management, i.e. the wild and cultivated populations. Plants grown from seeds in fruits purchased at local markets were also analysed, and were considered here as from wild populations, if (i) the people selling the fruits claimed that they had been collected from local wild chiltepin populations and (ii) after their genetic characterization based on the polymorphisms of nine microsatellite markers [35], those market populations were indeed shown to be related to the local wild populations.
Thus, for analyses of the pvr2/eIF4E1 gene, we considered a total of 25 populations, 16 wild and 9 cultivated, (S1 Table) from six biogeographical provinces of Mexico: Yucatan (YUC), Eastern side of the Sierra Madre Oriental (SMO), Altiplano Zacatecano-Potosino (AZP), Costa del Pacífico Sur (CPS), Costa del Pacífico (CPA), and Sonora (SON) [74]. A larger set of samples from populations growing in all the habitats (wild, cultivated and let-standing populations) [35] was used to evaluate Potyvirus incidence according to biogeographical province, habitat and year of sampling.
For analysis of the pvr2/eIF4E1 gene total nucleic acids were extracted from leaves as in González-Jara et al [35]. The pvr2/eIF4E1 gene is constituted of 5 exons of 278, 166, 126, 66 and 51 nucleotides (nt), respectively, separated by 4 introns of more than 3500 nt, 110 nt, 1143 nt and 83 nt, respectively [75]. To amplify both introns and exons of the pvr2/eIF4E1 gene, two different PCRs were run directly on the total nucleic acid extracts, using the Phusion High-Fidelity DNA Polymerase (New England Biolabs, MA, USA). The first PCR was performed with primers F-eIF4E.Full (ATGGCAACAGCTGAAATGGAG) and R-eIF4E.int1 (CCCCGAGAATCTTAGTAGCTCA), designed to amplify a 756 nt fragment including pvr2/eIF4E1 exon 1 and the 5’ most 403 nt of intron 1. Conditions for this PCR were 98°C for 30 sec, and 35 cycles of 98°C for 10 sec, 56°C for 30 sec and 72°C for 25 sec. The second PCR was performed using primers F-eIF4E.ex2 (TGCTTACAATAATATCCACCACCC) and R-eIF4E.3’UTR (CACAAGGTACTCAAACCAGAAGC), designed to amplify a 1848 nt fragment including the four other exons of pvr2/eIF4E1 and introns 2 to 4. Conditions for this PCR were 98°C for 30 sec, and 35 cycles of 98°C for 10 sec, 54°C for 30 sec and 72°C for 1 min. Primers F-eIF4E.Full and R-eIF4E.int1 were also used to obtain the full nucleotide sequence of the amplicon from the first PCR. To determine the nucleotide sequence of the amplicon from the second PCR, primers F-eIF4E.ex2, R-eIF4E.int3 (CCCCTTCATCTATAAGCATATTTC), F-eIF4E.int3end (GATGGTCTCAAGGGTTATGTGTC) and R-eIF4E.3’UTR were used, in order to obtain the complete sequence of exons 2, 3, 4 and 5, and of introns 2 and 4, and two partial sequences of intron 3 (5’ fragment: 293 nt; 3’ fragment: 547 nt). The pvr2/eIF4E1 coding sequence was then deduced from the exon sequences.
Sequence analyses identified plants heterozygous for the pvr2/eIF4E1 gene. Sequence determination in heterozygotes was done after RT-PCR amplification of pvr2/eIF4E1 coding sequences and/or cloning of the DNA amplicons in pCRII (TA Cloning Kit Dual Promoter, Invitrogen, Carlsbad, CA, USA). RT-PCR amplification of pvr2/eIF4E1 coding sequences was also used to identify the pvr2/eIF4E1 allele(s) present in virus-inoculated plants (see below). In this case, the RT step was performed with the SuperScript III Reverse Transcriptase (Invitrogen) according to the manufacturer’s recommendations using primer R-eIF4E.3’UTR, followed by a PCR amplifying the cDNA corresponding to the full coding sequence of pvr2/eIF4E1 with the primers F-eIF4E.Full and R-eIF4E.3’UTR (PCR conditions: 98°C for 30 sec, and 35 cycles of 98°C for 10 sec, 53°C for 30 sec and 72°C for 25 sec).
Nucleotide sequences were aligned to maintain the reading frame using CLUSTAL-W [76] as implemented in Mega 6 [77]. Differences in heterozygous plants at the pvr2/eIF4E1 locus, in haplotype richness and in resistance frequency between populations, regions or habitat were assessed by the analysis of contingency tables using the Fisher exact test. Genetic diversity within and between populations, biogeographical provinces or levels of human management were estimated using the Kimura 2-parameter model, with standard errors of each measure based on 1000 replicate bootstraps, as implemented in Mega 6. Differences in nucleotide diversity of the virus populations among biogeographical provinces and between habitats were tested by analysis of molecular variance (AMOVA), as implemented in Arlequin v. 5.3.1.2 [78]. Differences in dN/dS values were considered to be significant if the mean value of one estimate fell outside of the 95% CI values of another, indicating that these dN/dS values were drawn from different distributions. AMOVA calculates the FST index explaining the between-groups fraction of total genetic diversity. Significance of these differences was obtained by performing 1000 permutations. Tajima’s D (DT) and sliding window analyses were conducted using DnaSP v. 5.10 [79].
Mantel correlation tests between geographic and genetic distance matrices were performed to test the isolation-by-distance hypothesis [80] in wild chiltepin populations using the web service http://ibdws.sdsu.edu/~ibdws/ [81]. We used the geographic distance matrices obtained in González-Jara et al [35]. Geographical and genetic distances between pairs of populations were log transformed, and 1000 permutations were performed to assess the significance of the correlations.
We used the median-joining network method implemented in the Network version 4.611 software (available at www.fluxus-engineering.com) [82] to reconstruct the minimum spanning network (MSN) connecting all chiltepin pvr2/eIF4E1 alleles identified at the amino acid level. Phylogenetic relationships were reconstructed by the Neighbor-Joining method as implemented in Mega 6 [77] and incorporating the best-fitted nucleotide substitution model (F81 model) determined by jModelTest 0.1.1 [83]. The sequence of the Potyvirus susceptibility allele pot-1+ from tomato (Solanum lycopersicum, accession number AY723733) was used as outgroup. Phylogenies were also reconstructed by Maximum Likelihood and by Maximum Parsimony using Subtrees Pruning and Regrafting method as implemented in Mega 6 with similar results.
The ratio of non-synonymous (dN) to synonymous (dS) substitutions over the pvr2/eIF4E1 coding sequences from chiltepin populations was estimated by the Pamilo-Bianchi-Li method as implemented in Mega 6. The dN/dS ratio was also estimated at individual codons in the pvr2/eIF4E1 coding sequences, using different methods implemented in the HYPHY program (SLAC, Single Likelihood Ancestor Counting; FEL, Fixed Effects Likelihood; IFEL, Internal Fixed Effects Likelihood; REL, Random Effects Likelihood; FUBAR, Fast Unbiased Bayesian Approximation) [84–87] to determine whether each of the 228 codons of pvr2/eIF4E1 were under negative (dN/dS<1), neutral (dN/dS = 1), or positive (dN/dS>1) selection. These analyses were performed after confirmation of the absence of recombinant sequences in our dataset by two methods implemented in the HYPHY program (SBP, Single Breakpoint Recombination; GARD, Genetic Algorithms for Recombination Detection) [86] and using the tree topology previously obtained for pvr2/eIF4E1.
The Saccharomyces cerevisiae strain JO55 [cdc33-D LEU2 leu2 ura3 his3 trp1 ade2 (YCp33supex-h4E URA3)] [88], carrying a disrupted endogenous eIF4E gene (cdc33), was used as in Charron et al [23] to verify the functionality of the pvr2/eIF4E1 allelic variants identified in chiltepin populations. The coding sequence of the pvr2+ allele was cloned into the p424GBP/TRP1 glucose-dependent vector, and all pvr2/eIF4E1 allelic variants were obtained by mutagenesis of this construct using the QuikChange Site-Directed Mutagenesis Kit (Stratagene, Agilent Technologies, Santa Clara, CA, USA). Each construct was sequenced to confirm the presence of the introduced mutations and then independently used to transform S. cerevisiae strain JO55. After transformation, yeast cells were grown in appropriate selective nutrient drop-out media containing 2% glucose. Control transformations were performed with no DNA (untransformed yeast JO55) and empty p424GBP/TRP1 plasmids (negative controls), and with p424GBP/TRP1::At-eIF4E (eIF4E form of Arabidopsis thaliana, At4g18040) as a positive control. After transformation, yeast colonies were grown to stationary phase, were suspended in sterile water, and then were adjusted to an OD600nm of 5.10−2, 5.10−3, and 5.10−4 before spotting 10 μl aliquots onto the appropriate media in order to test for their ability to complement the lack of endogenous eIF4E at 30°C [89]. For each pvr2/eIF4E1 allelic variant, 3 independent colonies were randomly selected to perform the complementation assay.
The Matchmaker GAL4 two-hybrid system 3 (Clontech, Mountain View, CA, USA) was used according to the manufacturer’s recommendations to evaluate the interaction of the proteins encoded by the pvr2/eIF4E1 allelic variants with the potyviral VPg. The constructs previously developed by Charron et al [23] were used. The eIF4E1/pvr2+ coding sequence was cloned in-frame with the GAL4 activation domain into the pGADT7 vector (Clontech, Mountain View, CA, USA), and all pvr2/eIF4E1 allelic variants were obtained by mutagenesis with the QuikChange Site-Directed Mutagenesis Kit (Stratagene). All the constructs were sequenced to confirm the presence of the introduced mutations before yeast transformation. The VPg of PVY (avirulent isolate LYE84) [90] and of TEV (avirulent isolate HAT) [48] were cloned in-frame with the GAL4 binding domain into the pGBKT7 vector, respectively [23]. The pGADT7- and pGBKT7-derived vectors were transformed into AH109 and Y187 yeast strains, respectively, which contain two independent reporter genes, HIS3 and ADE2, to confer histidine and adenine auxotrophy, respectively, driven by hybrid GAL4 promoters. After yeast mating, double-transformed yeast colonies were grown to stationary phase, were suspended in sterile water, and then were adjusted to an OD600nm of 5.10−2 before spotting 10 μl aliquots onto various selective media including synthetic medium lacking leucine and tryptophan (hereafter named -LW) and medium lacking leucine, tryptophan and histidine (-LWH). Plates were incubated at 30°C, and yeast growth was checked daily from 2 to 7 days after spotting. The yeast growth on the selective–LWH medium reflects the pvr2/eIF4E1-VPg physical interactions. Empty pGADT7 and pGBKT7 vectors were used as negative controls and interaction between murine p53 and SV40 large T antigen as positive controls. Three independent yeast-two hybrid assays were performed, in which 3 independent colonies of each pvr2/eIF4E1-VPg combination were randomly selected.
For complementation and yeast-two hybrid assays, growth intensities were monitored with ImageJ software [91], and raw data were normalized to positive and negative controls and expressed as a percentage of the growth of the reference yeast colonies (transformed with p424GBP/TRP1::eIF4E1/pvr2+ for complementation assays, and co-transformed with pGADT7::eIF4E1/pvr2+ and pGBKT7::VPg-PVY for yeast two-hybrid assays) as previously described in Hébrard et al [92].
The secondary structure of eIF4E proteins used in this study from Capsicum annuum pvr2+ allele; Triticum aestivum, 2IDR; Pisum sativum, 2WMC; and the mammalian eIF4Es used as outgroup from Homo sapiens, PDB ID: 4DT6; Mus musculus, 1L8B [54,93–95] was predicted using the server NPS, which deduced the consensus secondary structure of protein from 12 different methods (http://npsa-pbil.ibcp.fr) [96].
The tertiary structure of all the pvr2/eIF4E1 alleles identified in chiltepin populations was modelled with the Iterative Threading ASSEmbly Refinement (I-TASSER) hybrid method [97–99]. Starting from an amino acid sequence, I-TASSER first generates 3D atomic models from multiple threading alignments and iterative structure assembly conducted by Monte Carlo simulations under an optimized knowledge-based force field. The lowest free-energy conformations are identified by structure clustering and final atomic structure models are constructed from the low-energy conformations by means of a two-step atomic-level energy minimization approach. The correctness of the models is assessed by a confidence score (C-score) and a measure of structural similarity (TM-score). In all cases, the 3D structures were constructed from scratch without resorting to previous models of other alleles. Among the five models predicted by I-TASSER, that having the best values of both C-score and TM-score was finally selected. The main pvr2+ structure had C-score = 0.09 (C-score is typically in the [−5, 2] range, with a higher value meaning a model with higher confidence) and TM-score = 0.73 ± 0.11 (a TM-score > 0.5 indicates a model of correct topology). For the remaining alleles, C-score ranged from -1.57 and +0.28 and TM-score ranged between 0.52 ± 0.15 and 0.75 ± 0.10 so that all the 3D models presented here for the different pvr2/eIF4E1 alleles may be considered as having significant confidence and being topologically correct.
The 3D model structures were first visualized and analyzed with Swiss-PdbViewer 4.1.0 [100], software which was also used for rendering van der Waals (VdW) surfaces, obtaining pairwise structural superpositions and computing the corresponding root mean square deviation (RMSD) values. All structure models of pvr2/eIF4E1 alleles showed an N-terminal unstructured segment spanning the first 45–50 residues in their amino acid sequences. To further assess this result, we applied the following predictors of protein disorder: DisEMBL [101], DISOPRED [102], and IUPred [103] to the amino acid sequence of the main pvr2+ allele. Given that they employ disparate algorithms based on rather different assumptions, their close agreement in predicting disorder for segments 1–44 (DisEMBL), 1–50 (DISOPRED), and 1–45 (IUPred) lend further support to the structural models generated by I-TASSER.
Poisson-Boltzmann (PB) electrostatic potentials mapped onto the protein surface of all the pvr2/eIF4E1 alleles were computed by solving the PB equation with APBS 1.4 [104] using AMBER99 [105] atomic charges and radii assigned with PDB2PQR 1.7 [106]. The nonlinear PB equation was solved at 298.15 K and 0.150 M ionic concentration in sequential focusing multigrid calculations in 3D meshes of 1603 or 1923 points with step sizes about 0.35 or 0.50 Å depending on the particular pvr2/eIF4E1 allele. Dielectric constants 4 for proteins and 78.54 for water were used. The output of PB electrostatic potentials thus computed were obtained in scalar OpenDX format and these numerical meshes were then mapped onto molecular surfaces of proteins and rendered with PyMOL 1.6 (PyMOL, Schrodinger, LLC). PB electrostatic potential values are given in units of kT per unit charge (k, Boltzmann's constant and T, absolute temperature).
All plants were grown under greenhouse conditions and transferred into growth chambers before inoculation (16h light/8h dark; 24°C/18°C). Chiltepin plants were mechanically inoculated at the cotyledon stage with PVY-LYE84 (pathotype PVY-0) and TEV-HAT [48,90] as previously described [107]. The C. annuum accessions Yolo Wonder (pvr2+ homozygous, susceptible to PVY-LYE84 and TEV-HAT) and Florida (pvr22 homozygote, resistant to PVY-LYE84 and TEV-HAT) were used as susceptible and resistant controls, respectively. Plants mock-inoculated with buffer were used as negative controls. Systemic infection was assessed by determining the presence/absence of symptoms on non-inoculated leaves and confirmed by DAS-ELISA using PVY or TEV antibodies.
Infection by Potyvirus species in natural chiltepin populations was detected by DAS-ELISA, using the complete kit of detection PSA 27200/0288 according to the manufacturer’s recommendation (AGDIA, Elkhart, IN, USA). This kit is based on the broad reactivity of a monoclonal antibody reacting to a highly conserved amino acid sequence on the coat protein of the Potyvirus genus. A total of 955 plants from 24 wild and cultivated populations were analysed in this way, plus 238 plants from let-standing populations. Differences in potyvirus incidence or symptom frequency in infected plants were assessed by the analysis of contingency tables using the Fisher exact test. The presence of virus in the ELISA-positive samples was confirmed by RT-PCR using the potyvirus-specific degenerated primers designed by Zheng et al [108], which amplify a region of the NIb gene (positions 7619–7968) highly conserved between Potyvirus species. Once the Potyvirus species was identified by NIb sequencing, species-specific primers bordering the VPg and the CP were designed. These primers were: for PepMoV, F-PepMoV.VPg: GTGCATCACCAGTCCAAGTCTT and R-PepMoV.VPg: CAGTCAACTTGCAAACAGTTTGG, F-PepMoV.CP: GCTGACTTGGCATCTGAAGGA and R-PepMoV.CP: TTCATCCCAGAGACCACATCAG; for TEV-like virus, F-TEVlike.VPg: GTATCATCCAAGACTTCAATCACCTGGAAAC and R-TEVlike.VPg: GATGTTGTGTGCCCATCAGATTCATTC, F-TEVlike.CP: CACAGCTTGCAGARGAAGGAAAGGC and R-TEVlike.CP: CTTAAAAGCGGAAAGCAAAGACACGC).
|
10.1371/journal.pbio.1001025 | Local Ca2+ Entry Via Orai1 Regulates Plasma Membrane Recruitment of TRPC1 and Controls Cytosolic Ca2+ Signals Required for Specific Cell Functions | Store-operated Ca2+ entry (SOCE) has been associated with two types of channels: CRAC channels that require Orai1 and STIM1 and SOC channels that involve TRPC1, Orai1, and STIM1. While TRPC1 significantly contributes to SOCE and SOC channel activity, abrogation of Orai1 function eliminates SOCE and activation of TRPC1. The critical role of Orai1 in activation of TRPC1-SOC channels following Ca2+ store depletion has not yet been established. Herein we report that TRPC1 and Orai1 are components of distinct channels. We show that TRPC1/Orai1/STIM1-dependent ISOC, activated in response to Ca2+ store depletion, is composed of TRPC1/STIM1-mediated non-selective cation current and Orai1/STIM1-mediated ICRAC; the latter is detected when TRPC1 function is suppressed by expression of shTRPC1 or a STIM1 mutant that lacks TRPC1 gating, STIM1(684EE685). In addition to gating TRPC1 and Orai1, STIM1 mediates the recruitment and association of the channels within ER/PM junctional domains, a critical step in TRPC1 activation. Importantly, we show that Ca2+ entry via Orai1 triggers plasma membrane insertion of TRPC1, which is prevented by blocking SOCE with 1 µM Gd3+, removal of extracellular Ca2+, knockdown of Orai1, or expression of dominant negative mutant Orai1 lacking a functional pore, Orai1-E106Q. In cells expressing another pore mutant of Orai1, Orai1-E106D, TRPC1 trafficking is supported in Ca2+-containing, but not Ca2+-free, medium. Consistent with this, ICRAC is activated in cells pretreated with thapsigargin in Ca2+-free medium while ISOC is activated in cells pretreated in Ca2+-containing medium. Significantly, TRPC1 function is required for sustained KCa activity and contributes to NFκB activation while Orai1 is sufficient for NFAT activation. Together, these findings reveal an as-yet unidentified function for Orai1 that explains the critical requirement of the channel in the activation of TRPC1 following Ca2+ store depletion. We suggest that coordinated regulation of the surface expression of TRPC1 by Orai1 and gating by STIM1 provides a mechanism for rapidly modulating and maintaining SOCE-generated Ca2+ signals. By recruiting ion channels and other signaling pathways, Orai1 and STIM1 concertedly impact a variety of critical cell functions that are initiated by SOCE.
| Store-operated Ca2+ entry is present in all cell types and determines sustained cytosolic [Ca2+] increases that are critical for regulating a wide variety of physiological functions. This Ca2+ entry mechanism is activated in response to depletion of Ca2+ in the endoplasmic reticulum (ER). When ER [Ca2+] is decreased, the Ca2+-sensor protein STIM1 aggregates in the ER membrane and moves to regions in the periphery of the cells where it interacts with and activates two major types of channels that contribute to store-operated Ca2+ entry: CRAC and SOC. While gating of Orai1 by STIM1 is sufficient for CRAC channel activity, both Orai1 and transient receptor potential channel 1 (TRPC1) contribute to SOC channel function. The molecular composition of SOC channels and the critical role of Orai1 in activation of TRPC1 have not yet been established. In this study, we demonstrate that TRPC1 and Orai1 are components of distinct channels, both of which are regulated by STIM1. Importantly, we show that Orai1-mediated Ca2+ entry triggers plasma membrane insertion of TRPC1 which is then gated by STIM1. Ca2+ entry via functional TRPC1-STIM1 channels provides additional increase in cytosolic [Ca2+] that is required for regulation of specific cell functions such as KCa activation. Together, our findings elucidate the critical role of Orai1 in TRPC1 channel function. We suggest that the regulation of TRPC1 trafficking provides a mechanism for rapidly modulating cytosolic [Ca2+] following Ca2+ store depletion.
| Store-operated Ca2+ entry (SOCE) is activated in response to a reduction of [Ca2+] in the ER. SOCE generates local and global [Ca2+]i signals that regulate a wide variety of cellular functions [1],[2]. The first store-operated Ca2+ channel to be characterized, the Ca2+ release-activated Ca2+ (CRAC) channel, has a high selectivity for Ca2+ versus Na+ and displays a typical inwardly rectifying current-voltage relationship. CRAC channel accounts for the SOCE in lymphocytes and mast cells [3]–[6] and has recently been detected in some other cell types [7]–[9]. Key molecular components of the channel are STIM1 and Orai1. STIM1 is an ER Ca2+ binding protein that has been established as the primary regulator of SOCE [10]–[12]. In response to store depletion STIM1 oligomerizes and translocates to ER/PM junctional domains where it aggregates into puncta. The site of these aggregates is the location where STIM1 interacts with and activates channels involved in SOCE [13]–[15]. Orai1 is the pore-forming subunit of the CRAC channel [16]–[18]. Following store depletion, Orai1, which is localized diffusely in the plasma membrane in resting cells, is recruited by STIM1 into the puncta and gated by interaction with a C-terminal region of STIM1 [19],[20]. While expression of this STIM1-domain induces spontaneous CRAC channel activation in extra ER/PM junctional domains, the site of the STIM1 puncta represents the cellular location where endogenous SOCE is activated by store depletion [21].
Store depletion also leads to activation of relatively non-selective Ca2+-permeable cation channels, usually referred to as SOC channels, that have been associated with SOCE in several other cell types [2],[22]–[25]. Despite more than a decade of studies, the molecular components of these channels have not yet been established and their function and regulation remain somewhat controversial. TRPC channels have been proposed as molecular components of SOC channels. Data in this regard are strongest for TRPC1 [2],[26]–[34] although TRPC3 and TRPC4 also appear to contribute to SOCE in some cell types [23],[25],[35]–[38]. Numerous studies show that disruption of TRPC1 attenuates SOCE and SOCE-dependent cell function [23],[26]–[34]. We have previously provided extensive data to demonstrate that TRPC1 is a critical component of SOC channels and SOCE in the human salivary gland cell line, HSG [30],[39]–[42]. Further, salivary gland acinar cells from TRPC1−/− mice display reduced SOCE and SOC channel activity, which account for loss of sustained KCa activation and, consequently, salivary fluid secretion [29]. However, the role of TRPC1 in SOCE has been questioned based on the lack of function of heterogously expressed channels [43]. Further, some tissues from TRPC1−/− mice do not display any changes in SOCE [44],[45]. The strongest evidence for the regulation of TRPC1 following store depletion has been provided by data demonstrating that STIM1 interacts with and activates TRPC1-SOC channels in response to Ca2+ store depletion [39],[42],[46]. SOC channels are attenuated by knockdown of endogenous STIM1 and spontaneously activated by expression of the STIM1 mutant, D76ASTIM1 [42],[46]. An important study showed that TRPC1 is gated by electrostatic interaction between STIM1(684KK685) and TRPC1(639DD640) [47].
An intriguing finding is that STIM1 alone is not sufficient for activation of TRPC1-SOC channels following Ca2+ store depletion. Functional Orai1 is also required since knockdown of Orai1 or expression of functionally deficit Orai1 mutants prevents TRPC1 activation [39],[48]. We have shown earlier that store depletion leads to the recruitment of a TRPC1/STIM1/Orai1 complex that is associated with the activation of SOCE [39],[42]. Thus, while STIM1 is the primary protein involved in SOC channel gating, both TRPC1 and Orai1 appear to contribute to SOC channel activity. There has been much debate about the essential role of Orai1 in TRPC1-SOC channel function and more specifically regarding whether TRPC1 and Orai1 contribute to a single SOC channel pore or whether Orai1 is a regulatory subunit of SOC channels. In this study we have assessed the critical role of Orai1 in regulation of TRPC1 function following intracellular Ca2+ store depletion and determined the contributions of TRPC1 and Orai1 to SOCE. We report that TRPC1 and Orai1 constitute two distinct channels that contribute to SOCE in HSG cells. Suppression of TRPC1 function unmasks the underlying CRAC channel function. Further, in response to store depletion, STIM1 mediates association of Orai1 and TRPC1 within ER/PM junctional domains. Ca2+ entry via Orai1/STIM1-CRAC channel triggers plasma membrane insertion of TRPC1 and gating is achieved by interaction with STIM1(684KK685) residues. Remarkably, while both Orai1 and TRPC1 contribute to [Ca2+]i increase following store depletion, they impact different cellular functions. Ca2+ entry mediated by TRPC1 is the primary regulator of KCa channel and partially contributes to NFκB activation while Orai1-mediated Ca2+ entry alone is sufficient for maximal NFAT activation and partial NFκB activation. Together these findings reveal the molecular events that determine activation of TRPC1 channels following store depletion. We suggest that local Ca2+ entry mediated by Orai1 determines plasma membrane insertion of TRPC1 while gating by STIM1 controls its activation. Thus, Orai1 and STIM1 not only determine Ca2+ signals generated by CRAC channels but by regulating TRPC1 channel activity rapidly modulate [Ca2+]i and thus significantly impact various cell functions.
Compared to SOCE in control HSG cells (transfected with vector or scrambled siRNA; black traces in Figure 1), knockdown of endogenous Orai1, STIM1, or TRPC1 attenuated thapsigargin (Tg)-stimulated Ca2+ influx by >90%, >80%, or >60%, respectively (Figure 1A). These conditions did not significantly affect internal Ca2+ release. Western blots (Figure S1A) demonstrate the effectiveness of TRPC1 knockdown in these cells. Ca2+ entry induced by Tg treatment of HSG cells was blocked by 1 µM Gd3+ and 20 µM 2APB (Figure S1B). Further, expression of TRPC1, TRPC1+STIM1, Orai1+STIM1, or TRPC1+STIM1+Orai1 increased Tg-stimulated Ca2+ entry (Figure S1G), which was also blocked by 1 µM Gd3+ and 20 µM 2APB (Figure S1C–F). Together, these data are consistent with our previous studies [42] that Orai1, STIM1, and TRPC1 contribute to endogenous SOCE in HSG cells. Additionally, the contributions of TRPC1, STIM1, and Orai1 to SOCE were not dependent on the level of stimulation (Figure S2). The relative decrease in SOCE induced by individual knockdown of the three proteins was similar in cells stimulated with 100 µM carbachol (CCh, a maximal stimulatory concentration) or 1 µM CCh (submaximal stimulatory concentration).
The contribution of TRPC1 and Orai1 to SOCE in HSG cells was further examined by using whole cell patch clamp technique [2],[16],[17],[40] to record the current generated by intracellular Ca2+ store depletion (Figure 1B). Consistent with our previous findings, Tg stimulation of cells resulted in activation of ISOC in HSG cells that is distinct from the typical ICRAC currents measured in RBL cells and T lymphocytes [40]. We have previously reported [40] that ISOC is a relatively Ca2+-selective cation current with Erev around +20 mV and pCa2+/pNa+ = 40 (ICRAC displays Erev>+60 mV and Ca2+/Na+ selectivity ≥400). Silencing of Orai1 expression blocked generation of ISOC while knockdown of TRPC1 by shRNA significantly reduced the amplitude of the inward current but induced more pronounced loss of the outward current. Thus the residual current detected in 6/10 shTRPC1 treated cells was more inwardly rectifying, i.e. more like ICRAC (Figure 1B, blue trace). These findings indicate the possibility that ICRAC in HSG cells can be masked by the larger relatively non-selective TRPC1-mediated current that is activated under the same conditions. The extent of TRPC1 knockdown would then determine the detection ICRAC. In the present set of experiments, 40% of the cells displayed ISOC or reduced ISOC. Our present data are somewhat contradictory to our previous finding that the residual current in Tg-stimulated submandibular gland acinar cells from TRPC1−/− mice was a much reduced transient current that was linear and did not display ICRAC-like properties (i.e. activation by low [2APB] or increase in DVF medium) [29]. We suggest that other TRPC channels or volume-regulated channels could account for the linear current. While further studies are required to determine the channel(s) involved in this residual current, our previous findings strongly demonstrate that TRPC1 contributes to SOCE and is critically required for salivary gland fluid secretion.
The two C-terminal residues of STIM1(684KK685) mediate gating of TRPC1 via electrostatic interaction with TRPC1(639DD640) residues [47]. Consistent with this, expression of a STIM1 mutant that lacks ability to gate TRPC1, STIM1(684EE685), induced suppression of SOCE in HSG cells while expression of WT-STIM1 resulted in a small increase in function (Figure 2A). Expression of the TRPC1 mutant that cannot be gated by STIM1, TRPC1(639KK640), induced a similar suppression of endogenous SOCE (Figure 2A, blue trace). Further, TRPC1 was not activated by store depletion when expressed with STIM1(684EE685) in HEK293 cells (Figure S3A), but when STIM1 and TRPC1 mutants were expressed together (i.e. “charge swap” between the proteins) there was recovery of SOCE (Figure S3A). Importantly, STIM1(684EE685) stimulated Orai1 similar to WT-STIM1 (Figure S3B). A key finding of this study, shown in Figure 2B, is that expression of STIM1(684EE685) resulted in generation of ICRAC in response to Tg-induced Ca2+ store depletion in >70% of HSG cells displaying currents. Together the data in Figures 1B and 2B suggest that ISOC in HSG cells is composed of a small Orai1-mediated ICRAC and a larger TRPC1-mediated non-selective current (note that we have not yet measured an isolated TRPC1+STIM1 current).
To conclusively demonstrate that endogenous Orai1 mediates ICRAC in HSG cells we expressed the STIM1-Orai1-activating region (SOAR) [20]. A large increase in basal Ca2+ entry (Figure 2C) and spontaneous ICRAC was seen in these cells (Figure 2D). SOAR-induced spontaneous SOCE was abolished by knockdown of endogenous Orai1 but was not affected by knockdown of endogenous TRPC1 (Figure 2C). In contrast, Tg-stimulated Ca2+ entry in SOAR-expressing cells was significantly reduced by knockdown of TRPC1 (Figure 2E, the residual Ca2+ entry reflects spontaneous Orai1-dependent Ca2+ influx). In aggregate, these data provide strong evidence that endogenous Orai1 mediates ICRAC without any contribution from TRPC1 while SOCE and ISOC display significant contribution from TRPC1. Importantly, the function of TRPC1 requires Orai1.
To identify the mechanism involved in regulation of TRPC1-SOC channels we examined the effect of intracellular Ca2+ store depletion on the surface expression of TRPC1. In resting cells the surface expression of TRPC1 (i.e. in the biotinylated fraction) was relatively low. Tg treatment of cells (Figure 3A, left panel, total TRPC1 and GAPDH are shown in input) significantly enhanced (about 3-fold, Figure 3C) the insertion of TRPC1 into the plasma membrane. An important finding of this study (Figure 3A) is that Tg-stimulated increase in plasma membrane insertion of TRPC1 was dependent on Orai1. Decreasing Orai1 expression or compromising Orai1 function by expression of the dominant negative mutant Orai1-E106Q (Figure 3A, middle and right panels, respectively, see Figure 3C for average data) severely reduced Tg-stimulated surface expression of TRPC1 without significantly affecting the resting level of TRPC1. To examine whether Ca2+ entry was involved in TRPC1 trafficking, biotinylation of TRPC1 was assessed in cells stimulated with Tg in nominally Ca2+-free medium or in normal Ca2+-containing medium with 1 µM Gd3+. Both conditions blocked the increase in the surface expression of TRPC1 induced by Tg (Figure 3B and C). These effects on TRPC1 trafficking were not due to loss of TRPC1/STIM1/Orai1 clustering, which was not affected in cells expressing Orai1-E106Q [39] or in the absence of external Ca2+ (unpublished data). The role of Orai1-mediated Ca2+ entry was more directly assessed by using Orai1-E106D, an Orai1 mutant that is permeable to Ca2+ in Ca2+-containing medium, but unlike the wild type channel, it is permeable to Na+ in nominally Ca2+-free medium. Tg treatment of cells expressing this mutant induced surface expression of endogenous TRPC1 in Ca2+-containing medium but not in Ca2+-free medium (Figure 3D). Finally, trafficking of TRPC1 was examined in HSG cells expressing STIM1(684EE685), which display ICRAC in response to Ca2+ store depletion (see Figure 2B). Although TRPC1 activation was suppressed in these cells, trafficking of the channel was not altered (Figure 3E). In aggregate these novel data suggest that Orai1-mediated Ca2+ influx is sufficient for plasma membrane insertion of TRPC1 but not activation; the latter depends on STIM1.
The mechanism involved in the clustering of TRPC1 with STIM1 and Orai1 was assessed by TIRFM. Ca2+ store depletion resulted in co-localization of YFP-TRPC1 and Orai1-CFP into puncta in the sub-plasma membrane region (Figure 4A, HA-STIM1 was co-expressed in these cells). Further, STIM1 co-clustered with both the channels following Tg stimulation of the cells (Figure 4B). As has been reported for Orai1, Orai1-TRPC1 clustering also required co-expression of STIM1 (unpublished data) and was not detected in cells when endogenous STIM1 expression was knocked down (Figure 4C). More significantly, co-IP of endogenous TRPC1 and Orai1 was abolished in cells treated with siSTIM1 (Figure 4D) but not in cells expressing STIM1(685EE685) (Figure 4E,F). TRPC1 clustering was not dependent on Orai1 since co-clustering of TRPC1 with STIM1 was unaffected by knockdown of Orai1 (Figure S4, compare data in A and B). Thus, STIM1 determines TRPC1 clustering in the sub-plasma membrane region following Ca2+ store depletion, and Orai1-mediated Ca2+ entry regulates its surface expression. Based on these findings we hypothesize that TRPC1 is present in recycling vesicles that traffic in and out of the plasma membrane region. Following store depletion when STIM1 clusters in ER/PM junctional domains, it interacts with TRPC1 possibly via the ERM domain [46] and increases the retention of TRPC1-containing vesicles. Concurrently, STIM1 also recruits Orai1 into the same regions, thus bringing the two channels in close proximity to each other. Ca2+ entry via Orai1 induces fusion of TRPC1-containing vesicles to the plasma membrane followed by gating of the channel by STIM1. Further studies are required to elucidate the mechanisms involved in trafficking and plasma membrane insertion of TRPC1.
We next examined whether relatively global or local [Ca2+]i increase regulates plasma membrane insertion of TRPC1. Figure 5A shows that loading HSG cells with 200 µM BAPTA-AM prior to Tg stimulation (details given in Methods) did not suppress trafficking of TRPC1 induced by Tg, although Tg-stimulated global [Ca2+]i increase was completely suppressed (Figure 5B, compare red trace with black trace, which shows [Ca2+]i increase in cells loaded with low [BAPTA-AM]). In addition, Tg-stimulated ISOC was not altered by replacing EGTA in the pipette solution with 10 mM BAPTA (Figure S5B,C), although the latter condition completely suppressed KCa activation in Tg-stimulated cells (Figure S5C, right panel). TRPC4 and TRPC5 are directly activated by elevation of intracellular [Ca2+]i [49], and a recent study demonstrated that Ca2+ entry mediated via Orai1 or other Ca2+ entry channels, including voltage-dependent channels, can directly enhance TRPC5 activity [50]. To determine whether [Ca2+]i increase directly activates TRPC1, whole cell current measurement was done with [Ca2+] in the pipette solution clamped to 0.1 µM or 1 µM. No current was detected with 0.1 µM Ca2+ (unless Tg was included in the external medium, Figure S5A, black and blue traces), 1 µM Ca2+ (Figure S5A, red trace), or up to 5 µM Ca2+ (unpublished data). Note that 1 µM Ca2+ induces >90% activation of TRPC4 and TRPC5 [49],[50]. These data also rule out possible contribution of other Ca2+-dependent cation channels to SOCE [51].
In aggregate, these data suggest that local Ca2+ entry via Orai1 determines plasma membrane insertion of TRPC1 and that [Ca2+]i elevation due to intracellular Ca2+ release is insufficient for triggering TRPC1 insertion. Further when cells were treated with Tg in a Ca2+ free medium for 5 min, there was no increase in TRPC1 expression in the plasma membrane until Ca2+ was added to the external solution (Figure 5C, right panel). As shown above, when cells were stimulated with Tg in a Ca2+-containing medium (Figure 5C, left panel), TRPC1 insertion in the plasma membrane was enhanced. Surprisingly, subsequent removal of Ca2+ from the external solution (for 10 min) did not change the level of TRPC1 in the surface membrane. Functional consequences of these treatments are shown in Figure 5E–F. In this experiment, HSG cells were treated with Tg in Ca2+-free medium prior to whole cell current measurements in DVF medium. Typical inwardly rectifying ICRAC with rapid inactivation was detected in these cells (Figure 5E), consistent with the lack of TRPC1 insertion in the plasma membrane under these conditions. However, when pre-treatment was done in Ca2+-containing medium, ISOC was detected in the DVF medium (Figure 5F). Note that the ISOC in DVF was relatively sustained, again consistent with the stable biotinylation of TRPC1. In aggregate, the findings presented above suggest that Orai1-mediated Ca2+ entry triggers insertion of TRPC1 in the plasma membrane, followed by activation of the channel by STIM1. Thus while channel insertion into the plasma membrane appears to depend on local increases in [Ca2+]i, TRPC1 internalization does not strictly depend on a decrease in [Ca2+]i. Further studies will be required to determine the exact molecular mechanisms involved in internalization of TRPC1.
The data presented above demonstrate that two STIM1-gated channels, Orai1 and TRPC1, are activated in response to internal Ca2+ store depletion in HSG cells. To establish the relative contributions of these channels in SOCE-mediated Ca2+ signaling, we examined three SOCE-activated mechanisms: KCa channel, NFκB, and NFAT. Figure 6A demonstrates that expression of STIM1(684EE685) in HSG cells induced a slow, much diminished (>80% reduction), and transiently activated KCa current compared to that in control cells. As shown above (Figure 2B), only CRAC channel activation was seen in cells expressing this STIM1 mutant. Thus, Orai1-mediated Ca2+ entry does not appear to be sufficient for activation of KCa activity following Tg stimulation. Further, NFκB activation (Figure 6B) was significantly decreased by the knockdown of TRPC1 expression, and predictably knockdown of Orai1 induced an even greater loss of activity. Significantly, expression of SOAR did not lead to much activation of NFκB. Remarkably, TRPC1 had minimal contribution to the regulation of NFAT since knockdown of Orai1 but not TRPC1 suppressed NFAT activation (Figure 6C). Thus, Orai1-mediated Ca2+ entry is sufficient for regulation of NFAT and for partial stimulation of NFκB, but not for KCa activation. In contrast, TRPC1-mediated Ca2+ entry regulates KCa channel activity and contributes to NFκB signaling, but not NFAT activation.
Similar to the findings in HSG cells, KCa activity was severely reduced in acinar cells from submandibular glands of TRPC1−/− mice, which could account for loss of salivary fluid secretion in these animals [29]. While our current findings suggest that Orai1+STIM1 dependent regulation of TRPC1 would be very critical for regulating salivary gland function, functional interaction between these proteins will depend on their precise localization within acinar cells, as is required in HSG cells (Figure 4B). We have previously reported that TRPC1 is localized in the basal and lateral regions of submandibular gland acinar cells [29],[52] and that TRPC1 and STIM1 co-IP following stimulation of acini by either Tg or CCh [53]. To determine possible physiological relevance of the present findings, we examined the localization of TRPC1, Orai1, and STIM1 in submandibular glands excised from resting and pilocarpine-stimulated mice (tissue was fixed in vivo in mice following pilocarpine injection and after an increase in saliva secretion was detected). In the samples from unstimulated mice, endogenous Orai1 was prominantly detected in the apical and lateral regions of submandibular gland acini (Figure S6A, upper panels, green signal, Orai1 signal shown by white arrows), co-localization of Orai1 with the luminal membrane protein AQP5 is also shown (red signal, right panel). STIM1 showed diffused localization within the acinar cells from unstimulated mice (Figure S6B, red signal, upper panel). Consistent with our previous findings, diffuse localization of TRPC1 was detected in the basal and lateral regions (green signal, upper panels, the same sections were labeled for STIM1 and TRPC1). In samples obtained from stimulated mice, Orai1 and AQP5 localization did not change (Figure S6A, lower panels). However, a dramatic translocation of TRPC1 and STIM1 to the basal and lateral membrane regions was seen with relative decrease in intracellular staining (Figure S6B, lower panels, see white arrows). Thus stimulation induces co-localization of STIM1, Orai1, and TRPC1 in the lateral membrane region of cells. While further studies are required to determine whether sufficient Orai1 is present in the basolateral membrane to regulate TRPC1, our data strongly suggest that regulation of TRPC1 by STIM1 and Orai1 is feasible within the lateral membrane region of salivary gland acinar cells. Our findings are generally consistent with the strong co-localization of Orai1 and STIM1 in the lateral membrane region of stimulated pancreatic acinar cells [54]. STIM1 was also localized in the basal membrane and co-localized with heterologously expressed, but not endogenous, Orai1, in these cells. This study suggested that localization of Orai1 and STIM1 in the lateral membrane was consistent with the proposed site of Ca2+ entry in exocrine acinar cells [55]–[57].
The findings described herein address several important and as-yet unresolved questions regarding the molecular components of TRPC1-SOC channel, the mechanism involved in regulation of the channel in response to store depletion, and its contribution to SOCE. We report that the previously described ISOC [39],[40],[58], which is stimulated by store depletion and dependent on TRPC1, STIM1, and Orai1, is a sum of Orai1/STIM1-mediated ICRAC and TRPC1/STIM1-mediated non-selective cation current. Our findings suggest that the latter relatively larger current masks the underlying ICRAC since suppression of TRPC1 function either by shTRPC1 or by expression of the STIM1(684EE685) mutant, which does not gate TRPC1, facilitates detection of ICRAC. Further, SOAR-activated ICRAC required Orai1 but not TRPC1. Thus Orai1 and TRPC1 are components of two distinct channels. These findings provide strong argument against the possibility that TRPC1 and Orai1 contribute to the same channel pore or that Orai1 is a regulatory, non-conducting, subunit of TRPC channels [59].
We also report that Orai1-mediated Ca2+ entry triggers plasma membrane recruitment of TRPC1. These data reveal a novel function for Orai1 that can explain its critical requirement in the activation of TRPC1 channels following Ca2+ store depletion. We show that Ca2+ store depletion leads to enhanced surface expression of TRPC1, which is blocked when Ca2+ is removed from the external medium or SOCE is inhibited by addition of Gd3+. Knockdown of endogenous Orai1 expression or expression of non-functional Orai1 mutants (Orai1-E106Q) also lead to loss of TRPC1 in the plasma membrane. Notably, in cells expressing Orai1-E106D, TRPC1 trafficking is supported in Ca2+-containing medium but not Ca2+-free medium. Together, these findings provide strong evidence that surface expression of TRPC1 is determined by the Ca2+ permeability of Orai1 and that TRPC1 is gated by STIM1 and not directly by [Ca2+]i increase. Presently we cannot conclusively rule out the involvement of possible downstream signaling pathway(s) activated by Orai1-mediated Ca2+ entry.
The data presented above also reveal important aspects of TRPC1, Orai1, and STIM1 clustering that are critical in the regulation of TRPC1 within the same ER/PM junctional domains where Orai1 is regulated by STIM1. We show that in response to store depletion TRPC1 co-clusters with STIM1 and Orai1. More importantly while Orai1 is not required for clustering and association of TRPC1-STIM1, localization of STIM1 in the ER/PM junctional domains is critical for recruitment and association of Orai1 and TRPC1. Thus far there are no data to show that TRPC1 and Orai1 directly interact with each other, although both channels interact with STIM1. STIM1 interacts with Orai1 via the SOAR domain, which also leads to gating of the channel. In the case of TRPC1 while the C-terminal 684KK685 residues of STIM1 are involved in gating the channel, the ERM domain [46] could interact with the channel and serve as a scaffold to retain TRPC1 within the ER/PM junctional regions. We suggest that interaction with STIM1 allows the channels to be localized in close proximity to each other, facilitating Orai1-mediated Ca2+ entry to locally regulate plasma membrane insertion of TRPC1. However, our data show that internalization of TRPC1 is apparently not dependent on [Ca2+]i (Figure 5C). Thus, TRPC1 can remain active provided the Ca2+ stores are depleted and STIM1 is localized in the peripheral domains. Based on our data, we suggest the following sequence of events in the activation of TRPC1: (i) Ca2+ store depletion leads to translocation of STIM1 to ER/PM junctional domains and recruitment of Orai1 (localized within the plasma membrane) and TRPC1 (likely localized in intracellular trafficking vesicles), (ii) Orai1 is activated by STIM1 and Ca2+ entry via Orai1 triggers exocytosis of TRPC1, and finally (iii) STIM1 gates plasma membrane TRPC1 (depicted in the model shown in Figure 7).
We also demonstrate the unique contributions of TRPC1 and Orai1 to SOCE. Remarkably, different cellular functions are regulated when Orai1 alone is activated compared to conditions when both channels are activated. Our data suggest that TRPC1 augments the [Ca2+]i increase resulting from Orai1-mediated Ca2+ entry. Consistent with this, TRPC1-mediated Ca2+ entry is required for KCa function and contributes to NFκB activation, both of which require relatively higher [Ca2+]i, but not for NFAT activation, which can be activated at lower [Ca2+]i (see Figure 7) [29],[60]. Interestingly, the requirement of TRPC1 for KCa activity is similar to our previous finding that submandibular gland acinar cells from TRPC1−/− mice display loss of sustained KCa activity, which accounts for the decrease in fluid secretion in these glands. We have previously shown that TRPC1 is localized in the basal and lateral regions of acinar cells [29],[52] and that TRPC1 and STIM1 associate following stimulation of acini [53]. Since Orai1 is critical for TRPC1 function, localization of these proteins in the salivary gland acinar cells is a key determinant for the functional interaction between them. Feasibility for the interaction of the three proteins and regulation of TRPC1 in the gland is demonstrated by our data (Figure S6), showing that following agonist stimulation Orai1, TRPC1, and STIM1 are strongly co-localized in the lateral membrane region of acinar cells while TRPC1 and STIM1 also appear to be colocalized in the basal region. In salivary gland acinar cells agonist stimulation leads to [Ca2+]i elevation, which is first detected in the apical region of the cells and then spreads to basal and lateral regions, irrespective of the level of stimulation [55],[61]. Although further studies will be required to confirm the presence of Orai1 in the basal membrane region of acini, co-localization of TRPC1, Orai1, and STIM1 in the lateral membrane region of stimulated cells supports our suggestion that Orai1 can regulate TRPC1 function in this region and thus modulate SOCE.
In conclusion, the data described above reveal novel insight into the molecular components and regulation of TRPC1-SOC channels. Our findings provide strong evidence that TRPC1 and Orai1 constitute distinct SOC and CRAC channels, respectively, both of which are gated by STIM1 in response to store depletion and contribute to SOCE in the same cell. The critical step in the activation of TRPC1 is its insertion into the plasma membrane, which is governed by Orai1-mediated local Ca2+ entry. In addition to gating TRPC1 and Orai1, STIM1 also mediates the association of the two channels within discrete ER/PM junctional domains, which is the site for SOCE [19],[21]. The three proteins are also co-localized in the membrane region predicted to be the site of SOCE in acinar cells [56],[57], thus highlighting the potential physiological relevance of our findings. Importantly, TRPC1 augments Ca2+ entry mediated by Orai1-CRAC channels and is required for activation of KCa channels and NFκB, but not NFAT, signaling. As has been suggested, the amplitude, frequency of oscillations, or spatial patterning of [Ca2+]i changes determines the regulation of different cell functions [1],[4],[51],[55],[60],[62]. Although further studies are required to elucidate exactly how TRPC1 alters the primary [Ca2+]i signals generated by Orai1, the present data suggest that regulation of TRPC1 trafficking can provide a mechanism for rapidly modulating [Ca2+]i. STIM1 is emerging as a versatile ER Ca2+ sensor that regulates multiple target proteins in response to Ca2+ store depletion. In addition to activation of Orai1 and TRPC channels, STIM1 has been reported to inhibit Cav1.2 channels [63],[64] and activate adenylyl cyclase [65], both of which depend on Ca2+ store depletion. While regulation of TRPC1 and Cav1.2 require association of the channels with Orai1 within ER/PM junctional domains, Orai1 function does not appear to be involved in STIM1-dependent inhibition of Cav1.2. Thus Orai1 and STIM1 by coordinating the regulation of other ion channels and signaling components can modulate [Ca2+]i and critically impact SOCE-mediated Ca2+ signaling and a variety of cellular functions.
HSG cells were cultured in MEM medium, supplemented with 10% heat-inactivated fetal bovine serum, and 1% penicillin/streptomycin. Sequences for the siOrai1, siSTIM1, and shTRPC1 targeting to human Orai1, STIM1, and TRPC1, respectively, were similar to previously described sequences [42]. All siRNA duplexes were obtained from Dharmacon. Lipofectamine RNAiMAX (Invitrogen) was used for siRNA transfection while Lipofectamine 2000 was used for other plasmids. Cells were typically transfected 24 h after plating and experiments were performed 48 h post-transfection. All other reagents used were of molecular biology grade obtained from Sigma Aldrich unless mentioned otherwise.
Fura-2 fluorescence was measured in single HSG cells cultured for 24 h in glass bottom MatTek tissue culture dishes (MatTek Corp. Ashland, MA) and transfected as required; experiments were done 48 h post-transfection. Cells were loaded with 5 µM Fura-2 (Invitrogen) for 30 min at 37°C. Fluorescence was recorded using a Till Photonics-Polychrome V spectrofluorimeter and MetaFluor imaging software (Molecular Devices). Each fluorescence trace (340/380 nm ratio) represents an average from at least 50–150 cells from >6 individual experiments. Student's t test was used to statistically evaluate the data.
Coverslips with HSG cells were transferred to the recording chamber and perfused with Ca2+ containing standard external solution (Ca2+-SES) with the following composition (in mM): NaCl, 145; KCl, 5; MgCl2, 1; CaCl2, 1; Hepes, 10; glucose, 10; pH 7.4 (NaOH). The patch pipette had resistances between 3 and 5 milliohms after filling with the standard intracellular solution that contained the following (in mM): cesium methane sulfonate, 145; NaCl, 8; MgCl2, 10; Hepes, 10; EGTA, 10; pH 7.2 (CsOH). For KCa measurements, pipette solution contained 150 mM KCl, 2 mM MgCl2, 1 mM Mg-ATP, 5 mM Hepes, 0.1 mM EGTA, and pH 7.2, potassium hydroxide. Osmolarity for all the solutions was adjusted with mannose to 300±5 mosM using a vapor pressure Osmometer (Wescor, Logan, UT). All electrophysiological experiments were performed in the tight-seal whole cell configuration at room temperature (22–25°C) using an Axopatch 200B amplifier (Molecular Devices). Development of the current was assessed by measuring the current amplitudes at a potential of −80 mV, taken from high resolution currents in response to voltage ramps ranging from −90 to 90 mV over a period of 100 ms imposed every 2 s (holding potential was 0 mV) and digitized at a rate of 1 kHz. Liquid-junction potentials were less than 8 mV and were not corrected. Capacitative currents and series resistance were determined and minimized. For analysis, the current recorded during the first ramp was used for leak subtraction of the subsequent current records. Thapsigargin (Tg 1 µM), dissolved in the bath solution, was used to stimulate the cells. DVF solution contains (mM): NaCl 165; CsCl 5; EDTA 10; HEPES 10; glucose 10; pH 7.4 (NaOH). Cells were pretreated with 1 µM Tg for 10 min either in Ca2+ containing or Ca2+ free medium before whole cell configuration was achieved. Cells were switched to DVF 1 min after achieving whole cell configuration in Ca2+ free external medium.
Transfected HSG cells were washed with phosphate-buffered saline (PBS) and lysed in radioimmunoprecipitation assay (RIPA) protein extraction buffer (50 mM Tris-HCl, 150 mM NaCl, 0.1% sodium dodecyl sulfate (SDS), 0.5% sodium deoxycholate, 1% Triton X-100, 2 mM EDTA, 1 mM dithiothreitol (DTT), pH 7.4) supplemented with Complete Protease Inhibitor Cocktail tablets (Roche Diagnostics). Where indicated, cells were first stimulated for 5 min with 1 µM Thapsigargin (Tg), lysates was then centrifuged at 12,000 x g for 30 min at 4°C, and the supernatant was collected and analyzed by SDS-PAGE and Western blotting (50 µg of protein were loaded per lane). Protein concentrations in the lysate was adjusted to 2 mg/ml and incubated with 10 µg/ml IP antibody. Immunoprecipitates were released by incubating in SDS-sample buffer and resolved in 4%–12% NuPAGE gels (Invitrogen) followed by Western blotting. Anti-STIM1 (Cell signaling technology, Danvers, MA), anti-Orai1 (Open Biosystems, Huntsville, AL), anti-GAPDH (Abcam Inc, Cambridge, MA), and Anti-TRPC1 antibody [42] were used at 1∶1000, 1∶1000, 1∶10000, and 1∶400 dilution, respectively.
Cells were transfected with vector or scrambled control as required. For stimulation experiments, cells were pretreated with 1 µM Thapsigargin in the presence (+Ca2+) or absence (−Ca2+) of extracellular calcium, and incubation time was 5 min or otherwise as indicated. The reaction was stopped by adding ice-cold quenching solution. In BAPTA-AM loading experiments, cells were pretreated with 200 µM BAPTA-AM (Invitrogen) in SES containing 100 µM extracellular Ca2+ for 30 min at 37°C. Treated cells were then incubated for 20 min with 1.5 mg/ml Sulfo-NHS-LC-Biotin (Pierce) in 1XPBS (pH 8.0) on ice. Following biotin labeling, cells were washed and harvested in RIPA buffer using the same protocol as described above. Biotinylated proteins were pulled down with NeutrAvidin-linked beads (Pierce) and detected by Western blotting. Band intensities of surface proteins were obtained using Image J software.
NFκB-Luc, NFAT-Luc, and hRLuc-TK were obtained from Promega. HSG cells were transfected with the indicated constructs with either NFκB or NFAT reporter gene, and co-transfection with the Renilla luciferase gene (hRLuc-TK) driven by the TK promoter was used to control for cell number and transfection efficiency. Transfected cells were stimulated as described in [60]. Luciferase activity was measured with the Dual-Glo Luciferase Assay System (Promega). For each condition, luciferase activity was measured with four samples taken from duplicate wells with a 96-well automated luminometer (Turner Biosystems). Results are represented as the ratio of firefly to Renilla luciferase activity.
An Olympus IX81 motorized inverted microscope (Olympus) was used as described previously [42] using 447, 514, and 568 nm lasers for excitation of CFP, YFP, and mCherry, respectively, and a TIRF-optimized Olympus Plan APO 60x (1.45 NA) oil immersion objective and Lambda 10-3 filter wheel (Sutter Instruments) containing 480-band pass (BP 40 m), 525-band pass (BP 50 m), and 605-band pass (BP 52 m) filters for emission. Images were collected using a Hamamatsu EM C9100 camera (Hamamatsu) and the MetaMorph imaging software (Molecular Devices). MetaMorph was also used to measure the fluorescence intensity before and after stimulation with Tg. Briefly, regions of interest were selected to obtain the values for their fluorescence intensities during a time course experiment. These values were then plotted using the Origin 8 software (OriginLab).
Balb/c mice were anesthetized and injected subcutaneously with either saline (Resting) or 0.5 mg of pilocarpine/kg (Stimulated). After the saliva secretion was observed in stimulated mice, the animals were perfusion fixed with 10% buffered formalin and immediately euthanized. Salivary glands were excised and embedded in paraffin for histologic processing. Slides of paraffin sections were deparaffinized and rehydrated. Sections were unmasked by microwaving samples for 10 min in a microwave pressure cooker (NordicWare) in 1 mM EDTA, pH 8.0, containing 0.05% Tween 20. After cooling, sections were blocked either with 0.5% BSA in PBS (for direct conjugates) or with 10% donkey serum in PBS (for samples using secondary antibodies). After blocking for 30 min at room temperature, primary antibodies were applied and incubated at 4°C overnight. For samples using two or more rabbit host primary antibodies, direct-conjugation with a fluorescent tag using Invitrogen's Zenon labeling kit was used. For antibodies requiring secondary antibody labeling, donkey anti-rabbit Alexa conjugates were used (Invitrogen). A negative control using normal rabbit IgG at the same concentration as specific primaries was included for both methods. After labeling with primary antibodies only, samples were washed extensively and incubated with secondary antibodies for 1 h at room temperature, washed, and mounted with VectaShield mounting medium containing DAPI. Zenon conjugated samples were washed extensively and mounted with cover slips as above. Images were collected by using a Leica Confocal microscope and MetaMorph software (Molecular Devices, Sunnyvale, CA).
Data analysis was performed using Origin 8 (OriginLab). Statistical comparisons were made using student's t test. Experimental values are expressed as means ± SD or SEM. Differences in the mean values were considered to be significant at p<0.01.
|
10.1371/journal.pntd.0006041 | Molecular characterization and phylogenetic relatedness of dog-derived Rabies Viruses circulating in Cameroon between 2010 and 2016 | Rabies is enzootic among dog populations in some parts of Cameroon and the risk of human rabies is thought to be steadily high in these regions. However, the molecular epidemiology of circulating Rabies Virus (RABV) has been hardly considered in Cameroon as well as in most neighboring central African countries. To address this fundamental gap, 76 nucleoprotein (N) gene sequences of dog-derived RABV were obtained from 100 brain specimens sampled in Cameroon from 2010 to 2016. Studied sequences were subjected to molecular and phylogenetic analyses with reference strains retrieved from databases. The 71 studied Africa-1 isolates displayed 93.5–100% nucleotide (nt) and 98.3–100% amino-acid (aa) identities to each other while, the 5 studied Africa-2 isolates shared 99.4–99.7% sequence similarities at nt and aa levels. Maximum Likelihood based phylogenies inferred from nucleotide sequences confirmed all studied RABV isolates as members of the dog-related species 1 of the Lyssavirus genus. Individual isolates could be unambiguously assigned as either the Africa-1 subclade of the Cosmopolitan clade or the Africa 2 clade. The Africa-1 subclade appeared to be more prevalent and diversified. Indeed, 70 studied isolates segregated into 3 distinct circulating variants within Africa-1a lineage while a unique isolate was strikingly related to the Africa-1b lineage known to be prevalent in the neighboring Central African Republic and eastern Africa. Interestingly, all five Africa-2 isolates fell into the group-E lineage even though they appeared to be loosely related to databases available reference RABV; including those previously documented in Cameroon. This study uncovered the co-circulation of several Africa-1 and Africa-2 lineages in the southern regions of Cameroon. Striking phylogenetic outcasts to the geographic differentiation of RABV variants indicated that importation from close regions or neighboring countries apparently contributes to the sustainment of the enzootic cycle of domestic rabies in Cameroon.
| Rabies has been repeatedly reported among dog populations in Cameroon, especially in Yaounde, its capital city. However, the relative rates and genetic variability of Rabies Virus (RABV) variants circulating among dog populations in Cameroon are still to be documented. This study aimed to estimate the frequency and genetic diversity of RABV isolates originating from rabid dogs in the southern regions of Cameroon from 2010 to 2016. Overall, 76 of the 100 dog-derived RABV isolates sampled in Cameroon from 2010 to 2016 were successfully characterized. Our findings revealed that studied isolates belonged to the dog-related species 1 of the Lyssavirus genus, specifically 70 Africa-1a, 1 Africa-1b and 5 Africa-2 group-E lineages. The general phylogenetic pattern suggested an in-country geographic differentiation of the circulating RABV variants. This apparent geographic differentiation was contradicted by striking outcasts indicating importation from close or distant regions. Overall, this study uncovered the co-circulation of several Africa-1 and Africa-2 lineages in some southern regions of Cameroon, thus providing base-line molecular data that would be of interest for future stages of implementation of the rabies surveillance and control plan that is being setup in Cameroon.
| Rabies is a neglected lethal neurological disease which has a case-fatality rate of almost 100% [1–3]. It causes an estimated 59,000 human deaths primarily in developing and low-income countries where the disease is endemic in animal populations [1,3]. Bites from rabid domestic dogs account for over 99% of the human rabies cases [3,4], most of which occur in Asia and Africa [5,6]. Post-exposure prophylaxis (PEP) efficiently prevents disease development in humans bitten by rabid animals when administrated immediately. Unfortunately, PEP is often unavailable in all settings or not affordable in many developing countries [3].
Canine rabies has been shown to be endemic in Cameroon and relatively higher frequencies of rabid dogs have been reported in urban settings compared to rural areas [7,8]. In the absence of a multiannual active national surveillance and control strategy, the actual burden of human and animal rabies in Cameroon is likely underestimated as in other African countries [3,9–12]. Some rabies control interventions, such as yearly discount of pet vaccination and irregular radio communication campaigns, are conducted in Cameroon but their actual impact remains unknown [8,9,13].
The etiological agent of rabies, the Rabies Virus (RABV), of the genus Lyssavirus and family Rhabdoviridae, has a single-stranded RNA genome of approximately 12 kb in length and of negative polarity [14]. The RABV genome consists of five genes encoding the nucleoprotein (N), the phosphoprotein (P), the matrix protein (M), the glycoprotein (G) and the large protein which is the polymerase (L). These five genes N, P, M, G, and L are separated by intergenic regions of variable lengths [14,15]. Like other RNA viruses, RABV displays high rates of mutation due to the lack of proofreading activity of the L protein [16]. RABV is the only virus among the 16 known Lyssavirus species found worldwide in a wide range of mammalian reservoirs of the orders Chiroptera and Carnivora [17–20]. It has been recently demonstrated that individual gene or complete genome sequences of RABV isolates segregate into two major phylogenetic clusters gathering bat- and dog-derived RABV isolates respectively. Within these clades, isolates fall into several major clades [21].
Bat-derived RABV isolates have been shown to circulate specifically in the New World mainly among bats and, to a lesser extent, in some terrestrial carnivores such as skunks (Mephitis mephitis) and raccoons (Procyon lotor) [22–25]. Conversely, dog-specific RABV isolates have been documented worldwide mainly among domestic dogs, but also among wild-living carnivores comprising foxes and raccoon dogs in Europe [26], foxes in the Middle East [27], raccoon dogs and ferret-badgers in Asia [28–30], skunks, foxes, coyotes and mongooses in the Americas [23,24], African civet and mongooses in Africa [31,32]. Within the divergent dog-specific cluster, six major well-defined clades, respectively assigned as Africa-2, Africa-3, Arctic-related, Asian, Cosmopolitan and Indian clades, have been documented [21,33].
In particular, molecular studies in Africa uncovered the presence of Cosmopolitan, Africa-2, and Africa-3 clades. All these clades comprise classical RABV species that segregate into several subclades and lineages varying by geographic area, virus variability, and reservoir species in Africa [21,33–37]. Two major subclades are defined by field RABV isolates of the Cosmopolitan clade: Africa-1 and Africa-4. Africa-4 has been recently identified in northern Africa [21,34] while Africa-1 subclade has been shown to circulate in the northern, eastern and southern parts of Africa [33,38]. Africa-1a lineage has been suggested to have a very broad distribution across Africa. It is predominant in northern and eastern Africa [38–40] and has also been previously reported in Cameroon, Gabon, Equatorial Guinea, Ghana and Madagascar [38,41,42]. Africa-1b lineage circulates mainly in eastern and southern Africa [38,39,43,44].
Africa-2 lineages are uninterruptedly found across West and Central Africa [35–37,42] and has been shown to co-circulate with the Africa-1 lineages in Nigeria and Central African Republic [36,38,42,44–46]. Although Africa-1 and Africa-2 lineages have been documented in several domestic and wild carnivore species, domestic dogs are virtually the only population essential for maintaining canid variants in some parts of Africa [47]. Conversely, wild canids have been suggested to contribute to the sustainment of canine rabies cycles in specific geographic locations in South Africa, Namibia and Zimbabwe [32,48–50].
The third Africa-3 clade is well adapted to mongooses [31,51,52] and is sustained through an independent epidemiological cycle (distinct from that of dog RABV) within viverrid species in southern Africa [31,32,51,53].
Although RABV has been continuously reported in Cameroon [7,8,54], its molecular epidemiology among dog populations have not yet been documented. Less than five genomic sequences of dog-related Africa-1 and Africa-2 RABV originating from Cameroon have been described in previous studies [21,36,38,44]. Based on very limited sequence data available so far in Cameroon, it was thought that there may be an in-country geographic differentiation of dog-derived RABV; with Africa-1 isolates found in the Center region including the capital city, Yaounde, and Africa-2 isolates detected in the northern part of Cameroon.
The purpose of this study was to provide insight into the frequency, genetic variability and phylogenetic relatedness of the RABV isolates derived from rabid dog in Cameroon. Interestingly, this study uncovered the co-circulation of both Africa-1 and Africa-2 in two southern regions of Cameroon and indicated that they circulate in close proximity between neighboring administrative regions of Cameroon as well as between Cameroon and neighboring countries.
From January 2010 to December 2016, a total of 163 animal specimens were analyzed for rabies diagnosis at CPC, comprising 159 specimens originating from domestic dogs and 4 specimens from other animal species (1 cat, 1 cow, 1 monkey and 1 pig). Overall, 65.4% (104/159) of all dog specimens analyzed were from the Center region amongst which 61.5% (64/104) originated from the capital city, Yaounde where CPC is located. All specimens from cat, cow, monkey and pig were found rabies-negative whereas 66.0% (105/159) of dog specimens were confirmed rabies-positive. The number of dog specimens analyzed from 2010 to 2016 as well as annual rates of positive samples were variable as depicted in Fig 1. Overall, 100 of the 105 rabies positive specimens were available for molecular characterization in this study. They originated from the Center region (68 samples amongst which 46 from Yaounde), East region (1 sample), Littoral region (3 samples), North West region (8 samples), West region (14 samples), South West region (3 samples) and South region (3 samples) (Table 1).
Of the 100 rabies-positive cases whose brains specimens were analyzed, 76 were confirmed by the molecular analyses performed in this study. The remaining 24 cases without molecular confirmation included 7 cases whose sequences were unexploitable (because of chromatograms with superimposed peaks and/or short reads), 4 cases which showed very weak amplification signals (insufficient for sequencing), and 13 cases which did not amplify. Comparison of the newly determined sequences with homologous sequences obtained from databases identified all 76 RABV isolates as strains of the Lyssavirus species 1 and specifically as Africa-1 or Africa-2 lineages. While 71 isolates were identified as belonging to the Africa-1 lineage displaying 93.5–100% nt and 98.3–100% aa identities to each other, 5 RABVs were closely related to Africa-2 isolates sharing 99.4–99.7% sequence similarities at nt and aa levels. Comparison between the studied and few database available Africa-1 RABV sequences from Cameroon showed 93.3–95.2% nt and 93.6–95.4% aa identities. The same analysis showed less sequence divergence between the studied and reference Africa-2 isolates from Cameroon: 96.4–99.0% nt and 94.5–99.7% aa sequence identities.
Africa-1 isolates, representing 71 of the 76 sequences obtained, were more prevalent and distributed across the southern regions of Cameroon (5 out of the 7 southern regions) (Fig 2). In contrast, Africa-2 isolates, detected from 2010 to 2014, were geographically restricted to the two neighboring South West and North West regions (Fig 2 and Table 1). This indicates that Africa-1 and Africa-2 RABV co-circulate in some regions of the southern part of Cameroon.
We analyzed the phylogenetic relationships of the nucleoprotein gene sequences (1040 nt) of the studied RABV isolates with database available homologous sequences representing RABV lineages originating from a wide geographic range in Africa (S1 Table). The resulting Maximum Likelihood phylograms confirmed all newly sequenced RABV isolates from Cameroon as dog-related Lyssavirus species 1.
Within the divergent clade of Cosmopolitan RABV, all 71 studied isolates fell in the Africa-1 subclade (Fig 3). None of them grouped with the recently reported Africa-4 subclade defined by RABV isolates from Egypt (in North Africa) and Israel (in Middle East) [34]. Within the Africa-1 subclade, the studied isolates segregated into two distinct lineages with a remarkable association to the geographic origin. One isolate (14V-4292) detected in 2014 in Garoua Boulai (East region) fell within the Africa-1b lineage with previously described isolates from Central African Republic (Fig 3). This isolate displayed 99.9% nt and 100.0% aa sequence identities with its closest match (GenBank N° KT119710) detected in 2008 in the Western part of Central African Republic. Identification of Africa-1b lineage in the East region of Cameroon is consistent with its high prevalence in eastern Africa and in Central African Republic [21,41,44]. This indicates the potential role of intercountry circulation in the sustainment of the enzootic cycle of rabies in Africa.
The other 70 studied isolates of the Africa-1 subclade grouped in the Africa-1a lineage along with previously described isolates from diverse geographic origin in North Africa (Algeria, Ethiopia, Nigeria, and Morocco) and Central Africa (Cameroon, Gabon and Equatorial Guinea) (Fig 3). Within the Africa-1a lineage, studied isolates fell into three distinguishable groups. The first group, (i), was defined by all viruses originating from Yaounde and other districts of the Centre region, indicating some association between the phylogenetic pattern and the regional origin of the RABV isolates from Cameroon. This first reliable group (bootstrap value of 96%) included sequences of RABV previously documented in the Centre region of Cameroon from 1992 to 2010 (Fig 3). However, the apparent feature of in-country geographic differentiation of the RABV was contradicted by several isolates: 12V-007 and 13V-4215, originating respectively from Bafoussam and Batcham in the West region; and 14V-4199 from Douala in the Littoral region. Indeed, they were strikingly related (99.8 to 100% nt and 100.0% aa identity) to their counterparts originating from the Centre region (Fig 3). The second distinguishable group, (ii), gathered Africa-1a isolates originating from the West and North West regions while third group (iii) was defined by Africa-1a isolates from the West, South West and North West regions (Fig 3). These results indicate that RABV circulate in close proximity between geographically close regions.
Interestingly all five newly sequenced Africa-2 RABV belonged specifically to the group-E (Fig 4). This Africa-2 group-E also comprised RABV strains previously reported in Cameroon and neighboring countries including Central African Republic, Chad, Niger and Nigeria [36,37]. Within the Africa-2 group-E, the newly described isolates were closely related to each other but were loosely related to databases available isolates, including those originating from the Northern regions of Cameroon from 1987 to 1994 (Fig 4). This study revealed that Africa-2 isolates circulate in two southern regions of Cameroon whereas the most prevalent Africa-1 lineage have been documented in five of the seven southern regions.
This study confirms previous reports suggesting continuous circulation of the RABV in Cameroon; especially in the capital city, Yaounde [7,8,54]. In the absence of a multiannual national surveillance and control plan, data on the actual burden of human and animal rabies in Cameroon is certainly underestimated as previously reported in comparable settings in Africa [3,9–12]. This fundamental gap prevents substantial conclusion about the geographical and temporal variation of rabies incidence in Cameroon (Figs 1 and 2).
This study is the first to provide data from elaborate phylogenetic analysis of RABV from Cameroon. Three clades of RABV have been previously documented as being specific to Africa: Cosmopolitan (Africa-1 and Africa-4 subclades), Africa-2 and Africa-3 clades [56]. In this study no isolate of the Africa-3 and Africa-4 was identified. This finding is consistent with the facts that Africa-3 and Africa-4 variants have been shown to be specific to southern [31,51,52] and northern Africa [34], respectively.
Based on the few previous molecular data, it was hypothesized that Africa-2 was exclusively found in the northern part of Cameroon whereas Africa-1 isolates were reported only in the southern regions. This study uncovered RABV of the Africa-2 group-E lineage in two southern regions of Cameroon (North West and South West regions) (Figs 2 and 4). Surprisingly, Africa-2 was the less prevalent lineage in this study whereas it has been shown to be uninterruptedly widespread in western and central Africa, including neighboring countries of Cameroon (Niger, Nigeria, Chad and Central African Republic) [35–37,42,44,45].
This study revealed Africa-1 of the Cosmopolitan clade as the most prevalent RABV lineage circulating in the southern regions of Cameroon (Table 1 and Fig 3). This observation is substantially true for the Centre region of Cameroon where all 51 RABV isolates that could be identified were assigned as Africa-1a lineage. Concerning specifically the North West and South West regions, which share borders with Nigeria, it might be possible to find that Africa-1 and Africa-2 co-circulate with comparable rates if more RABV isolates from these regions were characterized. The Africa-1 sub-clade of the RABV is predominant in the northern, eastern and southern parts of Africa [39,40,57], and have been shown to co-circulate with Africa-2 in Central African Republic [44,45,58] and Nigeria [35,36]. Interestingly, this study provides substantial evidence of co-circulation of Africa-1 and Africa-2 isolates of RABV in at least two regions of the southern part of Cameroon.
In accordance with previous reports, individual RABV sequences fell into a variety of groups, in association with the geographic origin. There was an apparent in-country geographic differentiation of the RABV, however few odds were observed. Similar findings suggesting region-specific variants of the RABV have been documented in some African countries [40,44], thus confirming genomic sequence relatedness as useful marker of intra- and inter-countries RABV dissemination among domestic dogs’ populations. An outstanding application of that marker is provided by the recent finding by Bourhy et al., suggesting that the maintenance of the enzootic cycle of rabies at local geographic level in Bangui is more likely driven by human-mediated waves of spread rather than by continuous dispersion in a relatively large and homogenous dogs’ population [45].
A limitation to this study was the fact that no specimen originated from the three northern regions of Cameroon (Fig 2). Furthermore, restricted geographic range covered by the study was associated to the fact that only few specimens originated from the East, Littoral and South regions while as high as 68.0% (68/100) of all specimens were from Yaounde and its neighborhoods (Fig 2). These shortcomings prevent final conclusion to be drawn on the relative rates and genetic diversity of RABV variants co-circulating in Cameroon. In particular, it remains unknown whether the apparently most prevalent Africa-1 lineage of the Cosmopolitan clade circulates in the northern regions of Cameroon (Adamoua, North and Far North regions) (Fig 2). However, our findings suggest that the Cosmopolitan subclade, Africa-1, circulate extensively in the Centre region of Cameroon, and in Yaounde in particular (Fig 2). Meanwhile the Africa-1a lineage was remarkably more frequent, Africa-1b was represented by only one isolate (Fig 3). Given the relatively high rate of Africa-1b RABV reported in the neighboring Central African Republic [44,45,58], it could be hypothesized that the unique Africa-1b identified in this study was introduced from Central African Republic. Accordingly, it has been recently suggested that RABV of the Africa-1 and Africa-2 variants circulate along the trunk roads between Cameroon and the city of Bangui. However, it is not possible to rule out the direction of RABV dissemination across the border between Cameroon and Central African Republic without extensive sampling in both sides.
In contract to the unique Africa-1b isolate, that was strikingly related to their counterparts from Central Africa, Africa-2 viruses were reliably separate from Africa-2 group C previously found along the trunk roads linking the Bangui city to Cameroon [44]. Africa-2 isolates from this study fell within Africa-2 group E and were only loosely related to their counterparts previously reported in the northern part of Cameroon as well as in neighboring countries (Figs 2 and 4). No Africa-2 group-C was found in this study despite the fact that this lineage has been reported in the northern part of Cameroon; that was not covered by this study.
Although it is tempting to explain away failure to efficiently amplify some RABV isolates by low RABV load in 17 studied rabies-positive specimens, the presence of potentially divergent Lyssaviruses among domestic dogs cannot be wiped out. Remarkably, none of the studied isolates displayed close phylogenetic relationships either with the divergent shrew-derived Mokola Lyssavirus strain 86100CAM (GenBank N° EU293117) from Cameroon, or with the Africa-4 and Africa-3 clades which are specific to northern [34] and southern [32,43,52] Africa, respectively. One explanation for the failure to efficiently amplify potential divergent variants of the RABV in this study may be that they were so divergent from the more prevalent variants that they could be refractory to amplification with generic primers used in this study. More powerful experimental approaches (including the use of divergent primers’ systems, virus propagation in cell cultures or high throughput sequencing) will be helpful for the complete assessment of the genetic landscape of Lyssaviruses in the studied specimens’ collection. Whole genome sequencing will also help to differentiate genetically-related isolates; thus providing more insights into the spatial dynamics of RABV epidemics.
This study is the first to tackle the molecular epidemiology of RABV isolates in Cameroon. We uncovered the presence of diverse lineages and variants of RABV co-circulating among dog populations in Cameroon. Striking phylogenetic evidence of outcasts to the apparent in-country geographic differentiation of RABV variants provided further support to the idea that the movements of rabid animals may be involved in the spread of dog rabies at least in urban areas. Molecular data reported here constitute potential baseline that would be interesting for the design, optimization and evaluation of rabies surveillance, prevention and control during the future stages of rabies elimination in Central Africa.
Animal specimens were collected with the approval of the Cameroon Ministry of Livestock, Fisheries and Animal Industries within the framework of routine rabies surveillance in Cameroon. All studied brain specimens originated from naturally infected rabid dogs enrolled by the Cameroon’s government veterinary services. No specimens were obtained from an experimental procedure nor animals used for experimental purposes.
The Republic of Cameroon is a Central African country (Fig 2); sharing borders towards the east with Nigeria, towards the west with Central African Republic, towards the north with Chad and Niger, and towards the south with Equatorial Guinea, Gabon and Congo. Cameroon is characterized by diverse ecosystems that are correlated, and thus attributed to, the patterns of rainfall and geological topology. The highlands of the West and North West regions define an area rich in volcanic lands having an average altitude of ≥ 1,100 meters. The southern rain forest of Cameroon is located in the maritime and equatorial zones (Center, East, Littoral, South and South-West regions) while its northern regions (Adamaoua, North and Far North) are progressively dominated by the savannah and steppe. There is no geographical features that may represent barriers to rabies spread in Cameroon. According to the 2010 estimates, the population of the Cameroon is at 19,406,100 people with 10,091,172 and 9,314,928 inhabitants in urban and rural settings, respectively [59]. In particular, the Center region has approximately 2,638,648 inhabitants in urban settings as compared to 887,016 people in rural settings.
This was a retrospective and transversal study based on the biological collection of brain specimens collected from rabid domestic dogs. Originally, heads of domestic dogs suspected of rabies were obtained from both private and public veterinary services. Domestic dogs were suspected of rabies if they displayed at least two of the following signs and symptoms: unprovoked aggression, foaming at the mouth, paralysis, incoordination, hoarse bark, hydrophobia, weakness, seizures, or loss of appetite. Brain specimens were collected during necropsy performed on dogs’ heads submitted for rabies diagnosis at the Centre Pasteur du Cameroun (CPC) located in the capital city, Yaounde. Laboratory confirmation of rabies at CPC was based on the detection of RABV nucleocapsid antigen in brain specimens by direct Fluorescent Antibody Test (dFAT) using rabbit IgG antibodies (Bio-Rad, Marnes-la-Coquette, France) [8]. Brain specimens negative for FAT were further confirmed by virus isolation on Murina neuroblastoma cell cultures as previously described [60]. After rabies diagnosis, remaining brain samples were kept frozen at -80°C. A total of 100 specimens derived from 100 rabid dogs were available and were thus considered in this study. All studied specimens originated from the 7 southern regions of Cameroon from 2010 to 2016 (Table 1 and Fig 2).
Each brain specimen was crushed in PBS (10% weight/volume) and 250 μL of supernatant resulting from clarified brain suspension was subjected to RNA extraction using TRIzol LS (Invitrogen, Paris, France), as recommended by the manufacturer's instructions. RNA samples were stored at -80°C prior to analysis.
Complementary DNA (cDNA) synthesis was performed in a final volume of 20 μL using pd(N)7 random primers and AMV reverse transcriptase (Promega). Briefly, 7 μL of purified RNA was incubated at 65°C for 10 minutes with 2 μL of RNase- and DNase-free water, 100 ng of random primers (1 μL) and 10 nmol. of each deoxynucleotide triphosphate (1 μL). Reaction tubes were then transferred on ice for at least 2 minutes and completed with 3 μL of RNase- and DNase-free water, 4 μL of AMV 5X Buffer, 40 U of RNAsin (1 μL) and 10 U of AMV reverse transcriptase (1 μL). Resulting 20 μL reaction mixtures were incubated 10 min at 25°C, 90 min at 42°C and 5 min at 95°C.
A 1485-base pairs DNA fragment encompassing the entire 1353 nucleotides (nt) of the N gene of RABV was amplified by reverse transcription-nested polymerase chain reaction (RT-nPCR) using previously described consensus oligonucleotide primers. First round PCR was performed with the primers pair RHN1 (5’-ACAGACAGCGTCAATTGCAAAGC-3’, nucleotides (nt) 28–52) and N8m (5’-CAGTCTCYTCNGCCATCTC-3’; nt 1584–1568) [21,26,61] in a final volume of 50 μL containing: 5 μL of cDNA, 34.5 μL of RNase- and DNase-free water, 5 μL of 10X PCR Buffer, 200 μM of each dNTP, 2 mM of MgCl2, 25 pmol of each primers and 2.5U of Taq DNA polymerase (Invitrogen, Cergy-Pontoise, France). The thermocycler profile was as follows: 5 min at 95°C followed by 35 cycles of 30 s at 95°C, 30 s at 56°C and 2 min at 72°C, and a final elongation at 72°C for 10 min. Second round PCR was carried out from 2 μL of the PCR product using the primers pair N127 (5’-ATGTAACACCTCTACAATGG-3’, nt 55–74) and 304 (5’-GAGTCACTCGAATATGTC-3’; nt 1539–1516) [21,62] under the same experimental conditions.
PCR products were analyzed by migration on Gelgreen-stained agarose gels and reveled on an ultraviolet transilluminator.
Amplicons were purified using the QIAquick PCR Purification kit (Qiagen, Courtaboeuf, France) following the manufacturer’s protocol. Purified amplicons were subjected to direct double strands sequencing using nested PCR primers, the BigDye terminator v3.1 kit (Applied Biosystems) and the ABI Prism 3140 automated sequencer (Applied Biosystems).
Consensus sequence editing, multiple sequences alignments and pairwise sequence comparisons were carried out with the CLC Main Workbench 5.7.2 software (CLC bio, Aarhus, Denmark).
GenBank accession numbers for the nucleocapsid gene sequences of RABV determined in this study have been assigned as MF537505 to MF537580.
To determine the phylogenetic relatedness of the newly sequenced RABV isolates, sequences were originally aligned with all relevant reference sequences available from online databases and originating from Cameroon and neighboring countries as well as representative sequences from other parts of Africa. Based on initial trees obtained, the alignment was downsized by removing duplicate sequences and by splitting the alignment into two datasets. We used Smart Model Selection [63] to determine the best-fit model of nucleotide substitution based on the Bayesian Information Criterion. This revealed that the General Time Reversible model with proportion of invariable sites plus gamma-distributed rate heterogeneity (GTR+I+Γ4) was optimal for the Africa 1 related dataset while the K80 model with proportion of invariable sites plus gamma-distributed rate heterogeneity (K80+I+Γ4) was the most suitable for the Africa 2 related dataset. Phylogenetic trees using individual datasets were then estimated by the maximum likelihood (ML) method available in PhyML 3.0 [55] using SPR branch-swapping. The reliability of individual nodes on the phylogenetic trees was estimated using 1,000 bootstrap pseudoreplicates.
|
10.1371/journal.pntd.0006533 | Current challenges and implications for dengue, chikungunya and Zika seroprevalence studies worldwide: A scoping review | Arboviral infections are a public health concern and an escalating problem worldwide. Estimating the burden of these diseases represents a major challenge that is complicated by the large number of unapparent infections, especially those of dengue fever. Serological surveys are thus required to identify the distribution of these diseases and measure their impact. Therefore, we undertook a scoping review of the literature to describe and summarize epidemiological practices, findings and insights related to seroprevalence studies of dengue, chikungunya and Zika virus, which have rapidly expanded across the globe in recent years.
Relevant studies were retrieved through a literature search of MEDLINE, WHOLIS, Lilacs, SciELO and Scopus (2000 to 2018). In total, 1389 publications were identified. Studies addressing the seroprevalence of dengue, chikungunya and/or Zika written in English or French and meeting the inclusion and exclusion criteria were included. In total, 147 studies were included, from which 185 data points were retrieved, as some studies used several different samples. Most of the studies were exclusively conducted on dengue (66.5%), but 16% were exclusively conducted on chikungunya, and 7 were exclusively conducted on Zika; the remainder were conducted on multiple arboviruses. A wide range of designs were applied, but most studies were conducted in the general population (39%) and in households (41%). Although several assays were used, enzyme-linked immunosorbent assays (ELISAs) were the predominant test used (77%). The temporal distribution of chikungunya studies followed the virus during its rapid expansion since 2004. The results revealed heterogeneity of arboviruses seroprevalence between continents and within a given country for dengue, chikungunya and Zika viruses, ranging from 0 to 100%, 76% and 73% respectively.
Serological surveys provide the most direct measurement for defining the immunity landscape for infectious diseases, but the methodology remains difficult to implement. Overall, dengue, chikungunya and Zika serosurveys followed the expansion of these arboviruses, but there remain gaps in their geographic distribution. This review addresses the challenges for researchers regarding study design biases. Moreover, the development of reliable, rapid and affordable diagnosis tools represents a significant issue concerning the ability of seroprevalence surveys to differentiate infections when multiple viruses co-circulate.
| Arthropod-borne viruses (arboviruses) are among the most important of the emerging infectious disease public health problems facing the world. The actual impact of arboviruses worldwide remains unknown, and estimating the true burden of these diseases represents a current challenge. Serological surveys are the most reliable tool for estimating the impact of arboviruses outbreaks in a given territory, and the results of such surveys have implications for potential mitigation measures such as vaccination. We undertook a thorough review of the literature produced from 2000 to March 15, 2018, addressing the seroprevalence of dengue, chikungunya and/or Zika to describe and summarize methodological approaches and map the geographical distribution of seroprevalence studies for these three viruses worldwide. A total of 185 studies addressing the seroprevalence of dengue, chikungunya and/or Zika were included in the review. Most of the studies were exclusively conducted on dengue (66.5%), but 16% were exclusively conducted on chikungunya, and 7 studies were exclusively conducted on Zika; the remainder were conducted on multiple arboviruses. Our study reveals that a wide range of methodological designs were applied regarding population, recruitment and/or laboratory testing. This study also highlights the high seroprevalence heterogeneity between continents and within a given country for dengue, chikungunya and Zika viruses. The results underscore existing gaps in seroprevalence studies distribution worldwide and the need to develop the most sensitive and specific diagnosis tool to provide recommendations for future serological studies.
| Arboviral infections have become a significant public health problem with the emergence and re-emergence of arboviral diseases worldwide in recent decades. Arboviruses are considered emerging or re-emerging pathogens based on their geographic spread and increasing impact on susceptible populations. For instance, dengue virus (DENV) infection, once rare, is now estimated to be the most common arboviral infection globally, with transmission occurring in at least 128 countries and with nearly 4 billion people at risk [1,2]. Over the period 2000–2010, an unprecedented increase in the number of cases was reported in the Americas, circulating all four serotypes (DENV1-DENV2-DENV3-DENV4) and reaching the highest record of cases ever reported over a decade [3]. DENV is now hyperendemic in many parts of the tropics and subtropics. The recent emergence of chikungunya virus (CHIKV) in the Caribbean in 2013 and its rapid spread to 45 countries and territories in North, Central, and South America highlight its high potential for epidemics [4]. In the aftermath of this emergence, Zika virus (ZIKV) aroused global attention due to its rapid spread since its first detection in May 2015 in Brazil to 22 other countries and other territories in the Americas [5].
Given the increasing number of cases; geographic spread; and health, social and economic impact of arboviral outbreaks, estimating their true burden represents a crucial issue but remains a difficult task. In their acute stages, arboviral infections cause a broad spectrum of disease, ranging from asymptomatic infection to severe disease, which can lead to misclassification in case reporting, especially when several arboviruses co-circulate [6]. Furthermore, surveillance systems, which generally rely on clinicians, hospitals and laboratory reports, are appropriate for helping detect outbreaks promptly but are not designed to estimate the real disease burden and tend to underestimate the total number of cases. In fact, because of the nature of arboviral infections with 75%, between 3 and 25% and 80% of asymptomatic cases for DENV, CHIKV and ZIKV respectively [1,7,8] and because healthcare seeking can vary greatly based on access to care, surveillance data alone can be unreliable [9].
Accordingly, some studies have estimated the burden of DENV outbreaks using a range of empirical or extrapolative methods and disease-modeling approaches [1,10,11]. However, the most reliable data for empirical assessments are drawn from seroprevalence studies, which are often lacking. In fact, these seroprevalence surveys are expensive and difficult to perform; such surveys require important logistical resources, including a large workforce (e.g., supervisors, technicians, physicians, nurses or phlebotomists, epidemiologists, statisticians, and field investigators) and biological support (e.g., sufficient freezer space for sample storage and reagents and kits for testing). Moreover, establishing good and reliable tests for arboviruses is an important task for public health institutions, especially when symptoms are difficult to distinguish from other common febrile illnesses and when cross-reactivity is observed [12]. The problem of cross-reactivity, as a result of the co-circulation of multiple arboviruses belonging to the same family in the same area, requires additional tests and thereby increases overall cost, time and labor [13].
However, data on arboviruses prevalence rates are essential for understanding their geographical distribution as well as their contribution to global morbidity and mortality. Such information is critical for determining the optimal allocation of the limited resources available for disease control and evaluating the impact of prevention policies and strategies such as vaccination. The rationale for conducting serological studies is straightforward; these studies provide surveillance that complements traditional symptom-based and laboratory-based surveillance. Serological studies provide an alternative approach for monitoring immunity levels in a population and do not require that people be tested during the short period when they are symptomatic [14]. In our research, seroprevalence can be defined as the frequency of individuals in a given population presenting evidence of a prior infection based on serological tests or a combination of serological and virological tests.
Seroprevalence studies can be conducted using multiple designs and among various populations involving a general population or specific or relevant population subgroups.
The general population concept is widely used in seroprevalence studies, but few studies provide a clear definition, and ambiguities related to the definition exist in the context of almost every country. Here, we present a definition that will be used throughout the review to classify serologic surveys according to the study population. A “general population” refers to the people (without any ethnic, socio-economic or health status restrictions) who inhabit a given area, usually in terms of political or geographical boundaries. The area may be quite small in size and population (e.g., a village of one hundred people) or quite large (e.g., a nation of one million people). A general population survey involves the collection of data to characterize all, or nearly all, people living in the area. Because of financial and logistical constraints, the data are typically collected from a representative sample of people residing in that area through a combination of personal interviews, administered on site using a standardized questionnaire, and blood samples drawn by skilled personnel (doctor, nurse or phlebotomist). Although surveys of the general population may gather data about inhabitants of all ages, lower and/or upper age limits are typically placed on eligible respondents, especially when blood samples are needed.
In contrast to general population surveys, specific population surveys focus on subgroups, (e.g., pregnant women, school children, blood donors, and patients). These subgroups are defined by membership in or contact with some social institution or by the presence of exposure. Furthermore, regardless of the type of population, because a census is resource-intensive, random sampling is highly recommended as a cost-effective method for obtaining seroprevalence estimates that are representative of the target population. Convenience sampling, such as selecting administrative units or schools that are easy to sample, is expected to result in bias. The reason is that administrative units selected because of convenience may not be generalizable to the larger population [9].
Seroprevalence studies can also use different designs, including cross-sectional, prospective, and retrospective designs, and can refer to cohort or case-control studies.
In the context of emerging and re-emerging arboviral diseases worldwide, we undertook a scoping review of the literature to describe and summarize the epidemiological practices, findings and insights related to seroprevalence studies reported worldwide over the recent period of 2000 to 2017, which was marked by an unprecedented increase in the number of arboviruses cases registered across the globe.
A literature review group (CFr, CFl) developed the protocol for conducting this literature review based on the Preferred Reporting Items of Systematic Reviews and Meta-Analyses (PRISMA) statement.
Screening was first conducted through an online MEDLINE (United States National Library of Medicine) search for English- or French-language literature published between January 2000 and March 2018. Between November 2016 and March 2018, we searched several electronic databases with reference to the expanded Medical Subject Headings (MeSH) thesaurus, using the following search terms: [“arbovirus” or “arbovirus infection” or “dengue” or “chikungunya” or “zika”] AND [“seroepidemiologic studies” or “seroprevalence” or “seroepidemiology” or “serosurvey”]. The databases included the following: MEDLINE, World Health Organization Library database (WHOLIS), Latin American and Caribbean Health Sciences Database (Lilacs), Scientific Electronic Library Online (SciELO) and Scopus. A free search was also conducted through the Google search engine. Additional studies were identified through manual searches of the reference lists of identified papers. No attempt was made to identify unpublished studies. After deleting duplicates, the literature review group systematically screened the title, abstract and full text of each study for the inclusion and exclusion criteria.
Articles were excluded if (i) the studies were published before January 1, 2000, or after March 15, 2018; (ii) the studies were published in languages other than English or French; (iii) the study sample included febrile patients, hospitalized patients, suspected or confirmed cases, or HIV or malaria patients because they are likely to provide biased estimates of seroprevalence, as well as if the study sample included immigrants, military personnel, travelers, or relief workers; and (iv) they were prospective/retrospective cohort studies that did not provide a baseline seroprevalence, because these study designs are likely to be associated with a specific first objective that only rarely focuses on determining seroprevalence rates.
We included cross-sectional and cohort studies analyzing samples from the general population, pregnant women, blood donors, age-specific subgroups, healthy volunteers and school children as possible sources of information about arboviruses seroprevalence.
Data from the selected sources were collated and summarized using a table consisting of a series of Excel spreadsheets. Eligible articles were abstracted for publication metadata, settings, design, population sampling approach, sample size, laboratory assays, age categories, seroprevalence rates, ethical approval and reported biases. When a study used several separate samples (e.g., from different countries or different study populations or age group), it was separated, and each sample was considered a unique data point. Duplicate citations were removed. When articles were not available or did not provide sufficient information, we contacted the authors for additional information.
We identified 265 unique studies reporting the seroprevalence of dengue, chikungunya or Zika that were eligible for full-text review (Fig 1). Among these studies, 18% (n = 48) were prospective or retrospective cohort or case-control studies, among which 16 studies provided a seroprevalence at baseline and enrolled participants according to our inclusion criteria. With respect to the study populations, 39.6% (n = 105) of these studies targeted febrile patients, hospitalized cases, suspected or confirmed cases, malaria or HIV patients, travelers, immigrants, relief workers or military personnel. Incomplete information was available for three studies. In total, 118 studies did not fulfill the inclusion criteria. Ultimately, the review was based on 185 data points from 147 unique studies (Fig 1). A description of the included studies is available in S1 Appendix.
The majority of the studies were exclusively conducted on dengue [15–112] (n = 123), with 16.% exclusively conducted on chikungunya (n = 29) [113–136] and 12% conducted on both dengue and chikungunya (n = 23) [137–154]; furthermore, seven studies were conducted on Zika [8,155–160], one study was conducted on both dengue and Zika [161] and two studies were conducted on both viruses [162].
Overall, as shown in the maps in Fig 3A, the studies were primarily conducted in inter-tropical areas, with some disparities within this region. We identified data from eight world regions, including eight studies from North America [15,22,55,60,60,71], three from Europe [65,126,134], 12 from Oceania [8,29,32,33,114,161,163,165], 38 from Central America and the Caribbean [20,24,24,36–38,48,48,53,58,63,91, 94,95,97,98,100,103,105,106,111,117,117,118,120,121,121,141,158], 21 from Latin America [17,19,19,19,23,25,28,54,56,72,73,78–80,101,102,108,128,160,162], 44 from Africa [18,21,26,26,31,31,31,47,52,57,68,74,76,81,86,88,107,116,120,120,124,127,130–132,132,133,137,140,143,144,144–149,149,149,149,151,153,156,158], and 59 from Asia [16,27,30,34,35,39,41–45,45,46,49–51,59,61,62,62,62,62,64,66,67,69,70,75,77,82–85,87,89,89,90,92,93,96,99,104,109,110,112,113,115,117,123,125,129,135,136,138,139,141,150,152,152].
Dengue studies were primarily conducted in the Americas (39%) and in Asia (33%) (Fig 3B). The countries most heavily involved in the implementation of the surveys over the past two decades were Brazil (12 studies) [17,19,19,19,23,28,78–80,101,108,128], Singapore (ten studies) [16,35,62,62,62,62,82,92,93,96], Thailand (eight studies) [40,46,66,67,87,90,139,150] and India (seven studies) [34,45,45,69,75,138,152].
Chikungunya studies were primarily conducted in Africa (46%) and Asia (24%). The most represented countries were Kenya, with six studies, and India [123,125,138,152], Madagascar [128,128,128,128] and French Polynesia [114,163], with four studies (Fig 3C).
Finally, Zika studies were conducted in Oceania, the Caribbean, Africa and Latin America, with three studies in French Polynesia [161,165,165], one in Micronesia [8], one in the French Indies (Martinique) [159], one in Zambia [156], one in Cameroon [158], one in French Guiana [160] and one in Bolivia [162] (Fig 3D).
The inclusion criteria restricted the analysis to studies published between January 2000 and March 2018; however, 14 studies were conducted before 2000 (Fig 4).
DENV seroprevalence studies were conducted between 1989 and 2017. Their distribution over the last decade indicated two peaks, one in 2004 and one in 2009–2010. The number of studies observed in 2004 might be enhanced by the re-emergence of CHIKV in Africa and the large DENV epidemic in Reunion Island. In 2010, the first phase III clinical trial for the now available tetravalent vaccine was initiated. This event may have encouraged seroprevalence studies to provide data for future vaccine programs. There were difficulties in interpreting the distribution of DENV studies with respect to the study year and location given the expansion of the virus in the Pacific, Southeast Asia, Africa, the Americas and the Middle East before the 1990s [166]. Moreover, at the time of this study, many countries were hyper-endemic with the co-circulation of four serotypes and with repeated epidemics every three to five years.
The first exclusive CHIKV seroprevalence study was conducted in 2004 in Kenya, with the re-emergence of the virus causing a large outbreak in 2004 [130]. This study was followed in 2005 by two studies, one in the Grande Comoros Island [131], where an outbreak occurred, and in Mayotte before the 2006 epidemic [132]. In 2006, four studies were conducted on Reunion Island and Mayotte during and after the 2006 epidemic [120,120,132,133] and in Benin, where no cases have been reported [116]. In 2007, three studies were conducted: one in Malaysia [129] (after the 2006 outbreak) when the virus subsequently spread to Asia, one in Gabon before the 2007 outbreak [153] and one in Italy [126] when CHIKV was imported to Europe, causing an outbreak. In 2008, two studies were conducted in India and Malaysia, where two outbreaks occurred [115,125]. In 2009, two studies were conducted in India and Kenya [123,124], and in 2010, a study was conducted in Congo after a 2010 CHIKV outbreak [127]. In 2013, CHIKV emerged in the Americas, and in 2014 and 2015, five studies were conducted in the Caribbean [118,118,119,121] (Saint-Martin, Guadeloupe, Martinique and Puerto Rico) and Central America [122] (Nicaragua) during an outbreak in Saint Martin and post-outbreak in the other locations. One study was conducted in Vietnam in 2015, where little was known about CHIKV transmission and where dengue is endemic [136]. The last study was conducted in 2016 in Brazil in a post-outbreak context [128].
There were seven ZIKV studies. The first study was conducted in Yap Island during the 2007 outbreak [8], and the second study was conducted in Zambia in 2013 [156], where no information on ZIKV was available. Two studies were conducted in French Polynesia in two distinct populations during and after the 2014 outbreak and one in 2015 [155]. Another study was conducted in Cameroon in 2015 [158], and one study was conducted in 2016 in Martinique (West Indies) [159]. Finally, the last study was conducted in French Guiana during the ZIKV outbreak in 2016 [160].
All seroprevalence data are presented in S1 Appendix.
Arboviral infections are common causes of disabling fever syndromes worldwide. In many countries, the concomitant co-circulation of dengue, chikungunya and Zika viruses represents a major recent public health and biomedical challenge. Prior to the introduction and subsequent spread of CHIKV and ZIKV in the Americas, dengue was the predominant arboviral infection worldwide. In this context of emerging and re-emerging arboviral diseases worldwide, estimating the burden of these diseases represents a major challenge to more efficient planning for disease control and reducing the risk of future re-emergence of arboviruses. Several affected countries face challenges in estimating the burden of arboviruses. Nonetheless, estimating the true burden of arboviral infections remains a difficult task given the large number of unapparent infections, especially those of dengue fever [1]. Serological surveys are thus required to identify the distribution of these diseases and measure their epidemic impact. A recent estimate indicated that the number of cases affected by any of these three arboviruses dramatically increased after 2013, reaching over 3.5 million by the end of 2015 in the Americas [169].
This review emphasizes several aspects of arboviruses epidemiology and describes current challenges and implications for dengue, chikungunya and Zika seroprevalence studies worldwide.
Overall, our results highlight the highly heterogeneous nature of study designs and serological tests used in arboviral seroprevalence studies. Seroprevalence surveys have the benefit of not being affected by surveillance system sensitivity or symptomatic case reporting rates but still have several limitations inherent to the adopted methodology. Selection biases, defined in our review as a distortion in the seroprevalence rate, may occur due to the use of a non-probabilistic sampling frame or poor field worker practices, such as replacing a selected household with one that is easier to reach [170]. Furthermore, the use of serum samples collected for various purposes frequently hinders the representativeness of the population sample and, consequently, that of the provided estimations. Even if the use of convenience samples is a good strategy for increasing the volume of serological data produced, the potential biases such sampling introduces must be considering during the analysis process to produce valid results. These limitations in the literature underscore the challenge of estimating global prevalence in the absence of nationally representative age-specific databases. Whenever surveys are conducted, all efforts to ensure high-quality collected data should be made. In particular, probabilistic samples should be used, and the sample size and number of clusters should be selected appropriately to rigorously estimate population seroprevalence rates.
The review also highlights the variety of serological tests used to measure antibodies activities. ELISA tests are the most common diagnostic method (used in more than half of the studies included in this review). Moreover, we noted that more than half of the studies that performed IgG ELISA tests used the indirect method, which is recommended, as it allows for the detecting of lower levels of antibodies than the direct method does and is thus more sensitive [171]. Most commercially available diagnostic IgG ELISAs that are adjusted to measure past arboviral exposure tend to have high sensitivity but suffer from low specificity due to high cross-reactivity with other arboviruses (flaviviruses or alphaviruses) circulating in a given geographical area or with Japanese encephalitis (JE) and YFV recommended immunization [172,173]. The resulting false positives could lead to information bias that can be overcome through control with neutralization tests. Both tests provide complementary results because one test is a biochemical assay (ELISA) measuring antibodies binding to the antigen and the other is a biological assay measuring antibodies’ capacity to neutralize an infecting virus. Only neutralization tests measure the biological parameters of in vitro virus neutralization and are the most virus-specific serological tests [174]. Indeed, for seroepidemiological studies, neutralization tests remain the “gold standard” for confirming and serotyping DENV infections in regions where two or more flaviviruses are co-circulating. These tests, however, are time-consuming, labor-intensive and expensive and are not as amenable to testing large numbers of sera as the ELISA is. When neutralization tests are not performed to complete results from IgG ELISA, country-specific contexts, including the presence of other circulating flaviviruses or alphaviruses and immunization programs for JE and YF, must be considered when interpreting seroprevalence results.
Although we stratified seroprevalence data by assay to allow for comparisons, more than half of the ELISA tests were performed using different commercial kits and in-house assays with variable sensitivities and specificities. Moreover, differences in assay formats, usage of antigen, and detection systems make it difficult to estimate the performances value of each individual assay by proper comparison [175]. In a multicenter evaluation using a commercial assay, it was shown that the sensitivities and the specificities varied between studies depending on the serum samples of the respective collaborating centers used for the performance evaluation [176]. These variations reported from several studies [177] indicated the need to develop the most sensitive and specific diagnosis tool to provide recommendations for future serological studies.
Although the diversity of study designs and serological tests used in the selected seroprevalence studies represents a major limitation for the comparison of seroprevalence rates by geographical region, our literature search highlights the highly heterogeneous seroprevalence of DENV and CHIKV worldwide as well as the significant variability among regions in the same country. The review also clearly shows that seroprevalence was the highest in island environments for both arboviruses. Some of these variations may stem from methodological differences, as well as the choice of study population, sample size and diagnostic test. This heterogeneity may also reflect differential exposure to mosquitoes. Indeed, disease transmission can substantially differ between regions characterized by different environmental and climatic determinants of vector density. For instance, in Kenya, alphavirus antibodies, specifically those against CHIKV, were detected only in children from the Kisumu District (lowlands) and not in children from the Nandi District (highlands) [144]. Geographic and climatic differences between these two regions could provide evidence for varying environmental factors related to arboviruses transmission risk. For instance, mosquito vectors are not as prolific in the colder climate of the highlands, whereas the lowlands offer warmer and wetter areas for mosquito development and could provide an appropriate environment for mosquito vectors and subsequent arboviral transmission. Moreover, the seroprevalence heterogeneity is related to the different transmission dynamics of these arboviruses, including force of infection, reproductive number and others factors such as strain [178].
The review highlights some factors associated with arboviral seroprevalence. We noted an increase in seroprevalence among older people in 44% of DENV studies and 18.5% of CHIKV studies mainly conducted in endemic areas, whereas only one ZIKV study obtained this result. The epidemiological context of affected countries appears to be associated with the relationship between age and seroprevalence. For instance, in an emergence context, few studies reported this association, as the target population is naïve. However, some CHIKV studies not conducted in an endemic area reported this association, which may suggest that age is associated with level of exposure. We noted that 14 studies found an association between gender and seroprevalence. These findings suggest that there may be gender-related differences; however, these discrepancies require further exploration, as the health gender gap may stem from various patterns that affect exposure to mosquitoes [179]. We also observed that, regarding DENV and CHIKV, this association varied between countries, and no study reported an association between gender and ZIKV seroprevalence. However, given the transmission issues associated with ZIKV, we can expect that in future studies, sexual transmission will correlate with higher seroprevalence among women. Although few studies revealed an association between ethnicity and seroprevalence, this finding can be related to background prevalence in the country of origin, in combination with increased early life exposure before migration or exposure during travel to their region of origin post migration. Moreover, these associations might also be partly explained by increased susceptibility related to the lower socioeconomic position of certain ethnic groups. The relationship between health and socio-economic status is well documented, and research has revealed a graded association in which people of lower socio-economic status have much worse health outcomes than those of higher socio-economic status [180,181]. Finally, 20 studies also reported an association between environmental factors and seroprevalence, as certain environmental conditions, such as house structure or objects collected in the yard, are more hospitable to A. aegypti [182].
Although the proportion of seropositivity depends on the diagnostic method used, it also relies on study planning; if a serosurvey is conducted long after the end of an outbreak, the signal for the antibodies may be lower than in a study conducted close to the end of an outbreak. Our results indicate that DENV seroprevalence in the Americas was higher than that in Asia, which is surprising because dengue has been endemic in Southeast Asia for decades [183], and the Asian burden, including the Western Pacific, accounted for nearly 75% of the global burden worldwide [184]. Analysis revealed that studies conducted in the Americas were performed significantly more frequently in an outbreak or post-outbreak context (p<0.01) via IgM ELISA, which could explain the discrepancy between America and Asia.
Our review included 185 studies worldwide according to well-defined inclusion criteria. The findings indicate that the distribution of our studies follows the same pattern observed for the expansion of their vectors [185]. Moreover, all of the studies reflect areas of arboviral circulation in an epidemic pattern. However, the maps clearly demonstrated where seroprevalence survey data are lacking and identified potential places for implementing future seroprevalence studies; the maps also highlighted places in tropical regions where no data are available, especially for CHIKV and ZIKV, which are considered emergent or re-emergent viruses. Some countries located in the tropics and subtropics were not represented, although they are considered at risk of transmission by the WHO; this is especially true for Africa, where few studies were conducted and where the epidemiology of these arboviruses is under-exploited. In addition, a recent review suggested that dengue transmission is endemic to 34 countries in all regions of Africa [186].
The temporal distribution of Chikungunya studies followed the timeline of CHIKV outbreaks during its rapid expansion since 2004, suggesting that ZIKV surveys, in the context of its recent emergence in the Americas, may be currently in process and, if not, that such surveys should be addressed rapidly.
Serological surveys provide the most direct measurement for defining the immunity landscape for infectious diseases, but they remain difficult to implement. Overall, dengue, chikungunya and Zika serosurveys have followed the expansion of these arboviruses, but there remain gaps in their distribution. Serological studies can address future challenges in identifying trends in arboviruses transmission over time, and age-specific antibody prevalence rates can be used to estimate when major changes in transmission occurred.
|
10.1371/journal.pgen.1002481 | A Pathogenic Mechanism in Huntington's Disease Involves Small CAG-Repeated RNAs with Neurotoxic Activity | Huntington's disease (HD) is an autosomal dominantly inherited disorder caused by the expansion of CAG repeats in the Huntingtin (HTT) gene. The abnormally extended polyglutamine in the HTT protein encoded by the CAG repeats has toxic effects. Here, we provide evidence to support that the mutant HTT CAG repeats interfere with cell viability at the RNA level. In human neuronal cells, expanded HTT exon-1 mRNA with CAG repeat lengths above the threshold for complete penetrance (40 or greater) induced cell death and increased levels of small CAG-repeated RNAs (sCAGs), of ≈21 nucleotides in a Dicer-dependent manner. The severity of the toxic effect of HTT mRNA and sCAG generation correlated with CAG expansion length. Small RNAs obtained from cells expressing mutant HTT and from HD human brains significantly decreased neuronal viability, in an Ago2-dependent mechanism. In both cases, the use of anti-miRs specific for sCAGs efficiently blocked the toxic effect, supporting a key role of sCAGs in HTT-mediated toxicity. Luciferase-reporter assays showed that expanded HTT silences the expression of CTG-containing genes that are down-regulated in HD. These results suggest a possible link between HD and sCAG expression with an aberrant activation of the siRNA/miRNA gene silencing machinery, which may trigger a detrimental response. The identification of the specific cellular processes affected by sCAGs may provide insights into the pathogenic mechanisms underlying HD, offering opportunities to develop new therapeutic approaches.
| Huntington's disease (HD) is a neurodegenerative disorder caused by an abnormal CAG expansion in the Huntingtin gene (HTT), resulting in an expanded polyglutamine track in the HTT protein. Longer CAG expansions correlate with an earlier more severe manifestation of the disease that produces choreic movement, behavioural and psychiatric disturbances, and dementia. Although the causative gene is widely expressed, neuropathology is characterized by striatal and cortical atrophy. HTT interacts with proteins involved in transcription, cell signaling, and transport. The pathogenic role of mutant HTT is not fully understood. This study shows that CAG expanded HTT RNA also contributes to neuronal toxicity. Mutant HTT RNA gives rise to small CAG-repeated RNAs (sCAGs) with neurotoxic activity. These short RNAs interfere with cell functions by silencing the expression of genes that are fully or partially complementary, through a mechanism similar to that of microRNAs. These findings suggest that a small RNA–dependent mechanism may contribute to HD neuronal cell loss. The exhaustive identification of the target genes modulated by sCAGs may lead to a better understanding of HD pathology, allowing the development of new therapeutic strategies.
| Huntington disease (HD), a dominantly inherited neurodegenerative disorder, is caused by an abnormal CAG expansion within the first exon of the Huntingtin gene (HTT), leading to an expanded polyglutamine (polyQ) track in the HTT protein. HTT is ubiquitously expressed in the cytoplasm of most cells in the body, with higher expression levels in brain and testis [1], [2]. However the disease shows a selective pattern of neurodegeneration, with clear effects in the cerebral cortex, and a more pronounced neuropathology in the striatum [3], [4].
The number of CAG repeats influences the severity and the age of onset of the disease. Longer expansions associate with a more severe form and an earlier manifestation of the disease [5].
It has been widely reported that the polyQ expansion in the HTT protein leads to protein aggregation and cell toxicity [6], a mechanism thought to be primarily involved in several neurological disorders caused by CAG repeats [7]–[10]. However, whether the mutant HTT aggregates are pathogenic, incidental or neuroprotective is still controversial. It has been shown that mutant HTT aggregates may function as sinks where essential proteins are sequestered [11], compromising cell survival [12]. Other studies show that increased levels of diffuse mutant HTT are responsible for neuronal cell death [13]. In agreement with the two possibilities, the activation of autophagy, reduce both soluble mutant protein and aggregate levels, and reduces toxicity [14], [15].
In addition to the widely described pathogenic role of expanded polyQ tracks, several studies have also shown that different neurodegenerative disorders caused by trinucleotide repeat expansions may involve RNA-mediated mechanisms [16], [17]. These include the sequestration of RNA-binding proteins by the expanded trinucleotide repeats, and activation of a variety of pathways such as RNA interference (RNAi) and protein misfolding pathways. The understanding of how expanded-repeat RNAs confer neurotoxicity is crucial to developing effective treatments.
A neurotoxic effect for CAG-expanded transcripts has been recently demonstrated in Drosophila models of Ataxin-3 [18] and Myotonic Dystrophy [19]. In the later, the authors propose a pathogenic role of siRNAs derived from complementary sense and anti-sense expanded (CUG/CAG) transcripts. In line with this, double-stranded CAG/CUG repeat RNA produced by bidirectional transcription induces neurodegeneration and movement disorder in Drosophila model [20]. This neurotoxic effect is largely dependent on Dicer activity and linked to the formation of (CAG)7mers. In addition, other studies describe that trinucleotide repeated transcripts form secondary structures [21] that can be cleaved by Dicer in vitro [22], [23] resulting in the generation of trinucleotide repeated short RNAs. Together, these data suggest that different mechanisms lead to the formation of aberrant small RNAs in trinucleotide expansion diseases.
Huntington's disease like 2 (HDL2) is caused by a CTG.CAG expansion in the JPH3 gene, and the neuropathologic outcome and clinical features largely resemble HD. CUG expansions in the JPH3 gene correlates both with the formation of RNA foci and cell toxicity, suggesting RNA mediated toxicity [24], [25]. RNA pathogenic mechanisms have been little explored in HD. Expanded HTT transcripts are retained in the nucleus of human HD fibroblasts and co-localize with the MBNL1 protein [26], a splicing factor involved in the pathogenesis of CTG/CAG expanded transcripts [27]. In addition, mutant HTT protein alters microRNA (miRNA) biogenesis [28], and a strong miRNA deregulation is observed in HD brains [29]–[32], which may contribute to the aberrant gene expression observed in HD.
Here we provide evidence for a pathogenic role of the mutant HTT RNA. CAG-expanded HTT RNA can be processed to generate CAG-repeated short RNAs with neurotoxic activity. We show that expanded HTT toxic effect is dependent on RNA-induced silencing complex (RISC) and further demonstrate that expanded HTT participates in posttranscriptional gene silencing of genes containing pure and interrupted CTG repeats. This, together with HTT polyglutamine toxicity, may contribute to the neurodegeneration pattern observed in HD.
To evaluate the contribution of CAG-expanded RNA in HD pathogenesis, we generated vectors expressing unexpanded and CAG-expanded forms of exon 1 of human HTT (HTT-e1). HTT-e1 constructs containing 23 CAG repeats (23*CAG) were used as wild-type (unexpanded) model. For the expanded HD mutation, we generated HTT-e1 constructs containing 80 CAG repeats (80*CAG). Each set of vectors was produced as a form that could be translated into protein, and as a variant lacking the translation initiation codon, that was only expressed as RNA. Due to the reduced size of HTT-e1, the different variants were cloned into a pIRES-GFP expression vector. This strategy allowed the monitoring of the transfected cells avoiding the generation of a GFP fusion protein that could lead to artefactual localizations (Figure 1A, 1B and Figure S1). A recent study reveals that RNA transcripts with expanded CAG repeats can be translated in the complete absence of a starting ATG [33]. Thus, we evaluated whether the constructs lacking translation initiation codon expressed polyglutamine, using the anti-glutamine monoclonal antibody 1C2 (Figure 1B). The different HTT-e1 constructs were efficiently expressed, as shown by PCR amplification of HTT-GFP (Figure 1B left panel). However, we only detected a polyglutamine track in the constructs containing the ATG starting codon, suggesting that repeat-associated non-ATG translation (RAN translation) is not compatible with the type of vector used to clone the different HTT-e1 forms, at least for polyglutamine production. Since RAN translation can occur in all frames [33], the possibility that CAG expansion produce homopolymeric polyalanine and polyserine proteins cannot be ruled out. It is worth mentioning that 1C2 antibody does not allow quantitative comparison of the levels of 23*CAG-Prot versus 80*CAG-Prot; thus, the differences in the intensity of the 1C2 detected bands is a consequence of the number of glutamines in each HTT-e1, expressed vector (Figure 1B right panel).
We transiently transfected these four different HTT-e1 expressing vectors in differentiated human neuroblastoma cells (SH-SY5Y) as a post-mitotic neuronal cell model. Transfection experiments revealed that CAG expansion in HTT mRNA was sufficient to induce a dramatic cytotoxic response in differentiated SH-SY5Y cells (Figure 1C). Cell toxicity assays demonstrated that both CAG-expanded constructs (translated and non-translated forms) drastically affected neuronal cell viability, only differing in the timing of the response, that was earlier for the 80*CAG-RNA construct. However, a expanded polyglutamine expressing vector using CAA instead of CAG repeats (80*CAA), induced a mild toxic effect at the latter time-point that did not reach statistical significance (Figure 1C). These results suggest that the toxic effect induced by the expanded polyglutamine tract is specific for expanded CAG. The HTT RNA toxicity was further confirmed with the analysis of early and late apoptotic markers. The results obtained revealed that the expression of CAG-expanded HTT-e1 RNA is sufficient to induce nuclear condensation (Figure 1D) and caspase 9 activation (Figure 1E), processes previously reported to occur in HD brain samples [34], [35]. On the contrary, 80*CAA expressing vector induced milder caspase 9 activation. These data point to a direct link between the toxic effect of expanded HTT RNA and an intrinsic apoptotic process.
Transcripts containing long hairpin structures composed of CNG repeats are Dicer targets [22], [23]. The resultant sRNA products may trigger aberrant gene silencing with putative downstream detrimental effects. To test whether mutant HTT-e1 toxicity was associated to sRNA related mechanisms, we isolated the sRNA fraction (<100 nt) from cells expressing the 23*CAG and 80*CAG forms of HTT-e1, and transfected these sRNA into differentiated human SH-SY5Y neuroblastoma cells. Cell viability assays demonstrated that the sRNA population obtained from cells expressing 80*CAG-PROT and 80*CAG-RNA constructs, induced a remarkable cell death response (Figure 2A), compared to the sRNA population originated from cells expressing the 23*CAG control constructs. These results indicate that the expression of expanded HTT-e1 RNA is sufficient to deregulate the sRNA profile, thereby impairing neuronal viability. Furthermore, the sRNA fraction of cells expressing 80*CAA-PROT failed to induce cell toxicity, suggesting that the sRNA detrimental effect is linked to expanded constructs containing CAG repeats.
In agreement with previous studies demonstrating the generation of trinucleotide-repeated sRNA from triplet-expanded transcripts [19], [23], the expression of CAG-expanded HTT RNA led to the generation of CAG-repeated sRNAs (sCAG), of around 21 nt long (Figure 2B). The identity of these products was further confirmed by direct sequencing of the PCR products (Figure S2) and northern blotting (Figure S3). However, cells expressing the CAA expanded construct failed to produce sCAG, suggesting that the production of these species is not an experimental epiphenomenon.
Variable penetrance for alleles carrying 36–39 CAG repeats has been noted, but the disease appears fully penetrant when the repeat numbers are above 40 [5]. To confirm the sRNA toxicity in HTT carrying a moderate number or repeats, we generated HTT constructs with 35, 40 and 48 CAG repeats (Figure S4). We performed transfection experiments using the sRNA fractions of cells expressing HTT vectors with 23*CAG (normal), 35*CAG (normal), 40*CAG (pathogenic) 48*CAG (pathogenic) and 80*CAG (model for juvenile HD) and subsequently determined cell viability. The sRNA fraction isolated from 40*CAG, 48*CAG and 80*CAG expressing cells induced a significant toxic effect (Figure 2C). Furthermore, the severity of the toxic effect in differentiated SH-SY5Y driven by the sRNA fractions was associated to the length of the CAG stretch, as previously described for the full protein [36] (Figure 2C). In addition, the pools of sRNAs isolated from 40*CAG, 48*CAG and 80*CAG expressing cells contained progressively increasing amounts of sCAGs when compared with that of the 23*CAG and 35*CAG expressing cells (Figure 2D). These results suggest that sRNAs derived from moderately expanded HTT are sufficient to induce a detrimental response and further indicate that expansions above 40*CAG repeats are enough to produce significantly increased amounts of sCAG and a parallel toxic effect.
To analyse the role of sCAG products in HTT sRNA toxicity, we then co-transfected sRNAs pools derived from cells expressing 23*CAG or 80*CAG vectors along with either antisense RNA oligonucleotides that specifically block the action of sCAG (anti-sCAG), or scrambled inhibitors as negative controls (Scrambled sRNA inhibitors). The toxic effect of 80*CAG- versus 23*CAG-derived small RNA was not affected in cells transfected with a scramble siRNA. However anti-sCAG significantly decreased the detrimental effect of the HTT-e1 expanded constructs (Figure 2E). We therefore propose that the generation of sCAG is a key element in the toxicity mediated by CAG-expanded HTT-e1.
We next examined whether sCAG were detected in different brain areas of the R6/2 HD mouse model, a transgenic line that over-expresses the exon 1 of human HTT with more than 100 repeats, and recapitulates many of the key features found in patients with HD [37]. R6/2 mice of 8 weeks of age exhibited deficits in coordination and activity, striatal atrophy, HTT-aggregate accumulation and down-regulation of striatal-neuron integrity markers [38]. RT-PCR analysis revealed increased sCAG levels in the cortex and striatum of R6/2 mice with respect to their wild-type littermates (Figure 3A), two brain areas preferentially affected in HD. However, no differences in the expression of sCAG species were detected in the cerebellum and hippocampus of R6/2 mice. These results suggest the existence of region specific mechanisms modulating sCAG biogenesis and/or stability in the R6/2 mouse model.
sCAG levels were subsequently analysed in post-mortem brain samples from HD patients and control subjects. RT-PCR analyses confirmed an increase of sCAG in the frontal cortex and caudate regions from HD samples (Figure 3B and Figure S5). The PCR products were sequenced, and sCAG species of 18 nt length were found in both control and HD brain samples. However, sCAG species of 21 nt-long were only detected in HD human brain samples (Figure S6).
To further validate the pathogenic role of sCAG in human brain samples, we isolated the sRNA fractions from control and HD frontal cortex and caudate, and transfected them into differentiated SH-SY5Y cells. HD-sRNAs reproduced the toxicity exerted by the expanded HTT-e1 sRNAs (Figure 3C). Furthermore, anti-sCAG dramatically diminished the toxic effect of HD-derived sRNA, supporting a pathogenic role of sCAG species produced in HD brains.
In CAG-repeat expansion diseases, Dicer-dependent mechanisms result in the formation of sCAG with putative functions in pathogenic gene silencing [19]. We therefore investigated whether RNAi machinery is involved in the generation and function of sCAG in HD.
To that end we performed Dicer knockdown experiments in differentiated SH-SY5Y cells that were subsequently transfected with HTTe1-expressing vectors. Dicer depletion prevented the generation of sCAG (Figure 4A and Figure S7) and efficiently mitigated 80*CAG RNA toxicity in SH-SY5Y cells, as indicated by the decrease in LDH release and the inhibition of caspase 9 cleavage (Figure 4B). This result suggest that the toxic effect of the 80*CAG-derived sRNA is caused by a major pathogenic pathway triggered by sCAG.
Since the generation of sCAG was largely dependent on Dicer, we determined Dicer levels in several brain areas of control and R6/2 mice (Figure S8). Dicer expression was significantly decreased in the hippocampus and cerebellum of R6/2 mice while no differences in sCAG levels were detected in these areas (Figure 3A), suggesting that this could be a factor modulating sCAG generation from mutant HTT.
To explore the potential mechanisms of HTT sRNA toxicity and sCAG deleterious effect, we next examined the relationship between 80*CAG toxicity and Ago2 activity, a key factor in miRNA/siRNA gene silencing [39], [40]. Cell viability assays revealed that the toxic effect of sRNA pools originated from cells expressing 80*CAG-PROT and 80*CAG-RNA was diminished in cells depleted of Ago2 (Figure 4C). This result indicates that Ago2 is an important player in the pathogenic effect of 80*CAG-derived sRNA species.
The initiation of a sCAG-mediated gene silencing process requires the incorporation of sCAG into RISC. To test whether sCAG could be loaded into the Ago2 silencing complex, we transfected the HTT-expressing constructs into SH-SY5Y cells stably expressing Flag-Ago2. We performed immunoprecipitation (IP) assays using anti-Flag antibodies for Ago2 IP or anti-V5 antibodies as negative control (Figure S9) and RNA bounded to immunoprecipitated Flag-Ago2 was isolated. The analysis of the Ago2-associated sRNA revealed that sCAG generated from mutant HTT RNA efficiently bound to the Ago2 complex (Figure 4D).
These results, along with the protective role of anti-sCAG, suggest that sCAG initiate a transcriptome-dependent detrimental response through Ago2-mediated gene silencing mechanisms. To evaluate the direct role of Ago2 in the toxic effect of expanded HTT-e1 we restored Ago2 levels in cells depleted of Ago2, and determined cell death (Figure 4E). Restoration of Ago2 levels by the co-transfection of a Flag-Ago2 expressing vector significantly re-established HTT toxicity (Figure 4E).
In humans, the Ago subfamily consists of Ago1, Ago2, Ago3 and Ago4 that guide both siRNAs and miRNAs to complementary sites on target RNAs to modulate their expression [41]. We therefore asked whether Ago2 was the critical mediator on HTT sRNA toxicity or other Ago proteins could be participating as well. Given that Ago3 and Ago4 are not significantly expressed in SH-SY5Y cells (data not shown), we analyzed Ago1 contribution in HTT toxicity. The toxic effect of expanded HTT-e1 was significantly decreased in cells with reduced levels of Ago1, suggesting that mutant HTT effect is also mediated by Ago1 (Figure S10). Since Ago2 is the only member of the Ago family with endonucleolytic activity [40], [42], the results linking both Ago2 and Ago1 with HTT toxicity suggest that sCAG may be modulating gene expression through target mRNA degradation and/or translational inhibition, as described for miRNAs [43].
To validate the possible detrimental effect of sCAG in human cells, a synthetic 21-nt long CAG-repeated siRNA, [(CAG)7 siRNA], was delivered to a panel of primary human cell lines including, breast (HMEC), bladder (UROTSA) and pancreatic cells (HPDE). Differentiated SH-SY5Y cells were used as a neuronal model. (CAG)7 siRNA impaired cell viability at variable levels in different cell types. Although these results indicate that (CAG)7 siRNA detrimental effect is not restricted to SH-SY5Y cells, this cell model displayed significant higher sensitivity to (CAG)7 (Figure 5A and Figure S11). We performed additional experiments in SH-SY5Y cells following several differentiation protocols that result in differential cell morphology (Figure 5B and Figure S12). These assays demonstrated a correlation between the type of differentiation of SH-SY5Y cells and the sensitivity to (CAG)7 which supports a transcriptome-dependent response in sCAG-mediated toxicity.
To validate the gene silencing activity of sCAG, and determine whether a full or partial complementary with the target genes was needed, we generated firefly luciferase-expressing vectors carrying a (CTG)14 stretch in the luciferase 3′UTR. We also developed constructs with the sequence (CAG)14, which offer an interrupted binding to sCAG. In an attempt to evaluate the consequences of an imperfect matching, we also cloned the sequences 5′-TCCGTGCTGAGCCTGCCTGTCGTCTGTG-3′ and 5′-TGCTAGTATCAGATCTGCTGTGGAATTG-3′, present in the genes ADORA2A and MEIS2 respectively. These two genes are downregulated in affected brain areas of HD patients and brains from the R6/2 mouse model [44]. Furthermore, in silico analysis of the sCAG and MEIS2 or ADORA2 duplex stability using RNA hybrid suggests that MEIS2 and ADORA2 could be putative targets of sCAG.
HeLa cells were co-transfected with the different combinations of HTT-expressing vectors and the luciferase vectors, and luminescence was measured 24 hours after transfection. Expanded HTT RNA was able to moderately silence luciferase expression in a construct containing a CTG14 sequence in the 3′UTR, compared to control luciferase vectors, and the non-expanded forms of HTT-e1 (Figure 6A). These experiments suggest that sCAGs derived from expanded HTT are involved in post-transcriptional silencing of genes containing CTG repeated tracks. In addition, we also detected a moderate reduction of luciferase activity in constructs harbouring the sequence CAG14, suggesting that, expanded-HTT-e1 targets genes with CAG repeats, although the mechanism related may differ from the canonical miRNA/siRNA silencing pathways.
Interestingly, HTT construct expressing 80*CAG moderately decreased the expression (10% of reduction) of the reporters containing ADORA2A and MEIS2 regions (Figure 6A). This result indicates that full complementary between sCAG and its target genes are not needed to induce gene silencing. Therefore, sCAG may behave as siRNA molecules, but also as miRNA-like species, and offer an additional explanation for the broad gene expression deregulation observed in HD [44]. This possibility was further confirmed by RT-PCR quantification of ADORA2 and MEIS2 expression in SH-SY5Y cells transfected with the HTT-e1 expressing vectors. (Figure 6B and 6C). The results obtained reproduced the decrease observed in the luciferase assays. Accordingly, the expression of CAA expanded constructs, which failed to generate sCAG, didn't affect ADORA2A or MEIS2 expression levels. We also evaluated if the 80*CAG construct silenced the expression of genes containing a CUG tract, including DMPK, ASTN2 and ZFR (Figure S13). The expanded HTT-e1 induced variable silencing of the different genes that did not correlate with the number of CUG repeats. DMPK is the transcript with higher number of CUG repeats in the 3′UTR with 13 CUG repeats; ASTN2 presents a moderate number of consecutive CUG repeats and ZFR transcripts contains a region harboring 4*CAG immediately followed by 5*CTG. The variability in the dowregulation response suggests that the number of CUG repeats it's not a key factor in mutant HTT-e1 silencing activity.
We evaluated a possible enrichment in CTG regions (of a minimal size of 7) either in the full transcript or in the 3′-UTR of HD downregulated genes. For this analysis we considered the downregulated genes (<−1,2 downregulation and p<0,05), upregulated genes (>1,2 upregulation and p<0,05) and the group of genes that did not show significant expression deregulation, provided in the study by Hodges et al [44]. No significant enrichment in genes containing CTG regions was detected in the downregulated, upregulated or non-regulated genes (X-square p>0,05), suggesting that the overall mRNA gene expression deregulation was dependent on several pathogenic factors besides sCAG-mediated gene silencing.
We next asked whether sCAG could be inducing gene silencing by target mRNA degradation or by translation inhibition. The levels of MEIS2 protein were analyzed in differentiated SH-SY5Y cells transfected with normal and expanded HTT-e1. Cells were lysated 24 hours after transfection, a time point in which CAG-expanded HTT RNA toxicity was validated. Given that neural cells are more sensitive to HTT-e1 expression and cell death can be detected 21 h after transfection, MEIS2 levels were normalized by Actin and also referenced to GFP expression, which indicates the percentage of transfected living cells at the time of the analysis. Figure 6D shows MEIS2 protein levels after performing this analysis, confirming that CAG-expanded HTT-e1 induce a reduction in MEIS2 levels by 10%, in agreement with the luciferase reporter assays and mRNA quantification. The decrease in MEIS2 protein levels is similar to the reduction in MEIS2 mRNA level, which may suggest that mRNA degradation is the main mechanism in the particular case of MEIS2 post-transcriptional gene silencing. However, an exhaustive study should be performed to fully identify sCAG targets and characterize the mechanisms of gene silencing in each particular case.
The latest evidences suggest that RNA detrimental effects contribute to neurodegeneration in a number of trinucleotide repeat expansion diseases. However, these processes have not been extensively addressed in HD, where pathogenesis has been traditionally thought to involve the mutant HTT protein. Our results suggest an RNA pathogenic mechanism in HD that involves the aberrant generation of sCAG RNA species with an inherent toxic effect in a neuronal cell model. We have shown that the generation of sCAG species from expanded HTT exon 1 is largely dependent on Dicer, in agreement with previous studies showing that triplet repeats formed by CNG units adopt hairpin structures that become sliced to sCNG by dicer [22], [23]. In addition, it has become apparent that most of the expanded repeat disease loci have transcription occurring from both strands, raising the possibility that the complementary repeat RNAs form double-stranded structures susceptible to be processed by Dicer. Recently, a natural antisense transcript for HTT (HTTAS) has been described, covering the exon-1 CAG repeat [45]. Although HTTAS is under the control of a weak promoter, it is expressed throughout the brain and other tissues. Therefore, the production of sCAG in HD brains shown in the present study and in fibroblasts of HD patients [22] may originate both from HTT expanded hairpin structures and HTT/HTTAS double stranded RNAs.
Importantly, CAG repeat lengths above the threshold for complete penetrance (40 or greater) generated increased amounts of sCAG compared with non-pathogenic repeat lengths. Furthermore, our data suggest that the generation of sCAG correlated with the length of the repeat, being sCAG levels progressively higher in cells transfected with HTT-exon-1 constructs harboring 40, 48 and 80 CAG repeats, respectively. This correlated with a gradually increasing detrimental effect driven by the small RNAs fraction of cells expressing HTT-e1 with 40, 48 or 80 CAG repeats, respectively. These results agree with the increased severity of the disease in HD cases presenting extremely long CAG expansions in the HTT gene [46].
The amount of sCAG products was not equivalent in different brain areas in a HD mouse model, where increased sCAG levels were detected in the more affected areas. Our data suggest that decreased levels of Dicer could contribute to explain the lack of sCAG increase in the hippocampus and cerebellum of R6/2 mice. However, It is worth mentioning that Dicer activity is subject to regulation that affects the accumulation of miRNAs and probably sCAG. Recent work has identified a battery of proteins that regulate processing either interacting with Dicer or with miRNA precursors, being the activity of some regulators restricted to specific miRNA families [47]. In this context, whether Dicer is particularly active in the cortex and the striatum under basal conditions and/or in HD, the possible mechanisms modulating Dicer activity in specific areas and/or diseased state and its relevance to human disease are open questions that deserve specific research.
Our data indicate that the toxic effect of the sRNA fraction generated by expanded HTT is dependent on Ago proteins and is abolished by anti-sCAG. Furthermore, increased levels of sCAGs were found in Ago2 immunoprecipitates of cells expressing expanded HTT-e1, suggesting that sCAG-driven gene silencing may underlie HTT-RNA toxicity. In agreement with RISC-dependent mechanisms, expanded HTT-e1 constructs moderately silenced genes showing pure and interrupted CUG tracks, complementary to sCAG. However, we did not detect a significant enrichment in mRNAs harboring CUG-tracks among those found to be downregulated in human HD brain samples [44]. This suggests that gene expression perturbation in HD brains may reflect primary and secondary pathogenic triggers. In addition, the possibility that sCAG may act through translational repression, a gene silencing mechanism also described for miRNAs [43], cannot be ruled out.
Interestingly, expanded HTT induced similar silencing when using CAG repeats as the target sequence in the luciferase assay. The main structural requirements for gene targeting in miRNA-RISC mediated gene expression regulation are well defined for the most expressed miRNAs, including seed region perfect pairing in the 3′-UTR of the target genes [48]. However, knowledge about the determinants governing gene targeting is far from complete. In fact, targeting can occur through sites other than the 3′-UTR and seed region base pairing is not always required [49]. Whether imperfect base pairing between the CAG tracks in the small RNA and the target genes is compatible with the location and configuration of the sCAG-RISC complex, is an interesting question that should be specifically addressed.
In addition, since trinucleotide repeats have been shown to bind proteins, additional functions for sCAGs should be considered. Gene expression modulation by miRNAs recently included a decoy function, where miRNAs bind to proteins that regulate gene expression, thus modulating their activity [50]. The characterization of the sCAG binding proteins that could have consequences in gene expression regulation may shed light to possible additional RNA related pathogenic mechanisms.
In summary, we propose a pathogenic RNA dependent mechanism in HD by which sCAG produced over a threshold are neurotoxic. In HD, this mechanism may complement other RNA dependent processes including miRNA deregulation [28]–[32] and possible alterations in alternative splicing driven by MBNL1 sequestration [26], [51] (Figure 7). The detrimental effect may depend not only on the amounts of sCAG generated, but also on the target transcriptome and factors modulating RISC function. These aspects may contribute to sCAG variable vulnerability in different human cells observed in the present study. sCAG induced pathogenesis may underlie common phenotypes in triplet repeat diseases showing CAG expansions in different coding RNAs (leading to polyglutamine expansions in several proteins) and untranslated RNAs [19]. The identification of the specific sCAG-targeted genes and the cellular processes affected by sCAG should pave the way for the development of new therapeutic approaches for HD and other CAG-repeat expansion diseases.
Human Mammary Epithelial Cells (HMEC) were maintained in MEBM medium supplemented with Bullet-kit (Lonza), Human Pancreatic Duct Epithelial Cells (HPDE) were cultured in KSFM medium (Invitrogen) supplemented with epithelial growth factor (0.1–0.2 ng/mL) and bovine pituitary extract (25 µg/mL). UROTSA cells were maintained in RPMI medium (Invitrogen) supplemented with 10% FBS (Fetal Bovine Serum, Invitrogen). HeLa cells and SH-SY5Y neuroblastoma cells were maintained in Dulbecco's Modified Eagle's Medium (DMEM, Invitrogen) supplemented with 10% FBS, 2 mM L-glutamine, 100 units/ml penicillin and 100 µg/ml Streptomycin (GIBCO, Invitrogen). In the case of SH-SY5Y cells, FBS was heat inactivated for 45 min at 56°C prior to use. Unless otherwise indicated, SH-SY5Y cells differentiation was performed culturing the cells in the standard growing medium containing 10 µM retinoic acid (RA) during four days. The media was then replaced by fresh medium containing 80 nM of 12-O-tetradecanoylphorbol-13-acetate (TPA) during five additional days [52] Different neuronal differentiation protocols are provided in Figure S8).
Different forms of the exon 1 of the HTT gene (HTT-e1) differing in the CAG repeat length (23*CAG-, 35*CAG-, 40*CAG-, 80*CAG- or 80*CAA-PROT; and 23*CAG-, 80*CAG-RNA) were synthesized by Geneart. Flanking EcoRI restriction sites were added during the synthesis that were used to sub-clone the HTT-e1 variants into the pIRES2-EGFP vector (BD Biosciences, Clontech). Not-translatable constructs lack the translation initiation codon (AUG) and the second methionine (AUG) found in HTT exon1 (Figure 1A).
All the transfection experiments were performed using Lipofectamine 2000 (Invitrogen), according to the manufacturer's instruction and at a 60% cell confluence. (CAG)7 (5′CAGCAGCAGCAGCAGCAGCAG-3′) and control, scrambled siRNA (5′-GCGACGUUCCUGAAACCAC-3′) were purchased from Dharmacon and were administered at a final concentration of 50 nM, unless otherwise indicated. The anti-sCAG small RNA (LNA modified 5′-(CTG)7), and scrambled sequences (LNA modified 5′-GTGTAACACGTCTATACGCCCA-3′) were ordered from Exiqon. Both anti-sCAG and the corresponding scrambled inhibitor were transfected at a final concentration of 60 nM. Transfections with sRNA pools were performed using 35 ng of each sRNA pool per well (quintuplicates, 96wells multiwell). Dicer, Ago1 and Ago2 knockdown experiments were performed by a double transfection procedure; consisting in the transfection of the Scrambled, Ago2 or Dicer siRNA in the first assay (50 nM), and the co-transfection of the siRNA and HTT construct 48 hours later at 75 nM and 400 ng, respectively, in MW6 plates. Dicer siRNA (5′- GCUCGAAAUCUUACGCAAAUA-3′), Ago1 siRNA (5′- CAUCAGGACUGUUGAGUAA -3′) and Ago2 siRNA (5′-GCACGGAAGUCCAUCUGAA-3′) were purchased from Dharmacon. A siRNA against the 3′UTR of Ago2 (siAgo2-3′UTR: 5′-GGAAATATGGTTTGCTAAA-3′) was used in the HTT toxicity rescue experiments (Figure 4E). Transfection efficiency in experiments using siRNA or sRNA pools was determined at each experimental condition using siGLO transfection indicator (Dharmacon). Transfection conditions were optimized in order to obtain similar transfection efficiencies (∼90%) in all the cell lines analyzed.
SH-SY5Y cells were transfected with a Flag/HA-AGO2 expressing vector (Flag-tagged Ago2 expression vector was kindly provided by Prof. R. Shiekhattar). The plasmid encodes for a neo-resistance marker and transfected cells were grown in the presence of 800 µg/ml of Geneticin (G418, Gibco Laboratories) for 10–14 days. Single clones were selected to generate monoclonal cell lines. Expression of Flag/HA-AGO2 protein was checked by western blot and immunofluorescence in several cell clones.
For protein extraction, cells were rapidly rinsed with ice-cold PBS and solubilized with a lysis buffer described in [53]. Cells were then scraped off, incubated on ice for 15 min and centrifuged at 14000 rpm for 15 min. Samples were resolved in 10% SDS-PAGE gels and transferred to nitrocellulose membranes using the iBlot Dry Blotting System (Invitrogen). Membranes were blocked for 1 h with 10% skimmed milk in TBS (Tris-HCl, pH 7.5, 10 mm; NaCl, 100 mm) containing 0.1% Tween-20 (TBS-T). Membranes were incubated at 4°C and overnight with primary antibodies (diluted in TBS-T). After washing with TBS-T, membranes were incubated for 45 min at room temperature with the appropriate secondary antibodies (diluted in TBS-T), and then washed again with TBS-T. Detection was performed by ECL Western blotting detection reagent (Amersham Bioscience). Chemiluminescence was determined with a LAS-3000 image analyzer (Fuji PhotoFilm Co., Carrollton, TX, USA). Primary antibodies were anti-polyQ (MAB1574, 1∶1000, Millipore), anti-GFP (1∶2000, Molecular Probes, rabbit), anti-PARP (1∶5000, BD Pharmigen, mouse), anti-cleaved caspase 9 (1∶1000, Cell Signaling, rabbit), anti-Dicer (1∶500, Abcam, mouse), anti-Ago2, (1∶500, Abnova, clone 2E12-1C9). Anti-Ago1 antibody (1∶1000, rat) was kindly provided by Dr. G. Meister . Anti-GAPDH (1∶4000, Chemicon, mouse), anti-α-Actin (1∶5000, Chemicon, mouse) and anti-α-Tubulin (1∶50000, Sigma, mouse) were used as loading controls. Secondary antibodies were HRP-conjugated anti-mouse, anti-rat and anti-rabbit (1∶2000, DAKO)
SH-SY5Y cells grown on coverslides were rinsed several times with PBS and fixed for 20 min at room temperature with 4% paraformaldehyde in PBS. After rinsing, cells were permeabilized for 20 min in 0.5% Triton-X-100 in PBS. Non-specific binding sites were then blocked by incubating for 1 h in PBS containing 0.2% Triton X-100 and 10% FBS. Incubation with primary mouse anti-PolyQ antibody (1∶2000, Millipore, clone 5TF1-1C21) was carried out overnight at 4°C in PBS containing 0.2% Triton-X-100 and 1% FBS. After washing, coverslides were incubated with secondary anti-mouse IgG Alexa 555 IgG (Molecular Probes) at a dilution of 1∶1000 for 1 h at room temperature. After washing, coverslides were mounted in Vectashield-DAPI solution, and cells visualized under a Leica microscope (DMR). Images were captured using a digital camera (Leica DC500).
Differentiated SH-SY5Y cells were transfected in 96 well plates and cell viability was determined 24 hours post-transfection with the 3-(4,5- dimethythiazol-. 2-yl)-2,5-diphenyl tetrazolium bromide (MTT) assay. MTT was added to cell culture media at 0.5 mg/mL final concentration and incubated for 40 minutes at 37°C. Cells were then lysed with 100 µL of DMSO upon medium removal and absorbance was measured at 550 nm. In each experiment, determinations were performed in tetraplicates.
Lactose dehydrogenase (LDH) released from dying cells was determined using the LDH assay (Cytotox 96, Promega) according to the manufacturer's protocol, at different time-points following transfection (see figure legends). Absorbance was recorded at 490 nm. LDH determinations were performed in quintuplicate.
Cell death was also determined with the simultaneous staining of alive and dead cells using fluorescein diacetate (FDA) and propidium iodide (PI), respectively in a double staining procedure. Cells were rinsed with PBS 1× and then incubated for 45 s at 22–25°C with 15 mg/ml FDA (Sigma) and 4.6 mg/ml PI (Molecular Probes, Inc., Eugene, OR, USA) in PBS. The staining solution was replaced by PBS 1× and the stained cells were immediately examined under a Leica microscope.
Ago2 Immunoprecipitations assays and the extractions of the Ago2-bounded RNA were carried out as described previously [54]. Flag-Ago2 immunoprecipitation was performed using ANTI-FLAG M2 affinity gel (Sigma). ANTI-V5 affinity gel (Sigma) was used as a negative control for Ago2 IP. sCAG levels were determined by polyadenylation and RT-PCR. SH-SY5Y cells transfected with 100 nM (CAG)7 were used as a positive control.
Total RNA from cells or brain tissues was extracted using miRNeasy Mini kit (Qiagen). Small RNA species (<100 nt), were fractionated by size-exclusion column chromatography using Microcon Y-10 (Millipore) according to the manufacturer's instructions.
Total RNA was treated with TURBO DNA-free kit (Ambion). In vitro polyadenylation reactions were carried out using 1 µg of total RNA or 100 ng of sRNA enriched fraction and poly(A) polymerase (Ambion) for 1 h at 37°C in the presence of ATP (1 mM). Samples were then annealed with a polyT-adapter primer (5′-CGAATTCTAGAGCTCGAGGCAGGCGACATGGCTfGGCTAGTTAAGCTTGGTACCGAGCTCGGATCCACTAGTCCTTTTTTTTTTTTTTTTTTTTTTTTTAC-3′) prior to RT reaction. Specific primers recognizing the adapter and sCAG allowed the amplification of specific products by RT-PCR.
sCAG expression levels in cells transfected with the non-expanded or expanded HTT-e1 were analyzed by RT-PCR or densitometry of the PCR amplified products. Total RNA, polyadenylated total RNA or sRNA was retrotranscribed using the Superscript III RT kit (Invitrogen).
Equal amounts of cDNA were mixed with SYBR Green PCR mix (Roche). Five pmol of the forward primer (designed on the CAG repeat sequence) and reverse primer (based on the adaptor sequence) were used in each reaction. Amplification was done under the conditions of 15 sec at 95°C and followed by 55 cycles consisting in 1 min at 60°C and 2 min at 72°C in a LightCycler 480 Real-Time PCR System (Roche). The entire experiments were repeated three times on independent RNA preparations. RNU66 expression was used as a reference small RNA. Values were also referenced to the GFP levels, which refers to the number of transfected living cells at a particular time. β-Actin was the endogenous reference gene for GFP normalization. Data are presented as the ratio between the normalized expression of sCAG (sCAG/RNU66) or a particular gene (gene/β-Actin) and the normalized expression of GFP (GFP/β-Actin).
sRNA qRT-PCRs were performed with the following set of primers: sCAG Forward: 5′-CAGCAGCAGCAGCAGCAG-3′, sCAG Reverse: complementary to the polyT adapter after polyadenylation (5′-CGAATTCTAGAGCTCGAGGCAGG-3′); RNU66 forward: 5′-GTAACTGTGGTGATGGAAATGTG-3′; RNU66 reverse: 5′- GACTGTACTAGGATAGAAAGAACC-3′; RNU6B forward: 5′-CGCTTCGGCAGCACATATAC-3′; RNU6B reverse: 5′-TTCACGAATTTGCGTGTCAT-3′.
mRNA qRTPCR were performed using the following primer sets: GFP forward: 5′-TGCAGTGCTTCAGCCGCTAC-3′; GFP reverse: 5-TCGCCCTCGAACTTCACCTC-3′; DMPK forward: 5′-TGGGCTACTCCTACTCCTG-3′; DMPK reverse: 5′- AGCTGTTTCATCCTGTGGG-3′; ASTN2 forward: 5′- GACATTCTACACGGAGCAGTAC-3′; ASTN2 reverse: 5′- GTGAGTGGACAAGACATCTGG-3′; ZFR forward: 5′- TGGGACTCAAAATCAGCTACG-3′; ZFR reverse: 5′- TGGTTCTGTTGATGGAATGGG-3′; β-Actin Forward: 5′-CTGGAACGGTGAAGGTGACA-3′; β-Actin Reverse: 5′-GGGAGAGGACTGGGCCATT-3′.
Regular detection of GFP and HTT-e1 expression was performed using the following set of primers GFP Forward/GFP Reverse, and HTT forward//pIRES-GFP reverse (5′-GTCCCTCAAGTCCTTCCAGC-3′/5′-GAACTTCAGGGTCAGCTTCG-3′).
Gene expression analysis of ADORA2A and MEIS2 genes were carried out using Taqman assays (assay ID: Hs00169123_m1* for ADORA2A and assay No: Hs00542638_m1* for MEIS2). Data were normalized using MRIP (assay ID: Hs00819388_m1) as an endogenous reference gene. Amplification was done under the conditions: 15 sec at 95°C and followed by 55 cycles consisting in 1 min at 60°C and 2 min at 72°C on the ABI PRISM 7000 Detection system (Applied Biosystems). The entire experiments were repeated four times on independent RNA preparations.
qPCR results were analyzed using the 2−delta delta Ct method.
The levels of the precursors and mature forms of miR-16 and miR-29 in normal cells and cells with decreased levels of Dicer were determined by polyadenylation and RT-PCR in total RNA, as previously described. The following oligonucleotides were used: for miR-16-1: 5′-TAGCAGCACGTAAATATTGGCG-3′; for miR-29a: 5′- TAGCACCATCTGAAATCGGTT-3′.
CAG PCR products were run on a 15% polyacrylamide gel and visualized by SybrSafe staining (Invitrogene). PCR products were purified and ligated into pGEMT-easy vector. The sequencing reactions of the vectors were carried out using the Big Dye 3.1 Termination Cycle Sequencing Kit and DNA Sequencer (ABI3100) from Applied Biosystems.
Total RNA (30 µg) or small RNA (<100 nt long, 4 µg) were resolved in a 15% acrylamide-7.5 M urea gel and transferred to Hybond-N+ membranes (Amersham Bioscience) in 0.5× Tris-buffered EDTA at 200 mA overnight at 4°C. The membranes were UV cross-linked and heated at 80°C for 1 h. LNA probes (Exiqon) and oligoribonucleotide probes against (CAG)7 repeats (5′-CTGCTGCTGCTGCTGCTGCTG-3′) were labelled with γ-32P-dATP using Optikinase (USB Corp.). LNA probe complementary to RNU6B was used as loading control (5′-CACGAATTTGCGTGTCATCCTT-3′, Exiqon) and an oligonucleotide probe complementary to GFP (5′- GAACTTCAGGGTCAGCTTGC) was used to detect the expression of the different pIRES-HTT-e1-GFP vectors (HTT-e1-IRES-GFP transcripts with a length of around 1.5 Kb). Oligonucleotide probes hybridisation and washings were performed at 50°C using PerfectHyb Plus buffer (Sigma). The membranes were exposed to Fuji Imaging plates, scanned with a FLA-5000 PhosphorImager (Fuji PhotoFilm Co.) and quantified with ImageJ software.
A series of firefly luciferase-based reporter constructs were used for quantitative measurement of sCAG-mediated post-transcriptional gene silencing in genes containing (CUG)7/(CAG)7 stretches.
The putative target sequences were obtained by the annealing of oligonucleotides with the desired sequence, containing an XbaI restriction site at their 5′ end. The resulting double stranded DNA fragments were cloned downstream of the firefly luciferase reporter gene in the pGL4.13 vector (Promega) using XbaI restriction site.
The oligonucleotides used were: 5′-CTAG(CTG)14-3′ and reverse 5′-CTAG(CAG)14 for genes containing (CUG)n repeats; forward 5′-CTAG(CAG)14-3′ and reverse 5′-CTAG(CTG)14-3′ for genes containing (CAG)n repeats; forward 5′-CTAGTCCGTGCTGAGCCTGCCTGTCGTCTGTG-3′ and reverse 5′- CTAGCACAGACGACAGGCAGGCTCAGCACGGA-3′ mimicking a CUG rich region located in ADORA2A gene; forward 5′-CTAGTGCTAGTATCAGATCTGCTGTGGAATTG-3′ and reverse 5′-CTAGCAATTCCACAGCAGATCTGATACTAGCA-3′for a CTG containing region of MEIS2 gene .
HeLa cells were seeded at 1.3×104 cells/well in 96-well plates and 24 h later they were co-transfected with the following set of vectors: HTT-e1 constructs (40 ng), Firefly reporter constructs (24 ng) and Renilla reporter plasmid pGL4.75 (3 ng). The pGL4.13 vector without 3′UTR insertion was used as negative control for gene silencing. The (CAG)7 mimic was used as a positive control for silencing effect of CUG enriched stretches.
The activity of Firefly and Renilla luciferases was determined 24 h after transfection using the Dual-Glo™ Luciferase Assay System (Promega). Relative reporter activity was obtained by normalization to the Renilla luciferase activity. Each experiment was done in triplicate, and at least three independent experiments were performed for each construct.
Hemizygous male mice transgenic for exon 1 of the human Huntingtin gene with a greatly expanded CAG repeat (R6/2 mice) [37] were purchased from The Jackson Laboratory (Bar Harbor, code B6CBA-Tg(HDexon1)62Gpb/1J; 155–175 CAG repeats). The colony was maintained by back-crossing R6/2 males with (CBA×C57BL/6J) F1 females. Mice were sacrificed at 8 weeks of age, and brain samples were snap-frozen and subsequently stored at −80°C until use. Those 8 week-old R6/2 mice exhibited various hallmarks of HD-like disease, such as motor symptoms (deficits in coordination and activity), neuropathological deficits (striatal atrophy and huntingtin-aggregate accumulation) and molecular-pathology alterations (down-regulation of striatal-neuron integrity markers) [38].
Brain samples corresponding to the frontal cortex (FC) and the striatum (dorsal caudate, CA) of HD patients and controls were obtained from the Institute of Neuropathology and the University of Barcelona Brain Bank. CAG expansions ranged from 41 CAG repeats to 62 CAG repeats in the HD samples (control samples harbored less than 23 CAG repeats). The neuropathological examination in HD cases revealed severe atrophy of the caudate and putamen, cerebral cortical atrophy. This was accompanied by marked neuronal loss and astrocytic gliosis. Individual neurons in the cerebral cortex and striatum exhibited ubiquitin-positive intranuclear inclusions typical of diseases with CAG triplet expansions. HD cases were categorized as stage 4 following Vonsattel classification.
Animal handling procedures was conducted in accordance with Directive 86/609/EU of the European Commission.
Brain samples of HD patients and controls were obtained from the Institute of Neuropathology and the University of Barcelona Brain Bank, after the informed consent of the patients or their relatives and the approval of the local ethics committee. Ethical issues and legislation as defined by the European Union and national law. All activities were conducted with the approval of responsible ethical committees. The following general guidelines apply:- The Charter of Fundamental Rights of the EU; - Directive 2004/23/EC of the European Parliament and of the Council of 31 March 2004 on setting standards of quality and safety for the donation, procurement, testing, processing, preservation, storage and distribution of human tissues and cells; - Directive 95/46/EC of the European Parliament and of the Council of 24 October 1995 on the protection of individuals with regard to the processing of personal data and on the free movement of such data.
In each experiment “n” refers to completely independent experiments. Statistical analyses were performed using the two-tailed unpaired t-student's test for single comparisons (p<0,05) and the analysis of variance (ANOVA) when multiple pair-wise conditions were compared, where ad-hoc tests were addressed with the Bonferroni correction. The ANOVA test included an interaction term in the cases were the aim was to evaluate whether specific proteins modulate HTT-e1 response. Unless specifically indicated, p-values withstand Bonferroni correction.
|
10.1371/journal.pbio.1001670 | NECAP 1 Regulates AP-2 Interactions to Control Vesicle Size, Number, and Cargo During Clathrin-Mediated Endocytosis | AP-2 is the core-organizing element in clathrin-mediated endocytosis. During the formation of clathrin-coated vesicles, clathrin and endocytic accessory proteins interact with AP-2 in a temporally and spatially controlled manner, yet it remains elusive as to how these interactions are regulated. Here, we demonstrate that the endocytic protein NECAP 1, which binds to the α-ear of AP-2 through a C-terminal WxxF motif, uses an N-terminal PH-like domain to compete with clathrin for access to the AP-2 β2-linker, revealing a means to allow AP-2–mediated coordination of accessory protein recruitment and clathrin polymerization at sites of vesicle formation. Knockdown and functional rescue studies demonstrate that through these interactions, NECAP 1 and AP-2 cooperate to increase the probability of clathrin-coated vesicle formation and to control the number, size, and cargo content of the vesicles. Together, our data demonstrate that NECAP 1 modulates the AP-2 interactome and reveal a new layer of organizational control within the endocytic machinery.
| Clathrin-mediated endocytosis is the major entry portal for cargo molecules such as nutrient and signaling receptors in eukaryotic cells. Generation of clathrin-coated vesicles involves a complex protein machinery that both deforms the membrane to generate a vesicle and selects appropriate cargo. The endocytic machinery is formed around the core endocytic adapter protein AP-2, which recruits clathrin and numerous accessory proteins to the site of vesicle formation in a temporally and spatially controlled manner. Yet it remains elusive as to how these interactions are regulated to ensure efficient vesicle formation. Here we identify the endocytic protein NECAP 1 as a modulator of AP-2 interactions. We show that NECAP 1 and AP-2 form two functionally distinct complexes. In the first, NECAP 1 binds to two sites on AP-2 in such a manner as to limit accessory protein binding to AP-2. Recruitment of clathrin to vesicle formation sites displaces NECAP 1 from one of these sites, leading to the formation of a second complex in which NECAP 1 and AP-2 cooperate for efficient accessory protein recruitment. Through these interactions, NECAP 1 fine-tunes AP-2 function and the two proteins cooperate to increase the probability that a vesicle will form and to determine the size and cargo content of the resulting vesicle.
| Clathrin-mediated endocytosis is the major vesicular entry route in mammalian cells. The formation of endocytic clathrin-coated vesicles (CCVs) depends on a complex protein machinery that deforms the planar plasma membrane into small, cargo-laden vesicles that are released into the cytosol [1],[2]. The endocytic machinery is organized around AP-2, a heterotetrameric protein complex in which the N-terminal regions of its two large subunits, α and β2, together with the two smaller subunits, σ2 and μ2, form a large globular domain referred to as the trunk [3]. The C-terminal region of each large subunit forms a small, bi-lobed globular domain termed the ear, and the two ears connect to the trunk through flexible linkers. Each of these regions mediates a specific function of the complex, allowing AP-2 to control the recruitment of a myriad of endocytic accessory proteins, cargo, and clathrin to PI(4,5)P2-rich sites of CCV formation at the plasma membrane [1],[2].
At the protein level, the endocytic machinery is built on the basic principle that short peptide motifs in unstructured or loosely structured regions of one protein mediate low-affinity interactions with a globular domain in a second protein [4]–[6]. In isolation, each interaction is of very low affinity and easily reversible; however, each protein has the potential to simultaneously interact with a number of binding partners to create an interaction network that stabilizes the protein coat around the forming vesicle.
Perhaps the least understood step of CCV formation is vesicle initiation, with two main models proposed to explain how new clathrin-coated pits (CCPs) are nucleated. In one, the FCHo complex, formed by FCHo1/2, Eps15, and intersectin, is recruited to PI(4,5)P2-rich sites at the plasma membrane, marking these sites for subsequent recruitment of clathrin/AP-2 [7],[8]. In the other, pits initiate by arrival of clathrin/AP-2 to PI(4,5)P2-rich sites with the FCHo complex recruited subsequently [9]. It is likely that in either scenario, clathrin/AP-2 complexes will need to be linked to the FCHo complex; however, the mechanisms that allow for efficient interconnection of the two complexes remain elusive.
Each endocytic accessory protein contributes to one or more specific aspects of vesicle formation such as membrane deformation, cargo recruitment, vesicle size control, and vesicle scission; thus, each needs to gain access to vesicle formation sites in the correct temporal order [2]. Many proteins target these sites through interactions with the globular ear domains of AP-2. During the course of vesicle formation, the β2-ear transitions from accessory protein binding to recruiting clathrin in conjunction with the β2-linker [10]–[12], whereas the α-ear serves as the main interface for accessory protein recruitment throughout the process [13]–[15]. All α-ear binding partners use one or more of three distinct peptide motifs to bind the ear; DPF/W and FxDxF motifs target the platform subdomain of the α-ear, while WxxF motifs bind the sandwich subdomain [6],[16]. Yet the mechanisms that modulate α-ear interactions allowing for appropriate recruitment of accessory proteins remain unknown.
In many cases, cargo selection and concentration in forming vesicles also depend on AP-2 [1]. Some cargo, such as the transferrin receptor, interact directly with the trunk region of AP-2. Other cargo, such as the low-density lipoprotein receptor, interact with alternate cargo adaptors, a heterogeneous subclass of accessory proteins that in turn interact with AP-2 and/or clathrin to bridge their cargo to the endocytic machinery [1],[17]. Therefore, mechanisms that control accessory protein recruitment to forming vesicles also provide a means to modulate vesicle cargo.
We identified NECAP 1 and 2 as CCV-enriched proteins using subcellular proteomics [18],[19]. The two proteins are only 62% identical at the amino acid level, and they function in distinct membrane trafficking processes. Knockdown (KD) of NECAP 2, a ubiquitously expressed isoform, does not appear to influence clathrin-mediated endocytosis but instead inhibits endosomal sorting (our preliminary data). NECAP 1, which functions in clathrin-mediated endocytosis [19], is primarily expressed in neurons but is also readily detected in cultured cell lines (Figure S1), which offer an easy-to-manipulate system to probe for its mechanistic role in this process. Through a WxxF motif at its C-terminus, NECAP 1 binds with high affinity to the α-ear sandwich subdomain of AP-2 [20]. The NECAP 1 N-terminus is well conserved with NECAP 1 orthologs in other species and encodes a globular PH fold, termed PHear [21]. PHear binds to accessory proteins harboring FxDxF motifs, similar to the platform subdomain of the AP-2 α-ear [21]. NECAP 1 binding to the α-ear sandwich site positions PHear in proximity of the α-ear platform, suggesting that NECAP 1/AP-2 complexes act synergistically to control the recruitment of FxDxF-containing accessory proteins to forming vesicles. In the current study we set out to examine the mechanistic role of NECAP 1 in endocytosis and we now report that NECAP 1 works cooperatively with AP-2 to recruit endocytic accessory proteins that control the number, size, and cargo content of CCVs.
The N-terminal region of NECAP 1 (aa 1–178) is highly conserved in NECAP 1 orthologs in other species, and residues 1–133 encode a PH-like domain [21]. We termed the domain PHear, as it has a PH fold and interacts with FxDxF motifs similar to the AP-2 α-ear [21]. Since the conservation extends beyond the C-terminal border of PHear (aa 129–178, termed Ex, Figure 1A), we reasoned that Ex could also have an important functional role. Ex shows no structural organization by NMR, on its own (Figure 1B) or in tandem with PHear [21]. Interestingly, when the entire conserved region of NECAP 1 (aa 1–178, termed PHear–Ex, Figure 1A) was used in affinity-selection assays from rat brain lysate, mass spectrometry of the isolated proteins identified multiple subunits of AP-2 (unpublished data). As the C-terminal WxxF motif that mediates NECAP 1 binding to the α-ear is not located within PHear–Ex (Figure 1A), these data indicate that PHear and/or Ex provide a second mechanism for NECAP 1 to engage AP-2. Western blot analysis confirmed that endogenous AP-2 is affinity-selected by GST–PHear–Ex (Figure 1C). AP-2 binding is stronger with PHear–Ex than PHear alone (Figure 1C), suggesting the presence of two AP-2 binding sites in the conserved N-terminus of NECAP 1. This was confirmed using PHear and Ex in isolation, which both bind AP-2 (Figure 1D). The observation that AP-2 binding to PHear–Ex is greater than the combined binding of PHear and Ex in isolation (Figure 1D) indicates that the two binding sites cooperate in AP-2 interaction. This is underscored by the fact that Flag-tagged PHear–Ex co-immunoprecipitates endogenous AP-2 from 293-T cell lysates while Flag-tagged PHear alone does not (Figure 1E). Thus, NECAP 1 has three distinct AP-2 binding sites, the high affinity WxxF motif at the C-terminus, and two lower affinity AP-2 binding sites in the conserved N-terminus, one in PHear and one in Ex.
To better define the interaction of PHear and Ex with AP-2, we first mutated K154 and G156, the only two amino acids within Ex that are invariant throughout evolution (Figure 1F). This double point mutation reduced AP-2 binding of PHear–Ex to levels seen with PHear alone (Figure 1G), further supporting the presence of a site in Ex that interacts with AP-2. As for PHear, we tested an array of previously generated PHear variants that contain mutations disrupting interactions with FxDxF motifs [21]. With only a few exceptions, mutations that reduce or eliminate binding to FxDxF motif proteins such as amphiphysin I and II also interfere with AP-2 binding (Figure 1H). Therefore, PHear uses an overlapping interface for binding to FxDxF motif proteins and AP-2, indicating that during CCV formation, NECAP 1 may transition between different roles by changing PHear binding partners.
To better understand the multiple PHear–Ex interactions, we first sought to identify the binding site(s) for PHear and PHear–Ex in AP-2. In affinity selection assays, GST-α-ear shows no binding to PHear or PHear–Ex, despite strong binding to full-length NECAP 1, which contains the C-terminal WxxF motif (Figure 2A). Coupled with the fact that NECAP 1 does not bind the AP-2 β2-ear [19], it becomes clear that PHear and Ex do not target the canonical AP-2 ear interfaces.
We next sought to identify the subunit of the AP-2 heterotetramer involved in PHear–Ex binding. Overlays (far-Western blots) provide a powerful means to identify and study protein–protein interactions. For example, synaptojanin was originally identified based on its interaction with the SH3 domains of Grb2 in overlay assays [22]. We thus performed overlay assays on coat proteins stripped from purified CCVs using soluble GST–PHear and GST–PHear–Ex, and both constructs revealed a strong signal at approximately 110 kDa, near the apparent MW of the two large AP-2 subunits α and β2 (Figure 2B). To distinguish between these subunits and further map the site, we immunoprecipitated Flag-tagged deletion variants of α and β2 containing either the trunk and linker or the linker and ear (Figure 2C) and used these in overlays. Both PHear and PHear–Ex interact with β2 variants with no binding to the α variants (Figure 2D). We next used variants containing β2 trunk alone, trunk and linker, linker and ear, and ear alone (Figure 2C). PHear and PHear–Ex only bind β2 variants containing the linker (Figure 2E). Throughout these experiments, the presence of Ex did not change the binding pattern of PHear–Ex compared to PHear. Unfortunately, we are unable to map the Ex binding site on AP-2 as Ex in isolation shows limited AP-2 binding (Figure 1D).
Mapping the site of PHear interaction to the β2-linker by overlay allowed us to confirm the interaction using in-solution affinity selection assays. GST–β2-linker specifically bound NECAP 1 from brain extracts with binding similar to that seen for clathrin, the only other protein known to interact with the β2-linker (Figure 2F). NMR analysis reveals that the β2-linker binds a site on PHear that overlaps with the interface for FxDxF motifs (Figure 2G–I) and titrations of β2-linker with PHear reveal a mean KD of 480 µM (Figure 2J,K). While the affinity of PHear for the β2-linker is low in isolation, full-length NECAP 1 has two additional binding sites on AP-2 that provide avidity effects, the low affinity site in Ex and the high affinity WxxF motif at the C-terminus (Figure 1C).
Using C-terminal deletion variants of the β2 trunk and linker in overlay assays, we determined that critical residues for PHear binding are located in the linker between amino acids 633 and 640 (Figure 2L,M). Intriguingly, this area overlaps with the binding motif for the terminal domain of clathrin heavy chain (Figure 2L) [23]. Together, these data indicate that PHear binds the β2-linker of AP-2 and that clathrin and NECAP 1 may compete for access to the β2-linker.
To test for competition between clathrin and NECAP 1 for AP-2 binding, we performed NMR studies with purified, 15N-labeled PHear in the presence of β2-linker and increasing concentrations of clathrin terminal domain. β2-linker causes chemical shift changes in PHear as it interacts with residues in the PHear binding site and these chemical shift changes are reversed to the ligand-free position in the presence of increasing concentrations of terminal domain (Figure 3A). We next performed more classical in-solution competition experiments. AP-2 was partially purified from stripped CCVs using gel filtration chromatography and bound to GST-clathrin terminal domain immobilized on Sepharose beads. Addition of increasing concentrations of purified PHear–Ex causes a dose-dependent decrease in the binding of AP-2 to clathrin terminal domain (Figure 3B), confirming that PHear–Ex and clathrin compete for binding to the β2-linker.
In summary, the data presented in Figures 1–3 indicate that NECAP 1 and AP-2 form two distinct complexes. In the first, the WxxF motif targets the α-ear, while PHear targets the β2-linker. This complex would likely be pre-endocytic because PHear occupies the clathrin-binding site in the β2-linker (Figure 3C). Once vesicle formation is initiated, clathrin competes PHear off the β2-linker, leading to the formation of the second complex, in which NECAP 1 remains bound to the sandwich domain of the AP-2 α-ear through the high-affinity WxxF motif, while PHear and the α-ear platform domain are positioned to cooperate in accessory protein recruitment (Figure 3C). Consistently, we detect cooperation of PHear and α-ear in binding to the FxDxF proteins amphiphysin I and II and synaptojanin 170 (binding is greater when the two domains are mixed than the sum of binding to the two domains on their own) (Figure 3D).
To address which steps of vesicle formation are dependent on the cooperation of NECAP 1 and AP-2, we knocked down NECAP 1 in COS-7 cells (Figure 4A), which express NECAP 1 at levels similar to other cultured cell lines (Figure S1) and tested for changes in vesicle formation at the plasma membrane. Two obvious alterations were observed following NECAP 1 KD: (1) a reduction in the number of AP-2 puncta at the plasma membrane (Figure 4B,C) and (2) an increase in AP-2 signal intensity in a large proportion of the remaining AP-2 puncta (Figure 4B). These AP-2 structures still co-localize with clathrin (Figure 4D), suggesting that they are functional vesicle formation sites. Endocytic vesicles labeled with transferrin following 1 min of uptake are also fewer in number and brighter following NECAP 1 KD (Figure 4E), indicating that the changes occurring at the level of the plasma membrane are maintained in the early endocytic pathway.
There is a direct correlation between the immunofluorescence signal intensity of coat proteins such as AP-2 and clathrin and the size of the forming structure [24]–[26]. To examine if the increase in AP-2 signal intensity reflects an increase in the size of vesicle formation sites, we used 3D superresolution microscopy (Figure 4F–J, Figures S2 and S3, and Movies S1, S2, S3, S4). Quantification of the diameters of deeply invaginated CCPs in x, y, and z confirms that NECAP 1 KD causes a size increase in all three dimensions (Figure 4F–H). There is also an increase in the number of AP-2 signals detected per vesicle formation site (Figure 4F,G,I), confirming that the increase in AP-2 signal intensity seen by confocal microscopy serves as a reliable readout of increased size. In addition to CCPs, some cell lines form large planar clathrin-coated plaques but only at the membrane opposed to the glass surface (bottom) [27]. To address whether NECAP 1 KD leads to an increase in CCP size or to a shift towards coated plaques, we compared the dimensions of AP-2-labeled structures on the membrane contacting the coverslip (bottom) and on the cell surface facing the culture medium (top). Both locations show the same three-dimensional increase in diameter (Figure 4J, Figure S3), indicating that NECAP 1 is needed to control the size of CCPs. To further validate this result, we performed EM analysis (Figure 4K,L). We measured the depth of the pits and placed them into three bins of 0–50 nm, 50–100 nm, and 100+ nm depth. In all three bins, the mean width of the pits was significantly increased in the NECAP 1 KD cells (Figure 4L). Notably, we found no evidence for clusters of CCPs, indicating that NECAP 1 KD causes an increase in pit diameter and not an increased clustering of pits.
We next used EM to investigate if the increase in the size of vesicle formation sites translates into larger CCVs. Indeed, the absence of NECAP 1 causes a population-wide shift towards larger CCVs (Figure 4M). The EM data also confirm that despite the increase in size, the clathrin-coated structures still complete the vesicle formation process (Figure 4M).
Controlling the size and number of CCVs is critical in all cell systems but nowhere more so than in the presynaptic nerve terminal. Synaptic vesicles are the smallest known transport vesicles and are reformed by clathrin-mediated endocytosis following neurotransmitter release [28]–[32]. Size control is especially important for synaptic vesicles as their size determines the neurotransmitter content [33]. NECAP 1 is expressed at highest levels in brain and is enriched in purified CCVs and synaptic vesicles [19],[34]. KD of NECAP 1 in primary hippocampal neurons leads to a reduction in synaptic vesicle number and an average increase of 12–14% in synaptic vesicle diameter (Figure S4), which translates to a change in vesicle volume of nearly 50%. Thus, NECAP 1 serves a similar function in the endocytic machineries of nonneuronal cells and neurons. We assume that the effects on vesicle number and size following NECAP 1 KD are indirect as NECAP 1 functions together with AP-2 to coordinate the recruitment of endocytic accessory proteins during vesicle formation.
To determine how NECAP 1 is involved in controlling the number and size of CCVs, we performed rescue experiments with NECAP 1 wild-type and point mutants. One essential feature of NECAP 1 is its ability to interact with the AP-2 α-ear and a NECAP 1 variant in which the high affinity C-terminal WxxF motif is inactivated fails to rescue the NECAP 1 KD phenotypes (Figure 5A,B). In contrast, re-expression of wild-type NECAP 1 restores the number of vesicle formation sites (Figure 5A,B) and also leads to lower AP-2 intensity levels in these structures. To address the importance of PHear-mediated interactions, we tested a mutant (R95E) that disrupts PHear binding to FxDxF motifs and the AP-2 β2-linker (Figure 1H). This mutant fails to rescue the NECAP 1 KD phenotype (Figure 5A,B) despite its ability to target the AP-2 α-ear as seen by its co-immunoprecipitation with AP-2 (Figure 5C). A NECAP 1 mutant in which the AP-2-binding site in Ex was mutated (K154A/G156S) (Figure 1G) rescues the phenotype (Figure 5A,B). This suggests that the ability of Ex to enhance AP-2 interactions with the NECAP 1 N-terminus is dispensable during vesicle formation. Within full-length NECAP 1, Ex has by far the lowest affinity to AP-2 and if at all, may only play a role in pre-endocytic NECAP 1/AP-2 interactions. Finally, expression of wild-type NECAP 2 [19] does not rescue the NECAP 1 KD phenotype (Figure 5A,B), demonstrating that the two mammalian NECAP isoforms are functionally divergent. This is consistent with our observations that NECAP 2 is not involved in clathrin-mediated endocytosis but instead functions in endosomal sorting (our unpublished data).
Our findings indicate that NECAP 1 functions to modulate the ability of AP-2 to recruit accessory proteins during vesicle formation. We thus set out to determine which accessory proteins might be responsible for the alterations in vesicle size and number seen following NECAP 1 KD. In respect to vesicle size, we tested for an effect of NECAP 1 KD on AP180/CALM. CALM is a clathrin- and AP-2-binding endocytic protein, and the increase in vesicle size in NECAP 1 KD cells is reminiscent of a phenotype observed in CALM KD cells [35]. Similarly, functional disruption of the neuronal isoform of CALM, AP180 leads to an increase in the size of synaptic vesicles [36]–[40]. Both CALM and AP180 contain FxDxF motifs that interact with NECAP 1 [21], and NMR binding studies demonstrate that an FxDxF-motif peptide derived from AP180 binds to the canonical interface on PHear (Figure 6A). The increase in vesicle size following NECAP 1 KD could thus result from reduced levels of CALM [35] and indeed, in the absence of NECAP 1, less CALM is observed at sites of vesicle formation (Figure 6B,C). Overexpression of CALM in NECAP 1 KD cells could provide a means to increase CALM at these sites and thus allow the endocytic machinery to rebalance. Indeed, CALM expression in the absence of NECAP 1 led to a decrease in the size of vesicle formation sites as judged by AP-2 intensity (Figure 6D). Thus, NECAP 1 is required for efficient recruitment of CALM to vesicle formation sites where CALM regulates vesicle size, either directly or indirectly.
CALM also plays a role in cargo selection by CCPs, even though the cargo-specific effects remain poorly understood. For example, CALM serves as a cargo adapter for R-SNAREs [40]–[42]. In addition, KD of CALM reduces clathrin-dependent endocytosis of amyloid precursor protein and EGF receptor without influencing transferrin receptor endocytosis [43],[44]. Given that NECAP 1 KD reduces the level of CALM recruited to forming vesicles, we hypothesized that NECAP 1 KD would lead to a selective disruption of cargo entry. Indeed, NECAP 1 KD decreases the clathrin-dependent internalization of EGF by over 40% (Figure 6E,F), similar to the reduction in EGF internalization seen upon CALM KD [44], while transferrin endocytosis and recycling was not altered (Figure 6G–I). The changes in vesicle size and cargo we observe upon NECAP 1 KD are thus likely a result of the reduced levels of AP180/CALM recruited during vesicle formation.
To better understand how NECAP 1 KD causes a decrease in vesicle number, we tested for effects on the FCHo complex. The members of the FCHo protein family form tri-partite complexes with the endocytic accessory proteins Eps15 and intersectin [7],[8]. FCHo expression levels directly correlate with the number of vesicle formation sites and successful endocytic events, while destabilization of FCHo complexes interferes with vesicle formation [8],[9]. Interestingly, both FCHo1 and 2 interact with PHear and this binding is reduced in the PHear variant R95A (Figure 6J), which also shows reduced binding to FxDxF proteins and AP-2 (Figure 1H). Testing a range of deletion variants revealed that PHear binds the central linker region of FCHo1 (Figure 6K), which is flanked by the N-terminal FCH domain and the C-terminal mu homology domain. NECAP 1 thus provides a parallel mechanism to Eps15 to interlink the FCHo complex with AP-2, helping to stabilize and maintain normal numbers of vesicle formation sites. Given the interaction between NECAP 1 and FCHo1/2, we tested for an effect of NECAP 1 KD on the FCHo complex using FCHo2 as a marker. NECAP 1 deletion causes a decrease in the number of FCHo2 puncta at the plasma membrane (Figure 6L,M), revealing that NECAP 1 is needed to efficiently stabilize the FCHo complex at the membrane. Given the correlation between FCHo expression levels and number of vesicle formation sites [8],[9], we reasoned that FCHo overexpression would increase the number of formation sites in NECAP 1 KD cells. Indeed, overexpression of FCHo1 in NECAP 1 KD cells efficiently increased the number of AP-2 puncta at the plasma membrane (Figure 6N).
Together, these data demonstrate that the cooperation of NECAP 1 and AP-2 allows for efficient protein recruitment into the endocytic protein network to control important aspects of CCV formation such as vesicle number, cargo content, and size.
Most proteins of the endocytic machinery are categorized as accessory proteins. In general, accessory proteins are recruited to sites of vesicle formation in a temporally controlled manner to facilitate specific steps—for example, the FCHo complex and/or clathrin/AP-2 [7]–[9] initiate vesicle formation at PI(4,5)P2-rich spots on the plasma membrane, epsin and clathrin are involved in membrane deformation during invagination, and a late burst of dynamin recruitment is required for vesicle scission [2]. As such, efficient vesicle formation depends on recruiting the correct set of accessory proteins in sufficient amounts at specific times. However, we have only rudimentary insights into how recruitment of accessory proteins is coordinated and regulated during CCV formation.
AP-2 is a central hub of the endocytic machinery, coordinating the recruitment of clathrin, cargo, and numerous accessory proteins. The data presented here identify NECAP 1 as a modulator of AP-2 interactions (Figure 7). NECAP 1 engages the sandwich domain of the AP-2 α-ear through its C-terminal WxxF motif, while PHear engages the β2-linker at a site overlapping with the binding site for the terminal domain of clathrin. Engagement of the β2-linker by clathrin frees PHear for synergistic interactions along with the platform part of the α-ear for recruitment of endocytic accessory proteins (Figure 7).
Recent studies with endogenously tagged proteins suggest that cargo recognition is not required to stabilize vesicle formation sites [45]. Therefore, the rate and efficiency of vesicle initiation is the major and perhaps only step that controls the endocytic capacity of a cell. However, the low number of molecules involved still hampers our ability to directly study these early events of vesicle formation in great detail. In one model of vesicle initiation, AP-2 stochastically samples the plasma membrane for PI(4,5)P2-rich sites, where it coordinates the recruitment of other endocytic proteins [9]. An alternative model [8] favors the idea that vesicle initiation sites are determined by the recruitment of the FCHo complex to PI(4,5)P2-positive sites at the plasma membrane, with subsequent incorporation of AP-2 and clathrin. If one considers vesicle initiation as a short temporal window of opportunity in which the stochastic association of AP-2 and the FCHo complex with PI(4,5)P2 brings these factors together to allow for their interaction, the success rate of vesicle formation would depend on stabilizing these factors at the membrane long enough to nucleate the formation of the endocytic protein network. NECAP 1 is ideally positioned to accomplish this stabilization, and our data demonstrate that NECAP 1 is required to maintain normal numbers of vesicle formation sites. When the NECAP 1/AP-2 complex is recruited to vesicle initiation sites, the NECAP 1 PHear can release from the β2-linker and cooperate with AP-2 to engage the FCHo complex in a manner parallel to that of Eps15, thereby increasing the probability of coincidence detection that triggers vesicle formation (Figure 7). It is tempting to speculate that recruitment of clathrin to initiation sites functions as the switch to trigger vesicle formation, given that the clathrin terminal domain competes PHear off the β2-linker, thereby promoting the cooperation of PHear and α-ear in FxDxF protein binding. Concomitantly, the β2-ear and linker are now fully accessible and can recruit and polymerize clathrin at the site of vesicle formation, providing a scaffold for the growing protein network. The fact that FCHo overexpression in NECAP 1 KD cells overcomes the KD phenotype and rebalances vesicle numbers further attests to the necessity of stabilization factors such as NECAP 1 to concentrate accessory proteins in amounts sufficient for efficient vesicle formation.
The composition of the endocytic machinery also controls cargo recruitment. Some cargo is recognized by alternate adaptors, accessory proteins with the ability to link their cargo to AP-2 and clathrin [1]. Interestingly, the endocytic machinery also has an inherent ability to translate cargo content into vesicle size adaptation—for example, vesicles that internalize ligand-bound low-density lipoprotein receptors or viruses increase in size to harbor such large cargo [2],[24]. On the other hand, synaptic vesicles are recycled such that their characteristic small size is maintained. For neuronal transmission, this is important for two reasons: first, smaller vesicles can be formed faster and this might be needed to maintain synaptic vesicle pools during times of high neuronal activity. Second, the size of a vesicle directly determines vesicle volume, which in turn determines the quantal amount of neurotransmitter, with implications for the strength of synaptic transmission and synaptic potentiation [46]. Consistently, alterations in synaptic vesicle size have been demonstrated to cause changes in neurotransmission in nonmammalian and mammalian systems alike [39],[44],[46],[47]. The predominantly neuronal expression pattern of NECAP 1, together with the fact that NECAP 2 does not function in clathrin-mediated endocytosis, suggests that a tightly controlled and efficient recruitment of accessory proteins may be most critical at the synapse. However, the molecular mechanisms that determine vesicle size and couple cargo to size regulation remain elusive.
In nonneuronal cells, deletion of CALM causes an increase in vesicle size [35]. Similarly, deletion or mutation of the neuronal isoform AP180 leads to larger synaptic vesicles [36]–[40], demonstrating the importance of these accessory proteins for vesicle size control even though the precise mechanism of their function remains unclear. Both proteins also have cargo-specific effects, AP180 and CALM recruit synaptobrevin during synaptic vesicle recycling, and CALM also promotes EGF receptor and amyloid precursor protein internalization in nonneuronal cells [38],[40],[43],[44],[47]. Our study reveals that the cooperation of AP-2 and NECAP 1 is crucial for efficient CALM recruitment and suggests that decreased CALM levels due to NECAP 1 ablation cause the formation of larger vesicles. This is confirmed by the fact that CALM overexpression in NECAP 1 KD reverses the phenotype and allows for the formation of small vesicle formation sites in the absence of NECAP 1. It remains formally possible that the size increase in vesicle formation sites observed following NECAP 1 KD results from clustering of vesicle formation sites or an increase in endocytic hot spots [48], but our EM analysis of vesicle formation sites does not support these possibilities.
Live-cell imaging reveals virtually identical recruitment profiles of CALM and NECAP 1 during vesicle formation [49], strongly supporting the idea that NECAP 1 and CALM share a common function. Moreover, NECAP 1 KD specifically reduces clathrin-dependent endocytosis of EGF receptor with no effect on transferrin receptor internalization, reminiscent of the phenotypes specific for CALM KD [44]. Together, these data demonstrate the importance of NECAP 1 for efficient accessory protein recruitment to sites of vesicle formation and reveal how balancing the levels of endocytic proteins during vesicle formation in turn controls vesicle size and cargo selection.
Mouse monoclonal antibodies against CHC (clone 23), α-adaptin (clone 8, for Western blotting), γ-adaptin (clone 88), and EEA1 (clone 14) were from BD Transduction Laboratories (Lexington, KY). Mouse monoclonals against α-adaptin (AP.6, for immunofluorescence) and Flag (M2) were from Thermo Scientific and Sigma (St-Louis, MO), respectively. Rabbit polyclonal antibody against c-myc (A-14) and goat polyclonal antibody against CALM (C-18) were from Santa Cruz and polyclonal antibodies against amphiphysin I/II (1874), clathrin light chains, and NECAP 1 and 2 have been described previously [19],[20],[50],[51]. The antibody against synaptojanin 170 was a generous gift of Dr. Pietro De Camilli (Yale University). AlexaFluor 633–conjugated human transferrin (T-23362) and biotinylated EGF complexed to Texas Red-Streptavidin (E-3480) were from Invitrogen and Cy3-conjugated human transferrin (009-160-050) was from Jackson ImmunoResearch (West Grove, PA). The mCherry tag was detected by Western blot using a mouse monoclonal antibody against RFP (abcam, Cambridge, MA, ab65856). The synthetic amphiphysin I peptide was purchased from the HHMI/Keck Biotechnology Resource Laboratory, Yale University. The synthetic peptides for AP180 and the β2-linker were purchased from Sheldon Biotechnology Center at McGill University.
The following constructs were described previously: GST–NECAP 1, GST–α-ear, and Flag–PHear–Ex (termed aa1–178) [19], Flag–NECAP 1 and Flag–NECAP 1 W272A, F275A (termed AVQA) [15], GST–PHear, GST–PHear–Ex (termed aa1–178), GST–Ex (termed aa129–178), and the GST–PHear point mutants [21]. cDNA clones for NECAP 1 (gi:27229051), NECAP 2 (gi:13384759), α-adaptin (gi:163644276), β2-adaptin (gi:71773037), FCHo1 (gi:255683302), and FCHo2 (gi:30854355) were used as PCR templates, and point mutations as needed were introduced using the megaprimer procedure [52]. For bacterial expression, inserts were subcloned into pGEX-4T1 or pGEX-6P1. For mammalian expression, inserts were subcloned into pcDNA3 with integrated Flag- or myc-tags (described in [20],[53]). For NMR studies using purified Ex, a PCR-amplified DNA fragment encoding amino acids 128–178 was subcloned into pPROEX-HTb for the expression of N-terminally His-tagged protein. For NECAP 1 KD, NECAP 1–specific target sequences for human and rat protein were designed using the Block-iT RNAi Designer (Invitrogen) and subcloned into pcDNA6.2/GW-EmGFP-miR (Invitrogen) following the manufacturer's instructions. The EmGFP-miR cassette was then amplified by PCR and subcloned into the pRRLsinPPT vector to generate the microRNA expression vectors. The number given in the name of each KD virus corresponds to the first nucleotide position targeted in the mRNA. The control virus has been described previously [54]. For generation of rescue/protein expression viruses, the microRNA part of the EmGFP expression cassette in pRRLsinPPT was replaced by a polylinker, which was subsequently used to clone Flag-tagged variants of NECAP 1 and 2 in frame with EmGFP such that EmGFP–Flag–NECAP fusion proteins were expressed. Expression vectors for mCherry-tagged FCHo1 and 2 were from addgene (Cambridge, MA, plasmids 27690 and 27686, respectively). The construct for GST-clathrin terminal domain was a gift of Dr. James Keen (Thomas Jefferson University, Philadelphia, PA).
HEK-293-T and COS-7 cells were maintained in DMEM High Glucose (Invitrogen) containing 10% FBS (PAA Laboratories Inc.), 100 U/ml penicillin, and 100 µg/ml streptomycin (both Invitrogen).
For expression of VSVG pseudotyped virus, HEK-293-T were seeded with 107 cells/plate on 15 cm plates in 25 ml of regular culture medium, with six plates for each virus. The following day, each microRNA or protein expression vector was co-transfected with a packaging mix (containing pMD2.g, pRSV-Rev, and pMDLg/pRRE, Addgene) using calcium phosphate. After 8 h, the medium was removed from each plate and replaced with 15 ml of collection medium per plate (regular medium supplemented with 1× nonessential amino acids (Gibco) and 1 mM sodium pyruvate (Gibco)). At 24, 36, and 48 h posttransfection, the medium was removed from each plate and for the 24 and 36 h time point, replaced with 15 ml collection medium. The supernatants for each construct and each collection were combined and stored at 4°C until the end of the collection procedure. The supernatants were then filtered through a 0.45 µm PES membrane and the virus was concentrated by centrifugation (8 h at 17,000× g), and the resulting pellets were resuspended in DMEM in 1/2,000 of the original volume. Concentrated virus was stored at −80°C until use. To determine the virus titer, HEK-293–T cells were plated in 24-well plates with 40,000 cells/well in regular medium. At 10–14 h postplating, the medium was replaced with regular medium containing varying amounts of concentrated virus. The next day, 1 ml regular medium was added to each well. Three days after transduction, the medium was replaced with 300 µl PBS to allow better visualization of the GFP fluorescence, and transduction efficiency was calculated based on the percentage of GFP-positive cells for the different virus dilutions. The MOIs used for COS-7 cells and primary hippocampal neurons are arbitrary MOIs based on the HEK-293–T cell transduction rate.
Statistical tests and posttests used are indicated in the figure legends where appropriate. N indicates the number of independent repeats analyzed, and n indicates the size of the total pool analyzed, if applicable.
For KD studies in COS-7 cells, cells were plated on the day of transduction. For transduction, the culture medium was replaced by DMEM High Glucose (Invitrogen) supplemented with 2% heat-inactivated FBS, 100 U/ml penicillin, 100 µg/ml streptomycin, and 6 µg/ml polybrene (Sigma), and viruses were added at an MOI of 10. The next day, media was replaced with fresh culture medium and the cells were cultivated until assays were performed 6 d after transduction. In some cases, COS-7 KD cells were transfected using jetPRIME (Polyplus Tranfection) following the manufacturer's instruction 5 d after transduction and processed for immunofluorescence following overnight incubation. For rescue studies, KD cells were plated on coverslips on day 5 after transductions. On the same day, cells were transduced with rescue viruses using an MOI of 4 as described above. The media was replaced early the next morning, and cells were processed for immunofluorescence 24 h after the second transduction.
For analysis of KD COS-7 cells, cells plated on poly-L-lysine-coated coverslips were processed for immunofluorescence following standard protocols 6 d after transduction with for COS-7 cells. COS-7 cells transduced with protein expression viruses or transfected with expression constructs were processed 1 or 2 d after manipulation. Images were analyzed using Image J (National Institutes of Health, Bethesda, MD).
Control and NECAP 1 KD cells (nt220) were plated with 40,000 cells/well on poly-L-lysine-coated 8-well Lab-Tek II chambered coverglasses #1.5 (cat. no. 155409). The following day, cells were fixed with 2% paraformaldehyde for 10 min at RT, processed for immunofluorescence detection of endogenous AP-2 using Alexa647-conjugated secondary antibodies, and stored in PBS until imaging. Images were recorded with a SR 200 microscope (Vutara, Inc., Salt Lake City, UT) based on the Biplane FPALM approach [55]. The system features four laser lines (405, 488, 561, and 647 nm) for excitation and activation of single fluorescent molecules. Speckle-free illumination with an even intensity distribution is realized by a specialized beam homogenizer. Images of fluorescing molecules are recorded with a 60×/1.2NA Olympus water immersion objective on an Photometrics Evolve 512 EM-CCD camera with the gain set at 50. Each acquisition consisted of 7,000 frames recorded at a speed of 40 frames/s, which encompassed a 20×20 µm field of view. The maximum powers used for the readout laser (647 nm) and activation laser (405 nm) were 4 and 0.05 kW/cm2, respectively.
The calibration entails experimentally calculating the point spread function (PSF) in three dimensions. This was done using 100 nm Tetraspeck beads. The analysis and rendering were done using Vutara's SRX localization and visualization software, based on an enhanced implementation of Juette et al. [55] and Mlodzianoski et al. [56]. Data were analyzed by the Vutara SRX software (version 3.16). In short, particles were identified by their brightness from the combined images taken in both planes simultaneously. If a particle was identified in multiple subsequent camera frames, data from these frames were combined for the specific identified particle. Background can be optionally removed based on the observed signal in the frames before or after the frames in which a particle was observed. Identified particles were then localized in three dimensions by fitting the raw data in a customizable region of interest (typically 16×16 pixels) centered around each particle in each plane with a 3D model function that was obtained from recorded bead datasets. The recorded fields are aligned automatically by computing the affine transformation between the pair of planes. Sample drift can be corrected by cross-correlation of the determined localized particles [56] or tracking of fiduciary markers. Fit results were stored as data lists for further analyses. The SRX software allows the 3D display of localized particles as solid shaded spheres or as an accumulation of transparent gaussian kernels. Alternatively, 2D slices through the 3D volume in any of the three main directions can be shown.
COS-7 cells transduced with control and NECAP 1 KD viruses were starved in DMEM High Glucose overnight. For microscopic analysis, cells were chilled on ice for 30 min and then incubated with Cy3-transferrin (200 µg/ml) in ice-cold DMEM on ice for 1 h. Cells were washed with cold PBS and incubated in prewarmed culture medium at 37°C for the times indicated. At each time point, a sample of cells was chilled on ice, surface-bound transferrin was removed by acid wash (0.2 M acetic acid, 0.5 M NaCl), followed by a PBS wash. The cells were fixed with 4% PFA and processed for immunofluorescence detection of marker proteins if indicated. For FACS analysis, cells were chilled on ice for 30 min and then incubated with Alexa633-transferrin (200 µg/ml) in ice-cold DMEM on ice for 1 h. Cells were washed with cold PBS and incubated in prewarmed culture medium at 37°C for the times indicated. At each time point, a sample of cells was chilled on ice, surface-bound transferrin was removed by acid wash (0.2 M acetic acid, 0.5 M NaCl), followed by a PBS wash. The cells were removed from the plate in 1 ml of PBS by pipetting, filtered through a cell strainer, and analyzed by flow cytometry on a FACSCalibur (Becton Dickinson). For EGF internalization assays, control and NECAP 1 KD cells were maintained in regular culture medium at 37°C. The regular medium was replaced with prewarmed DMEM High Glucose medium containing 2 ng/ml of Texas Red-labeled EGF for 2.5 min at 37°C. The cells were then chilled on ice and surface-bound EGF was removed by acid wash, followed by a PBS wash. The cells were fixed with 4% PFA, and internalized Texas Red-EGF was detected by fluorescence microscopy. For comparison, parallel sets of cells were processed to detect Cy3-transferrin (200 µg/ml) internalization under these condition.
COS-7 cell monolayers were washed in 0.1 M sodium cacodylate buffer (Electron Microscopy Sciences) and fixed in 2.5% glutaraldehyde (Electron Microscopy Sciences) in sodium cacodylate buffer overnight at 4°C. The following day the cells were washed in 0.1 M sodium cacodylate buffer and incubated in 1% osmium tetroxide (Mecalab) for 1 h at 4°C. The cells were dehydrated in a graded series of ethanol/deionized water solutions from 50%–100%. The cells were then infiltrated with a 1∶1 and 3∶1 Epon 812 (Mecalab)∶ethanol mixture, each for 30 min, followed by 100% Epon 812 for 1 h for embedding in the wells, and polymerized overnight in an oven at 60°C. The polymerized blocks were trimmed and 100 nm ultrathin sections cut with an UltraCut E ultramicrotome (Reichert Jung) and transferred onto 200-mesh copper grids (Electron Microscopy Sciences) having formvar support film. Sections were poststained for 8 min with 4% uranyl acetate (Electron Microscopy Sciences) and 5 min with lead citrate (Fisher Scientific). Samples were imaged with a FEI Tecnai 12 transmission electron microscope (FEI Company) operating at an accelerating voltage of 120 kV and equipped with a Gatan 792 Bioscan 1k61k Wide Angle Multiscan CCD Camera (Gatan, Inc.). Vesicle and pit size was measured using Image J (National Institutes of Health, Bethesda, MD).
PHear GST fusion protein and Ex His-tag fusion protein were expressed in the E. coli strain BL21. To generate uniformly 15N-labeled protein, the cells were grown in M9 minimal media containing 15NH4Cl. Bacteria were induced at 30°C for 4 h using IPTG at a final concentration of 1 mM once OD at 600 nm reached 0.8. The GST-fusion protein was purified, cleaved with thrombin in PBS, and thrombin was removed using benzamidine-Sepharose. The protein samples were further purified by S75 gel filtration equilibrated with PBS. The Ex His-tag fusion protein was purified by Ni-charged chelating Sepharose in 8 M guanidium chloride in PBS and eluted in 2 M guanidinium chloride in PBS with 500 mM imidazol. The sample was then purified by gel filtration using a S75 column equilibrated with PBS, immediately cleaved by TEV protease, then purified by reverse phase HPLC using a C18 column. The purified protein was lyophilized to afford a white powder. The NMR samples contain freshly prepared buffer with 25 mM sodium phosphate pH 7.2, 75 mM NaCl, 0.5 mM EDTA, and 3 mM DTT. All NMR experiments were performed at 30°C using a Bruker DRX 600-MHz spectrometer. Spectra were processed by NMRPipe [57] and analyzed by NMRview [58]. NMR titrations were carried out by acquiring 1H-15N heteronuclear single quantum correlation (HSQC) spectra on 250 µL of 15N-labeled protein at a concentration of 0.1–0.2 mM. Subsequent spectra were taken after the addition of an unlabeled ligand. Analysis of peptide binding to the PHear domain was carried out by comparison of chemical shifts for backbone amide signals in 15N–1H HSQC spectra. Weighted average shifts ((Δδ15N)2+(Δδ1H)2)0.5 were used to identify binding site residues. The NMR assignments for the NECAP 1 PHear domain were previously determined [16]. Peptides for AP-180 (Ac-VDIFGDAFAAS) and the β2-linker (Ac-SQGDLLGDLLNLDLPPVN) were purified by reverse phase C18 HPLC, lyophilized, and dissolved in NMR buffer to make peptide concentrations of 4.1 mg/mL and 5.1 mg/mL, respectively. 15N-labeled PHear at a concentration of 0.15 mM was titrated with unlabeled peptides at relative concentrations of 2∶1, 1∶1, 1∶2, 1∶4, 1∶8, and 1∶16. To observe that the PHear domain competes with clathrin for access to the β2-linker, 15N-labeled PHear together with unlabeled β2-linker peptide were titrated with unlabeled clathrin terminal domain (aa 1–579). The relative concentrations of PHear, β2-linker, and clathrin terminal domain were 1∶0∶0, 1∶4∶0, 2∶8∶1, 1∶4∶1, and 1∶4∶2, respectively.
CCVs were purified from adult rat brain using buffer A (100 mM MES, pH 6.5, containing 1 mM EGTA, and 5 mM MgCl2) as described previously [59]. For the extraction of coat proteins, CCVs were centrifuged at 200,000× g for 15 min, the pellet was resuspended in 0.5 M Tris pH 7.0, 2 mM EDTA, and incubated for 30 min on ice. The samples were centrifuged at 200,000× g and the supernatant fraction was resolved by SDS-PAGE, transferred to nitrocellulose and processed for Western blotting or overlay assays.
For affinity selection assays, soluble cell and rat brain extracts for affinity selection assays were prepared in 10 mM HEPES, pH 7.4, 1% Triton X-100, 50 mM NaCl, 0.83 mM benzamidine, 0.23 mM phenylmethylsulphonyl fluoride, 0.5 µg/ml aprotinin, and 0.5 µg/ml leupeptin and incubated for 1 h at 4°C with GST fusion proteins precoupled to glutathione-Sepharose. For AP-2 co-immunoprecipitation assays from 293-T cell lysates transfected with Flag-tagged NECAP 1 variants, the NaCl concentration in the buffer was reduced to 33 mM and lysates were incubated with protein G-agarose alone (mock) or with protein G-agarose and 10 µg Flag-antibody (M2) for 1 h at 4°C. Co-immunoprecipitation studies of endogenous AP-2 from brain cytosol and solubilized membrane fraction were performed as described previously [19]. Proteins were resolved by SDS-PAGE and analyzed by Western blotting. Overlays (far-Western blots) were performed as described previously [60].
CCVs were purified from rabbit brain (Pel-Freez, 250 g) and coat proteins were prepared for gel filtration as previously described [61]. Coat proteins were separated on a Hiprep 26/60 Sephacryl S-300 HR size exclusion column and eluted at 0.5 ml/min. Fractions of 1 ml were collected, AP-2-containing fractions identified by Western blot, and stored at −80°C until use. For competition assays, one 1 ml fraction was precipitated with ammonium sulfate, spun for 30 min at 10,000× g, and the resulting pellet dissolved in binding buffer (10 mM HEPES, pH 7.4, 1% Triton X-100, 50 mM NaCl, 0.83 mM benzamidine, 0.23 mM phenylmethylsulphonyl fluoride, 0.5 µg/ml aprotinin, and 0.5 µg/ml leupeptin). Equal aliquots of the purified AP-2 solution were incubated for 1 h at 4°C with 40 µg of purified GST-clathrin terminal domain or GST alone precoupled to glutathione beads. Unbound proteins were removed by washing three times with 1 ml of binding buffer. Equal fractions of GST-clathrin terminal domain were incubated 1 h at 4°C with either 1 ml of binding buffer alone or with 1 ml of binding buffer supplemented with varying amounts of purified PHear–Ex. Unbound proteins were removed as before and the samples were resolved by SDS-PAGE, and AP-2 binding to the clathrin terminal domain was analyzed by Western blot.
|
10.1371/journal.pgen.1006001 | Cooperative Wnt-Nodal Signals Regulate the Patterning of Anterior Neuroectoderm | When early canonical Wnt is experimentally inhibited, sea urchin embryos embody the concept of a Default Model in vivo because most of the ectodermal cell fates are specified as anterior neuroectoderm. Using this model, we describe here how the combination of orthogonally functioning anteroposterior Wnt and dorsoventral Nodal signals and their targeting transcription factors, FoxQ2 and Homeobrain, regulates the precise patterning of normal neuroectoderm, of which serotonergic neurons are differentiated only at the dorsal/lateral edge. Loss-of-function experiments revealed that ventral Nodal is required for suppressing the serotonergic neural fate in the ventral side of the neuroectoderm through the maintenance of foxQ2 and the repression of homeobrain expression. In addition, non-canonical Wnt suppressed homeobrain in the anterior end of the neuroectoderm, where serotonergic neurons are not differentiated. Canonical Wnt, however, suppresses foxQ2 to promote neural differentiation. Therefore, the three-dimensionally complex patterning of the neuroectoderm is created by cooperative signals, which are essential for the formation of primary and secondary body axes during embryogenesis.
| The sea urchin embryo is similar to vertebrate embryos in that the default cell fate is potentially neurogenic, and normal development restricts the neural fate to the narrow area that locates at the anterior/dorsal region of the embryo. Because maintaining the default neural fate to the anterior/dorsal region is required for embryos to precisely integrate information from both the primary anterior-posterior and secondary dorsal-ventral body axes, these axes must be mutually linked by some mechanisms. In this study, we describe how the combination of orthogonally functioning signaling pathways regulates their targeting transcription factors expressing at the anterior neuroectoderm to restrict and pattern the default neurogenic region. By loss-of-function experiments using sea urchin embryos, we revealed that canonical and non-canonical Wnt pathways regulate the anterior neuroectoderm patterning along the primary axis, and TGF-ß signals control the patterning of the neuroectoderm along the secondary axis. In addition, we showed that the crosstalk between the Wnt and TGF-ß pathways was of importance in regulating the neuroectoderm patterning. As the default cell fate in some deuterostome embryos, including embryonic stem cells, is neurogenic, our findings could be widespread mechanisms to coordinate the remaining and/or suppressing developmental programs along different embryonic axes because Nodal and Wnt signals are critical in establishing early developmental polarities in many embryos.
| Embryonic cells of some animals tend to be differentiated into neuroectoderm cells/neural progenitors unless they receive an extrinsic signal, so-called default model [1,2]. This characteristic is also applicable to mammalian embryonic stem cells and induced pluripotent cells (e.g., [3,4]). Therefore, normal development in such organisms can be rephrased as molecular mechanisms that repress the initial neuroectodermal fate and drive them to be differentiated into different cell types. Transforming growth factor-ß (TGF-ß) family members are one group of well-described signaling molecules that play essential roles in determining non-neuroectodermal cell fates. Among these, Chordin and Noggin, which were initially reported as neural inducers, function in protecting the initial neuroectodermal fate at the dorsal side in vertebrates from invading bone morphogenetic protein (BMP) signals that are expressed at the ventral side and that specify a non-neuroectodermal fate [2]. Wnts, another type of secreted signaling molecule, also have functions in repressing the initial anterior neuroectodermal fate. In vertebrates, posteriorly functioning Wnt inhibits anterior neuroectoderm specification genes, such as otx2, and leads to the specification of posterior neuroectoderm [5]. Together, these secreted signaling molecules that regulate body axis formation act to suppress the initial neuroectodermal fate during early embryogenesis. However, despite a large number of these non-neuroectodermal signals, embryos still maintain the neurogenic region in its proper size and location. In addition, within the remaining initial neurogenic ectoderm, each terminal cell differentiation is precisely controlled to organize the complicated neural network, i.e., the patterning of the neurogenic ectoderm is highly sophisticated in the restricted neuroectoderm of normal embryos.
It has been suggested that the pre-signaling cell fate in sea urchin embryos is also anterior neuroectoderm, called the animal plate (AP). This is shown by an experiment in which the earliest canonical Wnt (cWnt) signal is inhibited by injecting the intracellular domain of cadherin (Δcad) to interfere with the nuclear localization of ß-catenin, resulting in most of the ectoderm of the injected embryos becoming specified as AP and differentiating into serotonergic neurons as well as other types of neurons and non-neural cells (Fig 1A: [6,7]). The expanded AP in the early cWnt-deficient embryos lacks patterning, and the serotonergic neurons are therefore dispersed throughout the AP. In contrast, the restricted AP in normal embryos differentiates into serotonergic neurons only at the dorsal/lateral edge, i.e., there are no serotonergic neurons observed at the ventral edge and central part (anterior end) of the AP (Fig 1A), even though most of the cells in the anteriorly restricted neurogenic region have the potential to be serotonergic neurons [8]. Nodal-BMP2/4, via Smad2/3-1/5/8 signaling along the dorsoventral axis, is one of the signaling networks that regulates the specification of cell fate and patterning in this region (reviewed in [9]), but their target transcription factors remain unclear. In summary, the developmental features of the AP of the sea urchin embryo are the following: 1) the serotonergic neural fate is executed only at the dorsal/lateral edge of the neuroectoderm, 2) the anterior end (i.e., the central part) of the AP does not differentiate serotonergic neurons, and 3) no serotonergic neurons appear at the ventral edge of the AP. Although information regarding the morphological and phenomenological characteristics of the development of the AP in sea urchin embryos has accumulated, the details of the molecular mechanisms that perform the intrinsic system of serotonergic neural fate specification at the dorsal/lateral edge of the neuroectoderm and that suppress the neural fate in other regions must still be defined. Thus, we have focused on the functional regulation between the signaling molecules and the transcription factors that control the patterning of the AP in the sea urchin embryo.
Serotonergic neurons in the sea urchin embryo are differentiated within the AP by nearly bilateral patterning (Fig 1A: [9,10]). Among the transcription factors that are zygotically expressed in the AP, the earliest is foxQ2. Based on its expression pattern and previous experimental data, FoxQ2 is present in all AP cells during early embryogenesis (Fig 1B–1E: [11,12]), and it is essential for the specification of most of the cell types in the AP region, including the serotonergic neurons and apical tuft of Hemicentrotus pulcherrimus (Fig 1G and 1H: [8,13]) and Strongylocentrotus purpuratus [12]. However, FoxQ2 mRNA disappears from the dorsal/lateral edge of the neuroectoderm, where the serotonergic neurons are differentiated (Fig 1F, arrowhead: [13]), and the protein cannot be detected in differentiating serotonergic neurons (Fig 1G–1J). In addition, FoxQ2 plays an essential role in the formation of the apical tuft cilia through the maintenance of the ankAT-1 gene in later stages [14]. Because apical tuft cells are not serotonergic, these data suggest that FoxQ2 is first required for the specification of the most of the AP cells [12] but that the expression is subsequently suppressed in the cells at the dorsal/lateral edge of the AP, in which the serotonergic neural fate is executed. Thus, identifying the regulatory mechanisms of FoxQ2/foxQ2 patterning along the dorsoventral axis must be one of the keys to understanding how the initial neurogenic ectoderm is patterned during sea urchin embryogenesis.
Homeobrain (Hbn: LC064116 for Hemicentrotus pulcherrimus Hbn) is a paired-like homeobox gene that is classified into the homeobrain-like (hbnl) family [15]. The gene expression patterns of hbnl family members have been reported in the fruit fly [16], sandworm [15], sea urchin [17,18] and sea anemone [19], but the hbn gene has not been identified in chordate genomes. The hbn expression pattern was first investigated in Drosophila melanogaster, where it initially appears in the anterior dorsal head primordium, which forms portions of the brain, and then in the ventral nerve cord during later stages. In sandworms (Capitella sp. I), hbn expression was detected in the developing brain, as in fruit flies, and in its larval eyes. In the sea anemone Nematostella vectensis, hbn is expressed throughout the blastoderm except for around the blastopore, and its expression is excluded from the aboral pole, where the apical tuft and the subsequent neurogenic region are formed. The expression pattern of hbn in sea urchin embryos was reported during the genome sequencing of Strongylocentrotus purpuratus [17,18,20,21]. In those studies, hbn was initially expressed in the animal pole region during the early blastula stage, and, at later stages, it appeared outside of and then disappeared from the AP, where foxQ2 was expressed. Despite reports on the existence of the gene in some species, the control of its expression and its molecular function has not been investigated in all animals. Here, we focus on the function of Hbn because it is expressed in the same region as the foxQ2 gene during the early specification stage of the AP. Then, we describe the roles of Hbn in the specification of serotonergic neurons and report that the regulation of hbn and foxQ2 expression by Wnt and TGF-ß signals are essential for the precise patterning of the embryonic AP in the sea urchin H. pulcherrimus.
In adding to FoxQ2, we focused on the function of Hbn, another AP-specific factor. hbn is initially expressed throughout the AP (Fig 2A), as previously described in different species [18]. During subsequent developmental stages, the expression of hbn progressively fades from the ventral half of the AP and appears at the dorsal/lateral ectoderm, adjacent to the AP, in the early gastrula stage (24 h) (Fig 2B and 2C). At the late gastrula stage (30 h), its expression is restricted to the dorsal/lateral ectoderm and it completely disappears from the anterior end of the AP (Fig 2D). In addition to its dorsal/lateral ectodermal expression, hbn appears at the upper lip region in the prism stage (36 h, Fig 2E, arrow), where it remains, at least until the pluteus stage (48 h Fig 2F, arrow). To compare the expression pattern of hbn with that of foxQ2 or tryptophan 5-hydroxylase (tph), a serotonin synthase gene, we employed two-color fluorescence in situ hybridization. foxQ2 and hbn expression nearly overlapped in the AP region of the unhatched blastula (Fig 2G), but hbn gradually faded from the region and a portion of the dorsal/lateral ectoderm that was adjacent to the AP began to express hbn, which resulted in the expression pattern of hbn being ‘shifted’ toward the dorsal side away from the AP (Fig 2H and 2I). By the late gastrula stage, hbn expression had completely disappeared from the foxQ2 area (Fig 2J). Double staining of hbn and tph in the pluteus stage showed that the serotonergic neurons were differentiated at the edge of the AP, which was adjacent to the hbn-expressing region (Fig 2K and 2K’).
Hbn morphants developed into pluteus larvae without detectable defects in morphology or developmental timing, except for a defect in the elongation of the anterolateral arms at 72 h (cf. Fig 2N with 2L) and at 96 h (cf. S1C with S1A Fig). In addition, Hbn morphants have significantly fewer serotonergic neurons than normal embryos while non-serotonergic neurons at the AP and ciliary band are almost normal (Fig 2M, 2O, S1A–S1D and S1G Fig). Because the development of serotonergic neurons is affected by several signals from outside of the AP [9], we employed a Δcad-injected embryo to accentuate Hbn function under conditions where all other known signals were eliminated [22,23]. A Δcad-injected embryo, in which the initially fated AP contains a greatly increased number of serotonergic neurons (Fig 2P), is an appropriate system to analyze the intrinsic function of genes that are expressed within it, as was previously reported [6]. When Hbn was knocked down in Δcad-injected embryos, the development of serotonergic neurons was strongly inhibited, as was observed in normal morphants (Fig 2P and 2Q). These phenotypes were specific because they were also obtained when using a second morpholino (Hbn-MO2) that targeted a non-overlapping sequence in the mRNA (S1E and S1F Fig), and the microinjection of an mRNA encoding Hbn protein partially rescued the morpholino knockdown effect under the early cWnt-deficient condition (S2A–S2J Fig). These results indicate that Hbn is required for the development of serotonergic neurons in the AP.
To identify the step in which Hbn is involved during the development of serotonergic neurons, we examined Hbn morphants for the expression of tph, which is a terminal differentiation marker, and zinc finger homeobox 1 (zfhx1; [8]) and forebrain embryonic zinc finger (fez), which are early neural markers [14]. In Hbn morphants, tph was not expressed in the neuroectoderm (Fig 2R and 2S, arrowheads show tph-cells in the control), which indicated that the transcription of tph required Hbn function. zfhx1 and fez are downstream of FoxQ2 but are independent of each other. Hbn morphants expressed neither of these genes in the AP (Fig 2T–2W, arrowheads show the expression patterns of each gene in the control embryos). As expected, zfhx1 expression in the lateral ganglion was not affected in Hbn morphants (Fig 2U, arrows). In adding to the previous report, which showed that the entire AP region had the potential to produce serotonergic neurons [8], these data indicated that Hbn is required for the specification of serotonergic neurons.
The change in the foxQ2 expression pattern along the dorsoventral axis suggested that its expression may be regulated by or depend upon TGF-ß signals because cell fate specification along the secondary, dorsal-ventral axis of sea urchin embryos was determined by TGF-ß family members such as Nodal and BMP2/4 [23–25]. Therefore, we examined whether the Nodal pathway is involved in the regulation of foxQ2 expression throughout development. In Nodal morphants, the size of the foxQ2 region was smaller than that of control embryos at the hatched blastula stage (18 h) (cf. Fig 3G with 3B, quantification of foxQ2 region was shown in P, Q), but they were invariant in unhatched blastulae (12 h) (Fig 3A, 3F and 3Q). The protein localization of FoxQ2 in hatched blastula also showed the same size as its mRNA pattern, and the immunochemical signal in Nodal morphants was weak (cf. Fig 3H–3J with 3C–3E, between arrowheads), which indicated that Nodal is required for maintaining foxQ2 expression during the blastula stages. In contrast, in Lefty morphants, in which Nodal proteins are located throughout the ectoderm [26], the size of the AP in the hatched blastula stage was significantly wider than that in control embryos (Fig 3L–3O and 3Q), which indicated that misexpressed Nodal interferes with the restriction of the neuroectoderm during blastula stages. The difference in the foxQ2-mRNA positive region in controls, Nodal morphants and Lefty morphants measured with the angle from posterior pole was supported by the data, in which we counted the number of FoxQ2-protein positive cells in 18 h stages (Fig 3R). Based on these data, Nodal maintains the expression of foxQ2 during blastula stages, and this mimics the process that occurs on the ventral side of the AP during normal development.
Next, to investigate what controls hbn expression along the secondary axis, we observed its pattern in the embryos, in which the TGF-ß signals responsible for secondary axis formation were disturbed [23,25]. In Nodal morphants, hbn expression was shifted uniformly to the AP-adjacent region by the early gastrula stage (24 h) (cf. Fig 4H, 4I with 4C, 4D), which suggests that Nodal suppresses the expression of hbn on the ventral side of normal embryos. However, quantitative PCR (qPCR) data indicated that the amount of hbn mRNA was not significantly changed in the morphants or even in Nodal-misexpressed embryos (S3 Fig), suggesting that the function of a strong hbn inducer was missing in both Nodal morphants and misexpressed embryos. As a result of the uniform shifting of hbn, zfhx1-positive cells were distributed around the foxQ2 region in Nodal morphants (Fig 4I and 4J, asterisks), whereas in normal embryos, the precursor cells of serotonergic neurons were present at the dorsal edge of the AP (Fig 4D and 4E). In Lefty or BMP2/4 morphants, in which the dorsal ectoderm is missing and the ventral ectoderm surrounds the AP [24,26], but in which hbn expression at the unhatched blastula stage (12 h) is not changed (cf. Fig 4K, 4P with 4A), the expression patterns became obscure after hatching and never showed clear shifting towards the edge of the AP region, unlike in control or Nodal morphants (Fig 4K–4N and 4P–4S). These results suggest that Lefty and BMP2/4 are required for maintaining the strong expression of hbn after the hatched blastula stage. In fact, because the significant decrease of hbn mRNA was observed only in BMP2/4 morphants by qPCR (S3D Fig), we concluded that BMP2/4 is essential for the maintenance of hbn on the dorsal side. Furthermore, when the clear shifting of hbn expression towards the edge of the AP was almost entirely missing in these morphants, no zfhx1 expression was observed at the AP region (Fig 4N, 4O, 4S and 4T), resulting in the loss of all serotonergic neurons, as previously reported [27]. Taken together, the findings show that Hbn plays a role in an intrinsic system that determines the initial neural fate at the dorsal/lateral edge of the AP and that its expression patterns are highly regulated by TGF-ß signals along the dorsoventral axis.
The next question was what further restricts the foxQ2 region further to the anterior end without Nodal signaling (Fig 3G and 3H). To investigate this question, we focused on the cWnt pathway because the inhibition of the early cWnt pathway interfered not only with AP restriction but also with AP patterning, resulting in the serotonergic neurons being differentiated in a dispersed manner (Fig 1A). Because the effect of an exogenous cadherin fragment (Δcad) on blocking cWnt [22] might be short-lived due to the short half-life of the injected mRNA, we blocked the function of low-density lipoprotein receptor-related protein 6 (LRP6: LC064120 for H. pulcherrimus LRP6), a co-receptor that acts with the Wnt receptor Frizzled (Fzl) to mediate the cWnt pathway. Based on previous reports, LRP6 mRNA is expressed maternally, lasts throughout embryogenesis [28] and is present in all cells during the early stages [29]. The same observations were made in H. pulcherrimus embryos (S4A–S4F Fig). Although the localization of LRP6 mRNA was uniform in embryos, the protein was missing in the ingressed and future mesodermal region at the mesenchyme blastula stage (S4G and S4H Fig). In LRP6 morphants, the protein-detection level was significantly decreased (S4I and S4J Fig; each image was captured in the same microscopic condition), and mesenchyme ingression was normal but no endoderm invagination was observed (Fig 5A and 5B). This result occurred because the mRNA and likely, protein of LRP6 were present maternally, and LRP6-MO could not block the early cWnt, unlike Δcad injection. In LRP6 morphants, the foxQ2 expressing region was significantly wider than that of control embryos at 24 h (Fig 5A–5C), and the dorsal-ventral polarity in the ectoderm was normal based on nodal and hnf6 expression patterns (S4K–S4O Fig; [23]). The morphant retained the apical tuft, which should disappear by 72 h during normal development (Fig 5D and 5E; [13]), and, intriguingly, its essential regulatory gene, foxQ2, was also still detected at 96 h (Fig 5F and 5G). This result indicates that LRP6-mediated cWnt signaling is required for suppressing foxQ2 expression in the AP.
Despite the wider foxQ2 region, LRP6 morphants had no differentiated serotonergic neurons at 48 h (Fig 5H and 5I). This result was likely derived from the maintenance of FoxQ2 in each cell in the expanded AP. This was confirmed in 48 h Δcad-embryos and Δcad-LRP morphants because a number of serotonergic neurons and no FoxQ2 protein were observed in the former, while strong FoxQ2 signal and no serotonergic neurons were observed in the latter (Fig 5J and 5K). Taken together, these results indicate that an LRP6-mediated signal seems to be involved in the disappearance of FoxQ2 from the AP and for the precise differentiation timing of serotonergic neurons.
Among the Frizzled receptors for the Wnt pathway in sea urchin embryos, it was reported that Fzl1/2/7 and Fzl5/8 are expressed in the ectoderm [30], and Fzl5/8 is likely the only Fzl receptor whose function we can analyze during the modification of the restricted AP region because Fzl1/2/7 morphants lose the entire AP from the very beginning of its formation [31]. Thus, we investigated whether Fzl5/8 mediates the LRP6-based cWnt pathway that controls the disappearance of FoxQ2 from AP cells. In Fzl5/8 morphants at 56 h, the expression of foxQ2 was maintained (Fig 5M), whereas control embryos lost the foxQ2 message at this stage (Fig 5L). This result suggested that Fzl5/8 functions in controlling the cWnt at the AP region.
The disappearance of hbn from the anterior end occurred normally in Nodal morphants, and the hbn-expressing region surrounded the central part of the AP (Fig 5N), which suggests that the disappearance of hbn expression is independent of the dorsoventral axis formation by TGF-ß signals (Fig 4). To confirm this finding and to investigate the involvement of the cWnt pathway as an upstream factor of TGF-ß signals during hbn clearance, we employed Δcad-embryos, which lack all known zygotic signals including early cWnt and TGF-ß signals [22,23]. In these embryos, hbn is expressed throughout the entire region during the early stages (Fig 5O and 5P). Then, hbn disappears from the central portion of the AP by 30 h (Fig 5Q; [18]). The hbn-negative region is progressively expanded, and the hbn-expressing region is observed only in the squamous epithelia in the posterior half of 48 h Δcad-embryos (Fig 5R). This result supports previous data from a different species, S. purpuratus [18]. This disappearance pattern of hbn in Δcad-injected embryos was spatially similar to that observed in the anterior end area of normal embryos, which suggests that the spatial control of the disappearance of hbn expression from the anterior end of the AP is independent of cWnt and TGF-ß signals.
Focusing on the hbn expression pattern in Δcad-embryos in detail, we found that the disappearance of the gene from the anterior end was delayed. In normal embryos, hbn began to be diminished from the anterior end of the AP by 18 h (Fig 2H), but it did not disappear until 30 h in Δcad-embryos (Fig 5Q). In addition, because the cWnt pathway likely regulates the disappearance of foxQ2 as mentioned above, we investigated the function of LRP6 on hbn regulation. hbn gene expression remained in the entire anterior half in LRP6 morphants at 24 h, at when the clearance of the gene was quite obvious in the controls (Fig 5S and 5T). However, hbn had disappeared from the AP by 50 h, as in normal embryos (Fig 5U and 5V). This was confirmed by the result from double fluorescent in situ hybridization for foxQ2 (Fig 5W). In addition, the disappearance of hbn from the AP region is independent of the absence of LRP6 function (S4P Fig). Based on these observations, the LRP-mediated cWnt signal is not required for the disappearance of hbn from the AP, but it is required for the control of the timing of its clearance.
To find the ligands for cWnt signaling in suppressing foxQ2, we focused on later-expressed Wnts in this study because the early Wnts that function in endomesoderm formation might have indirect effects on AP regulation. Based on the temporal expression profile previously reported, Wnt3 (LC064118 for H. pulcherrimus Wnt3), Wnt6 (LC0641198 for H. pulcherrimus Wnt6) and Wnt7 (LC064119 for H. pulcherrimus Wnt7) are expressed relatively late [28,32]. In H. pulcherrimus, wnt3 is expressed during the cleavage stage but not after the blastula stage, according to qPCR, whereas wnt6 and wnt7 were expressed after hatching (Fig 6A). Based on perturbation experiments, it is suggested that Wnt7 functions as a ligand for the cWnt pathway, and Wnt6 for non-cWnt pathways, and those reasons will be explained in this section for Wnt7 and in the next section for Wnt6. wnt7 was expressed broadly at 20 h and was abundantly expressed in the AP. The broad expression of wnt7 and its strong expression in the AP were invariable until 30 h (Fig 6B–6E).
In Wnt7 morphants, foxQ2 mRNA was still expressed in the AP region even at 96 h (Fig 6F and 6H) and FoxQ2 protein remained to be detected in AP cell nuclei at the same stage (Fig 6G and 6I). Because FoxQ2 persisted, the differentiation of serotonergic neurons was extremely delayed. The number of serotonergic neurons at 72 h was significantly smaller than that in controls (Fig 6J). Because FoxQ2 persistence and missing serotonergic neurons were similar characteristics observed in LRP6 morphants, these results suggest that Wnt7 functions as a ligand of the cWnt pathway in mediating the differentiation of serotonergic neurons through the suppression of FoxQ2 expression.
Although our data so far have suggested that the cWnt pathway is involved in AP patterning through suppressing foxQ2 expression, nuclear ß-catenin was not observed in the region until at least the 8th cleavage stage [22]. Because the antibody that recognizes nuclear ß-catenin in H. pulcherrimus is not available, we performed a TCF-luciferase reporter system (Top-Flash) assay to measure the level of cWnt signal [33,34]. Δcad-injected embryos have only approximately 30% Top-Flash activity compared to controls at 24 h (Fig 6K). This decreased activity tends to recover as the embryos grow due to the degradation of exogenous Δcad-mRNA and/or protein. However, without LRP6 or Wnt7 functions, Top-Flash activity remains low, and the scores of the activity are significantly lower than those in controls at both 24 h and 42 h (Fig 6K). These data support that in the AP region cWnt functions through the Wnt7-LRP6 pathway in suppressing foxQ2 expression. In contrast, Wnt7 morphants had normal clearance of hbn expression from the anterior end (Fig 6L and 6M), supporting the idea that cWnt is not involved in controlling hbn expression patterns.
Because one of the interesting questions in this study is that of which signaling pathway crosstalks with Nodal signaling during the regulation of AP patterning (Fig 3), a cWnt pathway regulated by Wnt7 might be the candidate. In fact, when the Wnt7 function was blocked, the foxQ2 region was wider than in normal embryos at the blastula stage (S5A, S5B and S5D Fig). The excessive restriction of the foxQ2 region that was observed in Nodal morphants (S5A and S5C Fig) was rescued in Nodal-Wnt7 double morphants (S5A and S5E Fig), suggesting that Wnt7 is the factor that restricts foxQ2 expression to the anterior end and that the Nodal signal inhibits a Wnt7-mediated signaling pathway during the blastula stages.
We next focused on the function of non-cWnt signals on AP patterning. As it is downstream of early cWnt signals from the posterior side, a c-Jun N-terminal kinase (JNK) signal functions in the restriction of the AP to the anterior end [31]. To examine whether a JNK signal also plays a role in AP patterning, we applied a JNK inhibitor from 2–4 cell stages and analyzed the expression patterns of foxQ2 at the desired stages. As was previously reported, the restriction of foxQ2 to the anterior end was inhibited in the absence of JNK function at 24 h (Fig 7A and 7C). foxQ2 disappearance from the AP was delayed in JNK-inhibited embryos, but the remaining signal was weak, and its area was very small at 60 h (Fig 7B and 7D). Unlike the cWnt pathway, the JNK pathway seem to be weakly involved in the maintenance of foxQ2 expression because the timing of the initial differentiation of the serotonergic neurons was slightly delayed (48 h, Fig 7E and 7F), but a number of serotonergic neurons were differentiated in the expanded AP one day later [31]. We next focused on the function of non-cWnt signals on hbn expression, and used a JNK inhibitor and analyzed the expression patterns of hbn. hbn clearance from the anterior end at 24 h was not observed in JNK-inhibited embryos (Fig 7G and 7I, arrowheads). In addition, the clearance was intriguingly not observed in JNK-inhibited embryos, even in the later stages (cf. Fig 7J with 7H; arrowheads). Together, these results suggest that a JNK signal acts as a part of non-cWnt signaling and that it mainly plays a role in the clearance of hbn from the anterior end of the AP.
As mentioned above, Fzl5/8 is likely the only Fzl receptor whose function we can analyze during the modification of the restricted AP region [31]. Here, we investigated whether Fzl5/8 mediates the non-cWnt pathway that controls the clearance of hbn from the central part of the AP. In Fzl5/8 morphants at 56 h, the expression of hbn was maintained (Fig 7L) whereas all control embryos had no hbn mRNA (Fig 7K). This result suggested that Fzl5/8 mediates the non-cWnt signal at the AP region.
To investigate the ligands for non-cWnt signaling in hbn patterning, we focused on Wnt6 because Wnt7 was not involved in the regulation of hbn expression (Fig 6). wnt6 is expressed in the veg2 endoderm region and is not obvious at the ectoderm (Fig 7M–7P). The morphology of Wnt6 morphants did not resemble a normal pluteus stage even at 56 h and had a straight archenteron and no pluteus arms (Fig 7Q and 7S). Focusing on the hbn expression pattern, we found that it did not disappear from the central part of the AP region in Wnt6 morphants (Fig 7S). This result was confirmed by experiments in embryos that were doubly injected with Δcad and Wnt6-MO, in which hbn was broadly expressed in the expanded AP (S6I and S6K Fig). To clarify this finding, we employed Nodal morphants; without Nodal function, hbn “shifting” to the periphery of the AP is more obvious (Fig 7R). The morphants, in which Nodal and Wnt6 were simultaneously knocked down, showed no hbn disappearance from the central part of the AP (Fig 7T). This result clearly indicated that Wnt6 is required for hbn suppression in the AP. Because hbn was maintained in the AP in Wnt6 morphants, serotonergic neurons were differentiated not at the edge of the region but at the anterior end of the AP (cf. Fig 7V, asterisk with U, arrow). These data are supported by the serotonergic neural patterns in later stage, in which the differentiated serotonergic neurons gather at the anterior end in Wnt6 morphants even if there is no Nodal inhibition (Fig 7W and 7X). In addition, the nuclear localization of FoxQ2 had already disappeared at 60 h, as in normal embryos (Fig 7U and 7V), indicating that Wnt6 is not strongly involved in the regulation of the FoxQ2 expression pattern, which is mediated by LRP6/cWnt signaling. Although it is difficult to distinguish, these results suggest that Wnt6 functions as one of the players in the non-cWnt pathways that regulate AP patterning, especially in the control of hbn expression.
Because the expression patterns of foxQ2 and hbn are complementary after the gastrula stage in normal embryos, FoxQ2 is another candidate for suppressing hbn expression. To examine this possibility, we investigated hbn expression patterns in FoxQ2 morphants. The disappearance of hbn occurred normally in the morphants (S7A–S7F Fig), which suggests that FoxQ2 and its downstream genes do not regulate the suppression of hbn expression at the AP in later embryos. Furthermore, in Hbn morphants, the expression pattern of foxQ2 was the same as that in normal embryos (S7G–S7J Fig), indicating that FoxQ2 and Hbn are mutually independent.
Here, we reveal the molecular mechanisms that control the patterning of the anteriorly restricted neurogenic AP ectoderm along the anteroposterior and dorsoventral axes in the sea urchin embryo, i.e., how embryos pattern the initial neuroectoderm to let the specific neurons differentiate only at the correct location (Fig 8). Although a number of genes that are expressed at the neurogenic ectoderm were identified during the genome sequencing project of S. purpuratus [17,18,35], those analyses were not sufficient to explain the molecular pathways that regulate AP formation and neural differentiation. Our data show that the expression of transcription factors inside AP must be precisely controlled by the intrinsic and/or extrinsic TGF-ß and Wnt molecules and that this regulation is essential for the development of AP and neurons. We also reveal the function of Hbn in specifying the initial neuroectodermal fate. To our knowledge, this is the first study revealing the molecular function of Hbn in any animals, although the expression pattern of hbn has been reported in several species. With Hbn function, we must consider the function of FoxQ2 in the specification and differentiation of neurogenic AP. FoxQ2 is initially required for the specification of most of cell types in the AP by the mesenchyme blastula stage [12], but it is not required later for neural differentiation because its function is to maintain apical tuft gene expression [13]. Therefore, the key to understanding the molecular mechanisms that maintain and suppress the initial neuroectodermal fate is the regulation that controls these two transcription factors in sea urchin embryos.
It was previously reported that serotonergic neurons in the neurogenic AP of the sea urchin embryo are formed at the dorsal/lateral edges of the region [9,10] and that the differentiation of serotonergic neurons at the ventral side is suppressed by Nodal, which is expressed in the ventral ectoderm [6]. In this study, we revealed that hbn expression is suppressed by Nodal on the ventral side but maintained by BMP2/4 on the dorsal side. hbn expression is eliminated from the animal pole, likely by the non-cWnt pathway mediated by Wnt6/JNK after the blastula stage, which will be discussed below, and its pattern forms a horseshoe-like shape (Fig 8A). This pattern was not reported in another sea urchin, S. purpuratus [18], but, in H. pulcherrimus, it is clear that the expression is missing at the ventral side of the normal AP. The loss of Nodal function supports this observation because Nodal morphants have a ring-like shape of hbn expression around the neurogenic ectoderm (Fig 4). Because the expression of foxQ2 is also under the control of secondary axis formation by the Nodal and BMP2/4 pathways (this study; [14]), we need to know whether Nodal and/BMP2/4 regulation is direct or indirect by further experiments, including chromatin immunoprecipitation analysis aimed at uncovering the cis-regulatory modules of foxQ2 and hbn.
In vertebrates, TGF-ß signaling also functions in the neural plate patterning along the dorsoventral body axis [36]. For example, bmp2 and bmp7 expressing outside of the neural plate are necessary for the development of noradrenergic neurons through the induction of the homeodomain protein, phox2a, in zebrafish embryos [37]. Nodal, on the other hand, is required for suppressing the precocious acquisition of forebrain characteristics in mouse embryos [38]. Our data indicated that sea urchin embryos use the similar mechanisms to pattern the neuroectoderm, controlling the timing and location of the differentiation of serotonergic neurons. In addition, because there are other types of neurons present on the ventral side of the AP in sea urchin embryos [7], future investigations regarding the relationship between in-cell factors characterizing those neurons and TGF-ß signaling coming from the outside of the neurogenic AP region will lead us to understand the conserved mechanisms of neural patterning throughout the animal kingdom.
Persistent FoxQ2 and apical tufts in the AP region in LRP6 morphants strongly indicate that the cWnt pathway is required to suppress FoxQ2 and exert a neural fate, although it has been reported that early cWnt, visualized with the nuclear localization of ß-catenin, was observed only at the posterior half of the embryo until the gastrula stage [22]. In addition, our results suggest that Wnt7 works as a ligand in the LRP-cWnt pathway in the AP that suppresses FoxQ2 with precise timing (Fig 6). Although we could not rule out the possibility that Wnt7 functions indirectly from outside of the AP of normal embryos, based on its expression patterns, the strong expression of wnt7 at the thickened AP of normal (Fig 6) and Δcad embryos (S6E and S6H Fig) suggested that it plays an intrinsic role in FoxQ2 suppression within the AP. The difference between LRP6 morphants (abundant FoxQ2 and no serotonergic neurons) and Δcad embryos (less FoxQ2 and a number of serotonergic neurons) might be attributed to the lifetime length of exogenous Δcad mRNA and/or protein. Because the injected mRNA can last approximately 24 hr (e.g., S8A and S8B Fig), only early cWnt but not later cWnt is suppressed in Δcad embryos. This idea was well supported by Top-Flash assay (Fig 6). Of course, we cannot completely rule out the possibility that Δcad alone is not sufficient to block all cWnt, even during the early stages. The future TOP-Flash assay during the early stages can answer this question. Because foxQ2 expression is at a gradient from anterior tip to periphery (Fig 1; [11]), the area of the biggest effect of foxQ2 removal by Wnt7/cWnt might be the edge of the AP region, resulting in serotonergic neurons starting the differentiation process at the position. A previous report [31] showed that foxQ2 was expressed at the posterior half, where this gene is never detected by in situ hybridization in normal embryos, when Axin was misexpressed, and they implicated that foxQ2 is originally expressed throughout the embryo and early cWnt at the posterior half suppressed it. Thus, it is expected that the mechanisms suppressing foxQ2 expression were also applied in the AP region after it is restricted to the anterior end.
The results from JNK inhibition led us to consider that non-cWnt is involved in neurogenesis in the AP of the sea urchin embryos (Fig 7). Additionally, data regarding its ligand, Wnt6, supported our JNK result (Fig 7). However, it is still not clear how Wnt6 mediates non-cWnt signaling in the AP region. Because it has been reported that wnt6 is zygotically expressed at the vegetal plate and functions in endomesoderm formation [30,39], it is possible that the signal is indirect. However, our data using Δcad and Wnt6-MO indicated that Wnt6 could function within the AP region even though the expression of mRNA at that location is faint (S6 Fig). More detailed studies of its distribution and function with protein level will be necessary to understand the complete mechanisms of Wnt6 function. In contrast to the Wnt7 data, the normal disappearance of FoxQ2 in Wnt6 morphants indicated that Wnt6 did not function as a ligand of the cWnt pathway in the AP. The non-cWnt includes pathways other than JNK, planar cell polarity (PCP) and Ca2+ pathways [40], suggesting that those pathways are also involved in AP patterning in the sea urchin embryo. In fact, as a ligand for the non-cWnt pathway, Wnt6 might not be sufficient because JNK inhibition led to some foxQ2 remaining in the AP, indicating that the JNK pathway weakly acts in suppressing FoxQ2 and affects the precise control of the timing and the location of serotonergic neurons (Fig 7). It was reported that the JNK pathway functions in restricting the AP region, represented by foxQ2 and hbn expression, to the anterior end during blastula stages [31], but the Wnt6 morphant had no expanded AP and tended to have a more restricted hbn region, supporting the idea that other ligands for non-cWnt signaling function during anterior neuroectoderm formation.
Our results suggest that cWnt and non-cWnt signaling function in repressing FoxQ2 and Hbn, respectively, but we cannot completely rule out the possibility that each pathway affects both types of repression. This is because LRP6 morphants showed slightly delayed hbn clearance and JNK inhibition allowed some foxQ2 to remain in the AP region. Because cWnt and non-cWnt antagonize each other in some biological processes [41], this cross-reaction might be normal in AP formation in sea urchin embryos. We must also consider the functional combination of frizzled receptors, secreted frizzled-related protein, and Dickkopfs (Dkks) to determine the complete involvement of the Wnt pathways in AP patterning. In fact, the restriction of the AP to the anterior end is managed through their combination with other Wnt ligands, such as Wnt1 and Wnt8 [31], and the patterning of anterior structure, including neuroectoderm, is also regulated those factors in other deuterostomes [42–44]. In contrast, both LRP6 and Wnt7 morphants failed to finish the restriction of the AP region by the blastula stages (Figs 5 and 6), similar to Wnt1 and Wnt8 morphants [31], suggesting that those factors affect the early cWnt events that restrict AP to the anterior end. Our data using Fzl5/8 morphants suggested that this frizzled functions as a receptor for both cWnt and non-cWnt signals as was observed in other systems [45]. This relationship strongly supports the idea that cWnt and non-cWnt cross-react with each other through sharing the frizzled receptor during AP patterning, although we cannot rule out that other types of frizzled, which we did not analyze in this study, are more essential for each pathway. Adding to our knowledge of the involvement of these molecules, biochemical analyses to reveal ligand-receptor associations will be conducted in the future to understand the complete pathway regulating AP formation in the sea urchin embryo.
As mentioned in the Introduction, the anterior neural fate in vertebrates is restricted by Wnt signaling from the posterior side [5]. Posterior Wnt signals are also reported in invertebrates, such as sea urchins [31,46,47] and amphioxus [48,49]. These species commonly use Wnt signals to establish posterior identities during early development. In this study, we found that wnt7 is expressed in the AP region and required for the differentiation of serotonergic neurons as a ligand for the cWnt pathway (Fig 6), indicating that the sea urchin embryos utilize cWnt signaling at both the posterior and anterior ends. In addition, anterior and posterior cWnt share a function, repressing foxQ2 expression (in this study, [12,31]). The simple mechanism that cWnt suppresses foxQ2 with shifting the functional timing, early at the posterior and later at the anterior ends, enables embryos to have a complicated body plan along the anterior-posterior body axis: FoxQ2 initially specifies the AP region only at the anterior end and later it disappears from the AP to let the serotonergic neurons differentiate within it. We accidentally found this both-end cWnt signaling because we blocked the early cWnt at the posterior end using exogenous mRNA encoding Δcad, which has a short life-time. If we permanently and completely inhibit some of the components of the cWnt pathway, it might be difficult to recognize the later functioning anterior cWnt. It is possible that this type of anterior cWnt commonly functions in other systems during early development because in vertebrates it was reported that the wnt7 family was expressed in the developing anterior neuroectoderm region [50].
One of the most interesting findings in this study is that the dorsoventral Nodal pathway might interfere with the anteroposterior Wnt pathway during the embryogenesis of the sea urchin. Because a number of studies previously revealed the molecular mechanisms of cell fate specification along these embryonic axes in many species, we have now accumulated the information to determine, for example, how the anteroposterior axis is formed by the bicoid gradient in the fly [51–53] or how left-right asymmetry is created by Nodal flow in mice [54,55]. However, because embryos must control cell fate specification along all three body-axes in our three-dimensional world with precise timing, the formation of the body axes should not be independent of each other. Thus, the information from processes that occur along each axis should be integrated with a high degree of sophistication and affect cell-fate specification during each step of embryogenesis. We have previously reported that a single transcription factor links anteroposterior-dorsoventral axis formation in the sea urchin embryo and that it regulates the timing of the onset of specification of the secondary axis downstream of primary axis formation [12]. By combining our results with previous reports, we propose the following five combinational signaling steps that regulate serotonergic neuron formation at the dorsal/lateral edge of the AP in the sea urchin embryo (Fig 8A and 8B): 1) posterior cWnt/non-cWnt signaling restricts the AP, which is specified by early FoxQ2, to the anterior end [6,31], 2) Wnt7/cWnt suppresses late FoxQ2, which induces the apical tuft cilia and represses neural fate, at the edge of the restricted AP along the anterior-posterior axis, 3) dorsoventral Nodal suppresses Wnt7/cWnt to maintain late FoxQ2 at the ventral side, 4) Wnt6/non-cWnt suppresses the neural specifier Hbn, preventing its expression in the anterior end, and 5) BMP2/4 strongly maintains the expression of neural specifier Hbn at the dorsal side whereas Nodal suppresses it at the ventral side. After the AP is restricted, the regulation of the expression of two opposing functional transcription factors, the neural specifier Hbn and late neural suppressor FoxQ2, is accomplished by the molecular mechanisms of neural patterning in the AP. Crosstalk between Wnts along the primary axis and Nodal along the secondary axis is carried out during the process of suppressing the serotonergic neural fate at the ventral side of the AP. Although suppressing the specifier of neurons, Hbn, at the ventral side seems to be sufficient, maintaining the suppressor, FoxQ2, within the same region is a great supporting system for embryos to ensure the removal of the neural fate.
Our data suggested that the effect of Nodal suppressing cWnt at the AP region seems to reach to the dorsal edge (Fig 3). However, because it is reported that Nodal can diffuse to short range in AP region [56], we do not know how the Nodal pathway controls the sizing of the entire AP through interactions with cWnt signaling in the AP. As wnt7 is expressed abundantly in the AP (Fig 6), the Nodal pathway might affect its expression regulation even though Nodal itself can bind the receptor in a few rows at the ventral edge of the AP. In addition, the downstream factors of Nodal signaling, e.g., Lefty and BMP2/4, can diffuse to the AP region [12,14,23], and they might interact with cWnt pathway directly or indirectly to regulate the size control of the AP.
In S. purpuratus, it was indirectly implicated that Nodal regulates the expression of foxQ2 by controlling transcription factors, Not, and Emx. Their relationships seem to be complicated along the spatiotemporal patterning [57,58], and none of them have yet been analyzed in H. pulcherrimus. However, our data on Nodal loss-of-function and gain-of-function are quite reproducible during those stages (Fig 3), and even in S. purpuratus Nodal morphants had a smaller AP, judging from the distribution of serotonergic neurons [18], supporting the idea that Nodal maintains the foxQ2 expression in the sea urchin embryos. The precise cis-regulatory analysis of the foxQ2 expression pattern in the future will lead us to understand the detailed molecular mechanisms of how TGF-ß signals control AP patterning.
Because we have not yet succeeded in completely dissociating single cell-cultures, we cannot, strictly speaking, conclude that the default fate of sea urchin cells is the neuroectoderm or neurons. However, if the earliest known signal, cWnt, which functions in the posterior half at the beginning of the 8–16 cell-stage [59], is blocked, almost the entire region develops into neurogenic AP [31], suggesting that the initial or pre-signaling fate of sea urchin cells is anterior neuroectoderm. Within this expanded pre-signaling AP, embryos differentiate a number of serotonergic neurons that are scattered throughout the AP, in which other cells are produced with long, immotile, apical tuft cilia [12,13]. Removing Notch signaling from the AP promotes an increased number and the clustering of serotonergic neurons, indicating that lateral inhibition in the AP is another signal that inhibits the serotonergic neural fate in the sea urchin embryo [8]. FoxQ2 is required to exert the serotonergic neural fate initially [12]. However, FoxQ2 is a bifunctional transcription factor that is required early for the specification of most of the cell types in the anterior neuroectoderm, and then it must be removed from the cell late, which subsequently takes on a serotonergic neural fate [8,12–14]. As an initial specifier, the function of FoxQ2 might be similar to that of Zfp521 in mouse embryos. Zfp521 is zygotically expressed by a cell-intrinsic mechanism to exert the initial neural fate [60]. In sea urchins, however, the initial specifier is substituted with a second one, Hbn, because the function of FoxQ2 becomes another one that is against serotonergic neural differentiation after the blastula stage. Taken together, the unknown mechanisms that initially induce foxQ2 and/or hbn at the anterior end of the sea urchin embryo are the substances that determine the initial cell fate, which will be clarified in the future through the analysis of cis-elements of the foxQ2 and hbn genes. So far, none of functional data of Hbn in other systems have been published, but understanding these mechanisms will lead us to answer the question of what is truly the default cell fate in the sea urchin embryo as well as in other organisms.
Embryos of Hemicentrotus pulcherrimus were collected around Shimoda Marine Research Center, University of Tsukuba, and around the Marine and Coastal Research Center, Ochanomizu University. The divergence time between H. pulcherrimus used in this study and S. purpuratus used in most previously described studies was estimated to be 7.2–14 million years ago [61], and the developmental time-course, gene expression patterns, and reported phenotypes in gene-knockdown and/or misexpressed experiments are almost the same. The gametes were collected by the intrablastocoelar injection of 0.5 M KCl, and embryos were cultured in glass beakers or plastic dishes that contained filtered natural seawater (FSW) at 15°C. Cell-permeable JNK inhibitor I, (L)-Form (Merck Millipore, Billerica, MA, USA), was used at 50 μM from the two-cell stage to desired stages [31]. For the control experiment, we added same volume of dimethyl sulfoxide (DMSO), which is used for dissolving the JNK inhibitor.
In whole-mount in situ hybridization, embryos were fixed with 3.7% formaldehyde-sea water (SW) overnight at 4°C. After 7 x 7 min washes in MOPS buffer (0.1 M MOPS, pH 7.0, 0.5 M NaCl, 0.1% Tween-20), MOPS buffer was substituted with hybridization buffer (HB: 70% formamide, 0.1 M MOPS, pH 7.0, 0.5 M NaCl, 0.1% Tween-20, 1% BSA), and specimens were pre-hybridized at 50°C for 1 h. Subsequently, pre-hybridization HB was substituted with fresh HB containing Dig-labeled RNA probes (0.4 ng/μl final concentration), and samples were incubated at 50°C for 5–7 days. After washing in MOPS buffer for 7 min x 7 times at room temperature (RT), for 1 h x 3 times at 50°C, and for 7 min x 2 times at RT, samples were blocked with 1–5% skim milk (Nacalai Tesque, Tokyo, Japan) in MOPS buffer for 1 h at RT and thereafter incubated with anti-Dig antibody conjugated with alkaline phosphatase (Roche, Basel, Switzerland; 1:1,500 dilution) overnight at 4°C. Tissue was washed with MOPS buffer for a half day with several buffer exchanges. Dig signal was detected with NBT/BCIP (Promega, Madison, WI, USA). For two-color fluorescent in situ hybridization, Dig-labeled and FITC-labeled probes were simultaneously applied to HB and detected with anti-Dig and anti-FITC POD-conjugated antibodies, respectively (Roche), followed by the Tyramide signal amplification plus system (TSA-plus; Perkin Elmer, Waltham, MA, USA). After blocking in 1–5% skim milk, specimens were incubated with 1:1,000 diluted anti-Dig POD-conjugated antibody for 1 h at RT, washed with MOPS buffer for 7 min x 7 times at RT, and treated with tetramethylrhodamine TSA-plus for 10 min at RT. Then, samples were washed three times with MOPS buffer, and the remaining POD function was quenched by 0.5% sodium azide in MOPS buffer for 30 min at RT. After washing, we repeated treatment of the samples with anti-FITC antibody and the FITC TSA-plus system. The size of the foxQ2-expressing region is quantified with the angle from the posterior end (Fig 3P). The angle was measured using ImageJ and Student’s t-test was applied to each quantification to judge whether their differences were significantly meaningful. The graph was drawn with software R [62].
In whole-mount immunohistochemistry, embryos were fixed with 3.7% formaldehyde-SW for 10 min at RT. After washing with PBST (137 mM NaCl, 2.7 mM KCl, 10 mM Na2HPO4, 1.76 mM KH2PO4, pH 7.4, 0.1% Tween-20) for 7 min x 7 times, samples were blocked with 1–5% lamb serum in PBST for 1 h at RT and incubated with primary antibodies (dilutions: serotonin (Sigma-Aldrich, St. Louis, MO, USA) 1:2,000, synB [7] 1:100, LRP6 (Sigma) 1:1,000, FoxQ2 [14] 1:100 and c-myc (Sigma) 1:1,000) overnight at 4°C. Antibodies were washed off with PBST for 7 min x 7 times, and the samples were incubated with the secondary antibodies (1:2,000 diluted anti-rabbit IgG conjugated with Alexa 488 and/or 1:2,000 diluted anti-mouse IgG conjugated with Alexa 568 (Thermo Fisher Scientific, Waltham, MA, USA)) for 2 h at RT. The specimens were observed using a Zeiss Axio Imager.Z1 that was equipped with an Apotome system (Zeiss, Oberkochen, Germany) and an Olympus FV10i confocal laser scanning microscope (Olympus, Tokyo, Japan). The optical sections were stacked and analyzed using ImageJ and Adobe Photoshop. Panels and drawings for the figures were made using Microsoft PowerPoint. The number of FoxQ2-positive cells was counted under the fluorescent microscope (IX70, Olympus). Student t-test was applied on each quantification to judge whether their differences were significantly meaningful.
The morpholino (Gene Tools, Philomath, OR, USA) sequences and the in-needle concentration with 24% glycerol were as follows:
Hbn-MO1 (0.7 mM): 5’- AAAATGAACGGAACAAGTCCAGTGT -3’,
Hbn-MO2 (2.0 mM): 5’- TAGGAGAACCAACGACCGCCGTCAT -3’,
Nodal-MO (0.2 mM): 5’- AGATCCGATGAACGATGCATGGTTA -3’,
Lefty-MO (0.4 mM): 5’- AGCACCGAGTGATAATTCCATATTG -3’,
FoxQ2-MO (0.2 mM): 5’- TCATGATGAAATGTTGGAACGAGAG -3’,
BMP2/4-MO (0.4 mM): 5’- GACCCCAATGTGAGGTGGTAACCAT -3’,
LRP6-MO1 (1.9 mM): 5’- GAAAGGTTTCAAGGCAGCCCATTTC -3’,
LRP6-MO2 (1.5 mM): 5’- TGCCGTTGACTAAATATCATCTACA -3’,
Wnt6-MO1 (3.8 mM): 5’- ACGTGTCCACTCCATCTTGTAATAC -3’,
Wnt6-MO2 (1.9 mM): 5’- TCGTCCAGCGATTTAATAAAGAGCT -3’,
Wnt7-MO1 (3.8 mM): 5’- ATAACCACACCAAgTTgggCCgCAT -3’, and
Wnt7-MO2 (1.9 mM): 5’- GCTCAGCGATGCCCGATGGATAAAA -3’.
Two non-overlapping morpholinos that blocked the translation of Hbn, LRP6, Wnt6 and Wnt7 were used to confirm the specificity of their function. For negative control experiments, we injected 24% glycerol into eggs.
mRNAs were synthesized from linearized plasmids using the mMessage mMachine kit (Thermo Fisher Scientific) and injected at the indicated concentrations in 24% glycerol in needles: hbn-mRNA (0.1 μg/μl), Δ-cadherin (0.3–0.6 μg/μl; [22]), and myc-mRNA (0.1 μg/μl). Microinjections into fertilized eggs and into one blastomere at the two-cell stage were performed as previously described [13].
Quantitative PCR (qPCR) was performed as previously described [13,63] with some modifications. The total RNA from 100 embryos of H. pulcherrimus was isolated, and reverse transcription was performed using the Realtime Ready Cell Lysis kit and Transcriptor Universal cDNA Master (Roche). GoTaq qPCR Master Mix (Promega) was used for PCR carried out with a Thermal Cycler Dice Real Time system (Takara, Shiga, Japan). Primer pairs used for qPCR were the following:
Wnt3-qF1; 5’- TATATCCGGCAAACAGGTCC -3’,
Wnt3-qR1; 5’- TCTTCTCCCTCGGAACTGAA -3’,
Wnt6-qF1; 5’- GACCTGCTGGAAGAAAATGC -3’,
Wnt6-qR1; 5’- GGGCTGTTTGACCGTATCAT -3’,
Wnt7-qF1; 5’- CATGGTGTTTCAGGTTCGTG -3’,
Wnt7-qR1; 5’- TCCTAGTTCGTTTGGCCTTG -3’,
COI-qF1; 5’- CCGCATTCTTGCTCCTTCTT -3’, and
COI-qR1; 5’- TGCTGGGTCGAAGAAAGTTG -3’.
The relative concentrations of each mRNA were normalized with mitochondrial COI Ct values.
Top-Flash plasmid M50 Super 8xTOPFlash (Addgene plasmid # 12456) and M51 Super 8xFOPFlash (TOPFlash mutant) (Addgene plasmid # 12457) were gifts from Dr. Randall Moon. DNA fragments containing TCF/LEF-binding sites with Firefly Luciferase gene were amplified by KOD-Fx DNA polymerase (TOYOBO, Tokyo, Japan) with RVprimer3 and EBV_rev_primer set and injected at 20 ng/μl in a needle into the fertilized eggs with a carrier EcoRV-digested H. pulcherrimus genomic DNA at 10 ng/μl. The signal was obtained from 20–40 embryos for each experiment (three independent batches) using the Bright-Glo Luciferase Assay System (Promega). The luminescence was detected with the LB941 Multimode Reader TriStar (Berthold Technologies GmbH & Co.KG, Bad Wildbad, Germany) for 60 sec. The Top-Flash signal was normalized to the Fop-Flash level for each experiment.
|
10.1371/journal.pgen.1007865 | LACK OF SYMBIONT ACCOMMODATION controls intracellular symbiont accommodation in root nodule and arbuscular mycorrhizal symbiosis in Lotus japonicus | Nitrogen-fixing rhizobia and arbuscular mycorrhizal fungi (AMF) form symbioses with plant roots and these are established by precise regulation of symbiont accommodation within host plant cells. In model legumes such as Lotus japonicus and Medicago truncatula, rhizobia enter into roots through an intracellular invasion system that depends on the formation of a root-hair infection thread (IT). While IT-mediated intracellular rhizobia invasion is thought to be the most evolutionarily derived invasion system, some studies have indicated that a basal intercellular invasion system can replace it when some nodulation-related factors are genetically modified. In addition, intracellular rhizobia accommodation is suggested to have a similar mechanism as AMF accommodation. Nevertheless, our understanding of the underlying genetic mechanisms is incomplete. Here we identify a L. japonicus nodulation-deficient mutant, with a mutation in the LACK OF SYMBIONT ACCOMMODATION (LAN) gene, in which root-hair IT formation is strongly reduced, but intercellular rhizobial invasion eventually results in functional nodule formation. LjLAN encodes a protein that is homologous to Arabidopsis MEDIATOR 2/29/32 possibly acting as a subunit of a Mediator complex, a multiprotein complex required for gene transcription. We also show that LjLAN acts in parallel with a signaling pathway including LjCYCLOPS. In addition, the lan mutation drastically reduces the colonization levels of AMF. Taken together, our data provide a new factor that has a common role in symbiont accommodation process during root nodule and AM symbiosis.
| Symbiosis between plants and beneficial microbes such as nitrogen-fixing bacteria and arbuscular mycorrhizal fungi has enabled plant colonization of new environments. Root nodule symbiosis with nitrogen-fixing rhizobia enables sessile plants to survive in a nitrogen-deficient environment. To establish the symbiosis, host plant cells need to accommodate rhizobia during nodule development, a process mediated by a plant-derived intracellular structure called the infection thread (IT). In this study, we show that LACK OF SYMBIONT ACCOMMODATION (LAN) is involved in intracellular rhizobia accommodation in the model leguminous plant Lotus japonicus. LjLAN encodes a putative subunit of Mediator complex, a multiprotein complex that has a fundamental role as an activator of gene transcription. Mutation analysis suggests that LjLAN is required for root hair IT formation, which enables swift and efficient rhizobial accommodation. Moreover, we show that LjLAN is required for symbiosis with arbuscular mycorrhizal fungi. These data add a new component to the molecular mechanism relevant to the establishment of root nodule and arbuscular mycorrhizal symbiosis.
| Legumes can establish a symbiotic association with nitrogen-fixing bacteria through the formation of symbiotic root nodules. Nodulation is initiated by the rhizobia-derived lipo-chitooligosaccharidic nodulation (Nod) factors that trigger transient increases in calcium influx levels accompanied with calcium oscillation in the rhizobia-attached root hair cells, initiating dedifferentiation of the underlying cortical cells [1–3]. Studies using two model legumes, Lotus japonicus and Medicago truncatula, have revealed the basically conserved molecular mechanism that results in the progress of Nod factor signaling. In L. japonicus, Nod factor is recognized by two LysM receptor-like kinases NOD FACTOR RECEPTOR 1 (LjNFR1) and LjNFR5 [4–6], which induce a downstream signaling cascade. The Nod factor signaling pathway includes SYMBIOSIS RECEPTOR-LIKE KINASE (LjSYMRK), nucleoporins and cation channel proteins [7–11]. While loss-of-function mutations in components of the signaling pathway confer a complete nodulation deficiency phenotype, recent studies show that constitutive activation of either of LjNFR1, LjNFR5 or LjSYMRK can induce spontaneous nodule formation in the absence of rhizobia [12, 13]. This indicates that at least these three kinases each possess a necessary and sufficient role for nodulation. Following the calcium oscillation, the L. japonicus CALCIUM CALMODULIN-DEPENDENT PROTEIN KINASE (LjCCaMK)/M. truncatula DOES NOT MAKE INFECTIONS 3 (MtDMI3) phosphorylates the transcription factor (TF) LjCYCLOPS/M. truncatula INTERACTING PROTEIN OF DMI3 (MtIPD3) [14–18]. Phosphorylated LjCYCLOPS then induces the L. japonicus RWP-RK type TF, NODULE INCEPTION (LjNIN), by directly binding to its promoter region [14, 19]. A number of nodulation-related genes now have been identified as direct targets of Lj/MtNIN, including genes encoding the NUCLEAR FACTOR (NF)-Y subunits [20]. Root cortical proliferation is induced by constitutive expression of either of phosphorylated LjCYCLOPS, LjNIN or LjNF-Y subunits in the absence of rhizobia [14, 20, 21], indicating that induction of the LjCYCLOPS>LjNIN>LjNF-Y hierarchical transcription cascade is sufficient to initiate nodulation. Several data indicate that cytokinin signaling is another essential regulator of nodulation, and that Lj/MtNIN is a downstream component of the cytokinin signaling pathway, as indicated by findings that functional cytokinin receptor is required for rhizobia- and cytokinin-dependent Lj/MtNIN induction [22, 23]. It was recently shown that MtNIN directly binds to the promoter region of the CYTOKININ RESPONSE 1 (MtCRE1) gene encoding a cytokinin receptor and promotes its expression at root cortex [21, 24]. This result indicates that there is a positive feedback loop between MtNIN and cytokinin signaling. In addition to its activating role in nodulation, in some contexts LjNIN can negatively regulate nodule organogenesis through direct activation of CLE-ROOT SIGNAL 1 (LjCLE-RS1) and -RS2 that function as putative root-derived signals in long-distance inhibitory signaling of nodulation [23].
Accommodation of rhizobia within host cells is indispensable for the establishment of root nodule symbiosis; therefore, proliferating cortical cells need to be invaded by rhizobia at the appropriate time during nodulation. In L. japonicus and M. truncatula, the rhizobial invasion process starts from the tip of the root hair associated with root hair curing. Rhizobia invade proliferating cortical cells through a plant-derived intracellular tube-like structure called the infection thread (IT), and are finally released into host cells by endocytosis [25–27]. The signaling cascade initiating nodule organogenesis is also essential for the rhizobial invasion process, because in most cases root-hair IT formation is severely retarded if key proteins in the signaling pathway are mutated. A recent study demonstrated that, in addition to Nod factor, rhizobia-derived exopolysaccharides have a crucial role in the rhizobial accommodation process via interactions with the EXOPOLYSACCHARIDE RECEPTOR 3 (LjEPR3), a LysM receptor-like kinase that is paralogous to LjNFR1 [28]. The Nod factor signaling seems to have a role to induce the LjEPR3 expression at the epidermis. Overall, one signaling pathway achieves two qualitatively and spatially different phenomena, that is, rhizobial root hair accommodation at the epidermis and nodule organogenesis at the cortex. Studies using an epidermal-specific expression system indicated that this can be explained by a difference in the tissue-specific requirements of the genes involved in Nod factor signaling [29, 30]. In addition, cell-to-cell communication between the epidermis and cortex may be involved [31]. In terms of transcriptional regulation, a direct target of LjNIN, NODULATION PECTATE LYASE (LjNPL) has been implicated in the degradation of plant cell walls, and is required for normal root-hair IT formation [32]. Thus, LjNIN may participate in rhizobia accommodation through activation of genes relevant to root-hair IT formation, such as LjNPL. MtNF-Y subunits seem to be involved in rhizobial accommodation processes as well as nodule organogenesis. LjNIN can also directly induce LjEPR3 expression; the LjNIN>LjEPR3 cascade appears to control rhizobia infection process [33]. In particular, ETHYLENE RESPONSIVE FACTOR REQUIRED FOR NODULATION 1 (MtERN1) that encodes a TF involved in root-hair IT formation together with its close homologue MtERN2, was shown to be a direct target of MtNF-Y subunits [34–36]. Moreover, LjCYCLOPS has a role directly inducing LjERN1 expression [37]. In addition, recent studies show that epidermal cytokinin signaling appears to have a negative role in root hair IT formation [38–40]. Despite these advances in our understanding of the molecular mechanism of nodule organogenesis and the rhizobia accommodation process, our understanding of the mechanism remains incomplete, indicating that further components await discovery.
Symbiosis between plants and arbuscular mycorrhizal fungi (AMF) is another widely observed plant-microbe mutual relationship known as AM symbiosis. The plant regulatory pathway for AM symbiosis has been shown to share some components, called common symbiosis pathway (CSP) genes, of its genetic pathway with root nodule symbiosis [3, 41]. Based on current data, the role of CSP genes is thought to mostly relate to generate calcium signaling and make a read-out, which occurs commonly during the two symbioses. Both symbioses are strongly impaired by a mutation in the CSP genes such as LjSYMRK, LjCCaMK and LjCYCLOPS. In AM symbiosis the LjCCaMK-LjCYCLOPS module responds to calcium oscillation, transmitting a signal to the downstream pathway, that results in the formation of symbiotic organs such as the arbuscule. LjCYCLOPS/MtIPD3 physically interacts with Lj/MtDELLA to form the LjCCaMK/MtDMI3-LjCYCLOPS/MtIPD3-Lj/MtDELLA complex that directly induces the REDUCED ARBUSCULAR MYCORRHIZA 1 (Lj/MtRAM1) GRAS-type TF during AM symbiosis, which is required for arbuscule branching [42–44].
In the present study, we identify a L. japonicus mutant with a mutation in the LACK OF SYMBIONT ACCOMMODATION (LjLAN) gene. Observations of rhizobia infection/invasion patterns together with nodulation foci show that in lan mutant a developmental program of nodulation proceeds in the absence of root-hair IT formation, where rhizobia enter into roots through an intercellular invasion system. The LjLAN gene encodes a protein that is putatively orthologous to Arabidopsis MEDIATOR 2/29/32 (AtMED2/29/32) constituting a Mediator complex. Moreover, the lan mutation reduces symbiosis with AMF. These data suggests LjLAN acts as a putative transcriptional regulatory module required for the establishment of both root nodule and AM symbiosis.
To better understand the molecular mechanisms associated with the control of nodulation, we undertook a screen for nodulation-deficient mutants from EMS-treated L. japonicus wild-type (WT) MG-20 plants. From this screen we isolated a mutant with a mutation in the gene that we named lack of symbiont accommodation (lan) based on the nodulation-deficient phenotype. F1 plants derived from a cross between lan and the WT MG-20 parental line showed normal nodulation. In the F2 population, normal-nodulation and nodulation-deficient plants segregated in an approximately 3:1 ratio (58 normal-nodulation and 18 nodulation-deficient plants). Thus, the lan mutation is inherited as a recessive trait. In L. japonicus, mature nodules can be characterized by several morphological and physiological indicators, including nodule size, color, lenticel formation, and nitrogen fixation activity. In WT plants, formation of mature nodules was recognizable at the latest 14 days after inoculation of Mesorhizobium loti (dai) (Fig 1A, 1C, 1E and 1F). In contrast, in the lan mutant, no mature nodules were formed at the corresponding stage (Fig 1B and 1E). Formation of mature nodules could be observed at 21 dai, and their number gradually increased over time (Fig 1D and 1E), although the number was consistently lower than WT. Analysis of acetylene reductase activity per plant showed that nodules formed on the mutant roots at a later stage, such as 35 dai, were comparable to those of WT (Fig 1F). Therefore, in terms of nitrogen fixation activity, the mutant nodules formed at the stage appeared to be functional.
In order to characterize the effect of the lan mutation on root-hair IT formation and early nodulation, we used two fluorescent-based markers to visualize infection and nodulation foci. A M. loti strain expressing DsRED was used to mark root-hair ITs. During nodule development, a preferential auxin response is observed in proliferating cortical cells and bulge of nodule primordia [45–47]. Thus, we tried to quantify the sites of nodulation foci (cortical cells proliferation and nodule primordia) based on the expression of a reporter gene under the control of auxin responsive element DR5. To visualize the nodulation foci in lan mutant, we produced DR5:GFP-NLS/lan plants by crossing DR5:GFP-NLS/WT transgenic plants [45] with the lan plants. In DR5:GFP-NLS/WT plants, the formation of root-hair ITs was recognizable at 4 dai, and cortical cells located under some of the ITs started to proliferate (Fig 2A–2D, 2M–2P, 2U and 2V). In contrast, root-hair ITs were barely observed in the DR5:GFP-NLS/lan plants during the corresponding time scale (Fig 2E–2H and 2U). In the DR5:GFP-NLS/lan plants, although root-hair ITs were almost undetectable at all time points tested, we found some sites of auxin response, which implied cortical cell proliferation and the formation of nodulation foci (Fig 2I–2L, 2Q–2T, 2U and 2V). The number of nodulation foci gradually increased over time after inoculation (Fig 2V). In most cases, the occurrence of nodulation foci was accompanied with bright DsRED signals suggesting the accumulation of rhizobia at the surface of developing nodules. These results indicate that in the lan mutant the nodulation developmental program can be initiated in the absence of root-hair IT formation. Some mutants impaired in root-hair ITs formation tend to develop an excess number of small uninfected nodule primordia [48–50]. Even in the later nodulation stage such as 45 dai, the formation of such small uninfected nodule primordia were not observed in the DR5:GFP-NLS/lan plants (S1A Fig). In addition, inoculation of M. loti nodC mutants, which could not synthesize functional Nod factors, did not result in making any nodules in the lan mutant as well as WT (S1B Fig). Thus, the nodulation in the lan mutant depends on Nod factor signaling.
During nodulation, a series of calcium oscillations, defined as calcium spiking, in responsive cells is induced in response to the rhizobia-derived Nod factor [51, 52]. A normal calcium spiking pattern could be observed in the lan root hair cells following application of purified Nod factor (S2 Fig), indicating that in the lan mutant, nodulation signaling upstream of the calcium spiking response is unaffected.
In L. japonicus DR5:GFP-NLS/WT plants, rhizobia use the root-hair IT-mediated intracellular invasion system to enter into roots (S3A and S3C Fig) [53]. In DR5:GFP-NLS/lan plants, despite strongly impaired root-hair ITs formation (Fig 2U), cortical cell proliferation is induced, which results in the formation of nitrogen-fixing nodules (Figs 1E, 1F and 2V), raising the question of how rhizobia enter into roots in the lan mutant. The accumulation of rhizobia on the epidermis of nodule primordia suggested that rhizobia might enter developing nodules through intercellular invasion system as was previously reported in other L. japonicus mutants (Fig 2Q–2T and S3B and S3D Fig) [11, 54]. Thus, in order to clarify rhizobial localization in nodules, we examined sections of nodules. In the mutant nodules, a dense population of rhizobia was observed in some intercellular spaces (Fig 3A–3D). This bacteria localization pattern is reminiscent of that defined as pocket of intercellular bacteria seen in the several L. japonicus mutants, where rhizobia enter nodules predominantly through intercellular invasion system [11, 50, 54–56]. In WT nodules rhizobia enter nodule cells through cortical-ITs (Fig 3C) [56]. On the other hand, we could not determine the presence of cortical-ITs in the mutant nodules. An observation of lan mutant nodule sections of relatively later stage showed that the number of rhizobia-colonized cells were evidently reduced compared with WT (Fig 3E and 3F). In WT nodules, rhizobia-colonized cells were tightly packed at the inner region of nodules (Fig 3E). On the other hands, in the lan mutant, clusters of uninfected cells were located between rhizobia-colonized cells (Fig 3F). Thus, the lan mutation can affect rhizobia accommodation process throughout nodule development.
To understand the molecular function of LjLAN, we first sought to isolate the gene by a positional cloning approach. This mapped the LjLAN locus to a region between the simple sequence repeat (SSR) markers TM0216 and TM0135 on chromosome 3 (S4 Fig). Subsequent genome-resequencing of the lan mutant identified an A-to-T nucleotide substitution that occurs in the acceptor site of an intron located upstream of the gene, chr3.CM0112.280.r2.d (S5 Fig). In the mutant, the nucleotide substitution causes the production of two transcripts smaller than that of WT (Figs 4A and S5). We sequenced each mutant transcript, and found that in both cases intron mis-splicing spliced out a DNA region encompassing the original initiation codon of the gene. In addition, in the lan mutant no coding sequence was predictable in the locus. Thus, it is reasonable to suppose that the lan mutation causes a complete loss of function of the gene. The mutant two transcripts were detectable all time points tested after inoculation (S6A Fig). The lan mutation reduced the expression of the gene (S6B Fig). To verify if this gene is responsible for the lan mutation, a 5.8-kb genomic fragment containing the WT gene was introduced into the mutant by Agrobacterium rhizogenes-mediated hairy root transformation. The introduction of the fragment into the mutant rescued the phenotype, resulting in the formation of normal number of nodules at 14 dai (Fig 4B–4F), and normal root-hair ITs formation (Fig 4C–4F). The LjLAN gene encodes an uncharacterized protein of 145 amino acids that is putatively orthologous to AtMED2/29/32, a putative subunit of the Mediator complex (S7 Fig). It is generally thought that the Mediator complex, which consists of a large number of subunits, plays a role as a bridge between promoter-bound TFs and RNA polymerase II to activate gene transcription [57–59]. Indeed AtMED2 was shown to be required for the recruitment of RNA polymerase II [60]. AtMED2 could rescue the lan mutation when it was constitutively expressed by LjUBQ promoter (S8 Fig), suggesting that the LjLAN has a function similar to AtMED2. The phylogenetic analysis identified a homologue of LjLAN in L. japonicus, which was designated as LjLAN LIKE (S7 Fig); the similarity and identity values are respectively 94.6% and 81.7%.
The expression pattern of the LjLAN and LjLAN LIKE gene remained constant in some vegetative and reproductive organs investigated (S9A and S9B Fig). To gain insights into the role of the LjLAN gene during nodulation, we examined the time course expression pattern after inoculation of M. loti. LjLAN expression was largely constant during nodulation when whole roots were assayed by RT-qPCR (S9C Fig). However, an approximately 2-fold induction of LjLAN expression was detected in root segments where proliferating cortical cells were enriched (S9C Fig). Furthermore, reporter gene analysis using ProLjLAN:GUS plus construct showed that during nodulation the GUS activity was detectable at epidermis with curled root hairs, proliferating cortical cells and nodule primordia (S10 Fig). The GUS activity was also observed at lateral roots.
The lan mutant used for above-mentioned analyses has Miyakojima MG-20 genetic background. We obtained a plant with Gifu B-129 genetic background in which a retrotransposon, LOTUS RETROTRANSPOSON 1 (LORE1) [61, 62], was inserted in the middle region of coding sequence of LjLAN gene, causing an occurrence of premature stop codon in the mutant (S11A and S11B Fig). Consequently, we found that the plants have the truncated protein of LjLAN lacking C-terminal part of it (S11B Fig). Unexpectedly, the LORE1-tagged mutant showed normal nodulation phenotypes (S11C and S11D Fig). In order to interpret the observation, we raised two possibilities. First, the effects of lan mutation was observable in an ecotype-specific manner. The second possibility was that the truncated LjLAN that was produced in the LORE1-tagged mutant was functional. To verify them, we introduced modified LjLAN (LjLANΔC), in which amino acid residues constituting C-terminal part of LjLAN were deleted (S11B Fig), into lan mutant. LjLANΔC could rescue the lan mutation to the extent of same level of the introduction of control intact LjLAN (S8 Fig). We then created stable transgenic plants with nucleotide deletions or insertions in the middle region of coding sequence of LjLAN gene by CRISPR-Cas9 genome-editing system. In the transgenic plants, the frame-shifted mutations caused the deletion of amino acid residues constituting C-terminal part of LjLAN (S11B and S11E Fig). The nodulation phenotypes of the transgenic plants were indistinguishable from WT plants (S11F and S11G Fig). These results indicate that C-terminal part of LjLAN is not essential for the LjLAN function. Therefore, the lack of phenotype of the LORE1-tagged mutant can be explained by the retention of LjLAN function rather than an ecotype difference. The LORE1-tagged mutation did not affect the expression of LjLAN (S6B Fig).
After decoding calcium spiking followed by rhizobial infection, LjCYCLOPS has an important role in root nodule symbiosis, as it regulates both rhizobial infection and nodule organogenesis through induction of different downstream target genes [14, 37]. cyclops mutants retain nodulation to some extent [63], providing an accessible baseline for screen for second mutations influencing the cyclops nodulation defects. We then created lan cyclops double mutant. Of note, the lan cyclops double mutant plants showed a complete non-nodulating phenotype, different from each single mutant (Fig 5).
To gain insight into the potential relationship between LjLAN and LjCYCLOPS with respect to gene expression, we investigated the two nodulation-related genes expression, LjNIN and LjNF-YA. LjNIN, a direct target of LjCYCLOPS, has a pivotal role in the transcriptional cascade that is required for both nodule formation and rhizobial infection [19, 20], and LjNF-YA has been shown to be a direct target of LjNIN [20]. Confirming previous reports, we found that expression of LjNIN and LjNF-YA was strongly induced throughout nodulation stages investigated (Fig 6A and 6B) [19, 23, 31]. We found that in the lan and cyclops mutants the induction level of LjNIN was consistently weaker than that in WT along the time course after inoculation (Fig 6A). However, although the lan and cyclops mutation suppressed LjNF-YA induction at 1 and 7 dai, the induction level in lan and cyclops roots at 14 dai was largely comparable to that in WT roots of the corresponding stage (Fig 6B). Furthermore, in the lan cyclops double mutant, the expression of the two genes were strongly impaired at all time point tested as well as ccamk mutant. Together with lan cyclops nodulation phenotype, these results indicate that LjLAN acts in parallel with LjCYCLOPS for the control of key nodulation-related genes expression.
In order to clarify the potential impact of the LjLAN gene on the control of AM symbiosis, the lan mutant were inoculated with Rhizophagus irregularis. The level of AMF colonization of hyphae and arbuscules of the mutant at 21 dai was significantly lower in comparison with that in WT (Fig 7A–7D). The lower level of AMF colonization was maintained even if the plants were grown for a long time such as 28 and 35 dai following inoculation with R. irregularis (Fig 7A and 7B). In the lan mutant, R. irregularis tended to colonize in the lateral roots rather than primary roots (Fig 7E). The introduction of WT LjLAN gene into the mutant by A. rhizogenes-mediated hairy root transformation rescued the phenotype relevant to AM symbiosis (Fig 7F and 7G). In the hairy root system, although the defects in AM symbiosis was rescued compered with empty vector control, the colonization level was lower than normal root system. This may be due to the difference in root system. Overall, these results suggest that LjLAN is required for the establishment of AM symbiosis. LjLAN and LjCYCLOPS appear to have additive role for the control of AM symbiosis, as the double mutation of lan and cyclops had an additive effect on the AM symbiosis (S12 Fig). To gain insight into the phenotype of AM symbiosis from marker genes expression, the expression of LjSbtM1, LjRAM1 and LjPT4 were next investigated. Similar to previous reports [44, 64, 65], the three genes were specifically and strongly activated by AMF infection in WT plants (Fig 8A–8C). In the lan mutant, induction levels of LjSbtM1 and LjRAM1 were weaker than those in WT, but the LjPT4 level was largely unaffected (Fig 8A–8C). AMF colonization was normal in the LORE1-tagged mutant (S13 Fig). Expression of LjLAN itself seemed to be unaffected by AMF infection (S9D Fig).
In addition to the effect on root nodule and AM symbiosis, the LjLAN expression in non-symbiotic organs suggested that the role of LjLAN might not be restricted to the control of plant-microbe symbiosis (S9A Fig). We then examined the effect of the lan mutation on shoot and root growth by growing the plants in the soil that contained enough nutrients in the absence of rhizobia and AMF. The shoot and primary root lengths in the lan mutant was shorter than WT (S14A–S14C Fig). In addition, shoot branching tended to be promoted in the mutant (S14A Fig). These results suggest that LjLAN has a role in the control of overall plant development. The shoot and root phenotypes of LORE1-tagged mutant was indistinguishable from WT plants (S15 Fig).
Mediator is a multiprotein complex that has a fundamental role as an integrator of gene transcription, and governs diverse regulatory processes in plants including development, phytohormone signaling, and responses to biotic and abiotic stress [57–59]. The involvement of Mediator complex in such pleiotropic aspects seems to be achieved by assigning respective Mediator subunits specific functions. In this study, we showed that a nodulation-deficiency phenotype was caused by the mutation of a gene encoding a protein putatively homologous to AtMED2/29/32 subunit of Mediator complex. AtMED2 is required for the recruitment of RNA polymerase II, indicating that AtMED2 has an actual component of the complex [60]. We also demonstrated that AtMED2 could rescue the lan mutation. Thus, the functions of LjLAN and AtMED2 seem to be conserved. To the best of our knowledge, this is the first report describing the identification of a Mediator subunit that is involved in plant-microbe symbiosis. Mediator complex subunits are arranged into four modules; the head, middle and tail modules form the core part of Mediator complex, and the kinase module is separable. AtMED2/29/32 is considered as a tail module-type Mediator subunit. The function of AtMED2/29/32 appears to be pleiotropic and it has a role in abiotic stress signaling related to cold and redox, and phenylpropanoid biosynthesis [60, 66, 67]. Arabidopsis MED25/PHYTOCHROME AND FLOWERING TIME 1 (PFT1), which is a member of the tail module, is one of the best characterized Mediator subunits. AtMED25/PFT1 mediates pleiotropic phenomena, including flower and root development, jasmonate signaling, and salinity and water stress by interacting with key TFs acting in specific regulatory processes [58, 68]. Upon stress or developmental stimuli, plants synthesize jasmonate isoleucine, which enables interaction between AtMED25/PFT1 and AtMYC TFs, achieving transcription of jasmonate-responsive genes [69]. In auxin signaling a compositional change in Mediator complex, that includes AtMED13 and AtMED25, upon auxin stimuli enables Arabidopsis AUXIN RESPONSE FACTOR 7 (AtARF7) and AtARF19 to activate expression of downstream genes [70].
As LjLAN is a putative orthologue of AtMED2/29/32, an expected molecular function of LjLAN may be related to mediate gene transcription through interactions predominantly with TF in response to an environmental cue. Then what kind of TF and environmental cue can be involved in this machinery? To date, studies using L. japonicus and M. truncatula have identified several TFs involved in nodulation, such as LjCYCLOPS/MtIPD3, Lj/MtNIN, Lj/MtNF-Y subunits, Lj/MtNODULATION SIGNALING PATHWAY 1/2 and Lj/MtERN1/2 [3, 37, 71, 72]. However, the largely severe nodulation phenotype of mutants of these TFs, does not resemble the lan nodulation phenotype, although we cannot rue out the possibility that relatively milder lan nodulation phenotype may be explained by partial functional redundancy of other Mediator subunits with LjLAN. The arrested nodulation phenotype of cyclops is partly similar to the lan nodulation phenotype [63], but the analysis of lan cyclops double mutant suggests that LjLAN and LjCYCLOPS act in a parallel rather than in a same genetic pathway. Given the normal calcium spiking in the lan mutant, LjLAN-mediated transcriptional machinery may act downstream of calcium signaling in parallel with CSP pathway including LjCYCLOPS for the control of nodulation-related gene expression (S16 Fig). Thus, the data so far obtained suggest that LjLAN may interact with unidentified TF(s) rather than known ones. However, we cannot rule out the possibility that lan phenotype is due to overall low transcription of key symbiotic genes. An identification of interacting proteins of LjLAN based on the analysis of protein-protein interactions will be undoubtedly needed to verify the possibilities. With respect to the potential environmental cue in this machinery, it seems reasonable to propose that rhizobia infection may be a preferential cue. As the pattern of symbiotic calcium spiking is normal in the lan mutant, a more specific cue may be produced downstream of this signal. In an example of a plant-pathogen interaction, oomycete downy mildew pathogen can attenuate salicylic acid-triggered immunity in Arabidopsis by imposing the interaction between its effector and AtMED19a [73]. Hence, it is possible that a rhizobia-derived factor may directly affect plant Mediator complex to control plant gene transcription relevant to nodulation.
As described above, the Mediator complex is involved in different aspects of plant development and environmental responses. Although in this study we put particular emphasis on the role of LjLAN in plant-microbe symbiosis, it is possible that LjLAN is involved in overall plant development because shoot and root growth were affected by the lan mutation under nutrient sufficient conditions. In L. japonicus stable transformation, we use an A. tumefaciens-medited transformation, where tissue cultures undergo callus formation and shoot regeneration processes. While we were successful in making the transgenic plants with deletion in C-terminal part of LjLAN, we failed to create complete knockout plants of lan by aiming to mutate N-terminal part of LjLAN. In addition, in a stable transformation to complement non-symbiotic phenotype of lan, no regenerated plants were obtained. Therefore, based on these findings, we reason that null mutations of LjLAN are likely to affect callus formation and/or shoot regeneration processes. The non-symbiotic phenotype of lan may provide an intriguing scenario, where a general component of transcriptional machinery had been recruited to the specific functional context during the evolution of plant-microbe symbiosis. To verify this, detailed molecular function and non-symbiotic role of LjLAN need to be elucidated as an important next study.
In L. japonicus WT plants, rhizobia enter into roots through the intracellular invasion system, that principally depends on the formation of root-hair ITs. The Nod factor signaling pathway has a crucial role in this process by regulating root-hair ITs formation. Generally, defects in the signaling pathway cause complete loss of root-hair ITs formation that is accompanied by no nodule formation. While L. japonicus has adopted root-hair ITs-mediated intracellular rhizobia accommodation system, the intercellular invasion system can be used in the case where some nodulation-related factors are mutated [11, 50, 54–56]. For example, in nfr1 nfr5 symrk spontaneous nodule formation 1 (snf1) quadruple mutants, intercellular rhizobial invasion takes place despite apparently no root-hair ITs formation, which leads to the formation of functional nodules [56]. The snf1 plant is a gain-of-function mutant of LjCCaMK, in which spontaneous cortical cell proliferation occurs [16]. This observation indicates that Nod-factor receptors (LjNFR1/5) and LjSYMRK may not be essential to the intercellular invasion process. Furthermore, proliferating cortical cells may need to preexist in order to allow rhizobia to intercellularly enter into roots. In the lan mutant, formation of root-hair ITs is strongly compromised, but it is likely that rhizobia can intercellularly enter into roots, as functional nodules are formed. As we could not determine if the lan mutation affects cortical-ITs formation, it remains unknown how rhizobia are finally released into nodule cells in the mutant. Due to the delay in nitrogen-fixing nodules, the lan mutant exhibit growth defects in a nitrogen-depleted condition until they obtain benefit from symbiotic nitrogen fixation. The delayed nodulation phenotype is thought to be a common feature of some L. japonicus plants, where the intercellular rhizobial invasion is used to accommodate rhizobia in roots [11, 50, 54–56]. Based on the lan phenotype, we propose that the predominant role of LjLAN is to initiate swift and efficient production of nitrogen-fixing nodules by promoting root-hair IT-mediated intracellular rhizobial accommodation. In other words, LjLAN may have a role in preventing protracted and less effective nodulation caused by intercellular rhizobial invasion. Analysis of L. japonicus root hairless mutants indicates that the intercellular invasion system can be adopted in the plants lacking root hairs [54]. It is unlikely that the intercellular invasion phenotype of lan is caused by such physical defects, because root hairs are normally formed in the mutant (S14D and S14E Fig).
It is hypothesized that root-hair IT-mediated intracellular invasion is an evolutionarily advanced invasion system, whereas intercellular rhizobial invasion is the basal pathway [56]. Among the various plants that have an ability to perform root nodule symbiosis, it is estimated that 75% of plants use intracellular invasion and the remaining 25% of plants use the root-hair independent intercellular invasion system [74]. Interestingly, a plant such as Sesbania rostrata has a dual mode invasion system where both a root hair-independent intercellular invasion and a root hair-dependent invasion can be used, depending on whether the soil is flooded or dry [75]. Future detailed analysis of LjLAN may contribute to our understanding of the genetic basis and the evolution and diversity of the rhizobial invasion system.
In addition to root nodule symbiosis, the phenotype of the lan mutant during AM symbiosis suggests that LjLAN is also required for symbiosis with AMF. In contrast to the lan nodulation phenotype in which formation of functional nodules eventually takes place, the lan mutation continues suppressing the establishment of AM symbiosis. As the pattern of symbiotic calcium oscillation was normal in the lan mutant, the lan mutation seems to affect the progression of AM symbiosis downstream of calcium signaling. The expression of the AM-inducible genes, LjSbtM1, LjRAM1 and LjPT4 is generally suppressed by mutation of the CSP genes so far identified [44]. In contrast, while the induction levels of LjSbtM1 and LjRAM1 are reduced by the lan mutation, that of LjPT4 is largely unaffected. Currently, it remains almost completely unknown why these genes show different expression patterns in the lan mutant. However, based on the loss-of-function phenotypes of each gene, LjSbtM1 and LjRAM1 are required for initiation and/or growth of arbuscules [44, 64]. In contrast, a major role of legume PT4 seems to be associated with phosphate transport [76]. It is unclear if PT4 is directly involved in arbuscules development. Such differences in the molecular function of three genes might underlie different gene expression patterns depending on the context; while canonical CSP pathway regulates both arbuscules developmental program and phosphate transport by inducing the three genes, the LjLAN-mediated pathway may only regulate arbuscules developmental program by inducing LjSbtM1 and LjRAM1.
AMF accommodation can employ both the intercellular and intracellular dual invasion system [77]. A specialized structure called the prepenetration apparatus (PPA) mediates the intracellular invasion of AMF and has been suggested to share structural similarities with IT [78–80]. Given that LjLAN has a conserved role between root nodule and AM symbiosis, the predominant role of LjLAN in IT formation indicates that LjLAN also may be involved in intracellular AMF accommodation by mediating PPA formation. Future studies investigating the role of LjLAN during AM symbiosis should place particular emphasis on investigating if PPA formation is involved in AMF accommodation. Because of the lack of evidences, it is currently difficult to integratedly interpret the molecular function of LjLAN in root nodule and AM symbiosis. However, based on lan phenotype, LjLAN-mediated transcriptional regulatory system could be associated with regulation of genes acting symbiont infection processes. As cell cycle activation such as nuclear enlargement and endreduplication commonly occurs during both symbiont infections, the genes involved in this process may be target genes of LjLAN-mediated regulatory system.
The Miyakojima MG-20 and Gifu B-129 ecotype of L. japonicus was used as the WT in this study. The lan mutant was isolated as a result of screen for nodulation-deficient mutants using the M2 generation of WT plants that had been mutagenized with 0.4% ethylmethane sulfonate (EMS) for 6 hours. The LORE1-tagged line of lan (Plant ID: 30008618) was obtained from Lotus Base (https://lotus.au.dk). A description of the DR5:GFP-NLS plants and cyclops-6 has been published previously [45]. ccamk-14 mutant with MG-20 background was newly identified in this study. For the analysis of root nodule symbiosis, plants were grown with or without M. loti MAFF 303099 as previously described [81]. M. loti nodC mutant was obtained from LegumeBase (https://www.legumebase.brc.miyazaki-u.ac.jp/top.jsp). For the analysis of the AM symbiosis, plants were grown with or without R. irregularis (DAOM197198; PremierTech) as previously described [65].
The nitrogenase activity of nodules was indirectly determined by measuring acetylene reductase activity (nmol/ h per plant) as previously described [82].
The leaves of the lan mutant were ground with liquid nitrogen using a mortar and pestle. Genomic DNA was isolated using a DNeasy Plant Mini Kit (Qiagen). The quality of purified genomic DNA was evaluated by a Quant-iT dsDNA BR Assay Kit (Invitrogen). For whole-genome shotgun sequencing of the lan mutant, we performed paired-end sequencing with HiSeq 2000 (Illumina). After fragmentation of the isolated genomic DNA, an Illumina library with a mean insertion length of 350-bp was constructed using TruSeq Nano DNA LT Sample Preparation Kit (Illumina) following the manufacturer’s instructions. These libraries were subsequently sequenced 101 bp from both ends, yielding 6.25 gigabase (Gb) of raw data. After the removal of adaptor sequences and low quality reads (Phred quality score ≥ 20 in < 90% of the bases), 5.97 Gb of high quality sequences remained. The remaining reads were mapped against L. japonicus genome assembly build 2.5 using the Bowtie software [83]. The median value of per-base sequence depth was 18.3 and the genome coverage was 90.2%. The resulting data in the sam format were converted into bam format using Samtools [84]. Genome-wide SNPs were called from the bam files using Samtools and Bedtools [85]. A SNP that is specific to the lan mutant was found by examining the mapped region harboring the LjLAN locus with Integrative Genomics Viewer program (https://www.broadinstitute.org/igv/).
The primers used for PCR are listed in S1 Table. For the complementation analysis, a 5.8-kb genomic DNA fragment including the LjLAN candidate gene was amplified by PCR from WT genomic DNA. This fragment including 4.4 kb of sequence directly upstream of the initiation codon, was cloned into pCAMBIA1300-GFP-LjLTI6b [45]. The coding sequences (cds) of LjLAN and LjLANΔC were, respectively, amplified by PCR from template cDNA prepared from WT L. japonicus. The cds of AtMED2 was amplified by PCR from template cDNA prepared from Arabidopsis Col-0 plants. They were cloned into the pENTR/D-TOPO vector (Invitrogen). The insert was transferred into pUB-GW-GFP [86] by the LR recombination reaction. To obtain the ProLjLAN:GUS plus construct, first an artificially-synthesized GUS plus gene was cloned into pENTR/D-TOPO vector to create the vector pENTR-gus plus. The GUS plus gene in pENTR-gus plus was introduced into a vector pCAMBIA1300-GW-GFP-LjLTI6b [87] by the LR recombination reaction to create the vector pCAMBIA1300-GUS plus-GFP-LjLTI6b. Next, 4.4 kb of sequence directly upstream of the initiation codon of LjLAN was amplified by PCR and cloned upstream of GUS plus gene of pCAMBIA1300-GUS plus-GFP-LjLTI6b to create the vector pCAMBIA1300-pLjLAN-GUS plus-GFP-LjLTI6b. For the analysis of calcium spiking, we used a construct in which nuclear-localized yellow-chameleon (YC2.60) was expressed under the control of the LjUBQ promoter [81]. The recombinant plasmids were introduced into A. rhizogenes strain AR1193 [88] and were transformed into roots of L. japonicus plants by a hairy-root transformation method as previously described [45].
To create CRISPR-Cas9 construct of LjLAN, targeting site in the gene was designed using the CRISPR-P program (http://cbi.hzau.edu.cn/crispr/) [89]. Oligonucleotide pairs (S1 Table) were annealed and cloned into a single guide RNA (sgRNA) cloning vector, pUC19_AtU6oligo, as previously described [90]. Then, the sgRNA expression cassette prepared in pUC19_AtU6oligo was excised and replaced with OsU3:gYSA in pZH_gYSA_FFCas9, an all-in-one binary vector harboring a sgRNA, Cas9, and an HPT expression construct, as previously described [90]. The recombinant plasmid was introduced into A. tumefaciens strain AGL1 and was transformed into WT L. japonicus MG-20 plants by a stable transformation method as previously described [82].
The primers used for PCR are listed in S1 Table. Total RNA was isolated from each plant tissue using the RNeasy Plant Mini Kit (Qiagen) or the PureLink Plant RNA Reagent (Invitrogen). First-strand cDNA was prepared using the ReverTra Ace qPCR RT Master Mix with gDNA Remover (Toyobo). Real-time RT-PCR was performed using a Light Cycler 96 System (Roche) or a 7900HT Real-Time PCR system (Applied Biosystems) with a THUNDERBIRD SYBR qPCR Mix (Toyobo) according to the manufacturer’s protocol. The expression of LjUBQ was used as the reference. Data are shown as mean±SD of 3–4 biological replicates.
Sequence data from this article can be found in the GenBank/EMBL data libraries under the following accession numbers: LjLAN, LC171403; LjLAN LIKE, LC194237. Data of short reads from the lan genomic DNA has been deposited in the DNA Data Bank of Japan Sequence Read Archive under the accession number DRA004948.
|
10.1371/journal.pcbi.1004190 | Metrics for Assessing Cytoskeletal Orientational Correlations and Consistency | In biology, organization at multiple scales potentiates biological function. Current advances in staining and imaging of biological tissues provide a wealth of data, but there are few metrics to quantitatively describe these findings. In particular there is a need for a metric that would characterize the correlation and consistency of orientation of different biological constructs within a tissue. We aimed to create such a metric and to demonstrate its use with images of cardiac tissues. The co-orientational order parameter (COOP) was based on the mathematical framework of a classical parameter, the orientational order parameter (OOP). Theorems were proven to illustrate the properties and boundaries of the COOP, which was then applied to both synthetic and experimental data. We showed the COOP to be useful for quantifying the correlation of orientation of constructs such as actin filaments and sarcomeric Z-lines. As expected, cardiac tissues showed perfect correlation between actin filaments and Z-lines. We also demonstrated the use of COOP to quantify the consistency of construct orientation within cells of the same shape. The COOP provides a quantitative tool to characterize tissues beyond co-localization or single construct orientation distribution. In the future, this new parameter could be used to represent the quantitative changes during maturation of cardiac tissue, pathological malformation, and other processes.
| Biological tissues are highly organized on multiple length scales. In tissue engineering, recreating the in vivo architecture is an important aspect of in vitro experiments. An ability to quantify organization of cellular constructs, specifically the correlation in their orientations, would greatly enhance both our understanding of the function of each construct and provide a tool to evaluate engineered tissues. In this work, we have developed a parameter to evaluate the correlation and consistency of construct orientation. We have extensively characterized this co-orientational order parameter analytically, validated it using synthetic data, and demonstrated its use with experimental data of Z-lines and actin fibrils architectures in engineered cardiac tissues. As long as the orientation angles and location of constructs are known, the parameter can quantify both orientational correlation and consistency. Thus, the co-orientational order parameter has a wide scope of application both in biology and potentially outside of it.
| The architecture and organization of the cytoskeleton components in cells, the cells in tissues, and cellular ensembles in organs affect function at each of these physiological scales [1–4]. The study of architecture is therefore key to understanding how the cellular microenvironment potentiates function, and may provide new insights in the study of physiological mechanisms. Furthermore, for proper function, different components of the cytoskeleton, cell, or tissue need to co-localize and orient properly with respect to each other [5, 6]. Quantifying the degree of orientation of cells and subcellular components, both relative to themselves and to other components, is thus crucial for evaluating the quality of engineered tissues [7].
The problem of describing the organization of biological structures is twofold: first, the orientation of the constructs needs to be quantified from the available images, and second, a metric needs to be applied to summarize the overall organization. The quantification of orientation from biological images is in principle straightforward and can be either done manually [8] or with a variety of computational algorithms [6, 9–11]. As far as the second problem is concerned, summarizing the overall organization after image analysis involves selecting a metric, which is more controversial. As a result, a wide variety of metrics are utilized in the bio-imaging field. For example, some assume that the parameter can be described as the standard deviation of a truncated Gaussian, or normal, distribution [12]. Others use the von Mises circular distribution [9, 10], which is a wrapped normal distribution. However, cellular and cytoskeleton distributions are often non-Gaussian, and their being non-Gaussian may be of crucial importance [13].
An alternative metric, the Orientational Order Parameter (OOP), has been developed in the field of liquid crystals. The OOP is a mathematical construct developed to quantify the degree of order in anisotropic medias [14]. Mathematically, the OOP is equivalent to resultant vector length from the circular distribution with a period of π [15]. In biology, the OOP has been successfully employed to characterize organization of bacteria [16], fibroblasts [17], vascular smooth muscle [18], actin fibrils alignment in valve endothelial cells [19], and Z-lines in cardiac muscle [20]. However, there is a lack of a robust correlation metric that has been characterized for use with biological images. The suite of correlation parameters provided by circular statistics are either too limited to be used with cytoskeleton organization or so complex the results are hard to connect back to biological phenotypes [21]. Other correlation metrics are also not ideal for correlating orientations of the cytoskeleton components, and to date no metric has been developed or specifically characterized for this purpose [22–24].
In this work, we develop a new parameter with similar mathematical framework as OOP that will characterize both consistency of orientation of a single component and correlation of orientation of two components. As an example, we apply this Co-Orientational Order Parameter (COOP) to compare orientation of Z-lines and actin filaments in a neonatal rat ventricular myocyte (NRVM) monolayer. Lastly, we show how the COOP can be used to measure the consistency of building a cardiomyocyte on a triangular island of extracellular matrix (ECM).
One of the main goals of this work was to develop a metric to quantify the correlation between the orientation of different biological constructs within the cell or tissue. In designing the new metric, we aimed to overcome the challenge of analyzing orientation of multiple pseudo vectors, i.e. the metric needed to be symmetric with the period of π for both vectors. The OOP was designed to analyze the organization of pseudo vectors, and has become a standard parameter for use in liquid crystals [20]. The OOP ranges from zero, for isotropic, to one, for aligned mediums (S1 Fig.), and it has been applied to various biological systems [14, 16, 20]. However, the OOP was not designed to evaluate the correlation of orientation of coupled constructs such as actin fibrils and Z-lines. The first step in creating the COOP was to formally define the problem. Let the first construct be P, a set of pseudo vectors p i ⃗, and the second construct be Q, a set of pseudo vectors q i ⃗ (Fig. 1A). The order tensor and OOP of each field is:
𝕋 K = 2 k i , x k i , x k i , x k i , y k i , x k i , y k i , y k i , y - 𝕀 = { M e a n o r d e r t e n s o r } , (1)
O O P K = max eigenvalue ( 𝕋 K ) = { O r i e n t a t i o n a l o r d e r p a r a m e t e r o f K } , (2)
where K = {P, Q}, k i ⃗ = { p i ⃗ , q i ⃗ }, and 𝕀 is the identity matrix. To construct the new metric, we defined a new field F (a set of pseudo vectors f i ⃗):
f i , x = p i ⃗ · q i ⃗ = p i , x q i , x + p i , y q i , y = cos ( θ ) , (3)
f i , y = | p i ⃗ × q i ⃗ | = p i , x q i , y - p i , y q i , x = sin ( θ ) . (4)
Physically, field F represent the angle (θ) between the two biological constructs, p i ⃗ and q i ⃗. The metric was then calculated similarly to the OOP:
𝕋 P Q = 2 f i , x f i , x f i , x f i , y f i , x f i , y f i , y f i , y - 𝕀 = { M e a n t e n s o r o f t h e s y s t e m } . (5)
C O O P P Q = max eigenvalue ( 𝕋 P Q ) = C o - o r i e n t a t i o n a l o r d e r p a r a m e t e r o f t h e s y s t e m . (6)
The analytical solution of the COOP is:
C O O P P Q = 2 f i ⃗ · n ^ 2 - 1 = 2 f i ⃗ · n ^ 2 - 1 = 2 f i ⃗ · n ^ 2 - 1 , (7)
where n ^, the director, is the eigenvector associated with the maximum eigenvalue of mean tensor 𝕋PQ. The director represents the mean angle between the two constructs. Alternatively, the COOP can be written in the expended form:
C O O P P Q = ⟨ f i , x 2 ⟩ + ⟨ f i , y 2 ⟩ - 1 + ⟨ f i , x 2 ⟩ - ⟨ f i , y 2 ⟩ 2 + 4 ⟨ f i , x f i , y ⟩ 2 . (8)
The COOP was designed to range between zero and one. Here we present a series of theorems that illustrate the various properties of the COOP.
Theorem 1: Demonstration of COOP symmetry. Symmetry is an important characteristic of both the OOP and the COOP because it alleviates calculation errors that may arise when there is a random choice of signs for the pseudo vectors (Fig. 1B). Symmetry can be easily shown for OOP (S1 Supplemental Text). As can be seen from the anyaltical solution of the COOP (Equation 7), it is only necessary to demonstrate the symmetry of { f i ⃗ · n ^ } 2 in π to demonstrate the pseudo-symmetry of the COOP.
Table 1 shows the eight possible symmetry permutations. All of these can be reduced to the same equation with no difference in sign, which proves pseudo-vector symmetry of the COOP. We also demonstrated that the COOP is symmetric to the switch of P and Q:
f i , x ′ = q i , x p i , x + q i , y p i , y = f i , x and f i , y ′ = q i , x p i , y - q i , y p i , x = - f i , y , (9)
C O O P Q P = ⟨ f x 2 ⟩ + ⟨ ( - f y ) 2 ⟩ - 1 + ⟨ f x 2 ⟩ - ⟨ ( - f y ) 2 ⟩ 2 + 4 ⟨ - f x f y ⟩ 2 = C O O P P Q . (10)
Symmetry also plays an important role in interpreting the COOP director. There are four valid results for an angles between two pseudo-vectors: θ0, −θ0, π−θ0, θ0−π (Fig. 1B). For any symmetry permutation the director will correspond to one of these four angles. However, it is essential that the translation from n ^ to θ0 is handled with this symmetry in mind.
Theorem 2: Field rotation does not affect COOP. To verify that rotation of Q with respect to P does not affect COOP (Fig. 1C), let
p i ⃗ = cos ( α ) , sin ( α ) and q i ⃗ = cos ( β ) , sin ( β ) . (11)
If each q i ⃗ was rotated by angle ν, the rotated field Qrot would be defined by:
q ⃗ i , r o t = cos ( β + ν ) , sin ( β + ν ) = cos ( β ) cos ( ν ) - sin ( β ) sin ( ν ) , sin ( β ) cos ( ν ) + cos ( β ) sin ( ν ) . (12)
Then the rotated field Frot is:
f ⃗ i , r o t = [ f i , x cos ( ν ) - f i , y sin ( ν ) , f i , y cos ( ν ) + f i , x sin ( ν ) ] . (13)
As the angle ν is constant:
⟨ f i , r o t , x ⟩ = ⟨ f i , x ⟩ cos ( ν ) - ⟨ f i , y ⟩ sin ( ν ) and ⟨ f i , r o t , y ⟩ = ⟨ f i , y ⟩ cos ( ν ) + ⟨ f i , x ⟩ sin ( ν ) . (14)
In combining Equation (13) and (8), all terms with ν cancel or are reduced to cos2(ν)+sin2(ν) = 1. As a result COOProt = COOP. Thus we have proven that field rotation does not affect COOP, and without loosing generality we can assume n ^ p = n ^ q = [ 1 , 0 ]. This proves that the mean angle between fibers cannot be used to evaluate the correlation of orientations.
Theorem 3: The same field compared to itself gives COOP of 1. We next proved that the same field compared to itself would obtain a COOP of one (Fig. 1D). Imagine two sets of pseudo vectors distributed in a 2D space, p i ⃗ and q i ⃗. Let
p i ⃗ = cos ( α ) , sin ( α ) , and q i ⃗ = cos ( α ) , sin ( α ) . (15)
Then, the field f i ⃗ of this system can be written as:
f i , x = p i , x q i , x + p i , y q i , y = cos 2 ( α ) + sin 2 ( α ) = 1 and (16)
f i , y = p i , x q i , y - p i , y q i , x = cos ( α ) sin ( α ) - sin ( α ) cos ( α ) = 0 . (17)
The mean tensor:
𝕋 P Q = 2 f i , x f i , x f i , x f i , y f i , x f i , y f i , y f i , y - 𝕀 = 2 1 0 0 0 - 𝕀 = 1 0 0 - 1 . (18)
Therefore, the COOP of constructs P and Q:
C O O P P Q = max eigenvalue 1 0 0 - 1 = 1 . (19)
We obtained a COOP of 1 thus constructs P and Q are a perfectly co-oriented system which is expected as P = Q.
Theorem 4: Uncorrelated COOP limit. For any given pair of fields with OOPP and OOPQ there is a range of possible COOP values, with a maximum range of zero to one. To aid in interpreting the meaning of the COOP, we derived the value for the COOP (COOPu) as a function of the OOPs for the maximally uncorrelated system. For the COOP to be uncorrelated the two fields needed to have no correlation between their orientations (Fig. 1E). If and only if p i ⃗ and q i ⃗ are assumed independent and n ^ p = n ^ q (Theorem 2), the analytical solution Equation (7) can be re-written as:
C O O P u = 2 ( p i , x q i , x + p i , y q i , y ) 2 - 1 = ( 2 p i , x 2 - 1 ) ( 2 q i , x 2 - 1 ) . (20)
Using the classical probability distribution property:
C O O P u = ( 2 p i , x 2 - 1 ) ( 2 q i , x 2 - 1 ) = ( 2 p i , x 2 - 1 ) ( 2 q i , x 2 - 1 ) . (21)
Based on the definition of the OOP the uncorrelated COOP is:
C O O P u = ( 2 p i , x 2 - 1 ) ( 2 q i , x 2 - 1 ) = O O P P O O P Q = { U n c o r r e l a t e d C O O P } . (22)
COOPu is the lowest limit of the COOP if and only if the two constructs are independent. However, if P and Q are correlated in multiple ways it is possible to achieve a smaller value of COOP.
Theorem 5: Anti-correlated COOP. To prove that COOPu is not necessarily the minimum COOP, we constructed an example where P was composed of p 1 ⃗ and p 2 ⃗, and construct Q was composed of q 1 ⃗ and q 2 ⃗:
p 1 ⃗ = cos ( α ) , sin ( α ) and p 2 ⃗ = cos ( - α ) , sin ( - α ) . (23)
q 1 ⃗ = cos ( α + π 2 ) , sin ( α + π 2 ) and q 2 ⃗ = cos ( - α ) , sin ( - α ) . (24)
Construct P has an OOPP = cos(2α) and construct Q has an OOPQ = ∣sin(2α)∣. The vector f ⃗ is:
f 1 , x = 0 , f 2 , x = 1 and f 1 , y = 1 , f 2 , y = 0 . (25)
The mean tensor and COOP of this system are:
𝕋 P Q = 0 0 0 0 and C O O P = 0 . (26)
Using Equation (22) we obtained:
C O O P u = O O P P · O O P Q = cos ( 2 α ) sin ( 2 α ) . (27)
Unless, α ≠ { 0 , π 4 , π 2 } · n for n = {1, 2, …}, COOPu>0. Thus, the COOP can be lower than COOPu. Such a case can be graphically imagined if you have constructs correlated in two different ways in the same cell (Fig. 1F), similar to a single cardiomyocyte with punctate alpha-actinin at the ends and well defined Z-lines in the middle. This demonstrates that, similarly to the OOP, the COOP will not fully capture second order correlations (S1 Supplemental Text).
Theorem 6: Correlated COOP limit. For the upper limit, we derived the value of the COOP as a function of the OOPs for the maximally correlated constructs. To determine the correlated COOP (COOPc), we assumed the two fields, P and Q, were almost identical, except that Q was rotated by a random noise angle θ (Fig. 1G). We can assume, without loosing generality, n ^ p = n ^ q = [ 1 , 0 ] (Theorem 2), and P is better organized (i.e. OOPP>OOPQ, Theorem 1). Using the analytical solutions, the OOPP and COOP were rewritten as:
O O P P = 2 cos 2 ( α ) - 1 and C O O P = 2 cos 2 ( θ ) - 1 . (28)
As θ was assumed to be random noise generated, θ and α are independent, and therefore:
O O P Q = 2 cos 2 ( α - θ ) - 1 = ( 2 cos 2 ( α ) - 1 ) ( 2 cos 2 ( θ ) - 1 ) + 4 cos ( α ) sin ( α ) cos ( θ ) sin ( θ ) = ( 2 cos 2 ( α ) - 1 ) ( 2 cos 2 ( θ ) - 1 ) + 2 cos ( α ) sin ( α ) 2 cos ( θ ) sin ( θ ) = ( 2 cos 2 ( α ) - 1 ) ( 2 cos 2 ( θ ) - 1 ) + sin ( 2 α ) sin ( 2 θ ) = O O P P C O O P c . (29)
Solving for COOPc and rewriting it in a more general form we obtained:
C O O P c = m i n ( O O P P , O O P Q ) m a x ( O O P P , O O P Q ) . (30)
This is the upper limit of COOP if the two constructs are correlated but are subject to random biological variance (noise). This would not be the limit in a system where the variance is not random.
Theorem 7: Ultra-correlated COOP. We also showed that the correlated COOP is not necessarily the maximum. To prove this we defined P and Q as:
p i ⃗ = cos ( α i ) , sin ( α i ) and p ⃗ n + i = cos ( - α i ) , sin ( - α i ) (31)
q i ⃗ = cos ( α i + θ i ) , sin ( α i + θ i ) and q ⃗ n + i = cos ( - α i - θ i ) , sin ( - α i - θ i ) (32)
for i = {1, …, n}. Thus, each set of angles is paired in decreasing order (Fig. 1H). OOPP and OOPQ are defined as:
O O P P = 1 n ∑ i = 1 n cos ( 2 α i ) where α i : ∑ i = 1 n cos ( 2 α i ) ≥ 0 (33)
O O P Q = 1 n ∑ i = 1 n cos ( 2 α i + 2 θ i ) where α i , θ i : ∑ i = 1 n cos ( 2 α i + 2 θ i ) ≥ 0 . (34)
Assuming ∑ i = 1 n cos ( 2 θ i ) ≥ 0 the mean tensor is:
𝕋 P Q = 2 n ∑ i = 1 n cos 2 ( θ i ) 0 0 - 1 n ∑ i = 1 n ( 2 cos 2 ( θ i ) - 2 ) . (35)
Knowing C O O P = 1 n ∑ i = 1 n cos ( 2 θ i ) and the COOP c = O O P Q O O P P:
C O O P - C O O P c = 1 n ∑ i = 1 n cos ( 2 θ i ) - ∑ i = 1 n cos ( 2 α i + 2 θ i ) ∑ i = 1 n cos ( 2 α i ) . (36)
If the expression in Equation (36) is positive, then COOP ≥ COOPc. An example of such a case is when θ is constant and Equation (36) becomes:
C O O P - C O O P c = ∑ i = 1 n sin ( 2 α i ) sin ( 2 θ i ) ∑ i = 1 n cos ( 2 α i ) . (37)
Based on conditions in Equation (33) and (34), 0 ≤ α i ≤ π 2 and 0 ≤ θ i ≤ π 2, which implies sin(2αi) ≥ 0, cos(2αi) ≥ 0, and sin(2θi) ≥ 0 for all i. Therefore, every term in Equation (37) is positive. If P is not perfectly aligned (i.e. αi ≠ 0), and if P and Q are not identical (i.e. θi ≠ 0), then COOP>COOPc. Thus, there is an ultra-correlated COOP that can be greater than COOPc. Qualitatively, the ultra-correlated system is similar to the correlated example, except the q ⃗ vectors are not random within the noise, but are arranged maximum to minimum angles (Fig. 1H).
Normalized COOP. Based on Theorems 4–7 we can divide the range of COOP into three regions (Fig. 1I): anti-correlated, normal, and ultra-correlated. The boundaries of these regions (COOPu and COOPc) are determined by the organization of the two constructs (OOPp and OOPq) and will slide along the overall range [0, 1]. Experimentally, it might be more relevant to know how close the COOP is to the uncorrelated and correlated boundaries as these carry biological implications. We therefore defined a Normalized COOP, which is a measure of how close a parameter is to COOPu or COOPc:
Normalized COOP = C O O P - C O O P u C O O P c - C O O P u . (38)
The normalized COOP is negative when it is anti-correlated, zero when the system is uncorrelated, one when it is correlated, and greater than one when it is ultra-correlated (Fig. 1I). Any parameter has limitations. For example, the OOP of the red line-segments in the anti-correlated schematic (Fig. 1F) would be equivalent to an isotropic organization (OOP ∼ 0) even though the red line-segments are well organized. Similarly, the COOP cannot be used to identify correlations in tissues that have very non-trivial spatially dependent correlations (more complex than the ultra-correlated case (Fig. 1H)). We believe that most biological constructs for which the COOP has been developed will never exhibit this behavior. However, if the COOP is ever found to be statistically significantly greater than COOPc, it will be essential to re-evaluate the applicability of the parameter.
Estimated maximum tolerable error and minimum sample size. To interpret the information provided by the COOP, we would need to know which region our tissue falls under. Our ability to do so will be limited by the error inherent in any measurement and the width of the normal COOP range. COOPu approaches the maximum (COOPu = 1) as both OOPp and OOPq approach one (Fig. 1J(i)). COOPc, however, approaches its maximum when the two order parameters are close to being equal (Fig. 1J(ii)). This shows, for example, that if OOPp = 0 and OOPq = 0, then the normal region of the COOP ranges [0, 1]. However, if OOPp>0 and OOPq = 0, the normal range does not exist and a COOP>0 would indicate an ultra-correlated system. If the error in the system is so large that there is no statistically significant difference between the boundaries (COOPu and COOPc), it would not be possible to differentiate between the regions. Therefore, for the parameter to be useful, the maximum allowable error and minimum sample size have to be experimentally realistic. To estimate the error and sample size, we calculated the propagation of error in COOPu and COOPc, and used them in the student t-test to calculate statistical significance.
Assuming OOPP and OOPQ are normally distributed with the standard deviation σOOPP and σOOPQ, respectively, and OOPP and OOPQ are independent, then the variance:
σ C O O P u 2 = ∂ C O O P u ∂ O O P P 2 σ O O P P 2 + ∂ C O O P u ∂ O O P Q 2 σ O O P Q 2 = O O P Q 2 σ O O P P 2 + O O P P 2 σ O O P Q 2 , (39)
and the standard deviation:
σ C O O P u = O O P Q 2 σ O O P P 2 + O O P P 2 σ O O P Q 2 . (40)
Assuming OOPP>OOPQ, the variance for the correlated COOP:
σ C O O P c 2 = ∂ C O O P c ∂ O O P P 2 σ O O P P 2 + ∂ C O O P c ∂ O O P Q 2 σ O O P Q 2 = O O P Q O O P P 2 2 σ O O P P 2 + 1 O O P P 2 σ O O P Q 2 . (41)
Note that ∂ C O O P c ∂ O O P P and ∂ C O O P c ∂ O O P Q 2 do not exist when OOPP = OOPQ. However, using Theorem 1, we formulated the estimate for the the standard deviation of COOPc:
σ C O O P c = O O P Q O O P P 2 2 σ O O P P 2 + 1 O O P P 2 σ O O P Q 2 : O O P P > O O P Q 1 O O P Q 2 σ O O P P 2 + O O P P O O P Q 2 2 σ O O P Q 2 : O O P P < O O P Q . (42)
Then to estimate maximum allowable error, we assumed σOOPP = σOOPQ, and tested the null hypothesis of COOPu = COOPc. We calculated the t-value and degrees of freedom to complete the t-test. We assumed significance for a p-value less than 0.05 for the two sample t-test. There are two variables that impact significance as a function of OOPs: the error (σOOP) and the sample size (N). For visualization, we calculated the maximum allowable experimental error (max(σOOP)) at sample size of four (Fig. 1J(iii)). The maximal allowable error at N = 4 was at its highest, max(σOOP) = 0.18, for OOPP = OOPQ = 0.60, which is greater than experimental error reported for OOP organization [13]. While the maximal allowable error approaches zero as OOP→0, it rapidly increases for OOP>0.
We also determined the minimum sample size that provides statistical significance (p<0.05) as a function of construct organization for σOOP = 0.04. The minimum sample size ranges from two to infinity (Fig. 1J(iv)). Naturally, when COOPc→COOPu, the min(N) → ∞ as it is not possible to find them significantly different. For σOOP = 0.04, the minimum sample size is between two and five for most of the OOP values. For convenience, we also calculated the minimum sample size for a range of higher experimental errors (S2 Fig.). Obviously higher error would necessitate more samples to maintain significance. However, errors reported for OOP experimentally correspond to normal requirements in sample size. The errors and sample sizes were confirmed to be experimentally realistic, thus we next moved to testing the parameter with synthetic and experimental data.
Limiting cases. The first step to validate the new parameter was to construct four limiting cases that should lead to specific COOP values. We constructed a custom MatLab code that could be interfaced with experimental or synthetic (computer generated) data. The code was used to confirm the COOP for four synthetic limiting cases. The first test was to compare two perfectly aligned constructs (Fig. 2A(i)). It was clear from the image that these two constructs were perfectly correlated, and we therefore expected COOP = 1. We first confirmed that both constructs were perfectly aligned with OOPP = OOPQ = 1 (Fig. 2A(ii)). Analytically, it is clear that COOP = 1, by first applying Theorem 2 and then Theorem 3. When the synthetic data for this condition was analyzed by the code, the COOP was confirmed to be one (Fig. 2A(iii)). It is worthwhile to note that the director generated by the COOP code corresponds to the angle, θ0, between the constructs.
For the second limiting condition, one construct was perfectly organized, while the second was completely disorganized (Fig. 2B(i)). In this case, there was no correlation, and we expected COOP = 0. Analytically, even if we supposed that the two constructs were maximally correlated (COOP = COOPc) and applied Theorem 6 based on OOPQ = 1 and OOPP = 0 (Fig. 2B(ii)), the COOP = 0. This was confirmed by the result of the synthetic case (Fig. 2B(iii)).
The third case considered two isotropic constructs that are completely uncorrelated (Fig. 2C(i)-(ii)). We expected the COOP for uncorrelated constructs to be zero, and analytically this was confirmed by Theorem 4. These findings were also confirmed by the results of the synthetic data (Fig. 2C(iii)).
The fourth, most intriguing case, was of two isotropic constructs, which were perfectly correlated (Fig. 2D(i)-(ii)). We expected the COOP to be one as there was perfect correlation, and this was proven by Theorem 6. Synthetically we showed the COOP = 1, and again the director gave the average angle, θ0, between constructs. The limiting conditions validated the parameter for four simple cases. However, the interpretation of the COOP gets more complex when neither construct is perfectly aligned or isotropic.
COOPu and COOPc demonstrated with synthetic data. To understand the limits of the COOP parameter, we constructed a series of cases with different organizations by truncation of Gaussian distributions with specified standard deviations (Fig. 3A and 3B). We created synthetic data for the uncorrelated case (dark blue Fig. 3C) by generating two separate random number sets using the appropriate truncated Gaussian distribution for each. For the correlated case (brown Fig. 3C), we generated the first, more organized, data set by the same method. Then the noise was generated such that when it was added to the first data set, the new set would have the target distribution. Both methods of creating P and Q data sets lead to the same desired individual distributions (Fig. 3A and 3B). To visualize the results we constructed a slider, sketched in Fig. 1I, for each case (Fig. 3C) color-coded to indicate the boundaries, COOPu (dark blue) and COOPc (brown), as well as the three regions: anti-correlated (light blue), normal (ranging from dark blue to brown), and ultra-correlated (bright green). We expected that when OOPP = 1 (Fig. 3A(i)), COOPu = COOPc = OOPQ for all OOPQ values, which was confirmed by the results of the synthetic data (Fig. 3C(i)-(iii)). Also, if one OOP = 0 (Fig. 3B(i)) and the other OOP ≠ 0, the COOPu = COOPc = 0. This was also confirmed with synthetic results (Fig. 3C(i) and 3C(iv)). For a case where OOP ≠ {0 or 1} the normal range of the COOP was greater if the OOPs were smaller and closer to each other, which can be seen by comparing Fig. 3C(v) and Fig. 3C(vi). These synthetic results confirm the results of Fig. 1, which showed that the COOP parameter is best applied in situations where the correlation is not obviously dictated by the organization of the individual constructs.
The COOP was designed as a tool to evaluate correlations of orientation in experimental samples. We used actin fibers and sarcomeric Z-lines in NRVMs to validate the parameter and code. The tissues were stained for α-actinin and phalloidin to identify Z-line and actin fibril directions, respectively (Fig. 4A). The program we used to identify the direction of the construct was based on a finger-print identification code [6, 25, 26], and it assigned a direction to every non-empty pixel in the image. However, computing correlation of constructs based on individual pixels introduced too many errors. Indeed, most images of cytoskeleton constructs are obtained via immunostaining and imaging. The accuracy of the COOP will be a direct consequence of image quality. If the images are of poor quality (poor contrast, dead cells, etc.), it will not be possible to accurately extract the construct direction data, and thus, the COOP will not be accurate. However, if it is possible to accurately extract directionality data, the COOP can be used. The images we have collected for the proof-of-principle data set are representative of the images normally used to study tissue architecture [6, 20]. The images may be resolvable to different degrees, and image acquisition procedures can introduce inaccuracies at smaller scales. For example, during the collection of this data, it is customary to ensure that each channel results in sharp images of the corresponding construct, such as actin and Z-lines. This sometimes requires focusing on slightly different planes, and as a result, the images of the same field-of-view could be slightly off-set from each other. This along with other imaging inaccuracies lead us to develop a procedure to average out these small errors by calculating the direction of each construct within grid-squares. While any consistent small grid can be used for comparing results across multiple tissues, we recommend picking a grid size that would correspond to a natural biological unit. In this case, the image was partitioned into a grid (Fig. 4B), which was chosen such that at minimum, two Z-lines could be expected to fit within each grid-square, ensuring that at least one “sarcomere” complex is within each square (S3 Fig.). The spacing between Z-lines in NRVM tissues is 1.9–2.1 μm [27], thus we chose a grid size close to 4.2 μm (∼ 30 pixels). For each grid (i), we calculated the average direction for both constructs (i.e. Z-lines or actin fibrils) (Fig. 4C). To account for varying densities and partial grid-squares at the edges of the images, the area density (ρ) was calculated for each grid-square using the number of non zero angles in the grid-square divided by its area. Each grid-square (i) was assigned a weight (Wi) based on the OOP and density of constructs (ρ):
W i = O O P P , i · O O P Q , i · ρ P , i · ρ Q , i . (43)
Thus, partial grids with low densities have small weight factors whereas full grids have high weight factors (length of arrows in Fig. 4D). Additionally, grid-squares with better alignment have higher weights than grid-squares with isotropic organization, which prevents loss of consistency.
Z-lines and actin fibers were expected to be perfectly correlated within the sarcomeres of the healthy cardiac muscle tissue. However, histological samples may not be perfect, with some example imperfections identified in the zoomed-in image of Fig. 4C, which correspond to the non-perpendicular arrows identified in Fig. 4D. Note that the lengths of the arrows correspond to the weights assigned to each square, so these imperfections can significantly alter the resulting COOP. To test the parameter, we took four coverslips with 10 fields of view imaged for each and identified noiseless regions with minimal imperfections (Fig. 4E-F). Before any angle detection was done, ImageJ was used to merge fields of view containing Z-lines and actin fibrils, and regions with minimal imperfections were chosen by an experienced user. Only merged images that contained four or more regions of minimal imperfections were used (Fig. 4E). This was done prior to the organizational analysis to eliminate bias. The implementation of the code is summarized in a flow chart (S4 Fig.).
The OOPs for actin fibrils and Z-lines were essentially the same for both the raw and masked images. We measured no significant difference in the values of COOPu or COOPc between the full and noiseless data sets (Fig. 4G). There was also no significant difference between COOP and COOPc for the masked, noiseless images. In contrast there was a significant difference between COOP and COOPc (p<0.001) for the raw images. This is reflected in the Normalized COOP which is 1 for the noiseless images and less than 0.5 for the full data (Fig. 4H). This illustrates that while the parameter is capable of capturing the expected correlation in a sarcomere between actin fibril and Z-line orientation, cardiac tissues may have imperfections that result in a lower value of COOP. The mean angle for the noiseless images also shows the expected perpendicular correlation between orientations of sarcomere Z-lines and actin fibrils with minimal error (Fig. 4H). While, they were also approximately perpendicular in the raw images, the error was greater than for the noiseless images.
The COOP can also be used to evaluate the consistency of construct orientation within cells of the same shape. As an example we seeded NRVMs on triangular islands, and stained the samples for nucleus, Z-lines, actin fibrils, and fibronectin (Fig. 4I). For analysis we ensured that the fibronectin islands lined up for all five cells (S5 Fig.). Then the Z-line (solid bars) or actin fibril (dashed bars) images were compared in a pairwise manner (Tables 2 and 3). This showed an additional experimental confirmation of Theorem 3: COOP = 1 for each cell when it was compared with itself (Fig. 4I). The COOP was calculated for the same grid as in the isotropic images. The results showed that although there is an overall consistency between cells, i.e., myofibrils were bundled along the edges of the triangle, the orientation was not fully consistent at a smaller length scale (grid size). Indeed, the average COOP for ten pair-wise comparisons (bold in Tables 2 and 3) of both Z-lines and actin fibrils is less than 0.5 (Fig. 4K). This demonstrated another potential use of this parameter.
Colocalization, a process that analyzes the spatial overlap between two biological constructs, has been key in discovering cellular mechanisms that rely on the proximity of constructs [28, 29]. For example, second-order stereology has been used to analyze spatial arrangements of constructs in images. Noorafshan et al. used second-order stereology to examine the correlation between the spatial arrangements of cardiomyocytes and microvessels [22]. Their method involved pair correlation and cross-correlation functions to determine positive or negative correlation at different distances. However, neither simple colocalization nor second-order stereology do not analyze the relative orientations of biological constructs.
In this work, we have developed a new parameter, COOP, to characterize how two tissue components align with respect to each other. The COOP would allow for investigation of mechanisms or functions that rely on not only spatial proximity, but also specific organizational schemes. To properly interpret the meaning of the parameter values, we characterized it through a series of analytical theorems. As a result, we defined three regimes—normal, ultra-correlated, and anti-correlated—that have biological implications. After validating the parameter with synthetic data, we demonstrated its use with experimental images by showing that perfect portions of cardiac tissues have the expected correlation of the orientation between sarcomeric z-discs and actin fibrils. The reduction in the COOP for un-masked (raw) data suggests that the defects in the architecture will be distinguished by our new method (Fig. 4G-H). The code we have developed can also be used to calculate the mean angle between constructs thus allowing for tracking of mean angle changes as a function of experimental conditions. Furthermore, the parameter can be used to calculate the consistency of orientational organization to help evaluate the importance of orientational order.
To compare organization between different experimental conditions, it is necessary to have a robust metric. The best metrics place the least number of constraints on the distribution of orientations. For example, the standard deviation is not an appropriate metric for quantifying orientation distribution of Z-lines as they are not distributed normally. The OOP works with all types of distributions, and it has an additional benefit of being symmetric to pseudo-vectors [14]. As the COOP was designed with similar math, it shares the same benefits as the OOP such as pseudo-vector symmetry.
In general, there are multiple ways to use mathematical functions to analyze the properties of images. For example, Feng et al. use normalized cross-correlation to compare two images with a possible rotation or change in scale [23]. Their method involves identifying a relatively small number of points of interest and matching the comparison based on them. The Feng et al. method is insensitive to the rotation of the whole image (i.e. rotation plus translation), while the COOP method is insensitive to the rotation of all vectors without translation. The normalized cross-correlation is a powerful tool, but not appropriate whole cell architectural metric, the COOP is therefore useful for comparing consistency in similarly shaped cells with matching ECM islands, but cannot be used to identify the same cell that has been re-scaled and rotated. Another example of mathematical tools for image analysis is a set of a non-parametric circular statistics tests such as Watson’s U n 2 test, which is designed to evaluate the probability that a sample comes from a specific distribution or that two samples come from the same distribution. For instance, non-parametric circular statistics has been utilized to evaluate if a pattern of migration of different objects is the same [30, 31]. However, these tests do not consider the location of each sample pair, thus while the results can correspond to the COOP in very special situations these parameters address fundamentally different questions. Thus, these circular statistics tests are not a good tool to evaluate orientational correlation of co-localized pseudo-vectors.
There are specific cases where the COOP will correspond to other parameters. For example, the OOP has been used to quantify the organization of the bacterial population in a channel with respect to the channel direction [16]. Indeed, this is equivalent to a rudimentary case of the COOP where one of the constructs, the channel, is perfectly organized (Fig. 3C(ii)-(iii)). The COOP is more general in that it can be used when neither construct is perfectly organized. Circular statistics tool-sets include some correlation metrics [24], such as the circular correlation coefficient [21] which corresponds to the COOP in the same case. Specificically the circular correlation coefficient can only be used for uniform distributions (i.e. isotropic tissues). In that special case, the circular correlation coefficient and the COOP converge to the same equation (S1 Supplemental Text). However, the more general vector formulation of the circular correlation coefficient, while not constrained to a uniform distribution, is very complex, and thus cannot be easily characterized the way we have done for the COOP. This circular correlation coefficient would not be a convenient metric for cytoskeleton or cellular orientation quantification. The COOP can be calculated so long as two sets of angular distributions and their locations are known, and it has been extensively characterized. Thus, this new parameter can be used with a multitude of biological systems.
In a healthy, properly functioning cell or tissue, the cytoskeleton needs to be organized in an intricate manner. In disease, loss of this organization leads to reduction in function, such as the myofibril organization changes in dilated cardiomyopathy [32–34]. However, proper organization of one element in a cell or tissue is not sufficient. The multiple constructs have to be properly organized with respect to each other, and that organization can have biological implications. This has been shown to occur during maturation of myocytes where the α-actinin is initially punctate and parallel to actin fibrils, but, in mature cardiomyocytes, becomes part of the newly formed sarcomeric Z-lines, which are perpendicular to actin fibrils [35]. In this case the relative orientation of α-actinin staining and actin fibrils indicates the progressive maturation of myofibrils. An additional example where orientation of different constructs affects each other is when the organization of the extracellular matrix can be used to control the architecture of cells [2, 13, 36]. Conversely, cells have been shown to change the orientation of the extracellular matrix fibrils [37]. Another instance of organization correlation can be found in the different cell types and collagen fibrils within heart valves [38]. The common use of such metrics as the OOP and COOP for biological and medical sciences will allow for a quantitative evaluation of tissue engineered substrates from a variety of cell sources. Combining such metrics with histology will create a universal evaluation metric between in vitro and in vivo systems favorably impacting our ability to design replacement tissues, to create in vitro drug testing platforms, and to evaluate pathological reports in the clinic.
All animals were treated according to the Institutional Animal Care and Use Committee of UCI guidelines (Animal Experimentation Protocol permit number 2013-3093-0). This protocol met the guidelines for the use of vertebrate animals in research and teaching of the Faculty of Arts and Sciences of UCI. It also followed recommendations of the NIH Guide for the Care and Use of Laboratory Animals and was in accordance with existing federal (9 CFR Parts 1, 2 & 3), state, and city laws and regulations governing the use of animals in research and teaching.
To facilitate the calculation of the COOP we created a custom MATLAB code. The code was designed to have an input of angles for P and Q organized such that the information of which pseudo-vectors are paired was not lost. The code outputs were OOPP, OOPQ, COOP, COOPu, COOPc, n ^, and θ.
Synthetic data of isotropic constructs for limiting conditions (Fig. 2) was generated using a random number generator (rand) that provides a uniform distribution of at least 106 random values in MATLAB. Each construct used in testing COOPu and COOPc (Fig. 3) contained 108 random numbers (MATLAB function normrnd) that were normally distributed with the specified mean and standard deviation. We have included the codes to create synthetic data as supporting codes (S1 Code and S2 Code).
Microcontact printing and ECM patterns. To make the substrates 25 mm glass coverslips were coated with PDMS (Ellsworth Adhesives, Germantown, WI) and cured for 12 hours in a 60°C oven. To create triangular myocytes we utilized a microcontact printing procedure similar to that described by Tan et al [39]. A mask with the desired pattern was designed using Adobe Illustrator (Adobe Systems Incorporated, San Jose, CA) and made by Front Range Photomask (Palmer Lake, CO). The mask was used to make a silicone wafer (Integrated Nanosystems Research Facility, Irvine, CA). A polydimethylsiloxane (PDMS) stamp, cast from a silicon master, was used to contact transfer the extracellular matrix (ECM) protein fibronectin (FN) (Fisher Scientific Company, Hanover Park, IL) onto a UV-sterilized (UVO, Jelight Company, Inc. Irvine, CA) PDMS-coated coverslip. Fabricated substrates underwent one 10 minute pluronics (250g of Pluronics F-127, Sigma-Aldrich, Inc., Saint Louis, MO) wash and three rinses of phosphate buffer-saline (PBS) (Life Technologies, Carlsbad, CA). To make isotropic substrates, UV-sterilized PDMS-coated coverslip were coated with FN for 10 minutes and underwent three PBS washes. The substrates were stored at 4°C prior to NRVM seeding.
Cardiomyocyte culture. Cell cultures of NRVMs were prepared from two-day old Sprague-Dawley rats (Charles River Laboratories, Wilmington, MA). A mid-sternal incision was made in order to expose the heart of the neonatal rat for dissection. Ventricular tissue was removed and rinsed in a Hanks balanced salt solution buffer (Life Technologies, Carlsbad, CA) and placed in 1 mg/mL trypsin solution (Sigma-Aldrich, Inc., Saint Louis, MO) to be shaken overnight (12 hour incubation) at 4°C. The next day, isolated tissue was dissociated into individual cells by treatment with four separate washes of 1 mg/mL collagenase type II (Worthington Biochemical, Lakewood, NJ) for two minutes at 37°C. Isolated cardiomyocytes were resuspended in M199 culture medium (Invitrogen, Carlsbad, CA) supplemented with 10% heat-inactivated Fetal Bovine Serum, 10 mM HEPES, 20 mM glucose, 2 mM L-glutamine, 1.5 μM vitamin B-12 and 50 U/ml penicillin. The cell solution was filtered with a 40 μm filter (Thermo Fisher Scientific, Waltham, MA), and the remaining cells were pre-plated multiple times to eliminate fibroblast contamination. Immediately after purification, myocytes were plated on substrates (prepared as detailed above) at a density of 106 or 105 cells per well in a standard six-well plate for confluent or sparse cultures, respectively. These were incubated at 37°C with a 5% CO2 atmosphere. Seeded cultures underwent a wash with PBS 24 hours after plating to remove unattached and dead myocytes. They were then cultured in 10% serum media for another 24 hours at which point the media was changed to 2% serum media. After a total of 72 hours in culture, the samples were fixed and immunostained.
Fixing, immunostaining and imaging. After 3–4 days in culture, confluent monolayers of cardiomyocytes were fixed with 4% paraformaldehyde (PFA) (VWR, Radnor, PA) with 0.01% Triton X-100 (Sigma-Aldrich, Inc., Saint Louis, MO) for 10 min, and rinsed three times with PBS in 5-min intervals. Cardiomyocytes were stained with nuclei acid-sensitive dye 4’, 6’-diaminodino-2-phenylinodole (DAPI) (Life Technologies, Carlsbad, CA) for chromatin, FITC-phalloidin (Alexa Fluor 488 Phalloidin, Life Technologies, Carlsbad, CA) for actin, monoclonal mouse sarcomeric anti-α-actinin (Sigma-Aldrich, St. Louis, MO), and polyclonal rabbit anti-human fibronectin (Sigma-Aldrich, St. Louis, MO) and incubated for a total of 1–2 hours at room temperature. Secondary staining was applied using tetramethylrhodamine-conjugated goat anti-mouse IgG antibodies (Alexa Fluor 633 Goat anti-mouse, Life Technologies, Carlsbad, CA) and goat anti-rabbit IgG antibodies (Alexa Fluor 750 goat anti-rabbit, Life Technologies, Carlsbad, CA) for a 1–2 hour incubation. After each incubation period, coverslips were rinsed three times with PBS for 5–10 min. Each coverslip was then mounted onto a microscope slide preserved with prolong gold antifade reagent (Life Technologies, Carlsbad, CA). The images were collected using an IX-83 inverted motorized microscope (Olympus America, Center Valley, PA) with an UPLFLN 40x oil immersion objective (Olympus America, Center Valley, PA) and a digital CCD camera ORCA-R2 C10600-10B (Hamamatsu Photonics, Shizuoka Prefecture, Japan). For isotropic monolayers, at least ten fields of view were collected for every sample.
‘Noiseless” image generation. A macro was created in ImageJ that allows the user to select regions without imperfections in an image. The regions that were not selected became masked, resulting in a series of images with only regions of interest displayed in the new masked images for every channel imaged (for example: DAPI, m-cherry, GFP)(Fig. 4). The masked images could then be analyzed using the same codes used for raw images.
Calculating construct angles. To determine construct angles, we adapted a previous MATLAB code that detects ridges of a fingerprint [20, 25, 26]. This code was used to detect Z-lines and actin fibers in the images. In the code, a binary mask applied to the image determined the constructs and a filter was applied to clean up the constructs that were identified in the images. The code took pixel information from the images and for every non-empty pixel in the image, a pseudo-vector was calculated and used to determine the OOP for Z-lines and actin fibrils, as well as a new set of pseudo-vectors for each square in the grid (Fig. 4B, D) These new pseudo-vectors were then utilized to calculate the COOP between two constructs (i.e. Z-lines and actin fibrils) or two cells (Fig. 4).
Statistics. To calculate the average angle between the constructs (⟨θ0⟩) and the standard deviation of those angles (σθ0) across multiple conver-slips, it is essential to keep in mind that the angle period is π. The simplest way, but not the only way, to generate ⟨θ0⟩ and σθ0 is to calculate the director of the director pseudo-vectors n ^ ALL of each cover-slip. Meaning that in Equation (1) k i ⃗ = n ^ i where i is the cover-slip, and the n ^ ALL is the the eigenvector of the tensor from Equation (1) that corresponds to the eigenvalue from Equation (2). The angle for each cover-slip is then determined as follows
θ 0 , i = a r c c o s ( n ^ i ) for π 4 < a r c c o s ( n ^ A L L ) < 3 π 4 a r c s i n ( n ^ i ) for a r c c o s ( n ^ A L L ) > 3 π 4 or a r c c o s ( n ^ A L L ) < π 4 . (44)
The ⟨θ0⟩ and σθ0 are the average and standard deviation of θ0,i for all cover-silps. Obviously if the COOP →0 then this procedure is useless as the angles will be inconsistent between the constructs. However, this procedure is a convenient way to determine the range in which it is most convenient to report the angle (i.e. 0→π or − π 2→π 2). Additionally, there are two possible angles that can be calculated (Fig. 1B). We chose to provide the clock-wise angle from p ⃗ to q ⃗, but it is also possible to calculate only the acute angle instead. To compare the COOP, COOPc, and COOPu in the analysis of the experimental data, the one way ANOVA with the Student-Newman-Keuls test was used.
|
10.1371/journal.ppat.1007402 | HEXIM1-Tat chimera inhibits HIV-1 replication | Transcription of HIV provirus is a key step of the viral cycle, and depends on the recruitment of the cellular positive transcription elongation factor b (P-TEFb) to the HIV promoter. The viral transactivator Tat can displace P-TEFb from the 7SK small nuclear ribonucleoprotein, where it is bound and inactivated by HEXIM1, and bring it to TAR, which allows the stalled RNA polymerase II to transition to successful transcription elongation. In this study, we designed a chimeric inhibitor of HIV transcription by combining functional domains from HEXIM1 and Tat. The chimera (HT1) potently inhibited gene expression from the HIV promoter, by competing with Tat for TAR and P-TEFb binding, while keeping the latter inactive. HT1 inhibited spreading infection as well as viral reactivation in lymphocyte T cell line models of HIV latency, with little effect on cellular transcription and metabolism. This proof-of-concept study validates an innovative approach to interfering with HIV transcription via peptide mimicry and competition for RNA-protein interactions. HT1 represents a new candidate for HIV therapy, or HIV cure via the proposed block and lock strategy.
| HIV remains a major health issue, with still no vaccine or cure available, and lifelong antiretroviral treatment required for the always-increasing number of people living with the virus. Combination antiretroviral therapy inhibits HIV replication, but the persistence of latently infected cells remains a challenge. In this study, we developed a new approach to inhibiting HIV transcription with a chimera derived from host and viral proteins involved in the regulation of HIV gene expression. We fused a domain from the viral transactivator Tat to two domains from the host cell transcription regulator HEXIM1. The chimera (HT1) binds to TAR, inhibits P-TEFb, and prevents Tat transactivation of the HIV promoter. Cellular genes are not impacted. When stably expressed by lymphocyte T cells, the chimera potently inhibits HIV replication and reactivation from latency, which makes it a promising candidate for therapy or cure by a block and lock approach.
| Treatment with combination antiretroviral therapy (cART) leads to efficient suppression of HIV replication, but HIV persistence in latently infected cells remains an obstacle to cure [1]. Even under cART, residual HIV replication can arise and ultimately lead to the emergence of replicative resistance mutations and viral escape. Targeting diverse steps of the viral life cycle is the most efficient way to prevent viral escape. Currently, viral entry, reverse transcription, integration and maturation steps have been targeted by cART [2]. However, no efficient transcription inhibitor is clinically available, though multiple strategies–such as TAR decoys [3] or dominant-negative Tat [4]—have been explored to prevent expression of the integrated provirus.
Blocking transcription would not only add another therapeutic target, but also prevent sporadic reactivation of integrated HIV [5] that may contribute to HIV persistence, reservoir replenishment and chronic inflammation [6–8]. Suppressing residual HIV transcription is also the goal of the emerging block and lock HIV cure strategies [9–11], which aim at deepening HIV latency so that integrated proviruses remains permanently locked in the infected cells. Several latency promoting agents (LPAs) have been proposed, such as didehydro-cortistatin A [10], curaxin 100 [11], ruxolitinib and tofacitinib [12]. More studies are needed to determine whether a permanent state of latency can actually be reached after LPA treatment is interrupted. This would validate block and lock strategies as a path to a functional cure and/or a faster reservoir decay, a process that is probably delayed by residual replication and cell proliferation [13].
HIV expression is dependent on the viral transactivator Tat, which brings the cellular positive transcription elongation factor B (P-TEFb) to the HIV promoter [14, 15]. P-TEFb is comprised of cyclin T1 (CycT1) and cyclin-dependent kinase 9 (CDK9) [16, 17], and is required for transcription elongation, both for HIV and host gene expression. In cells, most of P-TEFb is sequestered in the 7SK small nuclear ribonucleoprotein complex (7SK snRNP), which includes a non-coding 7SK snRNA and proteins HEXIM1, LARP7 and MePCE [18, 19]. In 7SK snRNP, the transcriptional regulator HEXIM1 binds to the 7SK snRNA through a RNA-binding arginine rich motif (ARM, residues 150–162, see Fig 1A), and to P-TEFb through its CycT1 binding domain (TBD, residues 250–359 [20, 21] and its central inhibitory domain (ID, residues 200–211, Fig 1A). This ID includes a PYNT motif (202Pro-203Tyr-204Asn-205Thr), which masks CDK9’s substrate-binding site and is critical for its inactivation [22, 23]. Importantly, HEXIM1’s TBD acts synergistically with ID on CDK9 inhibition [24].
Without recruitment of active P-TEFb to the HIV promoter, RNA polymerase II (RNAPII) is stalled after having only transcribed the short transactivation response element (TAR) RNA, located at the 5’ end of all viral transcripts [15, 25]. The Tat activation domain (AD, residues 1–48, see Fig 1A) binds CycT1, while its central ARM binds to 7SK RNA, thus displacing P-TEFb from the 7SK snRNP and releasing it from HEXIM1 inhibition [26]. Tat and CycT1 also form a cooperative binding surface for TAR, where the central ARM region of Tat (residues 51–57) binds to the bulge region of TAR and the Tat-TAR recognition motif of CycT1 binds to the central loop of TAR [27]. These interactions allow P-TEFb to be recruited to the RNAPII early elongation complex that is stalled at the HIV transcription start site. There, CDK9 phosphorylates transcriptional inhibitory complexes NELF and DSIF as well as RNAPII C-terminal domain (CTD), resulting in stimulation of transcription elongation [5, 28].
Interestingly, Tat/TAR/P-TEFb interaction structurally mimics that of HEXIM1/7SK/P-TEFb, and the amino-acid sequences of Tat and HEXIM1 ARMs are almost identical [22]. Since the former is a strong transcription activator for HIV, while the latter is a potent inhibitor of P-TEFb, we sought to create a HIV-specific transcription inhibitor by taking advantage of the structure similarities of these two complexes. We designed chimeras that derive from critical functional domains of Tat and HEXIM1, by combining the P-TEFb binding N-terminal domain of Tat to the acidic and/or central basic domains of HEXIM1 that respectively inhibit P-TEFb and bind RNA. A small HEXIM1-Tat chimera, HT1, inhibited HIV transcription by preventing the recruitment of active P-TEFb to TAR, with only little off-target effects on cellular genes. This proof of concept study demonstrates the feasibility of designing highly specific transcriptional inhibitor chimeras.
We screened a collection of chimeras (Fig 1A and S1 Fig) derived from the ARM and/or ID of HEXIM1 fused to the AD of Tat (Tat1-48). These chimeras include: HT1 with both HEXIM1 domains of interest (Hex150-220), HT2 with only HEXIM1 ARM (Hex150-177) and HT3 with only HEXIM1 ID (Hex178-220).
Luciferase (Luc) reporter assays were performed to titrate the potency of each chimera to inhibit Tat-dependent gene expression from the HIV promoter (Fig 1B, upper schema). Effector plasmids included pHT1-3 and pTat, which respectively expressed a Myc-epitope tag chimera (m:HT1-3) and a Flag-epitope tagged Tat (f:Tat). pHT1, pHT2 or pHT3 was transiently co-transfected in 293T cells with pTat and a Luc reporter gene under the control of the HIV promoter (P-HIV, expressed from the plasmid pLTR-Luc) (Fig 1B, upper panel). HT2, which does not include HEXIM1 ID, failed to inhibit Luc expression from the HIV promoter, even when the ratio of pHT2 to pTat was 2:1 (Fig 1B, middle panel). HT3, which included the ID, induced up to a 2-fold decrease in Luc expression from the HIV promoter at the 2:1 ratio (Fig 1B, middle panel). Finally, HT1 lead to a 4-fold decrease in Luc expression at a 2:1 ratio (Fig 1B, middle panel). The stronger inhibitory effect by HT1 when compared to that by HT3 suggested that HEXIM1 ARM also contributes to the inhibition of Tat-induced Luc expression from the HIV promoter. Importantly, mutating the PYNT motif, which is critical for HEXIM1's CDK9 inhibition, abolished the inhibitory effect of HT1 (HT1.PNND in S1 and S2 Figs, bar 3), suggesting that HT1's inhibitory effect on HIV transcription is mediated by CDK9 inhibition. Adding different lengths of flexible peptide linkers (GGGGS) slightly, but insignificantly improved the inhibitory effects of HT1 (S1 and S2A Figs, bars 4 and 5). Also, reversing the order of peptide motif (1TH in S1 Fig) also abolished HT1's inhibitory effect (S2A Fig, bar 7), indicating that precise spatial arrangement of these motifs is required. Similar results were obtained when co-transfecting pHT1-3 and pTat in NIH1 cells, which stably carry an LTR-Luc reporter gene [29] (S2B Fig). Thus, we selected HT1 for further investigation.
Levels of protein expression from pHT1 and pTat in 293T cells were confirmed by western blotting (WB, Fig 1B, lower panel) and suggested that the expression of Tat may be slightly reduced upon co-transfection of pHT1. To rule out any bias in the inhibitory titration of HT1, we thus used pNL43ΔenvLuc, a defective HIV molecular clone from which Tat is expressed from the HIV promoter and not from a separate plasmid. In this model, transient expression of HT1 led to a 3-fold decrease in Luc expression (Fig 1C). Taken together, these results suggest that HT1 potently inhibits gene expression from the HIV promoter.
We next investigated the mechanisms by which HT1 can prevent HIV gene expression. Since both Tat AD and HEXIM1 ID interact with P-TEFb, we performed a series of co-immunoprecipitations to determine whether HT1 interacted with P-TEFb and competed with Tat for P-TEFb binding. m:HT1 and f:Tat were transiently expressed in 293T cells, and both bound to CycT1 as demonstrated using anti-Myc or anti-Flag antibodies (Abs) for co-immunoprecipitation (Fig 2A, lane 3 in upper panel and 4 in middle panel, respectively). Moreover, when co-expressing a fixed amount of f:Tat and increasing amounts of m:HT1 (Fig 2A, lanes 5–7, lower panel), the amounts of CycT1 co-immunoprecipitating with f:Tat decreased (Fig 2A, lanes 5–7, middle panel), suggesting a competition between HT1 and Tat for P-TEFb binding.
Next we investigated whether HT1, which contains HEXIM1 ID, inhibited the kinase activity of CDK9. m:HT1 was transiently expressed in 293T cells and immunoprecipitated using anti-Myc Ab. Co-immunoprecipitated P-TEFb was subjected to an in vitro kinase assay with ATP and recombinant GST-CTD proteins as a substrate. Similarly, m:Tat was expressed to co-immunoprecipitate P-TEFb as a positive control. Immunoprecipitated CDK9, and phosphorylated GST-CTD (CTD-P) were detected by WB using anti-CDK9 and anti-Ser2P Abs, respectively (Fig 2B). A larger amount of CDK9 was co-immunoprecipitated by HT1 than by Tat, while more CTD-P was detected with Tat than with HT1 (Fig 2B, left panel, lanes 3 and 4). Relative kinase activity associated with HT1 and Tat was calculated by CTD-P band intensity normalized with CDK9, which revealed that the kinase activity of P-TEFb was decreased 3.1 fold when bound to HT1, compared to control (Fig 2B, right panel). This suggests that once bound to P-TEFb, HT1 can inhibit the kinase activity of CDK9. Consistently, addition of another non-inhibitory P-TEFb-binding motif (PID) from Brd4 to HT1 decreased the ability to inhibit HIV transcription (PID.HT1 in S1 and S2 Figs, bar 6).
Finally, since HEXIM1 ARM resembles the TAR-binding domain from Tat, we investigated whether HT1 could bind to HIV TAR. m:HT1 or m:Tat was co-expressed in 293T cells with TAR RNA, expressed under RNA Polymerase III dependent H1 promoter [3], and immunoprecipitated using anti-Myc Ab or IgG as a control. TAR RNAs co-immunoprecipitated with HT1 or Tat were quantified by RT-qPCR analysis. Both HT1 and Tat immunoprecipitated TAR (Fig 2C, upper left panel, lanes 2 and 3), though m:Tat immunoprecipitated 2.54 fold more TAR than m:HT1. Since the level of m:Tat expressed in cell lysates was higher than that of m:HT1 (Fig 2C, lower left panel, lanes 2 and 3), we normalized the amount of immunoprecipitated TAR by the protein levels of m:HT1 and m:Tat, which suggested that similar amounts of m:HT1 bound at least as much TAR as m:Tat (Fig 2C, right panel). m:HT3, which lacked the ARM domain from HEXIM1 (Fig 1A), failed to bind TAR (S3A Fig). To test whether HT1 affects Tat-TAR interactions, increasing amounts of m:HT1 were also co-expressed with a fixed amount of f:Tat and TAR (Fig 2D). RNA immunoprecipitation assays were then performed using anti-Flag Ab followed by TAR RT-qPCR, and indicated that the amounts of TAR RNA co-immunoprecipitated with f:Tat decreased progressively when expression of m:HT1 increased (Fig 2D, upper panel). WB confirmed that the amounts of HT1 used for this assay did not impact on f:Tat expression (Fig 2C, lower panel), suggesting that the decrease in co-immunoprecipitated TAR was due to HT1 competing with Tat for TAR binding. As expected, HT1 also bound to endogenous 7SK snRNA (S3B Fig).
These results confirm that HT1 competes with Tat for P-TEFb binding and keeps its kinase subunit CDK9 inactive, which reduces the amount of P-TEFb that is available for Tat to bring to TAR. Moreover, HT1 also binds to TAR and prevents Tat from binding it, consistent with the observation in Fig 1B of a more potent inhibition by HT1 than by HT3. Two mechanisms are thus combined that prevent Tat from bringing active P-TEFb to TAR for successful HIV transcription elongation.
Since P-TEFb is a major transcription factor and regulates the expression of many cellular genes, we next assessed whether HT1 could impair host cell gene expression through inactivation of the kinase activity of CDK9. We first investigated how HT1 impacted the mRNA and protein expression levels of HEXIM1, a bona fide target of P-TEFb [30]. Increasing amounts of m:HT1 were transiently expressed in 293T cells and did not change the expression level of HEXIM1 protein (Fig 3A) and mRNA (Fig 3B).
The specificity of HT1 was further investigated by mRNA-seq analysis in 293T cells. Only 48 genes were differentially expressed upon ectopic expression of HT1 (26 up-regulated and 22 down-regulated with padj < 0.1, shown in red in Fig 3C, upper panel), while knocking out CycT1 as a control impacted 1673 genes (shown in red in Fig 3C, lower panel). A third of the genes impacted by HT1 expression corresponded to up-regulated non-coding RNAs, including 7SK (fold-change = 1.1, padj = 9.6E-06), an effect that may be due to a stabilization of these RNAs.
Finally, the impact of HT1 on cell growth was assessed using three T cell lines (CEM, MOLT4, and MT4) infected by a lentivirus expressing a triple Flag-epitope tagged HT1 (3f:HT1). Polyclonal population of HT1-expressing cells (C-HT, MO-HT, and MT-HT, respectively) was selected by puromycin and confirmed by WB (Fig 3D). Total viable cell count over time showed no difference in cellular growth rate between HT1-expressing and control cells (Fig 3E). This confirmed that HT1 was specific to HIV inhibition, and that the few off-target effects had little impact on the metabolism of the cells.
In this study, we developed a new approach to block HIV transcription. We designed a chimera (HT1) containing the RNA-binding (ARM) and CDK9-inhibitory (ID) domains from the transcription regulator HEXIM1, and the P-TEFb-binding domain from the viral transactivator Tat (AD). Consistent with the respective properties of these domains in the context of their original proteins, HT1 competed with Tat for P-TEFb- and TAR-binding, and kept P-TEFb inactive. As a consequence, HT1 prevented Tat from bringing active P-TEFb to TAR for successful transcription elongation, as confirmed by the potent inhibition of HIV gene expression and replication. The use of a Tat-derived domain also conferred HT1 a high level of specificity, with little impact on host gene transcription and metabolism.
In this proof-of-concept study, diverse T-lymphocyte derived cell line models of HIV infection, including latent infection models, were preferred over primary CD4+ T cells in which production and selection of HT1-expressing cells would be challenging. We used plasmid or lentiviral delivery of HT1, which would require major technical adjustments to translate to a primary cell model. The very low amounts of transcription factors in these cells, including P-TEFb, would impair HT1 expression and inhibitory effect–even though the titrations in Fig 1B and Fig 1C suggest that effective HIV inhibition may be reached even at low HT1 expression levels. Another strategy would be to transduce the peptide into primary cells, which will also require optimization due to the oxidation-prone Cystein-rich Tat domain included in the chimera [33]. Although beyond the scope of the present study, determining the optimal conditions for successful delivery will thus be needed for any clinical application of HT1.
Despite this limitation, we validated a new strategic approach in HIV therapy: we used the fundamental knowledge in the structure and function of proteins involved in Tat-dependent HIV transcription for a logical design of an inhibitory peptide. ID of HEXIM1 contains the PYNT motif which is critical for inhibition of CDK9 kinase activity [20]. However, the C-terminal CycT1-binding domain (TBD) also contributes to the CDK9-inhibition by HEXIM1 [24]. Since the AD of Tat has a higher affinity to CycT1 than HEXIM1 [26], replacing HEXIM1's TBD with Tat AD was expected to make HT1 able to compete with HEXIM1 for P-TEFb binding. Adding HEXIM1 ARM to HT1 also made it able to compete with Tat ARM for TAR binding. The three domains were thus needed for HT1 to efficiently compete with HEXIM1 and Tat, so that it potently inhibits P-TEFb-induced HIV transcription elongation. In addition, spatial arrangement of these domains was critical for the inhibitory activity (S2 Fig), suggesting requirement for a precise placement of ID between the RNA- and P-TEFb-binding domains, as is naturally the case in HEXIM1. However, and as opposed as HEXIM1, HT1 did not need the C-terminal coiled-coil domain that is required for HEXIM1 dimerization and inhibitory activity [34]. Indeed, we showed that HT1 successfully competed with Tat in both Tat/TAR/P-TEFb and HEXIM1/7SK/P-TEFb complexes, and kept bound P-TEFb inactive. HT1 was especially efficient in competing for TAR binding (Fig 2D), which may be due to higher affinity for RNA through a longer and more basic ARM than Tat’s. Since an efficient Tat/TAR/P-TEFb interaction involves the Tat-TAR recognition motif of CycT1, HT1 competing for P-TEFb binding may also contribute to the competition for TAR-binding. Our study hence demonstrated that HT1 properties precisely match assumptions derived from each piece of biochemical data on Tat/TAR/P-TEFb and HEXIM1/7SK/P-TEFb interactions. Precise design can therefore make peptide therapy more specific, and thus better tolerated than the small molecules most often used in cART, as confirmed by the only minor off-target effects of HT1 with no impact on cell growth. Reactivation assays from two distinct HIV latency models showed that HT1 efficiently inhibited HIV reactivation by a broad range of latency reversal agents. This efficiency was further confirmed in spreading HIV infection assays, where HT1 not only competed with Tat, but prevented Tat to be expressed from the first round of transcription of newly integrated proviruses. This positive feedback loop of inhibition resulted in efficient inhibition of HIV replication after multiple rounds of infection.
Multiple therapeutic applications are possible for HT1, from prevention to therapy and cure. HT1 domains were derived from the human protein HEXIM1 and from HIV-1 Tat, neither of which are immunogenic, suggesting that clinical use of the chimera would be well tolerated. A key focus should be on investigating feasible cell delivery and route of administration. In an era of multiple and potent cART options, acceptability of a potentially injection-based treatment would mostly depend on the frequency and duration of administration, and may especially be fit for HIV cure application as opposed to long-term use as cART. Diverse options are now at hand in the fast evolving field of peptide therapeutics [35], including injection or alternative delivery routes such as oral or transdermal [36]. Gene therapy should also be considered, since we showed a positive inhibitory feedback loop in cells that stably expressed HT1 prior to HIV infection. In this model, HIV was virtually put into direct latency in newly infected cells, which in combination with blocking reactivation from pre-existing latently infected cells, could contribute to achieving a functional cure.
Finally, the design of such chimeras can be finely tuned to block other transcription factors that depend on P-TEFb. The CDK9-inhibiting module from HEXIM1 could indeed be combined with functional domains from transactivator targets other than Tat, allowing this strategy to be applied to other pathologies, including inflammation or cancer. This study thus paves the way to multiple applications of transcription-targeted inhibition peptide therapy.
Myc-epitope tagged HT1-3 (m:HT1-3) and Tat (m:Tat), and Flag-epitope tagged Tat (f:Tat) were inserted in the mammalian expression vector pEF-Bos [37], under the control of the EF1alpha promoter (plasmids referred to as pHT1-3 or pTat). For stable expression of HT1, triple Flag-epitope tagged HT1 (3f:HT1) was cloned in a 3rd generation lentivirus gene expression vector (VectorBuilder). In this vector (pLHT1), HT1 and BFP were cloned under the control of the EF1a promoter, and a puromycin-resistance gene was also cloned under the control of mPGK promoter.
Human Embryonic Kidney (HEK) 293T (ATCC) or 293T CycT1 knockout (KO) cells were grown in DMEM containing 10% Fetal Bovine Serum (FBS), at 37°C and 5% CO2. 293T CycT1 KO cells were created using the CRISPR/Cas9 system. All-in-one Cas9/sgRNA plasmid pSpCas9 BB-2A-Puro encoding sgRNA specific for CycT1 (sgRNA sequence: AAGCAGATTGGCCGCCTGC) (GenScript) was transfected into 293T cells using Lipofectamine 2000 (Invitrogen). After 48 hrs, untransfected cells were selected against by puromycin treatment for another 72 hrs. After puromycin selection, cells were cultured in normal media, and further cloned by limiting dilution. CycT1 KO clones were selected by analyzing CycT1 protein expression by WB using at least two different anti-CycT1 Abs (SCBT SC10750 and SC271348). Genomic CycT1 sequences were also analyzed to confirm that both alleles contained mutations that cause truncation of CycT1 protein at the N-terminus.
Human T-lymphocyte cell line CEM, MOLT4, MT4 cells (ATCC), and HIV-latently infected Jurkat cell clones J-Lat 9.2 [32] (given by Dr. Eric Verdin) or 2D10 [31] (given by Dr. Jonathan Karn) cells were grown in RPMI containing 10% FBS, at 37°C and 5% CO2. J-HT, L-HT and D-HT cells were produced by respectively infecting Jurkat, J-Lat 9.2 or 2D10 cells with a lentivirus (LHT1) produced by co-transfecting pLHT1 (see above), pMD2.G and psPAX2 (Addgene ID 12259 and 12260, respectively) in 293T cells using Polyplus transfection kit (Jetprime). 48 hrs after transfection, the supernatant from 293T cell culture was filtered, ultracentrifugated, and used for infection of CEM, MOLT4, MT4, J-Lat 9.2 or 2D10 cells (C-HT, MO-HT, MT-HT, L-HT and D-HT, respectively) in the presence of 2 μg/mL polybrene. C-HT, MO-HT, MT-HT, L-HT and D-HT cells were cultured in RPMI containing 10% FBS for 48 hrs, then puromycin (1 μg/mL) was added to the medium for antibiotic selection. Integration and stable expression of pLHT1 in C-HT, MO-HT, MT-HT, L-HT and D-HT cells was confirmed by BFP expression using FACS, and by WB using an anti-Flag Ab, as described below. Polyclonal populations of HT1-expressing cells were used throughout the study.
pHT1-3 and/or pTat and/or empty pEF.Bos vector (see above) and a Luciferase reporter construct (whether pLTR-Luc or pNL43ΔenvLuc, both described in [38]) were transfected using Lipofectamine 2000 (Invitrogen) in 2.0E+05 293T cells. In Fig 1B transfection amounts were 0, 50, 100 or 200 ng/mL pHT1-3, 100 ng/mL pTat and 50 ng/mL pLTR-Luc, qsp 500 ng/mL using empty vector pEF-Bos. In Fig 1C transfection amounts were 0, 50, 100, 200 or 400 ng/mL pHT1 and 50 ng/mL pNL43ΔenvLuc, qsp 500 ng/mL using empty vector pEF-Bos. After 48 hrs, the cells were washed using PBS, lysed with Passive Lysis Buffer (Promega), and analyzed for Luciferase expression using D-Luciferin (BD Monolight) on an EG&G Berthold LB 96V microplate luminometer.
For detection of transiently expressed proteins, a total of 2.5 μg DNA, including pHT1-3 and/or pTat and/or empty pEF.Bos vector (see above), was transfected using Lipofectamine 3000 (Invitrogen) in 8.0E+05 293T cells. After 48 hrs, the cells were washed using PBS, and lysed using RIPA buffer containing 150 mM KCl and a protease inhibitor mixture (Thermo Fisher Scientific). Lysates were incubated for 10 min at 95°C in 2X Laemmli buffer (BioRad) supplemented with 5% DTT, and used for SDS-PAGE analysis in a 15% resolving gel. The proteins were then transferred to a PVDF membrane (BioRad), which was blocked in 5% milk in TBS and probed with specific primary Abs, which include anti-Myc (ab32 from mouse and ab9106 from rabbit, Abcam), anti-Flag (F7425 from mouse and F3165 from rabbit, Sigma-Aldrich), anti-GAPDH (GA1R, MA5-15738, Invitrogen), anti-actin (ab8227, Abcam), anti-CycT1 (SC-10570, SCBT), anti-HEXIM1 (25388, Abcam), anti-CDK9 (SC-484, SCBT), anti-phospho CTD (Ser 2) (ab5095, Abcam). Membranes were washed five times and incubated with peroxidase-conjugated secondary Abs, which include ECL mouse IgG HRP-linked whole Ab (NA9310, GE Healthcare) and ECL rabbit IgG HRP-linked whole Ab (NA9340, GE Healthcare). After five washes, membranes were incubated in Western-Lightning Plus-ECL (Perkin Elmer) and visualized using the Odyssey Fc imaging system (Li-Cor).
For detection of stably expressed proteins, 5E+06 cells were washed using PBS and processed as described above.
pHT1 and/or pTat and/or empty pEF.Bos vector (see above), were transfected using Lipofectamine 3000 (Invitrogen) in 3.5E+06 293T cells. Transfection amounts were as follow: lane 1, 500 ng/mL pHT1 and 500ng mL pTat; lane 3, 500 ng/mL pHT1; lane 4, 500 ng/mL pTat; lanes 5–7: 500 ng/mL pTat and 250, 500 or 1,000 ng/mL pHT1; in all lanes, pEF-Bos qsp 1.5 μg/mL total DNA. After 48 hrs, the cells were washed using PBS, and lysed using RIPA buffer containing 150 mM KCl and a protease inhibitor mixture (Thermo Fisher Scientific). Lysates were precleared for 2 hrs at 4°C using protein G-sepharose beads (Life Technologyies, and the precleared lysates were incubated overnight at 4°C with 4 μg of the indicated primary Ab or control mouse IgG1 (MI10-102, Bethyl). Protein G-sepharose beads were added to the lysates and incubated for 2 hrs at 4°C, washed five times with RIPA buffer, resuspended in 2X Laemmli buffer (BioRad) supplemented with DTT and incubated for 10 min at 95°C. Supernatants were subjected to WB as described above.
A total of 30 μg pHT1 or pTat was transfected in 4E+06 293T cells using Lipofectamine 2000 (Invitrogen). 48 hrs after transfection, cells were lysed with buffer A (20 mM HEPES-KOH pH7.8, 0.3 M KCl, 0.1% Nonidet P-40 and 0.2mM EDTA), and m:HT1 or m:Tat proteins were immunoprecipitated by using Myc-Trap A kit (Chromotek). Immunoprecipitation with mouse IgG-coupled protein G sepharose was used as a negative control. After washing three times with buffer A, beads were washed once with CTD kinase buffer (20 mM HEPES-NaOH, 7.5 mM MgOAc, 2%Glycerol, 0.1 M KOAc, 2 mM DTT). Beads were then incubated with 30 μL CTD kinase buffer containing 0.25 μg of purified recombinant GST-CTD proteins (Sigma) and 50 mM ATP for 60 min at 37°C. Kinase reactions were terminated by adding 30 mL 2x Laemmli sample buffer (BioRad), and heating the mixture for 5 min at 95°C. Supernatants were subjected WB, and phosphorylated GST-CTD proteins (P-CTD) were detected by anti-phospho CTD (Ser 2, ab5094, Abcam). Co-immunoprecipitated P-TEFb (CDK9) was also detected by WB. Band intensities of P-CTD and CDK9 were quantified by LiCor imaging system, and relative kinase activities were calculated as P-CTD band intensities normalized by CDK9 band intensities. Six replicate experiments were performed.
pHT1 and/or pTat and/or empty pEF.Bos vector (see above) and pU16TAR [3] (a generous gift from Dr. John Rossi at City of Hope) for TAR RNA expression from an RNAPIII promoter, were transfected in 4.0E+06 293T cells using Lipofectamine 2000 (Invitrogen). In Fig 2C, transfected amounts were 500 ng/mL pU16TAR and 1.5 μg/mL pEF-Bos, pHT1 or pTat. In Fig 2D, transfected amounts were 500 ng/mL pU16TAR, 1.0 μg/mL pTat and 0.5, 1.0 or 1.5 μg/mL pHT1, plus pEF-Bos qsp 3 μg/mL total DNA. The cells were lysed in buffer A on ice for 20 min. Cell lysates were centrifuged at 14,000 rpm for 5 min at 4°C, and supernatants were collected. Cell lysates were then precleared with protein G-Sepharose beads (Invitrogen) and divided into two aliquots. Each aliquot was incubated with 1 μg of normal-rabbit IgG or anti-Myc (ab9106, Abcam) Abs precoupled with Protein-A dynabeads (Life Techonologies), or anti-Flag (M2)-conjugated magnetic beads (Sigma) for 2–4 hrs at 4°C. Beads were washed five times with buffer A. RNA was then extracted with TRIzol (Invitrogen), followed by DNase I treatment (Turbo DNAfree kit, Ambion). Reverse transcription quantitative PCR (RT-qPCR) analyses were performed using Superscript III First Strand synthesis system (Invitrogen) with 1 μL random hexamers (Invitrogen) and 1 μL RNaseOUT (Invitrogen), and then sensiFAST Lo Rox kit (Bioline), to quantify TAR RNA enriched in the immunoprecipitates. The same sets of qPCR analyses using the samples without reverse transcription confirmed that the DNA contamination from transfected plasmid reporters was negligible. Sequences for specific primers for TAR RNA or 7SK snRNA are as follows: TAR, 5’-CTTACTCTGTTCTCAGCGACA-3’ (forward) and 5’-CAACCTTCTGTACCAGCTTACT-3’ (reverse), 7SK,5'-GAGGGCGATCTGGCTGCGACAT-3' (forward) and 5'-ACATGGAGCGGTGAGGGAGGAA-3' (reverse) [29]. Known concentrations of the pU16TAR plasmids were used as standards to determine the copy number of TAR RNA by qPCR analysis. Data are shown by relative TAR enrichment by calculating values obtained with anti-Myc- or anti-Flag-immunoprecipitations divided by values obtained with IgG controls.
A total of 2.0 μg DNA, including pHT1 and/or empty pEF.Bos vector (see above), was transfected using Lipofectamine 3000 (Invitrogen) in 8.0E+05 293T cells. After 48 hrs, the cells were washed using PBS, lysed using TRIzol (Invitrogen), and RNA was purified using DirectZol RNA kit (zymo research) followed by DNase I treatment (Turbo DNA free Ambion). 500 ng RNA was used for reverse transcription using Superscript III First Strand synthesis system (Invitrogen) with 1 μL random hexamers (Invitrogen) and 1 μL RNAseOUT (Invitrogen). Resulting cDNA was then used for HEXIM1 and GAPDH qPCR using sensiFAST Lo Rox kit (Bioline) and the following primers: 5’-CACCAGCGATGACGACTT-3’ (forward) and 5’-TCATGTTCTGCAGGCTCT-3’ (reverse) for HEXIM1, 5’-ACCACAGTCCATGCCATCAC-3’ (forward) and 5’-TCCACCACCCTGTTGCTGTA-3’ (reverse) for GAPDH. Each condition was tested in triplicate experiments, and results are shown as mean HEXIM1 mRNA count relative to GAPDH, normalized to the mean relative count in the control condition (pHT1 = 0 ng/mL).
A total of 2.0 μg DNA (pHT1 or empty pEF.Bos vector, see above), was transfected using Lipofectamine 3000 (Invitrogen) in 8.0E+05 293T cells. After 48 hrs, transfected 293T cells and CycT1 KO 293T cells were washed using PBS, lysed using TRIzol (Invitrogen), and RNA was purified using DirectZol RNA kit (Zymo research) followed by DNase I treatment (Turbo DNA free Ambion). 500 ng RNA was used to prepare mRNA libraries using QuantSeq 3’ mRNA kit (Lexogen) for Illumina. High Sensitivity DNA kit (Agilent) was used for quality control of the libraries using a Bioanalyzer (Agilent). Libraries were sequenced using a HiSeq4000 (Illumina), and the reads were analyzed using the differential expression pipeline from BlueBee, using empty vector transfected 293T cells as a control for HT1-expressing 293T cells and for CycT1 KO 293T cells. Each condition was tested in triplicate experiments.
5.0E+05 2D10 or D-HT cells were treated with PMA (10 nM), SAHA (5 μM) or JQ1 (5 μM) and 5.0E+05 J-Lat 9.2 or J-HT cells were treated with PMA (100 nM). After 24h, the cells were washed with PBS containing 2% FBS, then fixed in 2% paraformaldehyde. A BD LSRII FACS analyzer was then used to determine the percentage of GFP-expressing cells as a proxy for HIV reactivation. Each condition was tested in triplicate experiments.
Wild type HIV-1 molecular clone pNL43 (given by Dr. Oliver Fackler, University of Heidelberg) was transfected in 293T cells using Polyplus transfection kit (Jetprime). 72 hrs after transfection, the supernatant from the 293T cell culture was filtered and used for infection of 1.0E+06 cells in the presence of 2 μg/mL polybrene. HIV infection was completed by spinoculation for 90min at 2,500 rpm. After 24 hrs, the cells were extensively washed in PBS and resuspended in 5 mL fresh medium. Day 0 supernatant (1mL) was collected before the cells were cultured further. The cells were passaged on Days 2 and 4, and 1mL supernatant was collected before each passage. HIV replication was measured by detection of Gag p24 in the supernatants by ELISA using HIV-1 p24 antigen capture assay (ABL). Each condition was tested in triplicate experiments.
Replication-defective HIV Env-pseudotyped HIV-1 was produced as described above by co-transfecting pNL43ΔenvLuc and pHXB2-env (given by Dr. Emilie Battivelli, Buck Institute for Research on Aging) in 293T cells and collecting supernatants. The same amount of virus-containing supernatant was mixed with cells (1.0E+06) in the presence of 2 μg/mL polybrene. Cells were further incubated overnight, washed once, and resuspended with fresh media. 24 hours later, cells were harvested and genomic DNA was isolated using QIAamp DNA Blood Mini kit (QIAGEN). 100ng of DNA was subjected to qPCR analysis with specific primers: 5’-ACCCTGAACTAGCAGACCAACT-3’ (forward) and 5’-ACACTAGGCAAAGGTGGCTT-3’ (reverse) for HIV (H9 and H10 in [39]) and 5’-TCAAGTGGGGCGATGCTGGC-3’ (forward) and 5’- TGGGGGCATCAGCAGAGGGG-3’ for genomic GAPDH. Each condition was tested in triplicate experiments and results are shown as mean HIV DNA count relative to GAPDH.
|
10.1371/journal.pbio.1001331 | The Core Apoptotic Executioner Proteins CED-3 and CED-4 Promote Initiation of Neuronal Regeneration in Caenorhabditis elegans | A critical accomplishment in the rapidly developing field of regenerative medicine will be the ability to foster repair of neurons severed by injury, disease, or microsurgery. In C. elegans, individual visualized axons can be laser-cut in vivo and neuronal responses to damage can be monitored to decipher genetic requirements for regeneration. With an initial interest in how local environments manage cellular debris, we performed femtosecond laser axotomies in genetic backgrounds lacking cell death gene activities. Unexpectedly, we found that the CED-3 caspase, well known as the core apoptotic cell death executioner, acts in early responses to neuronal injury to promote rapid regeneration of dissociated axons. In ced-3 mutants, initial regenerative outgrowth dynamics are impaired and axon repair through reconnection of the two dissociated ends is delayed. The CED-3 activator, CED-4/Apaf-1, similarly promotes regeneration, but the upstream regulators of apoptosis CED-9/Bcl2 and BH3-domain proteins EGL-1 and CED-13 are not essential. Thus, a novel regulatory mechanism must be utilized to activate core apoptotic proteins for neuronal repair. Since calcium plays a conserved modulatory role in regeneration, we hypothesized calcium might play a critical regulatory role in the CED-3/CED-4 repair pathway. We used the calcium reporter cameleon to track in vivo calcium fluxes in the axotomized neuron. We show that when the endoplasmic reticulum calcium-storing chaperone calreticulin, CRT-1, is deleted, both calcium dynamics and initial regenerative outgrowth are impaired. Genetic data suggest that CED-3, CED-4, and CRT-1 act in the same pathway to promote early events in regeneration and that CED-3 might act downstream of CRT-1, but upstream of the conserved DLK-1 kinase implicated in regeneration across species. This study documents reconstructive roles for proteins known to orchestrate apoptotic death and links previously unconnected observations in the vertebrate literature to suggest a similar pathway may be conserved in higher organisms.
| Clinical success in reconnecting neurons damaged by injury will require detailed molecular understanding of how mature axons respond to being severed. To decipher intrinsic molecular pathways that stimulate axon regeneration, we use the small transparent model, Caenorhabditis elegans, in which individual labeled axons can be laser-severed without damage to neighboring tissue, and regrowing axons can be observed directly in the living animal. We find that the apoptotic protein CED-3, well known for its developmental roles in cell death, also unexpectedly acts in a beneficial role to promote regeneration of severed axons. Initial post-surgery outgrowth is impaired in a ced-3 mutant, suggesting that CED-3 is involved in the early steps of axonal regeneration. The activation of CED-3 caspase in this context occurs independently of major cell death regulatory pathways, but efficient regeneration does require the caspase activator CED-4/Apaf-1, the conserved regeneration kinase DLK-1, and calreticulin-dependent calcium fluxes within the severed neuron. Our data suggest a novel conserved pathway for neuronal reconstruction, and call into question the practice of blocking caspases to treat neuronal injury in the clinic.
| In the injured vertebrate central nervous system (CNS), neurons often survive and sprout but encounter extrinsic and intrinsic barriers to functional regeneration [1], with devastating consequences for victims. The successful repair of neurons severed by accident or surgery is an obvious goal of modern regenerative medicine. A more detailed understanding of the fundamental molecular mechanisms of neuronal regeneration within a physiological context will be required for design of novel and effective therapies that could shift treatment goals from palliative care to restoration of function.
Considerable understanding of regeneration responses consequent to neuronal injury has been generated via study of vertebrate models in vivo and in vitro. More recently, laser technology advanced the precision of in vivo investigation to the single axon level by enabling the axotomy of individual processes in genetic model organisms [2],[3]. Moreover, the opportunity to test individual gene activities for roles in regeneration biology in whole animal context, and now to conduct high throughput genetic and pharmacological screens for such activities [4]–[6], is contributing to rapid advances in dissection of molecular mechanisms involved in neuronal regeneration. Although very much a work in progress, the emerging picture suggests regeneration may employ mechanisms conserved across species [5]. For example, in Caenorhabditis elegans, like in other models, physical disruption of an axon triggers an intracellular calcium spike [7],[8]. Calcium waves can originate from extracellular sources via voltage-gated calcium channels and may be amplified by release from internal stores. Elevation of calcium concentration activates signaling pathways, notably cAMP and MAPK DLK-1 pathways [8]–[10], which control growth cone formation and subsequent axonal elongation through cytoskeleton and membrane remodeling. Many details of the complex mechanisms involved remain to be established, the accomplishment of which might inspire strategies for directed neuronal repair.
With an initial interest in whether neurons might activate death pathways to eliminate the dissociated fragments generated by axon severing, we performed femtosecond laser microsurgeries on individual C. elegans neurons that lacked cell death proteins. To our surprise, we found that dissociated fragments often persisted for significant amounts of time. Moreover, CED-3 caspase, the essential core executioner protease in apoptosis [11], rather than being needed for cell fragment elimination, instead acts beneficially to promote early events in neuronal regeneration. ced-3 mutations affect early regenerative dynamics with the consequence of slowing initial outgrowth and delaying the physical reconnection of the regenerating axon to the severed distal segment, although ced-3 deficiency does not change long-term regeneration outcome. Core apoptotic proteins CED-3 and CED-4 are mobilized via a regulatory mechanism distinct from that involving known apoptotic regulators but which requires calcium flux and regeneration kinase dlk-1. Our data pull together disconnected observations in the literature to suggest that caspases act via a conserved mechanism to promote regenerative responses in injured neurons.
With an initial interest in whether neurons might activate death pathways in soma or dissociated fragments in response to severe physical injury such as axon severing, we performed femtosecond laser microsurgeries on individual GFP-visualized C. elegans neurons. We find that ALM mechanosensory neurons and D-type motor neurons rarely die after laser axotomy in adult C. elegans. Moreover, the severed dissociated processes generally persist for several days post-surgery (Figure S1) and can remain functional, as axotomized animals were touch-sensitive 6 h after surgery and remained so up to at least 1 wk post-surgery (see data note in Materials and Methods). As observed previously [2],[3],[9],[12],[13], severed processes display substantial regeneration from the soma-proximal side, with the severed stump regenerating a structure that first extends multiple spike-like filopodia and then directs further axonal extension (see Movie S1 for a typical depiction of wild type (WT) regeneration). At 24 h, roughly one third of axotomized ALM axons grew back to track along the severed distal process (see below), and the remaining axons displayed dramatic outgrowth with long and branched processes (Figure 1ai–iii, v). We also noted limited regrowth responses from the end of the severed soma-distal side (see below and Movies S1 and S2).
As one approach toward quantitation of the regeneration response, we measured total new outgrowth length of the proximal fragment 24 h following laser surgery for those neurons that did not regrow back into the original severed process (Figure 1aiii, v; those processes that did track back to the old distal process could not be measured as the new process could not be distinguished from the old persisting process). Somewhat unexpectedly, four independent mutants of ced-3 caspase, the central apoptosis executor protease required for all C. elegans programmed cell deaths [11], showed markedly reduced regenerative ALM outgrowth in this timeframe (Figure 1aiv, vi, 1b), a phenotype also exhibited by ced-3 D-type motor neurons (Figure 1c). ced-3 regenerative defects in severed ALM neurons diminished with time and were no longer apparent at 3 d post-surgery (Figure 1d). Most severe ALM deficits in ced-3 mutants occur in L4 larvae, although significant regeneration differences are apparent in young adults (Figure 1e). Notably, mutant phenotypes in the ced-3(n2433) active-site point mutant, which is deficient in in vitro protease activity [14], indicate that caspase activity itself is necessary for efficient axonal regeneration.
Axonal regeneration involves a complex interplay of biochemical activities within the injured neuron and interactions of the neuron with signals and structures in its environment. Thus, caspases might act directly in injured neurons, in the synaptic partners that provide guidance cues, or in the surrounding tissue (hypodermis for touch neurons) to set up conditions permissive for regeneration. To address whether CED-3 caspase activity is required within the severed neuron to facilitate regeneration, we expressed ced-3 in the mechanosensory neurons of the ced-3(n2433) mutant. Although like others [15] we found that expression of caspase transgenes is most often associated with cell toxicity, making the generation of transgenic lines extremely challenging, we identified one low copy number transgenic line with only moderate touch neuron loss (Figure S2). We found that the regeneration defect induced by ced-3(n2433) was rescued by specific expression of ced-3 in the mechanosensory neurons (Figure 1f), supporting that CED-3 acts in the damaged neuron for regeneration. Of note, moderate overexpression of ced-3 in wild-type neurons did not trigger enhanced regeneration (Figure 1f), which suggests that CED-3, though necessary, may not be sufficient to promote efficient regeneration. However, because it may be difficult to achieve CED-3 cellular expression levels that permit optimal repair rather than cell death (Figure S2), whether CED-3 might have the capacity to drive regeneration on its own remains unclear.
To evaluate ced-3 impact on regeneration in greater detail, we acquired time-lapse images of regrowing neurons for the first 5 h following laser axotomy. We accomplished this using nematode immobilization techniques that are stable over long time periods without the use of harsh anesthetics (see Materials and Methods and Figure S3) [16],[17]. We found that both the rate and extent of new outgrowth were dramatically reduced in ced-3 mutants during the initial 5 h following laser axotomy, with total outgrowth reduced by ∼45% and the average outgrowth rate reduced by 55% (Figure 2a). Higher resolution analysis of initial regenerative dynamics in WT and ced-3 mutants revealed three striking phenotypes in regenerating ced-3 neurons that impact the sprouting of short, often transient, exploratory filipodia-like processes that dominate during this early stage of outgrowth: (1) there is a significant delay in outgrowth onset after axotomy, with first signs of re-growth appearing after 91±13 min on average in ced-3 mutant axons compared to 43±8 min characteristic of WT axons (Figure 2b); (2) the number of sprouts initiated in ced-3 mutants is greatly diminished 0–5 h post-surgery, with the greatest effect observed during the initial 0–45 min (Figure 2c); and (3) ced-3 extensions often appear defective or stunted, resulting in short, wide, persistent bleb-like outgrowths that are distinctly different from the transient, dynamically active filopodia-like extensions of WT neurons (Figure 2d,e, Movies S2 and S3). These dramatic defects in the initiation of regrowth responses to axotomy in ced-3 contrast with overall outgrowth scores 3 d post-surgery, which no longer show differences from WT (Figure 1d). We conclude that the CED-3 apoptosis caspase impacts very early events in post-axotomy filipodia extension but is not essential to regrowth per se, suggesting that, like in other C. elegans regeneration studies [18], additional gene activities may act in parallel to promote regeneration.
ced-3 mutant neurons are not generally defective in developmental growth cone formation or guidance. In ced-3 mutants, we observed that developmental growth cones of migrating VD motor neurons in L1 larvae exhibit wild-type behaviors when they contact a new surrounding tissue: rounded in the hypodermis, and anvil-shaped when contacting the lateral nerve cord or body wall muscle cells (Figure S4) [19]. In addition, when we examined the AVM touch neuron projection to the ventral nerve cord (VNC) (a model for in vivo regenerative axon guidance [2]) by laser dissecting the AVM process half way to the VNC, we found that the ced-3(n2433) mutant shows the same ability to reach the ventral nerve cord 24 h post-surgery as the wild type (WT: 63%±9.3% reach the VNC, N = 27 [2]; ced-3(n2433): 64.3%±9.1%, N = 28, no statistical difference by t test). Together, these observations suggest that ced-3 defects in early filopodia extension dynamics and outgrowth might be limited to injury responses, although detailed quantification of developmental outgrowth and guidance needs to be accomplished before relative roles in development versus injury can be definitively assigned.
Interestingly, our high-resolution time lapse studies also revealed that the distal part of the axotomized axon, disconnected from the cell body, exhibited regrowth attempts by blebbing and extending exploratory processes initially similar in appearance to those in the proximal end (Figure 2d purple arrow, Movies S2 and S3). However, in the ced-3 mutant 0–5 h post-surgery, growth from the distal side of the laser cut was both delayed in onset (58±13 min in WT versus 111±22 min in ced-3 distal termini, p<0.05) and diminished in extent (1.8±0.2 exploratory processes in WT versus 1.3±0.2 in ced-3, p<0.05) (Figure S5). These initial regenerative responses of axon segments separated from the cell body must therefore be driven by ced-3 proteins or transcripts already present in the injured axon [20]. Thus, it appears that a nucleus-independent mechanism of CED-3 caspase activation lies in wait in healthy processes prior to injury.
In C. elegans, injured neurons can reconnect to reestablish the cytoplasmic connection of the proximal axon with the dissociated distal region of the axon [8],[13]. To generate a more complete picture of the consequences of ced-3 deficiency, we assayed regenerative capacities of those WT and ced-3 mutant neurons that tracked back to the dissociated process (i.e., those not counted for overall outgrowth due to coincidence of old and new processes) using a cytoplasmic reconnection assay. To score for reconnection, we adapted a fluorescence transfer protocol for use with GFP (see Materials and Methods for details) [21]. In our assay, we isolated a segment of the previously severed fragment by introducing a second cut more distal to the initial injury/potential reconnection site; we then selectively photo-bleached GFP within this distal segment (Figure 3a). Rapid recovery of GFP fluorescence within this segment revealed free diffusion of GFP from the non-photobleached regenerating proximal axon into the formerly severed fragment, and thus a re-established cytoplasmic connection.
To assay regeneration phenotypes of those neurons that regrew to come in proximity to the dissociated process, we compared WT and ced-3 mutant neurons for restored cytoplasmic continuity. We found that ced-3 mutant neurons were somewhat diminished in their capacity to rapidly track back to the dissociated fragment (Figure 3b), but, of the neurons that grew back to, and appear to be in contact with, the dissociated distal process at 12 h, 92%±8% of WT versus 20%±18% of ced-3 processes successfully reconnect (p<0.05 Fisher's exact test) (Figure 3c). When we sum data for all axotomies at 12 h post-surgery, 34%±8% of total WT ALM axons severed were reconnected at this time point, as compared to 4%±4% of ced-3 mutant axons (Figure 3d). As is true for the outgrowth phenotype, reconnection can approach WT levels after a significant time lag (Figure 3e). We conclude that a consequence of ced-3 caspase inactivation is delayed reconnection. Although the reconnection defect might be an indirect consequence of slow initial outgrowth, it is clear that CED-3 caspase deficiency impairs both initiation of axonal regeneration and reparative timing. In cultured Aplysia neurons, the time to reconnection can influence long-term function of the neuron [22], so the speed to reconnection might hold physiological relevance in invertebrate physiology.
A pressing question raised by the discovery of the role of CED-3 caspase in post-axotomy neuronal responses is whether other apoptotic pathway components modulate neuronal regeneration. During C. elegans developmental apoptosis, the expression of EGL-1 (BH3 domain only protein) inhibits CED-9 (Bcl-2 family member), releasing CED-4 (apoptosis protease activating factor-1 Apaf-1 homolog), which in turn activates CED-3 caspase [23]; CED-8 modulates the timing of developmental apoptosis [24]. Physiological germline apoptosis requires ced-9 transcription directed by the lin-35 Rb ortholog [25], and under conditions of radiation stress, both the C. elegans BH3-only domain proteins EGL-1 and CED-13 are needed for CED-3-dependent apoptosis [26]. To address how CED-3 caspase might be activated by axotomy, we tested roles of known apoptosis regulators in regeneration using the amount of 24 h outgrowth as a measure. We found that ced-4(n1162) and ced-4(n1416) mutants displayed diminished regeneration similar to ced-3(n2433), establishing that CED-4 functions in axonal regeneration as well as in apoptosis (Figure 4a). The double mutant ced-4(n1162); ced-3(n2433) is impacted to the same degree as either single mutant, suggesting that ced-3 and ced-4 work in the same pathway to influence regenerative outgrowth (Figure 4a). We also found that expression of our one minimally toxic ced-3 transgene in the touch neurons partially rescued the ced-4(n1162) defect, consistent with ced-3 acting downstream of ced-4 in axonal regeneration (the same as the order of CED-4 and CED-3 action in apoptosis) (Figure 4a). As with ced-3, regenerative defects of ced-4 mutant animals were no longer apparent after 3 d (Figure 4b). We conclude that ced-4 is needed for efficient regeneration and acts upstream in the same pathway as ced-3.
Other known upstream regulators of apoptosis, including loss-of-function (lf) allele ced-9(n2812), gain-of-function (gf) allele ced-9(n1950), egl-1 lf mutants egl-1(n1084n3082) and egl-1(n986), the egl-1; ced-13 double mutant lacking both C. elegans BH3-only domain proteins, and lin-35(n745), did not affect regeneration proficiency, revealing an alternative regulatory mechanism for CED-4 and CED-3 activation in the response to axotomy (Figure 4c). Likewise, because the ced-8(n1891) mutation did not impact regeneration, we conclude that the delayed regeneration response in ced-3 mutants is unlikely to be the consequence of timing-regulator ced-8 action in axonal regeneration. Overall, our data reveal an unexpected reconstructive role for the core apoptotic proteins CED-3 and CED-4 that is mobilized via a novel regulatory mechanism distinct from known apoptosis regulatory pathways.
The DLK-1 p38-like MAPK pathway has been shown to play a critical role in C. elegans neuronal regeneration [8]–[10]. Our detailed phenotypic analysis of ced-3 suggests action early in axonal regeneration, influencing initial exploratory sprouting (Figure 2), and similarly, the dlk-1 mutant has a drastic reduction in primary growth cone formation consequent to axotomy [9]. We therefore addressed whether DLK-1 might act together with CED-3 and CED-4 in the same molecular pathway, or alternatively, might act in parallel. Using our femtosecond laser and immobilization protocol, we find that the single mutant dlk-1(ju476) displays ALM regenerative outgrowth similar to that of ced-3 mutants, with a ∼50% reduction as measured at the 24-h time-point but wild-type regeneration proficiency at 3 d (Figure 5a,b). In the dlk-1(ju476) mutant background, weak regeneration of touch neurons severed in the adult contrasts with total block of regeneration of D-type motoneurons severed at the L4 larval stage, as we measured no regeneration outgrowth following axotomy in 22/22 D-type motoneurons (unpublished data) [8]–[10], underscoring that different molecular mechanisms might control regeneration in different cell types or developmental stages and, more specifically, that multiple redundant pathways may influence regeneration in adult ALM neurons. Interestingly, the double mutant dlk-1(ju476); ced-3(n2433) exhibited ALM regeneration impairment similar to that of single mutants, both at 24 h and at 3 d post-surgery (Figure 5a,b), suggesting action in the same pathway. Additionally, the double mutant dlk-1(ju476); ced-4(n1162) showed the same regeneration defect as the single mutants at 24 h (Figure 5a), further genetic evidence in support of action in the same pathway. Finally, expression of ced-3 in the touch neurons did not ameliorate regeneration deficiencies in dlk-1 mutants (Figure 5a), suggesting that ced-3 may act upstream of dlk-1 to promote early events in regeneration of ALM touch neurons in adult C. elegans.
Kinase KGB-1 of the (JNK) MAPK pathway has recently been shown to operate in parallel to DLK-1 to promote axon regeneration [18]. We find that although mutant kbg-1(um3) is defective in ALM regeneration, the double mutant kgb-1(um3) ced-3(n2433) is significantly more impaired in overall regrowth scores than either of the kgb-1 or ced-3 single mutants (Figure S6). Our data suggest that, similar to dlk-1, ced-3 acts in a separate regeneration pathway from kgb-1. Together, these studies define two parallel processes, one involving ced-4, ced-3, and dlk-1, and the other involving kgb-1, that act in ALM axon regeneration.
Calcium signaling is known to play a fundamental role in the neuronal responses to damage and subsequent recovery, with acute cellular insult inducing large intracellular calcium transients important for regrowth [7],[8]. To address whether calcium signaling could play a role in the CED-3/CED-4 molecular pathway during regeneration, we performed in vivo measurements of cytoplasmic calcium levels in the touch neuron cell soma during laser axotomy using two versions of the genetically encoded fluorophore cameleon (see Materials and Methods). Laser axotomy of WT neurons initiates an immediate (within <3 s) and dramatic increase of cellular calcium levels reported by cameleon-based FRET (Figure 6a). In two independent crt-1 mutants, which lack the ER calcium-binding chaperone calreticulin known to contribute to cellular calcium homeostasis, we found neuronal damage-induced calcium signals are reduced by ∼50% (Figure 6a, left panel). By contrast, no dramatic defect in calcium responses was detected in either ced-3(n2433) or ced-4(n1162) mutants as compared to WT (Figure 6a, right panel). Thus, ER calcium stores modulated by CRT-1 influence early calcium fluxes in response to axotomy, but CED-3 and CED-4 do not influence early calcium changes in the injured neuron.
To genetically address the requirement for calcium in regeneration, we tested the crt-1 mutants for total regenerative outgrowth and found a significant deficiency for both at the 24 h time point, which was no longer apparent after 3 d (Figure 6b,c). We noted that cytoplasmic expression of calcium-binding cameleon YC2.12, which might sequester some intracellular calcium, diminished overall outgrowth, consistent with a role for calcium in directing re-growth responses (Figure S7). By contrast, expression of cameleon YC3.60 that has a lower calcium binding affinity does not appear to affect regeneration (see Materials and Methods). Note that despite a dampening effect of cameleon YC2.12 on total regenerative outgrowth, the relative differences between WT and crt-1 mutants in regeneration of ALM mechanosensory neurons 24 h post-surgery were maintained in cameleon-expressing lines.
We examined early regeneration phenotypes in crt-1(bz29) using high resolution video analysis and found that, similar to ced-3, the first signs of re-growth in axotomized crt-1 mutant neurons appeared with a significant delay (Figure 6d) and that numbers of exploratory processes were highly reduced over both the 0–45 min and 0–5 h time periods post-axotomy (Figure 6e). Our combined genetic and imaging results implicate calcium changes that are activated by injury, and dependent upon calreticulin, in initiation of regeneration. Data follow recent findings that correlate reduced calcium transients resulting from nerve damage with diminished neuronal regeneration in C. elegans (see note in Materials and Methods on some experimental differences) [8].
To probe the relationship between crt-1 and ced-3 in regeneration, we compared regenerative capacity in the ced-3(n2433) single mutant, the crt-1(bz29) single mutant, and the ced-3(n2433); crt-1(bz29) double mutant. We find that regeneration deficits at 5 h and at 24 h in the double mutant were similar to those in single mutants, consistent with the possibility that CED-3 and CRT-1 act via the same pathway to influence initiation of regeneration (Figure 6b,d,e). We also found that expression of our one minimally toxic ced-3 transgene in the touch neurons partially rescued the crt-1(bz29) defect (Figure 6b), suggesting that CRT-1/calcium elevation might act upstream of CED-3 activation during axonal regeneration. This is in agreement with the calcium imaging data in Figure 6a showing a defect in calcium signaling in the crt-1 mutants but not in the ced-3 or ced-4 mutants. Finally, the double mutant dlk-1; crt-1 showed similar defects to the single mutants (Figure 6b), consistent with the action of dlk-1 in the same pathway as crt-1/ced-4/ced-3. Taken together, our data are consistent with a model in which crt-1 could act to influence intracellular calcium signals needed for CED-4-dependent localized CED-3 activation and efficient regeneration initiation promoted in part via kinase DLK-1.
Here we document novel roles of core apoptosis executors in the initiation of process regrowth in axotomized neurons. CED-3 caspase activity within the injured neuron promotes rapid remodeling and outgrowth, often resulting in efficient reconnection. C. elegans apoptosis executor CED-3 contributes to early regenerative events via a process genetically implicated to include CED-4 and calreticulin. The DLK-1 kinase might act downstream in the crt-1/ced-4/ced-3 pathway. The definition of reconstructive roles for the core apoptosis executor CED-3 holds implications for regenerative medicine strategies.
High resolution video microscopy time course studies during the first 5 h post-axotomy revealed that in wild type proximal processes, distinctive filopodia-like extensions can appear within minutes, leading to active growth cones and extensive outgrowth. In ced-3, these responses are slowed and often appear markedly defective—ced-3 processes take longer to initiate outgrowth, there are fewer filopodia generated, and there is less overall outgrowth. One particularly striking phenotype is that severed processes in ced-3 mutants can appear to produce extensions that do not mature into filopodia—instead, ends persist as rounded blebs that lack structure and do not extend (Movie S3, Figure 2e). Previous in vitro screens have identified C. elegans cytoskeletal proteins, such as actin, tubulin, and myosin chains, as potential CED-3 targets [27]; and caspases can cleave mammalian cytoskeletal proteins and their regulators [28],[29]. Thus, although critical targets in the regrowth mechanism remain to be identified, one possibility is that CED-3 activity might induce structural rearrangements needed for efficient filopodia production by cleaving cytoskeletal proteins.
Eventually, both total regenerative outgrowth and reconnection to the severed distal fragment reached WT levels at longer time points in ced-3 mutants, 3 d post-surgery (Figure 1d and Figure 3e). This outcome is consistent with a model in which the CED-3 caspase plays a role in the kinetics of a single regeneration pathway; alternatively, other pathways may run in parallel to promote regeneration and these other pathways may eventually compensate for ced-3 defects. Given the complex processes that influence regeneration in C. elegans [6],[18] and mammalian systems [30], and our genetic data that suggest kinase KGB-1 acts in parallel to caspase CED-3, the contribution of multiple pathways to regeneration seems like a probable scenario. The dramatic deficits in the initiation and early outgrowth dynamics suggest that CED-3 plays a prominent role during this critical stage of regeneration. Because dlk-1 is needed for early growth cone formation [9], it exhibits similar outgrowth defects to ced-3, the dlk-1 ced-3 double mutant shows similar regenerative outgrowth defects as single mutants, and elevated expression of ced-3 does not ameliorate the dlk-1 mutant deficit, we suggest that conserved kinase DLK-1 may be an integral downstream component of this early-acting mechanism.
Interestingly, our studies reveal that both the proximal process (remaining in contact with the nucleus) and the dissociated distal process (devoid of a nucleus) exhibit early regrowth efforts, generating dynamic filopodial extensions. Changes in the dissociated end have been noted in another C. elegans regeneration study [10]. As growth cones have been observed to extend from isolated processes in injured cultured vertebrate neurons [31]–[33], this phenomenon might represent another conserved element of the injury response. We find that in C. elegans, the regenerative response in the dissociated end is significantly diminished when ced-3 is lacking, and thus ced-3-dependent responses can occur independently of a nucleus and new transcription. These observations suggest that CED-3 protein might persist at low levels in an inactive form in healthy axons, evidence for which has been previously noted in touch neurons [15] and suggested for other non-apoptotic caspase paradigms [34]. Low basal level caspase activity can modulate motility in some cell types [35],[36] and might contribute to regeneration in this case. Alternatively, ced-3 transcript distributed throughout healthy processes might be translated at the injury site upon transection, as rapid local translation of other messages has been documented at injury sites in in vitro mammalian culture models [20] and in C. elegans [10] (including dlk-1). Regardless of activation strategy, it appears that C. elegans neurons can rapidly employ CED-3 activity when regenerative repair growth is needed.
Wild type regenerating C. elegans neurons are capable of rapidly locating and re-fusing with the dissociated distal process (34.4% successfully reconnected at 12 h; >50% reconnected by 72 h). We observed that ced-3 mutant processes are diminished in reconnection at the 12-h time point—fewer neurons overall reconnect (3.8% for ced-3 versus 34.4% for WT), and of neurons that do successfully track to the distal severed process, fewer ced-3 ends successfully reconnect (20.0% for ced-3 versus 91.7% for WT). Eventually, severed neurons do grow and reconnect in the ced-3 mutant background. Thus, ced-3 is not essential for reconnection, but rather plays a role in promoting rapid reconnection. The phenomenon of reconnection raises the question as to whether process breaking is a natural in vivo challenge in the development and/or function of neurons such that a protective mechanism of repair has evolved. Interestingly, severed Aplysia neurons also reconnect in culture, with failure to reconnect associated with electrophysiological dysfunction of the proximal neuron [22]. Rapid reconnection might thus be physiologically important for restoring or maintaining the function of the injured neuron.
We tested multiple regulators of C. elegans somatic or germline apoptosis for an effect on neuronal regeneration but find that regrowth is not influenced by CED-9/Bcl-2, BH3 domain proteins EGL-1 and CED-13, or the LIN-35 germline apoptosis regulator. Likewise, regeneration responses are not altered in a ced-8 mutant in which the progression through apoptosis is slowed. These data support that CED-3 caspase must be regulated by a novel mechanism that transpires independently of known apoptosis regulatory pathways.
The one other apoptosis protein needed for efficient regeneration is Apaf-1/CED-4, which our genetic analysis suggests acts upstream in the same pathway as ced-3. In apoptosis, CED-4 oligomerizes to form the apoptosome structure that facilitates procaspase cleavage [37]. It is possible that a similar reaction occurs in the regenerative response, although this process would likely need to be tightly regulated to prevent apoptosis (see below). ced-4 has been documented to execute some functions independently of ced-3 [38]–[40], and one instance of non-apoptotic cell death in neurons knocked down for mitochondrial coenzyme Q involves both CED-3 and CED-4 [41]. To our knowledge, however, our finding is the first report of CED-3 and CED-4 co-function in a pro-survival mechanism.
If CED-4 activates CED-3 caspase activity, a key question becomes how CED-4 might become proficient to do so consequent to axotomy. Interestingly, the CED-4 protein contains two regions that exhibit similarities to EF-hand calcium binding domains [42]. Our data and that of others [8] document local and transient elevation of calcium within the damaged neuron, and also show that limiting calcium signals from the ER or plasma membrane can diminish regeneration. Thus, one model for CED-3 activation in regeneration could be that calcium transients resulting from nerve damage, amplified by CRT-1, might locally activate CED-4, which in turn activates CED-3. Consistent with this model, we find that calcium dynamics in response to axotomy are disrupted in crt-1 mutants, but are normal in ced-4 and ced-3 mutants.
Once CED-3 becomes activated, its proteolytic functions must be tightly regulated to prevent apoptosis. Indeed, the need for a delicate balance is evident by the extreme difficulty and resulting cell death that we and others have encountered with introducing caspase transgenes, which most often kills cells (Figure S2) [15]. If maintained high calcium is needed for continued activation, local calcium transients initiated by membrane lesion might confer regulation; it is also possible that a mechanism exists for very low level basal level activity [35],[36]. Two elegant examples of localized caspase activation/regulation for developmental functions in Drosophila are the pruning of dendrites in the restructuring nervous system [43],[44] and the differentiation of spermatids [45]. Mammalian caspases have been shown to function in cell differentiation, cell migration, olfactory neuron development [46], and modulation of long-term depression in the brain [47]. Our findings on CED-3 and CED-4 roles in axon repair extend thinking on how proteins known to orchestrate apoptotic cell death can also contribute to pro-life functions [45].
CED-3 is the executor caspase for all C. elegans apoptosis, yet CED-3 clearly influences regenerative neuronal repair. The growth protein GAP-43 and the transcription factors p53 and c-jun can have dual roles in both promoting neuronal death and regeneration following axonal injury [48], raising the possibility that recruitment of cell death machinery in localized axonal regrowth might be a feature shared across phyla. Indeed, our data intersect with independent findings in culture models that suggest a mechanism similar to what we propose for C. elegans regeneration might influence regrowth in vertebrate neurons. Studies on in vitro vertebrate neuronal culture showed that caspase-3 is rapidly activated within 5 min of application of the guidance cues netrin-1 and LPA on growth cones to promote chemotropic responses [49] and that addition of caspase-3 inhibitors hinder growth cone formation after axotomy [20]. Moreover, calreticulin expression has been found to be dramatically induced in mammalian growth cones [50]. Together, these studies raise the possibility that localized deployment of caspases and calreticulin activity in axonal regeneration may be conserved in higher organisms. If caspases are found to promote mammalian axonal regeneration, regulated activation of caspase-promoted regrowth/reconnection might be used to promote functional repair in regenerative microsurgery or in injury therapy; at the same time, the use of anti-caspase therapeutics to limit neuronal loss following nerve damage [51] might be reconsidered.
Laser surgery was performed as described earlier [52]. A Ti:sapphire laser system (Cascade Laser, Eclipse Pulse Picker, KMLabs, Boulder, CO or Mantis PulseSwitch Laser systems Coherent Inc., Santa Clara, CA) generated a 1 kHz train of ∼100 fs pulses in the near infrared (∼800 nm). The beam, focused to a diffraction limited spot (using either a Nikon 100×, or 60×, 1.4 N.A. microscope objective), resulted in vaporization and tissue disruption with pulse energies ranging from 5–15 nJ. Visual inspection of the targeted neuron immediately following brief laser exposure (∼100–500 ms) confirmed successful axotomy. In some cases multiple laser exposures were necessary to generate a visual break in the nerve fiber. For 24 h regeneration measurements, C. elegans were temporarily anesthetized on 2% agar pads containing 3 mM sodium azide to allow for laser surgery, subsequently rescued, and then re-anesthetized 24 h later for imaging.
We targeted ALM axons 20 µm from the cell body unless otherwise stated and D-type motor neurons 20 µm up from the ventral nerve cord along the ventral-dorsal commissure. For length measurements, we calculated the total outgrowth of a neuron by summing lengths of the multiple outgrowth branches (for example, the green traces in Figure 1av), excluding very short branches (those <5 µm long). For ALM 24 h and 72 h regeneration outgrowth analysis (Figures 1, 4, 5, and 6 and Figures S6 and S7), regrowth of the proximal end only was monitored. The outgrowth of the distal end was measured only in data presented in Figure S4. The integrated lines zdIs5[pmec-4gfp] [53] and rnyIs014[pmec-4mCherry unc-119(+)], a gift from K. Nehrke, U. Rochester Medical School, were used for ALM surgeries in young adults. D-motor neuron surgeries were performed in L4 larvae using oxIs12[punc-47gfp] [54]. Note, at 72 h, outgrowth measurements in different genetic backgrounds (data in Figures 1d, 4b, 5b, and 6c) showed no statistical difference from one another by one-way ANOVA.
About 35% of WT axons reconnect within 24 h. For scoring of regenerative growth, we focused on instances in which we could get an accurate measurement of the total length and excluded these potential reconnection events from outgrowth scores, as in those cases we could not distinguish new growth from old process (the old process persists and does not lose GFP signal). Wild-type data were generated in six distinct groups, taken months apart, in strains ZB154 (zdIs5[pmec-4gfp]) and KWN177 (rnyIs014[pmec-4mCherry unc-119(+)]) (each group consisting of experiments run on the same or adjacent days using the same reagents). We found no significant difference between 24 h regenerative outgrowth of different WT groups and different transgenic markers (by one-way ANOVA). Wild-type data from all groups were therefore pooled together to give the wild-type measurement reported in the figures.
As measured in the time-lapse analysis (see below), the effect of the ced-3 mutation in the first 5 h is quite striking, featuring a deficit in exploratory processes, stunted sprout morphology not seen in WT regenerating neurons, and a general delay in response. Although conducting 5 h scoring in all genetic studies might have maximized phenotypic differences, the 24 h time point was used to evaluate relative regeneration in most genetic comparisons (most of these studies were conducted prior to the 5 h measurements that revealed early action of ced-3). Because statistically significant differences were still apparent at 24 h, and 5 h high resolution video microscopy required laborious analysis of individual movies, it would have been impractical to redo all genetic analyses for the 0–5 h time points.
Light touch to the anterior part of C. elegans body is sensed through a pair of ALM neurons and the AVM neuron. Interestingly, when both ALM axons were cut 20 µm from the cell body (n = 10) and AVM was ablated, we found that anterior touch was not significantly reduced, suggesting that the severed distal processes, in contact with post-synaptic interneurons to mediate the escape behavior, maintain the capacity for touch transmission. When we cut both ALM axons >200 µm from the cell body (just posterior to the nerve ring, where critical contacts to interneuron targets are concentrated), touch sensitivity was diminished. Although some axons from this axotomy distance seemed to have a directed regeneration, we did not find evidence of restored touch sensitivity even several days post-axotomy, consistent with previous reports [12].
Some of our scores of extent of regeneration defects differ quantitatively from some published studies. Differences may be attributed to a number of factors, including different anesthetic techniques, neuron type studied, age, and laser surgery technique [2],[8],[9],[12]. A few teams previously reported a delay of ∼10 h in the formation of the growth cone [2],[12]. However, using nematode immobilization techniques that do not require harsh anesthetics (microfluidic devices (Figure S3), as well as a 10% agarose preparation, see below), we observed no such delay: neurons often displayed initial growth within minutes of the laser damage with average initial growth time <1 h (Figure 2b). We observed more robust regeneration in ALM neurons compared to that of D-motor neurons, which may explain some of our differing results with published dlk-1 mutant strains (we consistently find reduction to ∼50% 24 h regrowth rather than no regrowth; minimal but non-zero regeneration in dlk-1 mutant strains has been reported in the PLM neurons in other studies [8],[10]). Our laser ablation technique utilizing a 1 kHz femtosecond pulse train at ∼800 nm is specifically designed to deliver precise ablation with minimal collateral damage to the animal and target neuron. Other techniques using MHz femtosecond pulse trains and conventional UV lasers produce larger regions of ablation and therefore more significant damage to the targeted neuron. Although some studies have indicated that postsurgical neuronal regeneration is unaffected by laser ablation technique [12], under certain conditions this may not hold true, leading to possible discrepancies in details among experiments. Despite these technical differences in the field and quantitative differences in extent of regeneration reported, basic conclusions have held across the field.
A significant proportion of axotomized neurons grew back to the dissociated fragment and could not be monitored for total outgrowth. To determine how ced-3 caspase disruption altered regeneration outcomes in this fraction of axotomized neurons, we scored for reconnection. ALM axons were severed 20 µm from the cell body in young adults using an zdIs5[pmec-4gfp] marker to visualize processes. 12 h, 24 h, or 72 h post-surgery, neurons were inspected by eye. Neurons for which the regenerative outgrowth of the proximal axon segment appeared to track to (i.e., be in close contact with) the dissociated distal segment (Figure 3a) were further assayed for reconnection using the following photo-bleaching experiment: (a) An initial image of the neuron was recorded (frame 1, see Figure 3a, panels iii and iv). (b) Using the laser, a second cut (yellow arrow in Figure 3a, panels v and vi) was made along the distal segment ∼40 µm from the initial cut point (red arrow) and ∼20 µm from any potential reconnection points. This effectively isolated the distal segment, where there is potential reconnection, from the rest of the process. This was important to prevent GFP refilling from the distal side. (c) The relevant segment (i.e., between the two cut points) was selectively bleached using standard high intensity UV illumination and a restricted illumination field. A second image was acquired immediately after bleaching (frame 2, see Figure 3a, panels v and vi). (d) After 15 min a third picture was acquired (frame 3, see Figure 3a, panels vii and viii).
GFP fluorescence level in each frame was measured as the average intensity along ∼15 µm of the nerve process starting at the second cut point (white brackets in Figure 3a, panels vii and viii), minus the background fluorescence measured adjacent to the nerve process (note, the same portion of the process was analyzed in each successive frame). Percent recovery was calculated as intensity increase between frames 3 and 2, relative to the intensity decrease between frames 1 and 2: percentage recovery = (if3–if2)/(if1–if2) (where if1 is the fluorescence intensity measured in frame 1, etc.). Over the short recovery time, recovery of GFP intensity indicates diffusion of non-bleached GFP into the isolated segment through a new connection point with the regenerating neuron (Figure 3a, panel vii). If there is no reconnection, the segment is truly isolated and GFP fluorescence does not recover (Figure 3a, panel viii). Control experiments, performed by severing the axon of an intact neuron twice and immediately photo-bleaching the isolated unconnected segment, gave an average “recovery” background after 15 min of 3.05%±0.55%. We therefore set a cutoff for successful reconnection at >7.37% recovery (2 sigma from the control average). The percent of reconnection at 12 h as well as additional measurements are given in Figure 3b–d.
Time-lapse movies following laser surgery were acquired using two methods of worm immobilization: (1) microfluidic devices, the design, fabrication, and use of which followed previously described methods [17], and (2) a preparation of stiff 10% agarose pads and polystyrene microspheres as described earlier [16]. Laser surgery was performed by manual alignment, but subsequent imaging was computer-automated to allow simultaneous time-lapse imaging of up to 10 regenerating neurons in separate C. elegans. Initially movies were generated using microfluidic devices at lower resolution (30 min/frame, ×40 magnification). These data were eventually pooled with that from higher resolution movies (see below) to generate the time-lapse outgrowth data shown in Figure 2a (n = 43 for WT, n = 40 for ced-3). For all movies, outgrowth in each frame was measured as the contour length along the new axon growth, with branches <1 µm long excluded. At each time point, mean outgrowth values were calculated across all regenerating neurons of that strain type. Regression fits to the data displayed in Figure 2a (by the least squared error method, KaleidaGraph, Synergy Software, and restricted to pass through the origin) were used to generate the outgrowth rates displayed in the insert. These rates therefore measure the average total outgrowth of the neurons, which at this stage is largely dominated by the creation and retraction of numerous filopodial extensions rather than the elongation of an individual branch.
To generate an accurate account of the initial regenerative dynamics (the number and timing of exploratory processes displayed in Figure 2b–e, Figure 6d,e and Movies S2–S3), higher resolution movies (10 min/frame or 15 min/frame, ×60 magnification) were generated in two ways. Microfluidic devices were used as described above, with the addition of 0.05% tetramisole in the surrounding buffer. The tetramisole worked to partially paralyze the worms [19] in order to keep them still enough for automated re-imaging under high magnification for long time periods. Worms were also immobilized for imaging without anesthetics, using stiff 10% agarose pads and polystyrene microspheres [16]. Data were collected for 5 h post-surgery and images were analyzed by eye (counting number and timing of exploratory processes). Figure 2b,c and Figure 6d,e show the results of data pooled together from the two preparations, as no statistical difference was found between results from the microfluidic devices and stiff agarose protocols.
We quantified calcium dynamics as changes in ratiometric fluorescence emission between the cyan and yellow fluorescent protein components of cameleon, in the same manner as described previously [55],[56]. Two versions of cameleon were employed: YC2.12 [57] and YC3.60 [56],[58]. For measurements within the crt-1 mutants we used the bzIs17[pmec-4YC2.12+lin-15(+)] allele expressing cameleon YC2.12 from the mec-4 promoter [57]. Because of apparent close linkage between the ced-3 and the bzIs17[pmec-4YC2.12+lin-15(+)] allele, we used a second allele expressing cameleon YC3.60 under the mec-4 promoter, bzIs158[pmec-4YC3.60], for measurements in the ced-3 and in the ced-4 mutant backgrounds. Images were taken every 3 s with a 300 ms exposure time. The response of an individual neuron was measured as an integration of the fluorescence signal across the entire cell soma. For the YC2.12 measurements, animals were immobilized on a 2% agar pad containing 0.05% tetramisole. For the YC3.60 measurements, the 10% agarose preparation, described above, was used. Differences in the wild type calcium response between YC2.12- and YC3.60-expressing strains could be due to a number of factors including the larger dynamic range and lower calcium affinity of YC3.60, and the different worm immobilization techniques. For these reasons we compared calcium measurements only across genetic backgrounds expressing the same cameleon variant. Likewise, our measured intracellular calcium signals differ with that of others [8] due to a number of possibilities including differing neuron type, calcium reporter, position of cut relative to the cell body, and the portion of cell analyzed. Strains expressing cameleon YC2.12 displayed a deficit in regeneration compared to non-cameleon strains at the 24 h time point (Figure S7). Although we observed a general reduction in overall regenerative outgrowth for all strains expressing the calcium-binding cameleon YC2.12, the ∼50% relative reduction in outgrowth compared to WT control is maintained in the crt-1 mutant in the presence or absence of cameleon YC2.12, so basic conclusions on the requirement for crt-1 are not compromised by the use of the cameleon YC2.12 reporter (Figure S7). The WT strain expressing cameleon YC3.60 showed no significant defect in regenerative outgrowth at the 5 h time point.
Details of statistical analysis are stated in the figure legends. In general, for comparisons between two measurements a two-tailed Student's t test was used to show statistical significance (direct t tests are indicated by brackets where they are not otherwise obvious). For group comparisons involving multiple strains (i.e., all strains within one figure panel unless otherwise indicated) the Dunn-Sidak group comparison method was used. Statistical tests were implemented using MATLAB (The MathWorks, Inc.). Outgrowth rates in Figure 2a insert were calculated by regression fits to the data as described above.
Strains were grown at 20°C on NGM agar seeded with Escherichia coli OP50 as a food source [59]. The wild type strain was C. elegans N2 Bristol. Standard genetic techniques were used to generate compound mutant strains. The active site point mutation allele ced-3(n2433) was used in all compound mutant strains (see Table 1).
The true ced-3 null allele has not been formally defined, although many loss-of-function mutants have been described in detail [60]. All four ced-3 alleles studied are strong loss-of-function. The n2433 allele encodes a point mutation that alters the caspase active site and shows weak semi-dominance regarding apoptosis; the encoded substitution generates a mutant CED-3 that has no detectable protease activity in vitro [14]. We also studied regeneration in ced-3(n2452) (a 17 Kb deletion also disrupting four other putative genes: C48D1.1, F58D2.2, F58D2.4, and F58D2.1), ced-3(n717) (mutation of the conserved acceptor site of intron #7), and ced-3(n2888) (early stop codon).
The crt-1(ok948) deletion mutant deletes all but the first 21 amino acids, including the stop codon. crt-1(bz29) encodes a stop codon at position 28 and lacks immunoreactivity [61]. These crt-1 alleles have been suggested to be functional null alleles.
The dlk-1(ju476) allele is a 5 bp insertion at G631 [62]; this allele has been cited to act as a null allele for axonal regeneration [9].
Plasmids were constructed using standard genetic techniques. The pmec-4mCherry vector was constructed by amplification of mCherry sequence improved for expression in C. elegans [63] using the following primers: 5′-GGGATCCATGGTCTCAAAGGGTGAAGA-3′ and 5′-GGAATTCTTATACAATTCATCCATGCC-3′. The PCR fragment generated was cloned into pmec-4GFP [64], replacing GFP using BamHI and EcoRI sites.
For the construction of pmec-4ced-3, ced-3 cDNA was amplified from a pool of C. elegans cDNA using primers 5′-GGATCCATGATGCGTCAAGATAGAAGGA-3′ and 5′-CAATTGTTAGACGGCAGAGTTTCGTGC-3′ and cloned into pCR2.1 using TOPO TA cloning kit (Invitrogen). For further cloning purposes, the HindIII site of ced-3 cDNA was inactivated while introducing the silent mutation A to G at position 609 on the cDNA giving pCRced-3(A609G) (QuikChange II Site-Directed Mutagenesis Kit). The GFP fragment of pmec-4GFP [64] was replaced with ced-3(A609G) from pCRced-3(A609G) using BamHI and MfeI sites. A fragment containing the mec-4 promoter fused to ced-3(A609G) cDNA from the previous vector was introduced using ApaI and HindIII sites into pDP#MM016b bearing unc-119(+) [65] and giving the pmec-4ced-3 vector construction.
The pmec-4ced-3 vector and the pDP#MM016b [65] vector bearing unc-119 gene were used for bombarding unc-119(ed3) animals as described [66]. Generated transgenic lines were bzIs122[pmec-4ced-3 unc-119(+)] and bzIs123[unc-119(+)], named Is[pmec-4ced-3] and Is[unc-119(+)], respectively, in the figures presented for this study. Strains were outcrossed once before further genetic constructions. Note that the line generated exhibited evidence of some touch neuron loss (Figure S2) and that numerous repeated attempts at generation of transgenic expression of C. elegans caspase genes were unsuccessful. This is likely due to the toxicity of elevated ced-3 expression. Note that although we obtained published lines overexpressing dlk-1 on extrachromosomal arrays, transgenic lines were consistently sick and array transgenes were lost at a very high frequency, precluding our ability to test dlk-1 overexpression in ced-3 mutants.
|
10.1371/journal.ppat.1005576 | Transformed Recombinant Enrichment Profiling Rapidly Identifies HMW1 as an Intracellular Invasion Locus in Haemophilus influenzae | Many bacterial species actively take up and recombine homologous DNA into their genomes, called natural competence, a trait that offers a means to identify the genetic basis of naturally occurring phenotypic variation. Here, we describe “transformed recombinant enrichment profiling” (TREP), in which natural transformation is used to generate complex pools of recombinants, phenotypic selection is used to enrich for specific recombinants, and deep sequencing is used to survey for the genetic variation responsible. We applied TREP to investigate the genetic architecture of intracellular invasion by the human pathogen Haemophilus influenzae, a trait implicated in persistence during chronic infection. TREP identified the HMW1 adhesin as a crucial factor. Natural transformation of the hmw1 operon from a clinical isolate (86-028NP) into a laboratory isolate that lacks it (Rd KW20) resulted in ~1,000-fold increased invasion into airway epithelial cells. When a distinct recipient (Hi375, already possessing hmw1 and its paralog hmw2) was transformed by the same donor, allelic replacement of hmw2AHi375 by hmw1A86-028NP resulted in a ~100-fold increased intracellular invasion rate. The specific role of hmw1A86-028NP was confirmed by mutant and western blot analyses. Bacterial self-aggregation and adherence to airway cells were also increased in recombinants, suggesting that the high invasiveness induced by hmw1A86-028NP might be a consequence of these phenotypes. However, immunofluorescence results found that intracellular hmw1A86-028NP bacteria likely invaded as groups, instead of as individual bacterial cells, indicating an emergent invasion-specific consequence of hmw1A-mediated self-aggregation.
| Many bacteria are naturally competent, actively taking up DNA from their surroundings and incorporating it into their genomes by homologous recombination. This cellular process has had a large impact on the evolution of these species, for example by enabling pathogens to acquire virulence factors and antibiotic resistances from their relatives. But natural competence can also be exploited by researchers to identify the underlying genetic variation responsible for naturally varying phenotypic traits, similar to how eukaryotic geneticists use meiotic recombination during sexual reproduction to create genetically admixed populations. Here we exploited natural competence, phenotypic selection, and deep sequencing to rapidly identify the hmw1 locus as a major contributor to intracellular invasion of airway epithelial cells by the human pathogen Haemophilus influenzae, a trait that likely allows bacterial cells to evade the immune system and therapeutic interventions during chronic infections. Genetic variation in this locus can strongly modulate bacterial intracellular invasion rates, and possession of a certain allele favors adhesion and self-aggregation, which appear to prompt bacteria to invade airway cells as groups, rather than as individuals. Overall, our findings indicate that targeting HMW1 could block the ability of H. influenzae to invade airway cells, which would make antibiotic therapy to treat chronic lung infections more effective. Furthermore, our new approach to identifying the genetic basis of natural phenotypic variation is applicable to a wide-range of phenotypically selectable traits within the widely distributed naturally competent bacterial species, including pathogenesis traits in many human pathogens.
| Genetic mapping in bacteria historically relied on screening mutant libraries for loss-of-function mutations, followed by laborious isolation and identification of the disrupted loci. Recent innovations in mutagenesis approaches like TnSeq can accelerate the process and aid in characterizing gene function (e.g. [1,2]), yet such approaches have some limitations. For example: (a) many classes of genetic variation are not evaluated, (b) a suitable loss-of-function screen is typically required, and (c) such techniques ignore naturally occurring within-species phenotypic variation. An alternative is to emulate eukaryotic quantitative genetics approaches, which rely on sexual reproduction to map genetic variation. Rather than isolating the loci responsible for a specific phenotype with disruptive mutations, the QTL (quantitative trait locus) mapping approach identifies the loci and alleles that are directly relevant to phenotypic expression in natural populations.
Bacteria do not reproduce sexually, but genetic transfer mechanisms are widespread, and diverse bacterial species (including many important human pathogens) are naturally competent, able to actively take up and recombine homologous DNA from their surroundings into their chromosomes [3,4]. The value of this genetic transfer mechanism to researchers was seen as early as 1944 by Avery et al., when naturally competent Streptococcus pneumoniae were used to show that DNA is the genetic material, or “the transforming principle” [5]. But only recently has exploiting natural competence (and other gene transfer mechanisms) become a practical means to investigate the genetic basis for natural phenotypic variation, as massively parallel sequencing technologies have become cost effective [6–9].
The Gram-negative bacterium Haemophilus influenzae has a well-characterized natural competence mechanism (reviewed in [4]) and illustrates how experimental natural transformation can be useful for genetic mapping. Under nutrient limitation, cells actively take up double-stranded DNA from their environment through their cell envelope, and this DNA can replace homologous segments of the chromosome by recombination. In the laboratory, a competent H. influenzae cell will rapidly replace ~0.1–3% of its chromosome with genomic DNA from a divergent H. influenzae strain [8,10]. These replacements typically involve multiple independent recombination tracts and contain 100s to 1,000s of donor-specific single-nucleotide polymorphisms (SNPs) that span dozens of genes. Insertions and deletions readily transform as parts of longer recombination tracts, albeit with less efficiency than SNPs; this can add or remove whole genes and operons. Thus, a single transformation experiment can give millions of independently transformed recombinants containing all (or nearly all) of the genetic variation that distinguishes the donor and recipient strains [8]. Such pools can be screened or selected for donor phenotypes, and the donor-specific variation in isolated recombinants can be identified by DNA sequencing. A previous small-scale screen of <100 transformed recombinant clones identified a bacterial QTL with a ~10-fold effect on transformability itself, though this used laborious quantitative assays of individual clones that had already been sequenced [8].
Though typically a commensal of the nasopharynx, H. influenzae—especially in nonencapsulated or nontypeable forms (NTHi)—can cause middle ear infections (otitis media), community-acquired pneumonia, exacerbations of chronic obstructive pulmonary disease (COPD), conjunctivitis, and sometimes more severe invasive diseases [11]. Infections often persist and recur despite host production of bactericidal antibodies and the use of antibiotic therapy. Our understanding of the molecular mechanisms involved in the progression and persistence of H. influenzae infections remains limited, but identical strains have been repeatedly isolated from the lungs of COPD patients in serial clinic visits, suggesting that H. influenzae has traits that promote chronic infection [12,13].
Current evidence indicates that H. influenzae is a facultative intracellular pathogen, and host cell invasion may allow bacterial cells to temporarily evade the immune system and therapeutic interventions [14,15]. H. influenzae invades a variety of cell types [16–20], and viable NTHi have been found within host cells of adenoid tissues and bronchial biopsies [21,22]. After intracellular invasion, H. influenzae cells remains non-proliferative and resides within membrane-bound vacuoles with features of late endosomes [23,24] or freely within the cytoplasm [25], and intracellular bacteria eventually die after persisting for variable lengths of time [26]. While several host factors have been identified as important for intracellular invasion, much less is known about the bacterial factors responsible for this process.
We chose the intracellular invasion phenotype as a model to test the genetic mapping strategy described above, which we have named “transformed recombinant enrichment profiling”, or TREP (summarized in Fig 1). We chose this phenotype because entry of H. influenzae into airways cells: (a) is easily selectable in lab culture, (b) displays wide phenotypic variation between clinical isolates, and (c) is likely to be an important factor in chronic infections. Application of TREP rapidly identified the hmw1 adhesin as a factor crucial for intracellular invasion.
Three criteria were used to choose the donor and recipient strains: (a) substantially higher invasiveness of the donor over the recipients, (b) high natural transformability of the recipients, and (c) available genome references with many genetic markers distinguishing the donor from the recipients (S1 Text and S1 Fig). Three strains were assayed: the standard laboratory strain (Rd KW20, hereafter referred to as Rd) and two clinical isolates from pediatric patients with otitis media (Hi375 and 86-028NP). All three have complete genome sequences available [27–29]. Rd and Hi375 are highly transformable, while 86-028NP is not [30,31]. Rd is known to be a poor invader of several epithelial cell lines [32]; Hi375 has previously been used in studies of intracellular invasion [15,24]; and 86-028NP is known to be highly virulent in a chinchilla model of otitis media [33–36]. Antibiotic resistant derivatives of all three strains were produced to allow subsequent tracking of transformation events and genetic background, yielding strains Rd SpcR, Hi375 StrR, and 86-028NP NovR NalR, hereafter referred to as RdS, HiT, and NpNN, respectively (S1 Table).
Intracellular invasion frequencies were evaluated by gentamicin protection assays with A549 airway epithelial cells, and quantified as gentamicin-protected bacterial colony forming units (CFU) relative to the original inoculated CFU (hereafter “Invaders/CFU”). Gentamicin protection assays found that NpNN is a highly efficient invader of A549 cells with ~10−2 invaders/CFU, whereas both HiT and RdS invade at ~100-fold and ~1,000-fold lower frequencies, respectively (Fig 2A; one-way ANOVA p<<0.01, and p<<0.01 for all three pairwise comparisons by post hoc testing using Tukey’s HSD).
Several controls were performed, both for assays of the three parental strains and during the serial passage experiments described below: (a) To ensure that spontaneous gentamicin resistance was not responsible for survival in gentamicin protection assays, equivalent bacterial suspensions were treated in the absence of A549 cells, yielding no viable CFUs (limit of detection <10−8), and gentamicin-protected CFUs recovered as colonies on plates remained gentamicin sensitive. (b) To ensure gentamicin treatment was complete, culture supernatants of infected A549 cells treated with gentamicin were plated, rendering no viable CFUs (limit of detection <10−8). (c) To ensure that introduction of selectable markers had no effect on intracellular invasion, these strains were compared to their progenitors, finding that they had comparable phenotypes (S2A Fig, one-way ANOVA p-values > 0.1 within each strain background).
Adhesiveness of the three parent strains to A549 cells was assayed similarly to invasiveness, except that the gentamicin treatment was omitted; “Adherents/CFU” was calculated as the CFU that remained associated with A549 cells after incubation and washing, relative to CFU of the input inoculum. The NpNN strain was the most adherent, with ~10% of the infecting cells remaining associated with A549 cells (Fig 2B). The HiT strain was intermediate (~10-fold lower than NpNN), whereas RdS had ~100-fold lower adherence (p<<0.01 by ANOVA and all for three pairwise comparisons). As with invasion, antibiotic resistant derivatives had adhesiveness comparable to that of the progenitors (S2B Fig).
A calibration experiment was used to test the strength of the experimental selection applied by the gentamicin protection assay. Mixtures of a high invasion strain (86-028NP NovR) and a low invasion strain (Rd StrR) were passaged twice: Bacterial cells were used to infect A549 cells and invaders were recovered after gentamicin treatment. The recovered colonies were then pooled and used for a second infection. This found that the highly invasive strain easily out-competed the poorly invasive one, even when starting at a 1 to 10,000 disadvantage, dominating the population after the second infection (S3 Fig). The high enrichment was not due to a growth rate advantage of 86-028NP, which has a slower doubling time than both recipients [30]. These data demonstrate that even rare recombinants could be highly enriched from complex pools using serial selection.
The experimental design for isolating bacterial intracellular invasion genes by TREP is depicted in Fig 1, summarized below, and described in detail in subsequent sections. (a) Donor genomic DNA from NpNN was used to transform naturally competent cells of two low invasion strains, RdS or HiT. (b) Pools of ~105 recombinant clones were enriched for those that conferred increased invasiveness by serial passages through A549 cells by gentamicin protection. Material from each cycle was stored to use for replicate quantitative assays and DNA extractions. (c) Genomic DNA from pools was sequenced to high coverage, and donor-specific allele frequencies were calculated at diagnostic SNPs. Sequencing and alignment statistics are summarized in S2, S3 and S4 Tables, and further details are presented in S1 Text and Materials and Methods.
By profiling genome-wide donor allele frequencies from independent experiments, a gene responsible for high invasion by NpNN was rapidly identified in both recipients, RdS or HiT. Individual clones were isolated after the fourth cycle of selection for genome sequencing and phenotypic analysis of recombinants and mutant derivatives. The results reveal a novel role for the adhesin-encoding hmw1A gene in intracellular invasion, beyond its previously described role in adhesion [37,38].
Recombinant pools were made by incubating high molecular weight genomic DNA from NpNN with naturally competent cultures of either RdS or HiT (Fig 1A) [39]. Transformation by the two antibiotic resistance markers (NovR and NalR) indicated that ~15% of cells in each culture were competent and predicted that a given donor-specific SNP would be found in ~1% of NovR or NalR colonies, consistent with previously measured values (S5 Table) [8]. Selection for donor-specific antibiotic resistance alleles ensured elimination of untransformed recipient cells, and selection for recipient-specific resistances limited cross-contamination observed in preliminary experiments (Materials and Methods). This procedure generated four separate sets of recombinant clones (for the RdS recipient: SpcR NovR and SpcR NalR; for the HiT recipient: StrR NovR and StrR NalR). For each set, ~105 colonies were harvested into pools, thoroughly mixed, and stored prior to infections. We thus predicted that each pool would have millions of cells for which (nearly) any given donor-specific variant would be present. Sequencing of the initial recombinant pools provided several lines of evidence that the TREP approach would be practical (S1 Text): First, a single round of selection for the donor-specific antibiotic resistance alleles was sufficient to map the resistances to single nucleotide resolution (S4A and S4B Fig and S6 Table). Second, recombinant pools contained low frequency donor-specific alleles across the genome, both SNPs and and structural variants (S4C Fig and S7 Table).
The four initial recombinant pools (Pool 0s) were used in separate infections of A549 cells (Fig 1B). Serial passages of these pools using gentamicin protection resulted in >100-fold increased invasion after only three cycles of enrichment (S5 Fig). Replicate invasion assays using material from the initial enrichments confirmed the dramatic increase in intracellular invasion by pools after serial enrichment (Fig 3A, one-way ANOVA p<<0.001, Tukey’s HSD for comparisons of Pool 0 to Pools 2–4 were p<<0.001, but p>0.2 among Pools 2–4). Comparison with donor invasiveness values measured in parallel showed that the HiT recombinants in the two Pool 4s were not significantly less invasive than the donor (p = 0.082); the two Pool 4s using the RdS recipient remained marginally less invasive (p = 0.025). These data suggest that after the third serial enrichment (Pool 4), donor-specific genetic variants conferring invasiveness were at or near fixation.
Increased invasion was not due to selection for de novo mutations that confer increased invasiveness, nor due to changes in bacterial gentamicin sensitivity (controls detailed above). Such events were unlikely given that the number of cell generations across the experiment was <100 in total. Furthermore, a control experiment using untransformed RdS or HiT cultures found no significant increases in invasiveness over 5 serial cycles of selection with either recipient strain (S6 Fig; p>0.1 by one-way ANOVA), strongly suggesting that de novo mutations conferring increased invasiveness were not captured in these experiments.
Since higher adhesion might concomitantly increase intracellular invasion, we tested how adhesion was affected by the serial selections (Fig 3B). This found that the RdS pools showed substantial progressive increases in adhesiveness (one-way ANOVA p<<0.001, Tukey’s HSD p<<0.001 for comparisons of Pool 0 to Pools 2–4, but p > 0.2 for comparisons among Pools 2–4). The HiT pools trended towards increasing adhesion over serial enrichments, but for this experiment, no significant change was observed (p>0.1). Both pairs of pools still had significantly lower adhesiveness than NpNN cultures run in parallel (p<0.05 for all comparisons against NpNN).
In sum, RdS and HiT recombinants acquired loci or alleles that enhanced intracellular invasion. Adhesion increased, though to a lesser extent. This suggests that, while adhesion might be a prerequisite for invasion, its increase may be insufficient to explain the elevated invasion displayed by invasiveness-enriched recombinants.
Sequencing of genomic DNA across pools and serial passages (Fig 1C) showed that the four recombinant pools became progressively less complex, ultimately resulting in a total of only six recombinant clones dominating the four pools (out of ~4x105 total). This suggests that the causative alleles transformed competent cells at lower rates than typical SNPs (~1,000-fold; S5 Table). The change in complexity was particularly apparent at the antibiotic-selected sites, where donor allele frequencies shifted from a smooth decline on either side of the resistance alleles to sharply demarcated donor segments (stretches of contiguous donor-specific variation) supporting the presence of only 1 or 2 antibiotic resistance-spanning segments dominating each pool (S7 Fig).
Coincident with decreasing complexity at the antibiotic-selected loci, serial enrichment also increased the frequency of donor segments in other intervals (Figs 4 and 5). At the end of selection, 1–2 recombinant clones dominated each pool, each carrying several donor segments (one segment in each clone spanning the antibiotic resistance allele).
Overlapping donor segment intervals between independent recombinants and pools are the best candidates for carrying invasion loci (i.e. the purple circles in Figs 4 and 5). In principle, donor-specific genetic variation found in only some clones (those seen with intermediate frequencies in the pools) could potentially modulate intracellular invasion, but “hitchhiking” segments that are not associated with invasion are expected to occur, since previous work has shown that competent cells typically take up and recombine multiple donor DNA molecules. Similarly, independent recombination tracts carrying the same invasion locus are expected to typically have independent recombination breakpoints [8,10].
Sequencing of invader-enriched recombinant pools identified a single donor locus that was enriched to near-fixation in all four TREP experiments: hmw186-028NP (Fig 6). For the RdS recipient, a narrow interval reached near-fixation for both NovR and NalR-selected pools (Figs 4 and 6A; Rd coordinates 1,744,519–1,750,336 nt, accompanied by a short nearby interval at 1,760,431–1,760,794 nt with lower levels of enrichment). The Rd gene annotations in this interval have no obvious connection with host cell interactions, but the donor strain carries a large operon here that is absent from Rd: hmw1ABC, which is found in ~60% of H. influenzae strains [40–43] and located between yrbI and HI1680 (also known as NTHI1986 in the 86-028NP genome). Depth-of-coverage analysis and reciprocal mapping of sequence reads to the donor genome confirmed the insertion of hmw1ABC86-028NP into the invasion-enriched RdS recombinant pools. Three independent recombination-mediated insertions of hmw1ABC86-028NP dominated the two invasion-enriched pools: one recombinant clone in the NovR pool and two in the NalR pool, as indicated by the distinct recombination breakpoints flanking the insertion (Fig 4). Additional recombination tracts were detected distant from the putative invasion locus and the antibiotic resistance marker, but these were unique to one of the three Pool 4 clones, as expected for random “hitchhiking” recombination events.
Selecting for invaders from the HiT recombinant pools likewise enriched for donor segments containing hmw186-028NP adjacent to the yrbI gene (Figs 5 and 6B, S9 Table). Only two donor-specific intervals reached fixation in HiT NovR: one spanning gyrB as expected, and another that completely spanned only the hmw1A86-028NP gene carrying only short segments of the flanking genes (Hi375 genomic coordinates 1,178,771–1,183,728 nt). In HiT NalR, a long donor segment spanning this same interval reached ~75%.
In contrast to RdS, the HiT recipient already possesses the hmw1 operon and its paralogous operon hmw2, but the locations of the two adhesin genes (hmw1A and hmw2A) are swapped relative to their location in the donor NpNN. This is made evident through a comparison of the binding domains at the two hmw adhesin-encoding loci in HiT, NpNN, and the prototypic HMW-positive strain 12 (the strain that HMW adhesins were originally identified in and where they have been most extensively characterized, aka R2846) (Table 1). In strain 12 and the donor NpNN, hmw1 is adjacent to the yrbI gene (NTHI1982) and hmw2 is nearby the radA gene (NTHI1453). However, for Hi375, the radA-adjacent hmw adhesin has a binding domain with 100% amino acid identity to the yrbI-adjacent hmw1A gene from strain 12. Thus, whereas the hmw186-028NP operon was inserted into invasion-selected RdS recombinants, recombination events that increased HiT invasiveness were the result of an allelic substitution of hmw2Hi375 for hmw186-028NP.
The presence of the paralogous hmw locus nearby the radA gene caused an alignment artifact from multiply mapping sequence reads (Hi375 coordinates ~1,585–1,595 kb). Highly variable donor allele frequencies were seen across this interval ranging from ~0% to ~50% in Pool 8 for both HiT pools (S8 Fig). Furthermore, no donor-specific variation was detected flanking the radA-adjacent hmw1AHi375 locus, in contrast to donor variation flanking the recombinant yrbI-adjacent hmw1A86-028NP locus, which is expected for recA-mediated homologous recombination. Conclusive evidence for radA-proximal donor variation as a read alignment artifact is provided by allele-specific PCR assays on the isolated clones and mutants, as described below (S9 Fig).
Collectively, these data strongly support a role for hmw186-028NP in the increased intracellular invasion seen in enriched recombinants, though they do not strictly rule out a role for flanking donor-specific variation, particularly in the yrbI gene, since donor-specific variation in this gene was also near fixation in the invader-enriched recombinant pools. Comparison of sequence variants called between Pool 0s and Pool 8s identified that no novel mutations were fixed during serial selections, consistent with low per base mutation rates and the control experiment with untransformed recipients shown in S4 Fig.
Four colonies from Pool 4 were isolated from each of the four TREP experiments and further analyzed. Each of the 16 clones was assayed in triplicate for its invasion and adhesion phenotypes (Fig 7). Subsequent genome sequencing (summarized in S3 Table) revealed a total of only six distinct genotypes: three clones from each recipient background (Fig 7A; S9 Table). This is consistent with predictions from the pool data and shows that the isolated clones represented all the high frequency recombinants observed in Pool 4. Intracellular invasion and adhesion were strongly enhanced for all 16 isolated clones compared to recipient controls (Fig 7B and 7C, Tukey’s HSD p<<0.001 for all six comparisons). Immunoblot analysis confirmed expression of HMW1A86-028NP protein in two recombinant clones, RdS genotype B (S10A Fig) and HiT genotype E (Fig 8A).
Not all of the recovered genotypes had identical phenotypes. Genotype D colonies (HiT NalR from Pool 4) were significantly more invasive and adherent than genotype E and F colonies (one-way ANOVA p<<0.001, Tukey’s HSD gives p<0.001 for comparison of D against E or F, but p>0.1 for comparison of E and F). This indicates that donor-specific variation present in genotype D but absent from genotypes E and F slightly enhances adherence and invasion, albeit substantially less so than the hmw1ABC locus; the causative variation responsible remains unknown, but includes donor alleles of a QseBC-like two-component system (S9 Table, “D-specific” segments).
The sequencing of Pool 4 clones further supported a read alignment artifact at the radA-adjacent hmw1AHi375 locus. Colonies had been collected after re-streaking to ensure they represented single clonal lineages. Despite this, all three HiT clone genotypes (D, E, and F) from Pool 4 showed a variable mixture of recipient- and donor-specific alleles across the radA-proximal hmw operon (Fig 6). In contrast, donor-specific allele frequencies observed at the yrbI-adjacent hmw1A86-028NP were near fixation and accompanied by flanking donor-specific variation. As a final confirmation that this mixed signal was not due to merodiploidy or other complex genetic effect, allele-specific PCR assays were conducted that distinguished all four hmw adhesin genes (the yrbI-proximal hmw1A86-028NP and hmw2AHi375 genes and the radA-proximal hmw2A86-028NP and hmw1AHi375 genes), and these PCR assays unambiguously showed that HiT recombinants had replaced their yrbI-adjacent hmw2Hi375 allele with the hmw186-028NP allele, whereas the radA-adjacent hmw1Hi375 alleles were unchanged (S9 Fig).
Despite the unambiguous acquisition of hmw186-028NP in all recombinant clones, other donor-specific variation in the recombinant clones could conceivably be responsible. Since recombination tracts that carried hmw1ABC86-028NP also carried flanking donor-specific SNPs, we directly tested for a contribution by variation in the flanking interval. We cloned the 86-028NP alleles of the two genes upstream of hmw1ABC: kpsF and yrbI, encoding arabinose-5-phosphate isomerase and Kdo 8-phosphatase, respectively. Donor-specific variation in yrbI in particular might contribute to intracellular invasion, since the minimum interval that overlapped between all four experiments contained donor-specific variation in this gene (Fig 6). The resulting HA-tagged pSU20 plasmid (pSU20-kpsF-yrbI-HA) was then electroporated into Rd. Confirming expression from the plasmid, a ~19.3-kDa full-length YrbI86-028NP-HA protein was detected in whole cell extracts by immunoblot with an anti-HA antibody (S10B Fig). Strains Rd, Rd pSU20, and Rd pSU20-kpsF-yrbI-HA were tested for intracellular invasion into A549 cells. No significant difference was observed among strains (S10C Fig, one-way ANOVA p-value = 0.29), thereby excluding a significant role for these flanking loci in intracellular invasion.
Genetic confirmation of a role for hmw1A86-028NP in intracellular invasion (rather than other donor-specific variation acquired by recombinants) was performed using the HiT recipient and one of the invasive HiT recombinant clones (strain P551 with genotype E, hereafter called rHiT). We generated a panel of knockouts of the genes encoding HMW adhesins in the HiT and rHiT strains, either the locus adjacent to yrbI (hmw2AHi375 or hmw1A86-028NP) or the locus nearby radA (hmw1AHi375). In the case of rHiT, the double mutant was also produced (both hmw1A86-028NP and hmw1AHi375 deleted). Western blot analysis confirmed expression of the expected HMW adhesins in each strain (Fig 8A).
All mutant strains were assayed for invasion in parallel with NpNN, HiT, and rHiT controls (Fig 8B, one-way ANOVA p<<0.001). We hypothesized that knocking out hmw1A86-028NP in the recombinant rHiT would show a strong defect in intracellular invasion, but that knocking out either hmw gene in the HiT recipient would have little or no effect. Indeed, deletion of hmw1A86-028NP from rHiT reduced invasion frequencies 56-fold (Tukey’s HSD p<<0.01) down to HiT recipient levels. By contrast, deletion of the radA-proximal hmw1AHi375 had no significant effect in either strain background, nor did deletion of either locus in the HiT recipient (Tukey’s HSD p>0.5 for all comparisons).
These results confirmed the role of hmw1A86-028NP in the increased intracellular invasion of the HiT recombinant, and suggested that the HiT alleles of the HMW adhesins do not appreciably contribute to the HiT strain’s ability to invade A549 cells. Parallel adhesion assays of the knockout panel found qualitatively similar results (Fig 8C, one-way ANOVA p<<0.001). The recombinant rHiT was significantly more adherent than the HiT recipient (4.7-fold increase, Tukey’s HSD p<0.001)—comparable to adhesion by the donor NpNN (p = 0.99)—whereas deletion of hmw1A86-028NP from rHiT brought adhesion down 3.3-fold to HiT recipient levels (Tukey’s HSD p = 0.0014 versus rHiT, p = 0.99 versus HiT).
The effect of hmw1A86-028NP on adhesion is >10-fold lower than its effect on intracellular invasion, such that significant increases in adhesion had not been detected in the original pool experiments and were only marginally significant in adhesion assays with the isolated recombinant clones (see above). Nevertheless, these results indicate that hmw1A86-028NP might contribute to intracellular invasion in part through an indirect effect of increasing adherence. While this is one contributing factor, the immunofluorescence data reported below indicate an unexpected intracellular invasion phenotype that cannot readily be explained by increased adherence alone.
In the course of working with the parental strains, we noted that when cultures were left standing on the bench, NpNN (and 86-028NP) settled more quickly than RdS (and Rd), denoting a clumping or self-aggregation phenotype. Because self-aggregation could modulate bacterial-host cell interplay, we quantitatively tested this phenotype for a panel of strains: the three parents, an invasive recombinant RdS clone (P540, genotype B, hereafter rRdS), the rHiT recombinant, and the rHiTΔhmw1A86-028NP mutant. This clearly demonstrated that hmw1A86-028NP plays a major role in the high self-aggregation seen in the recombinants (Fig 9; one-way ANOVA p<<0.001 at the t = 140 min time point and higher). Both recombinants settled quicker than the recipient strains (Tukey’s HSD p<0.001) but were indistinguishable from NpNN (p>0.6). While the HiT recipient settled substantially faster than RdS, it and the rHiTΔhmw1A86-028NP mutant settled slower than NpNN (p<0.01). These observations raised the question of how this clumping or self-aggregation phenotype modulates adhesion and intracellular invasion.
To directly assess intracellular location of bacterial cells, we used immunofluorescence microscopy (Fig 10). As expected, the donor (NpNN) and two recombinant (rRdS and rHiT) strains infected A549 cells at high rates, whereas the recipients (RdS and HiT) and the mutant recombinant rHiTΔhmw1A86-028NP infected A549 cells at substantially lower rates (Table 2). Co-localization with the endosomal marker Lamp-1 was substantial for all strains, indicating bona fide intracellular invasion, rather than gentamicin resistance or occlusion by A549 cells. These results confirm: (a) that both recipients (including RdS) successfully enter A549 cells, albeit at low rates; (b) that donor and recombinants have substantially higher invasion rates than recipients; and (c) that hmw1A86-028NP is responsible for elevated intracellular invasion rates in the rHiT recombinant.
The number of bacteria per infected cell was also distinct between strains. Cells infected by either recipient or the rHiTΔhmw1A86-028NP mutant were infected with <10 bacteria/cell, whereas cells infected by the donor strain had >10 bacteria/cell. Although visually similar to the NpNN donor, scoring indicated that the rRdS and rHiT recombinants had an intermediate phenotype (Fig 10, Table 1). Whereas gentamicin protection assays with the recombinants gave invasion rates that approached donor levels, these data suggest that additional unidentified donor-specific factors besides hmw1A86-028NP may augment the ability of bacteria to invade.
Unexpectedly, bacterial invaders had a distinct pattern of co-localization with Lamp-1-positive endosomes for the donor and two recombinants. When infected by bacteria of either recipient strain or the rHiTΔhmw1A86-028NP mutant, Lamp-1-positive subcellular compartments typically enclosed single bacterial cells, as previously observed [24]. In contrast, cells of the donor or either recombinant had enlarged Lamp-1-positive subcellular compartments that surrounded groups of bacteria. The size of these groups varied, with ~5–50 bacteria per compartment (Fig 10).
These results indicate that increased intracellular invasion by the donor and recombinants relates—at least in part—to internalized bacteria localizing as groups in the same subcellular compartment, rather than as individual cells. Previous work showed that H. influenzae invaders do not replicate [24] and, in our assays, the time between infection and observation was sufficiently brief that the groups of intracellular bacteria seen are unlikely to reflect intracellular replication. Hi375 has previously been seen to clump at the surface of A549 cells prior to invasion, but with only single bacteria undergoing internalization [24]. Thus, our self-aggregation and immunofluorescence results strongly suggests that groups of hmw1ABC86-028NP bacteria remain clumped during epithelial cell entry, thereby increasing overall invasion rates above and beyond the indirect effect of elevated adhesion, and also explaining the observed Lamp-1 reorganization around groups of bacteria.
Previous studies of hmw1 focused on its role in adhesion, including detailed characterization of an Rd derivative carrying the hmw1 operon from strain 12 with only minimal flanking variation from strain 12 [44]. We used the Rd hmw1strain12 strain to independently test for the role of hmw1 in intracellular invasion. We evaluated invasiveness and adhesiveness of Rd, Rd hmw1strain12, and strain 12 using A549 cells, as described above. Rd hmw1strain12 had intermediate levels of both invasion and adhesion between both parents (Fig 11, p-values<0.01 for one-way ANOVA and all three comparisons by Tukey’s HSD).
To test how these results depended on the specific protocol or cell type used, we evaluated invasion and adhesion following an alternative protocol in both A549 and Chang cells [18]. Rd hmw1strain12 had significantly higher adhesion and invasion than Rd for both cell types, though the effect was much stronger with Chang cells (S11 Fig, p-value < 0.01 for all three comparisons, except p = 0.054 for adherence to Chang cells by Rd hmw1strain12 and Strain 12). Rd hmw1strain12 invaded Chang cells nearly as well as strain 12. In contrast, when infecting A549 cells, Rd hmw1strain12 had an intermediate phenotype. These results indicate that hmw1’s contribution to invasion depends on the host cell type and confirm a more general role for hmw1 in intracellular invasion beyond that seen for the 86-028NP allele.
In contrast to bacterial pathogens with well-characterized intracellular life styles like Salmonella, Listeria, Legionella or Brucella [45–47], the mechanism and functional role of intracellular invasion in H. influenzae has been less well understood. It has been suggested that intracellular invasion of airway epithelial cells by non-typeable H. influenzae (NTHi) allows the bacterium to evade the immune system (antibodies, surfactant, antimicrobial peptides, galectins, professional phagocytes) and therapeutic interventions (antibiotics, anti-inflammatory agents), and to facilitate access to essential nutrients [14,48–50]. Thus, entry into airways cells may better equip bacterial cells for survival during long-term infections, particularly in the context of chronic infections that are often treated with intense antibiotic regimes.
Potential factors that contribute to invasion include those known to facilitate H. influenzae’s interactions with host cell surfaces. Among these are bacterial surface proteins that participate in H. influenzae binding to extracellular matrix proteins, mucin, or epithelial cells (including P5, OapA, PE, Hap, Hia, the HMW1 and HMW2 adhesins, IgA1 proteases A and B, and type IV pili [26,33,38,51–54]). Some factors, such as IgaA1 protease, appear to be more directly involved in H. influenzae entry into epithelial cells [26]. The PE and Hap adhesins have also been implicated in H. influenzae entry into epithelial cells [55,56]. While adherence to host cells may be a prerequisite for invasion, we lacked information on the specific involvement of adhesins or other factors that modulate intracellular invasion by H. influenzae. TREP was designed to identify invasion-promoting genes in an unbiased manner, but we were nonetheless surprised to isolate the well-characterized adhesin-encoding hmw1 operon.
Gain of hmw186-028NP by a poorly invading strain naturally lacking both hmw1 and hmw2 (RdS) dramatically enhanced both adhesion and invasion, showing that hmw186-028NP is sufficient to confer high adhesion and invasion levels. When a recipient strain already possessing both hmw1 and its paralog hmw2 (HiT) was transformed, allelic substitution of the hmw2AHi375 allele with hmw1A86-028NP also strongly enhanced invasion. Characterization of recombinant clones and mutants lacking functional hmw186-028NP confirmed these results. Important roles for other donor-specific variation carried by the transformed recombinants can be excluded, except that one of the HiT recombinants (genotype D) showed significantly higher adhesion and invasion than the others, suggesting at least one additional invasion-promoting factor, albeit a smaller contributor; further analysis will be of interest in future studies.
The striking intracellular phenotype we observed by immunofluorescence for hmw186-028NP strains—in which groups of bacteria occupy engorged intracellular vesicles—suggests that increased adhesion alone is insufficient to fully explain the role of hmw1 in intracellular invasion. We instead suggest that elevated invasion is an emergent property of HMW1-mediated self-aggregation; whereas adhesion is increased as an indirect result, we speculate that invasion by bacterial groups directly enhances overall invasion rates. Alternatively, possession of hmw186-028NP may increase independent entry by bacteria into cells, followed by subsequent aggregation of bacterium-containing vesicles. Ruling out self-aggregation per se, recent results show that deletion of the Hap autotransporter from Hi375 (naturally present in both 86-028NP and Hi375 but absent from Rd) eliminates self-aggregation, but epithelial cell adhesion and invasion were not significantly affected [57]. Altogether, we propose that, whether or not it can truly be called an “invasin”, allelic variation in HMW1A affects invasion by increasing adhesion and also directly through a novel mechanism that allows for entry by groups of aggregated bacterial cells (model summary in Fig 12).
The hmw1 and hmw2 operons are found in ~60% of H. influenzae isolates, and they always co-occur, despite being at different chromosomal loci (one adjacent to HI1679 in Rd and the other to HI1598) [40,58,59]. Co-occurrence of the hmw1 and hmw2 loci in all tested clinical isolates suggests that the laboratory-created hmw1-only strains studied here must be at some unknown fitness disadvantage in nature.
The hmw1 and hmw2 operons are phase-variable, and expression is inversely correlated with the number of 7-bp tandem repeats found within their promoter regions [60–63]. Though subtle expression variation was not ruled out, western blot analysis indicated that HMW1 adhesin levels were mostly unchanged in recombinants (Fig 8A; 16 repeats upstream of hmw186-028NP but 17 upstream of hmw2Hi375). This indicates that allelic variation in the hmw coding sequences is likely responsible for differences in adhesion and invasion.
HMW1A and HMW2A display wide amino acid diversity both within and between isolates, with the region of lowest sequence identity in the host cell binding domain, which has been predicted to affect tissue tropism and immune evasion [42,59,63–67]. Phylogenetic analyses of the HMW adhesin binding domain has revealed four distinct sequence clusters, and the majority of sequences belonging to one of two dominant sequence clusters [41]. Of note, 86-028NP and strain 12 hmw1A binding domains belong to clusters 4 and 2, respectively, which might contribute to their strong and intermediate phenotypes; future studies using mosaic proteins with binding domains from distinct clusters and using multiple human cell lines could identify any clade-specific functions for HMW proteins.
The two hmw operons encode high molecular weight non-pilus adhesins (HMW1/HMW2) [37,64], along with two co-factor proteins encoded by the downstream genes. The co-factors are required for proper surface localization of HMW adhesins, and the paralogs of the co-factors are functionally interchangeable [37,58]. The hmw1B/hmw2B genes encode outer membrane pore-forming translocators that export HMW1 and HMW2 to the cell surface [68]. The hmw1C/hmw2C genes encode glycosyltransferases responsible for adding mono-hexose or di-hexose residues at asparagines in conserved NX(S/T) motifs of HMW1 and HMW2 [69], likely involved in stabilizing the adhesins during or after their synthesis [70]. Due to the high diversity in HMW adhesin sequences, differential glycosylation patterns might in part be responsible for distinct activities of different alleles.
Another key distinction between the HMW1 and HMW2 adhesins is that the former recognizes sialylated glycoprotein receptors on cultured human epithelial cells [71]. HMW1 confers high adherence to Chang, Hep-2, HaCaT and NCI-H292 cells mediated by interactions with α-2,3 N-linked sialic acids. By contrast, HMW2 confers adherence to HaCaT and NCI-H292 cells via a sialic acid-independent mechanism [59,67,71]. This suggests that the role of hmw1A86-028NP in intracellular invasion may involve specific interactions between H. influenzae cells and sialylated glycoprotein receptors, both on the bacterial cell surface to mediate self-aggregation and possibly also specific host sialylated glycoproteins on epithelial cells. Glycoproteins play important roles in many cellular activities, and new methods for investigating their expression and sialylation states are being developed and applied to multiple cell types including A549 [72], opening new avenues to identify host glycoproteins hijacked by bacterial proteins such as HMW adhesins during the infection process.
To our knowledge, this is the first report of an involvement for hmw1A in H. influenzae intracellular invasion and, more strikingly, we further found that intracellular invasion is modulated by allelic diversity at hmw1A. Finding that hmw186-028NP results in clumps of intracellular bacteria was unexpected and indicates that increased self-aggregation and adhesion per se are not sufficient to explain its effects, offering new avenues of investigation. Bacterial factors contributing to adhesion are already potential targets for antimicrobial therapies, and the additional role of HMW1 in intracellular invasion further increases its attractiveness as a target. Understanding the relationship between hmw1 allelic variation and within-host adaptive evolution poses interesting challenges for future studies.
To better understand intracellular invasion by H. influenzae, we have successfully employed a gain-of-function genetic mapping strategy, TREP, which takes advantage of within-species phenotypic variation, natural competence, and deep sequencing. In total, our experiment isolated six highly invasive recombinants from a total of ~400,000 independent recombinants. Thus, while transformation rates of SNPs was much higher (e.g. ~0.4% for the antibiotic resistance alleles), this approach was able to isolate even very rare recombinants, in the case of the Rd strain requiring the insertion of a particularly long operon (>9kb). TREP proved to be a rapid method to map genes. Once the donor and recipient were assayed and the effectiveness of the selections was determined, the total hands-on time was only about six weeks, from generating the recombinant pools, performing the serial selections, extracting DNA, making libraries and sequencing, with serial selections comprising the most time-consuming step. Owing to the strong selection used, we found that strains that have slight advantages in invasion were able to overtake the pools after serial selection. Thus, it was crucial to add selectable markers to our recipient background and to ensure that the serial selections were performed without the donor strain or any other strains assayed in parallel.
The TREP method holds great promise for studying a wide range of traits that show natural phenotypic variation in other naturally competent species, which includes many virulence traits and pathogens important to human health. In contrast to screening/selecting clones transformed by plasmids [73], TREP does not depend on dominance, a suitable vector, nor is it restricted to monogenetic traits. The approach should be readily applicable to any selectable trait in any bacterial species for which a natural competence protocol has been developed, and the number of such species continues to grow. Similar approaches have recently been reported in other organisms, for example to identify conjugation genes in Mycobacterium and causative alleles responsible for antibiotic resistance in Streptococcus [7,74]. Importantly, TREP is a general genetic mapping strategy agnostic to the type of variation (i.e. SNPs or whole loci can be identified), and we expand the utility of the transformation-based genetic mapping to include quantitative differences that go beyond absolute phenotypic differences (i.e. resistance versus sensitivity) by incorporating serial selection.
Traditionally, microbial experimental evolution studies rely on “hard” selective sweeps, in which newly arising beneficial mutations fix in a laboratory population [75]; more recently this has also included experimental evolution of pathogenic traits [76,77]. But “soft” sweeps, in which pre-existing genomic variation recombines within/into a population [78], may also play an important role in the adaptation of naturally competent species to new environments [79]. Allowing for introgression of natural variation has been used in experimental evolution in sexual eukaryotes (e.g. [80,81]), but its role in bacterial adaptation has been explored only recently and only in the context of standard experimental evolution studies that start with clonal populations [82,83]. Here, we found that rare recombinants generated in a single round of natural transformation could reach fixation after a small number of serial selections, illustrating the powerful contribution of natural competence to adaptation.
Finally, applying TREP to understand bacterial pathogenesis could use large “zoos” of donor genomic DNAs, rather than single donor-recipient combinations. This would better mimic the situation in chronic infections, where diverse polyclonal infections are common, and it would more fully sample the genomic diversity of these organisms in single experiments. However, caution must be exercised with such an approach: depending on the organism and trait under study, this could inadvertently generate novel hyper-virulent strains by combining multiple pathogenicity factors from different genetic backgrounds; a similar ethical concern has already been raised for studying pathogens using gain-of-function mutagenesis [84,85].
Bacterial strains and plasmids used are listed in S1 Table, and all PCR primers used are in S10 Table. General culturing and manipulation of Haemophilus influenzae followed standard methods [39]. Strains were grown at 37°C with 5% CO2, on chocolate agar or brain heart infusion (BHI) supplemented with 10 μg/ml hemin and 10 μg/ml β-nicotinamide (sBHI). Antibiotics were added as required: novobiocin (Nov) at 2.5 μg/ml, nalidixic acid (Nal) at 3 μg/ml, spectinomycin (Spc) at 25 μg/ml, streptomycin (Str) at 100 μg/ml, chloramphenicol (Cm) at 2 μg/ml, and erythromycin (Erm) at 11 μg/ml. Escherichia coli strains were grown at 37°C on Luria Bertani (LB), and Cm at 30 μg/ml was added as required.
The spectrum and distribution of recombinants in transformed pools that carry invasion alleles/loci depends upon the number of loci involved, their genetic interactions, the rates of recombination at those loci, and the experimental environment. To maximize the chance of enriching invasive recombinants from transformed pools: (a) We selected for a donor-specific marker (either NovR or NalR) after transformations to ensure that recombinant clones were not derived from non-competent cells in the original culture [8,10]. (b) We selected for a recipient-specific marker (SpcR or StrR) to limit cross-contamination. (c) Colonies were pooled, so that each independent recombinant in the pools was represented by many (>106) cells. (d) We maximized the complexity of the recombinant pools emerging from the first round of selection. (e) We progressively increased the frequency of invasive recombinants by serial selection, using a pool of the total CFU output from the gentamicin protection assay as the infecting material for the next cycle of selection.
Recipient strains were made naturally competent using the standard protocol [39], except scaled up to 10 ml (~1010 CFU). Briefly, exponentially dividing cells growing in rich medium (sBHI) were transferred to starvation medium (MIV) for 100 min. Purified genomic DNA from the donor was incubated with naturally competent cultures at a concentration of ~1 genome per cell, or ~2 μg / 109 CFU / ml, for 30 min on a roller drum at 37°C, followed by a 1:5 dilution into sBHI and further incubation for 80 min to allow for expression of resistance alleles. Cultures were diluted and plated on sBHI-agar ± antibiotics to measure transformation and co-transformation frequencies (as NovR or NalR resistant colonies / CFU). Percent competence was calculated as (NovR NalR / CFU) / (NovR/CFU * NalR/CFU), as previously described [8,10,90]. To generate high complexity pools of recombinants, we plated 0.75 ml of a 10−2 dilution to 20 large petri dishes (20 cm diameter): 10 containing Nov and 10 with Nal (plus Spc or Str, depending on the recipient). This yielded ~104 resistant colonies per plate. Colonies from each set of 10 plates were scraped into a single 10 ml sBHI pool, titrated by dilution and plating to sBHI+antibiotics, and immediately stored as 1.25 ml aliquots in 15% glycerol at -80°C. This generated a total of four pools with an initial complexity of ~105 independent recombinants each, two for each recipient (Rd KW20 SpcR and Hi375 StrR), selected for either the NovR or the NalR donor allele, as well as a second antibiotic to select for the appropriate recipient background.
The carcinomic human alveolar basal epithelial cell line A549 (ATCC CCL-185) was maintained in RPMI 1640 medium supplemented with 10 mM Hepes, 10% heat-inactivated fetal calf serum (FCS) and antibiotics (penicillin 100 units/ml and Str 0.1 mg/ml) in 25 cm2 tissue culture flasks at 37°C in a humidified 5% CO2 atmosphere. Chang cells (ATCC CCL-13) were cultured under the same atmospheric conditions in Minimum Essential Medium Eagle supplemented with 10% FCS and 1x MEM non-essential amino acid mixture (Sigma). Cells were seeded to 6×104 or to 1.2×105 cells / well in 24- or in 6-well tissue culture plates, respectively, for 32 h, and then serum starved for 16 h before infection. A ~90% confluence was reached by the time of infection. Adhesion and intracellular invasion assays in 24-well plates were conducted as previously described [15,24], starting with H. influenzae cells scraped from chocolate-agar plates (freshly grown for 16 h at 37°C with 5% CO2) into PBS and adjusted to OD600 = 1. A small aliquot of this adjusted suspension was diluted and plated on sBHI-agar to titrate the input CFU.
For invasion assays, A549 cells were incubated with 0.2 ml of each adjusted bacterial suspension for 2 h, washed 3 times with PBS and incubated for 1 h with RPMI 1640 containing 10% FCS, Hepes 10 mM and gentamicin 200 μg/ml to kill extracellular bacteria (the bacterial isolates used all had minimum inhibitory concentrations of < 5 μg/mL), washed 3 times with PBS, and human cells were lysed with 300 μl of PBS-saponin 0.025% for 10 min at room temperature. To quantify intracellular invasion frequencies, lysates were serial diluted and plated onto sBHI-agar with appropriate antibiotics. Recovered CFU was divided by the input CFU to calculate “Invaders/CFU”. Unless otherwise indicated, all infections were carried out in triplicate on three separate occasions. Adhesion assays were carried out similarly, excluding gentamicin treatment to calculate “Adherents/CFU”. For adhesion assays, cells were incubated with 0.1 ml of each adjusted bacterial suspension for 30 min. Wells were then washed 5 times with PBS and lysed as above. An alternative method was also used for both invasion and adhesion for comparisons of Rd, Strain 12, and Rd/HMW1Strain12 following a previously published protocol [18]. The primary differences were the lack of a serum starvation step, a low-speed centrifugation step to quickly bring bacteria into contact with the monolayer, and a lower MOI.
To enrich for recombinants carrying donor-specific invasion alleles, we performed eight serial selections for invasive clones for each recombinant pool. To maximize the complexity of the initial recombinant pools (Pool 0), one frozen aliquot per pool (~1010 CFU of ~105 independent recombinants per aliquot) was used to infect three wells of A549 cells seeded onto 6-well plates. Pool 0 aliquots were first recovered by thawing, pelleting, resuspending in 5 ml sBHI, and incubating on a roller drum for 60 min at 37°C under 5% CO2. Cultures were pelleted prior to proceeding with the invasion protocol, performed as described above, scaled up to cells seeded on 6-well plates (in 4 ml EBSS, with 0.8 ml of bacterial adjusted suspension / well), starting with resuspension of pellets in PBS, and ending with the total lysate plated on sBHI-agar (+appropriate antibiotics) at varying dilutions. This allowed measurement of intracellular invasion frequency and provided material for the next cycle.
For each subsequent serial invasion cycle, all CFU were scraped off plates into PBS and thoroughly mixed before normalizing to OD600 = 1 and proceeding with the infection. For all cycles, unused material was stored as (i) 15% glycerol stocks at -80°C for repeat assays and isolation of individual clones, and (ii) as a pellet at -20°C for DNA extractions (except for the RdS Pool 1 material, for which none was left over). In practical terms, this serial infection procedure was repeated for four enrichment cycles, at which point recovered pools were frozen as 15% glycerol stocks to allow for a new set of confluent A549 cells to grow up; frozen stocks were then restarted to carry out four additional cycles of selection. Untransformed recipient controls were run in parallel to exclude potential issues related to cell seeding. At Pool 4, several single gentamicin-protected clones per enrichment were isolated on sBHI-agar plates and stored at -80°C in 15% glycerol for adhesion and invasion assays, as well as clone sequencing.
To monitor YrbI-HA expression, whole cell extracts from strain Rd alone, Rd carrying pSU20, and Rd carrying pSU20-Pr::kpsF-yrbi-HA were prepared from bacterial cultures grown to OD600 = 0.9 in sBHI containing Cm, when required. YrbI-HA expression was analyzed by western blot with a primary rabbit anti-HA antibody (Sigma) diluted 1:4000, and a secondary goat anti-rabbit IgG (whole molecule, Sigma) antibody conjugated to horseradish peroxidase, diluted 1:1000.
To investigate HMW adhesin protein expression in strains NpNN, RdS, rRdS, HiT, HiTΔhmw1AHi375, HiTΔhmw2AHi375, rHiT, rHiTΔhmw1A86-028NP, rHiTΔhmw1AHi375, rHiTΔhmw1A86-028NPΔhmw1AHi375, whole cell extracts were prepared from bacterial suspensions recovered from overnight grown chocolate-agar plates and adjusted to OD600 = 1 in PBS. HMW1A expression was analyzed by western blot with a primary guinea pig anti-HMW1A (gp85) antibody diluted 1:2000 [91], and a secondary goat anti-guinea pig IgG (Santa Cruz) antibody conjugated to horseradish peroxidase, diluted 1:5000.
H. influenzae cells were scraped from chocolate-agar plates freshly grown for 16 h at 37°C with 5% CO2 into PBS solution, and adjusted to OD600 = 0.45 in a 35 ml volume, and left standing at room temperature for at least 260 min. OD600 readings were performed at regular time intervals on 500 μl aliquots gently collected from the top of each bacterial suspension. Four independent experiments were performed for each strain.
A549 cells were seeded on 13 mm circular coverslips in 24-well tissue culture plates. Cells were infected at an MOI ~1:8 (5 μl) of each adjusted bacterial suspension for 2 h, and infected cells were incubated in RPMI 1640 containing 10% FCS, Hepes 10mM and gentamicin 200 μg/ml for 1 h. Cells were washed three times with PBS and fixed with 3.7% paraformaldehyde (PFA) in PBS pH 7.4 for 15 min at room temperature. Immunofluorescence staining was carried out as previously described [24]. H. influenzae cells were stained with a rabbit anti-NTHi serum (raised against a pool of strains Hi375, 2019, and 398 [24]) diluted 1:600. Late endosomes were stained with mouse monoclonal anti-human Lamp-1 H4A3 antibody (Developmental Studies Hybridoma Bank) diluted 1:100. DNA was stained with Hoechst 33342 (Invitrogen) diluted 1:2500. Donkey anti-rabbit conjugated to Cy2 and donkey anti-goat or donkey anti-mouse conjugated to Rhodamine secondary antibodies (Jackson) were diluted 1:100.
Samples were analyzed with a Carl Zeiss Axioskop 2 plus fluorescence microscope and a Carl Zeiss Axio Cam MRm monochrome camera. We quantified: (a) the percentage of infected cells, counting at least 250 cells per sample; (b) the number of bacteria per infected cell in at least 250 cells per sample type, scoring <10 bacteria/cell or >10 bacteria/cell; (c) co-localization of bacteria and Lamp-1—an NTHi-containing vacuole (NTHi-CV) was considered positive for Lamp-1 when the marker was detected throughout the area occupied by the bacterium, or around/enclosing the bacterium. To determine the percentage of bacteria that co-localized with Lamp-1, all bacteria located inside a minimum of 150 infected cells were scored in each experiment. Results were calculated from two independent experiments.
Genomic DNA was extracted from the donor and recipients, stored pools, and isolated clones by phenol/chloroform extraction as described [8]. Purity and quality were evaluated by Nanodrop spectrophotometry (Thermo Scientific) and agarose gel electrophoresis, and quantification was performed with Qbit fluorometry prior to sequencing library construction. Multiplexed sequencing libraries were produced using the Nextera XT kit following manufacturer recommendations (Illumina). Paired-end sequencing (2x101nt) was conducted on an Illumina HiSeq in RapidRun mode over several independent runs/lanes. Raw base call data (bcl) was converted into FastQ format (Illumina version 1.8) using the bcl2fastq conversion software from Illumina (version 1.8.3, setting—no-eamss). For recombinant clones, paired-end sequencing (2x151nt) was conducted on an Illumina MiSeq, which automates demultiplexing to provide raw FastQ files. Properties of the genomic DNA samples and sequencing statistics (including donor and recipient controls) are in S2 and S3 Tables.
The genome sequence references for the donor and recipients were: 86-028NP (NC_007146.2) [28], Rd_KW20 (NC_000907.1) [27], and Hi375 (CP009610.1) [29]. For the Rd genome, all non-ACGT bases were first converted to Ns (some non-N ambiguous IUPAC nucleotide characters lead to errors running samtools mpileup). Reads from control strains were used to identify variation between the derivative strains’ genomes and their deposited parental reference sequences (as described below). For all raw Illumina sequence processing, paired-end reads were trimmed of adapter sequences with Trimmomatic (v0.32) [92] and overlapping pairs were merged with COPE (v1.1.3; simple-connect mode) [93]. Next, reads were mapped using bwa mem (v0.7.8) with default settings [94], duplicates were marked with SamBlaster (v0.1.14) [95], and aligned reads were sorted and compressed using SamBamba (v0.4.6) [96]. Subsequent steps filtered out reads with a mapping quality = 0, which excludes multiply mapping reads that align equally well to different reference genome coordinates.
For donor and recipient controls, as well as recombinant clones, single-nucleotide polymorphism (SNP) and small indel variant calling used samtools mpileup and bcftools view (v0.1.19) [97]. Variant frequency calling from recombinant pools used a python script (available at https://github.com/photonchang/allelecount/) to count reads supporting each of the 4 bases at each reference position directly from samtools mpileup output (for base calls with quality score >10), and subsequent parsing used linux commands (mostly awk). BedTools (v2.19.1) [98] was used for subsetting (using the intersect tool) with the variants detected between donor and recipient genomes. Variant tables were first corrected for “self” variants identified between reads and their own reference (with the exception of resistance-associated markers). This allowed calculation of recipient-specific, donor-specific, and erroneous base frequencies (i.e. bases with neither donor nor recipient identity). Manual validation of recombination breakpoints and clone assignments used the Integrative Genomics Viewer (v2.3.1) [99]. Identification of novel alleles that had approached fixation compared the variants called from Pool 8 reads to those from Pool 0 reads (using samtools mpileup and bcftools view). Due to systematic alignment artifacts that arise when mapping donor reads to recipient genomes in regions of high divergence, putative novel variation that was also identified only in reciprocal alignments of control reads (“unreliable” SNP positions) was excluded, leaving no observed fixed new mutations.
Significant differences in invasion, adhesion, and self-aggregation phenotypes among strains and pools were evaluated using one-way ANOVA with post hoc hypothesis testing using Tukey’s HSD (“honest significant differences”). Invasion and adhesion frequencies were first log-transformed prior to testing to account for the highly unequal variances observed between strains/pools that were quantified at distinct plating dilutions. Pairwise student’s t-tests with untransformed data and Bonferroni correction gave qualitatively similar results. Plotting used the R statistical programming language including add-on packages seqinr, genoplotr, ggplot2, and Rcolorbrewer.
All sequence data were deposited at NCBI under BioProject PRJNA308311. BioSample accessions are included in S2 and S3 Tables. Parental strains were submitted to the SRA as BAM files aligned to their own reference sequence. Recombinant pool and clone data were submitted to the SRA as BAM files aligned to the appropriate recipient reference sequence (Hi375 or Rd KW20).
|
10.1371/journal.pcbi.0030221 | STDP in a Bistable Synapse Model Based on CaMKII and Associated Signaling Pathways | The calcium/calmodulin-dependent protein kinase II (CaMKII) plays a key role in the induction of long-term postsynaptic modifications following calcium entry. Experiments suggest that these long-term synaptic changes are all-or-none switch-like events between discrete states. The biochemical network involving CaMKII and its regulating protein signaling cascade has been hypothesized to durably maintain the evoked synaptic state in the form of a bistable switch. However, it is still unclear whether experimental LTP/LTD protocols lead to corresponding transitions between the two states in realistic models of such a network. We present a detailed biochemical model of the CaMKII autophosphorylation and the protein signaling cascade governing the CaMKII dephosphorylation. As previously shown, two stable states of the CaMKII phosphorylation level exist at resting intracellular calcium concentration, and high calcium transients can switch the system from the weakly phosphorylated (DOWN) to the highly phosphorylated (UP) state of the CaMKII (similar to a LTP event). We show here that increased CaMKII dephosphorylation activity at intermediate Ca2+ concentrations can lead to switching from the UP to the DOWN state (similar to a LTD event). This can be achieved if protein phosphatase activity promoting CaMKII dephosphorylation activates at lower Ca2+ levels than kinase activity. Finally, it is shown that the CaMKII system can qualitatively reproduce results of plasticity outcomes in response to spike-timing dependent plasticity (STDP) and presynaptic stimulation protocols. This shows that the CaMKII protein network can account for both induction, through LTP/LTD-like transitions, and storage, due to its bistability, of synaptic changes.
| Learning and memory have been hypothesized to occur thanks to synaptic modifications. The efficacy of synaptic transmission has been shown to change as a function of correlated activity between presynaptic and postsynaptic neurons. Long-lasting synaptic modifications can occur in both directions (long-term potentiation (LTP) and long-term depression (LTD)). Recent experiments suggest that these synaptic changes are all-or-none switch-like changes. This would mean that only two stable states of synaptic transmission efficacy exist, i.e., a low state, or “switched off”, and a high state, or “switched on”. LTP would correspond to switching on the synapse and LTD to switching off. We propose a realistic biochemical model of protein–protein interactions which exhibits two stable states. We then investigate conditions under which the model exhibits transitions between the two stable states. We show that experimental stimulation protocols known to evoke LTP and LTD lead to corresponding transitions in the model. This work supports the idea that the investigated intracellular protein network has a role in both induction and storage of synaptic changes, and hence in learning and memory storage.
| Synaptic plasticity is thought to underlie learning and memory, but the mechanisms by which changes in synaptic efficacy are induced and maintained over time are still unclear. Numerous experiments have shown how synaptic efficacy can be increased (long-term potentiation, LTP) or decreased (long-term depression, LTD) by spike timing of presynaptic and postsynaptic neurons [1,2], presynaptic firing rate [3,4], or presynaptic firing paired with postsynaptic holding potential [5]. These experiments have led to phenomenological models that capture one or several of these aspects [6–14]. However, these models tell us nothing about the biochemical mechanisms of induction and maintenance of synaptic changes. The question of the mechanisms at the biochemical level has been addressed by another line of research work originating from early work by Lisman (1985) [15]. Models at the biochemical level describe enzymatic reactions of proteins in the postsynaptic density (PSD) [15–19]. These proteins form a network with positive feedback loops that can potentially provide a synapse with several stable states—two, in the simplest case—providing a means to maintain the evoked changes. Hence, synapses in such models are similar to binary switches, exhibiting two stable states, an UP state with high efficacy, and a DOWN state with low efficacy. The idea of binary synapses is supported by recent experiments on CA3-CA1 synapses [20–22].
One of the proposed positive feedback loops involves the calcium/calmodulin-dependent protein kinase II (CaMKII) kinase-phosphatase system [15–19]. CaMKII activation is governed by Ca2+/calmodulin binding and is prolonged beyond fast-decaying calcium transients by its autophosphorylation [23]. Autophosphorylation of CaMKII at the residue theronine-286 in the autoregulatory domain (Thr286) occurs after calcium/calmodulin binding and enables the enzyme to remain autonomously active after dissociation of calcium/calmodulin [24] (see Materials and Methods). In turn, as long as CaMKII stays activated it is reversibly translocated to a postsynaptic density (PSD)-bound state where it interacts with multiple LTP-related partners structurally organizing protein anchoring assemblies and therefore potentially delivering α-amino-3-hydroxyl-5-methyl-4-isoxazole-propionate acid (AMPA) receptors to the cell surface [23,25–28]. The direct phosphorylation of the AMPA receptor GluR1 subunit by active CaMKII enhances AMPA channel function [29,30]. The network involving CaMKII is particularly appealing in terms of learning and memory maintenance since N-methyl-D-aspartate receptor (NMDA-R)-dependent LTP requires calcium/calmodulin activation of CaMKII, potentially expressed by the phosphorylation level or the number of AMPA receptors, or both [19,27,28,31–33]. However, the role of CaMKII beyond LTP induction remains controversial [34–36]. Finally, there is experimental evidence for the involvement of proteins associated with CaMKII activity (cyclic adenosine monophosphate (cAMP)–regulated protein kinase A (PKA), protein phosphatase 1 (PP1), and calcineurin) in LTP and LTD [37–40]. We emphasize that multiple mechanisms supporting LTP/LTD induction and expression are likely to be present in synapses of different regions—we focus here on synapses for which the above statements have been shown to apply, e.g., the CA3-CA1 Schaffer collateral synapse (see review by Cooke and Bliss [41]).
Modeling studies have shown that a system including CaMKII and associated pathways could be bistable in a range of calcium concentrations including the resting level—a necessary requirement for the maintenance of long-term changes [15,17,18,42]. In such models, the two states correspond to two stable phosphorylation levels of the CaMKII protein for a given calcium concentration, i.e., a weakly (DOWN) and a highly phosphorylated state (UP). A transition from the DOWN to the UP state which could underlie long-term potentiation (LTP) can be induced by a sufficiently large and prolonged increase in calcium concentration. However, the opposite transition which could underlie depotentiation or LTD only occurs under unrealistic conditions, for example decrease of calcium concentration below resting level. Furthermore, it has not been considered how these biochemical network models behave in response to calcium transients evoked by experimental protocols that are known to induce synaptic plasticity such as STDP, which has been shown to rely on kinase (CaMKII) and phosphatase (calcineurin) activation [43]. Rubin et al. reproduce experimental results on STDP using a model detector system which qualitatively resembles the protein network influencing CaMKII, but this model does not exhibit bistability [44]. Other studies on biochemical signal transduction pathways including CaMKII showed that the AMPA receptor activity can reproduce bidirectional synaptic plasticity as a function of calcium [45,46]. However, realistic stimulation protocols were not investigated in these models, and again they do not show bistability.
In this paper, we consider a realistic model of protein interactions associated with CaMKII autophosphorylation through calcium/calmodulin and dephosphorylation by protein phosphatase 1 in the PSD. We first study the steady-state phosphorylation properties of CaMKII with respect to calcium and changing levels of PP1 activity. Conditions are elaborated for which the system allows for “LTP” and “LTD” transitions in reasonable ranges of calcium concentrations. We then demonstrate the ability of the CaMKII system to perform LTP- or LTD-like transitions in response to STDP stimulation protocols. We expose the CaMKII system to calcium transients evoked by pairs of presynaptic and postsynaptic spikes with a given time lag and show that short positive time lags evoke transitions from the DOWN to the UP state and short negative time lags lead to transitions from the UP to the DOWN state. We demonstrate furthermore that the CaMKII model qualitatively reproduces experimental plasticity outcomes for presynaptic stimulation protocols. Finally, we consider the transition behavior in response to purely presynaptic or postsynaptic spike-pair stimulation protocols.
We investigate in this paper a realistic model for the protein network of the postsynaptic density, focusing on the pathways affecting the phosphorylation dynamics of CaMKII localized in the PSD. The model describes the calcium/calmodulin-dependent autophosphorylation of CaMKII. Phosphorylation of a CaMKII subunit by its neighboring subunit requires calcium/calmodulin to bind to the substrate subunit. The catalytic subunit is active if bound to Ca2+/calmodulin, or phosphorylated (see Figure 1A–1E). Dephosphorylation of phosphorylated CaMKII subunits by PP1 in the PSD is implemented according to the Michaelis-Menten scheme. We also take into account how calcium/calmodulin influences PP1 activity via a protein signaling cascade. PP1 is inhibited by phosphorylated inhibitor 1 (I1). The phosphorylation level of inhibitor 1 is in turn controlled by the balance between a pathway phosphorylating I1 (through cAMP–PKA) and a pathway dephosphorylating I1 (through calcineurin). Therefore, calcineurin activation by calcium/calmodulin increases PP1 activity, while calcium/calmodulin-dependent activation of the cAMP–PKA pathway decreases PP1 activity (Figure 1F). Finally, we model postsynaptic calcium and postsynaptic membrane potential dynamics induced by presynaptic and postsynaptic spikes in order to investigate the effects of spike-induced calcium transients on the dynamics of the system. Details of the model can be found in the Materials and Methods section.
In this and the following section we investigate how the steady-state values of the total concentration of phosphorylated CaMKII subunits, Sactive, depend on the concentration of calcium and the dephosphorylation activity. We also study how the steady-state behavior changes with the number of interacting subunits in the cluster. We start by exploring Ca2+/calmodulin-stimulated autophosphorylation of CaMKII at a fixed dephosphorylation activity. This will allow us later to better understand how the parameters of the signaling cascade controlling dephosphorylation activity affect the phosphorylation behavior of CaMKII. To do this, we set the PP1 dephosphorylation activity to a constant, independent of the calcium concentration (this is equivalent to removing the red lines in Figure 1F except for the interaction between CaMKII and PP1). The PP1 dephosphorylation activity is the product of k12, the maximal dephosphorylation rate, and D, the free PP1 concentration (see Equation 6).
Figure 2A shows the steady-state concentration of phosphorylated CaMKII subunits as a function of the calcium concentration for 2, 4, 6, and 8 functionally connected subunits in the CaMKII cluster. The graphs show that in all cases there exists a range of calcium concentration for which the system is bistable (region between the diamond and the circle in the case of the six-subunit model). In the bistable region, three steady-states are present. The top and the bottom steady-states (depicted by the thick full lines) are stable, whereas the intermediate one (dashed thin lines) is unstable. The branch of unstable steady-states separates the basins of attraction of the highly and the weakly phosphorylated stable steady-states. This means that the system will converge to the UP state if it is initially above this line, while it will converge to the DOWN state if it is below this line. As in other studies on CaMKII bistability, the bistable phosphorylation behavior emerges from the combination of strong cooperativity of CaMKII autophosphorylation and the saturation of the per-subunit dephosphorylation rate, k10 (see Equation 6 in Materials and Methods ), at high phosphorylation levels [15,17,18]. This saturation arises from the Michaelis-Menten approach employed to describe dephosphorylation, which is valid if the enzyme (PP1) is present in small amounts compared to the substrate (phosphorylated subunits). This is plausible since the CaMKII protein is localized at high concentrations in the PSD [24,27,47].
Figure 2A demonstrates that the increasing saturation of the per-subunit dephosphorylation rate with increasing number of interacting subunits in the holoenzyme ring plays a crucial role in the extent of the bistable region. Whereas the difference between the two- and the four-subunit model is very pronounced, increasing further the number of subunits has less and less impact on the size of the bistable region—it still increases substantially when this number goes from four to six, but there is almost no noticeable difference between the six- and the eight-subunit model. The effect of the number of subunits on the extent of the bistable region is mainly due to an increase in the range of stability of the UP state with increasing subunit number, since the stronger saturation of the per-subunit dephosphorylation rate becomes apparent in the highly phosphorylated state only (see Equation 6). On the other hand, the range of stability of the DOWN state is essentially unaffected by the number of subunits, since it is mostly controlled by the balance between the dephosphorylation rate and the probability of the initiation step to occur. Interestingly, experimental data indicate that the number of functionally coupled subunits in a CaMKII holoenzyme ring is six [48–50], which could be a good compromise between having both a relatively small number of subunits and a large bistability range. In the following, we consider exclusively a model with six subunits.
How the location and the extent of the bistable region changes with respect to the PP1 dephosphorylation activity is shown in Figure 2B. The curves depict the boundaries of the bistable region for the six-subunit model in the PP1 activity—calcium concentration plane, for three values of the total calmodulin concentration, CaM0 (indicated by the three different colors). For each value of CaM0, the colored area shows the bistable region in which the UP and the DOWN states coexist. Above the colored area, only the DOWN state is present, while below that area only the UP state is present. The resting calcium concentration can be included in the bistable region, provided the PP1 activity is chosen accordingly (e.g., (k12·D) = 6.648 μM/s for CaM0 = 0.1 μM in Figure 2B).
The right-hand boundary of the bistable region in Figure 2A corresponds to a down-to-up switching threshold: if the calcium concentration increases persistently above this level, the CaMKII will converge from a weakly phosphorylated to a highly phosphorylated state (down-to-up switching). Hence, we define the range above this right-hand bifurcation point “LTP window”. It corresponds to high calcium concentrations, consistent with experimental data on the range of calcium concentrations leading to LTP.
The available experimental data also suggest that (i) at resting calcium concentrations, no transitions should occur (both UP and DOWN states should be stable), (ii) for intermediate calcium concentrations (higher than resting concentration, but lower than the down-to-up switching threshold), LTD transitions should occur. This would happen if the UP state was no longer stable in an intermediate range of calcium concentrations—in such a scenario, the UP state would be stable in two disconnected regions, one around resting calcium concentration, and the other one at high calcium concentrations. The region where the UP state would not be stable could be called “LTD window” since the system would exhibit LTD (up-to-down switching) whenever the calcium concentration stays in that region for a sufficiently long time, i.e., the CaMKII would converge from a highly phosphorylated state to a weakly phosphorylated state in this range of calcium. The scenario depicted in Figure 2 seems at odds, however, with this picture.
How can the steady-state picture of Figure 2A be modified to obtain such an LTD window? A possible scenario is to take into account the protein signaling cascade governing PP1 dephosphorylation activity in a calcium/calmodulin-dependent manner (see Figure 1F). In this way, the active concentration of PP1, D, changes with calcium, and the region of bistability is no longer defined by a horizontal line in Figure 2B. Rather, the location and extent of the LTD and the LTP windows are given by the intersections of the curve describing how the steady-state PP1 activity changes with calcium concentration with the curves specifying the location of the left- and the right-hand boundary of the bistable region in the PP1 activity–calcium concentration plane.
Figure 3B shows an example in which the steady-state PP1 activity (k12 · D) versus calcium concentration curve (purple line) intersects the bifurcation lines (red lines) four times, such that an LTD window emerges in a range of intermediate calcium concentrations. As Figure 3B shows, this can be obtained whenever the PP1 activity has a sufficiently large peak at some intermediate calcium concentrations. This peak has to be such that in a range of calcium concentrations, PP1 activity is above both bifurcation lines (region between intersection points 2 and 3 in Figure 3B). As discussed above, only the DOWN state is stable in this region. This PP1 peak is in turn obtained due to PKA activating at higher calcium concentration than calcineurin, since the balance between calcineurin and PKA activity determines the level of PP1 inhibition via inhibitor 1 (see Figures 1F and 3A). Hence, the peak in steady-state PP1 concentration at intermediate calcium concentrations is due to a relative increase in calcineurin activity with respect to PKA activity in this range (compare Figure 3A and 3B). To include the calcium resting concentration, Ca0 (marked by the vertical thin line in Figure 3), in a region of bistability, PP1 activity at Ca0 has to reside in between the bifurcation lines. The fourth intersection point defines the down-to-up switching threshold, i.e., the left-hand boundary of the LTP window (see point 4 in Figure 3B and 3C). The range of bistability between points marked 3 and 4 in Figure 3C emerges from the declining PP1 activity (purple line in Figure 3B) crossing the ascending range of bistability (red shaded area in Figure 3B). These opposing trends lead to a narrow range of bistability at high calcium concentrations in the example presented here since the intersections of both define the borders of the bistable region.
In practice, the location of these four intersection points can be chosen by adjusting parameters describing the calcium/calmodulin-dependent activation of PKA and calcineurin activity (see Materials and Methods section for more details). We can obtain four such parameters (PKA base and maximal activity
and kPKA, respectively, the PKA half activity concentration KPKA and the PKA Hill coefficient nPKA (see Equation 9)) by simultaneously solving four equations, i.e., one for each of the four intersection points 1, 2, 3, and 4 at Ca = 0.09, 0.22, 0.36, and 0.37 μM (see Figure 3B). Figure 3A displays the resulting instantaneous calcium/calmodulin-dependent phosphorylation vPKA and dephosphorylation vCaN rates of inhibitor 1 (see Equation 23 in Materials and Methods) leading to the steady-state PP1 concentration scenario shown in Figure 3B by the full purple line. The parameters obtained through this procedure can be found in Table 1B). Table 1B also shows the ranges of values of each parameter for which the above-described behavior is qualitatively observed. These ranges are obtained varying each parameter while keeping the remaining three constant. This shows that the system is relatively robust to parameter changes. It reacts most sensitively to changes in
, whose value can be varied by about 14% in both directions, while the other parameters can be varied over a range of about 100% and even ∼800% for kPKA. Note that the choice of the maximal calcineurin activity, which basically controls the height of the PP1 peak shown in Figure 3B, depends also on constraints discussed in the following section.
The system is also robust to changes in the total CaMKII concentration, as shown in Figure 3C where we compare the bifurcation diagrams for CaMKII0 = 16.67 μM (blue line) and 8.33 μM (green line) provided the PP1 activity is rescaled accordingly by using
1/s in the latter case (other parameters remain unchanged). Note that both values of CaMKII0 cover a range of CaMKII concentration that encompasses experimental estimates (∼10 μM) for the PSD [24,27,47]. The dashed lines in Figure 3B show the position of the bistable region (dashed red lines) and the steady-state PP1 activity (dashed purple line) in the PP1 activity-calcium plane for CaMKII0 = 8.33 μM and
1/s (see Table 1A and 1B for other parameters).
To summarize the results so far, the model behavior is such that: (i) the calcium resting concentration is included in a region of bistability, giving rise to two stable steady-states—DOWN and UP—at resting conditions (marked by the diamond and the cross in Figure 3C, respectively); (ii) in a region of intermediate calcium concentrations, only a weakly phosphorylated steady-state exists (the LTD window, between filled circles marked 2 and 3); (iii) conversely, at high calcium concentrations, only the highly phosphorylated steady-state is stable (the LTP window, beyond filled circle marked 4). This scenario is now qualitatively consistent with experimental data. Note that, in contrast with Zhabotinsky (2000), our model does not require an unrealistic high phosphorylated inhibitor 1 concentration at resting calcium concentration to have a stable highly phosphorylated CaMKII state (compare green line in Figure 3B and [17] at this concentration).
Up to this point, we have investigated the steady-states of the CaMKII kinase-phosphatase system as a function of the intracellular calcium concentration. In experimental conditions, however, synaptic modifications are evoked by calcium transients resulting from experimental stimulation protocols inducing synaptic plasticity. Hence, the occurrence of transitions between weakly and highly phosphorylated states in the model needs to be examined in response to such calcium dynamics. Here, we explore in which conditions the spike-timing dependent plasticity (STDP) protocol as well as presynaptic or postsynaptic stimulation protocols alone induce such transitions.
For the STDP protocol, we use a standard repetitive stimulation protocol (60 pairs at 1 Hz, see experiments by Bi and Poo [2]). Each stimulation pair consists of a presynaptic spike at time tpre and a back-propagating postsynaptic action potential occurring at time tpost = tpre + Δt. In experimental conditions, LTD is evoked for short negative Δts, while LTP is evoked for short positive Δts [1,2,51–53].
Figure 4 shows the time course of calcium concentration transients evoked by one pair of a presynaptic spike and a back-propagating action potential (BPAP) at different Δts. An isolated postsynaptic spike generates a calcium transient of amplitude ΔCapost, due to opening of calcium channels induced by the depolarization caused by the BPAP. Likewise, an isolated presynaptic spike generates another calcium transient of amplitude ΔCapre, due to NMDA channel opening. Below, we will vary systematically the size of ΔCapre, keeping the ratio constant, ΔCapost / ΔCapre = 2 [54]. See Materials and Methods for details of the model.
What happens when presynaptic and postsynaptic spikes are sufficiently close together so that their respective calcium transients overlap?
When the parameters of the model are chosen accordingly, the model reproduces qualitatively the experimental results in response to the STDP stimulation protocol: (i) short positive Δt stimulation protocols move CaMKII from the weakly phosphorylated state to the highly phosphorylated state. Starting from the UP state, no transition occurs. (ii) A system at rest at the UP state is switched to the DOWN state by short negative Δt protocols, whereas the same protocol does not evoke transitions from the DOWN to the UP state. (iii) Large positive and negative Δts do not evoke transitions between the DOWN and the UP states. We show in Figures 5–7 (red lines) the behavior of the model for parameters shown in Table 1A–1C, with kCaN = 18 1/s and ΔCapre = 0.17 μM.
Figure 5 shows the dynamics of the system for the whole stimulation protocol, and until the system has reached the final steady-state except for Figure 5A and 5B, which depicts the time course of the calcium concentration for one spike pair presentation only. Figure 5C and 5D shows the time course of active PP1, while Figure 5E and 5F shows the dynamics of phosphorylated CaMKII subunit concentration. The left column shows the dynamics of the system when it is initially in the DOWN state (low concentration of phosphorylated CaMKII subunits) for two representative time lags (Δt = 15 ms and 100 ms), while the right column shows the dynamics when it is initially in the UP state, again for two representative time lags (Δt = −50 and −10 ms).
To understand why the system exhibits transitions in specific ranges of Δt, is it first crucial to examine how PP1 activation depends on Δt. For the value of ΔCapre chosen here, PP1 activation is the largest at short negative time differences, since such values of Δt maximize the time spent by the system in the range of calcium concentrations close to PP1 peak activation (see Figure 3B). On the other hand, PP1 activation is minimal for short positive time lags since Ca goes transiently to high concentrations and spends a short time at intermediate values. Let us now focus on the situation in which the system is initially in the DOWN state. During the stimulation protocol, two situations can arise. For short positive time differences (as, for example, the 15 ms case shown in Figure 5), the increase in PP1 activity is low and insufficient to counterbalance the large increase in the concentration of phosphorylated CaMKII subunits, since high calcium transients strongly favor the autophosphorylation process which outweighs the low dephosphorylation activity. Hence, the system reaches a high phosphorylation level during the stimulation protocol and converges gradually toward its equilibrium value in the UP state thereafter. On the other hand, for negative and large positive time lags, the increase in PP1 activity is large enough to counterbalance the calcium/calmodulin triggered autophosphorylation, i.e., CaMKII stays dephosphorylated and remains in the DOWN state (see for example the 100 ms case in Figure 5).
When the system is initially in the UP state, the concentration of phosphorylated CaMKII subunits again depends on the competition between dephosphorylation by PP1 and autophosphorylation progress during the protocol. Again, we have two possible outcomes of the protocol: either the PP1 concentration becomes large enough such that the system gets sufficiently dephosphorylated and moves in the basin of attraction of the DOWN state during the stimulation protocol (this occurs for example for the −10 ms case shown in Figure 5); or it is not large enough and autophosphorylation prevails, i.e., the system remains in the basin of attraction of the UP state. For the parameter set used in Figure 5, this happens for large negative and positive time lags.
Another way of visualizing the dynamics during and after the STDP protocol consists in plotting the trajectory of the system in the concentration of phosphorylated CaMKII subunits Sactive–PP1 activity plane. This is done for several values of Δt in Figure 6A and 6B. The DOWN and the UP stable steady-states of the CaMKII phosphorylation level are shown by the diamond and the cross, respectively (located at the intersections of the Sactive and (k12 · D) nullclines). In Figure 6A the system is initially in the DOWN state, whereas it is initially in the UP state in Figure 6B. In both Figure 6A and 6B, the end of the stimulation protocol corresponds to the point at which PP1 and Sactive stop to oscillate, and there is a sharp turn of the trajectories in the plane. In this plane, the separatrix (dotted black line) marks the boundary between the basins of attraction of both stable steady-states. Depending on the position of the system at the end of the stimulation protocol relative to this separatrix, the system relaxes either to the UP or the DOWN state. The separatrix is obtained by adjusting numerically Δt to be at the boundary between the regions in which a transition to the UP (respectively, DOWN) state occurs or not.
The outcomes of the deterministic STDP protocols for Δt values from −100 to 150 ms are summarized in Figure 7A (red line). We consider a large population of independent synapses submitted to the same protocol, in which initially half of the synapses are in the DOWN state and the other half in the UP state. Figure 7A shows the relative change in the fraction of synapses in the UP state as a function of Δt (+1 means all synapses initially in DOWN where switched to UP; 0 means no change; −1 means all synapses in UP have switched to DOWN). There is a range of values of Δt (from 10 to 16 ms) for which all synapses initially in the DOWN state switch to the UP state (LTP). LTD, or up-to-down transitions of the synapses initially in the UP state, is observed in a range of Δt values from −14 to −2 ms (see red line in Figure 7A).
The CaMKII kinase-phosphatase system in the PSD is composed of a few molecules only (∼30 CaMKII holoenzymes [56]), hence stochastic fluctuations potentially play an important role (see [57]). The CaMKII system is also exposed to fluctuating calcium transients stemming from stochastic neurotransmitter release, stochastic channel opening, and the stochastic nature of neurotransmitter as well as calcium diffusion [54,58]. It is therefore necessary to investigate the dynamic behavior of the CaMKII system in the presence of noise. Here we choose for simplicity to introduce fluctuations in calcium transients exclusively.
Two sources of noise are introduced in the calcium dynamics simulations: (i) the NMDA receptor maximum conductance is drawn at random at the occurrence of each presynaptic spike, and (ii) the maximum conductance of the voltage-dependent calcium channel is drawn at random at the occurrence of each postsynaptic spike. Both conductances are drawn from binomial distributions similar to those measured in experiments [58,59] (see Materials and Methods for more details). Some examples of noisy calcium transients are shown in the inset of Figure 7A (dashed lines; the full line depicts the average transient).
Again, we consider a large population of independent synapses exposed to stochastic stimulation protocols. N = 300 independent synapses are simulated, 150 initially in the DOWN and 150 in the UP state. Applying the stimulation protocol leads to stochastic transitions between UP and DOWN states. Figure 7A shows the relative change in the fraction of synapses in the UP state as a function of Δt, for kCaN = 18 1/s and 20 1/s. For example, a relative change of −0.8 for Δt = −15 ms (kCaN = 20 1/s case) means that 120 of the synapses in the UP state (out of the 150) switched to the DOWN state in response to this protocol, while none of the 150 synapses in the DOWN state experienced a down-to-up transition during the Δt = −15 ms stimulation.
The variability in maximum NMDA and CaL current conductances results in a variability of calcium transients around the mean transients. The consequence of this variability in calcium transients is that, while the shape of the PP1 level versus Δt is qualitatively unchanged, the PP1 level reached during stimulation protocols is significantly reduced. This is due to the fact that the variability in calcium transients decreases the time spent by the system at calcium concentrations which maximize PP1 buildup. Hence, the probability that the PP1 level is high enough to make the system switch to the DOWN state becomes small for kCaN = 18 1/s (the value used in deterministic simulations). Consequently, the up-to-down transition probability at short negative time lags is low and LTD is effectively absent in this case (see green line with squares in Figure 7A). However, the LTD probability becomes larger as kCaN is increased. Figure 7A shows an up-to-down switching probability of about 0.93 for short negative time lags with kCaN = 20 1/s (at Δt = −10 ms). It also shows that for this value of kCaN there exists a small but finite probability of eliciting LTD transitions for large positive and large negative time lags, due to variability in calcium transients. The range of values of kCaN for which the LTD probability for short negative Δt is larger than 0.5 AND the LTD probability for large positive Δt is smaller than 0.5 is 19 1/s < kCaN < 20 1/s. To summarize, down-to-up transitions occur robustly for a large range of parameters at short positive time lags. On the contrary, the range of short negative values of Δt for which UP to DOWN switches are observed is less robust to noise (see Discussion).
The role of protein phosphatases in synaptic plasticity has been investigated through the application of phosphatase inhibitors during the presentation of stimulation protocols inducing synaptic changes [21,37,43,60]. These experiments have shown that phosphatase inhibitors prevent LTD while sparing LTP. We investigate the effect of phosphatase inhibitors in our model by gradually reducing the dephosphorylation activity of PP1 and study the changes in the steady-states of the phosphorylated CaMKII subunit concentration and in the transition behavior.
Since the steady-state PP1 concentration is given by Dsteady-state = D0 / (1 + (I0k13vPKA) / (k-13vCaN)) (see Materials and Methods), scaling down D0 corresponds to a reduction of the steady-state PP1 activity given by the purple line in Figure 3B. Consequently, the intersections between the boundaries of the bistable region (given by the red lines in Figure 3B) and the PP1 activity change. In other words, the locations and ranges of the LTD and the LTD windows change as a function of the level of PP1 inhibition. Scaling down the total PP1 concentration leads to a diminution of the size of the LTD window and to the emergence of a second LTP window at low calcium concentrations (see the 80% case shown by the green line in Figure 8A and 8B). Decreasing further protein phosphatase strength makes the LTD window disappear and a large LTP window emerges starting at low calcium concentrations (see 60% and 40% cases in Figure 8A and 8B). Finally, reducing the PP1 concentration below ∼40% results in a loss of the stability of the DOWN state at resting calcium concentrations, leaving the UP state as the only stable steady-state for all calcium concentrations.
Figure 8C shows how LTP/LTD transitions are affected by reduced total PP1 concentration, when the model is exposed to the STDP stimulation protocol with noisy calcium transients. Consistent with experiments, reducing the PP1 concentration by 20% leads to a loss of LTD transitions (see green line in Figure 8C), while increasing the range of Δt for which LTP transitions are observed. Further reduction of the PP1 concentration to 60% and more results in up-to-down transitions for all time differences Δt (see red line in Figure 8C).
Experiments on STDP show that presynaptic or postsynaptic spikes alone at the same stimulation frequency (1 Hz) do not evoke any plasticity (see for example [55]). To check the behavior of the model in this situation, we expose the CaMKII system to either 60 presynaptic or postsynaptic spikes of different frequencies and show the transitions results for the deterministic calcium transient case in Figure 7B by red lines. 60 presynaptic spikes alone do not evoke any transitions at low frequencies (1–3 Hz). For presynaptic stimulations in the range 4–18 Hz, up-to-down transitions occur, and for frequencies equal and larger to 19 Hz the CaMKII system is switched from the DOWN to the UP state (see full red line in Figure 7B). The transition outcomes change dramatically if stimulation occurs exclusively with postsynaptic spikes. 60 postsynaptic spikes do not evoke transitions up to a stimulation frequency of 84 Hz (see dashed red line in Figure 7B). Above 85 Hz, transitions from DOWN to UP occur. Note that the spike pairs during the STDP spike-pair stimulation protocol employed above are presented at a frequency of 1 Hz only, i.e., presynaptic or postsynaptic spikes alone at this frequency do not evoke transitions, consistent with experiments [55]. Note also that the model does not incorporate frequency-dependent attenuation of EPSPs and BPAPs. Attenuation of BPAPs at high frequencies could prohibit down-to-up transitions in the post protocol at any frequency.
We also expose the CaMKII system to fluctuating calcium transients evoked by presynaptic or postsynaptic frequency stimulations. The implementation of calcium transient noise is exactly as for STDP spike-pair protocols above. The average relative changes in the fraction of synapses in the UP state for these stimulations are shown for varying frequencies in Figure 7B for kCaN = 18 1/s and 20 1/s (N = 300 synapses). Presynaptic stimulations at frequencies between ∼2 and ∼16 Hz evoke a net increase of synapses in the DOWN state, while stimulation above ∼16 Hz lead to LTP transitions. Again, no up-to-down transitions are observed with postsynaptic stimulation alone, while stimulation frequencies above ∼50 Hz yield a net increase of the number of synapses in the UP state.
Another simple generalization of the STDP protocol consists in exposing the system to purely presynaptic spike pairs, or purely postsynaptic spike pairs. Spike pairs with a fixed inter-spike interval Δt are presented 60 times at varying frequencies. We investigate the transition behavior of the model for varying inter-spike intervals and for different presentation frequencies. This is a protocol for which plasticity outcomes have, to our knowledge, not yet been characterized.
Presynaptic spike pairs lead to up-to-down transitions for all values of Δt at a frequency of f = 1 Hz, consistent with the fact that presynaptic stimulation of single spikes at 2 Hz evokes such transitions (see Figure 7B). On the other hand, purely postsynaptic spike pairs evoke down-to-up transitions in a very narrow range of values of Δt (from 3 to 8 ms) at f = 1 Hz. In other words, postsynaptic spike pairs have to be presented sufficiently closely in time for the phosphorylation changes to sum up, so that the system converges to the UP state. Decreasing the spike-pair presentation frequency f leads to transitions in narrower ranges of Δt for the presynaptic and the postsynaptic protocol (e.g., f = 0.5 Hz; presynaptic spike pairs with 0 < Δt ≲ 300 ms lead to up-to-down transitions, postsynaptic spike pairs with 3 ≲ Δt ≲ 6 ms evoke down-to-up transitions). At f = 0.1 Hz, there are no longer any transitions in the purely presynaptic stimulation protocol and only a small down-to-up transition probability exists for postsynaptic spike pairs (3 ≲ Δt ≲ 4 ms; unpublished data).
The difference in transition outcomes between presynaptic and postsynaptic spike-pair stimulations can be understood by inspecting the calcium transients evoked by both stimulation protocols. The maximum calcium amplitude reached by pairs of postsynaptic spikes is much larger than the calcium amplitude evoked by presynaptic spikes (maximum amplitude for Δt = 10 ms is ∼0.45 μM for presynaptic spike pairs and ∼0.7 μM for postsynaptic spike pairs with the parameters given in Table 2). On the other hand, pairs of presynaptic spikes evoke calcium transients which last much longer than postsynaptic pairs of spikes (compare the different time scales in Figure 4). The high calcium transients evoked by postsynaptic spike pairs strongly activate the cAMP–PKA pathway and therefore suppress PP1 activity. This suppression, together with the strong CaMKII autophosphorylation due to high calcium concentrations, leads to down-to-up transitions in response to closely spaced postsynaptic spike pairs. Even single postsynaptically evoked calcium transients reach calcium levels sufficiently high to activate the cAMP–PKA pathway. This explains why purely postsynaptic stimulation at varying frequencies does not go through an LTD range (see Figure 7B and compare the small PP1 buildup in the inset in Figure 5D in response to the postsynaptically evoked calcium transient in the Δt = −50 ms protocol). In contrast, the long-lasting calcium transients evoked by purely presynaptic spike pairs make the system spend a lot of time in calcium ranges maximizing PP1 buildup. This leads to strong dephosphorylation of CaMKII by PP1 which cannot be counterbalanced by moderate autophosphorylation evoked by intermediate calcium levels. Hence up-to-down transitions are evoked for closely spaced presynaptic spike pairs. Again, this explains also why ongoing presynaptic stimulation at different frequencies evokes LTD at low presentation frequencies before the calcium transients are adding up sufficiently to activate the cAMP–PKA pathway and evoke strong autophosphorylation (this happens above f ≈ 16 Hz in Figure 7B).
The model with the parameter set discussed until this point reproduces qualitatively experimentally observed transition outcomes of the STDP protocol. We now discuss how changing parameters affect the transition behavior. Numerical investigations of the model show that two characteristics of the CaMKII system dynamics are crucial: (i) how the level of PP1 activity at the end of the stimulation protocol depends on Δt; (ii) the time course of autophosphorylation and dephosphorylation of CaMKII, and of PP1 buildup, influences the number of spike pairs during the stimulation protocol necessary to evoke down-to-up- or up-to-down transitions. We focus here for the sake of simplicity on deterministic calcium transients.
We have shown above that a peak in steady-state PP1 activity at moderate calcium concentrations occurs if the cAMP–PKA pathway activates at higher calcium concentrations than the calcineurin pathway (see Figure 3A). Here, we address the question of how the dynamics of the PP1 activity during the stimulation protocol changes as a function of the balance between the activation of both pathways. Changing the NMDA-R mediated calcium amplitude ΔCapre and the BPAP evoked calcium response, keeping their ratio constant (ΔCapost = 2 · ΔCapre ), and also keeping the parameters of the protein signaling cascade constant, allows us to change the balance between the activation of both pathways and to get an insight into what controls the dependence of the PP1 activity level on Δt. Figure 9A and 9C show the PP1 activity level immediately after the presentation of one and 60 spike pairs, respectively, as a function of Δt, for different values of ΔCapre. Note that the dependence of PP1 activity with respect to Δt after the presentation of one spike pair (Figure 9A) is qualitatively preserved after the entire stimulation protocol of 60 spike-pair presentations (Figure 9C).
Figure 9B represents the change in PP1 activity induced by a single spike pair, computed from Equation 34, as well as contributions of the PKA and calcineurin pathways to this change. The dashed lines in Figure 9B show the contribution of the cAMP–PKA pathway to the change in PP1 activity (second term in the integral of Equation 34). This contribution is negative, since this pathway decreases PP1 activity. Due to the high half activation calcium concentration of vPKA(C) (see blue line in Figure 3A), the cAMP–PKA pathway is sensitive to high calcium elevations only. Hence, the negative contribution of this pathway increases drastically when the calcium amplitude ΔCapre increases, since the calcium transients spend more time in the range of cAMP–PKA activation. In response to the supralinear superposition of the NMDA-R and the BPAP evoked currents at short positive time differences, this pathway ensures a low level of PP1 activity in this range.
The dotted lines in Figure 9B show the contribution of the calcineurin pathway to the change in PP1 activity. This contribution is positive, since this pathway increases PP1 activity. The calcineurin pathway activates at lower calcium concentrations than the PKA pathway (see red line in Figure 3A; integral of the first part of Equation 34), and therefore this pathway is sensitive to the time spent by the system at intermediate and high calcium levels. This calcineurin contribution starts to increase at negative time differences (when calcium transients induced by pre- and post-synaptic spikes start to interact), reaches a peak close to Δt = 0, and then decreases slowly with Δt.
The sum of the two contributions yields the net change in PP1 activity (full lines of Figure 9B). For ΔCapre = 0.17 μM, the value chosen in the rest of the paper, the PP1 change versus Δt curve shows first a peak at negative Δt (due to increase in calcineurin activity in this range), followed by a trough at positive Δt (due to the strong increase in PKA activity in this range). There is a secondary peak of PP1 change at larger values of Δt (∼100 ms) because calcineurin activity decays more slowly with Δt than PKA activity. However, this peak is smaller than the peak at negative Δt, which explains why LTD is observed at short negative Δt but not large positive ones.
Changing the size of the calcium transients potentially changes qualitatively the shape of this curve because it affects the time spent by the system in different calcium concentration ranges. For example, decreasing the size of the calcium transients weakens considerably the PKA pathway, leading to an increase in PP1 activity for negative as well as positive values of Δt. On the other hand, increasing the calcium transients leads to a strengthening of the PKA pathway relative to the calcineurin pathway, leading to a much smaller peak in the PP1 change curve at short negative Δt. This peak eventually vanishes for large enough ΔCapre ≥ 0.4 μM (unpublished data).
Since transitions are a result of an unbalance between autophosphorylation and dephosphorylation mediated by PP1, the Δt range for which transitions are evoked or prevented can therefore be controlled by means of the calcium amplitude. If the calcium amplitude is decreased in the model, no transitions are observed any more (e.g., for ΔCapre = 0.15 μM). On the other hand, increasing the calcium amplitude extends the Δt range for which up-to-down and down-to-up transitions are evoked (ΔCapre = 0.18 μM, LTD range: [−21... −3] ms and LTP range: [3...33] ms; unpublished data). These predictions could be checked experimentally by changing the external calcium concentration and therefore changing the calcium influx evoked by presynaptic and postsynaptic spikes.
To summarize, there exists a range of ΔCapre for which the PP1 level at the end of the stimulation protocol as a function of Δt exhibits a maximum for short negative Δts and is low enough to be outweighed by autophosphorylation for short positive Δts. This is a requirement for a system to exhibit LTD-like transitions at short negative time intervals only, and LTP-like transitions at short positive time intervals only. However, these qualitative features of the PP1 activation versus Δt curve are not sufficient to ensure that STDP protocol stimulations with short negative time lags lead to transitions from the UP to the DOWN state only. In addition, (i) the absolute level of PP1 activity for short negative Δt stimulations must be high enough to evoke up-to-down transitions; (ii) at the same time, the total PP1 level has to be low enough such that for large negative and large positive time lag stimulations the system remains in the UP state and that for short positive Δt protocols autophosphorylation prevails over dephosphorylation leading to down-to-up transitions. These two criteria can be met by changing the maximal calcium/calmodulin-dependent calcineurin activity kCaN, which changes the amplitude of the peak of the PP1 vs Ca2+ steady-state curve at moderate calcium concentrations (purple lines in Figure 3B). Consequently, this parameter allows us to control the PP1 level attained during the stimulation protocol for all Δts. In particular, the range 16.6 ≤ kCaN ≤ 18.1 1/s fulfills the two requirements above (Figures 5 and 6 and the red as well as the green lines in Figure 7 use kCaN = 18 1/s).
The autophosphorylation rates k6, k7, k8, the maximal dephosphorylation rate k12, and the total PP1 concentration D0 determine the velocity of autophosphorylation as well as dephosphorylation dynamics of CaMKII and the dynamics of the PP1 response during exposure to the STDP protocol. We introduce scaling parameters R and Q such that varying R and Q does not change the steady-state behavior of the CaMKII system (see Figure 3B and 3C) nor the maximum PP1 activity reached during the stimulation but only the dynamics of the system. Both scaling parameters are varied extensively in order to investigate their impact on the transition behavior of the model, i.e., 0.002 ≤ R ≤ 2 and 0.083 ≤ Q ≤ 1.67.
R is chosen such as to control the dephosphorylation kinetics, while leaving the PP1 activity, given by the product (k12 · D), constant. This leaves the steady-state behavior intact since it depends on this product only. Hence, in the following simulations, k12 and D0 are replaced by
and
, where k12 and D0 are the “control” parameters listed in Table 1B. R controls how fast the dephosphorylation dynamics responds to calcium transients, since the PP1 buildup during the presentation of the stimulation protocol and the decay dynamics thereafter depend on the value of D but not on k12 (see Equation 31 in Materials and Methods). Figure 10A shows the PP1 activity time course for three different values of R during and after the STDP stimulation protocol with Δt = 15 ms, i.e., for R = 0.002, 0.078, and 1.
Q scales the autophosphorylation rates k6, k7, and k8 together with the maximal dephosphorylation rate k12 as
(with x = 6,7,8, and 12), where all the rates kx take the values listed in Table 1B. This corresponds to a rescaling of the y-axis in Figure 3B, i.e., the points of intersection between the bistable range (red shaded areas) and the PP1 activity (purple lines) are kept fixed and therefore the steady-state concentration of phosphorylated CaMKII subunits (Figure 3C) is left unchanged. We illustrate the impact of changes in Q on the dynamics of Sactive in Figure 10B for three different Qs and R = 1 as well as three different Rs and Q = 1. Since the temporal evolution of Sactive is a result of the competing autophosphorylation and dephosphorylation progress, the choice of both scaling parameters influences Sactive dynamics (see Figure 10B). Note that R = 1 and Q = 1 is used everywhere in this paper, except for the results discussed in this section and shown in Figure 10.
Increasing R accelerates the convergence of PP1 toward a steady-state oscillation. In Figure 10A, this happens after ∼20 and ∼30 spike-pair presentations for R = 1 and R = 0.078, respectively. This constant value is not attained during the 60 s stimulation protocol with R = 0.002 at all. Reaching such a steady-state behavior is needed for the system to be robust to changes in the number of spike-pair presentations. Indeed, if the PP1 activity is still in the raising phase at the end of the stimulation protocol (as in the case R = 0.002, see blue line in Figure 10A), then more spike-pair presentations would lead to a higher PP1 level and therefore to up-to-down transitions for a drastically wider range of Δt values if the system is initially in the UP state. On the other hand, less spike-pair presentations would not give rise to any transitions at all. Figure 10C and 10D give an insight on how the Δt range for which transitions occur depends on R, Q, and the number of spike-pair presentations. For R = 1 and Q = 1 (kCaN = 18 1/s, ΔCapre = 0.17 μM), the range of Δt values evoking down-to-up transitions saturates beyond 50 spike-pair presentations, whereas the range resulting in up-to-down transitions becomes essentially insensitive to the number of spike-pair presentations beyond ∼150 spike pairs (see green regions in Figure 10C and 10D; both depict the same results). Increasing R further does not lead to any significant changes in the ranges of transitions compared to the case R = 1 (compare green and blue regions in Figure 10C for R = 1 and R = 2, respectively, and see Materials and Methods). Decreasing R slows down the convergence toward a stable range of time lags evoking down-to-up or up-to-down transitions (compare R = 1, green regions, and R = 0.078, red regions, cases in Figure 10C). This is due to the slower convergence of PP1 activity to its oscillatory behavior around a constant value during the stimulation protocol (see Figure 10A). The examples in Figure 10A show furthermore that the smaller R, the slower the decay of PP1 activity after the stimulation protocol. When R = 0.078 the PP1 dephosphorylation activity decay after the stimulation protocol is so slow that large positive time lag stimulations evoke transitions from UP to DOWN (see upper red shaded region in Figure 10C, Q = 1, kCaN = 18 1/s, ΔCapre = 0.17 μM). Up-to-down transitions at large positive time lags appear in the range up until 200 spike-pair presentations for R ≲ 0.202.
Similar arguments hold for the scaling parameter Q that controls CaMKII autophosphorylation and dephosphorylation dynamics. Indeed, the degree of CaMKII subunit phosphorylation should also reach a steady oscillation around a constant value during the presentation of spike pairs, for the system to exhibit robust behavior (see the R = 1,Q = 1.5 and R = 1,Q = 0.083 cases in Figure 10B). The larger the Q, the faster autophosphorylation (through an increase of k6, k7, and k8; Figure 1C–1E) and dephosphorylation (through an increase of k12, Equation 6) proceed. Therefore, less spike-pair presentations are required to evoke transitions and the Δt ranges leading to down-to-up- or up-to-down transitions saturate at smaller numbers of spike-pair presentations (see Figure 10D). For Q = 0.083, no transitions are observed at all in the range from 1 to 200 spike-pair presentations (see blue line in Figure 10B). Down-to-up transitions appear after 60 spike-pair presentations for 1≲ Q (kCaN = 18 1/s, R = 1) whereas up-to-down transitions occur at lower Q values (see (R = 1,Q = 0.83) case in Figure 10D). For 1.25 ≲ Q, dephosphorylation progress wins over autophosphorylation and leads to up-to-down transitions for large positive time lags (see red regions for R = 1,Q = 1.67 in Figure 10D).
The minimal stimulation protocols of Petersen et al. and O'Connor et al. on single hippocampal CA3-CA1 synapses evoke step-like all-or-none transitions of synaptic transmission efficacy [20,21]. These experiments suggest that individual synapses store information in a digital manner. The model presented here exhibits two stable steady-states of the CaMKII phosphorylation level at resting calcium conditions. This idea dates back from the pioneering paper by Lisman (1985) [15] and has since been investigated by modeling studies of increasing biochemical realism [15,17,18,42,57,61]. Miller et al. showed that the highly phosphorylated state in a system composed of a realistic number of CaMKII holoenzymes can remain stable on very long time scales (years) in the presence of protein turnover [57]. Hayer and Bhalla found that CaMKII bistability is preserved in the context of translocation and localization in the PSD of the protein [42]. The UP state in their study possesses a predicted lifetime of the order of tens of hours which is supported by experimental studies reporting a prolonged localization of the CaMKII at its postsynaptic site following LTP stimulation [25,62,63].
Experimentalists have reported several types of synaptic increase or decrease. For example, LTD (decrease of efficacy from “basal” strength) and depotentiation (decrease of efficacy after potentiation) have often been considered as two distinct processes. Some of the differences between the two can be reconciled in our model by considering that a “basal” condition is likely to be a mix of synapses in the UP and the DOWN state. Hence, a LTD protocol will decrease synaptic strength by provoking up-to-down transitions in some synapses that were initially in the UP state. On the other hand, in depotentiation protocols the initial conditions are different, because a larger fraction of the synapses are in the UP state. However, some studies indicate that depotentiation and LTD might operate through different molecular mechanisms [64–66]. A more complex model than the one proposed here would be necessary to account for these experimental data.
As in previous models, our model exhibits LTP for high enough calcium concentrations. Unlike previous models however, it possesses an “LTD window”, where the system makes a transition from the highly phosphorylated to the weakly phosphorylated state, under plausible conditions. There are three requirements for LTP and LTD transitions to occur at realistic calcium concentrations in our model (see Figure 3C).
(I) The steady-state concentration of phosphorylated CaMKII subunits has to exhibit a bistable behavior, i.e., a highly and a weakly phosphorylated state should coexist in a range of calcium. This is the case if phosphatase activity saturates at high CaMKII phosphorylation levels. In turn, this is ensured if the phosphatase is present in small amounts compared to CaMKII, which itself is enriched at high concentrations in the PSD [24,27,47]. We show that bistability is a property of the CaMKII system which is robust to variations in the number of interacting subunits but most effectively expressed in a ring of biological size with six interacting subunits (see Figure 2A).
(II) The phosphatase activity at resting conditions has to allow for two stable CaMKII phosphorylation states. Bistability at Ca0 is a robust property of the model over a large range of values of most of the protein signaling cascade parameters (see Table 1B). However, this requirement constrains the calcium-independent activities of the calcineurin and the cAMP–PKA pathways and is the reason why the model presented here is sensitive to changes in PKA base activity
, i.e., varying
more than 14% from the value given in Table 1B leads to a loss of bistability at resting conditions.
(III) The “LTD window” emerges from an elevated phosphatase activity in the range of intermediate calcium concentrations. There are two possible realizations of the cAMP–PKA pathway for such an “LTD window” to arise: (i) if the PKA activity is assumed to be calcium-independent, the PP1 activity curve (purple lines in Figure 3B) would show a Hill-function–like behavior. However, a CaMKII versus calcium bifurcation diagram qualitatively similar to Figure 3C could still be obtained. How such a scenario would affect the behavior of the system in response to the STDP protocol is still to be clarified. (ii) If the cAMP–PKA pathway is calcium/calmodulin-dependent as chosen here (see also [61]), the PP1 activity can be coupled to the calcium concentration such that a peak emerges at intermediate calcium concentrations. Several lines of experimental evidence support the inclusion of such a calcium-dependent cAMP–PKA pathway which promotes LTP by blocking phosphatases in the model: the induction of hippocampal LTP is blocked by inhibiting cAMP-dependent protein kinase A or inhibition of postsynaptic kinases in general and is facilitated in a PKA-dependent manner by inhibiting calcineurin [38,60,67]; a rapid increase in PKA activity accompanies the early phase of LTP in afferent fibers between hippocampus and prefrontal cortex [68]; calcium-stimulable forms of cAMP exist which indirectly control PKA activity [69]. For the CaMKII system to exhibit the “LTD window” with a calcium/calmodulin-dependent cAMP–PKA pathway, the model predicts that the cAMP–PKA pathway should activate at higher calcium concentrations compared to the calcineurin pathway, as this is required for the peak of phosphatase activity to emerge.
Another way to assess the coupling of the protein signaling cascades to PP1 activity and to CaMKII is to check what the model predicts if we block different parts of the pathways and compare it to experimental results. We can implement the blockade of the calcineurin or the cAMP–PKA pathways in the model by removing the calcium/calmodulin-dependence of the calcineurin or the cAMP–PKA pathways, since inhibitor 1 is also dephosphorylated by the calcium-independent protein phosphatase 2A [70,71] and phosphorylated by the calcium-independent protein kinase G [72]. Blocking the calcium/calmodulin-dependent part of the calcineurin pathway (i.e., kCaN = 0) leads to facilitation of LTP, and the reverse transition (LTD) is prevented. On the contrary, blocking the calcium/calmodulin-dependent part of the PKA pathway (i.e., kPKA = 0) facilitates LTD and prevents LTP. Transitions in either one of both directions can be evoked since bistability at resting conditions is preserved in both cases. All these model predictions are consistent with experimental assays inhibiting either the calcineurin [37,43,60] or the cAMP–PKA pathway [38]. If either the calcineurin or the cAMP–PKA pathways are completely abolished in the model, i.e., both the calcium-independent and the calcium-dependent parts are suppressed (i.e.,
or
), the system becomes locked in the UP or the DOWN state, respectively. Under these conditions, bistability is not present at resting calcium concentrations, i.e., no transitions can be evoked in a stable fashion. This also means a change in basal synaptic transmission since all synapses in the system will converge to one of the two stable states. Along the lines of the argumentation above, this situation would correspond to a scenario in which all proteins de- or phosphorylating inhibitor 1 are inhibited. Inhibiting completely protein phosphatase 1 activity, i.e., setting PP1 activity to zero, results in locking the system to the UP state for all calcium concentrations in our model. However, other calcium-independent phosphatases such as protein phosphatase 2A and 2C are known to dephosphorylate CaMKII [73]. Adding such phosphatases to the model would lead to bistability even in the absence of PP1. Such a scenario would be consistent with experiments which have shown that LTD but not LTP requires the activation of PP1 [37,39,74]. Our model indeed predicts a progressive diminution of the LTD window and an enlargement of the LTP window as a function of PP1 inhibition. In response to the STDP protocol, LTD disappears first when phosphatase activity is decreased as suggested by experimental results [67]. Reducing the phosphatase activity further results in down-to-up transitions for all Δts before the stable DOWN state disappears if the total PP1 concentration is reduced below 40%.
In addition to the “LTD window” at intermediate calcium concentrations, our model possesses a second region of bistability between the “LTD window” and the “LTP window” (see region between points 3 and 4 in Figure 3C). This region is not present in previous models and can be seen as a region of no changes. Starting from the DOWN or the UP state, calcium elevations to this range do not evoke any transition. A similar region of calcium concentrations in between LTP and LTD calcium levels leading to no plasticity is found experimentally by Cho et al. and discussed by Lisman as “no man's land” [75,76].
We have shown that the model can qualitatively reproduce plasticity outcomes in response to the STDP protocol. In our model as in previous models [9,10,77,78], the only signal driving synaptic changes is the dynamics of the calcium concentration, consistent with current experimental data [3,55,79–82]. However, previous modeling studies that use either the maximum amplitude of the calcium signal or simple readout mechanisms of the entire calcium dynamics reproduce only partially STDP results [9,10,77,78]. In particular, it has proven difficult to prevent the appearance of a second LTD range at large positive Δts. Shouval and Kalantzis show that stochastic properties of synaptic transmission can markedly reduce the LTD magnitude in this range [83]. Karmarkar et al. hypothesize that two functionally distinct calcium pools trigger different readout mechanisms for LTP and LTD in order to overcome this difficulty [9]. Here, we show that the compound calcium signal from VDCCs and NMDA-Rs combined with a complex readout mechanism is sufficient to account for experimental STDP data; in other words, the two calcium influxes do not have to be separated. This is due to the highly cooperative CaMKII autophosphorylation and the protein signaling cascade influencing PP1 activity, which provide a strongly nonlinear detector system, which is sensitive enough to translate differences in the time course of the calcium concentration into observed plasticity outcomes. Finally, CaMKII phosphorylation level changes need to sum over several pairs of spikes in order to observe LTP- or LTD-like transitions, as suggested by experiments on STDP [4,43,84–86]. These changes combine in a highly nonlinear fashion in our model, going beyond simple summation of pairwise interactions. In particular, a minimal number of spike pairs is needed to observe any plasticity, as shown in Figure 10. This number depends on the kinetics of autophosphorylation and dephosphorylation dynamics in the model. Froemke et al. (visual cortex slices) and Wittenberg and Wang (hippocampal slices) showed that LTP (causal spike pairings) requires only a few spike-pair presentations whereas the appearance of LTD (anti-causal pairings) requires a longer period of stimulation (∼100 spike pairings) [87,88]. Figure 10C and 10D show the faster saturation of the time lag range evoking LTP compared to the one evoking LTD, consistent with those experimental results. Interestingly, Wittenberg and Wang see a second LTD range at large positive time differences emerging after sufficiently long stimulation (70–100 spike pairings; compare second LTD window at positive Δt emerging at high spike-pair presentation numbers in Figure 10C and 10D for (R = 0.078, Q = 1) and (R = 1, Q = 1.67), respectively).
Rubin et al. recently proposed a model based on pathways resembling the CaMKII kinase-phosphatase system, which reproduces experimental STDP outcomes but does not exhibit bistability [44]. In that model, high, short-lasting calcium levels evoke LTP, low and prolonged calcium elevations above a certain threshold evoke LTD, and intermediate calcium levels act like a “Veto” preventing LTD induction. The durations for which their detector system has to be exposed to respective calcium levels are consistent with our findings. The competition between the PP1 buildup level and the autophosphorylation progress implements naturally the concept of the veto in our model. This balance between PP1 activity and autophosphorylation changes with Δt and defines the transition outcome: (short negative Δts) high PP1 accumulation and intermediate autophosphorylation of CaMKII evoked by linear interactions of the calcium influxes lead to LTD; (short positive Δts) low PP1 activity together with strong autophosphorylation of CaMKII as a result of supralinear calcium summations produce LTP; (all other cases) intermediate PP1 concentrations and weak to intermediate autophosphorylation arouse no changes. In particular, the stronger cAMP–PKA pathway activation due to higher calcium elevations for large positive Δt protocols can be seen as a realistic veto preventing LTD transitions to occur in this range. The differential activation of competing pathways at different calcium levels receives further support by recent experimental studies [89].
We observe a larger extent of the range of Δt values evoking LTD compared to the LTP range for R = 1 and Q = 1 in the noiseless case (see Figure 10D, green regions). In our model, the LTD range can be either larger, or smaller, than the LTP range, depending on various parameters such as noise, R, and Q. For large noise levels, the LTD range is generally smaller than the LTP range, while experimental data seems to indicate the opposite trend (compare blue line in Figure 7A and [2,4,85]). Investigating extensively how the parameters of the system change the extent of the LTP and LTD ranges goes beyond the scope of this study. In any case, the range of Δts leading to up-to-down transitions cannot be extended beyond the range of interaction between both calcium influxes. Hence, we expect the LTD range to become larger if this interaction is extended, e.g., due to nonlinear buffer dynamics [44] or the recruitment of additional protein signaling cascades [55,75,90]. Furthermore, BPAP attenuation and broadening has been shown experimentally to affect the STDP results and could change the balance between the ranges of time lags evoking LTP and LTD in our model [87,91]. Our model predicts that the range of time lags evoking LTP in response to the STDP protocol can be increased by amplifying PKA activity. On the other hand, increasing the strength of the calcineurin pathway shifts down horizontally the entire STDP curve (Figure 7A), leading to LTD transitions at all Δts (unpublished data).
Our model also reproduces qualitatively experimental transition results evoked by a purely presynaptic stimulation protocol [3]. Low stimulation frequencies evoke LTD and high frequencies LTP with a transition from LTD to LTP at 16–17 Hz in our simulations (compare with [3] where the transition happens at around 10 Hz but 900 presynaptic spikes are presented, instead of 60 here), for the same parameters that fit qualitatively the STDP data. Our model furthermore predicts that postsynaptic frequency stimulation evokes LTP at frequencies above 50 Hz (see Figure 7B). Interestingly, this type of stimulation does not evoke transitions from UP to DOWN at any frequency. However, we expect this form of plasticity to be strongly dependent on the extent and the time course of BPAP amplitude suppression. We also exposed the CaMKII system to purely presynaptic or postsynaptic spike-pair stimulation protocols. Since presynaptic spikes evoke long-lasting calcium transients and postsynaptic spikes high but fast-decaying calcium elevations, the Δt ranges for which transitions can be observed in the two cases are markedly different. In particular, our model predicts pairs of postsynaptic spikes should elicit down-to-up transitions only if spikes are very closely spaced, and only when the frequency of the pair is large enough (3 ≲ Δt ≲ 8 ms for f = 1 Hz). In the case of presynaptic spike pairs occurring at 1 Hz, the model predicts depression or up-to-down transitions for all values of Δt. The Δt ranges for which transitions occur become smaller if the presentation frequency of the spike pairs is reduced (presynaptic spike pairs at 0.5 Hz: down-to-up transitions for Δt < 300 ms; postsynaptic spike pairs at 0.5 Hz: up-to-down transitions for 3 ≲ Δt ≲ 6 ms). Nevian and Sakmann found that three postsynaptic spikes at 50 Hz, repeated 60 times at 0.1 Hz, do not evoke any synaptic changes [55]. We find a similar outcome in our model, but predict that an increase in frequency and/or decrease in the burst inter-spike interval should lead to potentiation. This prediction is, however, again sensitive to the extent of summation in calcium in between spikes. If the calcium transients evoked by the back-propagating action potentials do not accumulate, but the second BPAP evokes a calcium transient with the same amplitude as the first one, no down-to-up transitions are observed in the model. We have also investigated the transition behavior of the CaMKII system in response to spike triplets [43], and our model reproduces qualitatively such data provided short-term depression (STD) is added, as in [92] (unpublished data).
In conclusion, our model possesses two stable states of CaMKII activation, which could represent the core mechanism of binary synaptic strength maintenance. We furthermore show that it is possible to reproduce qualitatively experimental STDP results on LTP- and LTD-like transitions. These two results taken together suggest that the CaMKII-associated protein network could account for storage and induction of synaptic changes. Our model therefore predicts that the CaMKII protein also plays a major role in LTD, namely that CaMKII gets dephosphorylated during LTD induction. Experiments addressing the role of CaMKII in LTD provide controversial results. Sajikumar et al. showed that LTD in hippocampal CA1 neurons is blocked by CaMKII inhibition during induction but the application of the CaMKII inhibitor (KN-62) after the induction had no impact on LTD [93]. In other experiments, LTD has been shown to occur in the presence of CaMKII inhibitors during LTD induction in hippocampal cultures and slices [21,43]. Such inhibitors bind to CaMKII and block its activation by calmodulin (inhibitor KN-62, which is known not to inhibit the autophosphorylated kinase; used in [43]) or interact with the ATP-binding site of CaMKII (K252a used in [21]) [24]. In the presence of each of both inhibitors, CaMKII can still get dephosphorylated by PP1. We predict that LTD will no longer occur if the CaMKII–phosphatase interaction is disrupted. However, LTD experiments on the hippocampus, the somatosensory cortex, as well as the perirhinal cortex of rats suggest that the metabotropic glutamate receptor (mGluR) pathway is also involved in LTD [55,75,90,94]. The biochemical cascades emerging from mGluR activation could in principle make the occurrence of LTD transitions more robust. The negative coupling of group II mGluRs with the cAMP–PKA pathway [95–97] is consistent with the idea presented here, that LTD requires a shift in kinase–phosphatase balance in favor of phosphatases. Overall, we suggest the dynamics of the global calcium time course play a crucial role for the sign of synaptic changes alongside the crosstalk between signaling cascades that include the one considered here.
Calcium binding to calmodulin. Calmodulin contains four calcium binding sites, two at the C- and two at the N-terminal domain. Calcium binding happens in a cooperative manner in each one of these pairs [98]. The following scheme describes the macroscopic binding of calcium to calmodulin, i.e., we take into account the number of bound calcium ions only, regardless of the occupied microscopic binding sites:
Here M is free intracellular calmodulin and Ci (i = 1,2,3,4) with C4 ≡ C denotes the calcium/calmodulin complex with i bound calcium ions. Calmodulin target proteins including CaMKII are partially activated by calmodulin with two, three, or four calcium ions bound. However, CaMKII autophosphorylation rates induced by calmodulin bound with two or three calcium ions are much smaller than with calmodulin bound with four calcium ions [99]. Hence, we consider for simplicity in the model that only calmodulin bound with four calcium ions is able to phosphorylate CaMKII. Since the binding of calcium by calmodulin is fast (with binding rates of the order ∼1000 (μM/s)−1 [100,101]), we assume reaction Equation 1 to be in equilibrium with the calcium concentration. The macroscopic dissociation constants of successive calcium binding are taken from Linse et al. (see Table 1A for parameters [100]). The total concentration of calmodulin is CaM0 = M + C1 + C2 + C3 + C. Experimental studies suggest that the total available level of calmodulin in neurons is CaM0 ≈ 10 μM [98,102,103]. Here, we use a smaller value due to the vast number of target proteins of calmodulin besides CaMKII, and the sequestration of calmodulin by neurogranin in spines under resting conditions (see Table 1 and [102,104]). For simplicity, we do not consider the dynamics of calmodulin sequestration by neurogranin, which has been suggested to provide calmodulin during LTP protocols [105]. Assuming a calmodulin bath is an effective way to implement dissociation of calmodulin–neurogranin complexes, which provides calmodulin to the PSD during autophosphorylation and phosphatase/kinase activation. Italic style symbols in this manuscript refer to concentrations of the respective element or protein.
Autophosphorylation of CaMKII. The calcium/calmodulin-dependent protein kinase II (CaMKII) holoenzyme has 12 domains, grouped into two clusters each with six functionally coupled subunits [48,49]. CaMKII is activated by Ca2+/calmodulin binding to its subunits. Ca2+/calmodulin binding to adjacent subunits in the subunit ring stimulates intersubunit autophosphorylation at the residue theronine-286 in the autoregulatory domain (Thr286). Autophosphorylation increases CaMKII affinity for Ca2+/calmodulin and prolongs activation beyond dissociation of Ca2+/calmodulin. In turn, as long as CaMKII stays activated it can bind to the NMDA-R and phosphorylate exogenous substrates [24,49]. For simplicity, some aspects of CaMKII function are not accounted for in the model. Any differences between the CaMKIIα and −β isoforms are not considered. The binding of calcium/calmodulin and protein phosphatase 1 to a subunit is assumed to be independent of the state of neighboring subunits. The autophosphorylation at Thr305 and Thr306 is not considered.
CaMKII autophosphorylation is an intersubunit process during which one subunit acts as substrate and the neighboring subunit as catalyst. For autophosphorylation to take place, calmodulin has to be bound to the substrate subunit [49]. Autophosphorylation at Thr286 or binding of calmodulin each disable the autoinhibitory domain, therefore the catalytic subunit can be in one of the following states: (i) bound with calmodulin, (ii) phosphorylated and bound with calmodulin, or (iii) phosphorylated only (for an illustration see Figure 1C–1E) [27].
The chemical reaction schemes in Figure 1A–1E show schematically how binding of calmodulin and autophosphorylation is represented in the model. Reactions in 1A and 1B show calcium/calmodulin complex binding to dephosphorylated- or phosphorylated subunits, respectively. Autophosphorylation steps where the catalytic subunit is bound with calcium/calmodulin, phosphorylated and bound with calcium/calmodulin, or phosphorylated only are illustrated in Figure 1C, 1D, and 1E, respectively. The intersubunit autophosphorylation is likely to be a directed interaction in the ring and is here assumed to proceed in a single direction [27].
For the autophosphorylation steps depicted in Figure 1C, 1D, and 1E to occur, the substrate subunit must bind the calcium/calmodulin complex C. Let γ be the probability that a dephosphorylated subunit S binds with C, i.e., γ = SC / (S + SC) (SC stands for a dephosphorylated subunit bound with C); and γ* the probability that a subunit phosphorylated at Thr286, S*, binds with C, i.e., γ* = S*C / (S* + S*C) (S*C stands for a phosphorylated subunit bound with C). Assuming reactions in Figure 1A and 1B to be in equilibrium and using the Law of Mass Action yields SC = S · C / K5 and S*C = S*· C / K9, respectively, where K5 = k−5 / k5 and K9 = k−9 / k9 are the dissociation constants of reactions shown in Figure 1A and 1B, respectively. These assumptions lead to:
The probability that reaction in Figure 1C takes place in a unit time between two subunits in the single direction is k6γ2. Correspondingly, the probability for reaction in Figure 1D to occur in a unit time is k7γγ* and for reaction in Figure 1E to occur is k8γ (1 – γ*). This probabilistic description of autophosphorylation allows us to describe a given subunit by two possible states only, i.e., whether or not a subunit is phosphorylated at Thr286. Note that with a six-subunit ring this yields 14 macroscopic distinguishable activation states (see below). This ansatz does not account for calmodulin consumption during the process of autophosphorylation, assuming a bath of calmodulin. Similar approaches have been used in the investigations of Okamoto and Ichikawa as well as Zhabotinsky in CaMKII models exhibiting bistability [17,18], and in other studies describing in detail CaMKII autophosphorylation, but do not exhibit bistability [106–109].
Dephosphorylation of CaMKII. PP1 is the only protein phosphatase that dephosphorylates CaMKII associated with the postsynaptic densities [73]. The dephosphorylation of a free, phosphorylated subunit, and a phosphorylated subunit bound with the calcium/calmodulin complex are described according to the Michaelis-Menten scheme:
where D denotes the concentration of active PP1. Note that dephosphorylation happens independently whether a subunit is bound with the calcium/calmodulin complex or not. Assuming the
and
formations are at equilibrium, i.e.,
, we can use the standard Michaelis-Menten equation [110] to obtain the per-subunit rate of dephosphorylation, k10,
where the Michaelis constant KM is given by KM = (k−11 + k12) / k11 and Sactive is the total concentration of phosphorylated CaMKII subunits,
, where mi is the number of phosphorylated subunits of the macroscopic activation state i (see below for more details). The per-subunit rate of dephosphorylation, k10, is proportional to the amount of available phosphatase, D. The dephosphorylation rate per subunit declines if a lot of subunits are phosphorylated and the phosphatase activity remains constant, i.e., if Sactive is high and D constant. This saturation of k10 leads to the bistable behavior of the CaMKII phosphorylation level (see “Bistability of the CaMKII system with constant PP1 activity” section).
Applying the Law of Mass Action and taking into account the geometry of the CaMKII six-subunit ring, its autophosphorylation and dephosphorylation by PP1 is described by the following system of coupled, ordinary differential equations for concentrations of CaMKII with different numbers of phosphorylated subunits
Here Si refers to the concentration of the 14 (i = 0,...,13) macroscopic distinguishable activation states of the CaMKII protein. The subscript in the second column shows the geometrical order of Thr286 phosphorylated sites in the CaMKII ring, 1 refers to a phosphorylated subunit, 0 to a dephosphorylated subunit. Attention should be drawn to the fact that, for example, S5, S6, S7, and S8, all have three phosphorylated subunits, i.e., all of them have the same macroscopic level of activation, i.e., m5 = m6 = m7 = m8 = 3. However, in terms of symmetry all four have to be distinguished since at S5 the phosphorylated sites are adjoined, S111000, whereas at S8 they are separated by a dephosphorylated subunit, S101010, for example. With regard to this difference, the propagation of autophosphorylation is different for both, the phosphorylation step shown in Figure 1C can occur on two pairs of subunits at S5 but cannot occur at S8 at all. Taking into account that the different autophosphorylation steps, depicted in Figure 1C–1E, happen with different probabilities leads to differing occurrences of the activation states Si (with i = 0...13). Note that we used the fact k7 = k8 and simplified k7γγ* + k8γ (1 – γ*) to k7γ. k10 is the per-subunit rate of dephosphorylation (see above).
, with n = 13 for 14 macroscopic distinguishable activation states of the six-subunit CaMKII ring, yields the CaMKII protein mass conservation,
. 2CaMKII0 gives the total concentration of functionally independent CaMKII clusters of six subunits and CaMKII0 the total CaMKII concentration since one holoenzyme comprises two six-subunit rings. Note that the number of macroscopic distinguishable activation states is 3, 6, 14, and 36 for the two, four, six, and eight subunit models, respectively.
The dephosphorylation activity of PP1 is indirectly governed by calcium/calmodulin via inhibitor 1 (I1), i.e., phosphorylated inhibitor 1 inhibits PP1 [111,112]. Inhibitor 1 itself is phosphorylated by cyclic AMP-dependent protein kinase A (PKA) and protein kinase G and dephosphorylated by the phosphatase calcineurin and protein phosphatase 2A [37,38,60,70–72]. A simple realization of this protein signaling cascade is given by
where IG refers to dephosphorylated I1, I denotes phosphorylated inhibitor 1 (I1P), D is free PP1, and DI stands for inhibited PP1, i.e., PP1 bound with phosphorylated inhibitor 1. See Figure 1F for a scheme of the protein signaling cascade.
The balance between inhibitor 1 phosphorylation (vPKA)- and dephosphorylation rate (vCaN) is calcium/calmodulin-dependent. The enzymatic activity of calcineurin can be described by a Hill equation [113]. The PKA activity is also known to be calcium/calmodulin-dependent via cyclic AMP [69], but there is no data characterizing this dependency. We chose to describe both by a Hill equation
with a calcium/calmodulin-independent base activity (
) which also accounts for protein kinase G phosphorylation (x = PKA) and protein phosphatase 2A dephosphorylation (x = CaN). kx is the maximal, calcium/calmodulin-dependent activity, Kx the half activity concentration, and nx denotes the Hill coefficient.
Applying the Law of Mass Action and taking into account protein phosphatase 1 conservation yields
where I0 and D0 = D + DI refer to the total I1 and PP1 concentration, respectively. The concentration of dephosphorylated, free inhibitor 1, I0, is treated like a bath, assuming a rapid exchange between the PSD with the spine volume and between the spine and the parent dendrite, as in [17,57]. Therefore, inhibitor 1 is not conserved in Equation 24 due to this bath assumption.
As can be seen in Figure 5C and 5D, the change in PP1 activity, as well as the change in I1P concentration (unpublished data), during the presentation of one spike pair is small. We therefore separate both variables into two terms, a constant value and a small time-dependent change, i.e., D(t) → D* + ɛδD(t) and I(t) → I* + ɛδI(t), where D* and I* are the values before the spike-pair presentation, and δD(t) and δI(t) describe the changes during the presentation. Since these small changes are exclusively driven by changes in vCaN(t) and vPKA(t), we consider the time-dependent part of both rates as small compared to k13 and k−13, i.e.,
and
. Inserting these expressions in Equations 24 and 25 yields at zero order in ɛ the steady-state values D* and I*. The first-order equations in ɛ are
The Eigenvalues of the homogeneous system of Equations 26 and 27 are
Since vCaN is much smaller than k13D*, k13I*, or k−13, we expand the two Eigenvalues around vCaN. This yields a fast and a slow Eigenvalue since λ+ is zero at leading order. The Eigenvalues become
With the initial conditions δD(0) = δI(0) = 0, the solution for the inhomogeneous system (Equations 26 and 27) becomes
with A1 = −(k13D* + vCaN + λslow) / (λfast − λslow), B1 = (k13D* + vCaN + λfast) / (λfast − λslow), A2 = −k13D* / (λfast − λslow), and B2 = k13D* / (λfast − λslow). S(τ) is the inhomogeneous part in Equation 26, i.e., S(τ) = −δvCaN (τ)I* + δvPKA(τ)I0. The first term in Equations 32 and 33 describes the fast dynamics of both variables and allows D and I to follow on a fast time scale the calcium transient. After the spike-pair presentation, this term decays rapidly with the time constant λfast. The second term determines the slow dynamics of the system and therefore gives rise to a slow buildup, which decays after the spike-pair presentation with the slow time constant λslow. Since (λslow · t) is small at the scale of single presentations, we obtain for the slow dynamics
δD(t) is shown in Figure 9B as a measure for the slowly decaying PP1 buildup after the presentation of one spike pair. D* + δD(t) is compared with the PP1 activity obtained from numerical integration of Equations 24 and 25 after one spike pair in Figure 9A. Note that the product of δD, D*, and D with k12 is shown in Figure 9A and 9B.
In the section “STDP protocol with deterministic calcium transients,” we point out that an increase in R beyond the value of 1 does not significantly affect the dynamics of the PP1 response, which is basically determined by
(see paragraph above). This can be understood by considering
(Equation 31), if D*≪ I*, k−13 / k13, its denominator, will be controlled by k13I* and k−13 only, and changes in D* will have no impact on the PP1 dynamics.
To investigate how the model behaves when realistic calcium transients are applied to it, we use the following model for postsynaptic calcium and postsynaptic membrane potential dynamics. We focus on a single spine compartment, and do not simulate the backpropagation of the action potential from its initiation site to the spine. Instead, we model the action potential dynamics directly at the spine.
Postsynaptic membrane potential. The postsynaptic membrane potential is modeled using the Hodgkin-Huxley formalism in a single compartment. The reference volume for the membrane potential and the calcium dynamics model is taken to be a postsynaptic spine (Vspine ≈ 1 μm3, rspine ≈ 0.5 μm). The dynamics of the membrane potential V follows the differential equation
where Cm is the whole cell capacitance of 0.1 nF, Ix (x = L, Na, K, NMDA, CaL, AMPA) are the ionic currents listed below. An action potential is evoked by a 1 ms depolarizing pulse current Istim of 3 nA.
Postsynaptic calcium dynamics. The model of the calcium dynamics involves the two main sources of postsynaptic calcium influx in the spine: NMDA receptors (NMDA-R) and voltage-dependent calcium channels (VDCC) [114]. Extrusion, diffusion, and slow buffering is accounted for by a single exponential decay, yielding the following equation for the time course of the intracellular calcium concentration
where Ca is the free, intracellular calcium concentration, τCa = 12 ms refers to the single exponential time constant of the passive decay process [54], Ca0 is the calcium resting concentration, and ζ = 2.59 · 104 m2 μM/C converts the ion currents into concentration changes per time for a spine of volume ≈1 μm3. Ix (x = NMDA, CaL) are the ionic currents listed below. Scaling parameters βNMDA = 1/1000 and βCaL = 1/100 take into account both the immediate uptake of calcium by intracellular buffers (∼99%, [115]) and the fact that only about ∼10% of the NMDA-mediated current is carried by calcium ions (see [116,117] and below).
Noisy calcium transients. To investigate stochastic effects, we add two possible noise sources to the model: (i) NMDA receptor maximum conductance drawn at random at each presynaptic spike and (ii) maximum conductance of the voltage-dependent calcium channel drawn at random at each postsynaptic spike. Both conductances are drawn from binomial distributions characterized by the total number of channels Ntot and the opening probability per channel po. Each presynaptic or postsynaptic spike gives rise to an integer number, no, of NMDA or CaL channel openings, respectively. We assume that the channels open independently of each other. The single channel conductance gsingle is chosen so that the mean calcium amplitudes are as stated above. To account for the stochasticity of calcium ions influx, Gaussian noise with zero mean and a variance scaled with no is added to (no · gsingle). The parameters of the NMDA and CaL maximum conductance distributions are adjusted such that they fit the experimental findings of single spine measurements by Mainen et al. and Sabatini and Svoboda, respectively [58,59] (see Table 2 for parameters).
Ion currents dynamics. The description and the parameters of the ionic currents are taken from Poirazi et al. (ICaL) as well as Purvis and Butera (INa, IK) [118,119].
Leak Current: The leak current is given by
where gL is the leak conductance. The leak potential is adjusted such that the resting potential is −70 mV.
The ionic currents listed here have the general form Iionic = gy(V – Eionic). Eionic is the reversal potential for the respective ions carried, g refers to the maximum conductance of each current, and y is the product of one or more gating variables. y determines the dynamics of the ion currents regulated by voltage-dependent activation and inactivation variables which are described according to
Here x∞(V) is the steady-state voltage-dependent (in)activation function of x, and τx(V) is the voltage-dependent time constant. In terms of this formalism, the respective ion currents are given by:
Sodium current:
Delayed-rectifier Potassium current:
Voltage-dependent calcium current (high-voltage activated L-type):
AMPA current: excitatory postsynaptic potentials are mainly mediated by the AMPA receptor current given by
with τAMPA = 2 ms,
ms, αs = 1 1/ms, and αx = 1 (dimensionless) [120,121]. sAMPA is a single exponentially decaying gating variable with a finite rise time (the time-to-peak is ≈0.2 ms). At each occurrence of a presynaptic spike at time tk, the variable xAMPA is increased by one (the sum on the right-hand side of Equation 51 goes over all presynaptic spikes occurring at times tk).
NMDA current: the current mediated by the NMDA receptor is described by
where the voltage dependence of the magnesium block is given by
The voltage dependence is controlled by the extracellular magnesium concentration [Mg2+] = 1.0 mM [122]. The dimensionless gating variable sNMDA obeys the same types of equations as s and x of the AMPA current (Equations 50 and Equation 51, respectively) but with τNMDA = 80 ms and
ms [121] (the time-to-peak is ≈8 ms).
The maximum leak conductance and the whole cell capacitance yield a membrane potential time constant τm of 20 ms, according to the equation τm = Cm / gL. The AMPA receptor conductance gAMPA is chosen such that a single presynaptic spike evokes a maximal depolarization of 1 mV at −70 mV. gNMDA and gCaL are chosen such that the amplitudes of the NMDA-R mediated and the action potential–evoked calcium transients in the spine are realized as stated in the text. The ratio of ∼2 between the BPAP evoked calcium transient amplitude (ΔCapost) and the NMDA-R mediated contribution (ΔCapre) is as measured by Sabatini et al. [54]. Note that the VDCC- and the NMDA-mediated calcium currents in the calcium dynamics (Equation 36) are multiplied by the scaling parameters βNMDA and βCaL, which account for fast calcium buffering and for the fractional calcium current through NMDA-Rs of ∼10 %. The calcium reversal potential, ECa, is used to describe the fractional calcium current through NMDAs in the calcium dynamics (Equation 36), whereas the reversal potential of the compound sodium, potassium, and calcium ion current, ENMDA, mediated by NMDA-Rs, is employed in the voltage equation (Equation 35).
Parameters of the model. The model describing the interactions between proteins contains a large number of parameters (25). In some cases, we used experimentally determined values (see Table 1A for a list of those parameters). Other parameters are not (or poorly) determined experimentally. These parameters were varied systematically or were determined by the constraints we impose on the model (see Table 1B). Finally, a few parameters were set on the basis of previous modeling studies or set to an arbitrary value, in cases in which changing this value does not alter the results of the model (see Table 1C).
The calcium-dependent steady-state concentration of phosphorylated CaMKII subunits depends heavily on the choice of the parameters defining the PKA pathway (
,kPKA, KPKA, nPKA). These parameters are adjusted in order to obtain the “LTD” and the “LTP window” at specific intervals of calcium concentration (see section “LTD window” in a model with Ca-dependent PP1 activity via protein signaling cascade including PKA and calcineurin). The maximal calcineurin activity kCaN is used to adjust the PP1 level evoked during the STDP stimulation protocol (see section “STDP protocol stimulation with stochastic calcium dynamics”).
The total calmodulin concentration (CaM0) is smaller than the value found in experimental studies, due to the reasons given above (see “Calcium binding to calmodulin” paragraph). KM is taken from the modeling study of [17]. Equation 24 and 25 give the steady-state PP1 concentration, Dsteady-state = D0 / (1 + (I0k13vPKA) / (k−13vCaN)). Hence, Dsteady-state depends on I0, vPKA, and vCaN through the single variable
. This means that out of the five parameters I0,
, kCaN,
, and kPKA, the steady-state PP1 concentration depends on three independent combinations of those parameters, e.g.,
,
, and
. Thereby, two out of these five parameters can be set arbitrarily. The total I1 concentration and the calcineurin base activity, I0 and
, are set to the values given in Table 1C and are kept constant throughout all investigations, while the remaining three parameters kCaN,
, and kPKA are obtained by constraints imposed on the model (see Table 1B). On the other hand, the dynamics of the protein signaling cascade depends on all five parameters. We address this issue via the scaling parameter R which influences the PP1 response dynamics (see “STDP protocol stimulation with stochastic calcium dynamics” section).
The parameters describing postsynaptic calcium and postsynaptic membrane potential dynamics are taken from previous modeling studies [59]. We systematically vary the calcium amplitudes evoked by a presynaptic (ΔCapre) and a postsynaptic spike (ΔCapost), keeping their ratio constant, ΔCapost = 2 · ΔCapre (see the section “Effect of Kinetics of Autophosphorylation and Dephosphorylation on the Number of Spike-Pair Presentations Needed for Transitions”).
We solve the system of coupled, ordinary differential equations with a fourth-order Runge-Kutta method with adaptive stepsize control. This has been implemented in a C++ program. We used XPPAUT by G. Bard Ermentrout (http://www.pitt.edu/~phase/) for the steady-state calculations of the CaMKII system. |
10.1371/journal.pgen.1007702 | Boundaries mediate long-distance interactions between enhancers and promoters in the Drosophila Bithorax complex | Drosophila bithorax complex (BX-C) is one of the best model systems for studying the role of boundaries (insulators) in gene regulation. Expression of three homeotic genes, Ubx, abd-A, and Abd-B, is orchestrated by nine parasegment-specific regulatory domains. These domains are flanked by boundary elements, which function to block crosstalk between adjacent domains, ensuring that they can act autonomously. Paradoxically, seven of the BX-C regulatory domains are separated from their gene target by at least one boundary, and must “jump over” the intervening boundaries. To understand the jumping mechanism, the Mcp boundary was replaced with Fab-7 and Fab-8. Mcp is located between the iab-4 and iab-5 domains, and defines the border between the set of regulatory domains controlling abd-A and Abd-B. When Mcp is replaced by Fab-7 or Fab-8, they direct the iab-4 domain (which regulates abd-A) to inappropriately activate Abd-B in abdominal segment A4. For the Fab-8 replacement, ectopic induction was only observed when it was inserted in the same orientation as the endogenous Fab-8 boundary. A similar orientation dependence for bypass activity was observed when Fab-7 was replaced by Fab-8. Thus, boundaries perform two opposite functions in the context of BX-C–they block crosstalk between neighboring regulatory domains, but at the same time actively facilitate long distance communication between the regulatory domains and their respective target genes.
| Drosophila bithorax complex (BX-C) is one of a few examples demonstrating in vivo role of boundary/insulator elements in organization of independent chromatin domains. BX-C contains three HOX genes, whose parasegment-specific pattern is controlled by cis-regulatory domains flanked by boundary/insulator elements. Since the boundaries ensure autonomy of adjacent domains, the presence of these elements poses a paradox: how do the domains bypass the intervening boundaries and contact their proper regulatory targets? According to the textbook model, BX-C regulatory domains are able to bypass boundaries because they harbor special promoter targeting sequences. However, contrary to this model, we show here that the boundaries themselves play an active role in directing regulatory domains to their appropriate HOX gene promoter.
| The three homeotic (HOX) genes in the Drosophila Bithorax complex (BX-C), Ultrabithorax (Ubx), abdominal-A (abd-A) and Abdominal-B (Abd-B), are responsible for specifying cell identity in parasegments (PS) 5–14, which form the posterior half of the thorax and all of the abdominal segments of the adult fly [1–3]. Parasegment identity is determined by the precise expression pattern of the relevant HOX gene and this depends upon a large cis-regulatory region that spans 300 kb and is subdivided into nine PS domains that are aligned in the same order as the body segments in which they operate [4–6]. Ubx expression in PS5 and PS6 is directed by abx/bx and bxd/pbx, while abd-A expression in PS7, PS8, and PS9 is controlled by iab-2, iab-3, and iab-4 [7–10]. Abd-B is regulated by four domains, iab-5, iab-6, iab-7 and iab-8, which control expression in PS10, PS11, PS12 and PS13 respectively [6,11,12].
Each regulatory domain contains an initiator element, a set of tissue-specific enhancers and Polycomb Response Elements (PREs) and is flanked by boundary/insulator elements (Fig 1A;[13]. BX-C regulation is divided into two phases, initiation and maintenance [14,15]. During the initiation phase, a combination of gap and pair-rule proteins interact with initiation elements in each regulatory domain, setting the domain in the on or off state. In PS10, for example, the iab-5 domain, which regulates Abd-B, is activated by its initiator element, while the more distal Abd-B domains, iab-6 to iab-8 are set in the off state (Fig 1B). In PS11, the iab-6 initiator activates the domain, while the adjacent iab-7 and iab-8 domains are set in the off state. Once the gap and pair-rule gene proteins disappear during gastrulation, the on and off states of the regulatory domains are maintained by Trithorax (Trx) and Polycomb (PcG) group proteins, respectively [16,17].
In order to select and then maintain their activity states independent of outside influence, adjacent regulatory domains are separated from each other by boundary elements or insulators [18–24]. Mutations that impair boundary function permit crosstalk between positive and negative regulatory elements in adjacent domains and this leads to the misspecification of parasegment identity. This has been observed for deletions that remove five of the BX-C boundaries (Front-ultraabdominal (Fub), Miscadestral pigmentation (Mcp), Frontadominal-6 (Fab-6), Frontadominal-7 (Fab-7), and Frontadominal-8 (Fab-8)) [6,17,19,20,22,23,25,26].
While these findings indicate that boundaries are needed to ensure the functional autonomy of the regulatory domains, their presence also poses a paradox [27,28]. Seven of the nine BX-C regulatory domains are separated from their target HOX gene by at least one intervening boundary element. For example, the iab-6 regulatory domain must “jump over” or “bypass” Fab-7 and Fab-8 in order to interact with the Abd-B promoter (Fig 1A). That the blocking function of boundaries could pose a significant problem has been demonstrated by experiments in which Fab-7 is replaced by heterologous elements such as scs, gypsy or multimerized binding sites for the architectural proteins dCTCF, Pita or Su(Hw) [25,29–31]. In these replacements, the heterologous boundary blocked crosstalk between iab-6 and iab-7 just like the endogenous boundary, Fab-7. However, the boundaries were not permissive for bypass, preventing iab-6 from regulating Abd-B.
A number of models have been proposed to account for this paradox. One is that BX-C boundaries must have unique properties that distinguish them from generic fly boundaries. Since they function to block crosstalk between enhancers and silencers in adjacent domains, an appealing idea is that they would be permissive for enhancer/silencer interactions with promoters (Fig 1B). However, several findings argue against this notion. For one, BX-C boundaries resemble those elsewhere in the genome in that they contain binding sites for architectural proteins such as Pita, dCTCF, and Su(Hw) [23,31–35]. Consistent with their utilization of these generic architectural proteins, when placed between enhancers (or silencers) and a reporter gene, BX-C boundaries block regulatory interactions just like boundaries from elsewhere in the genome [19,36–42]. Similarly, there is no indication in these transgene assays that the blocking activity of BX-C boundaries are subject to parasegmental regulation. Also arguing against the idea that BX-C boundaries have unique properties, the Mcp boundary, which is located between iab-4 and iab-5, is unable to replace Fab-7 [31]. Like the heterologous boundaries, it blocks crosstalk, but it is not permissive for bypass. A second model is that there are special sequences, called promoter targeting sequence (PTS), located in each regulatory domain that actively mediate bypass [43–45]. While the PTS sequences that have been identified in iab-6 and iab-7 enable enhancers to “jump over” an intervening boundary in transgene assays, they do not have a required function in the context of BX-C and are completely dispensable for Abd-B regulation [18,30].
A third model (Fig 1C) is suggested by transgene “insulator bypass” assays [46,47]. In one version of this assay, two boundaries instead of one are placed in between an enhancer and the reporter. When the two boundaries pair with each other, the enhancer is brought in close proximity to the reporter, thereby activating rather than blocking expression. Consistent with a possible role in BX-C bypass, these pairing interactions can occur over large distances and even skip over many intervening boundaries [48–51]. The transgene assays point to two important features of boundary pairing interactions that are likely to be relevant in BX-C. First, pairing interactions are specific. For this reason boundaries must be properly matched with their neighborhood in order to function appropriately. A requirement for matching is illustrated in transgene bypass experiments in which multimerized binding sites for specific architectural proteins are paired with themselves or with each other[52]. Bypass was observed when multimerized dCTCF, Zw5 or Su(Hw) binding sites were paired with themselves; however, heterologous combinations (e.g. dCTCF sites with Su(Hw) sites) did not support bypass. A second feature is that pairing interactions between boundaries are typically orientation dependent For example, scs pairs with itself head-to-head, not head-to-tail [52].
If both blocking and bypass activities are intrinsic properties of fly boundaries then the BX-C boundaries themselves may facilitate contacts between the regulatory domains and their target genes (Fig 1C). Moreover, the fact that both blocking and bypass activity are non-autonomous (in that they depend on partner pairing) could potentially explain why heterologous Fab-7 replacements like gypsy and Mcp behave anomalously while Fab-8 functions appropriately. Several observations fit with this idea. There is an extensive region upstream of the Abd-B promoter that has been implicated in tethering the Abd-B regulatory domains to the promoter [53–56] and this region could play an important role in mediating bypass by boundaries associated with the distal Abd-B regulatory domains (iab-5, iab-6, iab-7). Included in this region is a promoter tethering element (PTE) that facilitates interactions between iab enhancers and the Abd-B promoter in transgene assays [57,58]. Just beyond the PTE is a boundary-like element, AB-I. In transgene assays AB-I mediates bypass when combined with either Fab-7 or Fab-8. In contrast, a combination between AB-I and Mcp fails to support bypass [59,60]. The ability of both Fab-7 and Fab-8 to pair with AB-I is recapitulated in Fab-7 replacement experiments. Unlike Mcp, Fab-8 has both blocking and bypass activity when inserted in place of Fab-7 [30]. Moreover, its’ bypass but not blocking activity is orientation-dependent. When inserted in the same orientation as the endogenous Fab-8 boundary, it mediates blocking and bypass, while it does not support bypass when inserted in the opposite orientation.
In the studies reported here we have tested this model by replacing the endogenous Mcp boundary with heterologous boundaries. Mcp defines the border between the set of regulatory domains that control abd-A and those that control Abd-B expression (Fig 1A). Unlike the boundaries that are within the Abd-B regulatory domain (e.g. Fab-7 or Fab-8), Mcp is not located between a regulatory domain and its target gene. Instead, it defines the boundary between regulatory domains that target abd-A and those that target Abd-B. For this reason, we expected that it does not need bypass activity. Consistent with this expectation, we find that multimerized dCTCF binding sites fully substitute for Mcp. A different result is obtained for the Abd-B-associated boundaries, Fab-7 and Fab-8. Both boundaries are (for the most part) able to block crosstalk between the abd-A regulatory domain iab-4, which specifies A4 (PS9) and the Abd-B regulatory domain iab-5, which specifies A5 (PS10). Their blocking activity is orientation independent. However, in spite of blocking crosstalk, these replacements still inappropriately induce Abd-B expression in A4 (PS9), causing the misspecification of this segment. For the Fab-7 replacements, this occurred in both orientations, while for the Fab-8 replacement ectopic induction was only observed when it was inserted in the same orientation as the endogenous Fab-8 boundary. We present evidence showing that the boundary replacements activate the Abd-B gene in A4 (PS9) by inappropriately targeting the iab-4 domain to the Abd-B promoter. In addition to altering the specification of A4 (PS9), the Fab-7 replacements induce novel transformations of A5 and A6. These findings indicate that when Fab-7 is inserted into the BX-C in place of Mcp, it perturbs Abd-B regulation in several segments besides PS9.
The Mcp boundary is defined by 340 bp core sequence that has enhancer blocking activity in transgene assays [36] and blocks crosstalk between iab-6 and iab-7 when substituted for Fab-7 [31]. Located just distal to the boundary is a PRE that negatively regulates the activity of the iab-5 enhancers [61]. We used CRISPR to delete a 789 bp DNA segment including the Mcp boundary and the PRE and replace it with an eGFP reporter flanked by two attP sites (McpattP) (S1 Fig). The presence of two attP sites in opposite orientation allows integration of different DNA fragments by recombination mediated cassette exchange (RMCE; [62]) using the phiC31 integration system [63].
The Mcp boundary marks the division between the set of regulatory domains that control the abd-A and Abd-B genes (Fig 1A). The iab-4 domain is just proximal to Mcp, and it directs abd-A expression in PS9. The iab-5 domain is on the distal side and it regulates Abd-B in PS10. A boundary in this position would be needed to block crosstalk between iab-4 and iab-5; however, neither of these domains would require the intervening boundary to have bypass activity. On the proximal side, iab-4 must bypass the putative Fab-3 and Fab-4 boundaries in order to activate the abd-A promoter, while on the distal side, iab-5 must bypass Fab-6, Fab-7 and Fab-8 in order to activate Abd-B. If this expectation is correct, a generic boundary that has blocking activity but is unable to direct iab-4 to the abd-A promoter or iab-5 to the Abd-B promoter should be able to substitute for Mcp. To test this prediction (Fig 2), we introduced either the iab-5 PRE itself (McpPRE) or the PRE in combination with four dCTCF sites (McpCTCF). In Fab-7 replacement experiments four dCTCF sites in combination with the iab-7 PRE blocked crosstalk between the iab-6 and iab-7 domains, but did not allow the iab-6 domain to regulate Abd-B expression in PS11 [30].
Abd-B is a master regulator of pigmentation in the male abdominal A5 and A6 segments and it controls the expression of the yellow and tan genes which are involved in melanin synthesis [64–66]. In flies carrying the null y1 allele, the tan gene is still expressed and the pigmentation in A5 and A6 is light brown-yellow not black [66,67]. In order to be able to recover recombinants and also to monitor the blocking activity of the replacement sequence and the on/off state of the iab-5 domain, we used a y1 genetic background and included a minimal yellow (mini-y) reporter in our Mcp replacement construct (S1 Fig). The mini-y reporter consists of the cDNA and the 340 bp yellow promoter and lacks the wing, body and bristle enhancers of the endogenous yellow gene. As a result, activity of the mini-y reporter depends upon proximity to nearby enhancers. Expression of the mini-y reporter was examined in the y1 background.
Based on previous studies [5,21,68], the expression of this reporter should be determined by the activity state of the iab-5 domain. When iab-5 is shutoff by Pc-G dependent silencing in PS9 and more anterior parasegments, the mini-y reporter will also be silenced. When iab-5 is turned on in PS10 and more posterior parasegments, the mini-y reporter will be expressed. This parasegment-specific regulation of the reporter activity will be reflected in the segmental pattern of black melanin pigmentation in the adult cuticle. In replacements in which blocking activity is compromised, mini-y will be expressed in PS9 and in adults the A4 tergite will be black, just like the A5 and A6 tergites. In contrast, in replacements that have blocking activity mini-y will be silenced in PS9, but active in PS10 and PS11. In this case, A5 and A6 will have black pigmentation, while the stripe of pigmentation along the posterior edge of the A4 tergite will be light yellow brown, as only the tan gene will contribute to pigmentation in this segment.
When we replaced the Mcp deletion by the iab-5 PRE alone (McpPRE) the mini-y reporter was active not only in A5 (PS10) and more posterior segments, but also in A4 (PS9).
As shown in Fig 2, the pigmentation in A4 is black like that in A5 indicating that the reporter is expressed in both segments (Fig 2). This finding shows that, similar to classical Mcp deletions, the McpPRE replacement does not have blocking activity. In these Mcp deletions iab-5 is ectopically activated in PS9 by the iab-4 initiator. It then drives Abd-B expression in PS9 resulting in a gain-of-function (GOF) transformation of parasegment identity from PS9 to PS10. We used two approaches to test whether this was true for the McpPRE replacement. In the first, we excised the mini-y reporter and introduced an X chromosome with a wild type yellow (y+) gene. Since Abd-B directly regulates y+ expression in the abdomen [65,67], a transformation of PS9 into PS10 should be accompanied by a PS10-like pattern of pigmentation. Fig 2 shows that this is the case. We also examined the pattern of Abd-B protein expression in the embryonic CNS. In wild type embryos Abd-B is not expressed is PS9, while it is expressed at low levels in PS10. As shown in Fig 3A, similar levels of Abd-B protein are detected in PS9 and PS10 in the McpPRE replacement.
As predicted, a quite different result is obtained when we combined the iab-5 PRE with multimerized dCTCF sites. Expression of the mini-y reporter in the McpCTCF replacement was restricted to A5 (PS10) and A6 (PS11) as would be expected if the multimerized dCTCF sites block crosstalk between the iab-4 and iab-5 domains so that iab-5 is silenced by Pc-G factors in PS9 (Fig 2). The same pigmentation pattern is observed for the endogenous yellow in the Δmini-y derivative of McpCTCF, indicating that Abd-B is not turned on ectopically in PS9. This conclusion is confirmed by antibody staining experiments (Fig 3A). Thus, unlike replacements of Fab-7, a generic boundary can fully substitute for Mcp.
We next tested whether the Fab-7 boundary can substitute for Mcp. The Fab-7 region consists of a minor (HS*) and three major (HS1, HS2 and HS3) nuclease hypersensitive sequences [17,21,22,41,42]. Unlike Mcp or other known or suspected boundaries in BX-C, dCTCF does not bind to Fab-7 [33,69]. Instead, Fab-7 boundary function depends upon two BEN domain protein complexes, Elba and Insensitive, the C2H2 zinc finger protein Pita, and a large multiprotein complex, called the LBC [31,70–73]. In addition to a boundary function, the HS3 sequence can also function as a PRE (iab-7 PRE; [73,74]. In previous studies, we found that a combination of HS1+HS2+HS3 can functionally substitute for the complete Fab-7 boundary in vivo and we used this sequence (named for simplicity F7) for the Mcp replacements (Fig 4). Although Fab-7 has only limited orientation dependence in its endogenous context [30,73], we inserted the HS1+HS2+HS3 sequence in both the forward (same as endogenous Fab-7) and reverse orientations in the Mcp replacement platform as indicated in Fig 4. The phenotypic effects of the Fab-7 replacement inserted in the forward orientation, McpF7, are considered first.
Like the McpCTCF replacement, the mini-y reporter is turned on in A5 (PS10) and A6 (PS11) in McpF7 males, and the tergites in both of these segments are black. However, McpF7 differs in two respects from McpCTCF. First, there are one or two small patches of darkly pigmented cuticle in the A4 tergite (marked by the arrow). These patches are variable and appear to be clonal in origin. This finding indicates that the blocking activity of McpF7 is incomplete, and that the mini-y reporter and thus the iab-5 domain is ectopically activated by the iab-4 domain in a small number of PS9 cells. Second, instead of a stripe of yellow-brown pigmentation along the posterior margin, nearly the entire A4 tergite is covered in yellow-brown pigmentation. This pattern of pigmentation is not observed in A4 in y1 males carrying the McpCTCF replacement and the mini-y reporter (Fig 2) or for that matter in control wild type y1 males (see Fig 4). The presence of the yellow-brown pigmentation throughout most of the A4 tergite suggests that cells in this segment (PS9) are not properly specified. This is the case. When the mini-y reporter was excised and replaced by the endogenous X-linked y+ gene, the A4 tergite has a black pigmentation like A5 and A6 (Fig 4). Since expression of the yellow gene is controlled by Abd-B, this observation indicates that Abd-B must be ectopically activated throughout A4. Antibody staining experiments of the CNS in McpF7 embryos indicate that this inference is correct (Fig 3B).
A simple interpretation of these findings is that McpF7 is unable to block crosstalk between iab-4 and iab-5 and, as a result, iab-5 is ectopically activated in PS9 cells and inappropriately drives Abd-B expression. However, such an interpretation is inconsistent with the expression pattern of the mini-y reporter; it is only activated in small clones in the A4 tergite and not in the entire A4 tergite. By way of comparison, the dark black pigmentation generated by the reporter in McpPRE, which has no boundary activity, is clearly quite different from the yellow-brown pigmentation observed for the reporter in McpF7. In this respect, McpF7 resembles McpCTCF in that the iab-5 domain must be shut off by Pc-G silencing in (most) PS9 cells. This would imply that the iab-4 regulatory domain (or one of the other abd-A domains that is turned on in PS9 cells) must be responsible for ectopically activating Abd-B expression in PS9. Moreover, this would mean that the mechanism underlying the misspecification of A4 (PS9) in the McpF7 replacement differs from that in McpPRE or Mcp1 where iab-5 is not properly silenced in PS9 cells.
There are other abnormalities in McpF7 replacement indicating that it has complicated and novel effect on Abd-B expression. In wild type males, the A6 sternite has a banana shape and no bristles, while the A5 and A4 sternites resemble isosceles trapezoids and are covered with bristles. While the A4 and A5 sternites in McpF7 males still have bristles, they are split into two connected lobes that resemble the banana shape of the A6 sternite. These morphological abnormalities indicate that the Fab-7 replacement induces a weak GOF transformation of both A4 (PS9) and A5 (PS10) towards an A6 (PS11) identity. This type of transformation is not observed in Mcp boundary deletions, nor it is observed in the McpPRE replacement.
Further evidence of A4/A5→A6 transformation can be seen in the pattern of trichome hairs in the tergites. In wild type flies, the A4 and A5 tergites are covered with trichomes, while trichomes are only found along the anterior and ventral margins of the A6 tergite (see darkfield image in Fig 4). In the McpF7 replacement, there are large regions of the A4 and A5 tergite that are devoid of trichomes. There are even anomalies in A6: the band of trichomes along the anterior margin is absent. Similar alterations in cuticular phenotypes are observed in McpF7 females (S2 Fig). These findings indicate that the normal regulation of Abd-B is disrupted in several parasegments when Mcp is replaced by the Fab-7 boundary.
In its endogenous context, the functioning of Fab-7 is weakly orientation dependent. For this reason, we anticipated that the reverse Mcp replacement, McpF7R, would give a similar though milder spectrum of phenotypic effects. Fig 4 shows that this is the case. In y+ background, large regions of the A4 tergite have a black pigmentation like A5 and A6. The ectopic activation appears to be weaker than in the McpF7 replacement as there are regions in A4 in which the endogenous yellow gene is not turned on. Also, and unlike McpF7, there are no bald patches in the A4 trichomes, while the sternite appears to have a normal isosceles trapezoid shape. However, the novel transformations seen in McpF7 in the more posterior segments A5 (PS10) and A6 (PS11) are still evident. The A5 tergite is not completely covered with trichomes, while the trichomes along the anterior margin of A6 are absent. The A5 sternite is also misshapen. Thus, like McpF7, introducing a reversed Fab-7 boundary in place of Mcp disrupts Abd-B regulation in PS9 and also in other parasegments. Since the pattern of mini-y expression in McpF7R indicates that the iab-5 domain is silenced in (most) PS9 cells, the iab-5 regulatory domain can’t be driving Abd-B expression in this parasegment. Instead, misexpression of Abd-B in PS9 is likely driven by the iab-4 domain. This possibility will be considered further below.
In previous Fab-7 replacement experiments we found that a 337 bp fragment (F8337) spanning the Fab-8 boundary nuclease hypersensitive site is sufficient to fully rescue a Fab-7 boundary deletion [30]. In the direct (forward) orientation this fragment not only blocks crosstalk but also supports bypass. However, when the orientation of the Fab-8 boundary is reversed, bypass activity is lost, while blocking is unaffected. Since F8337 appears to have full boundary function, we inserted this fragment in both orientations next to the iab-5 PRE in the Mcp deletion (McpF8 and McpF8R).
The effects of the Fab-8 replacement in the reverse orientation, McpF8R, will be considered first. Like the McpCTCF replacement, McpF8R blocks crosstalk between iab-4 and iab-5 and the mini-y reporter is off in A4 (Fig 5). The fact that mini-y is not expressed in PS9 also means that iab-5 is silenced as it should be in PS9 cells. After the deletion of the mini-y reporter and introducing a wild type y+ allele, the pigmentation in the adult male abdomen is equivalent to that in wild type flies. The morphological features of McpF8R tergites and sternites also resemble those in wild type flies or the McpCTCF replacement and there is no indication of the other abdominal transformations seen in the Fab-7 replacements. Consistent with the phenotype of the adult cuticle, the pattern of Abd-B expression in the embryonic CNS resembles wild type (Fig 3B). Thus, the McpF8R replacement fully substitutes for the endogenous Mcp boundary.
A different result is obtained when the F8337 sequence is inserted in its normal forward orientation. Like the reverse orientation McpF8R, McpF8 efficiently blocks crosstalk between iab-4 and iab-5 and the mini-y reporter is not activated in A4 (PS9). On the other hand, like the Fab-7 replacements (McpF7 and McpF7R) most of the A4 tergite is covered in a yellow brown pigmentation instead of the normal stripe of yellow brown pigmentation along the posterior margin of the tergite that is seen in y1 males. Moreover, when the reporter is excised and the y1 allele replaced by the wild type y+ gene, nearly the entire A4 tergite is black. Consistent with the induction of y+ expression in A4, Abd-B is active in PS9 in the embryonic CNS (Fig 3B). The GOF transformation of A4 (PS9)→A5 (PS10) is not the only anomaly in McpF8 flies. While there does not seem to be any misspecification of the tergite or sternites in A5 (PS10), the line of trichomes along the anterior margin of the A6 tergite is disrupted or absent altogether indicating that there are some abnormalities in the temporal and/or special pattern of Abd-B expression in PS11.
In the Fab-7 replacement experiments, the relative orientation of the Fab-8 boundary was thought to be important because it determined whether the chromatin loops formed between the replacement boundary and the AB-I element and/or the PTE sequence upstream of the Abd-B transcription start site were circle loops or stem loops [30,75]. In the forward orientation circle loops are expected to be formed and in this configuration, the downstream iab-5 regulatory domain is brought into close proximity with the Abd-B promoter. In the reverse orientation, iab-6 and iab-7 are predicted to form stem loops, and this configuration would tend to isolate the iab-5 regulatory domain from the Abd-B promoter.
It seemed possible that a similar mechanism might be in play in the Fab-8 replacements of Mcp. In the forward orientation (McpF8), the iab-4 regulatory domain would be brought into close proximity to the Abd-B gene, activating its ectopic expression in A4 (PS9). In the opposite orientation, the spatial relationship between the iab-4 domain and the Abd-B promoter would not be conducive for activation. In this case, Abd-B would be off in A4 (PS9). A strong prediction of this model is that the inappropriate activation of Abd-B in PS9 in the McpF8 replacement should depend on a functional iab-4 domain.
To test this prediction, we used CRISPR (see S3 Fig) to delete a 4,401 bp sequence (iab-4Δ) that spans the putative iab-4 initiation element in flies carrying the McpF8 replacement. The iab-4Δ sequence was selected based on the clustering of multiple binding sites for transcription factors controlling segmentation of the embryo. [76]. Fig 5 shows that the ectopic activation of y+ in A4 in McpF8 flies was eliminated by the iab-4Δ deletion. Moreover, Abd-B was not activated in A4 (PS9) in the embryonic CNS of iab-4Δ McpF8 embryos (Fig 3B). Interestingly, the loss of trichomes along the anterior margin of the A6 tergite in McpF8 also seemed to depend on a functional iab-4 domain. As can be seen in Fig 5, the trichome pattern in the A6 tergite of iab-4Δ McpF8 flies resembled that of wild type.
Boundaries flanking the Abd-B regulatory domains must block crosstalk between adjacent regulatory domains but at the same time allow more distal domains to jump over one or more intervening boundaries and activate Abd-B expression. While several models have been advanced to account for these two paradoxical activities, replacement experiments argued that both must be intrinsic properties of the Abd-B boundaries. Thus Fab-7 and Fab-8 have blocking and bypass activities in Fab-7 replacement experiments, while heterologous boundaries including multimerized dCTCF sites and Mcp from BX-C do not. One idea is that Fab-7 and Fab-8 are simply “permissive” for bypass. They allow bypass to occur, while boundaries like multimerized dCTCF or Mcp are not permissive in the context of Fab-7. Another is that they actively facilitate bypass by directing the distal Abd-B regulatory domains to the Abd-B promoter. Potentially consistent with an “active” mechanism that involves boundary pairing interactions, the bypass activity of Fab-8 and to a lesser extent Fab-7 is orientation dependent.
In the studies reported here we have tested these two models further. For this purpose we used the Mcp boundary for in situ replacement experiments. Mcp defines the border between the regulatory domains that control expression of abd-A and Abd-B. In this location, it is required to block crosstalk between the flanking domains iab-4 and iab-5, but it does not need to mediate bypass. In this respect, it differs from the boundaries that are located within the set of regulatory domains that control either abd-A or Abd-B, as these boundaries must have both activities. If bypass were simply passive, insertion of a “permissive” Fab-7 or Fab-8 boundary in either orientation in place of Mcp would be no different from insertion of a generic “non-permissive” boundary such as multimerized dCTCF sites. Assuming that Fab-7 and Fab-8 can block crosstalk out of context, they should fully substitute for Mcp. In contrast, if bypass in the normal context involves an active mechanism in which more distal regulatory domains are brought to the Abd-B promoter, then Fab-7 and Fab-8 replacements might also be able to bring iab-4 to the Abd-B promoter in a configuration that activates transcription. If they do so, then this process would be expected to show the same orientation dependence as is observed for bypass of the Abd-B regulatory domains in Fab-7 replacements.
Consistent with the idea that a boundary located at the border between the domains that regulate abd-A and Abd-B need not have bypass activity, we found that multimerized binding sites for the dCTCF protein fully substitute for Mcp. Like the multimerized dCTCF sites, Fab-7 and Fab-8 are also able to block crosstalk between iab-4 and iab-5. In the case of Fab-7, its’ blocking activity is incomplete and there are small clones of cells in which the mini-y reporter is activated in A4. In contrast, the blocking activity of Fab-8 is comparable to the multimerized dCTCF sites and the mini-y reporter is off throughout A4. One plausible reason for this difference is that Mcp and the boundaries flanking Mcp (Fab-4 and Fab-6) utilize dCTCF as does Fab-8, while this architectural protein does not bind to Fab-7 [33].
Importantly, in spite of their normal (or near normal) ability to block crosstalk, both boundaries still perturb Abd-B regulation. In the case of Fab-8, the misregulation of Abd-B is orientation dependent just like the bypass activity of this boundary when it is used to replace Fab-7 [30]. When inserted in the reverse orientation, Fab-8 behaves like multimerized dCTCF sites and it fully rescues the Mcp deletion. In contrast, when inserted in the forward orientation, Fab-8 induces the expression of Abd-B in A4 (PS9), and the misspecification of this parasegment. Unlike classical Mcp deletions or the McpPRE replacement described here, expression of the Abd-B gene in PS9 is driven by iab-4, not iab-5. This conclusion is supported by two lines of evidence. First, the mini-y reporter inserted in iab-5 is off in PS9 cells indicating that iab-5 is silenced by PcG factors as it should be in this parasegment. Second, the ectopic expression of Abd-B is eliminated when the iab-4 regulatory domain is inactivated.
Our results, taken together with previous studies [30,59,60], support a model in which the chromatin loops formed by Fab-8 inserted at Mcp in the forward orientation brings the enhancers in the iab-4 regulatory domain in close proximity to the Abd-B promoter, leading to the activation of Abd-B in A4 (PS9). In contrast, when inserted in the opposite orientation, the topology of the chromatin loops formed by the ectopic Fab-8 boundary are not compatible with productive interactions between iab-4 and the Abd-B promoter. Moreover, it would appear that boundary bypass for the regulatory domains that control Abd-B expression is not a passive process in which the boundaries are simply permissive for interactions between the regulatory domains and the Abd-B promoter. Instead, it seems to be an active process in which the boundaries are responsible for bringing the regulatory domains into contact with the Abd-B gene. It also seems likely that bypass activity of Fab-8 (and also Fab-7) may have a predisposed preference, namely it is targeted for interactions with the Abd-B gene. This idea would fit with transgene bypass experiments, which showed that both Fab-7 and Fab-8 interacted with an insulator like element upstream of the Abd-B promoter, AB-I, while the Mcp boundary didn’t [59,60].
Similar conclusions can be drawn from the induction of Abd-B expression in A4 (PS9) when Fab-7 is inserted in place of Mcp. Like Fab-8, this boundary inappropriately targets the iab-4 regulatory domain to Abd-B. Unlike Fab-8, Abd-B is ectopically activated when Fab-7 is inserted in both the forward and reverse orientations. While the effects are milder in the reverse orientation, the lack of pronounced orientation dependence is consistent with experiments in which Fab-7 was inserted at its endogenous location in the reverse orientation. Unlike Fab-8 only very minor iab-6 bypass defects were observed. In addition to the activation of Abd-B in A4 (PS9) the Fab-7 Mcp replacements also alter the pattern of Abd-B regulation in more posterior segments. In the forward orientation, A4 and A5 are transformed towards an A6 identity, while A6 is also misspecified. Similar though somewhat less severe effects are observed in these segments when Fab-7 is inserted in the reverse orientation. At this point the mechanisms responsible for these novel phenotypic effects are uncertain. One possibility is that pairing interactions between the Fab-7 insert and the endogenous Fab-7 boundary disrupt the normal topological organization of the regulatory domains in a manner similar to that seen in boundary competition transgene assays [77]. An alternative possibility is that Fab-7 targets iab-4 to the Abd-B promoter not only in A4 (PS9) but also in cells in A5 (PS10) and A6 (PS11). In this model, Abd-B would be regulated not only by the domain that normally specifies the identity of the parasegment (e.g., iab-5 in PS10), but also by interactions with iab-4. This dual regulation would increase the levels of Abd-B, giving the weak GOF phenotypes. Potentially consistent with this second model, inactivating iab-4 in the McpF8 replacement not only rescues the A4 (PS9) GOF phenotypes but also suppresses the loss of anterior trichomes in the A6 tergite.
The backbone of the recombination plasmid was designed in silico and contains several genetic elements in the following order: [MCS5]-[attP]-[3xP3-EGFP-SV40polyA]-[attP]-[FRT]-[MCS3]. This DNA fragment was synthesized and cloned into pUC57 by Genewiz. The two multiple cloning sites MCS5 and MCS3 were used to clone homology arms into this plasmid. The orientations of the two attP sites are inverted relative to each other and serve as targets for фC31-mediated recombination mediated cassette exchange [62]. The 3xP3-EGFP reporter [78] was introduced as a means to isolate positive recombination events. The Flp-recombinase target FRT [79] was included for the deletion of the selectable mini-yellow marker after recombination mediated cassette exchange.
Homology arms were PCR-amplified from y w genomic DNA using the following primers: CCTGCCGACTGAACGAATGC and ACGCCCTGATCCCGATACACATAC for the proximal arm (iab-4 side; 3967 bp fragment), and GCGTTTGTGTGTAGTAAATGTATC and AAAGGCCAACAAAGAACACATGGACG for the distal arm (iab-5 side; 4323 bp fragment). A successful homologous recombination event will generate a 789 bp deletion within the Mcp region (Genome Release R6.22: 3R:16’868’830–16’869’619; or complete sequence of BX-C: 113821–114610 [4]).
The recombination plasmid was injected into y w vas-Cas9 embryos together with two gRNAs containing the following guides: GCTGGCTTTTACAGCATTTC and GCTTTGTTACCCCTGAAAAT. Concentrations were as described in Gratz et al.[80]. The injected embryos were grown to adulthood and crossed with y w partners. Among the few fertile crosses, one produced many larvae with a clear GFP-signal in the posterior part of their abdomens. This observation suggested that these animals had integrated the recombination plasmid and that the 3xP3-EGFP reporter acts as an enhancer trap for Abd-B specific enhancers. GFP positive larvae were isolated and grown to adulthood. Emerging males showed the expected Mcp phenotype. Also, and as expected for a reporter located in the BX-C, no fluorescence signal could be detected in their eyes, indicating that the 3xP3-EGFP reporter is silenced in eye cells where the 3xP3 promoter is usually active. The planned homologous recombination event could later be verified by PCR and sequencing. We will refer to it as McpattP.
12 EGFP- and Mcp-positive candidate males were individually crossed with y w virgins. Only 2 were fertile. The sterility of others may be caused by presence of off-targets as a frequent non-specific result of CRISPR/Cas9 editing. Starting from the two fertile crosses, 2 independent balanced stocks could be obtained according to established crossing schemes. One of them was used to obtain a y w M{vas-integrase}zh-2A; McpattP/TM3,Sb stock for recombination mediated cassette exchange. Because of poor survival rates in injection experiments, the McpattP chromosome was also temporarily combined with Dp(3;3)P5, Sb (y w M{vas-integrase}zh-2A; McpattP/ Dp(3;3)P5, Sb). By selection we obtained homozygous McpattP line that was subsequently used for fly injections.
For generating dsDNA donors for homology-directed repair we used pHD-DsRed vector that was a gift from Kate O'Connor-Giles (Addgene plasmid # 51434). The final plasmid contains genetic elements in the following order: [iab-4 proximal arm]-[attP]- [lox]- [3xP3-dsRed-SV40polyA]-[lox]- [iab-4 distal arm]. Homology arms were PCR-amplified from yw genomic DNA using the following primers: TTTGAATTCTTCCAGACACGCATCGGG and AAACATATGCTTGCTATCGACCCTCCTC for the proximal arm (846 bp fragment), and AATACTAGTCTCGGAAAGGGAAGAAGTTC and TACTCGAGCCGCTAAAGGACGTTCTGC for the distal arm (835 bp fragment). A successful homologous recombination event will generate a 4401 bp deletion within the iab-4 region (Genome Release R6.22: 3R:16,861,368..16,869,768; or complete sequence of BX-C [4]: 120073–115673).
Targets for Cas9 were selected using “CRISPR optimal target finder”–the program from O'Connor-Giles Lab. The recombination plasmid was injected into McpF8 vasa-Cas9 embryos together with two gRNAs containing the following guides: ATAGCAAGTAGGAGTGGAGT and GAACTTCTTCCCTTTCCGAGCGG. Concentrations were as described in Gratz et al. (2014). Injectees were grown to adulthood and crossed with y w; TM6/MKRS partners. Flies with clear dsRed-signal in eyes and the posterior part of their abdomens were selected into a new separate line. The successful integration of the recombination plasmid was verified by PCR.
3 day adult flies were collected in eppendorf tubes and stored in 70% ethanol at least 1 day. Then ethanol was replaced with 10% KOH and flies were heated at 70°C for 1–1.5h. After heating flies were washed with dH2O two times and heated again in dH2O for 45min. Then the digested flies were washed with 70% ethanol and stored in 70% ethanol. The abdomen cuticles were cut from the rest of the digested fly using fine tweezer and a needle of an insulin syringe and put in a droplet of glycerol on a glass slide. Then the abdomens were cut longitudinally on the dorsal side through all of the tergites with the syringe. To spread the cuticles flat on the slides cuts may be done between the tergites. Than the cuticles were flattened with a coverslip. Photographs in the bright or dark field were taken on the Nikon SMZ18 stereomicroscope using Nikon DS-Ri2 digital camera, processed with ImageJ 1.50c4 and Fiji bundle 2.0.0-rc-46.
Primary antibodies were mouse monoclonal anti-Abd-B at 1:100 dilution (1A2E9, generated by S.Celniker, deposited to the Developmental Studies Hybridoma Bank) and polyclonal rabbit anti-Engrailed at 1:1000 dilution (a kind gift from Judith Kassis). Secondary antibodies were goat anti-mouse Alexa Fluor 647 (Molecular Probes) and anti-rabbit FITC conjugated (Jackson Research) at 1:2000 dilution. Stained embryos were mounted in the following solution: 23% glycerol, 10% Mowiol 4–88, 0.1M Tris-HCl pH 8.3. Images were acquired on Leica TCS SP-2 confocal microscope and processed using GIMP 2.8.16, ImageJ 1.50c4, Fiji bundle 2.0.0-rc-46.
|
10.1371/journal.pcbi.1003135 | Genetic Selection for Context-Dependent Stochastic Phenotypes: Sp1 and TATA Mutations Increase Phenotypic Noise in HIV-1 Gene Expression | The sequence of a promoter within a genome does not uniquely determine gene expression levels and their variability; rather, promoter sequence can additionally interact with its location in the genome, or genomic context, to shape eukaryotic gene expression. Retroviruses, such as human immunodeficiency virus-1 (HIV), integrate their genomes into those of their host and thereby provide a biomedically-relevant model system to quantitatively explore the relationship between promoter sequence, genomic context, and noise-driven variability on viral gene expression. Using an in vitro model of the HIV Tat-mediated positive-feedback loop, we previously demonstrated that fluctuations in viral Tat-transactivating protein levels generate integration-site-dependent, stochastically-driven phenotypes, in which infected cells randomly ‘switch’ between high and low expressing states in a manner that may be related to viral latency. Here we extended this model and designed a forward genetic screen to systematically identify genetic elements in the HIV LTR promoter that modulate the fraction of genomic integrations that specify ‘Switching’ phenotypes. Our screen identified mutations in core promoter regions, including Sp1 and TATA transcription factor binding sites, which increased the Switching fraction several fold. By integrating single-cell experiments with computational modeling, we further investigated the mechanism of Switching-fraction enhancement for a selected Sp1 mutation. Our experimental observations demonstrated that the Sp1 mutation both impaired Tat-transactivated expression and also altered basal expression in the absence of Tat. Computational analysis demonstrated that the observed change in basal expression could contribute significantly to the observed increase in viral integrations that specify a Switching phenotype, provided that the selected mutation affected Tat-mediated noise amplification differentially across genomic contexts. Our study thus demonstrates a methodology to identify and characterize promoter elements that affect the distribution of stochastic phenotypes over genomic contexts, and advances our understanding of how promoter mutations may control the frequency of latent HIV infection.
| The sequence of a gene within a cellular genome does not uniquely determine its expression level, even for a single type of cell under fixed conditions. Numerous other factors, including gene location on the chromosome and random gene-expression “noise,” can alter expression patterns and cause differences between otherwise identical cells. This poses new challenges for characterizing the genotype–phenotype relationship. Infection by the human immunodeficiency virus-1 (HIV-1) provides a biomedically important example in which transcriptional noise and viral genomic location impact the decision between viral replication and latency, a quiescent but reversible state that cannot be eliminated by anti-viral therapies. Here, we designed a forward genetic screen to systematically identify mutations in the HIV promoter that alter the fraction of genomic integrations that specify noisy/reactivating expression phenotypes. The mechanisms by which the selected mutations specify the observed phenotypic enrichments are investigated through a combination of single-cell experiments and computational modeling. Our study provides a framework for identifying genetic sequences that alter the distribution of stochastic expression phenotypes over genomic locations and for characterizing their mechanisms of regulation. Our results also may yield further insights into the mechanisms by which HIV sequence evolution can alter the propensity for latent infections.
| Non-genetic heterogeneity is a ubiquitous feature of cellular gene expression that can significantly impact the genotype–phenotype relationship. Even under highly controlled culture conditions, a clonal population of cells may demonstrate a broad range of expression levels for a given gene [1]–[4]. At least some of this variability, often termed ‘noise’, is believed to arise from the intrinsically stochastic nature of the biochemical processes involved in gene expression [5], [6]. Studies that couple quantitative experimentation with mathematical modeling have begun to reveal the mechanisms by which non-genetic variability is generated and moderated [7], finding that noise: differentially impacts the expression of functional classes of genes [8], [9]; can be propagated, amplified, or attenuated by gene regulatory circuits [10], [11]; and is subject to selective pressure [12]–[15]. Stochastically-generated expression variability is increasingly appreciated to have important phenotypic consequences in diverse cellular settings, including bacterial evasion of antibiotic treatment [16], multi-cellular development [17], cancer development and progression [18], and viral latency [19], [20].
Recent evidence demonstrates that the chromosomal position of a gene, or its genomic context, affects both its mean expression level and expression noise [21]–[24]. One mechanism by which genomic context modulates gene expression is by specifying the dynamics of the local chromatin state, which can impact multiple neighboring genes [3], [25], [26]. Additionally, endogenous genes can sample different genomic environments through translocation and recombination, impacting diverse biological processes including species evolution, organism development, and cancer [27], [28]. Human retroviruses, such as human immunodeficiency virus-1 (HIV), also sample genomic environments through semi-random integration into the host genome, which in turn affects viral replication [29]. Thus, genomic context impacts cellular phenotypes and offers additional dimensions of selectable variation that shape the architecture and evolution of eukaryotic genomes, as well as the retroviruses that invade them.
Stochastic gene expression phenotypes that are modulated by genomic context present new challenges for quantifying the genotype–phenotype relationship. In particular, understanding how genomic context and gene sequence cooperate to alter gene expression dynamics requires quantifying how the sequences of regulatory elements alter the distribution of expression phenotypes over the set of genomic environments sampled by a gene. Gene regulatory networks may further alter gene expression phenotypes by amplifying or minimizing noise in gene expression through positive and negative feedback. Thus, when a genetic mutation is linked to a change in the distribution of stochastic phenotypes over genomic contexts, a further challenge is to identify the underlying mechanism that drives this change.
In this study, we identify promoter mutations that modulate context-dependent stochastic phenotypes in a lentiviral human immunodeficiency virus-1 (HIV) model system and investigate the mechanisms by which they impact viral gene expression. HIV exhibits a high degree of genetic variability due to its high replication rates [30] and the error-prone nature of reverse transcription [31], [32]. Following semi-random integration into the genome of host CD4+ T cells [29], HIV usually establishes a productive infection, but in rare cases can adopt a non-replicating but reversible latent phenotype, such as when an infected activated T cell transitions to a memory T cell [33], [34]. Latently infected cells do not express virus and thus cannot be effectively targeted by current therapeutics [35]; however, latent HIV can reactivate after long delays, leading to renewed viral spread [36]. Consequently, latent infection represents the single greatest obstacle to fully eradicating HIV in patients [37]. Importantly, a number of studies have demonstrated that genomic context and non-genetic variability play important roles in determining the replication-versus-latency decision of integrated HIV within a cell [19], [21], [22], [26]. Thus, HIV provides an ideal system for studying the interplay between gene sequence, genomic environment, and stochastic gene expression.
The virally encoded transcriptional activator Tat plays an essential role in HIV expression dynamics and the replication-versus-latency decision. The nascent HIV transcript forms a RNA hairpin, termed the HIV transactivation response element (TAR loop), that causes RNA polymerase II (RNAPII) to stall [38]. Tat binds to the TAR loop and in turn recruits the positive elongation factor b (p-TEFb), which phosphorylates RNAPII to relieve the stall and complete a cycle of transcription [39]. Transcript processing and translation then results in production of viral proteins, including more Tat. Thus, Tat enhances HIV transcriptional efficiency in a strong positive-feedback loop [40] that is necessary for viral gene expression from proviruses that immediately initiate replication or from latent infections that reactivate [41], [42].
We have previously demonstrated that an in vitro model of the HIV Tat positive feedback loop can generate a diverse range of stochastic phenotypes by sampling genomic contexts. These stochastic phenotypes include bimodal expression behaviors where non-expressing and highly expressing cells co-exist in a single clonal population [20], [43] and random switching between these two expression states occurs with significant delays. Noise in basal viral gene expression in the absence of Tat varies systematically over genomic integrations [21], [22], and its amplification by Tat feedback provides a possible mechanism to explain the diverse phenotypes generated in the presence of Tat. We have hypothesized that stochastically-driven delays in activation for some viral integrations are an intrinsic property of Tat positive feedback, and that these delays may provide a sufficient time window to establish latent infections in vivo when coupled to host-cell dynamics such as the transition to a memory T cell [20], [43]. Thus, HIV sequence mutations that affect the frequency of stochastic phenotypes in vitro may affect the frequency of latent infections in vivo. While isolated examples of promoter mutations that control context-dependent stochastic phenotypes have been investigated for HIV [43], no study has yet systematically identified such mutations or analyzed the mechanisms by which the distribution of phenotypes is modulated.
Here, we designed a forward genetic screen to select for HIV promoter mutations that increase the fraction of genomic integrations that result in stochastic gene expression phenotypes. Our screen identified important mutations in a number of core promoter regions, including Sp1 and TATA transcription factor binding sites. Through single-cell experiments, we confirmed that our strongest hits – point mutations in Sp1 site III and in the TATA box – increased the frequency of stochastic phenotypes several fold. We further demonstrated experimentally that the Sp1 mutation altered basal expression dynamics in the absence of Tat, and also impaired transactivated gene expression in the presence of Tat. Computational analysis demonstrated that the changes in basal expression observed for the Sp1 mutant could contribute significantly to the enrichment in stochastic phenotypes in the presence of impaired Tat feedback, if the mutation affected Tat-mediated amplification differentially across genomic contexts. Our analysis thus demonstrates a methodology for identifying genetic elements that affect the distribution of context-dependent stochastic phenotypes and the mechanisms by which they function. Our findings may also contribute to understanding how mutational selection could alter the frequency of latent HIV infection.
To quantitatively study stochastic gene expression of HIV infections as a function of genomic context, we adapted a full-length HIV NL4-3-based LTR lentiviral packaging platform [44] by introducing stop codons into all viral proteins except Tat and by replacing Nef with GFP (sLTR-Tat-GFP; Figure 1A). This minimal viral system, referred to in this study as wild type (WT), is similar to a model vector used previously in which Tat and GFP are expressed from a bicistronic lentiviral vector under control of the same LTR promoter [20], [43]. However, the new sLTR-Tat-GFP vector more closely mimics HIV gene expression, with Tat produced as a splice product of two exons as in natural HIV infection. The leukemic Jurkat T cell line was infected with sLTR-Tat-GFP at a low multiplicity of infection (MOI<0.1), such that the majority of infected cells (>95%) contained a single integrated provirus. The infected, GFP+ cells were then isolated by fluorescence activated cell sorting (FACS) after stimulation with tumor necrosis factor-α (TNFα) and cultured for ten days so that the population relaxed to a steady-state GFP expression profile. The resulting polyclonal or “bulk-infected” cell population showed bimodal gene expression, which indicated the presence and absence of Tat positive feedback in different cellular infections (Figure 1B), as observed with the previously studied bicistronic lentiviral vector [20], [43].
Bimodal Tat–GFP expression in the bulk-infected population arises from a mixture of integration events that result in either high or low gene expression, as well as individual integrations that result in variable or stochastic gene expression. To separate these contributions to the overall bulk distribution, we sorted individual cells – each containing a single (different) genomic integration of the provirus – from low, mid, or high ranges of GFP expression (Figure 1B). We then expanded these individual sorted cells to yield 125 single-integration clonal populations and subsequently quantified their GFP expression phenotypes by flow cytometry. Consistent with earlier studies [20], [43], a diverse spectrum of clonal GFP expression phenotypes was observed, including narrow single peaks of low or high GFP expression (referred to here as Dim and Bright distributions, respectively), as well as wide and/or bimodal distributions (Figure 1C). The wide/bimodal clonal distributions occurred with higher frequency within populations sorted from the mid-GFP range (Figure S1) and included both cells that are Bright, representing Tat-transactivated expression that would support viral replication, and cells that are Dim, representing low levels of basal expression that may be related to viral latency. Analogously, earlier work showed that when Dim cells are sorted from the bulk multi-integration population, a fraction eventually activated and migrated into the Bright range, and vice-versa [20], [22], [43]. We collectively refer to these stochastic viral gene expression phenotypes as “Switching” and consider them to be a model for latent infections that can randomly “switch” from an inactive state to a productive state.
Given HIV's rapid mutation rate [30]–[32], an interesting question is how changes in the viral promoter could affect the relative frequency of different expression phenotypes over the set of genomic environments that are sampled through infection and viral integration, and in particular whether specific mutations could increase the frequency of Switching phenotypes. As a first step in addressing this question, we developed objective, feature-based clustering criteria to classify gene expression behavior for a clonal population as Switching, Dim, or Bright. In this classification, cut-off values were manually selected for nine GFP-distribution measures that reflect expression heterogeneity, such as bimodality, width, and skewness (Table S1 and Figure S2). Distributions with a value exceeding the cut-off for any one of these features were labeled as Switching (details of methods described in Text S1). By applying these criteria uniformly to our initial collection of single-integration clones (Figure 1C), we estimated the fraction of integrations in our system that led to a Switching phenotype to be 8.2% (Figure 1D). We developed an alternate estimate of the Switching fraction based on sampling single-integration clones sorted only from the mid-GFP range and extrapolating to the full population (see Text S1). This method resulted in a similar Switching fraction estimate of 8% (Figure 1D), and was thus used in the remainder of our study for increased experimental efficiency.
We next developed a stochastic model of HIV transcription and amplification by the Tat positive feedback loop to aid our intuition concerning the underlying gene expression dynamics that may account for the observed variation in HIV expression phenotypes (Figure 2A). We previously built a model of basal LTR promoter-driven gene expression in the absence of Tat, which probabilistically described the processes of gene activation, transcription, and translation [22]. Our analysis suggested that basal transcription from the LTR occurs in short, infrequent bursts, and we found that the size of these transcriptional bursts strongly correlated with mean gene expression from different viral integration positions [22]. Here, we extended this basic model to include Tat expression from the LTR, and Tat positive feedback on transcription from the LTR, by assuming a Michaelis-Menten-like dependence of transcriptional burst size and burst frequency on Tat concentration (full model description included in Text S1). The assumption that Tat positive feedback enhances the frequency of transcriptional bursts from the LTR is consistent with observations that Tat interacts with transcription factors involved in gene activation [45], [46], and the assumption that Tat increases transcriptional burst size is based on observations that Tat enhances elongation by recruiting p-TEFb [39].
The model is specified by two basal transcription parameters, which set the average size and frequency of transcriptional bursts that occur in the absence of Tat, and three feedback parameters that describe transcriptional amplification in the presence of Tat. Two of these feedback parameters, which specify the average size and frequency of transcriptional bursts at saturating Tat concentrations (full transactivation), were set to give approximately a 100-fold increase in transcription rate at saturating Tat concentrations [40]. The third feedback parameter, which specifies the Tat concentration at half maximal binding, was set to approximately the top of the mid range of our bulk expression distributions (Figure 1B). The remaining model parameters (including degradation and translation rates) were set as in previous work [22]. The model, which was solved numerically for steady-state protein distributions, reproduced each of our major experimental expression phenotypes over different ranges of parameter values (Dim, Bright, and Switching (Figure 2B).
We qualitatively analyzed the relationship between transcriptional dynamics and expression phenotype in our model by generating a series of phase diagrams. These phase diagrams fix the Tat feedback parameters in our model as described above, and then systematically scan over basal transcription parameters, which are known to vary over genomic integrations [21], [22]. By applying our experimental criteria for Dim, Bright, and Switching phenotypes to our simulated distributions, we drew boundaries separating combinations of basal transcription parameters that lead to distinct expression phenotypes in our model (Figure 2C).
Interestingly, near the range of model parameters that generate Switching phenotypes, small changes in basal transcription that occur in the absence of Tat result in large changes in phenotype when amplified by Tat feedback (Figure 2B). Additionally, we found that Switching phenotypes exhibit delayed activation of gene expression. That is, if a simulated population of cells with model parameters corresponding to a Switching phenotype is initialized in the Dim state, a time-scale of one to many weeks is required for half of the population to cross a threshold of gene expression intermediate between Dim and Bright states (Figure 2B). This is in contrast to a Bright steady-state phenotype initialized in the Dim state, which will cross an intermediate expression threshold on a time scale of days (corresponding to the time scale of protein dilution in our cells). The delayed activation observed for the Switching phenotype is approximately the time scale over which an activated CD4+ T cell may transition to a memory state, and memory T cells are a primary reservoir of latent HIV infection in vivo [33], [34]. Thus, the delayed transcriptional activation exhibited by a Switching phenotype could substantially increase the opportunity for the memory state transition to occur in an infected T cell before viral production, and may therefore increase the probability of a latent infection.
The general relationship between Switching phenotypes and delayed activation is highlighted by superimposing a measure of distribution activation time on the phenotypic information in our phase diagrams (Figure 2C). Delayed activation results when transactivation depends on the probabilistic (infrequent) occurrence of multiple transcriptional bursts that are larger and/or more closely spaced than occur on average. In our model, such behavior occurs at intermediate values of basal transcriptional burst size and frequency, which are typically the same values that specify Switching phenotypes (additional discussion in Text S1). Our model thus supports the hypothesis that Switching phenotypes also exhibit delayed activation, which may underlie the establishment of latent HIV infections [20], [22], [43].
Finally, we note that Switching phenotypes also exhibit delayed deactivation of gene expression as compared to Dim clones when initiated in a Bright state. Although delayed deactivation is not relevant to the establishment of latent infections in vivo (due to the fact that viral replication would kill the host cell and block any possible memory state transition before deactivation could occur), it is possible to observe this behavior in our in vitro model. Thus, we hypothesized that probabilistic delays in both activation and deactivation can be used to select for Switching phenotypes in our in vitro system.
We exploited the delayed activation/deactivation of gene expression associated with Switching phenotypes to design a forward genetic screen to identify LTR promoter mutations that increase the prevalence of Switching phenotypes, and which could thus potentially influence the fraction of latent infections. We prepared a library of HIV-1 vectors in which the WT LTR promoter was subjected to random point mutations via error-prone PCR (Figure 3A) [47]. The ∼105 member library had an average mutation rate of 0.6%, such that each position of the 634 base-pair promoter was mutated hundreds of times across the library. We packaged the library into our model vector, infected Jurkat cells, and isolated cell populations containing single viral integrations as described for the WT vector above. The resulting bulk population of singly infected cells, which was heterogeneous in both LTR sequence and viral integration position, was subjected to two alternate phenotypic screens. First, we implemented an ‘activation’ screen, in which infected cells with low GFP expression (low GFP gate) were isolated by FACS and allowed to grow for 5 days, at which point cells that had switched to high GFP expression (high GFP gate) were selected again by FACS. Second, a ‘deactivation’ screen reversed the order, selecting for high GFP expression first and low second (Figure 3A). We refer to the fraction of cells selected in these screens as the activating and deactivating fraction, respectively.
To confirm that our activation screen effectively selected for clones with a Switching phenotype, we applied the activation screen to the WT virus and randomly selected a sample of single cells from the activating fraction, which were then expanded to clonal populations for analysis. Remarkably, nearly 54% of these clones (22 out of 42) showed Switching phenotypes, as compared to only 8% from the original population and 19% from the mid-sorted population (Figure S1), confirming the effectiveness of the screen.
We thus implemented a larger scale analysis to identify viral promoter mutations that favor Switching phenotypes. Specifically, we performed multiple rounds of infection and FACS-based screening as described above to average the behavior of promoter sequences across different integration positions and thus identify genotypes that give rise to a higher fraction of Switching phenotypes across genomic contexts. After each round of infection, we recovered the viral LTRs from the genomic DNA of the selected populations (by PCR), re-cloned them into the sLTR vector, repackaged virus to produce a new library of selected promoters, and infected a new population of Jurkat cells (Figure 3A).
After four rounds of selection, the fraction of activating cells increased 6-fold compared to the original library (p<0.001, t-test on triplicate measurements) and 2-fold compared to the WT promoter (p<0.01; Figure 3B). The fraction of deactivating cells increased by a factor of 1.7 compared to the original library (p<0.04) and by a factor of 3 relative to WT (p<0.002; Figure 3C). Interestingly, the median GFP expression of the Tat-transactivated population (Bright peak in the bulk GFP histogram) was significantly lower for the unselected library than for WT, and it continued to decrease with each round of selection in both screens (Figure 3D–E). Importantly, the bulk gene expression distributions of the selected promoters also displayed an increased weight in the mid range of GFP expression (Figure 3F–G), which we had found to be enriched in integrations that demonstrate a Switching phenotype for the WT promoter. Altogether, these results indicate that our dynamic screens for activation and deactivation effectively selected for mutations that increased the fractions of activating and deactivating cells, which is a hallmark of the Switching phenotype.
To analyze the LTR promoter mutations that were enriched by the activation and deactivation screens, approximately 90 clones were sequenced from each selected library and compared to a control set of promoters from the unselected library. The average mutation frequency per position in the selected libraries was approximately 1.1% (as compared to 0.6% for the unselected library), but the distribution of mutation frequencies was long-tailed, with some positions mutated in as many as 20% of the promoters for a given screen (Figure 4A). We first analyzed how mutations were distributed across the LTR for the combined screens by comparing the mutation frequency for each regulatory region of the LTR with the average mutation frequency over the whole promoter [48] (Figure 4A). For both screens, mutations were most significantly enriched in the 78 base-pair core promoter region (p<0.0001, Chi-squared test), which includes transcription factor binding sites required for efficient promoter activation [48]. In contrast, mutation rates were not increased above those in the initial library in the enhancer region, the U5 region downstream of TAR, and in the TAR region itself, possibly reflecting the essential role of the TAR loop secondary structure to enable efficient gene expression [49]. The remaining regions displayed increased mutation frequencies that did not differ significantly from the average increase across the entire promoter for both selected libraries.
We next compared the mutation frequency at each position in the core promoter to the mutation frequency for the same base type in the unselected library (Figure 4B). We identified two positions in Sp1 site III, one position in Sp1 site II, and two positions in the TATA box with significant mutation rates in both screens (Table S2), with additional Sp1 and TATA positions significantly mutated in one of the two screens. The top hit was in Sp1 site III (position 4 of the 10 bp site, p<0.0001). Selection for this mutation is consistent with our previous results demonstrating that simultaneous mutation of positions 3 and 4 in Sp1 site III, which had been shown to eliminate binding of Sp1 [50], also increased delayed activation and deactivation in infected Jurkat cell populations [43]. The next strongest hit was in the TATA box (position 2 of the 8 bp site, p = 0.0005). The A to G mutation observed most frequently in our selected libraries has been previously shown to reduce the affinity of the TATA binding protein (TBP) for the TATA box [51]. Notably, mutations at positions 3 and 4 of the TATA box, which are considered critical for TBP binding and thus TATA function [51], [52], were not enriched in either screen.
Altogether, for the activation screen we found that 40% of the sampled sequences had mutations in Sp1 site III, and 25% had TATA mutations; for the deactivation screen, 20% had mutations in Sp1 site III, and 20% had TATA mutations (Figure 4C). All of these mutation frequencies were well above those for the same regions in the unselected library. Together, these results suggest the importance of Sp1 site III (and to a lesser extent the TATA box) in controlling stochastic gene expression and Switching fractions.
To directly analyze how the point mutations identified in our screen affect gene expression, we generated vectors with a single point mutation at position 4 of the Sp1 site III (Sp1 mutant) or at position 2 of the TATA box (TATA mutant) (Table S3), and infected Jurkat T cells as previously described. The TATA mutation increased both the activating and deactivating fractions of the infected population by approximately 2.5-fold relative to WT (p<0.01; Figure 5A–B), and the Sp1 mutation increased the activating fraction 1.5-fold (p<0.03; Figure 5A) and the deactivating fraction almost 7-fold relative to WT (Figure 5B, p<0.001). Both point mutations also significantly decreased Tat-mediated gene expression and increased expression in the mid-range of fluorescence (Figure 5C), mirroring the bulk expression phenotype of the full library after selection, and consistent with previous studies [43], [53], [54].
We next quantified Switching fractions for both mutants by sorting approximately 80 single-integration clones from the mid-range of GFP in the bulk populations as previously described for the WT virus (Figure 1). The Switching fractions increased from 8% for the WT virus to 25% for the TATA mutant and 46% for the Sp1 mutant (Figure 5D). These results confirm that increased activation and deactivation in the bulk infection for these mutants reflect an increased frequency of single-integration clonal Switching phenotypes (p<0.01, bootstrap method).
We next considered how promoter mutations might alter transcriptional dynamics to increase the fraction of infections that generate Switching phenotypes. For this analysis, we chose to focus on the Sp1 point mutation, because this point mutation exists in naturally occurring HIV isolates, while the TATA mutation was not found (as determined by searching the Los Alamos HIV sequence database, http://www.hiv.lanl.gov). Furthermore, our previous work also demonstrated a role for Sp1 site III in regulating Switching phenotypes [43].
Our earlier work demonstrated that basal transcription (i.e. in the absence of Tat) varies significantly with integration position of the LTR [22]. Therefore, we hypothesized that Sp1 may modulate phenotypic distributions by directly affecting basal transcription. To test this hypothesis, we introduced stop codons into the first Tat exon of the lentiviral vector backbones of the WT and the Sp1 mutant promoter and infected Jurkats as described above (Figure 6A). Bulk-infection expression distributions for both Tat-null vectors demonstrated substantial overlap with autofluorescence controls, but with a strong right skew towards higher fluorescence. Notably, a small but significant decrease in mean GFP expression was observed for the Sp1 mutant promoter compared to WT (p<0.05), consistent with previous studies [53], [54]. Additionally, clonal cell populations expanded from each bulk population had monomodal, wide, right-skewed distributions (Figure S3) and displayed high levels of noise across clonal expression means (Figure 6B), consistent with previous results for the WT LTR promoter [22].
To infer the underlying transcriptional dynamics of our Tat-null clones, we systematically fit their GFP distributions using our model (Figure 2A with transactivation removed), following our earlier analysis of WT basal expression dynamics [22]. The sets of clonal WT and Sp1 distributions were all best accounted for by a bursting dynamic, whereby short infrequent transcriptional bursts generate large basal expression heterogeneities (see Text S1 and [22] for further discussion). The basal transcription dynamics for each clonal population were fully quantified by a best-fit basal transcriptional burst size and burst frequency. Transcriptional burst sizes were found to vary from a few to tens of transcripts, and to be strongly positively correlated with mean expression level across different integration positions for both the mutant and for the WT vector (Figure 6C). In contrast, typical transcriptional burst frequencies were on the order of a few events per cell division time, and demonstrated little correlation with mean gene expression levels over integration positions (Figure 6D). These findings are consistent with our earlier analysis of the WT promoter [22].
Although the Sp1 mutant and WT promoters share the same qualitative basal expression dynamics, regression analysis revealed that the Sp1 mutant demonstrated an increased positive correlation between basal burst frequency and clonal expression mean, with burst frequencies decreased for Dim clones (Figure 6D; p = 0.04). Thus, the selected Sp1 mutation does not change the qualitative bursting mode of transcription from the HIV LTR, but it does appear to modestly alter how the dynamics vary quantitatively across integration positions.
We returned to our model to explore if the small changes in basal transcriptional dynamics quantified experimentally with our Tat-null vector could contribute significantly to the increased Switching fraction observed for the Sp1 mutant in the presence of Tat (Figure 5D). The phase diagrams developed for the WT promoter (Figure 2C) specify the predicted expression phenotype for every combination of basal transcriptional burst size and burst frequency parameters for fixed Tat feedback. Thus, model phase diagrams can be used to predict the Switching fraction that would result from a given probability density with which the virus samples basal transcriptional parameters through its sampling genomic locations via infection and integration, under the assumption of fixed Tat feedback. We used our experimental data to estimate the probability density with which the WT and Sp1 mutant promoters sampled combinations of basal transcription parameters (see Text S1 for details), and then calculated model-predicted Switching fractions by integrating this sampling density over the Switching region of the phase diagram (Figure 7A).
We found that the changes in basal transcriptional dynamics observed for the Sp1 mutant – particularly the increased sampling of lower transcriptional burst frequencies, which specify noisier basal transcription – indeed resulted in higher model-predicted Switching fractions compared to WT for all sets of feedback parameters analyzed. In particular, for a set of feedback parameters that specify a model-predicted Switching fraction of 12% for the WT basal parameter sampling density, the model predicted a Switching fraction of 22% for the Sp1 mutant sampling density (Figure 7B). Thus, we conclude that changes in Sp1 basal transcription dynamics can result in a substantial increase in the fraction of genomic integrations that lead to a Switching phenotype in the presence of Tat feedback.
In addition to altering basal expression, mutations in Sp1 site III weaken Tat positive feedback, as demonstrated in our experiments (Figure 5C) and in previous work [53]; however our model had not yet accounted for this observation. We therefore explored if weakening Tat positive feedback in the model would maintain the predicted Switching fraction enrichment that arises from altered basal transcription, or even enhance it to more fully account for the nearly 6-fold enrichment observed in our experiments. In contrast to these expectations, we found that decreasing Tat-driven fold-amplification of basal transcription in the model typically decreased predicted Switching fractions (Figure 7B), a result which can be explained by our model phase diagrams (Figure 7A). Notably, weakening feedback shifts phenotypic boundaries to the right (towards larger basal transcriptional burst sizes), transforming Bright integrations to Switching, and Switching to Dim. The resulting Switching region typically enclosed a smaller fraction of the viral basal parameter sampling density, which is highly right skewed and heavily weighted at lower basal transcriptional burst sizes. Thus, our analysis suggests that the Sp1 site mutation specifies a more complex perturbation of the Tat positive feedback loop that differentially affects Bright and Dim integrations, rather than one that uniformly attenuates expression amplification over genomic integrations.
A biological mechanism by which the Sp1 site mutation could differentially affect Bright and Dim integrations is by impairing transcriptional reinitiation. In the bursting model of transcription, each gene activation event can drive multiple cycles of transcription, requiring multiple rounds of RNAPII binding and transcription-complex formation (i.e. reinitiation). In the absence of Tat, the rate-limiting step in HIV-LTR transcription is RNAPII stalling at the TAR hairpin that forms after transcriptional initiation [38]. Therefore, moderate impairment of the reinitiation rate via mutation would be masked during basal transcription, or for integrations that inefficiently activate Tat feedback. However, at higher concentrations of Tat, when the TAR-loop stall is no longer rate limiting, impaired reinitiation would significantly attenuate full Tat transactivation, and the effect would be more pronounced for Brighter genomic integrations. Because Sp1 and p-TEFb interact in vivo to activate HIV transcription [46], [55], [56], a mutation in the Sp1 site could plausibly alter transcriptional reinitiation if it disrupted recruitment of p-TEFb.
To investigate this possibility, we extended our model to include a ‘reinitiation’ step between each transcript production event (rescaled model parameters included in Figure 7 legend and full model description and equations included in Text S1). The effective transcript production rate in this extended model depends on both an elongation rate, which varies over genomic integrations, and a reinitiation rate, which is fixed (but may be altered through mutation). The elongation rate specifies the variation of the basal and transactivated transcription rates over genomic integrations, while the reinitiation rate specifies the maximal value at which the transcription rate saturates as a function of elongation rate. In this extended model, we found that a moderate decrease in the transcriptional reinitiation rate had little effect on the phenotypic boundaries of our phase diagrams (Figure 7B), but significantly weakened Tat-transactivated expression from Bright integrations (Figure 7C), consistent with our experimental observations (Figure 5C). As a result, predicted Switching fractions were preserved, though they were not further enhanced to the level observed experimentally. Thus, moderate impairment of transcriptional reinitiation could account for the observed attenuation in Tat-mediated gene expression (Figure 7B), while preserving (but not increasing) the Switching fraction enhancement that was predicted for the observed changes in Sp1 mutant sampling of basal transcription parameters.
Amplification of HIV gene expression noise via Tat positive feedback results in a wide range of noise-driven phenotypes that vary across the diverse host genomic environments sampled during HIV infection. Here, using an in vitro cell-based HIV model system and a novel forward genetic screen, we identified LTR promoter mutations that increase the frequency of the Switching phenotype, a model for latent viral infections. Two key features of our screen are 1) its dynamic nature, which selects for stochastic phenotypes that ‘switch’ between quiescent and highly expressing states; and 2) integration randomization, which applies selective pressure on mutations affecting the fraction of integrations that specify Switching phenotypes rather than on the integration positions themselves. These features reflect the time-varying selective pressure that is likely applied by a dynamic immune system and therapy schedule, and the integration randomization that occurs when a viral lineage is propagated by new infections in vivo. Our forward genetics approach enabled the systematic identification of promoter elements that affect the Switching fraction, and complements prior reverse genetics approaches that analyzed how specific mutations affect gene expression and phenotype [43], [53], [57]. The screen identified strongly selected mutations in Sp1 and TATA transcription factor binding sites within the core transcriptional regulatory region of the HIV LTR, and we confirmed that these mutations led to higher frequencies of Switching phenotypes across integration positions.
Our study was enabled by the development of a computational model that described how promoter-driven expression fluctuations are propagated via Tat positive feedback to generate the wide range of expression phenotypes in our system. We used this model to investigate features of Tat feedback that generate stochastic phenotypes, to formulate hypotheses concerning the mechanisms by which these features may be varied through mutation, and to study the implications and consistency of these hypotheses with our experimental data.
The Tat transactivation circuit – an essential and conserved feature of the HIV virus across clades – is characterized in our model by positive feedback loops that enhance both the size and frequency of transcriptional bursts. HIV gene expression phenotypes range from Dim to Bright as the kinetic parameters of the circuit are varied, with intermediate parameter values generating the stochastic Switching phenotypes that our screen was designed to select. These Switching phenotypes, which we have suggested may serve as a model for latent infection [20], [22], [43], are characterized by Tat-amplified transcriptional fluctuations that drive stochastic switching between quiescent and highly expressing states (Figure 2). Importantly, all of the transcriptional and regulatory processes described in our model – and their underlying kinetic parameters – can be modulated by genomic environment. Thus, a viral sampling of genomic environments that range from repressive to permissive can tune the steady-state behavior of Tat positive feedback circuit to generate a distribution of expression phenotypes that span from Dim to Bright, with intermediate integrations generating Switching phenotypes [41]. In this way, the possibility of stochastically-generated latent phenotypes at a subset of viral integrations may be an intrinsic feature of the Tat circuit and its sampling of host-cell genomic environments, and the virus may tune the fraction of integrations that specify this phenotype through mutation.
Guided by our model analysis both here and in previous work [22], we hypothesized that the Sp1 mutation may alter the prevalence of Switching phenotypes by modulating basal transcription dynamics. Although the underlying basal bursting dynamic of the WT promoter was essentially preserved in the Sp1 mutant (Figure 6), we were able to detect modest quantitative differences in the sampling of basal expression dynamics over integration positions. Our computational analysis confirmed that these small differences in basal expression for the Sp1 mutant could be amplified in the presence of Tat feedback to produce substantial increases in the Switching fraction (Figure 7B).
The selected Sp1 mutant also demonstrated weaker Tat-transactivated expression, and we further used our model to investigate how this could affect the Switching fraction. Our model analysis demonstrated that weakening Tat feedback proportionately for all integrations would decrease, rather than increase, Switching fractions (Figure 7B). Thus, accounting for an increased Switching fraction in the presence of weaker Tat feedback required a mechanism by which the selected mutation could differentially affect basal and transactivated expression, which we suggested could be accomplished through impaired transcriptional reinitiation. A revised computational model that included impaired transcriptional reinitiation could thus account qualitatively for both trends observed experimentally for the Sp1 mutant: an enhanced Switching fraction accompanied by attenuated Tat-transactivated expression (Figure 7B–C). However, we note that our model still does not quantitatively account for the full increase in Switching fraction observed experimentally for the Sp1 mutant (Figure 5D). A complete explanation might thus require identification of additional mechanisms that differentially affect Tat transactivation across genomic integrations and a more detailed characterization of how the selected mutations perturb the transcription parameters sampled by the virus over genomic integrations.
Multiple studies have demonstrated that mutations in the Sp1 sites of the HIV LTR can significantly reduce HIV Tat-mediated transactivation, while minimally affecting basal expression (for those cases in which it was measured) [53], [55], [58], [59]. Although the detailed mechanisms by which Sp1 regulates HIV expression remain unknown, there is evidence that Sp1 recruits P-TEFb in vivo to release the stalled RNAPII from the promoter proximal region and activate transcriptional elongation of HIV [46], [55], [56]. To our knowledge, a role for Sp1 in transcriptional reinitiation has not been directly tested. However, if Sp1 participates in recruitment of P-TEFb, then lower affinity Sp1 binding (caused by promoter mutation) may destabilize the P-TEFb complex in the promoter active state and thus lower the rate of transcriptional reinitiation (κr in our model).
Interestingly, TATA mutations in the HIV LTR also substantially reduce Tat-mediated transactivation without affecting mean basal expression from the HIV LTR [53], [54], [60], [61], similar to observations by others and us for Sp1 mutation. Although we did not explore the mechanisms underlying mutation of the TATA box, an increase in the half-time of transcriptional reinitiation (1/κr in our model) has been measured directly for a mutation at site 2 of the TATA box [62]. Furthermore, TATA box mutations that decreased reinitiation also correlated with decreased stability of the TBP:TFIIA (general transcription factor) complex on the DNA, suggesting that retention of general transcription factors at the promoter is a primary determinant of the reinitiation rate [63]. Our results motivate a future experimental study that directly measures if reduced transcriptional reinitiation provides a mechanistic explanation for the differential effect of Sp1 and TATA box mutations on basal and Tat-transactivated HIV transcription, as observed here and in many previous studies [53], [55], [58], [59].
In vivo, infected CD4+ T cells that have transitioned to a memory state form a primary reservoir of latent infection [33], [34]. However, HIV does not efficiently establish infection in resting memory CD4+ T cells [64], [65], and activated CD4+ T cells typically die within days after infection [30]. Therefore, we hypothesize that transcriptional delays, such as those associated the Switching phenotype in our in vitro system and that occur on a similar time scale to the memory-state transition, could delay viral production and thus increase the time window during which the memory-state transition could occur post-infection. Thus, viral mutations such as the Sp1 and TATA mutations identified in our study, which result in an increased fraction of viral integrations demonstrating transcriptional delays, could lead to an increase in the fraction of memory T cells that harbor a latent infection.
If this in vitro model of latency has in vivo implications, then our results suggest that there may be enrichment for viruses with an Sp1 and/or TATA box mutation in the latent reservoir. Although we are unaware of any direct evidence of enrichment for either Sp1 or TATA box mutations in the latent pool, there is evidence that viruses with an Sp1 site III mutation are enriched during the course of disease progression [66] and that viruses with impaired Tat activity are enriched in latent reservoirs [67]. These studies are suggestive that some viral mutations, particularly ones affecting Tat transactivation as demonstrated in our study, may create favorable conditions for establishing latent infections. Interestingly, these studies suggested that lower transcriptional activity may underlie the propensity of these viruses to establish a latent infection, but our results suggest it is instead the increased probability for transcriptional delay that potentiates latent infection. A related and testable hypothesis is that the three HIV subtypes (D, F and H) with mutations in Sp1 site III may demonstrate an increased propensity for latency and thus give rise to larger latent reservoirs relative to subtype B infection. To our knowledge, there is no study that has examined the relative sizes of the latent viral reservoirs for different HIV subtypes, and therefore this may be an important translational study that is motivated by our work.
In conclusion, our study provides an integrated experimental and computational framework for identifying genetic sequences that alter the distribution of stochastic expression phenotypes over genomic locations and for characterizing their mechanisms of regulation. Our results also may yield further insights into the mechanisms by which HIV sequence evolution can alter the propensity for latent infections.
HEK293T cells (ATCC) were cultured in IMDM (Mediatech) and Jurkat clone E6 cells (ATCC) were cultured in RPMI (Mediatech). All media was supplemented with 10% FBS (Gibco) and 100 U/ml penicillin+100 mg/ml streptomycin (Gibco). Jurkat cell concentrations were maintained between 2×105 and 2×106 cells/ml in 5% CO2 at 37°C.
We modified a full-length single-LTR packaging platform described previously in which HIV Nef was replaced with GFP [44]. Multiple stop codons were introduced into all viral proteins except Tat (psLTR-Tat-GFP; Table S4) using Quickchange site-directed mutagenesis (Stratagene). To generate Tat-null sequences, additional stop codons were introduced into the first exon of Tat (psLTR-TatKO-GFP). The LTR promoter library was amplified in an error-prone PCR reaction described previously [47] using Taq DNA polymerase with 2% MnCl2. The resulting promoter library was cloned into the psLTR-Tat-GFP by restriction digest with PmeI and KasI. Following each round of selection, the genomic DNA from the selected cells was isolated using a QiaAMP DNA Micro Kit (Qiagen) and the LTR promoters of the integrated proviruses were amplified with primers that retained the PmeI and KasI restriction digest sites for cloning. Single point mutations in the LTR were introduced with Quickchange site-directed mutagenesis (see Table S3 for sequences) and each mutant LTR was sequenced and subcloned back into the parental plasmid to avoid unintended mutations. All psLTR-Tat-GFP and psLTR-TatKO-GFP plasmids were packaged and harvested in HEK 293T cells with helper plasmids (pcDNA3 IVS VSV-G, pMDLg/pRRE, pRSV Rev, and pCLPIT-tat mCherry) as previously described [20], [68]. Harvested lentivirus was concentrated by ultracentrifugation to yield between 107 and 108 infectious units/ml. To titer, Jurkat cells were infected with a range of vector concentrations and six days post infection, gene expression of infected cells was transactivated by stimulation with 20 ng/ml PMA (Sigma) and 400 nM TSA (Sigma). After stimulation for 18–24 hours, GFP expression was measured by flow cytometry, and titering curves were constructed by determining the percentages of cells that exhibited GFP fluorescence greater than background levels.
Jurkat cells were infected with the sLTR-Tat-GFP virus at an MOI of ≤0.1 and cultured for 7–10 days. Cells were stimulated with 20 ng/ml TNF-α (Peprotech) for 18–24 hours and GFP+ cells were sorted on a MoFlo Cell Sorter (Cytomation). Sorted cells were cultured for 10 days. For the activation screen, cells were sorted from the off peak (bottom third of the full range of GFP expression), cultured for 5 days, and then selected as positive for enrichment if the cells activated above the mid-point of the expression range. For the inactivation screen, cells were sorted from the bright peak (top third of the full range of GFP expression), cultured for 5 days, and then selected as positive for enrichment if the cells inactivated below the mid-point of the expression range. Flow cytometry data analysis was performed with FlowJo (Tree Star, Inc.).
For LTR-Tat-GFP infections, single cells were selected from the region of interest (bottom third of the expression distribution for off cells, mid third of the expression distribution for bimodal cells, and top third of the expression distribution for bright cells). For the LTR-Tat-null vector, single cells were selected from either the top 10% or 18% of the GFP expression distribution and sorted into each well of a 96-well plate on a MoFlo Cell Sorter (Cytomation). Clonal cells were cultured for 2–3 weeks and then analyzed on an FC500 flow cytometer (Becton Dickenson).
Fluorescence histograms for single-integration clonal sLTR-Tat-GFP infections were labeled as Switching if they exceeded specified cut-offs in any of the following distribution features: inter-quartile range, cube root of 3rd central moment, peak separation and dip for bimodal distributions, and the product of distribution weight in approximately the lower third and upper half of our cytometer log fluorescence range. Feature cut-offs were specified by visualizing the full set of clonal distributions using k-means clustering based on 8 distribution features normalized by inter-quartile range (those mentioned, and mean log fluorescence, distribution weight in the lower 3rd of the bulk fluorescence range, and distribution 4th central moment) using 20 clusters and a Euclidean distance, implemented in Matlab (The Mathworks). Sorting clusters separately by each feature centroid allowed identification by eye of features and cut-off values beyond which all distributions could be labeled as Switching. This approach extended our by-eye intuition from distributions whose phenotype could be unambiguously scored by eye to those whose phenotype was ambiguous (see Text S1 for further details). Key results, such as Switching fraction enrichment for our analyzed mutants, were robust to variation of feature cut-off values.
Switching fractions, over the full set of genomic integrations, were estimated from mid-sorted sub-samples, via an application of Bayes theorem:where S is the event that an infected cell contains a Switching integration and M is the event that the cellular fluorescence is in the range of the sorting gate (i.e. mid range). The conditional probability, , was estimated as the fraction of clones from a given mid sort that were labeled as Switching ( where is the total number of clones analyzed from the mid sort and is the number that were labeled as Switching). The probability that a cell expresses fluorescence in the range of the sort, , was estimated by the distribution weight of the bulk multi-integration population in the sort range. , the distribution weight in the sort region for the full population of Switching integrations, was estimated from our mid-sorted set of Switching clones as:where the are individual distribution weights of the mid-sorted Switching clones in the sort region. Uncertainties in Switching-fraction estimation were calculated based on a bootstrap approach [69]. Further details are provided in Text S1.
Our model of the transactivation circuit considers each reaction as a Markov process, proceeding with fixed probability per unit time (full model details in Text S1). For any fixed set of parameter values, the model was solved to obtain predicted steady-state protein distributions across a clonal population of cells by approximating and numerically integrating the master equation for the system [70] in time until a stationary distribution was achieved. Protein numbers were convert to cytometer RFU by scaling, and distributions were convolved with a measured autofluorescence profile for comparison with experimental distributions, following [22].
Tat-null distributions were fit to the transactivation model with feedback from Tat removed based on the first 6 distribution moments (see Supplemental text for further details). Transcriptional bursting was assumed, so that transcriptional burst size () and burst frequency () were the only model fit parameters, with the remaining model parameters calibrated following [22]. The assumption of transcriptional bursting was checked by systematically varying the active-state duration () and refitting the model at each value. Consistent with [22], the best fits were always found in the transcriptional bursting regime (). All analysis was done using in-house code written in Matlab (The Mathworks).
Statistical significance of differences between means in triplicate experiments was assessed using a 2-sided t-test. Pearson Chi-squared statistics were calculated for the appropriate contingency tables to assess differences in mutation rates between libraries marginally and by regulatory region, and at individual positions along the promoter, after controlling for base type in the WT (parent) sequence. All quoted raw p-values for post-hoc analysis remain significant at the level for Type I error after Bonferroni correction, and corresponding global tests were always significant at least at this level. Equality of regression coefficients was assessed by partial F-test, and differences between individual regression parameters were assessed by t-test in post-hoc analysis. Confidence intervals for experimental Switching fraction estimates, and p-values for their differences, were estimated using a bootstrap procedure. Contingency table analysis was conducted using SAS/STAT software version 9.1 for Windows, Copyright 2012 SAS Institute Inc. All other computational analysis was performed using Matlab (The Mathworks).
|
10.1371/journal.pgen.1007692 | TIMP3 and TIMP1 are risk genes for bicuspid aortic valve and aortopathy in Turner syndrome | Turner syndrome is caused by complete or partial loss of the second sex chromosome, occurring in ~1 in 2,000 female births. There is a greatly increased incidence of aortopathy of unknown etiology, including bicuspid aortic valve (BAV), thoracic aortic aneurysms, aortic dissection and rupture. We performed whole exome sequencing on 188 Turner syndrome participants from the National Registry of Genetically Triggered Thoracic Aortic Aneurysms and Cardiovascular Related Conditions (GenTAC). A gene-based burden test, the optimal sequence kernel association test (SKAT-O), was used to evaluate the data with BAV and aortic dimension z-scores as covariates. Genes on chromosome Xp were analyzed for the potential to contribute to aortopathy when hemizygous. Exome analysis revealed that TIMP3 was associated with indices of aortopathy at exome-wide significance (p = 2.27 x 10−7), which was replicated in a separate cohort. The analysis of Xp genes revealed that TIMP1, which is a functionally redundant paralogue of TIMP3, was hemizygous in >50% of our discovery cohort and that having only one copy of TIMP1 increased the odds of having aortopathy (OR = 9.76, 95% CI = 1.91–178.80, p = 0.029). The combinatorial effect of a single copy of TIMP1 and TIMP3 risk alleles further increased the risk for aortopathy (OR = 12.86, 95% CI = 2.57–99.39, p = 0.004). The products of genes encoding tissue inhibitors of matrix metalloproteinases (TIMPs) are involved in development of the aortic valve and protect tissue integrity of the aorta. We propose that the combination of X chromosome TIMP1 hemizygosity and variants of its autosomal paralogue TIMP3, significantly increases the risk of aortopathy in Turner syndrome.
| BAV is the most frequent congenital heart defect, occurring in about 1–2% of the population with 70% of cases occurring in males. BAV increases risk for thoracic aortic aneurysm (TAA) and early death. Approximately 30% of individuals with Turner syndrome have BAV/TAA, making this an important population for the study of this disease. Given that individuals with Turner syndrome are missing a complete or partial second sex chromosome, it is presumed that X chromosome genes are involved in causing the defect. This is consistent with the bias towards occurrence in euploid males. However, not everyone with Turner syndrome has a BAV, so we hypothesized that autosomal genes may also play a role. Using whole exome sequencing we have shown that deleterious variation in TIMP3 is associated with BAV and indices of TAA. We further found that there is a synergistic interaction between loss of the X chromosome gene, TIMP1, and deleterious variation in TIMP3 that significantly increases that risk. TIMP1 and TIMP3 play roles in aortic valve morphogenesis and in stabilizing the aortic wall, loss of which leads to TAA. Hence our findings have implications for understanding the cause of BAV/TAA in all populations and as a potential therapeutic target.
| Turner syndrome is the most common sex chromosome aneuploidy, where ~50% have a complete monosomy X and ~48% have either a partial loss, rearrangement, or mosaicism of a second X chromosome.[1] The remaining ~2% have a partial or mosaic Y chromosome. Although Turner syndrome can be compatible with life, less than 1% of Turner syndrome fetuses survive.[2] The majority of prenatal deaths are due to cardiovascular defects.[3] Live born females with Turner syndrome share a constellation of phenotypes including primary ovarian insufficiency, short stature, lymphedema, webbed neck, skeletal deformities, neurocognitive disability, and a high incidence of congenital cardiovascular malformations. In particular, they are at a greatly increased risk for having left heart obstructions including hypoplastic left heart syndrome, BAV, coarctation of the aorta, and TAA.[4] Heart defects are the major cause of premature death. The degree to which a second sex chromosome is retained is the primary determinant of the morbidity and mortality in Turner syndrome, an observation that strongly implicates X chromosomal genetics in the pathology of acquired and congenital cardiovascular disease.[5, 6] In depth studies have shown that BAV, coarctation of the aorta, and risk for aneurysm are linked to the short arm of the X chromosome (Xp).[7, 8]
BAV is a congenital malformation where the aortic valve is comprised of two leaflets as opposed to the normal three leaflet configuration. BAV is associated with lifelong heart disease including valve calcification, stenosis, aortic endocarditis, and thoracic aortic dilation (TAD) that has a high risk of progression to aneurysm, dissection and rupture, and premature death. It is the most common congenital heart malformation occurring in about 2% of the general population where it is predominantly found in males, which comprise about 70% of all BAV cases.[9] However, despite the prevalence in the population, little is known about the etiology of BAV. There is clearly a genetic component as 10–40% of BAV is familial.[10] BAV and aortic aneurysm are thought to have a common genetic etiology.[11] Mutations in NOTCH1[12], GATA5[13], and NKX2.5[14] have been identified as the causative factor in some families with inherited BAV, but the majority of cases remain unexplained. The sex bias in euploid BAV indicates that having two X chromosomes may be protective. In Turner syndrome the incidence of BAV is increased by at least 50-fold over that seen in the euploid population.[15] This suggests that the lack of a second X chromosome predisposes both males and Turner syndrome females to have BAV and TAA, a condition known as BAV aortopathy.
Although there is a paucity of information about the etiology for BAV, a great deal is known about the pathogenic events underlying TAA and dissections associated with BAV. Numerous studies have shown significantly increased expression of matrix metalloproteinases (MMPs) and decreased expression of TIMPs in aneurysmal tissue.[16] This is significant because the role of MMPs is to degrade extracellular matrix (ECM); an activity that is inhibited by TIMPs. It is thought that in aneurysms the ECM in the aortic wall becomes degraded by MMPs, which weakens the aorta allowing it to succumb to hemodynamic stress thereby enlarging the diameter and thinning the aortic wall. In particular, increased expression of MMP2 and MMP9, which degrade the collagen and elastin components of the aortic wall, and a decrease in TIMP1, which inhibits MMP2 and MMP9 activity, have been implicated in the pathogenesis of aortic aneurysms.[16] In addition, an increased MMP9/TIMP1 ratio has been shown to be elevated in chronic aortic dissection, demonstrating a persistent role for ECM degradation.[17]
Deficiency of the second sex chromosome contributes to aortopathy in Turner syndrome, but its loss is not sufficient to cause disease since ~50% of women with Turner syndrome have a normal aortic valve and aortic dimensions. We hypothesized that autosomal genetic variation sensitized by sex chromosome deficiency causes aortopathy in Turner syndrome. To address this hypothesis we used whole exome sequencing to identify autosomal genetic variation associated with BAV and TAD in Turner syndrome. We used TAD as an indicator of aneurysm formation. This study of a discovery cohort of 188 and a replication cohort of 53 individuals with Turner syndrome identified an exome-wide significant association between TIMP3 (MIM: 188826) and BAV/TAD. Furthermore, investigation of the TIMP3 paralog, TIMP1 (MIM: 305370), revealed that having more than one copy of the Xp chromosome gene TIMP1 was protective against BAV/TAD. Combinatorial analysis shows a synergistic effect between having a single copy of TIMP1 plus the TIMP3 risk allele and the occurrence of BAV/TAD. Knowledge of a direct link between TIMP family-gene expression and aortopathy points the way to the development of novel biomarkers for disease progression and therapies to combat catastrophic aortic dissection and rupture in Turner syndrome.
The presence of BAV was associated with a higher aortic root (AR) z-score (mean AR z-score in BAV 1.29 ±1.59, versus no BAV 0.31 ± 1.08, p = 0.0002, mean difference = 0.98; Fig 1A). BAV was also associated with a significantly higher ascending aorta (AAO) z-score (mean AAO z-score in BAV 2.04 ± 1.99, versus no BAV 0.61 ± 1.18, p<0.0001, mean difference = 1.44; Fig 1B).
SKAT-O analysis revealed that variants in TIMP3 on chromosome 22 achieved exome-wide significance for association with BAV and TAD. TIMP3 was associated with the occurrence of BAV when it was used as the sole dichotomous phenotype (p = 1.58x10-6; Fig 2A), with the significance level increasing by an order of magnitude when BAV and AR z-scores were evaluated as covariates (p = 2.27x10-7; Fig 2B). This demonstrates a TIMP3-driven association between BAV and aortic enlargement in Turner syndrome. The quantile-quantile plots showed that there was no departure from observed vs. expected p-values (S1 Fig). Targeted exome sequencing of TIMP3 in a replication cohort also showed a significant association of TIMP3 variants with BAV and AR z-scores using SKAT-O (p = 0.038; Table 1).
There were a total of four variants identified in TIMP3 in the discovery cohort (Table 2). Of the four variants, rs11547635 was determined to be the SNP predominantly driving the association based on the increased allele frequency in cases compared to controls (p = 0.001, chi-squared) and evidence that the variant is deleterious based on the CADD score of 16.67. This is above the recommended deleterious significance cutoff of 15, which indicates that is in the top 5% of all damaging variants in the human genome. On the gene level, TIMP3 has a GDI PHRED score of 0.449, placing in the top 10% of genes intolerant of mutations. The lead driving SNP encodes a synonymous C>T transition at p.Ser87 in exon 3. Another SNP, rs9862, which is a synonymous variant at p.His83 is always present along with the p.Ser87 variant in the BAV cases in this study. Importantly, these variants, which have been studied in various types of cancer are associated with reduced TIMP3 plasma levels.[18–20] In combination the two variants disrupt two core ETS1 binding consensus sequences and prevent ETS1 binding, which is thought to be the basis of the reduction in expression.[20] Our discovery that known deleterious variants in TIMP3 are significantly associated with BAV and TAD of the aortic root in Turner syndrome fits well with the known role for TIMPs in protection against aortopathy. Nearly 25% of our Turner syndrome cohort carry these SNPs, making them a significant risk genotype. The two additional TIMP3 variants, rs149161075 and rs369072080, are rare and occur only in cases in this study.
Analysis of all of the genes on Xp identified TIMP1 as the top gene meeting our aortopathy criteria, which includes the potential for escape from X-inactivation, no Y chromosome or autosome homologues, and expression in the aorta. The list of all of the genes that met the criteria is shown in Table 3, ranked according to the likelihood that they could contribute to aortopathy. The list of all Xp genes and their characteristics can be found in S1 Table. TIMP1 polymorphically escapes X inactivation,[21] has partial functional redundancy with TIMP3[22], and is highly expressed in the aorta with nearly 10-fold higher expression than any of the other genes (GTExPortal). In addition, it is the only Xp gene that meets these criteria and has a known role in aortic valve development.[23] TIMP1 is also the only gene on Xp with a known association with aortic aneurysms in both humans and mouse models. Timp1 mouse models are susceptible to the development of aortic aneurysms[24, 25] and TIMP1 is known to be reduced in TAA in humans.[16, 26] Additionally, overexpression of Timp1 prevents aneurysm degradation and rupture in a rat model.[27] We therefore hypothesized that reduced copy number of TIMP1 in Turner syndrome increases the risk for BAV/TAD. Using BAV as the only variable the analysis revealed that subjects with only one copy of TIMP1 have a 4.50 increased odds of having a BAV than those who have greater than one copy (p = 0.0009, 95% CI = 1.9–11.8, Fig 3A). When BAV with TAD was studied as the outcome, having only one copy of TIMP1 increased these odds substantially (OR = 9.76, p = 0.029, CI = 1.91–178.80, Fig 3B).
To determine the specificity of the association between TIMP3 rs11547635 and TIMP1 copy number for having BAV/TAD or other phenotypic features, we compared cases with or without rs11547635. Height, weight, blood pressure, body surface area, the presence of webbed neck, broad chest, primary ovarian insufficiency, hypertension, or lymphedema occurred with equal frequency in subjects with or without the rs11547635 SNP (Table 4). On the other hand BAV, BAV with TAD, coarctation of the aorta, and any aortic disease occurred with significantly higher frequency in the group with rs11547635, indicating that it is specifically associated with aortopathy. TIMP1 copy number associations were similar but also included systolic blood pressure, lymphedema and webbed neck (Table 5).
We investigated the combinatorial effect of TIMP1 and TIMP3 variation on the outcome of BAV alone, and BAV with TAD. This analysis shows that the combination of having only one copy of TIMP1 and being a carrier of TIMP3 rs11547635 specifically increases the odds for having a BAV by nearly twenty-fold (OR = 18.00, 95% CI = 5.19–74.89, p<0.001) and also for having a BAV with TAD (OR = 12.86, 95% CI = 2.57–99.39, p = 0.004) compared to the group with no rs11547635 and >1 TIMP1 (Table 6).
Turner syndrome, like all genetic syndromes, is characterized by a primary inherent defect that sensitizes downstream modifier genes to breach a pathologic threshold. Thus, a single triggering event is capable of unleashing a myriad of phenotypic variations. Consistent with this disease model, we found that in Turner syndrome hemizygosity of TIMP1 due to lack of a complete second X chromosome is associated with genetic variation of its paralogue, TIMP3 on chromosome 22, synergistically heightening the risk for BAV and TAD, which is the first sign of aneurysm formation.
Given the detailed understanding of the fundamental role of MMPs in thoracic aortic disease, the results of this study have clear biological relevance. In the euploid population there is a significant reduction in TIMP1 and TIMP3 expression in BAV-associated TAA and a highly significant increase in MMP2 and MMP9, which are both regulated by TIMP1 and TIMP3.[16] This results in a considerable MMP/TIMP imbalance in aneurysms compared to control aortas.
We propose that hemizygosity for TIMP1 is the X chromosome basis for increased susceptibility for BAV and aortopathy in Turner syndrome. This coupled with a SNP-driven decrease in TIMP3 expression synergistically increases risk for both BAV and BAV with TAD. This is consistent with our hypothesis that a gene or genes on Xp interact with autosomal variants that are benign unless expressed on a genetically sensitized background such as that in Turner syndrome. The inherent decrease in TIMP1 in Turner syndrome subjects missing a complete second copy of the X chromosome sensitizes those individuals to decreased TIMP3 expression. In addition, a global methylation profile for Turner syndrome found that the Turner syndrome X chromosome has a unique methylation pattern when compared to the X chromosome of euploid males.[11] Notably, TIMP1 tends to be hypermethylated in Turner syndrome,[28] which suggests that the expression level may be decreased even beyond the reduction in copy number.
Importantly, TIMP1 and TIMP3 have functional redundancy in the aorta. Both exercise inhibitory control over MMP2 and MMP9, which are the two MMPs associated with degradation of the aortic wall. We propose that decreased TIMP1 expression due to a reduction in copy number sensitizes the aorta to MMP-induced damage, but protection is conferred by the expression of TIMP3. Decreased expression of both negates that protection making the aortic wall vulnerable to degradation which can lead to TAD and aneurysm. In addition, TIMPs 1 and 3 are expressed in the aortic valve, where they play a role in valve remodeling,[23] which is a critical activity in the development of the tricuspid aortic valve. This fundamental link between BAV pathogenesis and downstream TAD provides a previously unrecognized mechanism for the heightened risk for aortopathy in Turner syndrome. In a study of 18 women with TS and aortic dissection, 6 cases were available for biochemical analysis, and that study showed a skewed ratio of collagen I to collagen III (normally 30:70%) with 60% collagen I and only 30% collagen III,[29] which could well be the end result of an altered MMP/TIMP activity.
As with all studies of this nature there are some limitations and caveats. The exome sequencing was done on DNA isolated from peripheral blood, so the molecular karyotypes reflect the chromosome composition in that tissue. It is possible that the karyotype in other tissues such as the developing heart may differ, particularly with respect to mosaicism. In addition, our analyses did not include potential effects of the autosomal rearrangements found in some of the study subjects. These were genetically heterogeneous and often in single individuals, so it is unlikely that they would significantly affect the results of this study. Another limitation is that this study did not assess any potential influence of maternal genetic effects, nor did we assess the parent-of-origin of the retained X chromosome.
There is no clear explanation for the strikingly higher prevalence of aortopathy in euploid men compared to women. And, the larger questions regarding the role of the sex chromosome genes in the differential susceptibility to common diseases has received little attention. Bellott and colleagues proposed that dosage differences between X chromosome genes and homologous ancestral genes retained on the Y chromosome may account for phenotypic differences between men and women.[30] Our data supports another model where expressed genes that escape inactivation on the second X chromosome and that are also absent from the Y chromosome (like TIMP1) play a role in the frequently observed sex bias in disease.
In conclusion, we propose that aortopathy in Turner syndrome results from an inherent dysregulation of the TIMP/MMP ratio. This imbalance increases risk for both congenital cardiovascular defects and later onset aortic disease. Beyond Turner syndrome, the lack of a second copy of TIMP1 in euploid males may also explain the increased risk for BAV/TAD compared to euploid females. The findings of this study represent a significant advance in the understanding of the mechanisms underlying aortopathy in Turner syndrome.
The Turner syndrome cohort was accessed from the National Registry of Genetically-Triggered Thoracic Aortic Aneurysms and Related Conditions (GenTAC).[31] GenTAC study subject recruitment was approved by the institutional review board for each member of the GenTAC investigative team, and informed consent for participation in associated research studies was obtained for each study subject. The project was approved by the Oregon Health & Science University institutional review board. The Danish cohort was approved by the Central Denmark Region Ethical Scientific Committee (#2012-500-12) and registered at ClinicalTrials.gov (#NCT01678274).
GenTAC spent a decade recruiting study subjects with conditions related to thoracic aortic aneurysms, including collection of biospecimens, rigorous evaluation and documentation of clinical data, and collection of follow-up data for longitudinal studies. The majority of subjects enrolled in GenTAC had aorta imaging studies that provide information on aortic dimensions and evaluation of aortic valve status. All images, such as echocardiograms, CT and MRT studies were collected clinically, but transferred to the GenTAC imaging core (ICORE) for re-evaluation by a single cardiac imaging expert for consistency of measurements and interpretation.[32] The discovery cohort for this study was composed of Turner syndrome study subjects of Northern European (non-Finnish) descent. Inclusion criteria included a diagnosis of Turner syndrome, self-reported race as white, ethnicity as non-Hispanic, evaluation for a diagnosis of BAV, and availability of aortic dimension measurements and body morphometrics. The diagnosis of BAV was based on clinical images and interpretations. An additional 53 study subjects from a prospective study in Denmark were used as an independent replication cohort.[33, 34]
For the purposes of this study we defined aortopathy (cases) as those having a BAV with or without TAD. In keeping with clinical norms for Turner syndrome, a thoracic aortic dimension z-score ≥ 1.9 was used as the definition of TAD as an indicator of aneurysm formation. All study subjects were confirmed for a diagnosis of Turner syndrome based on either clinical karyotype or exome sequence-based karyotyping. Subjects were phenotyped for presence of a BAV or a normal aortic valve. Our final Turner syndrome discovery cohort was composed of 88 cases (Turner syndrome with BAV) and 100 controls (Turner syndrome with no BAV). Within this cohort, 113 subjects had aortic root (AR) dimensions and 106 subjects had ascending aorta (AAO) dimensions. For the replication cohort 14 had a BAV and 39 had a normal aortic valve. For all subjects AAO and AR diameters were converted into z-scores using methodology that was specifically developed for children and adults with Turner syndrome to correct for the altered longitudinal growth in Turner syndrome.[35] Briefly, the regression equations and coefficients were used to calculate expected aortic dimensions based on body surface area (BSA, Haycock formula) for each individual in the study with a measurement (Eqs 1&2). The z-scores were calculated by comparing expected aortic dimensions to actual aortic dimensions and incorporating the mean squared error (MSE; Eq 3).[35] Expected aortic dimension data points and lines were generated for each z-score.
Equations:
AorticRootequation:(expected)2=(1.035+(0.589*BSA)+(−0.129*BSA2))2
(1)
AscendingAortaequation:(expected)2=(0.942+(0.593*BSA)+(−0.122*BSA2))2
(2)
Z-scoreequation:=(√(actualdimension(cm)−√(expecteddimension(cm))/(√(MSE))
(3)
The BSA (m2) vs. AR or AAO (cm) for BAV cases (triangle) and BAV controls (square) were plotted. Overlaid on the same plot are the polynomial trend lines for z = 0, z = 1, z = -1, z = 2, z = -2, z = 3, z = -3 (S2 Fig).
In total, 215 genomic DNA samples isolated from peripheral blood were submitted for exome sequencing and the exome capture kit Roche Nimblegen SeqCap EZ was used to prepare the sequencing libraries. Whole exome sequencing (WES) was performed by the NHLBI Resequencing & Genotyping Service at the University of Washington (D. Nickerson, US Federal Government contract number HHSN268201100037C). In summary, 16 samples failed post-sequencing QC and 199 samples passed post-sequencing QC. The average read depth for the targeted exome was 71X, with 86% of the target regions covered at greater than 20X. Reads were mapped to the hg19 UCSC genome build using the Burrows-Wheeler aligner, version 0.7.10. Variants were called using the GATK best practices pipeline, where in the 199 samples, 195,034 variants were called. BAM files and VCF files were transferred to the Maslen lab for evaluation.
Data cleaning and filtering was performed using PLINK v1.90b3g[36], which 1) removed any variants with less than 99% genotyping rate, where 6,815 variants were removed; 2) removed individuals with more than 5% missing genotypes, where no individuals were removed; 3) excluded markers that fail the Hardy-Weinberg equilibrium test using a threshold of 1.0x10-6, where 2,334 variants were removed. A principal components analysis (PCA) was performed using the R package SNPRelate to calculate the eigenvectors (EVs) for each subject.[37] Data were prepared for PCA analysis by taking common SNPs (MAF >5%) and pruning out SNPs in linkage disequilibrium with an r2 > 0.2, stepping along five SNPs at a time within 50kb windows. We plotted EV1 vs EV2 to look for population outliers (S3A Fig). Population outliers were removed and the analysis was repeated a total of four times until no more outliers remained (S3B Fig). In total, 11 subjects were detected at EV1 < -0.3 and EV2 > 0.3 and were removed from the dataset. Additionally, we use the first three eigenvectors as covariates in most downstream analysis. The final dataset contained 185,885 variants across 188 subjects, providing a total genotyping rate of 0.998084.
To enhance the probability of identifying an exome-wide significant signal a gene-based burden test, the optimal sequence kernel association test (SKAT-O), was used to evaluate the data.[38] This analysis clusters variants into genes for a gene by phenotype analysis, which improves signal strength for exome data from smaller cohorts as it reduces the multiple testing burden. This state-of-the-art approach is particularly useful for studies of rare disorders such as Turner syndrome. The 185,885 variants which passed QC from the WES pipeline were assigned to their respective genes using hg19_refGene. Variants were allowed to be in more than one gene since the test compares gene burden in the same gene, not between different genes. All analyses included the first three principal component eigenvalues as covariates to adjust for any underlying population structure. First, SKAT-O was used to test for an association with the dichotomous BAV status. Second, SKAT-O was used to test for an association with BAV and aortic diameter z-scores as a proxy for TAD evaluated as a continuous variable. For each analysis, a quantile-quantile (Q-Q) plot was generated to look for departure of the observed p-values from the expected p-values.
Combined Annotation Dependent Depletion (CADD) scores were used as a tool for scoring the deleteriousness of the genetic variants identified in exome sequencing data. PHRED-scaled CADD scores integrate multiple annotations into a single metric that outperforms other commonly used algorithms of this type. A CADD score ≥20 indicates that a variant is among the top 1% most deleterious variants in the human genome. We used the recommended cutoff score of ≥15 as our threshold for considering a variant to be likely deleterious. The allele frequency of each variant was queried in the Exome Aggregation Consortium (ExAC) database of exome data from over 60,000 unrelated individuals, from which we used the European non-Finnish population.[39] All variants with alleles that were overrepresented in cases were validated by Sanger sequencing. For the replication cohort, we performed targeted Sanger sequencing of all TIMP3 exons and followed the same SKAT-O association test as described above.
X and Y chromosome information from the WES data was used to assess the presence of any second sex chromosome. X and Y SNP plots were generated for each study subject and compared to control reference plots to define the second sex chromosome status for each individual.[40] Alternate allele frequencies from the exome variant calls were used to create SNP plots. Briefly, the alternate allele frequencies were calculated for all variants on the X chromosome and sorted by position for each subject. In R, scatter plots were generated and evaluated for the presence of a second X chromosome. This was repeated for the Y chromosome and the Integrative Genome Viewer (IGV) was used to confirm the presence of Y chromosome reads.[40] Reference plots of a control female with 46,XX karyotype, a control female with 45,X karyotype, and a control male with 46,XY karyotype were generated (S4A Fig). We then generated X and Y chromosome plots for each subject in this study.
In these plots, the X-axis is sorted by position on the X chromosome and the Y-axis is the alternate (ALT) allele frequency. As expected for a 46,XX karyotype, some SNPs are homozygous for the ALT allele (1.0), homozygous for the reference (REF) allele (0.0), or heterozygous for the ALT/REF allele (0.5). In contrast, a 45,X karyotype only has SNPs that are homozygous for the ALT allele, or homozygous for the REF allele because only one copy is present. The 46,XY karyotype looks similar to the 45,X plot, but has SNPs heterozygous for the ALT/REF allele clustered in the captured pseudoautosomal (PAR) region. The presence of any Y chromosome material was confirmed using IGV. Available clinical karyotypes were compared to molecular karyotypes generated from the SNP data and basic second sex chromosome status groups were created to categorize the study subjects. While the majority of subjects were true monosomy 45,X, examples of other karyotypes included Xp deletions, Xq deletions, Xq isochromosomes, and X chromosome rings for their second X chromosome (S4B Fig); mosaicism for the second X chromosome, either 45,X/46,XX or 45,X/47,XXX, or mosaicism for Y chromosome material (S4C Fig), although we were unable to quantify the Y mosaicism level based on the plots.
While most plots were straight forward in their interpretation, some were more complicated. In those cases, the clinical karyotype was relied upon. To assess the mosaicism observed in a large number of subjects, a model was created to predict the percent 45,X mosaicism based on alternate allele frequencies (S5 Fig). The equations from each model were used, where y is the percent 45,X mosaicism and x is the alternate allele frequency. The average of the upper and lower predicted values was used as the final estimate of 45,X mosaicism. The molecular karyotypes and estimated TIMP1 copy number based on the percentage of cells with a second X chromosome are shown in Table 7.
Genes on Xp were evaluated to identify candidates likely to contribute to aortopathy. We hypothesized that an aortopathy gene would be found on Xp, would escape X inactivation in euploid females,[41–43] would be expressed in the aortic wall, and would not be a pseudogene, or have a Y homologue.
To calculate the magnitude of the association between BAV status and aortic z-score, a linear regression model was fit where BAV was the predictor and aortic z-score was the response variable. This was performed separately for both AR z-score and AAO z-score. The mean differences and 95% confidence intervals were generated to accompany p-values. Boxplots for each AR and AAO z-scores were plotted against BAV status.
To investigate if the TIMP3 paralog TIMP1 was associated with BAV, a general logistic regression model was performed where TIMP1 copy number was the categorical predictor, 1 copy and >1 copy of TIMP1 were the variables, and BAV status or BAV with TAD was the response variable with no BAV serving as the reference. Odds ratios and 95% confidence intervals were generated to accompany p-values.
To investigate the combination of the TIMP3 variant rs11547635 and TIMP1 as risk factors for the presence of a BAV or BAV with TAD, four groups were formed: 1) no TIMP3 rs11547635 and >1 copy of TIMP1, 2) with TIMP3 rs11547635 and >1 copy of TIMP1, 3) no TIMP3 rs11547635 and only 1 copy of TIMP1, and 4) with TIMP3 rs11547635 and only 1 copy of TIMP1. Separate general logistic regression models were created to compare these four groups in order to determine their associations with BAV, or the combination of BAV and TAD. Odds ratios and 95% confidence intervals were generated to accompany p-values.
Other physical attributes of Turner syndrome were studied to determine if any were also associated with the TIMP3 rs11547635 risk allele. Continuous variables (height, weight, body surface area, systolic blood pressure, and diastolic blood pressure) were analyzed using a Student’s two-sample t-test, where the means of those with or without TIMP3 rs11547635 were compared. Categorical variables (lymphedema, broad chest, webbed neck, primary ovarian insufficiency, hypertension, coarctation of the aorta, bicuspid aortic valve, and any aortic risk factor) were analyzed using a Chi-squared test with Yate’s correction or Fisher’s exact test as appropriate. The same analysis was done using TIMP1 copy number as the variable.
|
10.1371/journal.pntd.0000056 | A Dominant Clone of Leptospira interrogans Associated with an Outbreak of Human Leptospirosis in Thailand | A sustained outbreak of leptospirosis occurred in northeast Thailand between 1999 and 2003, the basis for which was unknown.
A prospective study was conducted between 2000 and 2005 to identify patients with leptospirosis presenting to Udon Thani Hospital in northeast Thailand, and to isolate the causative organisms from blood. A multilocus sequence typing scheme was developed to genotype these pathogenic Leptospira. Additional typing was performed for Leptospira isolated from human cases in other Thai provinces over the same period, and from rodents captured in the northeast during 2004. Sequence types (STs) were compared with those of Leptospira drawn from a reference collection. Twelve STs were identified among 101 isolates from patients in Udon Thani. One of these (ST34) accounted for 77 (76%) of isolates. ST34 was Leptospira interrogans, serovar Autumnalis. 86% of human Leptospira isolates from Udon Thani corresponded to ST34 in 2000/2001, but this figure fell to 56% by 2005 as the outbreak waned (p = 0.01). ST34 represented 17/24 (71%) of human isolates from other Thai provinces, and 7/8 (88%) rodent isolates. By contrast, 59 STs were found among 76 reference strains, indicating a much more diverse population genetic structure; ST34 was not identified in this collection.
Development of an MLST scheme for Leptospira interrogans revealed that a single ecologically successful pathogenic clone of L. interrogans predominated in the rodent population, and was associated with a sustained outbreak of human leptospirosis in Thailand.
| A sustained outbreak of human leptospirosis occurred in northeast Thailand between 1999 and 2003, the basis for which was unknown. Leptospirosis is a potentially serious infection cause by bacteria known as Leptospira; infection usually occurs following environmental exposure to pathogenic Leptospira shed in the urine of an infected animal. The purpose of this study was to obtain bacterial isolates from humans with leptospirosis around the time of the Thai outbreak for genotyping, and to relate these to the maintenance host animal. To achieve this, a bacterial typing scheme (multilocus sequence typing, MLST) was developed for L. interrogans, the major cause of human disease. This approach has the advantage over existing typing schemes in that the data generated are amenable to detailed evolutionary analysis, and are readily comparable via the internet. Our results demonstrated the emergence of a dominant clone of L. interrogans serovar Autumnalis; this was the major cause of human disease during the outbreak, and was found in a maintenance host which was defined as the bandicoot rat.
| Leptospirosis is a zoonotic infection caused by pathogenic members of the genus Leptospira. Human disease is usually acquired following environmental exposure to Leptospira shed in the urine of an infected animal [1],[2]. Infection is acquired during occupational or recreational exposure to contaminated soil and water, organisms gaining entry to the accidental human host via abrasions or less commonly the conjunctiva [1]. Disease may also be acquired through direct contact with infected animals, and occurs in farmers, veterinarians and abattoir workers [1]. The disease has a worldwide distribution but is most common in tropical regions where incidence peaks during the rainy season [1],[2]. Clinical manifestations are broad ranging and follow a biphasic pattern in which a septicemic phase lasting around one week is followed by an immune phase during which antibodies are raised and organisms localize in tissues and appear in urine. Much disease is sub-clinical or mild, but patients reaching medical attention usually have an acute febrile illness associated with one or more of chills, headache, myalgia, conjunctival suffusion, and abdominal symptoms which can include nausea, vomiting and diarrhea [1]. Leptospirosis has been described as anicteric or icteric; the former represents 85–90% of cases and is associated with a good prognosis, while the latter may be associated with multisystem disease involving particularly the kidneys, lung and heart, with a reported mortality rate of 5–15% [1].
Leptospirosis is an emerging infectious disease in Thailand [3],[4]. Before 1996, the number of cases reported to the Department of Disease Control (DDC) was approximately 200 per year. Leptospirosis was sporadic and reported mainly from central and southern regions. A marked change occurred in the subsequent decade, with a year-on-year rise from 398 cases in 1996 to a peak of 14,285 cases in 2000. This was followed by a continuous decline with 2,868 cases reported during 2005 [5]. Reporting in Thailand is voluntary and probably represents a small proportion of true cases. There was also a shift in the geographical distribution, with the majority of cases being reported in the northeast. One explanation for the outbreak is that it was related to the emergence of a biologically successful clone of Leptospira. This possibility is supported by a study of 44 leptospiral strains obtained from humans during three outbreaks in Brazilian urban centers, in which typing using arbitrarily primed PCR demonstrated that 43 isolates exhibited very similar fingerprints suggestive of a clonal population of L. interrogans [6]. In addition, during a large urban outbreak in Brazil, L. interrogans serovar Copenhageni was isolated from 87% of cases with positive blood cultures [7]. Although it is currently unclear to what extent genetic relatedness can be informed by serotype alone, this observation is consistent with the majority of cases being caused by the expansion of a single outbreak clone.
The aim of this study was to define the molecular epidemiology of Leptospira strains isolated from humans during the Thai outbreak, and to relate this to the maintenance animal host. To achieve this, an MLST scheme was developed for L. interrogans, the major cause of human disease. This approach has the advantage over existing typing schemes in that the data generated are amenable to detailed evolutionary analysis. MLST data are also readily comparable via the internet, and establishment of an MLST scheme therefore paves the way for future studies. Our results confirm the emergence of a dominant clone of L. interrogans serovar Autumnalis; this was the major cause of human disease, and was found in a maintenance host which was defined as the bandicoot rat.
A prospective study was undertaken at Udon Thani General Hospital in northeast Thailand to identify patients with leptospiremia. This 1,000 bed provincial hospital serves a predominantly rural population, >80% of whom are rice farmers and other agricultural workers who are repeatedly exposed to rats and water contaminated by rat urine. Patients were recruited during consecutive months from October 2000 to December 2002, then for four months during each rainy season (July to October inclusive) during 2003 and 2005. The reason for this pattern of recruitment is that leptospirosis is predominantly a rainy season disease. Consecutive adult patients (≥15 years) presenting with fever (>37.8°C) of unknown cause were recruited following informed and written consent. Patients with a blood smear positive for malaria parasites or other definable infections such as pneumonia or urinary tract infection were excluded. The clinical features of leptospirosis are broad ranging and similar to other acute febrile illnesses common to this geographic area such as scrub typhus and dengue fever. In view of this, all adult patients presenting with acute undifferentiated fever were cultured to detect leptospiremia. A 10 ml blood sample was drawn on the day of admission into a sterile tube containing 250 units of sodium heparin for Leptospira culture. The study protocol was approved by the Ethical Committee of the Ministry of Public Health, Royal Government of Thailand.
A further 24 unselected isolates cultured from the blood of patients with leptospirosis presenting to hospitals in 8 additional provinces in Thailand during the rainy seasons of 2003 and 2004 were obtained from strain collections. These provinces were: Lumpang (situated in the north), Yasothon, Nakhon Ratchasima, Maha Sarakahm and Loei (northeast), Ratchaburi (central), Rayong (east) and Chumphon (south).
Seventy six reference strains representative of the species L. interrogans, L. kirschneri and L. borgpetersenii were obtained from the WHO/FAO/OIE Collaborating Center for Reference & Research on Leptospirosis, Australia, or National Institute of Health, Thailand.
A total of 1,126 rodents were trapped in Nakhon Ratchasima, northeast Thailand during 2004 by the National Institute of Health, Thailand. Animals were identified and cultured for Leptospira, as described previously [8]. Ten animals were culture positive (Bandicota indica 8, Bandicota savilei 1, and Rattus rattus 1), while all samples from Rattus exulans, Rattus losea, Mus cervicolor, Mus caroli and Sancus murinus were culture negative for Leptospira. Eight unselected isolates remained viable and were evaluated in this study; of these, 6 were isolated from B. indica (greater bandicoot rat) and 1 each was isolated from B. savilei (lesser bandicoot rat) and Rattus rattus (black rat).
Culture of leptospires from human blood was performed using EMJH supplemented with 3% rabbit serum and 0.1% agarose, as described previously [9]. Positive cultures were sent to the WHO/FAO/OIE Collaborating Center for Reference & Research on Leptospirosis, Australia for serovar identification using the cross agglutinin absorption test (CAAT) [10]. Definitive identification of species was undertaken by amplification and sequencing of the near full-length 16S rRNA gene. Primers were designed to anneal to conserved regions of genes from pathogenic species L. interrogans, L. kirschneri, L. borgpetersenii, L. santarosai, L. alexanderi and L. fainei. The primers (f - 5′ GTTTGATCCTGGCTCAG 3′ and r -5′CCGCACCTTCCGATAC 3′) amplified a 1,483 bp PCR product which was sequenced in its entirety using internal primer pairs (primers available on request).
Genomic DNA was extracted using the High Pure PCR Template Preparation Kit (Roche Applied Science, Germany). In a pilot study, 14 housekeeping loci were selected using the whole genome sequence of L. interrogans serovar Lai strain 56601 (11 loci situated on chromosome I and 3 loci on chromosome II; loci and primer sequences available on request). Primers were designed using PrimerSelect software (DNASTAR Inc., Wisconsin,USA), and synthesized by Sigma-Proligo (Proligo Singapore Pty Ltd). These were evaluated using 30 clinical or reference strains belonging to species L. interrogans, L. kirschneri or L. borgpetersenii, using standard MLST methodology [11] (data not shown). Each of the 14 gene fragments were amplified by PCR, purified and sequenced using a MegaBACE 500 sequencer and DYEnamic ET Dye Terminator Cycle Sequencing Kit (Amersham Biosciences, England). Seven loci were then selected based on performance of primers, number of alleles at a given locus and distribution of strain numbers between the alleles. These loci were pntA, sucA, fadD, tpiA, pfkB, mreA, & glmU, which are located on chromosome I with the exception of fadD. Primer sequences are shown in Table 1. Amplifications were performed in 25-µl total volumes of PCR reaction mix contained 1–10 ng of genomic DNA, 5 pmol of each primer, 200 µM dNTP, (eppendorf, Germany), 1.5 mM of MgCl2, 1.25 unit of Taq DNA polymerase (Promega, USA) and 1× buffer. A PTC-200 Peltier Thermal Cycler (MJ research, USA) was used to perform PCR with an initial denature step at 94°C for 5 minutes, followed by 30 cycles of 94°C for 10 seconds, 52°C (mreA, pfkB, pntA, sucA, and tpiA), or 50°C (fadD and glmU) for 15 seconds, 72°C for 50 seconds, then 72°C for 7 minutes. PCR product size ranged from 555 bp to 638 bp; the sequence start and end points used to define each MLST locus are shown in Table 1. MLST was performed for the remaining isolates using these 7 loci. Following the standard MLST protocol, each allele was assigned a different allele number and the allelic profile (string of seven integers) was used to define the sequence type (ST). A leptospira mlst website was established to provide public access to these data, and to provide a resource to other investigators who can use this to assign the ST of further strains. This can be accessed at http://leptospira.mlst.net.
DNA sequences for the 16S rRNA gene have been deposited in the GenBank database with the accession numbers shown in Table S1.
The number of leptospirosis cases reported to the Department of Disease Control, Thailand between 1990 and 2005 is shown in Figure 1. An increase in cases of leptospirosis was also observed by clinicians working in northeast Thailand during 1999 (personal communication, Dr R. Limaiboon, Udon Thani Hospital). A prospective study was commenced at Udon Thani Hospital in mid-October 2000 to identify and culture suspected cases and isolate the causative Leptospira.
A total of 1,658 patients were recruited in Udon Thani, of whom 115 were culture positive for Leptospira. The number of cases of culture proven leptospirosis was greatest during 2001 (there were only 2 study months during 2000), followed by a decline to the end of the study in 2005 (Figure 1). There was a significant reduction over time in the proportion of patients presenting with fever who were leptospiraemic (chi-squared for trend = 15.3, p<0.0001). This case load pattern mirrors the number of cases reported to the Department of Disease Control. Our data provides additional confirmatory evidence for a true increase in leptospirosis in northeast Thailand during the putative outbreak.
The pathogenic strains of Leptospira obtained from patients presenting to Udon Thani hospital were characterized to determine whether the increased disease incidence was related to one or a number of different circulating bacterial clones. Of the 115 isolates obtained from patients in Udon Thani, 104 were available for molecular characterization. 16S rRNA sequencing was performed on at least one representative of each ST. One hundred isolates were L. interrogans, 3 were L. borgpetersenii and 1 was L. kirschneri (Table S1). Three strains (L. borpetersenii serovar Javanica) failed to amplify at five or six MLST loci but were identical to each other at glmU; these strains are not considered further. The 101 isolates from Udon Thani corresponded to 12 STs but a single ST predominated, with ST34 accounting for 77 (76%), all of which were identified as serovar Autumnalis. Of the remainder, 8 isolates belonged to ST46, 4 isolates were ST49, and the remaining nine sequence types consisted of one or two isolates (Figure 2a, Table S1). A single isolate defined as ST41 was also serovar Autumnalis but this strain is unrelated to ST34, showing divergence at all seven alleles. Thus two strains sharing the same serovar can be distantly related [1], possibly because serovars may arise independently in differently lineages by evolutionary convergence, or by horizontal gene transfer.
To explore the role of the dominant clone ST34 in the putative outbreak of leptospirosis, the proportion of Leptospira ST34 was determined for each year of the study in Udon Thani (Figure 1). This demonstrated that the dominance of ST34 declined over time, being replaced by a range of other sequence types (Table 1). The proportion of clinical Leptospira isolates that were ST34 fell from 85% in 2000/2001 to 64% in 2002/2003 and 56% in 2004/2005 (years combined because of small numbers in 2004 and 2005 - χ2 for trend = 6.61, p = 0.01).
To define the extent to which ST34 was distributed across Thailand, a further 24 unselected isolates obtained in 2003 and 2004 from human cases of leptospirosis from across the country were evaluated. The total proportion of isolates corresponding to ST34 was 17/24 (71%) (two strains were non-typable L. borpetersenii). The geographic distribution was as follows: Lumpang, 1/2 isolates; Rayong, 1/1 isolate; Chumphon, 2/4 isolates; Loei, 9/9 isolates; Ratchaburi, 1/1 isolate; Yasothon, 1/1 isolate; Nakon Ratchasima, 2/5 isolates; and Maha Sarakham, 0/1 isolate. This is not significantly different from the proportion of ST34 in isolates from Udon Thani in the same years (Fisher's exact p = 0.37), and confirms that the outbreak clone ST34 was widely distributed throughout Thailand and formed the predominant virulent strain at the time of the outbreak. A further six STs were identified in this collection, four of which were not observed in the Udon Thani collection. These data provide further support for the picture of a single dominant clone (ST34) associated with an increased incidence of human disease, within a “background” population of higher genotypic diversity. One strain (ST22 obtained in Lumpang province) was serovar Autumnalis, but showed divergence at 5/7 alleles from ST34.
To determine whether a link could be identified between ST34 and a maintenance host, 8 isolates available from rodents captured in northeast Thailand were characterized. Seven strains (from B. indica (6) and B. savilei (1)), were L. interrogans ST34. This confirms the predominance of the outbreak strain in a maintenance host, which in this case appears to be the bandicoot rat. The remaining isolate from R. rattus was L. interrogans, ST49 which was also isolated from human cases in Udon Thani in 2001/2 (n = 4) and Nakhon Ratchasima in 2004 (n = 1). This does not exclude the possibility of additional maintenance hosts, but rodents trapped in agricultural areas reflect the species to which farmers are commonly exposed.
To place the Thai isolates within a global context, we selected a total of 76 reference strains representative of the species of the Leptospira strain population in Thailand but recovered from diverse geographical sources (L. interrogans 65, L. borgpetersenii 3, L. kirschneri 8) (Table S1). From our Thai sample of 123 clinical isolates, 16 STs were identified (0.13 ST per isolate; 5 strains of L. borgpetersenii being non-typable by MLST). In contrast, MLST revealed 59 STs for 73 reference strains (0.81 ST per strain), revealing that the reference strains are far more diverse, and that only a small fraction of the global diversity was recovered in the Thai sample. The reference L. borgpetersenii strains did not amplify at all seven loci, and so are again scored as non-typable. The largest clones within the reference collection were ST17 and ST37 (both with 4 isolates); one ST contained 3 isolates, 7 STs contained two isolates and the remaining 49 sequence types had one representative strain (Figure 2a, Table S1). ST34 was not represented in the reference collection. One strain was serovar Autumnalis (Akiyami A, ST27) but this was unrelated to ST34 and was much more similar to the non-ST34 Autumnalis strain isolated in Udon Thani (ST41). This analysis indicates that the strains causing human disease in Thailand are more clonally restricted than reference strains from variable hosts and geographical locations, and that the population genetic structure of L. interrogans is highly diverse when considering non-ST34 isolates. This further supports the argument that the predominance of ST34 during the Thai outbreak does not reflect a clonal population structure, and is consistent with a temporary selective advantage.
A phylogenetic analysis was performed to shed light on the emergence of ST34. All 204 typable strains (excluding 8 non-typable L. borgpetersenii isolates) were evaluated to identify the close relatives of ST34. Figure 2 shows two neighbour-joining trees based on the concatenated sequences of the seven MLST genes (3165-bp). Figure 2a was constructed using all 204 isolates. There was a clear distinction between the two species L. interrogans and L. kirschnerii which was also noted in loci individually (not shown). ST34 isolates accounted for almost half of the tree, illustrating the numerical dominance of this clone. As the branching order of this tree is unclear, Figure 2b shows a neighbour joining tree for just the STs highlighted in Figure 2a. Of the four Autumnalis STs, ST27 and ST41 appear closely related in both Figure 2a and 2b, but unrelated to the other Autumnalis STs ST22 and ST34. This latter pair appears to be closely related in Figure 2a, but Figure 2b reveals this is an artifact of the poorly resolved topology of this tree. The different clones sampled from Thailand in this study did not form a single cluster but were dispersed throughout the tree. This suggests that they have not all diverged from a single common Thai ancestor. The lack of evidence for strong geographical structure is consistent with high rates of migration via the rodent (or possibly human) host. Figure 2b identifies ST29 (reference strain Bangkinang 1) as a close relative of ST34; this was isolated from a human in Indonesia. Other close relatives of ST34 are also reference strains from Indonesia and Malaysia (not shown), although the significance of this is unclear as the tree is not robustly supported.
Human outbreaks of leptospirosis are well documented in the literature, as are clusters of cases linked by specific water-related activities or occupations [1],[2]. Outbreaks in Thailand and elsewhere are often linked to climatic events such as flooding and the concomitant increase in human exposure to environments contaminated by Leptospira. The precipitous increase in reported cases of leptospirosis in Thailand commencing in 1999, followed by the sustained incidence during the ensuing years, could not be explained by persistent climatic change or sequential episodes of regional flooding. Changes in reporting practice can lead to marked changes in the perceived disease incidence, although this does not explain the marked rise and fall in reported cases over time. An alternative explanation is that this was associated with the presence of a biologically successful clone of pathogenic Leptospira. In this study, we developed and applied robust typing methods to provide several lines of evidence in support of this hypothesis. This clone is likely to harbour an adaptive (competitive) advantage, albeit transiently. Possible explanations include a selective advantage for ST34 in the maintenance host (the bandicoot rat) leading to a higher bacterial load and higher shedding from urine, or a survival advantage once shed into environment, such as increased resistance to desiccation. Both possibilities are amenable to testing in the laboratory setting. Alternatively, ST34 may have a greater propensity to cause human disease compared with other circulating clones. Although difficult to test, the finding that ST34 co-existed in the environment with a large number of other STs but caused most disease would be supportive of this hypothesis. The virulence of ST34 as reflected by severity of human disease was not assessed in patients presenting to Udon Thani hospital, since the comparator group was small and caused by 11 other STs. The emergence of ST34 may have predated the outbreak, and this is difficult to refute since no strains were available from the period prior to the outbreak. However, the decline in frequency of ST34 as a cause of leptospirosis over time is consistent with the suggestion that there is a direct link between the clone and the outbreak.
Previous studies of human outbreaks have largely relied on serological methods to confirm clinical cases and to define indirectly the infecting isolate [1]. The standard serological method (microscopic agglutination test, MAT) provides a broad idea of serogroups responsible for leptospirosis in a given geographic area, but in one study the predominant serogroups at a titer of ≥100 correctly predicted less than 50% of serovars [13]. Arbitrarily primed PCR has been used successfully to study human outbreaks in Brazil [6], and to characterize 40 isolates recovered from humans between 1995 and 2001 on the Andaman and Nicobar Islands in India, 32 of which were a clone with a fingerprint matching that of L. interrogans sensu stricto [14]. Here, we use the more discriminatory and robust method of MLST to identify clusters of closely related isolates. The use of multiple gene loci is essential, as frequent recombination within the population would make inferences based on single gene loci unreliable [15]. This study clearly demonstrates the advantages of bacterial isolation in that it permits detailed typing studies to characterize local populations and outbreaks.
The MLST scheme presented here was developed primarily to characterize the isolates responsible for the outbreak of leptospirosis unfolding in Thailand in the early 2000s (i.e. L. interrogans and the closely related L. L. kirschneri), and is not designed for the characterization of the genus as a whole. Nevertheless, the scheme presented here demonstrates the utility of MLST for Leptospira for characterizing isolates from a clinical perspective. For more taxonomic or genus-wide evolutionary studies, or for disease caused by other Leptospira species, the primer sequences could be refined in order to broaden the phylogenetic range over which they amplify, or alternatively the loci used by Ahmed et al. may be employed [16].
In conclusion, our observations provide strong support for the hypothesis that the ST34 clone was associated with the 1998–2003 outbreak of leptospirosis in northeast Thailand. The existence of this strain collection now provides a unique opportunity to study the basis for pathogenicity and disease acquisition. |
10.1371/journal.pntd.0001045 | Evaluation of Spatially Targeted Strategies to Control Non-Domiciliated Triatoma dimidiata Vector of Chagas Disease | Chagas disease is a major neglected tropical disease with deep socio-economical effects throughout Central and South America. Vector control programs have consistently reduced domestic populations of triatomine vectors, but non-domiciliated vectors still have to be controlled efficiently. Designing control strategies targeting these vectors is challenging, as it requires a quantitative description of the spatio-temporal dynamics of village infestation, which can only be gained from combinations of extensive field studies and spatial population dynamic modelling.
A spatially explicit population dynamic model was combined with a two-year field study of T. dimidiata infestation dynamics in the village of Teya, Mexico. The parameterized model fitted and predicted accurately both intra-annual variation and the spatial gradient in vector abundance. Five different control strategies were then applied in concentric rings to mimic spatial design targeting the periphery of the village, where vectors were most abundant. Indoor insecticide spraying and insect screens reduced vector abundance by up to 80% (when applied to the whole village), and half of this effect was obtained when control was applied only to the 33% of households closest to the village periphery. Peri-domicile cleaning was able to eliminate up to 60% of the vectors, but at the periphery of the village it has a low effect, as it is ineffective against sylvatic insects. The use of lethal traps and the management of house attractiveness provided similar levels of control. However this required either house attractiveness to be null, or ≥5 lethal traps, at least as attractive as houses, to be installed in each household.
Insecticide and insect screens used in houses at the periphery of the village can contribute to reduce house infestation in more central untreated zones. However, this beneficial effect remains insufficient to allow for a unique spatially targeted strategy to offer protection to all households. Most efficiently, control should combine the use of insect screens in outer zones to reduce infestation by both sylvatic and peri-domiciliated vectors, and cleaning of peri-domicile in the centre of the village where sylvatic vectors are absent. The design of such spatially mixed strategies of control offers a promising avenue to reduce the economic cost associated with the control of non-domiciliated vectors.
| Chagas disease is one of the most important parasitic diseases in Latin America. Since the 1980's, many national and international initiatives have contributed to eliminate vectors developing inside human domiciles. Today's challenge is to control vectors that are non-adapted to the human domicile, but still able to transmit the parasite through regular short stay in the houses. Here, we assess the potential of different control strategies applied in specific spatial patterns using a mathematical model that reproduces the dynamic of dispersion of such ‘non-domiciliated’ vectors within a village of the Yucatan Peninsula, Mexico. We show that no single strategy applied in the periphery of the village, where the insects are more abundant, provides satisfying protection to the whole village. However, combining the use of insect screens in houses at the periphery of the village (to simultaneously fight insects dispersing from the garden and the forest), and the cleaning of the peri-domicile areas of the centre of the village (where sylvatic insects are absent), would provide a cost-effective control. This type of spatially mixed strategy offers a promising way to reduce the cost associated with the repeated interventions required to control non-domiciliated vectors that permanently attempt to infest houses.
| Chagas disease, also called American trypanosomiasis, is caused by the protozoan parasite Trypanosoma cruzi, which is primarily transmitted to humans by blood-sucking bugs of the Triatominae subfamily. The disease is endemic throughout Latin America, where it is one of the most important parasitic diseases with large socioeconomic impact. According to various estimates, the prevalence rate in humans varies between 0.1 and 45.2% (with an average of 1.4%), 8 to 15 million people are infected with T. cruzi (with 40–50,000 yearly new cases), and 28–75 million individuals are at risk of infection [1]–[3]. The disease causes about 12,500 deaths a year, and is responsible for premature disabilities of workers that are estimated to cost 670,000 disability-adjusted life years lost [4].
Although international initiatives have been launched to reduce transmission of Chagas disease, especially through vector control and screening of blood or organ donors [5], there are still large regions with active vector transmission [6]. One of the main explanations for this is the transmission caused by non-domiciliated triatomines [7]. These vectors are not able to reproduce and develop in the domestic habitat, and thus constitute typical ‘sink’ domestic populations sustained by peri-domestic and/or sylvatic ‘source’ populations [8]. Non-domiciliated vectors tend to jeopardize the efficacy of vector control by insecticide spraying in the domestic habitat because of the re-infestation of treated houses [9], [10], [11]. This situation has been described for several vector species of triatomines as T. brasiliensis and T. pseudomaculata in Brazil [12], T. mexicana in central Mexico [13] and T. dimidiata in the Yucatan Peninsula of Mexico and Belize [14], [15]. Accordingly, the risk of transmission associated with non-domiciliated vectors is now identified as a major challenge for the future of Chagas disease control [16], [17], [18], and a key objective is to evaluate the efficacy of classical or alternative control strategies to reduce their abundance.
Identifying optimal strategies can hardly be achieved through laboratory or field experiments, since testing a broad enough number of alternatives would require very large human and financial investments [11], [19]. Alternatively, mathematical models have proven to be very effective at evaluating the relative merit of various alternative strategies to control parasitic diseases [11; and references therein]. In addition, identifying optimal strategies clearly requires a detailed understanding of the vector spatial and temporal infestation dynamics. Valuable insights into such spatio-temporal dynamics can be gained using the framework of meta-population theory combined with presence/absence data [19]–[21]. Although appealing, the use of more elaborated models that include quantitative information on local population sizes requires even more data than the meta-population model sensus stricto [22].
In previous contributions, we developed spatially explicit population dynamics models that were able to reproduce and to predict the spatial and temporal dynamics of T. dimidiata house infestation observed at the village scale in the Yucatan Peninsula, Mexico. These models provided us with indirect estimates of the origin and characteristics of dispersal of these triatomines [23], [24]. Individuals found inside houses in the Yucatan Peninsula originated in similar proportions from both sylvatic and peri-domestic habitats, dispersed over rather small distances (40–60 m per displacement) and were strongly attracted to houses [24]. Remarkably, the observed and predicted dynamics showed an heterogeneity in transmission risk both in time, with a peak of vector abundance during March–June [14], [25], and in space, with much higher abundance of insects in the periphery of the village reflecting the influence of the sylvatic habitat [11], [26]. The temporal optimization of insecticide spraying with respect to this pattern has already been investigated at the scale of one house [11], but the spatial micro-scale heterogeneity suggests that interventions could also be spatially targeted. Such interventions would focus on the periphery of the village, where bugs were found more abundant. While temporal heterogeneity adds constraints on control strategies (i.e. the timing of intervention has to match the seasonality of house infestation, [11]), spatial heterogeneity could have beneficial consequences for control activities as it might allow to reduce the overall surface (or number of houses) to be treated and thus allow to reduce the cost associated with control. Properly assessing whether such spatial design is relevant requires evaluating not only the efficacy of control in the treated areas, but also the impact of the control interventions in the untreated areas of the same village.
In this contribution, we aimed to build on our understanding of the temporal optimization of control strategies [11], as well as our previous spatial modelling [23], [24] to evaluate the potential of several strategies. We first focus on conventional strategies — namely indoor insecticide spraying, use of door/window insect screens and peri-domicile management — that have been used to control vectors of different diseases as well as T. dimidiata [27]–[29]. We further look at the potential of insect lethal traps that are currently extensively investigated for the control of a variety of vector species [30], [31]. Finally, since we have previously found that T. dimidiata was directly attracted to houses [24], a control alternative could be to eliminate this house attractiveness, and the potential of such a strategy was also explored.
We aimed to set up a spatial population dynamics model able: (1) to reproduce and predict the temporal variations of vector abundance in all the houses of one village in the absence of control, and (2) to spatially represent various control strategies. We adapted previous population dynamic models [23], [24], and combined them with a mathematical description of the control strategies that we aimed at evaluating. The resulting model predicts the temporal variations in vector abundance in every house of the village as a function of survival, reproduction and dispersal of the triatomines, and the effect of the above control strategies on the demographic processes at each point of the village. It was then used for the evaluation of the efficacy of spatially targeted interventions based on each of those strategies. Model predictions in absence of control were fitted through a maximum likelihood approach to a first set of spatio-temporal data describing house infestation dynamics by T. dimidiata within a village in the absence of vector control. We tested the predictive value of the resulting parameterized model on a replicate data set, corresponding to the infestation dynamics observed in the same village the following year.
The description of the effect of the different control strategies was then added to the model, and the resulting framework was used to explore the efficacy of control interventions whose spatial coverage was progressively increased from the border to the centre of the village. The efficacy of each intervention was evaluated as the percentage of reduction in the yearly abundance of vectors in the village, in comparison with the expected abundance in the absence of control intervention that we evaluated from the model with no control. Efficacy was also related to the consented effort, as measured by the number of households, where control strategies were applied (either in the domestic or peri-domestic habitats). We performed a sensitivity analysis to each survival, reproduction or dispersal parameter of the model to ensure the robustness of our conclusions on the efficacy of the various interventions within the confidence region associated with the maximum likelihood estimate of model parameters. We further conducted a sensitivity analysis to different parameters of the model that described the efficacy of each of the strategies as measured by their impact on the survival, reproduction or dispersal of triatomines.
The spatio-temporal pattern of house infestation was observed in the rural village of Teya, Yucatan, Mexico over a two-year period from August 2006 to October 2008 [26]. All houses were identified and geo-referenced with a handheld global positioning system (GPS). Insects were collected by a standardized methodology based on community participation [32], and data were imported into a geographic information system (GIS) database (ArcView 3.2 -Environmental Systems Research Institute, Redlands, CA, USA) to produce maps of observed triatomine abundance in the houses over 2-week intervals [32]. Participating families provided oral consent prior to their participation, as written consent was waived because the study involved no procedures for which written consent is normally required outside of the research context. Consent was logged in field notebooks. All procedures, including the use of oral consent, were approved by the Institutional Bioethics Committee of the Regional Research Centre “Dr. Hideyo Noguchi”, Universidad Autonoma de Yucatan.
We set up a GIS-based Spatially Explicit Model (GIS-SEM) as such modelling provides a suitable framework to investigate spatial population dynamics in real landscapes by importing GIS data on a grid representing the area under study [33]. Our GIS-SEM model was based on Cellular Automaton (CA) formalism [34]. It consisted of a grid of cells representing the village of Teya, and allowed the calculation of the temporal variations of the vector abundance in cells, referred to as state variables, according to both local rules describing birth and death processes of bugs within cells, and dispersal rules that allow accounting for walking and flight movements between neighboring cells. This model was similar in essence to the models built by Barbu et al. [24], but with two necessary adaptations. First, the local and dispersal rules were described in a deterministic rather than stochastic manner to reduce the complexity of the model and shorten the simulation time. Second, the time unit of the model was changed from 15 days to a day to allow specifying the effect of control on a daily basis.
A deterministic CA such as the one intended here is defined as a quadruple Q = (A,S,V,f), where A is the grid of cells arranged uniformly to represent the studied area; S is the set of values that can be taken by the state variables; V is the neighborhood function that allows identifying the set of neighboring cells V(c) that contribute to the change of the state variable of any given cell c by the mapping:(1)with v denoting the size of the neighborhood; and where f is the function describing the local and dispersal rules and thus specifies how the set of neighboring cells V(c) changes the state of the cell c from one time step to another:(2)with N(c,t), the state variable that tracks the status of cell c at time t.
Maximum likelihood estimates (MLE) of the parameters of the model with no control were obtained using the spatio-temporal data sets describing T. dimidiata infestation dynamics of the village of Teya between mid-September 2006 and mid-September 2007. Model predictions were fitted to the observed number of bugs in each cell of the 24 maps describing the average biweekly distribution within the village. The log likelihood (LLH) value was then calculated as follows:(6)where log denotes the natural logarithm, X(c,t) is the statistical variable corresponding to the number of adults in cell c, O(c,t) the observed abundance in this cell, and θ is a set of parameters of the model. Probabilities were defined assuming a zero-inflated Poisson distribution to take into account an excess of null abundance in the data set [39], possibly due to the non-participation of a proportion (w) of householders, with w = 0.7 as before [24].
The parameters θ of the model were identified using a genetic algorithm run at the super-computing centre ‘Institut du Développement et des Ressources en Informatique Scientifique (IDRIS)’ located at Orsay, France (http://www.idris.fr/ - Project IDRIS 112290). Genetic algorithms search for solutions using techniques inspired by natural evolution. The interested reader can find a detailed description of such methods and the typical terminology we adopted below in [40]. The algorithm considered the 8 parameters of the model (Sd, Sp, d, Kp, Ks, D, σ and H) to be estimated as independent quantitative traits with a continuum of alleles representing possible trait values within biologically relevant domains. The fitness function corresponded to the LLH value defined with respect to the GIS-SEM model with no control described above. The fittest individuals were selected to produce offspring through free recombination and unbiased mutations. The variance of the effect of the mutations was dynamically adapted to the variance in the parental population. All codes were written in C/MPI.
Confidence intervals were calculated by establishing the profile likelihood for each parameter , and by using these relationships to determine the 1−α confidence region defined as:(7)where is the MLE of parameter and stands for the (1−α)th quantile of the distribution on 1 degree of freedom [41].
The ability of the parameterized model to predict other infestation dynamics was tested by comparing its prediction to the spatio-temporal distribution of bug abundance in a second year of infestation of the same village. A Poisson regression between observed and predicted abundances was performed after data were pooled over 3-month periods (starting in mid-September) and within three distance categories: 0–80 m, 81–200 m and >200 m from the bush area outside the villages [24], [26]. The McFadden's likelihood ratio index was used as a pseudo R-squared.
Because the spatial distribution of bugs follows a spatial gradient with higher abundance at the periphery of the village [24], [26], the control strategies were applied to a ring of cells located at the border of the village, the size of this ring increasing progressively until the intervention covered the whole village (Figure 1). The efficacy of any given spatially targeted strategies was measured in terms of yearly bug abundance both in the whole village and in the different concentric rings. This allowed us to quantify the relationship between the effort in terms of control coverage and the global efficacy, and to simultaneously assess the consequences of interventions in the various parts of the village. The efficacy of intervention was evaluated using the set of parameters' estimates providing the best fit to the data. It was complemented by a sensitivity analysis of the corresponding results to the parameter's estimates. Each parameter was then independently set to the boundary values of its confidence interval, i.e. and , while keeping the others to their MLE.
We evaluated the efficacy of five types of control strategies applied individually, including indoor insecticide spraying, door and window insect screens, peri-domicile cleaning, triatomine lethal traps located in the peri-domestic habitat, and housing improvement to reduce house attractiveness to bugs. The effect of each strategy on bug survival, reproduction and/or dispersal was modelled as described below (see also supplementary methods — Text S1 — for the mathematical changes that were made to the model to include these effects).
Indoor insecticide spraying was modelled by reducing vector survival in each treated house as before [11]. The control-induced mortality was calculated with respect to the residual dose of insecticide that we adjusted daily, and to the lethality of the dose as expected from a typical sigmoid dose-response relationship. Assuming that the control-induced and natural mortalities act independently (i.e. to survive one of the two causes of death does not affect the probability to survive the second one), we combined them multiplicatively to define the overall survival probability.
We considered a spray rate of 50 mg.m−2 of pyrethroid insecticide at the beginning of the infestation season (since it was previously shown to be the optimal timing for spraying [11]), the half-life of the insecticide was set to 38 days, and the lethal doses 50% and 90% were fixed to 32.2 mg.m−2 and 182.4 mg.m−2 [11]. A sensitivity analysis to insecticide dose was performed predicting the effect of spraying at 100, 200 and 300 mg.m−2.
Door and window screens were considered as physical barriers impeding the arrival of a proportion of the non-domiciliated vectors into the domestic habitat, and were thus modelled by lowering immigration into the houses by a factor of bug exclusion r set at 85% and constant over time [11], [42]. Again, a sensitivity analysis was conduced by considering r equals to 70, 80 and 90%. Because the efficacy of screens is likely to depend on the behavioral response of dispersal bugs failing to enter houses because of screens, and because no information was available in the literature about such a response, we considered three alternative assumptions. Bugs that could not enter into houses were considered: (1) to stop dispersing and die, or (2) to stop dispersing for one day before starting again with no learning in their dispersal behavior (and thus possibly attempting to enter the same house), or (3) to go on dispersing while avoiding the house they could not enter.
Peri-domicile cleaning was assumed to eliminate all bug colonies established in this habitat for the rest of the current year. This reduced immigration from the cleaned sites, but did not have any effect on individuals that originated from other areas and may pass through the peri-domiciles where this control strategy was applied. In addition, we performed a sensitivity analysis by considering that cleaning removes only 60% and 80% of insects established in the peri-domestic habitat.
Manipulation of houses' attractiveness to bugs was achieved by decreasing H from its estimated value to 1, the value for which houses are no more attractive than the peri-domestic and sylvatic habitats. This represents the strongest possible effect and allows evaluating the maximal potential for this strategy; a sensitivity analysis for the intermediate values of H was then performed.
Triatomine lethal traps in the peri-domestic habitat were assumed to attract and kill triatomines into the cells where they are positioned according to an additional parameter Htrap that measured the trap attraction. As for the study of the control of houses' attractiveness we first wanted to evaluate the maximal potential of this strategy. The density of traps was then fixed at 2 traps per household, and attraction was set to a constant level Htrap = 12, almost twice the attraction of houses. Sensitivity analysis was then performed for different density of traps, in the range 5 traps per household to 1 trap for 10 households, and trap attraction, in the range 1 to 50.
The model predictions fitted very well the yearly spatio-temporal dynamics of infestation observed in the village of Teya between mid-September 2006 and mid-September 2007. The correlation between observed and simulated spatio-temporal data indicated that the model reproduced well both the seasonal variations in triatomine densities, and the spatial spread of bugs from the border to the centre of the village (Figure 2A, McFadden's likelihood ratio index = 0.93). Importantly, the model parameterized with the data on this first year was able to predict the observed spatial and temporal dynamics of bug abundance in the following year (Figure 2B, McFadden's likelihood ratio index = 0.67). We note that while our model tends to predict well high abundances, predictions at lower vector abundances seem less precise. However, this is rather inconsequential since predicting fine variations in space and time at low abundances is of little relevance for our ultimate objective of evaluating control strategies.
The convergence of the presented results with a previous study, that used a stochastic model [24], also showed that the selected local and dispersal rules (see Definition of the function f including the local and dispersal rules) were reliable in their ability to both reproduce and anticipate the spatio-temporal dynamics of these non-domiciliated vectors. Likelihood profile confidence intervals gave further information on the estimated parameters of these rules (Table 1). Those confidence intervals were quite narrow around the MLE. The lower and upper boundaries were typically located at less than 30% of the MLE of each of the parameters, indicating that larger changes in one of the parameter estimates would no longer allow properly reproducing the data. The survival rates in the domestic and peri-domestic habitats were very close to 0.2 and 0.9, respectively; the numbers of insects immigrating from the colonies established in the sylvatic and peri-domestic habitats were in the range 150–260 insects for 15 days; there was nearly a 1∶1 ratio between immigration from the sylvatic and peri-domestic habitats; the attraction to the house was always at least 5 times higher than attraction to the peri-domestic area, and the optimal (and mean) distance of dispersal was between 50 and 60 meters (Table 1). All of those results were consistent with and supported our previous conclusions that insects found in houses came in roughly similar proportion from the sylvatic and peri-domestic habitats and that they disperse over rather small distances and with a strong attraction to the domestic habitat [24]. Overall, our spatial model with no control thus offered a good framework where spatially targeted control strategies could be evaluated.
We investigated the efficacy of the five strategies considered independently by applying them to concentric rings defined from the border of the village and whose size was increased until a complete coverage of the village was reached. For each strategy, we calculated its efficacy, measured as the post-intervention reduction of bugs' abundance in the whole village, in function of the extent of village zones treated, i.e. the effort in terms of the control intervention (Figure 3A–E). We also calculated the effect of the interventions in each concentric village area, including those without control intervention (Figure 3F–J). Finally, we performed a sensitivity analysis to the parameter values by independently replacing the MLE with the upper and lower values of each profile likelihood based confidence interval (Table 1).
The first key point is that all the results obtained with each of the five strategies were only weakly sensitive to changes in demographic parameters values. Such changes indeed lead to no qualitative change in the form of the relationships (Figure 3). As expected, parameters with the strongest effect depend on the control strategy considered. Maximal changes were obtained when changing survivals (Sp, Sd) for insecticide spraying, immigration rates (d) for screens and outdoor traps, houses' attraction (H) for the control of houses' attractiveness and the number of individuals leaving colonies (Kp, Ks) for peri-domestic cleaning (results not shown). However, these effects were systematically lower than 5% on both treated (Figure 3A–E) and untreated areas (Figure 3F–J). The results obtained are thus very robust to variations of the parameters of the model with no control, and we will thus further describe only the results obtained with the MLEs.
Indoor insecticide spraying in the whole village allowed the reduction of total bug abundance over a year by about 70% for one year (Figure 3A). The relationship between the proportion of treated houses and global efficacy was a slightly convex diminishing return curve, so that half of the maximal decrease could be obtained by spraying only the first two external zones of the village (a third of the houses). We also evaluated the local efficacy of insecticide in untreated village zones at the forefront of the treated areas. Independently of the number of village areas sprayed, the use of indoor insecticide only reduces the vector abundance in the treated area; it has a negligible effect on neighboring untreated areas (Figure 3F). To increase the dose applied allowed the predicted levels of vector reduction to reach higher levels (doses of 100, 200 and 300 mg.m−2 lead to a 79%, 85% and 87% maximal control efficacy, respectively; data not shown), with no change in the main conclusion: Insecticide spraying in only the first two outer zones allowed for half of the maximal control efficacy.
Door and window insect screens applied to all the houses of the village decreased the total vector abundance by about 80% when bugs that could not enter into houses were assumed to go on dispersing (assumptions 2 and 3, the former including possible attempts at entering again the house they just failed to infest) (Figure 3B). As for insecticide spraying, there was a slightly convex diminishing return between the number of treated zones and efficacy. Accordingly, limiting the intervention to the first two zones at the periphery of the village (a third of the village houses) again led to half of the maximal reduction in abundance. Under the two assumptions not including the death of the insects failing to enter the houses [2]–[3], the analysis of insect screens' local efficacy indicated that while infestation was well controlled in houses with screens, the control had a detrimental effect on the immediate non-equipped neighbor: an increase of up to 40% in vector abundance was estimated in the most proximate untreated village zone (Figure 3G). This negative effect on neighboring areas disappeared for untreated areas more than 3 zones away from the treated one. On the other hand, when the vectors were assumed to die when failing to enter a house (assumption 1), the effects of screens were significantly different. In this case, vector abundance was reduced slightly further (up to 90%) when screens were used in all the houses of the village (Figure 3B upper dotted black line), and the control strategy then had no negative effect on untreated neighboring houses (Figure 3G upper dotted black line). To vary the efficacy of screens produced only small linear changes in the global efficacy. Under assumptions 2 and 3, a reduction factor r of 70%, 80% and, 90% lead to a 51%, 64% and 80% maximal control efficacy, while under assumption 1, a reduction factor r of 70%, 80% and 90% led to a 73%, 82% and 91% maximal control efficacy; data not shown. The above conclusions are consequently very robust to variations of r, which is thought to be in the range 80–90% in the field [42].
Peri-domicile cleaning reduced total bug abundance by up to 62% for one year when performed in the whole village (Figure 3C). The increase in efficacy with increasing coverage was a concave relationship with a slightly increasing return. Because of the lower efficacy of peri-domicile cleaning at the periphery of the village, intervention in at least the first 3 zones (60% of the village peri-domestic surface) was required to reach half of the maximal reduction in abundance. Interestingly, when peri-domicile cleaning was performed only in some parts of the village it had an important beneficial effect on untreated neighboring houses. The vector abundance in the two closest non-treated zones was reduced by 40% and 15% respectively (Figure 3H). Lowering the rate of colonies' destruction by peri-domicile cleaning, which was initially set to 100%, lowered the total efficacy in an almost perfectly linear way, but again had no effect on the above qualitative conclusions. Typically, assuming that only 80% or 60% of colonies are removed by cleaning peri-domiciles allowed for a maximal control efficacy of 50% (≈62%×80%) and 37% (≈62%×60%), and in both cases intervention in the first 3 zones was needed to get half of these outcomes.
Manipulation of houses' attractiveness was found 60% effective when applied to the whole village and when such attraction was completely eliminated, so that domestic habitat was no more attractive than the peri-domestic and sylvatic habitats (H = 1) (Figure 3D). Half of the maximal efficacy could be reached by an intervention targeted on the first two zones of the village representing a third of the village houses. However, such strategy had an important negative impact on the abundance of bugs in non-manipulated neighboring houses when applied to parts of the village (Figure 3I). Indeed, the lack of attraction of manipulated houses resulted in an increase of over 50% and 30% in bug abundance in the next two untreated village zones. Importantly, sensitivity analysis of intermediate values of reduction in house attractiveness indicated that efficacy of the intervention was rapidly lost as H was incompletely reduced: the maximal efficacy was of 40%, 17% and less than 5%, for H values of 2, 4, and 6, respectively (Figure 4).
Insect lethal traps were found potentially able to reduce global vector abundance by up to 72% when considering a high density (two traps per household) and a high attractiveness (Htrap = 12, nearly twice the attractiveness of houses) and 100% of lethality (Figure 3E). Under these conditions, an important diminishing return was observed since to install traps in the first zone of the village (27% of the peri-domestic surface) allowed to attain half of the maximal efficacy. Furthermore, this strategy had substantial positive effects on the 4–5 neighboring areas without traps, where insects' abundance was decreased by 50%, 30%, 15% and 7%, respectively (Figure 3J).
However, reducing the attraction factor of each trap had an important effect at the village scale as the global control went down from 72% to 55% when attraction of individual traps was reduced from Htrap = 12 to Htrap = 5, a value similar to the attractiveness of houses (Figure 5). On the contrary, to increase attraction to higher levels had almost no effect whatever the number of traps considered. To lower the number of traps also had a strong detrimental effect, and the reduction of bug abundance due to control was never found larger than 30% when the number of traps was dropped to 1 trap for 10 households (Figure 5). A nearly 100% control efficacy at the village scale was reached only when more than 2,500 traps were used in the village, which represent about 5 traps per household within the village.
Although the elimination of transmission of Chagas disease was targeted by the WHO for the year 2010 [4], there are still large regions with active vectorial transmission mostly due to non-domiciliated triatomines [6]. These vectors do not constitute permanent colonies inside houses, so that domestic populations actually are typical ‘sinks’ sustained by peri-domestic and/or sylvatic ‘source’ populations [8]. The risk of transmission associated with these non-domiciliated vectors is thus now identified as a major challenge for the future of Chagas disease control [5], and a key objective is to evaluate the efficacy of classical or alternative control strategies to reduce their abundance. Since non-domiciliated insects infesting houses typically come from the sylvatic and peri-domestic habitat [11], to evaluate the potential of various strategies requires a good understanding of the village infestation dynamics in absence of control. In this perspective, spatial population dynamic models able to reproduce and predict the dispersion of individuals from these two non-domestic habitats are valuable tools.
Taking advantage of previous field and modelling works on well-studied populations of non-domiciliated triatomines in villages of the Yucatan Peninsula, Mexico, we performed the first attempt to evaluate the efficacy of putative control strategies applied spatially. We identified triatomines' dispersal characteristics through a selection model approach based on maximum likelihood estimates [37], [38]. The best deterministic model and the associated estimates of the dispersal characteristic identified here were found very similar to the ones identified in a similar approach, but based on stochastic models [24]. In addition, just as the previous more complex stochastic model, our deterministic model reproduced and predicted very well the spatio-temporal dynamic of the village infestation. The present study thus confirmed that the selection model approach is a well-adapted strategy to simultaneously obtain indirect estimates of triatomines dispersal, hard to quantify in the field [19], and robust GIS-based Spatially Explicit Models (GIS-SEM) able to reproduce and predict the dynamic of infestation in the absence of control. Such a model is required for the evaluation of the efficacy of putative control strategies; to this end we combined our selected model with a representation of different strategies to evaluate their potential.
We found that indoor insecticide spraying and insect screens applied to the entire village were able to reduce yearly vector abundance in the whole village by 70 and 80%. Interestingly, in both cases, half the maximal effect was obtained while interventions were limited to the first two outer zones of the village. This mostly reflected the higher abundance of insects typically found in houses in the periphery of the village, where the vectors dispersing from both the peri-domestic and sylvatic habitats contribute to domestic infestation [23], [24], [26].
Although global efficacy was roughly similar for these first two strategies, a possible difference between them could be on their effect on untreated neighboring households. Indeed, insect screens were shown to impose some additional infestation on nearby untreated houses when vectors were allowed to go on dispersing after failing to infest a protected house. However, this negative effect was not present when vectors were assumed to systematically die after their first attempt to infest a protected house. Interestingly, the latter scenario is qualitatively consistent with a field trial conducted in a village of the north of the Yucatan Peninsula, in which the use of impregnated curtains and windows screens in some houses seems to reduce bug abundance in nearby untreated houses [42]. This may be due to some knockout effect of the low dose insecticide used for impregnation, or to a poor energetic status and/or exhaustion of bugs that could prevent re-departure after a flight/walk to intent infest a first house.
Particularly in these conditions, and even if more empirical and modelling studies are needed to quantify vector dispersal at the individual scale, our results do support the idea of a spatially targeted use of insect screens to control the higher bug abundance at the periphery of the village as it maximizes the overall reduction in transmission risk at the level of the entire village. The most cost-efficient intervention would then be to treat the houses located in the first two outer zones (about 33% of the total houses of the village) to obtain around 50% bug abundance decrease in the entire village.
The weak effect of insecticide spraying on the neighboring houses shown in this study is also consistent with a field trial [42]. Treating the first two outer zones would allow obtaining about 40% decrease in total bug abundance in the entire village but with no efficiency on untreated areas. Those results suggest that the cost associated to the temporary effect of insecticide spraying on non-domiciliated vectors demonstrated at the house scale [9], [11], can only weakly be compensated for by spatially targeted strategies that would exploit the typical gradient of abundance due to the immigration of sylvatic bugs [26], [43].
Peri-domicile cleaning appears to be an interesting alternative strategy having the potential to substantially reduce vectors abundance inside the treated zones and to exert a positive influence on untreated areas. By eliminating all the colonies established in the backyards, a perfect cleaning of the peri-domiciles provided a 60% reduction of bug abundance in the village, although it provided a substantially lower efficacy at the periphery of the village, compared to the efficacy of residual insecticides and insect screens. This lower predicted efficacy at the village scale and in the outer zones is due to the absence of impact of this strategy on insects dispersing from the sylvatic habitat. It is also consistent with previous estimates indicating that infesting bugs come from both peri-domestic and sylvatic sites, and that both sources need to be controlled [24], [44].
Interestingly, the positive effect of this strategy on nearby households with no intervention confirmed results of a previous field trial where peri-domicile cleaning (elimination of unnecessary objects of the peri-domicile followed by insecticide spraying) also reduced infestation in neighboring houses without intervention [42]. Accordingly, peri-domicile cleaning could valuably be used to significantly reduce bug abundance, especially in the centre of the village where the majority of non-domiciliated vectors found in houses come from the peri-domestic habitat [24]. Because it targets specifically peri-domestic vectors, such a strategy could lead to a substantial level of control when combined with insect screens in the periphery of the village.
The manipulation of house attractiveness was explored here as a potential novel vector control intervention based on the rationale that triatomines were found to be directly attracted to the houses [24]. We found that such an intervention could reduce domestic infestation by up to 60% when applied to the entire village. However, when applied to only a fraction of the houses, we show that it would induce an increased infestation of neighboring untreated areas as bugs no longer attracted to manipulated houses tend to disperse to nearby domiciles. Control intervention based on this strategy should thus preferentially be implemented in all the houses of the village, and feasibility would then rely on the kind of modifications to be done in the domestic habitat to limit attraction.
The actual determinants for house attractiveness to bugs are still unknown, but if light is proven to be a key factor [28], [45], [46], the use of devices limiting the diffusion of the light may be considered. Nevertheless, it is important to emphasize that the effect of the intervention is rapidly lost if the reduction in the attractiveness is only partial. This strategy would thus be of little interest if a nearly complete reduction in house attractiveness to bugs could not be achieved. Thorough research on the mechanism and factors of triatomine dispersal toward houses would then be needed to allow the implementation of such a strategy.
Triatomine lethal traps were also tested in an attempt to keep bugs away from the houses. Such traps were estimated effective if they were highly attractive and lethal, and used at very high densities; in these conditions they would also have a marked beneficial effect on neighboring houses without traps. The attractiveness of potential traps such as yeast-baited traps is difficult to estimate, but available studies suggest an attractiveness H in the range of 2–3, i.e. rather less that the attractiveness of houses evaluated at 6–7 [47]–[49]. In such conditions, the use of 5 traps per household, which would represent about 2,500 traps in the whole village, only allows for about 30% reduction of triatomines abundance in the village.
In addition, traps were assumed to be of constant efficacy in our model, which seems to be highly unlikely in practice, as it would raise the issue of the periodic maintenance/renewal of the traps depending on their half-life. It thus seems that the performance of potential outdoor traps would need to be dramatically improved to become a viable strategy for non-domiciliated triatomine control.
Overall, this study has shown that control strategies applied at the periphery of a village can contribute to reduce infestation in untreated, more central houses, but only in limited proportions. Typically, insecticide or insect screens used in the first two outer zones of the village, which represents 33% of the households, would only reduce vector abundance in the whole village by 40–50%. In these conditions, spatial targeting of strategies based on either insecticide spraying or insect screens applied to houses in the two outer zones of a village, combined with peri-domicile cleaning in the centre, would provide optimum vector control at the lowest cost (Table 2). Essentially, such mixed strategy would remove peri-domestic colonies where they are the major source of vectors, and impede the insects to enter houses where they also come from non-manageable sylvatic colonies.
At first, the costs of combining insecticide spraying or insect screens with peri-domicile cleaning seem roughly equivalent. However, the seasonal pattern of house infestation requires, for insecticide spraying, the dispatch of a large number of spraying teams to cover an entire region within 2 months [11], generating additional costs of transportation and logistics [27].
Thus, a combination of insect screens in the periphery and peri-domicile cleaning in the centre would be the most cost-effective and sustainable strategy to be implemented in the Yucatan Peninsula. The design of such spatially mixed strategies of control offers a promising avenue to reduce the economic cost associated to the repeated intervention intrinsically associated with the permanent re-infestation of houses by non-domiciliated vectors [9], [11].
|
10.1371/journal.ppat.1003381 | Murinization of Internalin Extends Its Receptor Repertoire, Altering Listeria monocytogenes Cell Tropism and Host Responses | Listeria monocytogenes (Lm) is an invasive foodborne pathogen that leads to severe central nervous system and maternal-fetal infections. Lm ability to actively cross the intestinal barrier is one of its key pathogenic properties. Lm crosses the intestinal epithelium upon the interaction of its surface protein internalin (InlA) with its host receptor E-cadherin (Ecad). InlA-Ecad interaction is species-specific, does not occur in wild-type mice, but does in transgenic mice expressing human Ecad and knock-in mice expressing humanized mouse Ecad. To study listeriosis in wild-type mice, InlA has been “murinized” to interact with mouse Ecad. Here, we demonstrate that, unexpectedly, murinized InlA (InlAm) mediates not only Ecad-dependent internalization, but also N-cadherin-dependent internalization. Consequently, InlAm-expressing Lm targets not only goblet cells expressing luminally-accessible Ecad, as does Lm in humanized mice, but also targets villous M cells, which express luminally-accessible N-cadherin. This aberrant Lm portal of entry results in enhanced innate immune responses and intestinal barrier damage, both of which are not observed in wild-type Lm-infected humanized mice. Murinization of InlA therefore not only extends the host range of Lm, but also broadens its receptor repertoire, providing Lm with artifactual pathogenic properties. These results challenge the relevance of using InlAm-expressing Lm to study human listeriosis and in vivo host responses to this human pathogen.
| Co-evolution of microbes with their hosts can select stringently specific host-microbe interactions at the cell, tissue and species levels. Listeria monocytogenes (Lm) is a foodborne pathogen that causes a deadly systemic infection in humans. Lm crosses the intestinal epithelium upon the interaction of its surface protein InlA with E-cadherin (Ecad). InlA-Ecad interaction is species-specific, does not occur in wild-type mice, but does in transgenic mice expressing human Ecad and knock-in mice expressing humanized mouse Ecad. To study listeriosis in wild-type mice, InlA has been “murinized” to interact with mouse Ecad. Here, we demonstrate that in addition to interacting with mouse Ecad, InlAm also uses N-cadherin as a receptor, whereas InlA does not. This artifactual InlAm-N-cadherin interaction promotes bacterial translocation across villous M cells, a cell type which is not targeted by InlA-expressing bacteria. This leads to intestinal inflammation and intestinal barrier damage, both of which are not seen in humans and humanized mouse models permissive to InlA-Ecad interaction. These results challenge the relevance of using InlAm-expressing Lm as a model to study human listeriosis and host responses to this pathogen. They also illustrate that caution must be exercised before using “murinized” pathogens to study human infectious diseases.
| Co-evolution of microbes with their hosts can select stringently specific host-microbe interactions at the cell, tissue and species levels [1]. Species-specific host-microbe interactions, which are the rule rather than the exception, pose a challenge for the use of laboratory animal models to study human pathogens, including Listeria monocytogenes (Lm), the etiological agent of listeriosis, a deadly foodborne infection. Lm is able to actively cross the intestinal barrier, reach the systemic circulation and cross the blood-brain and placental barriers, leading to its dissemination to the central nervous system and the fetus [2].
The mouse is a genetically amenable model that is widely used to investigate human diseases [3], [4]. To obtain a mouse model in which the pathogenic properties of a given pathogen are similar to what is observed in human, species specificity can be circumvented by humanizing the mouse by transgenesis [5], [6], [7], [8], knock-in [9], knock-out [10] or xenograft techniques [11]. One can also adapt the pathogen to the mouse by multiple passages on cell lines [12], [13] or in vivo [14], or specifically “murinize” a pathogen ligand so that it interacts with the mouse ortholog of a species-specific human receptor [15], [16].
The Lm surface protein InlA interacts with E-cadherin (Ecad) and mediates Lm entry into epithelial cells, which express this adherens junction protein [17], [18]. Cadherins constitute a family of calcium-dependent cell adhesion receptors. Ecad is expressed mainly in epithelia, whereas N-cadherin (Ncad) is found primarily in neuronal cells and endothelial cells together with VE-cadherin [19], [20]. Ncad can also be coexpressed with Ecad in epithelial cells [21]. Importantly, Ncad has been reported to not act as a receptor for InlA, and so far Ecad is the only known classical cadherin acting as a receptor for InlA [18]. In contrast to Ecad from human, guinea pig, rabbit and gerbil, mouse Ecad (mEcad) and rat Ecad are not recognized by InlA and do not promote bacterial entry [9], [22]. The interaction of InlB, another Lm invasion protein, with its host receptor is also species-specific [23]. InlB recognizes the hepatocyte growth factor receptor Met of human, mouse, rat and gerbil but not that of guinea pig and rabbit [9], [23], [24].
Two mouse lines have been established to study InlA-Ecad interaction in vivo: a transgenic mouse line expressing human Ecad (hEcad) in enterocytes (hEcad Tg) [6], and a humanized mEcad knock-in mouse line (E16P KI) with an E16P amino acid substitution which enables mEcad to interact with InlA without affecting Ecad homophilic interactions and allows Lm internalization [9], [22]. Using these two humanized mouse models, we have demonstrated that InlA mediates Lm crossing of the intestinal epithelium upon targeting of luminally-accessible Ecad around goblet cells [6], [9], [25], and that InlA and InlB act interdependently to mediate the crossing of the placental barrier [9]. Epidemiological investigations have confirmed the relevance of these experimental findings, and shown that InlA is implicated in Lm crossing of human intestinal and placental barriers [9], [26].
In 2007, Wollert et al. engineered a genetically modified InlA with the purpose of increasing its binding affinity to hEcad [16]. Two amino acid substitutions in InlA, S192N and Y369S, were shown to enhance InlA binding affinity to hEcad [16]. Neither S192N nor Y369S substitution has been observed in the more than 500 Lm isolates InlA sequences we have checked (our unpublished results). Wollert et al. published that this increased affinity for hEcad translates into an increased bacterial entry into human epithelial cells (Caco-2) [16]. Importantly, Wollert et al. also showed that this modified InlA binds the extracellular cadherin domain 1 (EC1) of mEcad in solution with a comparable affinity to that of the wild-type (wt) InlA for hEcad EC1 [16]. They hypothesized that this interaction would allow Lm expressing this “murinized” InlA (InlAm) to cross intestinal barrier and would render wt mice orally permissive to Lm infection, a phenotype which is mediated by InlA in permissive models [6]. In support of this hypothesis, Wollert et al. found an increased intestinal, spleen and liver bacterial loads of wt mice orally inoculated with Lm expressing InlAm, yet only after 3 to 4 days post infection, which is later than in models permissive to InlA-Ecad interaction [6], [9], [16]. Moreover, the ability of InlAm to mediate mEcad-dependent Lm internalization into host cells has never been tested. In addition, InlAm unexpectedly promoted pronounced inflammation and intestinal epithelial cell damages in wt mice [16], whereas wt InlA mediates the crossing of the intestinal barrier without inducing significant intestinal response and tissue damage in hEcad transgenic mice [6], [27].
This prompted us to investigate the detailed properties of InlAm in cultured cells, as well as the in vivo cell and tissue tropisms of bacteria expressing InlAm, as compared to that of its isogenic parental Lm strain that expresses wt InlA. Here, we demonstrate that InlAm promotes bacterial entry not only into mEcad-positive but also into mEcad-negative mouse cells. We show that InlAm-mediated entry into mEcad-negative cells is mouse Ncad (mNcad)-dependent. Importantly, InlAm-mNcad interaction allows bacteria to specifically target Ncad-positive villous M cells in vivo, a cell type which is not targeted by Lm in humanized mouse models permissive to InlA-Ecad interaction. This leads to enhanced intestinal inflammatory responses and disruption of the intestinal barrier integrity, both of which are not observed in Lm-infected humanized mice and human listeriosis. Together, these results demonstrate that the murinization of InlA not only extends Lm host range, but also broadens its receptor repertoire, consequently changing Lm cell tropism and enhancing host immune responses to Lm. These results challenge the relevance of using InlAm-expressing Lm to study human listeriosis and in vivo host responses to this human pathogen.
We first investigated whether the increased affinity of InlAm to hEcad translates into an enhanced invasion of hEcad-expressing cells, as proposed by Wollert et al. [16]. To this end, we assessed InlAm-dependent entry into LoVo cell, a human epithelial cell line expressing hEcad [22]. Lm wt strain and Lm expressing InlAm (Lm-inlAm) invaded LoVo cells at similar levels (Figure 1A). Because Lm can be internalized by InlA-independent pathways such as InlB-Met, we transferred either inlA or inlAm onto the chromosome of Listeria innocua (Li), a naturally non-invasive and non-pathogenic Listeria species, in which heterologous expression of inlA has been shown to confer invasiveness [17], [18], [28]. Li expressing either InlA (Li-inlA) or InlAm (Li-inlAm) were equally invasive in LoVo cells (Figure 1B). These results indicate that contrary to what is reported by Wollert et al. [16], the increased affinity of InlAm to hEcad does not translate into an increased level of bacterial entry. Both Li-inlA and Li-inlAm recruited hEcad when incubated with LoVo cells, suggesting that hEcad is involved in both InlA- and InlAm-mediated entries (Figure 1E, upper panel). Because purifed InlAm interacts with the purified EC1 domain of mEcad, Wollert et al. have proposed, although not tested, that InlAm would mediate bacterial entry into mEcad-expressing cells [16]. We therefore tested the ability of InlAm to promote bacterial entry into the mouse epithelial cell line Nme, which expresses mEcad [29]. InlAm promoted bacterial entry into mEcad-expressing Nme cells, although to a lower level than InlA in hEcad-expressing LoVo cells (Figure 1C and D). Li-inlAm also recruited mEcad during cell invasion, whereas as expected, Li-inlA does not (Figure 1E, lower panel). Together, these results show that (i) the increased affinity of InlAm to hEcad does not enhance bacterial entry into hEcad-expressing cells, and (ii) the murinization of InlA confers to Lm an enhanced ability to be internalized into mEcad-expressing cells [16].
Monk et al. have reported that Lm-inlAm invades mouse CT26 cells more efficiently than Lm [13]. Strikingly, CT26 cells do not express mEcad (Figure 2A) [30], yet we confirmed that InlAm mediates bacterial entry into these cells (Figure 2B). Because classical cadherins exhibit a high level of conservation in their EC1 domains (Figure S1A), we tested whether Li-inlAm would recruit another classical cadherin than mEcad in CT26 cells. We labeled CT26 cells with a pan-cadherin antibody, which recognizes the cytoplasmic domain of classical cadherins [31]. CT26 cells were strongly stained with the pan-cadherin antibody (Figure S1B), indicating that they likely express classical cadherin proteins. Furthermore, this pan-cadherin-immunoreactive protein was recruited in CT26 cells by Li-inlAm but not Li-inlA (Figure S1B). Immunoblotting and immunostaining revealed that CT26 cells express Ncad (Figures 2C and D), a classical cadherin known to be expressed in endothelial cells, neurons and some transformed epithelial cells [20]. Importantly, Li-inlAm, but not Li-inlA, recruited Ncad in CT26 cells (Figure 2D). We next tested other cell lines for Ncad expression. We found that Nme cells (which also express mEcad and are permissive to InlAm-mediated entry), human HeLa cells, and guinea pig 104C1 cells all express Ncad (Figure 2C). As in CT26 cells, InlAm promoted bacterial entry into HeLa and 104C1 cells, although these two cell lines do not express Ecad and are therefore not permissive to InlA-dependent entry (Figure S2) [23]. These results suggest that the murinization of InlA confers to this protein the ability to interact with Ncad from different species, and to enter into host cells expressing Ncad.
To investigate if mNcad serves as a receptor for InlAm-mediated entry into CT26 cells, CT26 cells were treated with mNcad-specific siRNAs or scrambled control siRNAs. Treatment of CT26 cells with mNcad siRNAs led to a reduced expression of mNcad which correlated with a significantly decreased InlAm-dependent entry (Figures 3A and B). To directly assess the ability of mEcad and mNcad to act as receptors for InlAm, we used the BHK21 cell line, which is of hamster origin and does not express any known classical cadherin [32], and transfected this cell line with plasmids encoding either hEcad, mEcad or mNcad. As expected, both InlA and InlAm mediated bacterial entry into hEcad-expressing cells (Figure 3C). Moreover, InlAm mediated entry into mEcad-expressing cells, whereas as previously shown, InlA did not (Figure 3C) [22]. Most importantly, we also demonstrated that InlAm mediated bacterial entry into Ncad-expressing cells, whereas, as previously shown, InlA did not (Figure 3C) [18].
To investigate whether the InlAm receptor repertoire extends to other members of classical cadherins, we tested the ability of mouse P-cadherin (mPcad) and VE-cadherin (mVEcad) to serve as receptors for InlAm (Figure S1A). Neither mPcad nor mVEcad acted as a receptor for InlAm or InlA (Figure 3C). Taken together, these data confirm that InlA exhibits a species-specific and narrow repertoire for Ecad and mediates entry into hEcad- but not mEcad-expressing cells, and demonstrate that by widening InlA species spectrum from human to mouse Ecad, murinization of InlA extends its receptor repertoire to Ncad.
In order to investigate if these in vitro results translate into an in vivo phenotype, and study in particular the cell tropism of InlAm-expressing bacteria, we investigated Ncad luminal accessibility at the intestinal epithelium level, which is the portal of InlA-mediated entry of Lm. In contrast to luminally-accessible Ecad which is mostly observed as rings surrounding goblet cells [25], mNcad was accessible on the apical pole of villous M cells (Figure 4, Movie S1), but not M cells of Peyer's patches (Movie S2) in wt mice. The expression of luminally-accessible Ncad was also detected on the apical pole of villous M cells in E16P KI mice (Figure S3, Movie S3). These results suggest that InlAm may allow bacteria to target villous M cells upon mouse oral inoculation.
To specifically investigate whether InlAm-expressing bacteria target cells that express luminally-accessible Ncad, we inoculated orally wt mice with Li-inlA or Li-inlAm, and for comparison we inoculated humanized E16P KI mice orally with Li-inlA. As expected from our recent results [25], Li-inlA were found in goblet cells 5 hrs post oral inoculation of E16P KI mice (Figures 5C and D). In contrast, Li-inlAm targeted both goblet cells (Figures 5A and D) and villous M cells (Figures 5B and D, Movie S4) in wt mice. We next performed a detailed quantification of the location of bacteria in the intestinal epithelium (i.e. goblet cells, villous M cells, other epithelial cells). This demonstrated that, contrary to InlA, which targets almost exclusively goblet cells in E16P KI mice (82%), InlAm preferentially targets villous M cells (56%) in wt mice, and to a lower degree goblet cells (34%) (p<0.001, χ2 test analysis) (Figure 5D). In agreement with these results obtained with Li-inlAm, Lm-inlAm also targeted both goblet cells (Figures S4A and D, S5A, Movie S5) and villous M cells (Figures S4B and D, S5B, Movie S6) in both wt and E16P KI mice, in contrast to Lm which exclusively targeted goblet cells, only in E16P KI mice (Figures S4C and D, S5C, Movie S7). Together, these results demonstrate that while InlA- and InlAm-Ecad interactions both contribute to the targeting of goblet cells, InlAm-mNcad interaction allows bacteria to target villous M cells, a cell type which is not targeted when InlA interacts only with its native receptor Ecad.
To investigate the impact of InlAm-mNcad interaction on the infection process, we inoculated orally wt and E16P KI mice with Lm-inlAm or Lm. In Lm-infected E16P KI mice in which InlA-Ecad interaction is functional, InlA promoted Lm invasion of the small intestinal tissue and bacterial dissemination to spleen and liver as early as 2 days post infection (dpi) (Figure 6). In contrast, in Lm-inlAm infected wt mice, in which both InlAm-Ecad and InlAm-Ncad interactions are functional, Lm bacterial loads in the small intestinal tissue, spleen and liver were not significantly increased at 2 dpi compared to Lm-infected wt mice, but were at 4 dpi (Figure 6). This delayed systemic dissemination was also observed when comparing Lm-inlAm to LmΔinlA in E16P KI mice (Figure S7). These results demonstrate that, although promoting Lm crossing of the wt mouse intestinal barrier, InlAm delays bacterial systemic dissemination relative to InlA in E16P KI mice, and therefore alters the kinetics of Lm infection in vivo.
Given the changes in infection kinetics induced by InlAm, and the artifactual route of translocation taken by InlAm-expressing bacteria at the intestinal epithelium level, we investigated whether InlAm-Ncad-mediated targeting of villous M cells would have an impact on host responses. Strikingly, oral inoculation of Lm-inlAm led to a significant neutrophil recruitment in wt (Figures 7A and B), E16P KI (Figures S8A and B) and hEcad Tg mice (Figures S8A and B), which was not observed with Lm in E16P KI (Figures 7A and B) and in hEcad Tg mice (Figures S8A and B). Importantly, neutrophil infiltration correlated only with InlAm-mediated invasion, and did not reflect bacterial load in the villi, which was actually the highest in Lm-infected humanized mice, in which no neutrophil infiltration was observed (Figures 7A–C, S8A–C). Moreover, a significant increase in IFN-γ and IL-1β expression was observed in the intestinal tissue of wt mice infected with Lm-inlAm, whereas no significant increase was observed in Lm-infected wt and humanized mice (Figures 7 D and E). Together, these results indicate that InlAm-Ncad-mediated intestinal invasion per se leads to exacerbated host responses compared to InlA-Ecad-mediated intestinal invasion, and are not a reflect of enhanced bacterial tissue invasion.
We next assessed intestinal barrier integrity upon infection by testing the intratissular diffusion of biotin administered intraluminally (see Material and Methods) [33]. In wt and humanized mice infected by Lm for two days, biotin localized exclusively to the luminal side of the small intestine (Figures 7F and S8D). In contrast, although the intestinal villi of Lm-inlAm infected wt and humanized mice were not heavily infected, biotin accessed the lamina propria (Figures 7F and S8D). These findings indicate that InlAm-Ncad-mediated intestinal invasion leads to a disruption of intestinal barrier integrity. Together, these results demonstrate that the murinization of InlA profoundly modifies the pathogenic properties of Lm by altering its intestinal portal of entry, host intestinal responses and intestinal barrier integrity.
InlA interaction with Ecad allows Lm translocation across the intestinal epithelium and is therefore a critical event in the development of systemic listeriosis, one of the deadliest foodborne infections in human. Because InlA does not interact with mEcad, the discovery and characterization of this key step were made in species permissive to InlA-Ecad interaction (guinea pig, gerbil) and humanized mouse models (hEcad Tg and E16P KI mouse lines) [6], [9]. A genetically engineered Lm strain expressing a murinized InlA (InlAm) enabling interaction with mEcad in vitro has been proposed to constitute an attractive alternative model to study human listeriosis in wt mice [16]. A practical advantage of this latter system is that it can be readily used to infect several different mouse lines. However, a systematic study comparing the properties of Lm expressing InlAm to that of its isogenic parental strain has not been performed, neither in vitro nor in vivo.
Here we show that InlAm is able to recruit mEcad and mediate mEcad-dependent entry into cultured cells. We also show that InlAm mediates entry into goblet cells of wt mice, which express luminally-accessible mEcad. These results confirm that the S192N and Y369S substitutions confer to InlA a phenotype in wt mice which is observed in humanized mice permissive to InlA-Ecad interaction [25].
Importantly, we also uncover that InlAm is able to recruit Ncad and mediate Ncad-dependent internalization. This artifactual interaction translates in vivo into InlAm-dependent targeting of villous M cells, intestinal inflammatory responses, disruption of intestinal barrier integrity and delayed bacterial systemic dissemination in wt mice, as well as in humanized mice. Such stricking phenotypes are not observed in humanized mice orally-inoculated with wt Lm, suggesting that they depend on InlAm-Ncad interaction and invasion of villous M cells, but not on InlAm-Ecad interaction and invasion of goblet cells (Figure 8). It is important to note that these phenotypes are also present in E16P KI and hEcad Tg mice infected with Lm-inlAm, indicating that intestinal inflammation is a direct consequence of InlAm-mediated intestinal invasion, and proving that the absence of inflammation in Lm-infected humanized mice is not a side effect of mouse humanization, but is a genuine property of InlA-dependent intestinal invasion. These results are in agreement with the observation by Wollert et al. that infection with Lm-inlAm leads to severe intestinal inflammation and tissue damage in wt mice [16], and with our earlier observation that InlA has little impact on Lm intestinal responses in mice permissive to InlA-Ecad interaction [6], [27]. This indicates that the murinization of InlA, in addition to broadening the host range of Lm, also extends its receptor repertoire to another member of the classical cadherin family, Ncad, therefore modifying its cell tropism, host responses and the dynamics of infection.
The engineering of InlAm was based on the rational protein design of a modified InlA that would increase InlA-hEcad binding affinity [16]. Indeed, S192N and Y369S substitutions in InlA lead to a 6,700-fold increase in the binding affinity of InlA to hEcad [16]. Here we have shown that this does not translate into increased invasion of hEcad-expressing cells. Before drawing this conclusion, we ensured that the BHK21 cell line we used does not express other cadherins than the one we intended to study. A possible reason for the observed increased level of invasion of Lm-inlAm in Caco-2 cells observed by Wollert et al. is the coexpression of Ecad and Ncad in these cells [21]. These results suggest that InlA-hEcad interaction, although it is of relatively low affinity (KD = 8±4 µM) [16], has been naturally selected to mediate an optimal level of infection.
We have shown that InlB, another major invasion protein of Lm, does not play a significant role for the crossing of the intestinal barrier [23]. In contrast, InlB has been reported to promote Lm expressing InlAm to invade intestinal villi [34]. Our results shed light onto these apparent contradictory results and raise the possibilty that InlAm-Ncad mediated invasion of villous M cells may involve the InlB pathway.
Shigella flexneri, the etiological agent of bacillary dysentery is associated with strong polymorphonuclear infiltration, severe local inflammation, disruption of intestinal barrier integrity, yet no systemic dissemination [35], [36]. In contrast, listeriosis in human and humanized mice is characterized by the paucity of intestinal symptoms, the absence of polymorphonuclear intestinal infiltration, little local inflammation, the absence of intestinal barrier disruption, but systemic dissemination [6], [27], [36], [37]. We have demonstrated that Lm-inlAm triggers pro-inflammatory response and disrupts epithelial integrity in intestinal tissue of wt and humanized mice, and exhibits a delayed systemic dissemination, compared to Lm-infected humanized mice. These observations strongly suggest that the targeting of villous M cells by InlAm-expressing bacteria triggers pro-inflammatory host responses which contain bacterial invasion but lead to intestinal epithelium damages. This fits with the observation that antigen delivery via villous M cells stimulates immune reponses [38]. Like InlAm, Als3 is a Candida albicans invasin that binds both Ecad and Ncad to invade host cells [39]. Candida albicans has been shown to favor gut inflammation and promotes food allergy accompanied by gut epithelial barrier hyperpermeability, the underlying mechanisms of which are so far unclear [40], [41]. Our study indicates that Candida albicans may use Als3 to target Ncad-positive villous M cells, and thereby trigger intestinal inflammation. The specific functions of villous M cells remain poorly understood, yet villous M cells are a particularly abundant constituent of the intestinal epithelium. Our results show that InlAm- and Als3-expressing microorganisms would be particularly instrumental to study villous M cell functions.
Repeated infection of mice in vivo or mouse cells in vitro allows the obtention of “murinized” pathogens adapted to the mouse. Despite the great adaptability of microbes, evolutionary constraints limit pathogen variability [42]. A mutation beneficial under certain environmental conditions may end up as disadvantageous in another, highlighting the fine-tuning of host-microbe interactions. The structure-based rational design of InlAm was proposed as a subtle and elegant way to electively “murinize” a microbial ligand with least impact on the pathogen. However, we provide here evidence that the rationally designed InlAm has gained the unfortunate ability to interact with another surface protein than its cognate receptor Ecad. Even though InlAm mediates Lm crossing of the intestinal barrier, a phenotype which is strictly dependent on InlA-Ecad interaction, the way by which Lm crosses the intestinal barrier in an InlAm-dependent manner differs from what observed with wt Lm in humanized mice and humans, as does the resulting infection process. This illustrates that murinization of human-specific pathogens, although an elegant and rational approach, may unfortunately mislead rather than ease the understanding of human infectious diseases' pathophysiology. Caution must therefore be exercised before engineering and using “murinized” pathogens to study human infectious diseases.
Bacterial strains, plasmids and primers are listed in Table S1. Note that the sequences of inlA, inlAm in Lm and in Li were confirmed by sequencing, as well as the integration sites of inlA and inlAm in Li and the deletion site of inlA in Lm. Listeria and Escherichia coli strains were respectively cultivated in BHI and LB at 37°C with shaking at 180 rpm. To deliver plasmids into Li, E. coli S17-1 (colistin and nalidixic acid sensitive) cells were transformed with the plasmids followed by conjugation with Li (colistin and nalidixic acid resistant). Mammalian cell lines used in this study were routinely cultured at 37°C in 5% CO2. Except for the culture medium for BHK21 which was supplemented with 5% fetal bovine serum, all the cell culture media were supplemented with 10% fetal bovine serum. Human epithelium LoVo cells were cultured in F12K nutrient GlutaMax medium. Mouse epithelium Nme cells were cultured in DMEM GlutaMax medium supplemented with 10 µg/ml insulin. Mouse CT26 and guinea pig 104C1 cells were cultured in RPMI1640 GlutaMax medium supplemented with HEPES buffer and sodium pyruvate. Human HeLa cells were cultured in MEM GlutaMax medium. Hamster BHK21 cells were cultured in GMEM GlutaMax medium supplemented with tryptose phosphate buffer and HEPES buffer. All the culture medium and related chemicals were purchased from Gibco (Invitrogen). Transient transfection of mammalian cells was performed with jetPRIME transfection kit (Polyplus transfection). The scrambled (sc-37007) and mouse Ncad specific siRNAs (sc-35999) were purchased from Santa Cruz. For the transfection of siRNAs, mouse CT26 cells were seeded into the 24-well plates for 1 day and then transfected with scrambled siRNAs (25 nM) or mNcad-specific siRNAs (25 nM) followed by 1 day incubation and replacement of transfection medium with growth medium another 1 day of incubation before infection. For the transfection of plasmid DNAs, BHK21 cells were transiently transfected with pcDNA3 expression vector harboring the cDNAs of each cadherin (1 µg DNA for each well in a 24-well plate) followed by 2 days incubation before infection.
The strategy to express inlA or inlAm in Li is as described based on integrative plasmid pAD containing a constitutive promoter [43]. The primers EagI_UTRhly-F and UTRhly-R were used to amplify the hly 5′ UTR of Lm EGDe. Full length of inlA and inlAm were amplified from the genomic DNA of Lm EGDe and Lm-inlAm, respectively, with the primers UTRhly_inlA-F and SalI_inlA-R2. The resulting PCR products were ligated to hly 5′ UTR by splicing-by-overlap-extension (SOE) PCR. The final SOE PCR products, containing the entire hly 5′ UTR sequence fused to the start codon of the inlA (hly 5′ UTR-inlA) or inlAm, (hly 5′ UTR-inlAm), were then cloned in pCR-Blunt (Invitrogen) and verified by sequencing. Plasmids containing correct sequence and pAD-cGFP were digested by EagI and SalI. The backbone of pAD-cGFP was ligated with hly 5′ UTR-inlA and hly 5′ UTR-inlAm to form pAD-inlA and pAD-inlAm.
The mouse N-cadherin (mNcad) cDNA was bought from Open Biosystems (Thermo Scientific) and the cDNAs of mouse P-cadherin (mPcad) and mouse VE-cadherin (mVEcad) were from Riken Fantom Clones (Dnaform). To form pcDNA3-mNcad, mNcad cDNA was cloned into EcoRI-NotI site of pcDNA3. The plasmids pcDNA3-mPcad and pcDNA3-mVEcad were constructed by inserting mPcad and mVEcad cDNAs into NotI-KpnI site of pcDNA3, respectively.
Cell suspensions from confluent monolayers were seeded at a concentration of 5×104 cells per well in 24-well tissue culture plates and grown for 40–48 hr in an antibiotics-free medium at 37°C. Lm and Li strains were grown to OD600 at 0.8 and 0.6 in BHI, respectively. Bacterial culture were then washed with PBS and diluted in cell culture medium without serum. Bacterial suspensions were added to the cells at a multiplicity of infection (MOI) of approximately 50 and incubated for 1 hr. Following wash with complete medium, 10 µg/ml of gentamicin was added to kill the extracellular bacteria for 1 hr. The cells were then washed by complete medium and PBS, and homogenized in PBS supplemented with 0.4% Triton X-100, followed by serial dilution and colony forming units (CFUs) counting. For cadherin recruitment assay, the procedure was the same as the invasion assay except that the cell attachment buffer (HEPES 20 mM, NaCl 150 mM, glucose 50 mM, MgCl2 1 mM, CaCl2 2 mM, MnCl2 1 mM, 0.1% BSA) was used for infection and PBS (Ca2+/Mg2+) (Gibco) was applied to wash the non-attached bacteria stringently followed by fixation.
Eight to 10-week old C57BL/6 female mice (JANVIER) and isogenic mEcad E16P KI female mice were food restricted overnight but allowed free access to water. Lm culture was prepared as described [6], and inoculated with a feeding needle intragastrically [44]. Mice were then immediately allowed free access to food and water. All the procedures were in agreement with the guidelines of the European Commission for the handling of laboratory animals, directive 86/609/EEC (http://ec.europa.eu/environment/chemicals/lab_animals/home_en.htm) and were approved by the Animal Care and Use Committee of the Institut Pasteur, as well as by the ethical committee of “Paris Centre et Sud” under the number 2010-0020.
Preparation of tissue sections and whole mount tissues were as described [9], [25]. The following antibodies and fluorescent probes were used for immunostaining and Western blot: anti-hEcad clone HECD-1 mouse monoclonal antibody (Invitrogen), anti-mEcad clone ECCD-2 rat monoclonal antibody (Invitrogen), anti-β-actin clone AC-15 mouse monoclonal antibody (Sigma), anti-Ncad clone 32/N-cadherin mouse monoclonal antibody (BD), anti-Ncad clone GC-4 mouse monoclonal antibody (Sigma), anti-pan cadherin clone CH-19 monoclonal antibody (Sigma), anti-M cell clone NKM 16-2-4 rat monoclonal antibody (Miltenyl Biotec), R6 anti-Li rabbit polyclonal antibody and R11 anti-Lm rabbit polyclonal antibody [45], Rat anti-mouse Ly-6G (BD), wheat germ agglutinin (WGA) conjugated with Alexa Fluor 647 (Jackson ImmunoResearch), Alexa Fuor 488 goat anti-rabbit (Invitrogen), Alexa Fluor 488 or Alexa Fluor 546 goat anti-mouse (Invitrogen), Alexa Fluor 647 donkey anti-rat (Jackson ImmunoResearch), Alexa Fluor 546 goat anti-rat (Invitrogen), Cy3-conjugated streptavidin (Jackson ImmunoResearch) and Hoechst 33342 (Invitrogen).
Biotin was used as a molecule to address the integrity of intestinal epithelium as described previously [33]. Briefly, 2 mg/ml of EZ-link Sulfo-NHS-Biotin (Pierce) in PBS was slowly injected into the lumen of ileum loop via the open end adjacent to cecum immediatedly after removal of the entire ileum. After 3 min, the loop was opened followed by PBS wash and 4% paraformaldehye fixation.
Four mice for each condition were sacrificed 2 days post infection. 1 cm-long of ileal loop of each animal was applied for RNA extraction. The RNA isolation, reverse transcription and quantitative real time PCR (qRT-PCR) were performed as described [46]. Primers used for qRT-PCR were pre-designed, validated RT2 qPCR primer pairs (SABioSciences, Qiagen) as follows: IFNG (IFN-γ, PPM03121A), IL1B (IL-1β, PPM03109F) and GAPDH (PPM02946E).
Values are expressed as mean + SD. Statistical comparisons were made using the unpaired Student's t test, Mann-Whitney u test or the χ2 test as indicated. p values<0.05 were considered significant. Significant differences are marked with an asterisk for p<0.05, two asterisks for p<0.01, three asterisks for p<0.001 and four asterisks for p<0.0001.
|
10.1371/journal.pntd.0002875 | High Parasitological Failure Rate of Visceral Leishmaniasis to Sodium Stibogluconate among HIV Co-infected Adults in Ethiopia | Antimonials are still being used for visceral leishmaniasis (VL) treatment among HIV co-infected patients in East-Africa due to the shortage of alternative safer drugs like liposomal amphotericin B. Besides tolerability, emergence of resistance to antimonials is a major concern.
This study was aimed at assessing the clinical outcome of VL-HIV co-infected patients when treated with sodium stibogluconate (SSG).
Retrospective patient record analysis of VL-HIV co-infected patients treated at a clinical trial site in north-west Ethiopia was done. Patients with parasitologically confirmed VL and HIV co-infection treated with SSG were included. The dose of SSG used was 20 mg Sb5 (pentavalent antimony)/kg and maximum of 850 mg Sb5 for 30 days. The clinical outcomes were defined based on the tissue aspiration results as cure or failure, and additionally the safety and mortality rates were computed.
The study included 57 patients treated with SSG and by the end of treatment only 43.9% of patients were cured. The parasitological treatment failure and the case fatality rate were 31.6% and 14.0% respectively. SSG was discontinued temporarily or permanently for 12 (21.1%) cases due to safety issues. High baseline parasite load (graded more than 4+) was significantly associated with treatment failure (odds ratio = 8.9, 95% confidence interval = .5-51.7).
SSG is not only unsafe, but also has low effectiveness for VL-HIV patients. Safe and effective alternative medications are very urgently needed. Drug sensitivity surveillance should be introduced in the region.
| The co-infection of VL and HIV is a very challenging clinical problem especially in the East-Africa region. Though liposomal amphotericin B is the recommended treatment option for VL-HIV co-infection, it is often not available in practice in Ethiopia. Thus several patients are still being treated with antimonials that are infamous for their toxicity. In this study, we describe the results of such antimonial treatment in a series of 57 patients. The effectiveness of antimonials was found to be low and, in comparison to previous studies, declining. There is an urgent need to assure the availability of safer and more effective alternative medications for VL-HIV. It seems also wise to start a surveillance scheme for drug susceptibility in leishmania parasites as our results may relate to emerging antimonial resistance in the East-Africa region.
| Visceral leishmaniasis (VL), also called kala-azar, is a vector-borne disseminated infection caused by the Leishmania donovani spp. complex, a protozoan parasite that predominantly affects tissue macrophages. Overt disease is lethal without treatment. The zoonotic form, with dogs as the main reservoir, is caused by Leishmania infantum and is found mainly in the Mediterranean basin and Latin America. The anthroponotic form is caused by Leishmania donovani and is prevalent in the Indian subcontinent (with estimated 300,000 cases/year) and East Africa (30,000 cases/year). Within East Africa, Sudan and Ethiopia are most affected [1]. Like most other neglected tropical diseases, VL has traditionally ranked low on the national and international health agenda [2]–[5].
HIV has been identified as one of the emerging challenges for VL control. HIV infection dramatically increases the risk of progression from asymptomatic Leishmania infection to VL and VL accelerates HIV disease progression [6], [7]. At the global level, the highest burden of VL-HIV co-infection is found in north-west Ethiopia, where up to 40% of VL patients can be co-infected with HIV [8].
Treatment of HIV co-infected individuals poses particular challenges with poor treatment response and recurrent relapse [9]–[12]. For over four decades, antimonials have been in use and still are the mainstay of therapy in East-Africa for immunocompetent individuals either alone or in combination with paromomycin [13]. Several studies demonstrated more than 90% effectiveness among VL patients without HIV [11], [14]–[17]. However, following reports of high rates of toxicity of antimonials in HIV-infected individuals [18], liposomal amphotericin B has been recommended as the preferential first line treatment for this patient group [13], [19]. Although the cost of liposomal amphotericin B has undergone substantial reductions for low income countries, there remain issues of availability of this drug in the public sector in several resource-constrained settings in East-Africa [20], [21]. In Ethiopia, the country with the highest co-infection rate globally, liposomal amphotericin B is in short supply. As a result, clinicians often still have to resort to antimonials for treating co-infected individuals, reserving the available liposomal amphotericin B for the most complicated or severe cases.
Besides poor tolerability, emergence of drug resistance is a major concern with antimonials, as has been witnessed in Bihar, India [22]. Whereas initial studies on antimonials reported relatively low rates of treatment failure in HIV-infected individuals in East-Africa [17], more recent retrospective cohort studies conducted in Ethiopia seemed to indicate relatively high failure rates [11], [14]. However, these reports came from settings where parasitological confirmation of diagnosis and treatment response was not systematically performed. Moreover, with the roll-out of Anti-Retroviral Therapy (ART) in East-Africa and better survival of HIV-infected patients, the VL patient population has gradually changed, with an increasing proportion of HIV patients presenting with VL relapse [23]. Data on efficacy and tolerability of antimonials in this patient population are very scarce.
Hereby, we report on the clinical outcomes in a cohort of adult VL-HIV co-infected patients recruited at the Leishmaniasis Research and Treatment Center (LRTC) of University of Gondar (UoG) Hospital, Ethiopia.
Ethical approval was obtained for this study from the institutional review board of the University of Gondar. All data were analyzed anonymized.
The study was conducted at the LRTC in the UoG that is located in north-west Ethiopia close to the main VL endemic focus in the country. The LRTC was founded by the Drugs for Neglected Diseases initiative (DNDi), and is now part of the governmental health system and is used as a clinical trial site. In addition to VL research, the center provides free VL treatment and care to all patients with leishmaniasis. Patients present to the center either spontaneously or are referred from other health institutes in the catchment area. Several international aid agencies including the World Health Organization (WHO), Médecins Sans Frontières (MSF) and DNDi support VL treatment and care within Ethiopia, but shortage of anti-leishmanial drugs remains a frequent problem in clinical practice. Pentavalent antimonials and liposomal forms of amphotericin B, and more recently paromomycin, are the main drugs to treat VL. The combination regimen SSG and paromomycin was started as first line therapy for non-HIV VL cases at the site in September 2012. While liposomal amphotericin B is recommended for VL in HIV patients, due to the inadequate supply, it is often reserved for more severe cases such as patients with organ dysfunction. Miltefosine is infrequently available. As a result, most HIV co-infected VL patients are being treated with SSG.
In November 2011 the LRTC initiated a clinical trial on the use of secondary prophylaxis to prevent relapse in HIV co-infected VL patients that is currently ongoing (http://clinicaltrials.gov/show/NCT01360762). All VL-HIV co-infected patients presenting at LRTC are screened for enrolment in this clinical trial and their findings and initial treatment responses are documented in individual patient record files. We reviewed these patient records and considered those who were initially started on treatment with antimonials between November 2011 and January 2013 as eligible for the present retrospective cohort study on safety and effectiveness of antimonials.
A VL diagnosis was suspected based on the WHO clinical case definition: prolonged fever, weight loss and splenomegaly in a patient from an endemic area or with a travel history. In patients who met this case definition we confirmed the VL diagnosis by microscopic detection of the parasite in tissue aspirates (spleen, bone marrow, lymph node). The parasite load was graded as 6+: >100 parasites per field; 5+: 10–100 parasites per field; 4+: 1–10 parasites per field; 3+: 1–10 parasites per 10 fields; 2+: 1–10 parasites per 100 fields; 1+: 1–10 parasites per 1000 fields and 0: 0 parasites per 1000 fields as viewed with a 100x oil-immersion lens [13].
Sodium stibogluconate (SSG) (at a dose of 20 mg Sb5 (pentavalent antimony)/kg for 30 days) was used for the initial treatment of these patients. The maximum SSG dose used was 850 mg Sb5/day, to avoid the toxicity related to high doses in HIV infected patients [24], [25].
All patients were admitted to the LRTC and monitored for their treatment response using clinical and laboratory parameters. The main clinical parameters monitored during treatment were fever, appetite, fatigue, weight and spleen size. In addition, the patients were checked daily for new complaints especially related to adverse reactions. Blood chemistry and hematology tests were done weekly and a test of cure (ToC) by tissue aspiration and microscopic evaluation for parasites was done systematically at the end of SSG treatment (day 30). Electrocardiography was done based on the presence of cardiac symptoms. Treatment was extended for patients with a positive ToC at the end of SSG treatment. Those who i. showed a clinical response, ii. had a parasite load reduction of two log or more (“slow responders”) and iii. a parasite grade below 4 received a treatment extension with the same drug, while those who had less than two logs parasite reduction (“non-responders”) or a parasite grade of 4 or above were treated with an alternative drug (most commonly liposomal amphotericin B with or without miltefosine). While this was the principle followed, the shortage of medications often affected the practice. Depending on the parasite grade (high or low), ToC was repeated on day 15 or day 30 of the treatment extension with further extension of treatment when ToC was positive.
At LRTC provider-initiated testing and counselling for HIV is offered to all hospitalized patients. The HIV diagnosis is based on the national algorithm with two serial positive rapid test results; KHB (Shanghai Kehua Bio-engineering Co-Ltd, Shanghai, China) followed by STAT-PAK™ (Chembio HIV1/2, Medford, New York, USA). In case of discrepancy between the two tests, Uni-Gold™ (Trinity Biotech PLC, Bray, Ireland) is used as a tie breaker. As VL is considered a stage IV-defining illness in HIV patients [13], [19], all patients are given ART as soon as they were stabilized from their acute illnesses. ART regimens follow the national guidelines: tenofovir-lamivudine-efavirenz; zidovudine-lamivudine-efavirenz; or zidovudine-lamivudine-nevirapine [26]. Second-line ART consists of protease inhibitor-based combination regimens.
Initial cure is defined as a negative ToC at the end of the standard treatment [13]. A positive ToC indicates treatment failure, which consists of two types: slow-response or non-response. Clinical improvement (resolution of symptoms and signs such as fever, improvement in appetite, weight gain, gaining physical strength and regression of the spleen) with a reduction in parasite load by more than two logs without complete parasite clearance is defined as slow-response. If the parasite load remains the same or if the reduction is less than two logs, this is defined as non-response. SSG discontinuation as an outcome refers to permanent discontinuation of SSG due to intolerance.
Besides the initial treatment outcomes (at the end of the standard 30 days SSG treatment), treatment outcomes were also ascertained at the end of the entire treatment course (end of treatment outcomes), integrating eventual SSG extensions, or treatment changes for reasons of SSG intolerance or initial treatment failure.
Adverse reactions that required discontinuation (temporary or permanent) of the treatment were considered serious adverse drug reactions and included in the analysis. Reasons for interruption included clinical or biochemical pancreatitis; acute renal injury; or other severe conditions that were considered probably related to SSG like cardiotoxicity or worsening of bone marrow function. Clinical pancreatitis was defined as persistent vomiting, abdominal pain and raised serum amylase levels. Biochemical pancreatitis was defined as asymptomatic grade 4 serum amylase levels (using the Common Terminology Criteria for Adverse Events (CTCAE v 4.0) system [27]. Acute renal injury requiring SSG interruption was defined as acute increase of serum creatinine to more than 2 mg/dl (normal values 0.6 to 1.1 mg/dl). While some patients with toxicity improved after a temporary interruption (less than five days) and were able to continue with SSG, those with poor/slow recovery were shifted to other treatment regimens.
We used a structured data collection format to extract information from the individual patient chart records. Socio-demographic data, weight, episode of leishmania, CD4 count, hematology and blood chemistry values and treatment outcome data were recorded. Count data were summarized as frequency (%), and numerical variables as median values with Inter-Quartile Range (IQR). To assess associations we computed Crude Odds Ratios (OR) with a 95% Confidence Interval (CI). For individuals with either initial cure or failure, the medians of within patient changes during treatment in laboratory and clinical parameters were compared using the Wilcoxon rank-sum test. P-values<0.05 were considered statistically significant. Statistical analysis was done using STATA 11 software.
From a total of 84 HIV co-infected VL patients treated at LRTC during the study period, 57 received SSG as first line regimen. The others were treated with liposomal amphotericin B (6 patients), liposomal amphotericin B plus miltefosine (7 patients) or the SSG and paromomycin (14 patients) combination regimen, see Figure 1.
Of the 57 patients treated with SSG, 56 (98.2%) were male; the median age was 32 (IQR 28–36) years. While 33 (57.9%) were admitted for primary VL, 24 (42.1%) were relapses, and out of them 12 (50%) had more than two previous VL episodes. The majority were malnourished, with huge spleen and anemia (Table 1).
Most patients (n = 36; 63.2%) were already on ART at the time of VL diagnosis. The median CD4 count was 61 (IQR 35–101) cells/µL, and 39 (73.6%) of them had a CD4 count less than 100 cells/µl. Most patients (n = 48; 90.6%) had a CD4 count less than 200 cells/µl. The leishmania parasite load in tissue aspirates was graded as +4 or above for 41/56 (73.2%) individuals (Table 1).
Amongst the 57 patients starting antimonial treatment, SSG was permanently discontinued due to intolerance in five (8.8%) cases. In addition, seven (12.3%) patients temporarily interrupted SSG (for a maximum of five days) due to adverse reactions but re-instituted after stabilization. At the end of the standard SSG treatment course, 25 (43.9%) achieved initial cure while eight (14.0%) patients died during SSG therapy. One patient left the hospital against medical advice. There were 18 (31.6%) cases with initial treatment failure (positive ToC). Ten of these were slow responders while the eight others were non-responders.
For the 18 cases with parasitologically confirmed SSG treatment failure, treatment was extended beyond one month. A repeat course of SSG was used for 14 of these patients (for an additional 15 to 90 days). The other 4 were given liposomal amphotericin B (n = 2) or a liposomal amphotericin B and miltefosine combination (n = 2). For the five cases that permanently discontinued SSG due to intolerance, liposomal amphotericin B was given and four of them eventually got cured. At the global assessment made at the end of drug treatment, 43 of the 57 (75.4%) patients had achieved parasite clearance and were considered cured, three (5.3%) had treatment failure and nine (15.8%) died. The outcome was unknown for two (3.5%) individuals.
Serious adverse reactions were observed in 14 of the 57 patients (12 in the first month and 2 more during extension of treatment) mainly due to pancreatitis (n = 3), renal failure (n = 3) or both (n = 6). Cardiotoxicity and the combination of hepatitis and bone marrow suppression were observed in individual cases. Out of the nine deaths, five occurred after developing both pancreatitis and renal failure, and one after acute renal failure. The other three deaths were related to the presence of additional co-morbidities (malnutrition, sepsis and tuberculosis).
Table 2 shows the key clinical and laboratory parameters in relation with treatment response. While tissue parasite grading decreased more significantly in cured patients, there was a more pronounced increase in hematocrit and a greater reduction in spleen size in failure cases.
Although VL relapse cases and those on ART were more likely to fail, only high tissue parasite grading was significantly associated with treatment failure (OR = 8.9, CI = 1.5–51.7), Table 3 and Figure 2). Individuals with low CD4 counts (<100 cells/µl), malnourished patients (BMI <18.5 kg/m2), patients presenting with primary VL, patients having a large spleen size (≥10 cm) on admission and a higher tissue parasite load (≥4) were more likely to die, but these differences did not reach statistical difference (Table 3).
This study showed high toxicity and very low effectiveness of SSG when used as a primary treatment option for VL-HIV co-infection. A 30-days drug course of SSG led to treatment failure in close to one out of three patients, an assessment based on parasitological data. SSG-related toxicity was common, probably contributing to death in six cases (i.e. 75% of the deaths). These findings indicate the urgent need of wider availability of alternative and safer drugs such as liposomal amphotericin B for this specific population.
Previous studies from the region have reported initial failure rates of SSG in the range of 2.3% to 14.1% [11], [14], [16]. Several factors might have contributed to the higher frequency observed in this study. In contrast to most other studies, proportionally more patients on ART and/or (multiple) VL relapse cases, identified as risk factors for treatment failure, were included [23]. High baseline tissue parasite grade – the strongest risk factor in our study – was also relatively common. Third, the parasitological response was not systematically assessed in the other studies, which could have led to under-diagnosis of failure. Moreover, a maximum dose of 850 mg SSG was used in this treatment center for this group of patients due to the experiences of exacerbated toxicity with higher doses in other regions with a similar patient population [25]. Whether parasite drug resistance, as observed in India [22], played a role is currently unclear. Although the change in parasite grading was lower in failing patients compared to cured ones, the difference was not statistically significant. Drug sensitivity testing of the isolated parasites is currently underway.
Safety of SSG has always been a concern. However, limited studies addressed the experience in the treatment setting concerning the adverse effects of SSG in the region. Vomiting has been reported as a common adverse event in most of the studies ranging from 8% to 38% [14], [16], [17], though the underlying cause of the vomiting is not clearly addressed. A quarter of the patients in this study were suffering from adverse reactions that required either temporary or permanent discontinuation of the SSG. Pancreatitis and renal failure were found to be the serious adverse reactions that were probably the main causes of death necessitating the need to close monitoring during treatment. This implies that SSG should only be used in a setting where monitoring of renal functions and pancreatitis is possible. As electrocardiographic monitoring was not systematically done, the cardiotoxicity might be under estimated.
The case fatality rate during treatment (14%) is in the previously observed range, (6.8%–33.3%) [11], [28]. The risk factors for mortality identified in our study are in line with the previous observations.
The strengths of this study were that it was conducted in a dedicated leishmania treatment and research center where tissue aspiration is routinely performed in HIV co-infection, and standardized treatment protocols and data collection tools are in place. Limitations include the relatively small sample size, and the missing information for some laboratory tests. Moreover, HIV-1 viral load testing and electrolytes measurements could not be conducted.
The small sample size in this study did not allow for a thorough study of risk factors for treatment failure, and did preclude any control for confounding in multivariable analysis. We observed that patients with high parasite load tended to respond slowly or not. Extension of treatment (using antimonials or liposomal amphotericin B) beyond 30 days helped to increase the cure rate from 44% to 75%. Patients who received prolonged duration of treatment seem to have tolerated SSG. These few patients were also with better clinical situations (higher hemoglobin and better spleen regression) which can be a source of bias. High SSG toxicity was observed with the daily maximal dose limited to 850 mg. It should be noted that higher dose and prolonged therapy could be at the expense of increasing toxicity. Higher daily dose of antimonials should be discouraged for HIV co-infected patients given the high toxicity observed at the current dose. Case-by-case decisions on dose and duration of therapy need to take into consideration the parasite load at the time of diagnosis and safety issues.
The small sample size in this study did not allow for a thorough study of risk factors for treatment failure, and did preclude any control for confounding in multivariable analysis. We observed that patients with high parasite load tended to respond slowly or not. Extension of treatment (using antimonials or liposomal amphotericin B) beyond 30 days helped to increase the cure rate from 44% to 75% with minimal additional toxicity. Given the small sample size, the safety of extending SSG in individuals tolerating the first month of SSG remains to be confirmed. On the other hand, high SSG toxicity was observed with the daily maximal dose limited to 850 mg. Consequently, a higher daily dose of antimonials should in general be discouraged for HIV co-infected patients. Case-by-case decisions on dose and duration of therapy need to take into consideration the parasite load at the time of diagnosis and safety issues.
Patients with VL relapse and those developing VL while on ART tended to fail SSG treatment. With further expansion and access to ART in this region, such difficult to treat patients might gradually become more prevalent. Additional studies focusing on treatment in this patient group should be conducted. Despite ART, these patients had profound immune deficiency as seen from their CD4 cell level. Additionally, most patients failed SSG treatment while on ART showing the need for additional treatment strategies. Early screening and treatment or primary prophylaxis are possible options yet to be explored. The reason behind the better response in hematocrit and spleen regression in the treatment failure group may be related to the longer treatment period. Patients with better hematocrit may survive longer than the severely anemic ones, and face repeated relapses.
This study re-confirms that SSG is not safe in patients with HIV co-infection and additionally shows that its effectiveness is low - and potentially declining - in the region [29]. However, despite the recommendations in the national and international guidelines, SSG remains in use due to shortage of alternative safe and effective medications. Even with the price reductions and donation programs of liposomal amphotericin B, there is still a limited and irregular supply in Ethiopia. We call on all stakeholders to urgently take measures to ensure a stable access to liposomal amphotericin B for all high risk groups – including HIV co-infected patients – in which treatment with antimonials leads to unacceptable high rates of toxicity and/or failure. A clinical trial evaluating high dose liposomal amphotericin B or a combination with miltefosine in HIV co-infection is about to be initiated in Ethiopia. If found effective, ensuring the availability of these drugs should be a priority.
This study provides evidence of high treatment failure (failure of parasite clearance) and toxicity in HIV co-infected VL patients treated with SSG. Assuring availability of safer and more efficacious treatment options for HIV co-infected VL patients in this region needs urgent attention. Furthermore, the need for drug resistance surveillance is called upon.
|
10.1371/journal.pntd.0000915 | Sex, Subdivision, and Domestic Dispersal of Trypanosoma cruzi Lineage I in Southern Ecuador | Molecular epidemiology at the community level has an important guiding role in zoonotic disease control programmes where genetic markers are suitably variable to unravel the dynamics of local transmission. We evaluated the molecular diversity of Trypanosoma cruzi, the etiological agent of Chagas disease, in southern Ecuador (Loja Province). This kinetoplastid parasite has traditionally been a paradigm for clonal population structure in pathogenic organisms. However, the presence of naturally occurring hybrids, mitochondrial introgression, and evidence of genetic exchange in the laboratory question this dogma.
Eighty-one parasite isolates from domiciliary, peridomiciliary, and sylvatic triatomines and mammals were genotyped across 10 variable microsatellite loci. Two discrete parasite populations were defined: one predominantly composed of isolates from domestic and peridomestic foci, and another predominantly composed of isolates from sylvatic foci. Spatial genetic variation was absent from the former, suggesting rapid parasite dispersal across our study area. Furthermore, linkage equilibrium between loci, Hardy-Weinberg allele frequencies at individual loci, and a lack of repeated genotypes are indicative of frequent genetic exchange among individuals in the domestic/peridomestic population.
These data represent novel population-level evidence of an extant capacity for sex among natural cycles of T. cruzi transmission. As such they have dramatic implications for our understanding of the fundamental genetics of this parasite. Our data also elucidate local disease transmission, whereby passive anthropogenic domestic mammal and triatomine dispersal across our study area is likely to account for the rapid domestic/peridomestic spread of the parasite. Finally we discuss how this, and the observed subdivision between sympatric sylvatic and domestic/peridomestic foci, can inform efforts at Chagas disease control in Ecuador.
| Trypanosoma cruzi is transmitted by blood sucking insects known as triatomines. This protozoan parasite commonly infects wild and domestic mammals in South and Central America. However, triatomines also transmit the parasite to people, and human infection with T. cruzi is known as Chagas disease, a major public health concern in Latin America. Understanding the complex dynamics of parasite spread between wild and domestic environments is essential to design effective control measures to prevent the spread of Chagas disease. Here we describe T. cruzi genetic diversity and population dynamics in southern Ecuador. Our findings indicate that the parasite circulates in two largely independent cycles: one corresponding to the sylvatic environment and one related to the domestic/peridomestic environment. Furthermore, our data indicate that human activity might promote parasite dispersal among communties. This information is the key for the design of control programmes in Southern Ecuador. Finally, we have encountered evidence of a sexual reproductive mode in the domestic T. cruzi population, which constitutes a new and intriguing finding with regards to the biology of this parasite.
| Chagas disease, caused by the protozoan Trypanosoma cruzi, is the most important parasitic infection in Latin America [1]. An estimated 10 million people carry the infection, while another 90 million live at risk [2]. This vector-borne zoonosis causes severely debilitating and potentially deadly disease in more than a third of infected people [3]. Mucosal or abrasion contact with the infected faeces of hematophagous triatomine bugs constitutes the major mode of transmission [2].
Chagas disease is endemic to several regions in Ecuador, including the warm inter-Andean valleys of the southern province of Loja, where the main vectors are Rhodnius ecuadoriensis, Triatoma carrioni, Panstrongylus chinai, and Panstrongylus rufotuberculatus [4], [5]. Loja Province is currently targeted by the Ecuadorian Chagas Disease Control Program. Complementing disease prevention efforts, recent progress has been made in understanding local vector dynamics [5]–[7]. However, parasite molecular epidemiology could also play a role in guiding effective intervention measures.
Molecular diversity was first recognised in T. cruzi in the early 1970s [8]. Six major genetic subdivisions, known as discrete typing units (DTUs), are currently recognized (TcI–TcVI [9]), with distributions loosely defined by geography, transmission cycle, and ecology [1]. TcI predominates in northern South America, causes significant human disease [10], [11] and occurs in both domestic and sylvatic cycles of parasite transmission. Of major interest to those planning sustainable control strategies in this region is the extent to which these cycles are connected [12]–[14]. The provision of such data relies on the evaluation of molecular diversity ‘hidden’ at the sub-DTU level [15]–[17].
Hypervariable molecular markers, like microsatellites, have given new and unprecedented insight into the population genetics of other important parasitic zoonoses [18]–[22]. For the first time, specific hypotheses regarding parasite dispersal and reproduction can be addressed. However, the validity of molecular epidemiological data depends heavily on study design. Numerous confounders, including biased sampling (e.g., sampling only one host in a heteroxenous transmission system [23]), population subdivision in both space and time (leading to Wahlund effects [24]), and low sample size all influence the estimation of key population genetic parameters. Historically, such biases have acted as an impediment to obtaining useful epidemiological information from parasite molecular data, and, particularly in T. cruzi, to resolving the frequency of sex in natural populations.
Here we present microsatellite data for 10 variable loci amplified from a large number of TcI isolates collected from domestic, peridomestic, and sylvatic hosts and vectors in and around several adjacent communities in Loja Province, Ecuador. We evaluate evidence for genetic subdivision between transmission cycles, anthropogenic dispersal of parasites between communities, and panmixia among a subset of strains.
Sixteen communities in Loja Province, southern Ecuador, were sampled (Figure 1). These communities were located at altitudes less than 2,200 m and were representative of the ecological diversity of the province. Trypanosomes were isolated from triatomines and small mammals (rodents and opossums) captured at domestic (within dwellings), peridomestic (near dwellings and/or associated with human activities, e.g., crop stores, chicken roosts, wood and rock piles), and sylvatic (more than 20 meters from dwellings) foci (Table S1). Written informed consents from the head of the houses for domiciliary bug searches and capture of mammals near houses were obtained. These documents have been approved by the institutional review board from National Institute of Health (NIH), Ohio University (OU) and Pontifical Catholic University of Ecuador (PUCE). Vertebrates were euthanized to obtain samples; all procedures were carried out in strict accordance with the protocol approved by the Ohio University Institutional Animal Care and Use Committee (IACUC). The Ohio University IACUC adheres to the guidelines in the United States Government Code of Federal Regulations (CFR), Title 9, Chapter 1, Subchapter A- Animal Welfare Parts 1–3 and the United States Health Research Extension Act of 1985, Public Law 99–158 “Animals in Research”.
Trypanosome species was determined by PCR amplification of the kinetoplast minicircle region as in Vallejo et al. [25]. Discrete Typing Units (DTU) genotyping was achieved by assaying a combination of three nuclear loci as described by Lewis et al. [26].
Ten previously identified polymorphic microsatellite loci were studied (Table S2) [16]. These loci are distributed across seven T. cruzi chromosomes and include two groups of physically linked markers [27]. Allelic products were amplified using previously described reaction conditions [16]. Allele sizes were determined using an automated capillary sequencer (AB3730, Applied Biosystems, UK) in conjunction with a fluorescently tagged size standard and were manually checked for errors. All isolates were typed “blind” to control for user bias. By reference to a representative panel of strains, no cross reactivity was identified between T. rangeli and the microsatellite primers used in this study. Population-level genetic diversity was assessed using sample size corrected allelic richness (Ar) in FSTAT 2.9.3.2 [28] and number of private (population specific) alleles per locus (PA). FIS, a measure of the distribution of heterozygosity within and between individuals, was estimated per locus per population in FSTAT 2.9.3.2. FIS can vary between −1 (all loci heterozygous for the same alleles) and +1 (all loci homozygous for different alleles). FIS = 0 indicates Hardy-Weinberg allele proportions. The extent of population subdivision between isolates from different transmission cycles was estimated using (FST) in ARLEQUIN v3.1 and statistical significance assessed via 10,000 random permutations of alleles between populations [29]. Similarly, within-population subdivision was examined in ARLEQUIN v3.1, in this case using a hierarchal Analysis of Molecular Variance (AMOVA). Population-level heterozygosity indices were also calculated in ARLEQUIN v3.1 and associated significance levels for p values derived after sequential Bonferroni correction to minimise the likelihood of Type 1 errors [30]. Individual-level pair-wise distances were estimated using DAS (1-proportion of shared alleles at all loci / n) [31] under an IAM and δμ2 [32] under an SMM in MICROSAT [33]. DAS values form the basis of the dendrogram in Figure 2. To accommodate multi-allelic loci, a script was written in Microsoft Visual Basic to make multiple random diploid re-samplings of each multilocus profile (software available on request). Individual-level genetic distances were calculated as the mean across multiple re-sampled datasets. A Mantel's test for the effect of isolation by distance within populations (pair-wise genetic vs. geographic distance) was implemented in Genelax 6 using 10,000 random permutations [34]. Linkage disequilibrium indices, pair-wise (RGGD) and multilocus (IA), were calculated in LINKDOS [35] and MULTILOCUS1.3b [36], respectively. Multiple diploid re-samplings were also made to evaluate the influence of multi-allelic loci on IA, the results of which are shown in Table 1. Assignment of individuals to populations was made by reference to the topology of the DAS derived tree. Secondarily, this model-free population assignment was corroborated using STRUCTURE (Figure S1) [37]. Sample affiliations are listed in Table S1.
Eighty-one isolates of T. cruzi were obtained from triatomines and mammals. All were genotyped as TcI. Kinetoplast analysis detected the presence of mixed infection with T. rangeli in nine isolates from the sylvatic environment (Table S1). In total, a dataset of 1,637 alleles was derived across all loci, excluding missing data (Table S3). Multiple (≥3) alleles were observed at 3.08% of loci.
We evaluated patterns of clustering and subdivision among parasite strains in the Loja samples based upon their microsatellite profiles. To identify genetically distinct groups we relied on three lines of evidence: neighbor-joining analysis based on pair-wise genetic distance; model-based population assignment (STRUCTURE); and the statistical significance of the fixation index FST.The deepest and most robust (56.5%) internal branching within the neighbor-joining tree constructed from pair-wise genetic distance values (DAS) supported the delineation of two populations (Figure 2 and Table 1). No pattern or diversification by host or vector was observed within these populations. The observed bipartite subdivision was unaffected by the presence of multi-allelic loci (100% congruence, Figure 2) and was used as a means to define the populations examined in later analyses (See Table 1). Sample allocation between these two populations was exactly corroborated by the optimal number of clusters (k) derived using STRUCTURE software as defined by Evanno et al. [37] by Δk (Figure S1). One population, henceforth called LOJADom/Peri, was predominantly composed of isolates from domestic and peridomestic foci, the other, henceforth LOJASylv, of isolates from the sylvatic environment. Estimates of genetic subdivision (FST) between a priori populations (transmission cycle defined) corroborated this pattern of dispersal. No evidence for subdivision existed between domestic and peridomestic isolates (FST = 0.027, p = 0.354), whereas subdivision between these populations (grouped) and sylvatic samples was pronounced (FST = 0.212, p<0.0001). Naturally, reassignment of outliers to their “correct” genetic groups according to neighbor-joining and STRUCTURE analyses further inflated the latter estimate (FST LOJADom/Peri−LOJASylv = 0.397, p<0.0001). These outliers are evidence for some, albeit limited, parasite dispersal between domestic/peridomestic transmission cycles and sylvatic transmission cycles as evident in Figure 1 and 2.
Following the identification of two genetically distinct groups of parasite strains circulating in this endemic area, the genetic diversity of each was evaluated and compared. Estimates of allelic richness (Ar) did not demonstrate dramatic difference between LOJADom/Peri and LOJASylv (Table 1); both populations showed considerable genetic diversity. More private alleles per locus (PA) were found in the larger and marginally more diverse sylvatic population (PA = 2.0; Table 1). In conjunction with its apparent genetic distance from other South American TcI populations (Figure 2), the lack of private alleles within LOJADom/Peri (PA = 0.8) suggests diversification of this population from a local source.
In light of the role played by transmission cycles in structuring the local parasite population, we compared the rate of parasite dispersal within LOJADom/Peri with that within LOJASilv. This rate is inversely proportional to the amount of spatial structure (or isolation by distance (IBD)) in the population. Interestingly, tests for IBD among individuals from LOJADom/Peri and LOJASylv showed statistically significant and epidemiologically important differences between these two populations. Infinite allele models (IAMs) of microsatellite mutation intrinsically overestimate genetic distances between closely related isolates as compared to stepwise mutational models (SMMs). To circumvent possible bias we chose to test for IBD under both. Strong evidence for spatial structure in LOJASylv existed regardless (DAS−RXY = 0.265, P<0.0001; δμ2−RXY = 0.177, p = 0.001). Among isolates from LOJADom/Peri, no spatial structure was evident from either measure (DAS−RXY = 0.100, p = 0.164; δμ2−RXY = −0.05, p = 0.384). Results are summarised in Figure 3 and strongly suggest more rapid parasite dispersal among domestic and peridomestic foci than that occurring between sylvatic locales at the same spatial scale.
Several approaches were employed to estimate the rate of genetic recombination within the parasite populations identified in Loja. Multiple indicators suggested frequent sex among trypanosomes of LOJADom/Peri. Pair-wise inter-locus linkage (RGGD) was infrequent (5.5%; Table 1) even among physically linked loci (3/4 physically linked locus pairs, those on the same chromosome, were not statistically linked) and despite abundant allelic diversity available within this population for inter-correlation (the statistical power of RGGD drops dramatically with decreasing population-level genetic diversity). Infrequent pair-wise linkage is consistent with the lack of significance attributable to the index of association (IA) (median p = 0.13, P≥0.05 in 89% of 1000 resampled diploid datasets; Table 1), and with the null hypothesis of random mating that must be assumed. Additionally, tests for deficit or excess of heterozygosity in this population showed no significant deviation from Hardy-Weinberg expectations, reflected by mean values for the inbreeding coefficient (FIS) across loci that approximate zero (Table 1). Finally, repeated multilocus genotypes, indicative of clonal reproduction, were absent from this population while present in LOJASylv. Other aspects of LOJASylv diversity pointed to predominant clonality, especially strong pair-wise (38.5% of locus pairs) and multilocus linkage (IA P<0.001) in all diploid resampled datasets (Table 1), but also strong deviation from Hardy-Weinberg levels of heterozygosity under all three measures employed (Table 1). Consistent with spatial structure identified in this population, however, an AMOVA conducted across isolates from San Jacinto and Bramaderos, which make up the majority of LOJASylv strains (Figure 1 and Table S1), did demonstrate significant but weak FST (FST = 0.173, P<0.0001, 16,000 permutations), evidence that a Wahlund effect could be depressing heterozygosity. Correspondingly, estimates of linkage disequilibrium might also be somewhat inflated by subdivision in this population [38], and it is difficult to reject the possibility that recombination may occur in the sylvatic populations at a micro-geographic scale.
This study constitutes a first attempt to understand the population dynamics of T. cruzi at a local scale using high-resolution molecular markers. The sample includes isolates from different transmission cycles, vectors, hosts, and adjacent communities. This arrangement aims to minimise sample bias and maximise the resulting molecular epidemiological inference. However, all field studies are affected by the natural abundance of hosts and vectors in different transmission cycles, and we cannot claim a perfect dataset. Nonetheless, we can report strong evidence for parasite diversification by transmission cycle, human involvement in parasite dispersal, and the possibility of sex in one parasite population.
The presence of the T cruzi lineage I in southern Ecuador is consistent with reports of this DTU throughout northern South America [10], [39], [40]. In our study, as in other studies, sub-DTU level diversity of the parasite occurred independently of vector and host [16], [17], [41]. Instead, we found evidence that transmission cycle (domestic, peridomestic, or sylvatic) is likely to be the major driver behind parasite differentiation, apparently a phenomenon common to T. cruzi populations across much of northern South America [15], [16] but never before studied on a local scale. On the basis of our data, we suggest that widespread, internationally distributed TcI subgroups associated with specific transmission cycles may not exist. A lack of connectivity between LOJADom/Peri and domestic TcI from Venezuela, VENDom, (Figure 2) exemplifies this. Furthermore, clear cross-propagation of parasites between transmission cycles (Figure 2) and few private alleles in LOJADom/Peri (Table 1) suggests that these domestic groups are likely to emerge and diversify from local sylvatic sources.
T. cruzi is the only stercorarian trypanosome of medical importance [42]. Natural transmission efficiency by this route (contamination with vector feces) is very low. The rate of transmission from infected Triatoma infestans to humans in Argentina, for example, is estimated at approximately one in 650 bites [43]. As with R. prolixus in Venezuela [14], R. ecuadoriensis, a major disease vector in Loja, occurs at high frequency in both domestic and sylvatic locales [7]. Our data suggest that even if vector invasion from sylvatic foci is common, as in Venezuela [14], associated transmission of parasites to domestic foci is too infrequent to break up population subdivision. Where cross-propagation does occur, circumstantial evidence incriminates synanthropic mammals as the link between transmission cycles. Didelphis marsupialis infected with parasites from the LOJADom/Peri group were found at both peridomestic (Isolate Numbers (IN) 9 and 13, Figure 2) and sylvatic locales (IN 6 and 17, Figure 2). Furthermore, a R. rattus individual captured at a peridomestic site was found infected with a LOJASylv strain (IN 31, Figure 2). Finally P. chinai and T. carrioni adults and nymphs, so far thought to be exclusively domestic and peridomestic triatomine species in Loja (IN 27,28,58,68 and 81, Figure 2) [5], were found infected with a LOJASylv strain, likely as a result of contact with invasive sylvatic mammals. This blurring of the lines between transmission cycles is likely to mirror local environmental change, where human activity is driving land-use transformation.
Parasite sampling in Loja was undertaken across an area only 50 km in radius (Figure 1). However, this area encompassed several ecological zones punctuated by high mountains (>2,500 m in elevation) and deep interconnecting valleys. Spatial genetic diversification among sylvatic isolates is an expected outcome given such barriers to host and vector migration (Figure 1 and 3). Conversely, parasites belonging to the LOJADom/Peri group lack this signature, a finding possibly linked to rapid anthropogenic dispersal in the form of infected individuals, livestock, or passively transported vectors and/or small peridomestic mammals.
T. cruzi has, until recently, been a paradigm for clonal population structure in pathogenic organisms [44], [45]. However, the presence of naturally occurring hybrids [46], mitochondrial introgression [46], and a capacity for genetic exchange in the laboratory [47] has challenged this dogma. The frequent observation of linkage disequilibrium in T. cruzi may partially stem from cryptic population subdivision (temporal, spatial, and/or genetic) to which linkage statistics are intrinsically sensitive [38]. Frustratingly, if assignment software with intrinsic Hardy-Weinberg assumptions (e.g., STRUCTURE [48] or BAPS [49]) is used to account for subdivision prior to linkage analysis, the resulting populations will be sorted to maximise adherence to Hardy-Weinberg allelic frequencies, with artifactual sexuality a possible result [21], [50]. Fortunately, in our study, the status of LOJADom/Peri as a stable deme is corroborated by distance-based, model-free assignment, as well as STRUCTURE. In conjunction with Hardy-Weinberg allele frequencies at individual loci, we consider, therefore, that linkage equilibrium among isolates from LOJADom/Peri represents strong evidence for frequent genetic exchange among field isolates of T. cruzi. We believe that the relatively small sample size of LOJADom/Peri (n = 18) does not affect this conclusion, partly due to the ample genetic diversity present in this popualtion, but also because the lack of spatial subdivision in this group suggests frequent contact and opportunity for mixis. Thus it constitutes exactly the group of strains between which genetic exchange might be expected to occur. We cannot rule out the possibility that genetic exchange may also occur in the sylvatic cycle, if the role that substructure found in LOJASylv played in inflating linkage statistics IA and RGGD could be taken into account. However, more focused high-density sample collection from multiple individual localities would be required to address such a hypothesis. Furthermore, we cannot infer the cellular mechanism of genetic recombination events on the basis of these data. Hardy-Weinberg allelic allelic frequencies are consistent with classical meiosis. However, the lack of haploid life stages so far observed in T. cruzi are not consistent with classical meiosis, nor are the genetic exchange events so far observed in vitro [47].
Molecular epidemiology at this scale has an important guiding role to play in Chagas disease control programmes. Future efforts in Loja province must account for inter-domiciliary and inter-community parasite dispersal. This includes sustained surveillance and coordinated region-wide spraying campaigns to eliminate local vector re-invasion sources and community education to target passive triatomine dispersal routes. It is also clear that the role of synanthropic mammals cannot be overlooked as these represent an important potential link between sylvatic and domestic foci.
We have shown that microsatellite markers, adequate sample sizes, and associated population statistics give fundamental insight into the genetic exchange in T. cruzi. Our results, skewed toward samples from the vector, intuitively imply that the vector may be a site of genetic exchange, as is the case for T. brucei [51] and Leishmania major [52]. The data also indicate, not surprisingly, that the majority of events probably occur within a T. cruzi lineage between epidemiologically linked strains, and these events have therefore historically been difficult to detect. The intriguing mechanisms of genetic exchange in T. cruzi warrant further investigation of their functional, adaptive, and epidemiological significance.
|
10.1371/journal.pgen.1002043 | A Missense Mutation in PPARD Causes a Major QTL Effect on Ear Size in Pigs | Chinese Erhualian is the most prolific pig breed in the world. The breed exhibits exceptionally large and floppy ears. To identify genes underlying this typical feature, we previously performed a genome scan in a large scale White Duroc × Erhualian cross and mapped a major QTL for ear size to a 2-cM region on chromosome 7. We herein performed an identical-by-descent analysis that defined the QTL within a 750-kb region. Historically, the large-ear feature has been selected for the ancient sacrificial culture in Erhualian pigs. By using a selective sweep analysis, we then refined the critical region to a 630-kb interval containing 9 annotated genes. Four of the 9 genes are expressed in ear tissues of piglets. Of the 4 genes, PPARD stood out as the strongest candidate gene for its established role in skin homeostasis, cartilage development, and fat metabolism. No differential expression of PPARD was found in ear tissues at different growth stages between large-eared Erhualian and small-eared Duroc pigs. We further screened coding sequence variants in the PPARD gene and identified only one missense mutation (G32E) in a conserved functionally important domain. The protein-altering mutation showed perfect concordance (100%) with the QTL genotypes of all 19 founder animals segregating in the White Duroc × Erhualian cross and occurred at high frequencies exclusively in Chinese large-eared breeds. Moreover, the mutation is of functional significance; it mediates down-regulation of β-catenin and its target gene expression that is crucial for fat deposition in skin. Furthermore, the mutation was significantly associated with ear size across the experimental cross and diverse outbred populations. A worldwide survey of haplotype diversity revealed that the mutation event is of Chinese origin, likely after domestication. Taken together, we provide evidence that PPARD G32E is the variation underlying this major QTL.
| A central but challenging objective in current biology is to dissect the genetic basis of quantitative traits. Numerous quantitative trait loci (QTL) have been uncovered in model and farm animals, providing unexpected insights into the biology of complex traits. However, only a few causal variants underlying the QTL have been explicitly identified. By using a battery of genetic and functional assays, we herein show that a major QTL effect on pig ear size is most likely caused by a single base substitution in an evolutionary conserved region of the PPARD gene. The protein-altered mutation is of functional significance and explains a proportion of variation in ear size across diverse pig breeds. A worldwide survey showed that the mutant allele for increased ear size was derived from a common ancestor in Chinese pigs, likely after domestication. These findings establish, for the first time, an essential role of PPARD in ear development and highlight the great potential of naturally occurring mutations in farm animals to gain insights into mammalian biology. Moreover, the knowledge of the PPARD causal mutation adds to the limited list of quantitative trait genes and quantitative trait nucleotides characterized in domesticated animals.
| The external ear is part of the auditory system and plays a vital role in collecting sound as the first step in hearing. Multiple congenital anomalies have been documented for human external ears. For instance, microtia, characterized by a small and abnormally shaped outer ear, occurs in approximately one in 8,000–10,000 births. However, only in a minority of cases has a genetic or environmental cause been found [1]. The domestic pig services as not only an agriculturally important animal for meat production but also an important large-animal model for human medicine [2]. Thousands of years of selective breeding has created diversity of phenotypes in pigs, such as ear size in Erhualian and White Duroc breeds. Erhualian is the most prolific pig breed and exhibits unusually large and floppy ears as breed character (Figure 1). Historically, the large-ear feature of Erhualian pigs had been favored by owners for the traditional sacrificial culture [3]. White Duroc is one of worldwide-popular boar line and has small and erect ears (Figure 1). We have created a four-generation White Duroc × Erhualian resource population, in which phenotypic traits related to ear size have been recorded in 1,027 adult F2 animals and 560 adult F3 individuals (Table S1). We mapped a major QTL for ear size around 58 cM on SSC7 (Figure S1) using a genome scan on the White Duroc × Erhualian cross [4], which confirmed the previously reported QTL affecting ear size in a Large White × Meishan F2 resource population [5]. The significant QTL had a small confidence interval of 2 cM and explained more than 40% of phenotypic variance. The aim of this study was to identify the genetic determinant underlying this major QTL.
To fine map the QTL, we genotyped 1,027 adult F2 animals and their 68 parents and 19 grandparents in the White Duroc × Erhualian cross using additional 17 SNP markers and 11 microsatellite markers in the QTL region. A final set of 33 markers covering the QTL region were then explored to deduce the QTL genotypes of F1 sires by the marker-assisted segregation analysis as proposed previously [6]. We determined QTL genotypes of all 9 F1 sires (Figure S2). All 9 Q-bearing chromosomes for increased ear size shared a haplotype of ∼1.2 Mb flanked by markers HMGA1 – TULP1. The shared haplotype was distinct from q-bearing chromosomes (Figure 2). These observations strongly suggest that the QTL is located in the 1.2-Mb interval.
Given the extremely divergent ear size phenotypes between Erhualian and White Duroc animals, we assumed that Q and q alleles were alternatively fixed in Erhualian and White Duroc founder animals; hence all Erhualian founder sows could share a chromosomal segment carrying the Q allele for increased ear size. To test this assumption, we reconstructed haplotypes of all 19 founder animals (2 sires and 17 dams) using 50 markers (15 microsatellites and 35 SNPs) in the QTL region. Almost all Erhualian founder sows shared a haplotype of ∼750 kb within the refined 1.2-Mb interval (Figure 2). As predicted, this shared haplotype was associated with increased ear size and presumably Q-bearing chromosomes. Two Erhualian founder sows carried a distinct haplotype (denoted as Eq), which was unexpected because it was contrast with our initial assumption. We then conducted a statistical analysis of F2 animals in the White Duroc × Erhualian cross. The results revealed that the Eq chromosome had an effect on decreased ear size similar to the White Duroc chromosome (Dq) and significantly different from the Erhualian Q-bearing chromosome (EQ). The least-squares means (± s.e.) of ear weight were 323.07±4.55 for EQEQ and 266.66±18.9 for EQEq (P = 0.04); 264.71±3.52 for DqEQ and 236.98±17.12 for DqEq (P = 0.06, Table 1). The shared EQ chromosome allowed us to refine the location of the major QTL to the 750-kb interval between markers UHRF1BP1 and TULP1 (Figure 2).
Historically, Erhualian pigs had undergone selection for ear size because pigs with extraordinary large and floppy ears were favored for the ancient sacrificial culture in the Taihu region of East China [3]. Reduced genetic variation in the critical region containing the QTL was therefore predicted. To define the region of reduced genetic variation, we collected 211 animals representing all lineages in 3 Erhualian nucleus populations, 216 animals from 6 Chinese indigenous breeds and 119 independent animals from 3 Western worldwide-popular commercial breeds. Using these samples, we genotyped 6 microsatellite and 32 SNP markers in the 750-kb region. We found that 18 adjacent markers in a 630-kb region between markers UHRF1BP1 and FANCE showed dramatically reduced polymorphisms in all Erhualian pigs with nearly all major allele frequencies of more than 0.90. Notably, the 18 markers in the 630-kb region are monomorphic in the Erhualian nucleus population from Xishan county (n = 72). In comparison, the genetic polymorphisms of these markers were maintained in other Chinese, Western breeds, and wild boars (Figure 3). The 630-kb region showing strong selective-sweep effects on Erhualian pigs was therefore predicted to contain the responsible locus. We further genotyped the 18 markers in the 630-kb region on 188 adult animals of Sutai pigs. This breed was developed after 18-genereation selection from a Duroc (50%) × Erhualian (50%) cross in 1986 [7], meaning that the breed has undergone 18 generations of meiosis reducing the extent of linkage disequilibrium between QTL and linked markers. The Erhualian-originated haplotype of 630 kb showed significant (P = 0.009) association with increased ear size compared with other chromosomes in Sutai pigs (Figure S3), thereby supporting the conclusion that this region harbors the causative gene.
The 630-kb region encompasses 9 annotated genes (ANKS1A, DEF6, FANCE, PPARD, SCUBE3, TAF11, TCP11, UHR1BP1 and ZNF76) in the human homologous region. RT-PCR was performed to detect expression levels of these genes in ear tissues of piglets. Four genes including PPARD, FANCE, TAF11 and ZNF76 were highly expressed, whereas transcripts of other genes were almost absent in ear tissues (data not shown). Of the 4 genes, PPARD (peroxisome proliferator-activated receptor delta) is a ligand-modulated transcription factor belonging to the nuclear receptor superfamily and plays crucial roles in diverse biologically important processes [8]. For instance, PPARD play a pivotal role in modulating cell differentiation in both keratinocytes and sebocyte of skin [9]. PPARD also serves as a key regulator in fat metabolism; it triggers fat burning and enhances energy uncoupling in adipose tissues and skeletal muscle [10]–[12]. Moreover, PPARD is a key player in Wnt/β-catenin pathway [13], which has essential roles in diverse cellular activities including chondrocyte proliferation and differentiation [14]. The external ear is composed of skin, cartilage, connective tissues and fat. Given its crucial role in skin homeostasis, cartilage development and fat metabolism, PPARD stood out as a prime positional candidate for the major QTL. We monitored the relative mRNA expression of PPARD in ear tissues of Erhualian and Duroc pigs at four different ages by real-time RT-PCR. The expression levels were higher in samples at early ages (days 0, 45 and 90) compared with adult samples (day 300). However, no significant difference of expression levels was found in ear tissues between large-eared Erhualian and small-eared Duroc pigs (Figure S4).
To search for causative mutations, we first sequenced the entire coding region of the PPARD gene using ear mRNA of two White Duroc and two Erhualian animals and identified only one nonsynonymous mutation. The G to A mutation caused a glycine to glutamic acid substitution at codon 32 (GU565977) in the conserved intrinsically disordered domain of the PPARD protein predicted by SMART (http://smart.embl-heidelberg.de/). The intrinsically disordered domain is a distinctive and common characteristic of eukaryotic hub proteins like multifunctional nuclear receptors and serves as a determinant of protein interactivity [15]. Comparison of amino acids of this protein domain across mammals revealed that glycine is well conserved in mammalian PPARDs (Figure 4), while the derived glutamic acid occurs only in alleles increasing ear size in pigs. We thus speculated that the nonconservative substitution probably changes the PPARD interactivity with other protein partners and consequently affects the gene's regulation function. Genotypes of F1 sires (9 heterozygotes) and F0 animals (17 homozygotes and 2 heterozygotes) at the mutation site were 100% concordance with their QTL genotypes. The potentially altered function and QTL concordance of PPARD G32E corresponded to the hypothesis that this SNP may be the causative mutation underlying the major QTL.
PPARD is involved in the Wnt/β-catenin signaling pathway that regulates diverse cellular functions. In the nucleus, PPARD interacts with β-catenin binding to TCF/LEF transcription factors that stimulate transcription of target genes important for multiple cellular activities including cartilage development and organogenesis [14]. To demonstrate functional significance of PPARD G32E, we cotransfected the 293T cells with the lentiviral expression vectors of wild-type or mutant PPARD and a TCF/LEF-driven luciferase reporter construct. A Renilla luciferase expression vector was used for the normalization of transfection efficiency. Overexpression of mutant PPARD led to a 40% decrease (P<0.05 compared with the wild-type treatment) in TCF/LEF reporter activity (Figure 5A), indicating the G32E mutation mediates down-regulation of β-catenin downstream genes. To examine a direct functional role of PPARD G32E in target genes of β-catenin, we treated pig ear-derived primary fibroblast cells with the lentiviral PPRAD expression vectors and monitored the mRNA levels of β-catenin and its known downstream (c-myc) [16] and upstream (Sox9) [17] genes along with GAPDH as a loading control by real time quantitative RT-PCR. The mRNA levels of β-catenin and c-myc were reduced respectively by 4.1-fold and 11.5-fold (P<0.001) in mutant PPARD transfectants compared with the cells transfected with wild-type PPARD. Western blot analysis showed that both β-catenin and c-myc protein levels were decreased by the mutant PPARD treatment (Figure 5B), thereby confirming the results of mRNA and luciferase reporter analyses. Sox9 mRNA expression in mutant PPARD transfectants was only slightly decreased to 1.1-fold of the wild-type PPARD treatment; the result was validated by Western blot (Figure 5B). GAPDH was used as a protein loading control for total cell lysate, which was not affected by both wild-type and mutant PPARD treatments (Figure 5B). Altogether, we conclude that PPARD G32E is a functional variant that mediate down-regulation of β-catenin and its target gene expression in the Wnt/β-catenin signaling pathway. Wnt/β-catenin signaling has been firmly demonstrated to suppress adipogensis [18]–[19]. The fact that PPARD is a key modulator of lipid production in the skin [9] and that PPARD G32E inhibits β-catenin expression led us to assume that the mutation stimulates lipid production and storage that are required for enlarged ear size.
To confirm the effect of PPARD G32E on ear size, we performed a standard association test, a marker-assisted association test and an F-drop test [20] in the White Duroc × Erhualian cross. The SNP showed greatly significant (P<0.0001) association with ear weight and ear size in the standard association test. In the marker-assisted association test, the SNP was more significant (P<0.001) for these traits compared with the QTL effect. After fitting this polymorphism in the QTL model, the great QTL effect disappeared with F-value drop rations of less than 0.03 (Table S2). These results were in agreement with the hypothesis that the SNP is the causative mutation for the major QTL affecting ear size. Nevertheless, we cautioned the results because variants closely linked with a causative mutation also lead to strong association in F2 resource populations due to the high level of linkage disequilibrium between founder breeds [20].
To obtain additional supporting evidence, we further genotyped the G32E mutation on 667 mature pigs from 4 Chinese local breeds (Erhualian, Hang, Yushan Black and Bama Xiang) and 3 synthetic commercial lines (Sutai, Suzhong, Sujiang) with phenotypic data of ear size. These populations show a wide range of ear size and segregate for the mutation. The association analyses confirmed the effect of PPARD G32E on ear size. The 32E allele was significantly associated with increased ear size across the tested breeds (P<0.05; Table 2). Chinese local pig breeds have low levels of linkage disequilibrium extending up to only 0.05 cM [21]. The concordantly significant association across Chinese breeds thereby strengthened the hypothesis that PPARD G32E is the responsible locus for ear size. The effects of PPARD G32E differ in their magnitude in the tested breeds; one reason is that the effects are context-dependent and are influenced by different genetic backgrounds and environments. Another possibility is that PPARD G32E is only responsible for part of the effect on ear size in Erhualian pigs.
To reveal the ancestral state and allele frequency of PPARD G32E in diverse pig breeds, we genotyped the mutation in a panel of 1,166 animals representing 31 domestic breeds and Chinese and European wild boars. Overall, the derived 32E allele for increased ear size occurred at high frequencies (>0.80) in Chinese breeds with large and floppy ears. In contrast, the 32G allele for normal ear size was fixed in all wild boars, European local and commercial breeds, and occurred at low frequencies (<0.30) in Chinese indigenous breeds having small and erect ears. These results indicated that PPARD G32E may occur in Chinese pigs after domestication. We detected only one heterozygote in European local breeds (Table 3). The animal was from Large Black pigs that exhibit large and floppy ears and have been influenced by Chinese breeds brought into England in the late 1800′s [22].
We further analyzed the genetic variability and haplotype structure around the G32E mutation in a worldwide pig panel. A total of 868 animals representing 34 breeds were genotyped for 32 SNPs in a 77-kb region of PPARD. Again, the Erhualian breed showed a selective sweep signal as it had negative classical selection statistics Tajimas D and much smaller nucleotide variability (πN) compared with other Chinese local breeds and Western commercial breeds (Table S3). Especially, the genetic variability at the 32 loci was wiped out in the Erhualian population from Xishan. Moreover, we plotted a distribution of the frequency of the derived 32E allele (PA) against Tajimas D index to elucidate the existence of directional selection for the G32E mutation. When PA = 0, Tajimas D was highly variable across breeds, likely due to demographic and/or sampling effects. In stark contrast, Tajima's D took highly negative values when PA >0.8 in Erhualian and other Chinese large-eared breeds as expected in a classical directional selection (Figure S5). We reconstructed 16 major haplotypes with frequencies larger than 0.01 from the 32 SNPs genotyped. Of the 16 haplotypes, only one carried the derived 32E allele; it was at high frequencies in Erhualian pigs and intermediate frequencies in some floppy-eared Chinese breeds whereas absent in Western pigs and wild boars (Table 4). The NJ phylogenetic tree illustrated that the typical haplotype of Erhualian pigs was generally divergent from other haplotypes (Figure S6). These observations supported the assumption that the G32E mutation has a unique origin in Chinese breeds likely after domestication and has undergone selection in Erhualian pigs. We calculated linkage disequilibrium measures (r2) between all pairs of loci and inferred haplotype blocks. Three and two haplotype blocks were identified in the PPARD region for Chinese indigenous pigs and Western commercial breeds, respectively. Only a single nevertheless larger block that spanned 53 kb and contained the G32E SNP was found in Erhualian pigs, reflecting a selection hitching effect (Figure S7). The G32E SNP was in high disequilibrium with very few of the SNPs analyzed (two with r2>0.8), and there was no observable trend between physical distance and disequilibrium measures for the G32E SNP and the rest of loci (Figure S8).
The elucidation of the genetic basis of multifactorial traits in domestic animals is still a big challenge, and few successful examples have been reported [23]–[27]. In this study, a battery of genetic and functional assays obtained diverse pieces of supporting evidence that the PPARD G32E substitution underlies the major QTL effect on ear size on SSC7. (1) The shared haloptypes of 9 F1 sires segregating for the QTL spanned a region of ∼1.2 Mb containing PPARD. (2) All Erhualian founder chromosomes shared a ∼750 kb segment spanning PPARD that were associated with the Q allele for increased ear size. (3) Erhualian pigs showed an obvious selective sweep signal in a 630-kb region encompassing PPARD; the signal was concordant with the breeding history of the breed. (4) The 630-kb haplotype showed similar QTL effect on increased ear size in Sutai pigs that were developed after 18-generation selection in the Erhualian × Duroc cross. (5) Of the 4 genes expressed in ear tissues within the critical region, PPARD stood out a prime candidate for its established essential roles in skin homeostasis, cartilage development and fat metabolism. (6) Only one missense mutation (G32E) was identified in PPARD using White Duroc and Erhualian founder animals. The mutation caused a nonconservative amino acid change at the conserved intrinsically disordered domain and was of functional significance. (7) The G32E SNP was concordant with QTL genotypes of F0 and F1 animals in the White Duroc × Erhualian cross. (8) The G32E SNP showed strikingly significant association with ear size across the experimental cross and diverse outbred populations. (9) The derived allele for increased ear size occurred at high frequencies only in Chinese floppy-eared breeds. Altogether, these data led us to conclude that G32E in the PPARD gene has an important contribution to ear size in pigs. The results establish, for the first time, a direct and novel role of PPARD in ear development and may be of relevance for the pathogenesis of external ear abnormalities in humans.
The genomic region harboring PPARD G32E is of great interest in pig genetics, because significant QTL for diverse traits related to growth, carcass length, skeletal morphology and fat deposition have been consistently evidenced in the region using the current resource population and different crosses between Chinese Meishan and commercial breeds [28]–[33]. The overlapping QTL for multiple traits in the region led us to assume that there might be a single critical gene having pleiotrophic effects on these traits. We herein showed the causality of PPARD G32E for the QTL affecting ear size in the critical region. Given that PPARD serve as a crucial and multifaceted determinant of diverse biological functions including fat metabolism, cartilage development, chondrocyte proliferation and differentiation [8], [10]–[14], we thus speculate that PPARD is a strong candidate of the multiple significant QTL on SSC7 and that PPARD G32E might have pleiotropic effects on growth, carcass and fatness traits in pigs. Further investigations will be performed to validate the hypothesis in the future.
All animal work was conducted according to the guidelines for the care and use of experimental animals established by the Ministry of Agriculture of China.
Microsatellite markers in the mapped interval were mined from the pig genome assembly (Build 9.2) at http://www.ensembl.org/Sus_scrofa/Info/Index and were genotyped using standard procedures. Primers for amplification of microsatellite markers are given in Table S4. QTL genotypes of F1 boars in the White Duroc × Erhualian intercross were determined by marker-assisted segregation analysis as described previously [6]. Briefly, a Z-score was calculated for each F1 sire; the score is the log10 of the H1/H0 likelihood ratio where H1 assumes that the boar is heterozygous at the QTL (Qq), while H0 postulates that the boar is homozygous QQ or qq. Boars were considered to be Qq when Z >2, QQ or qq when Z <−2, and of undetermined genotype if −2<Z<2. The pedigree and management of the intercross population with phenotypic data of ear size have been described elsewhere [4]. Haplotypes of founder animals were reconstructed with the SimWalk2 program.
To detect the effects of a putative selection sweep on the genetic variation in Erhualian pigs compared with control animals, we analyzed the microsatellite and SNP genotypes of 211 Erhualian pigs and 335 control animals representing 10 different breeds (Hetao Large-Ear: 56; Laiwu: 32; Yushan Black: 31; Wuzhishan, 32; Dianan Small-Ear: 31; Tibetan: 34; White Duroc: 12; Duroc: 29; Large White: 39; Landrace: 39). SNP markers were genotyped using the ABI SNapshot protocol or PCR-RFLP assays. All primers are given in Table S4.
Total RNA was extracted from pig tissues using the Rneasy Fibrous Tissue Mini Kit (Qiagen). To analyze expression of candidate genes in ears, products from the first strand-complementary DNA synthesis (TaKaRa) were amplified with primers given in Table S5. The quantification of the PPARD transcripts was performed by the comparative Ct method (2−ΔΔCt) using the primers and TaqMan probes shown in Table S5. Real-time PCR was done with the Universal PCR Master Mix using an ABI7900 instrument (Applied Biosystem). All samples were analyzed in triplicate. The β-actin gene was used as the internal reference gene.
The entire coding region of porcine PPRAD was re-sequenced using ear mRNA of two White Duroc and two Erhualian animals. Primer pairs listed in Table S6 were used to generate overlapping PCR amplicons. All PCR products were purified using the NucleoSpin Extract II kit (Macherey-Nagel) and sequenced using the same primers. The sequence traces were assembled and analyzed for polymorphisms using the SeqMan program (DNASTAR). The PPARD G32E mutation was genotyped using the ABI SNapshot protocol. A 385-bp DNA fragment was amplified with the F2/R2 primer pairs (F2: 5′-CGG CTG TTT TAC AGG AAG GA-3′; R2: 5′- CTG CAC TCA GAC CCA GAT GA-3′). SNapshot reactions were performed with Multiplex Ready Reaction Mix (Applied Biosystem) and an extension primer (5′-TTT TTT TTT TGC TGG AGG GAA GCG AGT GCT CTG GT -3′) using an ABI 3130XL DNA Analyzer (Applied Biosystem).
The coding region of pig PPARD was amplified with primers PPARD-Age-I-F (5′- GAG GAT CCC CGG GTA CCG GTC GCC ACC ATG GAG CAG CCG CCG GAG-3′) and PPARD-Age-I-R (5′- TCA TCC TTG TAG TCG CTA GCG TAC ATG TCC TTG TAG-3′). The amplified cDNA was gel-purified and digested with AgeI and NheI (NEB). The restricted fragments were cloned to pGC-FU-EGFP-3FLAG lentiviral expression vector (Genechem). The sequence and orientation of the insert were verified by DNA sequencing. The expression of His-tagged PPARD in cultured cells was confirmed by Western blot analysis with anti-His antibody. The human 293T cells were infected with the lentiviral expression constructs of pig wild-type and mutant PPARD. The infected cells were seeded at a concentration achieving 80% confluence in 96-well plates 18 h before transfection. The cells were transiently transfected with TCF/LEF-Luc reporter vector (Cignal, SAB) along with a control Relina luciferase vector using Lipofectamine plus reagent. The cell lysates were obtained with 1× reporter lysis buffer (Promega) 48 h after transfection. The luciferase activity was assayed in a Berthold Auto Lumat LB953 luminometer (Nashua, NH) by using the luciferase assay system from Promega. The relative luciferase activity was normalized to the Relina luciferase activity in each sample.
The pig ear-derived fibroblast cells were transfected with pGC-FU-EGFP-3FLAG lentiviral expression vector (Genechem). Five days post-transfection, 1×106 cells were harvested for qPCR and Western blot analysis. Total RNA was extracted from harvested cells using Trizol (Invitrogen). Two µg of total RNA was synthesized into cDNA with M-MLV reverse transcriptase (Promega) and oligo d(T). Real time PCR was performed on the cDNA using the SYBR Premix Ex Taq (TaKaRa) and primers listed in Table S7 in a TP800 Real Time System (TaKaRa). The quantification of transcripts was performed by the comparative Ct (2−ΔΔCt) method. All values were reported as mean ± S.D. of triplicate assays of each cDNA sample. Rabbit anti-PPARD (Sigma), mouse anti-β-catenin (Abcam), rabbit anti-c-myc (Cellsignaling), mouse anti-Sox9 (Abcam) and mouse anti-GAPDH (Santa Cruz) antibodies were used in Western blots in a routine way. The specific immunoreactive bands were visualized using an ECL plus kit (GE Healthcare) and quantified with the Molecular Imaging Software (Kodak).
The entire White Duroc × Erhualian resource population was genotyped for the PPARD G32E mutation. Association of the mutation with ear size and weight was evaluated using standard association, marker-assisted association and F-drop test as described previously [20]. Association analyses were also performed on 667 animals representing 7 different breeds. Photographs were taken for one ear of each animal after the ear was fixed and covered with a ruler as an internal reference of the size. Ear size was calculated using the Qwin software (Laica). Significance was evaluated by the t-test in the GLM procedure of SAS 9.0.
Genomic DNA pools of White Duroc (n = 2) and Erhualian (n = 2) animals were amplified with primers given in Table S6. All PCR products were purified with the Qiagen protocol and sequenced using the same PCR primers, revealing a subset of SNP markers in the genomic region of porcine PPARD. SNP markers were genotyped by iPLEX SEQUENOM MassARRAY platform. SNP genotype calls were filtered and checked manually, and aggressive calls were omitted from the dataset. Population genetics parameters including the mean number of pairwise differences across loci (πN), Tajimas D, Fu and Li's D were estimated with DnaSP v5 [34]. Haplotypes were reconstructed with PHASE v2 [35]. Haplotype phylogenetic tree based on p-distance were drawn using MEGA4 [36]. The Haploview v4.1 program [37] was used to calculate linkage disequilibrium measures (r2 and D') and to identify haplotype blocks.
|
10.1371/journal.ppat.1006398 | FAS-associated factor-1 positively regulates type I interferon response to RNA virus infection by targeting NLRX1 | FAS-associated factor-1 (FAF1) is a component of the death-inducing signaling complex involved in Fas-mediated apoptosis. It regulates NF-κB activity, ubiquitination, and proteasomal degradation. Here, we found that FAF1 positively regulates the type I interferon pathway. FAF1gt/gt mice, which deficient in FAF1, and FAF1 knockdown immune cells were highly susceptible to RNA virus infection and showed low levels of inflammatory cytokines and type I interferon (IFN) production. FAF1 was bound competitively to NLRX1 and positively regulated type I IFN signaling by interfering with the interaction between NLRX1 and MAVS, thereby freeing MAVS to bind RIG-I, which switched on the MAVS-RIG-I-mediated antiviral signaling cascade. These results highlight a critical role of FAF1 in antiviral responses against RNA virus infection.
| Type I interferon-mediated antiviral response is critical for controlling virus infections. However, interferon-mediated immune responses need to be tightly regulated to maintain host immune homeostasis. Recently, molecules involved in regulating interferon-mediated innate immune response are the subject of much research. Among these, the first protein to be identified as a negative regulator of MAVS was the nucleotide-binding domain and leucine-rich repeat containing family member, NLRX1. NLRX1 associates with MAVS to inhibit antiviral signaling by interrupting virus-induced RLR-MAVS interactions. Interestingly, we found that FAF1 interacts with NLRX1 in response to RNA virus infection and this interaction inhibits binding of MAVS to NLRX1, which in turn switches on RIG-I mediated antiviral immune responses. As results, we showed that FAF1gt/gt mice, which deficient in FAF1, and FAF1 knockdown immune cells were highly susceptible to RNA virus infection and showed low levels of inflammatory cytokines and type I interferon (IFN) production. Our findings suggest that FAF1 is a crucial regulator that induces the antiviral innate immune responses against RNA virus infection.
| FAS-associated factor 1 (FAF1) was originally identified as a member of the FAS death-inducing signaling complex [1]. FAF1 harbors several protein interaction domains, including FAS-interacting domains (FID), a death effector domain-interacting domain (DEDID), and multi-ubiquitin-related domains, which interact with ubiquitinated target proteins and regulate their proteolysis [2]. Although FAF1 initially demonstrated to have Fas induced apoptotic potential [3], it also has diverse biological functions such as regulation of NF-κB signaling, chaperone activity and proteosomal degradation by ubiquitination. [2,4–7].
Early recognition of invading viruses by host cells is critical to antiviral innate immunity. Invading viruses trigger type I interferon-mediated antiviral responses and induce production of effector proteins that inhibit completion of the virus cycle and virus dissemination in vivo [8–12]. Germline-encoded pattern recognition receptors (PRRs) within the innate immune system sense signature molecules expressed by pathogens, known as pathogen-associated molecular patterns (PAMPs). To date, PRRs are classified into three families: retinoic acid inducible gene (RIG)-I-like receptors (RLRs), Toll-like receptors (TLR), and the nucleotide oligomerization domain (NOD) and leucine-rich repeat and pyrin domain-containing (NLRP) proteins [8,13]. RLRs such as RIG-I and melanoma differentiation-associated gene-5 (MDA-5) are important molecules that detect viral RNA in the cytosol. In uninfected cells, RIG-I exists in an auto-repressed conformation in which the caspase activation and recruitment domains (CARDs) are not available for binding to induce downstream signal transduction [14]. Upon recognition of viruses, particularly RNA viruses, RIG-I is activated and undergoes self-dimerization and structural modifications that permit CARD-CARD interactions with the downstream adapter molecule, mitochondrial antiviral signaling protein (MAVS; also known as IPS-1, VISA, and Cardif) [15–20]. Then it activates type I interferon responses via downstream signaling molecules TBK1/IKKi and IRF3, and NF-κB activation via IKK, to elicit inflammatory responses [21–26].
However, interferon- or NF-κB-mediated immune responses need to be tightly regulated to maintain host immune homeostasis, otherwise the uncontrolled immune response can be deleterious, or even fatal, to the host [27–32]. Hence, molecules involved in regulating interferon-mediated innate immune response are the subject of much research. Indeed, mechanisms that regulate RIG-I-mediated antiviral signaling, which is tightly controlled by a series of positive and negative regulators, have been reported [13,33,34]. Among these, NLRX1, a member of the nucleotide-binding domain and leucine-rich-repeat-containing (NLR) protein family, resides on the outer mitochondrial membrane and interfere CARD-CARD interactions between MAVS and RIG-I to negatively regulate antiviral interferon signaling [35–38]. However, during virus infection, the mechanism which controls type I interferon (IFN) signaling via modulating the MAVS and NLRX1 interaction, needs to be investigated more in detail. Here, we show that FAF1 is a positive regulator of the NF-κB and type I interferon signaling pathways during RNA virus infection. FAF1 competitively binds to NLRX1, thereby disrupting its interaction with MAVS and ultimately amplifying the downstream antiviral immune response.
To examine the biological function of FAF1, we performed experiments using FAF1+/+ and FAF1gt/gt mice after confirmed by genotyping (S1 Fig, panels A-B-C). First, mice were infected with the of vesicular stomatitis virus (VSV) Indiana strain (VSV-Indiana) via tail-vein injection and their survival was monitored to determine susceptibility to viral infection (Fig 1, panel A). Knockdown of FAF1 rendered mice significantly more susceptible to lethal VSV infection. A plaque assay and quantitative real-time polymerase chain reaction (qRT-PCR) was conducted to measure the amount of VSV in spleen, lung, liver, and brain tissues at 24 hr and 6 days post-infection (hpi and dpi) (Fig 1, panels B-C and S1 Fig, panel D). Organs from FAF1gt/gt mice contained higher amount of virus than those from FAF1+/+ mice. This suggests that the virus replicates more actively in FAF1gt/gt mice than in FAF1+/+ mice, resulting increased mortality. Additionally, serum samples were collected at different time points after mice were infected with green fluorescent protein (GFP)-tagged VSV (VSV-GFP) (Fig 1, panels D-E) or treated with Poly (I:C) (S1 Fig, panel E). The serum of FAF1gt/gt mice contained more replicating virus and lower levels of IFN-β and IL-6 than that of FAF1+/+ mice, indicating that knockdown of FAF1 suppresses cytokine secretion upon virus infection. Moreover, peripheral blood mononuclear cells (PBMCs) from both groups of mice injected with VSV-GFP via tail-vein were collected and measured to check mRNA encoding IFN-related genes expression at 24 hpi (Fig 1, panel F). PBMCs from FAF1gt/gt mice expressed lower levels of mRNA encoding IFN-related genes than those from FAF1+/+ mice. These results provide in vivo evidence that FAF1 knockdown affects type I IFN mediated signaling and antiviral immunity.
BMDMs were isolated from the bone marrow of FAF1+/+ and FAF1gt/gt mice and infected with VSV-GFP or GFP tagged H1N1 influenza virus (A/PR8/8/34; PR8-GFP). Virus replication was higher in BMDMs of FAF1gt/gt than in those of FAF1+/+ mice at 12 and 24 hpi (Fig 2, panel A). To determine the reason for the increased viral replication in BMDMs of FAF1gt/gt mice, IL-6 and IFN-β levels were analyzed after 12 and 24 hr of VSV-GFP and PR8-GFP infection or Poly (I:C) treatment (Fig 2, panel B). BMDMs of FAF1gt/gt mice produced less IL-6 and IFN-β than BMDMs of FAF1+/+ mice. Next, BMDCs and PBMCs were isolated from FAF1+/+ and FAF1gt/gt mice, and stimulated with VSV-GFP, PR8-GFP or Poly (I:C). Virus titers and cytokine secretion were then compared (S2 Fig). BMDCs and PBMCs isolated from FAF1gt/gt mice harbored greater amounts of virus and secreted lower levels of cytokines than BMDCs and PBMCs of FAF1+/+ mice. These data suggest that immune cells within the BMDMs, BMDCs, and PBMCs populations from FAF1gt/gt mice show inhibited type I IFN signaling, which facilitates viral replication.
To find out whether FAF1 has a similar effect after infection with a DNA virus, BMDMs were isolated from FAF1+/+ and FAF1gt/gt mice and infected with GFP tagged Herpes Simplex virus 1 (HSV-GFP) (S2 Fig, panels C-D). There was no difference in the observed levels of cytokine secretion or virus replication between BMDMs from the two groups of mice. This confirms that FAF1 has no role in DNA virus-stimulated type I IFN signaling. Taken together, these data suggest that FAF1 positively regulates type I IFN signaling in response to infection by RNA viruses.
To examine the effect of FAF1 on virus replication in vitro, we prepared FAF1 knockdown MEFs from FAF1gt/gt mice. FAF1 knockdown was confirmed by immunoblot analysis (S3 Fig, panel A). Virus titers and cytokine levels were measured at 12 and 24 hpi with VSV-GFP, PR8-GFP (Fig 3, panels A-B) or GFP tagged New castle disease virus (NDV-GFP) (S3 Fig, panels B-C). The amount of GFP expressed by cells following viral infection was examined by fluorescence microscopy and quantitated using a fluorescence modulator. FAF1 knockdown MEFs showed increased GFP expression. The virus titer was also higher in FAF1 knockdown MEFs than in wild-type (WT) MEFs (Fig 3, panel A and S3 Fig, panel B). Supernatants from FAF1 knockdown MEFs contained less IL-6, IFN-α, and IFN-β than those from WT MEFs (Fig 3, panel B and S5 Fig, panel C). Moreover, supernatants from FAF1 knockdown cells contained lower levels of cytokines than those from WT cells after stimulation with Poly (I:C) or 5’ppp-dsRNA (Fig 3, panel C). Taken together, these data suggest that knockdown of FAF1 inhibited the immune responses by reducing IFN secretion in response to viral infection, thereby facilitating virus replication. Additionally, FAF1-reconstituted MEFs were prepared and expression of FAF1 was confirmed by immunoblotting (S3 Fig, panel D). Virus titers and cytokine levels in FAF1 knockdown MEFs and FAF1-reconstituted MEFs were compared after virus infection (S3 Fig, panels E-F-G). FAF1-reconstituted cells showed reduced viral replication and higher cytokine secretion than FAF1 knockdown MEFs, demonstrating that reconstitution of FAF1 restores induction of type I IFN signaling.
To exclude the possibility that positive regulation of type I IFN signaling by FAF1 is a cell type-specific phenomenon, knockdown FAF1 murine macrophage cell line was prepared by infecting a lentivirus harboring FAF1 shRNA (small hairpin RNAs) or transfecting FAF1 siRNA (small interfering RNA) to RAW264.7. First, reduced FAF1 expression was confirmed by immunoblot analysis (S4 Fig, panel A). Viral titers and cytokine levels were evaluated after VSV-GFP or PR8-GFP infection (Fig 4, panels A-B and S4 Fig, panels B-C) and treatment with Poly (I:C) or 5’ppp-dsRNA (Fig 4, panel C). Consistent with our previous results, viral titers were higher and cytokine levels were lower in both shRNA and siRNA FAF1 knockdown RAW264.7 cells than in control (scramble) cells. Additionally, THP-1 cells (a human immune cell line) were transfected with siRNA targeting FAF1, and virus replication and cytokine levels were measured after virus infection (S4 Fig, panels D-E-F-G). The results were similar to those for FAF1 knockdown RAW264.7 cells. To confirm these results, we generated stable FAF1 overexpressing RAW264.7 cells and overexpression was confirmed by immunoblot analysis (S5 Fig, panel A). FAF1-overexpressing RAW264.7 cells infected with PR8-GFP, VSV-GFP (Fig 4, panels D-E) or NDV-GFP (S5 Fig, panels B-C) showed lower levels of viral replication and higher levels of IL-6, IFN-β, and IFN-α production than control RAW264.7 cells. Treatment with Poly (I:C) or 5’ppp-dsRNA yielded consistent results with the virus infection experiments (S5 Fig, panel D). Additionally, to find out whether FAF1 has no effect to DNA virus infection in RAW264.7 cells, similar to HSV infection in BMDMs, GFP tagged adenovirus (Adeno-GFP) were infected to control and FAF1 knockdown (S6 Fig, panel A) or overexpressing (S6 Fig, panel B) RAW264.7 cells. Accordance with the results of HSV-GFP in BMDMs, Adeno-GFP experiment also showed no difference in virus replication and cytokine secretion levels between control and FAF1 knockdown or overexpressing cells. Taken together, these results suggest that, irrespective of the cell type, FAF1 positively regulates type I IFN secretion upon RNA virus infection, and not upon DNA virus infection.
Moreover, to confirm whether enhanced VSV-GFP replication in FAF1 knockdown RAW264.7 cells and MEFs was due to repressed IFN secretion by knockdown of FAF1 and not due to intrinsic block to replication of RNA viruses, we infected VSV-GFP to FAF1 knockdown RAW264.7 cells and MEFs in the presence of an anti-IFNAR blocking antibody (IFNAR Ab) and determined VSV-GFP replication level (S7 Fig, panels A-B). As shown in the results, IFNAR Ab treated control cells showed almost two to three times higher virus replication level compared with non-treated control cells due to the IFNAR blocking effect of IFNAR Ab. On the contrary, in FAF1 knockdown cells, virus replication levels slightly enhanced after treatment of IFNAR Ab, which indicated that IFNAR Ab could not exhibit IFNAR blocking effect prominently in FAF1 knockdown RAW264.7 cells, since type I IFN secretion was already suppressed by knockdown of FAF1. These results suggests that enhanced or reduced virus replication depends on knockdown or overexpression of FAF1 due to the regulation of type I IFN secretion by FAF1. Furthermore, from our results of Poly (I:C) and 5’ppp-dsRNA stimulation studies, we could anticipate that FAF1 regulates type I IFN signaling through RIG-I-MAVS signaling pathway, as those stimulants induce type I IFN signaling pathway by activating RIG-I. To investigate whether FAF1 regulates type I IFN secretion not through TLR7 and TLR9, we also stimulated the TLR7 and TLR9 by their agonists, imiquimod and ODN2395, respectively to FAF1 knockdown RAW264.7 cells (S7 Fig, panel C). According to data, the IL-6 and IFN-β secretion levels of control and FAF1 knockdown RAW264.7 cells were similar, which indicate that FAF1 regulates type I IFN signaling through RIG-I mediated pathway, and not via the TLR7 and TLR9 mediated pathway.
To further examine the effects of FAF1 on the antiviral signaling cascade, we next examined virus-induced phosphorylation of IRF3, p65, STAT1, p38, and TBK1. Cells were stimulated with PR8-GFP and samples were collected at indicated time points. Whole cell lysates (WCL) were prepared and analyzed by immunoblotting (Fig 5, panels A-B and S8 Fig, panel A). First, scramble and FAF1 knockdown RAW264.7 cells were infected with PR8-GFP, and phosphorylation levels of the indicated proteins were examined (Fig 5, panel A). Protein phosphorylation in both scramble and FAF1 knockdown RAW264.7 cells was initiated at 8 hpi, and increased until 16 hpi, however, at later time points the levels of protein phosphorylation detected in FAF1 knockdown cells were lower than those in scramble RAW264.7 cells. By contrast, FAF1-overexpressing RAW264.7 cells showed higher levels of phosphorylation at early time points than control cells (Fig 5, panel B). These results provide strong evidence that FAF1 activates the type I IFN signaling pathway. In addition, we examined phosphorylation of target proteins in FAF1 knockdown and FAF1-reconstituted MEFs after PR8-GFP infection (S8 Fig, panel A). The results showed that higher phosphorylation of these signaling proteins occurred at early time points in FAF1-reconstituted MEFs than in FAF1 knockdown MEFs. mRNA encoding IFN-related gene expressions were measured to determine whether type I IFN-related gene transcriptions were affected by FAF1 protein knockdown and reconstitution (Fig 5, panels C-D and S8 Fig, B-C). Lower mRNA expression levels were observed in FAF1 knockdown MEFs than WT MEFs (Fig 5, panel C), and significantly higher levels were noted in FAF1-reconstituted MEFs compared to FAF1 knockdown MEFs (S8 Fig, panel B). Furthermore, we examined the expression of mRNA encoding IFN-related genes in BMDMs and PBMCs isolated from FAF1+/+ mice and FAF1gt/gt mice after infection with PR8-GFP or VSV-GFP (Fig 5, panel D and S8 Fig, panel C). Consistent with our previous findings, low levels of IFN-related gene transcription was observed in BMDMs and PBMCs of FAF1gt/gt mice. To examine FAF1 expression levels in response to viral infection, the levels of FAF1 mRNA were measured in BMDMs, RAW264.7, THP-1, HEK293T, HeLa and A549 cells after PR8-GFP infection. As shown in S9 Fig, panel A, FAF1 mRNA expression levels were increased after viral infection, however, this increase varied according to cell type. Results from these experiments led us to postulate that FAF1 is a positive regulator of the type I IFN signaling pathway.
Previous studies in our laboratory focused on identifying binding partners for NLRX1. A large scale pull-down assay using HEK293T cells overexpressing the GST-tagged N-terminal domain (amino acids (aa) 1–225) of NLRX1 followed by mass spectrometry analysis identified that FAF1 is a binding candidate for NLRX1 (Fig 6, panel A). Immunoprecipitation of GST-tagged NLRX1 followed by immunoblotting with an anti-FAF1 antibody showed that NLRX1 interacted with FAF1 (Fig 6, panel B). Additionally, V5-tagged FAF1 was pull-down from HEK293T and RAW264.7 cell lysates, and NLRX1 was visualized by immunoblotting with an NLRX1 antibody (Fig 6, panel C). To further confirm whether FAF1 directly interacts with NLRX1, in vitro binding assay was performed using purified GST tagged FAF1 (Fig 6, panel D). Incubation of GST tagged FAF1 with recombinant His tagged NLRX1 followed by immunoblotting with anti-His antibody showed the thick band corresponding to the NLRX1 protein size, which indicates direct binding between FAF1 and NLRX1. Moreover, as shown in Fig 6, panel E, confocal microscopic visualization of overexpressed V5-tagged FAF1 and Flag-tagged NLRX1 in HEK293T cells or overexpressed V5-tagged FAF1 and endogenous NLRX1 in FAF1-reconstituted MEFs showed overlapping of NLRX1 and FAF1 spots, confirming their co-localization. To examine endogenous protein binding upon virus infection, FAF1 was immunoprecipitated from PR8-GFP or H1N1-infected HEK293T or RAW264.7 cells using an anti-FAF1 antibody, followed by immunoblotting with an anti-NLRX1 antibody, and band corresponds to NLRX1 was detected (Fig 6, panel F). Similar results were obtained from BMDMs of FAF1+/+ mice after infection of PR8-GFP or VSV-GFP (Fig 6, panel G). These results suggested that FAF1 interacts with NLRX1.
To further elucidate the interaction between FAF1 and NLRX1, the NLRX1 domains responsible for the interaction with FAF1 were analyzed using GST-labeled NLRX1 domain constructs (Fig 7, panel A). This experiment revealed that amino acids (aa) 1–327 of NLRX1 are required for the interaction with FAF1. Moreover, we found that the FAF1 binding site within NLRX1 overlapped with the binding site for MAVS (aa 75–556) (Fig 7, panel B). For further confirmation, another two NLRX1 fragments were constructed (aa 556–975 and 75–975) to check whether FAF1 can bind with MAVS binding region of NLRX1 (S10 Fig, panel A). FAF1 bound to aa 75–975 of NLRX1, however, did not bind to aa 556–975 of NLRX1. This indicates that FAF1 binds with the aa 75–556 region of NLRX1 which binds with MAVS. Based on these findings, we hypothesized that FAF1 and MAVS compete for binding to NLRX1. To test this, a competition assay was performed by transfecting HEK293T cells with V5-tagged FAF1 in a dose-dependent manner (Fig 7, panel C). The results confirmed reduced binding between MAVS and NLRX1, meanwhile increased binding between FAF1 and NLRX1. Thus, FAF1 inhibits the interaction between NLRX1 and MAVS by binding to NLRX1 competitively. To investigate time-dependent changes in the interaction between FAF1 and NLRX1 after virus infection, RAW264.7 cells and BMDMs were infected with H1N1 or PR8-GFP and then harvested at different time intervals. Immunoprecipitation using an anti-NLRX1 antibody, followed by immunoblotting with an anti-FAF1 antibody, revealed that FAF1 bound to NLRX1 in RAW264.7 cells and BMDMs at early time points (4 and 2 hpi, respectively) (Fig 7, panel D). Additionally, the interaction between NLRX1 and FAF1 or NLRX1 and MAVS were examined in infected cells (Fig 7, panel E). The interaction between NLRX1 and MAVS in non-infected cells was markedly reduced after viral infection in a time-dependent manner, importantly, there was a corresponding increase in FAF1-NLRX1 binding. These results suggest that FAF1 interacts with NLRX1 and inhibits binding between MAVS and NLRX1. For further confirmation of this mechanism, we checked whether knockdown of NLRX1 abolishes the antiviral effect of FAF1 on type I IFN signaling. FAF1 and NLRX1 was knockdown in HEK293T cells using FAF1 and NLRX1 specific siRNA (S11 Fig, panel A), and then VSV-GFP was infected to the cells. First, we confirmed antiviral effect of FAF1 in HEK293T cells, as shown in S11 Fig, panel B-C. In knockdown of NLRX1, increased cytokine secretion and reduced VSV-GFP replication were observed. Interestingly, knockdown and overexpression of FAF1 had no effect to virus replication and cytokine secretion levels in NLRX1 knockdown condition. These results correlated with our proposed mechanism that FAF1 regulates NLRX1 mediated type I IFN signaling upon virus infection by interacting with NLRX1. Indeed, FAF1 competes with MAVS for binding to NLRX1, leading to disassociation of NLRX1 from MAVS, MAVS is then free to interact with RIG-I and initiate type I IFN signaling.
MAVS is the key adaptor protein for RLR-mediated signaling [15,39]. The RLR signaling pathway is initiated by recognition of distinct species of viral RNA by RIG-I or MDA5, and activated RLRs bind MAVS via CARD-mediated interactions [19]. MAVS then recruits downstream signaling molecules and, eventually, induces production of type I IFNs and proinflammatory cytokines [20,40]. Although activation of MAVS plays a role in inducing type I IFN to limit virus spread, it must be tightly modulated to prevent excessive cellular immune responses that may have a detrimental effect on the host [28,41]. After viral infection, MAVS is regulated negatively or positively by different mechanisms, including mitochondrial dynamics [42,43], post-translational modifications [44,45], or protein-protein interactions [35,46,47]. With respect to protein-protein interactions, the first protein to be identified as a negative regulator of MAVS was the nucleotide-binding domain and leucine-rich repeat containing family member, NLRX1 [35–38]. Although conflicting results have been reported with regard to the function of NLRX1 as a negative regulator of RLR-mediated antiviral signaling [48–50], it is believed that NLRX1 associates with MAVS on the mitochondrial membrane to inhibit antiviral signaling by interrupting virus-induced RLR-MAVS interactions [35–37,51]. However, the mechanism that regulates NLRX1 during virus infection remains poorly characterized.
FAF1, a member of the ubiquitin regulatory X (UBX) family, potentially interacts with diverse proteins and functions as a negative and/or positive regulator in variety of biological possesses, including apoptosis [1,3], tumor growth [2,4–7], protein degradation [2,6,52] and chaperone activity [53]. Here, we provide several lines of evidence showing that FAF1 is a positive regulator that modulates the type I interferon signaling pathway in response to RNA virus infection. First, FAF1gt/gt mice were more susceptible to infection by VSV as they were permissive to high rates of virus replication and mounted weak antiviral immune responses. Second, knockdown of endogenous FAF1 in immune cells or MEFs from FAF1gt/gt mice reduced RNA virus-induced IFN-β and proinflammatory cytokines production and increased viral replication. Third, overexpression of FAF1 in immune cells or MEFs promoted RLR-mediated antiviral response against RNA virus infection but not DNA virus infection. Fourth, FAF1 interacts with NLRX1 in response to RNA virus infection or RLR stimulation, and aa 1–327 of NLRX1 are responsible for the interaction between FAF1 and NLRX1. Finally, FAF1 interacts with NLRX1 at the early time points after RNA virus infection; this interaction inhibits binding of MAVS to NLRX1, which in turn switches on RIG-I mediated antiviral immune responses. Taken together, these findings indicate that FAF1 is a crucial regulator that induces the antiviral innate immune responses against RNA virus infection.
Recently, Song et.al., reported that FAF1 negatively regulates virus-induced IFN-β signaling and the antiviral response by inhibiting the translocation of active phosphorylated IRF3 from the cytosol to the nucleus [54]. However, this result contradicts that presented herein, and we clearly demonstrated that FAF1 acts as a positive regulator of the type I IFN signaling pathway during RNA virus infection. In particular, we identified the physiological role of FAF1 in innate immune responses against viral infection in FAF1+/+ and FAF1gt/gt mice. FAF1gt/gt mice were more susceptible to infection with VSV than FAF1+/+ mice, resulting high mortality in FAF1gt/gt mice due to a high viral load in the organs. After virus infection or Poly (I:C) stimulation, FAF1gt/gt mice showed lower levels of cytokine production (IL-6 and IFN-β) than FAF1+/+ mice, which strongly supporting an impaired antiviral immune response, especially with respect to type I IFN signaling.
We also evaluated antiviral responses in several cell types. BMDMs, BMDCs, and PBMCs isolated from FAF1gt/gt mice, as well as FAF1 knockdown RAW264.7 cells and MEFs showed higher viral replication and lower IL-6 and IFN-β production, suggesting reduced antiviral and inflammatory responses due to suppression of FAF1. However, reconstitution of FAF1 in FAF1 knockdown MEFs restored the antiviral function of FAF1 by recovering production of these cytokines. Consistent with this, overexpression of FAF1 in RAW264.7 cells resulted enhanced antiviral responses. These results strongly support the involvement of FAF1 as a positive regulator of RNA virus-mediated type I IFN signaling. Here, we also examined the phosphorylation of IRF3 (the main integral component of the type I IFN response), p65 (a subunit of NF-κB), STAT1, p38, and TBK1 after the induction of IFN responses by PR8-GFP. Together with this, significantly decreased type I IFN, ISGs, and antiviral mRNA transcript levels were observed in BMDMs and PBMCs of FAF1gt/gt mice compared with those in FAF1+/+ mice after infection with VSV and PR8.
Interestingly, our large scale co-immunoprecipitation data demonstrated that FAF1 is one of the binding partners for NLRX1 (Fig 6, panel A), which negatively regulates type I IFN signaling by modulating the interaction between RIG-I and MAVS [35–38]. We confirmed that FAF1 binds to and co-localizes with NLRX1. Moreover, we performed domain studies to better understand the mechanism underlying the interaction between FAF1 and NLRX1, and to identify which domain of NLRX1 binds to FAF1. The results showed that the MAVS binding site within NLRX1 (aa 75–556) [35] overlaps with the binding site for FAF1 (aa 1–327). Hence, we postulate that the mechanism by which FAF1 positively regulates IFN signaling probably operates through FAF1-mediated disassociation of NLRX1 from MAVS. This supports our time course binding studies, which showed that FAF1 binds to NLRX1 at early time points (2 and/or 4 hpi) after virus infection. We also examined the time course binding of NLRX1 and MAVS and compared it with that of NLRX1 and FAF1 after the virus infection. Similar to previous studies showing constitutive interaction between RIG-I and MAVS in NLRX1-deficient cells [38], we found that NLRX1 bound to MAVS in the absence of virus infection. This suggests that under normal conditions NLRX1 blocks the interaction between MAVS and RIG-I. However, after virus infection, FAF1 appears to displace MAVS from NLRX1 by competitive binding to NLRX1. Thus, we postulate that FAF1 stimulates type I IFN signaling pathway by sequestering NLRX1 from the RIG-I-MAVS-mediated pathway. Furthermore, we confirmed that knockdown of NLRX1 abolishes the FAF1 mediated effects on type I IFN signaling, which supports our proposed mechanism. Nevertheless, several controversial reports are related with NLRX1 [48–50], indicating more complicate mechanisms might involve on the role of NLRX1 on MAVS-dependent antiviral responses that unexplored yet, and FAF1 could be one of the key participant molecule in this mechanism which needs to be elucidated in future.
Recently, Guo et al. demonstrated that NLRX1 is a negative regulator of host innate immune responses to DNA viruses by sequestering the DNA-sensing adaptor, STING, from TANK-binding kinase 1 (TBK1) [55]. However, in this study, we found no differences in cytokine secretion levels or viral replication levels after HSV infection, suggesting that FAF1 may not modulate type I IFN production via STING-mediated sequestration of NLRX1 upon DNA virus infection. Hence, the present data indicate that FAF1 targets NLRX1 to regulate type I IFN production upon infection with RNA virus only. Moreover, the upstream signaling molecule that activates FAF1 after RNA virus infection and the reason that FAF1 only regulates NLRX1 upon RNA virus invasion remains unclear.
In summary, we showed that FAF1gt/gt mice are highly susceptible to RNA virus infection and show defective innate immune responses both in vitro and in vivo. Upon RNA virus infection, FAF1 binds competitively to NLRX1, thereby preventing it from binding to MAVS; this frees MAVS to interact with RIG-I and switch on the antiviral signaling cascade. These results suggest a plausible and novel mechanism by which FAF1 positively regulates type I IFN signaling and increases our understanding of the molecules that control type I IFN signaling and antiviral immune responses.
All animal experiments were managed in strict accordance with the Guide for the Care and Use of Laboratory Animals (National Research Council, 2011) and performed in BSL-2 and BSL-3 laboratory facilities with the approval of the Institutional Animal Care and Use Committee of Bioleaders Corporation (Reference number BLS-ABSL-14-009).
C57BL/6 FAF1+/+ and FAF1gt/gt mice were kindly provided by Dr. Eunhee Kim (Department of Biology, Chungnam National University, Korea) [56]. Mice (6–7 weeks of age) were infected with vesicular stomatitis virus (VSV) Indiana strain (VSV-Indiana; 2 × 108 pfu (plaque forming unit) per mouse) or green fluorescent protein (GFP)-tagged VSV (VSV-GFP; 2 × 108 pfu per mouse) via tail vein injection. Mice infected with VSV-Indiana were observed daily to measure the mortality until 12 dpi. Organs and sera of mice infected with VSV-Indiana or VSV-GFP were collected at indicated time points to measure virus titers by plaque assays and qRT-PCR as described below, and levels of mouse IFN-β (PBL interferon source) and mouse IL-6 (BD biosciences) were measured by ELISA. Poly (I:C) (Invivogen; 200 mg per mouse) was injected intravenously via tail vein, and the sera were collected to measure mouse IFN-β and IL-6 levels by ELISA at indicated time points. Peripheral blood mononuclear cells (PBMCs) were isolated from whole pheripheral blood of FAF1+/+ and FAF1gt/gt mice infected with VSV-GFP (4 × 108 pfu per mouse) via tail-vein injection at 24 hpi as described below. Total RNAs from PBMCs were extracted and were used for qRT-PCR analysis as described below.
The femurs and tibias were isolated from euthanized C57BL/6 mice (4–6 weeks of age) aseptically. After removing muscles, the bones were flushed with Dulbecco’s Modified Eagle’s medium (DMEM, Gibco) using syringe (26G × ½ needle) to extrude bone marrow at least 3 times. After centrifugation of bone marrow, pellet was re-suspended with 1.0 ml of ammonium-chloride-potassium (ACK) lysis buffer (Gibco) to lyse the red blood cells, and supernatant was aspirated from the white cell pellet after centrifugation. Cells were cultured with DMEM supplemented with 10% heat-inactivated fetal bovine serum (FBS, Gibco) and 1% antibiotic-antimycotic (Gibco) (10% FBS DMEM) and 10 ng/ml granulocyte-macrophage colony-stimulating factor (GM-CSF) for 5 days to obtain Bone Marrow-Derived Macrophages (BMDMs). Additionally, cells were cultured for 6 days by adding 100 ng/ml IL-4 (Invivogen) to the above media to prepare Bone Marrow-Derived Dendritic Cells (BMDCs). Whole peripheral blood obtained from mice was diluted with roswell park memorial institute (RPMI) 1640 medium (Gibco) and PBMCs were isolated by Histopaue-1077 (Sigma). Isolated PBMCs were washed 3 times and cultured in 10% FBS and 1% antibiotic-antimycotic included RPMI1640 medium (10% FBS RPMI).
FAF1 knock-down murine embryonic fibroblasts (MEFs) provided by Dr. Eunhee Kim (Department of Biology, Chungnam National University, Korea) [56], mouse leukaemic monocyte macrophage (RAW264.7; ATCC TIB-71), human embryonic kidney 293 (HEK293T; ATCC CRL-11268), human epithelial cervix adenocarcinoma (HeLa; ATCC CCL-2) and adenocarcinomic human alveolar basal epithelial (A549; ATCC CCL-185) cell line were grown and maintained in 10% FBS DMEM at 37°C and 5% CO2. Human acute monocytic leukemia (THP-1; ATCC TIB-202) cell line was grown and maintained 10% FBS RPMI.
FAF1 tagged with V5 expression plasmid (pIRES-FAF1-V5) was constructed by inserting the FAF1 complete ORF which was amplified from pFLAG-CMV-2/hFAF1 plasmid [6] to the pIRES-V5 vector between AflII and EcoRI site. NLRX1 inserted to pIRES plasmid was kindly donated by Dr. Jae U. Jung (Department of Molecular Microbiology and Immunology, University of Southern California, USA), and GST tagged NLRX1 full (aa 975) and 6 fragments (aa 1–156, 157-.327, 386–674, 675–975, 556–975 and 75–975) were constructed by cloning into the pEBG vector between BamHI and NotI site.
For stable overexpressing cell line preparation, pIRES-V5 vector or pIRES-FAF1-V5 was transfected to RAW264.7 and HEK293T with Lipofectamine 2000 (Invitrogen) according to manufacturer’s protocol. Cells stably expressing pIRES-V5 and pIRES-FAF1-V5 were selected with 2 μg/ml puromycin (Gibco) containing 10% FBS DMEM for 2 weeks.
VSV-GFP, GFP tagged Herpes Simplex virus 1 (HSV-GFP) and Adenovirus (Adeno-GFP) were propagated in Ceropithecus aethiops epithelial kidney (Vero; ATCC CCL-81) cells. GFP tagged H1N1 influenza virus (A/PR8/8/34; PR8-GFP) and Newcastle disease virus (NDV-GFP) were propagated in embryonated chicken eggs. Culture medium was replaced by DMEM supplemented with 1% FBS right before virus infection, and the viruses were added into the medium with indicated MOI. After 2 hr incubation, extracellular virus was removed and replace with 10% FBS DMEM or RPMI. Poly (I:C) was transfected with Lipofectamine 2000 into MEFs or treated to RAW264.7 cells. 5’- triphosphate double-stranded RNA (5’ppp-dsRNA, Invivogen) was transfected into both cell lines with Lipofectamine RNAiMAX (Invitrogen). Imiquimod (Invitrogen) and ODN2395 (Invitrogen) were treated to RAW264.7 cells.
Oligonucleotide sequences of FAF1-specific shRNA cloned into the pGIPZ lentiviral vector expressing GFP was purchased at Open Biosystems. (http://www.openbiosystems.com). Lentiviruses were produced using transient transfection of packaging plasmids (psPAX2 and pMD2.VSV-G purchased from Addgene) into HEK293T cells using Lipofectamine 2000. Media supernatant containing the virus particles were collected after 72 hr, filtered (0.45 μm filter, Millipore) and infected to the RAW264.7 cells with 8μg/ml polybrene (Sigma). Culture medium was replaced after the transduction process (after 12hr) with fresh puromycine-containing medium every 2 days until resistant colonies could identified. Similarly, control cells were prepared by infecting lentivirus which was produced with pGIPZ lentiviral vector expressing GFP.
To knockdown the FAF1 or NLRX1 gene expression, siRNA oligonucleotide duplexes for targeting mouse FAF1 (si-mFAF1-S 5’-UGUUUCCCUGGGACCAUCU-3’ and si-mFAF1-AS 5’-AGAUGGUCCCAGGGAAACA-3’), human FAF1 (si-hFAF1-S 5’-CAGUAGAUGAGUUAAUGAU-3’ and si-hFAF1-AS 5’-AUCAUUAACUCAUCUACUG-3’) or human NLRX1 (si-hNLRX1-S 5’-GAGGAGGACUACUACAACGAU-3’ and si-hNLRX1-AS 5’- AUCGUUGUAGUAGUCCUCCUC-3’) was transfected to cells (si-mFAF1; RAW264.7 cells, si-hFAF1; THP-1 and HEK293T cells and si-hNLRX1; HEK293T cells) using Lipofectamine RNAiMAX according to the manufactures protocol.
To measure virus titer, supernatant of homogenized organs (VSV-Indiana), cells (VSV-GFP, NDV-GFP and HSV-GFP) or freezed-thawed cells (PR8-GFP and Adeno-GFP) which collected at indicated time points were serially 10-fold diluted and inoculated to Vero cells in 1% FBS containing media. After incubation for 2 hr at 37°C, cells were overlaid with DMEM containing 1% agarose (Sigma). Cultures were incubated at 37°C, 5% CO2 for 48 hr, plaques were visualized with crystal violet. Virus titer was calculated using the number of plaques and the dilution factor. GFP expression levels were measured using a fluorescence modulator (GloMax-Multi detection system; Promega) to digitize.
ELISA was used to detect the production of pro-inflammatory cytokines and type I interferon from cells. After infection, treatment and transfection of stimulants, cell supernatant was collected and analyzed cytokine production levels. Mouse IFN-α (PBL interferon source), mouse IFN-β (PBL interferon source), mouse IL-6 (BD biosciences) and mouse TNF-α (BD biosciences), human IFN-β (PBL interferon source) and human IL-6 (BD biosciences) were used for analysis according to manufacturer’s protocol.
At 48 hr post-transfection of indicated plasmids, cells were harvested and lysed with radio-immunoprecipitation assay (RIPA) lysis buffer (50 mM Tris-HCl, 150 mM NaCl, 0.5% sodium deoxycholate, 1% IGEPAL, 1 mM NaF, 1 mM Na3VO4) supplemented with protease inhibitor cocktail and phosphatase inhibitor cocktail (Sigma) and sonicated to prepare the whole cell lysate (WCL). WCL were precleared with Sepherose 6B (GE Healthcare Life Science) at 4°C at least for 2 hr. Precleared lysates were incubated with 50% slurry of glutathione-conjugated Sepharose (GST) beads (Amersham Biosciences) for GST pull-down, and for immunoprecipitation of anti-V5 and NLRX1, lysates were incubated with anti-V5 or NLRX1 antibody (1.0 μg/ml) for 12 hr and then protein A/G plus agarose beads (Santacruz) were added. Immunoprecipitates were collected by centrifugation, washed with lysis buffer in different washing conditions.
Additionally, WCL in control, FAF1 knockdown and overexpressing RAW264.7 and FAF1 knockdown and reconstituted MEFs infected with PR8-GFP during indicated time points were subjected to immunoblotting to analyze protein phosphorylation levels using respective antibodies.
For all the immunoblot analysis, samples were separated by SDS-PAGE and transferred onto a PVDF membrane (Bio-rad) using Trans-Blot semi dry transfer cell (Bio-rad). Membranes were blocked for 1 hr in tris-buffered saline containing 0.05% tween 20 (TBST) containing 5% bovine-serum albumin (BSA). After overnight-incubation at 4°C with antibodies, membranes were washed with TBST. Membranes were incubated at room temperature with 1:3000 dilutions of horseradish peroxidase-conjugated secondary antibodies. Membranes were developed with western blotting detection reagents (GE healthcare, ECL select Western Blotting Detection Reagent).
The antibodies used in this study were as follows: anti-GST (Santacuze, #SC-138), anti-V5 (Invitrogen, #46–0705), or anti-IRF3 (Abcam, #ab25950), anti-phospho-IRF3 (Ser 396) (Cell signaling, #4947), anti-NF-κB p65 (Cell signaling, #4764), anti-phospho-NF-κB p65 (Ser536) (Cell signaling, #3031), anti-STAT1 (Cell signaling, #9175), anti-phospho-STAT1 (Cell signaling, #9167), anti-phospho-p38 (Cell signaling #9216), phospho-TBK1 (Cell signaling #5483), anti-NLRX1 (Proteintech, #17215-1-AP) and anti-His (Santacuze, #SC-1803) antibodies. The anti-FAF1 monoclonal antibody was provided by Dr. Eun-hee Kim (Department of Biology, Chungnam National University, Korea). The anti-interferon-α/β receptor (IFNAR) (25 μg/ml; Leinco Technologies) was pre-incubated in RAW264.7 cells and MEFs for 1 hr before VSV-GFP infection to block IFNAR.
Total RNA was isolated from cells and tissues from the organs using RNeasy Mini Kit (Qiagen) and cDNA synthesis was performed using ReverTra Ace kit (TOYOBO). cDNAs were then quantified with gene specific primer pairs using QuantiTect SYBR Green PCR kit (Qiagen) on a Rotor-Gene Q (Qiagen) and relative expression of mRNA was normalized to GAPDH mRNA expression using delta-delta CT method. Gene specific primer pairs were referred in Table 1.
The cells were seeded into collagen-coated chamber slides (LabTek, Nunc), 1 day prior to the experiments. Following day, the cultured cells were washed with phosphate buffer saline (PBS) and fixed with 4% paraformaldehyde for 20 min, then permeabilized through incubation for 20 min with 100% methanol at -20°C. The fixed cells were first incubated with 2% FBS diluted in PBS for 1 hr to block non-specific binding of antibodies. V5 and NLRX1 were detected through incubation with the primary antibodies (1:100 diluted in 2% BSA) for 12 hr at 4°C. After 3 times PBS containing 0.05% tween 20 (PBST) washing, the secondary antibodies (1:100 diluted in 2% BSA; Alexa 488 goat anti-rabbit IgG (Invitrogen), Cy3-conjugated donkey anti-mouse IgG (The Jackson Laboratory) were added and the cells were incubated for 1 hr at room temperature. Three times PBST washing followed by 10 min incubation with 1 μg/ml DAPI (Sigma-Aldrich) containing 0.01% RNase A, the nuclei were visualized, and then the slides were mounted with mounting solution (VECTOR) to check under fluorescence microscopy. Images were acquired from Nikon C2 Plus confocal microscope (Nikon) consisting of a Nikon Eclipse Ti inverted microscope with a confocal scanning system (Nikon) in conjunction with C-HGFIE precentered fiber illuminator (Nikon). FITC and TRITC fluorescence was detected using the 488 nm and 561 nm laser line of a Sapphire driver unit (Coherent), respectively, and DAPI fluorescence was detected using 405 nm laser line of a CUBE laser system (Coherent). The image data were analyzed using NIS-Elements microscope imaging software program (Nikon).
HEK293T cells were transfected with an empty GST vector (GST) or with the GST-NLRX1-N-terminal region containing vector (aa 1–225; GST-NLRX1-N). Cells were harvested after 48 hr and after cell lysis, proteins in the cell lysates were immunoprecipitated with GST beads and separated by 4–15% Nu-PAGE gels (Invitrogen), followed by silver staining [57]. Protein bands present exclusively in GST-NLRX1-N lane were excised from the gel and identified by mass spectrometry.
GST and GST tagged FAF1 (GST-FAF1) were expressed and purified using GST beads. The purified GST-FAF1 was incubated with recombinant His tagged NLRX1 (NovoPro) in binding buffer (50 mM Tris-HCl, 150 mM NaCl, 1% IGEPAL and protease inhibitors) at 4°C for 3 hr with gentle rocking. After centrifugation, collected beads were washed five times with binding buffer, and bound proteins were subjected to SDS-PAGE followed by immunoblotting with GST and His antibodies.
Statistical analysis was performed using GraphPad Prism software version 6 for Windows (GraphPad Software). All the data were from at least of two independent experiments and data are shown as mean ± SEM. The means values of all the in vitro experiments were compared by Student’s t test. Log Rank test and Mann-Whitney test was subjected for in vivo survival data analysis. Comparisons between multiple time points were analyzed by one way analysis of variance (ANOVA). In all experiments, p values of less than 0.05 were considered statistically significant. *p<0.05, **p<0.01 and ***p<0.001
|
10.1371/journal.pcbi.1006450 | A neuromechanistic model for rhythmic beat generation | When listening to music, humans can easily identify and move to the beat. Numerous experimental studies have identified brain regions that may be involved with beat perception and representation. Several theoretical and algorithmic approaches have been proposed to account for this ability. Related to, but different from the issue of how we perceive a beat, is the question of how we learn to generate and hold a beat. In this paper, we introduce a neuronal framework for a beat generator that is capable of learning isochronous rhythms over a range of frequencies that are relevant to music and speech. Our approach combines ideas from error-correction and entrainment models to investigate the dynamics of how a biophysically-based neuronal network model synchronizes its period and phase to match that of an external stimulus. The model makes novel use of on-going faster gamma rhythms to form a set of discrete clocks that provide estimates, but not exact information, of how well the beat generator spike times match those of a stimulus sequence. The beat generator is endowed with plasticity allowing it to quickly learn and thereby adjust its spike times to achieve synchronization. Our model makes generalizable predictions about the existence of asymmetries in the synchronization process, as well as specific predictions about resynchronization times after changes in stimulus tempo or phase. Analysis of the model demonstrates that accurate rhythmic time keeping can be achieved over a range of frequencies relevant to music, in a manner that is robust to changes in parameters and to the presence of noise.
| Music is integral to human experience and is appreciated across a wide range of cultures. Although many features distinguish different musical traditions, rhythm is central to nearly all. Most humans can detect and move along to the beat through finger or foot tapping, hand clapping or other bodily movements. But many people have a hard time “keeping a beat”, or say they have “no sense of rhythm”. There appears to be a disconnect between our ability to perceive a beat versus our ability to produce a beat, as a drummer would do as part of a musical group. Producing a beat requires beat generation, the process by which we learn how to keep track of the specific time intervals between beats, as well as executing the motor movement needed to produce the sound associated with a beat. In this paper, we begin to explore neural mechanisms that may be responsible for our ability to generate and keep a beat. We develop a computational model that includes different neurons and shows how they cooperate to learn a beat and keep it, even after the stimulus is removed, across a range of frequencies relevant to music. Our dynamical systems model leads to predictions for how the brain may react when learning a beat. Our findings and techniques should be widely applicable to those interested in understanding how the brain processes time, particularly in the context of music.
| Humans have the ability to estimate and keep track of time over a variety of timescales in a host of different contexts ranging from sub-seconds to tens of seconds or more [1, 2]. On the millisecond to second time scale, for example, numerous studies have shown that humans can accurately discriminate shorter intervals from longer intervals [3, 4]. On a longer timescale, we utilize a form of time estimation that can span hours, days or years [5]. Many such examples involve the brain making a calculation over a single event, so-called “interval timing” [6, 7]. Humans can also track timing that involves multiple or repeated events. For example, we instinctively move to the beat of a piece of music through a form of sensorimotor synchronization, so-called beat-based timing [8–11]. Doing so involves identifying an underlying beat within a piece of music and coordinating the frequency and timing of one’s movements to match this beat.
Understanding how humans perceive a beat has been an active area of research for quite some time. Beat perception refers to our ability to extract a periodic time structure from a piece of music. It is a psychological process in which beats can be perceived at specific frequencies, even when the musical stimulus does not specifically contain that frequency [12]. In a recent study by Nozaradan et al. [13], brain activity was found to entrain to the beat frequency of a musical rhythm. Additionally, participants with strong neural entrainment exhibited the best performance when asked to tap to the rhythm [13]. Various parts of the brain have been identified as being active during beat perception. Grahn and Brett reported that basal ganglia and the supplementary motor area showed increased activity for beat-based tasks, and as such, postulated that these areas mediate beat perception [14]. Interestingly, fMRI studies of participants asked to lie still with no movement while listening to music revealed that the putamen, supplementary motor area, and premotor cortex are active [15]. Thus although no external movement may be occurring, various motor areas are nevertheless active when the brain is representing a passage of time. From the theoretical perspective, error-correction [16–22], entrainment [12, 23, 24], and Bayesian [25–27] models have been proposed to account for the ability to perceive a beat.
Many beat perception studies have involved finger tapping while listening to a piece of music or a metronome [13, 28–32]. However, humans can also mentally conjure a beat in the absence of motor movement and external stimuli. These observations, in part, lead us to ask what neural mechanisms might be responsible for detecting, learning and generating a beat. We define beat generation as the brain’s process of construction and maintenance of a clock-like mechanism that can produce repeatable, essentially constant, time intervals that demarcate a beat. In this formalism, the brain may be monitoring the firing times of neurons involved in beat generation to match event times of an external source. Beyond such reactive synchronization, we suggest that beat generation can occur as a strictly internally created and self-driven phenomenon in the absence of external stimuli.
Models of beat generation should take into consideration several empirically observed features taken from human finger tapping studies in the presence of isochronous tones (evenly spaced in time). First, the model’s output should rapidly synchronize with the external tone sequence. Second, a model should mimic the human ability to continue tapping even after the stimulus is removed [33], a property known as synchronization-continuation. Third, a model should quickly resynchronize after a tempo change or perturbation (deviant or phase shift) to the tone sequence. Fourth, ideally a model should be capable of addressing the phenomenon of negative mean asynchrony (NMA), the reported tendency of humans to tap on average prior to tone onsets. Although the cause of, extent of, and even the existence of NMA are still in dispute, new models can potentially provide insight into the phenomenon.
Two primary modeling frameworks for addressing the above mentioned properties have been proposed: entrainment models and error-correction models. Entrainment models rely on principles of dynamical systems. Mathematically, entrainment refers to an external forcing that sychronizes a set of oscillators to a specific frequency. In the context of beat perception, entrainment models posit the existence of oscillators that resonate and entrain to the underlying periodicity creating an oscillation whose spectral profile matches that of the sound sequence. These models have been used to explain various beat-related phenomena including the emergence of pulse and meter [24] and the missing pulse percept [12]. These oscillator models are typically abstract mathematical formulations and, although generic in structure, presuppose a formulation in which the system is poised near to oscillatory-destabilization of a steady state, a Hopf bifurcation [24]. Error-correction models, on the other hand, are formulated at an algorithmic level to understand how a motor movement, such as a finger tap, can be synchronized to an isochronous tone sequence [18–22]. Errors between the current tap and tone times and between the current intertap time and stimulus period are used to adjust the timing of the next finger tap. Error-correction models provide different algorithmic ways in which to make an adjustment (see [29] for a review), but typically do not propose mechanisms for how a set of neurons would estimate and correct for the error.
In this paper, we introduce a neuromechanistic framework that can be used to construct neuronal network models that are capable of learning and retaining isochronous rhythms. In its simplest form, the network consists of a single, biophysically-based, beat generator neuron (BG), a periodic brief stimulus and a time-interval computation mechanism based on counting cycles of a gamma oscillation. The BG does not directly receive input from the external stimulus and is thus not being entrained by it. Instead, the BG learns (within a few cycles) the frequency of the stimulus thereby allowing the BG to continue oscillating at this frequency, even in the absence of the stimulus. Our approach combines ideas from entrainment based, information processing based and interval timing based models. In part, it extends the heuristic two-process model [18–21] to a neural setting, by developing a neural system that learns the period and phase of the periodic stimulus, in order to bring its spikes into alignment with the stimulus tones.
A central feature of our model is the concept of a gamma counter. Gamma rhythms (30-90 Hz) are ubiquitous throughout the human nervous system [33, 34]. Here we utilize roughly 40 Hz gamma oscillations to form two discrete-time clocks that count the number of gamma cycles between specific events. This idea is similar in spirit to pacemaker-accumulator models, which also use counting mechanisms [6, 7, 35]. In our case, one clock counts gamma cycles between successive onsets of a stimulus, sometimes called the interonset interval in behavioral studies. Another neuronal clock counts gamma cycles between successive spikes of the BG, the interbeat interval. A comparison is made via a putative gamma count comparator (GCC) between these two counts and this information is sent to the BG, consistent with recent studies of neuronal circuits that can count discrete events and compare counts with stored information [36–38]. Our BG neuron possesses plasticity and uses the difference in count to adjust an intrinsic parameter so that it learns the interonset interval of the stimulus. The gamma counters also provide information to the BG about its firing phase relative to stimulus onset times, thereby allowing for the possibility of synchronization with zero-phase difference of the BG spikes with the stimulus. The coupling of the stimulus and BG counts through the GCC is both nonlinear and non-periodic. This contrasts with coupling between stimulus and oscillator in many entrainment and information processing models where either the phase or period is directly targeted for change. Further, in such models either the coupling is periodic or the update rules are linear or vice versa [39]. Our model updates are neither periodic or linear. We note that the neuronal clocks that count cycles need not operate exclusively in the range of 40 Hz. The comparison mechanism that we describe will work for any sufficiently fast frequency oscillator.
In this paper, we will show how the BG model learns and holds an isochronous beat over a wide range of frequencies that includes the band of 0.5 Hz to 8 Hz which is relevant for beat generation and perception. Using mathematical analysis and a continuous time clock, we explain first how the BG learns to period match. We then show how this extends to both period and phase for the discrete time clock counters. As will be seen, the discrete time clocks give rise to a natural bounded variability of spike times of the BG even when it is holding a beat. We will test our model using standard paradigms which parallel behavioral studies that are designed to study specific aspects of beat perception and production, such as NMA [13, 40], synchronization-continuation [41] and fast resynchronization times [42, 43]. The BG’s mean firing time exhibits negative mean asynchrony and we describe how this NMA can be manipulated. The model resynchronizes rapidly in response to perturbations. For an abrupt change in tempo we predict and explain with the model an asymmetry in the resynchronization time for increases versus decreases in tempo. As with tempo changes, our model predicts an asymmetry in the resynchronization time due to a phase advance versus delay of the stimulus sequence. An asymmetry also arises for a transient perturbation, a single deviant onset timing in the stimulus sequence. We explain these effects by understanding how our model incorporates linear, but discrete step, error-correction to invoke non-linear changes in frequency of the BG. In turn, we develop a set of testable predictions for human behavior that help to contrast our proposed model framework from existing ones.
The main components of our model consist of a periodic stimulus with an associated neuron, S, whose spikes mark each stimulus onset, a neuronal model for the beat generator, BG, and a gamma count comparator, GCC, which acts as a type of neural integrator as well as error detector. These components are linked together as shown in Fig 1A. The output from the spiking neuron S and of the BG are sent to the GCC. There a comparison is made which is then sent via a period learning rule, LRT, and a phase learning rule, LRϕ, to adjust Ibias which subsequently changes the instantaneous frequency of the BG. The term Ibias is taken here to represent the drive to the BG that governs its frequency. It could be considered as a parameter internal to the BG, or can be more generally associated with summed synaptic input that drives the BG. In either case, it is a term that regulates the BG’s excitability.
The BG in our model can be described using biophysical, conductance based equations and is required to have only two specific properties. The first is that it possesses a sharp (voltage) threshold. A spike of the BG occurs when the voltage increases through this threshold. The second requirement is that the BG has a slow process that governs the time between its spikes. A simple model that possesses these basic properties is the leaky integrate and fire (LIF) model which we first use to describe our analytic findings. Our simulation studies utilize a biophysical model motivated by models of delta waves in sleep, as we require a similar frequency range for the BG. To that end, we chose voltage-gated currents similar to those from an idealized model for sleep spindle rhythms of thalamo-cortical relay cells, namely for the slow wave of relay cells in burst mode [44]. The slow wave can be generated by the interplay of a T-type calcium current, and Ih current and a leak current. Here, the BG has a persistent sodium INaP, T-type calcium ICaT, sag Ih and leak IL currents. In the text, we refer to this as the INaP model; for a full description and the equations see Appendix. The analytic results also hold for the INaP model, but as the analysis is more complicated, showing so is outside the scope of this paper. For either the LIF or INaP models, parameters are chosen that allow for a wide range of intrinsic frequencies of the BG up through 8 Hz as Ibias is varied as quantified by a neuron model’s Ibias versus frequency relationship which is presented in the Results. This range of frequencies is appropriate for speech and music. The time between successive spikes of the BG is called the interbeat interval and is denoted by IBIBG.
The voltage in the LIF model evolves according to
v ′ = I bias - v τ (1)
where v is a dimensionless variable representing voltage, Ibias is the drive to the neuron and τ is the membrane time constant. The LIF model has a spike and reset condition which makes it discontinuous. When the voltage reaches one at t = ts, it is instantaneously reset to the value 0; if v ( t s - ) = 1, then v ( t s + ) = 0. When Ibias > 1 oscillations exist. In this case, the LIF model is rhythmic with period given by
T = τ log e ( I bias I bias - 1 ) . (2)
The period of BG given by Eq (2) can be adjusted to any positive value by appropriately adjusting Ibias.
For both the LIF and INaP models, the specific nature of the stimulus is not modeled, only the onset is of interest here. We limit our simulations to a range between 1 and 6 Hz, which corresponds to an interstimulus interval ranging from 1 s down to 166 ms. There is no theoretical or practical problem to extend the model outside of this range, as further addressed in the Discussion. We utilize a neuron S to faithfully transform the stimulus sequence into spikes. The interonset interval, IOIS, is then defined as the time between successive S spikes. The model for S is not important provided that it is set to be an excitable neuron that fires quickly in response to input; see the Appendix for equations.
The gamma count comparator, GCC, in our model utilizes two generic oscillators with frequency sufficiently larger than that of both the stimulus and the BG. Here it is taken to lie in the gamma range at roughly 40 Hz (Fig 1B). We choose the oscillators to be identical, though this is not a requirement of the model. To avoid integer values, both have a frequency of 36.06 Hz (period 27.73 ms); see the Appendix for details. We let γBG be a variable that counts the number of gamma cycles between consecutive BG spikes and γS be a variable that counts the number of gamma cycles between consecutive spikes of S. At each spike of BG or S, the appropriate counter is reset to zero. We stress that γBG and γS are integers, but, in general, the periods that they are estimating are not integer multiples of a gamma cycle (27.73 ms). Hence, although the stimulus period may be constant, the gamma counts may vary from cycle to cycle. The difference γBG − γS provides an estimate of how different or aligned the frequencies of BG and S are. For example, if S oscillates at 5 Hz and the BG is initially oscillating at 3 Hz, then the GCC would count roughly 8 cycles between S spikes and 13 cycles between BG spikes. In this case, the GCC determines that the BG is oscillating too slowly and sends a speed up signal to the BG. Alternatively if the BG were initially oscillating at 6 Hz, then the GCC counts roughly 6 cycles and sends a slow down signal. In general, speeding up or slowing down of the BG is achieved by changing Ibias. At each spike of the BG, the period learning rule adjusts Ibias by a fraction of the difference between the gamma oscillator count of cycles between S and BG. This learning rule (LRT) assigns at each BG spike
LR T : I bias → I bias + δ T ( γ B G - γ S ) , (3)
where the parameter δT is independent of period. This simple rule is enough to align the frequencies of S and BG, the details of which will be explored through the derivation and analysis of a one-dimensional map. However, this frequency matching rule does not provide the beat generator with any information about its firing phase relative to stimulus onset.
To align the phase of the BG to the stimulus onset, we formulate a second learning rule. We define the current count of the BG, CCBG, as the number of gamma cycles from the last BG spike to the current S spike; see Fig 1B. Then at each S spike define ϕ = CCBG/γS to be the phase of BG firing. We use the phase to determine if the BG fires “before” or “after” S at each cycle. In a rhythmically active network, the concept of whether BG fired before S is somewhat ambiguous. We define the BG to be “before” the stimulus if it fires in the second half of the stimulus period ϕ ∈ (0, 0.5). In this case we say that the BG is too fast and needs to slow down. Conversely, if ϕ ∈ (0.5, 1), the BG is said to fire “after” S and needs to be sped up. At each S spike, we update Ibias with the second part of the learning rule (LRϕ)
LR ϕ : I bias → I bias + δ ϕ q ( ϕ ) ϕ | 1 - ϕ | (4)
where δϕ is independent of period and phase and q(ϕ) = sgn(ϕ − 0.5), with q(0.5)<0. Thus if ϕ = 0 (or 1), there is no change to Ibias. But if the BG fires before S (ϕ ∈ (0, 0.5)), then q(ϕ)<0 and Ibias is decreased to slow down the BG. The opposite occurs if the BG fires after S. The absolute value keeps the last term positive as ϕ can become larger than 1. For example, during transitions from high to low frequency, CCBG can exceed γS. The quadratic nature of LRϕ is chosen so that the maximum change occurs for phases near 0.5. With this two-part learning rule, the BG learns both the period and phase of S. Both parts of the rule are implemented concurrently so that the process of period and phase alignment occurs simultaneously. The two rules LRT and LRϕ target the value Ibias. In the Results section, we describe how these changes to Ibias, in turn, affect the frequency of the BG which then affects the period and phase of oscillations.
Given the discreteness of our gamma counters, the BG learns to fire a spike within a suitably short window of time of the stimulus onset, an interval equal to plus or minus one gamma cycle. We define this concept as one gamma cycle accuracy. For the earlier described choice of parameters, this amounts to ±27.73 ms from stimulus onset. We address two important and related concepts: synchronization to the beat and holding a beat. In our model, synchronization to the beat refers to the process by which the BG brings its spike times within one gamma cycle accuracy of a specific stimulus frequency. Holding a beat refers to the ability of the BG to maintain synchronized firing at a specific frequency over a specified stretch of time. We will say that BG has synchronized to the stimulus if three consecutive BG spikes each fall within one gamma cycle accuracy in time of a stimulus onset. The BG is said to be holding a beat for as long as it continues to remain synchronized with the stimulus onset.
In the presence of an isochronous stimulus, the BG displays what we shall call stationary behavior. This refers to the pattern of spike times of the BG in response to a fixed frequency stimulus. Despite there being no source of noise in our model, the discrete nature of the gamma count comparator allows the BG’s spike times during stationary behavior to naturally display variability. Thus, during stationary behavior, while the BG’s spike times typically land within one gamma cycle accuracy of stimulus onset they can also fall outside this window. The variability of the BG’s spikes arises because the gamma counters and learning rules adjust Ibias in discrete steps whenever γS ≠ γBG or ϕ ≠ 0. What this means is that during stationary behavior, the BG does not converge to a limit cycle oscillation (periodic orbit). The variables that govern the dynamics of the BG do not periodically return to the same values, but instead can vary by small amounts from cycle-to-cycle. In practice, these small differences affect the exact spike times of the BG, creating the variability. Furthermore, as the gamma counts are not exact representations of the period, they may be equal even when IOIS and IBIBG are unequal. This amounts to additional variability in the BG’s spike times relative to the spike times of S.
We will determine the time that it takes for the BG to resynchronize its spikes to stimulus onset after a change to the stimulus. Resynchronization is declared similarly to synchronization in that the BG is required to fire three consecutive spikes each of which must lie within one gamma cycle accuracy of a stimulus onset. The resynchronization time is then taken as the time of the first synchronized spike. In all studies, we begin with the BG displaying stationary behavior at a specific frequency. Because of the variability present in stationary behavior, the resynchronization times will depend on the initial conditions at the moment that the change to the stimulus profile is enacted. We will compute mean resynchronization times and standard deviations over 50 realizations, each of which differs by a small change in the initial condition of the BG at the time that the stimulus profile is changed. All simulations are carried out in MATLAB with a standard Euler solver (Euler-Maruyama when noise is introduced).
We provide a short outline of the results that follow. We start with a demonstration of how the BG learns to synchronize to an isochronous stimulus sequence. We then describe how the BG learns a period by first utilizing a continuous time version of the gamma counters to derive a one-dimensional map. The discrete gamma counters are then used to describe period and phase matching. Next, we present the basic behaviors of the BG describing its response under both stationary (isochronous stimuli) and transient (tempo changes, phase shifts and deviants) conditions. The section concludes with a brief description of the effects of parameter changes and noise.
An oscillatory neuronal model spikes with a period that is quantifiable by its frequency versus Ibias relation (f-I). This relationship is obtained from the reciprocal of (2) for the LIF model and computed numerically for the INaP models (Fig 2A). The blue (red) curve depicts the f-I curve for the LIF (INaP) model. In the LIF model, the interspike interval is governed by the difference between Ibias and the spiking threshold, as well as the parameter τ. In the INaP model, the interspike interval is determined by an interplay of the various non-linear currents (Fig 2B). In particular, the ICaT and IL currents provide basic excitability to the model, the INaP current allows for spikes once a voltage threshold is crossed and the Ih current provides a slow depolarization of the membrane allowing the neuron’s voltage to gradually reach spiking threshold. Thus the primary determinant of the interspike interval is the time constant of the Ih current. An important point regarding the f-I relations is that they are both strictly increasing. Hence, there is exactly one value of Ibias that yields a specific frequency. The learning rules we use make discrete changes to Ibias. Thus, there is little chance of adjusting Ibias to the exactly correct value. Instead, the learning rules adjust Ibias so that it stays within a small window of the correct one. The frequency relations increase steeply from frequency equal to zero. Therefore, at low frequencies, larger changes in frequency can result from small changes in Ibias. The same is also true for the INaP model at frequencies in the 3 to 8 Hz range. It is important to note that any implementation of a BG model with a monotone increasing f-I relation will produce the qualitative results described below. However, the quantitative details will certainly depend on the slope and non-linearities of these relations that are produced by different ionic currents and parameters. For example, changes to Ibias in the LIF at frequencies above 1 Hz lead to linear changes in BG frequency. Alternatively, for the INaP model a change of Ibias from say 5 to 10 produces a smaller change in BG frequency than changes from 10 to 15.
The BG learns to oscillate at a frequency by adjusting its bias current through the set of plasticity rules LRT and LRϕ (Fig 3). The BG is initially set to oscillate at 2 Hz with Ibias = 9.06. At t = 0 ms, we adjust the stimulus frequency to 4.65 Hz and activate the period learning rule LRT (Fig 3A). Notice how the cycle period of the BG increases on a cycle-by-cycle basis until it matches the stimulus period. This results from the value of Ibias iteratively increasing over the transition, based on the difference γBG − γS. The first change to Ibias does not occur until t = 500 ms, which is the time the BG naturally would fire when oscillating at 2 Hz, since LRT updates (purple curve in the bottom panel of Fig 3A) are only made at spikes of the BG. At around t = 2.25 s, the value of Ibias falls within one gamma cycle accuracy as depicted by the blue band but continues to adjust. Note that Ibias does not settle down to a constant value. Instead it changes by ±δT whenever |γBG − γS| = 1. Additionally, since LRT contains no phase information, the spikes of the BG are not synchronized in phase with those of S. At t = 4.2 s (20 cycles of the stimulus), the stimulus is completely removed, and the BG continues to oscillate at roughly 4.65 Hz. This shows that the BG has learned the new frequency and does not require periodic input to make it spike. There are still adjustments to Ibias because the BG continues to compare γBG with the last stored value of γS. This example demonstrates how the BG oscillates at a learned frequency rather than through entrainment to external input.
When both learning rules operate together, the BG learns both the correct period and phase. Starting with the same initial conditions as in Fig 3A, now at t = 0 ms both LRϕ and LRT are turned on (Fig 3B). Note the very rapid synchronization of the BG’s spikes with the stimulus onset times. The middle panel shows how Ibias grows much more quickly when both rules are applied. The first update is due to the phase learning rule at the third stimulus spike, at t = 433.5 ms, which is earlier than in the previous example. This causes enough of an increase in Ibias for the BG to immediately fire, which causes an update due the period learning rule. The lower panel shows how the two rules LRT and LRϕ contribute to the change in Ibias. Note that the learning is not sequential with period learning preceding phase learning or vice versa. Rather, period and phase learning occurs concurrently. Below we shall describe in more detail each of these learning rules and their role in synchronizing the BG with the stimulus.
The dynamics of how the period learning rule LRT matches the interbeat interval of the BG, IBIBG, with the interonset interval of S, IOIS, can be explained in terms of an event-based map. Each spike of the BG is treated as an event and we define a map that updates the bias current Ibias on a cycle-by-cycle basis. To derive the map, we first use exact time differences, in effect, equivalent to a continuous time-keeping mechanism. This will allow the map to possess a parameter-dependent asymptotically stable fixed point. For simplicity of presentation, we use the LIF model to derive the specifics of the map. We will then discuss how those findings inform simulations of the INaP model for the discrete gamma count case.
Assume that the stimulus sequence occurs with a fixed period T* corresponding to a specified IOIS, and that the BG is initially oscillating with an interbeat interval of T0. This IBIBG corresponds to a specific value I0 of Ibias given by solving (2). Ibias is then updated to I1 by comparing T0 to T*. In turn, this produces a new cycle period T1 and so on. In general, the continuous time version of LRT updates Ibias at each firing of the BG as follows:
I n + 1 = I n + δ T ( T n - T * ) = I n + δ T ( τ log e ( I n I n - 1 ) - T * ) , (5)
where the second line is obtained by substituting Eq (2) evaluated at In for Tn. Error-correction models also take the form of an iteration scheme, but typically target the next cycle period for adjustment, i.e. Tn and Tn+1 would replace In and In+1, respectively, in the first equation of (5). In contrast, the adjustment in our model is made to the biophysical parameter Ibias (In) which then has a subsequent effect on the cycle period (Tn).
Eq (5) defines a one-dimensional map which can be expressed as In+1 = f(In), where f(I) denotes the right-hand side. A fixed point of the map satisfies I* = f(I*) whose stability can be determined by checking the condition |f ′(I*)| < 1. A fixed point of the map corresponds to a case where the IBIBG of the BG is equal to the IOIS of S. Stability of the fixed point implies that the learning rule is convergent. Note that for any T*, there is a unique fixed point of the map which satisfies I* = 1/(1 − exp(−T*/τ)). This means that any stimulus period can in practice be learned by the BG, provided that the fixed point is stable. A simple calculation shows that |f ′(I*)| < 1 provided 0 < δT < 2I*(I* − 1)/τ. For fixed δT, as the stimulus frequency gets smaller, I* converges to 1, and as a result the term 2I*(I* − 1)/τ goes to zero. This expression provides the insight that convergence for lower stimulus frequencies requires taking smaller increments in the learning rule. This finding carries over to any f-I relation that is steeply sloped at low frequencies.
Parameter dependence and the ensuing dynamics of the map are readily illustrated graphically (Fig 4). The one-dimensional map has a vertical asymptote at I = 1, a local minima at I = ( 1 + 1 + 4 δ T τ ) / 2 and a slant asymptote of I − δT/T*. The graph intersects the diagonal at exactly one point, and the slope of the intersection determines the stability as calculated above. For increasing stimulus frequency, with δT and τ fixed, the map’s graph shifts upward (Fig 4A) and the fixed point moves to larger values of Ibias. Note, for low stimulus frequency (here, 1 Hz) the fixed point is unstable. The update parameter δT does not change the value of the fixed point I*, but affects the stability (Fig 4B). As δT increases, the slope at the intersection decreases through 0, then through -1, at which point stability is lost. If the stimulus frequency changes (eg, 2 Hz to 5 Hz), Ibias changes dynamically as the BG learns the new rhythm. The learning trajectory corresponds to the cobweb diagram on the map (Fig 4C, black dashed lines and arrows). Each adjustment of Ibias occurs at a spike of BG and allows it to speed up for the next cycle. In this example, it takes only a few cycles for the BG to learn the new rhythm. The transition from 5 to 2 Hz (Fig 4C, red dashed lines and arrows) demonstrates the asymmetry in convergence for similar sized changes of opposite directions. Here the convergence for the decrease in frequency occurs over fewer cycles because the value of δT chosen yields a fixed point at 2 Hz with near zero slope. The smaller in magnitude the slope, the less the number of cycles needed to converge. This result suggests that certain preferred frequencies can exist for specific choices of parameters.
In contrast to the idealized continuous-time learning rule, the gamma count-based case does not lead to updates that converge to zero. An interesting illustration is seen after an IOIS has been learned and the stimulus is turned off. Small updating persists (e.g., Fig 3A, bottom panel). Just after the turn-off, the IBIBG is less than the last stored IOIS (γBG < γS). So the BG is too fast, and at the next BG spike the period rule LRT activates and decreases Ibias by δT producing a new, longer IBIBG. Not immediately, but after a while (just after t = 5s) a difference in gamma counts again arises. This time LRT increases Ibias, shortening IBIBG and so on. These changes are all due to LRT as the phase learning rule LRϕ can never be invoked since the stimulus is no longer present.
The phase learning rule LRϕ considers the current BG gamma count, CCBG, at each firing of S. As a result, the BG has information about its phase at each stimulus onset. We use a learning rule function ϕ|1 − ϕ| that has maximal effect at ϕ = 0.5 and no effect at ϕ = 0 and 1. This is similar to a logistic function that attracts dynamics towards ϕ = 1; see also [45] for a similar mathematical rule used in a different biological context. In our case ϕ = 0 is equivalent ϕ = 1, so our learning rule LRϕ utilizes a sign changing function q(ϕ) = sgn(ϕ − 0.5), q(0.5) = −1 to stabilize ϕ = 0 as well. This will allow convergence via either phase increase or decrease towards synchrony. At each S spike-time, the BG is sped up (if ϕ ∈ (0.5, 1)) or slowed down (if ϕ ∈ (0, 0.5)) by adjusting Ibias until the phase reaches a neighborhood of 0 or 1. This, in conjunction with LRT which equalizes the IOIS and IBIBG, brings about synchronization. Note that when the BG fires within one gamma cycle accuracy of S, ϕ = 0 or 1. In that case, there is no update to Ibias. Thus as with LRT, because of the discreteness of the learning rule updates, the value of Ibias is brought into close proximity of the value of Ibias that produces a specified rhythm but need not become exact. The rapid synchronization results shown earlier in Fig 3 hold for a large range of stimulus frequencies. Under certain conditions, it is possible to derive a two dimensional map that tracks how Ibias and ϕ change on a cycle-by-cycle basis. Though it is outside the scope of this paper, an analysis of the map shows that stable period and phase matching can be achieved for each IOIS, if δT and δϕ are not too large. In practice these two parameters should be chosen so that the changes due to LRT, δT(γBG − γS), and LRϕ, δϕ ϕ|1 − ϕ|, are of the same order of magnitude.
If the IBIBG is at least one gamma cycle longer than the IOIS, then ϕ > 1. We could have restricted the phase to be less than one by periodically extended LRϕ beyond the unit interval, but this would allow for stable fixed points at ϕ = 2, 3, 4, etc. Instead, the learning rule utilizes an absolute value around the term 1 − ϕ to keep it positive. This introduces an asymmetry in the resets of Ibias. For example if ϕ = 0.1 then ϕ(1 − ϕ) = 0.09 but if ϕ = 1.1, then ϕ|1 − ϕ| = 0.11. Thus when the BG fires after the S spike after a long IBIBG (e.g. ϕ = 1.1) then Ibias is increased more than it is decreased if it fires before the S spike (e.g. ϕ = 0.1). As a result, the learning rule favors the BG firing before the stimulus onset, as if in anticipation. This is more pronounced at lower frequencies where the slope of the f-I curve is much steeper than linear. This issue is explored in more detail in the following sections.
The phase learning rule LRϕ adjusts Ibias as opposed to directly affecting the phase of the BG, for example, via a perturbation and reset due to the phase response curve (PRC). Using a PRC to adjust phase would lead to a situation of entrainment rather than learning. Indeed, with a PRC, bringing the value of Ibias to within one gamma cycle accuracy to achieve a specific frequency is rarely reached. Thus if the stimulus were to be removed, a BG with a PRC-based phase rule would fail to continue spiking at the correct target frequency, i.e. it would fail in a synchronization-continuation task.
To hold a beat, the BG must fire spikes within a time window of one gamma cycle accuracy of stimulus onset. As discussed earlier, the discreteness of the gamma counters and comparator causes the BG spike times to naturally display variability. Thus the BG must at each firing compare its period and phase relative to stimulus onset times and make necessary corrections. Holding a beat is an example of stationary behavior of the BG in response to a constant frequency stimulus (Fig 5). In this typical example, here shown at 2 Hz, each spike of the BG is aligned to the closest spike of S and then a timing error equal to the BG spike time minus S spike time is computed. The value of Ibias hovers around the dashed black line Ibias = 9.06 which is the value that produces exactly a 2 Hz oscillation (Fig 5A, upper). The spike times of the BG jitter around those of S, and thus, the timing error is poised around zero (Fig 5A, lower). While holding a beat, these differences fall within a single gamma cycle time window (dashed gray lines at ±27.73 ms). During some time windows (pink shaded region in Fig 5B) no updating of Ibias occurs (lower time course), but the timing differences progressively decrease and become more negative. The BG spike times are drifting relative to onset times because the IBIBG is slightly smaller than IOIS, but close enough that γBG = γS. The drift represents the fact that during this interval, the BG is not entrained to the stimulus. The slope of the timing error in this interval is shallower when Ibias is closer to 9.06, as shown for example in the time window between t = 72 and 76 s. During some intervals (e.g. shaded blue region), the update rules are actively trying to keep the period and phase of the BG aligned with S. Although the timing errors are not large in this case, the counts γBG and CCBG differ from γS thus causing LRT and LRϕ to be invoked. During either drifting or corrective behavior, the BG spike times occur closely in time with the stimulus onsets. Note that drifting and corrective behaviour continue to occur after the stimulus is turned off, but only the period rule will be active (see Fig 3). Many studies have pointed towards drifting dynamics in human tapping experiments [46–48]. Dunlap first noted this behavior, stating that the errors tend to get progressively more negative/positive, until a correction occurs and causes a change in direction [46].
Although the dynamics of the BG are deterministic, they are sensitive in quantitative detail to changes in initial conditions. This is because the learning rules LRT and LRϕ ignore timing differences less than one gamma cycle. To get a more general sense of the fluctuations in BG firing times, we ran a simulation for 1000 stimulus cycles and calculated error distribution plots (spike time of BG minus spike time of S). This was performed at six different stimulus frequencies in steps of 1 Hz (Fig 6). There are several points to note. First, at all frequencies, the error distribution shows negative mean asynchrony [49, 50]. In other words, the actual time of the beat generators firing, on average, preceded the time of the stimulus onset. Second, the variance in the error distribution shows some frequency dependence, particularly with the standard deviation increasing at slower frequencies. Further, the standard deviation increases as the frequency decreases down to 0.5 Hz. We also found that the standard deviation increases with frequency in the 6-8 Hz range. Accurately tapping at rates above ∼ 4 Hz is extremely difficult, hence, no tapping studies exist for this frequency range to either corroborate or contradict our result. However, Drake et al. [51] found a U-shaped dependence of the subject’s tempo discrimination ability on frequency in the range 1-10 Hz, consistent with our result. We additionally note that increasing δT and δϕ leads to increased variability and larger NMA at all frequencies.
In our model, the interplay of the BG neuron’s f-I curve and the learning rules, LRT and LRϕ, are responsible for the frequency dependent results. In particular, at both low (∼ 0.5Hz) and high (∼ 5-8Hz) frequencies the f-I curve is steeper than in the intermediate range. Hence, equidistant changes of Ibias around a mean value result in different changes in instantaneous frequency. The negative mean asynchrony arises from the non-linearity in the f-I curve and the asymmetry in the phase learning rule LRϕ. As states earlier, LRϕ pushes the BG to fire before the stimulus.
As demonstrated in Fig 3, the BG is able to quickly learn a new frequency. This learning can be quantified as a resynchronization of the BG’s spike times with the new stimulus onset times. As previously stated, we declare the BG to be resynchronized if three consecutive spikes each fall within one gamma cycle accuracy of an S spike. We computed the resynchronization times as a function of several parameters including initial and final stimulus frequency (Fig 7 shows one example). From a fixed initial stimulus frequency, we changed the stimulus frequency to different values within the range 1 to 6 Hz and computed resynchronization times. In one such case, the stimulus frequency is decreased from 3 to 2 Hz (Fig 7A). The change is applied at t = 0 s (gold star) and the BG takes about four seconds (eight stimulus cycles) to synchronize to the new frequency (depicted by the shaded region). During the transient, the learning rules LRT and LRϕ drive Ibias down in order to slow the BG down (lower panels of Fig 7A). Adjustments due to LRT occur whenever the BG spikes. For the first second after the change in frequency, BG spikes at roughly its initial 3 Hz rate. The S neuron spikes at t = 0.5 s, which resets γS. But the BG is not aware of this new larger value of γS until it fires at around t = 0.66 s. At this point, γS > γBG and the period learning rule LRT decreases Ibias. Adjustments due to LRϕ occur whenever the stimulus neuron S spikes, which now occur at the slower 2 Hz rate. These adjustments depend on the current phase of the BG and are seen at times to increase Ibias, but at other times to decrease it. Within two seconds, both rules have succeeded in bringing Ibias within one gamma cycle accuracy of the 2 Hz target value (dashed black line inside blue band in middle panel). Aligning the spike times then takes a few more seconds. In contrast, an increase in stimulus frequency can lead to much shorter resynchronization times (Fig 7B). In the transition from a 3 to 4 Hz stimulus frequency, the BG only takes about one and a half seconds (six stimulus cycles) to synchronize. The phase learning rule LRϕ plays a more prominent role as it is invoked more often due to the increase in stimulus frequency. These examples illustrate two important properties of the resynchronization process. First, the two learning rules act concurrently to adjust Ibias, but are asynchronous in that LRT adjustments occur at different times than those of LRϕ (see the lower panels of either Fig 7). Second, adjustments to Ibias are not periodically applied, they occur at a BG or S spike, and only the S spikes occur periodically.
Resynchronization times increase with decreasing frequency, but are nearly constant and mostly flat for increasing frequency (Fig 7C). Decrements from initial to final frequency lead to slower convergence than equally-sized increments. This follows from the slope of the f-I curve being steeper while increasing from 3 Hz than when decreasing. For the slowest stimulus frequency (1 Hz) oscillation, we have included a broader measure of synchrony, defined by BG spike times falling within 5% of the interonset times, i.e. ± 50 ms around S spike times. This definition is consistent with the stationary behavior shown in Fig 6 where many of the BG spikes fall outside of one gamma cycle accuracy. With this broader measure of resynchronization, the average number of cycles and standard deviation of the resynchronization to 1 Hz rhythm are reduced. Although resynchronization times are longer for frequencies decrements, the number of stimulus cycles for resynchronization do not show major differences for increments and decrements, except for the 1 Hz case (the mean number of cycles for resynchronization are reported beside each data point).
The resynchronization process occurs stereotypically depending on whether there is an increase or decrease in frequency (Fig 7D). We calculated the average cycle-by-cycle time differences for 50 realizations of the resynchronization process from 3 Hz to the target frequency with the standard deviation shown in the shaded region. Decreases (increases) in frequency show initial time errors that are positive (negative). This is due to our spike alignment process (see Fig 7 caption). Each curve is non-monotonic and, except for the 6 Hz curve, has an under- or over-shoot that transiently takes the curve outside the band of one gamma cycle accuracy (horizontal grey line). The average resynchronization times in Panel C are shown as the time at which the curve reenters this band, i.e. the time at which the timing errors become consistently less than one gamma period. Consistent with our prior results, the standard deviation bands are largest for the 1 Hz curve and relatively similar and small for the other curves.
Resynchronization also occurs when a phase shift of the stimulus sequence occurs. Now consider the 2 Hz case for which the IOIS is 500 ms. A phase advance will occur if we shorten one IOIS to be less than 500 ms and then return the remainder of sequence to the original IOIS of 500 ms. A phase delay is the opposite, where a single IOIS is elongated. We define the phase ψ of the shift to lie within (−0.5, 0.5) where negative values represent advances and positive values represent delays. If the phase shift falls within one gamma cycle of the normal onset time, the BG is likely to initially ignore it since no change in the gamma counts will occur. But for larger valued phase shifts, resynchronization will need to occur (Fig 8). As an example, resynchronization for a positive phase shift at ψ = 0.4 (Fig 8A) is much quicker than the corresponding negative phase shift ψ = −0.4 (Fig 8B). The reason for this is how LRT changes Ibias in either case. A negative phase shift causes the BG to increase its frequency in response to the temporarily shorter IOIS, followed by a return to a lower frequency. A positive phase shift causes the opposite, a transient decrease in the BG frequency followed by an increase. As we have shown earlier, resynchronization times are shorter when the target frequency is larger (Fig 7). Hence, the model predicts that resynchronization times should be shorter for positive phase shifts (Fig 8C red). The mean timing errors (standard deviation shaded) for different phase shifts (Fig 8D) are stereotypical in much the same way as the timing errors for tempo changes. In the current context, the timing errors start out large and then systematically reduce until they fall within one gamma cycle accuracy. The graph clearly shows that the resynchronization after positive phase shifts is faster than after negative shifts, as negative phase shifts exhibit an overshoot.
Another case where we see resynchronization is the introduction of a temporal deviant where a single S spike occurs at an unexpected early or late time. Unlike phase shifts in which a single IOIS changes, a single deviant causes both the IOIS before and the IOIS after the deviant to change. An early (late) deviant causes a shorter (longer) interonset interval, followed by a longer (shorter) than normal one, followed by a return to the standard IOIS. The model’s response is different for early versus late deviants. For an early deviant, the phase learning rule LRϕ is invoked at the time of the early deviant. This is then followed by the period learning rule LRT at the next BG spike. Both of these signal the BG to speed up. For a late deviant, however, the BG spikes when it normally would have. At that time, it has no new information about its phase or about the value of γS. Therefore there is no change to Ibias. When the late deviant arrives, it now causes LRϕ to send a slow down signal to BG. But any potential changes due to LRT have to wait one full BG cycle to be invoked. Thus, the model reacts quicker to an early deviant than to a late deviant. Thus we predict shorter resynchronization times for earlier deviants (Fig 8C blue). Additionally, because of the need to adjust to two different IOISs, we predict that resynchronization times due to deviants will be longer than those for comparably sized phase shifts where only a single IOIS is changed.
The model results are robust to perturbations and can operate over a range of parameters. To assess this we ran several simulations, where we varied intrinsic parameters of the BG and the gamma counter speeds. For example, the maximal conductance for IL, INaP and ICaT was varied by up to 10% and we measured the subsequent performance across a range of periods. This did not affect the ability of the BG to learn the correct period and phase, because the f-I relation remained qualitatively unchanged. At a quantitative level though, the range of Ibias values which yield gamma cycle accuracy will differ and the BG neuron may have a different preferred frequency. These results indicate that the BG does not require fine tuning of parameters to learn a rhythm.
Next we allowed the speed of the gamma counters for S and BG to be different. We kept the gamma counter for IOIS at 36.06 Hz while we varied the gamma counter frequency for the BG counter by up to 10%. In all cases, across a range of frequencies there was little qualitative difference in the BG’s ability to learn the correct frequency. This is not surprising as the discreteness of the gamma counts allows for similar values of the counts despite there being differences in counting speed. Note that a faster gamma counter for the BG tends to lead to earlier firing times relative to stimulus onset times for the parameters we have chosen. In this case, LRT tends to increase Ibias since γBG is larger than γS. On the other hand LRϕ decreases the same quantity and it is the parameter-dependent balance between the two rules that determines how much earlier on average the BG fires. A slower BG gamma counter has the opposite effect. These results imply that the extent of NMA can be manipulated by changing the counter frequency of the BG and can even be transformed to a positive mean asynchrony if the BG gamma counter is too slow.
To assess the effects of noise, we introduced stochasticity into the gamma counters (see Appendix for details). This acts to jitter the gamma periods, but for modest noise this will only cause the gamma count to discretely change by at most plus/minus 1. Since the BG is monitoring its period and phase at each spike and stimulus event, it quickly adjusts to counteract these potential changes. We also see an increase in the standard deviation of the timing error, across all frequencies, during stationary behavior, as well as an increase in NMA. While this widened the distributions (as seen in Fig 6), approximately 90% of the timing errors remain within one gamma cycle accuracy, apart from at 1 Hz where only 60% of the distribution lies within one gamma cycle accuracy (80% lie within 5% accuracy). Finally, although not explicitly modeled here, one could introduce intrinsic noise in the BG, for example a noisy spike threshold or ionic conductance. This small amount of noise would not change the IBIBG by more that a single gamma cycle and, as above, should not change the BG’s ability to synchronize to the external rhythm. In general, noise makes the model fit better with tapping data, exhibiting more variability and larger NMA. Given that human tapping data contains both motor and time keeper noise that our model does not attempt to disentangle (our model does not have an explicit movement component), we did not address this further.
We presented a modeling framework that begins to address how a neuronal system may learn an isochronous rhythm across a range of frequencies relevant to speech and music. We showed how a biophysical conductance-based model neuron, the beat generator (BG), adjusts its spiking frequency and aligns its spike times to an external, metronome-like stimulus. Our model employs two gamma frequency oscillators to estimate the number of oscillatory cycles between certain salient events. We posit a mechanism that compares the states of these independent counts to inform the BG to either increase or decrease its instantaneous frequency and adjust its relative phase. With this idealized paradigm, we showed that the BG quickly learns to hold a beat over a range of frequencies that includes, but is not limited to, 1 to 6 Hz. Further, we showed how the BG reacts within a few cycles to changes in tempo, phase shifts (permanent realignment of the stimulus sequence) and the introduction of deviants (temporary misalignment of a single stimulus event). Of particular note, the BG displays an asymmetry in reacting to changes to the rhythm. It adapts more quickly when the tempo is increased as opposed to decreased; correspondingly, it reacts faster to phase delays than phase advances, but slower to late deviants than early deviants. Importantly in our model formulation no direct input from the stimulus to the BG is provided. This implies that the BG is learning the correct period and phase rather than being entrained to them. Secondly, no explicit or exact time intervals are required to be calculated, implying that the BG does not need specific mechanisms to exactly track time. Instead, in order to tune the BG, one needs only to know, in some rough sense, whether the BG’s spikes are happening too fast or slow relative to stimulus frequency and too early or late relative to the stimulus onset. Finally, because of the discrete nature of the gamma counters, the BG dynamics are robust to modest parameter changes and noise.
Beat perception as described in many previous studies [52, 53] refers to the ability of an individual to discern and identify a basic periodic structure within a piece of music. Beat perception involves listening to an external sound source as a precursor to trying to discern and synchronize with the beat. Alternatively, we might ask how do we (humans) learn and then later reproduce a beat in the absence of any external cues. Such issues and questions lead us to consider what neuronal mechanisms might be responsible for producing an internal representation of the beat. At its most basic level, we refer to this as beat generation, and a neuronal system that does so we call a beat generator. Different than beat perception, beat generation is envisioned to be able to occur in the absence of an external cue. A BG is a neural realization of an internal clock that can be used as a metronomic standard by other internally driven processes that depend on time measurements. While demonstration of a beat involves a motor action (tapping, clapping, vocalizing, head bobbing), the BG could include a general representation of a motor rhythmicity but the specific motor expression (say, foot tapping) may not be an integral part of the BG.
Our formulation proposes that time measurement for beat perception and the beat generator model are oscillator-based. In this view, a beat can be learned and stored as a neuronal oscillator (cell or circuit). The frequency range of interest, 1-6 Hz, is relatively low compared to many other neuronal rhythms, but similar to those seen in sleep. We rely on faster (gamma-like) oscillators to provide clocklike ticks and we assume two counters and a comparator circuit can be used for adjusting the BG period and phase to match with the stimulus. Conceptually, counting and comparing with a target period are essential features of the algorithmic (or sometimes called, information processing timekeeper) approach, falling into the class of error-correction strategies; see [16, 17, 20–22, 54] for examples of two-process models. These models suggest mechanisms used by humans to bring their movements into alignment with a rhythmic stimulus. They do not, however, provide a biological framework for these mechanisms. We provided a neuronal implementation of the BG in the form of an oscillator with a tunable biophysical knob and two learning rules; the BG is a continuous-time dynamical system, a realizable neuronal oscillator. It does not require a separate reset mechanism. The implementation also does not require a separate knob for phase correction; the two learning rules both make adjustments/corrections to the same parameter, Ibias, and they are ongoing whenever a stimulus is present. We propose this BG as the internal clock—an oscillator that learns a beat and keeps it.
A different class of oscillator models for beat perception relies on large networks of neuronal units [12, 24, 28]. The units’ intrinsic frequencies span the range of those that are relevant in speech and music. In the neural entrainment models of Large and collaborators, different units within the ensemble respond by phase-locking to the periodic stimulus. Units with intrinsic frequencies near that of the metronome will entrain 1:1 while those with higher intrinsic frequencies entrain with different patterns, such as 2:1. Dominant responses are found at harmonics and sub-harmonics of the external input. Amplitude, but not precise timing relative to stimulus features (say, stimulus onset times), are described in these models. The framework is general although the identities of neuronal mechanisms (synaptic coupling or spike generation) are not apparent as the description is local, based on small amplitude perturbation schemes around a steady state and the coupling is assumed to be weak. The approach is nonlinear and provides interpretations beyond those of linear models, e.g. it identifies a beat for complex input patterns even if the beat/pulse is not explicitly a component of the stimulus [12].
Our model cannot be described as entrainment in the classical sense. Entrainment occurs when an intrinsically oscillatory system is periodically forced by an external stimulus to oscillate and, in the present context, to phase lock at the forcing frequency (or some subharmonic) that may differ from its endogenous frequency. Our BG neuron is not entrained by the stimulus but rather it learns the frequency of the stimulus. The BG’s frequency is adapted indirectly through the control parameter in order to match with the stimulus. The influence of the stimulus on the BG diminishes as learning proceeds. In fact, in the continuous time version when the frequency and phase are eventually learned, the BG no longer requires the stimulus; it will oscillate autonomously at the learned frequency if the stimulus is removed or until the stimulus properties change. In the discrete time version, even after the stimulus and BG periods and phase agree (to within a gamma period accuracy) modest adjustments are ongoing to maintain the rhythm. In contrast, for an entrainment model, the oscillator’s parameters are fixed. The stimulus does not lead to a change in the oscillator’s intrinsic properties. For a transient perturbation, the dynamics of resynchronization are according to an entrainment unit’s phase response curve, which instantaneously changes the current phase of the oscillator. In contrast, the BG’s response to transient inputs impacts the parameter Ibias invoked by adjustments according to either or both of the period and phase learning rules. Our model is further distinguished from entrainment models in that the BG strives for zero phase difference but in an entrainment setting there is typically a phase difference between the stimulus and the units. Finally, for an entrainment model the coupling from stimulus to oscillator is periodic. In our model, the influence of a periodic stimulus is delivered both periodically (via LRϕ) and aperiodically (via LRT).
Although humans can learn to accurately estimate time intervals [1], little is known about the neural mechanisms used to generate these estimates. For beat generation, we are positing an ability to estimate time intervals (e.g., between stimulus onset events) in real time in an ongoing and flexible manner. We introduced the idea of gamma counters to perform such measurements. These counters provide a rough estimate of elapsed time that can be used to compare the internal representation of an interval with that of an external cue. The model then produces a finer representation of the interval by adjusting the BG’s spike time and period. There is growing evidence for the existence of counting mechanisms within neuronal systems. For example, Rose and collaborators have demonstrated that neurons in the auditory mid-brain of anurans (frogs and toads) count sound pulses in order to make mating decisions [36, 55]. These neurons have been called ‘interval counting neurons’ because they respond only after a threshold number of pulses have been counted provide that those pulses are spaced in time intervals of specific lengths [37]. In a very different context, it has been recently demonstrated that mossy fiber terminals in rat hippocampus have the ability to count action potentials, an ability cited as improving the reliabilty and accuracy of information transfer [38].
The discreteness of the gamma counter, used in our model, leads to variability in the BG spike times, allowing the model to exhibit negative mean asynchrony (NMA). This is consistent with finger tapping experiments which show that humans display variability in their tap times relative to an isochronous stimulus and tend to, on average, tap before the stimulus 13]. As discussed earlier, the NMA as shown in Fig 6, is rather modest, however, changes in parameters will lead to larger NMA. In contrast, replacing the gamma frequency oscillators with continuous time clocks, which exactly determine time intervals, leads to perfect phase alignment, ϕ = 0 (no NMA). Thus, our work posits the existence of discrete time clocks as a potential source of intertap time variability.
The gamma counters also provide an upper bound on the stimulus frequency which can be reliably learned by the BG neuron. For the 36 Hz clocks used here, this limit is roughly 9 Hz. After this point the phase rule overcorrects, transiently increasing Ibias to a value corresponding to a much larger frequency. We stress that this upper bound is dependent on the specific gamma frequency, and faster clocks may be used to keep track of shorter intervals. An interesting experimental study would be to look at the EEG power spectrum while subjects listen to periodic stimuli and monitor whether the gamma band activity changes with stimulus presentation rate.
Many interval timing models involve accumulation (continuous time or counting of pacemaker cycles) with adjustment of threshold or ramp speed [6, 7] to match the desired time interval. Applications to periodic beat phenomena, say the metronome case, would include instantaneous resetting and some form of phase adjustment/correction [56, 57]. Algorithmic models may not specifically identify the accumulator as such, but instead refer to counters or elapsed time. Our BG model shares some features with interval models for beat production (as described in [9] and [58]), as the BG relies on counters and accumulators. Additionally, as described earlier, it shares features with entrainment models, as the BG is a nonlinear oscillator. In short, the BG is a hybrid.
Interval- and oscillator-based models are related. Even if not explicitly stated as such, in an interval model, the accumulator and its reset are equivalent to highly idealized models for neuronal integration, the so-called integrate-and-fire (IF) class of models [59]. For steady input, the state variable rises toward a target value (that is above the event threshold), rising linearly for a non-leaky IF model and with a decreasing slope for a leaky IF model (LIF), and is reset once the state variable exceeds the threshold. These IF/LIF models are dynamical system oscillators, and are also nonlinear by way of the reset mechanism. However, the time constant/integration rate required for beat applications is much longer/slower than in typical applications of IF models for neuronal computations where timescales of 10-30 ms are more common. These models have entrainment and phase-locking properties [60, 61] and they typically show a phase difference from the stimulus. Extended in this way, periodic in time, such an interval model can be recast as an entrainment model (see also [39]). As noted by Loehr et al. [39], differences between such interval and continuous oscillator models do appear in some circumstances. Adding a plasticity mechanism, say for the threshold or input drive, then allows learning of a period. We described how one may analyze the dynamics of such an LIF oscillator-like interval model in terms of a map (Fig 4). One could additionally add a phase correction mechanism as in two-process models in order to achieve zero-phase difference. This can be achieved in a LIF model, for example, by adjusting the reset condition after reaching threshold or by utilizing phase response curves. Our mechanism for phase correction differs from these approaches in that we target the excitability parameter Ibias for adjustment. This has the advantage that the BG learns the correct phase and period allowing it to continue to hold a beat after the stimulus is removed, similar to other two-process interval models.
The effects of noise on time estimation/production have been studied with interval models, cast as first passage time problems for accumulator models (drift-diffusion models) [62–64]. In that context, the issue of scalar timing is of significance [5, 63, 65, 66], however the time intervals of interest are typically longer than what one would find in a musical context. Wing and Kristofferson [16, 17] considered effects of noise and contrasted sensory noise with motor sources of noise, concluding that timekeeper noise was frequency dependent but motor noise was not. Whether or not scalar timing holds for short rhythmic intervals is unsettled. A number of tempo discrimination studies have failed to produce any evidence for frequency dependent errors for periods below 1000 ms [51, 67]. However, Collyer et al. [68] reports scalar timing in the distribution of tap times when tapping to an isochronous rhythm.
A distinctive interval model was developed by Matell and Meck [69, 70]—the striatal beat frequency (SBF) model. In this neuromechanistic description, the basic units are neuronal oscillators with different fixed frequencies. All oscillators are reset at t = 0; differences in frequencies of convergent units will eventually lead to collective near-coincidence (so-called beating phenomenon of non-identical oscillators) at a time that through learned choices (synapses onto coincidence detector units) can match the desired interval. It may be extended to the periodic case and considered for beat generation as discussed in [56, 57] although the brain regions involved may be different for explicit time estimation than for rhythmic prediction/reproduction [71, 72].
We consider here only the case of isochronous inputs. A natural next step is to consider more complex, non-isochronous stimulus sequences. Additionally, we have side-stepped questions of perception in order to focus solely on timing. Our BG model does not recognize variations in pitch or sound level. For example, if stimulus events were alternating in, say, sound level (as in [73]), our model, as is, would not capture the effects. An extension of our model involving pairs of stimulus and beat generator clocks for each sound level could conceivably address this shortcoming.
We have chosen a particular biophysical instantiation for the BG. The capabilities of learning and holding a beat over a range of frequencies depends only on the monotonic frequency dependence of the control (“learnable”) parameter and would not be compromised by variation of biophysical parameters. Some features of the BG dynamics (say, the degree and signatures of asymmetries in resynchronization for speeding up or slowing down) can be expected to depend on the specifics of, say, the relationship between Ibias and the intrinsic frequency, but we have not explored this in detail.
The learning rules LRT and LRϕ utilized in our study both target the excitability parameter Ibias with a simple goal to either speed up or slow down the BG so that it synchronizes with the stimulus. Alternatively, the drive could be provided as the summed synaptic input from a population of neurons afferent to the BG. The synaptic weights onto the BG and/or internal to the afferent population could be plastic and affected by our learning rules which in spirit are similar to spike time dependent plasticity rules [74]. Our model assumes significant increments of drive at each learning step, leading to fast learning. This may be relatable at a population scale to balanced network models, where fast learning can be achieved with smaller step changes due to the large number of synapses [75]. Currently, during synchronization continuation, our BG model retains its estimate of the most recent stimulus period, γS. We have not yet included a slow decay of this memory or a slow degradation of the BG rhythm. It is plausible that the addition of noise could lead to this slow drift after the stimulus is removed since, as we showed, noise does introduce additional variablity during stationary behavior.
We have not ascribed a location for the BG within a specific brain region. As a result, we have not addressed issues of sensorimotor synchronization (SMS) where sensory processing of a beat must be coordinated with the motor action that demonstrates the beat (e.g. finger tapping). Several models for SMS in the context of beat perception already exist, for example the two-layer error-correction model of Vorberg and Wing [76] and the entrainment model of Large et al. [12] described earlier. Van der Steen and Keller have developed the Adaptation and Anticipation Model (ADAM) [22], a type of algorithmic error-correction SMS model, and they noted a need for an extended ADAM that would incorporate dynamical systems principles. Our model could certainly be a starting point for such an endeavor. Patel and Iversen [77] proposed the Action Simulation for Auditory Prediction (ASAP) hypothesis. In their conceptual model, the motor system primes the auditory system to be able to process auditory input. In particular, ASAP proposes that the motor system is required for beat perception. Generally, these studies raise questions about whether the causal roles of sensory and motor systems can be disambiguated in the context of beat perception and beat generation. Addressing such questions from a modeling perspective is a natural next step.
Our model framework allows us to make several predictions, which are summarized here. First, the BG model succeeds at synchronization continuation [78]; it can hold a beat after the sound stimulus terminates. The BG will continue to oscillate with fine adjustments of its period as needed, according to LRT, to match that of the most recently stored IOIS of the stimulus (as in Fig 3). Error-correction models would also continue oscillating at the last stored frequency [20, 21]. In contrast, for an entrainment model, without a plasticity mechanism, the oscillators are likely to return to their original intrinsic frequencies after the stimulus is removed. However, Large et al. [12] have illustrated using their two-layer model, that the network can hold a beat if the units within the motor layer, have bistable properties, i.e. a unit may have a steady state (damped resonator) coexistent with an oscillator state. The main difference between our approach and others is that period error correction (LRT) occurs at the BG spike times rather than at stimulus onsets. As such, even after stimulus removal, comparison between the BG period and last stored stimulus period continues. Second, the time course of adjusting to a sudden tempo change occurs over seconds and has more or less monotonic phases of slowing down or speeding up (Fig 7). If the new sound stimulus is stopped during this transition, we predict from the model that the BG will still learn the new beat frequency. However, the phase of the BG will differ depending on when during the transition the stimulus is removed. This could be detected using EEG or perhaps a finger-tapping demonstration. This prediction differs from those made by traditional error-correction models which will cease making updates after the stimulus is turned off. Hence, these models may not reach the new frequency, and instead settle at an intermediate frequency. An entrainment model also relies on the external input to match its frequency. Thus if the sound stimulus is stopped during the transition, an entrainment model is likely to return to its basal frequency, and not learn the new one. Third, resynchronization should be faster after a phase shift of the rhythmic stimulus than after a single timing-deviant sound event. Lastly, the model predicts an asymmetry in the resynchronization process after phase shifts (advance versus delay), deviants (early versus late) and tempo changes in the stimulus sequence. We hesitate to assign whether the resynchronization time will be shorter or longer in these cases, as this property depends on the intrinsic dynamics of the BG as defined by its f-Ibias relationship. Similarly, the specific responses of an entrainment model for these cases depends on the phase response properties of the underlying oscillator model. Typically, these are one time adjustments to the phase of the oscillator, followed by a transient return to the entrained solution. Thus rather than elucidating concrete differences in predictions of our BG model versus others, instead consideration of different models may help to narrow the choice of biophysical currents that produce Ibias-f or phase response properties that allow neuron spike times to match empirical data.
There are several questions that we plan to address in our future modeling and behavioral studies. How sensitively do timing errors depend on variability of the gamma counters and, say, on stimulus frequency? To what degree can the BG model track modulations of the beat frequency? Different candidate beat generators that possess different ionic currents will have f-I relationships that are different than the ones presented here. How sensitively do the quantitative results described here depend on the shape of these curves? When considering an ensemble of beat generating neurons, how does coupling between these neurons shape the dynamics of learning? How could the model be enhanced to become predictive, to not just track modulation but to predict dynamic trends? Going beyond isochronous timing only, we plan to consider more complex rhythms. For example, suppose we consider the effect of shifting identically the timing of alternate stimulus tones. Eventually, after a sequence of modest shifts, the beat frequency would be halved although the number of stimulus events would be maintained but with a different temporal pattern. How is the transition of frequency halving executed dynamically? Perhaps there is a regime of shift values where beat determination is ambiguous, a possible regime of bistability. A different manipulation toward a complex stimulus could involve parametrically changing the sound intensity or pitch of alternate tones. Such cases will bring us toward questions of perception and auditory streaming together with beat perception.
The questions surrounding how we perceive and keep a beat are easy to pose but developing models for beat perception and generation present challenges. Our model is a first-pass attempt at formulating and analyzing a neuromechanistic model that can learn a beat. Our approach stems from a neurobiological and dynamical systems perspective to develop neuronal system-based models for beat learning and generation. The essential features involve neuro-based elapsed timekeepers, time difference comparators and a neural oscillator (cellular or circuit level) with some plasticity and learning rules. Looking ahead, one hopes for development of more general beat and rhythm pattern generators (for complex rhythmic sounds, music pieces) that can be stored in a silent mode and are both recallable and replayable.
|
10.1371/journal.ppat.1007191 | Para-cresol production by Clostridium difficile affects microbial diversity and membrane integrity of Gram-negative bacteria | Clostridium difficile is a Gram-positive spore-forming anaerobe and a major cause of antibiotic-associated diarrhoea. Disruption of the commensal microbiota, such as through treatment with broad-spectrum antibiotics, is a critical precursor for colonisation by C. difficile and subsequent disease. Furthermore, failure of the gut microbiota to recover colonisation resistance can result in recurrence of infection. An unusual characteristic of C. difficile among gut bacteria is its ability to produce the bacteriostatic compound para-cresol (p-cresol) through fermentation of tyrosine. Here, we demonstrate that the ability of C. difficile to produce p-cresol in vitro provides a competitive advantage over gut bacteria including Escherichia coli, Klebsiella oxytoca and Bacteroides thetaiotaomicron. Metabolic profiling of competitive co-cultures revealed that acetate, alanine, butyrate, isobutyrate, p-cresol and p-hydroxyphenylacetate were the main metabolites responsible for differentiating the parent strain C. difficile (630Δerm) from a defined mutant deficient in p-cresol production. Moreover, we show that the p-cresol mutant displays a fitness defect in a mouse relapse model of C. difficile infection (CDI). Analysis of the microbiome from this mouse model of CDI demonstrates that colonisation by the p-cresol mutant results in a distinctly altered intestinal microbiota, and metabolic profile, with a greater representation of Gammaproteobacteria, including the Pseudomonales and Enterobacteriales. We demonstrate that Gammaproteobacteria are susceptible to exogenous p-cresol in vitro and that there is a clear divide between bacterial Phyla and their susceptibility to p-cresol. In general, Gram-negative species were relatively sensitive to p-cresol, whereas Gram-positive species were more tolerant. This study demonstrates that production of p-cresol by C. difficile has an effect on the viability of intestinal bacteria as well as the major metabolites produced in vitro. These observations are upheld in a mouse model of CDI, in which p-cresol production affects the biodiversity of gut microbiota and faecal metabolite profiles, suggesting that p-cresol production contributes to C. difficile survival and pathogenesis.
| Clostridium difficile is a bacterium responsible for causing the majority of antibiotic associated diarrhoea outbreaks world-wide. In the United States of America, C. difficile infects half a million people annually. Antibiotics disrupt the natural protective gut microbiota, rendering people susceptible to C. difficile infection, which leads to potentially life-threatening disease and complications. C. difficile is transmitted by spores, which are able to survive in harsh environments for long periods of time. After initial treatment for C. difficile, up to 35% of patients develop the disease again, thus requiring additional and more successful treatment. Here, we use novel techniques to show that C. difficile produces a compound, p-cresol, which has detrimental effects on the natural protective gut bacteria. We show that p-cresol selectively targets certain bacteria in the gut and disrupts their ability to grow. By removing the ability of C. difficile to produce p-cresol, we show that it makes C. difficile less able to recolonise after an initial infection. This is linked to significant alterations in the natural healthy bacterial composition of the gut. Our study provides new insights into the effects of p-cresol production on the healthy gut microbiota and how it contributes to C. difficile survival and pathogenesis.
| Clostridium difficile is a Gram-positive spore-forming enteric pathogen and the leading cause of antibiotic-associated diarrhoea worldwide[1]. C. difficile infection (CDI) ranges from self-limiting diarrhoea to severe and life threatening pseudomembranous colitis[2]. C. difficile spores are the aetiological agent of CDI transmission and are resistant to desiccation, environmental stress, disinfectants and heat[3, 4]. These spores, present in both hospitals and the environment are transmitted via the faecal-oral route, contributing to both nosocomial and community acquired CDI [3]. Infection with C. difficile is frequently preceded by treatment with broad-spectrum antibiotics, which eliminate discrete taxa of the commensal intestinal microbiota resulting in dysbiosis and permitting colonisation by C. difficile. Certain bacterial taxa have been highlighted as important in the prevention of C. difficile colonisation[5–7]. Since restoration of microbial diversity can resolve recurrent infections, faecal transplantation is viewed as an effective treatment strategy[8]. However, a greater understanding of how C. difficile is able to influence the gut microbiota and disrupt intestinal homeostasis is a current imperative.
Human intestinal bacteria have been shown to ferment dietary-derived carbohydrates[9] and proteins[10], producing short chain fatty acids (SCFA), as well as an array of metabolites via fermentation of aromatic amino acids[11]. The secondary metabolites of this highly diverse microbial community have the potential to either positively or negatively influence many aspects of human health [12], with some demonstrated to possess toxic and carcinogenic properties [11, 13, 14]. Aromatic amino acids such as phenylalanine, tryptophan and tyrosine are important sources of phenyl metabolites. These metabolites can be absorbed in the small intestine or pass through to the colon[15] to be excreted in faeces. One such fermentation product, phenylacetic acid (PAA), is the most commonly detected secondary metabolite in healthy human faeces, with reported concentrations of 479 μM[15]. C. difficile ferments tyrosine, via p-hydroxyphenylacetate (p-HPA), to produce p-cresol. Para-cresol is a phenolic compound [16] that has been demonstrated to inhibit the growth of a range of bacterial species and other microorganisms[17, 18]. To date, the capacity to produce p-cresol has only been demonstrated in a select number of organisms[19, 20], including eighteen intestinal commensal species[11]. However, the in vitro production of p-cresol by these species was relatively low (ranging from 0.06–1.95 μg/ml)[11]. Furthermore, C. difficile can tolerate relatively high concentrations (1 mg/ml) of p-cresol[21, 22]. As such, the ability to synthesise and tolerate high concentrations of p-cresol has led to the hypothesis that it may provide C. difficile with a competitive advantage over other microorganisms.
The enzyme responsible for the decarboxylation of p-HPA is a member of the glycyl radical family, 4-hydroxyphenylacetate decarboxylase, which is encoded by three genes hpdB (CD630_01530), hpdC (CD630_01540) and hpdA (CD630_01550), which are co-transcribed in an operon. The hpdBCA operon is highly conserved in all the sequenced C. difficile isolates. We have previously shown that disruption of any of the three genes renders C. difficile incapable of synthesising p-cresol[22]. In this study, we demonstrate that production of p-cresol by C. difficile confers a fitness advantage over other intestinal bacteria both in vitro and in vivo, specifically those with a Gram-negative cell envelope. The treatment of human faecal samples with exogenous p-cresol significantly modified the cultivable bacteria therein, in a species-specific manner. Furthermore, a p-cresol deficient mutant showed a modest but significant reduction in viable counts in a relapse mouse model of CDI. Comparisons of the metabolic and 16S rRNA profiles identified variation in the biochemical and bacterial composition between mice infected with the C. difficile strain 630Δerm and the p-cresol deficient mutant (hpdC::CT) following infection and relapse. This is the first study to show that p-cresol production is a mechanism by which C. difficile confers a competitive advantage over other gut bacteria.
It has been hypothesised that p-cresol production provides C. difficile with a selective advantage over competitors in the human gut. To investigate this, we assessed the effect of exogenous p-cresol on the in vitro growth dynamics of selected intestinal commensal species (S1 Table) compared to C. difficile strain 630Δerm (Fig 1 & S2 Table). The data shows a clear pattern whereby sensitivity to p-cresol correlated with bacterial cell envelope structure. We observed that Gram-positive bacteria were significantly more tolerant to p-cresol than Gram-negative bacteria (Coefficient of variance (COV) = 0.599, p<0.001). Growth of the Gram-negative species, including members of the Bacteroidaceae (Bacteroides thetaiotaomicron) and Enterobacteriaceae (Escherichia coli, Klebsiella oxytoca and Proteus mirabilis) families were inhibited by the addition of exogenous p-cresol in a dose-dependent manner (Fig 1 & S3 Table) and demonstrated a significant decrease in cell growth compared to C. difficile (p<0.005). In contrast, the Gram-positive species including those from the Bifidobacteriaceae (Bifidobacterium adolescentis), Enterococcaceae (Enterococcus faecium) and Lactobacillaceae (Lactococcus fermentum) families displayed no significant reduction in growth rate, even at 0.1% (v/v) p-cresol (Fig 1). Interestingly, E. faecium displayed greater tolerance to p-cresol than C. difficile itself (COV = 0.6 p = 0.002, S3 Table).
We had previously constructed a ClosTron inactivation mutant in the hpdC decarboxylase gene (strain hpdC::CT), which renders C. difficile unable to produce p-cresol[22]. To investigate whether production of p-cresol contributes to fitness in vitro, we performed co-culture assays with 630Δerm and hpdC::CT cultured with a selection of intestinal commensal species, supplemented with exogenously added p-cresol. Brain heart infusion media with yeast extract (BHIS) was chosen for these co-culture experiments, as we have previously shown that intrinsic production of p-cresol under these conditions is negligible[22], therefore any observed effect could be attributed to the exogenously added p-cresol. We observed no difference in growth rate between 630Δerm and hpdC::CT in these conditions [22]. To establish the comparable growth conditions, each species was normalised to the same starting optical density (OD595 0.5) and starting CFU/ml was determined (S4 Table). The competitors were mixed in a 1:1 ratio at matched OD, and were grown for 24 hours in media supplemented with 0.05% (v/v) p-cresol. Viable counts for each species were determined by plating serial dilutions onto media supplemented with and without D-cycloserine and cefoxitin, facilitating differentiation between C. difficile and the competitor. When C. difficile 630Δerm was grown in co-culture with E. coli in the absence of exogenous p-cresol, E. coli was the dominant organism, represented by a significantly higher CFU/ml than C. difficile (8:1 ratio of E. coli to C. difficile) (COV = 1.02, p = 0.003; Fig 2, S5 Table). However, when the medium was supplemented with exogenous p-cresol, the relative proportion of C. difficile increased to a ratio of 1:1, representing an 8-fold increase in the number of viable C. difficile (Fig 2A) (COV = -1.38, p<0.001). A similar profile was observed when E. coli was co-cultured with the hpdC::CT mutant (Fig 2B) (COV = -0.27, p = 0.882). This suggests that 630Δerm and the hpdC::CT mutant displayed comparable fitness when grown in co-culture with E. coli. When C. difficile was grown in co-culture with a Gram-positive bacterium, E. faecium, C. difficile was significantly less abundant compared with the competitor (COV = 3.41, p<0.001). Here, we observed a ratio of 1:10 of C. difficile to E. faecium (Fig 2C & S5 Table). The growth dynamics of the C. difficile hpdC::CT mutant and E. faecium (Fig 2D) were also indistinguishable from the 630Δerm grown in competition with E. faecium (COV = -0.33, p = 0.283). However, when the medium was supplemented with p-cresol, the relative proportion of E. faecium increased significantly (COV = 1.44, p = 0.010). This suggests that the growth conditions were more permissive for E. faecium. However, this was not the case for all Gram-positive species tested. The relative ratio in co-culture of C. difficile (630Δerm and hpdC::CT mutant) to L. fermentum was not significantly altered by exogenous p-cresol (COV = -0.058, p = 0.818) (Fig 2E & 2F and S5 Table). These data indicate that p-cresol had a range of effects on growth dynamics depending on the Phylum of bacteria and their susceptibility to p-cresol.
We observed no difference in competitive fitness between wild type and p-cresol mutant when grown in BHIS supplemented with 0.5% (w/v) p-cresol. Therefore, we developed an additional in vitro competition assay to determine whether intrinsic p-cresol production by C. difficile conferred a competitive advantage over other intestinal commensal species. To achieve this we measured the growth rate of E. coli and K. oxytoca in monoculture and compared this to that of C. difficile (S1A Fig). Under these conditions, C. difficile reached exponential growth at a later time point than the other species tested. Furthermore, we have previously shown that p-cresol is detected in C. difficile cultures at around 4 hours (or OD595 0.5) [22]. In order to limit the dominance of competitor species, and ensure optimal p-cresol production, we grew C. difficile to exponential phase (OD595 0.6) before inoculating the medium with the competitor (at OD595 0.05). We also supplemented the growth medium with p-HPA to drive production of p-cresol. To determine a concentration of p-HPA that resulted in inhibitory p-cresol production, competitive co-culture experiments with C. difficile and E.coli were performed in a range of p-HPA concentrations (0.1%, 0.2% or 0.3% (w/v)) (Fig 3). When the medium was supplemented with 0.1% (6.5 mM) p-HPA we observed a ratio of 13:1 (E. coli:C. difficile) (Fig 3A). When the concentration of p-HPA was increased to 0.2% (13.1 mM), we observed a significant difference (p<0.001) in the ratio of E. coli: C. difficile (1:1), compared to the ratio in 0.1% p-HPA (13:1)(Fig 3A). Further increasing the concentration of p-HPA to 0.3% (19.7 mM) resulted in culture conditions that favoured C. difficile, reflected by a ratio of 1:4 (E. coli:C. difficile) (p<0.001)(Fig 3A). Thus, we observed a positive correlation between the proportion of p-HPA supplemented in the growth medium and the survival of C. difficile compared to E. coli (Fig 3A). To determine whether this effect was linked to the level of p-cresol production, we quantified p-cresol in these culture supernatants by High Performance Liquid Chromatography (HPLC) (Fig 3B). Fig 3B demonstrates that increasing p-HPA concentration correlated with a significant increase in p-cresol production (p<0.001). We observed 25 ±0.04 mM p-cresol when the growth medium was supplemented with 0.3% p-HPA (Fig 3B).
Next, we investigated whether the p-cresol mutant displayed reduced fitness when grown in competitive co-culture with other gut competitor species. Furthermore, we constructed a complement by expressing the hpdC and hpdA genes from a tetracycline-inducible promoter using a plasmid based system (generating strain hpdC::CT::phpdCA). We compared the growth of C. difficile strains 630Δerm, hpdC::CT and the complement (hpdC::CT::phpdCA), in competition with E. coli, K. oxytoca or B. thetaiotaomicron in media supplemented with 0.2% p-HPA (Fig 4). The number of viable counts for each species was determined as outlined above. When 630Δerm was grown in co-culture with E. coli, we observed a 1:1 ratio of C. difficile to E. coli (Figs 4A & 3A). This was consistent with the co-culture assays supplemented with exogenous p-cresol (Fig 2). However, competitive co-culture between hpdC::CT and E. coli resulted in a decrease in the relative proportion of C. difficile to ca. 25% of the total culture (1:4, C. difficile:E. coli). This indicates that the mutant was significantly less viable than the wild type (COV = -1.06, p<0.001). 630Δerm demonstrated comparable relative fitness in competitive co-culture with K. oxytoca (1:1 ratio) (Fig 3B). However, we observed proportionally fewer CFUs of 630Δerm when grown in co-culture with B. thetaiotaomicron (1:4 ratio) (Fig 3C). By contrast, hpdC::CT displayed reduced fitness relative to 630Δerm when grown in competition with both K. oxytoca (COV = -1.40, p<0.001) and B. thetaiotaomicron (COV = -0.79, p = 0.001). This fitness defect was restored when the complement was grown in competition with K. oxytoca and B. thetaiotaomicron (Fig 3). However, complementation of the hpdC mutation, did not restore C. difficile fitness to wild-type levels in competitive co-cultured with E. coli. Therefore, we quantified both p-cresol production and p-HPA utilisation by HPLC (Figs 4D, 3B & S2). Quantification of 630Δerm supernatants grown in both monoculture and competitive co-culture supplemented with 0.2% p-HPA revealed an average p-cresol concentration of 13.3 ±0.1 mM. In contrast, the concentration in supernatants of the complemented mutant (at 0.2% p-HPA) was only 4.8 ±0.2 mM, representing a significant 2.7 fold reduction (p<0.01). Therefore, we conclude that under competitive co-culture conditions, 4.8 ±0.2 mM p-cresol was sufficient to have a deleterious effect on the growth of K. oxytoca and B. thetaiotaomicron, but not on E. coli. Increasing the concentration of the transcriptional inducer (anhydrotetracycline) and p-HPA resulted in increased p-cresol production (from 4.8 ±0.2 mM to 15.6 ±3.9 mM) by the complement and restoration of the phenotype (S2 Fig). Here, the level of p-cresol production directly correlated with the concentration of the transcriptional inducer (S2C and S2D Fig). As expected, this suggests that complementation was more greatly influenced by transcript expression rather than availability of the p-HPA precursor. Furthermore, we observed no difference in growth rate between the three C. difficile strains at any tested concentration of anhydrotetracycline (S1 Fig). Taken together, these data suggest that production of p-cresol by C. difficile confers a competitive growth advantage over susceptible bacterial species (specifically, Gram-negative species) under our in vitro conditions.
To further understand the effect of p-cresol production on the interaction between C. difficile and these intestinal species, we characterised the metabolic content of both mono-culture and competitive co-culture supernatants using 1H nuclear magnetic resonance (NMR) spectroscopy. We analysed the culture supernatants described in Fig 4 and performed principal component analysis (PCA) to identify metabolic variation across the profiles. The scores plot from the PCA model comparing all profiles showed that the largest variation in the metabolic data (Principal Component 1 (PC1)) was between the C. difficile strain 630Δerm samples and those from the hpdC::CT strain (Fig 5A). The metabolic profiles from the complement samples clustered between 630Δerm and hpdC mutant samples. The loadings for PC1 describe the metabolites varying between the strains. This indicated that culture supernatants from 630Δerm contained significantly greater amounts of p-cresol and alanine compared to the other strains, but lower amounts of p-HPA, butyrate and isobutyrate (Fig 5B). This was consistent with the notion that p-HPA is being depleted in order to produce p-cresol. The 630Δerm samples clustered together regardless of whether the bacteria were grown in mono-culture or co-cultured with E. coli, K. oxytoca or B. thetaiotaomicron. In contrast, hpdC::CT and complement (hpdC::CT::phpdC-A) strains were separated in the second principal component (PC2) based on the competitive co-culture conditions. The loadings for PC2 indicated that the mono-cultured C. difficile and B. thetaiotaomicron competitive co-culture samples contained lower amounts of acetate compared to the competitive co-cultures from E.coli and K. oxytoca. This metabolic variation between strains and competitive co-culture conditions is summarised in the clustergram shown in Fig 5B, which was constructed from the Z-scores of 1H NMR peak integrals measured for each metabolite across all samples. The dendrogram shows that the 630Δerm metabolic profiles clearly cluster away from those of the other two strains (Fig 5) and the variation in p-cresol production between samples is apparent in the 1H NMR spectrum (S3 Fig). The dendrogram also showed that p-cresol and alanine clustered together as did butyrate and isobutyrate (Fig 5B). We also assessed the effect of altered p-HPA concentration on metabolic profiles. The PCA of metabolites produced in media supplemented with 0.1% and 0.2% p-HPA demonstrated no clear metabolic variation between these samples. However, we did observe clustering within the 0.1% p-HPA samples, driven by increased p-cresol and alanine (S4 Fig). These data suggest that under these growth conditions p-cresol is one of the most abundant metabolites in culture supernatants. This is reflected by both metabolic profiling and HPLC quantification, which correlates to p-cresol susceptibility observed in both competitive co-culture and monoculture of Gram-negative bacteria.
Our results have demonstrated that p-cresol production confers a fitness advantage over discrete bacterial species in vitro. Therefore, we sought to determine whether this was also true in vivo. Individually caged C57BL/6 mice were infected in parallel with 1x104 spores of C. difficile strain 630Δerm (n = 5) or the hpdC::CT mutant (n = 5) and compared to uninfected naïve control mice (n = 5), in a relapse model of infection. Mice were given cefoperazone in their drinking water for 10 days to stimulate gut dysbiosis, before infection by oral gavage with C. difficile spores (Fig 6A)[23, 24]. Stool samples were collected throughout the experiment for analysis. Twenty-eight days post-infection, mice were treated with vancomycin in their drinking water for 7 days to encourage recurrence of infection (Fig 6A). Infection was monitored by enumeration of spores isolated from faeces on C. difficile selective media (Fig 6B). We observed no significant difference in the number of spores recovered from faeces of mice infected with either 630Δerm or hpdC::CT following cefoperazone treatment, indicating that the hpdC::CT mutant and 630Δerm were equally competent at initial colonisation (Fig 6C). This was consistent with the notion that these strains demonstrate similar resistance to cefoperazone and vancomycin (S6 Table) and no differences in sporulation in vitro (S5 Fig). However, at day 7 post-infection we observed a modest difference in colonisation, represented by significantly fewer 630Δerm CFUs compared to the hpdC::CT mutant (p<0.05). Relapse was detected by enumeration of C. difficile spores post-vancomycin treatment. Three days following removal of vancomycin (D3R), spores were detected in all but one 630Δerm infected mouse and three out of five hpdC::CT mutant infected mice. By day 4 post-relapse (D4R) C. difficile spores were recovered from the faeces of all mice and we observed a significant reduction (p<0.05) in the number of spores recovered from the faeces of mice infected with the hpdC::CT mutant relative to 630Δerm infected mice (Fig 6D). Both infections followed a broadly comparable progression, however, we observed modest but significant differences in the number of spores recovered both pre- and post- relapse between 630Δerm infected and hpdC::CT infected mice at discrete time points. Interestingly during relapse, the spore density remained lower in hpdC::CT compared to 630Δerm, despite an initially higher CFU at D7 post-infection, indicating that these differences are a result of in vivo fitness.
Given that the hpdC::CT mutant displayed an altered colonisation profile during relapse and that p-cresol displays bacteriostatic properties against a number of species, this led us to postulate that production of p-cresol may alter the composition of the intestinal microbiota in such a way that favoured C. difficile re-colonisation. We isolated bacterial DNA from four key time points during the relapse model of CDI; day 7 post-cefoperazone treatment, immediately upon cessation of vancomycin treatment (D0R), day 2 post-relapse (D2R) and day 4 post-relapse (D4R), when all the mice were colonised with C. difficile (1.6 x 106 WT and 2.3 x 105 hpdC::CT mutant spores/g faeces). To assess the community structure of the microbiota, 16S rRNA sequencing was performed by sequencing the V5-V7 region of 16S rRNA gene. The data was grouped with distance-based similarity of 97% into operational taxonomic units (OTUs), using Greengenes and associated summaries and diversity analyses were performed in QIIME. Consistent with previous studies, the microbiota of cefoperazone-treated mice was dominated by Lactobacillaceae (Fig 7A and S7 Table)[24–26], which comprised 39.7% (mean relative abundance) of the total microbiota in 630Δerm infected mice. The microbiota was also populated by Bacteroidetes, including members of the S24-7 (an uncultured commensal of homeothermic animals[27]) (17.6%) and Paraprevotellaceae (1.2%) families. Furthermore, Firmicutes, including Staphylococcaceae (12.75%), other Clostridiales (7.2%), Lachnospiraceae (4.2%), Erysipelotrichaceae (3.7%), Ruminococcaceae (2.9%), Enterococcaceae (2.2%), Turicibacteraceae (1.2%), and Actinobacteria including Bifidobacteriaceae (1.9%) (Fig 7A) also contributed to the microbiota composition. However, animals infected with the hpdC::CT mutant demonstrated a significant increase in microbial diversity at D7 (ANOVA p<0.05), compared to 630Δerm infected and naïve mice (Fig 7B and S7 Table), which is also upheld with an ANOSIM population analysis p<0.05 (S6 Fig). Consistent with the notion that p-cresol prevents outgrowth of Proteobacteria, the majority of the families that were only present in the hpdC mutant infected animals were members of the Proteobacteria Phylum (S7 Table), albeit at low abundance.
Treatment with vancomycin significantly reduced diversity of both uninfected (naïve) and infected mice (630Δerm and hpdC::CT) (Fig 7B & 7C), resulting in a dramatic increase in the relative representation of Lactobacillaceae, specifically the Lactobacillus genus, which constituted ≥85% of the microbiota of all the mice examined (90% 630Δerm, 87% hpdC::CT mutant and 85% naïve at D0R). Principle component analysis demonstrated clustering of D0 and D2 post relapse (Fig 7C). At D2R, the diversity of the microbiota remained low. At D4R, partial recovery of the microbial diversity was observed (Fig 7A and 7B), which coincided with the detection of C. difficile spore in faeces (Fig 6D). At D4R there were distinct differences in population composition in the intestinal bacteria of these animals (Fig 7D). We observed an increase in spread on the PCA plot at D4R (Fig 7C), compared to the clustering observed at D0R and D2R. Although the mean relative proportions of Lactobacillaceae were similar (73.9% in 630Δerm and 72.3% in the hpdC::CT mutant infected animals), there were clear differences in the representation of other taxa, including other Firmicutes, Proteobacteria and Bacteroidetes (Fig 7D). In 630Δerm infected mice there was an increase in representation of Firmicutes from the Erysipelotrichales (16.5% 630Δerm and 0.1% hpdC::CT mutant), Bacillales (1.9% 630Δerm and 0% hpdC::CT mutant) and Clostridiales orders, and the Bacteroidetes (Fig 7D). Conversely, in the p-cresol mutant infected mice, we observed an increase in the representation of Proteobacteria, specifically, the Gammaproteobacteria of the Pseudomonadales (5.5% 630Δerm and 18% hpdC::CT mutant) and Enterobacteriales (0% 630Δerm, 6.75% hpdC::CT mutant) order, and the Betaproteobacteria of the Burkholderiales order (Fig 7D). Consistent with the notion that Gram-negative species were more susceptible to the effects of p-cresol, Gammaproteobacteria formed 26.2% of the total microbiome in hpdC::CT infected animals D4R, compared with 5.5% in 630Δerm infected mice (COV = 9.37, p = 0.023), suggesting that p-cresol may inhibit their outgrowth following treatment with vancomycin.
Our data suggest that p-cresol production by C. difficile influenced the composition of the mouse faecal microbiota. Therefore, we investigated whether these differences resulted in an altered biochemical profile. Stool samples collected throughout the duration of the initial mouse infection (at day 2 (D2), 4 (D4) and 7 (D7)) and during relapse (at days 0 (D0R) and 4 (D4R)) were analysed using 1H NMR spectroscopy (Fig 8). The PCA scores plot identified biochemical variation between the faecal profiles collected D2-D7 versus D0R-D4R in the control mice and those infected with the hpdC::CT mutant. The D2-D7 samples contained greater acetate compared to the relapse time points, and lower amounts of an unknown metabolite (δ 3.59, singlet). Vancomycin induced perturbations in the metabolic activity of the intestinal bacteria are likely to underlie these changes[28].
The faecal profiles from mice infected with the 630Δerm strain showed similar metabolic alterations to the control mice and those infected with the mutant strain during the initial infection (D2-D7). However, the response was different 4 days after relapse. At D4R, the faecal profiles from mice infected with the 630Δerm strain were more variable than the uninfected mice and those infected with the mutant strain and were similar in composition to the initial infection profiles (Fig 8).
We have shown that p-cresol production has deleterious effects on the outgrowth of Gram-negative bacterial species both in vitro (Figs 1–4), and in an in vivo mouse infection model (Fig 7). Thus, we sought to determine the effect of exogenously added p-cresol on biodiversity of the human microbiome. Healthy human stool samples were taken from donors ranging from 60–65 years old, who had not received antibiotic treatment in the last 3 months, eliminating possible perturbations by antibiotic therapy. We measured the effect that exogenously-added p-cresol (at 0.1% and 0.3%) had on faecal microbiota, compared to a Phosphate Buffered Saline (PBS) control. Differential plating revealed that the facultative anaerobes were particularly sensitive to p-cresol at both 0.1% (COV = -0.61, p = 0.006) and 0.3% (COV = -1.82, p<0.001), represented by a significant reduction in viable counts (Fig 9). The Bacteroides fragilis group was also significantly reduced after exposure to both 0.1% (COV = -1.29, p = 0.009) and 0.3% (COV = -4.39, p<0.001) p-cresol (Fig 9). The total anaerobes and lactose-fermenting Enterobacteriaceae were also significantly reduced after exposure to 0.3% p-cresol (COV = -1.48, p<0.001, COV = -2.36, p<0.001, respectively) (Fig 9). Consistent with the mouse model of CDI, p-cresol at 0.1% had a limited effect on the survival of Lactobacillus (COV = -0.045, p = 0.890) and Bifidobacterium species (COV = -0.100, p = 0.642). However, a significant decrease in survival was observed for both groups when they were incubated in 0.3% p-cresol (p<0.01). In line with our in vitro co-culture data (Fig 2), Enterococcus species present in human faecal samples were not adversely affected by the addition of p-cresol (Fig 8), even at the highest concentrations tested (COV = 0.48, p = 0.873).
Given that we observed a clear distinction in the nature of species that displayed tolerance to p-cresol, we reasoned that the cell envelope would be an obvious target for its mode of action. Phenolic compounds that target membranes typically induce a rapid loss of low molecular weight compounds from within the cell as a result of increased membrane permeability[29–31]. Thus, we used the release of inorganic phosphate (Pi) as a metric for determining membrane integrity in the presence of p-cresol. Initially, we compared the release of phosphate from E. coli and C. difficile strains (630Δerm and hpdC::CT) in increasing concentrations of p-cresol (Fig 10A). We observed a significant increase in the amount of phosphate released by E. coli compared to C. difficile (COV = 0.868, p = 0.005). Only 16% of the total intracellular phosphate of C. difficile was released upon contact with p-cresol. Furthermore, the hpdC::CT mutant displayed a similar phosphate release profile to 630Δerm C. difficile (COV = 0.201, p = 0.444), which was not significantly different (Fig 10A). This indicates that disruption of p-cresol production had little bearing on p-cresol tolerance, under these conditions. Our data suggests that cells can tolerate p-cresol up to a threshold level (~0.4% v/v), after which the amount of phosphate released becomes saturated. Therefore, we selected a concentration of 0.3% (v/v) to determine phosphate release over time in a selection of Gram-positive and Gram-negative gut bacteria. Membrane integrity was measured in the presence of p-cresol by comparing the amount of p-cresol induced phosphate release, to the total intracellular phosphate, which was determined by boiling cell suspensions for 15 minutes. Fig 10 demonstrates that species with a Gram-positive cell envelope display greater tolerance to p-cresol than Gram-negative species, represented by significantly less phosphate release (COV = -2.478, p<0.001), (Lactobacillales: E. faecium (p = 0.005) and L. fermentum (p = 0.003), the Bifidobacteriales: B. adolescentis (p = 0.01) and the Clostridiales: C. difficile (p<0.01)) corroborating previous observations. Conversely, the Gram-negative Gammaproteobacteria: P. mirabilis, E. coli and K. oxytoca released their total intracellular pool of phosphate over the course of the assay. P. mirabilis and K. oxytoca released 68% and 60% of their total phosphate respectively immediately upon contact with p-cresol (Fig 10B). Both species released >90% of their total phosphate pool following 30 minutes contact with p-cresol. In contrast, no Gram-positive species analysed released their total intracellular phosphate pool over the course of the assay (Fig 10C). However, B. adolescentis released 63% of its total phosphate at 30 minutes compared to 20% for E. faecium, 27% for L. fermentum and 33% for C. difficile, indicating that B. adolescentis is more sensitive to p-cresol than other Gram-positive species. Prolonged exposure to p-cresol resulted in other Gram-positive species releasing a greater portion of their intracellular pool of phosphate, however, the level of Pi released by C. difficile never exceeded its initial level of release (Fig 10). In conclusion, our data demonstrate a clear correlation between bacterial cell envelope structure and susceptibility to p-cresol.
The indigenous microbiota has been shown to form an ecological barrier that prevents the ingress of pathogenic bacteria such as C. difficile [32]. However, the specific components of the intestinal microbiota that facilitate colonisation resistance are only recently becoming clear [5–7, 25, 33–36]. Both the treatment with broad-spectrum antibiotics and the availability of specific metabolites has been shown to play a role in the expansion of particular bacterial species within the human microbiota [37, 38]. Here, we present compelling evidence that C. difficile may directly modify the intestinal microbiota through production of p-cresol. We demonstrate that C. difficile displays a greater degree of tolerance to p-cresol compared to other common intestinal species, including the Gammaproteobacteria: E. coli, K. oxytoca and P. mirabilis, as well as the Bacteroidetes, B. thethaiotaomicron. We show that these bacterial species are susceptible to the effects of both endogenous and exogenous p-cresol, which was reflected in reductions of viable counts when these intestinal microbiota species were grown in competitive co-culture with C. difficile. Using a plasmid based complementation system to restore the expression of the p-HPA decarboxylase, we have shown that p-cresol production by C. difficile must exceed 5 mM to elicit a significant alteration in competitive growth dynamics. We have shown that C. difficile is able to utilise all the available p-HPA supplemented in the growth medium, which results in the production of up to 25 ±0.04 mM p-cresol in vitro (Fig 3), which is 1000-fold more than the amount of p-cresol produced from tyrosine metabolism by other organisms cultured from the intestinal microbiota (range 0.06–1.95 μg/ml)[11]. There is evidence that p-HPA is present in the human colon and detected in healthy human stool samples at 19 μM[15], therefore C. difficile can potentially utilise free tyrosine and p-HPA to produce p-cresol in vivo.
We expanded our investigation of the influence of p-cresol on the growth of other bacterial species, to identify other metabolites influencing growth. In particular, alanine, p-cresol, acetate, butyrate, isobutyrate and p-HPA were the six main metabolites that were differentially modulated in mono-culture and co-culture of C. difficile with intestinal bacteria. The abundance of these metabolites in vitro was altered in the presence of the p-cresol mutant compared to C. difficile strain 630Δerm. Acetate and butyrate are the most common end products of fermentation in the gut[39]. C. difficile can use amino acids as the sole energy source via Stickland fermentation, in which amino acid acceptors (such as glycine, proline and hydroxyproline) are reduced in a paired metabolism with electron donors (such as leucine, isoleucine or alanine). This can result in the conversion of alanine to acetate [40]. However, in C. difficile monoculture, we did not observe an inverse association between acetate and alanine (Fig 5B), suggesting that C. difficile is not utilising alanine in stickland fermentation under nutrient rich conditions (in BHIS media). This suggests that the competitor species, may have been responsible for the increased utilisation of alanine in co-culture with the p-cresol mutant and complement, where the competitor is more abundant. The reduction of the Stickland acceptors glycine and proline in C. sticklandii and C. difficile requires two selenium dependant reductases, glycine reductase and D-proline reductase [40, 41], highlighting the importance of selenium in growth and metabolism in C. difficile, particularly in glycine reduction and selenocysteine production [40, 42]. Nutrient availability has been linked to virulence in C. difficile in a number of different ways, via the global transcriptional regulators CodY, CcpA, PrdR and Rex, which are involved in overlapping cellular processes including toxin production, amino acid biosynthesis, stickland fermentation, nutrient transport, fermentation and cell membrane components[43]. The hypervirulent C. difficile strains (RT027 and RT078) have also developed the ability to metabolise low concentrations of trehalose, via acquisition of a single point mutation in the trehalose repressor (treA), which increases virulence of these ribotypes in vivo[44]. C. difficile is also capable of utilising ethanolamine as a carbon source[45] and the ethanolamine genes are upregulated in vivo in the presence of B. thetaiotaomicron when animals were fed on a standard polysaccharide diet [38].
Cysteine is involved in amino acid and energy metabolism in C. difficile [46], modulating processes such as carbon transfer, electron transport, butyric acid and butanol production. Cysteine results in increased levels of intracellular tyrosine [47]. Cysteine also down-regulated 4p-hydroxyphenylacetate-3-hydroxylase [48], which may reduce p-HPA availability for the p-HPA decarboxylase and thus would decrease flux to p-cresol. Therefore, cysteine-regulated pathways may result in increased p-cresol production. This is consistent with the notion that the production of butyrate and p-cresol are inversely regulated.
C. difficile and other opportunistic gut bacteria have developed metabolic strategies that differ in response to environmental signals, one such strategy is the production of short chain fatty acids [38]. Studies using a simplified gnotobiotic mouse model, have shown that succinate produced by B. thetaiotaomicron is used by C. difficile to produce butyrate, boosting C. difficile titres [38]. However, in our experimental conditions (nutrient rich conditions) succinate was detected at very low levels by 1H NMR spectroscopy. Butyrate has anti-inflammatory properties [49] and is produced by a diverse array of bacterial phyla [50]. Yet, butyrate stimulates C. difficile toxin production in the absence of rapidly metabolised carbohydrates (e.g. glucose) [48]. The production of butyrate from acetyl-CoA or succinate by C. difficile is suppressed by the transcriptional regulators CcpA, CodY and Rex [51], however, if proline is limited then alternative pathways for NAD+ regeneration are used including glycine reductase, alcohol dehydrogenase and butyrate production from acetyl-CoA and succinate are induced [51]. We observed lower butyrate and isobutyrate concentrations in the 630Δerm cultures (mono- and co-cultures), which implies that C. difficile is unable to synthesise butyrate under these conditions. However, we observed high butyrate and isobutyrate concentrations in co-culture with the p-cresol mutant and complement, where the competitor is more abundant, suggesting that the competitors are responsible for the increase in butyrate in co-culture.
Inter-C. difficile strain variation in the metabolic profiles included altered abundance of alanine, isobutyrate, p-cresol and p-HPA. We observed an increase in the production of p-cresol in monocultures and co-cultures containing the C. difficile 630Δerm and an absence of p-cresol in all cultures with the hpdC::CT mutant, consistent with the inability of the mutant to synthesise p-cresol (Fig 5B). The complement produced an intermediate amount of p-cresol (Figs 5B and S3) and therefore clustered between the wild type and mutant in the PCA plot, suggesting that p-cresol and p-HPA were the main metabolites driving separation between C. difficile strains. This observation was corroborated when p-cresol production was increased in co-culture by increasing the supply of p-HPA, which resulted in a decrease in the competitor relative to wild-type C. difficile.
In this study, we present evidence that p-cresol production by C. difficile prevents outgrowth of discrete taxa of bacteria in vitro and that p-cresol production may modulate composition of the mouse microbiota. Our data showed clear differences in the composition of the microbiota of mice infected with the p-cresol mutant compared with the 630Δerm strain pre- and post-relapse in our infection model (Fig 7). Recovery of the microbial community to its pre-dysbiotic state is often a slow process and, consequently, susceptibility to C. difficile colonisation can be increased for weeks and even months following cessation of antibiotics[52, 53]. Both the 630Δerm and the hpdC::CT mutant successfully colonised mice at the initial infection stage. However, we observed increased microbial diversity in hpdC::CT infected mice, at day 7 post-infection. This diversity was largely driven by OTUs that each constituted <0.1% of the total microbiota. Despite relatively low abundance of these OTUs, they could have important consequences for the microbial ecosystem. These Families include Corynebacteriaceae, Propionibacteriaceae (both Actinobacteria), Bradyrhizobiaceae, Burkholderiaceae, Comamonadaceae, Oxalobacteraecea, Rhodocyclaceae, Bdellovibrionaceae and Enterobacteriaceae (all Proteobacteria). Consistent with the notion that p-cresol prevents outgrowth of Proteobacteria, the majority of these Families were members of the Proteobacteria Phylum. Vancomycin treatment reduced microbial diversity, altered the metabolic content of the stool samples, and resulted in a microbiome that was susceptible to relapse with C. difficile. The remaining microbial community was dominated by Lactobacillaceae (consistent with previous publications [24]), which was insufficient to restore colonisation resistance following vancomycin withdrawal. Upon cessation of vancomycin treatment, we observed an expansion of microbial diversity (D2R and D4R). There were clear differences in microbiota composition at D4R between animals infected with the p-cresol mutant and 630Δerm C. difficile. The second most abundant class present in the microbiome of hpdC mutant-infected mice was Gammaproteobacteria (26.2%). In contrast, the second most abundant class in animals infected with 630Δerm C. difficile was the Erysipelotrichia (16.5%). Other studies have shown that without FMT, dysbiosis is maintained in mice with two main OTUs, Lactobacillus and Turicibacter (Erysipelotrichia order)[24, 35], which we observed in the 630Δerm infected mice, but not mice infected with the p-cresol mutant. We have shown in vitro that Gammaproteobacteria, including K. oxytoca, E. coli and P. mirabilis are more sensitive to p-cresol than C. difficile, while Gram-positive bacteria from the Lactobacillales family are more resistant to p-cresol. This is particularly pertinent as the majority of the microbiota post-vancomycin treatment was comprised almost exclusively of Lactobacillus and an increased expansion of the Gammaproteobacteria was only seen in p-cresol mutant-infected mice. The faecal metabolic profiles from all animals post-vancomycin treatment (D0R) were clearly distinct from those collected post-infection (D2, D4 and D7; Fig 7). However, the metabolic profiles of mice infected with the 630Δerm strain 4 days after withdrawal of vancomycin were more variable than those infected with the mutant strain. The 630Δerm infected mice also had a metabolic signature more similar in composition to those samples collected post-infection (D2, D4 and D7). This suggests that the biomolecular perturbations following re-establishment of infection with 630Δerm were more closely related to those observed with the initial infection. In contrast, the hpdC::CT-infected mice had profiles more closely related to the uninfected mice.
To complement the in vitro co-culture assay and mouse model of CDI we assessed the effect of exogenous p-cresol on the human microbiota using ex vivo healthy human faecal samples. We observed a reduction in the number of viable total anaerobes, facultative anaerobes and lactose fermenting enterobacteriacea (LFE). The LFE are comprised of the Gammaproteobacteria E. coli, Klebsiella spp, Enterobacteria spp, Citrobacter spp and Serratia spp. This observation corroborates previous findings that demonstrate a significantly reduced viability of Gram-negative bacteria by in vitro growth kinetic analysis, competitive co-culture and in a mouse model of CDI. In contrast, the Gram-positive bacteria isolated from the ex vivo healthy human faecal samples (Bifidobacteriaceae, Lactobacillales and Enterococcaceae) were consistently less sensitive to p-cresol in the assays we performed.
In this study, we have shown that C. difficile displays a greater degree of tolerance to p-cresol when compared to a selection of other common intestinal bacterial species. Our data suggest a clear distinction between the fundamental properties of the organisms susceptible to the p-cresol, whereby Gram-positive species displayed greater tolerance than Gram-negative species. We demonstrate that p-cresol affects the integrity of surface barriers resulting in a concentration-dependant leakage of small molecules such as phosphate. Similar effects have been observed with m-cresol and chloro-cresol on bacterial cell membranes[29]. p-cresol was recently shown to inhibit proliferation of colonic epithelial cells and induce necrotic leakage of protons through the inner mitochondrial membrane[54]. Pseudomonas putida strain P8, which has the capacity to degrade p-cresol, modifies its fatty acid composition by increasing the abundance of 9-trans hexadeconoic acid and decreasing the abundance of 9-cishexadeconoic acid when grown in the presence of sub-lethal concentrations of phenol[30, 55, 56]. Previous work has demonstrated that sublethal concentrations of phenolics, including p-cresol, resulted in an increase in the degree of saturation of cell membrane lipids, which is thought to counteract the increase in membrane fluidity [30, 57].
In conclusion, we demonstrate that the production of p-cresol by C. difficile alters the composition and recovery of diversity in the intestinal microbiota. A p-cresol deficient mutant has a reduced ability to compete with other intestinal microbiota species in vitro. We have shown that the effect of p-cresol is more detrimental to the growth of Gram-negative bacteria, differentially inhibiting proliferation of various bacterial Phyla. Exposure to p-cresol resulted in release of cellular phosphate, suggesting that it disrupts cell envelope integrity. This study provides evidence that p-cresol production by C. difficile provides it with a competitive survival advantage over other intestinal bacterial species.
C. difficile strains 630Δerm[58] and hpdC::CT[22] have been previously described. The intestinal microbiota species used in the study were obtained from Mark Wilcox and Simon Baines at the University of Leeds isolated from a gut soup model of CDI (S1 Table). All bacteria were cultured in pre-reduced Brain Heart Infusion (BHI) (Oxoid), supplemented with 0.5% (w/v) yeast extract (BHIS) and 0.05% (w/v) L-cysteine (Sigma), at 37°C and under anaerobic conditions. For growth rate analysis, our collection of gut bacteria was grown in 100 ml tissue culture flasks with shaking at 50 rpm, 37°C and under anaerobic conditions. Pre-reduced growth media was supplemented with 0.1%, 0.05% and 0.01% (v/v) p-cresol as indicated. OD595 was determined every hour for 8 hours with a final reading at 24 hours, growth curves were performed in triplicate. A p-cresol complement strain (hpdC::CT::phpdCA) was made, using an inducible plasmid based system derived from pRPF185[59] (S2 Table). The hpdCA genes were PCR amplified and cloned downstream of a tetracycline inducible promoter (ptet) in pRPF185 [59] to produce the plasmid phpdCA. This was then conjugated into the hpdC::CT mutant to create a complement[60]. This was performed alongside an empty plasmid pLDempty, which was derived from pRPF185[59], to contain the ptet promoter, but without a gene (S2 Table). This was transferred into C. difficile using competent E.coli CA434[60] into both 630Δerm and hpdC::CT mutant as controls. Linear regression analysis was performed using Stata15; data was transformed using Log10 to approximate a normal distribution. The data was mined to determine if there was a significant difference in; a) strains, the growth of all bacteria strain compared to the reference strain C. difficile strain 630, b) p-cresol concentration compared to the BHIS untreated control, c) the Gram-negative bacteria compared to the Gram-positive bacteria. The COV indicates whether the growth is higher (positive number) or lower (negative number) than the reference and the p-value indicates the probability, a minimum cut off of p<0.05 was used throughout for significance (S3 Table).
Culture supernatant from mono-culture and co-culture experiments in media supplemented with 0.1 and 0.2% (v/v) p-HPA were filter sterilised. Samples were diluted into 400 μL of phosphate buffer (pH 7.4, 100% D2O, 3 mM of NaN3, 1 mM of 3-(trimethyl -silyl)-[2,2,3,3-2H4]-propionic acid (TSP) for the chemical shift reference at δ0.0) according a 1:2 ratio. Samples were transferred to 5 -mm tubes for 1H nuclear magnetic resonance (NMR) spectroscopic analysis, which was performed on a Bruker 600 MHz spectrometer (Bruker Biospin, Karlsruhe, Germany) at 300K (26.85°). The parameters of the acquisition were as previously reported for urine[61]. Each spectrum was acquired with 4 dummy scans followed by 32 scans. Spectra were automatically phased, baseline corrected and calibrated to the internal standard (TSP) using Topspin (Bruker Biospin, Karlsruhe, Germany). The processed spectral data was imported into Matlab (version R2014a, The Mathworks Inc.). The region δ4.84–4.76 was removed to eliminate the residual water signal. Principal Components Analysis (PCA) was performed using pareto scaling, due to the significant intensity of the acetate signal (δ 1.92, single. Based on the PCA loadings, spectral peaks contributing to the principal components were integrated using an in-house script. These metabolite peak integrals were used to construct a clustergram in Matlab using the clustergram script.
Primary cultures were inoculated from a single colony of C. difficile strains (630Δerm, hpdC::CT, hpdC::CtphpdCA) into pre-reduced BHIS broth and grown to an OD595 0.3 on a shaking platform at 50 rpm. These were inoculated 1/100 into pre-equilibrated BHIS broth which was incubated statically for 72 h under anaerobic conditions at 37°C. Total counts (vegetative cells and spores) and spore counts were then determined using CFU assays in 1X PBS (1/10 dilutions from 0 to -5). All dilutions were plated onto BHIS plates supplemented with 1% taurocholate. The spore counts were performed by heat inactivation of vegetative cells at 65°C for 20 minutes, these were then serially diluted and CFU counts determined on BHIS taurocholate plates. All experiments were performed with duplicate technical replicates and triplicate biological replicates. All data was analyzed in Excel, plotted in GraphPad Prism 7 and statistical analysis was performed in Stata15 using regression analysis p<0.05 were considered significantly different (S5 Fig).
Female C57BL/6 mice (Charles River; 7–9 weeks old) were kept in independently ventilated cages under sterile conditions. As outlined by Theriot et al.[62], mice were treated with cefoperazone in the drinking water (50 mg/litre) for 10 days to disrupt their normal microflora, rested for two days, before they were infected with 104 C. difficile spores by oral gavage. After 28 days, vancomycin was added to the drinking water for 7 consecutive days (400 mg/litre) to induce relapse of CDI. Fresh faecal samples from individually infected mice were collected throughout the time course to be utilised for determining the C. difficile load, 16S rRNA sequencing of the microflora and metabolite profiling by 1H spectroscopy. Stool samples were plated onto C. difficile selective plates to determine the bacterial load (CFU/g). Statistical analysis was performed using a one tailed Mann Whitney U test, p<0.05 were considered significantly different.
All animal procedures were performed at Royal Holloway in accordance with the Home Office project license PPL 70/8276, that enables work to be conducted under the UK “Animal (Scientific Procedures) Act 1986”. This work was approved by the Royal Holloway, University of London Ethics Committee.
Healthy human donor faecal samples were collected and processed using different healthy donors who had not received antibiotic treatment in the preceding 3 months in accordance with the University of Hertfordshire Ethics committee guidelines and approval (UH Ethics Approval Number: aLMS/SF/UH/00103). All donors provided informed written consent.
DNA was extracted from faecal samples using a combined method based on phenol:chloroform:isoamyl alcohol extraction, ethanol precipitation and FastDNA SPIN kit for soil (MPBiomedicals). Briefly, an equal mass of faecal material was suspended in 50 mM Tris-HCl pH7.5, 10 mM EDTA, homogenised and bacterial cells were lysed using a FastPrep-24 Classic Instrument (MPBiomedicals). Nucleic acid was extracted using a standard phenol:chloroform:isoamyl alcohol procedure, followed by ethanol precipitation and was suspended in nuclease free dH2O. Faecal DNA was subsequently purified using the DNA binding matrix from the FastDNA SPIN kit for soil with minor modifications of the manufacturer’s instructions. Briefly, DNA samples were added to sodium phosphate buffer, MT buffer and protein precipitation solution supplied in the kit and this was added directly to the binding matrix. DNA was subsequently purified according to the manufacturer’s instructions and eluted in DNase-free water. Library preparations for the MiSeq were performed as outlined in Rosser et al[63]. Briefly, an amplification step was used to add Illumina compatible adaptors, with a unique 12 bp individual barcodes for each sample, with an extra pad and linker sequence. The V5-7 regions of the 16S rRNA genes were then amplified using 785F: 5ʹ-GGATTAGATACCCBRGTAGTC-3ʹ, 1175R: 5 ʹ-ACGTCRTCCCCDCCTTCCTC-3ʹ primers, where the reverse primer (1175R) contained the individual error-corrected barcode: 25 μl reactions were comprised of 1x Molzym PCR buffer, 0.025 μM Moltaq (Molzym), 200 μM dNTPs (Bioline), 0.4 μM forward and reverse primer, 2 μl DNA and nuclease free water (Bioline)[63]. Cycling parameters for each reaction were 94°C x 3 min, then 30 cycles of 94°C x 30 s, 60°C x 40 s, 72°C x 90 s and final extension at 72°C for 10 min. Samples were purified and normalised using a SeqPrep normalisation plate kit (Invitrogen), and quantified using a Qubit2.0 (Life technologies), and further purified using 0.6 X Agencourt AMPure Beads (Beckman Coulter), a selection of samples were run on an Agilent high sensitivity DNA chip (Agilent Technologies), samples were quantified again using a Qubit 2.0 (Life Technologies), and were pooled in equimolar solution, then diluted to a 2 nM library, with 10% PhiX control and loaded into the MiSeq run cartridge in accordance with the manufacturer’s instructions (Illumina). The MiSeq runs produced 250 bp paired end reads, with a 12 bp individual index for each sample. The sequence reads generated were de-multiplexed and quality filtered using QIIME (version 1.9.1 [64]) following the standard pipeline to assign Illumina reads to operational taxonomic units (OTUs) using the Greengenes database [65]). Associated summaries and diversity analyses were also performed in QIIME. Subsequent analyses were performed in R [66] and visualised with ggplot2 [67]. We selected families to include in our 16S plots (Fig 7A) if they had mean proportion of greater than 1% in any of the 12 day/type phenotype combinations. Box plots (Fig 7B) and PCA (Fig 7C) were calculated from the full 16S rRNA sequence dataset at the family level. ANOSIM analysis was used to identify variation in species abundance and composition between strains 630Δerm and hpdC::CT, as well as between time points D7, D0R, D2R and D4R. Significant differences were indicated with a circle p<0.001.
Faecal samples were defrosted and mixed with 400 μL of phosphate buffer (pH 7.4, 100% D2O, 3 mM of NaN3, 1 mM of 3-(trimethyl-silyl)-[2,2,3,3-2H4]-propionic acid (TSP) for the chemical shift reference at δ0.0) and Zirconium beads (0.45 g ±0.1). The samples were vortexed and then homogenised with a FastPrep-24 Classic Instrument (MP BIOMEDICALS) (30 sec per cycle, speed 6.0, 2 cycles). After a centrifugation (13,000 xg, 15 min), 180 μL of the supernatants were collected and transferred in 3-mm tubes for 1H nuclear magnetic resonance (NMR) spectroscopic analysis, which was performed on a Bruker 600 MHz spectrometer (Bruker Biospin, Karlsruhe, Germany) at 300K (26.85°). The parameters of the acquisition were as previously reported for urine[61]. 4 dummy scans followed per 64 scans were acquired for each spectrum which were then imported into Matlab (version R2014a, The Mathworks Inc.). The region δ4.82–4.76 was removed to eliminate residual water signal. All spectra were normalised according probabilistic quotient method and automatically aligned. Principal Components Analysis (PCA) was performed with mean-centring and Pareto scaling.
Frozen culture supernatants were defrosted on ice and were mixed in a 1:1 ratio with methanol: water, transferred to HPLC tubes and processed immediately by HPLC. Mouse faecal samples were defrosted, and added to a 2 ml screw cap tube containing 2 mm beads. These were weighed before and after addition of the faecal sample. To these, 400 μl 1:1 methanol:water was added to the pellet, then ribolysed twice using a FastPrep-24 Classic Instrument at speed 6.0 m/s for 30 sec. Tubes were transferred to ice and centrifuged at 14000 xg for 20 minutes. 250 μl of the supernatant was transferred to a clean sterile HPLC tube and were transferred immediately for HPLC. Each experiment was performed in triplicate. All HPLC equipment, software, solvents, columns and vials were from Thermo Fisher Scientific, UK. Separations were performed utilising an Acclaim 120, C18, 5 μm Analytical (4.6 x 150 mm) and the mobile phase consisting of ammonium formate (10 mM, pH 2.7) and menthol (v/v; 40:60) at a flow rate of 2 ml/min. p-HPA and p-cresol were detected by the photo-diode array detector (UV-PDA; DAD 3000) set at 280 nm. Peak identity was confirmed by measuring the retention time, spiking the sample with commercially available p-HPA and p-cresol and determination of absorbance spectra using the UV-PDA. A calibration curve of each compound was generated by Chromeleon (Dionex software) using known amounts of the reference standards (0–100 mg/ml) in methanol/water (v/v; 1:1) injected onto the column to determine the amount in the samples. The lower limit of detection was determined for p-HPA to be 0.03 mg/ml, for p-cresol to be 0.02 mg/ml. The concentration in mM was determined in Excel, using the molecular weight of the compounds and the quantity in mg/ml. The data was analysed in GraphPad Prism7 and statistical analysis was performed in Stata15 using linear regression analysis.
Healthy human donor faecal samples were collected and processed using three different healthy donors who had not received antibiotic treatment in the preceding 3 months. Faecal samples (5 g) were emulsified in sterile pre-reduced PBS (50 ml) and faecal material was coarse filtered by passing the 10% emulsion through sterile muslin cloth to remove larger particle matter and leave a bacterial suspension. Faecal emulsions were incubated for 1 hour and 30 minutes in 1X PBS, or PBS containing 0.1% (v/v) p-cresol or 0.3% (v/v) p-cresol. Samples were then sedimented by centrifugation at 14000 x g for 5 minutes and the supernatant were removed. Pellets were resuspended in 1 ml 1X PBS and viable counts (CFU/ml) were performed on differentially selective agar, both anaerobically and aerobically. Each experiment was performed in triplicate. Serial 10-fold dilutions of re-suspended faecal emulsions in sterile pre-reduced peptone water were inoculated onto: fastidious anaerobe agar (total anaerobes), nutrient agar (total facultative anaerobes), kanamycin aesculin azide agar (Enterococci), LAMVAB agar (Lactobacilli), Beeren’s agar (Bifidobacteria), MacConkey agar number 3 (lactose-fermenting Enterobacteriaceae), Bacteroides bile aesculin agar (B. fragilis group), and total viable counts were determined in triplicate, and normalised to the starting CFU.
Release of cellular phosphate was investigated using a Colorimetric Phosphate Assay Kit (Abcam). The assay involved treating samples with ammonium molybdate and malachite green which forms a chromogenic complex with phosphate ions which can be detected at a wavelength of 650 nm. An overnight culture of each bacterial strain was sedimented by centrifugation and re-suspended in Tris-buffered saline (TBS, 50 mM Tris-HCl pH7.5, 150 mM NaCl) and subsequently washed two further times to remove traces of the growth medium. OD595 was determined and cell suspensions were normalised to an OD595 of 1.0. Five hundred microliter aliquots of cell suspension were sedimented by centrifugation and re-suspended in either TBS alone or TBS + p-cresol. Cell suspensions were incubated for the indicated time, cells were sedimented and 30 μl of supernatant was removed and added to 170 μl H2O and 30 μl ammonium molybdate and malachite green reagent. Absorbance was read at 650 nm in a 96-well microtitre plate reader. Phosphate release was determined by normalising the optical density from cell suspensions incubated with p-cresol against cell suspensions that were incubated with TBS alone. The assay was performed under anaerobic conditions except for spectrophotometry and sedimentation steps, for which tubes and flasks were sealed with parafilm to prevent oxygen infiltration. The maximum intracellular phosphate pool was determined by boiling a 500 μl cell suspension (OD595 1.0) for 15 minutes. All assays were performed in triplicate.
|
10.1371/journal.pbio.1001094 | Gene Gain and Loss during Evolution of Obligate Parasitism in the White Rust Pathogen of Arabidopsis thaliana | Biotrophic eukaryotic plant pathogens require a living host for their growth and form an intimate haustorial interface with parasitized cells. Evolution to biotrophy occurred independently in fungal rusts and powdery mildews, and in oomycete white rusts and downy mildews. Biotroph evolution and molecular mechanisms of biotrophy are poorly understood. It has been proposed, but not shown, that obligate biotrophy results from (i) reduced selection for maintenance of biosynthetic pathways and (ii) gain of mechanisms to evade host recognition or suppress host defence. Here we use Illumina sequencing to define the genome, transcriptome, and gene models for the obligate biotroph oomycete and Arabidopsis parasite, Albugo laibachii. A. laibachii is a member of the Chromalveolata, which incorporates Heterokonts (containing the oomycetes), Apicomplexa (which includes human parasites like Plasmodium falciparum and Toxoplasma gondii), and four other taxa. From comparisons with other oomycete plant pathogens and other chromalveolates, we reveal independent loss of molybdenum-cofactor-requiring enzymes in downy mildews, white rusts, and the malaria parasite P. falciparum. Biotrophy also requires “effectors” to suppress host defence; we reveal RXLR and Crinkler effectors shared with other oomycetes, and also discover and verify a novel class of effectors, the “CHXCs”, by showing effector delivery and effector functionality. Our findings suggest that evolution to progressively more intimate association between host and parasite results in reduced selection for retention of certain biosynthetic pathways, and particularly reduced selection for retention of molybdopterin-requiring biosynthetic pathways. These mechanisms are not only relevant to plant pathogenic oomycetes but also to human pathogens within the Chromalveolata.
| Plant pathogens that cannot grow except on their hosts are called obligate biotrophs. How such biotrophy evolves is poorly understood. In this study, we sequenced the genome of the obligate biotroph white rust pathogen (Albugo laibachii, Oomycota) of Arabidopsis. From comparisons with other oomycete plant pathogens, diatoms, and the human pathogen Plasmodium falciparum, we reveal a loss of important metabolic enzymes. We also reveal the appearance of defence-suppressing “effectors”, some carrying motifs known from other oomycete effectors, and discover and experimentally verify a novel class of effectors that share a CHXC motif within 50 amino acids of the signal peptide cleavage site. Obligate biotrophy involves an intimate association within host cells at the haustorial interface (where the parasite penetrates the host cell's cell wall), where nutrients are acquired from the host and effectors are delivered to the host. We found that A. laibachii, like Hyaloperonospora arabidopsidis and Plasmodium falciparum, lacks molybdopterin-requiring biosynthetic pathways, suggesting relaxed selection for retention of, or even selection against, this pathway. We propose that when defence suppression becomes sufficiently effective, hosts become such a reliable source of nutrients that a free-living phase can be lost. These mechanisms leading to obligate biotrophy and host specificity are relevant not only to plant pathogenic oomycetes but also to human pathogens.
| For more than 150 years, attempts to culture downy mildews, powdery mildews, and rusts on artificial nutrient media have been unsuccessful. The terms obligate parasitism and obligate biotrophy are used to denote organisms that live in such an obligatory association with living hosts [1],[2]. Recent research on the obligate biotroph powdery mildew fungus Blumeria graminis or downy mildew oomycete Hyaloperonospora arabidopsidis reveals a close correlation between the biotrophic life style and massive gene losses in primary and secondary metabolism [3],[4]. Obligate biotrophs form an intimate haustorial interface with parasitized cells. Haustoria are differentiated intercellular hyphae, but little is known about their functionality and evolution beyond their involvement in nutrient uptake [5],[6].
The obligate biotroph oomycete Albugo laibachii is a member of the Chromalveolata, which incorporates Dinophyta, Ciliophora, Heterokonts (containing the oomycetes), Haptophyta, Cryptophyta, and Apicomplexa (which includes human parasites like Plasmodium falciparum and Toxoplasma gondii [7],[8]).
Within the oomycetes, A. laibachii belongs to a lineage known as peronosporalean, which includes the hemibiotrophic pathogen of potato Phytophthora infestans [9] and the necrotroph pathogen Pythium ultimum [10]. Within this lineage, obligate biotrophy evolved twice independently in white blister rusts (Albuginales) and downy mildews (part of the Peronosporaceae) [11]. The downy mildew pathogen H. arabidopsidis and A. laibachii are both pathogens of the model plant Arabidopsis thaliana [12]. While both show similar infection structures within the host [13],[14], A. laibachii releases motile zoospores from asexual spores and sexual oospores, while H. arabidopsidis lacks all motile stages [4],[15]. Both pathogens are regularly found to co-infect plants and sporulate on the same leaf [16].
A remarkable consequence of infection by Albugo sp. is enhanced host plant susceptibility to other parasites to which the host is resistant in the absence of Albugo infection, and also impairment of cell death mechanisms [16]. Albugo sp. infect 63 genera and 241 species [17], including economically important Brassica rapa (canola), B. juncea (oilseed mustard), and B. oleracea (cabbage family vegetables) [18],[19]. Recent analysis of oomycete evolutionary history [11] suggest that Albugo is more closely related to necrotrophs such as Pythium than to downy mildews, and thus provides a unique system to study the evolution and consequences of biotrophy, and to identify new defence-suppressing effectors and their host targets.
Since prolonged culture of pathogen strains can result in genetic changes [20], we sequenced a fresh highly virulent field isolate of A. laibachii. The strain was selected from a heavily infected Ar. thaliana field plot (Norwich, United Kingdom) [21], and strains were single zoospore purified. Isolate Norwich 14 (Nc14) was determined as A. laibachii [19] and used for further analyses. In contrast to Nc14, A. laibachii isolate Em1 (formerly Acem1, A. candida East Malling 1 [19]) is an established Albugo strain that was collected 15 y ago [16],[22],[23], and we resequenced this strain. Both strains show identical ITS (internal transcribed spacer of ribosomal RNAs) and COX2 (cytochrome C oxidase subunit II) sequences. To ensure that sequence differences observed between these strains are of biological relevance not just the result of background mutations, we tested the host range for both isolates on 126 Ar. thaliana accessions and identified 12 that show resistance to only one of the A. laibachii isolates (Table S1). Nc14 is virulent on more accessions than the Em1 isolate is (Table 1).
The A. laibachii Nc14 genome was sequenced using Illumina 76-bp paired reads with ∼240-fold coverage (Figure 1). In order to assemble the diploid heterozygous genome, an assembly pipeline was developed using Velvet [24] as primary assembler and Minimus [25] as meta-assembler (Figure S1). Short read assembly programs are sensitive to heterozygous positions depending on read depth and kmer-length. Reads not aligning to bacterial or plant sequence in public databases were used to estimate the genome size as ∼37 Mbp. Using the estimated genome size, 50% of the resulting assembly is contained in 164 contigs with an N50 of 56.5 kbp. A comparative analysis of contig size classes versus frequency indicates that 90% of the assembled genome shows a high degree of continuity in only 585 contigs, while 10% of the genome is fragmented in 3,231 contigs (Figure 2A). Read depth indicates that this 10% of the genome shows elevated levels of nucleotide coverage that are likely to comprise unresolved repeats (Figure 2B). Aligning Illumina cDNA reads from different stages of infection to reveal transcriptionally active regions in the assembly shows that few transcripts arise from the unresolved repetitive regions of the genome (Figure 2D), suggesting that the gene space of a genome can be reliably defined using Illumina-only approaches. A CEGMA [26] analysis revealed a high degree of completeness of assembly of core eukaryotic genes, as well as a continuity within the core genes comparable to high-quality Sanger read assemblies (Figure S2; Table S2). We designed 32 primer pairs for regions between 0.6 and 5 kb based on our assembly (Table S3). Thirty-one genomic regions could be amplified and were Sanger sequenced from both ends. All PCR products had the predicted size, and sequences showed 100% identity to the genome assembly.
The mitochondrial draft genome was assembled in a separate attempt because of its high repeat content and therefore higher coverage compared to the core genome. The assembled genome comprises 26.7 kb in 11 contigs and shows a high degree of synteny to the P. infestans mitochondrion Ia [27] and the Py. ultimum mitochondrion [10] (Figure S3). Considering the node coverage of the Velvet primary assembly (∼150×), 15.6 kb of the mitochondrial genome have >300× node coverage and seem to be duplicated. This might indicate, comparable to the Py. ultimum mitochondrion genome [10], that ∼50% of the genome is duplicated, leading to an estimated genome size of ∼43 kb. While the highly repetitive tRNAs are not resolved within the A. laibachii mitochondrial genome, regions of high synteny between the Py. ultimum and the P. infestans mitochondrial genome are found in ribosomal proteins and subunits of the NADH dehydrogenase as well as cytochrome C oxidase.
Approximately 22% of the A. laibachii Nc14 genome assembly consists of repetitive regions (Figure 3; Tables S4 and S5). The majority of repeats are represented by transposable elements (96%), while 4% of all repeats are A. laibachii-specific (Table S5). Compared to other obligate biotrophs, the number of repeats is low. H. arabidopsidis, for example, with an estimated genome size of 100 Mb, contains ∼43.3% repeats [4], while transposable elements account for 64% of the ∼120-Mb Bl. graminis (powdery mildew) genome [3]. We identified 45 contigs carrying telomeric repeats; amongst these, 25 contigs have telomeric repeats located at one end of a contig. We therefore postulate that the A. laibachii Nc14 genome is distributed over 12 or 13 chromosomes (Table S6). tRNA genes are difficult to resolve because of their high copy number [28]. Within our Illumina assembly, 153 tRNA genes were detected with 48 distinct anticodons (Figure S4; Table S7). Our ability to resolve all these repeats within the Illumina short read assembly illustrates its quality.
Based on read depth, both Nc14 and Em1 isolates possess ∼6 Mbp of hemizygous or highly heterozygous regions (6.2 and 5.6 Mbp for Nc14 and Em1, respectively) (Figure 1B and 1D) as well as ∼13,000 heterozygous loci (13,116 and 13,523 for Nc14 and Em1, respectively) (Figure 2C). Remarkably, most of the hemizygous/highly heterozygous regions are shared between Nc14 and Em1.
Compared to other sequenced oomycetes like P. infestans (240 Mbp), H. arabidopsidis (100 Mbp), or even Py. ultimum (42.8 Mbp), A. laibachii has a highly compact genome structure (Figure 4A). Approximately 50% of the A. laibachii genome assembly matched cDNA reads, and transcriptionally active regions are further clustered, resulting in transcriptional hot spots and silent genomic regions (Figure 4B).
A reference set of 13,032 gene models was generated incorporating cDNA reads from different stages of infection (Figure S5A). From extensive cDNA sequencing of infected Arabidopsis leaves, approximately 20 M (∼1.5 Gbp) unique Illumina reads match the Nc14 genome assembly but not Ar. thaliana TAIR 9.0, and these were used to generate training sets for ab initio gene predictions and as evidence sets for consensus gene prediction. In all, 88.3% of all gene models are supported by at least three cDNA hits.
For validation of these gene models, a set of 860 annotated core eukaryotic orthologous groups (KOGs) [29] was compiled and tested. In all, 75% of these groups are present in the current annotation. For comparison, 78% of KOGs were present in P. infestans, 73% in H. arabidopsidis, 42% in Pl. falciparum, and 85% in Ar. thaliana (Figure S5B). In addition, 49.9% of all gene models show Pfam support, resulting in 2,505 Pfam domains, and 803 genes were functionally assigned to pathways using ASGARD [30] and manual annotation. Transcriptional units show an even more compact, clustered occurrence than P. sojae or P. ramorum and an occurrence pattern clearly different from that of P. infestans [9] (Figure 4C).
From our annotations using ASGARD we identified major enzymes of the lipopolysaccharide biosynthesis pathway, as have been described for P. infestans [31]. These analyses revealed, in addition, the possibility that A. laibachii is able to synthesize brassinosteroids. We identified potential homologues to the Ar. thaliana brassinosteroid biosynthesis genes Dwf4 and DET2 (Table S8). Although ASGARD identified homologues of Br6ox, D2, and CPD, manual annotation revealed that assigning function to members of the superfamily of cytochrome P450 enzymes in A. laibachii is difficult based on homology alone (Table S8). It has been hypothesized that the frequency of functionally redundant genes is reduced in obligate biotrophs, as reported for Bl. graminis [3]. Combining ASGARD and manual annotation we identified the absence of the whole steroid biosynthesis pathway, and, like other oomycetes, A. laibachii probably relies on the host as a source of sterols. We hypothesize that A. laibachii would need to take up campesterol from the plant as a precursor for brassinosteroid synthesis.
During evolution, plastids of both red algae and green algae were transferred to other lineages by secondary endosymbiosis. How often and when secondary endosymbiosis occurred is difficult to address but of importance to clarify the origin of chromalveolates and their gain and loss of endosymbionts. There are two distinct hypotheses for what took place. The monophyletic hypothesis posits that a red alga was taken up only once, followed by repeated losses of this algal genome, giving rise to the highly divergent group of chromalveolates [32]. An alternative and more common view hypothesizes polyphyletic origins of the Chromalveolata, with in some cases multiple events of secondary endosymbiosis [33]–[35].
Molecular divergence of A. laibachii from other species within the Chromalveolata was assessed by examining the percentage of amino acid identity between orthologous gene pairs (Figure 5). These analyses demonstrate that the green alga Chlamydomonas reinhardtii, the brown alga Ectocarpus siliculosus, and the diatom Phaeodactylum tricornutum show the same distribution of percentage amino acid identity to A. laibachii Nc14 regarding the cumulative frequency of orthologous pairs. In contrast, previous systematic analyses suggested that brown algae and diatoms are the closest relatives of oomycetes and that secondary endosymbiosis occurred with a red alga [32], although there are suggestions that oomycetes diverged before this event [36]. Using a set of >1,700 genes that are of “green” origin (from green algae) or “red” origin (from red algae) and that have been integrated into the diatom nuclear genome [37], we found more oomycete genes that show significant BLAST hits to green algae than to red algae (34 “green” compared to five “red”) (Figure S6; Table S9). These findings are consistent with the results published by Moustafa et al. [37] for diatoms. In a separate approach we identified genes showing high similarity between oomycetes, green algae, and red algae that are absent from diatoms (32 “green”; 11 “red”) (Tables S10 and S11). This result might indicate the presence of all these genes in a common ancestor, followed by loss or expansion of the gene family depending on adopted live style. To address this question, we further analysed genes absent from A. laibachii Nc14 and studied their presence/absence in three other oomycetes, Pl. falciparum, and the brown alga E. siliculosus (Table S12). The majority of genes absent from A. laibachii Nc14 are absent from other oomycetes and from Pl. falciparum but are present in the brown alga. These genes are involved in the photoautotrophic, aquatic life style of diatoms and algae, such as a sodium/bile acid cotransporter, a haloacid dehalogenase-like hydrolase, fatty acid biosynthesis genes, a zeaxanthin epoxidase and a fucoxanthin chlorophyll a/c binding protein. In contrast to the genes lost, we found that certain gene families like aspartic proteases or proteases containing MORN (membrane occupation and recognition nexus) repeats [38] show expansion in A. laibachii Nc14 compared to in diatoms. Although our results fit the hypothesis of a common ancestor, we cannot exclude horizontal gene transfer and uptake of an endosymbiont after the divergence between a brown algal ancestor and an oomycete ancestor, given the low number of diagnosed genes that we could analyse.
Potentially green-algae-derived proteins carrying MORN repeat domains (Figure S7) are involved in the complex process of internal budding in apicomplexans [39], which may be similar to the zoospore formation of oomycetes within oospores or zoosporangia or gamete formation in diatoms [40]. While oomycetes with a motile zoospore stage like A. laibachii and P. infestans carry the MORN repeat proteins, these proteins are absent in the non-motile H. arabidopsidis and absent in the non-motile red alga Cyanidioschyzon merolae [41]. We therefore hypothesize that loss of this gene of hypothetical green algal origin could have led to the evolutionary loss of the whole flagellum apparatus in H. arabidopsidis [4]. However, we cannot rule out that depletion of any major flagellar protein could have caused evolutionary loss of the whole flagellum apparatus. Inspection of the flagellar inner arm dynein 1 heavy chain alpha, which is absolutely necessary for flagellum function, reveals that genomic regions carrying flagellar inner arm dynein 1 heavy chain alpha genes show a high degree of synteny between oomycetes like Py. ultimum and A. laibachii. In contrast, a syntenic region in H. arabidopsidis shows replacement of the flagellar dynein by Mariner- or Gypsy-like transposable elements (Figure S8).
Since within the peronosporalean lineage, biotrophy evolved twice independently [11], we compared A. laibachii with the other obligate biotroph H. arabidopsidis [4], hemibiotroph P. infestans [9], and necrotroph Py. ultimum [10] (Figure 5; Tables S13 and S14). We found that H. arabidopsidis is the most diverged from A. laibachii. H. arabidopsidis shares the fewest (4,826) orthologous genes with A. laibachii, versus the average of 5,722 in A. laibachii/P. infestans and A. laibachii/Py. ultimum comparisons. Meanwhile, H. arabidopsidis genes show the highest amino acid identity with the genes of P. infestans, on average 73% of amino acid identity between all single copy orthologous pairs.
Py. ultimum shares the highest number of orthologous genes with A. laibachii (5,910 pairs). P. ultimum proteins also have a slightly higher percentage of amino acid identity with A. laibachii proteins than with other oomycetes (Figure 5). Yet, Py. ultimum itself is closer to H. arabidopsidis and P. infestans than to A. laibachii, sharing with them more orthologous genes with higher mean amino acid identity.
These analyses support the hypothesis that A. laibachii and H. arabidopsidis evolved biotrophy independently; genes missing in one or the other genome compared to the necrotroph Py. ultimum or hemibiotroph P. infestans may be correlated with biotrophy (Table S15). One of these genes is that for molybdenum-cofactor-dependent nitrate reductase. Nitrate reductase catalyzes pyridine-nucleotide-dependent nitrate reduction for nitrogen acquisition [42]. Both biotroph pathogens have a set of transporters showing homology to amino acid transporters, but other uptake mechanisms or sources could also enable nitrogen acquisition from their hosts [43]. While H. arabidopsidis lost only the nitrate reductase, A. laibachii also lost the sulphite oxidase and the whole molybdopterin (a cofactor required for nitrate reductase and sulphite oxidase function) biosynthesis pathway. In Pl. falciparum, which shows a high degree of adaptation to parasitism, nitrate reductase, sulphite oxidase, and the whole molybdopterin biosynthesis pathway are also missing. Most likely the loss of the two Mo-containing enzymes and the Mo-cofactor biosynthesis is the outcome of biotrophy and not the reason for biotrophy, though conceivably there may have been selection against this pathway if other nitrogen or sulphate sources are less energy-consuming and therefore enhance fitness during parasitism. Molybdenum has been reported to interfere with function of chaperones like Hsp90 [44],[45]. Avoiding the uptake of molybdenum might prevent this Hsp90 inhibition and increase fitness on Ar. thaliana accessions with high molybdenum levels like Col-0 [46]. H. arabidopsidis therefore could be in a less advanced stage of host adaptation compared to A. laibachii and Pl. falciparum.
Besides biotrophy, the formation of haustoria and haustorium-like structures evolved several times in peronosporalean biotroph and hemibiotroph pathogens. Haustoria in fungi are sites of enhanced nutrient uptake [47] and metabolism, such as thiamine biosynthesis [48]. In the oomycetes, all haustorium-forming species have lost the thiamine biosynthetic pathway. We infer that haustorial oomycetes obtain thiamine from the host.
We therefore hypothesize that evolution to biotrophy is initiated not by gene loss, but rather from the ability to build a haustorium and therefore differentiate a sophisticated interface with a host. The critical step to adopting biotrophy is likely to be efficient defence suppression to enable persistence of functioning haustoria; subsequent loss of biosynthetic pathways is likely to be secondary.
Well-adapted human pathogens like Pl. falciparum and plant pathogenic fungi like Ustilago maydis have small secretomes (320 [49] and 426 [50] proteins, respectively) compared to necrotrophic fungi like Aspergillus fumigatus (up to 881 proteins [51]). We found that the same is true for oomycetes. Using SignalP [52] to predict potential secretion signal peptides and MEMSAT [53] to predict transmembrane (TM) domains, we identified 2,473 (2,136 without TM domains) potentially secreted proteins in the hemibiotroph P. infestans and 1,636 (1,222 without TM domains) in the necrotroph Py. ultimum. For H. arabidopsidis only 1,350 (1,054 without TM domains) and for A. laibachii 949 (672 without TM domains) were identified. Analysing the secretome for pathogenicity-related proteins like proteases, glucosyl hydrolases, and potential elicitins or lectins reveals a significant reduction in the H. arabidopsidis and A. laibachii secretome (Tables 2 and S16). We postulate that biotrophs reduce their activation of host defence by reducing their inventory of secreted proteins, particularly cell wall hydrolyzing enzymes.
The ability to establish a sophisticated zone of interaction like the parasitophorous vacuole in Pl. falciparum or the haustorium in oomycetes and fungi requires sophisticated host defence suppression [54], which is predominantly achieved via secreted proteins delivered into the host cell [55],[56]. The A. laibachii secretome comprises 672 secreted proteins without TM domains. Genetically identified oomycete avirulence (Avr) proteins are secreted proteins that have signal peptide and RXLR motifs [57],[58]. In many oomycete genomes the RXLR motif is over-represented and positionally constrained within the secreted protein [59]. We identified 25 RXLR and 24 RXLQ effector candidates in the A. laibachii secretome. To determine the likelihood that RXLR or RXLQ motifs occur merely by chance in the A. laibachii secretome based on amino acid content, we performed in silico permutation of the motifs (Figure 6A and 6B). We concluded that the RXLR and RXLQ motifs were not likely to occur merely by chance, and that the likelihood of occurrence by chance is higher in the proteome as a whole than among secreted proteins. It was shown for P. infestans that effectors are often located in gene-depleted repetitive regions of the genome [9]. We therefore investigated RXLR candidate proteins in highly repetitive regions of the genome. We identified two RXLRs, one in a highly conserved repeat region with ∼10 repeats in Nc14 and one in a more diverged repeat region with >80 repeats within the genome. The first region also exists in A. laibachii isolate Em1; the diverged repeat of the second identified region exists but without the RXLR gene-containing region (Figure S9). There are 563 RXLR effector candidates identified in P. infestans [9], so RXLR effectors are less likely to be relevant for A. laibachii virulence.
Similar conclusions can be drawn for the CRN protein family, which shows expansion in P. infestans [9],[60] but not A. laibachii, where only three members of the CRN family could be identified with signal peptides. Eight additional CRN-like proteins were identified where no signal peptide has been predicted.
To identify new classes of effectors in the Albuginales clade, the secretome of A. laibachii was computationally screened for genes either showing heterozygosity or showing nucleotide polymorphisms between Nc14 and Em1. We identified a new class carrying a “CHXC” motif by inspection of the first 80 amino acids after the signal peptide cleavage site. CHXC candidates are significantly enriched within the secretome (Figure 6C). Comparisons of the N-terminal part of the CHXC proteins revealed additional conserved amino acids, particularly a glycin at +6 to the CHXC motif (Figure 6D).
In host–pathogen interactions, intraspecies comparisons enable the search for virulence alleles that undergo positive selection and fixation within the population [61],[62]. Secreted proteins with close contact to the host cell, such as effector proteins, often show enhanced levels of positive selection [63],[64]. By comparing the two A. laibachii isolates Nc14 and Em1, we identified a significantly higher frequency of non-synonymous to synonymous mutations within the predicted secretome compared to the rest of the proteome. Our analyses showed that this was particularly true for heterozygous positions and less convincing for homozygous SNPs (Table S17). Genes that are highly conserved between species, like KOGs, showed comparable non-synonymous and synonymous substitution rates, with a slight excess of synonymous mutations. There are significantly more genes within the KOGs showing a non-synonymous/synonymous ratio less than 1 than genes with values greater than 1. Comparing this to candidate effector classes like RXLRs, RXLQs, and CHXCs reveals that in particular the CHXCs show significantly higher frequencies of non-synonymous to synonymous mutations. This supports the idea that the CHXC sub-class of secreted proteins is under positive selection, similar to other described oomycete effectors like ATR1 or ATR13 from H. arabidopsidis [57],[65].
Further to this we identified Nc14 genes absent or highly diverged from the Em1 complement. We defined a gene as absent or highly diverged if >10 bp showed 0 coverage in the Em1 alignment. Out of the 672 secreted proteins without TM domains, we identified seven as absent from Em1 (1.04%). We also detected two with a predicted TM domain (0.73%) that are absent from Em1. Regarding all gene models, 96 were absent (0.74%). This finding is a further indication for a greater selection pressure on secreted than on non-secreted proteins, as has been found in species or interspecies comparisons in Phytophthora sp. [66] and Ustilago/Sporisorium [67].
We tested A. laibachii effector candidates (one CHXC, one RXLR, and one CRN effector candidate) for their host delivery efficiency using a P. capsici–Nicotiana benthamiana translocation assay [68]. Briefly, N-terminal domains of candidate effectors were fused to the P. infestans Avr3a effector domain, transformed into P. capsici, and tested for whether they confer translocation of Avr3a into N. benthamiana carrying R3a, resulting in avirulence. Statistical analyses of the delivery efficiency (Figure 7) clearly indicate that the A. laibachii CRN3 N-terminus and CHXC9 N-terminus are as efficient as the Avr3a N-terminus in Avr3a translocation, while the RXLR1 N-terminal domain is less efficient. An alanine replacement construct of the CHXC motif supports the importance of this motif for delivery efficiency. The Avr3a C-terminus alone confers a low basal delivery level without the need for the N-terminal enhancer. These findings reveal the potential of the CHXC proteins to be delivered into the host cell, similar to RXLRs and CRNs, though the delivery mechanism for all these effector classes requires further investigation.
To assay the effectors for virulence function, we used Pseudomonas syringae pv. tomato (Pst) DC3000 luciferase [69] carrying “effector detector vector” (EDV) constructs to deliver effectors into the plant cytoplasm via type III secretion [70] (Figure 8). Tests on Ar. thaliana Nd-0 plants revealed that several selected A. laibachii RXLRs, CRNs, and CHXCs enhance virulence compared to a non-functional AvrRps4 (AvrRps4[AAAA]). On Ar. thaliana Col-0, in contrast, the CRN and one RXLR (RXLR1) do not enhance virulence while RXLR2 and CHXCs still do. These tests indicate that CHXCs carry the capacity to enhance virulence in phytopathogenic bacteria, perhaps by suppression of host resistance mechanisms [54],[70]. These virulence assays together suggest that A. laibachii uses at least three different major effector classes.
To try to identify the evolutionary source of CHXCs, we investigated enrichment of CHXC-motif-containing proteins in the secretomes of P. infestans, Py. ultimum, H. arabidopsidis, Saprolegnia parasitica, Thalassiosira pseudonana (diatom), Pl. falciparum (Apicomplexa), E. siliculosus (brown alga), C. merolae (red alga), Ch. reinhardtii (green alga), Volvox carteri (green alga), and Ar. thaliana. Only A. laibachii contained a significant enrichment of CHXCs in its secretome. Although not significantly enriched, both the fish pathogen S. parasitica and the land plant Ar. thaliana contained more than ten CHXC proteins carrying potential secretion signals (14 and 11, respectively) (Figure S10). In contrast to CHXC-containing proteins, almost all inspected organisms show a high number of CXHC-containing potentially secreted proteins; a common CXHC protein is protein disulphide isomerase (Table S18).
Given that A. laibachii CHXCs show the closest clustering with S. parasitica, V. carteri, Ch. reinhardtii, and Ar. thaliana CHXCs (Figure 9), conceivably this candidate effector class evolved from an ancestral green-alga-derived gene.
Whatever their origin, we conclude that CHXC proteins are present in all organisms analysed but evolved effector function only in Albuginales and possibly Saprolegniales. In Albuginales, one N-terminal sub-class of CHXCs (CHxCLx(4)Gx(5–6)L) shows significant expansion, with 23 members, while other CHXCs are distinct from this clade. S. parasitica CHXCs are distinct from this major A. laibachii clade and therefore remain to be tested in future experiments.
The A. laibachii genome assembly sheds light on the evolution of biotrophy since it allows the first comparison, to our knowledge, of two oomycete obligate biotroph pathogens (A. laibachii and H. arabidopsidis) that evolved biotrophy independently. In addition, A. laibachii shows the highest overall amino acid identity to the necrotroph pathogen Py. ultimum and the hemibiotroph P. infestans. One of the striking results of this comparison is that all organisms able to build haustoria have lost their thiamine biosynthesis pathway, presumably because thiamine is easily obtained from hosts via the haustorial interface. A closer interface requires effective host defence suppression. We therefore hypothesize that the evolution of biotrophy involves a series of steps: step 1, involving progressively more effective effectors to suppress defence, step 2, attenuated activation of defence by reduction in the inventory of cell wall hydrolyzing enzymes, resulting in, step 3, weak selection to maintain certain biosynthetic pathways if the products of the pathways can be directly obtained from the host. This results in progressively more comprehensive auxotrophy and culminates in irreversible biotrophy (Figure 10).
An infected leaf was harvested from an Ar. thaliana plant grown in a heavy infected field plot in Norwich (UK; 52.6236,1.2182) [21] in December 2007. Zoosporangia were washed off the leaf surface and used to infect Ar. thaliana Ws-0-eds1 plants. After 1 wk one pustule was punched out, and spores were placed on ice for 30 min to release zoospores. Unhatched zoosporangia were removed by filtration, and zoospores were diluted to ∼10 zoospores/ml and sprayed on Ar. thaliana Ws-0 plants (∼100 µl/plant). This procedure was repeated 4× until spores were bulked up on Ar. thaliana Ws-0 plants. Zoosporangia were harvested using a home-made cyclone spore collector [71].
Zoospores were suspended in water (105 spores/ml) and incubated on ice for 30 min. The spore suspension was then sprayed on plants using a spray gun (∼700 µl/plant), and plants were incubated in a cold room in the dark over night. Infected plants were kept under 10-h light and 14-h dark cycles with a 20°C day and 16°C night temperature.
High molecular weight DNA was extracted from zoosporangia using a phenol/chloroform-based purification method after grinding in liquid nitrogen, adapted from [72]. Library preparation for Illumina sequencing was performed as described [28]. All data were generated using paired-end reads. 800 bp and 400 bp paired-end sequencing libraries were constructed, and 8.8 Gbp of usable data were generated (for read and insert length, see Figure 1A).
Figure 1A lists all reads after purification from plant and bacterial contamination as well as all reads aligned to the assembly. In summary, 91.6% of all reads can be aligned to the contigs, suggesting 2.8 Mbp missing from the assembly. Since 32.7 Mbp are in the assembly, the genome can be estimated to 35.5 Mbp. In another approach considering all reads and their read length, 8.8 Gbp (∼7% correction for lower quality of second read pair) were generated, which would lead to an expected coverage of the 32.7 Mbp genome of ∼270×. The mean coverage using single copy genes (glycolysis and TCA) is 240×. Considering the 2.5 Mbp of repeats (Figure 1B, right side, coverage underestimated) with an average coverage of 1,086×, which is ∼4.4 times more than the mean coverage of the contigs, this repeat region corresponds to 10.9 Mbp. In contrast to this, the genome contains ∼6.2 Mbp of hemizygous regions (Figure 1B, left side, coverage overestimated). These calculations suggest a genome size of ∼43 Mbp, given all repeats resolved, or an effective genome size of ∼37 Mbp.
A. laibachii–infected Ar. thaliana Ws-0 plants were harvested 0 (after cold room, see plant inoculation), 2, 4, 6, 8 and 10 d after infection. Total RNA was extracted using TRI Reagent RNA Isolation Reagent (Sigma), and Dynabeads (Invitrogen) were used to enrich for mRNA. First and second strand cDNA synthesis was performed according to manufacturer's instructions using the SMART cDNA Library Construction Kit (Clontech), and cDNA was normalized using the Trimmer kit from Evrogen. cDNA samples were mixed in equal amounts and fragmented using a Covaris sonicator (Covaris). Illumina libraries were prepared as described for fragmented genomic DNA [28].
Data for comparative genomics were downloaded from the sources listed in Table 3.
First Velvet [24] was used, running different kmer-lengths and different sequencing library subsets (kmer-length: 23, 31, 41, 45, 49, 55, 61, 67, and 73; subsets: 400-bp insert only, 800-bp insert only). N50 number and length were determined for each of the assemblies, and the best assembly was selected as the matrix to be used with the Minimus2 genome merge pipeline [25]. For the current assembly the 400-bp only subset with kmer-length 61 was used as matrix, and for kmer-lengths 49, 55, 61, 67 and 73, all 400- and 800-bp assemblies were added (Minimus parameters: consensus error <0.001; minimum identity >99%; 20-bp maximum trimming). A set of genes showing high heterozygosity was used to ensure that contigs were properly joined. Parameters were changed through several rounds, and minimum overlap, in particular, was lowered from 100 bp to 15 bp. An overlap of 15 bp was found to be the optimum for difficult heterozygous regions. After each Minimus assembly, all reads were back aligned to the contigs using MAQ aligner [73]. Regions showing less than 3× average coverage were removed, and redundant fragments were removed using BLASTN with an e-value cut-off of 1e−20 and 99.9% identity. After this step a next round of Minimus was started, with changing minimum overlap in steps of 20 bp down from 100 bp. Below 20 bp steps were changed by 5 bp (See Figure S1 for work flow).
Since it is impossible to cultivate obligate biotrophs under sterile conditions, plant and bacterial contaminations were removed by using BLAST against genome sequences of the host plant Ar. thaliana (TAIR 9.0), fungal genomes (Neurospora crassa), oomycetes (H. arabidopsidis), and diverse bacterial genomes (Xanthomonas sp. and Pseudomonas sp.).
To identify heterozygous loci, Illumina reads were aligned using MAQ, and the SNP detection pipeline was used according to the manual, with default parameters and minimum coverage greater than 180× for the Nc14 alignment and greater than 20× for the Em1 alignment. From the MAQ SNP file, positions were selected where two bases are possible and maximum coverage was less than 350×.
Assembled repetitive elements were identified using the RepeatScout program (http://bix.ucsd.edu/repeatscout/) with a seed size of 14. The frequency of elements and their location in the assembly were estimated with RepeatMasker using a library of repetitive elements built up by RepeatScout. A sequence was considered to be repetitive if it occurred in the genome assembly on at least three different contigs. The resulting library was searched for the sequences homologous to the known transposon elements using TBLASTX (e-value cut-off of 1e−5) and a database of transposons, RepBase [74]. Consensus repeats that matched predicted Nc14 protein coding genes were filtered out. The remaining consensus repeats that do not match any sequences deposited in the NCBI database or any known transposon element and that do not overlap with Nc14 protein coding genes represent either Albugo-specific repeats or simple repeats.
tRNA genes were predicted with the program ARAGORN [75] using first default parameters and second options allowing introns in the gene sequences.
CEGMA was used according to the manual [26] with a local installation.
For the combined ABySS [76] and Oases [77] assembly, adaptor sequences from the SMART kit cDNA synthesis were removed for the ABySS assembly, and the ABySS program was used according to the manual. Different kmer-lengths were tested, and a length of 61 used for the final assembly.
Untrimmed cDNA sequences were assembled using Velvet and a kmer-length of 51, 57, 61, and 71. Oases was used for the final assembly of the contigs according to the manual, using default parameters.
MUMmer in maxmatch mode was used to combine all ABySS and Velvet assemblies. Redundant contigs were removed using BLAST.
Since the assembled cDNA is not strand specific but orientation is needed for gene prediction, cDNA 5′ tags were generated by Illumina sequencing (E. Kemen, A. Balmuth, J. D. Jones, unpublished data). Using Bowtie aligner [78], cDNA 5′ tags were aligned onto the assembled cDNA and, based on tag counts, orientated in the 5′ to 3′ direction.
To map assembled cDNA against the genome, either BLAT [79] in trimT and fine mode or PASA [80] with default settings was used.
Illumina reads were directly mapped to the genome using the Bowtie aligner, in “best” mode and with strand correction (strandfix mode). Pileup files were generated using bowtie-maqconvert and maq pileup allowing four mismatches per 76-bp read. To incorporate this data as hints files for gene prediction, regions with greater than 3× coverage were extracted.
To generate a reliable gene set to train further programs, GeneMark [81] was used for ab initio gene prediction. ORFs plus 50 bp on the 3′ end and 50 bp on the 5′ end were extracted, and Illumina-sequenced cDNA was aligned to the ORFs using Bowtie. Gene models were selected if the coverage within the ORF didn't drop below three. This dataset with more than 2,000 genes was used as “traingenes” for the automated training program provided with the Augustus package (autoAug.pl). The trained Augustus program was then used for gene prediction including the combined Oases/ABySS-assembled cDNA (mapped using BLAT) as evidence. Default parameters (extrinsic.ME.cfg) were used for all predictions.
For consensus gene predictions with P. infestans, SGP2 was used according to the manual [82].
ASGARD [30] alignments were converted into GFF files to be used for consensus predictions.
Consensus gene models were generated using Evigan [83]. cDNA from assemblies and alignments was converted into GFF files and combined with Augustus, GeneMark, SGP2, and ASGARD predictions. The genome was than screened for gene-free regions, and Augustus gene predictions were added if available. In a third round, regions that did not contain consensus gene models or Augustus gene models were extracted, and GeneMark annotations were added if available.
A set of genes was further tested by 5′ and 3′ RACE to validate start and stop sites.
Molecular divergence of A. laibachii from other species was assessed by examining the percentage of amino acid identity between orthologous gene pairs [75].
Orthologous pairs were identified using the OrthoMCL program with an e-value cut-off of 1e−5 [84]. Alignments of protein pairs were performed with MUSCLE [85].
Amino acid identity was calculated only for the single copy genes by either excluding alignment gaps from calculations or taking gaps into account. The results show similar trends, so we present only results for the calculations when alignment gaps were excluded.
The total number of orthologous groups identified between species and the number of one-to-one orthologous pairs, as well as a mean amino acid identity, are shown in Table S7. In the comparison of T. gondii and A. laibachii, we found few orthologous pairs represented by the single copy genes (23 pairs); therefore, we excluded this pair of species from the analyses of sequence divergence.
We also estimated the levels of amino acid identity for the core eukaryotic genes (orthologous genes shared by all examined species); these data are presented in Table S8.
To identify A. laibachii genes with sequence similarity to green- or red-algal-derived diatom genes, a set published by Moustafa et al. [37] was used. All A. laibachii proteins showing homology to genes identified by Moustafa et al. [37] were further blasted (BLASTP) against the Ch. reinhardtii gene set, the E. siliculosus gene set, the U. maydis gene set, and the Fusarium oxysporum gene set with an e-value cut-off of 1e−20. Genes were considered to be green-alga-derived only if the protein was absent from U. maydis and F. oxysporum but present in Ch. reinhardtii, and was considered red-alga-derived if not in U. maydis or F. oxysporum but in E. siliculosus. The same analyses were performed on the Saccharomyces cerevisiae, Pl. falciparum, H. arabidopsidis, P. infestans, Py. ultimum, V. carteri, Ch. reinhardtii, C. merolae, C. merolae, Th. pseudonana, and Ph. tricornutum gene sets.
A. laibachii candidate genes with significant sequence similarity to green or red algae and other oomycetes (e-value cut-off of 1e−20) but not to fungi, brown algae, or diatoms were identified using the criteria in Table 4. Representative organisms for each group are as follows: green algae: V. carteri, Ch. reinhardtii; red algae: C. merolae, Galdieria sulphuraria; fungi: F. oxysporum; brown algae: E. siliculosus; diatoms: Ph. tricornutum, Th. pseudonana; oomycetes: P. sojae, Py. ultimum, H. arabidopsidis.
Homologues between oomycetes, fungi, brown algae, and diatoms were identified using OrthoMCL (e-value cut-off of 1e−20 or 1e−5) [37].
Synteny between multiple species was analysed using the Artemis Comparison Tool [86]. Alignments between genomic sequences were performed using TBLASTX with a score cut-off of 210. Annotations of P. infestans, Py. ultimum, and H. arabidopsidis were transferred using TBLASTN with an e-value cut-off of 1e−30. LTR_FINDER [87] was used to annotate long terminal repeats (LTRs) within the genomic sequences, and coordinates were manually added. Regions between LTRs were blasted against RepBase [74] to identify the presence and/or type of transposon.
Secreted proteins were predicted using a local installation of SignalP 3.0 [88]. Proteins were considered to be secreted if both the neural networks and hidden Markov model methods predicted the protein to have a signal peptide. Predictions of TM domains were performed after removing the predicted secretion signal. TM domains were identified using MEMSAT3 [89]. Proteins were considered to be without a TM domain with pnon-TM>0.0004 or, for high stringency, pnon-TM>0.01.
To identify new motifs, subsets of secreted proteins were selected and analysed using MEME [90] with default parameters. Identified motifs were tested against the whole gene set and the Swiss-Prot database using MOTIF Search. In a second step, motifs were selected only if they were positioned within 50 amino acids after the secretion signal.
Tests for over-representation of an identified motif were done using motif and sequence shuffling. Secreted proteins were predicted [88] as described in the previous section, and the signal peptide was removed prior to further analyses.
Each of the sequences without secretion signal was randomly shuffled 30 times. After each shuffling the sequences were screened for the motif in question. If the motif was identified after shuffling, the sequence was excluded from the next round. If the motif was never identified within the 30 times shuffling, the motif in the original protein was counted as “unique empirical”. All possible combinations of the amino acid sequence within the motif were calculated. For each of these permutations, the “unique empirical” proteins were calculated.
The 30 times shuffling was repeated 1,000 times to calculate background levels. Background levels were defined as how often a sequence was found again having the motif or the permutated motif. This was called “background (mean)”. Motifs that were above this background were considered for further analyses.
The second criterion was if a motif was significantly enriched in the secretome compared to all non-secreted proteins. For statistical validations we calculated the cumulative hypergeometric probability.
Candidates for further experiments were evaluated according to a ranking list. Maximum possible score was nine points, and the following scores were given: one point for being on a shorter, repetitive contig (≤3,000 bp) or end of contig, since we assumed that effector candidates might be in repetitive regions as shown for P. infestans effectors [9]; one point for having cDNA support; two points for being a short protein (≤400 amino acids); two points for carrying one of the identified motifs (RXLR, RXLQ, CHXC, CRN); one point for being expressed before day 10 after infection; one point for being expressed before day 4 after infection; and one point for showing SNPs in the Em1 comparison.
Candidate RXLR effectors were cloned from RXLR to stop; all other candidate effectors were cloned from SP cleavage site to stop into pENTR D-TOPO (Invitrogen) and mobilized into pEDV6 [70]. The resulting effector∶pEDV6 constructs were conjugated into Pst DC3000 luxCDABE [69] and Pst DC3000 ΔAvrPto/ΔAvrPtoB [92]. The contribution of an individual effector was assessed by spray inoculating 4- to 5-wk-old short day grown plants as previously described [93].
Growth of Pst DC3000 luxCDABE effector∶pEDV6 was calculated by measuring whole plant luminescence using a Photek camera system and normalizing this to plant fresh weight [69].
To assess the virulence of Pst DC3000 ΔAvrPto/ΔAvrPtoB effector∶pEDV6, bacterial colony counts were performed as previously described [94].
All Illumina sequence reads generated during this study have been submitted to the Sequence Read Archive at EBI and are accessible under the accession number ERA015557. Individual studies are available with accession numbers ERP000440 (Alias: albugo_laibachii_nc14_dna_sequencing, http://www.ebi.ac.uk/ena/data/view/ERP000440), ERP000441 (Alias: albugo_laibachii_nc14_cdna_sequencing, http://www.ebi.ac.uk/ena/data/view/ERP000441), and ERP000442 (Alias: albugo_laibachii_em1_dna_resequencing, http://www.ebi.ac.uk/ena/data/view/ERP000442).
All contigs and annotations are available through EBI or NCBI. The accession range is from FR824046 to FR827861 (3,816 contigs including annotations) and can be accessed through the ENA browser (http://www.ebi.ac.uk/ena/).
|
10.1371/journal.pgen.1006199 | Determining the Effect of Natural Selection on Linked Neutral Divergence across Species | A major goal in evolutionary biology is to understand how natural selection has shaped patterns of genetic variation across genomes. Studies in a variety of species have shown that neutral genetic diversity (intra-species differences) has been reduced at sites linked to those under direct selection. However, the effect of linked selection on neutral sequence divergence (inter-species differences) remains ambiguous. While empirical studies have reported correlations between divergence and recombination, which is interpreted as evidence for natural selection reducing linked neutral divergence, theory argues otherwise, especially for species that have diverged long ago. Here we address these outstanding issues by examining whether natural selection can affect divergence between both closely and distantly related species. We show that neutral divergence between closely related species (e.g. human-primate) is negatively correlated with functional content and positively correlated with human recombination rate. We also find that neutral divergence between distantly related species (e.g. human-rodent) is negatively correlated with functional content and positively correlated with estimates of background selection from primates. These patterns persist after accounting for the confounding factors of hypermutable CpG sites, GC content, and biased gene conversion. Coalescent models indicate that even when the contribution of ancestral polymorphism to divergence is small, background selection in the ancestral population can still explain a large proportion of the variance in divergence across the genome, generating the observed correlations. Our findings reveal that, contrary to previous intuition, natural selection can indirectly affect linked neutral divergence between both closely and distantly related species. Though we cannot formally exclude the possibility that the direct effects of purifying selection drive some of these patterns, such a scenario would be possible only if more of the genome is under purifying selection than currently believed. Our work has implications for understanding the evolution of genomes and interpreting patterns of genetic variation.
| Genetic variation at neutral sites can be reduced through linkage to nearby selected sites. This pattern has been used to show the widespread effects of natural selection at shaping patterns of genetic diversity across genomes from a variety of species. However, it is not entirely clear whether natural selection has an effect on neutral divergence between species. Here we show that putatively neutral divergence between closely related species (human and chimp) and between distantly related pairs of species (humans and mice) show signatures consistent with having been affected by linkage to selected sites. Further, our theoretical models and simulations show that natural selection indirectly affecting linked neutral sites can generate these patterns. Unless substantially more of the genome is under the direct effects of purifying selection than currently believed, our results argue that natural selection has played an important role in shaping variation in levels of putatively neutral sequence divergence across the genome. Our findings further suggest that divergence-based estimates of neutral mutation rate variation across the genome as well as certain estimators of population history may be confounded by linkage to selected sites.
| Determining the evolutionary forces affecting genetic variation has been a central goal in population genetics over the past several decades. A large body of empirical and theoretical work has suggested that neutral genetic variation within a species (diversity) can be influenced by nearby genetic variants that are affected by natural selection (reviewed in [1]). This can occur via two mechanisms. In a selective sweep, a neutral allele linked to a beneficial mutation will reach high frequency [2,3]. Selective sweeps reduce neutral genetic variation near regions of the genome that are directly affected by natural selection. The second process, background selection, also reduces neutral genetic variation [4–7]. Here, purifying selection that eliminates deleterious mutations also removes nearby neutral genetic variation. Many empirical studies have found strong evidence for the effects of background selection and selective sweeps affecting patterns of neutral genetic diversity (intra-species DNA differences) across the human genome. For example, several studies have reported a correlation between genetic diversity and recombination rate [8–13]. This correlation can be driven by selective sweeps and background selection because these processes affect a larger number of base pairs in areas of the genome with a low recombination rate than with a high recombination rate. Additionally, other studies found reduced neutral genetic diversity surrounding genes [12–17], which is consistent with the idea that there is more selection occurring near functional elements of the genome.
While the evidence for natural selection reducing genetic diversity at linked neutral sites is unequivocal, the effect of natural selection on linked neutral divergence between species (inter-species DNA differences) is less clear. Elegant theoretical arguments have suggested selection does not affect the substitution rate at linked neutral sites [18,19]. However, these theoretical arguments do not include mutations that arose in the common ancestral population, the population that existed prior to the split and formation of two descendant lineages. Such ancestral polymorphism has been shown to be a significant confounder in estimating population divergence times [20]. When also including ancestral polymorphism, it becomes less clear whether selection affects divergence at linked neutral sites.
Based on coalescent arguments, neutral polymorphism in the ancestral population will be affected by linkage to selected sites the same way as genetic diversity within a population (Fig 1). Presumably, neutral divergence between closely related species, with lots of ancestral polymorphism, could be affected by selection. Indeed, McVicker et al. [15] demonstrated that background selection could explain the variation in human-chimp neutral divergence across the genome. Additionally, Cruickshank and Hahn [21] found that divergence between recently separated species pairs was reduced in regions of low recombination and in “islands of speciation”. They attributed at least some of these patterns to selection affecting linked neutral sites.
However, the reduction in neutral diversity in the ancestral population is thought to have a negligible effect and/or be undetectable when considering neutral divergence from species with a very long divergence time [9,18] because there would be many opportunities for mutations to occur after the two lineages split (Fig 1). These neutral mutations that occur after the split would not be influenced by selection at linked neutral sites [18] and would dilute the signal from the ancestral polymorphism. Thus, it is generally believed that selection at linked neutral sites should not affect divergence between distantly related species. An example of this argument was presented by Hellmann et al. [9]. They argued that the positive correlation between human-baboon divergence and human recombination was due to mutagenic recombination, rather than selection affecting linked neutral sites, because of the long split time between humans and baboons (>20 million years). Reed et al. [22] suggested that though it is unlikely background selection by itself could explain the entire correlation observed by Hellmann et al., background selection may still contribute to divergence. However, there has been little quantitative investigation of the effect that selection has on divergence at linked neutral sites among distantly divergent species when including ancestral polymorphism.
In addition to conflicting conceptual predictions about the expected effect of selection on divergence at linked neutral sites, empirical studies also have been ambiguous. While some studies found no evidence for a correlation between divergence and recombination such as in Drosophila [23,24] or in yeast [25], other studies have reported correlations between divergence and recombination in Drosophila [26,27]. Further, positive correlations between human-chimpanzee divergence and human recombination rate [10,12,13], human-macaque divergence and human female recombination rate [28], or human-baboon divergence and human recombination rate [9] have been reported. Finally, even though there was evidence for a strong reduction in human-chimpanzee divergence and human-macaque divergence surrounding genes [15,28], McVicker et al. [15] attributed the reductions seen for human-dog divergence to variation in mutation rates. Thus, the degree to which divergence is affected by selection across species with different split times remains elusive.
Determining whether and how selection affects linked neutral divergence is critical to understanding the evolutionary forces influencing genetic variation and mutational processes. If selection in the ancestral population only has a limited effect on divergence, it would suggest correlations between recombination and divergence to be evidence of mutagenic recombination. This may further suggest the need to consider recombination rates when modeling variation in mutation rates across the genome [9,29–32]. Because mutations rates have been difficult to estimate reliably in humans [33,34], understanding the biological factors influencing them will be of paramount importance for obtaining improved estimates. If, on the other hand, selection can affect linked neutral divergence, reductions of linked neutral divergence surrounding genes would suggest an abundance of selection affecting linked neutral sites [35]. Selection affecting linked neutral diversity and divergence is at odds with the neutral and nearly neutral theories [36–38], which have been the prevailing views in molecular population genetics for the last several decades. It would also suggest the need to consider the effects of selection when estimating mutation rates from neutral divergence.
Here we aim to examine the effects of selection on linked neutral divergence for pairs of species with a range of split times. We first present evidence that neutral divergence is reduced at putatively neutral sites close to selected sites across a wide range of taxa, including those with split times as long as 75 million years ago. Factors such as hypermutable CpG sites, GC content, or biased gene conversion by themselves cannot explain these results. We then use coalescent simulations to explore whether models incorporating background selection in the ancestral population could generate the empirical patterns. We also present a theoretical argument as to how background selection can affect variation in neutral divergence across the genome, even for species with a long split time such as human and mouse. Finally, we show that purifying selection directly reducing divergence at putatively neutral sites cannot explain these findings unless a large fraction of the genome is directly under selection, or there is a substantial number of sites under selection in the human or mouse lineage that are not conserved across species. Even though we cannot formally reject the direct effects of purifying selection from driving some of these correlations, our empirical and simulation-based findings indicate that natural selection can indirectly affect neutral genetic divergence. In sum, the view that selection does not affect divergence at linked neutral sites between distantly diverged species should be re-considered.
We wished to test whether the genetic divergence at a linked neutral site is influenced by the indirect effects of natural selection. As such, we set out to obtain putatively neutral sites by removing sites that were potentially functional and under the direct effects of purifying selection. In particular, a site was considered putatively neutral if it was (1) located at least 5kb from the starting or ending position of an exon, (2) not located within a phastCons element that was calculated over different phylogenic scopes, (3) not alignable between human and zebrafish, and (4) not found within the top 10% of most conserved Genomic Evolutionary Rate Profiling (GERP) scores [39]. Criteria 2 and 3 remove sites that are likely to be conserved across species and therefore not neutral. We chose these filtering criteria following previous studies [12,13,40]. Additionally, we chose to remove the top 10% of sites having the most extreme GERP scores because previous work suggests <10% of the genome was under the direct effect of selection [39,41–49]. The putatively neutral sites close to genes show comparable levels of divergence to four-fold degenerate sites (S1 Fig, S1 Table). As four-fold degenerate sites are often used as a neutral standard in molecular evolution, the fact that they show similar levels of divergence as our putatively neutral noncoding sites argues that our putatively neutral sites are unlikely to be under additional direct effects of selection.
To understand the evolutionary factors affecting linked neutral divergence between closely related species, we examined human-primate divergence, particularly human-chimp divergence and human-orangutan divergence. First, we explored the relationship between neutral human-primate divergence and functional content, defined as the proportion of sites within a 100kb-window that overlapped with an exon or a phastCons region. We hypothesized that if natural selection contributes to the reduction of divergence at linked neutral sites, its effect would be more pronounced at regions with greater functional content [14]. This hypothesis predicts a negative correlation between functional content and neutral divergence. To test this, we divided the human genome into non-overlapping windows of 100kb and obtained putatively neutral divergence for each window as described above. We found a negative correlation between functional content and neutral divergence between pairs of closely related species (Spearman’s ρhuman-chimp = -0.235, P < 10−16, Spearman’s ρhuman-orang = -0.204, P < 10−16, Fig 2A and 2B, S2 Table).
We next examined the relationship between human-primate neutral divergence and broad-scale human recombination rate which we obtained from the deCODE genetic map [50]. While recombination has not been conserved throughout evolutionary history, the recombination rate at the broad-scale level (i.e. 100kb) was shown to be correlated between human and chimp [51,52]. We found a positive correlation between neutral human-primate divergence and human recombination rate (Spearman’s ρhuman-chimp = 0.234, P < 10−16, Spearman’s ρhuman-orang = 0.249, P < 10−16, Fig 2C and 2D, S3 Table), which indicates that neutral human-primate divergence is reduced in regions of low recombination rate. Additionally, when we stratified windows into those that were near genes and those that were far from genes based on the proportion of sites in each window that overlapped with a RefSeq transcript, we found that the correlation between divergence and recombination is stronger for windows with a higher overlap with RefSeq transcripts (S2 Fig). These observations indicate that neutral divergence is reduced at sites that are more tightly linked to those under the direct effect of selection, consistent with the hypothesis that natural selection indirectly reduces linked neutral divergence.
These two correlations are robust to the presence of multiple confounding factors. First, the correlations are robust to the choice of window size used for analysis as they persisted when using 50 kb windows (S2 and S3 Tables). Second, some features of the genome such as hypermutable CpG sites or GC content are known to correlate with genic content [10,12,13]. To test whether these features confounded the correlations found in our data, we repeated our analyses removing potential CpG sites by omitting sites preceding a G or following a C [15]. The correlations persisted after filtering out CpG sites (S2 and S3 Tables). We next computed partial correlations controlling for GC content. Similarly, we found that the correlations persisted (S2 and S3 Tables).
Biased gene conversion is an additional evolutionary force that has been shown to influence patterns of divergence [53,54]. In this process, double-strand breaks in the DNA in individuals heterozygous for AT/GC variants will be preferentially repaired with the GC allele, resulting in AT → GC substitutions occurring at a higher rate than GC → AT substitutions [54–56]. To control for the effects of biased gene conversion on this analysis, we filtered out sites that could be affected by removing any AT → GC substitutions genome-wide. The negative correlation between human-primate divergence and functional content did not change after controlling for biased gene conversion (S2 Table). Though the positive correlation between human-primate divergence and human recombination decreased after this filter (from 0.234 to 0.108), it still remained significant (S3 Table). Thus, the observed correlations are unlikely to be driven solely by choice of window size or mutational properties based on sequence composition. Because biased gene conversion appears to contribute to some of the correlation between divergence and recombination rate, subsequent analyses of this correlation use the divergence dataset filtered for biased gene conversion.
We next explored the evolutionary forces affecting divergence between more distantly related pairs of species, specifically human-mouse and human-rat. These species were predicted to have diverged approximately 75 million years ago [41] and, as such, current thinking would predict that natural selection would not affect linked neutral sites. Similar to what was seen for the closely related species, functional content is negatively correlated with neutral human-rodent divergence (Spearman’s ρhuman-mouse = -0.184, P < 10−16, Spearman’s ρhuman-rat = -0.149, P < 10−16, Fig 3A and 3B, S4 Table). This negative correlation persisted when using 50kb windows and also after accounting for the confounding factors of hypermutable CpG sites, GC content, and GC-biased gene conversion (S4 Table).
Since the broad-scale recombination rate at 100kb appears to have changed over the course of evolution of the species [57], we looked for other potential signatures of whether natural selection has affected linked neutral divergence. In particular, we examined the relationship between human-rodent divergence and the strength of background selection across the genome inferred from divergence within primates [15]. This strength of background selection is captured by the B-value, which represents the degree to which neutral variation at a given position is reduced by selection relative to neutral expectations. While McVicker et al. [15] concluded that divergence between primates was indeed reduced due to background selection, they did not consider human-mouse divergence in their analyses and did not model background selection within the human-dog ancestor. As such, there is no a priori reason why the B-values of McVicker et al. [15] should be related to human-mouse divergence.
Nevertheless, we found a positive correlation between human-rodent divergence and the B-values from McVicker et al. [15] (Spearman’s ρhuman-mouse = 0.445, P < 10−16, Fig 3C, S5 Table, Spearman’s ρhuman-rat = 0.402, P < 10−16, Fig 3D, S5 Table). The positive correlation between human-rodent divergence and B-values remained significant even after accounting for the confounding factors of CpG sites, GC content, and GC-biased gene conversion. Similarly, these correlations remained when using 50kb windows (S5 Table). Taken together, the empirical correlations are consistent with the hypothesis that natural selection has contributed to reducing neutral divergence at linked sites even between species with a long split time such as human and mouse.
To test whether a model including background selection in the ancestral population can explain the empirical observations regarding neutral human-primate divergence and neutral human-rodent divergence, we used a coalescent simulation approach. To a first approximation, the effect of background selection in a sample size of two chromosomes can be accounted for by scaling the ancestral population size by the strength of background selection [4,5,7,15,58–62]. Thus, we modeled the effect of background selection as a reduction in the ancestral population size using the B-values estimated in McVicker et al. [15]. Briefly, we first used ms [63] to generate genetic variation in the ancestral population where the ancestral population has size NaB. Then we simulated mutations that accumulated since the split between two species using a Poisson process. The total divergence was the sum of the mutations in the ancestral population and mutations accumulated since the split (see Methods). We modeled mutation rate variation by drawing a mutation rate for each window from a gamma distribution. We chose values for the parameters of the gamma distribution as well as the ancestral population size (Na) such that the mean and standard deviation of the simulated divergence across the genome and the correlation coefficients between divergence and other functional properties were similar to those seen empirically (S3 Fig, S6 and S7 Tables, Methods).
We first examined which models could generate the observed correlation between recombination and human-chimp divergence. Here we use the value of Spearman’s ρ estimated from the data after filtering out sites that could be affected by biased gene conversion (ρ = 0.108). When considering a model without background selection (i.e. B = 1 for all windows), the average value of Spearman’s ρ between human-chimp divergence and recombination rate was 0.042, and none of the 500 simulation replicates approached the value of Spearman’s ρ seen empirically (Fig 4A, white histogram). On the other hand, when modeling background selection using the McVicker B-values, the average Spearman’s ρ was 0.107 which was comparable to the Spearman’s ρ computed from empirical human-chimp divergence with human recombination after accounting for biased gene conversion (Fig 4A, gray histogram).
We then tested whether a model incorporating background selection could generate a positive correlation between neutral human-rodent divergence and B-values as observed empirically. We modified our simulation approach to account for the difference in generation time between human and mouse (see Methods). When considering models without background selection (i.e. B = 1 for all windows), the average value of Spearman’s ρ was 0.012, and none of the 500 simulation replicates approached the value of Spearman’s ρ seen empirically (Fig 4B, white histogram). However, when modeling background selection using the McVicker B-values, the average Spearman’s ρ was 0.446 which was comparable to the Spearman’s ρ computed from empirical human-mouse divergence and McVicker’s B-values (Fig 4B, gray histogram).
In sum, our results suggest that for a given set of parameters, a model with background selection in the ancestral population can generate the correlations observed in the empirical data (i.e. a positive correlation between neutral human-primate divergence and human recombination and a positive correlation between neutral human-rodent divergence and B-values) whereas neutral coalescent models cannot.
Current thinking argues that natural selection affecting linked neutral sites is not a plausible explanation for the reduction in neutral divergence between pairs of species with a long split time such as human-mouse or human-rat. Here, we outline a theoretical analysis of a simple two-locus model to gain intuition about how the mutation rate (μ), strength of background selection (B), and ancestral population size (Na) affect the degree to which background selection can affect divergence (Fig 5A).
If background selection has any effect on the variation in neutral divergence across the genome, this can only be due to its effect on divergence in the ancestral population, since deleterious mutations do not affect the fixation rate at linked neutral sites [18]. Recombination in the ancestral population results in a distribution of coalescent times within each locus, with an average coalescent time of t¯. We assumed that the recombination rate within each locus is large enough, such that there is no variation in t¯ for a fixed value of B, i.e. Var[t¯|B]≈0. This is a reasonable assumption as long as the window size and recombination rate are not too small. Recombination events cause the sequence to be broken into independent segments, such that for a total ρ > 10 (where ρ denotes the population-scaled recombination rate, 4Ner) the variance in t¯ approaches zero [64]. For an average 100kb window in the human genome (r = 10−8/bp, Ne = 10,000), ρ is 40 and thus this assumption holds true. Any difference in t¯ between loci is then only attributable to differences in background selection: E[t¯|B]=2NaB. Further, variation in ancestral (da) and total (dt) divergence results from a Poisson distributed number of mutations added to the genealogy, such that Var[da|B] = E[da|B] = 4NaBμL and Var[dt|B] = E[dt|B] = E[da|B]+2tsplitμL where L is the sequence length of a locus. The law of total variance can be used to compute the variance in total divergence across loci with varying levels of background selection:
Var[dt]=VarB[E[dt|B]]+EB[Var[dt|B]]
Thus, variance in total divergence can be decomposed into variance due to background selection and variance due to the mutational process. For simplicity, the first locus experiences no background selection (B1 = 1), and the second locus experiences some fixed amount of background selection (0 ≤ B2 ≤ 1). Under this model, we computed the variance due to background selection as:
VarB[E[dt|B]]=((E[dt|B=1] – E[dt|B=B2])/2)2.
We then computed the variance due to the mutational process as:
EB[Var[dt|B]]=(Var[dt|B=1]+Var[dt|B=B2])/2.
We assumed an old split time, such that the divergence that accumulated from present time to population split is similar to the human-mouse divergence (40%). Both loci have a sequence length (L) of 100kb. Our theoretical analysis of variance approach shows that with this old split time and assuming a low mutation rate of 1 x 10−9/bp, more than 20% of the variation in the divergence can be explained by background selection in the ancestral population with the following conditions: ancestral population size > 600,000 and B < 0.2 (Fig 5A, panel 1, blue, purple, and pink lines). Note that under these conditions, the proportion of divergence that accumulated in the ancestral population can be as low as 0.3% (Fig 5B, panel 1). However, the proportion of the variance in divergence that is attributable to the ancestral population is larger than 20% (Fig 5C, panel 1), mainly due to background selection leading to differences in t¯ between loci. With a larger mutation rate (2 x 10−8/bp), background selection results in a stronger effect on variation in divergence even when ancestral population size is relatively small (>50,000; Fig 5A, yellow line). When assuming a moderately large population size of 200,000, and a moderate strength of background selection (B = 0.75), then as much as 50% of variance in divergence can be explained by background selection (Fig 5C, light green line). Nonetheless, the proportion of divergence that accumulated in the ancestral population in this case is still only 3.4%. Collectively, even for old split times, where the vast majority of divergence accumulated after the population split, with certain assumptions about the ancestral population size, mutation rate, and strength of background selection, the variance in the divergence could be explained by background selection.
Because the theoretical model described above ignores regions of low recombination and only considers one pair of loci at a time, we used coalescent simulations (similar to what we outlined above) to examine whether background selection could generate the positive correlation between estimates of background selection in primates and divergence between distantly related species using more realistic models. Since we were not particularly concerned with any specific species, we simplified these simulations by setting the mutation rate to 2.5 x 10−8/bp.
We found that across all population sizes and split times examined, background selection generated a positive correlation between recombination and divergence as well as a positive correlation between divergence and B-values, even for pairs of species that split up to 100N generations ago (S4 Fig, black lines and dashed lines). This correlation remained strong even when the proportion of the divergence due to ancestral polymorphism was small. For example, for a pair of populations with tsplit = 100N generations and an ancestral population of size 50,000, only 1.53% of the divergent sites are due to ancestral polymorphism (S4 Fig, red lines). However, this model predicts a correlation of 0.211 between recombination and divergence and a correlation of 0.377 between recombination and B-values. Although ancestral polymorphism only contributes in a small way to the total divergence, the variance in the amount of ancestral polymorphism across the windows accounts for nearly 60% of the variance in divergence across different windows (S5 Fig, black lines). In general, the correlations decreased as both the split time increased and the size of the ancestral population decreased (S5 Fig). This behavior is expected as the contribution of the variance in levels of ancestral polymorphism to the variance in divergence decreases with increasing split time and decreasing ancestral population size (S5 Fig).
While we have shown under a variety of models that natural selection can affect putatively neutral divergence and generate the correlations that we observe empirically, other selective scenarios could explain these patterns. An alternative explanation for the empirical correlations reported in Figs 2 and 3 is that the filtering criteria we used to obtain neutral sites did not effectively remove all non-neutral sites. Therefore, the observed correlations could be due to the direct effects of purifying selection reducing genetic divergence. As sites under purifying selection may be located close to conserved functional elements and could conceivably result in low B-values, this is a potentially plausible explanation for our findings. As our current filters removed the 10% of the genome that was most likely under the direct effect of selection based upon the top 10% of GERP scores, we reasoned that additional sites under purifying selection would have elevated GERP scores relative to neutrality.
To test this hypothesis, we repeated our correlation analyses by first obtaining the neutral human-primate divergence and neutral human-rodent divergence using different GERP score cutoffs (i.e. 5% to 25%). When examining human and primate pairs, the correlation between neutral human-primate divergence and functional content decreased as a function of increasing GERP cutoff score (Fig 6A). Nevertheless, the negative correlation between neutral human-primate divergence and functional content remained significant even after removing any site whose GERP score fell within the top 25% of the distribution (Spearman’s ρhuman-chimp = -0.189, P < 10−16, Spearman’s ρhuman-orang = -0.122, P < 10−16, S6A and S6B Fig). On the other hand, the relationship between neutral human-primate divergence and human recombination rate were not affected by varying GERP score cutoffs (Fig 6B, S6C and S6D Fig).
When examining human and rodent pairs, we found that the negative correlation between human-rodent divergence and functional content decreased as a function of increasing GERP score cutoff. Further, the relationship became nonsignificant when filtering any site whose GERP score fell within the top 15th percentile (Fig 6C, S7A and S7B Fig). The positive correlation between neutral human-rodent divergence and McVicker’s B values decreased as a function of increasing GERP score cutoff, but remained significantly positive even after removing any sites whose GERP score fell within the top 25th percentile (Fig 6D, S7C and S7D Fig). Still, this latter pattern indicates that the direct effects of natural selection are unlikely to explain our findings, unless the selected sites are not in the upper 25% of the GERP score distribution.
To test whether background selection could explain these correlations when removing the 25% of the genome with the most conserved GERP scores, we used our coalescent simulation framework. These simulations match the empirical distribution of divergence across the genome (S8 Fig, S7 Table) and use the parameters given in S6 Table. For human-chimp divergence, none of the 500 neutral coalescent simulations resulted in a Spearman’s ρ between divergence and human recombination rate as large as observed empirically after filtering sites affected by biased gene conversion (S9A Fig, white histogram). On the other hand, simulations including background selection in the ancestral population generated a Spearman’s ρ between divergence and human recombination rate similar to what was observed empirically after filtering sites affected by biased gene conversion (S9A Fig, gray histogram). Similarly, for human-mouse divergence, while none of the 500 coalescent simulations using the neutral model could generate a Spearman’s ρ between divergence and McVicker’s B-values as large as the empirical correlation, models including background selection in the ancestral population could generate this correlation (S9B Fig).
Here we have examined patterns of divergence between pairs of species with various degrees of divergence. We document several signatures that are consistent with the action of natural selection reducing divergence at linked neutral sites. First, for all pairs of species considered, we find that neutral divergence is lowest in regions of the genome with the greatest functional content (Figs 2 and 3). This pattern may be expected if more selection occurs in regions of the genome with greater functional content. Second, human-primate neutral divergence strongly correlates with human recombination rate and the correlation persists after accounting for hypermutable CpG sites, GC content, and biased gene conversion. Regions of low recombination show lower levels of divergence, which is consistent with selection having a greater effect on linked neutral sites in regions of low recombination. The correlation between human-primate divergence and human recombination is higher in regions with greater overlap with RefSeq transcripts, indicative of a greater reduction in neutral divergence in regions near genes as opposed to far from genes (S2 Fig). Third, human-rodent neutral divergence strongly correlates with the strength of background selection estimated for primates. These correlations persist after accounting for CpG sites, GC content, and biased gene conversion. Importantly, coalescent simulations including background selection can generate several of these correlations. However, neutral coalescent models without background selection do not.
One interesting observation made was that while most of our correlation analyses were robust to the confounding effect of biased gene conversion, the correlation between human-primate neutral divergence and recombination rate was affected significantly by biased gene conversion. This suggests that while some of the correlation between recombination and divergence can be driven by biased gene conversion, it cannot explain the entire correlation. This result also argues that when testing for a correlation between divergence and recombination, the effect of biased gene conversion should be taken into account.
While we found that models incorporating background selection predict correlations comparable to the empirical data, in principle, several other evolutionary processes may be able to generate these patterns. First, selective sweeps in the ancestral population could reduce divergence just like background selection. Given that we are unlikely to be able to survey patterns of polymorphism in the human-mouse ancestor in more than two lineages, it will be difficult or nearly impossible to distinguish between these two types of selection at linked neutral sites. Thus, one should interpret our use of B-values as reflecting a reduction in divergence due to the combined effects of both background selection and selective sweeps, as suggested in McVicker et al. [15].
A second possibility is that the negative correlation between divergence and functional content as well as the positive correlation between divergence and B-values could be driven by variation in mutation rate across the genome. Indeed, McVicker et al. [15] attributed a positive correlation between B-values and human-dog divergence to the effects of variable mutation rates. However, for this mechanism to explain our results, it would require that mutation rates would have to be lower closer to genes and in regions of the genome thought to experience more background selection (i.e. in regions with lower B-values). There is some limited evidence of this effect in Arabidopsis where mutation rates are higher in regions of the genome with greater heterozygosity [65]. However, the extent to which these results apply to mammalian genomes remains unclear. Further, other studies in humans do not support the view that mutation rates are systematically lower in regions of the genome more subjected to selection. Recent estimates of the de novo mutation rate have not found any evidence of a reduction close to genes [32]. Further, Palamara et al. [66] found that their estimates of the mutation rate do not differ as a function of B-values. Variation in mutation rate across the genome, while inflating the variance in divergence across the genome, would not be predicted to generate correlations between B-values and divergence as well as the correlation between functional content and divergence. Thus, we can rule it out as the sole explanation for the empirical patterns seen in our study.
Further, mutagenic recombination is unlikely to explain the empirical patterns in our study because the correlation between divergence and functional content does not depend on recombination rate. The negative correlation between divergence and functional content remained strong when controlling for variation in recombination rates (S2 Table), suggesting our results are unlikely to be driven by mutagenic recombination. Nevertheless, our results do not rule out the possibility of mutagenic recombination and this topic certainly warrants further investigation.
Another possibility is that the reduction in neutral divergence near genes and in regions with lower B-values could be due to the direct effects of purifying selection removing variation from the population. Current evidence from a variety of comparative genomic studies suggests <10% of the genome is under purifying selection [39,41–49]. We attempted to mitigate the direct effects of purifying selection by employing a conservative set of filters in order to obtain putatively neutral sites. When removing the 10% of the genome that is most conserved, using a variety of conservation metrics, the correlations persisted, suggesting they were not driven by the direct effects of selection. However, when we removed the top 15% of sites with the most conserved GERP score, the correlation between human-rodent divergence and functional content disappeared. This finding suggests that either the GERP scores themselves are affected by background selection, or, instead, that this correlation is driven, in part, by the direct effects of purifying selection. However, in order for direct purifying selection to explain the correlation, either more of the genome (at least 15%) would have to be under selection than suggested by current estimates [39,41–49] or many of the sites in the top 15% most conserved GERP scores would have to be neutrally evolving. Additionally, the negative correlation between human-chimp divergence and functional content, the positive correlation between human-chimp divergence and recombination rate, and the positive correlation between human-mouse divergence and B-values, remained even after removing the 25% of the genome that is most conserved (Fig 6). This implies that even such a large amount of functional sites under selection cannot explain all of our results. Finally, an additional line of evidence suggesting that the putatively neutral sites close to genes are not subjected to the direct effects of purifying selection stems from the fact that they show similar levels of neutral divergence to four-fold degenerate sties (S1 Fig, S1 Table). Thus, our putatively neutral noncoding sites have levels of divergence comparable to those seen for sites solely subjected to background selection.
Additionally, our filters rely on functional annotations and conservation to remove functionally important sites directly under the effects of selection. It is formally possible that the direct effects of selection could generate the correlations seen in our study if there are sites under selection that were invisible to the conservation-based filters used in our study. This could occur if there are recently derived, lineage-specific functional elements under selection that cannot be picked up by conservation metrics, or if there are sequences subject to purifying selection in the ancestral population but subsequently became neutral and therefore were not conserved. While we cannot exclude such a scenario, current population genetic evidence provides, at most, limited support for such an explanation [44,46,47,67].
One limitation in this study is that we made many assumptions regarding the parameters used in the simulations such as the ancestral population size, generation times, and mutation rates over the last 5–7 million years between human and chimp and 75 million years between human and mouse. There is much uncertainty surrounding all of these parameters [41,68–72]. Overall, we used a set of parameters in which the simulated divergence dataset from the coalescent simulations matched closely with the mean and standard deviation of the empirical divergence dataset. This allowed us to assess whether a simple neutral model could result in the correlations as large as observed empirically or whether a model with background selection needed to be invoked. We utilized the coalescent simulations as a proof of concept and therefore, the parameters we used in these sets of simulations should not be taken as estimates of the true values. Estimation of these parameters (ancestral population size, mutation rate, split time, etc.) is beyond the scope of this study and certainly warrants further in-depth investigation.
Other studies have argued that selection will not affect linked neutral divergence between distantly related species because the genealogy in the ancestral population only comprises a small proportion of the total genealogy between one chromosome from each of the two species [9,18,27]. This means that ancestral polymorphism will only account for a small proportion of the total divergence between distantly related species. It was thought that the signature of selection reducing the genealogy in the ancestral population would be diluted by the mutations that occurred since the split. As such, there would be no detectable signature of selection. Our theoretical results and simulations show the proportion of ancestral polymorphism actually is a poor predictor of the correlation between divergence and recombination as well as between divergence and B-values. For example, consider a pair of species that split N generations ago with an ancestral population size of 25,000. In this model, 40% of the divergence is attributable to ancestral polymorphism (S4A Fig). Now consider a second pair of species that split 100N generations ago where Na = 200,000. Here <5% of the divergence is due to ancestral polymorphism (S4D Fig). Previous intuition suggests the effect of background selection would be stronger in the first pair of species because they split more recently and ancestral polymorphism makes a greater contribution to divergence. However, our simulations show the exact opposite pattern (S4A and S4D Fig). The correlation between B-values and divergence is higher in the model with the more ancient split (Spearman’s ρ = 0.610) than the one with the more recent split (Spearman’s ρ = 0.452). Similar results are seen for the correlation between recombination rate and divergence. The reason for this discrepancy is that the main driver of these correlations is not the average amount of ancestral polymorphism, but rather the contribution to the variance in divergence due to the variance in ancestral polymorphism. Even when ancestral polymorphism makes only a small contribution to the overall average divergence, a substantial amount of the variance in total divergence across the genome can still be explained by variance in ancestral polymorphism, particularly if the ancestral population size is large. Our theoretical results suggest that the variance in the amount of background selection in different regions of the genome can account for a lot of the variance in total divergence, even for species that split long ago. In sum, our theoretical results and simulations suggest that previous intuition has understated the importance of even small amounts of ancestral polymorphism on the variability of genome-wide patterns of divergence between species.
Our results have important implications for understanding patterns of genetic variation and divergence across genomes. First, our findings add to the growing literature suggesting the importance of background selection at shaping genome-wide patterns of variability across species [1,7,13,15,16,60,62,73–80]. Our new contribution to this literature is demonstrating that natural selection can affect neutral divergence, even between distantly related species. Second, our work suggests that estimators of mutational properties that rely on contrasting patterns of divergence across different parts of the genome that may be differentially affected by background selection may yield biased results. This effect has been studied within primates in greater detail in recent work [81]. Third, the fact that we detect evidence of background selection between distantly related species suggests that there is still some information about the distribution of coalescent genealogies across the genome. This distribution of coalescent genealogies can be exploited to obtain more reliable estimates regarding the human-mouse ancestral population size. While several methods exist to estimate ancestral demographic parameters from divergence [82–86], we suggest that these methods may be applicable for very distantly related species. Our finding that background selection can increase the variance in coalescent times across the genome suggests these methods as well as other statistical methods which seek to infer demographic history from the distribution of coalescent times across the genome, such as the PSMC approach [87], should account for the increased variance in coalescent times across the genome due to background selection. Not accounting for background selection could result in inferring spurious demographic events to account for the additional variance in coalescent times across the genome as has recently been suggested for positive selection [88]. Lastly, our results suggest a need for caution when using patterns of divergence to calibrate neutral mutation rates. Some of the variation in divergence across the genome may be due to varying coalescent times, further accentuated by selection, rather than differing mutation rates [20,89]. Future work could explore the extent to which selection at linked neutral sites can explain the discrepancies between different types of estimates of mutation rates [33,34].
We obtained the pairwise (.axt) alignments between human/chimpanzee (hg18/panTro2), human/orang (hg18/ponAbe2), human/mouse (hg18/mm9), human/rat (hg18/rn4), and human/zebrafish (hg18/danRer15) from the UCSC genome browser [90]. These alignments are the net of the best human chained alignments for each region of the genome [91]. For quality control, we excluded sites that (1) were missing in either of the species in the alignment, (2) were located within 10Mbp from the starting or ending position of a centromere, (3) were located within 10Mbp from the ending position of a telomere, (4) were located in repetitive elements.
We obtained the coordinate positions for the exons, RefSeq transcripts, and different phastCons measures calculated from different phylogenetic scopes [82] from the UCSC table browser [92] with the following specifications:
GERP scores were downloaded for hg18 from http://mendel.stanford.edu/SidowLab/downloads/gerp/. We used RS scores (range from -11.6 to 5.82) to obtain the conserved sites to remove. S8 Table summarizes the cutoffs we used.
For each window, we computed the recombination rate using the high resolution pedigree-based genetic map assembled by deCODE [50]. The B-value for each window was obtained from McVicker et al. [15]. Four-fold divergence was calculated by counting the number of between species differences that overlapped four-fold sites, divided by the total number of four-fold sites within each window. Functional annotation was done following Lohmueller et al. [13]. Briefly, we translated the Consensus Coding Sequence (CCDS) genes from the UCSC Genome Browser into proteins and determined which nucleotide changes did not alter the encoded amino acid. If transcripts overlapped, we retained the longest one.
To calculate the divergence between each pair of species, we divided the human genome into 100kb non-overlapping windows. For each window, we computed the total number of sites that passed the filtering criteria which resulted in the total number of neutral sites in each 100kb window. To reduce variation, we only considered windows in which the total number of eligible sites was greater than 10,000 (for analyses using 50kb as window size, we only considered windows in which the total number of eligible sites was greater than 5,000). Then we computed the divergence by tabulating the number of sites that are different between the two species being compared. To account for multiple mutational hits for the distantly related species pairs (human-mouse and human-rat), we applied the Kimura two-parameter model [93].
To compute Spearman's ρ, we used the cor function in R. We used the pcor function to calculate partial correlation [94].
To filter out possible hypermuteable CpG sites, we excluded sites that were preceded by a C or were followed by a G in hg18 [15]. To control for the effects of biased gene conversion, we removed all AT→GC substitutions across the genome.
We modeled background selection as a simple reduction in effective population size in the ancestral population [4,5,7,15,58–62]. This was done by scaling the ancestral population size Na, by the B-values. We used the B-values from McVicker et al. [15]. Each simulation replicate consisted of two parts. The first part modeled genetic variation in the ancestral population, and included the effects of background selection. For each window i, we simulated an ancestral recombination graph (ARG) with a population-scaled recombination rate 4NaBiri, where Na is the ancestral population size, Bi is the strength of background selection affecting window i, and ri is the recombination rate for window i. Mutations were added to the genealogy assuming a population-scaled mutation rate θ = 4NaBiμa,iLi, where μa,i is the ancestral per-base pair mutation rate for window i and Li is the number of successfully aligned neutral bases in window i. Simulations were done using the program ms [63]. Note, we included recombination in the ancestral population because it affects the variance in coalescent times across windows and this variance in coalescent times will in turn affect the variance in levels of divergence, which will ultimately affect the strength of the correlation between divergence and recombination. Thus, we aimed to capture this variance as accurately as possible. This part of the simulation generated the amount of divergence due to ancestral polymorphism, which we call da.
We then added the mutations that arose since (i.e. more recently than) the split. The divergence from the present time to the split time follows a Poisson distribution, where the rate parameter equals the expected divergence between two populations. For each window of the genome, ds was simulated using the rpois function in R. Finally, the total divergence within a window is the sum of divergence generated in the ancestral population (da) and the divergence generated since the two species split (ds).
For human chimp divergence (Fig 4A, S9A Fig), ds = 2tsplitμL where ds is the expected divergence from the present time to the split time in the divergence model, tsplit is the split time, μ is the mutation rate, and L is the length of each sequence. When computing both da and ds for human-chimp divergence in Fig 4A, we drew μ from a gamma distribution with shape = 16.82 and scale 1.7 X 10−10 (S6 Table). In S9A Fig, we drew μ from a gamma distribution with shape = 15.68 and scale 1.8 X 10−10 (S6 Table). These parameters were chosen to match the observed mean and standard deviation of the distribution of human-chimp divergence (after removing all AT to GC differences as such changes could be due to biased gene conversion) as well as the observed correlation coefficient between divergence and recombination rate (S3A, S3B, S8A and S8B Figs). The split times and ancestral population sizes are roughly comparable to previous estimates from genetic data [84,85,95,96].
Due to the differences in generation times and mutation rates between the human and mouse lineages, we modified our approach for these simulations (Fig 4B, S9B and S10 Figs). First, here ds = (tmouse μmouse+ thumanμhuman)L, where tmouse is the number of generations on the lineage leading to the mouse from tsplit till the present day, thuman is the number of generations on the lineage leading to human experienced from tsplit till the present day, μmouse is the mutation rate along the mouse lineage, and μhuman is the mutation rate along the human lineage. There is much uncertainty surrounding these parameters. However, the following values are broadly consistent with what has been reported previously and match the observed mean and standard deviation of human-mouse divergence (S3C, S3D, S8C and S8D Figs, S7 Table). First, we assumed tsplit = 75 million years ago. We then assumed mice have 1 generation per year, giving tmouse = 75 x 106 generations. We assumed humans have 25 years per generation, making thuman = 3 x 106. We then set μmouse = 3.8 x 10−9 per generation and μhuman = 3.75 x 10−8 per generation (S10 Fig). These estimates are broadly consistent with previous reports and allow for approximately twice as much divergence on the mouse lineage as compared to the human lineage [41].
For the simulations in Fig 4B, we assumed that μa was equal to 2 x 10−8 per generation, which is the average of μhuman and μmouse. We accounted for variation in mutation rates across different regions of the genome by drawing μa from a gamma distribution [97]. We kept the ratio of μa to μmouse constant across all windows of the genome. For example, μa / μmouse = 5.26. Then if μa,i is the rate for the ith region drawn from the gamma distribution, we set μmouse,i equal to μa,i / 5.26. A similar procedure was used to find μhuman,i. Note that for the simulations in S9B Fig, we used the average mutation rate of 2.7 X 10−8, but we kept the ratio of μa to μmouse and the ratio of μa to μmouse to be the same as the simulations in Fig 4B. Increasing the variance in the mutation rate across regions increased the variance in divergence across windows of the genome and decreased the correlation between divergence and the B-values. We then examined different values of Na and parameters of the gamma distribution that matched the observed mean and standard deviation of the distribution of human-mouse divergence as well as the observed correlation coefficient between divergence and B-values. The ancestral population size, shape, and scale parameters of the gamma distribution used for the simulations in Fig 4B and S9B Fig are reported in S6 Table. The simulated human-mouse divergence using these parameters matched closely with the empirical human-mouse divergence (S3C, S3D, S8C and S8D Figs, S7 Table).
|
10.1371/journal.pntd.0000276 | State–Space Forecasting of Schistosoma haematobium Time-Series in Niono, Mali | Much of the developing world, particularly sub-Saharan Africa, exhibits high levels of morbidity and mortality associated with infectious diseases. The incidence of Schistosoma sp.—which are neglected tropical diseases exposing and infecting more than 500 and 200 million individuals in 77 countries, respectively—is rising because of 1) numerous irrigation and hydro-electric projects, 2) steady shifts from nomadic to sedentary existence, and 3) ineffective control programs. Notwithstanding the colossal scope of these parasitic infections, less than 0.5% of Schistosoma sp. investigations have attempted to predict their spatial and or temporal distributions. Undoubtedly, public health programs in developing countries could benefit from parsimonious forecasting and early warning systems to enhance management of these parasitic diseases.
In this longitudinal retrospective (01/1996–06/2004) investigation, the Schistosoma haematobium time-series for the district of Niono, Mali, was fitted with general-purpose exponential smoothing methods to generate contemporaneous on-line forecasts. These methods, which are encapsulated within a state–space framework, accommodate seasonal and inter-annual time-series fluctuations. Mean absolute percentage error values were circa 25% for 1- to 5-month horizon forecasts.
The exponential smoothing state–space framework employed herein produced reasonably accurate forecasts for this time-series, which reflects the incidence of S. haematobium–induced terminal hematuria. It obliquely captured prior non-linear interactions between disease dynamics and exogenous covariates (e.g., climate, irrigation, and public health interventions), thus obviating the need for more complex forecasting methods in the district of Niono, Mali. Therefore, this framework could assist with managing and assessing S. haematobium transmission and intervention impact, respectively, in this district and potentially elsewhere in the Sahel.
| Adequate forecasting and early warning systems are based upon observations of human behavior, population, disease time-series, climate, environment, and/or a combination thereof, whichever option best compromises among realism, feasibility, robustness, and parsimony. Fully automatic and user-friendly state–space forecasting frameworks, incorporating myriad options (e.g., expert opinion, univariate, multivariate, and spatial-temporal), could considerably enhance disease control and hazard mitigation efforts in regions where vulnerability to neglected tropical diseases is pervasive and statistical expertise is scarce. The operational simplicity, generality, and flexibility of state–space frameworks, encapsulating multiple methods, could conveniently allow for 1) unsupervised model selection without disease-specific methodological tailoring, 2) on-line adaptation to disease time-series fluctuations, and 3) automatic switches between distinct forecasting methods as new time-series perturbations dictate. In this investigation, a univariate state–space framework with the aforementioned properties was successfully applied to the Schistosoma haematobium time-series for the district of Niono, Mali, to automatically generate contemporaneous on-line forecasts and hence, providing a basis for local re-organization and strengthening public health programs in this and potentially other Sahelian districts.
| Prevalent parasitic infectious diseases frequently evade the public health radar because infected individuals present with a clinical history that is characterized by a highly heterogeneous symptomatology. Schistosoma sp., also known as bilharzias, expose and infect more than 500 and 200 million individuals in 77 countries, respectively [1],[2]; however, only those with severe symptoms seek available treatment. Though sub-clinical Schistosoma sp. infection detrimentally impacts the health of infected individuals, the enormous impact of seemingly asymptomatic and mildly symptomatic infection remains difficult to quantify. Furthermore, Schistosoma sp. incidence continues to rise because of 1) numerous irrigation and hydro-electric projects, 2) steady shifts from nomadic to sedentary existence, and 3) ineffective control programs unable to cope with population growth. With the mounting evidence that Schistosoma sp. impose an enormous burden on, as well as their control have paramount importance to improve public health in, developing countries, intervention programs therein could benefit from parsimonious forecasting and early warning systems to enhance management and hazard mitigation of these parasitic infections [1]–[8].
Most individuals at risk of Schistosoma sp. infection reside between latitudes 36° N and 34° S where average fresh water temperatures range from 25 to 30° C [1], placing African states among the most affected countries. Schistosoma mansoni and Schistosoma haematobium account for most Schistosoma sp. infection in Africa [1],2. S. mansoni and S. haematobium cercarias enter the human circulation trans-cutaneously. Subsequently, adult forms mate, migrate, and lay eggs, which eventually lodge in the intestine (S. mansoni) or bladder (S. haematobium). Excreted eggs hatch under favorable aquatic conditions, releasing miracidia, which penetrate the intermediate snail host—in Africa, S. mansoni and S. haematobium infect Biomphalaria sp. and Bulinus sp. snails [8],[9]. Finally, mature cercarias emerge from their intermediate host to seek human reservoirs thus, perpetuating their life cycle [1]. Individuals infected with S. mansoni are usually asymptomatic or mildly symptomatic (rash, fever, aching, cough, diarrhea, and or gland enlargement). In serious infection, lodged S. mansoni eggs trigger a granulomatous immune response that may cause colonic obstruction, hemorrhages, portal hypertension, ascites, and life-threatening esophageal varicose. S. haematobium produces similar unspecific symptoms whereas its fully symptomatic form manifests primarily as terminal hematuria.
Moreau et al. [10] reported the pervasive endemicity of S. haematobium in West Africa, particularly in the Sahel (Figure 1)—i.e. the sub-Saharan region that spans the entire east-west African axis, bordering the Sahara desert to the north and the Savanna to the south [11]. Conversely, his collaboration demonstrated that the prevalence of S. mansoni is greater in Sudanese and Guinean savannas [10]. Along this line of investigation, several epidemiological studies have evaluated the Schistosoma sp. prevalence in Mali [12]–[15], which ranks among the poorest countries in the world, and which is transected by savannas, the Sahel, and the Sahara desert.
Traore et al. [12] reported a 55% overall S. haematobium prevalence, with a case distribution orbiting the 7–14 age-category, in the district of Niono (Segou Region) and Dogon Plateau, Mali; circa 50 and 30% of infected individuals presented with clinical symptoms and pathologic lesions, respectively. The surveys conducted by Keita et al. [13] demonstrated that the Schistosoma sp. prevalence (7–14 age-category) in the community health center (CSCOM) service area of Molodo, in the district of Niono, was 72, 68, and 51% for S. haematobium, S. mansoni, and co-infection, respectively. Finally, Medina et al. [11] reported that S. haematobium is the 5th most frequently diagnosed infectious disease, accounting for 2.5% of total CSCOM consultations in the district of Niono. The high prevalence of Schistosoma sp. in this district may be attributed to an extensive irrigation system that supports predominantly rice monoculture. Unfortunately, district communities not only ingest water from the irrigation scheme but also wash their belongings, bathe, excrete, and amuse themselves in the canals (Figure 2), considerably increasing exposure to Schistosoma sp. infection.
Notwithstanding the colossal scope of these parasitic infections in developing countries, only circa 0.5% of Schistosoma sp. investigations have attempted to predict their spatial and or temporal transmission distributions e.g. [1], [2], [16]–[18]—meriting special attention, Yang et al. [18] modeled both the spatial and temporal S. japonicum transmission dimensions in Jiangsu province, China. [The number of reports investigating Schistosoma sp. spatial and or temporal distributions roughly obtain via keyword-searching “schistosomiasis”, “Schistosoma”, “bilharzias”, “forecast”, “forecasting”, “prediction”, and keyword combinations at www.pubmed.com (09/25/2007). A meta-analysis is beyond the scope of this manuscript.] Regrettably, S. haematobium time-series (TS) forecasts are practically inexistent for Sahelian locations, such as Mali, where this neglected tropical disease tremendously deteriorate public health. Thus, the quest for robust S. haematobium TS forecasting methods to assist with preventing transmission, rapidly treating patients, as well as monitoring intervention impact must not be ignored.
In this longitudinal retrospective (01/1996–06/2004) investigation, the S. haematobium consultation rate TS for the district of Niono, Mali (Fig. 1), was fitted with general-purpose exponential smoothing (ES) methods—encapsulated within a state-space framework—to produce contemporaneous on-line forecasts. On-line forecasts imply that historical records are continuously supplied to the execution program, which automatically revises external predictions. Although this state-space framework ignores direct effects from climate, public health intervention, and irrigation on S. haematobium TS evolution, it accommodates seasonal as well as inter-annual TS fluctuations. The ES methods within this framework may capture prior non-linear interactions between disease dynamics and the aforementioned covariates, potentially obviating the need for more complex predictive approaches in the district of Niono, Mali. [An intuitive overview of this ES state-space framework is conveyed by Figure 3.] Therefore, not only does this analysis address the paucity of reported S. haematobium TS investigations but it also demonstrates that this state-space framework could assist with managing S. haematobium infection in this district and possibly elsewhere in the Sahel.
This longitudinal retrospective (01/1996–06/2004) S. haematobium TS investigation was conducted in the district of Niono, Mali (Fig. 1). Panel A in Fig. 1 is a satellite image that portrays Mali, with a projected population of 12 million in 2004 [19], along with its neighboring West African countries. Panel B—which corresponds approximately to an enlargement of the red demarcation in panel A—depicts the district of Niono (red rectangle), 330 km northwest of Bamako, 100 km north of the Niger River, in the Segou region. This district is a model location to test S. haematobium TS forecasting and early warning systems feasibility because its extensive irrigation network pervasively exposes communities to this neglected parasitic infection. Furthermore, the district of Niono shares epidemiological similarities with other regions in the Sahel where poverty- and disease-induced morbidity and mortality are rampant.
The review of monthly clinical consultation records from the district of Niono, Mali, is part of a larger study on climate and health (“Putting climate in the service of public health”) that was approved by the “Columbia University Medical Center Institutional Review Board” (New York, U.S.A.) and the “Ethics Committee of the Mali National Medical School” (Bamako, Mali). Patient privacy was protected from inadvertent (or deliberate) violations because consultation records reflect monthly summaries that lack information with which individuals may be identified [11]. The assembled monthly data set (01/1996–06/2004) comprises consultation records for 20 diseases, which were tabulated by gender and age categories, from 17 CSCOM service areas within the district of Niono [11],[19],[20]. However, only the S. haematobium TS was analyzed here—diarrhea, acute respiratory infection of the lower tract (ARI), and malaria TS forecasts, as well as preliminary frequency description of all 20 diseases, have already been reported [11]. Of note, Schistosoma sp. consultation records reported by Medina et al. [11] and analyzed herein reflect cases of S. haematobium–induced terminal hematuria in over 99% of consultations, as discussed later, for which a single dose of 40 mg/kg of prazinquatel was prescribed in most cases.
Monthly S. haematobium consultation records for the 17 CSCOM service areas, both genders, and all ages were amalgamated. Rather than interpolating missing observations with imputed CSCOM-specific monthly median values and excluding ineligible CSCOM service area TS [11], this amalgamated consultation rate TS, {yt}, was estimated by simultaneous adjustment of time-dependent nominator (cases) and denominator (population) observations, according to Equation 1(1)where Ct,η is the monthly number of CSCOM-specific S. haematobium-induced terminal hematuria consultations for both genders and all ages; Pt,η is the time-dependent population of each CSCOM service area, which was adjusted for a national annual population growth rate of 3.2% [11],[19],[20]; αt,η = 0 if Ct,η is missing for month t and CSCOM service area η, otherwise αt,η = 1; last, N = 17 is the total number of CSCOM service areas. The approximately random distribution of missing observations (∼17%) across months, years, and CSCOM service areas [11] ensures approximately unbiased {yt} estimation with Eq. 1, which holds as long as the denominator summation is positive. In simpler words, Eq. 1 estimates the monthly consultation rate for S. haematobium-induced terminal hematuria from CSCOM service areas for which records are available. Consultation rates and their forecasts are expressed as the monthly number of newly diagnosed S. haematobium–induced terminal hematuria cases per 1000 individuals in the district of Niono. Additional record details appear in Table 1, which was adapted from Medina et al. [11].
The amalgamated TS was fitted with ES methods, which are encapsulated within a state-space framework hereafter referred to as ETS for error (E), trend (T), and seasonal (S) components. The E component is either additive (A) or multiplicative (M); T and S components may be A, M, or inexistent (N); last, T may also be dampened additively (Ad) or multiplicatively (Md). For example, the ETS method MMN has E(M), T(M), and S(N) structures. Therefore, there are 30 possible ES combinations within this forecasting framework, comprising linear and non-linear ones. However, only the 15 ES methods with multiplicative error structures (heteroskedastic) were herein considered for TS analysis [21]–[32]. Not only do multiplicative error structures are conservative but they also yield more realistic 95% prediction interval (PI) values. Furthermore, a reduction in the number of ES methods evaluated also diminishes the expensive computational time.
The versatile and fully automatic ETS framework requires neither stationarity nor “strict” linearity to produce contemporaneous TS forecasts for variable time horizons (h) [21]–[32]. Consequently, it is extensively employed in, e.g., econometrics and inventory control where automatic forecasts are required for a large number of diverse TS. This forecasting framework, whose performance was recently and favorably compared to those of several forecasting techniques across thousands of TS [32], adapts to underlying alterations in disease dynamics and automatically revises forecasts on-line as new observations accumulate (Fig. 3). This adaptability is essential for epidemiological forecasting methods because interventions (e.g. medical and prophylactic treatment) almost ubiquitously perturb disease TS dynamics. An intuitive description of the ETS framework appears in Fig. 3; it is only succinctly described below because it has been meticulously derived elsewhere [21]–[31].
In the ETS framework, the expected mean of a forecasted observation, E[F(yt+h|It)], is conditioned on the information set (It) available at time t—i.e. these are external predictions. The information set It contains unobserved level (lt), trend (rt), and or seasonal (sm|t: month, m = [1, 12]) components, whichever pertinent, depending on the underlying ES method. Possible lower-frequency “harmonics”, i.e. inter-annual fluctuations, are handled by lt and rt components in the ETS framework because the limited temporal window (01/1996–06/2004) considered in this investigation precludes stable estimation of periodicity much longer than 12 months. The observed amalgamated TS is symbolized by {yt}, as previously defined, whereas unobserved TS components enter the vector xt, according to the general state and transition Equations 2 & 3, respectively:(2)(3)where,
For ES methods with multiplicative error structures, w(xt-1) and r(xt-1) have both the form of the expected mean of a forecasted observation, E[F(yt|It-1)]. Otherwise, w(xt-1) = E[F(yt|It-1)] and r(xt-1) = 1 for ES methods with additive error structures (not discussed hereafter). All ES methods rely on the adjustment of lt, rt, and or sm|t TS components with their corresponding smoothing control α, β, and γ values; furthermore, φ controls smoothing of rt-dampening if present. Basically, contemporaneous forecasts obtain via TS extrapolations whereby previous deviations between forecasts and their corresponding observations are exponentially adjusted with α, β, γ, and or φ. Large smoothing control values confer greater weights to recent information and effectively shorten the smoothing “memory”, i.e. the recent-past has a more pronounced influence on estimated components than does the distant-past [11], [21]–[31]. Three important remarks: 1) a single or multiple smoothing control values may be required depending on which TS components are present in the selected ES method; 2) although smoothing controls are symbolized with the same notation across distinct ES methods, their function may vary from one ES method to another because the relationship between TS components may also differ (e.g. multiplicative vs. additive rt); last, 3) the function of smoothing control values approximately parallels that of the bandwidth in a one-side Nadaraya-Watson exponential kernel.
Smoothing controls plus unobserved components are estimated for all ES methods within the ETS state-space framework using a maximum likelihood function analog [31]. Here, the general ETS constraints are: 0<α, φ<1; 0<β<α; and, 0<γ<1−α; strictly multiplicative error structures; multiplicative sm|t values add annually to 12 because m = [1, 12]; and, 36 months [≡3p] for initial training, the possible specification of longer intervals notwithstanding. Defaulted ETS constraints are specified for several reasons [21]–[31] among them to prevent the forecast execution program from producing unrealistic results.
Once each ES method within the ETS framework is optimized at time t, that which minimizes the Akaike's Information Criterion (AIC) is selected to generate the h-month horizon forecast path distribution. The h-month horizon forecast path distribution, F(yt+h|It), obtains via recursive iterations (Eqs. 2 & 3) of B = 1000 ordinary {εt} bootstrap-generated pseudo-TS [11],[31],[33]. With the accumulation of each new observation, ES methods within the ETS framework are re-optimized and the best-performing ES method is re-selected based on the AIC. Subsequently, F(yt+h|It) is again recursively generated from B = 1000 ordinary {εt} bootstrap-generated pseudo-TS. For example, observations from January 1996 to December 1998 initialize the ETS execution program (Fig. 3) that predicts consultation rates for January 1999 to May 1999, assuming h = [1, 5]. Once the January 1999 forecast plus the available TS (including the most contemporaneous observation of January 1999) become available to the on-line system (Fig. 3), the execution program cycles again and optimizes all considered ES methods, re-selecting the best-performing one (which may or may not be the same one employed prior to the arrival of the new observation). As a result, revised consultation rate predictions for February 1999 to June 1999 ensue. This process repeats ad infinitum (Fig. 3). The 95% PI values for the simulated F(yt+h|It) paths are estimated from distribution percentiles.
Although a full portrayal of the ETS framework (Eqs. 2 & 3) encapsulating the 15 considered ES methods [21]–[31] is beyond the scope of this investigation, those ES methods which have been selected at least once during this TS analysis are described in terms of E[F(yt|It-1)] and xt recursions (Table 2). [Table 2 caption also provides an ES method example explicitly written in matrix notation.] As discussed afterwards in the Results section, none of the selected ES methods (Table 2) is seasonal, reflecting the endemicity of the TS analyzed herein. For further details concerning the ETS framework, refer to, e.g., Hyndman et al. [27],[29],[31].
Standard accuracy and dispersion measures were employed in this analysis. Accuracy—which measures the forecasting competence—is defined here as the mean absolute percentage error (MAPE) between observed and forecasted TS values whilst infrequently reported PI values reflect the dispersion of forecast distributions; the dispersion of simulated F(yt+h|It) probability density functions were also summarized as the average coefficient of variance (). MAPE and values are calculated with Equations 4 & 5, respectively:(4)(5) and MAPE (external) values are expressed in percentage (%) as a function of the h-month horizon forecast. In Eqs. 4 & 5, T is the TS length and f = 3p−1+h reflects the actual time when the h-month horizon forecast begins. Large MAPE and values imply low accuracy and large dispersion, respectively, and vice-versa. The distinction between MAPE and PI (or ) values is an important one. The first assesses the competence, i.e. the skill, of the h-month horizon forecast; the latter only measures the dispersion of the h-month horizon forecast path distribution. Thus, PI (or ) values have paramount importance for calculating, e.g., the probability that a future observation will be smaller or greater than the expected forecast distribution mean by a certain margin. Likewise, the number of individuals at risk may be calculated for a specified probability.
This TS has not undergone Box-Cox transformations. Notice however, that TS frequently undergo such transformations prior to the forecasting analysis. Regardless, contemporaneous forecasts and standard accuracy measures (e.g. MAPE) must be (and were) superimposed onto and computed for, respectively, the originally observed TS because accuracy may be severely distorted in the transformed dimension—i.e. occasionally, forecasts may be simultaneously accurate and inaccurate in the transformed and original dimensions, respectively. All calculations were performed in R: A language and environment for statistical computing [30],[31].
This longitudinal retrospective (01/1996–06/2004) investigation analyzed the S. haematobium consultation rate TS for the district of Niono, Mali. In Figure 4, the observed amalgamated S. haematobium consultation rate TS is symbolized by black lines. The TS is excessively noisy from 1996 to 1999 when a sharp rise in consultation rates clearly ensues. From 2001 onwards, consultation rates decline because of large-scale prophylactic de-parasitation programs. Regardless, 2- to 5-month horizon forecasts clearly captured these inter-annual tendencies (Fig. 4)—red traces correspond to contemporaneous on-line 2-, 3-, 4-, and 5-month horizon forecasts (panels A, B, C, and D, respectively) whilst their 95% PI values are depicted in dots of the same color. Abscissa TS projections span 102 months (01/1996–06/2004) while ordinate scales represent the number of newly diagnosed (or forecasted) S. haematobium–induced terminal hematuria cases per 1000 individuals.
TS observations were continuously submitted to a family of general-purpose ES methods—encapsulated within the ETS state-space framework—to produce contemporaneous on-line forecasts (i.e. external predictions). Predictions were superimposed onto the original TS to allow visual evaluation of prediction accuracy. While this superimposition is absolutely essential, it might convey the false impression that forecasts represent internal predictions—this is not the case. Fig. 4 should be considered dynamically. As observations and forecasts become available to and from the on-line forecast execution program (Fig. 3), respectively, the actual graphing of forecasts (red traces) precede that of observations (black lines) by exactly h-month horizon.
Generally, the ETS framework accommodates seasonal and inter-annual fluctuations, producing reasonably accurate TS forecasts. Here, inter-annual fluctuations dominate the S. haematobium TS while seasonal oscillations are practically inexistent (Fig. 4). These fluctuations are intuited from the observed consultation rate TS (black lines), as well as implied by the absence of {st|m} vis-à-vis the presence of {lt} and or {rt} components in automatically selected ES methods (Table 2). Only 3 ES methods were automatically selected with the AIC during this S. haematobium TS forecasting analysis. These selected ES methods, which have been described in terms of E[F(yt|It-1)] and xt recursions (Table 2), are: the multiplicative error/ trendless/ aseasonal (MNN), multiplicative error/ damped additive trend/ aseasonal (MAdN), and multiplicative error/ damped multiplicative trend/ aseasonal (MMdN) ES methods. None of them are seasonal and hence exogenous forcing (e.g. climate covariates) was not invoked to improve predictions.
Table 3 lists the frequency (n) with which these ES methods were re-selected during the forecasted investigational period plus the method-specific median (and IQR: inter-quartile range) of pertinent smoothing control values. Smoothing control values are time-dependent because they are continuously re-estimated as new observations accumulate. Yet, their magnitude drifts little in this investigation. Hence, they were reported as median and IQR values. The MNN smoothing control α is obviously large since this method only has a level {lt} component, i.e. the MNN ES method lacks {rt} and {st|m} components as well as their corresponding smoothing control β, φ and γ values. For MAdN and MMdN methods, β≤α≪φ due to large dampening of minute rt components. As new observations accumulated, the automatic and criterial re-selection of ES methods conferred an additional layer of flexibility to the ETS framework and consequent TS forecasts. [Smoothing control values may differ in functional form across ES methods despite the retained notation (Methods).]
MAPE and values for 1- to 5-month horizon forecasts were circa 25 and 45%, respectively (Figure 5). values reflect the average dispersion of simulated F(yt+h|It) probability density functions whilst MAPE values measure the mean absolute percentage error between TS observations and their forecasts. Accuracy (MAPE; panel A) becomes approximately asymptotic as the h-month horizon forecast increases beyond 6 months because of a minute {rt} component irrespectively of the selected ES method, significant inter-annual {lt} fluctuations notwithstanding. As expected, dispersion (; panel B) increases as innovations propagate through longer stochastic h-month horizon forecast paths.
Schistosoma sp. expose and infect more than 500 and 200 million individuals in 77 countries, respectively. In the Sahel, S. haematobium is endemic and highly prevalent [2], [10]–[15]. The few reports evaluating S. haematobium transmission in Mali [10]–[15], particularly in the district of Niono (Fig. 1), suggest that forecasting S. haematobium consultation rate TS may locally assist with reducing morbidity. For instance, S. haematobium is the 5th most frequently diagnosed infection (the 6th commonest consultation etiology); it accounts for 2.5% of total CSCOM service area consultations [11],[20] with 50 to 75% community prevalence [12],[13] in the district of Niono. Paradoxically, temporal S. haematobium analyses are scarcely reported in the parasitic literature e.g. [16]–[18] probably because 1) this neglected tropical disease is endemic whereas most infectious disease TS forecasts usually attempt to detect epidemics, i.e. unexpected rises in consultation rate first moments, assisting with tailoring control measures; 2) S. haematobium TS tend to be excessively noisy, hindering analyses; finally, 3) long delays between S. haematobium infection and diagnosis challenge efforts to relate predicted high consultation rates to their potentially preventable sources. Notice that, though endemic, S. haematobium TS does fluctuate.
The ETS framework employed herein reasonably forecasted long horizons (Fig. 4), partially circumventing the limitations imposed by the S. haematobium TS noisy level and long delays between infection and diagnosis. Thus, this report addresses challenges in, and the scarcity of, S. haematobium TS forecasting reports with the flexible ETS framework (Fig. 3), which may locally assist with managing endemic S. haematobium transmission in the district of Niono, Mali. Here, accuracy (i.e. MAPE) and dispersion () for contemporaneous (“out-of-fit”) 1- to 5-month horizon S. haematobium consultation rate TS forecasts were circa 25 and 45%, respectively (Figs. 5). MAPE values assess the competence, i.e. the skill, of h-month horizon forecasts; (or PI) values measure the dispersion of h-month horizon forecast path distributions. The later has paramount importance for calculating, e.g., the probability that a future observation will be smaller or greater than the expected forecast distribution mean by a certain margin. Likewise, the number of individuals at risk may be calculated for a specified probability. The rarely considered 2nd moment forecasts (PI) may significantly assist authorities with risk and scenario analyses.
A comprehensive S. haematobium intervention strategy depends not only on prevalence, which has already been discussed in the Introduction section [10]–[15], but also on incidence measures. For instance, an abnormal rise in incidence should alarm authorities who are charged with investigating and containing hazard, ensuring that CSCOM service areas are able to handle patient demand, sensitize communities, control transmission, and monitor intervention impact. Thus, it is important to delineate some parallels between the S. haematobium consultation rate TS plus their forecasts analyzed herein (Fig. 4) and the unobserved incidence.
The monthly S. haematobium consultation rate is proportional to the unobserved monthly incidence TS—i.e. an increase in the monthly S. haematobium consultation rate most likely stems from a rise in the monthly incidence TS since the former is a fraction of the latter. The observed and forecasted consultation rate TS (Fig. 4) approximately reflect the monthly S. haematobium-induced terminal hematuria incidence because ∼95% of the Niono district population lives within 15 km of CSCOM facilities and hematuria alarmingly prompts patients to access available treatment. Although these records [19],[20] are unsuitable for estimating the exact S. haematobium incidence, it may be approximated to at least an order of magnitude higher than the observed consultation rate TS under mean-field, steady-state, stable age structure, constant population growth (3.2%), and overall prevalence (∼60%) assumptions. Consequently, the difference between the observed consultation rate (Fig. 4) and the estimated incidence TS described above (not shown) approximately reflects the S. haematobium incidence of asymptomatic and mildly symptomatic cases. The effective S. haematobium incidence depends on age as recurrent cercarial exposure induces partial-immunity [1]. S. haematobium-induced terminal hematuria consultations emanate primarily from the 7–14 age-category, which comprises 20 to 30% of the district population [11],[19],[20]. Thus, a rough population structure TS adjustment suggests that the actual and forecasted S. haematobium-induced terminal hematuria incidence is roughly 3 to 5 times higher in the aforementioned age-category.
The dependency of S. haematobium transmission on the environment is extremely important and cannot be understated. S. haematobium transmission depends on climate [1],[18], as well as natural (e.g. lakes) and artificial (e.g. irrigation schemes) water reservoirs [1],[2]. Despite these dependencies, covariates such as climate were not invoked to forecast the S. haematobium TS because it is endemic [10]–[13] and aseasonal (Fig. 4 and Tables 2 & 3) in the district of Niono, Mali. In this district, temperature and rainfall TS values guarantee S. haematobium transmission suitability throughout the year—i.e. transmission is not limited here by climate thresholds beyond which the S. haematobium life-cycle becomes unstable.
Unlike temperature, rainfall TS values exhibit large (inter-tropical convergence zone-mediated) inter-annual oscillations in the Sahel. These fluctuations prompt the local authority (Office du Niger) to accordingly adjust irrigation management, which inevitably and transiently alters S. haematobium transmission suitability in this district. In other words, rainfall precipitation only indirectly affects S. haematobium transmission in this district. For example, an augment in rainfall precipitation increases water availability. Consequently, the Office du Niger may relax water control to better irrigate drier areas while collaterally enhancing water-flow through typically well-served agricultural fields—S. haematobium transmission suitability could then simultaneously increase and decrease in the former and latter scenarios, respectively. Another major source of TS fluctuations stems from the constant exposure to, and behavioral risks associated with, the irrigation system (Fig. 2). These TS fluctuations are further aggravated by the influx of migrant workers from non-endemic areas. The variable clinical course of S. haematobium-induced terminal hematuria may also introduce stochastic fluctuations into this TS. Finally, the impact of large-scale prophylactic de-parasitation programs perturbs S. haematobium transmission as evidenced by sustained consultation rate declines from 2001 onwards (Fig. 4). Consequently, S. haematobium TS fluctuations in this district require forecasts, the endemecity of this neglected tropical disease notwithstanding.
Future studies should dedicatedly investigate the intricate roles of geography, climate, irrigation management, and human behavior (including migration) in the context of S. haematobium transmission ecology to improve forecasts and interventions in this district. Unfortunately, addressing the multidimensionality of this disease remains difficult owing to poor documentation. Until this information becomes available, the employment of univariate methods (e.g. ETS framework) to forecast S. haematobium-induced terminal hematuria incidence in the district of Niono seems adequate. This is consistent, for example, with the successful employment of univariate methods to forecast schistosomiasis TS in Dongting Lake, China [16], albeit with the admonition that these results cannot be indiscriminately generalized to any location.
Furthermore, this S. haematobium-induced terminal hematuria TS is aseasonal (Fig. 4 and Tables 2 & 3), which intuitively argues against the incorporation of seasonal climate covariates and corroborates the employment of univariate prediction methods. [The automatically selected MNN, MAdN, and MMdN forecasting methods (Tables 2 & 3) are very similar; they reflect the fact that the S. haematobium-induced terminal hematuria TS is aseasonal, quasi-trendless, with significant inter-annual fluctuations in the district of Niono, Mali.] S. haematobium transmission generally drifts slowly in response to also slow climate and or other environmental changes. As a result, the ETS framework has the benefit of time to adapt to perturbations in and revise forecasts for this fully-stable (endemic) S. haematobium TS. In other words, current observations mirror past disease dynamics and environmental interactions. Forecasting methods that capture these relationships through historical TS analysis often reflect prior and present interactions on post-sample (external) predictions. This is clearly not the case when the chaotic weather or a newly erected dam, for example, suddenly inundate large areas triggering outbreaks and epidemics (i.e. under unstable transmission conditions). While it is difficult to predict weather, environmental impact may be investigated with satellite technology; for example, Beck-Wörner et al. [34] successfully employed a hybrid of remotely-sensed and surveyed data from western Côte d'Ivoire to spatially predict S. mansoni infection risk. [Although the consultation records analyzed herein are resolved by 17 CSCOM service areas, spatial considerations were dismissed because the district of Niono occupies only ∼20 000 km2 (Fig. 1).]
Conversely, lagged weather- and or climate-based models are particularly powerful whenever disease transmission is unstable and epidemics are suddenly-triggered. For example, a weather-based Poisson regression (4th-degree polynomial distributed lag) was employed to model malaria TS in highly unstable regions of Ethiopia [35]. However, lagged weather- and or climate-based models not only demand extensive programming and expertise to reasonably specify the number of lags but they also require caution to avoid multicollinearity, problematic optimization, and lengthy TS requirements. Furthermore, lagged models, unlike ES methods, must be tailored to each disease because the optimum functional form of climate covariates is not obvious [35]–[38]. Weather events must be measured because predicting its chaotic nature with several weeks in advance is usually impossible. Predicting climate is not trivial and such predictions are typically too global to substantially add local forecasting accuracy. Otherwise, weather- and or climate-based models are crucial to: elucidate complex disease transmission behavior [37], forecast long horizons [38], and model infectious disease transmission in the spatial dimension [18],[36]. If the optimum functional form of climate covariates is unveiled [37] then reasonable forecasts yield [38]. While some form of lagged weather- and or climate-based model may be indispensable in certain cases [35]–[38], simpler ES alternatives may locally forecast fully- and or partially-stable disease TS, e.g. meso-endemic malaria [11] and endemic S. haematobium transmission in the district of Niono, Mali.
Like other forecasting approaches, ES methods perform reasonably well whenever disease transmission comprises relatively large event-probabilities during long investigational periods. Forecasting methods surrender when disease transmission depends on rare stochastic events (in highly-structured populations), each associated with minute (albeit finite) probabilities, governing unstable and transient disease dynamics. These highly-stochastic structured disease dynamics feature sudden epidemic resurgence and ample epidemic volume variability that are not easily investigated with univariate and most multivariate methods, often requiring more sophisticated approaches e.g. [39],[40].
The generality, reasonable performance, and operational simplicity of the ETS forecasting framework employed herein may appeal to those working towards infectious disease hazard mitigation. Computationally, recursive ES methods (Table 2), encapsulated within this framework, may be easily and automatically optimized, as well as operated, by non-statisticians in the public health sector [21]–[32]. They are often available as software procedures (e.g. SPSS and EViews), pre-written functions for programming environments (e.g. S-plus and the freely-available R language and environment for statistical computing), and scripts in classical programming languages (e.g. FORTRAN and C). This has been previously discussed in Medina et al. [11].
Moreover, ES methods adapt with an on-line training (Fig. 3) that exponentially discounts prior information, i.e. information from the recent-past is more relevant to forecasts than those from the distant-past. Its versatility reflects “density-estimation” of unobserved TS components (Methods) [41]. Owing to both adaptability and versatility, ES methods tend to accommodate intervention-induced perturbations (e.g., medical and prophylactic treatment) that inherently plague longitudinal retrospective disease TS investigations (Fig. 4) e.g. [3]–[7] as well as disease TS with distinct transmission modes [11], respectively. This is illustrated here by the S. haematobium monthly consultation rate declines from 2001 onwards (owing to large-scale prophylactic de-parasitation programs) and corresponding 2- to 5-month horizon TS forecasts, which captured these inter-annual tendencies (Fig. 4).
Forecasting and early warning systems for managing infectious diseases depend on human behavior, population, disease TS, climate, environment and or a combination thereof, whichever alternative best compromises among realism, feasibility, robustness, and parsimony. Nevertheless, forecasts do not obligatorily require exogenous covariates. Medina et al. [11] demonstrated how a robust univariate general-purpose ES method may produce contemporaneous (“out-of-fit”) forecasts for dissimilar diseases without disease-specific tailoring of the forecasting method. More recently, Chaves & Pascual demonstrated the importance of assessing the performance of several forecasting methods, including climate-based ones, in a systematic fashion [38]. Finally, the aforementioned ideas were successfully combined here to allow AIC-directed switches (as new TS observations accumulated and perturbations evolved) among 15 general-purpose ES methods within the ETS framework, further improving forecasts (Fig. 4 & 5 and Tables 2 & 3).
Sudden TS perturbations transiently limit the performance of this and other forecasting approaches. Like most forecasting approaches, particularly univariate ones, ES methods react only after initial TS fluctuations ensue. Thus, this limitation is not unique to ES methods employed herein. Introducing covariates may lessen this limitation if, and only if, the underlying covariate fluctuation is either measurable or predictable—this is often, but not always, the case. Furthermore, the deleterious effects of sudden, even if small, TS perturbations propagate through h-month horizon forecast paths. This phenomenon clearly surfaced in Fig. 4 (panels A, B, C, and D). As the horizon increased from 2- to 5-month, forecasts became progressively worse (Fig. 5) for sudden consultation rate TS fluctuations in 2001 (Fig. 4) as previously discussed.
A major limitation of all TS analyses, and this investigation is not exempt from it, consists of information unavailability. The intricate role of geography, rainfall, irrigation management, and human behavior (including migration) in the S. haematobium transmission ecology has not been extensively documented for this district. Thus, general, adaptable, and versatile univariate ES methods were employed herein to generate forecasts. Second, missing monthly consultation records could have potentially introduced bias in this monthly S. haematobium consultation rate TS. However, this is unlikely owing to the random distribution of missing records across CSCOM service areas, months, and years. As listed in Table 1, missing records distribute approximately normally across CSCOM service areas and approximately uniformly through the investigational period [11]. The percentage of missing monthly records in the amalgamated TS is circa 17%, generally less than 2% per year. The only exception manifests in the practically reconstructed year of 1997 that was employed for program initialization—nevertheless, this is minimally consequential because program initialization would otherwise reflect the customary (and arbitrary) “opinion of an expert” [11].
Changes in multiple dimensions (e.g. human behavior, population, disease TS, climate, and environment) will confer an ever-increasing role to infectious diseases forecasting and early warning systems. These predictive systems are based upon a single dimension or a combination thereof, whichever alternative best compromises among realism, feasibility, robustness, and parsimony. With the mounting evidence that S. haematobium—a neglected tropical disease—imposes an enormous burden on developing countries, public health programs therein could benefit from parsimonious forecasting and early warning systems to enhance management and control of this parasitic infection. Not only does this report address the paucity of S. haematobium TS forecasting investigations but it also advocates the usage of parsimonious state-space frameworks to forecast neglected tropical diseases. The ETS state-space forecasting framework employed herein generated reasonable 1- to 5-month horizon S. haematobium TS forecasts, obliquely capturing prior non-linear interactions between disease dynamics and exogenous covariates (e.g. climate) and hence, obviating the need for more complex predictive methods in the district of Niono, Mali. Thus, this and other e.g. [11], [21]–[32] results suggest that the remarkable performance of state-space forecasting methods since the 1960s may be capitalized by the public health sector, providing a basis for local re-organization and strengthening of intervention programs in this and potentially other Sahelian districts. The operational simplicity, generality, and flexibility of state-space frameworks, such as the one employed here, conveniently allow for: 1) unsupervised model selection without disease-specific methodological tailoring; 2) on-line adaptation to fluctuations in partially- and fully-stable disease TS; and, 3) automatic switches between distinct forecasting methods as new TS perturbations dictate. Generally, state-space approaches are malleable to the dynamic incorporation of covariates (e.g. climate), expert opinion, and even a spatial dimension as needed. Therefore, fully automatic and user-friendly state-space forecasting frameworks, incorporating myriad (e.g. univariate, multivariate, and spatial-temporal) options, could considerably enhance disease control and hazard mitigation in regions where vulnerability to neglected tropical diseases is pervasive and statistical expertise is scarce.
|
10.1371/journal.ppat.1003692 | Epigenetic Dominance of Prion Conformers | Although they share certain biological properties with nucleic acid based infectious agents, prions, the causative agents of invariably fatal, transmissible neurodegenerative disorders such as bovine spongiform encephalopathy, sheep scrapie, and human Creutzfeldt Jakob disease, propagate by conformational templating of host encoded proteins. Once thought to be unique to these diseases, this mechanism is now recognized as a ubiquitous means of information transfer in biological systems, including other protein misfolding disorders such as those causing Alzheimer's and Parkinson's diseases. To address the poorly understood mechanism by which host prion protein (PrP) primary structures interact with distinct prion conformations to influence pathogenesis, we produced transgenic (Tg) mice expressing different sheep scrapie susceptibility alleles, varying only at a single amino acid at PrP residue 136. Tg mice expressing ovine PrP with alanine (A) at (OvPrP-A136) infected with SSBP/1 scrapie prions propagated a relatively stable (S) prion conformation, which accumulated as punctate aggregates in the brain, and produced prolonged incubation times. In contrast, Tg mice expressing OvPrP with valine (V) at 136 (OvPrP-V136) infected with the same prions developed disease rapidly, and the converted prion was comprised of an unstable (U), diffusely distributed conformer. Infected Tg mice co-expressing both alleles manifested properties consistent with the U conformer, suggesting a dominant effect resulting from exclusive conversion of OvPrP-V136 but not OvPrP-A136. Surprisingly, however, studies with monoclonal antibody (mAb) PRC5, which discriminates OvPrP-A136 from OvPrP-V136, revealed substantial conversion of OvPrP-A136. Moreover, the resulting OvPrP-A136 prion acquired the characteristics of the U conformer. These results, substantiated by in vitro analyses, indicated that co-expression of OvPrP-V136 altered the conversion potential of OvPrP-A136 from the S to the otherwise unfavorable U conformer. This epigenetic mechanism thus expands the range of selectable conformations that can be adopted by PrP, and therefore the variety of options for strain propagation.
| Prions are infectious proteins, originally discovered as the cause of a group of transmissible, fatal mammalian neurodegenerative diseases. Propagation results from conversion of the host-encoded cellular form of the prion protein to a self-propagating disease-associated conformation. It is believed that the self-propagating pathogenic form exists in a variety of subtly different conformations that encipher prion strain information. Here we explored the mechanism by which prion protein primary structural variants, differing at only a single amino acid residue, interact with prion strain conformations to control disease phenotype. We show that under conditions of co-expression, a susceptible prion protein variant influences the ability of an otherwise resistant variant to propagate an otherwise unfavorable prion strain. While this phenomenon is analogous to the expression of genetically-determined phenotypes, our results support a mechanism whereby dominant and recessive prion traits are epigenetically controlled by means of protein-mediated conformational templating.
| Prion-mediated phenotypes and diseases result from the conformationally protean characteristics of particular amyloidogenic proteins. The prion state has the property of interacting with proteins in their non-prion conformation, thus inducing further prion conversion. The prion phenomenon has been described for a variety of different proteins involved in diverse biological processes ranging from translation termination in yeast, memory in Aplysia, antiviral innate immune responses [1], and most recently the action of the p53 tumor suppressor [2]. Since the prion and non-prion conformations have differing biological properties, the net result of this replicative process is protein-mediated information transfer, the characteristics of which vary from prion to prion. The ubiquity of prion replication indicates that this is a wide-ranging of means of information transfer in biological systems.
In the case of mammalian neurodegenerative diseases the prion state is pathogenic as well as transmissible. A hallmark of such conditions is the inexorable progression of pathology between synaptically connected regions of the central nervous system (CNS), consistent with advancing cell-to-cell prion spread. Experimental transmission in several settings has been convincingly demonstrated in the case of the amyloid beta (Aβ) peptide which features prominently in Alzheimer's disease (AD), the intracytoplasmic protein tau, also involved in AD as well as various neurodegenerative diseases referred to as taopathies, and α-synuclein, the primary constituent of Lewy bodies found in Parkinson's disease (PD) [1], [3].
The prototypic and best-characterized prion diseases are the transmissible spongiform encephalopathies (TSEs) of animals and humans, including sheep scrapie, bovine spongiform encephalopathy (BSE), chronic wasting disease (CWD) of cervids, and human Creutzfeldt-Jakob disease (CJD). TSEs result from conformational conversion of the host-encoded cellular form of the prion protein, PrPC, to the corresponding prion, or scrapie form, PrPSc. Since TSEs share numerous properties with nucleic acid-based pathogens, including agent host-range, stable strain properties, and the ability to mutate and respond to selective pressure, early researchers assumed a viral etiology for these diseases. While this is not the case, the unequivocal infectivity of TSEs set these prions apart. Their singular capacity to cause fatal neurodegeneration in genetically tractable animal models, and the ability to propagate and quantify infectivity, in vivo, in cell culture or cell-free conditions, provide unparalleled settings to elucidate general mechanisms and devise integrated therapeutic approaches for all diseases involving conformational templating [4].
TSEs have long incubation periods ranging from months to years, are invariably fatal, and currently incurable. While a variant of CJD (vCJD) is unequivocally linked to prions causing BSE [5], the zoonotic potential of other TSE's remains uncertain. Whereas all TSEs, including human genetic and sporadic forms, are experimentally transmissible, most are naturally infectious and frequently occur as unanticipated epidemics. Scrapie is one such example, and several iatrogenic epidemics have been reported. More than 1,500 sheep developed scrapie following administration of a scrapie-contaminated vaccine [6]. A similar recent event led to an ∼20-fold increase in the rate of scrapie in Italy [7].
Prion strain properties and the primary structure of PrP are the two major elements controlling prion transmission. Optimal disease progression appears to occur when the primary structures of PrPSc constituting the infectious prion, and substrate PrPC expressed in the host are closely related [8]–[10]. Underscoring the importance of primary structure on transmission, susceptibility and disease presentation are strongly influenced by several PRNP polymorphisms in humans and animals. For example, a strong association between susceptibility/resistance to natural scrapie is associated with the valine (V)/alanine (A) dimorphism at PrP residue 136 [11]. Prion strains are classically defined by differences in incubation times, and the neuropathological profiles they induce in the CNS. Seminal studies of mink prions [12], as well as studies of human prions in Tg mice [13] indicated that strain information is enciphered within the tertiary structure of PrPSc. While this remains the favored explanation for prion strain diversity, the mechanism by which primary and higher order PrPC and PrPSc structures interact to influence pathogenesis are not understood.
Our previous studies demonstrated that A at ovine PrP residue 136 is a component of the monoclonal antibody (mAb) PRC5 epitope [14]. This property allowed us to use PRC5 in this study to distinguish OvPrP-A136 from OvPrP-V136, affording the opportunity to monitor allele-specific OvPrP conversion during prion infection. To accomplish this, we engineered Tg mice expressing either OvPrP-A136 or OvPrP-V136, as well as Tg mice expressing both alleles in the same neuronal populations. Here, using a combination of in vivo and in vitro approaches, we address the mechanism by which this important disease susceptibility dimorphism influences scrapie strain-specific pathogenesis.
We created Tg mice expressing OvPrP encoding either A or V at residue 136. Using semi-quantitative Western and immuno dot blotting we ascertained that levels of expression in the CNS of Tg(OvPrP-A136)3533+/− and Tg(OvPrP-V136)4166+/− mice were close to that of PrP expressed in the CNS of wild type mice (Fig. 1A).
Both lines of Tg mice tolerated these levels of expression without spontaneously developing recognizable signs of disease (Table 1). In contrast, Tg mice of both genotypes intracerebrally (ic) inoculated with brain homogenates from scrapie-affected sheep succumbed to the neurological effects of prion disease following variable incubation periods (Table 1). Rapid disease onset occurred following inoculation of Tg(OvPrP-V136)4166+/− mice with SSBP/1 prions [15], [16], which consistently produced an ∼130 d mean incubation time. While SSBP/1 also caused disease in Tg(OvPrP-A136)3533+/− mice, mean incubation times were ∼230 to 280 d longer (Fig. 1B and Table 1). In contrast, CH1641 prions [17] induced disease in all inoculated Tg(OvPrP-A136)3533+/− mice with a mean ∼310 d onset of disease, whereas no disease was registered in Tg(OvPrP-V136)4166+/− mice after >560 d. These distinct transmission profiles are consistent with previously recognized strain differences between SSBP/1 and CH1641 scrapie prions [17]. Consistent with this notion, western blot analysis of proteinase K-treated brain extracts of diseased Tg(OvPrP-A136)3533+/− mice confirmed that the molecular profiles which distinguish PrPSc constituting SSBP/1 and CH1641 prions [18] were maintained upon transmission (Fig. 1C). These results demonstrate that Tg(OvPrP-A136)3533+/− and Tg(OvPrP-V136)4166+/− mice are capable of distinguishing scrapie strain-specific transmission patterns, and in turn that these properties are influenced by the A/V136 dimorphism.
Previous studies revealed a positive correlation between PrPSc conformational stability and the incubation times of mouse and cervid prions [19], [20], but not of hamster prions [21], [22]. We performed guanidine denaturation treatments on PrPSc in brain extracts of SSBP/1 infected Tg(OvPrP-V136)4166+/− mice with short incubation times and SSBP/1 infected Tg(OvPrP-A136)3533+/− mice with long incubation times. Analyses using mAb 6H4 revealed distinct stability curves for OvPrPSc-V136 and OvPrPSc-A136. The conformational stability of OvPrPSc-V136 was lower than OvPrPSc-A136 in the range of GdnHCl concentrations between 1 and 2 M, (Fig. 2A) and GdnHCl1/2 values were 1.78 and 2.17 respectively. This confirmed that the conformation of OvPrPSc-V136 produced in Tg(OvPrP-V136)4166+/− mice with rapid incubation times was less stable than OvPrPSc-A136 produced in Tg(OvPrP-A136)3533+/− mice with longer incubation times. We refer to these conformations as unstable (U) and stable (S), and to the rapidly and slowly propagating prions composed of these conformers as SSBP/1-V136(U), and SSBP/1-A136(S).
We then used histoblotting [23], a widely used method for characterizing strain-specific differences in PrPSc distribution [20], [24], with mAb 6H4 to characterize OvPrPSc-A136(S) and OvPrPSc-V136(U) deposition in the CNS. While OvPrPSc-A136(S) had a punctate pattern of accumulation throughout the midbrain, pons, and oblongata of slow incubation time Tg(OvPrP-A136)3533+/− mice (Fig. 3A), the neuroanatomical distribution of OvPrPSc-V136(U) in the same sections of rapid incubation time Tg(OvPrP-V136)4166+/− mice was distinctly different, being more intense and diffusely deposited than OvPrPSc-A136(S) (Fig. 3B). Since Tg(OvPrP-A136)3533+/− and Tg(OvPrP-V136)4166+/− mice were both engineered using the same cosSHa.Tet cosmid vector which drives expression from the PrP gene promoter, we conclude that these differences are not the result of expression OvPrPC-A136 and OvPrPC-V136 in different neuronal populations.
We used brain extracts of Tg(OvPrP-A136)3533+/− or Tg(OvPrP-V136)4166+/− mice as sources of OvPrPC-A136 and OvPrPC-V136 substrates for protein misfolding amplification (PMCA) [25] using SSBP/1. While neither template spontaneously converted to PrPSc in the absence of seeded prions (Fig. 4A and E), SSBP/1 reproducibly converted OvPrPC-V136 to OvPrPSc-V136 during a single round of PMCA (Fig. 4C). In contrast, conversion was not observed after a single round of PMCA when OvPrPC-A136 was used as template (Fig. 4B and Fig. 5A). We therefore used serial PMCA (sPMCA) [26] over 10 rounds to address whether conversion of OvPrPC-A136 to OvPrPSc-A136 might be detected after prolonged replication. Serial PMCA was performed in triplicate using equal amounts of PrP from three different Tg(OvPrP-A136)3533+/− or Tg(OvPrP-V136)4166+/− mouse brains using sheep SSBP/1 as seed. Apart from a slight but consistent decrease between rounds two and five, OvPrPSc-V136 production, detected using mAb 6H4, was sustained throughout rounds one to 10 (Fig. 4C). As expected, mAb PRC5 failed to detect OvPrPSc-V136 (Fig. 4G). In contrast, OvPrPSc-A136 was undetectable with mAbs 6H4 or PRC5 until round eight, after which levels decreased during rounds nine and 10 (Fig. 4B and F).
Having established that Tg mice expressing OvPrPC-V136 and OvPrPC-A136 propagate SSBP/1-V136(U) and SSBP/1-A136(S) prions with relatively rapid and slow incubation times respectively, we produced Tg(OvPrP-A/V136) mice expressing both OvPrPC-A136 and OvPrPC-V136 and inoculated them with SSBP/1 to examine whether disease developed with fast, slow or intermediate kinetics. Although more rapid than the ∼130 d onset of disease in Tg(OvPrP-V136)4166+/− mice (P = 0.0094), the mean 105±5 d onset of disease contrasted with the >400 d SSBP/1 incubation times observed in Tg(OvPrP-A136)3533+/− mice (Fig. 1B and Table 1).
Stability assessments using mAb 6H4 showed that the denaturation curves of OvPrPSc produced in the brains of diseased Tg(OvPrP-V136)4166+/− and Tg(OvPrP-A/V136) mice were superimposable over most of the range of GdnHCl concentrations (Fig. 2A), indicating that OvPrPSc produced in Tg(OvPrP-A/V) mice shared the conformation of OvPrPSc-V136(U) produced in SSBP/1 infected Tg(OvPrP-V136)4166+/− mice. In accordance with this notion, histoblotting using mAb 6H4 showed that the neuroanatomical distribution of OvPrPSc(U) in the brains of diseased Tg(OvPrP-A/V) mice mirrored the diffuse deposition of the OvPrPSc-V136(U) conformer located in similar sections of rapid incubation time Tg(OvPrP-V136)4166+/− mice (Fig. 3C).
While the rapid SSBP/1 incubation times, and properties of the converted PrPSc in diseased Tg(OvPrP-A/V) were consistent with propagation of SSBP/1-V136(U) prions, remarkably, western blotting of diseased Tg(OvPrP-A/V136) brain extracts with mAb PRC5 revealed substantial conversion of OvPrC-A136 to OvPrPSc-A136 (Fig. 2D). Densitometric comparisons of OvPrPSc levels using mAbs 6H4 (Fig. 2B) and PRC5 (Fig. 2D) allowed us to estimate relative conversion efficiencies of each allele product in the brains of SSBP/1 infected Tg(OvPrP-A/V136) mice. Using samples from diseased Tg(OvPrP-A136)3533+/− mice probed with mAbs 6H4 and PRC5 as normalizing controls for differences in the affinities of the two mAbs for OvPrP-A136, we estimated by Western or dot blotting that OvPrPSc-A136 comprised ∼45% of total PK-resistant PrP in the brains of diseased Tg(OvPrP-A/V136) mice.
We then used mAb PRC5 to determine the conformation of OvPrPSc-A136 among total OvPrPSc produced in the brains of diseased Tg(OvPrP-A/V136) mice. The 1.64 GdnHCl1/2 value of OvPrPSc-A136 produced under these conditions was distinct from that of OvPrPSc-A136(S) produced in long incubation time Tg(OvPrP-A136)3533+/− mice (GdnHCl1/2 = 2.11), and their non-superimposable PRC5 denaturation curves were significantly different in the range of 1.5–2.5 M GdnHCl (Fig. 2C). These findings demonstrated that the conformation of OvPrPSc-A136 in rapid incubation time Tg(OvPrP-A/V) mice was distinct from OvPrPSc-A136(S) produced in long incubation time Tg(OvPrP-A136)3533+/− mice. We refer to this novel conformation as OvPrPSc-A136(U), and to the resulting prions as SSBP/1-A136(U).
Histoblotting using mAb PRC5 confirmed the comparatively limited and punctate distribution pattern of OvPrPSc-A136(S) in the CNS of long incubation time Tg(OvPrP-A136)3533+/− mice (Fig. 3D) that we observed with mAb 6H4 (Fig. 3A). As expected, OvPrPSc-V136(U) in the CNS of diseased Tg(OvPrP-V136)4166+/− mice was refractory to detection by mAb PRC5 (Fig. 3E). We probed histoblots of the CNS from diseased Tg(OvPrP-A/V136) mice with mAb PRC5 to assess the appearance and distribution of OvPrPSc-A136(U). In contrast to the punctate deposits of OvPrPSc-A136(S) in long incubation time Tg(OvPrP-A136)3533+/− mice (Fig. 3A and D), OvPrPSc-A136(U) in Tg(OvPrP-A/V136) mice (Fig. 3F) acquired a diffuse deposition and a distribution pattern that was equivalent to OvPrPSc-V136(U) in Tg(OvPrP-V136)4166+/− mice (Fig. 3B). Consistent with the co-expression of each allele in identical cell populations of Tg(OvPrP-A/V) mice, spatial distributions of 6H4- and PRC5-reactive PrP coincided in all analyzed sections of Tg(OvPrP-A/V) mice (Fig. 3G and H).
To simulate the combined effects of OvPrP-A136 and OvPrP-V136 on PrP conversion in Tg(OvPrP-A/V136) mice in vitro, we mixed equal quantities of OvPrPC-A136 and OvPrPC-V136 in PMCA reactions seeded with SSBP/1. Under these conditions, similar to when OvPrPC-V136 was present in isolation (Fig. 4C), we observed early, reproducible conversion to OvPrPSc in round one (Fig. 4D). Probing of western blots with mAb PRC5 showed that OvPrPSc-A136 was a component of this converted material (Fig. 4H). Thus, similar to our observations in Tg mice, the presence of OvPrPC-V136 induced the relatively rapid conversion of OvPrPC-A136 to OvPrPSc-A136 by SSBP/1. Interestingly, subsequent conversion of both OvPrPC-A136 and OvPrPC-V136 diminished in rounds two to five, ultimately becoming undetectable through rounds six to 10 (Figs. 4D and H).
SSBP/1 was originally produced from a pool of diseased sheep brains from the positive selection line in the Neuropathogenesis Unit (NPU) Cheviot sheep flock, and has subsequently been passaged as a pool. We next compared the seeding properties of SSBP/1 with those of SSBP/1-A136(S) or SSBP/1-V136(U) prions derived from SSBP/1-infected Tg(OvPrP-A136)3533+/− or Tg(OvPrP-V136)4166+/− mice. We monitored conversion of OvPrPC-A136 or OvPrPC-V136 templates every two hours for a total of 12 h of PMCA. SSBP/1-V136(U) had the same PMCA properties as SSBP/1: both SSBP/1 and SSBP/1-V136(U) prions efficiently converted OvPrPC-V136 in isolation, but not OvPrPC-A136 in isolation; when both templates were present in the PMCA reaction, the presence of OvPrPC-V136 facilitated conversion of OvPrPC-A136 to OvPrPSc-A136 by SSBP/1 or SSBP/1-V136(U) prions (Figs. 5A and B). In contrast, SSBP/1-A136(S) prions converted either OvPrPC-V136 or OvPrPC-A136 templates to PrPSc when they were present in isolation, the latter being unequivocally confirmed to be OvPrPSc-A136 using mAb PRC5 (Fig. 5C); however, in the presence of both templates SSBP/1-A136(S) prion propagation was inhibited (Fig. 5C). The properties of prions derived from Tg(OvPrP-A/V) mice differed from SSBP/1, SSBP/1-A136(S) or SSBP1/-V136(U) prions. Like SSBP/1 and SSBP/1-V136(U), prions passaged through these mice efficiently converted OvPrPC-V136, but not OvPrPC-A136. However, unlike SSBP/1 and SSBP/1-V136(U), such prions failed to facilitate conversion of OvPrPC-A136 in the presence of OvPrPC-V136 (Fig. 5D).
Previous studies described the production of Tg mice expressing OvPrP, and reported their susceptibility to scrapie prions [27]–[32]. The most widely characterized models are tg338 mice expressing OvPrP-V136 [31], and Tgov59 [33] or Tgov4 [29] lines expressing OvPrP-A136. In the case of tg338 mice, the transgene was comprised of a bacterial artificial chromosome insert of 125 kb of sheep DNA, while in the case of Tgov59 and Tgov4 mice the neuron specific enoloase promoter was used to drive OvPrP expression. These lines are maintained on different heterogeneous genetic backgrounds, and CNS expression levels in tg338 mice are ∼8- to 10-fold higher than wild type, while Tgov59 and Tgov4 lines each over express OvPrP-ARQ at levels ∼2- to 4-fold higher than those found in sheep brain. Spontaneous neurological dysfunction has been reported in Tg lines over expressing OvPrP [27], [31].
Tg(OvPrP-A136)3533+/− and Tg(OvPrP-V136)4166+/− mice express transgene-encoded PrP, either slightly lower, or slightly higher than PrP levels normally expressed in the CNS of wild type mice. Since Tg(OvPrP-A136)3533+/− and Tg(OvPrP-V136)4166+/− lines were produced using the cosSHa.Tet cosmid vector which drives expression from the PrP gene promoter [34], we expected expression of OvPrPC-A136 and OvPrPC-V136 in identical neuronal populations, and therefore that both alleles are co-expressed in the same cells of Tg(OvPrP-A/V) mice. Finally, other than variable transgene insertion loci, both lines are otherwise sygeneic on an inbred Prnp0/0/FVB background.
Previous studies reported on Tg mice expressing OvPrP with V at 136, referred to as Tg(OvPrP)14882+/− mice, that were also produced in a Prnp0/0/FVB background using the cosSHa.Tet cosmid vector [32]. However, in that study, comparable Tg mice expressing OvPrP-A136 were not reported. Median SSBP/1 scrapie incubation times in Tg(OvPrP)14882+/−mice were 75 d, and this line expresses OvPrP at levels only slightly higher than Tg(OvPrP-V136)4166+/− mice. While we exercise caution when comparing results from mice produced by different groups, the otherwise similar properties of Tg(OvPrP)14882+/− and Tg(OvPrP-V136)4166+/− mice suggest that even slight differences in the levels of transgene expression can have significant effects on prion incubation time.
A clear link to codon 136 genotype and susceptibility/resistance to different sheep scrapie isolates has been described in multiple previous studies. Importantly, the influence of residue 136 on the transmission of SSBP/1 and CH1641 prions in Tg(OvPrP-A136)3533+/− and Tg(OvPrP-V136)4166+/− mice is in accordance with the properties of these isolates in sheep of various genotypes [17]. Generally, increased susceptibility to scrapie is associated with expression of OvPrP-V136, with A/A136 being the most resistant, and V/V136 the most susceptible genotypes. In the case of SSBP/1 incubation periods are ∼170 days in V/V136 sheep, while transmission to A/A136 sheep is relatively inefficient, with no disease recorded after >1000 days [35]. While SSBP/1 eventually transmits to Tg(OvPrP-A136)3533+/− mice with incubation times exceeding 400 days, the general effects of the A/V136 dimorphism on SSBP/1 transmission observed in sheep are recapitulated in Tg(OvPrP-A136)3533+/− and Tg(OvPrP-V136)4166+/− mice (Table 1). Similarly, CH1641, which propagates efficiently in A/A136 sheep [35], preferentially propagates in Tg(OvPrP-A136)3533+/− mice (Table 1). In previous studies, CH1641 transmitted to TgOvPrP4 mice with an ∼250 d mean incubation time [36].
Although SSBP/1 incubation times are prolonged in A/V136 compared to V/V 136 sheep [35], in our studies incubation times were shorter in Tg(OvPrP-A/V) than in Tg(OvPrP-V136)4166+/− mice. While the condition of A/V136 heterozygosity has not been previously modeled in Tg mice, this difference may result from double the levels of transgene expression in Tg(OvPrP-A/V) mice compared to Tg(OvPrP-A136)3533+/− and Tg(OvPrP-V136)4166+/− mice. Tg(OvPrP-A/V136) mice were derived by mating Tg(OvPrP-A136)3533+/+ with Tg(OvPrP-V136)4166+/+ mice, and therefore express greater total levels of OvPrP than Tg(OvPrP-A136)3533+/− and Tg(OvPrP-V136)4166+/− mice (Fig. 1A). Since the levels of OvPrP-V136 are equivalent in Tg(OvPrP-V136)4166+/− and Tg(OvPrP-A/V) mice, and we show that OvPrP-A136 also becomes available for conversion, this situation results in more available substrate for conversion. While previous studies revealed an inverse correlation between transgene expression levels and prion incubation times in Tg mice [8], whether shorter incubation periods in Tg(OvPrP-A/V136) mice than in Tg(OvPrP-V136)4166+/− mice reflect overall differences in PrPC expression levels remains uncertain. Differences in scrapie pathogenesis between mice and sheep may also reflect the influence of additional factors on disease in the natural host including other PRNP polymorphisms [37], [38], and different involvements of the lymphoreticular system in sheep compared to Tg mice.
Our observations in Tg mice expressing individual allele products suggested that rapid or prolonged SSBP/1 incubation times in Tg(OvPrP-V136)4166+/− and Tg(OvPrP-A136)3533+/− mice respectively, reflected preferential conversion by SSBP/1 prions of OvPrPC-V136, rapidly producing a relatively unstable OvPrPSc-V136(U) conformation that was diffusely deposited in the CNS, compared to the slower conversion of OvPrPC-A136 to the more stable OvPrPSc-A136(S) conformer which accumulated in the CNS with a punctate pattern (Figs. 1–3). Our results are consistent with the selection by the A/V136 dimorphism of SSBP/1-A136(S) and SSBP/1-V136(U) prions in Tg(OvPrP-A136)3533+/− and Tg(OvPrP-V136)4166+/− mice respectively. We also show that PMCA recapitulates the influence of the A/V136 polymorphism on the kinetics of SSBP/1 propagation observed in Tg mice. The general conclusions from these studies agree with previously published assessments of the mechanism of conformational selection by distinct PrP primary structures [39], [40].
Based on the rapid SSBP/1 incubation times in Tg(OvPrP-A/V136) mice, and shared conformational and distribution properties of OvPrPSc produced under these conditions with OvPrPSc-V136(U) in Tg(OvPrP-V136)4166+/− mice, we speculated that OvPrP-A136 played no part during the propagation of SSBP/1 prions in Tg(OvPrP-A/V) mice. To address this we used mAb PRC5 to exclusively monitor conversion of OvPrPC-A136. Surprisingly, in contrast to its relatively slow conversion when OvPrPC-A136 is expressed in isolation, co-expression with OvPrPC-V136 in Tg(OvPrP-A/V136) mice facilitated rapid conversion of OvPrPC-A136 to OvPrPSc-A136. The conformation and diffuse CNS distribution of the resulting OvPrPSc-A136(U) were equivalent to that of OvPrPSc-V136(U) and not OvPrPSc-A136(S). Collectively, these results lead us to conclude that once OvPrPSc-V136(U) is formed by conversion of OvPrPC-V136 by SSBP/1 prions, the resulting unstable conformation induces rapid conversion of OvPrPC-A136 to OvPrPSc-A136(U). That this outcome is dependent on allele co-expression within the host is demonstrated by the inability of the OvPrPSc-V136(U) conformer to template OvPrPC-A136 when it is expressed in isolation.
Effects of OvPrP genotype on the propagation of scrapie prions were not controlled during the isolation and propagation of SSBP/1. Passage of SSBP/1 through Tg mice therefore allowed us to generate prions composed solely of OvPrPSc-A136, OvPrPSc-V136, or mixtures of both, and to draw additional conclusions about the effects of the A/V136 dimorphism of prion propagation using PMCA. Similar to our observations in Tg mice, SSBP/1 failed to convert OvPrPC-A136 to OvPrPSc-A136 by PMCA, except in the presence of OvPrPC-V136 (Figs. 4 and 5). These conversion properties are shared with SSBP/1-V136(U) prions, but are distinct from SSBP/1-A136(S) prions, which show facile conversion of both OvPrPC-V136 and OvPrPC-A136. These results suggest that the SSBP/1-V136(U) is the dominant strain in the natural SSBP/1 isolate. Multiple parameters could account for this, including, but not restricted to, the effects of OvPrP genotype, for example as a result of exclusive propagation in sheep of the V136/V136 genotype, route of transmission in the infected sheep, and differential/selective prion replication in the lymphoreticular or central nervous systems of sheep. While our analyses indicate the presence of both OvPrP-V136 and OvPrP-A136 alleles in SSBB/1 (Table 1), it is important to note that SSBP/1 was derived from a pool of sheep brains of undefined genotypes. PCR approach precludes assessment of the extent to which alleles are present in a sample, raising the possibility that the one or other allele exists as a minor component in SSBP/1.
Our findings also suggest that PrPSc conformers may cross-inhibit PrP conversion. In case of SSBP/1-A136(S) prions, the presence of OvPrPC-V136 inhibited PMCA of OvPrPC-A136 (Fig. 5C). Also, while SSBP/1 seeding of PMCA reactions containing mixtures of OvPrPC-A136 and OvPrPC-V136 resulted in robust, reproducible conversion to OvPrPSc-A136 as early as round one (Fig. 4H), total PrPSc production was ephemeral with subsequent PrPSc formation diminishing during rounds two to five, and conversion ultimately becoming undetectable after round six. Since early PrPSc conversion was sustained out to round 10 when OvPrPSc-A136 was not produced (Fig. 4C and G), these results are consistent with inhibited conversion of OvPrPC-V136 to OvPrPSc-V136 by OvPrPSc-A136. While early (round one) PMCA conversion of PrPSc by SSBP/1 with either OvPrPC-V136 or mixtures of OvPrPC-V136 and OvPrPC-A136 correlates with early onset of disease following SSBP/1 infection of both Tg(OvPrP-V136) and Tg(OvPrP-A/V136) mice, the subsequent inhibitory effects of OvPrPSc-A136 observed in PMCA would be impossible to detect in vivo, since Tg(OvPrP-A/V) mice succumb to the lethal effects of early PrPSc accumulation. Consistent with an inhibitory effect of OvPrPSc-A136(U), prions from Tg(OvPrP-A/V) mice, while they converted OvPrPC-V136 in isolation, failed to convert OvPrPC-A136 to PrPSc in the presence of OvPrPC-V136 (Fig. 5D). Thus, the properties of prions from this defined genetic background differ from SSBP/1. We emphasize that, despite PCR data supporting the presence of OvPrP-A136 alleles in this isolate, SSBP/1 was derived from sheep of undefined OvPrP genotypes, rather than sheep with a defined heterozygous OvPrP-A/V136 genotype.
The inter-related effects of PrP primary and higher order structures on prion transmission were addressed in the Conformational Selection Model, which proposed that strains are composed of a range of PrPSc conformers, or quasi-species, and that only a subset of PrPSc conformations is compatible with each PrP primary structure [41]. While this model also took into account the effects of polymorphic variation on prion propagation, it did so only in the context of Tg mice expressing individual PrP allele products. Transgenetic studies of the human codon 129 methionine (M)/valine (V) polymorphism, and the analogous codon 132 M/leucine (L) polymorphism in elk, indicated that these dimorphisms acted to restrict or promote the propagation of particular prion strains [39], [40]. While the responses of Tg(OvPrP-A136)3533+/− and Tg(OvPrP-V136)4166+/− mice are consistent with this notion, that is selection of the U conformer by OvPrP-V136, and the S conformer by OvPrP-A136, our unprecedented ability to analyze allele specific conversion in infected Tg(OvPrP-A/A136) mice reveals a more complex mechanism where mixtures of PrP variants may assist or inhibit the propagation of strains under various conditions. For example, SSBP/1 or SSBP/1-V136(U) prions facilitate conversion of OvPrPC-A136 to OvPrPSc-A136(U) only in the presence of OvPrPC-V136. Expressed in isolation, conversion of OvPrPC-A136 is favored by the OvPrPSc(S) conformer. Our results demonstrate that co-expression of different polymorphic forms of PrP, which would be the norm in humans and animals, have profound effects on conformational selection of prion strains.
The results reported here address the molecular mechanisms associated with the phenomenon of prion strain over-dominance first observed by Dickinson and Outram [42], and subsequently reported in other settings involving co-expression of long and short incubation time PrP alleles [43]. While this phenomenon was reconciled at the time by the assumption that TSEs were caused by unidentified viral agents, our results now indicate that the suggestion raised by those studies, namely that over-dominance most likely resulted from physical interaction of allele products of the scrapie incubation time locus during infection, was prescient. Our results support a molecular mechanism involving cross templating of an otherwise resistant allele product by a dominant prion conformer, in this case OvPrPSc(U), which, we speculate, involves physical association of otherwise “susceptible” and “resistant” allele products. Consistent with the observations reported here, prion strain interference may also utilize similar mechanisms of conformational selection in a host expressing different PrP allele products infected with long and short incubation period strains with different PrPSc conformational stabilities [44], [45].
In conclusion, we have used a combination of transgenic, immunologic, and in vitro approaches to explore the mechanism by which PrP primary structure variations and the conformations enciphered by different prion strains interact to control TSE propagation. While our results support previous studies indicating that PrP susceptibility polymorphisms, expressed in isolation, act to restrict or promote the propagation of particular prion conformers, we now show that under conditions of allele co-expression a dominant conformer may alter the conversion potential of an otherwise resistant PrP polymorphic variant to an unfavorable prion strain. While such responses are analogous to the phenotypic expression of genetically determined heritable traits, dominant prion conformers act epigenetically by means of protein-mediated conformational templating. By expanding the range of possible conformations adoptable by a particular prion protein primary structure, such interactive effects provide a mechanism for promoting strain fitness, and, we speculate, strain diversification. While the precise number scrapie strains in sheep and goats remains uncertain, the description of at least 24 additional major sheep PRNP polymorphisms, and combinations thereof, is likely to have a significant influence on strain diversity.
All animal work was conducted according to the National Institutes of Health guidelines for housing and care of laboratory animals, and performed under protocols approved by the Colorado State University Institutional Animal Care and Use Committee, with approval number 11-2996A.
Sequences upstream of codon 44 of the OvPrP-A136 and V136 coding sequences were replaced with the corresponding sequence from mouse PrP. The resulting constructs contained the OvPrP coding sequence, except for addition of an extra residue for glycine at codon 31, and the mouse PrP N-terminal signal peptide instead of OvPrP signal peptide. Tg mice were generated by cloning the OvPrP-A136 and OvPrP-V136 expression constructs into the cosSHa.Tet cosmid vector [34], and microinjection of embryos from inbred Prnp0/0/FVB mice. Tg founders were identified by PCR screening of genomic DNA isolated from tail snips. Founder mice were mated with inbred Prnp0/0/FVB mice, and generally maintained with the transgene in the hemizygous state, with Tg mice identified by PCR screening of genomic DNA from weanlings. It was also possible to generate homozygous counterparts of each line, and Tg(OvPrP-A/V136) mice were generated by crossing homozygous Tg(OvPrP-V136)4166+/+ mice with homozygous Tg(OvPrP-A136)3533+/+ mice. We used immuno-dot blotting and Western blotting with mAb 6H4 (Prionics, Schlieren, Switzerland) to estimate the levels of OvPrP expression. Tg mice subsequently shipped to and maintained in Edinburgh were crossed onto the Prnp0/0/129Ola background [46].
SSBP/1 originated as a homogenate of three natural scrapie brains that were subsequently passaged mostly through Cheviot sheep at the Neuropathogenesis Unit (NPU), Edinburgh UK [15], [16]. CH1641 is a naturally infected cheviot sheep from the NPU flock [17]. The presence of OvPrP-A136 or OvPrP-V136 alleles in these samples was ascertained by restriction fragment length polymorphism analysis of the PCR amplified PRNP coding sequences.
Ten % mouse brain homogenates (w/v) were prepared in phosphate-buffered saline (PBS) lacking calcium and magnesium ions by repeated extrusion through 18- and 21-gauge needles. Sheep brain homogenates (10%) in PBS were prepared by repeated extrusion through 14-gauge, followed by 18- to 28-gauge needles in PBS. Total protein content was determined by bicinchonic acid (BCA) assay (Pierce Biotechnology, Inc.).
Anesthetized mice were inoculated intracerebrally with 30 µl of 1% (w/v) brain extracts prepared and diluted in PBS.
General health was monitored daily. Onset of prion disease was determined by observation of the progressive development of at least three of the following clinical signs: truncal ataxia, loss of extensor reflex, difficulty righting from a supine position, plastic tail, head bobbing or tilting, kyphotic posture, circling and paresis/paralysis. Animals were diagnosed when at least two investigators agreed with the manifestation of these signs. Incubation time is defined as the period between the time of inoculation to the day on which subsequently progressive clinical signs were initially recorded.
Brain homogenates containing 500 µg protein were digested with 400 µg/ml proteinase K (PK) in 0.4 M NaCl, 10 mM Tris–HCl, pH 8.0, 2 mM EDTA, pH 8.0, and 2% SDS at 55°C overnight. Genomic DNA was precipitated with isopropanol. The partial OvPrP coding sequence was amplified by PCR with the forward and reverse primers: 5′-GGACAGGGCAGTCCTGGA-3′, 5′-GTGATGCACATTTGCTCCACCACT-3′. PCR products were purified with QIAquick Gel Extraction kit (QIAGEN Science, MA, USA), digested with BspH I that only recognizes the OvPrP-V136 allele, and the products were resolved on a 1.2% agarose gel.
Tg mice were perfused with PBS/5 mM EDTA. Ten % brain homogenates (w/v) were prepared in PBS containing 150 mM NaCl, 1.0% Triton X-100, and the complete TM cocktail of protease inhibitors (Roche, Mannheim, Germany). Samples were clarified by brief, low-speed centrifugation. Protein concentrations of brain homogenates used as substrates for PMCA were adjusted to contain equivalent amounts of OvPrP-A136 or OvPrP-V136, based on the estimated relative levels of transgene expression. Substrates in which OvPrP-A136 and OvPrP-V136 were mixed were adjusted based on the estimated relative levels of transgene expression, so that approximately equal amounts of each allele product were present in the PMCA reaction. PMCA reactions were performed as described previously [20], [47] at a seed to substrate ratio of 1∶180. One cycle corresponded to 20 seconds of sonication followed by 30 minutes incubation at 37°C. Controls samples were incubated for the same duration at 37°C without sonication. Amplified and control samples were digested with PK at a final concentration of 0.33 µg/µl and analyzed on western blots using mAbs 6H4 or PRC5.
Brain homogenates and cell lysates were digested with 100 µg/ml or 30 µg/ml of PK respectively (Roche, Mannheim, Germany) in cold lysis buffer for 1 h at 37°C. Digestion was terminated with phenylmethylsulfonyl fluoride at a final concentration of 2 µM. Samples were boiled for 10 min in the absence of β-meracaptoethanol [14] and proteins were resolved by SDS-PAGE and transferred to polyvinylidenedifluoride Immobilon (PVDF)-FL membranes (Millipore, Billerica, USA). Membranes were probed with primary mAbs followed by horseradish peroxidase–conjugated anti-mouse secondary antibody (GE Healthcare, Little Chalfont, UK). Protein was visualized by chemiluminescence using ECL Plus (GE Healthcare, Piscataway, USA) and an FLA-5000 scanner (Fujifilm Life Science, Woodbridge, USA).
Brain homogenates containing 5 µg protein were incubated with various concentrations of guanidine hydrochloride (GdnHCl) in 96-well plates for 1 h at room temperature. Samples were adjusted with PBS to a final of concentration of GdnHCl of 0.5 M and transferred onto nitrocellulose (Whatman GmbH, Dassel, Germany) using a dot blot apparatus. After two PBS washes, the membrane was air-dried for 1 h, then incubated with 5 µg/mL PK in 50 mM Tris-HCl, pH 8.0, 150 mM NaCl, 0.5% sodium deoxycholate, 0.5% Igepal CA-630 for 1 h at 37°C. PK was inactivated with 2 mM PMSF. The membrane was incubated in 3 M guanidine thiocyanate in Tris-HCl, pH 7.8 for 10 min at room temperature. After four washes with PBS, the membrane was blocked with 5% nonfat milk in TBST for 1 h, and probed with mAbs 6H4 (1∶20,000) or PRC5 (1∶5000) overnight at 4°C, followed by HRP-conjugated goat anti-mouse IgG secondary antibody. The membrane was developed with ECL Plus and scanned with GE image quant 4000. The signal was analyzed with ImageQuant TL 7.0 software.
Histoblots were produced and analyzed according to previously described protocols [23]. Images were captured with a NikonDMX 1200F digital camera in conjunction with Metamorph software (Molecular Devices).
|
10.1371/journal.pgen.1000797 | Genotype and Gene Expression Associations with Immune Function in Drosophila | It is now well established that natural populations of Drosophila melanogaster harbor substantial genetic variation associated with physiological measures of immune function. In no case, however, have intermediate measures of immune function, such as transcriptional activity of immune-related genes, been tested as mediators of phenotypic variation in immunity. In this study, we measured bacterial load sustained after infection of D. melanogaster with Serratia marcescens, Providencia rettgeri, Enterococcus faecalis, and Lactococcus lactis in a panel of 94 third-chromosome substitution lines. We also measured transcriptional levels of 329 immune-related genes eight hours after infection with E. faecalis and S. marcescens in lines from the phenotypic tails of the test panel. We genotyped the substitution lines at 137 polymorphic markers distributed across 25 genes in order to test for statistical associations among genotype, bacterial load, and transcriptional dynamics. We find that genetic polymorphisms in the pathogen recognition genes (and particularly in PGRP-LC, GNBP1, and GNBP2) are most significantly associated with variation in bacterial load. We also find that overall transcriptional induction of effector proteins is a significant predictor of bacterial load after infection with E. faecalis, and that a marker upstream of the recognition gene PGRP-SD is statistically associated with variation in both bacterial load and transcriptional induction of effector proteins. These results show that polymorphism in genes near the top of the immune system signaling cascade can have a disproportionate effect on organismal phenotype due to the amplification of minor effects through the cascade.
| Genetic variation for resistance to infection is widespread among insects and other organisms. However, the extent to which this variation in resistance is mediated by changes in infection-induced gene expression is not known. In this study, we assayed expression of immune system genes and bacterial load after infection in a genotyped panel of lines of the model insect Drosophila melanogaster. We find that statistical associations between genetic variants and bacterial load tend to cluster in genes encoding proteins involved in microbial recognition. Variation in suppression of bacterial growth is also determined in part by genetic variation in the expression of downstream components of the immune system that function to directly kill bacteria, despite finding no genetic variation in any single of these effector gene significantly associated with phenotype. Instead, it appears that activity differences in upstream components of the pathway have a cascading effect that results in larger variation in the expression of coordinately regulated downstream effector genes. These results imply that the interactions among genes need to be taken into account when assessing the phenotypic consequences of genetic variation, as signaling cascades such as those in the immune response have the potential to amplify the phenotypic effects of minor genetic variation in individual genes.
| Drosophila, like other insects, use a generalized immune response to combat pathogens. Unlike vertebrates, the insect immune response consists solely of an innate response, with cellular and humoral (cell-free) arms [reviewed in 1]. Despite considerable knowledge of the molecular basis of the Drosophila immune response, and increasing understanding of the extent of natural genetic variation for immunocompetence in this system [2]–[4], relatively little is known about the role of network structure in shaping the phenotypic consequences of genetic variation.
Linking genetic variation in transcriptional regulation to differences in complex phenotypes has the potential to illuminate mechanistic aspects of genotype-phenotype associations. Passador-Gurgel and coworkers [5] identified several genes in which transcript levels significantly associate with survival times after exposure of D. melanogaster to nicotine. Other studies in Drosophila have identified transcriptional variation associated with male reproductive success [6], male body size [7], aggressive behavior [8] and locomotive behavior [9]. While in some cases it has been possible to show that genetically determined transcriptional differences are statistically correlated with phenotypic differences, these studies have generally not identified causal genetic variants. In Drosophila, linking genetic variation to phenotypic variation via transcriptional changes has proven difficult [10],[11]. The Drosophila immune system provides an ideal opportunity to examine the consequences of genetic variation and differences among lines in patterns of gene expression in the context of a well-defined network.
In Drosophila, the humoral response is initiated by the recognition of microbial cell wall component by proteins such as PGRPs and GNBPs [12]–[14]. These proteins activate two primary signaling pathways, the Toll and Imd pathways. The Toll pathway is primarily activated after infection by fungi and Gram-positive bacteria, whereas the Imd pathway is primarily activated after infection by Gram-negative bacteria [15],[16], although this specificity is not absolute [17],[18]. In addition to these primary signaling pathways, the JAK/STAT and JNK pathways are thought to play a role in immune response, largely as part of the general stress response and wound healing [19],[20]. Activation of the Toll and Imd signaling pathways leads to the translocation of NF-κB transcription factors (Relish, DIF, Dorsal) to the nucleus where they drive transcription of effector genes, which encode proteins that are directly involved in bacterial clearance, such as antimicrobial peptides. These effectors are then released into the hemolymph, where they act to directly kill invading microorganisms [21].
Previously, we have examined associations between bacterial load after infection with each of four different bacteria and genetic markers (SNPs and indels) in candidate genes on the Drosophila melanogaster second chromosome [2],[3]. Here, using markers in candidate genes on the third chromosome, we examine both bacterial load and gene expression phenotypes, testing associations between genotype, sustained bacterial load, and transcription level of approximately 400 known and putative immune system genes.
We examined a sample of 94 third-chromosome substitution lines for variation in bacterial load sustained 28 hours after infection with each of four different bacteria: Serratia marcescens, Providencia rettgeri, Enterococcus faecalis, and Lactococcus lactis (Figure 1). In order to assess the effect of different third chromosomes on bacterial load phenotypes, we compared the likelihood of the data under a statistical model that includes variation among genetic (third-chromosome) lines as a main effect to the likelihood of the data under a model that does not. Likelihood ratio tests reveal a large, highly significant effect of third chromosome line on phenotypic variation in bacterial load against all four bacteria (S. marcescens: χ2 = 128.42, d.f. = 1, P<2.2×10-16; P. rettgeri: χ2 = 263.88, d.f. = 1, P<2.2×10−16; E. faecalis: χ2 = 51.533, d.f. = 1, P = 7.04×10−13; L. lactis: χ2 = 35.391, d.f. = 1, P = 2.70×10−9). Genetic line explains 66.9% of the non-error variance (14.5% of the overall variance) for load sustained after S. marcescens infection and 58.3% (22.1%) for load sustained after P. rettgeri infection, but only 27.4% (7.2%) for E. faecalis and 19.5% (6.2%) for L. lactis (Table 1). Total variance in bacterial load is much higher for the two Gram-positive bacteria (E. faecalis and L. lactis), as is residual variance and the fraction of total variance explained by experimental factors, suggesting that these infections produce noisier data (Table 1). The smaller fraction of variance attributable to line after infection with these two bacteria presumably stems from stochastic events during initiation and establishment of infection. The overall mean load sustained after infection also varies among bacteria, ranging from a low of 2,186 colony forming units (CFU) per fly 28 hours after infection with S. marcescens to a high of 653,436 CFU per fly after infection with L. lactis. Correlations of line means between bacteria (measured as Spearman's ρ) are generally moderate and positive (Table 2). While the positive sign of correlations between bacteria suggests that some genetic lines may have generally better immune responses, the relatively small magnitude suggests substantial bacteria-specific effects.
We tested for statistical associations between bacterial load and genotypes at 137 polymorphisms in 26 genes and gene families on the third chromosome with known or suspected immune function. These included 6 antimicrobial peptide loci, 10 putative recognition loci (GNBPs and PGRPs), 8 known signaling loci, the Toll-like receptor Toll-9, and the iron-binding protein Transferrin 2 (Table 3). Our association test is based on mixed linear models: we assessed significance by comparing the observed model coefficient (effect size) for the marker in question to a null distribution generated from 5070 permuted data sets where phenotypes are randomly shuffled across lines while preserving linkage disequilibrium among genetic markers and correlations among bacterial loads after infection with different bacteria (see Methods for details of the permutation protocol).
We also tested for associations between SNP markers and the first principal component estimated from line means of bacteria after infection with each of the four different bacteria. This principal component is significantly positively correlated with load after infection with all four bacteria, suggesting that it represents a measure of general immune competence and/or general vigor. Results from this analysis recover statistical associations with markers that show significant associations with bacterial load measured after infection with multiple different bacteria, but do not uncover any additional general immune factors, and are not discussed further (Table S1). All statistical tests were implemented in R, as described in the Methods, and presented in Table S1.
Across all bacteria, 43 tests (7.85%) are significant at a nominal αof 0.05, and 12 tests (2.19%) are significant at a nominal α of 0.01; in both cases, we observe a significant excess of significant tests (α = 0.05: χ2 = 9.35, d. f. = 1, P-value = 0.0022; α = 0.01: χ2 = 7.84, d. f. = 1, P-value = 0.0051). Because some SNPs are in linkage disequilibrium and because bacterial loads across different pathogen challenges are weakly positively correlated, the 548 tests we conducted (137 markers by 4 phenotypes) are not likely to be independent. Thus, we also calculated the null distribution of significant SNPs based on permutations that preserve the correlation structure in the data (see Methods for additional details). We observe a mean of 28.6 significant tests under the null hypothesis at an α of 0.05, and a mean of 5.9 significant tests under the null hypothesis at an α of 0.01. In both cases, the number of significant tests we observe in the permuted data are significantly fewer than the values we observe in the real data (α = 0.05: 43 observed significant tests, P-value = 0.0323; α = 0.01: 12 observed significant tests, P-value = 0.0296).
Several markers in our dataset (8 and 2 at a nominal α of 0.05 and 0.01, respectively) are nominally associated with variation in multiple independent bacterial load phenotypes. Assuming all tests are independent, it is extremely unlikely that we would observe this number of SNPs associated with more than one bacterial load phenotype (α = 0.05: χ2 = 19.52, P-value (by simulation) = 0.00087; α = 0.01: χ2 = 45.42, P-value (by simulation) = 0.00299). To verify this conclusion in the face of non-independence among tests, we used a permutation approach to estimate the null distribution of the number of SNPs with two or more significant tests under the assumption of no genotype-phenotype associations (α = 0.05: P-value = 0.0118; α = 0.01: P-value = 0.0053; see Methods for details).
Significant tests at a nominal α of 0.01 (0.05) are not randomly distributed among bacteria: 83.3% (67.4%) of the significant cases represent associations between genotype and bacterial load after infection with Gram-negative bacteria (S. marcescens and P. rettgeri). Gram-positive bacterial load has higher residual error variance and higher experimental variance in our experiments (Table 1), which could lead to reduced power to detect associations with this phenotype. In order to test this hypothesis, we calculated power by simulation, assuming variances estimated from either the Gram-negative or Gram-positive bacteria in our study (see Methods for details). Although power is lower for our simulated Gram-positive data across a range of effect sizes and two assumptions about minor allele frequencies (Figure S1), if average effect size of associations is equal between the two bacterial types we would not expect to see such a substantial excess of Gram-negative associations. It is possible that the observed excess of associations with resistance to Gram-negative infection could be driven by a biological difference in the response of D. melanogaster to the specific Gram-negative and Gram-positive bacteria we employed in this study that results in less among-line variation in load after infection with these particular Gram-positive bacteria.
Nominally significant associations are also not evenly distributed within functional classes of the immune system. The proportion of tested markers that are associated with bacterial load phenotypes (at a nominal α of 0.05) significantly varies among functional classes (Figure 2; χ2 = 11.35, d. f. = 2, P = 0.0034). Markers in genes encoding recognition proteins have the highest proportion of significant associations with bacterial load (12.11% of tested markers in these genes are significantly associated with phenotype), followed by markers in genes encoding signaling proteins (5.36% of tested markers in these genes are significantly associated with phenotype). Markers in genes encoding effector proteins are rarely associated with differences in bacterial load (only 1.39% of tested markers in genes encoding effectors are significantly associated with phenotype). Average intralocus linkage disequilibrium is not significantly different among functional classes (data not shown), suggesting that this pattern is not driven by biases introduced by LD among SNPs. However, in order to rule out this possibility we generated a distribution for the fraction of significant associations in each of the three functional categories under the null hypothesis that there is no association between genotype and phenotype. Markers in genes encoding recognition proteins are significantly more likely to have significant associations (α = 0.05: 31 observed significant tests compared to a mean of 13.3 in the permutated data, P = 0.0016; α = 0.01: 10 observed vs. mean of 2.73 in permuted data, P = 0.0059). The same pattern does not hold, however, for markers in genes encoding signaling or effector proteins (α = 0.05: Psignaling = 0.492, Peffector = 0.965; α = 0.01: Psignaling = 0.500, Peffector = 1). Furthermore, while the average fraction of markers with significant associations at α = 0.05 (0.01) that are in recognition genes in the permuted dataset is 51.21% (50.96%), in the observed data it is 75.6% (83.3%).
Polymorphism at the GNBP75D locus, consisting of the genes GNBP1 and GNBP2, is particularly striking in the extent and significance of associations with resistance to Gram-negative bacteria (Figure 3). Seven of the 10 SNPs at this locus are nominally significantly associated with variation in bacterial load after infection with P. rettgeri, although average linkage disequilibrium is high at this locus (average pairwise r2 = 0.303; average pairwise D′ = 0.636). Four of those seven SNPs are also significantly associated with differences in bacterial load after infection with S. marcescens. These include one SNP in the 3′ UTR of GNBP2 (GNBP75D_1041), one SNP in the 5′ UTR of GNBP1 (GNBP75D_3350), and a pair of SNPs in the first intron of GNBP1 (GNBP75D_3696 and GNBP75D_3768). Notably, GNBP75D_3696 is one of two SNPs that is significantly associated with differences in bacterial load after infection with two different bacteria at a nominal α of 0.01.
The haplotype structure at the GNBP75D locus is unusual for D. melanogaster. Despite spanning more than 2 kb, the four SNPs mentioned previously are found in only 6 of the 16 possible haplotypes in 91 of the 94 genetic lines (the remaining three lines have unique haplotypes). There are two major haplotypes (A-A-A-T and C-G-G-A) at frequencies of 0.244 and 0.449 respectively. When the phenotypes of the lines that carry these two haplotypes are compared directly using nonparametric tests, the A-A-A-T haplotype has a significantly higher median bacterial load after infection with both S. marcescens (medianAAAT = 8.12, medianCGGA = 7.48, Mann-Whitney U P = 0.036) and P. rettgeri (medianAAAT = 13.16, medianCGGA = 12.26, Mann-Whitney U P = 0.000299).
Further support for a biologically meaningful association of genetic differences at the GNBP75D locus with phenotypic variation comes from an analysis of per-gene significance. Because of the high intra-locus LD, we applied permutation tests to assess significance of effects attributable to genetic variation at each gene. We generated a null distribution of the sum of the χ2 test statistics for each marker within a locus under the assumption of no association between genotype and phenotype, while controlling for confounding effects of correlation among markers within a locus, and compared the observed sum of the χ2 test statistics to this permuted distribution (see Methods for details; full results in Table S2). We find significant evidence for an association between markers in GNBP75D and bacterial load after infection with P. rettgeri (P = 0.00059; P = 0.015 after multiple test correction using the Holm method implemented in the R function p.adjust) and, more weakly, S. marcescens (P = 0.023; P = 0.577 after multiple test correction as above).
PGRP-LC is another recognition gene with repeatable evidence for a significant association between SNPs and bacterial load, albeit somewhat weaker than the evidence for the GNBP75D association. In this case, genotypes at two out of 10 SNPs are associated with variation in bacterial load against at least two different bacteria at a nominal α of 0.05, and a third marker has marginal significance. A SNP marker approximately 125 bp upstream of the transcriptional start site of PGRP-LC (PGRPLC_884) is associated with resistance against both E. faecalis (P = 0.0556) and S. marcescens (P = 0.038). A SNP marker in the third exon of splice variant PGRP-LC-RB (PGRPLC_5624; intronic in splice variants PGRP-LC-RA and PGRP-LC-RC) is associated with variation in bacterial load against L. lactis (P = 0.0481) and S. marcescens (P = 0.006), with the same allele associated with lower bacterial load against both bacteria. Another SNP marker in the fourth exon (in the PGRP domain) of PGRP-LC-RA (PGRPLC_6635; in the intron of PGRP-LC-RA and PGRP-LC-RC) is also associated with variation in bacterial load against L. lactis (P = 0.0075) and S. marcescens (P = 0.0095). These two SNP markers are in linkage disequilibrium (r2 = 0.193, P = 7.92×10−4), but neither is in significant linkage disequilibrium with the upstream marker.
Empirical and theoretical work [22]–[24] suggests that immune function may differ between the sexes, as males and females make different resource allocation decisions between immune defense and reproductive output. These observations lead to the hypothesis that the genetic basis of the immune response may depend on sex: indeed, these kinds of genotype by sex interactions have been observed for other quantitative traits in D. melanogaster [e.g., 25]. To test this hypothesis, we compared the likelihood of our observed bacterial load data under a model with a Sex by Marker interaction to the likelihood of the data under a model without such an effect (but retaining the main effects of Sex and Marker). To assess the significance of the resulting likelihood ratio test statistics, we used a null distribution of likelihood ratio test statistics calculated by permuting the data 1000 times.
We find little evidence for significant effects of marker by sex interactions on bacterial load. While 6.93% of tests are significant at a nominal α of 0.05, a weakly significant excess over the null expectation (χ2 = 4.32, d. f. = 1, P = 0.0377), only 0.91% of tests are significant at a nominal α of 0.01, which is not different from the null expectation (χ2 = 0.042, d. f. = 1, P = 0.8367). While it is possible that there are weak marker by sex interactions that we do not have the power to detect in this experiment, we believe that such effects are likely to be small compared to main effects of SNP across sexes. There is a clear main effect of sex: males have consistently lower bacterial loads irrespective of genotype, consistent with the hypothesis that the sex bias in immune function is phenotypically plastic in Drosophila, and depends on food and mate availability [22]. We have only examined variation on the third chromosome in this study; a similar studies of genes on the second chromosome also find little evidence for substantial sex by SNP interactions [2],[3]. However, a recent study of variation in X-linked immune genes suggests substantial sex by SNP interactions [26].
In order to understand the mechanistic basis of differences in immune phenotypes linked to genetic variation on the third chromosome, we measured gene expression of approximately 700 transcripts in males from a subset of 30 of the 94 phenotyped chromosome 3 substitution lines. Using custom-designed Illumina BeadChip microarrays, we measured transcript abundance under three different conditions (uninfected, 8 hours post S. marcescens infection [Sm-infected], and 8 hours post E. faecalis infection [Ef-infected], where S. marcescens and E. faecalis were chosen arbitrarily to represent Gram-negative and Gram-positive bacteria respectively). We selected the subset of assayed lines to be biased toward the tails of the phenotypic distribution in order to enhance our power to detect correlations between transcript abundance and phenotype. We normalized and log-transformed expression values as described in the Methods. For most analyses, we focused on the Ef-induced (Ef-infected minus uninfected) and Sm-induced (Sm-infected minus uninfected) samples.
In addition to quantifying the 329 genes with a known or putative immune function (including 172 genes with some characterized function and 157 genes predicted to have a role in immunity based on transcriptional induction after infection), our BeadChip microarrays include genes involved in metabolism (139) and sex/reproduction (164), as well as 69 probesets consisting of housekeeping gene controls, and genes involved in insecticide resistance. Full details of the BeadChip design are described in the Methods; the full list of genes are presented as Table S3 (probe sequences are available upon request from T. B. S.). For most analyses, we focus on the 329 immune genes on the BeadChips, although in some cases we use the other genes as controls.
Although with only 30 lines applied to the BeadChip arrays we have limited power to detect associations between SNPs and gene expression variation, we tested for significant associations by comparing a mixed model with a fixed effect of SNP to one with just a fixed intercept. Because permutations are not computationally feasible for the large number of tests required for this analysis, we assessed significance by comparing the likelihood ratio test statistic to a standard χ2 distribution. Overall, 3.55% (9.09%) and 2.98% (10.33%) of genotype-expression association tests are significant at a nominal α of 0.01 (0.05) in the Sm-induced and Ef-induced samples, respectively. In all cases it is highly improbable to obtain this many significant tests purely by chance under the assumption that regulation of expression of all genes is independent (Sm-induced, α = 0.01: χ2 = 6351, P<2.2×10−16; Ef-induced, α = 0.01: χ2 = 3833, P<2.2×10−16; Sm-induced, α = 0.05: χ2 = 5791, P<2.2×10−16; Ef-induced, α = 0.05: χ2 = 3416, P<2.2×10−16). The same pattern holds if we consider the absolute expression level in the Ef-infected, Sm-infected, and uninfected samples individually (data not shown).
Because we assumed the null distribution of the test statistic follows an asymptotic chi-square distribution, it is possible that the excess of significant P-values we observe is primarily due to mis-specification of the null distribution. We expect that polymorphisms in genes known to have a role in the immune system will be more likely to affect expression of immune-related genes than expression of other genes on the BeadChip. Indeed, for the Ef-induced sample, we see significantly more tests with both P<0.01 and P<0.05 among immune-related genes than other genes (P<0.01: 0.0325 vs. 0.0274; χ2 = 21.6874, d.f. = 1, P = 3.206×10−6; P<0.05: 0.0933 vs. 0.0889; χ2 = 5.6409, d.f. = 1, P = 0.01755), although this is not the case for the Sm-induced sample (but note that “non-immune” genes may still be responding transcriptionally to infection). Thus, while it appears that some of the genotyped SNPs in this study have significant effects on gene expression, particularly for the Ef-induced sample, limiting our experiment to 30 lines reduces our power to detect significant associations. Nonetheless, there are 304 and 350 associations between genotypes and induction of immune genes after E. faecalis and S. marcescens infection respectively significant at a 10% false-discovery-rate, which are presented in Table S4. Of particular note is the marker PGRPSD_494, which is associated with expression of 73 of the 329 immune genes we assayed. However, given the uncertainty in the true estimates of significance, we focus on overall qualitative patterns of genotype-expression associations.
Because a considerable amount is known about the transcriptional feedback relationships in innate immune networks, we can make some predictions about the expected direction of associations between genotypes and variation in gene expression of specific genes. Most generally, we expect that markers in upstream genes in the immune pathway should predict expression of downstream genes much more often than vice versa. For example, we believe that genetic differences in signaling genes could lead to differential expression of effector genes, but that genetic differences in effector genes do not result in feedback that influences transcription of signaling genes. For both the Ef-induced and Sm-induced samples, we consistently see an excess of associations between markers in upstream loci and gene expression of downstream loci relative to associations between markers in downstream loci and expression of upstream loci (Table 4). This pattern is consistently more significant between “adjacent” functional classes in the immune network, although the recognition/effector pair is also the case with the smallest number of tests and thus the lowest power.
The network structure argument also has implications for the distribution of cis and trans associations across expression of effector, signaling and recognition genes. Specifically, while there is no reason to believe that cis associations should be related to network structure, we hypothesize that downstream categories (particularly effector genes) will have significantly more trans associations than upstream categories. For both Ef-induced and Sm-induced samples, we find support for this hypothesis. In the Sm-induced sample, 4.32% of tests between trans markers and expression of effector genes are significant, compared to 2.75% for expression of signaling genes and 2.37% for expression of recognition genes (χ2 = 47.6607, d.f. = 2, P-value = 4.473×10−11). In the Ef-induced sample, 4.64% of tests between trans markers and expression of effector genes are significant, compared to 2.65% for signaling genes and 2.62% for recognition genes (χ2 = 64.5568, d.f. = 2, P-value = 9.587×10−15). These differences remain significant if trans tests are split into those that involve markers in the same functional class as the expression phenotype being measured and those that involve markers in different functional classes (data not shown). In neither case do we observe significant differences in the proportions of cis tests that are significantly associated with gene expression phenotypes (data not shown), although pooled across all markers we observe a higher proportion of significant cis tests that trans tests (Ef-induced: Fisher's Exact Test P = 0.02737, Odds Ratio = 1.99; Sm-induced: Fisher's Exact Test P = 0.08566, Odds Ratio = 1.69).
To dissect the role of crosstalk and cross-regulation between signaling pathways in the pattern of associations between gene expression and SNPs, we examined the number of significant associations between markers in signaling genes in either the Toll or Imd pathway and expression of signaling genes in other signaling pathways. On the BeadChips, we have representatives from the Toll, Imd, JAK/STAT, JNK, Ras, p38, and Notch signaling pathways. We compared the observed number of tests significant at α = 0.01 (excluding potential cis associations) to the expected number based on chance alone, using χ2 tests. For the Ef-induced sample, we observe a significant excess (over chance expectations) of associations between markers in signaling genes in the Toll pathway and induction of signaling genes in the Toll pathway (P = 1.32×10−13) and the JAK/STAT pathway (P = 3.05×10−14); we also observe an excess of significant associations between markers in signaling genes in the Imd pathway and induction of signaling genes in the Imd pathway (P = 0.00159) and the Toll pathway (P = 0.0292), although the latter is not significant after Bonferroni correction.
For the Sm-induced sample, we see a similar pattern. There is a significant excess of significant associations between markers in signaling genes in the Toll pathway and induction of signaling genes in the Toll pathway (P = 1.32×10−13), and to a lesser extent induction of signaling genes in the Imd pathway (P = 0.0341) and the JAK/STAT pathway (0.0496), although the latter two P-values do not survive a Bonferroni correction. Markers in signaling genes in the Imd pathway are significantly more likely than expected by chance to be associated with induction of signaling genes in the Imd pathway (P = 0.0219) and the JAK/STAT pathway (P = 0.00102) after infection with S. marcescens.
Because the numbers of markers in signaling genes represent a relatively limited sample, some caution should be used in interpreting these results. Nonetheless, these data suggest that, in addition to self-regulation of both the Toll and Imd signaling pathways by components of the pathway, there is some crosstalk between the Toll, Imd, and JAK/STAT pathways, although there seems to be relatively little crosstalk between either of the Toll or Imd pathways and the JNK pathway, at least at the time point we examined (8 hours after infection). Given genetic variation for flux through the pathway, these patterns of autoregulation and cross-regulation may have the effect of amplifying the phenotypic consequences of minor genetic variants.
Considerable recent interest has focused on identifying not just genetic markers that associate with quantitative variation in phenotypes, but also transcripts whose abundance correlates with phenotypes of interest [5],[10],[27]. These attempts have had mixed success, with some studies failing to find any significant correlations between transcript abundance and phenotype [e.g., 10] and others finding some evidence for significant associations [e.g., 5].
Here, we used a simple regression of the induction of immune-related transcripts against either E. faecalis bacterial load (for Ef-induced sample), S. marcescens bacterial load (for Sm-induced sample), or overall bacterial load (as measured by the first principal component from all four bacterial load measures) to attempt to detect expression-phenotype associations. In this analysis, induction correlates with bacterial load for very few transcripts. Only the induction of Attacin C and Drosocin after E. faecalis infection correlate with E. faecalis bacterial load at a false discovery rate of 10%. Induction levels after S. marcescens infection do not appear to correlate with S. marcescens load for any transcripts, although uninfected expression level of pole hole (D-Raf) associates with S. marcescens load at a FDR of 0.0035, the most significant transcriptional association in our dataset (Figure 4). Uninfected transcriptional levels of six genes (CG30088, phl, Thor, Keap1, Dif, IM1) significantly associate with a principal component measuring overall immune competence and/or general vigor, at an FDR of <10%. Interestingly, pole hole is necessary for the proliferation or survival of circulating hemocytes in D. melanogaster [28],[29] suggesting that flies with lower levels of phl transcription may have fewer hemocytes and will be less able to resist infection.
Our analysis suggests that naturally occurring variation in expression level of individual genes, measured as either induction after infection or as absolute expression in uninfected flies, is a weak predictor of bacterial load phenotypes. This result suggests that, unlike complete or nearly complete knockdowns of single genes, which can have dramatic effects on bacterial load, the differences in expression of immune genes among lines that is observed in natural populations has relatively subtle consequences. However, given the structure of the immune network, this observation may not be surprising. The immune system is a highly co-regulated system, in which small changes in expression of upstream components can be amplified among downstream genes, and multiple feedback loops provide for some measure of self-regulation of the system. Furthermore, correlated transcription of many effectors could indicate that the overall extent to which the immune system (in whole or in part) is transcriptionally activated after infection is more biologically relevant than variable levels of activation of any one gene. In order to test this hypothesis, we considered whether principal components obtained from the correlation matrix among transcriptional profiles of subsets of genes predict phenotype. As an added advantage, the method of principal components reduces the dimensionality of large datasets, improving power.
Our initial hypothesis is that the most important transcriptional determinant of phenotype is the extent to which effector proteins are induced after infection. To measure this, we initially constructed a set of principal components (PCs) from the 61 genes in our dataset with a known or putative “effector” function. These include antimicrobial peptides, components of the phenoloxidase cascade, lysozymes, putative iron-sequestration proteins, and some less-well-characterized genes such as the Turandots. For both the Sm-induced dataset and the Ef-induced dataset, the variance explained by the first principal component is substantially higher than the variance explained by any other, and so we have focused on the first PC when looking for correlations with phenotypes.
This first PC estimated from the effector genes in the Ef-induced sample is significantly positively correlated with E. faecalis bacterial load (Figure 5A; β = 74.8, F1,28 = 7.309, P = 0.01153), explaining just over 20% of the variance among lines in resistance to E. faecalis (r2 = 0. 207). This PC is dominated by negative loadings of several antimicrobial peptide genes (Mtk, DptB, AttC, Drs) and genes encoding several uncharacterized peptides known to be induced by infection (IM23, IM10, TotM, IM2, IM4, IM1). The full set of loadings is available as Table S5. Thus, this analysis suggests that genetic lines that induce antimicrobial peptides (and potentially related peptides) more strongly (i.e., have a lower PC1) sustain a lower bacterial load and thus are better able to resist infection.
We also examined the Sm-induced sample using a similar procedure. However, we do not see any correlation between the first PC from the effector genes in the Sm-induced sample and S. marcescens bacterial load (β = 7.819, F1,28 = 0.2491, P = 0.6216), despite the fact that the Sm-induced PC is quite similar to the Ef-induced PC in terms of loadings. Serratia marcescens is resistant to the antimicrobial effects of Cecropins [30], Drosocins, and Defensins [31] suggesting that this bacterium may be particularly resistant to Drosophila antimicrobial defenses and providing a plausible hypothesis for the lack of effect of variation in effector gene induction on variation in bacterial load. The resistance of S. marcescens to antimicrobial peptides may also explain the disproportionate effect of expression level of the hematopoetic gene pole hole on resistance to S. marcescens infections, as cellular immunity may be the main mechanism of D. melanogaster resistance to S. marcescens.
A major challenge of quantitative genetics in Drosophila has been to link genetic polymorphisms to phenotypes via differences in expression. In this study, we have shown correlations between transcript abundance and phenotype, as well as correlations between genotype and phenotype. To look for genotype-expression-phenotype correlations, we focused on the E. faecalis bacterial load phenotype and the Ef-induced expression sample, and asked whether any of the SNPs that have nominally significant correlations with bacterial load are also correlated with the effector induction PC1. Of the eight SNPs with at least nominal associations between genotype and phenotype (P<0.05), we find that one of them, PGRPSD_494, is also statistically associated with effector induction PC1 (Figure 5B; β = 0.0235, F1,27 = 11.4, P = 0.002237), explaining nearly 30% of the variance in this principal component (r2 = 0.297).
The PGRPSD_494 marker is a C/T polymorphism located approximately 500 bp upstream of the transcriptional start site of PGRP-SD. The T allele is associated with both a higher bacterial load after infection (Ef loadT-C = 0.6741; P = 0.02) and with lower induction of antimicrobial peptides (higher PC1; Figure 5C). PGRP-SD has been shown to have a role in the recognition of some Gram-positive bacteria, including E. faecalis [32], and our data suggest that naturally occurring variation in PGRP-SD may in fact mediate the strength of the transcriptional response to infection, and thus the ability of the fly to resist infection. This site does not appear to be significantly associated with induction or naïve expression of PGRP-SD in our data, but as mentioned previously it is associated with induction levels of 73 of the 329 immune genes we assayed. No other SNP in our dataset is associated with induction levels of more than 14 genes, and most are associated with induction levels of fewer than 10 genes.
In order to test whether there is differential activation of either the Toll or the imd signaling pathway in lines carrying alternate alleles of PGRP-SD, we selected 7 random lines carrying the T allele and 7 random lines carrying the C allele at the PGRPSD_494 marker for further study. We infected these 14 lines with E. faecalis (as described in the Methods) and then, at five time points post-infection, assayed expression of two antimicrobial peptides that are commonly used as read-outs for the two major immune signaling pathways in Drosophila: DptA for the Imd pathway, and Drs for the Toll pathway. We find that there is a significant time by allele interaction for Drs expression (Table 5), but not DptA expression (data not shown), suggesting that the dynamics of Toll pathway activation are significantly different depending on which PGRP-SD allele a given fly line carries. Specifically, we find that lines carrying the PGRPSD_494 ‘C’ allele sustain Toll pathway activation at higher levels that those carrying the PGRPSD_494 ‘T’ allele (Figure 6), consistent with the observation that fly lines carrying the ‘C’ allele both have higher expression of effectors (measured by the effector PC1 described above) and sustain lower bacterial loads. Taken together, these results suggest that allelic state at PGRP-SD has a significant impact on downstream transcript abundance via modulation of Toll pathway activation dynamics, which in turn leads to observable differences in immune phenotypes.
The pursuit of an understanding of underlying determinants of phenotypic variation in Drosophila has a long history [33]. More recently, the availability of high-throughput gene expression microarrays has generated interest in correlating variation in transcript abundance across genetic lines with differences in phenotypes [5]–[7],[10],[11]. However, datasets that include both genotype information and transcriptional variation have been rare [but see 10],[34]. In this paper, we have focused on attempting to predict immunocompetence in D. melanogaster from SNPs in candidate genes and transcript abundance, guided by the known structure of the innate immune network [1].
The strong context dependence of association test results leads us to focus on trends across functional classes of genes instead of individual statistical associations between markers and bacterial load phenotypes. We take advantage of the replication of our experiment across four different bacterial strains, as well as patterns of nominally significant associations within genes, to increase confidence in our associations. We identify two loci, both encoding proteins involved in bacterial recognition, that appear to harbor genetic variation that is consistently associated with differences in bacterial load phenotypes.
One of the these loci contains the closely linked genes GNBP1 and GNBP2. Several SNPs representing a single major haplotype are associated with differences in bacterial load after infection with both S. marcescens and P. rettgeri. It is somewhat unexpected to suggest a role for variation at GNBP1 in resistance against Gram-negative bacteria, as GNBP1 has only been shown to be involved recognizing Gram-positive bacteria and activating the Toll signaling pathway [35],[36]. However, the major haplotype spans both genes, making it impossible to determine the causal variant, and no definitive role for GNBP2 is known. Notably, however, the statistical support for an association between variation at this locus and bacterial load after P. rettgeri infection is particularly strong, and is significant even after strictly controlling family-wise error rates. The second of these loci encodes PGRP-LC, the major receptor in the Imd pathway [12], [37]–[39]. Three SNPs at this locus are associated with differences in bacterial load after infection with S. marcescens, E. faecalis, and L. lactis. The observation that variation in the Imd pathway, canonically thought to be principally involved in resistance to Gram-negative bacteria, appears to associate with differences in bacterial load after infection with Gram-positive bacteria suggests that the innate immune network is dynamic with extensive feedback, co-activation and crosstalk, consistent with previous work demonstrating synergistic activation of the immune response by the Toll and Imd pathways in combination [40]. This pattern is further demonstrated by the pattern of associations between genotype and gene expression: there are significantly more associations than expected by chance between SNPs in both the Toll and Imd pathways and signaling genes outside those pathways (particularly in the JAK/STAT pathway).
This study, combined with previous candidate-gene-based association studies between immunocompetence and polymorphisms on the second chromosome [2],[3], allows us to infer general patterns about the genetic architecture of immunocompetence in Drosophila. Most of the significant associations between SNPs and phenotype that we observe in this study are in genes encoding recognition proteins, primarily PGRPs and GNBPs, suggesting that variation in upstream components of the signaling network has substantial phenotypic consequences. Strikingly, we find a near-complete lack of significant associations, even without correcting for multiple tests, in antimicrobial peptides. In this study and in the previous studies, we genotyped 204 markers covering every known antimicrobial peptide in D. melanogaster. Only a single marker (CecC_1660), a noncoding SNP downstream of CecC, has a nominal P-value less than 0.05, and even that marker is unlikely to be a true association, as the association neither survives multiple test correction nor is observed in multiple experiments. Taken together, these studies provide convincing evidence that any functional effect of genetic variation in D. melanogaster AMP genes is far too small to be observed in experiments such as these. This observation supports the previous inference from genetic evidence that Drosophila AMPs are at least partially functionally redundant [41].
A different picture is painted when considering the effect of variation across lines in overall transcript abundance. Here, the total induction of effector genes (primarily AMPs and other induced peptides such as the Turandots) appears to correlate with bacterial load, at least after E. faecalis infection. Together, these observations suggest that while cis-acting variation in individual AMPs may be of little consequence overall for resistance to bacterial infection in D. melanogaster, the combined output of AMPs after infection is a critical determinant of resistance. Thus, genetic polymorphisms that influence expression of many downstream components of the pathway can potentially have large effects on resistance phenotypes, as appears to be the case for the PGRPSD_494 marker. We additionally note that in both the present and in our previous studies, SNP associations in candidate genes have failed to explain the entirety of the observed genetic variance. This indicates genetic variation for immunocompetence that maps to genes outside our candidate list, or to more complex (epistatic) interactions among genes.
The combination of genetic polymorphism, bacterial load phenotypes, and transcript abundance thus allows us to propose a model of the genetic architecture of immunocompetence informed by the structure of the innate immune network. Genetic variation in genes encoding proteins at the top of the network (such as recognition proteins) can be amplified by the pathway (as demonstrated by the association between variation at PGRP-SD and the Ef-induced effector PC1), leading to more significant associations with phenotype. However, genetic variation in genes encoding proteins at the bottom of the network, such as AMPs, has relatively little effect, as changes in any single effector protein do not seem to cause large enough effects on phenotype to be detectable in experiments of the scale we have performed. Since there appears to be relatively little feedback between SNPs in effector proteins and transcription of upstream genes (as demonstrated by the dearth of associations between effector SNPs and signaling gene transcripts), these effector SNPs probably have relatively little impact in trans. Overall, then, it is polymorphisms in upstream genes, and especially recognition genes, that lead to variation in abundance of effectors, and ultimately to fitness differences among lines (to the extent that resistance to infection correlates with overall fitness), while single mutations in antimicrobial peptides are likely to be of relatively little consequence.
This view of the evolutionary and fitness consequences of mutations in different components of the immune response is consistent with what is known about the evolutionary history of immune system genes in Drosophila. Population genetic and molecular evolutionary studies have suggested little evidence for adaptive evolution in antimicrobial peptides [42]–[44], which might be expected given the lack of evidence for fitness consequences attributable to segregating variation in these genes reported in this study and others [2],[3]. In contrast, we see significant evidence for adaptive evolution in upstream components of the immune system [42],[43],[45]; it is these genes that appear to harbor the population variation with the largest consequences for individual fitness. By combining expression data, genetic data, and knowledge of network structure, we can gain a much better understanding of the phenotypic consequences of genetic variation than any one component could provide alone.
We evaluated ninety-four lines of D. melanogaster for resistance to infection against each of four different bacteria. These lines are originally derived from a natural collection of wild-caught D. melanogaster from State College, PA by Anthony Fiumera. Each line in the panel is homozygous for an individual third chromosome isolated from the natural population and substituted into a common genetic background. The construction of these lines is described in more detail in Fiumera et al. [46]. The third chromosome is the largest D. melanogaster chromosome, containing about 44% of the euchromatic genome, including genes encoding proteins from all major functional classes of the immune system, and thus represents the most natural chromosome on which to focus our study. The D. melanogaster lines in this study were challenged with each of four different bacteria, two Gram-positive and two Gram-negative. The Gram-positive bacteria used are the E. faecalis and L. lactis strains described in Lazzaro et al. [2]. The Gram-negative bacteria used are the S. marcescens strain described in Lazzaro et al. [2], and Providencia rettgeri [47].
We ascertained markers to be genotyped by sequencing the complete coding region and 1–2 kb upstream of 25 candidate loci (listed in Table 3) from 8 lines. We selected loci to represent genes encoding relatively well-characterized proteins that encompass a range of immune functions. While using a candidate gene approach necessarily means that we will not sample every polymorphism that may be associated with phenotypic differences among lines, our primary goal of capturing sufficient polymorphism to test hypotheses about the role of network structure mediating genotype-phenotype associations is well served by such an approach. We assembled sequencing reads into contigs using Sequencher and manually identified SNPs and indels to assay in the full panel of 94 lines. We used three different methods for genotyping our panel of lines. Approximately half of the markers were genotyped using SNPlex (Applied Biosystems, Foster City, CA) and the remaining markers were genotyped using pyrosequencing assays, SNPstream (Beckman Coulter, Fullerton, CA), or fRLFP [48]. A small number of markers were genotyped with both SNPlex and pyrosequencing; for the rare cases where the genotype call disagreed, we used the SNPlex call. After genotyping, SNPs were filtered to produce a set of 137 usable markers (136 SNPs and 1 indel): markers with a minor allele frequency <0.05 were dropped, and only one marker (chosen at random) was kept from any pair with LD (measured by r2)>0.90. Annotation information for each SNP, including the genotyping method used to assay each SNP in the 94 lines, are presented as Table S6. Filtered genotype calls for each line are presented in Table S7. Linkage disequilibrium between each pair of genotypes is provided in Table S8.
We infected the 94 D. melanogaster lines in a complete-block design, with each line infected on each of three different days. On each day, each line was infected by one of 3 to 5 infectors at random, and a different infector infected each line on each day. Typically 2-3 replicates per line per sex were obtained on each day, for a total of 12–18 replicate data points for each D. melanogaster line. The entire experiment was repeated independently for each bacterial challenge. Flies were artificially infected by septic pinprick as described previously [2],[3]. Briefly, we pierced the thoraces of individual D. melanogaster aged 3–5 days post-eclosion with a 0.1-mm dissecting pin (Fine Science Tools, Foster City, CA) coated in liquid culture (OD600 = 1.0±0.2) of the bacterium of interest, delivering an average of 4×103 bacteria to each fly. Drosophila were maintained at 22°–24°C on a rich dextrose medium for the duration of the experiment. To measure bacterial load, we homogenized same-sex trios of flies 28 hours post-infection in 500 µl of sterile LB and then quantitatively plated the homogenates on standard LB agar plays using robotic spiral platers manufactured by Spiral Biotech (Bethesda, MD) and Don Whitley Scientific (Fredrick, MD). We incubated the plates overnight at 37°C and then estimated the concentration of viable bacteria in each homogenate using the colony counting systems associated with each plater. Prior to plating, we diluted homogenates of L. lactis 1000-fold, homogenates of P. rettgeri 100-fold, and homogenates of E. faecalis 10-fold, all in sterile LB, in order to correct for anticipated high bacterial loads. Our estimates of bacterial load per fly were transformed to correct for these dilutions before analysis. Mean bacterial load sustained by each line against each of the four bacteria is presented in Figure 1 and Table S7.
For some analyses, we generated a principal component from bacteria load line means after infection by each of the four bacteria using the prcomp() function in R. This principal component is positively correlated with load after infection with all four bacteria, suggesting it represents a common measure of immunocompetence across bacteria. However, it is also likely that this principal component captures some aspects of general vigor.
A number of recent studies have suggested that bacterial load sustained after infection and survival to infections are not strongly correlated in Drosophila melanogaster, suggesting that survival may be mediated in part by tolerance to bacterial loads [49]–[51]. In this study, we focus on resistance, as defined by bacterial load sustained 28 hours after artificial infection. Although knowledge of the molecular mechanisms that determine tolerance is increasing [52], there is not yet sufficient understanding of the underlying mechanistic basis for tolerance phenotypes to allow fruitful candidate gene association studies or to develop models based on network structure and functional attributes of candidate genes.
We selected 329 immune genes for inclusion on the custom Illumina BeadChips based on a number of criteria, including evidence for transcriptional regulation by infection in previous microarray experiments, genetic or molecular evidence for a role in immunity, and homology to known immune proteins in D. melanogaster or other organisms. The remaining 384 non-immune genes were selected either as controls or for other experimental reasons. Each gene is represented by two different probes, each of which is represented by an average of 30 beads on the array, giving an extremely high degree of technical replication. Given the number of samples assayed (as described below), we determined that genome-wide expression approaches were not practical; however, since numerous previous studies in D. melanogaster have identified a robust set of immune-regulated transcripts [15],[20],[53] we believe that a targeted expression approach represented by custom Illumina BeadChips captures the vast majority of genes whose expression is regulated by infection.
We selected a total of 30 lines for our expression analysis, biased towards the upper and lower tails of the phenotypic distribution. Males of each line were either infected with S. marcescens with E. faecalis, as described above, or left uninfected, and then frozen 8 hours after treatment. We chose to use an 8-hour post-infection timepoint as a compromise between earlier time points, where the transcriptional response to wounding could be confounding, and later time points that risked missing transcriptional events that would be relevant to bacterial load at 28 hours after infection. We extracted total RNA using Trizol (Invitrogen Corp., Carlsbad, CA) following standard protocols, then made cDNA and amplified RNA samples following the BeadChip protocol.
RNA samples were hybridized to BeadChips following standard protocols and scanned. After scanning, we normalized the data using the qspline method in the beadarray package for R. Mean probability of detection and signal intensity of control genes were used as hybridization quality control: for samples that failed to pass quality control checks, cDNA synthesis, RNA amplification, and hybridization were repeated from the original RNA extractions. Normalized induction after E. faecalis and S. marcescens infection (where induction is measured as log2 signal intensity for the infected sample minus log2 signal intensity for the uninfected sample), as well as unnormalized expression data from all treatments (Ef-infected, Sm-infected, Naïve) are presented as Dataset S1 and Dataset S2, respectively.
For quantitative PCR experiments, we sampled three replicates of 5–7 flies from each of 14 lines (7 carrying the C allele at PGRP-SD_494, 7 carrying the T allele, randomly selected) at five time points: uninfected (0 hours), 3 hours post-infection (with E. faecalis), 6 hours post-infection, 12 hours post-infection, and 24 hours post-infection. Flies were frozen in liquid nitrogen, RNA was extracted with Trizol, first strand cDNA synthesis was carried out, and qPCR TaqMan assays were run using standard protocols. We measured expression of three different genes: Drs, DptA, and Rp49. TaqMan probe and primer sequences, and reaction conditions, are available upon request from T.B.S. Data points with raw Rp49 CT values more than 1.5 times the interquartile range from the median were removed to eliminate samples with very little RNA or poor reverse transcription efficiency. Raw 1/CT values were normalized by Rp49 expression and values for each plate were mean-centered. Normalized expression of either Drs (Toll pathway) or DptA (Imd pathway) was then used as the response variable in the following second-order linear model:(1)where Y is normalized expression, Time (i = 0,3,6,12,24) is time after infection measured in hours, and PGRPSD (j = C, T) is allele at the PGRPSD_494 marker, and Line (k = 3F, 3E, 8A, 12E, 9D, 7C, 4C, 11F, 6E, 1C, 9E, 1E, 7D, 6H) is the genetic line and is treated as a random effect nested within PGRPSD. Because the response to time is not linear, we fitted a second-order model with a linear and quadratic time term, using the poly() function in R to estimate orthogonal polynomial terms.
In order to test for associations between genotype and phenotype, we analyzed the following model using the package lme4 in R 2.6.0,(2)where Y is bacterial load, Sex (i = 1,2) and Allele (j = 1,2) are main effects, and Line (k = 1,94), Day (l = 1,3), Infector (m = 1,5), and Plater (n = 1,2) are random effects. To assess significance, we compared the model coefficient for the Allele term to the null distribution obtained by permuting the genotype vector assigned to each line 5070 times and reanalyzing the data with the same model. The permutation approach was carried out as follows: for each row of the dataset, we have columns representing the four bacterial load phenotypes and the 137 genetic markers. For each permutation iteration, we randomize the phenotype vector with respect to the genotype vector, but do not shuffle relationships between among load phenotypes or among genetic markers. In this way, the permutation procedure preserves the correlations among bacterial loads and among genetic markers, but randomizes the association between genotype and phenotype.
For each permutation, we retain the estimated model coefficient (effect size), and the χ2 statistic for the test of the alternate and null (without an Allele term) model. Because the permutations shuffle the full genotype vector assigned to each line, rather than individual allele states, linkage relationships among markers are preserved in the permuted data. We use this fact to correct for linkage relationships among markers for many tests. Using the χ2 statistics from the permutated data, we can generate null distributions of P-values under the appropriate linkage conditions but assuming no significant genotype-phenotype associations.
We also use the χ2 statistics to estimate a combined probability of an association between all markers in a loci and a bacterial load phenotype. In this case, we sum the χ2 statistics for each marker in a loci for the permuted dataset, and use that distribution as a null distribution to compare the observed sum of χ2 statistics within each gene.
For our simulations to estimate the power of our experiment, we collapsed Day, Infector, and Plater terms into a single Experimental Error term, and then simulated 10,000 replicate datasets for each combination of Gram type (positive or negative), minor allele frequency (0.25 or 0.5) and Allele coefficient (0 to 1 in 0.1 increments). Each simulation assumes 3 replicates per experimental treatment (n = 3), per sex (n = 2), per line (n = 94), for a total of 18 data points per line and 1692 per simulation. This approximates our experimental conditions, with the caveat that the simulations assume no missing data and so will be an upper bound on our true power. Error terms are assumed to be normally distributed with a mean of 0 and variance equal to our estimated variance terms from Table 1, averaged across either Gram-positive or Gram-negative bacteria. To calculate power, we counted the number of tests significant at a nominal α of 0.01; significance was estimated by comparing the fit of a mixed linear model that included Line and Experimental Error as random effects and Allele and Sex as fixed effects to the fit of a null model that does not include a fixed effect of Allele.
In order to test for sex*marker interactions, we used a similar approach. In this case, we compared the likelihood of the data under the null model specified by equation (2) to likelihood of the data under the following alternative model:(3)where all terms are as described above. To assess significance, we compared the likelihood ratio test statistic obtaining by comparing the null and alternative models to the empirical null distribution of likelihood ratio test statistics obtained by analyzing 1000 permuted datasets in which the genotype vector assigned to each line was shuffled.
To test for associations between genotype and expression, we compared the likelihood of the data under the following linear model:(4)where Y is the normalized induction of a given gene (where induction is measured as log2 normalized signal intensity for the infected sample minus log2 normalized signal intensity for the control sample), Probe (j = 1,2) is a random effect representing the two probes on the array for each gene, and Allele (i = 1,2) is the fixed main effect of interest, to the likelihood of the data under the null model that retains the random effect of Probe but includes only a fixed intercept. As the number of tests is far too large for permutations to be computationally feasible, we used the anova() function in lme4 to assess the significance of the alternative model using a likelihood ratio test.
In order to test for correlations between transcript abundance and phenotype, we used two approaches. In the first approach we tested each transcript against phenotype individually, using a simple linear regression (with the model Load = Expression) and assessing significance assuming the standard null distribution for the F statistic. In the second approach, we generated principal components from a priori subsets of transcripts, using the prcomp() function in R, and then assessed the correlation between the first principal component and bacterial load using a simple linear regression.
To correct for multiple testing, we used an false-discovery-rate (FDR) and/or Holm familywise error rate control approach, as described in the Results section, implemented using the p.adjust() function in R.
|
10.1371/journal.pgen.1003619 | The C. elegans cGMP-Dependent Protein Kinase EGL-4 Regulates Nociceptive Behavioral Sensitivity | Signaling levels within sensory neurons must be tightly regulated to allow cells to integrate information from multiple signaling inputs and to respond to new stimuli. Herein we report a new role for the cGMP-dependent protein kinase EGL-4 in the negative regulation of G protein-coupled nociceptive chemosensory signaling. C. elegans lacking EGL-4 function are hypersensitive in their behavioral response to low concentrations of the bitter tastant quinine and exhibit an elevated calcium flux in the ASH sensory neurons in response to quinine. We provide the first direct evidence for cGMP/PKG function in ASH and propose that ODR-1, GCY-27, GCY-33 and GCY-34 act in a non-cell-autonomous manner to provide cGMP for EGL-4 function in ASH. Our data suggest that activated EGL-4 dampens quinine sensitivity via phosphorylation and activation of the regulator of G protein signaling (RGS) proteins RGS-2 and RGS-3, which in turn downregulate Gα signaling and behavioral sensitivity.
| All animals rely on their ability to sense and respond to their constantly changing environments to survive. C. elegans (small roundworms) depend heavily upon their ability to taste and smell chemical information in their soil environment to find food and avoid danger. While similar signal transduction pathways are implicated in both C. elegans and vertebrate chemosensation, there are still large gaps in our understanding of the mechanisms used to regulate signaling in these systems. We have identified a new role for the C. elegans cGMP-dependent protein kinase (PKG) EGL-4 in the negative regulation of nociceptive chemosensory signaling. Our data suggest that EGL-4 negatively regulates signaling and behavior by activating known inhibitors of G protein-coupled signal transduction, RGS proteins. Using C. elegans behavioral response to aversive stimuli as the readout for neuronal activity, we provide the first evidence for PKG regulation of RGS function in sensory neurons in any system.
| The ability to detect and avoid noxious stimuli in the environment is critical to an organism's survival. Nociceptive sensory systems mediate detection of harmful stimuli, allowing rapid initiation of protective behavioral responses. In the nematode Caenorhabditis elegans, the pair of polymodal nociceptive head sensory neurons termed ASH respond to a broad range of aversive stimuli, including soluble chemicals (e.g. the bitter tastant quinine, heavy metals and SDS), odorants (e.g. octanol), ions (e.g. Na+), osmotic stress and mechanosensory stimulation (nose touch) [1]–[9]. Because the ASH sensory neurons synapse onto command interneurons that drive backward locomotion via their connections with motor neurons, ASH activation elicits reversal and stimulus avoidance.
In general, olfaction and some forms of taste (including bitter) are mediated by G protein-coupled signal transduction pathways [10]–[12]. Signaling is initiated when the chemosensory ligand binds to a seven-transmembrane G protein-coupled receptor (GPCR). This induces a conformational change in the receptor that activates the associated heterotrimeric G proteins. Gα becomes active upon exchange of GDP for GTP. Once dissociated, the Gα-GTP and Gβγ subunits can activate distinct signaling cascades within the cell [13]. Although the C. elegans genome encodes >500 predicted functional chemosensory GPCRs [14], only one aversive chemical stimulus, dihydrocaffeic acid, has been functionally coupled to a receptor, DCAR-1 [15]. However, the C. elegans stimulatory Gα subunits ODR-3 and GPA-3 (both most similar to Gαi/o) are used by ASH to mediate avoidance of a variety of stimuli [7], [8], [16]–[18].
Regulator of G protein signaling (RGS) proteins are important negative regulators of G protein-coupled signal transduction. They bind to Gα-GTP and accelerate the intrinsic GTPase activity of the Gα subunits. Once GTP is hydrolyzed (back to GDP), signaling via Gα is attenuated [19], [20]. By dampening Gα signaling, RGS proteins help to protect cells from overstimulation. Mammalian RGS proteins have been implicated in the regulation of sensory signaling. For example, RGS9-1 plays an important role in regulating the light response of rod photoreceptor cells. Retinas isolated from knock-out mice lacking RGS9-1 function displayed a prolonged dim flash response [21], while overexpression of the RGS9-1 containing complex resulted in a faster light response recovery in the retina rod cells of transgenic mice [22]. In addition, RGS21 is coexpressed with T2R bitter receptors and T1R2 and T1R3 sweet receptors in rat taste bud cells [23]. RGS21 also coprecipitates with α-gustducin, the Gαi protein that is coupled to T2R bitter receptors [23]–[27]. Taken together, these observations suggest a potential role for RGS21 in the regulation of taste transduction.
C. elegans lacking RGS-3 function are defective in their response to a subset of strong sensory stimuli detected by the ASH sensory neurons [28]. Interestingly, the behavioral defects appear to be due to increased signaling in the sensory neurons that in turns leads to decreased synaptic transmission [28]. Although our previous study did not identify chemosensory hypersensitivity (e.g. enhanced sensitivity to dilute quinine) in rgs-3 mutant animals, we note that feeding status and, consequently, biogenic amine (e.g. dopamine and serotonin) levels modulate rgs-3 behavioral responses [28]. For example, rgs-3 animals responded to 100% octanol (odorant) and 10 mM quinine (tastant) when they were assayed in the presence of food (E. coli bacterial lawn), and were only defective when assayed after a short (10 minute) period of starvation [28]. Taken together, the sensitivity of C. elegans to an environmental stimulus is ultimately coordinated by proteins (e.g. RGS-3) that directly regulate the sensory G protein-coupled signaling cascade, in conjunction with signals modulated by nutritional status. All of the tastant avoidance experiments presented herein were performed 30 minutes after animals were removed from food, in contrast to the previous study focused on rgs-3, where “off food” assays were performed 10 minutes after animals were removed from the bacterial food source [28].
cGMP-dependent protein kinases (PKGs) are serine/threonine kinases that are activated upon the binding of cGMP [29], [30] and their function is implicated in a variety of cellular contexts. PKGI-null mice have altered physiological processes including impaired vasorelaxation [31]–[33] and increased platelet aggregation and activation [34] via the nitric oxide/cGMP/PKGI signaling pathway [35]. They also display impaired nociceptive flexion reflexes [36] and reduced inflammatory hyperalgesia [37]. Mice lacking PKGII function exhibit enhanced anxiety and alcohol consumption [38]. In addition, mammalian PKG can act as an indirect negative regulator of G protein-coupled signal transduction. Upon activation, PKGI phosphorylates serine/threonine residues on RGS2 in mouse vascular smooth muscle cells [39] and Rat1 fibroblast cells [40], RGS3 and RGS4 in rat diencephalic astrocytes [41] and RGS4 in rabbit gastric smooth muscle cells [42]. Phosphorylation by PKG causes each RGS protein to translocate from the cytosol to the plasma membrane, stimulating its binding to Gαq and consequently enhancing GTPase activity and the dampening of Gα signaling [39]–[42].
In C. elegans, EGL-4 is a cGMP-dependent protein kinase with physiological roles including egg laying, life span, cell growth, quiescence and dauer formation [43]–[46]. Additionally, EGL-4 function contributes to C. elegans sensory responses. For example, egl-4(lof) animals are strongly defective for chemotaxis to the AWA-detected odorant diacetyl [47]. In the AWC olfactory neurons, EGL-4 functions in the cytoplasm and nucleus to regulate short term and long term olfactory adaptation, respectively [48], [49]. EGL-4 also acts with KIN-29, a salt-inducible kinase, to regulate chemoreceptor gene expression [50], [51], contributing to behavioral plasticity and chemosensory signal-dependent development [52].
In mammalian systems, cGMP binding activates PKG [53]–[55] by relieving an inhibitory interaction that blocks the kinase domain [56]. C. elegans EGL-4 has two allosteric cGMP-binding domains in its amino-terminal half [46] and altering key residues within either or both cGMP-binding domains renders animals completely defective for AWC-mediated odor adaptation [49]. Interestingly, while EGL-4 requires intact cGMP binding domains to accumulate in the AWC nucleus [49], it requires reduction in cGMP levels to enter the AWC nucleus and promote adaptation [57]. In uterine epithelial cells, nuclear EGL-4 regulates gene expression [58]. Guanylyl cyclases (GCYs) produce cGMP, and 27 receptor-type GCYs and 7 soluble GCYs are encoded by the C. elegans genome [59]. Although several GCYs have been shown to function in C. elegans sensory neurons, no GCYs are known to be expressed in or function in ASH.
Herein we describe a new role for the C. elegans cGMP-dependent protein kinase EGL-4 as a negative regulator of nociceptive chemosensory signaling. Animals lacking EGL-4 function respond better than wild-type animals to dilute concentrations of several ASH-detected chemical stimuli, including the bitter tastant quinine. We provide evidence that a subset of GCYs function non-cell-autonomously to generate cGMP to stimulate EGL-4 phosphorylation and activation of RGS proteins in ASH to dampen G protein-coupled chemosensory signal transduction and behavioral sensitivity.
EGL-4 has not previously been implicated in nociceptive chemosensory signaling in C. elegans and egl-4(n479) loss-of-function (lof) animals effectively avoid the bitter tastant 10 mM quinine (Figure 1A). However, we noticed that egl-4(lof) animals appeared to respond to 10 mM quinine slightly better than wild-type animals (p<0.01). To determine whether EGL-4 might contribute to chemosensory signaling in a way that might not be fully revealed by this high concentration of quinine, which elicits a robust behavioral response even in wild-type animals, we challenged both genotypes with dilute (1 mM) quinine. At this lower concentration, egl-4(lof) animals were clearly hypersensitive, with significantly more egl-4(lof) animals responding to the dilute quinine than wild-type animals (Figure 1A). As loss of EGL-4 function led to quinine hypersensitivity, we reasoned that excess EGL-4 function could result in diminished quinine sensitivity. ad450 is a gain-of-function (gof) allele of egl-4. Consistent with our prediction, egl-4(gof) animals were no longer hypersensitive to quinine and responded worse than wild-type animals to 10 mM quinine (Figure 1A).
The ASH sensory neurons are the main cells used to detect quinine in C. elegans, but the ASK neurons also contribute [7]. EGL-4 is broadly expressed throughout the animal, including these sensory neurons [60]. To determine whether EGL-4 function in either the ASHs or ASKs is sufficient to regulate quinine response, the cell-selective promoters osm-10 (ASH) [3], srb-6 (ASH) [61] and srbc-66 (ASK) [62] were used to restore EGL-4 function in each neuron pair and animals were assayed for response to 1 mM quinine. Expression in ASH, but not ASK, returned the egl-4(lof) response to wild-type levels so that animals were no longer hypersensitive to the dilute stimulus (Figure 1B). To assess whether selective loss of EGL-4 function in the ASH sensory neurons could also lead to quinine hypersensitivity, we used the cell-specific RNAi approach of Esposito et al. [63]. The osm-10 [3] and srb-6 [61] promoters were used to co-express a non-coding fragment of egl-4 in both the sense and antisense orientations in the ASH neurons of otherwise wild-type animals. egl-4 knock-down using either promoter resulted in hypersensitivity to 1 mM quinine, similar to egl-4(lof) animals (Figure 1C). Taken together, our data suggest a role for EGL-4 in the negative regulation of quinine avoidance in the ASH sensory neurons.
In addition to the ASH-mediated avoidance of quinine, C. elegans avoid several additional bitter compounds [7]. We therefore tested whether EGL-4 selectively regulates quinine avoidance (Figure 1) or whether it could regulate bitter taste responses generally. We assayed the response of egl-4(lof) animals to the bitter tastants amodiaquine and primaquine (Figures 2A and 2B). egl-4(lof) animals were hypersensitive to amodiaquine, with more animals responding than wild-type across a range of concentrations (Figure 2A). However, egl-4(lof) animals responded to primaquine similarly to wild-type animals and did not appear hypersensitive (Figure 2B).
In C. elegans, as in mammals, bitter compounds signal through G protein-coupled receptor pathways [7], [10], [11]. The aversive odorant 1-octanol, which like quinine is detected primarily by the ASH sensory neurons [61], [64], also activates G protein-coupled signaling [16], [65]. To determine whether EGL-4 regulates octanol avoidance, animals were assayed for their time to reverse when presented with 100%, 30% or 10% octanol. At each concentration egl-4(lof) animals responded better than wild-type animals (Figure 2C), suggesting that EGL-4 normally dampens this olfactory response.
The ASH sensory neurons also detect soluble stimuli (copper and SDS) that are thought to signal in a G protein-independent manner [4], [6], [8]. To assess whether EGL-4 regulation of sensory signaling extends to these compounds, animals were tested for their avoidance response across a range of concentrations for each. In all cases, the response of egl-4(lof) animals was similar to that of wild-type animals (Figures 2D and 2E). We conclude that EGL-4 regulates response to a subset of G protein-coupled chemosensory responses, including the bitter tastants quinine and amodiaquine and the aversive odorant octanol, but does not regulate ASH sensitivity in general.
Upon prolonged exposure to attractive odorants, EGL-4 translocates from the cytoplasm into the nucleus of the AWC olfactory neurons to regulate long-term adaptation in a manner downstream of primary sensory transduction [49]. To determine where within the ASH sensory neurons EGL-4 might function to regulate quinine sensitivity, we used the osm-10 promoter [3] to express a functional GFP−EGL-4 fusion protein [49] and visualized its subcellular localization via GFP fluorescence. As was previously reported for naïve AWC neurons [49], [66], GFP−EGL-4 was distributed throughout ASH in wild-type animals, with expression seen in both the cytoplasm and the nucleus (Figure 3A).
To ascertain whether the cytoplasmic or nuclear pools of EGL-4 contribute to the regulation of quinine avoidance, we used the osm-10 promoter [3] to express two modified forms of GFP−EGL-4 in egl-4(lof) animals. Deletion of the endogenous nuclear localization sequence, GFP−EGL-4(ΔNLS), lead to cytoplasmic accumulation and had no effect on the ability of EGL-4 to rescue the quinine hypersensitivity of egl-4(lof) animals (Figure 3A). Conversely, egl-4(lof) animals expressing the constitutively nuclear NLS−GFP−EGL-4 remained hypersensitive in their response to 1 mM quinine (Figure 3A). Combined, these results suggest a primarily cytoplasmic role for EGL-4 in the negative regulation of quinine signaling.
Genetically encoded calcium indicators such as G-CaMP [67] and cameleon [68] can be used to monitor the activity of C. elegans neurons. In response to quinine, a transient rise in intracellular calcium levels has been seen in the ASHs of wild-type animals [8], [17], [65]. We imaged quinine-evoked calcium fluxes in wild-type, egl-4(lof) and egl-4(gof) animals expressing G-CaMP3 in ASH (sra-6p::G-CaMP3) [69], [70]. Consistent with their observed behavioral hypersensitivity (Figures 1 and 3A), egl-4(lof) animals displayed quinine-evoked ASH calcium fluxes of greater amplitude in response to 1 mM quinine, while expression of wild-type egl-4 in ASH, using the osm-10 promoter [3], rescued this elevated calcium flux (Figures 3B and 3C). Conversely, egl-4(gof) animals showed a diminished calcium flux in response to 10 mM quinine (Figures 3B and 3C), consistent with their decreased behavioral sensitivity to this concentration (Figure 1). Taken together, we suggest that the elevated calcium fluxes seen in the absence of EGL-4 function underlie the enhanced sensitivity to quinine.
PKGs have been shown to down-regulate Gαq signaling by directly phosphorylating and activating RGS proteins in mouse vascular smooth muscle cells, Rat1 fibroblast cells, rat diencephalic astrocytes, and rabbit gastric smooth muscle cells [39]–[42]. The C. elegans genome encodes twelve predicted RGS proteins [71]. To determine if the loss of RGS function also results in quinine hypersensitivity, similar to loss of EGL-4, we tested animals lacking each of the eight neuronally expressed RGS proteins for response to 1 mM quinine (Figure 4A). rgs-2(lof) and rgs-3(lof) animals displayed a statistically significant hypersensitive response, with more animals than wild-type responding to the dilute stimulus. Although an RGS-2 reporter construct did not show expression in ASH [72], RGS-3 is known to be expressed in and function in the quinine-detecting ASH sensory neurons [28]. We used the osm-10 promoter [3] to conduct cell-selective knock-down of rgs-2 and rgs-3 in the ASH sensory neurons of wild-type animals and in both cases, knock-down also resulted in hypersensitivity to dilute quinine (Figure 4B). Furthermore, consistent with a role in dampening sensory sensitivity, overexpression of RGS-2 or RGS-3 in the ASHs of wild-type animals decreased quinine response (Figure 4C).
If EGL-4 functions in the same pathway as RGS-2 and RGS-3 to regulate chemosensory signaling, then egl-4(lof);rgs-2(lof) and egl-4(lof);rgs-3(lof) double mutant animals should display a behavioral phenotype similar to egl-4(lof) animals and the hypersensitivity should not be additive. We found that the double mutant animals' response to dilute (1 mM) quinine was indistinguishable from animals lacking only EGL-4 function (Figure 4D), suggesting that they do function in the same regulatory pathway. To determine whether EGL-4 might function upstream of the RGS proteins, we utilized the egl-4(gof) animals, which display reduced sensitivity to quinine (Figure 1A). If RGS-2 and RGS-3 function downstream of EGL-4, then loss of either in combination with the egl-4(gof) allele should relieve the dampened sensitivity of egl-4(gof) mutants in response to 10 mM quinine. Indeed, egl-4(gof);rgs-2(lof) and egl-4(gof);rgs-3(lof) double mutants both responded to 10 mM quinine similarly to the rgs single mutants (Figure 4E). In addition, the double-mutant animals were hypersensitive to 1 mM quinine, similar to the rgs single mutants (Figure 4F).
The characterized consensus sequence of mammalian PKG, seen in 75% of targets surveyed, consists of (R/K)2–3-X-S*/T* [73]–[76]. A search for putative PKG phosphorylation sites in RGS-2 and RGS-3 using NetPhosK 1.0 server [77] revealed one site in each protein, serine 126 in RGS2 and serine 154 in RGS-3. Both of the predicted target serines were changed to alanine and the mutated constructs, expressed under the control of the osm-10 promoter [3], were compared to the wild-type cDNAs for their ability to rescue quinine hypersensitivity. If EGL-4 phosphorylates RGS-2 and RGS-3 to stimulate their activity, loss of the target phosphorylation site(s) in each RGS protein should preclude activation. This would result in an RGS protein that cannot rescue the hypersensitivity of the animals lacking the corresponding RGS protein. As shown in Figure 4G, ASH expression of wild-type RGS-2 or RGS-3 in the respective loss-of-function animals, using the osm-10 promoter [3], returned quinine sensitivity to the levels of wild-type animals. However, rgs-2(lof) animals expressing RGS-2(S126A) and rgs-3(lof) animals expressing RGS-3(S154A) remained hypersensitive to quinine, suggesting that these putative PKG phosphorylation sites are required for RGS-2 and RGS-3 function. We conclude that EGL-4 functions upstream of RGS-2 and RGS-3, and propose that phosphorylation by EGL-4 may be required for the function of these regulatory proteins in ASH.
A single point mutation (T276A) within the cGMP-binding domain of EGL-4 abolished its function in AWC-mediated adaptation [49]. Although no guanylyl cyclase (GCY) has been reported to be expressed in ASH, we reasoned that because EGL-4 is a PKG, cGMP binding is likely required for its function in the regulation of quinine avoidance. Although ASH expression of wild-type EGL-4 significantly rescued the egl-4(lof) quinine hypersensitivity, egl-4(lof) animals expressing EGL-4(T276A) remained hypersensitive to dilute quinine (Figure 5A). This suggests that cGMP binding is required for EGL-4 function in ASH-mediated avoidance behaviors. It also suggests that one or more GCYs may provide cGMP to regulate ASH function.
We assayed animals with loss-of-function alleles for 19 of the 34 GCYs encoded by the C. elegans genome [59] for response to dilute (1 mM) quinine (data not shown). Four GCY mutants, odr-1, gcy-27, gcy-33 and gcy-34, responded better than wild-type animals (Figure 5B). To determine whether the guanylyl cyclases ODR-1, GCY-27, GCY-33 and GCY-34 function upstream of EGL-4, we again utilized the egl-4(gof) animals, which display reduced sensitivity to 10 mM quinine (Figure 1A). A loss-of-function allele of each of the 4 GCYs was combined with the egl-4(gof) mutation and double mutant animals were assayed for quinine avoidance. Each double mutant remained less sensitive to both 10 mM and 1 mM quinine, similar to the egl-4(gof) single mutant animals (Figures 5C and 5D), supporting a role for ODR-1, GCY-27, GCY-33 and GCY-34 upstream of EGL-4 in the regulation of quinine sensitivity.
GCY expression has not previously been reported in ASH, consistent with our own analysis of odr-1, gcy-27, gcy-33 and gcy-34 GFP reporter constructs (data not shown). However, gcy-27 is expressed in the ASK head sensory neurons [59] that also contribute to quinine response [7]. To determine whether expression of any of these GCYs in ASH was sufficient to rescue the quinine hypersensitivity of the respective loss-of-function animals, we used the osm-10 [3] and srb-6 [61] promoters in cell-selective rescue experiments, as their expression overlaps in ASH. The hypersensitivity of odr-1(lof) animals was rescued by srb-6p::odr-1, but not osm-10p::odr-1 expression, and expression using neither promoter was sufficient to rescue gcy-33(lof) or gcy-34(lof) hypersensitivity (Figure 5E). Interestingly, the hypersensitivity of gcy-27(lof) animals was rescued by both osm-10p::gcy-27 and srb-6p::gcy-27, as well as by ASK-selective expression using the srbc-66 promoter [62] (Figure 5E). However, as these constructs contain gcy-27 genomic sequence, we cannot rule out the possibility that regulatory information within the introns could direct expression in addition cells beyond those predicted by the cell-selective promoters used. Taken together, and consistent with the reported expression patterns, these results suggest that the primary site of GCY function is in cells other than the ASHs, and that the cyclases may function in a non-cell-autonomous manner to provide cGMP to regulate EGL-4 function in ASH.
PKGs regulate the physiological responses of a variety of cell types, and have a wide range of known and predicted protein targets [35], [78]. As the C. elegans genome encodes only two PKGs, EGL-4 (also known as PKG-1) and PKG-2, it provides an excellent model environment in which to further elucidate the unique physiological roles of PKGs in different cellular contexts. In the AWC olfactory neurons, EGL-4 has been shown to regulate both short-term and long-term adaption. At the cellular level, adaptation is thought to serve as a protective measure against prolonged stimulation, as in the case of photoreceptor signaling for long-term light adaptation [79]. In C. elegans, short-term olfactory adaptation (<30 minutes) diminishes odorant sensitivity for up to 60 minutes and utilizes cytoplasmic EGL-4, whereas long-term adaptation to prolonged odor exposure (>80 minutes) requires the translocation of EGL-4 to the nucleus, where it can alter gene expression and decrease sensitivity to attractive scents for the lifetime of the animal (until the animal is refed) [48], [49], [66]. The cellular localization of EGL-4 depends on the levels of cGMP, as increased cGMP resulting from the loss of phosphodiesterases blocks odor-induced nuclear accumulation of EGL-4 [57], and loss of ODR-1 function leads to constitutively nuclear EGL-4 [57]. In addition, ectopic expression of constitutively nuclear EGL-4 mimicked the adapted state and decreased sensitivity to AWC-detected attractive odors [49]. However, a majority of mammalian PKG substrates are signaling proteins, suggesting an important role for PKGs in cytoplasmic cellular responses that may be closer to the initial stimulus signaling event [35], [78].
Similar to PKGI and PKGII in mammalian systems, whose targets include components of G protein-coupled signaling [39]–[42], the c-Jun N-terminal kinases (JNK) pathway [80] and the anti-apoptotic pathway that includes Bad and Akt [81], EGL-4 appears to function in or regulate multiple signal transduction pathways in C. elegans. For example, EGL-4 likely functions upstream of the TGF-β pathway SMAD transcription factors DAF-3 and DAF-5, as mutations in DAF-3 and DAF-5 suppress several egl-4 mutant phenotypes, including chemosensory, dauer formation, egg laying and body size defects [47]. In addition, EGL-4 has been implicated in C. elegans Notch signaling in sensory neurons, acting directly or indirectly downstream of Notch receptor activation. Animals lacking the DOS-motif Notch co-ligand OSM-11 do not effectively avoid 100% octanol, and this defect is partially suppressed by the gain-of-function egl-4(ad450) allele [82]. We now provide the first direct evidence for EGL-4 function in the ASH nociceptive neurons, and show that it acts to negatively regulate G protein-coupled signal transduction and dampen behavioral sensitivity to the bitter tastant quinine.
We found that egl-4(lof) animals respond better than wild-type animals to 10 mM quinine and also avoid dilute levels of quinine (1 mM) that wild-type animals do not respond to (Figure 1A). EGL-4 function in the two bilaterally symmetric ASHs was both necessary and sufficient to regulate quinine sensitivity (Figures 1B and 1C). Given the short time period that animals are given to initiate the avoidance response (four seconds), we reasoned that EGL-4's primary function in modulating quinine sensitivity is likely the regulation of signal transduction; nuclear translocation and transcriptional changes are unlikely to occur on this timescale. Indeed, in contrast to the necessary nuclear localization of EGL-4 to mediate long-term odor adaptation in AWC [49], cytoplasmic EGL-4 expression in the ASHs was sufficient to rescue the quinine hypersensitivity of egl-4(lof) animals (Figure 3A).
Calcium influx is essential for neuronal function and increased intracellular calcium levels trigger exocytosis of neurotransmitter-containing synaptic vesicles [83]. Consistent with their observed enhanced behavioral sensitivity, upon exposure to quinine egl-4(lof) animals exhibited an elevated ASH calcium flux, when compared to wild-type animals (Figures 3B and 3C). Combined, our data suggest that EGL-4 normally dampens ASH signaling. Interestingly, this is in contrast to the role of mouse PKGI in centrally located nociceptors in mice. In these cells, PKGI phosphorylation of IP3R(Ser1755) potentiates IP3R-mediated calcium release from internal stores [84]–[86]. This promotes stimulus-induced synaptic transmission between nociceptors and spinal-periaqueductal grey projection neurons, ultimately leading to the withdrawal reflex (to applied pressure, intrathecally administered NMDA or thermal stimuli) and development of hyperalgesia [84]–[88]. Thus, PKGI promotes an aversive behavioral response in mice, while EGL-4 dampens behavioral sensitivity to a nociceptive stimulus in C. elegans. This difference highlights the diversity of mechanisms by which PKGs can regulate signaling in different cellular contexts.
As the C. elegans response to quinine utilizes the Gα subunits ODR-3 and GPA-3 [7], [8], our results suggest a role for EGL-4 in the negative regulation of a G protein-coupled signal transduction pathway. We found that when C. elegans rgs-2(lof) and rgs-3(lof) animals were assayed 30 minutes after being removed from their bacterial food source, they also displayed a significant hypersensitivity phenotype (Figure 4), and RGS-2 and RGS-3 function in ASH was both necessary and sufficient to regulate quinine sensitivity. When the predicted PKG phosphorylation sites in RGS-2 and RGS-3 were changed to alanines, the mutated constructs failed to rescue quinine hypersensitivity. These findings support previous studies using mammalian smooth muscle, astrocyte and fibroblast cells [39]–[42], wherein PKGs can function as activators of RGS proteins, and provide the first evidence of this mechanism in nociceptive sensory neurons. It is also interesting to speculate that EGL-4 may target RGS proteins to regulate additional physiological processes, such as C. elegans egg laying [47], [72], [89]–[92]. We examined the amino acid sequences of all twelve C. elegans RGS proteins and identified predicted PKG phosphorylation sites in RGS-1, RGS-6, RGS-7, EGL-10 and EAT-16.
The sensitivity of C. elegans to environmental stimuli is dramatically and dynamically regulated by an animal's nutritional status, which influences signaling levels of biogenic amines such as serotonin and dopamine [14], [93]. In particular, the responses of wild-type animals to nociceptive stimuli diminish upon food deprivation [28], [64], [94]–[96]. rgs-3(lof) animals are defective in their avoidance of 10 mM quinine when assayed just 10 minutes after removal from food (Figure S1) [28]. This defective response is the result of elevated signaling in the ASHs in the absence of the negative regulator, which ultimately leads to decreased synaptic transmission [28]. We see rgs-3(lof) animals responding better than wild-type animals to quinine (10 mM and 1 mM) when they are assayed 30 minutes after removal from food (Figure 4). Similarly, rgs-2(lof) animals are defective in response to 10 mM quinine when assayed after only the short (10 minute) period of starvation (Figure S1), but are hypersensitive when assayed after being off food for 30 minutes (Figure 4). Interestingly, it is not until 45 minutes off food that rgs-2(lof);rgs-3(lof) double mutant animals reach the level of hypersensitivity seen in the single mutants at 30 minutes off food (Figure S1). It is possible that at the intermediate time off food (30 minutes) there is still “too much signaling” in the absence of both RGS-2 and RGS-3 to allow response to 1 mM quinine, similar to the elevated signaling that blocks the response of rgs-3(lof) single mutant animals to 10 mM quinine at 10 minutes off food [28]. The longer (45 minute) period off food may allow the increased signaling due to loss of both RGSs to attenuate over time, bringing it into a range that allows for the sensitized behavioral response. Because the time-course for quinine sensitivity differs between egl-4(lof) animals and rgs-2(lof);rgs-3(lof) animals, EGL-4 may have additional targets within the ASH sensory neurons. Alternatively, RGS-2 and RGS-3 may retain residual function in egl-4(lof) animals, such that loss of EGL-4 function does not affect cellular signaling as adversely as loss of the RGS proteins themselves.
PKGs require cGMP for their activation, and the levels of cGMP within a cell are modulated by production by guanlyl cyclases and breakdown by phosphodiesterases. Guanylyl cyclases are widely expressed in mammalian tissues and exist in two forms: soluble and transmembrane [97]. Mammalian soluble GCYs have been well studied in smooth muscle cells, where they are activated by nitric oxide (NO) to produce cGMP [98], [99]. Transmembrane GCYs have been well characterized for their role in natriuresis and phototransduction [100], [101]. While 34 GCYs are encoded by the C. elegans genome, the physiological roles of most are unknown. Our analysis revealed that EGL-4 requires cGMP binding in order to negatively regulate quinine sensitivity (Figure 5A), suggesting that a pool of cGMP is available to activate EGL-4 in the ASH nociceptive neurons. Moreover, animals lacking the function of the transmembrane guanylyl cyclases ODR-1 and GCY-27, or the soluble guanylyl cyclases GCY-33 and GCY-34, are hypersensitive in their response to dilute quinine (Figure 5B). However, these GCYs do not appear to function directly in the ASHs (Figure 6), suggesting that other neurons in the circuit may provide the cGMP that regulates ASH function, perhaps via GAP junctions between cGMP-generating neurons and ASH. For example, gcy-27 is expressed in ASK [59], which forms GAP junctions directly with ASH. Continued studies in C. elegans may yield new insights into nociceptive signaling in mammalian systems, where PKGI is known to function in central nociceptors [84]–[86].
Strains were maintained under standard conditions on NGM agar plates seeded with OP50 E. coli bacteria [102].
Strains used in this study include: N2 Bristol wild-type, MT1074 egl-4(n479), DA521 egl-4(ad450), CX10979 kyEx2865 (sra-6p::G-CaMP3;ofm-1p::gfp), FG414 (kyEx2865 sra-6p::G-CaMP3;ofm-1p::gfp), FG417 egl-4(n479);kyEx2865 (sra-6p::G-CaMP3;ofm-1p::gfp), FG413 egl-4(ad450);kyEx2865 (sra-6p::G-CaMP3;ofm-1p::gfp), FG454 egl-4(n479);udEx208 (osm-10p::egl-4;myo-3p::mCherry);kyEx2865 (sra-6p::G-CaMP3;ofm-1p::gfp), LX147 rgs-1(nr2017), LX160 rgs-2(vs17), LX242 rgs-3(vs19), LX533 rgs-6(vs62), FG105 rgs-10(ok1039), FG108 rgs-10/11(vs109), MT8504 egl-10(md176), LX1226 eat-16(tm761), FG329 egl-4(n479);rgs-2(vs17), FG330 rgs-3(vs19);egl-4(n479), FG275 egl-4(ad450);rgs-2(vs17), FG276 rgs-3(vs19);egl-4(ad450), FG269 rgs-3(vs19);rgs-2(vs17), FG376 rgs-3(vs19);egl-4(n479);rgs-2(vs17), VC2796 gcy-3(gk1154), RB1010 gcy-5(ok930), IK800 gcy-8(oy44), VC2675 gcy-15(gk1102), VC2450 gcy-17(gk1155), VC2321 gcy-18(ok3047);nT1[q1351], RB1909 gcy-19(ok2472), RB1935 gcy-20(ok2538), RB924 gcy-23(ok797), VC2375 gcy-25(gk1187), CZ3714 gcy-31(ok296), RB1048 gcy-32(ok995), AX1295 gcy-35(ok769), RB626 gcy-37(ok384), IK597 gcy-23(nj37);gcy-8(oy44);gcy-18(nj38), FG290 odr-1(n1936), FG280 gcy-27(ok3653), FG278 gcy-33(ok232), FG279 gcy-34(ok2953), FG294 egl-4(ad450);odr-1(n1936), FG288 gcy-27(ok3653);egl-4(ad450), FG289 egl-4(ad450);gcy-33(ok232) and FG285 egl-4(ad450);gcy-34(ok2953).
Germline transformations were performed as previously described [103]. For egl-4, rgs-2, rgs-3, odr-1, gcy-27, gcy-33, and gcy-34 rescue experiments, 50 ng/µl of pJM67 elt-2::gfp plasmid [104] was used as the co-injection marker, along with 50 ng/µl of the rescuing plasmid. For rgs-2 and rgs-3 overexpression, 250 ng/µl of the rgs plasmid was co-injected with 50 ng/µl of pJM67 elt-2::gfp plasmid [104]. For egl-4 rescue with G-CaMP3 expression, 5 ng/µl of myo-3p::mCherry plasmid (Yamamoto Lab) was used as the co-injection marker, along with 50 ng/µl of the osm-10p::egl-4 rescuing plasmid. Animals expressing the egl-4 rescuing array (udEx208) were then crossed with kyEx2865-expressing (sra-6p::G-CaMP3;ofm-1p::gfp) animals, and egl-4(n479) animals co-expressing both arrays were isolated. Cell-specific RNAi knock-down experiments were performed as previously described [63]. 25 ng/µl of pJM67 elt-2::gfp plasmid [104] was co-injected with 50 ng/µl of each PCR fusion product [63].
Neuronal calcium changes were recorded using the GFP-based fluorescent calcium reporter G-CaMP3 [69], [70] and based on described methods [109], [110]. Briefly, a microfluidic device similar to that used by [110] was fabricated by the Stanford Microfluidics Foundry and used to immobilize a kyEx2865 (sra-6p::G-CaMP3;ofm-1p::gfp)-expressing worm for imaging while streams of buffer or quinine (or buffer for controls) under laminar flow were alternatively presented to the nose of the worm. To reduce the influence of neuronal activation by blue light, we pre-exposed the worm to blue light for 5–8 seconds. The worm was imaged one minute later using a 40X air objective on an inverted Axiovert 200 microscope (Zeiss, Oberkochen, Germany). Images were captured with an exposure time of 20 milliseconds every 500 milliseconds with an ORCA-Flash 2.8 camera (Hamamatsu, Shizuoka Pref., Japan) and recorded over 15 seconds with µManager software [111]. The movies were analyzed using ImageJ (Rasband, W.S., ImageJ, US NIH, Bethesda, MD, USA). The mean signal intensities of the region of interest (ROI) and background were determined and the background corrected value taken for further analysis. The percent change in fluorescence intensity based on the average intensity of the first 3 frames (delta F/F0) was calculated. The ROI was centered on the cell body of the ASH neuron. After 5 seconds of imaging with M13 buffer (pH 7.4), the quinine stimulus was given for the last 10 seconds of imaging. Quinine was dissolved as a 1 or 10 mM solution in M13 and the pH was adjusted to 7.4 after dissolving. Ten worms for each condition were tested and averaged. A one-tailed unpaired Student's t-Test for each point in all conditions was performed. The maximum change in fluorescence before and after exposure to quinine was determined and used for statistics.
Well-fed young adults were used for analysis, and all behavioral assays were performed on at least three separate days, in parallel with controls. The number of transgenic animals assayed in each experiment is indicated within the figure legends, and in all cases n≥58 for non-transgenic animals. Response to the soluble tastant quinine was scored as the percentage of animals that initiated backward locomotion within 4 seconds of encountering a quinine drop placed on the agar plate in front of a forward moving animal [6], [7], [65]. Quinine was dissolved in M13, pH 7.4 [112]. For quinine avoidance assays, animals were tested 30 minutes after transfer to NGM plates lacking bacteria (“off food”). Response to octanol was scored as the amount of time it took an animal to initiate backward locomotion when presented with a hair dipped in octanol [3], [61]. For octanol avoidance assays, animals were tested 10–20 minutes after transfer to NGM plates lacking bacteria and assays were stopped at 20 seconds. All data is presented as ± standard error of the mean (SEM). The Student's two-tailed t-Test was used for statistical analysis, except for panels 4A and 5B, in which the one-way Anova with Tukey's Honestly Significant Difference (HSD) statistical analysis was used.
|
10.1371/journal.ppat.1002686 | Transmitted/Founder and Chronic Subtype C HIV-1 Use CD4 and CCR5 Receptors with Equal Efficiency and Are Not Inhibited by Blocking the Integrin α4β7 | Sexual transmission of human immunodeficiency virus type 1 (HIV-1) most often results from productive infection by a single transmitted/founder (T/F) virus, indicating a stringent mucosal bottleneck. Understanding the viral traits that overcome this bottleneck could have important implications for HIV-1 vaccine design and other prevention strategies. Most T/F viruses use CCR5 to infect target cells and some encode envelope glycoproteins (Envs) that contain fewer potential N-linked glycosylation sites and shorter V1/V2 variable loops than Envs from chronic viruses. Moreover, it has been reported that the gp120 subunits of certain transmitted Envs bind to the gut-homing integrin α4β7, possibly enhancing virus entry and cell-to-cell spread. Here we sought to determine whether subtype C T/F viruses, which are responsible for the majority of new HIV-1 infections worldwide, share biological properties that increase their transmission fitness, including preferential α4β7 engagement. Using single genome amplification, we generated panels of both T/F (n = 20) and chronic (n = 20) Env constructs as well as full-length T/F (n = 6) and chronic (n = 4) infectious molecular clones (IMCs). We found that T/F and chronic control Envs were indistinguishable in the efficiency with which they used CD4 and CCR5. Both groups of Envs also exhibited the same CD4+ T cell subset tropism and showed similar sensitivity to neutralization by CD4 binding site (CD4bs) antibodies. Finally, saturating concentrations of anti-α4β7 antibodies failed to inhibit infection and replication of T/F as well as chronic control viruses, although the growth of the tissue culture-adapted strain SF162 was modestly impaired. These results indicate that the population bottleneck associated with mucosal HIV-1 acquisition is not due to the selection of T/F viruses that use α4β7, CD4 or CCR5 more efficiently.
| Most new HIV-1 infections worldwide are caused by the sexual transmission of subtype C viruses, which are prevalent in Asia and southern Africa. While chronically infected individuals harbor a genetically diverse set of viruses, most new infections are established by single variants, termed transmitted/founder (T/F) viruses. This raises the question whether certain viral variants have particular properties allowing them to more efficiently overcome the transmission bottleneck. Preferential binding of the viral envelope (Env) to the integrin α4β7 has been hypothesized as one important feature of transmitted viruses. Here, we compared Envs from subtype C viruses that were transmitted to those that were prevalent in chronic infections for efficiency in utilizing α4β7, CD4 and CCR5 for cell entry and replication. We found that transmitted and chronic Envs engaged CD4 and CCR5 with equal efficiency, and that blocking the interaction between Env and α4β7 failed to inhibit replication of T/F as well as control viruses. While the search for determinants of transmission fitness remains an important goal, preferential CD4, CCR5 or α4β7 interactions do not appear to represent distinguishing features of T/F viruses.
| Mucosal transmission of HIV-1 is most often caused by a single variant from amongst the complex viral quasispecies in the infected donor [1]–[8]. After an eclipse phase of approximately two weeks during which virus is generally not detected in the blood, the progeny of this transmitted/founder (T/F) virus give rise to a productive systemic infection [9]–[15]. At a minimum, this significant population bottleneck selects for replication competent viruses, most of which use CCR5 as a coreceptor, since viruses that exclusively use CXCR4 are rarely transmitted [10], [16]. Whether other phenotypic traits are associated with enhanced mucosal transmission remains uncertain, though addressing this question is of importance because T/F viruses are the targets of vaccines, microbicides, and pre- and post-exposure prophylaxis.
Characterization of T/F virus properties is complicated by the challenges inherent in identifying acutely infected individuals, generating bona fide T/F molecular clones, procuring appropriate control viruses, obtaining sufficient numbers of samples to perform meaningful comparisons, and developing sufficiently sensitive in vitro assays to detect phenotypic differences that could impact transmission fitness in vivo. Almost all studies examining viral properties associated with mucosal transmission have focused on the viral envelope (Env) glycoprotein, most often in the context of viral pseudotypes [10], [17]–[20]. In addition, most initial studies examined viruses obtained weeks to months after infection from relatively few transmission events [21]–[23]. Given the rapidity with which HIV evolves in the face of immune pressures [24], “early” isolates could differ in important ways from true T/F viruses. Nonetheless, analyses of single genome amplification (SGA) derived T/F Env proteins and viruses have shown that mucosal transmission is associated with CD4+ T cell tropism and CCR5 use [10], [20], [25], [26] as well as a variety of signatures in the viral env gene [21], [27]–[32]. These include shorter variable loops, fewer potential N-linked glycosylation sites (PNGs) and, in some cases, enhanced sensitivity to neutralization by CD4 binding site (CD4bs) monoclonal antibodies (mAbs) [20]. More recently, it has been shown that the gp120 subunit of some Env glycoproteins can bind to, and signal through, the integrin α4β7 that is expressed on activated CD4+ T cells in the gut mucosa [33]–[35]. These findings have been taken to suggest that these interactions play an important role early in sexual transmission of HIV-1 [35], [36]. Specifically, it has been hypothesized that genetic signatures associated with transmission of certain subtype A and C viruses, including the absence of some PNGs in V1/V2 and C3/V4 regions, reflect selection for Envs that exhibit strong α4β7 binding and thus increased transmission fitness [35].
To explore the role of α4β7 interactions and other Env properties that might impact mucosal transmission, we employed SGA to generate a panel of T/F (n = 20) and chronic control (n = 20) Env constructs from geographically-matched individuals infected with subtype C viruses, the most prevalent HIV-1 lineage worldwide. To examine Env phenotypes in the context of replication competent viruses, we also produced full-length infectious molecular clones (IMCs) for six T/F and four chronic subtype C strains. Testing their biological activity in a variety of functional assays, we found no differences in the efficiency with which T/F and chronic Envs utilized CD4 or CCR5, mediated infection of primary CD4+ T cell subsets, or were neutralized by mAbs targeting the CD4bs. We confirmed that infection of α4β7-expressing CD4+ T cells by the prototypic subtype B strain HIV-1/SF162 could be partially inhibited by antibodies to α4β7 under some conditions as previously described [33], [34]. However, saturating concentrations of α4β7 antibodies had no inhibitory effect on infection of all-trans retinoic acid (atRA) stimulated CD4+ T cells from multiple donors by any of the T/F or chronic control viruses, even though most of their gp120 subunits are predicted to bind this integrin pair based on previously identified genetic signatures [35]. These findings indicate that the ability of some gp120 proteins to engage α4β7 may not be recapitulated by their native Env trimers on the surface of infectious particles, and thus suggests that interaction with this integrin pair is not critical for mucosal HIV-1 transmission.
Previous studies of T/F phenotypes focused almost exclusively on HIV-1 subtype B [10], [20], [37], [38]. To examine the extent to which these results are applicable to other subtypes, we focused in this study on the transmission properties of HIV-1 subtype C. To assess viral entry, we assembled a panel of 20 T/F Env clones, six of which have previously been described [39]. The remaining 14 clones were derived from 13 acutely infected individuals from South Africa and Zambia (8 males, 5 females) - nine of whom were sampled during the earliest stages of viral infection (Fiebig stages I and II [40]; Table 1). Plasma viral RNA was extracted, subjected to SGA and direct amplicon sequencing, and used to infer the T/F env sequences as previously described [12]. Consistent with earlier findings, infection was established by one or a limited number of viral variants. Of the 18 acutely infected individuals included in this panel, 14 acquired a single variant, while three others were infected with two variants and one was infected by four variants (Table 1 and Figure S1).
To generate an appropriate control group, we obtained 20 Env clones from individuals chronically infected with subtype C viruses (Table 1). Seven of these have previously been described [41], [42]. The remaining 13 were generated from chronically infected individuals (11 females, 2 males) enrolled in the CHAVI 001 cohort [43]. While T/F Envs were derived from individuals of both sexes (10 males; 10 females), chronic Envs were predominantly derived from female subjects (2 males; 18 females). To increase the probability of identifying functional env genes, we used SGA to generate up to 42 env gene sequences for each chronically infected individual (Figures S2 and S3). We then constructed phylogenetic trees to identify viruses that had undergone a recent clonal expansion as evidenced by clusters or “rakes” of closely related sequences (Figures 1, S2 and S3). We reasoned that the common ancestor of such clonally expanded “rakes” would be more likely to encode a fully functional env gene than a sequence chosen at random from the quasispecies. To approximate this ancestor, we cloned env amplicons whose sequences were either identical to the consensus sequence of the corresponding rake (n = 5) or encoded an Env that differed in a single amino acid residue (n = 2). For subjects from whom none of the env amplicons met these criteria, the rake consensus sequence was inferred and chemically synthesized (n = 6). This same approach had also been employed to generate the previously reported chronic Env constructs [41], [44]. Thus, all 20 chronic control Envs used in this study were derived from clonally expanded viruses.
Virus pseudotypes containing a luciferase reporter gene and bearing one of the T/F or chronic control Envs were produced in human 293T cells and then diluted serially on NP2/CD4/CCR5 and NP2/CD4/CXCR4 cells to assess coreceptor usage and to determine the linear range of the assay. All viral pseudotypes were functional, leading to infection of NP2/CD4/CCR5 cells at least 100-fold above Env-negative particles. In contrast, none of the T/F Envs and only one chronic control Env (4707.E1) mediated entry into NP2/CD4/CXCR4 cells at levels 10-fold above background. However, this Env did not mediate entry of GFP-encoding pseudoviruses into primary CD4+ T cells in the presence of the CCR5 antagonist maraviroc (data not shown). Thus, the small amount of CXCR4-dependent infection seen in NP2/CD4/CXCR4 cells is likely due to the over-expression of this coreceptor and does not reflect CXCR4 use on primary cells. Importantly, all T/F and chronic Env constructs were functional, thus validating our methods to correctly infer T/F as well as clonally expanded chronic control viruses.
Enhanced utilization of CD4 and CCR5 could influence virus transmission since changes in CD4 and CCR5 expression have been shown to impact infection by different HIV-1 strains [45]–[49]. A previous study of subtype B T/F and chronic viral Envs did not reveal differences in their utilization of CCR5 [20]. To determine the efficiency with which the newly derived subtype C T/F Envs utilized CD4 relative to the chronic Env controls, we compared their ability to infect affinofile cells, a 293T cell line that expresses CD4 and CCR5 under independently inducible promoters [50] (Figure 2). We induced CCR5 to maximal levels and induced CD4 to high or low expression levels (relative to primary human CD4+ T cells; Figure 2B) prior to infection. Infection levels of each pseudovirus in the CD4-low cells were then expressed relative to the values obtained in the CD4-high cells (Figure 2A). The macrophage-tropic JR-FL Env, which is known to mediate efficient entry into cells expressing low levels of CD4 [51], [52], was used as a control. Using this system, we found that virus pseudotypes expressing T/F and clonally expanded chronic Envs utilized CD4 with similar efficiency, while CD4 use by JR-FL was 10-fold more efficient than most of the other pseudoviruses (Figure 2A). These results demonstrated that the affinofile system is sufficiently sensitive to detect differences in CD4 utilization amongst different virus strains. Additionally, the results confirmed earlier studies of subtype B and C viruses, which indicated that the ability to use limiting levels of CD4 is not a major determinant of transmission fitness [18], [37].
To assess the efficiency of CCR5 use, we infected NP2/CD4/CCR5 cells in the presence of increasing concentrations of the CCR5 antagonist maraviroc and measured the IC50 value for each virus. We chose this approach over the use of affinofile cells since we have found that CCR5 expression levels cannot be controlled with sufficient precision at intermediate concentrations of the inducing reagent. Moreover, maraviroc titration should impact CCR5 availability to the same degree on all cells in the population. Therefore, maraviroc sensitivity is a surrogate for the efficiency of CCR5 use, provided that none of the Envs tested can use CCR5 when it is bound to maraviroc [53], [54]. This was true of our Env panel; all T/F and chronic Envs examined were equally sensitive to saturating concentrations of maraviroc with maximal percent inhibitions of >95%. Additionally, T/F and chronic Envs exhibited similar maraviroc IC50 values (median T/F = 2.22 nM; chronic = 1.67 nM; p = 0.45). Thus, enhanced CCR5 utilization efficiency does not account for the profound transmission bottleneck of both subtype B [20] and C (Figure 2) [18] infections.
CD4+ T cell subsets have different activation and coreceptor expression levels, and thus may be differentially susceptible to infection by T/F versus chronic control viruses [55]–[57]. Effector memory (TEM) and effector memory RA+ (TEMRA) cells predominate in mucosal effector sites where the transmission bottleneck likely occurs, while central memory (TCM) and naïve cells are more common in lymph nodes [58], [59]. Therefore, an enhanced ability to infect TEM and TEMRA cells could be linked to enhanced mucosal transmission. To explore this, we infected primary CD4+ T cells with GFP-expressing pseudoviruses and then stained for CCR7 and CD45RO to define naïve (CCR7+CD45RO−), TCM (CCR7+CD45RO+), TEM (CCR7−CD45RO+), and TEMRA (CCR7−CD45RO−) cells. As shown in Figure 3, we saw no differences in the abilities of subtype C T/F and chronic control Envs to mediate entry into these CD4+ T cell subsets. Similar to our observations for subtype B T/F and control Envs [20], most infected cells were TEM, but T/F Envs showed no preference for this cell type relative to chronic control Envs.
We recently reported that subtype B T/F Envs were more sensitive than chronic Envs to the CD4bs mAbs b12 and VRC01, and that this was attributable to increased binding of these antibodies to the native Env trimer [20]. To determine whether this was also true for subtype C, we performed neutralization assays with the same antibodies. No significant differences in neutralization sensitivity to b12, VRC01, PG9 or PG16 were observed for T/F and chronic control Envs (Figure S4). As expected, VRC01 generally neutralized subtype C Envs more potently than b12 [60]. Using 10 µg/ml of b12, only five Envs were inhibited by 50%, and the most sensitive Env was inhibited by 83%. Using the same concentration of VRC01, 25 Envs were inhibited by 83%, and 16 Envs were inhibited by 95%. Nonetheless, we noted a relationship between the sensitivity to both VRC01 and b12 and the efficiency of CD4 use. When Envs were divided into those that used CD4 most efficiently (top 50% regardless of whether they represented T/F or chronic controls) and those that used CD4 least efficiently, the Envs that used CD4 efficiently were more sensitive to CD4bs (Figure 4). In contrast, no relationship between CD4-use efficiency and neutralization sensitivity was observed for PG9, PG16, or purified immunoglobulin pooled from five individuals infected with subtype C HIV-1 (data not shown). Thus, subtype C T/F and chronic Envs used CD4 with similar efficiency, and there was a correlation between CD4 utilization and sensitivity to neutralization by CD4bs mAbs.
The gut-homing integrin α4β7 is expressed on activated CD4+ T cells in the gut [61], [62] and vaginal mucosa [63] and has been shown to bind the gp120 proteins from several recently transmitted subtype A and subtype C viruses [35]. In contrast, gp120 proteins from chronic viruses appear to bind α4β7 only rarely, although the subtype B HIV-1/SF162 strain is a notable exception [35]. Only a few studies have examined the effect of Env-α4β7 interactions on virus replication [33], [34], [64]. It has been shown that mAbs specific for α4β7 partially and transiently inhibit infection of α4β7-positive CD4+ T cells by HIV-1/SF162 at low inocula. Based on these studies, it has been suggested that engagement of α4β7 may enhance HIV-1 infection, especially in the context of mucosal transmission [35], [36].
If the ability of gp120 proteins to bind α4β7 is recapitulated by Env molecules present on virus particles, we reasoned that α4β7 engagement should enhance virus entry, especially at low multiplicities of infection, since binding to the cell surface is a rate-limiting step of virus infection in vitro [65], [66]. If so, then saturating levels of mAbs to α4β7 should suppress virus infection, as has been shown for the subtype B virus strain HIV-1/SF162 [33], [34]. To investigate this, we used a protocol previously developed by Arthos and colleagues in which human CD4+ T cells were stimulated with IL-2 (20 IU/ml), anti-CD3 (1.5 µg/ml) and atRA (10 nM). Under these growth conditions, α4β7 expression was enhanced and detected on 15–65% (median 32%) of CD4+ T cells, predominantly on effector memory cells, from six different donors (data not shown). One donor was non-responsive to atRA and expressed α4β7 on only five to six percent of CD4+ T cells at day six, so these cells were not used for subsequent experiments.
We next titrated two commercially available α4β7 mAbs, Act1 (specific for the α4β7 heterodimer) and 2B4 (specific for α4), both of which have been shown to inhibit gp120 binding and to suppress infection of atRA-treated CD4+ T cells by the laboratory adapted HIV-1/SF162 strain, using concentrations previously reported to be saturating for Act1 [33]. To determine if our infection and inhibition conditions were sufficiently sensitive, we used GFP reporter-expressing SF162 Env-containing pseudovirus as the positive control. Pseudovirus expressing the JR-FL Env served as the negative control, since the JR-FL gp120 does not bind α4β7 [35]. We failed to detect any inhibition of infection by either pseudovirus at a broad range of inocula using saturating concentrations of Act1 (Figure 5). In fact, Act1 treatment enhanced infection of SF162, JR-FL and VSV-G pseudoviruses by approximately 30% in cells from two different donors. However, gp120-α4β7 binding has recently been described to be critically dependent on Env glycosylation, with high mannose carbohydrates enhancing and complex glycans reducing α4β7 interactions [35]. We therefore reasoned that viruses derived from primary CD4+ T cells would be physiologically more relevant, since these cells produce Env proteins with predominantly high-mannose carbohydrates that support α4β7 binding [67].
To examine the impact of α4β7 blockade on the infectivity and growth kinetics of replication competent viruses, we generated full-length subtype C IMCs representing T/F (n = 6) and chronic control (n = 4) viruses (Table 2 and Figure S3). All but three of these had the LDI/V tripeptide motif in the V2 loop, which has been shown to play a key role in gp120-α4β7 binding (Figure S5) [33]–[35]. Moreover, the number of N-linked glycosylation sites in the V1/V2 region of these IMCs (range from 3 to 8) was comparable to that in gp120 proteins known to interact with α4β7 (range 3 to 9 for the strains SF162, 205F, QA203, and CAP88 [35]). Finally, all replication competent virus stocks were produced in primary human CD4+ T cells to ensure physiologically relevant Env glycosylation, processing and virion incorporation. Using these reagents, we infected atRA-treated primary CD4+ T cells with each virus strain, using a wide range (100-fold) of inocula in the presence of saturating concentrations of Act1. In three independent experiments, we found that Act1 consistently inhibited replication of an SF162 Env-containing molecular clone (NL4-3-SF162, gift from J. Arthos) at six days post-infection. This inhibition was greatest at the lowest multiplicity of infection (Figure 6 A and D). We also observed significant Act1-mediated inhibition of an NL4-3 construct that encoded the subtype B Env R3A [68], but again this was seen only at the two lowest virus inputs (Figure 6B). No inhibition of infection and replication was observed for YU-2 (Figure 6C), which expresses a gp120 that does not bind α4β7 [35].
If mucosal transmission selects for viruses that interact with α4β7, we reasoned that the replication of T/F IMCs would be inhibited when the Env-α4β7 interaction was blocked. Our subtype C IMC infection assays were powered to detect a 30% or higher decrease in virus (p24 antigen) production on day six, a time point when the largest effect on virus growth following α4β7 blockade was observed in previous experiments [34]. Two to four ELISA measurements were performed to monitor virus production in each of six replicate wells infected at different multiplicities using CD4+ T cells from three different donors. At the lowest virus inoculum, replication was undetectable in one to four of the six replicate wells from each of the ten subtype C viruses, indicating that virus was added at limiting dilution. Using Act1 at saturating concentrations, we observed no significant inhibition of replication of any subtype C virus (T/F or chronic) at any viral inoculum or time point post-infection (Figure 7), while replication of the positive control SF162 was reduced. In addition to Act1, we tested the α4 integrin-specific mAb 2B4 using SF162 and three of the subtype C viruses in CD4+ T cells from two donors and obtained similar results: a modest and transient inhibition of SF162, but no inhibition of the other viruses (data not shown). Finally, we tested seven subtype B T/F IMCs [37] using cells from a single donor and again observed no inhibition using the anti-α4β7 mAb Act1 (data not shown). Taken together, these results indicate that blocking the integrin α4β7 does not reduce the replication of T/F and chronic subtype C viruses in atRA-stimulated primary CD4+ T cells.
While Act1 failed to inhibit infection and/or replication by any of 13 T/F and four chronic control viruses, it significantly increased p24 production of five of the ten subtype C viruses (two of six T/F and three four chronic viruses; Figure 7). This was observed at multiple time points and with multiple viral inputs. To determine whether antibody binding to α4β7 could lead to enhanced cellular activation [69] and a resulting increase in virus production, we examined the expression of cellular activation markers. We found that neither Act1 nor 2B4 increased the expression of CD25, HLA-DR, Ki67, and CD69 at 1 hour, 2 days, and 5 days post-treatment, nor did these antibodies lead to an increase in CCR5 or α4β7 expression levels (Figure S6). However, we noted increased clumping of cells in both Act1 and 2B4 treated cultures [70], which raised the possibility that the enhanced virus production seen in some Act1-treated cultures could be due to increased cell-to-cell viral spread. We thus used high-speed cell imaging to examine cell-cell conjugates in CD4+ T cell cultures from a single donor that were infected with SF162 as well as three subtype C viruses (2 T/F and 1 chronic control) after treatment with Act1, 2B4 or a murine IgG1 isotype-control. Neither Act1 nor 2B4 increased expression of the high-affinity form of LFA-1, which is known to be upregulated by α4β7 engagement of gp120 [33]. However, we noted a significant increase in the expression of the cell-cell adhesion molecule ICAM-1 in cells exposed to 2B4 or Act1 compared to murine IgG1 (Figure S7A). Consistent with increased cell-cell spread, more cell-cell conjugates were virus-positive than predicted, with doublets being more than twice as frequently infected as singlets (p = 0.04) and triplets being more than three times as frequently infected as singlets (p = 0.06) (Figure S7B). Overall, these results suggest that Act1- and 2B4-mediated increases in cell-cell conjugates could facilitate more efficient spread and replication of some viruses in the absence of increased cellular activation.
The identification of viral traits that might enhance mucosal transmission is an important goal for vaccine development and other prevention strategies. A first step in characterizing such traits is the identification of T/F viruses, while a second step is the selection of appropriate chronic controls. Virological traits that are strongly associated with transmission, such as CCR5 use, should be readily identifiable when comparing T/F viruses to virtually any control group, while identifying more subtle phenotypes will greatly depend on the choice of control viruses, perhaps explaining discrepancies in genetic and phenotypic transmission signatures identified by different groups [17], [18], [21], [23], [27]–[30], [32]. Finally, the use of in vitro assays that recapitulate key steps in mucosal transmission are needed to identify properties unique to T/F viruses. Here, we have compiled a relatively large panel of both Envs and IMCs representing subtype C T/F and chronic control viruses, and developed a series of infection assays using virus pseudotypes, replication competent viruses, cell lines and primary human CD4+ T cells to improve our ability to identify viral phenotypes associated with transmission.
Env glycoproteins of HIV-1 can differ significantly in the efficiencies with which they utilize CD4 and the viral coreceptors, which in turn can impact viral tropism [45]–[49]. Given the variability in expression levels of entry cofactors on different cell subsets as well as between individuals [71]–[73], it is easy to envision several ways in which Env function could impact transmission efficiency at the level of virus entry. To date, several genetic Env signatures have been reported, with more compact variable loop structures and fewer PNGs being the most frequent findings [27]–[30], [32]. It is possible that these genetic traits impact Env function in ways that increase transmission fitness. However, to date no consistent T/F phenotype has been described. Thus, it is possible that mucosal transmission is a stochastic event where any reasonably functional R5 or dual tropic Env can initiate a productive infection [74]. However, it is also possible that the in vitro assays employed thus far have failed to reveal subtler or more transmission-specific phenotypic differences. The recent finding that some gp120 proteins from early HIV-1 infections can bind to the α4β7 integrin is consistent with this, although the ability of T/F viruses to productively interact with α4β7 was not explored.
To determine whether subtype C T/F viruses, which account for the great majority of new infections worldwide, utilize CD4 or CCR5 with enhanced efficiency, we tested both T/F and chronic Env constructs in pseudotyping assays. Consistent with previous results for subtype B [20] and subtype C [18], we failed to observe differences in both CCR5 and CD4 utilization. This is in contrast to findings by Etemad and colleagues who reported enhanced CCR5 utilization by Envs from individuals with chronic subtype A infection, although only V1–V5 fragments were tested in the context of chimeric viruses [23]. Similarly, Nawaz and colleagues found that gp120s from three subjects acutely infected with subtype A and C viruses bound to dodecameric but not monomeric CD4, while gp120s from subsequent time points of two of the same subjects bound to CD4 in both forms, suggesting an increase in CD4 affinity in later stages of infection [35]. These results may be specific to the particular Envs [35] or Env fragments [23] used, or due to the fact that gp120 and particle-associated Env trimers bind CD4 differently. In either case, current data utilizing a large number of Env constructs strongly suggests that the mucosal bottleneck is not the result of selective transmission of viruses with highly efficient CD4 or CCR5 use [17], [18], [20], or with increased efficiency of entry into particular CD4+ T cells subsets [20].
We also examined whether subtype C T/F and chronic Envs differed in their interaction with the gut homing integrin α4β7 as recently proposed [35], although previous data are almost entirely based on gp120 binding studies. In many ways, the α4β7 hypothesis is an attractive one. This integrin is expressed at high levels on activated CD4+ T cells in the gut [62] and cervicovaginal mucosa [63], both representing major sites of HIV replication early in infection [75], [76]. Moreover, intravenous administration of an anti-α4β7 mAb in rhesus macaques prior to and during acute infection with SIVmac239 resulted in decreased virus loads, perhaps by inhibiting trafficking of α4β7-positive T cells to the GI tract [77]. Finally, gp120-induced α4β7 signaling could promote virus replication through increased cell-to-cell adhesion. However, the ability of HIV-1 gp120 to bind α4β7 is far from universal - the commonly studied subtype B gp120s examined to date either do not bind to α4β7 or do so weakly, with the exception of SF162 [35]. Nonetheless, several gp120 proteins derived from early subtype A or subtype C infections have been shown to exhibit α4β7 binding capacity, and there is an obvious link between some α4β7 binding properties (fewer PNGs in the V1–V4 region) and genotypes associated with virus transmission in subtype C viruses [21], [27], [35]. While monomeric gp120 binds CD4, viral coreceptors and most broadly neutralizing antibodies, it differs from virion-associated Env trimers in important ways. Perhaps the best example is that numerous antibodies that bind to gp120 fail to neutralize the cognate Env trimer, consistent with both conformational differences and the fact that certain gp120 domains are sequestered in the oligomeric molecule. Thus, a key question that remains relatively unexplored is whether α4β7 binding by gp120 translates into an interaction by trimeric Env that influences virus infection and spread. To address this question, we concentrated on virus infection assays rather than gp120 binding experiments.
In our attempts to define the role of α4β7 in HIV-1 transmission, we were able to replicate a previous key finding, namely that saturating levels of antibodies to α4β7 modestly suppressed infection and replication by the prototypic subtype B strain HIV-1/SF162 [33], [34]. The inhibitory effects of α4β7 antibodies on SF162 infection were both transitory and most evident when low levels of virus input were used, which is precisely what would be expected if α4β7 functioned as an attachment factor [34], [35]. Attachment of virus particles to the host cell surface is a significant rate-limiting step to virus infection in vitro, but can be overcome in part by spinoculation [65], the inclusion of polycations that enhance viral binding [78], or the expression of virus attachment factors such as CD209 or CD209R [79], [80]. In the case of attachment factors, their ability to enhance infection is most pronounced when low levels of virus are used. Thus, our finding of a partial inhibition of SF162 replication in α4β7-positive T cells six days post infection at the lowest virus input is entirely consistent with previous reports and shows that our assays are sufficiently sensitive to measure the impact of α4β7 blockade on virus infection. Despite this, we found no inhibition of any T/F or chronic subtype B or C virus using cells from multiple donors and levels of virus empirically determined to be barely sufficient to establish a spreading infection. These findings are consistent with those of Pauls and colleagues, who found that a mAb to α4 used for the treatment of multiple sclerosis and Crohn's disease did not impact infection of atRA-treated CD4+ T cells by several HIV-1 strains, including two with the LDI/V tripeptide binding motif in the V2 region [64]. Since most of our T/F and chronic viruses possessed the α4β7-binding tripeptide motif as well as below average numbers of PNGs in the V1/V2 region, selection bias - i.e. the preferential inclusion of viruses that would be unlikely to interact with α4β7 - can also be excluded. Thus, we favor the hypothesis that not all Envs that can bind α4β7 in the form of gp120 necessarily do so as unliganded trimers.
Our failure to detect enhancement of viral infection of human CD4+ T cells by primary subtype B or C viruses, including T/F viruses, due to α4β7 interaction is by no means definitive, but does suggest that extrapolating results from gp120 binding assays to more complex virion infectivity studies may be misleading. It is possible that α4β7 interactions will be more important in other types of infection assays. In addition, we have not tested the ability of gp120 proteins derived from our viruses to bind to α4β7, although the relevance of such findings remains uncertain unless the corresponding trimeric Env exhibits similar properties. Our results demonstrate the importance of using replication-competent viruses to study properties associated with mucosal transmission. In contrast to single-round pseudovirus assays, experiments with IMCs are unbiased with respect to the genes that could influence fitness and enable detection of subtle differences following multiple rounds of replication. Thus, T/F and chronic IMCs are ideal reagents for future studies of phenotypes that may influence HIV-1 transmission.
This study was conducted according to the principles expressed in the Declaration of Helsinki. It was approved by the Institutional Review Boards of the University of Pennsylvania and Duke University. All subjects provided written informed consent for the collection of samples and subsequent analysis. Blood samples were obtained from 26 subjects infected with HIV-1 subtype C. A summary of their geographic origin and infection status is shown in Table 1. Blood specimens were collected in acid citrate dextrose, and plasma was separated and stored at −70°C.
The inference and cloning of T/F Envs and IMCs from SGA-derived viral sequences has been described (Figure S1) [10], [12], [13], [25], [37]. To ensure efficient expression of cloned subtype C Envs for pseudotyping, the sense primer used for amplification of the corresponding rev1-vpu-env cassette lacked the rev initiation codon (underlined) (5′-CACCGGCTTAGGCATCTCCTATAGCAGGAAGAA-3′) [39].
Since chronic HIV-1 infections represent complex quasispecies of genetic variants, it is impossible to predict, based on sequence analysis alone, which members of this quasispecies are functional and which are defective or partially defective. To generate biologically relevant chronic controls, we thus targeted viral variants for both Env and IMC construction that exhibited evidence of a recent clonal expansion. Viral RNA was extracted from the plasma of chronically infected individuals and subjected to SGA and direct amplicon sequencing as described [12], [13], with the following modifications: 5′ half genome amplification: 1st round sense primer 2010ForRC 5′- GTCTCTCTAGGTRGACCAGAT -3′, 1st round antisense primer 2010Rev1C 5′- AAGCAGTTTTAGGYTGRCTTCCTGGATG -3′, 2nd round sense primer 2010R1C 5′- TAGGTRGACCAGATYWGAGCC -3′ and 2nd round antisense primer 2010Rev2C 5′- CTTCTTCCTGCCATAGGAAAT -3′; 3′ half genome: 1st round sense primer 07For7 5′- CAAATTAYAAAAATTCAAAATTTTCGGGTTTATTACAG -3′, 1st round antisense primer 2.R3.B6R 5′- TGAAGCACTCAAGGCAAGCTTTATTGAGGC-3′, 2nd round sense primer VIF1 5′- GGGTTTATTACAGGGACAGCAGAG -3′ and 2nd round antisense primer Low2C 5′- TGAGGCTTAAGCAGTGGGTTCC -3′. Thermal cycling conditions were identical to [13] except that 60°C was used for primer annealing. Sequences were then aligned using ClustalW [81] and subjected to phylogenetic analysis using PhyML [82]. Phylogenetic trees were inspected for clusters of closely related viruses, or “rakes”, which are indicative of a recent clonal expansion. (Figures 1, S2 and S3). In five subjects (Table 1), at least one env amplicon was identical in sequence to the inferred “rake” consensus and thus selected for cloning using the pcDNA3.1 Directional Topo Expression kit (Invitrogen). In two subjects, observable “rakes” were limited to only two closely related sequences, which encoded Env proteins that differed by a single amino acid. In these cases, the amplicon that matched the within patient consensus at this ambiguous site was cloned. In the remaining six subjects, the consensus sequences of the clonal expansion “rakes” were chemically synthesized and cloned (designated .synR1 in Table 1). IMCs from chronically infected subjects (CH256, CH432, CH457, and CH534) were generated using the same approach. 3′ and 5′ half-genome SGA was performed using viral RNA from subjects with evidence of clonal expansion as determined by env sequencing. 3′ and 5′ half genome sequences were used to construct neighbor joining trees (Figure S3), and clusters of closely related sequences were selected for further analysis. A consensus sequence of the members of such “rakes” was generated using Consensus Maker (hiv.lanl.gov). 3′ and 5′ half genome sequences were confirmed to be identical in their 1,040 bp overlapping regions, chemically synthesized in fragments bordered by unique restriction enzymes, and ligated together to construct infectious proviral clones.
Virus pseudotypes were produced by co-transfecting 6 µg of pcDNA3.1+ containing the desired env clone with 10 µg of HIV-1 backbone (pNL43-ΔEnv-vpr+-luc+ or pNL43-ΔEnv-vpr+-eGFP (catalog no. 11100 from the NIH Aids Research and Reference Reagent program (ARRRP) contributed by Haili Zhang, Yan Zhou, and Robert Siliciano [83]) into 293T cells using the calcium phosphate precipitation method. 72 h post-transfection, the pseudovirus-containing supernatant was filtered through a 22 µm filter, aliquoted, and stored at −80°C. Pseudovirus used to infect primary CD4+ T cells was concentrated by ultracentrifugation through a 20% sucrose cushion. Pelleted pseudovirus was then resuspended in phosphate-buffered saline (PBS) in 1/100th the initial volume. All luciferase-encoding pseudoviral stocks were serially two-fold diluted and used to infect NP2 cells to define the linear range of the assay.
NP2 cells stably expressing CD4 and either CCR5 (NP2/CD4/CCR5) or CXCR4 (NP2/CD4/CXCR4) were infected with luciferase-encoding HIV-1 pseudoviruses by spinoculation in 96-well plates at 450× g for 90 min at 25°C. Cells were lysed with Brite-Glo (Promega) at 2 h post-infection and analyzed on a Luminoskan Ascent luminometer. To assess sensitivity to coreceptor inhibitor maraviroc, NP2/CD4/CCR5 cells were preincubated for 30 min with saturating concentrations of the CCR5 inhibitor maraviroc (1 µM; ARRRP catalog no. 11580; [84]) or the fusion inhibitor enfuvirtide (10 µg/ml) prior to infection. To assess sensitivity to broadly neutralizing mAbs (HIV-1 gp120 mAb IgG1 b12 (ARRRP catalog no. 2640) from Dennis Burton and Carlos Barbas [85]; HIV-1 gp120 mAb VRC01 (ARRRP catalog no. 12033) from Dr. John Mascola [60]; HIV-1 mAbs PG9 and PG16 (ARRP catalog no. 12149 and 12150) from IAVI [86]), viral pseudotypes were preincubated with 10 µg/ml of antibody for 1 hour at 37°C. Virus and antibody mixes were then used to infect NP2/CD4/CCR5 cells. All NP2 cell line infections were done in at least triplicate in at least three independent experiments using R5-tropic JR-FL as a positive control and Env-deficient pseudotypes as a negative control.
The ability of Env pseudoviruses to infect cells expressing low levels of CD4 was determined using affinofile cells, which are a modified 293T cell line that stably express CD4 and CCR5 under the control of independently inducible promoters [50]. 5×102 cells were plated in each well of a 96-well plate and allowed to grow for two days prior to infection. Cells were induced with 2 µM ponasterone, which induces supraphysiologic levels of CCR5 thus ensuring CD4 is the limiting factor in entry, and either 0.156 ng/ml (CD4-low) or 5 ng/ml (CD4-high) minocycline 18 hours prior to infection. Expression levels were monitored by quantitative FACS analysis [57]. At the time of infection, media was exchanged and 25 µl of luciferase-encoding pseudovirus was added. Cells were then spinoculated at 450 g for 90 minutes. Luciferase activity was measured three days post-infection. Each infection condition was done in at least triplicate in each of three independent experiments. Pseudovirus-containing vesicular stomatitis virus glycoprotein (VSV-G) which is CD4-independent and thus infects CD4-high and CD4-low cells equally, HIV-1 JR-CSF Env which requires high levels of CD4, and HIV-1 JR-FL Env which can utilize low levels of CD4 were included in all experiments as controls [87]. To calculate CD4-use efficiency, mean relative light units in CD4-low cells were divided by the value obtained in CD4-high cells. The signal-to-noise ratio was higher in affinofile cells than NP2/CD4/CCR5 cells. Nine Envs (5 T/F, 4 chronic) infected maximally-CD4-induced afffinofile cells less than 100-fold above background. For these Envs, we noted increased variability across independent assays. Additionally, the ratio of CD4-high to CD4-low infection was likely falsely elevated due to infection of CD4-low cells at background levels. These Envs are highlighted in Figure 2 and were excluded from subsequent analyses.
Maraviroc IC50 values were determined by pretreating NP2/CD4/CCR5 cells with 11 serial 3-fold dilutions of maraviroc, ranging from 5.9 µM to 0.1 nM, or no drug then spinoculating as above with luciferase-encoding pseudovirus and measuring luminescence 72 h post-infection. NP2 cells were chosen for this experiment because in the absence of CCR5 (NP2/CD4 only) these cells are highly restrictive to infection [88]; thus entry through potential alternative coreceptors when CCR5 is blocked by maraviroc is negligible. IC50 was determined using Prism 4.0 to determine the best fitting non-linear curve. Reported IC50 values are the mean of four independent experiments, with each drug concentration/pseudovirus condition performed in duplicate.
Primary human CD4+ T cells were purified by negative selection by the University of Pennsylvania's Human Immunology Core. Cells were stimulated with plate-bound anti-CD3 (clone OKT3) (eBiosciences) and anti-CD28 (clone 28.2) (BD Biosciences) and 20 U/ml recombinant interleukin-2 (IL-2) (aldesleukin, Prometheus Laboratories) in RPMI supplemented with 10% fetal bovine serum (FBS, Sigma-Aldrich). Three days after stimulation, cells were transferred to 96-well V-bottom plates. Five microliters of concentrated GFP-expressing pseudovirus was used to infect 6.7×105 cells in triplicate by spinoculating at 1,200× g for 2 hours. Cells were then transferred to new 24-well plates, and new medium containing 20 U/ml IL-2 was added. Three days post-infection, cells were stained for flow cytometry [20], [54].
A total of 1–2×106 cells from each condition were stained for flow cytometry. Incubations were done at room temperature in fluorescence-activated cell sorter (FACS) wash buffer (PBS, 2.5% FBS, 2 mM EDTA). Cells were first washed in PBS, then live/dead Aqua (Invitrogen) was added and incubated for 10 min. Next, anti-CCR7 IgM (BD) in FACS wash buffer was added and incubated for 30 min. Cells were then washed in FACS wash buffer before staining with anti-CD3–Qdot 655 (Invitrogen), anti-CD4–Alexa Fluor 700 (BD), anti-CD45RO–phycoerythrin (PE)-Texas Red (Beckman Coulter), and anti-IgM–PE (Invitrogen) for 30 min. Cells were then washed in FACS wash buffer and resuspended in 1% paraformaldehyde (PFA). Samples were run on an LSRII (BD) instrument and analyzed with FlowJo 8.8.6 (Treestar). Cells were gated as follows: singlets (FSC-A by FSC-H), then live cells (SSC-A by live/dead), then lymphocytes (SSC-A by FSC-A), then CD3+ cells (SSC-A by CD3), then memory markers (CCR7 by CD45RO).
To examine activation of cells treated with Act1, 2B4, isotype control murine IgG1, or no antibody, cells were pretreated with 33 nm of the specified antibody for 1 hour, 2 days, or 5 days. At each time point, 1×106 cells were washed in PBS, then live/dead Aqua was added for 10 minutes. Anti-CCR5-PE (BD), anti-CD4-PerCP-Cy5.5, anti-CD25-APC-Cy7, anti-HLA-DR-PE-Cy7, anti-α4β7-Alexa Fluor 680 (clone Act1) and anti-CD45RO-TexasRed-PE were added in FACS wash buffer for 30 minutes. Cells were then washed in FACS wash buffer and treated with cytofix/cytoperm buffer (BD) for 17 minutes. Anti-CD3-V450, anti-CD69-APC, and anti-Ki67-FITC in perm/wash buffer were added and incubated for 1 hour at room temperature. Cells were washed in perm/wash buffer, fixed in 0.1% paraformaldehyde, run on an LSRII instrument and data was analyzed with FlowJo. Live cells expressing CD3 and CD4 were analyzed for expression of activation markers.
Replication competent viral stocks were generated by transfecting a 10 cm dish 30% confluent with 293T cells with 6 µg of IMC DNA. Virus was harvested 72 hours post-transfection and filtered through a 45 micron low protein binding filter. 293T-derived HIV was then used to infect stimulated human CD4+ T cells. 18 hours after infection cells were washed twice to remove unbound 293T-derived virus. CD4+ T cell derived HIV was then harvested 11 days post infection, filtered through a 45 micron filter, aliquoted and frozen at −80°C. p24 antigen concentration of viral stocks was assessed by Alliance ELISA and high sensitivity alphaLISA (Perkin Elmer); these methods were in good agreement for all IMCs tested.
Freshly isolated human CD4+ T cells purified by negative selection were stimulated with 1.5 µg/ml anti-CD3 clone OKT3, 20 units/ml IL-2, and 10 nM atRA. atRA was resuspended in DMSO, filter sterilized, aliquoted in the dark, and immediately frozen at −80°C for no longer than one month. 24 hours after stimulation, media was removed, and new media with IL-2 and atRA was added. Media was changed and new IL-2 and atRA were added every two to three days. Efforts were made to precisely follow previously reported methods [33], [34]. Cells were infected six days post-stimulation for both pseudotype and replication competent HIV infection. Cells were pre-treated for 1 hour with 33 nM Act1 (an α4β7 heterodimer specific mAb; ARRRP catalog no. 11718 from Dr. A. A. Ansari [89]) at 37°C prior to infection. α4β7 expression and saturating mAb blockade was confirmed on the day of infection by flow cytometry. α4β7 expression was determined with Alexa Fluor 680-conjugated Act1 (Invitrogen). GFP-expressing pseudotype infections were performed as described above. SF162 and JR-FL pseudovirions were used as positive and negative controls, respectively [35]. For infections with replication competent HIV, virus made in CD4+ T cells was used to limit potentially non-physiologic properties of 293T-derived HIV. 2×105 stimulated atRA-treated CD4+ T cells were plated in 100 µl per well of a 96-well plate. After incubation with Act1 or media only, 100 µl of CD4+ T cell-derived virus was added at a neat, 1∶10, or 1∶100 dilution. Cells were infected for five hours at 37°C without spinoculation and then cells were washed four times to remove unbound HIV. Three, six, and nine days post-infection, media was changed and p24 antigen concentration was assessed in the supernatant using an alphaLISA high sensitivity kit (Perkin Elmer) read on a Synergy H4 plate reader (BioTek instruments). Either four or six replicate wells were used per condition, and each alphaLISA measurement was performed 2–4 times. AlphaLISA assays were performed in 25 µl volume in 384 well plates. Each plate contained an internal standard curve ranging four orders of magnitude with each standard concentration repeated in eight wells. The lower limit of detection for most assays ranged between 3 and 10 pg p24 per mL.
Purified CD4+ T cells were treated with atRA, anti-CD3, and IL-2, and infected as described for α4β7 blocking experiments. Nine days post infection, cells exposed to 10 µl of NL4-3-SF162, ZM249, CH162, CH256, and mock infected were pooled from six replicate wells, washed in PBS and then FACS wash buffer. Anti-CCR5-PE-Cy7, anti-CD54 (ICAM-1)-PE-Cy5, anti-LFA-1-PE (clone MEM-148), and anti-α4β7-Alexa Fluor 680 (clone Act1) were added in FACS wash buffer and allowed to incubate for 30 minutes. Cells were then washed in FACS wash buffer and treated with cytofix/cytoperm buffer for 17 minutes. Anti-Gag-FITC (clone KC-57) was added for 1 hour. Cells were fixed in 1% paraformaldehyde and DAPI was added 30 minutes prior to analysis. Samples were run on an ImageStream IS100 equipped with two cameras and 405, 488, and 658 nm excitation lasers. At least 20,000 images were collected per condition, and the upper limit of images classified as cells was set to 600 pixels to allow collection of cell-cell conjugates. Cells were gated as follows: singlets, doublets, and triplets (DAPI by brightfield area), then Gag+ (Gag intensity by Gag median pixel). Gag positivity was gated such that the percent of mock-infected Gag+ cells was equal in singlet, doublet, and triplet populations. A single Gag-expressing cell was sufficient for a doublet or triplet image to be considered Gag+, as confirmed by visual inspection of Gag+ images. LFA-1 and ICAM-1 expression were analyzed on nucleated (DAPI+), focused (brightfield gradient root mean square-high), singlets (brightfield aspect ratio ∼1) to ensure that all analyzed images were of high quality.
To test the hypothesis that T/F Envs as a group were different from chronic Envs in various functional measures, we used two-tailed Mann-Whitney tests. To test the hypothesis that Act1 treatment inhibited viral replication in CD4+ T cells, we again used two-tailed Mann-Whitney tests comparing the six biological replicates with and without Act1 treatment for each virus. No attempts were made to correct for multiple testing, largely because Act1 treatment did not have the expected effect on T/F viruses. Because six values from each group were compared, the minimum uncorrected p-value was 0.002. Applying the conservative Bonferroni correction would render all comparisons insignificant (α level of 0.05 divided by 39 tests = 0.001). Thus, we conclude that Act1 inhibits or enhances only when multiple input levels of the same virus show a consistent effect. To determine if cell-cell conjugates were infected more than expected by chance, the percent of Gag+ doublets was compared in a paired t-test to double the percent Gag+ singlets, and the percent of Gag+ triplets was compared to triple the percent Gag+ singlets.
All newly obtained HIV-1 sequences have been submitted to GenBank and are available under accession numbers listed in Table S1.
|
10.1371/journal.pcbi.1007224 | Depressive symptoms are associated with blunted reward learning in social contexts | Depression is characterized by a marked decrease in social interactions and blunted sensitivity to rewards. Surprisingly, despite the importance of social deficits in depression, non-social aspects have been disproportionally investigated. As a consequence, the cognitive mechanisms underlying atypical decision-making in social contexts in depression are poorly understood. In the present study, we investigate whether deficits in reward processing interact with the social context and how this interaction is affected by self-reported depression and anxiety symptoms in the general population. Two cohorts of subjects (discovery and replication sample: N = 50 each) took part in an experiment involving reward learning in contexts with different levels of social information (absent, partial and complete). Behavioral analyses revealed a specific detrimental effect of depressive symptoms–but not anxiety–on behavioral performance in the presence of social information, i.e. when participants were informed about the choices of another player. Model-based analyses further characterized the computational nature of this deficit as a negative audience effect, rather than a deficit in the way others’ choices and rewards are integrated in decision making. To conclude, our results shed light on the cognitive and computational mechanisms underlying the interaction between social cognition, reward learning and decision-making in depressive disorders.
| Blunted sensitivity to rewards is at the core of depression. However, studies that investigated the influence of depression on decision-making have often done so in asocial contexts, thereby providing only partial insights into the way depressive disorders impact the underlying cognitive processes. Indeed, atypical social functioning is also a central characteristic of depression. Here, we aimed at integrating the social component of depressive disorders into the study of decision-making in depression. To do so, we measured the influence of self-reported depressive symptoms on social learning in participants performing an online experiment. Our study shows that depressive symptoms are associated with decreased performance only when participants are informed about the actions of another player. Computational characterizations of this effect reveal that participants with more severe depressive symptoms differ only in the way they learn from their own actions in a social context. In other words, our results indicate that depressive symptoms are associated with a negative audience effect and thus provide new insights into the way social cognition and decision-making processes interact in depression.
| One of the core clinical symptoms of depression is anhedonia, which refers to a reduced motivation to engage in daily life activities (motivational anhedonia) and a reduced enjoyment of usually enjoyable activities (consummatory anhedonia) [1, 2]. In principle, this clinical manifestation could be explained by reduced reward sensitivity, both in terms of incentive motivation and in terms of reinforcement processes [3–5]. A direct prediction of this hypothesis is that depressive symptoms should be associated with reduced reward sensitivity in learning contexts both at the behavioral and neural level. However, while some studies do find evidence that depressive symptoms in the general population and in clinical depression are associated with blunted reward learning and reward-related signals in the brain [6, 7], others indicate no [8, 9] or mixed effects [5]. As a consequence, there is no strong consensus about which components of reward processing are most predictive of depressive symptoms in both the general population and clinical depression [5].
Another striking clinical manifestation of depressive symptoms is a marked decrease in social interactions. Depression is indeed associated with social risk factors, social impairments and poor social functioning [10]. Surprisingly, despite the importance of the socio-cognitive impairments that are often associated with elevated depressive symptoms, non-social aspects have received disproportionate attention. Furthermore, when social aspects are investigated the focus is often on emotional processing and theory of mind but not on how social information is integrated to produce efficient goal-directed behavior [11]. In the present study, our goal was to investigate whether the reward-learning deficit that is often associated with elevated depressive symptoms interacts with the social context [12].
According to social learning theory, a sizable amount of decisions are not directly shaped by people’s personal history of reward and punishments, but are rather acquired through social observation [13]. More specifically, this framework posits that human learning occurs mostly in social contexts, where subjects can be influenced by social cues (i.e. others’ choices and outcomes) [13, 14]. In order to test how depressive symptoms affect the integration of social cues during reinforcement learning, we administered a variant of a previously validated observational learning task on two independent samples of participants [14, 15]. Subjects also completed psychometric questionnaires assessing depression and anxiety (a co-morbid trait) symptoms. The task included a ‘Private’ learning condition, in which participants only had access to the outcome of their own choice, and two social conditions: the ‘Social-Choice’ condition in which participants had access to the demonstrator’s choice, and the ‘Social-Choice+Outcome’ condition in which participants had access to the demonstrator’s actions and their outcome (Fig 1A and 1B).
Our design allowed us to test several hypotheses concerning the relation between depressive symptoms and learning performance in private and social contexts. First, our design allowed us to test whether or not depressive symptoms degrade reward learning per se, as assumed by the standard account of depression as a reward sensitivity deficit. Second, by comparing the ‘Private’ and the ‘Social’ learning contexts, we could assess whether or not depressive symptoms are associated with a learning deficit in ‘Social’ contexts, as predicted by evidence of socio-cognitive impairments in depressive patients. Finally, thanks to computational analyses, we could precisely characterize the learning deficit in the ‘Social’ context either as a primary social learning deficit (i.e. impaired imitation) or as a secondary social learning (i.e. a negative audience effect).
An online experiment was particularly suited to test our hypothesis because—compared to laboratory-based experiments—it provides a more diversified pool of subjects, in terms of psychiatric traits and cognitive performance [16–19]. Specifically, we tested 50 participants in the general population and then ran a direct replication of the experiment on a second independent sample of 50 participants. In the main text, we report the meta-analytical p-values computed using a mixed effect meta-analysis. In the tables we present the results separately for each experiment and highlight the replication criteria proposed by the open science framework [20].
Levels of depressive and anxiety symptoms spanned a large range (Table 1) [21], with good internal consistency (Hospital Anxiety Depression scale—depression subscale: Cronbach’s alpha 85%; anxiety subscale: Cronbach’s alpha 84%). Participants were paired with a virtual demonstrator and performed a probabilistic reinforcement learning task in three contexts: a ‘Private’ condition, in which participants performed the task individually with no access to the demonstrator’s choices and outcomes, and two social conditions: the ‘Social-Choice’ condition in which participants had access to the demonstrator’s choices, and the ‘Social-Choice+Outcome’ condition in which participants had access to the demonstrator’s choices and their outcome. Overall, participants displayed robust instrumental learning and chose the most rewarded symbol above chance in all conditions (meta-analysis ‘Private’: MMETA = 0.65 ± 0.03, zMETA = 11.37, p < .001; ‘Social-Choice’: MMETA = 0.65 ± 0.03, zMETA = 11.83, p < .001; ‘Social-Choice+Outcome’: MMETA = 0.67 ± 0.03, zMETA = 12.45, p < .001; ± corresponds to the 95% confidence intervals; Fig 1C; See S1 Table for the results on the two samples separately).
Contrary to previous studies [14, 15], we used an online adaptive learning algorithm that determined the demonstrator’s behavior (Q-learning with learning rate = 0.5 and choice temperature = 10). As a consequence, the virtual demonstrators displayed realistic learning curves with some variability of performance (Fig 1C). We predicted that observational learning would result in a correlation between the participants’ and the demonstrator’s correct choice rate in a given learning session. As predicted, a higher correct choice rate for the demonstrator was associated with a higher correct choice rate for participants in both social conditions (‘Social-Choice’ condition: rMETA = .20 ± 0.07, zMETA = 2.89, p = .004; ‘Social-Choice+Outcome’ condition: rMETA = .20 ± 0.07, zMETA = 2.87, p = .004) but not in the private condition (rMETA = -.01 ± 0.11, zMETA = -0.05, p > .250; Fig 2A; see Table 2 for the results on the two samples separately).
In order to confirm that participants actually integrated the virtual demonstrator as a social partner, we measured the influence of participants’ rating of trustworthiness of the demonstrator’s face on social learning. An effect of perceived trustworthiness evaluations was found, such that participants who perceived the demonstrator’s avatar as more trustworthy had higher correct choice rates in the ‘Social-Choice’ (rMETA = .32 ± 0.13, zMETA = 2.54, p = .011) and in the ‘Social-Choice+Outcome’ conditions (rMETA = .29 ± 0.10, zMETA = 2.96, p = .003) but not in the ‘Private’ condition (rMETA = .11 ± 0.10, zMETA = 1.09, p > .250; Fig 2B). This effect of the social evaluation of the demonstrator’s avatar confirms that participants processed the information in a social context.
A significant effect of depressive symptoms was found such that the higher the depressive symptoms, the lower the rate of correct choices in the ‘Social-Choice’ condition only (rMETA = -.33 ± 0.10, zMETA = -3.47, p < .001; ‘Private’ condition: rMETA = .04 ± 0.16, zMETA = 0.16, p > .250; ‘Social-Choice+Outcome’ condition: rMETA = -.05 ± 0.10, zMETA = -0.48, p > .250; Fig 3A). However, a similar effect of anxiety, which is a comorbid trait of depression [22, 23], was found as a trend (rMETA = -0.18 ± 0.10, zMETA = -1.85, p = .065; Fig 3B). In order to better understand the effect of depressive symptoms on learning in social contexts, we ran a mixed linear logistic regression that included depressive and anxiety scores, taken as continuous between-subject variables (the regression also included a range of controls listed in Table 3). The analysis revealed a significant effect of depression scores such that the higher the depressive scores, the lower the rate of correct choices in the ‘Social-Choice’ condition compared to the ‘Private’ condition (zMETA = -2.85, p = .004; no other significant effect of depression and anxiety scores was evidenced: all ps > .250; Fig 3A). Importantly, the negative effect of depressive symptoms in the ‘Social-Choice’ condition was particularly robust, because it was found in both the discovery and the replication sample and in the blocks with stable and reversal contingencies (within-subject) (S2 Fig).
Finally, we tested whether the correct choice rates in the ‘Social-Choice’ condition identified participants with difficulties linked to depressive symptoms (i.e. scoring ≥ 8 on the HAD depression subscale [21]) from participants in whom these difficulties are absent. The classification analysis revealed that the performance in the ‘Social-Choice’ condition identified participants with depressive symptoms with good accuracy of 73 ± 1% and with good sensitivity, or True Positive Rate (82 ± 2%) but low specificity, or True Negative Rate (53 ± 3%) of the classifier (Fig 4A).
Although model-free analyses reveal a robust negative effect of depressive symptoms on learning in the ‘Social-Choice’ condition, they do not elucidate the cognitive mechanisms underlying this effect. Indeed, the effect of depressive symptoms could either be due to differences in social information processing, such as the demonstrator’s choices and outcomes (i.e. a primary social learning deficit) or to differences in the weighting of the information generated by participants’ own choices when social information is also available (i.e. a secondary social learning deficit or audience effect). These two hypotheses are hard to tease apart based on raw behavioral analyses, because both predict a reduced correct choice rate in the ‘Social’ conditions. Thus, to arbitrate between these two possibilities, we fitted a previously validated social reinforcement learning model [14, 24]. This model allows for biasing participants’ choice depending on the demonstrator’s choice in the ‘Social-Choice’ condition (i.e. imitation) and to update the value attributed to each symbol depending on the demonstrator’s outcome in the ‘Social-Choice+Outcome’ condition (i.e. vicarious trial-and-error). To directly assess the ‘socially induced individual learning deficit’ hypothesis [14], we allowed participants to have different individual learning parameters in the ‘Private’ (learning rate: αP ,temperature parameter: βP) and in the two social conditions (‘Social-Choice’ and ‘Social-Choice+Outcome’ conditions: αS , βS; Fig 5A).
More precisely, individual learning and decision-making were modeled with classical softmax (Eq 1) and delta-rule (Eq 2) functions, respectively governed by learning rate and choice randomness (or temperature) parameters:
Pt(st,at)=1/(1+e(ΔQt(st))*β)
(1)
Qt+1(st,at)=Qt(st,at)+αP*RPEt
(2)
Where RPEt is the reward prediction error calculated as follows (Eq 3):
RPEt=Rt−Qt(st,at)
(3)
During the ‘Social-Choice’ condition, the model assumes that the Demonstrator’s choice induces an ‘action’ prediction error (APEt; (Eq 4)), which measures how surprising the Demonstrator’s choice is, given the subject’s current estimate of the probability of selecting this option:
APEt=1−Pt(st,at)
(4)
The APEt is then used to bias choice probability (Eq 5) in the subsequent trial and the effect is scaled by a parameter κ ∈ {0–1}:
Pt+1(st,at)=Pt(st,at)+κ*APEt
(5)
Finally, in the ‘Social-Choice+Outcome’ trials, the model assumes that the demonstrator’s outcome induces an ‘observational’ reward prediction error (Eq 6), which is scaled by observational learning rate αO ∈ {0–1} (Eq 7):
OPEt=R(demonstrator)t−Qt(st,at)
(6)
Qt+1(st,at)=Qt(st,at)+αO*OPEt
(7)
To sum up, this computational model allowed us to address both primary social learning deficits (i.e. learning deficits captured by the parameters κ and αO, which are specific to social information) and secondary social learning deficits (i.e. learning deficits captured by the parameters βS and αS, which are specific to individual learning in contexts where social information is available).
As previously, we analyzed the model parameters fitted on participants’ actual behavior using correlations. Higher depression scores were specifically associated with lower learning rates in the ‘Social’ conditions (rMETA = -.25 ± 0.10, zMETA = -2.55, p = .011; all others, including anxiety: |zMETA| < 1.30, all ps > .190; Fig 5B–5D). These results where further confirmed by with structural equation modeling accounting for the correlation between the parameters (depression scores: zMETA = -2.61, p = .009; other ps > .188; Fig 4C). Interestingly, high depression scores were not solely associated with decreased learning rates in the ‘Social’ conditions, but also with decreased learning rates in the ‘Social’ conditions when controlling for the learning rates in the ‘Private’ condition (zMETA = -3.08, p = .002), which indicates that the presence of social information decreased the learning rate of the most depressed participants. To assess the complementary utility of computational measures, we tested whether the learning rate in the ‘Social’ conditions could identify participants with symptoms of depression (i.e. HAD depression subscale score equal or above 8 [21]). The difference in learning rates detected participants with depressive symptoms (score ≥ 8) with good accuracy (64 ± 1%), good sensitivity (64 ± 2%) and good specificity (65 ± 3%). A comparison between a classifier based on the model parameters and a classifier based on correct choice rates revealed that the model-based classifier was more specific to detect participants with higher symptoms of depression (t(198) = 5.86, p < .001), but was less sensitive (t(198) = -12.03, p < .001; Fig 4C) than the classifier based on correct choice rates.
Model-based analyses indicated that the severity of depressive symptoms specifically reduced individuals’ learning rate in ‘Social’ conditions (αS): a parameter that is used both in the ‘Social-Choice’ and in the ‘Social-Choice+Outcome’ condition. Model-free behavioral analyses showed that the learning deficit associated with depressive symptoms was specific to the ‘Social-Choice’ condition. To ascertain that this computational result was compatible with our model-free observation, we ran the same statistical analysis on simulated data [25]. Crucially, data simulated using the fitted parameters accurately recovered the decrease in performance associated with depression scores in the ‘Social-Choice’ condition compared to the ‘Private’ condition using the same mixed linear regression as on behavioral data (zMETA = -2.72, p = .007) as well as the blunted effect of depression scores in the ‘Social-Choice+Outcome’ condition compared to the ‘Private’ condition (zMETA = -1.74, p = .082). Therefore, it appears that, although depressive symptoms are associated with decreased learning rates in both social conditions, its detrimental effect is manifest only in the ‘Social-Choice’ condition. This is probably due to showing the demonstrator’s outcomes in the ‘Social-Choice+Outcome’ condition. This additional outcome information may compensate for the decreases learning rates with depressive symptoms. Confirming this intuition, our simulation analyses accurately recovered the absence of significant effect of depressive symptoms in the ‘Private’ condition (zMETA = -0.29, p > .250; S6 Fig). Thus, the simulations captured the specificity of the behavioral effect of depression scores and illustrate that our model provides an accurate description of the data.
As we were interested in the modulation of specific parameters by depression scores we tested whether our task allowed us to successfully retrieve a correlation between parameters in simulated datasets, an important quality check often referred to as ‘parameter recovery’ [25]. To do so, we ran 100 sets of simulations for each parameter, each simulating 100 participants, with the parameter of interest correlating with an arbitrary variable (defined as the depression scores) and the other parameters being randomly set for each participant in the range obtained by optimization on the total sample. The simulated data were then fitted using our social reinforcement-learning model. Overall parameter recovery was very good, especially for the parameters of the social conditions, with significant correlations were found in the 100% of the simulated datasets (average correlation coefficient of the parameters: r = 0.73 ± 0.01). Importantly, the recovery of the correlations was specific to the manipulated parameter with false alarms detected in less than 10% of the cases except for learning rate and choice temperature in the ‘Private’ condition (which was not our condition of interest) (Fig 5B). This result indicates that it is very unlikely that a correlation of one of our parameters with participants’ HAD depression scores is due to an effect of depression scores on another parameter.
In the present study we assessed reinforcement learning with a behavioral paradigm involving both private and social contexts, while concomitantly assessing depressive and anxiety symptoms in the general population. First, we replicate previous findings showing that participants integrate the demonstrator’s choices and outcomes, which is consistent with the idea that social learning processes (both in terms of imitation and vicarious trial-and-error) play a role in human reinforcement learning [14, 15, 26–28]. Second, we show that the severity of depressive symptoms is associated with a learning impairment that is specific to the learning context where participants are informed about the demonstrator’s choices (social context). This negative effect was robust to the inclusion of anxiety, and robust across experiments and outcome contingencies. Finally, computational analyses allowed us to characterize the effect of depressive symptoms as a secondary social learning deficit, i.e. a reduction of the learning rate in social contexts.
We found that depressive symptoms had a specific effect on imitation in the ‘Social-Choice’ condition. Crucially, the effect was robust to the inclusion of anxiety, which did not modulate performance in our task. That anxiety had no effect may come as a surprise given that previous studies have found that anxiety is associated with deficits in social and non-social reinforcement learning [29]. One possible explanation is that anxiety might be more strongly linked to classical fear conditioning than reward-based instrumental learning [30]. Depressive symptoms might thus undermine social reinforcement learning in instrumental and reward-maximization contexts, while anxiety might affect the same processes when outcomes are independent from the participants’ choices (i.e. Pavlovian learning) and when outcomes have a negative valence (aversive contexts).
Model-free analyses per se do not allow us to pinpoint the psychological mechanisms underlying the negative effect of depressive scores on correct choice rates in the ‘Social-Choice’ context. The absence of interaction between the demonstrator’s performance and depressive symptoms suggests that depressive symptoms did not lead participants to disproportionally follow ‘bad examples’ or to be insensitive to ‘good examples’. However, interpretations based on negative results are, at best, unsafe. To formally characterize the psychological mechanisms of the detrimental effects of depressive symptoms we thus turned to model-based analyses.
We fitted subjects’ choice with a slightly modified version of a previously validated social reinforcement-learning model [14]. As in standard algorithms, the model assumes that subjects learn option values via the calculation of a reward prediction error, that the values are moderated by a learning rate (αP) and that choices are generated via a soft-maximization process whose stochasticity is governed by a temperature (βP) [31]. In addition to this ‘private’ learning module, the model also displays sensitivity to social information: in the ‘Social-Choice’ condition the demonstrator’s choice biases the subsequent subject’s choice (the magnitude of this effect is governed by an imitation rate κ) and in the ‘Social-Choice+Outcome’ condition the demonstrator’s outcome is integrated into the subject’s value function with a vicarious learning rate (αO). Finally, we also allowed for different private learning rates and temperatures in the ‘Social’ contexts (αS and βS). This precise model parameterization allowed us to disentangle two different hypotheses concerning the drop in performance associated with depressive symptoms in the ‘Social-Choice’ condition. A correlation between depressive scores and imitation rates and/or vicarious learning rates would imply what we define a ‘primary’ social learning impairment (i.e. an impairment of the social learning processes per se). On the contrary, a correlation between the ‘Social’ context-specific learning rate and/or temperature would imply a ‘secondary’ social learning impairment (i.e. an impairment of the private learning processes in presence of social information). We found that depressive scores negatively correlated with the private learning rate in the social context (αS), thus indicating that the effect was consistent with a secondary impairment and was specific to the learning (as opposed to the decision) process. In other words, our computational results suggest that one possible way in which depressive symptoms affect learning in social contexts is conceptually similar to a negative audience effect [32, 33], where the presence of social signals (the demonstrator’s choices) induces a reduction of subjects’ instrumental performance.
From a methodological point of view, our study exemplifies how computational approaches can provide new insights on the way in which cognitive processes vary with clinical symptoms. Indeed, computational modeling demonstrated that the effect of depressive symptoms was selective of the way individual information was processed [34, 35]. It is worth noting that these conclusions were only allowed after a careful testing of the ability of our task to precisely identify which model parameter was influenced by depressive symptoms [25]. The exact cognitive and psychological mechanisms that mediate the negative effect of social signals in instrumental performance remain to be characterized. One possibility given that depressive symptoms are associated with lower cognitive functioning in general [36] is that the mere presence of others exacerbates these difficulties by capturing already scarcer attentional resources. Alternatively, negative perception of self and negative comparison to others are core symptoms of depressive symptoms [37]. Therefore, it is possible that the most depressed participants perceived their demonstrator’s behavior as more reliable, thus underweighting the information they acquired through their own experience.
Our results provide new evidence that depression-related reward learning deficits are highly context-dependent [3–5], and suggest that the difference in learning rates associated with depressive symptoms may only arise in social contexts [5, 9]. Crucially, our results suggest that supposedly neutral aspects of the experimental setup (such as whether or not the task is done in the presence or absence of an experimenter), may affect the results and explain inconsistent findings [38]. In line with recent propositions, our results also suggest that a deeper investigation of socio-cognitive impairments in depressive symptoms may provide important new insights [10, 11]. Following this idea, it would be particularly interesting to contrast the effect of depressive symptoms on learning when the information is socially (as in the current study) compared to asocially provided. Finally, we suggest that developing tools assessing reward learning outside and inside social contexts (characterized either by the presence of another player or by the social nature of the outcomes [39]) may prove useful to improve diagnosis and personalize treatments of depressive syndromes in the long term.
An obvious limitation of our study, is that we did not control for participants’ actual diagnosis and treatment, which may be problematic since medication interacts with decision-making in depression [40]. Therefore, our results would benefit from being replicated in carefully characterized population, while controlling for medication status and medical history. This replication would allow us to further measure the diagnostic value of our behavioral task and associated computational model-based analyses. Indeed, in the present study, we only tested its ability to detect participants with depressive symptoms as identified by a self-rated scale [21] . It would be particularly interesting to test whether our behavioral and computational measures improve existing self-assessments that detect clinically diagnosed cases of depression [41]. Finally, longitudinal designs will be required to assess whether or not our behavioral and computational measures present good test-retest reliability and reflect states or traits, and whether or not they predict the evolution of depressive symptoms to clinical diagnosis.
Our results have implications beyond their clinical relevance. Consistent with the ‘social learning theory’ participants imitated demonstrators’ choices (‘Social-Choice’ condition) and learned from their outcomes (‘Social-Choice+Outcome’ condition) [13, 14]. At the behavioral level, these two psychological processes were manifest in the fact that participants’ performance was modulated by the demonstrators’ performance. In particular, we found that participants observing a demonstrator performing ‘well’ performed better in the social compared to the private learning context. Importantly, the opposite was also true: participants observing low performing demonstrators displayed lower performance in the social compared to the private context. This latter result is in apparent contrast with the normative view that imitation should be biased toward successful individuals in order to be evolutionary adaptive [42–44]. This is also in contrast with recent empirical evidence using a very similar paradigm and showing that imitation rate is modulated by the actual performance of the demonstrator, so that demonstrators making random (i.e., non reward-maximizing) decisions are less imitated [15]. Two differences between the previous design and ours may explain this discrepancy. First, the previous study involved mild electric shocks (primary reinforcer), while our study involved abstract points to be converted into money (secondary reinforcer). More importantly perhaps, the previous design involved a between-subjects design with two groups of participants paired either with a consistently good or with a consistently bad participant, while in our experiments the performance of the demonstrator was allowed to fluctuate in a within-subject manner around an optimal behavior. Therefore, it could also be argued that our experiment is not well-suited for measuring demonstrators’ performance effects on participants’ imitation behavior as such effects require a relatively long and stable reputation building process [45, 46].
The question remains whether or not social learning in our task (imitation and vicarious trial-and-error) engaged domain-specific social cognitive module or domain-general information processing modules. In the absence of additional data (such as neuroimaging) we cannot provide a definitive answer. However, evidence from post-learning face ratings provides some clues [47]. We found a positive correlation between performance in the social contexts and the demonstrator’s judgment of trustworthiness. Even if we cannot infer a causal link and its direction from the post-learning face evaluation, these results suggest that a specific socio-cognitive module (face evaluation) correlated with instrumental performance, thus demonstrating the engagement of social information-specific processing and our reinforcement learning task.
Two independent cohorts of 100 American participants, similar in terms of reported age (mean reported age across the two cohorts: 33.39 ± 2.03) and of reported male/female ratio (mean reported male/female ratio across the two cohorts: 35%; see Table 1) were recruited via Amazon Mechanical Turk to participate in this online study. Each participant received a fixed 4$ amount for completing the 40-minute task to which a bonus earned during the experiment was then added (average bonus: 0.49$). Participant received a description of the study and signed an informed consent before starting the experiment. The study was approved by the the local Ethical Committee (Conseil d’évaluation éthique pour les recherches en santé–CERES n°201659) and is in accordance with the Declaration of Helsinki (World Medical Association, 2008). The first cohort corresponded to a ‘discovery experiment’ where we explored the relation between instrumental performance and clinical scores; the second cohort corresponded to a ‘replication experiment’ where we tested the robustness and replicability of the effect identified in the first experiment.
Participants performed the probabilistic instrumental learning task described in the Results section (Fig 1A and 1B). The task was programmed on Qualtrics and was composed of six learning blocks of 20 trials each. In each block, participants had to choose between two cues. Cues were characters of the agathodaimon font and were always presented in pair and only in one block per subject. The cue-to-condition attribution was randomized across subjects. Participants made their choice by pressing the E or P keys to choose the leftmost or rightmost symbol. Participants were given no explicit information on reward probabilities, which they had to learn through trial and error. In addition, they were encouraged to accumulate as many points as possible, with their final amount of points being translated into bonus money at the end of the experiment (conversion rate: 40 points equals 1$ bonus). In each pair, cues were associated with reciprocal reward probabilities (20/80% or 30/70%). For instance, in a 30/70% pair, the most rewarded cue provided a positive outcome (+1 point) 70% of the times and a negative outcome (-1 point) 30% of the time, while the less rewarded cue provided a negative outcome 70% of the time and a positive outcome 30% of the time. Participants had unlimited time to make their choice (Mean reaction time: 2.47 ± 0.88 s, no significant effect of depressive symptoms were found on the reaction times, all ps > .250).
Participants were told they had been paired with another player at the beginning of the experiment with whom they played in turn in each trial. In addition, it was indicated that there was no competition between them and the other player and that each player played for her/himself. As in previous studies [48], the behavior of the demonstrators was determined by a reinforcement learning algorithm (Q-learning) with a reasonable set of free parameters (𝛼 = 0.5, ß = 10; see below for a description for the Q-learning and its parameters). To avoid social perceptual biases, the other player was represented by a neutral avatar, chosen to be generally perceived as neither dominant or submissive nor trustworthy or untrustworthy [49]. Participants had to choose their own avatars in a set of other 16 identities (8 female, 8 male) at the beginning of the task. Participants performed this task in three different contexts with different amounts of social information: a ‘Private’ condition in which they did not have access to the demonstrator’s behavior, a ‘Social-Choice’ condition in which participants could see the demonstrator’s behavior but not their outcomes and a ‘Social-Choice+Observation’ in which participants could observe the demonstrator’s decisions and outcomes. Importantly, participants performed each condition (‘Private’, ‘Social-Choice’ and ‘Social-Choice+Outcome’) in separate blocks and each block was repeated twice. In the ‘Stable’ type of contingency, outcome probabilities were set at 30/70% and did not change during the block. In the ‘Reversal’ type of contingency, outcome probabilities were set at 20/80% and was inverted across cue after 10 trials (in average). Finally, at the end of the experiment, participants rated their demonstrator’s avatar on three personality traits (trustworthiness, dominance and competence) and completed the Hospital Anxiety and Depression Scale [21] as well as the Peters et al. Delusions Inventory, that was included in the exploratory analysis of the Discovery sample and then discarded in absence of any significant effect and its inclusion did not affect the effect of depression. The total procedure lasts approximatively 45 minutes.
The analyses were performed on all participants and trials. No exclusion criteria was applied.
|
10.1371/journal.pgen.1000800 | Understanding Gene Sequence Variation in the Context of Transcription Regulation in Yeast | DNA sequence polymorphism in a regulatory protein can have a widespread transcriptional effect. Here we present a computational approach for analyzing modules of genes with a common regulation that are affected by specific DNA polymorphisms. We identify such regulatory-linkage modules by integrating genotypic and expression data for individuals in a segregating population with complementary expression data of strains mutated in a variety of regulatory proteins. Our procedure searches simultaneously for groups of co-expressed genes, for their common underlying linkage interval, and for their shared regulatory proteins. We applied the method to a cross between laboratory and wild strains of S. cerevisiae, demonstrating its ability to correctly suggest modules and to outperform extant approaches. Our results suggest that middle sporulation genes are under the control of polymorphism in the sporulation-specific tertiary complex Sum1p/Rfm1p/Hst1p. In another example, our analysis reveals novel inter-relations between Swi3 and two mitochondrial inner membrane proteins underlying variation in a module of aerobic cellular respiration genes. Overall, our findings demonstrate that this approach provides a useful framework for the systematic mapping of quantitative trait loci and their role in gene expression variation.
| High-throughput genotypic and expression data for individuals in a segregating population can provide important information regarding causal regulatory events. However, it has proven difficult to predict these regulatory relations, largely because of statistical power limitations. The use of additional available resources may increase the accuracy of predictions and suggest possible mechanisms through which the target genes are regulated. In this study, we combine genotypic and expression data across the segregating population with complementary regulatory information to identify modules of genes that are jointly affected by changes in activity of regulatory proteins, as well as by genotypic changes. We develop a novel approach called ReL analysis, which automatically learns such modules. A unique feature of our approach is that all three components of the module—the genes, the underlying polymorphism, and the regulatory proteins—are predicted simultaneously. The integrated analysis makes it possible to capture weaker linkage signals and suggests possible mechanisms underlying expression changes. We demonstrate the power of the method on data from yeast segregants, by identifying the roles of new as well as known polymorphisms.
| DNA sequence polymorphisms that alter the activity of regulatory proteins can have considerable effect on gene expression [1]–[3]. With the advent of microarray and other genotyping technologies, it is now possible to examine the genome-wide effects of naturally occurring DNA sequence polymorphism on gene expression variation in segregating populations. For example, genotyping and expression data have been measured for 112 segregants obtained from a cross between the laboratory (BY) and wild (RM) strains of S. cerevisiae [1] and for 111 BXD mouse strain segregants [3].
Linkage analysis is commonly employed to identify DNA sequence polymorphism underlying gene expression phenotypes [1], [3]–[14]: the gene expression levels are treated as quantitative traits and the underlying DNA polymorphisms are called expression quantitative trait loci (eQTLs). Although standard linkage analysis successfully identifies eQTLs when applied to relatively small datasets, its utility in high-throughput eQTL analysis is limited due to the increased amount of background noise. To tackle this problem, a variety of methods take advantage of the modularity of biological systems and identify sequence polymorphisms that underlie an entire group of genes rather than single gene expression traits [4]–[6],[8],[10],[12]. Alternatively, a number of integrative approaches combine several data sources, including promoter binding data and sequence information, to improve the accuracy of eQTL identification [3],[14]. Several advanced methods capture not only sequence polymorphisms, but also the regulatory proteins underlying the expression changes. In those methods, the regulatory proteins are inferred concurrently with the linkage analysis, based on the approximation of regulatory protein activities by their mRNA expression level (e.g., [4],[12]).
In this study we devise a new method for characterizing the transcriptional response to DNA sequence variation. Called Regulatory-Linkage (ReL) analysis, it captures groups of genes together with their underlying DNA polymorphisms and their common regulatory mechanisms. The method (Figure 1A) takes as input genotyping and expression data for individuals in the segregating population, as well as a compendium of high-throughput transcription regulatory signatures. These regulatory signatures are gene expression profiles (selected from the literature) of strains mutated in particular regulatory proteins, such as transcription factors and chromatin modifiers. Our method produces a set of ‘ReL modules’, each consisting of a triplet: a small set of regulatory proteins, a group of target genes, and a genetic linkage interval. The target genes are jointly linked to the interval and share a common transcriptional control by the regulatory proteins. We say that the module's target genes are co-regulated by the module's regulatory proteins and are co-linked to the modules' linkage interval (Figure 1B).
The novelty of the current approach is twofold. All three components of the ReL modules – the groups of target genes, the underlying polymorphism and the regulatory proteins – are predicted simultaneously. Extant methods predict only two of the components simultaneously and add the third one in a separate pre- or post-processing step. Moreover, we integrate high-throughput gene expression data consisting of perturbations in a large variety of transcription factors. This integrated approach has several important benefits: First, the additional regulatory information makes it possible to capture weaker linkage signals. Second, the analysis focuses on groups of target genes that have a common regulatory protein and therefore avoids groups of genes that happen by chance to be co-linked to the same genomic interval. Third, the approach infers regulatory relations based on perturbations in a variety of regulatory proteins, thereby avoiding the approximation of protein activities by mRNA expression levels. Previous studies relied on this rough approximation to infer regulatory proteins concurrently with DNA polymorphisms (e.g., [4],[12]). Finally, the predicted regulatory proteins may suggest possible mechanisms through which genetic polymorphisms affect their target genes, providing initial interpretations of the ReL modules as part of the analysis.
Our analysis takes, as input, genotypic and expression data for a set of 112 individuals in a yeast wild-type segregating population. We organize these data as a linkage matrix, which presents the linkage (an eQTL likelihood score) between the expression level of each gene and each genetic marker (Figure 1A; see Methods). In addition, our procedure utilizes a compendium of ‘regulatory signatures’ that includes gene expression profiles from 283 different strains mutated in a variety of regulatory proteins [15]–[16]. In the following analysis, linkage relations are evaluated based on the linkage matrix, whereas regulatory relations are assessed by preferential over- or under-expression of target gene groups across regulatory signatures.
We aim to identify triplets of (i) target genes, (ii) linkage interval, and (iii) regulatory signatures, where the target genes are jointly linked to the linkage interval and co-expressed in the regulatory signature. The naïve approach of finding high-scoring triplets by evaluating all possible combinations is computationally infeasible even for relatively small datasets. To tackle this problem, our method proceeds heuristically in two stages. In the first stage, we organize the input as a higher order ‘ReL matrix’ across all genetic markers and regulatory signatures (Figure 1A). Each entry in the matrix indicates whether genes that are strongly linked to a particular marker are also over- or under-expressed in a particular regulatory signature. This statistical measure, referred to as ReL score, is calculated as follows: For each genetic marker, we partitioned the genes into two sets: genes with high linkage to the genetic marker and the rest of the genes. Given the regulatory signature, the ReL score measures the difference in the gene expression distribution between these two sets (see Methods).
We now use the observation that when a group of genes is co-regulated by several regulatory proteins and is jointly linked to the same linkage interval, the corresponding ReL sub-matrix will attain high scores. In accordance, the second analysis stage (Figure 1A) applies a biclustering algorithm on the ReL matrix to search for sub-matrices whose average scores are higher than randomly expected. In this work, we assume a single linkage interval underlying each sub-matrix. Accordingly, the ISA biclustering algorithm [17] was adapted to choose a single range of genetic markers (Methods).
The biclustering output is a set of sub-matrices, each scored by its average ReL scores, and specifies a set of regulatory signatures and a single linkage interval. For each high-scoring sub-matrix, referred to as ReL module, we attached additional attributes: (i) A set of regulatory proteins – the proteins that were mutated in the strains from which the module's regulatory signature was obtained. (ii) A group of target genes - genes that are both co-regulated by the module's regulatory proteins and co-linked to the module's linkage interval (see Methods). Since we focus only on trans-acting regulation, genes residing within or near the modules' linkage interval were excluded from the group of target genes. (iii) We hypothesize that the linkage interval contains a single gene that underlies the module's gene expression variation. We call this gene the causal regulator of the module. Among the genes within the linkage interval, we predict a plausible putative causal regulator (see Methods; Figure 1B).
In this analysis, we focus on the thirteen highest-scoring ReL modules (modules with ReL score >3). A comprehensive description of these modules is given in Table S1 and Table S2. Five additional modules were highly enriched in target genes residing in telomeric or subtelomeric regions of multiple chromosomes, and therefore were excluded from the analysis (Table S2; gene expression variation in telomeres has been discussed extensively elsewhere (e.g., [4])). Each of the identified ReL modules consists of at least 10 target genes. The modules comprise a total of 311 genetic markers, 82 different regulatory proteins, and 281 different target genes. Randomization analysis shows that the identified modules are highly unlikely to be generated at random (module size P-value<0.05, see Text S1 for details).
The identified ReL modules have no overlapping linkage intervals and only a few shared regulatory proteins: Eleven regulatory signatures are shared across two modules and no regulatory signature is shared across three or more modules. This is likely to be a consequence of our biclustering approach and the small number of modules. The little overlap allows us to organize the ReL matrix into a global map of ReL modules (Figure 2). The global map highlights the existence of ‘high intensity’ sub-matrices (modules). The map clearly shows that the high ReL scores within each module decrease drastically at the boundaries of its linkage interval and for regulatory signatures that are not part of the module.
Table 1 summarizes the ReL modules and their function. Modules are listed along with their key (best-scoring) regulatory protein, putative causal regulator, and the biological processes most enriched in the target genes (based on enrichment test; see Table S3). For example, the nucleobase biosynthesis module (module #6) predicts that uracil biosynthetic enzymes are linked to the causal regulator URA3 and regulated by the transcription factor Ppr1. Indeed, Ppr1 is a known transcription regulator of uracil biosynthesis genes, and the RM parental strain carries a deletion of URA3, a gene encoding one of the uracil biosynthetic enzymes (see details below).
All thirteen modules are significantly associated with a biological process (Table 1; eleven significant enrichments based on the GO database and two additional enrichments based on SGD, see Table S3). These significant enrichments give further support to the inferred ReL modules. For example, they justify the division of linkage interval II:352–697kb into two neighboring modules, #1 and #2 (linkage intervals II:352–376kb and II:489–697kb, respectively), since each module is characterized by a different biological process (‘ribosome biogenesis’ and ‘cytokinesis’, respectively; Table S3). Module #1 consists of 32 target genes, including ten ribosome biogenesis genes and only one cytokinesis gene. In contrast, module #2 consists of thirteen target genes with seven cytokinesis genes and no ribosome biogenesis genes (Table S1).
Among the genes residing within the linkage interval, the putative causal regulators (Table 1) were identified based on three criteria: (i) genes sharing the same biological process as the target genes, (ii) genes that have a physical interaction with at least one of the module's regulatory proteins, or (iii) proteins having a preferential binding to the promoter of the target genes (see Methods and Text S2 for a comprehensive description of causal regulator identification). For example, we have two indications that the causal regulator URA3 underlies gene expression variation in module #6. First, it takes part in the same biological process as the target genes (nucleobase biosynthesis), and second, it physically interacts with the module's regulatory protein Ppr1.
Out of the thirteen putative causal regulators, seven were previously confirmed (LEU2, URA3, AMN1, MAT, GPA1, HAP1, IRA2; [1], [6], [18]–[19]), thereby serving as positive controls. Two other putative causal regulators (ZAP1 and CAT5) were proposed previously but have not been tested [4],[6]. Two previously confirmed eQTLs (MKT1 and FLO8 [1],[5]) are not included in our ReL modules. Four putative causal regulators, RFM1, CRD1, TRM7 and TAN1 (modules #1, #5, #7, and #13), have not been previously identified.
The ReL analysis predicts regulatory relations between the modules' regulatory proteins and target genes. To demonstrate the quality of these predictions, we present their agreement with known, well-established transcriptional relations. Out of six known relations, ReL detects five relations whereas compared methods detect zero and four relations (see Text S3 for details). Interestingly, the nucleobase biosynthesis system was detected only by the ReL analysis.
The nucleobase biosynthesis system (module #6; Table 1) shows the unique ability of ReL analysis to recover not only the causal regulators, but also the regulatory proteins. The module's causal regulator is URA3, the target genes consist of URA1 and URA4, and the highest scoring regulatory protein is Ppr1. The module successfully captures the current biological knowledge about the uracil biosynthesis system. The RM parental strain carries a deletion of the URA3 gene, which is known to be linked to several members of the uracil biosynthesis pathway [1]. De-novo uracil biosynthesis is catalyzed by seven biosynthetic enzymes (Ura2,3,4,5,6,7,10). Four biosynthetic enzymes (Ura1,3,4,10) are subject to transcription regulation via the transcriptional activator Ppr1, whose activity is negatively regulated by uracil production rate [20]. The predicted effect of URA3 mutation on URA1,4 is highly likely to be mediated by Ppr1 activity: in the absence of Ura3 (RM variant), uracil production is reduced, causing Ppr1 activation (through the negative feedback) and, consequently, a transcriptional up-regulation of the uracil biosynthetic genes.
Notably, although most extant methods detect the nucleobase biosynthesis module, our approach is unique in inferring Ppr1 as the regulatory protein of the module (Text S3). This difference is not surprising, as most extant methods estimate Ppr1 activity by its mRNA level, whereas the actual activity is governed by uracil production rate. Taken together, the nucleobase biosynthesis module highlights the advantage of ReL analysis in predicting regulatory proteins based on causal information, without estimating protein activities with mRNA levels.
The sporulation module (module #13) shows our method's ability to reveal small modules. This module consists of only seventeen genes, eight of which encode meiosis- and sporulation-specific proteins (Figure 3A), linked to a locus on chromosome XV. Using previously reported mRNA expression patterns of all yeast genes through the sporulation time course, we found that these target genes are induced during mid-sporulation (Figure 3B). In agreement, the module's regulatory proteins are two DNA-binding proteins, Hst1 and Sum1, both required for transcriptional repression of middle sporulation-specific genes during vegetative growth and mitosis ([21], Figure 3A). Taken together, these results associate the module with transcription regulation of middle sporulation.
Hst1 and Sum1 are two subunits [1] of the Sum1p/Rfm1p/Hst1p tertiary repression complex controlling middle sporulation genes. RFM1 is a specificity factor that directs the Hst1p histone deacetylase to some of the promoters regulated by Sum1p [22]. Notably, Rfm1 lies in the modules' linkage interval; in fact, it is located within the peak of the interval (Figure 3C). It has an average eQTL likelihood score of 2.6 to its targets, and explains 27% of their gene expression variation. Segregants that inherited the linked locus from the wild RM showed higher expression of the sporulation module's targets than did segregants carrying the locus from the BY strain (Figure 3D). The BY parent carries two polymorphisms at the RFM1 locus: P247S and N227D. Sequence alignment of six yeast species [23]–[24] showed that the proline residue at position 247 is conserved whereas only the BY strain carries the P247S polymorphism; aspartic acid at position 227 is not evolutionarily conserved (data not shown). This observation suggests that the Ser247 impairs Rfm1 function, perhaps affecting the activity of the entire Sum1/Rfm1/Hst1 complex, leading to residual de-repression of mid-sporulation genes during vegetative growth. The linkage of RFM1 to expression variation has not been previously shown, probably since the signal could not be detected robustly for a small number of target genes. Our methodology overcomes this problem by exploiting the joint repressive effect of Hst1 and Sum1 during vegetative growth, enabling prediction of the genetic cause of variation in mid-sporulation genes.
The two respiration modules (#5 and #12) show the ability of our method to identify two distinct linkage intervals sharing the same target genes. The target genes of both modules are enriched with oxidative phosphorylation (P10−16 in #5, P10−41 in #12), and generation of precursor metabolites and energy (P10−13 in #5, P10−29 in #12), both of which are related to the process of aerobic cellular respiration, generating energy in the form of ATP (Table S3). The predicted causal regulators are CRD1 and CAT5 (modules #5 and #12, respectively), both required for normal respiration functionality and both residing within the peaks of the linkage interval on chromosomes IV and XV, respectively (Figure 4A; only CAT5 was previously proposed as a causal regulator [6]). Cat5 and Crd1 have an average eQTL likelihood score of 3.5 and 2.6 to their targets, respectively. Cat5 is required for biosynthesis of ubiquinone, an electron-carrying coenzyme in the electron transport chain. Cardiolipin is a phospholipid of the mitochondrial inner membrane, synthesized by the Crd1 cardiolipin synthase. Absence of cardiolipin in crd1 mutants results in decreased mitochondrial membrane potential and reduced respiration activity [25]. The target genes of the two modules show lower expression in segregants carrying the linked locus from the RM strain compared to the BY strain (Figure 4B).
Our results point to Swi3, but not to the common regulators of respiratory gene expression, as the key mediator of the CAT5-CRD1 effect. Swi3 is the sole predicted regulator of both respiratory modules (Table S1 and Table 1). Figure 4C demonstrates that indeed, the two respiratory modules are significantly over-expressed in the swi3 strain (t-test P10−8 and P10−22 in modules #5 and #12, respectively). Interestingly, the effect of swi3 deletion is stronger than the deletion effect of known respiratory transcriptional regulators, including Hap2/3/4/5, Mot3, Rox1, Aft1/2, and Cth1/2 (Figure S1). Swi3 is a subunit of the SWI/SNF chromatin remodeling complex, which is required for transcription of a diverse set of genes (e.g., mating-type switching and Gcn4 targets), but its specific role in respiratory gene expression has not been documented.
We next investigated the interrelations between the genetic variation in CAT5 and CRD1. To that end, we analyzed all genes that have high linkage (eQTL likelihood >2.5) to either CAT5 or CRD1. Interestingly, the linked genes have a strong overlap: out of the 62 genes linked to CAT5 and 29 genes linked to CRD1, twelve genes are linked to both regulators (hyper-geometric test P10−17) and contain mainly respiratory-related genes (11 of 12, Figure 5A and Table S4). Many of the linked genes are subunits of four respiration-related reactions: the electron transport chain, the citric acid cycle, ATP synthase, and mitochondrial carriers (in total, 15 of 29 in module #5 and 35 of 62 in module #12). Interestingly, the linked genes encode proteins that are non-randomly distributed across the various respiratory complexes: cytochrome c oxidase (Complex IV of electron transport chain) is exclusively encoded by genes linked to CAT5; the TCA cycle is composed of proteins encoded by the CRD1 linked group; and the genes encoding the ATP synthase complex and succinate dehydrogenase (Complex II of electron transport chain) are linked to both CAT5 and CRD1 (Figure 5B).
To test for possible genetic interactions, we compared the expression of the twelve overlapping linked genes in segregants carrying four possible combinations of the CRD1 and CAT5 alleles. Interestingly, we observed an additive effect of the CAT5-CRD1 genotypes (Figure 5C; compare also with Figure 4B). Whereas CAT5 and CRD1 alone explain 22% and 17% of gene expression variation, respectively, the combination of the two eQTLs CAT5-CRD1 explains 32% of the gene expression variation. Therefore, our results indicate that a genetic interaction between the eQTL pair CAT5 and CRD1 underlies the inheritance of genes required for normal respiration.
Our approach provides a high-resolution tool for identifying functional DNA polymorphisms that affect gene expression. Importantly, it also provides insights into the mechanisms by which genotypes underlie expression changes.
In our method, the regulatory signatures are gene expression profiles that were measured in rich medium under standard conditions on yeast cells carrying a single perturbation. The same methodology can be expanded to handle additional regulatory signature resources. For example, gene expression data measured under a variety of conditions may be included, disclosing modules that are inactive under standard conditions but active under particular extracellular stimuli. Furthermore, protein-DNA binding data, and data from double mutants might provide additional powerful information on ReL modules.
The ReL modules should be interpreted with caution. Genetic linkage does not necessarily imply causality. Two of the three criteria used for identifying the causal regulator are aimed to select among plausible hypotheses but do not demonstrate causality (see Text S2 for details). Additionally, the linkage interval might contain more than one causal polymorphism, whereas ReL analysis assumes a single causal regulator. In the case of two causal polymorphisms located at the same genomic region, ReL analysis might unify them into the same module or fail to detect one of them. Another point to consider is that the ReL modules do not provide an unbiased view of genome-wide genetic linkage. Since the modules are detected based on co-regulation in at least one regulatory signature, the resulting modules depend on the particular signatures included in the compendium. Further, some regulatory relations might be specific to a single regulatory signature, a short linkage interval, or a small number of target genes. ReL analysis may not have enough statistical power to generalize those focused relations into a module. Finally, our modules currently contain only a single linkage interval. Hence, ReL analysis might fail to detect the prevalent case where the target genes are influenced by a combination of multiple interacting loci. It might be possible to extend our framework to detect such interactions automatically. For all these reasons, our method may fail to identify certain correct modules despite a detectable causal polymorphism.
ReL analysis is likely to succeed in organisms other than yeast, including mouse and human. Several genotypic and gene expression datasets are available for these populations [26]–[27], and thus the most prominent obstacle is the lack of a large compendium of mammalian regulatory signatures. Such a resource, however, is likely to be compiled in the future, and the ReL methodology provides a good example of its usefulness. Text S4 provides a quantitative estimation of the number of regulatory signatures required for significant ReL analysis, highlighting the importance of a large compendium. As new technologies for cost-effective count of transcripts in perturbed cells become available (e.g., nCounter [28], shRNA-perturbation), it will be soon easier to obtain a large collection of mammalian regulatory signatures and apply our methodology to them.
When applied to the yeast system, our methodology reveals two intriguing ReL modules. First, we find that DNA polymorphism in RFM1 underlies gene expression variation of middle sporulation genes. Second, we show that both CRD1 and CAT5 underlie gene expression variation in aerobic cellular respiration genes. Further analysis reveals a novel genetic interaction (epistasis) between these two loci. It would be of great interest to explore whether the regulatory mechanisms uncovered here are conserved in other fungal genomes. The discovery here of previously uncharacterized modules and interactions in the well-studied segregating yeast population underscores the importance of large-scale integrated methods in genetic analysis.
We calculated the linkage (an eQTL likelihood score) of genotypic and expression data measured for 112 individuals in a yeast segregating population, as described previously [1]. The linkage matrix represents genetic markers versus genes, where each entry corresponds to the eQTL likelihood score between a given genetic marker and the expression of a given gene. The analysis was applied to all 2956 markers that were genotyped, and all 6230 genes whose gene expression was measured across the segregating population.
We formed a compendium of 283 high-throughput expression profiles obtained from strains mutated in various regulatory proteins [15]–[16]. The compendium includes only strains mutated in a single gene, and each mutant strain is represented by exactly one expression profile. The expression profiles are referred to as regulatory signatures.
Given a genetic marker and a regulatory signature, we evaluate whether genes that are tightly linked to the genetic marker are also over- or under-expressed in the regulatory signature. To that end, we partition the genes into two subsets: genes with high linkage to the genetic marker (denoted high-linkage genes), and the rest of the genes. The difference in the distribution of the regulatory signature values between the two subsets is evaluated using a t-test. The ReL score is the −log10 P-value of this t-test (all reported ReL scores are Bonferroni corrected). In our analysis, 11,166 of the 836,548 ReL scores (1.3%) were significant at P<0.001 (see Text S1). Given that the high-linkage (the rest) genes tend to have high (respectively low) regulatory signature values, the group of hit genes includes all those high-linkage genes whose values are above (respectively below) the average regulatory signature value. The hit genes are later used to calculate the target genes of the ReL modules.
The eQTL likelihood threshold, which distinguishes the high-linkage genes from the rest of the genes, was identified as follows: First, genes that are over-expressed and genes that are under-expressed in the regulatory signature are identified. For every possible eQTL likelihood threshold, we test for the over-representation of high-linkage genes in one of these expression groups using a hyper-geometric score (we consider all observed eQTL likelihood values as thresholds). The best score determines the eQTL likelihood threshold. The combination of hyper geometric score and the t-test is important for a robust evaluation. Unlike a t-test, the hyper-geometric test takes into account the amount of high-linkage genes, making sure that the eQTL likelihood threshold is not too high; on the other hand, unlike the hyper-geometric test, the t-test estimates the significance of difference between two distributions. Text S5 demonsrates the robustness of ReL analysis to small changes in the eQTL likelihood threshold.
The ReL matrix summarizes the ReL scores across all genetic markers and regulatory signatures. We set out to construct a group of co-regulated genes whose common transcription regulation involves both regulatory proteins and a causal regulator. In the ReL matrix, such an event appears as a sub-matrix with significant over-representation of high ReL scores. To identify those sub-matrices, the ISA biclustering algorithm [17] was adapted to work on the ReL matrix. ISA looks for any subset of columns and any subset of rows whose sub-matrix has high scores; the sub-matrix is subject to iterative improvements by adding or removing any column or row. Here we seek sub-matrices with a single range of consecutive genetic markers rather than any subset of markers. To that end, we modified the original ISA so that only markers at the boundaries of the current genetic marker range can be added or removed. On each ISA step, the genetic marker range is optimized efficiently using a dynamic programming algorithm. We start from all possible single entries as seed sub-matrices, and optimize each such seed independently of all others (see Text S6 for details). The resulting sub-matrix is called a ReL module. The ReL score of a module is the average ReL score of its entries.
A ReL module specifies a single range of genetic markers (referred as a linkage interval) and a set of regulatory signatures. For each ReL module, we further compiled the following information:
(i) Each module is associated with a set of regulatory proteins corresponding to the deletion mutants in the module's regulatory signatures. The ReL score of a regulatory protein is its average ReL score over the linkage interval.
(ii) As defined above, each entry of the ReL matrix is associated with a set of hit genes. The module's target genes are all hit genes included in at least 60% of the sub-matrix entries. Here we aim to investigate trans-acting regulation, and therefore, to avoid biases related to cis-acting regulation, genes residing within the linkage interval or less than 30 genes away from it were excluded from the set of target genes. In all thirteen modules under analysis, the original fraction of cis-linked genes was relatively small (Table S2).
Next, the function of the set of target genes is characterized by a hyper-geometric enrichment test using the GO biological process annotation (computed using the EXPANDER software [29]; all reported P-values are corrected for multiple testing). Given one or more significantly enriched biological processes for the same set of target genes, the best scoring process is termed the primary biological process of the module.
(iii) A causal regulator is a gene carrying a polymorphism in its promoter or coding region, which has a trans-acting effect on expression variation of other genes. For each ReL module, we aim to find one or a few putative causal regulators – genes contained within the linkage interval that are highly likely to be the causal regulators of the target genes. Following Tu et al. [7], we predict a putative causal regulator based on the following rules: The causal regulator either plays a role in the primary biological process of the module, or the yeast protein-protein and protein-DNA interaction network contains at least one direct link between the causal regulator and the module's regulatory proteins. Alternatively, the module shows statistical significant enrichment for targets of the causal regulator (see Text S2 for details).
Taken together, a full description of a module includes a set of regulatory proteins, a (small) set of putative causal regulators, and a set of target genes characterized by a primary biological process.
A program implementing our framework is available on the website: http://acgt.cs.tau.ac.il/ReL/.
|
10.1371/journal.pntd.0005673 | Flavivirus and Filovirus EvoPrinters: New alignment tools for the comparative analysis of viral evolution | Flavivirus and Filovirus infections are serious epidemic threats to human populations. Multi-genome comparative analysis of these evolving pathogens affords a view of their essential, conserved sequence elements as well as progressive evolutionary changes. While phylogenetic analysis has yielded important insights, the growing number of available genomic sequences makes comparisons between hundreds of viral strains challenging. We report here a new approach for the comparative analysis of these hemorrhagic fever viruses that can superimpose an unlimited number of one-on-one alignments to identify important features within genomes of interest.
We have adapted EvoPrinter alignment algorithms for the rapid comparative analysis of Flavivirus or Filovirus sequences including Zika and Ebola strains. The user can input a full genome or partial viral sequence and then view either individual comparisons or generate color-coded readouts that superimpose hundreds of one-on-one alignments to identify unique or shared identity SNPs that reveal ancestral relationships between strains. The user can also opt to select a database genome in order to access a library of pre-aligned genomes of either 1,094 Flaviviruses or 460 Filoviruses for rapid comparative analysis with all database entries or a select subset. Using EvoPrinter search and alignment programs, we show the following: 1) superimposing alignment data from many related strains identifies lineage identity SNPs, which enable the assessment of sublineage complexity within viral outbreaks; 2) whole-genome SNP profile screens uncover novel Dengue2 and Zika recombinant strains and their parental lineages; 3) differential SNP profiling identifies host cell A-to-I hyper-editing within Ebola and Marburg viruses, and 4) hundreds of superimposed one-on-one Ebola genome alignments highlight ultra-conserved regulatory sequences, invariant amino acid codons and evolutionarily variable protein-encoding domains within a single genome.
EvoPrinter allows for the assessment of lineage complexity within Flavivirus or Filovirus outbreaks, identification of recombinant strains, highlights sequences that have undergone host cell A-to-I editing, and identifies unique input and database SNPs within highly conserved sequences. EvoPrinter’s ability to superimpose alignment data from hundreds of strains onto a single genome has allowed us to identify unique Zika virus sublineages that are currently spreading in South, Central and North America, the Caribbean, and in China. This new set of integrated alignment programs should serve as a useful addition to existing tools for the comparative analysis of these viruses.
| Flaviviruses, including Zika and Dengue viruses, and Filoviruses, including Ebola and Marburg viruses, are significant global public health threats. Genetic surveillance of viral isolates provides important insights into the origin of outbreaks, reveals lineage heterogeneity and diversification, and facilitates identification of novel recombinant strains and host cell modified viral genomes. We report the development of EvoPrinter, a web-accessed alignment tool for the rapid comparative analysis of viral genomes. EvoPrinter superimposes alignment data from multiple pairwise comparisons onto a single reference sequence of interest, to reveal both similarities and differences detected in hundreds of selected viral isolates. Evoprinter databases provide easy access to hundreds of non-redundant Flavivirus and Filovirus genomes. allowing the user to distinguish between sublineage identity SNPs and unique strain-specific SNPs, thus facilitating analysis of the history of viral diversification during an epidemic. EvoPrinter also proves useful in identifying recombinant strains and their parental lineages and detecting host-cell genomic editing. EvoPrinter should serve as a useful addition to existing tools for the comparative analysis of these viruses.
| Flaviruses, including Dengue, Yellow Fever, Japanese Encephalitis and West Nile viruses, are significant public-health pathogens responsible for wide-spread epidemics. Recently, another member of this genus, Zika virus (ZIKV), has emerged as a global public health threat (reviewed in [1]. Two major ZIKV lineages have been recognized: an African lineage first detected in the Uganda Zika forest in 1947, and an Asian lineage, first isolated in South East Asia during the 1950s, that has since spread to the Americas (for review, [2, 3]). Phylogenetic analysis has revealed that both the African and Asian lineages can be further divided into distinct sublineages or groups [4, 5]. Recent studies have also shown that ongoing epidemics are accompanied by the continued diversification of viral sequences via accumulation of base substitutions and recombinant exchanges between related sub-groups [3, 6, 7]. Members of the Flavivirus genus have been grouped based on their vectors (reviewed in [8]). Mosquito-borne human pathogens include ZIKV, Yellow Fever virus, four Dengue virus species, St. Louis and Japanese encephalitis viruses, and West Nile virus, along with other highly diverse less-characterized groups for review, [8]. Although mosquitos are considered the primary vector for ZIKV transmission, recent studies have identified human to human transmission via sexual contact [9].
Analysis of Filovirus human outbreaks during the last 49 years, from the initial 1967 Marburg virus outbreak in Germany through the most recent 2014–15 Ebola virus epidemic in West Africa and in the Congo, indicates that these pathogens will continue to pose serious public health risks (reviewed in [10–12]. Ebola virus species involved in these outbreaks and other non-human infections include the Zaire, Sudan, Taï Forest, Reston and Bundibugyo species, with the Zaire strains responsible for the most extensive human outbreak [13, 14]. Likewise, multiple Marburg outbreaks have occurred in Kenya, the Congo, Angola, Uganda and South Africa (for review, [11, 15]. Studies indicate that each Filovirus genus may have its own particular transmission cycle that includes non-human primates, bats, rodents, domestic ruminants, mosquitoes and ticks (reviewed in [16]). While bats are considered the primary reservoir for many of these viruses [17, 18], studies on humans that survive acute Ebola/Zaire infections reveal the presence of persistent active virus within immune-privileged or tissue sanctuary sites [19].
Phylogenetic analyses of both Ebola and Marburg strains responsible for human and non-human primate hemorrhagic fevers reveal that genetically identifiable strains from distinct lineages are associated with individual outbreaks; during these outbreaks, evolving sublineages have emerged [20–27]. For example, sequence analysis of Ebola isolates collected during the 2014–2015 West African Zaire/Makona outbreak has revealed the presence of multiple distinct sublineages that can be temporally traced to an initial Guinea strain that diversified during its spread into Liberia and Sierra Leone [13, 28–31].
The availability of hundreds of Flavivirus and Filovirus genomic sequences is an important resource for acquiring insights into the evolution of these pathogens [32, 33]. Using current web-accessed alignment tools, when multiple viral genomes are compared, alignments are often difficult to visually assimilate given the large size of their readouts. For example, a ClustalW alignment [34] of 14 ZIKV strains produces a 51-page readout. In addition, web-accessed alignment programs restrict the number of viral isolates that can be compared in an individual alignment. To circumvent these limitations, we have developed a multi-genome alignment method that can superimpose hundreds of one-on-one alignments to reveal sequence polymorphisms and conservation as they exist within a sequence of interest [35, 36]. Individual one-on-one input:database alignments can also be accessed directly from the input-centric readouts.
The combined EvoPrinter/Clustal alignment algorithms described here access databases of hundreds of Flavivirus or Filovirus genomes, allowing the user to input a full or partial viral sequence to initiate a comparative analysis. EvoPrint readouts identify sequences shared by all selected strains, in addition to highlighting (through color-coding) unique base substitutions and those shared by subsets of database entries. EvoPrinter databases currently contains 1,094 Flavivirus entries including 148 ZIKV strains and 460 Filovirus genomes with 393 Zaire isolates from the recent West African Ebola outbreak. To demonstrate the utility of these comparative tools, we show how 1) alignment readouts highlight unique bases in both the input and database sequences; 2) multiple sublineages are identified within ongoing Florida, Dominican Republic, Puerto Rico, and Brazil ZIKV outbreaks; 3) SNP analysis of other ZIKV strains also reveals different Central American, Caribbean and Chinese sublineages; 4) novel Dengue2 and Zika recombinant viruses and their parental lineages were identified using differential SNP pattern screens; 5) SNP patterns differentiate between Ebola/Zaire sublineages; 6) host cell A-to-I hyper-editing within Ebola and Marburg genomes is identified by SNP profiling and 7) inter-species multi-genome Ebola virus alignments can identify ultra-conserved sequences.
Flavivirus and Filovirus EvoPrinter search and alignment algorithms allow for rapid one-on-one or multi-genome comparisons of either a user supplied viral sequence or a database sequence selected from hundreds of database genomes. By superimposing sequence homology data from either a single or an unlimited number of one-on-one alignments onto a selected reference sequence, EvoPrinter readouts provide an uninterrupted view of polymorphisms as they appear within the genomes of interest and allow direct access to individual alignments by expanding readout sequence lines. The following is a description of the EvoPrinter databases, alignment algorithms and readouts. The EvoPrinter programs and tutorials are found at: https://evoprinter.ninds.nih.gov/evoprintprogramHD/evphd.html.
Genomic sequences were curated from the NCBI/Genbank database [32], and additional information about virus strains was obtained from the Virus Pathogen Database [33]. To ensure that duplicate genomes do not interfere in the identification of uniquely shared sequences among different strains, redundant entries (detected by BLAST or Evoprinter alignments) were excluded. Database genome names contain the following information: species, NCBI designation, country of origin and year of isolation. When available, additional information is included in the names, such as lineage assignments, group designations and/or serotypes [14, 25, 37–44]. A lineage represents a set of genomes that differ from others within a species by a unique assemblage of sequence polymorphisms when compared to other species members. Different lineages are often marked by greater than 50 unique lineage-specific base differences.
In addition to FASTA formatted sequences, each entry was formatted for enhanced-BLAT (eBLAT) alignments to speed initial database searches [36, 45]. For eBLAT alignments, each genome was indexed into non-overlapping 11-mers, 9-mers and 6-mers and used to generate independent BLAT alignments that are superimposed to produce an eBLAT readout [36]. As of April 2017, the Flavivirus EvoPrinter database contains 1,094 non-redundant genomes that include the following: 574 Dengue (groups 1–4); 37 St. Louis Encephalitis; 115 West Nile; 110 Japanese Encephalitis; 70 Yellow Fever; 148 Zika; 8 Aroa-related; 7 Edgehill-related; 3 Entebbe-related; 3 Natya-related; 2 Spondweni-related; 12 Yaounde-related; 14 Insect-specific; 4 No Known Vector; 5 Seabird Tick-associated; and 8 Tick-borne genomes. Flavivirus groupings correspond to those previously described [8, 46, 47]. Databases will be updated when new genomes are submitted to NCBI.
The Filovirus database currently consists of 460 genomes that include 66 Marburg strains and 393 Ebola (371 Zaire, 10 Sudan, 7 Reston, 4 Bundybuygo, and 1 Taï Forest) isolates. Also included in the database is a single Cuevavirus genus strain, Lloviu Cuevavirus, isolated from European cave bats [48, 49].
To initiate the comparative analysis of a user-provided sequence, an eBLAT search is performed to identify database genomes that closely match the input sequence [36]. User-supplied sequences can range from 100 bases to complete genomes. Once the eBLAT search identifies the input species, one-on-one Clustal alignments using the alignment algorithms developed by [34] are generated between the input sequence and the intra-species database genomes. Although BLAT alignments are significantly faster than Clustal comparisons, aligning bases at or near sequence ends are often missed due to insufficient K-mer alignment target lengths. Pairwise alignments are then converted to distinguish between aligning bases (upper case) and non-aligning bases (lower case) within the input sequence for each comparison [45]. This input-centric format allows for the superimposition of alignment data from an unlimited number of pairwise comparisons [35, 36]. In addition, holding one-on-one alignment data in memory instead of multi-genome alignments allows for user-customized comparisons.
To achieve higher throughput volumes and processing speeds, we wrote a Java-based program that employs multithread parallel processing [50] to generate pairwise alignments concurrently. By random allocation of 144 computational threads, database search and alignment processing speeds are significantly enhanced using a Hewlett Packard 2.5GHz/512 GB RAM; 4 socket, 18-core processor server operating with the RedHat Enterprise Linux 6 operating system. User-provided Flavivirus sequences (including full genomes) are automatically aligned to all intra-species database genomes and, to speed up processing times, alignments to the larger 18 kb Filovirus genomes are done incrementally, with the initial alignment round to the top ten eBLAT scoring Filovirus genomes. Additional database genomes can then be added to include strains of interest.
From the genome selection tree, the user can select genomes for single or multi-genome comparisons with the input sequence. The genome selection page orders the one-on-one alignments, based on the number of base mismatches with the input sequence (least to most). The selection page allows the user to 1) view individual alignments with the input sequence by selecting the genome of interest; 2) view multi-genome superimposed alignments of all or a selected subset of genomes in order to either identify shared or conserved sequences via an EvoPrint readout or to highlight sequence differences by generating an EvoDifference print readout, and 3) initiate inter-species alignments. By moving back and forth between the genome selection page and alignment readouts, the user can quickly add or remove viral strains from the comparative analysis.
Sequence differences in multi-genome EvoDifference print readouts are color-coded to highlight base differences that are 1) unique to the input, 2) differ in only one of the database genomes, or 3) differ in two or more of the database entries (Fig 1). While sequence identity among the aligning genomes is indicated by gray-colored text in the EvoDifference print, conserved sequences within an EvoPrint readout are denoted by black text (Fig 2) and less conserved sequences are shown in gray font highlighted in green. In addition, bases that are unique to the input sequence and not present in any of the database genomes included in the EvoPrint readout are highlighted in red. The start and stop translation codons of open reading frames are highlighted when included in the alignment. For Flaviviruses, protein boundaries for the processed polyprotein are annotated (positions taken from the Virus Pathogen Resource [51]). Sidebars to the right of the readouts delineate protein encoding ORFs.
Sequence lines in both EvoDifference and EvoPrint readouts can be expanded to view the alignment details for each of the database genomes and, by selecting a virus strain listed in the readout, the user can view its one-on-one alignment with the input sequence. Amino acid alignments can also be viewed from one-on-one ORF alignments, to allow the user to assess whether nucleotide changes result in different encoded amino acids. A tutorial that details these alignment steps is available at the Flavivirus or Filovirus EvoPrinter websites via the EvoPrinter homepage (https://evoprinter.ninds.nih.gov/evoprintprogramHD/evphd.html).
As an alternative to a user-provided sequence analysis, a database genome can be selected as the input reference sequence for either individual or multi-genome alignments. EvoPrinter keeps a library of one-on-one alignment data between all Flavivirus database entries and a separate library for Filovirus database alignments that can be accessed for rapid comparative analysis. As with the user supplied input sequence search, database alignments are ordered on the genome selection page based on numbers of base mismatches compared to the input and individual alignments can be viewed by clicking on the database genomes.
To resolve different lineages and/or sublineages, the user should select ten or more genomes that have similar mismatch numbers with the input reference sequence and generate a multi-genome EvoDifference print. On the genome selection page, the bracketed numbers after the database name represent the number of base mismatches with the input sequence. In the readout, bases in black text indicate two or more database mismatches, and when these multi-genome differences are identical in two or more strains they frequently represent lineage or sublineage identity SNPs. In other words, when multiple strains have the same base substitutions these SNPs can be considered markers of lineage progression. By expanding readout lines that contain multiple mismatches, sublineages can be differentiated by their uniquely shared differences with the input (see Figs 3 and 4).
One-on-one EvoDifference print SNP patterns can be used to identify recombinant viruses and their parental lineages. Virus strains that are closely related to the input sequence, as revealed by low mismatch numbers (listed after the database genome name on the Genome Selection Tree), usually have randomly distributed base differences throughout their pairwise alignments with the input sequence. Discontinuity in mismatch scores between related database genomes, as seen by a sudden jump in score values, are often due to one of two reasons. First, a higher score can indicate a sublineage difference and in this case, the increased SNPs are randomly distributed throughout the alignment. Second, the higher score could indicate a recombinant exchange, and in this case, a cluster of high-density SNPs (a recombinant fragment from a more divergent minor parent) would be flanked by regions of lower SNP densities (from the major parent). Alternatively, if the recombinant is aligned with a member of the minor parental lineage, a significantly reduced low-SNP density region (corresponding to the above high SNP density cluster) is flanked by regions of higher SNP densities (from the major parent).
To identify members of the minor parental lineage, the database search is repeated using the region of the input sequence that generates the high SNP density cluster along with flanking sequences of the putative recombinant strain. If members of the minor parental lineage are present in the database, they will likely have the lowest mismatch numbers when compared to the other database genomes. By repeating the initial search using the complete or nearly complete recombinant genome and then comparing one-on-one alignments with members of both parental lineages, the genomic region that generates the high SNP density when aligned to a major parental lineage strain (Figs 5A and 6A) will show near identity within the corresponding region when aligned to a minor parental strain (Figs 5B and 6B). If a member of the minor parental strain is detected first, the members of the major parental lineage can be identified in database searches by using the low SNP density region plus its flanking higher SNP density regions and examining high mismatch scoring strains.
Differential SNP patterning can also be used to identify recombinant strains that are decedents of multiple rounds of recombinant exchanges with different partners. For example, if not all of the high SNP density clusters observed in the recombinant / major parental lineage alignment have corresponding “SNP clearings” when aligned to a member of the minor parental lineage, then the recombinant strain most likely is a mosaic of different recombination events involving multiple partners. To confirm putative recombinants, we recommend that additional recombinant detection programs be employed such as the Recombination Detection Program [52].
Both one-on-one and multi-genome EvoDifference prints of related Ebola or Marburg strains can be used to identify genomic sequences that have undergone A-to-I editing by host cell adenosine deaminases. When the conversion occurs within the replicative template of Filoviruses, the inosines are read as guanine residues resulting in T/U -> C substitutions in the negative stranded RNA genome (for review, [53]). In both one-on-one and multi-genome EvoDifference prints, hyper-editing appears as clusters of T or C unique substitutions depending on whether the editing occurred in the input sequence or database genome.
We have modified the EvoPrinter phylogenetic footprinting tool for the rapid comparative analysis of Flavivirus and Filovirus genomic sequences. Its alignment algorithms superimpose alignment data from an individual or up to hundreds of pairwise alignments, highlighting both sequence conservation and base differences within the user’s input sequence and database genomes. SNP pattern differences and conserved sequences can be viewed from readouts that highlight sequence differences (an EvoDifference print) or sequence conservation (an EvoPrint) (Figs 1 and 2). By expanding multi-genome readout lines, individual database alignments reveal SNPs that define lineages or sublineages, clusters of A-to-I host cell hyper-editing, and conserved sequence elements shared by all or a subset of database genomes. Differentially shared SNP patterns, identified in one-on-one EvoDifference print comparisons, also allow for the identification of recombinant viruses and their parental lineages.
Resolving sublineages during a viral outbreak or epidemic facilitates the identification of the genetic heterogeneity among viral isolates, identifies the spread of related strains to different countries, and allows for the detection of recombinant variants. Based on phylogenetic analysis, previous studies have identified major ZIKV groups: two African groups, consisting of West and East African sublineages [3] and a diverse Asian/Western hemisphere lineage (for review, [54]). The West African group contains isolates primarily from Senegal and Cote-d’Ivoire, while the East African sublineages can be further resolved into isolates from Uganda and the Central African Republic.
EvoDifference print readouts can be used to highlight sequence differences among related and evolutionary distant ZIKV strains (Fig 1). Using the capsid, pre-membrane and envelope encoding region from the Zika_KU321639.1_Brazil_2015 strain as the reference input sequence, one-on-one alignments with twenty-nine ZIKV database genomes were selected to identify 1) bases that are unique to the input, 2) bases that differed in only one of the database genomes, 3) sequences that differed in two or more database genomes, and 4) sequences shared by the input and all selected database genomes (Fig 1A). Alignment details and the color-coded names of the database isolates included in the comparative analysis can be viewed by selecting line numbers (Fig 1B). In this example, sequence line number 975 was expanded to highlight SNPs that are unique to the input or database genomes and shared SNPs. The expanded sequence line also highlights the greater SNP density of the more divergent African isolates (located below the horizontal line) when compared to the Asian isolates (above the line). Database genomes are ordered by their total number of base differences when aligned to the input sequence (least to most). The genome ranking and base differences are also part of the database selection page.
Differentially shared SNP patterns among multiple ZIKV isolates can be used to resolve individual sublineages. For example, when 525 bases of the Zika_KF383118.1_Senegal_2001 NS5 coding region are used to generate an EvoDifference print with database genomes from different African sublineages, their base differences with the input Senegal isolate or SNP profiles resolve different sub-groups (S1 Fig), that correspond to previously described sublineages [3, 4, 55–57].
Phylogenetic footprinting, identifying evolutionary conserved sequence elements using multi-genome alignment protocols, has become an important tool for resolving essential genomic information [35, 58, 59]. A significant advantage of EvoPrinter is the ability to rapidly change the cumulative evolutionary divergence stringency of a multi-genome comparison. By moving between the genome selection page and the EvoPrint readout, one can quickly add or remove viral strains from the analysis to reveal different levels of conservation of essential elements, as they exist within genomes of interest. For example, to identify previous characterized Ebola virus conserved transcriptional start and stop regulatory elements (for review [60, 61]), we generated a multi-genome EvoPrint of the Zaire_lin6_Kissidougou_GIN_C15_KJ660346.2_2014 strain that included 271 non-redundant genomes from 3 Ebola species (269 Zaire, 1 Bundibugyo and 1 Taï Forest) (Fig 2). In addition to resolving transcriptional regulatory elements that flank each of the seven Ebola virus genes, the divergence stringency of the EvoPrint is sufficient to highlight essential amino acid codons by revealing their less-conserved wobble positions and identify the transcription editing site within the GP gene (Fig 2). The EvoPrint also delineates less-conserved intergenic regions and the evolutionarily variable GP mucin-like domain encoding region [62] (Fig 2).
As with the Filoviruses, near-base resolution of essential information is obtained with Flaviviruses. A multi-genome EvoPrint was generated using the YellowFever_GQ379162.1_Peru_2007 NS3 encoding region as the input reference sequence, comparing it with 15 South American and African Yellow Fever strains selected from the Yellow Fever database (S2 Fig). Together the 15 strains provide a cumulative evolutionary divergence sufficient to resolve essential bases, as evident from the less conserved codon wobble positions (S2A Fig). Flavivirus SNP differences can also be accessed by expanding readout lines of multi-genome EvoPrints. The shared SNP profiles of different Yellow Fever Virus sub-groups (S2B Fig). correspond to previously identified phylogenetic tree groupings [63].
Shared SNPs that highlight differences between groups of viruses serve as ancestry informative markers for identifying sublineages (for review, [64]). We call these identity SNPs (ID-SNPs), since they represent lineage markers for descendants of an earlier parental strain and multiple shared ID-SNPs, or profiles can be used to resolve different sublineages and illuminate ancestral relationships among ZIKV strains during spreading epidemics. Most ID-SNPs highlight differences between a sublineage and all other strains outside of the sublineage that have maintained the same ancestral base at those nucleotide positions (Figs 3 and 4).
Phylogenetic tree comparisons of Asian/Oceania strains have revealed that the South American epidemic (first identified in Brazil) derives from a distinct sublineage that arose from an outbreak in French Polynesia in 2013 [4, 55–57] (for review [4, 65, 66]). Our SNP profiles of Brazilian isolates reveal that they can be further divided into at least four different subgroups based on non-overlapping ID-SNP patterns shared among 20 isolates (Fig 3 and S3 Fig). For example, when the Zika_KX447510.1_FrenchPolynesia_2014 strain is used as the input reference genome and aligned to 13 Brazilian isolates, 3 subgroups (Br1-3) (each represented by multiple isolates) were distinguished by 22 ID-SNPs that are positioned throughout the genome (Fig 3). When isolates from China, Ecuador, Florida, Dominican Republic, Puerto Rico, Suriname and French Guiana are included in the analysis, all five of the Florida isolates, all of the Ecuador, and two of three Dominican Republic strains share ID-SNPs with the first Brazilian subgroup (Br1) but not with the Br3 subgroup (Fig 3). The second Brazil sublineage (Br2) shares ID-SNPs with Florida isolates and with the Puerto Rico strains but not with Br1 or Br3 (Fig 3). The alignment also reveals that the Puerto Rico, Suriname, French Guiana and a single Dominican Republic isolate share ID-SNPs with the third Br3 Brazil subgroup but not with the Br1 sublineage. In addition, while isolates from Florida and Puerto Rico represent two distinct subgroups, the ID-SNP patterns of isolates from the Dominican Republic reveal that one isolate is related to the Puerto Rico subgroup while the other two share ID-SNPs with the Florida subgroup (Fig 3). Interestingly, pairwise alignments between the Dominican Republic isolate that is related to the Puerto Rico subgroup, the Zika_KX766028.1_DominicanRepublic_2016 strain, and any of the China Ch2 sublineage members reveal near identity, suggesting that the Ch2 sublineage may have originated from the Caribbean (Fig 3 and S4 Fig). This possibility is further strengthened by the observation the China Ch2 strains share many ID-SNPs with isolates from Puerto Rico, Dominican Republic, Suriname, French Guinea, and members of the Brazil Br3 sublineage. In addition, these observations are in agreement with Zhang et. al., who report the presence of highly diversified ZIKVs that have been most likely imported into China [67].
Comparative analysis of isolates from the recent southern Florida outbreak identify ancestral ID-SNPs that together suggest a progressive evolutionary divergence away from other related strains and other members of the Asian lineage. For example, an EvoDifference print of the Zika_KX832731.1_Florida_2016 isolate with 71 other Asian/Oceanian/Western hemisphere strains (both related and evolutionarily distant) revealed ID-SNPs that are shared among Florida and Dominican Republic isolates while all other strains have maintained the same ancestral base at those positions (Fig 4). Our analysis also identified ancestral ID-SNPs that are restricted to just a subset Florida and Dominican Republic strains and ID-SNPs that only distinguish a subset of Florida isolates from all other Asian lineage strains. Taken together, the different subgroups indicate that progressive, multi-generation base substitutions at different genomic positions are playing a significant role in ZIKV divergence. In addition, the multi-genome analysis demonstrated that the KX832731_Florida strain has recently acquired three unique SNPs that are not shared by any of the other Asian/Oceanian/South American strains (two of the three unique SNPs are red highlighted in Fig 4).
We have also used ID-SNP profiles to search for additional Western Hemisphere sublineages by examining pair-wise alignments of South/Central American and Caribbean isolates. Our screen identified two Central American sublineages, differentiated from the Brazil Br1-4 subgroups by combinations of 15 ID-SNPs (S3 Fig). These subgroups contained isolates from Mexico, Guatemala, Honduras, Panama and Columbia. Strains from Mexico fall into either the first or second central American group. Our comparative analysis also revealed that the single Martinique isolate, Zika_KU647676.1_Martinique_2015, most likely originated from a Mexican strain as it differs from the Zika_KU922960.1_Mexico_2016 isolate by only 4 bases.
To examine sublineage heterogeneity among Asian and Southeast Asian ZIKV strains, we searched for ID-SNPs that group isolates from different locations. As indicated above, our SNP pattern screen revealed two Chinese subgroups that are differentiated by 31 ID-SNPs (Fig 3 and S4 Fig). Using Zika_KU955589.1_China_2016 as the input reference genome, our multi-genome analysis revealed that the Chinese Ch2 subgroup shares many ID-SNPs with Western hemisphere isolates, while the first China subgroup (Ch1) constitutes a distinct (perhaps older) Asian sublineage (Fig 3 and S4 Fig). The French Polynesian strains share six ID-SNPs with the Ch1 subgroup and the Tonga strain shares eight ID-SNPs, suggesting that strains from Tonga and French Polynesia may be evolutionarily positioned between the Chinese Ch1 sublineage and Western hemisphere isolates (S4 Fig).
Genomic diversity among Flaviviruses is driven in part by homologous recombination between related strains, with their recombinant exchanges occurring in both protein encoding and noncoding sequences [3, 7, 68, 69]. Alignment programs that scan for changes in sequence homology within multiple genomes and methods that examine differential phylogenetic clustering using genomic sub-regions have been used to identify recombinants and locate approximate recombinant fragment boundaries [70, 71]. Evoprinter screens can also identify recombinants and resolve the approximate boundaries of their recombining fragments within parental lineages.
By examining a previously characterized Dengue2 recombinant, we show how SNP profiling can be used to identify recombinant strains and their parental sublineages (S5 Fig). Phylogenetic tree clustering analysis of the Dengue2_AF100466.2_Venezuela_ 1990 (Mara4) strain with other Dengue2 genomes revealed that Mara4 is the recombinant progeny of two distinct Dengue2 sublineages [71]. Differential phylogenetic clustering analysis revealed that the first ~500 bases of Mara4 are nearly identical to Dengue2 strains from Thailand, while the remaining genome is related to American strains [71]. Side-by-side EvoDifference SNP profile comparisons of the Mara4 recombinant with members of the parental sublineages (from Thailand and Jamaica; S5A and S5B Fig, respectively) demonstrate that the 5’ recombinant fragment originated from the minor parental Thailand sub-group (boxed region in S5 Fig). Note that, by convention, the strain that produces the highest SNP density within the recombinant region when aligned to the recombinant strain is designated as the major parental lineage, while the minor parental sub-group shares identity or near identity with the recombinant within the boundaries of the recombinant fragment. In this example, the differing parental SNP pattern boundaries are located at positions 594 (major parent) and 600 (minor parent), indicating that the recombinant exchange most likely occurred between bases 595 and 599 (S5 Fig).
One advantage of the genome SNP profiling is that recombinants and their parental lineages can be identified by differental SNP patterning. For example, Fig 5 identifies a novel Dengue2 recombinant strain. Side-by-side comparisons of SNP profiles generated from one-on-one EvoDifference prints of the Dengue2_GQ398269.1_PuertoRico_1994 strain with another Puerto Rico strain (Dengue2_KF955363.1_PuertoRico_1986) and with a New Guinea isolate (Dengue2_AF038403.1_NewGuinea_1988)–Fig 5A and 5B, respectively—revealed that the Puerto Rico_GQ398269.1 strain is the resultant progeny of a recombinant exchange between a member of a Puerto Rican subgroup (major parental sublineage) and a New Guinea sub-group member (minor parental sublineage) (Fig 5). The abrupt SNP density pattern change within the recombinant Puerto Rico/New Guinea strain alignment delineates an ~2,100 base region (spanning the NS2B and NS3 protein encoding sequences) that is identical in both the recombinant and New Guinea genomes (Fig 5B). Note that the higher density SNP cluster in the Puerto Rico (major parent)–recombinant strain SNP profile alignment corresponds to the region of identity shared between the recombinant and the minor parental strain (Fig 5A and 5B).
Using SNP profiling, we have sought evidence of recombination within Asian and African ZIKV lineages. Our initial screen of the China Ch-1 sublineage isolates revealed that many are nearly identical, however, the SNP profile generated when the Zika_KU963796.1_China_2016 strain was aligned to Zika_KU866423.1_China_2016 identified two genomic regions that have significantly higher SNP densities when compared to flanking sequences (Fig 6A). Further analysis that included other Asian strains revealed that when the KU866423.1 strain was aligned to a Cambodian isolate, Zika_JN860885.1_Cambodia_2010, their genomes are identical in the same two regions that displayed higher density SNP clustering in the above KU866423.1—KU963796.1 comparison, but differ significantly in sequences flanking these regions (Fig 6). The matching genomic positions that have converse high SNP density vs. sequence identity reveals that the KU866423.1 strain is the recombinant progeny of the two separate parental sublineages, one from China (major parent) and the other from Cambodia (minor parent). The one-on-one SNP pattern comparisons also revealed that the recombinant strain is the product of two genomic exchanges, with one occurring in sequences that code for NS3, NS4A and 2K proteins, while the other in-frame exchange occurred within the 3’ end of the NS5 coding region. Notably, when members of the parental lineages are aligned, their SNP profiles do not reveal any significant changes in SNP densities that would flag these as recombinant strains (Fig 6C).
African lineage ZIKV recombinant strains have been described previously [3]. Consistent with these observations, EvoDifference prints of available African ZKIV strains have identified multiple one-on-one alignments that display significant changes in SNP densities within different regions of their polyprotein encoding sequences (Fig 7). For example, using the Zika_KF383119.1_Senegal_2001 as the input reference genome and examining other African strains, significant changes were identified in SNP densities within different genomic regions. Our initial multi-genome comparisons identified a 139-base region within the NS5 coding region that significantly differs from sequences within the original 1947 Uganda Zika forest sentinel monkey isolate and two other strains from Senegal and the Central African Republic (S6A Fig). Expanding the readout lines revealed that the Uganda and Senegal isolates are identical to adjacent but non-overlapping portions of the KF383119.1_Senegal reference sequence, while a Central African Republic strain shares many of the sequence differences of both the Uganda and Senegal isolates (S6B Fig).
Examination of other African strains also revealed SNP clustering within this region and other significant changes in SNP densities outside of the NS5 coding region. Similar to the China/Cambodia recombinant, many of the high-to-low-to-high SNP density changes indicate multiple recombination exchanges have occurred within these viruses (Fig 7). For example, alignment of the KF383119.1_Senegal with Zika_KF383118.1_Senegal strains identified three additional clusters of sequence differences; most notably, a putative recombinant fragment that spans the capsid and envelope encoding sequence (Fig 7B). Also note, the SNP cluster in panel A (that spans the NS5 encoding sequence) and the high density SNP cluster within the same genomic region shown in panel B were adjacent but non-overlapping (also highlighted in S6B Fig.pdf). High density SNP clusters were also identified in an EvoDifference print of KF383118.1_Senegal and the LC002520.1_Uganda (Fig 7C), with the NS5 SNP cluster expanded to include both NS5 high density SNP clusters (Fig 7A and 7B). The juxtaposition of high and low SNP densities within the one-on-one comparisons highlight putative recombinant exchanges, with one of the aligning strains most likely belonging to the major parental sublineage (Fig 7A–7C).
An EvoDifference print of the African KF383119.1 strain with an evolutionarily distant African strain, Zika_KF383116.1_Senegal_1968, shows extensive divergence throughout their coding sequences, with the exception of the centrally located NS3 encoding sequence (bases 5227 to 5556) (Fig 7D). In addition, pairwise alignments with other African strains uncovered evidence of additional African recombinant exchanges. For example, one-on-one SNP profiles of KF383119.1 or KF383116.1 strains with another highly divergent Senegal strain, Zika_KF383120.1_Senegal_2000 revealed multiple significant changes in SNP densities (Fig 7E and 7F, respectively). Although the KF383120.1 strain is considered to be inactive, given the presence of an internal in-frame stop codon [3], recent phylogenetic analysis reveals that the KF383120.1 strain belongs to a distinct African sublineage that includes other closely related functional strains [4].
Filovirus database genomes are grouped according to their species and lineage designations [23, 25, 26, 37, 72]. A comparative analysis of the Zaire species identified seven lineages that make up three major groups: 1) the Kikwit (lin1), Gabon (lin2) and Mayinga (lin3) isolates taken together fall into a related group; 2) the Ilembe (lin4), Luebo (lin5) and Boende (lin7) together fall into a second group, and 3) the recent Zaire/Makona West African isolates (lin6) represent a more divergent lineage (Fig 8), in agreement with recent lineage designations [72]. Alignments of the other Ebola species revealed two Bundibugyo lineages, four Reston, six Sudan lineages and one Taï Forest (the sole sequence in this species). Consistent with previous studies, EvoDifference prints identified nine Marburg lineages [22, 73].
As an example of using SNP patterning to resolve different Filovirus lineages, we show how a multi-genome EvoDifference print of the Zaire_lin1_Kikwit_AY354458.1_1995 GP gene mucin-like domain encoding sequence [62] with other Zaire reference strains can identify different lineage-specific SNP patterns (Fig 8). By selecting a readout line number (line 7425 in this case), sequence differences are revealed among the different aligning lineage pairs, allowing an assessment of which bases conform to the input sequence and which are different and unique to a single lineage or shared by various other lineages (Fig 8B). Interestingly, most of the SNPs within the mucin domain are T/U->C substitutions and may be the result of host cell A-to-I RNA editing (discussed below).
Phylogenetic analyses coupled with retrospective epidemiological studies of the recent West African Ebola/Zaire outbreak revealed that the epidemic started in Guinea and spread to Sierra Leone and Liberia [28, 30, 74, 75]. During its rapid spread, base substitutions were identified that distinguished between early and late isolates [74], reviewed by [12]. To highlight the ability of EvoPrinter to identify subgroups, we illustrate how a multi-genome EvoDifference print, using the early isolate Gueckedou_GIN_C05_KJ660348.2_2014 genome as the input reference sequence, identified two subgroups within the Ebola/Zaire outbreak [marked by two identity SNPs at position 13,856 (A->G) and position 15,660 (T->C)], one represented by the Gueckedou subgroup (Guinea-1a), and a larger subgroup represented by the majority of Ebola/Zaire strains (S7A Fig). The accumulation of SNPs in Guinea-1a strains from Coyah and other locations illustrates the persistence of this early lineage over the course of the epidemic. The second identity SNP at position 15,660 (T->C), reinforces the hypothesis that the Coyah isolates, plus an isolate from Liberia, are part of the same early sublineage [12, 29, 30, 31].
Using a strain isolated during the later phase of the epidemic as the input reference sequence, identity SNPs were identified in the Sierra Leone-Guinea-3 sublineage [74, 76] (S7B Fig). Our analysis revealed an identity SNP at position 10218 (A->G), that marks isolates from Sierra Leone and Guinea. In addition, several Sierra Leone members of this subgroup are closely related to the reference sequence, while others are more distantly related, as seen by the presence of many SNP differences with the reference genome. Many sublineage A strains, described by [31] contain both an adenosine nucleotide at position 10218 and an additional A->G substution, at position 10273. Zaire strains with a G at position 10218 include the following: 1) all early sublineage Guinea-1a, 2) all Liberia strains, indicating their early origin during the course of the epidemic, 3) many Sierra Leone and Guinea isolates, both closely or distantly related to the reference sequence, and 4) a group of isolates from Guinea that contained an additional marker at position 10248 (T->C) (S7 Fig).
Host-cell adenosine deaminases that act on RNA (ADARs) modify RNAs by converting adenosine bases to inosines (for review, [53]). When ADARs edit a Filovirus replicative template, the viral polymerase interprets inosines as guanines, resulting in the negative stranded RNA genome having a cytosine instead of an uracil base at the modified or edited position. ADAR editing has been detected in both Marburg and Ebola isolates (for review, [44]).
Whole-genome EvoDifference prints of related Marburg strains have revealed multiple T/U -> C base substitution clusters within non-coding regions and in protein encoding sequences of individual strains (Fig 9 and S8 Fig). For example, the intra-lineage comparison of the Marburg_lin2_Popp_Cercopithecus_Human_Z29337.1_1967 strain with the Marburg_lin2_LakeVictoria_GQ433353.1_2011 isolate identified a cluster of T -> C base differences within the NP gene 3’UTR and flanking VP35 intragenic region (Fig 9A). All 40 base differences in the 556 base non-coding region (bases 2,282 through 2,838) were identified as T/U -> C substitutions in the Lake Victoria strain. Examination of Marburg strains revealed examples of putative T/U -> C base editing in the VP35 and VP40 ORFs. Our analysis identified two strains with clusters of T/U -> C within the VP35 coding sequences. The first, found in Marburg_lin3_Ang_KM2601523.1 is illustrated in S8A Fig. The twenty-three T/U -> C base substitutions span 289 bases, 19 of which are in the intragenic region between the NP and VP35 coding regions; four additional substitutions are within the VP35 ORF and two of these resulted in nonsynonymous amino acid changes. The second example of A to I editing within the VP35 ORF was found in Marburg_lin9_Kenya_EU500826.1_1987 (S8B Fig). The T/U -> C substitutions resulted in 5 amino acid changes. Although overlapping in distribution, the substitutions in these lin3 and lin9 strains occurred at different positions.
We also identified evidence of T/U -> C substitutions in VP40 coding sequences. Within Marburg_lin9_Kenya_EU500828.1, fourteen T/U -> C base substitutions were identified (S8C Fig); 13 bases fell within the VP40 coding sequence. These substitutions resulted in four nonsynonymous amino acid changes. A second example identified in the VP40 coding sequence of Marburg_lin9_Kenya EU500826.1_1987 is illustrated in S8D Fig. These substitutions resulted in three amino acid changes.
Our search of Ebola genomes also uncovered clusters of T/U -> C base substitutions in Zaire and Bundibugyo strains. Within three Zaire_lin6_Port Loko_2015 isolates, we found identical T/U -> C patterns in the 3’ UTR of their NP genes (Fig 9B). The fact that the 3 isolates have the same T/U -> C substitutions indicates that the editing most likely occurred in a previous generation of this subgroup and was not a product of in vitro cell culture passage. In the Bundibugyo lineage, evidence of A-to-I editing was detected within the GP Mucin domain encoding sequence of the Bundibugyo_lin1_Uga_FJ217161.1_2008 strain (Fig 9C). Prior to full genomic sequencing, the GP gene from this Bundibugyo strain was sequenced from a patient serum-derived PCR product [21], indicating that the putative editing occurred in vivo. Of note, four of the 12 T/U -> C substitutions in the Bundibugyo_lin1_Uga_FJ217161.1_2008 mucin-like domain encoding sequence result in amino acid codon changes, suggesting that A-to-I editing may contribute to the antigenic diversity of the Filovirus spike proteins.
The methodology and databases described here represent a new set of alignment tools for the rapid comparative analysis of a Flavivirus or Filovirus sequence. By superimposing alignment data of either one or up to hundreds of strains onto the user’s input sequence, uninterrupted readouts enable the following; 1) surveillance of lineage complexity within viral outbreaks, 2) the identification of unique base substitutions within the input sequence and/or database genomes, 3) the identification of recombinant strains, and 4) superimposed alignments highlight conserved sequence elements and allow for the identification of viral genomes that have been modified by host cell editing.
EvoPrinter should not be considered a stand-alone application for the analysis of Flavi or Filovirus evolution. We recommend that it’s search algorithms be used in conjunction with other tools that employ different sets of comparative analysis stratagies. For example, while EvoPrinter resolves sublineage markers and isolate-specific SNPs, other phylogenetic analysis programs provide information concerning lineage progression and diversification (e. g. [12–14]). Our strategy of detecting recombinants, using differential SNP patterns, is also complementary to other tools such as the multi-genome Recombination Detection Program that identifies recombinant fragments in graphic readouts [52]. When used together with these other tools, EvoPrinter should prove to be an important addition for the genetic surveillance of these evolving pathogens.
|
10.1371/journal.pgen.1004428 | BLMP-1/Blimp-1 Regulates the Spatiotemporal Cell Migration Pattern in C. elegans | Spatiotemporal regulation of cell migration is crucial for animal development and organogenesis. Compared to spatial signals, little is known about temporal signals and the mechanisms integrating the two. In the Caenorhabditis elegans hermaphrodite, the stereotyped migration pattern of two somatic distal tip cells (DTCs) is responsible for shaping the gonad. Guidance receptor UNC-5 is necessary for the dorsalward migration of DTCs. We found that BLMP-1, similar to the mammalian zinc finger transcription repressor Blimp-1/PRDI-BF1, prevents precocious dorsalward turning by inhibiting precocious unc-5 transcription and is only expressed in DTCs before they make the dorsalward turn. Constitutive expression of blmp-1 when BLMP-1 would normally disappear delays unc-5 transcription and causes turn retardation, demonstrating the functional significance of blmp-1 down-regulation. Correct timing of BLMP-1 down-regulation is redundantly regulated by heterochronic genes daf-12, lin-29, and dre-1, which regulate the temporal fates of various tissues. DAF-12, a steroid hormone receptor, and LIN-29, a zinc finger transcription factor, repress blmp-1 transcription, while DRE-1, the F-Box protein of an SCF ubiquitin ligase complex, binds to BLMP-1 and promotes its degradation. We have therefore identified a gene circuit that integrates the temporal and spatial signals and coordinates with overall development of the organism to direct cell migration during organogenesis. The tumor suppressor gene product FBXO11 (human DRE-1 ortholog) also binds to PRDI-BF1 in human cell cultures. Our data suggest evolutionary conservation of these interactions and underscore the importance of DRE-1/FBXO11-mediated BLMP-1/PRDI-BF1 degradation in cellular state transitions during metazoan development.
| The migratory path of DTCs determines the shape of the C. elegans gonad. How the spatiotemporal migration pattern is regulated is not clear. We identified a conserved transcription factor BLMP-1 as a central component of a gene regulatory circuit required for the spatiotemporal control of DTC migration. BLMP-1 levels regulate the timing of the DTC dorsal turn, as high levels delay the turn and low levels result in an early turn. We identify and characterize upstream regulators that control BLMP-1 levels. These regulators function in two ways, i.e. by destabilization of BLMP-1 through ubiquitin-mediated proteolysis and by transcriptional repression of the blmp-1 gene to down-regulate BLMP-1. Interestingly, blmp-1 also negatively controls these regulators. Our data suggest that a dietary signal input acts together with a double-negative feedback loop to switch DTCs from the “blmp-1-on” to the “blmp-1-off” state, promoting their dorsal turn. Furthermore, we show that some protein interactions in the circuit are conserved in C. elegans and humans. Our work defines a novel function of the conserved blmp-1 gene in the temporal control of cell migration, and establishes a gene regulatory circuit that integrates the temporal and spatial inputs to direct cell migration during organogenesis.
| Cell migration is important for organogenesis and development of animals. Numerous extracellular guidance cues and receptors for the spatial control of cell migration have been identified and characterized [1]. However, little is known about the temporal regulation of cell migration and how the spatial and temporal signals are coordinated to generate a specific and reproducible pattern of cell migration during development.
The bilobed gonad of C. elegans hermaphrodites develops from a four-cell primordium positioned in the ventral midbody [2]. The shape of the two symmetrical U-shaped gonadal arms is determined by the migratory paths of the two distal tip cells (DTC), leader cells found at the tip of each arm [3]. The DTCs undergo three sequential phases of migration and re-orient twice during the three larval developmental stages, thus providing a paradigm for the study of the spatio-temporal regulation of cell migration in vivo [2], [3]. In phase I during the L2 and early L3 stages, the anterior and posterior DTCs move centrifugally along the ventral body wall muscles towards the head or tail, respectively (Figure 1A). In phase II, they turn 90 degrees and move from the ventral to the dorsal muscles, then, during phase III, they again turn 90 degrees and move centripetally along the dorsal body wall muscles and halt in the mid-body. Both orthogonal turns occur during the late L3 stage. The timing of these turns is regulated by a redundant function of the heterochronic genes daf-12, dre-1, and lin-29 [4]. A single mutation in any of these three genes has no effect on DTC migration, but mutation of 2 or 3 of the genes delays the L3-specific DTC migration pattern, which fails to take place even in L4 or the adult. lin-29, daf-12, and dre-1 encode, respectively, a zinc-finger transcription factor, a steroid hormone receptor similar to the vertebrate vitamin D and liver-X receptor, and an F-Box protein of an SCF ubiquitin ligase complex [4]–[6], indicating that a complex mechanism, involving steroid hormone signaling, gene transcription, and protein degradation, is responsible for the temporal control of the dorsal turn. However, how these three genes function to do so is unclear.
The dorsal migration of DTCs is regulated by the guidance receptors UNC-5 and UNC-40 (a homolog of Deleted in Colorectal Cancer) [7]–[9], which drive DTCs to move away from the ventrally localized UNC-6 to the dorsal side [10], [11]. Dorsally localized UNC-129/TGF-β also promotes DTC dorsal migration through UNC-5 and UNC-40 receptors [12]. Mutations in these genes disrupt the ventral-to-dorsal migration of DTCs. unc-40 appears to be transcribed in the DTCs throughout their migration [7], whereas unc-5 is transcriptionally up-regulated at the time when the dorsal turn is initiated [13]. Precocious expression of unc-5 cDNA at the time when DTCs have not yet turned dorsalward induces unc-6-dependent precocious dorsal migration [13]. These data show that unc-5 is both necessary and sufficient for DTC dorsal migration and that the increase in unc-5 transcription is responsible, at least in part, for the initiation of DTC dorsal migration [8], [9], [13]. However, how unc-5 is temporally regulated to direct dorsal migration precisely at the late L3 stage is unclear.
In this study, we performed genetic screening for mutants defective in DTC migration and isolated blmp-1 mutants. Our results showed that blmp-1 is a heterochronic gene that acts with daf-12, dre-1, and lin-29 in a regulatory circuit to control the correct timing of DTC migration. We also showed that this regulatory circuit is, at least in part, conserved in C. elegans and humans.
To identify genes that are important for the spatiotemporal regulation of DTC migration, we performed genetic screening for mutants defective in DTC migration, as described in the Materials and Methods, and isolated alleles tp5 and tk41. A genetic complementation test showed that tp5 and tk41 were allelic, and that either allele failed to complement the previously identified mutation dpy-24(s71) (dpy stands for dumpy, shorter than wild-type) [14].
We mapped dpy-24(s71) to chromosome I, near stp124, by sequence tag site (STS) mapping [15] (Figure S1A). Three factor mapping using unc-40 and unc-75 and subsequent SNP mapping positioned dpy-24(s71) within the region between cosmids F45H11 and F37D6. Cosmids covering this region were microinjected into the dpy-24(s71) mutant, and cosmid F25D7 rescued the DTC migration defect (Figure S1B). The genomic DNA fragments corresponding to the 5 predicted open reading frames of F25D7 were individually amplified by long PCR and tested for their ability to rescue the dpy-24(s71) mutant and only F25D7.3, which contained a single blmp-1 gene, had a rescue effect (Figure S1B). In addition, F25D7.3 RNA interference (RNAi) phenocopied the dpy-24(s71) mutant (Table 1). These results demonstrated that F25D7.3 corresponded to the blmp-1 gene (named for its sequence similarity to mouse Blimp-1, see below) [16].
The blmp-1 gene encodes a protein with 27% identity to mouse B lymphocyte-induced maturation protein 1 (Blimp-1) and 26% identity to human positive regulatory domain I-binding factor (PRDI-BF1) (Figure S2). Both Blimp-1 and PRDI-BF1 are thought to act predominantly as transcription repressors and are essential for the terminal differentiation of B and T cells [17], [18]. As shown in Figure 2, BLMP-1, like Blimp-1 and PRDI-BF1, is predicted to contain a positive regulatory (PR) domain, a nuclear localization signal (NLS), and five Kruppel-type [(Cys)2-(His)2] zinc fingers. The zinc fingers of both Blimp-1 and PRDI-BF1 have been shown to bind to target DNA and are essential for their transcriptional repression activities [19]–[21].
Alleles s71, tk41, and tp5 have, respectively, a non-sense mutation in codon 281, 381, or 434, and are predicted to encode truncated proteins without zinc fingers (Figure 2 and S2). The deletion allele tm548, which was isolated by a reverse genetic approach (National Bioresource Project), has an 810 bp deletion, removing part of exon 3 and part of intron 3 (Figure 2 and S2) and may result in a truncated BLMP-1 protein containing the first 254 amino acids of BLMP-1 and 17 amino acids encoded by the third intron.
The four blmp-1 mutants, s71, tk41, tp5, and tm548, were found to have a similar set of defects, including a DTC migration abnormality (shown for blmp-1(s71) in Figure 1), a weak dumpy phenotype (shown for blmp-1(s71) in Figure S3A), and a partially penetrant embryonic lethality (shown for all four in Figure S3B). Wild-type blmp-1 genomic DNA rescued the dumpy phenotype and embryonic lethality of the blmp-1(s71) mutant (Figure S3B), showing that these defects were caused by the loss of blmp-1.
The DTC migration patterns of these mutants were varied, but shared the common feature that the mutant DTCs had a shorter centrifugal phase I migration path and executed the dorsal turn at a point closer to the mid-body than wild-type DTCs (Figure 1B–G, Table S1), suggesting either slower movement and/or precocious initiation of the dorsal turn. However, we timed the movement of the DTCs during phase I migration and found that the blmp-1 mutant DTCs did not migrate significantly slower than the wild-type DTCs (Table S2). In addition, some DTCs migrated obliquely with respect to the dorsal-ventral axis until they reached the dorsal muscle (Figure 1D); such a migratory route is probably due to the simultaneous execution of centrifugal phase I and dorsal phase II migrations, support for a precocious execution of the dorsal turn at the time when phase I migration normally occurs.
To examine whether the abnormal DTC migration pattern of the blmp-1 mutants was indeed caused by a precocious dorsal turn, we performed a time-course analysis of DTC migration in the wild-type and blmp-1(s71) mutant, using the division stages of the vulval precursor cell P6.p as temporal developmental markers, as described previously [22]. P6.p is generated in mid L1, undergoes three rounds of cell division during L3, and gives rise to eight descendants that constitute the vulva [23]. Figure 3A shows representative DIC images of wild-type (a, b, e, f) and blmp-1(s71) (c, d g, h) posterior gonadal arms in early L3 (top panels) and late L3 (bottom panels). Figure 3B shows that, in the wild-type, the DTCs in more than 90% of worms underwent ventral-to-dorsal migration at the four-P6.p cell stage and the DTCs in none made a dorsal turn before P6.p cell division, whereas, in the s71 mutant, the anterior DTCs in 36% of worms and the posterior DTCs in 66% of worms had turned dorsalward before P6.p divided. These results demonstrate that loss of blmp-1 causes a precocious initiation of DTC dorsal migration (early L3, rather than late L3) and that blmp-1 functions to prevent a precocious DTC dorsal turn.
To determine the localization of BLMP-1 and explore its function, we raised polyclonal antibodies against bacterially expressed recombinant BLMP-1 (Materials and Methods) and used the affinity-purified antibodies to stain whole-mount animals. The results showed that BLMP-1 was localized in the nucleus and was detected in hypodermal, vulval, and intestinal cells (Figure 4A), as well as DTCs (Figure 4B). We next stained the mutant embryos with the blmp-1 mutations (s71, tm548, tp5 and tk71), which are loss-of-function recessive alleles by genetic tests (Materials and Methods). Little or no signals were detected in these mutant embryos as shown in representative images in Figure 4C, demonstrating the specificity of the antibodies. Notably, BLMP-1 was seen in DTCs prior to the mid L3 larval stage (two P6.p-descent cells), but not during, or after, mid L3 stage, after the DTCs had undergone the dorsal turn (Figure 4B). This result and the precocious dorsal turn phenotype of the blmp-1 mutant support a model in which BLMP-1 functions in DTCs before mid L3 stage to prevent these cells from undergoing a precocious dorsal turn.
The localization of BLMP-1 in DTCs supports the model that BLMP-1 controls the timing of DTC dorsal migration in a cell-autonomous fashion. We then tested whether expression of blmp-1 cDNA in DTCs under the control of the lag-2 promoter Plag-2 is sufficient to rescue the DTC migration defect of the blmp-1 mutant. Plag-2 drives gene expression in DTCs, but not in body wall muscle or hypodermis, the structures on which DTCs migrate [24]. In blmp-1 mutants carrying the Plag-2 blmp-1 transgene, the percentage of worms with abnormal anterior or posterior DTC migration was reduced, respectively, from 82% to 20% and from 93% to 23% (Figure 1G), similar levels to those seen when blmp-1 cDNA was expressed under the control of the blmp-1 endogenous promoter Pblmp-1 (40% or 19% for the anterior and posterior DTC, respectively; Figure 1G). This result further supports a cell-autonomous role of blmp-1 in DTC migration.
Next, we tested whether blmp-1 may prevent dorsalward turning of DTCs by regulating a dorsal-ventral guidance system at the early larval stage. The dorsal migration of DTCs is regulated by the guidance receptors UNC-5 and UNC-40 (a homolog of Deleted in Colorectal Cancer) [7]–[9], which drive DTCs to move away from the ventrally localized UNC-6 [10]–[11]. The observations that blmp-1, unc-5 and unc-40 function cell-autonomously in the control of DTC migration [7], [9] (Figure 1G) suggest that blmp-1 may prevent DTC precocious dorsalward turning by regulating unc-5 and/or unc-40. unc-40 appears to be transcribed in the DTCs throughout their migration [7], whereas unc-5 is transcriptionally up-regulated at the time when the dorsal turn is initiated [13]. Using the transgene Punc-5(1 kb) gfp, in which an approximately 1 kb sequence upstream of the unc-5 coding sequence was used to drive the GFP reporter, we confirmed this unc-5 expression pattern, i.e. unc-5 was expressed during and after, but not before, the DTC dorsal turn (Figure S4). Interestingly, this temporal pattern of unc-5 transcription is complementary to that of BLMP-1 expression in DTCs, as BLMP-1 was present in DTCs only before the dorsal turn (Figure 4B). This raised the possibility that BLMP-1 inhibits unc-5 transcription and thus prevents DTCs from turning dorsalward. If this were the case, the correct timing of the disappearance of BLMP-1 from DTCs would alleviate this transcriptional inhibition and allow unc-5 transcription and the DTC dorsal turn. We tested this hypothesis by altering the temporal expression pattern of blmp-1 and examining their effects on unc-5 transcription.
At the early L3 stage when the P6.p cell has not yet divided, no wild-type worms contained DTCs that had turned dorsalward or showed any expression of the Punc-5(1 kb)::gfp transgene (Figure 5Ba), whereas, in some blmp-1(s71) mutants, DTCs at the same stage had performed the dorsal turn and displayed precocious unc-5 transcription (Figure 5Bb). Thus, loss of blmp-1 causes precocious unc-5 expression, which coincides with the precocious dorsalward turning of DTCs.
Next, we tested whether the blmp-1 precocious dorsal turn phenotype required unc-5 by examining and comparing the DTC dorsal migration patterns of the blmp-1 and unc-5 single mutants and the blmp-1; unc-5 double mutant. Approximately 48% anterior DTCs and 83% posterior DTCs failed to turn dorsalward in the unc-5(e53) mutant (Table S3). No precocious dorsal turn was observed in the blmp-1 (s71); unc-5 double mutant, showing that the unc-5(e53) mutation blocked the precocious dorsal turn phenotype of the blmp-1 (s71) mutant and that the blmp-1 precocious dorsal turn phenotype required unc-5.
Because BLMP-1 is present in the DTC before the dorsal turn, we examined the significance of blmp-1 down-regulation and its effect on the initiation of DTC dorsalward turning. To this end, we overexpressed blmp-1 using the transgene Pblmp-1::blmp-1::gfp, in which the fusion protein BLMP-1::GFP was expressed under the control of the promoter Pblmp-1. In the resulting transgenic line, 10% of transgenic worms had DTCs that displayed normal BLMP-1::GFP down-regulation, underwent a normal dorsal turn (Figure 5A, a, b), whereas 90% of transgenic worms showed persistent BLMP-1::GFP expression, even at the L4 stage, showing that BLMP-1::GFP was not appropriately down-regulated in these worms, and these worms showed no sign of dorsal turn (Figure 5A, d, e). This retarded turn phenotype was in contrast to the precocious dorsal migration phenotype caused by loss of blmp-1 (Figures 1C–G). Collectively, these results show that the BLMP-1 level is important regulation for DTC dorsal migration and that the timely disappearance of BLMP-1 allows the DTCs to turn dorsalward, hence switching their migration phase from centrifugal phase I migration to ventral-to-dorsal dorsal phase II migration.
To test whether blmp-1 overexpression might repress unc-5 expression in DTCs and thus resulted in the retarded turning phenotype, we expressed the transgene Pblmp-1::blmp-1::gfp in worms carrying the Punc-5(1 kb)::mCherry reporter. In the resulting transgenic line, 10% of transgenic worms had DTCs that displayed normal BLMP-1::GFP down-regulation, underwent a normal dorsal turn, and showed Punc-5(1 kb)::mCherry expression (Figure 5A, a–c), whereas 90% of transgenic worms showed persistent BLMP-1::GFP expression, and these worms showed no sign of dorsal turn or unc-5 expression (Figure 5A, d–f). Thus, constitutive blmp-1 expression in the late larval stage represses unc-5 transcription and blocks the DTC dorsal turn. These results confirmed the causal relationship of the reciprocal patterns of BLMP-1 expression and unc-5 transcription in DTCs and support the model in which BLMP-1 inhibits unc-5 transcription and thus prevents DTCs from turning dorsalward during early larval development.
The heterochronic genes daf-12, dre-1, and lin-29 function redundantly to specify the temporal identity of DTCs and prevent DTCs from undergoing retarded ventral-to-dorsal migration [4]. We tested the genetic interaction of blmp-1 with daf-12, dre-1, and lin-29 in the temporal control of DTC dorsal turn. For daf-12, we used the null allele rh61rh411. Because complete loss of dre-1 or blmp-1 caused lethality [4], [25] and the lin-29(n546); blmp-1(s71) double mutant was very sick and could not be maintained as homozygote, we used viable alleles and/or RNAi for dre-1, blmp-1 and lin-29 to analyze the genetic interactions of these genes. The precocious phenotype of the blmp-1(s71) mutant was partially suppressed by a single mutation of lin-29, dre-1, or daf-12 or combined mutations of any 2 or all 3 (Table 1). For example, 93% of blmp-1(s71) mutants had a precocious DTC migration defect, whereas 54% of the blmp-1(s71); lin-29(RNAi) double mutants, 73% of the blmp-1(s71); dre-1(dh99) double mutants, and 31% of the blmp-1(s71);daf-12(rh61rh41) double mutants displayed a precocious turn defect. This suggests that the blmp-1 precocious phenotype may require the activity of lin-29, dre-1, or daf-12, and that blmp-1 may function upstream of, or in parallel with, lin-29, dre-1, and daf-12 to prevent a precocious DTC dorsal turn during early larval development. Conversely, the blmp-1 mutation also partially suppressed the retarded dorsal migration of the double or triple mutants of lin-29, dre-1, and daf-12 (Table 1). For example, 98% of the dre-1(dh99); daf-12(rh61rh411) double mutants showed retarded DTC migration, whereas 12%, 73%, and 15% of the blmp-1(71); dre-1(dh99); daf-12(rh61rh411) triple mutants showed, respectively, wild-type, precocious, or retarded DTC migration. This suggests that the retarded phenotype of the lin-29, dre-1 and daf-12 double and triple mutants may require blmp-1 activity and that lin-29, dre-1, and daf-12 may function upstream of, or in parallel with, blmp-1 to prevent a delay in the DTC dorsal turn during the late larval stage. Thus, two distinct regulatory hierarchies in these heterochronic genes are likely employed to specify the temporal identities of DTCs at the early and late larval stages, and the switch from the “blmp-1-on” state to the “blmp-1-off” state during developmental progression may determine the timing of the DTC dorsal turn.
The observation that mutation of daf-12, dre-1, or lin-29 partially suppressed the blmp-1-related precocious DTC dorsal migration defect (Table 1) prompted us to examine whether blmp-1 negatively regulated the transcription of daf-12, dre-1, or lin-29. Using chromatin immunoprecipitation coupled with high-throughput DNA sequencing (ChIP-seq), the modENCODE Consortium has shown that BLMP-1 binds to the upstream sequence of lin-29, but not that of daf-12 or dre-1, at the early larval stage [26]. This suggested that BLMP-1 might negatively regulate lin-29 expression. To test this possibility, we analyzed and compared lin-29 transcription in the wild-type and blmp-1 mutants using the transcriptional reporter Plin-29::gfp, in which gfp was expressed under the control of the lin-29 promoter Plin-29 (Figure 5Ca). We confirm that the transcription of lin-29 starts at approximately mid L3 stage and continues during L4 [4], [27]. We found that knockdown of blmp-1 using RNAi resulted in precocious expression of Plin-29::gfp at the L2 stage (Figure 5Cb,c), showing that blmp-1 represses lin-29 transcription at the L2 stage.
To investigate the significance of this blmp-1-mediated lin-29 down-regulation during DTC phase I migration in L2, we precociously expressed lin-29 under the control of the Plag-2 promoter using the Plag-2::lin-29 transgene. In the wild-type animals carrying this transgene, 27% and 9.5% of the anterior or posterior DTCs, respectively, underwent a precocious dorsal turn (Figure 1G). This result shows the functional importance of blmp-1-mediated repression of lin-29 transcription in preventing precocious DTC dorsal migration during early larval development.
The observation that the constitutive expression of blmp-1 in DTCs throughout larval development prevented them from performing the dorsal turn (Figure 5A) highlights the importance of BLMP-1 down-regulation in promoting the dorsal turn. Two observations suggested that daf-12, lin-29, and dre-1 might be responsible for BLMP-1 down-regulation. First, like worms constitutively expressing blmp-1 (Figure 5Ad–f), the double and triple mutants of daf-12, lin-29, and dre-1 had a retarded DTC migration phenotype (Table 1). Second, loss of blmp-1 partially suppressed the retarded phenotype of the double and triple mutants (Table 1), showing that the defect in the double and triple mutants required blmp-1 activity. We therefore examined BLMP-1 levels in mutants defective in daf-12, lin-29 and/or dre-1 using immunostaining with anti-BLMP-1 antibody.
As mentioned above, in the wild-type control, BLMP-1 was not detected in DTCs at the L4 stage (Figure 4B) and a similar result was observed in the daf-12, dre-1, and lin-29 single mutants. In contrast, persistent BLMP-1 expression in L4 stage was seen in DTCs from the double mutants dre-1(dh99); lin-29(n546) (100% of 30 worms scored), dre-1(dh99); daf-12(rh61rh411) (20% of 60 scored), and lin-29(RNAi); daf-12(rh61rh411) (∼5% of 106 scored) (examples shown in Figure 4D). The intensity of the persistent BLMP-1 signal was much weaker in the lin-29(RNAi); daf-12(rh61rh411) double mutant than in the other double mutants and bleached too quickly to be photographed by our imaging system. The daf-12(rh61rh411) allele is molecular null, while lin-29(n546) and dre-1(dh99) are partial loss-of-function [4], [5]. These results suggest that DRE-1 acts together with transcription factor LIN-29 or, to a lesser extent, with transcription factor DAF-12 for efficient down-regulation of BLMP-1 and correct timing of the control of DTC dorsal migration.
To determine how BLMP-1 was down-regulated by lin-29, daf-29, and dre-1, we used two gfp reporters in an in vivo expression assay to examine whether blmp-1 expression was repressed at the transcriptional level by transcription factors LIN-29 and DAF-12 and at the post-transcriptional level by the F-Box protein DRE-1. We first generated a transcriptional reporter Pblmp-1::dgfp (Figure 6A), in which dGFP (destabilized GFP) was controlled by the Pblmp-1 promoter. dGFP contains a PEST sequence and has a shorter half-life than normal GFP [28] and therefore allows sensitive detection of promoter activity. In animals carrying the integrated Pblmp-1::dgfp transgene, the DTCs of 75.6% of worms expressed GFP in the early and mid L3 stages, while only 4.3% expressed GFP in the late L3 and L4 stages (Figure 6B and 6C), suggesting that blmp-1 transcription is strongly repressed during, and after, late L3 stage.
Next, we tested whether daf-12 or lin-29 was responsible for blmp-1 transcriptional repression by examining Pblmp-1::dgfp transgene expression in the absence of daf-12 and/or lin-29. As shown in Figure 6C, in early and mid L3, loss of either daf-12 or lin-29 did not affect the percentage of worms with DTCs expressing dgfp. However, during the late L3 and L4 stages, loss of lin-29, but not daf-12, increased the percentage of worms with DTCs expressing dGFP in late L3 and L4. For example, only 4.3% or 2.4% of wild-type or daf-12 mutant worms, respectively, had DTCs expressing dGFP, whereas 39.6% of lin-29(RNAi) worms had dGFP-expressing DTCs. This result shows a differential requirement for lin-29 (stronger) and daf-12 (weaker) for blmp-1 transcriptional repression during, and after, the late L3 stage (Figure 6C). In addition, 78.4% of the daf-12(rh61rh411); lin-29(RNAi) double mutants had DTCs expressing dGFP (Figure 6C). These results suggest that daf-12 plays a non-essential, but auxiliary, role in the repression of blmp-1 transcription.
We also tested the involvement of dre-1 in blmp-1 transcriptional repression. No effect on the Pblmp-1::dgfp transcription level was observed when the dre-1(dh99) mutation was introduced into the wild-type or the daf-12mutant (Figure 6C). Thus, dre-1 probably plays no role in blmp-1 transcriptional repression. Interestingly, blmp-1 seems to be required for its own expression. During early and mid L3 stages, approximately 75% of the wild-type worms had DTCs expressing dGFP from the Pblmp-1::dgfp transgene, but only 9% of the blmp-1(s71) worms had DTCs expressing dGFP.
Previous genetic and biochemical data have shown that the F-Box protein DRE-1 functions in an SCF ubiquitin ligase complex, which contains CUL-1 (a cullin scaffold protein), SKP-1 (a SKP1-like adaptor that binds to the F-Box protein), and RBX-1 (an RBX ring finger that bridges to the E2 ubiquitin conjugating enzyme), and is important for the temporal control of somatic and gonad development [4] and the timing of tail spike cell death [29]. This raised the possibility that DRE-1 may regulate BLMP-1 stability through ubiquitin-mediated proteolysis. Like dre-1 RNAi, RNAi for skp-1, rbx-1, or cul-1 caused a retarded DTC migration phenotype in the daf-12(rh61rh411) single mutant [4] (Table S4), consistent with a model in which DRE-1 targets protein(s) for proteolysis in an SCF ubiquitin ligase complex during the temporal regulation of the DTC dorsal turn.
Using the transgene Plag-2::gfp::blmp-1, in which BLMP-1 was tagged with GFP and expressed under the control of the Plag-2 promoter (Figure 7A), we next examined whether DRE-1 destabilized BLMP-1. Two independent transgenic lines carrying an extrachromosomal transgene array were generated. In line 1, approximately 27% of worms expressed GFP in DTCs during early L3 and the percentage decreased to 16.9% during mid L3 to L4, showing approximately 38% down-regulation, and a similar level of down-regulation (35.5%) was observed in line 2 (Figure 7Aa,b and C). In contrast, no down-regulation was observed in the control line carrying the integrated transgene qIs56[Plag-2::gfp], in which gfp was expressed under the control of the Plag-2 promoter (Figure 7Ba,b and C). These results show that BLMP-1 levels drop significantly when entering the mid L3 stage. Although the difference may be due to differences in sensitivity of the gfp reporters, given the previous result that blmp-1 transcriptional repression was detected in late, but not mid, L3 stage, this suggests that BLMP-1 degradation occurs approximately 3 hours earlier than transcriptional repression.
Next, we examined whether dre-1 was required for this decrease in GFP::BLMP-1 fusion protein levels. In two independent transgenic lines of the dre-1 mutant, the percentage of worms with DTCs expressing GFP::BLMP-1 during the early L3 stage was similar to that seen during the mid L3-to-L4 stage (Figure 7Ac,d, and C), showing that dre-1 is required for GFP::BLMP-1 degradation. Since loss of dre-1 completely blocked the decrease in GFP::BLMP-1 levels, it is unlikely that any other gene acts in a redundant fashion with dre-1 to regulate BLMP-1 stability.
The F-box protein of an SCF ligase complex binds to substrates and targets them for ubiquitin-mediated proteolysis [30]. To test the idea that BLMP-1 might be the target of DRE-1 in the SCFDRE-1 ligase complex, we examined whether DRE-1 and BLMP-1 physically interacted by co-immunoprecipitation in mammalian cell cultures. When Myc-tagged dre-1 and HA-tagged blmp-1 were transfected into HEK293T cells, then lysates were subjected to immunoprecipitation with anti-Myc antibody and Western blotting with anti-HA antibody, BLMP-1 was co-immunoprecipitated with DRE-1, while no signal was seen in the singly transfected controls (Figure 8A). The reciprocal experiment using immunoprecipitation with anti-HA antibody and Western blotting with anti-Myc antibody showed that DRE-1 was co-immunoprecipitated with BLMP-1 (Figure 8B). These results demonstrate that DRE-1 binds to BLMP-1 in HEK293T cells.
We next examined whether FBXO11 and PRDI-BF1, the respective human orthologs of DRE-1 and BLMP-1, also associate in HEK293T cells and found that Myc-tagged FBXO11 was co-immunoprecipitated with HA-tagged PRDI-BF1 in lysate of cells co-transfected with both, but not in lysate of cells transfected with only Myc-tagged FBXO11 (Figure 8C). This result shows association of FBXO11 and PRDI-BF1 in human cell cultures. In addition, HA-tagged PRDI-BF1 also pulled down endogenous CUL1 in co-immunoprecipitation experiments in HEK293T cells (Figure 8C), demonstrating the association of FBXO11, BLMP-1, and CUL1 in an SCF complex and suggesting that FBXO11 may regulate PRDI-BF1 stability through ubiquitin-mediated proteolysis. These results demonstrate a conserved interaction between DRE-1/FBXO11 and BLMP-1/PRDI-BF1 in both humans and C. elegans. The regulation of BLMP-1/PRDI-BF1 stability by DRE-1/FBXO11 in an SCF ubiquitin ligase complex may also be conserved in evolution.
In this study, we identified a conserved transcription factor, BLMP-1, as an essential component of the heterochronic hierarchy during gonad development, and provided evidence that BLMP-1 levels are critical for the timing of DTC dorsal migration during gonadogenesis. BLMP-1 was present in DTCs only before the dorsal turn and disappeared when the DTC was about to make the dorsal turn. Loss of blmp-1 resulted in precocious unc-5 transcription and DTC dorsal migration, whereas constitutive expression of blmp-1 delayed both events. These data show that blmp-1 controls the temporal identity of DTCs by preventing them from undergoing a precocious dorsal turn. In addition, our results provide a molecular mechanism by which daf-12, dre-1, and lin-29 promote the correct timing of DTC dorsal turn by timely down-regulation of blmp-1. In this model, lin-29 and daf-12 repress blmp-1 transcription, which abolishes the synthesis of blmp-1 mRNA, while dre-1 mediates BLMP-1 degradation. Together, these two negative regulatory systems result in the efficient elimination of blmp-1 activity, and thus alleviate the repression of the DTC dorsal turn.
Interestingly, daf-12, dre-1, and lin-29 appeared to contribute to BLMP-1 down-regulation to different extents. For example, in the dre-1; lin-29, dre-1; daf-12, and lin-29; daf-12 double mutants, 100%, 20%, or 5% of worms, respectively, showed persistent BLMP-1 expression beyond late L3 in the immunostaining experiment. This suggests that, of these three genes, loss of dre-1 has the strongest effect on BLMP-1 down-regulation and loss of daf-12 has the weakest effect. Using transcriptional and translational gfp reporters, it appeared that the DRE-1-dependent protein degradation process occurred at the mid L3 stage, slightly earlier than the transcriptional repression mediated by lin-29 and daf-12 (late L3 stage), although this difference could be due to different sensitivities of the gfp reporters. The notion that lin-29 is more significant than daf-12 in the repression of blmp-1 was further supported by the observation that, using the Pblmp-1::dgfp reporter, loss of lin-29 increased blmp-1 transcription in the late L3 to L4 stages, but loss of daf-12 had no detectable effect on blmp-1 expression (Figure 6C). Further experiments are necessary to test whether LIN-29 may inhibit blmp-1 transcription by directly binding to the blmp-1 genomic sequence.
It is intriguing that only 20% of the dre-1; daf-12 double mutants had DTCs with detectable persistent BLMP-1, whereas 98% of these double mutants showed the retarded phenotype (Table 1). It is possible that, in the mutant, BLMP-1 persists at a low level beyond the detection limit of our system, but sufficient to block lin-29 transcription and therefore abolish the DTC dorsal turn in the absence of daf-12. On the other hand, the retarded DTC migration phenotype of the lin-29; daf-12 double mutant may not be caused solely by persistent BLMP-1 at the undetectable low level, but also by the loss of both transcription activators lin-29 and daf-12, which function in a redundant fashion in the transcriptional activation of unc-5 (see below and Figure S5).
As shown in Figure 5 A, BLMP-1::GFP in worms carrying Pblmp-1::blmp-1::gfp was present and persisted beyond late L3 stage and the transgenic worms frequently showed a retarded DTC migration defect. A similar result was observed in worms carrying Plag-2::blmp-1::gfp (Huang and Wu, unpublished results). Interestingly, wild-type worms carrying Plag-2::blmp-1 had normal DTC migration (Figure 1G), suggesting that BLMP-1 in the transgenic worms was properly degraded beyond L3 stage. These data and the observation that GFP::BLMP-1 expressed from the transgene Plag-2::gfp::blmp-1 was degraded suggest that GFP tag at N-terminus or C-terminus make a difference in GFP fusion BLMP-1 degradation.
Using ChIP-seq, the modENCODE Consortium showed that BLMP-1 binds to the upstream sequence of unc-5 in vivo [26], suggesting that it might inhibit unc-5 expression by binding to the regulatory region of the gene. In addition, using the Punc-5(4.6 kb)::gfp transgene, we found that inactivation of either daf-12 or lin-29 slightly reduced unc-5 transcription, while loss of both genes completely abolished it (Figure S5). These results are consistent with the notion that the transcription factors DAF-12 and LIN-29 act together in a redundant fashion to activate unc-5 transcription and promote the DTC dorsal turn. The consensus DAF-12 binding sequence can also be identified in the unc-5 promoter (T.F. Huang and Y.C. Wu, unpublished results), raising the possibility that DAF-12 may directly bind to unc-5 for its transcriptional activation.
The DTC dorsal turn can be characterized by a specific transition in which the DTC moves irreversibly from phase I centrifugal migration into phase II dorsal migration. Phase I centrifugal migration occurs in the “blmp-1-on” state (high BLMP-1 and low UNC-5 levels) and phase II dorsal migration occurs in the“blmp-1-off” state (low BLMP-1 and high UNC-5 levels). A double-negative feedback loop, in which two genes mutually repress each other directly or indirectly, is commonly utilized to generate switch-like bistable responses during the progression of cellular and developmental processes [31]. Our data suggest that lin-29 and blmp-1 act in a double negative feedback loop that helps maintain DTCs in one of the bistable “blmp-1-on” and “blmp-1-off” states. On the basis of our results, we propose a molecular model for the switch-like process of the DTC dorsal turn. In L2 and early L3, BLMP-1 is expressed and represses lin-29 transcription (Figure 9). The double-negative feedback loop keeps DTCs in the “blmp-1-on” state, thus repressing DTC dorsal migration. In late L2, the hormones known as dafachronic acids (DA) are generated from dietary cholesterol and bind to the ligand binding domain of DAF-12 [32]. The DAF-12-DA complex promotes the L2-to-L3 transition [32], but is insufficient to repress blmp-1 transcription. During early to mid L3, dre-1 expression is initiated [4], which reduces BLMP-1 levels, probably through ubiquitin-mediated proteolysis. The decrease in BLMP-1 levels shifts the steady state toward low BLMP-1 levels and high LIN-29 levels (blmp-1 off). In late L3, accumulation of LIN-29 locks the DTCs in the “blmp-1-off” state through the negative feedback loop, thus allowing DTCs to switch to dorsal migration. This gene regulatory circuit integrates the temporal and spatial signals and coordinates with overall development of the organism to direct DTC dorsal migration during organogenesis.
Expression of lin-29 is negatively regulated by lin-42 [33]. LIN-42 is similar to the Period (Per) family of circadian rhythm proteins and functions as a member of the heterochronic pathway, regulating temporal cell identities [33], [34]. However, how lin-42 regulates the expression of lin-29 is not clear. lin-42 also genetically interacts with daf-12. During development, when conditions are unfavorable, due to starvation, crowding or high temperature, daf-12 promotes diapauses and formation of duaer larvae [5], [35], [36]. Previous studies show that daf-12 represses expression of lin-42 during dauer formation [37] and that lin-42 antagonizes the dauer-inducing signal from ligand-free DAF-12 [38]. However, whether daf-12 and lin-42 also regulate each other in the timing of DTC dorsalward turn during gonadaogenesis needs to be examined in the future.
In addition to ventral-to-dorsal migration, blmp-1 affects DTC migration along the AP axis. As shown in Figure 1D, some DTCs migrated obliquely with respect to the dorsal-ventral axis in blmp-1 mutants, suggesting simultaneous execution of centrifugal phase I and dorsal phase II movement. This result implies that the mutual exclusion of dorsalward and centrifugal migration in wild type animals was impaired in the blmp-1 mutants. Moreover, as shown in Figure 1E and 1F and Table S1, the reversal of migration direction was frequently observed in the centripetal phase III migration of blmp-1 mutants. Expression of unc-5 by the lag-2 promoter from transgene Plag-2::unc-5 resulted in DTC migration defects, including the reversal of the migration direction during the centripetal phase III migration (Huang, T.F. and Wu, Y.C., unpublished results), similar to those seen in blmp-1 mutants (Table S1). Further experiments are necessary to test whether the phase III DTC migration defect of blmp-1 mutants may be caused by an accumulative high level of unc-5 in the DTCs.
Previous studies have shown that blmp-1 mutants have a dumpy phenotype, a weak cuticle sensitive to oxidative stress, and defective adult alae, showing that blmp-1 is important for normal body length, oxidative stress resistance, and alae formation [39]. In addition, blmp-1 is important for pharynx and [40] male tail morphogenesis [41]. Under Nomarski optics, we confirmed the blmp-1 alae defect, as 40% of blmp-1(s71) mutants had incomplete alae and the rest had no alae in the adult stage (Figure S3C). The alae are continuous ridged cuticular structures synthesized by the lateral epidermal seam cells at the larval to adult (L/A) transition and serve as a specific marker of adult fate. The seam cells undergo asymmetric cell divisions at each larval stage and exit the cell cycle at the L/A transition. The blmp-1(s71) mutant was found to have a normal number of seam cells at the late L4 and early adult stage (Figure S3D), indicating that seam cell division was normal. These results and the abnormal alae phenotype suggest that blmp-1 is essential for the adult, but not the larval, fate of the epidermal seam cells.
The dre-1 mutant shows precocious formation of adult alae [4]. We therefore tested the genetic interaction between dre-1 and blmp-1 in the seam cell heterochronic hierarchy. About 8% of dre-1(dh99) mutants had adult alae at the early L4 stage (Figure S3D), confirming a role of dre-1 in preventing the adult fate of the seam cell [4]. At the adult stage, 35% of dre-1(dh99) mutants had incomplete alae (Figure S3Cc and D). Interestingly, the blmp-1; dre-1 double mutant had a weaker defect in alae formation than the blmp-1 single mutant, as 23% of double mutants and 60% of blmp-1 single mutants had no alae (Figure S3C and D). This result shows that dre-1 partially suppresses the blmp-1 alae abnormality and, therefore, blmp-1 may genetically act upstream of, or in parallel to, dre-1 in the specification of the adult seam cell fate, at least in terms of alae formation.
blmp-1(tm548 or s71) mutants have a weak uncoordinated movement phenotype [39]. Although blmp-1 expression has been previously observed in neurons using reporter transgene assays [26], we failed to detect BLMP-1 in neurons. It is possible that our anti-BLMP-1antibody system was not sensitive enough to detect a low amount of protein in neurons. Alternatively, the neuronal expression of blmp-1 revealed by the transgene may be ectopic and not reflect the endogenous expression pattern.
Like the blmp-1 mutations, lin-42 RNAi causes both a precocious DTC dorsal turn [33] and precocious lin-29 expression in L2, one stage earlier than in the wild-type [33]. However, the lin-42(RNAi) and blmp-1 mutations induce precocious dorsal migration at different larval stages, i.e. in L2 for lin-42(RNAi) [33] and early L3 in the case of blmp-1 (Figure 3). Similar to loss of blmp-1, precocious expression of lin-29 under the control of the lag-2 promoter caused the DTCs to undergo a precocious dorsal turn in early L3, but not L2. These results suggest that lin-42 regulates most, if not all, of the genes necessary for the DTC dorsal turn, while blmp-1 regulates the temporal expression of only a subset, including lin-29.
DRE-1 has two human orthologs FBXO10 and FBXO11, which are localized in the cytoplasm and nucleus, respectively [29]. FBXO10 and DRE-1 mediate, respectively, the degradation of anti-apoptotic protein Bcl2 or CED-9 to promote the death of a cell [29]. Recently, FBXO 11 has been shown to target the pro-oncogene product BCL6 for degradation and to be inactivated in diffuse large B cell lymphomas [42]. BCL6, a zinc finger transcription factor, regulates the transcription of a variety of genes involved in B cell development, differentiation, and activation [43], [44] and is overexpressed in the majority of patients with aggressive diffuse large B cell lymphoma [45]. Using the Clustal Omega multiple sequence alignment website http://www.ebi.ac.uk/Tools/msa/clustalo/, we found that Bcl6 shares sequence similarity with BLMP-1 (20%) and PRDI-BF1 (24%). These observations and our result showing that DRE-1/FBXO11 and BLMP-1/PRDI-BF1 are associated in human cell cultures indicate that DRE-1, FBXO10 and FBXO11 target different proteins for degradation in different cellular or developmental contexts, i.e. DRE-1 targets BLMP-1 and CED-9, FBXO10 targets Bcl2 and FBXO11 targets PRDF-BF1 and Bcl6. It will be interesting to determine how target specification is regulated.
C. elegans strains were cultured at 20°C on NGM agar inoculated with E. coli OP50, as described previously [46]. The N2 Bristol strain was used as the reference wild-type strain. The mutations used were as follows: LGI, blmp-1(s71, tk41, tm548, tp5), lin-29(n546); LGIII, unc-119(ed3); LGV, dre-1(dh99); LGX daf-12(rh61rh411). The blmp-1(tm548) mutant was generated and provided by the National Bioresource Project in Japan. Strains CB4856 and RW7000 were used for single-nucleotide polymorphism (SNP) mapping [15], [47]. Strain OS1841 carrying the transgene Plin-29::GFP was kindly provided by S. Shaham [48].
To test whether the tk41, tm548, s71 or tp5 mutation is recessive or dominant, homozygous mutant hermaphrodites were crossed to wild-type males carrying the integrated sur-5::gfp transgene. The cross-progeny hermaphrodites carrying the sur-5::gfp transgene were scored for the dumpy (Dpy) and DTC migration phenotypes using Nomarski microscopy. In these crosses, all heterozygous cross-progeny hermaphrodites were wild-type, showing that these mutations are recessive.
In complementation tests, the tk41, tm548 or tp5 homozygous hermaphrodites were crossed to s71 males carrying the sur-5::gfp transgene, the cross-progeny hermaphrodites with the transgene were scored for the Dpy and DTC migration phenotypes using Nomarski microscopy. In these crosses, all of the mutations tested failed to complement s71.
We positioned blmp-1(s71) on the basis of the Dpy phenotype within the region between snpF45H11 and snpY106g6h, which correspond to cosmids F45H11 and F37D6, respectively, using three-factor mapping with unc-40 and unc-75 and subsequent SNP mapping as previously described [47]. The following cross demonstrates that blmp-1 is located between unc-40 and snpY106g6h: from a blmp-1(s71)unc-75/++ (CB4856) hermaphrodite 37/37 Dpy non-Unc recombinant progeny segregated neither snpY106g6h nor snpF59C6, and 6/6 nonDpy Unc recombinant progeny segregated snpY106g6h and snpF59C6. From an unc-40 blmp-1 (s71)/++(CB4856) hermaphrodite 85/91 nonDpy Unc recombinant progeny segregated snpY106g6h and snpF59C6, 1/91 recombinant progeny segregated snpY106g6h and 5/91 recombinant progeny segregated neither snpY106g6h nor snpF59C6. The following cross demonstrates that blmp-1 is located right of snpF45H11: from an unc-40 blmp-1(s71)/CB4856 hermaphrodite 3/171 Dpy nonUnc recombinant progeny segregated snpF45H11.
We performed genetic screening for mutants defective in DTC migration under a dissecting microscope, as described in Nishiwaki, 1999, and isolated the mutation tk41. In an independent screen for mutants with a DTC migration defect using Nomarski microscopy, we isolated the mutation tp5.
The DTC migratory patterns of the wild-type and mutants were determined by observing the shape of the gonadal arms in the adult stage. Worms were mounted on a 4% agar pad containing 20 mM NaN3 and observed on a microscope with Nomarski optics. For the time course analysis of DTC migration, wild-type and blmp-1(s71) mutants at the indicated time points were collected and scored under the Nomarski microscope.
To obtain the 5′ end of blmp-1 cDNA, the first three exons were amplified by RT-PCR using a forward primer corresponding to the SL1 sequence and the reverse primer blmp-1_exon_3/r (see Table S5 for detailed information). The PCR product was fused with the blmp-1 cDNA fragment from the yk487b7 clone by fusion PCR [49], and the resulting product was cloned into the vector pGEM-T Easy (Promega) to generate pYW687, containing the full-length blmp-1 cDNA.
To construct Plag-2::blmp-1 (pYW798), blmp-1 cDNA was PCR-amplified from pYW687 using primers DPY-24-KpnI/f and DPY-24-KpnI/r and the product inserted into pPD49_26/Plag-2 via the KpnI site.
Two unc-5 transcriptional gfp constructs were generated. The Punc-5(4.6 kb) DNA fragment was PCR-amplified from C. elegans genomic DNA using primers Punc-5_4.6 kb/f and Punc-5_4.6 kb/r and inserted into the vector pGEM-T Easy. The Punc-5(4.6 kb) fragment of the resulting construct was then inserted into the vector pPD95.77 (A. Fire) via SphI/SalI sites to generate Punc-5(4.6 kb)::gfp. A similar approach was used to generate Punc-5(1 kb)::gfp. The Punc-5(1 kb) fragment was PCR-amplified from genomic DNA using primers Punc-5_1 kb/f and Punc-5_1 kb/r and cloned into the vector pGEM-T Easy, then the Punc-5(1 kb) fragment was excised from the resulting plasmid and inserted into the vector pPD95.75 (A. Fire) via the XmaI site. To generate Punc-5(1 kb)::mCherry, the gfp fragment of the plasmid Punc-5(1 kb)::gfp was replaced by mCherry cDNA.
The Punc-5(1 kb) fragment was expected to contain the regulatory region for proper unc-5 function in guiding DTC migration, as the Punc-5(1 kb)::unc-5::gfp plasmid is sufficient to rescue the DTC migration defect of the unc-5 mutant (our unpublished data). The Punc-5(1 kb)::unc-5::gfp plasmid was generated by fusion PCR by fusing two PCR-amplified products corresponding to Punc-5(1 kb)::unc-5(cDNA) and gfp::unc-54 3′ UTR from plasmid pU5HA (J. Culotti) or pPD95.75, respectively.
The Pblmp-1::dgfp fragment was generated by fusion PCR by fusing two PCR products corresponding to Pblmp-1 and the region containing dgfp and the unc-54 3′ UTR. Pblmp-1 was PCR-amplified from genomic DNA using primers D24-5end5kb and d24Pgfp/r, and the region containing dgfp and unc-54 3′UTR was PCR-amplified from pPD95.75PEST (pYW807), which contains dgfp, using primers GFP/f and D. The two PCR products were then mixed and fused by fusion PCR using primers d24-5kbf-nest and D′. The Pblmp-1::blmp-1::gfp fragment was generated by fusion PCR by fusing two PCR products corresponding to blmp-1 genomic DNA and the region containing gfp and unc-54 3′ UTR from the pPD95.75 plasmid [50]. blmp-1 genomic DNA was PCR-amplified from wild-type genomic DNA using primers D24-5end5kb and d24gfp/r, and the region containing gfp and the unc-54 3′ UTR was PCR-amplified from the pPD95.75 plasmid using primers GFP/f and D, then the two PCR products were fused by fusion PCR using primers d24-5kbf-nest and D′.
The primers used in this work are listed in Table S6.
Transgenic worms were generated by microinjection of the indicated plasmid, as described previously [51]. For genetic rescue experiments, cosmids and plasmids were microinjected into the indicated strains and the indicated phenotype of the stably transmitting lines scored using DIC optics. Pblmp-1::blmp-1::gfp (20 ng/µl) was co-injected with the unc-119 rescuing plasmid (100 ng/µl) [52] into the unc-119(ed3) mutant to generate tpEx49 transgenic worms. For the unc-5 transcription assay, Punc-5(4.6 kb)::gfp, Punc-5(1 kb)::gfp, or Punc-5(1 kb)::mCherry (20 ng/µl) was co-injected with the marker Pmyo-2::gfp (2 ng/µl), which expresses gfp in the pharynx [26], [53], as described previously [51]. To determine how blmp-1 was down-regulated during larval development, the indicated plasmids (20 ng/µl) were co-injected with Pmyo-2::gfp (2 ng/µl) into wild-type worms to generate transgenic worms, and the resulting transgenes were crossed to the indicated heterochronic mutants. For the blmp-1 cell autonomy assay, Plag-2::blmp-1 or Pblmp-1::blmp-1 (20 ng/µl) was co-injected with Pmyo-2::gfp (2 ng/µl) into wild-type worms to generate worms carrying the respective transgenes, then the blmp-1(s71) mutation was introduced into these transgenic worms by crossing to generate the blmp-1(s71) mutant carrying either the Plag-2::blmp-1 or Pblmp-1::blmp-1 transgene.
The transgenes used in this work were listed in Table S5.
The blmp-1 cDNA fragment from the plasmid pYW687 was digested with EcoRI and cloned into the vectors pGEX5X-1 (GE Healthcare) and pRSET B (Invitrogen) at their EcoR I sites to generate the respective constructs pYW802 and pYW691. The GST-BLMP-1 and HIS-BLMP-1 fusion proteins were present in inclusion bodies and were purified using standard methods [54]. Polyclonal rabbit antibodies against GST-BLMP-1 were generated and affinity-purified using 6His-tagged BLMP-1 as described previously [55] and the purified antibodies recognized GST-BLMP-1, but not GST, on a Western blot, indicating their BLMP-1 specificity. The immunostaining method was adopted from a previously published protocol [56], with slight modification. The animals were fixed for 1 h at 4°C in 80 mM KCl, 20 mM NaCl, 10 mM EGTA, 5 mM spermidine, 0.03 mM PIPES, 25% methanol, and 2% formaldehyde, then incubated overnight at 4°C with a 1∶100 dilution of the purified anti-BLMP-1 antibodies in PBS, 0.2% BSA, 0.5% Triton X-100 (PBT), then for 2 h at room temperature with rhodamine red-X (RRX)-conjugated donkey-anti-rabbit IgG antibodies (Jackson ImmunoResearch Laboratories; 1∶100 in PBT). For immunostaining of the lin-29(RNAi);daf-12(rh61rh411) mutants, the worms were mounted on a gelatin-chromic potassium sulfate-subbed slide and immunostained with anti-BLMP-1 antibodies, as described previously [57]. Worms were freeze-cracked and fixed in 95% ethanol for 10 min, then in 2% paraformaldehyde for 10 min, then incubated sequentially overnight at room temperature with purified anti-BLMP-1 antibodies (1∶1000 in PBT), followed by RRX-conjugated donkey-anti-rabbit IgG antibodies (Jackson ImmunoResearch Laboratories; 1∶1000 in PBT). They were then mounted with 2 µl of DABCO anti-bleaching reagent (Fluka) and 1 µl of DAPI (0.5 µg/ml) and observed using a Zeiss Axioplan 2 microscope equipped with a digital camera (AxioCam; Carl Zeiss, Inc.).
To knock down blmp-1, skr-1, cul-1, rbx-1, dre-1, or lin-29, RNA interference (RNAi) was performed either by feeding worms with bacteria expressing the double stranded RNA for the indicated genes or by injecting the double stranded RNA for the indicated genes as described previously [58], [59]. To generate the lin-29 RNAi construct, the region corresponding to exons 3–9 of lin-29 was PCR-amplified from the yk1430g04 plasmid, which contains lin-29 cDNA (Y. Kohara, Japan), and cloned into the L4440 vector. The blmp-1, srk-1, cul-1, and rbx-1 RNAi constructs were obtained from the Ahringer RNAi library.
HEK293T cells were grown in Dulbecco's modified Eagle's medium (Thermo Inc) containing 10% fetal bovine serum, 100 U/ml of penicillin, and 100 µg/ml of streptomycin (all from Life technologies) and were transfected with the indicated plasmids using Lipofectamine 2000 (Invitrogen). MG132 (20 µM) was added to the medium 2 h before the cells were harvested to block protein degradation by the proteasome.
Cells were harvested using pre-chilled PBS and lysed using 1% NP-40 lysis buffer with protease inhibitor (Roche). HA-tagged or Myc-tagged fusion proteins were immunoprecipitated by incubation overnight at 4°C with anti-HA or anti-Myc agarose beads (Sigma) and bound protein identified by Western blotting using antibodies against HA (Covance), Myc (BD Biosciences), or CUL1 (Zymed). Antibodies against tubulin (Sigma) or GAPDH (Abnova) were used on the loading controls.
|
10.1371/journal.pntd.0000664 | Development and Evaluation of a Sensitive PCR-ELISA System for Detection of Schistosoma Infection in Feces | A PCR-enzyme-linked immunosorbent assay (PCR-ELISA) was developed to overcome the need for sensitive techniques for the efficient diagnosis of Schistosoma infection in endemic settings with low parasitic burden.
This system amplifies a 121-base pair tandem repeat DNA sequence, immobilizes the resultant 5′ biotinylated product on streptavidin-coated strip-well microplates and uses anti-fluorescein antibodies conjugated to horseradish peroxidase to detect the hybridized fluorescein-labeled oligonucleotide probe. The detection limit of the Schistosoma PCR-ELISA system was determined to be 1.3 fg of S. mansoni genomic DNA (less than the amount found in a single cell) and estimated to be 0.15 S. mansoni eggs per gram of feces (fractions of an egg). The system showed good precision and genus specificity since the DNA target was found in seven Schistosoma DNA samples: S. mansoni, S. haematobium, S. bovis, S. intercalatum, S. japonicum, S. magrebowiei and S. rhodaini. By evaluating 206 patients living in an endemic area in Brazil, the prevalence of S. mansoni infection was determined to be 18% by examining 12 Kato-Katz slides (41.7 mg/smear, 500 mg total) of a single fecal sample from each person, while the Schistosoma PCR-ELISA identified a 30% rate of infection using 500-mg of the same fecal sample. When considering the Kato-Katz method as the reference test, artificial sensitivity and specificity rates of the PCR-ELISA system were 97.4% and 85.1%, respectively. The potential for estimating parasitic load by DNA detection in feces was assessed by comparing absorbance values and eggs per gram of feces, with a Spearman correlation coefficient of 0.700 (P<0.0001).
This study reports the development and field evaluation of a sensitive Schistosoma PCR-ELISA, a system that may serve as an alternative for diagnosing Schistosoma infection.
| Schistosomiasis is a neglected disease caused by worms of the genus Schistosoma. The transmission cycle requires contamination of bodies of water by parasite eggs present in excreta, specific snails as intermediate hosts and human contact with water. Fortunately, relatively safe and easily administrable drugs are available and, as the outcome of repeated treatment, a reduction of severe clinical forms and a decrease in the number of infected persons has been reported in endemic areas. The routine method for diagnosis is the microscopic examination but it fails when there are few eggs in the feces, as usually occurs in treated but noncured persons or in areas with low levels of transmission. This study reports the development of the PCR-ELISA system for the detection of Schistosoma DNA in human feces as an alternative approach to diagnose light infections. The system permits the enzymatic amplification of a specific region of the DNA from minute amounts of parasite material. Using the proposed PCR-ELISA approach for the diagnosis of a population in an endemic area in Brazil, 30% were found to be infected, as compared with the 18% found by microscopic fecal examination. Although the technique requires a complex laboratory infrastructure and specific funding it may be used by control programs targeting the elimination of schistosomiasis.
| Schistosomiasis affects 200 million people and about 779 million people live in endemic areas in the Middle East, South America, Caribbean, Southeast Asia and particularly sub-Saharan Africa [1]. Population- and treatment-based control programs have been successful in reducing the intensity of infection and severe morbidities associated with schistosomiasis; however, transmission remains active in highly endemic areas, and recurring low-level reinfection is likely to be associated with subtle but persistent morbidities such as anemia, malnutrition and diminished performance status [2]–[4]. In the presence of these conditions, the assessment of infection becomes less reliable since the currently used diagnostic methods are not sufficiently sensitive to accurately determine the prevalence of schistosomiasis or parasite burden in order to eventually achieve elimination of the disease [5], [6].
Microscopic demonstration of the parasite's eggs in feces or urine remains the most wide-spread tool for schistosomiasis diagnosis. The Kato-Katz technique [7] is currently the most used method for fecal examination because it is quantitative, relatively inexpensive and simple. A significant increase in the sensitivity of the method is gained by microscopic examination of multiple samples [8], [9], but this is a limiting procedure for field work.
To overcome the current limitations with respect to diagnosis, the simultaneous use of different diagnostic methods, such as antibody detection followed by stool examination of seropositive individuals, has been applied to monitor the human population and to identify the small number of infected people once morbidity control is achieved [6]. However, because antibody detection methods often cannot distinguish between current and past infection and may also present a high level of crossreactivity, molecular tools should be considered despite their higher cost and the requirement for special laboratory equipment [10].
Hamburger et al. [11] described a 121-base pair tandem repeat DNA sequence present in 12% of Schistosoma mansoni genome. This sequence has been successfully used in PCR-based approaches for the detection of the parasite in snails [12], monitoring of cercariae in water bodies [13] and diagnosis of human infection using fecal or serum samples [14] and, more recently, plasma samples [15]. In a population study, the prevalence of S. mansoni infection was determined to be 31% when three fecal samples were examined using the Kato-Katz technique, but the prevalence rose to 38% when the PCR technique developed by Pontes et al. [14] was employed using only one fecal sample [16]. The same result was observed by another group in a recent study assessing the marginal error of Kato-Katz examinations for diagnosis and cure evaluation of S. mansoni infection in areas of low endemicity [17].
Conventional PCR requires several steps after DNA amplification, including electrophoresis or blotting and hybridization, which are limited in the number of samples that can be conveniently analyzed. The PCR-enzyme-linked immunosorbent assay (PCR-ELISA) consists of an alternative process for large-scale screening that allows for semi-quantitative analysis. This technique combines an immunological method to quantify the PCR product directly after immobilization of biotinylated DNA on a microplate [18], [19]. The advantage of PCR-ELISA as compared to PCR-electrophoresis is that it makes use of standard equipment widely used for the processing of ELISAs, and the reagents used are easy to obtain commercially. Therefore, PCR-ELISA allows for the use of PCR-based DNA diagnosis for routine purposes in laboratories in less developed countries with fewer resources.
The aim of this work was to design a Schistosoma PCR-ELISA system as an improvement over the PCR assay previously developed by our group to detect S. mansoni DNA in human fecal samples [14]. The performance of the new assay was evaluated with fecal specimens from a Brazilian endemic area for S. mansoni infection and compared with the parasitological Kato-Katz technique for detection and estimation of the intensity of infection.
For use throughout the development of the Schistosoma PCR-ELISA system and for the estimation of its lower detection limit, S. mansoni eggs were obtained from the livers of Swiss albino mice 60 days after infection with 150 cercariae and stored at −20°C in 1.7% saline until use [20]. The animals were handled according to local and federal regulations, and the research protocol was approved by the Fiocruz Committee on Animal Research (License L-0118/09).
The number of eggs in saline suspension was quantified using a Neubauer chamber, and a solution containing approximately 2,000 eggs was used for DNA extraction. A negative fecal sample (verified to be free of S. mansoni eggs by the Kato-Katz technique and negative for the Schistosoma DNA detection by the PCR-ELISA system) was spiked with the egg-saline solution. The egg count was assessed by the Kato-Katz method, and approximately 500 mg of screened feces were used for DNA extraction.
Samples of genomic DNA from the human parasites S. haematobium, S. intercalatum and S. japonicum, and also from S. bovis, S. magrebowiei and S. rhodaini were provided by the Laboratório de Parasitologia Celular e Molecular (Centro de Pesquisas René Rachou, Fiocruz, Brazil) and used to assess the genus specificity of the Schistosoma PCR-ELISA system.
Two hundred and six people from Pedra Preta, Minas Gerais, Brazil, an endemic area for S. mansoni infection, participated in this study. The group was composed of 69 children (female/male: 33/36; age range of 1–17 years) and 137 adults (female/male: 66/71; age range of 18–86 years). Thirty-six healthy members of the laboratory staff participated as negative controls throughout the development of the assay. Written informed consent was obtained from all adult participants and from parents or legal guardians of minors. This study was approved by the Ethics Committee of the Centro de Pesquisas René Rachou, Fiocruz, Brazil (No. 14/2008).
For PCR-ELISA, fecal samples were collected and stored at −70°C until DNA extraction. All samples were evaluated for the presence of S. mansoni eggs by the Kato-Katz method. Twelve glass slides (41.7 mg/smear) of a single fecal sample were examined, resulting in a total sample weight of about 500 mg. Egg counts were expressed in eggs per gram of feces (epg), using the arithmetic mean of eggs counts obtained from the 12 slides multiplied by 24 [7]. The intensity of infection was calculated as the geometric mean of the individual egg counts.
Individuals with positive fecal examination results were treated with a single oral dose of praziquantel (50 or 60 mg/kg for adults and children, respectively), according to the recommendation of the Brazilian Ministry of Health.
Total DNA from 500 mg of each fecal sample and the DNA from a saline solution containing approximately 2,000 S. mansoni eggs were extracted using the QIAamp DNA Stool Mini Kit (Qiagen GmbH, Hilden, Germany), according to the manufacturer's recommendations and following the protocols “DNA Isolation from Large Amounts of Stool” and “Isolation of DNA from Stool for Pathogen Detection”. The heating step was performed at 95°C for 20 min to guarantee egg rupture. The concentration and purity of the DNA were determined spectrophotometrically by readings of A260 and A280 (Eppendorf, Hamburg, Germany).
The Schistosoma PCR-ELISA system consisted of a biotin 5′-labeled forward primer (5′-GATCTGAATCCGACCAACCG-3′), an unlabeled reverse primer(5′-ATATTAACGCCCACGCTCTC-3′) and the fluorescein 5′-labeled probe (5′-TGGTTTCGGAGATACAACGA-3′). The primers used were previously described [14] and designed to amplify a 121-bp tandem repeat DNA sequence (GenBank accession number M61098) found in the genome of S. mansoni [11].
To control for variation in the efficiency of DNA extraction and PCR-amplification, all clinical samples were evaluated with the human beta actin PCR-ELISA system. The primers used were Aco1 (5′-ACCTCATGAAGATCCTCACC-3′) and Aco2 (5′-CCATCTCTTGCTCGAAGTCC-3′), which were previously described to target the fourth exon of the human beta actin gene (ACTB) [21]. In this assay, the sense primer (Aco1) was biotinylated at the 5′ end, and a fluorescein 5′-labeled probe (5′-TCTCCTTAATGCACGCACG-3′) was designed together with the Schistosoma probe described above using the program Primer3-web 0.4.0 [22] and submitted to homology searches on the National Center for Biotechnology Information website with nucleotide BLAST program using database Nucleotide collection (nr/nt) and Megablast option. Amplification primers, biotinylated primers and probes were purchased from Integrated DNA Technologies, Inc. (Coralville, Iowa, USA).
For amplification, fecal DNA samples were diluted fivefold, and 2 µl were used as the template. The same volume was used to amplify S. mansoni egg-derived DNA, artificial S. mansoni egg-spiked fecal DNA and DNA from other Schistosoma species. PCR was carried out in a final volume of 20 µl containing 2 µl of GeneAmp 10X PCR Gold Buffer (150 mM Tris-HCl, pH 8.0, 500 mM KCl), 2.0 U of Amplitaq Gold (Applied Biosystems, Foster City, CA, USA), 0.1 µg/µl of BSA (Sigma, St. Louis, MO, USA), 0.5 µM of each primer, 1.5 mM MgCl2 and 200 µM of each deoxynucleoside triphosphate (Promega, Madison, WI, USA). The cycling programs, preceded by 12 min at 95°C to activate the HotStart Taq polymerase, consisted of 15 cycles of 95°C for 1 min, 63°C for 1 min and 72°C for 30 s; 12 cycles of 80°C for 1 min, 63°C for 1 min and 72°C for 30 s and 7 cycles of 80°C for 1 min, 65°C for 1 min and 72°C for 30 s, followed by a final elongation step at 72°C for 7 min. Positive controls based on DNA from S. mansoni eggs were included in all tests. Negative controls containing all of the elements of the reaction mixture except DNA were also included in each PCR assay as surveillance for contamination.
A 120-bp segment of the ACTB gene was amplified in a separate tube containing fecal DNA, following the amplification protocol described above, except that the MgCl2 concentration used was 2 mM. The cycling program, preceded by 12 min at 95°C, consisted of 35 cycles of 95°C for 20 s, 60°C for 30 s and 72°C for 1 min, followed by a final elongation step at 72°C for 7 min.
The chance of PCR contamination was minimized by physical separation of the starting materials and the amplified products in different rooms; the rooms contained laminar flux chambers with UV light, and sterile, disposable laboratory supplies were used.
Data were processed with SPSS statistical software package 13.0 for Windows (SPSS Inc., Chicago, IL, USA) and GraphPad Prism 3.0.3 software (San Diego, CA, USA). All quantitative variables were individually assessed with the one-sample Kolmogorov-Smirnov test for normality. Absorbance readings and arithmetical means of the number of eggs per gram of feces and template DNA concentration (both transformed into log scale) were analyzed by Pearson's parametric correlation coefficient or Spearman's nonparametric correlation coefficient. In order to determine the variability of the assays, intra-assay (repeatability) and inter-assay (reproducibility) precision levels were measured by comparing the means ± the S.D. and reported as coefficients of variation (CVs, S.D./mean X 100%). Sensitivity, specificity and 95% confidence intervals (CIs) were calculated using the OpenEpi Version 2.3 program [24]. Agreement beyond chance was assessed using the Kappa index and interpreted according to Landis and Koch [25]: 1.00-0.81 is excellent, 0.80-0.61 is good, 0.60-0.41 is moderate, 0.40-0.21 is weak and 0.20-0.0 is negligible. The X2 test was employed for the comparison of proportions. The level of significance was set at P<0.05.
The system's analytical sensitivity was evaluated using eight 10-fold serial dilutions ranging from 1.3 ng to 130 ag of genomic DNA extracted from a saline solution containing S. mansoni eggs. The limit of detection was 1.3 fg of genomic DNA, equal to that obtained by 6% polyacrylamide gel electrophoresis and silver staining, since the absorbance value for the sample with 130 ag of DNA was equivalent to the PCR negative control (0.093 and 0.064, respectively) (Figure 1A). The correlation between the numeric results (optical density readings) of the PCR-ELISA and the log number of the template DNA was significant, with a Pearson's correlation coefficient of 0.986 (P<0.0001) (Figure 1B). The analytical sensitivity of the assay was also evaluated with 10-fold dilutions of DNA extracted from a negative fecal sample spiked with an S. mansoni egg-saline solution determined by the Kato-Katz method to contain 1,534 epg. The Schistosoma PCR-ELISA system was consistently able to detect a sample estimated to contain 0.1534 epg.
The genus specificity of PCR-ELISA was assessed with purified DNA from S. mansoni, S. haematobium, S. bovis, S. intercalatum, S. japonicum, S. magrebowiei and S. rhodaini adult worms. Results (Figure 2) showed a ladder of PCR products due the amplification of the Schistosoma tandem-repeated unit, with the main DNA band of 110 bp present in all samples, as expected, and absorbance readings comparable to the positive control (S. mansoni DNA).
To analyze the repeatability of the Schistosoma PCR-ELISA system, three positive and three negative fecal samples from patients were assessed in four replicates in a single run. The intra-assay CVs for absorbance values for the negative samples were: 5.6%, 7.4% and 8.3%; the CVs for the positive samples were 1.9%, 3.6% and 4.2%. In addition, to measure the reproducibility, four replicates of the same samples were assessed on different days in four different assays. The inter-assay variations of absorbance values for the positive DNA samples were 3.8%, 7.7% and 15.9%; the variations for the negative DNA samples were 14.2%, 15.4% and 18.7%.
Fecal samples from 206 patients from an area in Brazil endemic for S. mansoni were analyzed by the Schistosoma PCR-ELISA system and also by the parasitological Kato-Katz technique. Comparison of the results obtained by the Schistosoma PCR-ELISA system and the Kato-Katz technique, based on examination of twelve slides (a total of 500 mg feces), is shown in Table 1. The geometric mean of the number of eggs per gram of feces estimated by the Kato-Katz technique for the positive samples was 18, which indicates a low intensity of infection [26]. The prevalence observed using the PCR-ELISA system (30%) was higher than that determined with the examination of twelve slides by the Kato-Katz technique (18%) (X2 = 8.81, P<0.003). The Kappa index of 0.663 indicates good agreement between the two methods. Analysis of discordant results showed that 25 samples were positive only by the Schistosoma PCR-ELISA system, and one sample was positive only by the Kato-Katz technique. This patient had very low egg output (8 epg).
Table 1 also shows a comparison of the results obtained by the Schistosoma PCR-ELISA system and the Kato-Katz technique, based on examination of two slides (a total of 83.4 mg feces), as is routinely done in epidemiological surveys. The Kappa index of 0.491 indicates a moderate agreement between the two methods. Analysis of discordant results showed that 36 samples were positive only by the Schistosoma PCR-ELISA system.
Diagnostic parameters were calculated by two different approaches: 1) taking two-slide Kato-Katz examination as the reference method for comparison or 2) considering twelve-slide Kato-Katz examination as the reference method. The sensitivity values of the Schistosoma PCR-ELISA system were high, regardless of the reference considered: 96.3% (95% CI, 81.7–99.3) for approach 1 and 97.4% (95% CI 86.5–99.5) for approach 2. Specificity values changed significantly depending on the reference used, being 79.9% (95% CI 73.4–85.1) for approach 1 and 85.1% (95% CI 78.7–89.7) for approach 2.
All 206 fecal samples analyzed by the ACTB PCR-ELISA system showed a positive result, ensuring that negative results correspond to true negative samples for the Schistosoma PCR-ELISA system rather than to a problem with sample degradation or PCR inhibition. Also, the remaining sixteen fecal samples of non-infected persons were evaluated to be positive for the ACTB PCR-ELISA system and negative for the Schistosoma PCR-ELISA system.
The ability of the Schistosoma PCR-ELISA system to estimate the parasitic load was assessed, and a Spearman's correlation coefficient of 0.616 (P<0.0001) was found when comparing values of absorbance readings and epg (transformed into log[epg+1]) to those determined by the Kato-Katz technique with 12 examined slides (Figure 3A). When the correspondence between the methods was evaluated, considering only positive and negative samples for both, the correlation coefficient observed was 0.700 (P<0.0001) (Figure 3B).
Results obtained with the Schistosoma PCR-ELISA system based on two levels of intensity of infection (1–100 epg and >100 epg), according to the Kato-Katz stool examination method, are shown in Table 2. The high sensitivity of the Schistosoma PCR-ELISA system was evidenced by the high number of positive samples in both groups. The assay also revealed the potential to be considered semi-quantitative, as the mean absorbance readings corresponded to the intensity of infection.
The control of schistosomiasis-related morbidity has become feasible due to the development of single-dose oral drugs such as oxamniquine and praziquantel, which are given to heavily infected patients (high worm burden) that were easily detected by field-applicable parasitological methods. With the transition to lower morbidity, there is need for more sensitive diagnostic methods. In an attempt to surpass this diagnostic limitation, a PCR-ELISA system was developed and evaluated as a new molecular assay for the diagnosis of Schistosoma infection. This system showed sound features, such as a low analytical limit of detection, genus specificity (and absence of cross-reactivity with other related parasites), good precision, high sensitivity, good specificity and also the potential for semi-quantitative analysis of parasitic load.
Evaluation of the analytical sensitivity of the Schistosoma PCR-ELISA system showed that it could accurately detect 1.3 fg of S. mansoni genomic DNA, which is equivalent to less than the DNA found in a single cell of this multicellular parasite, since its genome contains around 580 fg [27]. The high sensitivity of the assay is attainable due to the high copy number of the target sequence, which comprises approximately 12% of the S. mansoni genome [11]. In another approach, the analytical sensitivity was evaluated as the capability to detect DNA samples according to the egg count. The result obtained (0.1534 epg) corresponds to fractions of an egg.
The Schistosoma PCR-ELISA system was developed using primers designed by Pontes et al. [14], who demonstrated specificity by the absence of amplification when DNA from four related helminths (Ascaris lumbricoides, Ancylostoma duodenale, Taenia solium and Trichiuris trichiura) was used as templates. Further assay also showed no crossreactivity with Fasciola hepatica (data not shown). Additionally, the potential use of the previously described primers and probe and the current PCR protocol to detect genomic DNA from other Schistosoma species was assessed by the amplification of DNA extracted from worms of seven Schistosoma species (S. mansoni, S. haematobium, S. bovis, S. intercalatum, S. japonicum, S. magrebowiei and S. rhodaini). The results obtained confirmed what Hamburger et al. [11] partially addressed using a different set of primers that targets the internal part of the same DNA sequence: namely, this 121-bp tandem repeat is genus specific.
The precision of the Schistosoma PCR-ELISA system can be considered good, as the intra-assay variation was lower than 9%, and the inter-assay variation was around 3% to 19%. These percentages are similar to, or lower than, those calculated for other PCR-ELISA assay based on similar principles [28].
The choice of a reference test is crucial for the evaluation of any new testing method. The Kato-Katz technique is considered the test of choice to diagnose schistosomiasis in fecal samples and was chosen as a reference test for comparison with the Schistosoma PCR-ELISA system. Initially, for a traditional approach (approach 1 in this study), the comparison was made using two Kato-Katz slides, corresponding to 83.4 mg feces. The PCR-ELISA system was able to detect S. mansoni DNA in 36 fecal samples from patients with negative results by the Kato-Katz technique (two slides examined), demonstrating high sensitivity (96.3%), though lower specificity (79.9%). This specificity is likely to be artificial and incorrect, resulting from the less sensitive and inadequate reference method. In another approach (approach 2), twelve Kato-Katz slides were evaluated, corresponding to 500 mg of feces, the same quantity used to extract total DNA for analysis with the Schistosoma PCR-ELISA system. Again, the PCR-ELISA system detected more cases of infection with S. mansoni. Thus the sensitivity was high (97.4%), and the specificity was more satisfactory (85.1%). It is well documented that the Kato-Katz method lacks sensitivity if only a single fecal sample is examined, particularly in areas with high proportions of light-intensity infections. A small number of eggs unequally excreted over several days or patchily distributed may not be detected in the small amount of feces examined, negatively impacting on the method's sensitivity. In order to overcome these shortcomings, examination of an increased number of fecal samples as well as in the number of slides per specimen is required. There exists a general consensus that two Kato-Katz slides for each of three fecal samples yield enough sensitivity to obtain reliable data [29]-[32]. As this present study is part of a wider study in the endemic area of Pedra Petra, Minas Gerais, Brazil, three additional samples were collected on different days and analyzed by two Kato-Katz slides each (data not shown). From samples with discordant results between Kato-Katz analysis (twelve slides examined) and the Schistosoma PCR-ELISA, 12 out of 25 (48%) were positive in subsequent Kato-Katz examinations of additional samples. Therefore, a better explanation for these discordant results is that the Schistosoma PCR-ELISA system is more sensitive than the Kato-Katz technique considering the same amount of stool examined. That is, these cases correspond to S. mansoni-infected samples and not false-positive results. Although carryover contamination is often difficult to exclude as a cause of false-positive results in PCR-based diagnostic assays, the strict measures taken to avoid it throughout the entire handling and processing steps of the assay make the chances low.
Following both approaches, one sample was positive for the Kato-Katz technique and negative for the Schistosoma PCR-ELISA system. Since the parasitological method has an assumed specificity of 100%, this result can be most reasonably attributed to misdiagnosis by the PCR-ELISA. Inhibition of the amplification reaction by fecal compounds and/or DNA degradation during transportation of the sample from the field were considered the most likely technical causes. However, they were subsequently ruled out, as a positive result was obtained with the control ACTB PCR-ELISA system. Therefore, a better explanation seems to be the uneven distribution of a small quantity of eggs in the feces [33], since the sample had a very low parasite load (8 epg).
ELISA methods for the detection of PCR products provide an alternative to gel-based detection. Among other features, ELISAs prevent possible subjective interpretations of PCR results due to “nonspecific products” or “bands of unknown origin.” The main advantage of the procedure is the ability to process many samples in parallel (e.g., the 96-well microplate format), using instrumentation developed for processing ELISA for antibody detection. In addition, the procedure takes less than 2.5 hours to be performed (with the PCR Plate Detection Kit; Sigma).
The detection format of the commercial kit used here combines methodological features of PCR-ELISA methods described by Landgraf et al. [18] and Luneberg et al. [19]. Biotin-streptavidin binding was used for immobilization of amplification products on microtiter plates, taking advantage of the biotin moiety conjugated to one PCR primer. Subsequent hybridization was carried out with a fluorescein-labeled oligonucleotide probe, with the advantage of confirming the specificity of the test and allowing for detection by an anti-fluorescein-HRP conjugate.
Regardless of the fact that real-time PCR provides the best means to quantify PCR products, PCR-ELISA also allows for quantification, although it is usually considered a semi-quantitative technique, as it analyzes end-point amplification products. At the end-point, net synthesis is significantly reduced, inhibitory effects have accumulated and differences in initial starting template concentrations are masked [34]. These problems may be overcome by stopping the reaction in the exponential phase (as was done in the present study) or by requiring competitive amplification with very similar template DNA and primer pairs [18].
One classification used for epidemiological estimates of the intensity of infection of schistosomiasis on a community level considers the following groups: i) light, ≤100 eggs/gram of feces; ii) moderate, 101 to 400 eggs/gram of feces and iii) heavy, >400 eggs/gram of feces [26]. Even though the PCR-ELISA system developed in this study showed a satisfactory correlation coefficient between absorbance readings and egg counts per gram of feces, it seemed to be adequate to classify the samples according to the same parameters. However, as the performance of this new assay was evaluated in a low endemic setting, the number of samples with high egg counts was small, limiting statistical analysis. Thus, a statistically significant difference between mean absorbance values was observed only between the positive and negative groups (P<0.0001 for all). Nevertheless, an increase in the mean absorbance readings with increasing intensity of infection could be observed. Additional studies in areas of higher prevalence, resulting in a statistically significant number of patients with high egg counts per gram of feces, need to be carried out in order to validate the Schistosoma PCR-ELISA system's semi-quantitative potential.
The choice of a molecular method for diagnosing Schistosoma infection in endemic settings will depend on different factors, chiefly the available infrastructure and the cost-effectiveness. A control program may decide between performing multiple sample collections for the Kato-Katz examination or carrying out a single PCR-ELISA test, being the former less costly and the last more accurate. Considering only reagents, the cost of the PCR-ELISA system is roughly US$10 per fecal sample.
In conclusion, the Schistosoma PCR–ELISA system constitutes a precise tool for the diagnosis of Schistosoma infection, which may be particularly useful in low-prevalence settings and probably for post-treatment situations.
|
10.1371/journal.ppat.1007307 | A polymorphic residue that attenuates the antiviral potential of interferon lambda 4 in hominid lineages | As antimicrobial signalling molecules, type III or lambda interferons (IFNλs) are critical for defence against infection by diverse pathogens, including bacteria, fungi and viruses. Counter-intuitively, expression of one member of the family, IFNλ4, is associated with decreased clearance of hepatitis C virus (HCV) in the human population; by contrast, a natural frameshift mutation that abrogates IFNλ4 production improves HCV clearance. To further understand how genetic variation between and within species affects IFNλ4 function, we screened a panel of all known extant coding variants of human IFNλ4 for their antiviral potential and identify three that substantially affect activity: P70S, L79F and K154E. The most notable variant was K154E, which was found in African Congo rainforest ‘Pygmy’ hunter-gatherers. K154E greatly enhanced in vitro activity in a range of antiviral (HCV, Zika virus, influenza virus and encephalomyocarditis virus) and gene expression assays. Remarkably, E154 is the ancestral residue in mammalian IFNλ4s and is extremely well conserved, yet K154 has been fixed throughout evolution of the hominid genus Homo, including Neanderthals. Compared to chimpanzee IFNλ4, the human orthologue had reduced activity due to amino acid K154. Comparison of published gene expression data from humans and chimpanzees showed that this difference in activity between K154 and E154 in IFNλ4 correlates with differences in antiviral gene expression in vivo during HCV infection. Mechanistically, our data show that the human-specific K154 negatively affects IFNλ4 activity through a novel means by reducing its secretion and potency. We thus demonstrate that attenuated activity of IFNλ4 is conserved among humans and postulate that differences in IFNλ4 activity between species contribute to distinct host-specific responses to—and outcomes of—infection, such as HCV infection. The driver of reduced IFNλ4 antiviral activity in humans remains unknown but likely arose between 6 million and 360,000 years ago in Africa.
| Natural genetic variation and its influence on the outcome of viral infection is a topical area given the wealth of genetic data now available. However, understanding how clinical phenotype is affected by genetic variation at the molecular level is often lacking yet critical for any insight into immunity and disease. It is known that variants in the antiviral ‘interferon lambda 4’ (IFNL4) gene significantly influence outcome of hepatitis C virus (HCV) infection in humans. Counter-intuitively, those producing IFNL4 have greater risk of establishing chronic HCV infection, compared to individuals with an inactive variant, although the underlying mechanisms remain poorly understood. From a comprehensive screen of all natural human variants, we show that the most common form of IFNλ4 is less able to protect human cells from pathogenic virus infection than the equivalent protein from our closest living relative the chimpanzee. This is as a result of a single amino acid substitution that impedes its release from cells and reduces antiviral gene expression. Our observed differences in activity correlated with divergent host responses in HCV-infected livers from humans and chimpanzees. We suggest that human IFNL4 evolution places humans at a disadvantage when infected with pathogens such as HCV.
| Vertebrates have evolved the capacity to coordinate their antiviral defences through the action of proteins called interferons (IFNs) [1], which are small secreted signalling proteins produced by cells after sensing viral infection. IFNs bind to cell surface receptors, commencing autocrine and paracrine signalling via the ‘JAK-STAT’ pathway. Through this mechanism, IFNs induce expression of hundreds of ‘interferon-stimulated genes’ (ISGs) that establish a cell-intrinsic ‘antiviral state’ and regulate cellular immunity and inflammation [2,3]. Thus, IFNs are pleiotropic in activity and modulate aspects of protective immunity and pathogenesis [4].
Three groups of IFNs have been identified (types I–III), with the type III family (termed IFNλs) being the most recently discovered [5,6]. Emerging evidence highlights the critical and non-redundant role that IFNλs play in protecting against diverse pathogens, including viruses, such as norovirus [7], influenza virus [8] and flaviviruses [9]; bacteria [10]; and fungi [11]. While IFNλs induce nearly identical genes to type I IFNs, differences in signalling kinetics and cell-type specificity contribute to their specialisation [12,13]. Hence, as a consequence of selective expression of the IFNλ receptor 1 (IFNλR1) co-receptor on epithelial cells [13], type III IFNs play a significant role in defence of ‘barrier tissues’, such as the gut, respiratory tract and liver [reviewed in 14]; the second co-receptor for IFNλ is IL10-R2, which is expressed more broadly.
Although important for host defence, some IFNs are highly polymorphic [15]. In humans, a number of genetic variants in the type III IFN locus (containing IFNλs 1–4) have been identified and are associated with clinical phenotypes relating to viral infection [16–18]. Although many of these variants are in linkage disequilibrium, the major functional variant is thought to lie in the IFNL4 gene [19]. This causative variant is a single substitution/insertion mutation converting the ‘ΔG’ allele to a ‘TT’ allele (rs368234815), thereby yielding a frameshift which leads to loss of active human IFNλ4 (HsIFNλ4) [18]. Genome-wide association studies have convincingly demonstrated a seemingly counter-intuitive correlation between the IFNL4 ΔG allele and reduced clearance of hepatitis C virus (HCV) infection, i.e individuals who produce HsIFNλ4 clear HCV infection with reduced frequency in the presence or absence of antiviral IFN therapy [17,18]. Although IFNλ4 is highly conserved among mammals, the ‘pseudogenising’ TT allele of HsIFNλ4 has evolved under positive selection in some human populations suggesting that expression of the wild-type protein likely conferred a fitness cost during recent human evolution [20]. Expression of IFNλ4 is tightly controlled and reduced in human as well as Gorilla cells following viral infection compared to IFNλ3 [21]. The mechanism underlying the contribution of HsIFNλ4 to viral persistence in HCV infection is not well understood but is associated with enhanced ISG induction. Moreover, a common natural variant of HsIFNλ4 (P70S) [18], which has reduced signalling capacity, is also linked with improved HCV clearance [22]. Thus, there is a spectrum of HsIFNλ4 activity in humans as a consequence of natural variation that has a significant influence on chronic HCV infection, with wt HsIFNλ4 representing the protein with apparently the greatest antiviral activity.
Whether other human IFNλ4 variants exist in addition to P70S, which affect antiviral activity, has not been explored fully. In this study, we have examined human genetic data to identify other possible naturally occurring IFNλ4 variants and performed comparative analysis with mammalian orthologues in species closely related to humans. We provide evidence that the antiviral potential for the most common form of IFNλ4 in humans has attenuated activity due to a single amino acid substitution. In addition, we propose that acquisition of the attenuating substitution arose very early during human evolution but that some populations do encode a more active variant. Mechanistically, our data show that the reduced antiviral potential of human IFNλ4 results from a likely dual defect in secretion and potency.
Firstly, we undertook genetic and functional comparisons of natural human IFNλ4 coding variants present in the human population. We identified 15 non-synonymous HsIFNλ4 variants in the 1000 Genomes Project Database [23] (Fig 1A and S1 Data), including three previously described variants (C17Y, P60R and P70S; >1% global frequency, classified as ‘common’) [18]. The remaining 12 variants were classified as rare (<1% global frequency). The African population harboured the largest number of, as well as the most unique, variants. Interestingly, three rare variants (A8S, S56R and L79F) were shared exclusively between African and American populations, which may have arisen due to relatively recent movements of people perhaps through the transatlantic slave trade.
Variants were located in regions of functional significance in the HsIFNλ4 protein (Fig 1B and S1A–S1C Fig), such as the predicted signal peptide (amino acids 1–24), surrounding the single glycosylation site (N61) [both of which are required for secretion of active protein], and helix F that is predicted to interact with the IFNλR1 receptor (variants 151–158) [24]. Interestingly, the variants in helix F were clustered in the N-terminal portion of the predicted helix. Based on the above predictions, we hypothesised that some of these variants may have phenotypic effects on HsIFNλ4 function. Of note, no variants were found in helix D, which is predicted to contribute to interaction with the IL-10R2 receptor, nor on the IL-10R2-interacting face of the protein.
The functional impact of variation on HsIFNλ4 has only been assessed for the common P70S variant and so we sought to screen all other variants in activity assays. To determine whether variants affected HsIFNλ4 antiviral activity, they were introduced independently into an expression plasmid that produced HsIFNλ4 with a C-terminal ‘FLAG’ tag. Transient transfection of the expression plasmids into human ‘producer’ cells (HEK-293T cells) allowed harvesting of active HsIFNλ4 in the cell supernatant (referred to herein as conditioned media [CM]) thereby enabling analysis of the effects of variants on HsIFNλ4 production, glycosylation, secretion and potency; a similar approach has been successfully adopted previously to determine the relative activities of secreted HsIFNλ3 and HsIFNλ4 as well as a HsIFNλ4 variant that is not glycosylated [24]. We chose to screen the function of the panel of HsIFNλ4 variants on the interferon-competent hepatocyte cell line HepaRG cells [25].
Firstly, we investigated the antiviral activity of variants by titrating them against encephalomyocarditis virus (EMCV), a highly IFN-sensitive and cytopathic virus used to measure IFN-mediated protection [26] (Fig 1C and S2A Fig). We also measured their capacity to induce two major ISGs, MX1 and ISG15, by RT-qPCR (Fig 1D and 1E and S2B and S2C Fig) and validated the ISG15 mRNA data by determining production of unconjugated ‘mono’ ISG15 and high-molecular weight ISG15-conjugates (‘ISGylation’; S2D Fig). In addition, we constructed a series of negative controls (plasmids expressing EGFP and the frameshift TT variant of HsIFNλ4), a positive control (HsIFNλ3op) for comparative analysis to examine HsIFNλ4 activity, and three HsIFNλ4 variants, which do not occur naturally but were included as they could alter post-translational modification (N61A which ablates glycosylation) or potential receptor interactions (F159A and L162A located in helix F), respectively [27]. Negative controls (EGFP or the frameshift TT variant) gave very low induction of ISG15 and MX1 and no detectable antiviral activity in the EMCV assay whereas the positive control (HsIFNλ3op) was highly active in both assays (S2A–S2C Fig). The non-natural variants N61A and F159A almost abolished activity compared to wt HsIFNλ4 and HsIFNλ3op while L162A gave slightly less activity in the ISG induction assay but activity was reduced to a greater extent in the EMCV assay. In a previous report, ablating glycosylation at N61 substantially reduced activity of secreted HsIFNλ4 in an ISG induction assay [24]. Thus, our assay systems recapitulated findings from previous studies with similar assays and provided a range of activities to assess the impact of the natural HsIFNλ4 variants.
Our analyses on the natural HsIFNλ4 variants revealed that only three variants (P70S, L79F and K154E) consistently and substantially modulated antiviral activity and signalling compared to wt HsIFNλ4 (Fig 1C–1E). The impact of these variants was particularly pronounced in the EMCV assay that measures the dilution giving 50% activity over a large range of dilutions (Fig 1C). Our results confirmed previous observations on the lower activity of the P70S variant [22] and demonstrated that the rare L79F variant had a similar phenotype. By contrast, the K154E variant substantially enhanced antiviral activity and ISG induction.
These effects on activity for P70S, L79F and K154E did not arise from differences in the levels of HsIFNλ4 intracellular production or changes to glycosylation (S3A and S3B Fig). However, variants S56R and R60P (R60P is a common variant in Africa) did lead to marked reductions in the glycosylated form of HsIFNλ4 as demonstrated by the mean ratio of glycosylated:non-glycosylated protein (S3 Fig) but did not greatly alter their antiviral activity in contrast with our findings with the N61A non-natural variant, which abolished both glycosylation and antiviral activity of conditioned media (S2A–S2C Fig and S3A and S3B Fig). From this screen, we concluded that three non-synonymous variants in HsIFNλ4 (P70S, L79F and K154E), identified as either common or rare alleles in the human population, affect the antiviral activity of the protein.
Examining the global distribution of genetic variation can help understand its origins, evolution and functional consequences. P70S is a common variant that is found worldwide (in every population in the 1000 Genomes Database). By contrast, L79F and K154E are rare and, based on evidence in the 1000 Genomes Project Database, geographically restricted to West Africa/Americas, and central Africa, respectively (Fig 1A). We were able to obtain DNA from lymphoblastoid cell lines developed from the 2 individuals encoding the L79F variant; our analysis revealed that the West African subject carried the SNP for this variant but not the individual from the Americas see Materials and Methods).). Thus, we could verify the occurrence of the L79F variant in Africa but not in the Americas. From further interrogation of the 1000 Genomes Database [28], the HsIFNλ4 K154E variant was present in two individuals from different African rainforest ‘Pygmy’ hunter-gatherer populations (Baka and Bakola) in Cameroon (S4A Fig). The Bakola individual was homozygous for the ΔG allele, indicating that the K154E variant would be encoded on one of the functional ΔG HsIFNλ4 alleles. The Baka subject was heterozygous at rs368234815 (ΔG/TT) and thus only one allele (ΔG) would produce full-length HsIFNλ4, presumably K154E. The other allele (TT) is a pseudogene and would not lead to expression of functional HsIFNλ4. Each of the Baka and Bakola individuals also had additional non-synonymous HsIFNλ4 variants (V158I and R151P, Baka and Bakola individuals respectively); these variants were included in our functional screen of HsIFNλ4 variants but did not significantly alter activity (Fig 1C–1E and S2A–S2C Fig). From analysis of the Genome Aggregation Database (gnomAD) [29], the SNP variant resulting in the K154E substitution was found in a further 29 out of a population of 8,655 African individuals. Thus, this SNP is rare in the African population (0.003%) compared to the combined Baka/Bakola groups (20%). K154E was not found in other East or Southern African hunter-gatherer populations (such as Hadza and Sandawe) nor in the African San, who have the oldest genetic lineages among humans [30] (S4B Fig); it was also not identified in Neanderthal and Denisovan lineages (denoted as ‘archaic’ in S4B Fig). However, E154 is encoded in the IFNλ4 orthologue for the chimpanzee, Pan troglodytes (Pt), our closest mammalian species. Notably, the human TT allele encodes a potential K154 codon [18] suggesting that the E154K substitution arose in humans prior to IFNL4 pseudogenisation. Together with the fact that nearly all humans encode K154, these data suggest that the less active E154K substitution emerged early during human evolution after the divergence of our last common ancestor with chimpanzees.
Since a lysine residue encoded at positon 154 is unique to humans compared to other mammalian species (Fig 2A) [31], we compared wt HsIFNλ4 and its K154E variant to wt PtIFNλ4 and an equivalent ‘humanised’ PtIFNλ4 E154K mutant in both the EMCV and ISG induction assays as well as a CRISPR-Cas9 cell line in which the EGFP coding region had been introduced into the endogenous ISG15 gene upstream of and in-frame with the ISG15 open reading frame (ORF) (S5 Fig). This cell line offered advantages over other approaches since it facilitated measurement of ISG induction of an endogenous gene by assessing EGFP fluorescence across a range of dilutions of secreted IFNλs (S5B Fig).
Although intracellular expression levels of each IFNλ4 variant were similar, (Fig 2B), wt PtIFNλ4 was significantly more active than HsIFNλ4 in each assay and had approximately equivalent activity to the HsIFNλ4 K154E variant in signalling as well as antiviral assays (Fig 2C–2E). Converting PtIFNλ4 to encode the E154K variant significantly decreased activity to levels that were similar to those for wt HsIFNλ4 (encoding lysine at position 154). Extending the analysis to include rhesus macaque IFNλ4 (Macaca mulatta, MmIFNλ4) gave the same pattern whereby wt MmIFNλ4 with E154 had greater activity than its K154 variant. However, wt MmIFNλ4 was less active than either the human or chimpanzee IFNλ4 with E154 indicating that other genetic differences likely modified MmIFNλ4 activity in our assays. Consistent with the hypothesis that additional genetic differences affect susceptibility to E154K, introducing a lysine into the equivalent position of HsIFNλ3 had a much lesser effect on its activity compared to IFNλ4 (Fig 2C–2E). Overall, we observed a similar ~100-fold enhancement of activity for E154 over K154 for each of the IFNλ4 orthologues in anti-EMCV activity and EGFP IFN reporter induction. Thus, we conclude that wt HsIFNλ4 has attenuated activity principally because of a single amino acid change at position 154.
To broaden analysis of the impact of a lysine residue compared to a glutamic acid at position 154 in HsIFNλ4, antiviral assays were conducted with other human viruses that are less sensitive to exogenous IFN compared to EMCV, and on different cell lines. Specifically, we used HCV infection in Huh7 cells as well as infectious assays with influenza A virus (IAV) and Zika virus (ZIKV) in A549 cells against single high dilutions of each IFN. As controls, we also included the less active P70S and L79F HsIFNλ4 variants alongside HsIFNλ3op in these assays. Using the HCVcc infectious system in Huh7 cells, HsIFNλ4 K154E significantly decreased both viral RNA abundance compared to wt protein and exhibited a trend towards a lower number of infected viral antigen (NS5A)-positive cells (Fig 3A, upper and lower panels respectively). Furthermore, we performed assays examining HCV entry (HCV pseudoparticle system [HCVpp]), viral RNA translation and RNA replication (both assessed with the HCV sub-genomic replicon system). There was no significant difference in the efficiency of HCVpp infection between wt HsIFNλ4 and any of the three variants tested in the MLV-based pseudoparticle assay (Fig 3B, upper panel). However, we did observe a greater inhibition when the non-HCV E1E2-containing PPs were used, potentially reflecting the higher efficiency or different mode of entry of HCVpp entry compared to non-glycoprotein-containing retroviral PPs that could saturate an inhibitory response (Fig 3B, lower panel). wt HsIFNλ4 reduced HCV RNA replication compared to EGFP and introducing the K154E mutation into wt HsIFNλ4 gave a further significant reduction in replication. (Fig 3C, upper panel). However, primary translation of input viral RNA was not affected by HsIFNλ addition (Fig 3C, lower panel). To examine further the inhibitory effect of wt HsIFNλ4 and the K154E variant on viral RNA replication, Huh7 cells that constitutively expressed a HCV sub-genomic replicon [Tri-JFH1; 32] were treated with both forms of the protein over several passages (S6A Fig). Our results revealed a consistent decrease in HCV RNA levels over 8 passages spanning 25 days with the K154E variant exerting a greater inhibition on RNA replication compared to wt HsIFNλ4; consistent with this conclusion, fewer sub-genomic replicon-bearing cells survived treatment with HsIFNλ4 K154E than the wt protein (S6B and S6C Fig). Thus, HsIFNλ4 reduces HCV RNA replication and the K154E variant exerts greater potency against this stage in the virus life cycle. HsIFNλ4 K154E also reduced titers of IAV and ZIKV to a greater extent than wt protein in A549 cells (~10-fold; Fig 3D and 3E). Although this was only statistically significant in the context of IAV, a similar trend was evident with ZIKV for the K154E variant compared to wt HsIFNλ4. We found that the P70S and L79F variants consistently reduced the ability of wt HsIFNλ4 to protect against infection in most assays. Taken together, our data further confirmed the greater antiviral activity associated with converting a lysine residue at position 154 in HsIFNλ4 to a glutamic acid residue.
The enhanced antiviral activity of HsIFNλ4 E154 against multiple viruses in different cell lines suggested that this variant may differentially affect global transcription of antiviral ISGs. To test this hypothesis and examine the impact of HsIFNλ4 on global transcription, A549 cells were treated with wt and variant forms of HsIFNλ4 that had different antiviral activities and transcriptional changes were analysed by RNA-Seq at 24 hrs post stimulation (Fig 4). A549 cells were used because they recapitulate the functional differences in HsIFNλs as observed in other cell types and are widely used as a cell line model for epithelial antiviral immunity. The data revealed that K154E induced the broadest profile of significantly differentially-regulated genes (n = 273) compared with either the wt protein (n = 178) or the P70S variant (n = 115; Fig 4A–4C and S2 Data). The pattern of genes induced by the positive control HsIFNλ3op and HsIFNλ4 K154E were very similar (Fig 4B and 4C). From IPA pathway analysis, all HsIFNλs induced the same transcriptional programmes with differences in the overall significance of these pathways, most notably enhancement of the antigen presentation and protein ubiquitination pathways with the K154E variant (Fig 4D). Many of the differentially-expressed genes shared by HsIFNλ4 wt, K154E and P70S included known restriction factors with antiviral activity (e.g. IFI27, MX1, ISG15; Fig 4E) although the magnitude of induction was consistently greatest for HsIFNλ4 K154E (Fig 4F). There were also several ISGs that only achieved significant induction by K154E and HsIFNλ3op (e.g. IDO1, IRF1 and ISG20; Fig 4E and 4F). We predict that the apparent selectivity by IFNλ4 K154E results from the greater potency of this variant compared to wt. HsIFNλ4 and the P70S variant allowing genes to reach the significance threshold (Fig 4F). Enhanced production of antiviral genes in cells treated with HsIFNλ4 E154 would explain differences in antiviral activity against EMCV, HCV, IAV and ZIKV.
Direct in vivo validation of our transcriptomic findings alone on the enhanced activity of the K154E variant would require liver biopsy samples from either HCV-infected Pygmies or chimpanzees combined with equivalent samples from infected humans encoding wt HsIFNλ4. This was not possible since such tissue samples are not available from the Pygmy population infected with HCV and biopsies from acutely infected individuals are exceptionally rare. Moreover, chimpanzees are no longer used for experimental studies for ethical reasons. Therefore, we compared lists of reported differentially-expressed genes during acute HCV infection in humans and chimpanzees from the available literature. In the case of humans, there is only one report that analyses the transcriptional response in acute infection [33]. For chimpanzees, gene expression analysis is available from four independent studies [34–37] which include longitudinal data from serial biopsies. Therefore, all of the data was collated and we focused our comparisons on periods when human and chimpanzee biopsies were taken across the same time period after initial HCV infection (between 8 and 20 weeks post infection).
Comparative gene expression analysis revealed distinct host responses in humans and chimpanzees as well as overlapping differentially-regulated genes (Fig 5A and S3 Data). In chimpanzees, the transcriptional profile contained significantly expressed genes that were type I/III IFN-regulated ISGs known to restrict HCV infection (RSAD2, IFI27 and IFIT1) [2], as well as genes involved in antigen presentation and adaptive immunity (HLA-DMA and PSMA6). These genes were not significantly differentially expressed in humans, whose response was mainly directed towards up-regulation of pro-inflammatory genes (for example, CXCL10, CCL18 and CCL5) and metabolism genes (AKR1B10 and HKDC1) (Fig 5A and S3 Data). This was consistent with previous characterisation of the human acute response to HCV infection that failed to detect a major type I/III IFN signature but predominantly found a type II or IFN-gamma-mediated response [33]. From the available longitudinal data, the ‘chimpanzee-biased’ differentially-expressed genes were induced early in infection and remained significantly up-regulated during the acute phase following an early peak after infection (S7A and S7B Fig). Differences were also reflected in pathway analysis in terms of the most significant pathways and their overall levels of significance (S3 Data). For example, the ‘chemokine-mediated signalling pathway’ was upregulated in humans but not chimpanzees whereas the T cell receptor signalling pathway which was modulated in chimpanzees was not significantly altered in humans. Inspection of the raw data from humans indicated that many apparently ‘chimp-biased genes’ were expressed but did not reach significance in the original study. These genes were typically induced at a lower level in the human group when compared to averaged values for chimp studies across the similar time period (Fig 5B). Furthermore, there was a greater induction of antiviral ISGs in chimpanzees during chronic infection in comparison to humans although to a less pronounced effect (S7C Fig).
From examining the in vivo biopsy data, we identified a group of 29 chimpanzee-biased genes in liver biopsies that were seemingly up-regulated during acute infection to a greater extent compared to humans. Comparing this set of genes to those from the RNA-Seq transcriptomic data obtained in vitro (Fig 4) showed that the majority (17/29 genes) of the chimpanzee-biased genes were induced by HsIFNλ4 stimulation, with approximately half (8 genes) of those being significantly up-regulated to a great extent with K154E compared to wt, including MX1, IFITM1, IFIT1, IFIT3, TRIM22 and IFI44L (Fig 5C). Thus, there are similarities between our in vitro analysis and published in vivo studies that would correlate with differences in IFNλ4 activity between humans and chimpanzees.
Having established the greater antiviral potential for the E154 IFNλ4 variant and its apparent evolutionary relevance, we set out to determine the possible basis for its enhanced activity. No crystal structure for HsIFNλ4 is available but a homology model based on comparison with the IFNλ3 structure has been reported [24]. We expanded this predicted model based on both of the IFNλ1 and IFNλ3 crystal structures to explore the possible impact of K154E, P70S and L79F on IFNλ4 function (Fig 6A and S8 Fig; [38,39]). As has been previously described, the sequences in helix F, which binds to IFNλR1, are relatively well conserved [18,24]. The position equivalent to amino acid 154 in IFNλ4 is a glutamic acid in both IFNλ1 and IFNλ3 (amino acid position 176 in IFNλ1 and 171 in IFNλ3) and its side chain faces inward towards the opposing IL10R2-binding helices C and D (Fig 6A). The free carboxyl group of glutamic acid forms non-covalent intramolecular interactions with two non-linear segments on IFNλ1 and 3 (IFNλ1 residue K64, and in IFNλ3 K67 and T108). In IFNλ4, these E154-interacting positions are not conserved compared to IFNλ1/3 although homologous positions do exist with biochemically similar residues (IFNλ4 R60, and R98 that lies just upstream of the residue homologous to IFNλ3 T108).
To test whether the biochemical properties of glutamic acid at position 154 contribute to IFNλ4 activity, a panel of variants was constructed with biochemically distinct amino acids (R154, L154, A154, D154 and Q154). Firstly, intracellular expression of each variant at position 154 was approximately equivalent (Fig 6B). In signalling assays, the order of activity was E>Q/D>A>L>K>R (S9A Fig). We found a similar pattern in the EMCV antiviral assays except that L154 had the least activity (Fig 6C). We interpret these findings to conclude that E154 is biochemically the most favoured residue at this position with regards to antiviral potential, and that substitution of E154 to lysine results in the lowest potency for IFNλ4 activity. Interestingly, both Q154 and D154 had ‘intermediate’ activity compared to E154 and K154, suggesting that side chain length and negative charge are important to maximise the activity of IFNλ4.
In a final series of experiments aimed at giving further insight into the mechanism of action of IFNλ4 K154E, we compared the relative activities and abundance of different IFNλ4 variants in cell lysates (i.e. intracellular protein) and supernatants (i.e. extracellular protein). As wt HsIFNλ4 is poorly secreted into the supernatant from transfected cells in the absence of enrichment [24,40], IFNλ4 in CM was immunoprecipitated using an anti-FLAG antibody. In antiviral assays, the activity of human, chimpanzee and macaque E154 variants from cell supernatants, IP fractions and lysates was greater than the corresponding K154 variants in agreement with our earlier results (Fig 6D and Fig 2C–2E). Moreover, the D154 and R154 variants yielded patterns for cell lysates, cell supernatants and immunoprecipitated IFNλ4 protein such that D154 had intermediate activity between E154 and K154 while R154 had approximately equivalent activity to K154 (S9B Fig). Thus, each variant displayed a similar pattern of activity irrespective of the source of IFNλ4. From Western blot analysis, the E154 and K154 variants for each individual species were detected at similar levels in cell lysates (S9C Fig). The HsIFNλ4 D154 and R154 variants were expressed to slightly higher and lower levels respectively compared to E154 and K154 from humans. Paradoxically, we did not find the same pattern in IFNλ4 abundance for immunoprecipitated protein derived from cell supernatants. Thus, we were able to detect greater amounts of the E154 variants for human, chimpanzee and macaque IFNλ4 compared to their K154 variants (Fig 6E and S9D Fig). It was not possible to reliably detect macaque K154 or human R154 variants. HsIFNλ4 D154 had levels intermediate between the E154 and K154 variants. By Western blot and subsequent densitometry analysis, the relative abundance of IFNλ4 E154 and K154 variants in cell lysates for any species differed by 1.3 fold yet the approximate fold increase in antiviral activities were significantly greater and on average 16-fold. For the secreted IFNλ4 variants, we found that not only was there a higher abundance of E154 to K154 protein (9-fold), but activity was 41-fold higher for E154 than K154 variants, which results in a significant 3 to 4-fold rise in antiviral activity not explained by protein abundance (S9E Fig). With the exception of macaque K154, the FLAG antibody detected a putative breakdown product of about 11kDa in each of the samples, which we presume arose from cleavage by an unknown intracellular protease as it was also detected in cell lysates. The amount of this lower molecular weight product followed the same pattern as the full-length protein in that there was more with E154 than K154 thus cleavage does not explain differences in antiviral activity.
Towards the end of the study, the split NanoLuc reporter system became available [41], which comprised an 11 amino acid HiBiT tag that could replace the FLAG tag at the C-terminal end of the HsIFNλ4 variants and reconstitute luciferase activity (Fig 6F). The advantage of this approach was that the relative secretion of each variant could be determined by comparing enzyme activity from cell lysates and culture media in a highly-quantitative manner. Moreover, the antiviral activity of HiBiT-tagged HsIFNλ4 variants could be compared to their respective FLAG-tagged versions. To demonstrate the capacity of the system to quantify secretion, we generated HiBiT-tagged versions of wt HsIFNλ3 and HsIFNλ4 with and without N-terminal signal sequences, which would abrogate secretion. Removing the signal sequences from either HsIFNλ protein reduced secretion by fivefold (Fig 6F and 6G). The non-natural N61A variant that is not glycosylated was secreted ~2-fold less efficiently than wt HsIFNλ4. In agreement with our data using FLAG-tagged HsIFNλ4, the K154E variant of HsIFNλ4 was secreted about 3 times more efficiently than the wt protein (Fig 6F and 6G). In a subsequent screen of all HsIFNλ4 natural variants with the HiBiT tag system, we also validated the data shown in Fig 1C and S2 Fig (S10A–S10C Fig). Moreover, HiBiT-tagged HsIFNλ4 K154E was >10-fold more active than the wt form (S11C Fig). Thus, both FLAG- and HiBiT-tagged forms of K154E gave higher antiviral activity and were secreted more efficiently than wt HsIFNλ4.
Lastly, to examine secretion in more detail, cells that had been transfected with HiBiT-tagged HsIFNλ3 and HsIFNλ4 variants were treated with Brefeldin A (BFA) and Monensin to block ER and Golgi transport respectively (S11A–S11C Fig). Treatment with these inhibitors disrupted secretion of HsIFNλ3 by 12- (BFA) and 25-fold (Monensin). For the HsIFNλ4 K154E variant, each of the inhibitors blocked secretion by about 10-12-fold. In the case of wt HsIFNλ4, BFA inhibited secretion to a greater extent (~20-fold) than for either HsIFNλ3 or HsIFNλ4 K154E whereas Monensin decreased secretion to a similar extent for all three proteins. The explanation for the slightly greater inhibition of HsIFNλ4 by BFA requires further investigation. Overall, our data conclusively demonstrate that glutamic acid at position 154 promotes greater antiviral potential by enhancing both IFNλ4 secretion from cells and its intrinsic potency.
In this study we have identified further functional variants of human and non-human IFNλ4 that expand the spectrum of its activity. By comparing IFNλ4 from different species we demonstrate that the genus Homo evolved an IFNλ4 gene with attenuated activity (prior to the TT allele), and that the vast majority of extant humans carry an IFNλ4 variant with lower antiviral potential due to a mutation of a single highly-conserved amino acid residue (E154K). Human African hunter-gatherer Pygmies and chimpanzees encode a more active IFNλ4 (E154). We speculate that position 154 in IFNλ4 plays a key role in intramolecular interactions that may facilitate stabilisation of the protein thereby influencing its secretability and antiviral activity (S12 Fig)
Our analysis suggests that the Homo IFNλ4 orthologue acquired the E154K substitution, yielding a less active protein, after the genetic divergence of the hominid Homo and Pan ancestral lineages (estimated to be at most 6 million years ago in Africa [42]) but before human/Neanderthal divergence (~370,000 years ago, [43]). Subsequently, the IFNL4 gene acquired two further variants, the P70S and TT alleles that are now common in the human population [18]. Acquisition of each of these alleles either further reduced (P70S) or abolished (TT) IFNλ4 activity. Other rare variants have arisen in humans with little impact on HsIFNλ4 antiviral potential based on our in vitro assays, except for variants L79F and K154E, which lower and increase activity respectively. To us, the most intriguing of these variants is K154E, which was found in a high proportion of rainforest ‘Pygmy’ hunter-gatherers from west central Africa [28] but was rarely present in the African population. Since this variant was not present in the genetic data for San and Archaic Neanderthal and Denisovan human lineages, we speculate that these populations likely reacquired K154E following divergence of chimpanzees and humans. However, with the ever-increasing availability of genetic data from ancient and extant human populations, it may be possible to identify other populations carrying the E154 variant; in particular more details on the demographics of the 29 African individuals in the gnomAD database [29] may address whether there are additional specific groups who carry this more active variant and how it arose.
The factors responsible for divergent functional evolution of the IFNL4 gene within and between species are not known. It has been demonstrated that loss of IFNL4 has evolved under positive selection in some human populations thus we speculate that differences in exposure to certain pathogenic microbes has driven evolution of the E154 variant. On the one hand, type III IFN signalling enhances disease and impedes bacterial clearance in mouse models of bacterial pneumonia [44]. This suggests that IFNλ4 with a lower activity could be beneficial during non-viral infections although a link between IFNL4 genotype and bacterial infection in humans has not yet been made. Conversely, we postulate that the presence of more active IFNλ4 exemplified by E154 in Pygmies and chimpanzees may be linked to increased exposure to zoonotic viral infections in the Congo rainforest, such as pathogenic Filovirus infections [45].
For decades, experimental studies in chimpanzees have provided unique insight into HCV infection [46] but they do not present with identical clinical outcomes as human subjects. For example, chimpanzees have been reported to clear HCV infection more efficiently than humans [47], rarely develop hepatic diseases similar to humans [48], and are refractory to IFNα therapy [49]. Moreover, HCV evolves more slowly in infected chimpanzees, possibly due to a stronger immune pressure that reduces replication compared to humans [50]. In humans, IFNL4 genetic variants are associated with, and thought to regulate, each of these characteristics [18,51,52]. Although myriad factors could explain these phenotypic differences, including differences in antagonism of the immune response by HCV or changes in IFNL transcription for example, we propose that the greater antiviral activity of PtIFNλ4 compared to HsIFNλ4 contributes to the distinct responses to HCV infection in the two species.
Acute HCV infection in human cells in vitro and chimpanzees in vivo selectively stimulates type III over type I IFNs, which are effective at signalling in hepatocytes [53,54]. Notably, there is no apparent type I/III IFN gene expression signature in liver biopsies from humans with acute HCV infection [33]. Differences in IFN signalling during HCV infection have been postulated to explain the ability to control HCV infection in cell culture or following IFN-based therapy in humans [55,56]. Our comparative meta-analysis of the available literature revealed an apparent enhanced expression of ISGs with anti-HCV activity as well as genes involved in antigen presentation and T cell mediated immunity in chimpanzees compared to humans. Our analysis cannot be considered conclusive given the disparate nature of the retrospective studies used in our comparisons and therefore, there are a number of caveats (e.g. the expression levels of the IFNL genes; the relative expression of receptors in the two species). Nonetheless, we speculate that enhanced expression of ISGs in chimpanzee liver due to higher IFNλ4 activity could lead to greater control of viral infection by both inducing antiviral genes and by coordinating a more effective adaptive T cell response, which is critical for clearance and pathogenesis during HCV infection [57].
Based on the above speculation, we would predict that the response to HCV infection in chimpanzees may be similar in Pygmies with the K154E variant. A recent study in Pygmies from Cameroon, including the Baka and Bakola groups, showed low seroprevalence of 0.6% and no evidence of chronic HCV infection [58]. Interestingly, infection in non-Pygmy groups in Cameroon has a seroprevalence of ~17% [59]. One explanation for this difference could be higher IFNλ4 activity in populations with the K154E variant, which may enhance HCV clearance.
In our study of three primate orthologues, glutamic acid at position 154 in IFNλ4 provided greater antiviral activity and enhanced its ability to induce antiviral gene expression. A functional comparison of human and chimpanzee IFNλ4 orthologues has been explored previously but no significant differences in signalling activity were observed [31]. There are substantial differences in the methodologies used in our study and that of Paquin et al., which could explain our ability to detect divergent activity, for example the size of the tag attached to IFNλ4 and dose of protein used in assays.
Our observed functional differences between E154 and K154 did not correlate with levels of intracellular accumulation or glycosylation. However, we did find that the more active E154 variants for human, chimpanzee and macaque IFNλ4 were detected at higher levels in the immunoprecipitated fractions from cell supernatant (CM) compared to the K154 variants; the D154 and R154 variants also were detected at a lower level than E154. Interestingly, endogenous, wt HsIFNλ4 with K154 is not secreted to detectable levels compared to wt HsIFNλ3 [24,40,60]. Detection of low levels of secreted wt HsIFNλ4 by Western blot analysis requires exogenous expression of the protein and precipitation of material in the cell supernatant. Poor release of HsIFNλ4 was not due to differences in the signal peptide of HsIFNλ4 or HsIFNλ3 but secretion could be ablated if the single N-linked glycosylation site was mutated [24]. Our data indicate that position 154 regulates release of intracellular IFNλ4. Moreover, IFNλ4s with E154 are more potent than those with K154 when correcting for the difference in amounts of protein. This increase in potency for E154 was detected in both IP protein and lysates as well as HsIFNλ4 with different C-terminal tags. Moreover, we observed that the difference between E and K is greater in the cell released fraction than the cell lysate. The reason for this discrepancy could be explained by a number of factors that are outside of the scope of this study. Based on our modelling of the IFNλ4 structure and further mutational analysis, glutamic acid is apparently the optimal residue at position 154.
At the biochemical level, glutamic acid has the capacity to form electrostatic bonds with charged residues in the IFNλ4 protein and moreover it possesses a side chain which could contribute greater flexibility for such interactions. Notably, replacing glutamic acid with either aspartic acid or glutamine gave higher IFNλ4 activity than either non-polar or positively-charged residues. These potential E154-mediated interactions occur in the region of the protein devoid of cysteine-bonds likely making the interaction between helix F (IFNλR1-binding) and the loop connecting helices C and D (IL-10-R2-binding) particularly flexible. The putative greater structural stability facilitated by E154 may inherently increase the structural integrity of IFNλ4 making this variant more competent for secretion and more potent in signalling through the IFNλR1-IL10R2 surface receptor complex. Increased binding to IFN receptor complexes has been shown to enhance signalling by type I IFNs [61,62]. Further biophysical studies using highly-purified recombinant protein measuring affinity and avidity of HsIFNλ4 wt and K154E for each receptor molecule [as in 27,39] combined with studies on the mechanism of IFNλ4 release will help address these hypotheses.
To conclude, our study further supports a significant and non-redundant role for IFNλ4 in controlling the host response to viral infections yet one whose activity has been repeatedly attenuated during human evolution, commencing with E154K. Taken together, this provides the foundation for more detailed investigation into the mechanism of action of IFNλ4 and its overall contribution to host immunity in regulating pathogen infection.
All available human IFNL4 genetic variation along with associated frequency and ethnicity data for the human population were collected from the 1000 Genomes Database available at the time of study (June 2016) [23] (http://browser.1000genomes.org/index.html). The reference sequence for the human genome contains the frameshift ‘TT’ allele and so potential effects of variants on the HsIFNλ4 predicted amino acid sequence were identified manually following correction for the frameshift mutation (TT to ΔG). The effect of all single nucleotide polymorphisms (SNPs) on the open reading frame (ORF) was thus assessed and re-annotated as synonymous or non-synonymous changes resulting in the selection of coding variants reported here. Inspection of whole genome sequence data from African hunter-gatherers was carried out using previously published datasets [28]. We remapped the raw reads of six San individuals (four Juǀʼhoan and two ‡Khomani San) in the Simons Genomic Diversity Project [30] to the human reference genome (hg19) and conducted variant calling using the haplotype caller module in GATK (v3). Two Juǀʼhoan individuals were heterozygous at rs368234815 (TT/ΔG genotype, S2 Data). The genotypes of rs368234815 in Neanderthal and Denisovan were extracted from VCF files that were downloaded from http://cdna.eva.mpg.de/denisova/VCF/hg19_1000g/ and http://cdna.eva.mpg.de/neandertal/altai/AltaiNeandertal/VCF/. Neanderthal and Denisovan genetic data contained only ΔG alleles (S2 Data). Independent validation of the SNP variant that gives rise to the K154E substitution in HsIFNλ4 was obtained from gnomAD (http://grch37.ensembl.org/Homo_sapiens/Variation/Population?db=core;r=19:39737353-39738353;v=rs377155886;vdb=variation;vf=58909380 and https://www.ncbi.nlm.nih.gov/projects/SNP/snp_ref.cgi?rs=rs377155886). Amino acid sequences for mammalian IFNλ genes were obtained from NCBI following protein BLAST of the wt HsIFNλ4 polypeptide sequence. Multiple alignments of IFNλ amino acid sequences were performed by MUSCLE using MEGA7. Accession numbers of specific IFNλs used in the experimental section of this study were as follows: HsIFNλ1: Q8IU54; HsIFNλ3, Q8IZI9.2; and for IFNλ4: Homo sapiens AFQ38559.1; Pan troglodytes AFY99109.1; Macaca mullata XP_014979310.1; Pongo abelii (orangutan) XP_009230852.1, Bos taurus (cow) XP_005219183.1, Felis catus (cat) XP_011288250.1.
To validate the L79F variant, genomic DNA isolated from Epstein-Barr virus-immortalised B-cell lymphoblastoid cell lines from individuals (HG03095 and NA19658) identified through the 1000 Genomes Project and International HAPMAP project as probands with the L79F substitution (rs564293856 G>A SNP) were obtained from the Coriell Institute for Medical Research. PCR was used to generate an amplicon of ~300 base pairs corresponding to the region that includes rs564293856; amplicons were column-purified (Qiagen PCR-Clean Up kit, Qiagen). Internal forward and reverse primers within the amplicon were employed to determine the sequence of the region across SNP rs564293856 by Sanger sequencing (S13A and S13B Fig). The resulting chromatograms were inspected manually and the sequences at rs564293856 were called when identified using primers from both directions. HG03095 was confirmed as heterozygous for rs564293856 as predicted although we failed to detect the variant in NA19658; HG03095 and NA19658 were individuals who originated from Africa and America respectively.
The homology model of the HsIFNλ4 structure used in Fig 6 and S8 Fig was generated using the RaptorX online server (http://raptorx.uchicago.edu). The resultant HsIFNλ4 structural model was then structurally aligned with both HsIFNλ1 (PDB 3OG6) [38] and HsIFNλ3 (PDB 5T5W) [39]. Visualization, structural alignments, and figures were generated in Pymol (The PyMOL Molecular Graphics System, Version 1.8).
DNA sequences encoding the ORFs of HsIFNλ4, PtIFNλ4 and MmIFNλ4 (based on accession numbers above) were synthesized commercially with a carboxy-terminal DYKDDDDK/FLAG tag using GeneStrings or Gene Synthesis technology (GeneArt). As a positive control for functional assays, the HsIFNλ3 ORF was codon optimised (human) to ensure robust expression and antiviral activity, and is termed ‘HsIFNλ3op’; the exception to this were the HiBiT-tagged versions of HsIFNλ3 which were not codon optimised. All IFNλ4 coding region sequences were retained as the original nucleotide sequence without optimisation. Synthesized DNA was cloned into the pCI mammalian expression vectors (Promega) using standard molecular biology techniques. At each cloning step, the complete ORF was sequenced to ensure no spurious mutations had occurred during plasmid generation and manipulation. Single amino acid changes were incorporated using standard site-directed mutagenesis protocols (QuickChange site-directed mutagenesis kit [Agilent], or using overlapping oligonucleotides and Phusion PCR).
A549 (human lung adenocarcinoma), U2OS (human osteosarcoma), MDCK (Madin-Darby canine kidney) and HepaRG (non-differentiated human hepatic progenitor) cells were obtained from Chris Boutell; Huh7 (human hepatoma) cells were obtained from Charles Rice, Rockefeller University, USA; HEK293T (human embryonic kidney) cells were obtained from Sam Wilson; Vero (African Green Monkey kidney) cells were obtained from Arvind Patel. Cells were grown in DMEM growth media supplemented with 10% FBS and 1% penicillin-streptomycin except for HepaRG and genome-edited derivatives (generated during this study); these cells were cultured in William’s E medium supplemented with 10% of FBS, 1% penicillin-streptomycin, hydrocortisone hemisuccinate (50 μM) and human insulin (4 μg/mL). Huh7 cells harbouring the HCV JFH-1 sub-genomic replicon [Tri-JFH1; 32] were cultured in the presence of 500 μg/ml G418. All cells were grown at 37°C with 5% CO2. Cell lines were routinely tested for mycoplasma and no contamination was detected.
To inhibit secretion of HsIFNλs, cells were treated with BFA and Monensin at a final concentration of 5 μg/ml. BFA was purchased in DMSO-dissolved form (10 mg/ml) and Monensin sodium salt was dissolved in methanol to the same concentration as BFA; both inhibitors were purchased from Sigma-Aldrich (UK).
Plasmid DNA generated from bacterial cultures (GeneJET plasmid midiprep kit, ThermoScientific) was introduced into cells by lipid-based transfection using Lipofectamine 2000 or Lipofectamine 3000 (ThermoFisher) following manufacturer’s instructions. To produce IFN-containing conditioned media (CM) or measure protein production, HEK293T ‘producer’ cells were grown to near-confluency in 12 (~4 x 105 cells per well) or 6-well (~1.2 x 106 cells per well) plates and transfected with plasmids (2 μg) in OptiMEM (1–2 ml) overnight. At approximately 16 hours (hrs) post transfection (hpt), OptiMEM was removed and replaced with complete growth media (1–2 ml). CM containing the extracellular IFNλs was harvested at 48 hpt and stored at -20°C before use. Although antiviral activity was observed at 16 hpt, we chose 48 hpt to harvest CM to ensure robust production and secretion of each IFNλ. Intracellular IFNλs also were harvested from transfected cells at 48 hpt. CM was removed and replaced with fresh DMEM 10% FCS (2 ml) and then frozen at -70°C. To prepare cell lysates with IFNλ activity, plates were thawed and the cell monolayer was scraped into the media and clarified by centrifugation (5 minutes [mins] x 300 g) before use. CM or lysates were diluted in the respective growth medium for each cell line before functional testing as described in the text. Two-fold serial dilutions of CM were used in titration of anti-EMCV activity and ability to induce EGFP in an IFN-reporter cell line. Single CM dilutions of 1:4 (HepaRG and A549) or 1:3 (Huh7) were chosen based on initial experiments for gene expression and non-EMCV antiviral activity measurements to allow measurement of both high and low activity variants.
Immunoprecipitation of extracellular FLAG-tagged IFNλ4 present in the supernatant of transfected cells was carried out using an anti-FLAG M2 antibody-bound gel as described by the manufacturer’s guidelines (Sigma Aldrich). Immunoprecipitated IFNs were used in activity assays and for Western blot analysis. Briefly, resin with anti-FLAG antibody (40 μl) and supernatants (1 ml) were thawed on ice. Beads were washed repeatedly in ice cold buffer before being incubated with IFNs in CM for 2 hrs at 4°C while rocking. Bead-bound IFN was pelleted by centrifugation, washed and eluted with FLAG peptide (100 μl). Positive and negative controls were ‘BAP-FLAG’ and buffer only, respectively. Centrifugation conditions were 8,200 x g for 30 sec at 4°C. One quarter (25 μl) of total immunoprecipitated protein was loaded onto gels for Western blot analysis.
The FLAG tag at the C-terminal ends of IFNλs was replaced with the 11 amino acid HiBiT tag without a linker (amino acid sequence: N-VSGWRLFKKIS-C [41]) using PCR and subsequent cloning into the pC1 vector. To block secretion from cells, the first 26 amino acids of HsIFNλ3 and 23 amino acids of HsIFNλ4, corresponding to their respective N-terminal signal peptide sequences, were removed using a similar approach. The sequences of all plasmid constructs were verified. Luciferase activity was measured following transfection of plasmids using the Nano-Glo HiBiT Lytic Detection System and its Extracellular Detection System counterpart following manufacturers protocols (Promega, UK). Briefly, for measuring intracellular enzyme activity, supernatant was removed from transfected cells and lytic buffer was added directly to the cell monolayer before incubation for 10 minutes; all lysed material was transferred to a 1.5 ml Eppendorf tube and luciferase activity was measured in a luminometer. For determining extracellular activity, 100 μl of the CM was removed from the culture medium and mixed 1:1 with both the LgBiT and substrate for 10 minutes prior to measuring luciferase activity; the luciferase value for secreted HiBiT-tagged HsIFNλ was then calculated based on the total volume of culture medium. Ratios to determine the extent of secretion of each IFNλ were generated based on total intracellular and extracellular luciferase values.
Total cellular RNA was isolated by column-based guanidine thiocyanate extraction using RNeasy Plus Mini kit (genomic DNA removal ‘plus’ kit, Qiagen) according to the supplier’s protocol. cDNA was synthesised by reverse transcribing RNA (1 μg) using random primers and the AccuScript High Fidelity Reverse Transcriptase kit (Agilent Technologies); the recommended protocol was followed. Relative expression of mRNA was quantified by qPCR (7500 Real-Time PCR System, Applied Biosystems) of amplified cDNA. Probes for ISG15 (Hs01921425), Mx1 (Hs00895608) and the control GAPDH (402869) were used with TaqMan Fast Universal PCR Master Mix (Applied Biosystems). The results were normalised to GAPDH and presented in 2−ΔΔCt values relative to controls as described in the text. HCV genomic RNA was quantified by RT-qPCR as described previously [32].
IFN-competent cells (A549) were stimulated with IFN CM (1:4 dilution) in 6-well plates (~1.2 x 106 cells) for 24 hrs and global gene expression was assessed by RNA-Seq, using three biological replicates per condition. Sample RNA concentration was measured with a Qubit Fluorometer (Life Technologies) and RNA integrity (RIN) was determined using an Agilent 4200 TapeStation. All samples had a RIN value of 9 or above. 1.5 μg of total RNA from each sample was prepared for sequencing using an Illumina TruSeq Stranded mRNA HT kit according to the manufacturer's instructions. Briefly, polyadenylated RNA molecules were captured, followed by fragmentation. RNA fragments were reverse transcribed and converted to dsDNA, end-repaired, A-tailed, ligated to indexed adaptors and amplified by PCR. Libraries were pooled in equimolar concentrations and sequenced in an Illumina NextSeq 500 sequencer using a high output cartridge, generating approximately 25 million reads per sample, with a read length of 75 bp. 96.3% of the reads had a Q score of 30 or above. Data was de-multiplexed and fastq files were generated on a bio-linux server using bcl2fastq version v2.16. RNA‐Seq analysis was performed using the Tuxedo protocol [63]. Briefly, reads from 3 replicates per condition were aligned and junctions mapped against the human reference transcriptome hg38 using Tophat2 with the default settings except library type. Transcriptome assembly was performed using Cufflinks supplying annotations from the reference genome hg38 and the differential gene expression was calculated using Cuffdiff. Differential gene expression was considered significant when the observed fold change was ≥2.0 and FDR/q‐value was <0.05 between comparisons. Pathway analysis was carried out using Ingenuity Pathway Analysis [IPA] (Ingenuity Systems, Redwood City, CA, USA).
Cell growth media was removed and monolayers were rinsed once with approximately 0.5 ml PBS before lysis using RIPA buffer (ThermoFisher) containing protease inhibitor cocktail (1x Halt Protease inhibitor cocktail, ThermoFisher, or cOmplete, Mini, EDTA-free Protease Inhibitor Cocktail, Sigma Aldrich) for 10 mins at 4°C before being frozen at -20°C overnight. Lysates were collected into a 1.5 ml sample tube and clarified by centrifugation (12,000 x g for 15 mins). Samples (10 μl) from the soluble fraction were heated to 90°C for 10 mins with 100 mM dithiothreitol (DTT)-containing reducing lane marker at 90°C for 10 mins. Samples were run on home-made 12% SDS-PAGE gels alongside molecular weight markers (Pierce Lane marker, Thermofisher) before wet-transfer to nitrocellulose membrane. Membranes were blocked using a solution of 50% PBS and 50% FBS for 1 hr at room temperature and then incubated overnight at 4°C with primary antibodies in 50% PBS, 50% FBS and 0.1% TWEEN 20. Secondary antibodies were incubated in 50% PBS, 50% FBS and 0.1% TWEEN 20 for 1 hr at room temperature. Membranes were washed four times (5 mins each) following each antibody incubation with PBS containing 0.1% TWEEN 20. After the 4th wash following incubation with the secondary antibody, the membrane was washed once more in PBS (5 mins) and kept in ddH20 until imaging. Primary antibodies to the FLAG tag (1:1000) (rabbit, lot. 064M4757V, LiCor) and α-tubulin (1:10000) (mouse, lot. GR252006-1, LiCor) were used along with infra-red secondary antibodies (LI-COR) to anti-rabbit (donkey [1:10,000], 926–68073) and anti-mouse (donkey [1:10,000], C50422-05) to allow protein visualisation. Pre-stained, Pageruler Plus marker was used to determine molecular weights (ThermoFisher). Membranes were visualised using the LI-COR system on an Odyssey CLX and the relative expression level of proteins determined using LI-COR software (Image Studio).
An IFN reporter HepaRG cell line was generated to measure IFN activity by introducing the EGFP ORF fused to the ISG15 ORF separated by ribosome skipping sites by CRISPR-Cas9 genome editing. We chose to introduce EGFP in-frame to the N-terminus of the ISG15 ORF since it is a robustly-induced ISG. We also introduced the blasticidin resistance gene (BSD) for selection purposes. BSD, EGFP and ISG15 were separated using ribosome skipping 2A sequences (P2A and T2A). Transgene DNA was flanked by homology arms with reference to the predicted target site. Homology donor plasmids for CRISPR-Cas9 knock-in were generated through a series of overlapping PCR amplifications using Phusion DNA polymerase followed by sub-cloning into pJET plasmid. Plasmids for CRISPR-Cas9 genome editing (wt SpCas9) were generated using established protocols [64] in order to create plasmids that would direct genome editing at the 5’ terminus of the HsISG15 ORF (exon 2). pSpCas9(BB)-2A-Puro (PX459) V2.0 was a gift from Feng Zhang (Addgene plasmid # 62988). All sequences are available by request. HepaRG cells grown in 6 well dishes were co-transfected with CRISPR-Cas9 editing plasmids targeting the beginning of the ISG15 ORF in exon 2 (exon 1 contains only the ATG of the ORF), and homology donor plasmids described above (1 μg each) using Lipofectamine 2000 and the protocol described above. Transfected cells were selected using puromycin (Life Technologies) (1 μg/ml) and blasticidin (Invivogen) (10 μg/ml) until non-transfected cells were no longer viable. Selected cells were cloned by single cell dilution, expanded and tested for EGFP induction following IFN stimulation. Positioning of the introduced transgene was assessed by PCR amplification on isolated genomic DNA from individual clones (S5C Fig). Primers were designed to include one primer internal to the transgene and another external to the transgene and found in the target loci (sequences available on request). For use as an effective IFN reporter cell line, cells had to demonstrate robust induction of EGFP expression following stimulation with IFN and evidence of specific introduction of the transgene. This study uses clone ‘G8’ of HepaRG.EGFP-BSD-ISG15 cells. We have not tested whether there is a single transgene integration site or multiple ones nor confirmed that the EGFP produced following stimulation by IFNs results from the expression of the specifically-introduced transgene rather than off-target integration, which is theoretically possible. We do not predict this would affect the cells’ ability to act as a reporter cell line. For use in IFN reporter assays, stimulated cells (in 96 well plates stimulated for 24 hrs; ~5 x 104 cells per well) were washed, trypsinised and fixed in formalin (1% in PBS) at room temperature for 10 mins in the dark before being transferred to a round-bottomed plate and stored at 4°C in the dark until measurement of EGFP fluorescence. Non-stimulated cells were used as negative controls and the change in % EGFP-positive cells was assessed by flow cytometry using a Guava easyCyte HT (Merck Millipore). For fluorescence microscopy, EGFP induction was measured by indirect immunofluorescence of stimulated cells that were fixed and permeabilised on coverslips prior to antibody binding. An EGFP primary antibody (1:1000, rabbit ab290 Abcam) was used followed by a fluorescent anti-secondary antibody (1:500, Goat anti Rabbit Alexa-Fluor, Thermo Fisher, 568nm). Samples were counter-stained using DAPI and visualised with a confocal laser-scanning microscope (Zeiss LSM 710) under identical conditions.
Antiviral activity of IFNλs was determined using encephalomyocarditis virus (EMCV), influenza A virus (IAV; A/WSN/1933(H1N1)), Zika virus (ZIKV; Brazilian strain PE243) [65] and HCV (HCVcc chimeric clone Jc1) [66]. EMCV was grown on Vero cells followed by titration on U2OS cells by plaque assay. IAV stocks were generated on MDCK cells and titrated by plaque assay on MDCK cells with protease (TPCK-treated trypsin, Sigma Aldrich). ZIKV was titrated on Vero cells by plaque assay. For all plaque assays, cells were grown in 12 or 6-well plates to ~90% confluency before inoculation with serial 10-fold dilutions of virus stocks in serum-free Optimem. Inoculum remained on the cells for 2 hrs before removal and the monolayers were rinsed with PBS (1x) and semi-solid Avicell overlay (Sigma Aldrich) was added. For EMCV and IAV, 1.2% Avicell was used, diluted in 1x DMEM 10% FCS, 1% penicillin-streptomycin. For IAV titration, TPCK-treated trypsin was added (1 μg/ml). For ZIKV plaque assay, 2x MEM was used instead of 1x DMEM. HCVcc Jc1 was generated as described previously by electroporation of in vitro transcribed viral RNA into Huh7 cells and harvested at 72 hrs post electroporation. After filtration of the supernatant, HCVcc Jc1 stocks were titrated by TCID50 on Huh7 cells and stored at 4°C before use. HCVcc Jc1 TCID50 assays were performed using anti-NS5A antibody [67]. Infected cells at 72 hrs post infection were fixed and permeabilised with ice-cold methanol. Cells were rinsed in PBS, blocked with 3% FCS in PBS at room temperature and incubated overnight with mouse monoclonal anti-NS5A antibody (9E10) at 4°C. After removal of the antisera, cells were rinsed 3 times with PBS containing 0.1% TWEEN 20, and then incubated in the dark at room temperature for 1 hr with secondary antibody [Alexa-fluor 488nm anti-mouse (donkey)]. Cells were finally washed with PBS containing 0.1% TWEEN 20 and NS5A-expressing cells were visualized with a fluorescent microscope.
Cells stimulated with IFNλs were infected with viruses at the following multiplicities of infection (MOI): EMCV (MOI = 0.3; added directly to the media); IAV (MOI = 0.01); ZIKV (MOI = 0.01); HCVcc (MOI = 0.05). For IAV, ZIKV and HCVcc, the inoculum was incubated with cells for at 2 (IAV/ZIKV) or 3 hrs (HCVcc) in 0.5–1.0 ml serum-free Opti-MEM/DMEM at 37°C before removal. Cells were rinsed with PBS and then incubated with fresh growth media for the allotted time (24 hrs for EMCV, 48 hrs for IAV and 72 hrs for ZIKV and HCVcc). At the times stated for individual experiments, infected-cell supernatants were harvested and infectivity was titrated by plaque assay. IAV, ZIKV and HCVcc antiviral assays were all carried out in 12 well plates except for measurement of HCVcc infectivity by indirect immunofluorescence, which was measured in a 96 well plate. In the case of EMCV, a cytopathic effect (CPE) protection assay was employed to assess infectivity (26). Here, HepaRG cells were plated in a 96-well plates (~5 x 104 cells per well) and, when confluent, were incubated with two-fold serial dilutions of CM or lysate for 24 hrs before the addition of EMCV. At 24 hrs post infection with EMCV, media was removed, cell monolayers were rinsed in PBS and stained using crystal violet (1% in 20% ethanol in H20) for 10 mins. Crystal violet stain was then removed and stained plates were washed in water. The dilution of ~50% inhibition of EMCV-induced CPE was marked visually and the difference determined relative to wt HsIFNλ4.
Luciferase-expressing MLV pseudoparticles containing the E1 and E2 glycoproteins from JFH1 HCV strain were generated as described [68] along with their corresponding JFH1 E1-E2 deficient controls (particles generated only with MLV core) and used to challenge IFNλ-stimulated Huh7 cells. Huh7 cells grown in 96-well plates overnight (seeded at 4 x 103 cells per well) were stimulated with IFNλs for 24 hrs and transduced with HCVpp. 72 hrs later, cell lysates were harvested and luciferase activity was measured (Luciferase assay system, Promega) on a plate reading luminometer.
For HCV RNA replication assays, RNA was transcribed in vitro from a sub-genomic replicon (HCV-SGR) expressing GLuc (wild-type and non-replicating GND) [69]. In vitro transcribed RNA (200 ng) was transfected using PEI (1:1) into monolayers of Huh7 cells in 96-well plates overnight (seeded at 4 x 103 cells per well) that had been stimulated with IFNλs (24 hrs). At the specified time points, total supernatants (containing the secreted GLuc) from treated Huh7 cells were collected and replaced with fresh growth media. 20μl (~10% of total volume) was used to measure luciferase activity and mixed with GLuc substrate (1x) (50 μl) and luminescence (as relative light units, RLUs) was determined using a luminometer (Promega GloMax). Pierce Gaussia Luciferase Flash Assay Kit (ThermoFisher) was used and the manufacturer’s instructions were followed.
Previously published datasets of intrahepatic differentially-expressed genes from liver biopsies were used to compare human and chimpanzee transcriptomic responses to early HCV infection. At first, we used reported lists of differentially-expressed genes between humans and chimpanzees but further validated observations with raw data from human studies. Studies focusing on acute HCV infection (0 to 26 weeks) in humans and chimpanzees were acquired through manual literature search using Pubmed and gene lists were compiled. For chimpanzees, data was acquired from 4 studies [34–37] and one report was employed for human data [33]. The study by Dill et al. comprised single biopsy samples from each of six individuals, while in toto the chimpanzee studies combined data from ten animals with multiple, serial biopsies. All studies were carried out using similar Affymetrix microarray platforms except Nanda et al. who used IMAGE clone deposited arrays. Although similar microarrays measured different numbers of genes we focused on ‘core’ shared genes from chimpanzee studies. Humans were infected with HCV genotype (gt)1 (n = 2), gt3 (n = 3) and gt4 (n = 1) while chimpanzees were experimentally infected with HCV gt1a (n = 6), gt1b (n = 3) and gt2a (n = 1). The human dataset included individuals with IL28B rs12979860 genotypes T/T, C/T and C/C but no association between IL28B genotype and gene expression was noted (33). Gene names and fold-changes were manually converted to a single format (fold change rather than log2 fold change for example) to allow comparative analysis. Human biopsies were taken between two and five months after presumed infection following known needle-stick exposure, and serial chimpanzee biopsies were taken at different time points from between one week and one year after HCV infection. For comparative purposes, differentially-expressed genes in chimpanzees were included if they were detected during a time period overlapping with the human data. We identified a ‘core’ set of chimpanzee differentially-expressed genes (independently characterized in at least two studies) and compared them to the single human transcriptome study data at equivalent time points (between 8 and 20 weeks post-infection). This approach generated a set of core chimpanzee genes (genes found differentially-expressed in at least 2 studies, >2 fold change compared to controls and during the time frame compared to humans) for comparison with the human data. This is reflected in the ten-fold higher numbers of differentially-regulated genes found in the one human study compared to the ‘core’ (reduced) set assembled from four chimpanzee studies. To validate these findings, we used three studies of chronic infection for which data were available [56,70,71], two from humans employing RNA-Seq and one from chimpanzees using microarray measurement. Gene lists were extracted and a core human list was produced and compared to that from chimpanzees. For shared genes, the fold change values were compared for humans and chimpanzees. The ratio of chimpanzee induction to human induction was calculated.
These gene sets were compared to determine their degree of species-specificity or species-similarity using Venn diagram analysis (http://bioinfogp.cnb.csic.es/tools/venny/). The gene lists for humans and core genes for chimpanzees are shown in the S1 Data. For the chimpanzee-biased genes, mean expression values were determined at each time point from individual animals.
For non-transcriptomic analysis (transcriptomic analysis is outlined above), Graphpad Prism was used for statistical testing, which included Students’ T test and ANOVA and post-hoc tests (Dunnett’s test) where appropriate as described in figure legends. ****, p = <0.0001; ***, p = <0.001; **, p = <0.01; *, p = <0.05, are used throughout to denote statistical significance. Unless explicitly stated, in our analysis, the term biological replicate refers to measurements made from different wells during an experiment on the same day while independent experiment refers to measurements made from assays carried out on different days.
Ethical approval was not required for analysis of human or animal samples as no such material was used in this study. All human and animal genomic and transcriptomic datasets were obtained from either publicly-available data or published data. We note that all data collected in the original studies was subject to ethical approval, details of which are available in the original study manuscripts.
|
10.1371/journal.pbio.1002272 | Frontoparietal Structural Connectivity Mediates the Top-Down Control of Neuronal Synchronization Associated with Selective Attention | Neuronal synchronization reflected by oscillatory brain activity has been strongly implicated in the mechanisms supporting selective gating. We here aimed at identifying the anatomical pathways in humans supporting the top-down control of neuronal synchronization. We first collected diffusion imaging data using magnetic resonance imaging to identify the medial branch of the superior longitudinal fasciculus (SLF), a white-matter tract connecting frontal control areas to parietal regions. We then quantified the modulations in oscillatory activity using magnetoencephalography in the same subjects performing a spatial attention task. We found that subjects with a stronger SLF volume in the right compared to the left hemisphere (or vice versa) also were the subjects who had a better ability to modulate right compared to left hemisphere alpha and gamma band synchronization, with the latter also predicting biases in reaction time. Our findings implicate the medial branch of the SLF in mediating top-down control of neuronal synchronization in sensory regions that support selective attention.
| Directing attention to a part of visual space produces patterns of "brainwaves" or neuronal oscillations in the human visual cortex (the part of the brain at the back that processes incoming information from the eyes); oscillations at low frequencies are believed to help the brain block out irrelevant or distracting information, whereas high-frequency oscillations signal processing of relevant information. The instructions to increase or decrease these oscillations likely originate in the front part of the brain. In this study, we investigated the structural "highways"—bundles of white matter—that connect the front and back of the brain together. Not only did we show that these highways are asymmetric—i.e., some participants have a larger fiber bundle in the left hemisphere of their brains, and some in the right—we also showed that these asymmetries predicted whether subjects were better able to control the neuronal oscillations in their left or right hemispheres. This, in turn, predicted whether the participants were faster in detecting targets in the right or left half of the screen. Thus, we showed that these structural highways are important in helping the brain pay attention to parts of visual space.
| In order to operate in complex environments, it is necessary to selectively attend to relevant information while inhibiting distraction. It has been proposed that changes in neuronal synchronization implement the mechanism required for selective gating [1,2]. The increase in synchronization supports a gain increase [3] as well as information transfers to downstream regions by means of communication through coherence [4]. For instance, neurons in the monkey visual cortex activated by a given object show increased gamma-band (50–90 Hz) synchronization when attention is allocated to that object [1,5]. These results generalize to human electroencephalography (EEG) and magnetoencephalograhy (MEG) studies that have identified increased gamma band activity associated with selective attention [6–8]. Alpha oscillations on the other hand have been proposed to reflect active inhibition of distracting information. This is underscored by alpha oscillations (8–12 Hz) being relatively strong in regions anticipating distracting input [9–11]. Modulations in both the alpha and gamma band are predictive of performance in visual attention tasks [6,12–14]. Given that these neuronal oscillations are modulated by selective attention, they are under top-down control. The aim of this study is to identify the anatomical pathways supporting the top-down control of the oscillatory activity in sensory regions.
Cue-directed shifts of attention are believed to be subserved by the dorsal attentional network [15] consisting of the frontal eye field (FEF) and intraparietal sulcus (IPS), in contrast to the ventral attentional network that governs stimulus-driven attentional shifts [15]. Recent studies using transcranial magnetic stimulation (TMS) have implicated the dorsal network in providing top-down control of alpha [16–18] and gamma [18] oscillations. Communication within the dorsal network must be subserved by structural connections, and there is evidence that the development of frontoparietal white matter tracts co-occurs with recruitment of superior frontal and parietal cortex during attention and working memory tasks [19,20]. The superior longitudinal fasciculus (SLF), a network of white-matter fiber tracts consisting of medial, middle, and lateral branches [21], has recently been proposed to connect prefrontal control areas to posterior regions. In particular, the medial SLF branch (SLF1) projects to areas overlapping with the dorsal network—namely posterior superior frontal cortex in and near to the FEF and the IPS [21]. The lateral branch (SLF3) projects to nodes in the ventral network (inferior frontal gyrus and temporoparietal junction [21]), while the middle branch (SLF2) supposedly provides connections between the two networks. Individual differences in SLF2 volume have been shown to predict behavioral attentional biases [21,22]. Further, the number of SLF1 connections predicts the disruptive effects of FEF perturbation with TMS on visual task performance [23]. Given that individual differences in the SLF are behaviorally relevant, we hypothesize that the variance in these tracts also explains individual abilities to modulate alpha and gamma oscillations in sensory regions.
In the present study, we performed both MEG and high angular resolution diffusion imaging (HARDI) magnetic resonance (MR) measurements in the same subjects. Oscillatory brain activity was quantified from the MEG data while the subjects performed a cued spatial attention task requiring attention to the left or right visual hemifield. From the MR data, we used whole-brain spherical deconvolution tractography [24,25] to reconstruct the SLF branches. We hypothesized that the medial branch (SLF1)—connecting superior frontal to parietal cortex [21]—served as the structural pathway for controlling oscillatory brain activity in visual brain regions. Therefore, individual differences in SLF1 properties should predict individual ability to modulate visual cortical oscillations and thereby performance on a spatial attention task.
We acquired data from 26 subjects. These subjects performed a cued attention task in the MEG requiring shifts of attention to the left, right, or to both visual hemifields in order to identify the orientation of an upcoming target grating briefly presented 1,500 ms after the cue (Fig 1A). A second grating was always concurrently presented in the unattended hemifield. Analysis of the behavioral data using repeated-measures ANOVA confirmed that spatial cueing improved both accuracy and reaction time, respectively by 10% and 76 ms (Fig 1B; accuracy: F(1,25) = 42.077, p < 10−6; reaction time: F(1,25) = 110.114, p < 10−9). Direction of attention did not significantly alter these variables, and no interaction of direction with cueing was observed (p > 0.05 in all cases).
We first confirmed previous results demonstrating that both anticipatory alpha oscillations (defined as 8–12 Hz activity in a 1 s window prior to presentation of the target and distractor stimuli) and stimulus-induced gamma activity (defined as 50–90 Hz activity in a 400 ms window following target and distractor presentation) in occipital brain regions are modulated by direction of attention. Attentional modulation index (AMI) was calculated for each sensor j according to the formula AMIj = 100% * (PowerAttention left,j—PowerAttention right,j) / (PowerAttention left,j + PowerAttention right,j). The sensor-level analysis revealed a robust increase in gamma band activity in response to the target contralateral to the attended hemifield (Fig 2A and 2B). This finding is consistent with gamma band synchronization reflecting visual processing that is modulated by selective attention. The alpha band activity was strongly modulated in the cue-target interval and showed a relative decrease contralateral to the attended hemifield. The strong modulation during this delay is consistent with the notion that alpha band activity reflects the anticipatory allocation of attentional resources. No strong attentional modulation was observed in the intermediate beta-band or in other frequency bands.
To determine the underlying cortical sources of these modulations, we used a frequency domain spatial filtering technique (a beamformer approach [26]). To statistically quantify these modulations we used cluster-based permutation statistics [27], a method controlling for multiple comparison in space (see Materials and Methods). When comparing power values from “attention left” and “attention right” trials, we found robust modulations in the occipital cortex. When subjects were cued to the left, right occipital alpha power was lower than when they were cued to the right. The reverse pattern was observed in the left hemisphere (Fig 2C). These differences were greatest in the superior occipital cortex (MNI coordinates: left, −26 −92 38, right, 34 −82 44; associated clusters: left, p = 0.02; right, p = 0.0008, see S1 Fig). Conversely, when subjects were cued to the left, right occipital gamma power was higher than when they were cued to the right, and the reverse pattern was observed in the left hemisphere (Fig 2D). These differences were greatest in the middle occipital cortex (MNI coordinates: left hemisphere, −26 −94 16; right hemisphere, 34 −82 16; associated clusters: left, p = 0.002; right, p = 0.004, see S2 Fig). Consistent with the literature, both anticipatory alpha oscillations and stimulus-induced gamma band activity in occipital cortex are robustly modulated by spatial attention [6–12].
Next, we sought to relate individual differences in modulations of the gamma and alpha band activity to properties of the SLF. Spherical deconvolution tractography [25,28] was used to reconstruct the SLF branches from the diffusion data. Consistent with previous research [21,23], a network of three branches in each hemisphere was reconstructed (Fig 3A). For each of the three SLF branches, a hemispheric asymmetry index was computed (100% (volume_left–volume_right)/ (volume_left + volume_right); see Materials and Methods), quantifying whether each subject had greater tract volume in the left or right hemisphere. Nonoverlapping regions were identified as regions of interest (ROIs) in prefrontal cortex and then used for seeding the fiber tracking. This ensured that the fiber bundles were well separated. The medial SLF1 branches were defined as fibers passing through superior frontal gyrus, SLF2 as passing through middle frontal gyrus, and SLF3 as passing through precentral gyrus (see Materials and Methods). Replicating previous findings [21], the SLF3 was right-lateralized at the group level, whereas SLF1 and SLF2 did not show evidence of lateralization at the group level (see Fig 3B). Furthermore, a modulation asymmetry index was also calculated for each subject’s MEG data indicating whether—for both alpha and gamma oscillations—that subject displayed a stronger degree of power modulation with attention in the left or right hemisphere (ΔAMI = (- AMIleft,j)—AMIright,j; see Materials and Methods). We derived the alpha and gamma modulation values (ΔAMI) from the anatomical regions demonstrating strongest attentional modulation for each band, namely the superior occipital cortex for the alpha band and the middle occipital cortex for the gamma band (see Fig 2). Alpha and gamma asymmetry were not correlated with each other (r = -0.148, p = 0.47). We then correlated alpha and gamma asymmetry with the volumetric asymmetry of the three SLF branches.
Our main finding (Fig 4A, top panel) shows that gamma modulation asymmetry was strongly positively correlated with SLF1 hemispheric asymmetry (r = 0.596, r2 = 0.36, p = 0.0016, Spearman, significant at the p < 0.005 level after Bonferroni correction for three comparisons). This demonstrates that subjects who displayed relatively greater gamma modulation in the left hemisphere than in the right hemisphere also had relatively greater tract volume in the left than in the right hemisphere (and vice versa). No correlation was observed with SLF2 or SLF3 (in all cases p > 0.05 without Bonferroni correction).
Our second main finding (Fig 4B, top panel) shows that alpha modulation asymmetry was strongly negatively correlated with SLF1 hemispheric asymmetry (r = -0.503, r2 = 0.25, p = 0.0096, Spearman, significant at the p < 0.05 level after Bonferroni correction for three comparisons). This means that subjects who displayed relatively greater alpha modulation in the left hemisphere than in the right hemisphere also had relatively greater tract volume in the left than in the right hemisphere. The difference in the signs of the correlation is explained by alpha power decreasing and gamma power increasing contralateral to attention (see Materials and Methods for detailed explanation). No correlation was observed with SLF2 or SLF3 (in all cases, p > 0.1 without Bonferroni correction). This is evidence that individual differences in SLF1 hemispheric asymmetry predict individual differences in the top-down modulation of neuronal synchronization in both the alpha and gamma band.
To determine whether target-driven reorienting produced an asymmetry in the gamma band, we computed a reorienting index (RI) analogous to the AMI, according to the formula RIj = 100% * (PowerAttention both target left,j—PowerAttention both target right,j) / (PowerAttention both target left,j + PowerAttention both target right,j), for the gamma-band data in the post-stimulus window. This did not reveal a pattern of lateralized modulation, and no correlation was observed with any SLF branch (p > 0.1 without Bonferroni correction in all cases).
Having demonstrated a link between hemispheric asymmetry of SLF1 and both anticipatory alpha and stimulus-induced gamma band modulations in visual cortex, we further tested if these effects were predictive of subjects’ task performance. Accordingly, we quantified the degree to which subjects benefitted (in terms of reaction time and accuracy) from a left versus a right cue in comparison to the control condition with no spatial cue (see Materials and Methods). This hemifield specific asymmetry of the cueing benefit correlated with the hemispheric asymmetry of occipital gamma power modulation (ΔAMI; Fig 5A, r = -0.40, p < 0.05) but did not correlate with alpha power modulation (Fig 5B, r = 0.03, p = 0.89). The negative correlation value means that subjects with relatively stronger gamma modulation in the left occipital cortex than in the right occipital cortex benefitted more from a right cue than a left cue. This is fully commensurate with the notion that visual cortical gamma modulation in the hemisphere contralateral to target presentation boosts effective synaptic gain and thus enhances stimulus processing. No correlation was observed between accuracy benefit and hemispheric asymmetry of occipital gamma modulation (Fig 5C, r = 0.16, p = 0.45) or alpha modulation (Fig 5D, r = 0.31 p = 0.12).
Participants also performed a behavioral “landmark” task outside the MEG, designed to test spatial perceptual and motor response biases in the absence of directed attention [29–31]. Performance on this task was found not to correlate with hemispheric asymmetry of any SLF branch (see S1 Text and S3 Fig).
Although evidence exists for behaviorally relevant modulation of alpha and gamma oscillations in the occipital cortex [13,14], there is evidence that the top-down control signals that produce these modulations originate in the frontal cortex [16,18]. Given that gamma oscillations likely represent a general-purpose mechanism for effective communication [32], we further investigated whether SLF1 asymmetry predicted hemispheric asymmetry of gamma oscillations in prefrontal regions. To do this, we predefined two frontal ROIs: first, the FEF as defined by a meta-analysis of saccade studies [33] and, second, an adjacent region in the superior frontal cortex that has been identified as part of a frontoparietal network underpinning spatial attention and working memory [19,20]. To our surprise, hemispheric gamma modulation asymmetry (delta AMI) was found to correlate strongly with SLF asymmetry in the latter ROI (Fig 6A, r = -0.47, p = 0.017). Notably, the correlations in superior frontal cortex are negative, while they are positive in the occipital cortex. Fig 6B shows statistical maps of the correlation of SLF1 asymmetry with gamma asymmetry for every grid point. Grid points in the frontal cortex show negative correlations, and grid points in the occipital cortex show positive correlations. This means that those subjects with a greater left than right SLF1 volume actually displayed relatively greater gamma modulation in the right than left superior frontal cortex.
For the FEF as defined from the saccade literature, no correlation was observed with respect to hemispheric gamma modulation asymmetry (r = 0.35, p = 0.08). Neither ROI showed a correlation with hemispheric alpha modulation asymmetry (r = -0.33, p = 0.097, and r = 0.02, p = 0.92, respectively). No correlations were observed between SLF2 or SLF3 asymmetry and hemispheric alpha or gamma modulation asymmetry in the above ROIs (p > 0.15 in all cases).
Finally, we computed functional connectivity values between the superior frontal and occipital ROIs within the left and right hemispheres for each subject using power envelope correlations [34] and correlated the hemispheric asymmetry in functional connectivity with asymmetry of the SLF branches. No correlation was observed for the alpha or the gamma band data (all p > 0.1).
As reported in numerous studies, we have shown that stimulus-induced gamma band activity increases with spatial attention. Further, alpha oscillations decrease in anticipation of an upcoming stimulus. Importantly, we have now demonstrated a relationship between hemispheric asymmetry of the medial branch of the SLF (SLF1) and individual differences in the ability to exert top-down control over both anticipatory-alpha and stimulus-induced gamma oscillations. To our knowledge, this is the first evidence demonstrating that individual differences in frontoparietal white matter tracts predict the ability to modulate occipital cortical oscillations. This is strong evidence that the SLF1 is a structural pathway mediating top-down signals that control attentional modulations in visual cortex by modulating neuronal synchronization.
There is evidence suggesting that attention-modulated neuronal synchronization in the gamma band increases effective synaptic gain, and this synaptic gain increase enhances the impact of a neuronal population on connected downstream regions [1,35].Crucially, the ability to modulate gamma band activity in the present study was found to be predicted by the SLF1. Top-down signals from frontal cortex may thus serve to enhance gamma band synchronization and thus effective communication between visual cortex and downstream brain regions [2]. Emphasizing the relevance of these connections, hemispheric gamma band asymmetry was itself found to predict reaction times on the behavioral cueing task. This implies a causal chain by which a structural feature—hemispheric SLF1 asymmetry—can impact behavioral outcomes via its effect on neuronal dynamics. In contrast, no relationship was found between alpha oscillations and accuracy, in contrast to previous reports [36,37]. Gamma power has previously been shown to lock to the phase of ongoing alpha oscillations [38], suggesting an intimate relationship between bottom-up drive (indexed by the former) and pulsed inhibition (indexed by the latter). The present findings suggest that attentional modulation of alpha and gamma oscillations may not be related in such a simple fashion. The direct relationship between alpha and gamma oscillations should be a topic for future studies.
The relationship between hemispheric asymmetry of tract volumes and modulation of occipital cortical oscillations warrants further investigation. We propose that larger tract volume results in a higher fidelity of the top-down signal. A larger number of top-down connections from frontal control regions could result in a stronger propagation of the top-down signal by increased signal transmission. Tract volume is likely to depend on several factors including number of axons, proportion of myelinated axons, and axonal diameter [21]. Future work should therefore focus on identifying contributions of these factors to the effect on oscillatory modulation observed in the present study.
A previous HARDI study from Thiebaut de Schotten and colleagues found that SLF2 asymmetry predicted attentional task performance, whereas in the present study we found a relationship with SLF1. This is most likely explained by differences in the tasks. Although both studies used Posner paradigms [39], Thiebaut de Schotten and colleagues used 50% cue validity (Thiebaut de Schotten et al. [21], Supplemental Materials, page 12). Accordingly their subjects may have adopted a more stimulus-driven strategy engaging the ventral attentional network [15], consistent with the notion that the SLF2 supports communication between the dorsal and ventral networks [21]. The present study uses 100% valid cueing allowing preallocation of attention and likely engaging the dorsal attentional network. The present findings complement and extend these previous findings, demonstrating that in the context of high cue validity the dorsal network (and thus SLF1) is more strongly implicated.
The present study demonstrated that frontal top-down signals propagated via SLF1 impact visual cortical oscillations. Data from nonhuman primates implicate beta-band (18–34 Hz) oscillations in the FEF as controlling shifts of covert attention [40], and entrainment of 30 Hz activity in FEF using TMS has been shown to enhance visual perceptual sensitivity on a visual detection task in humans [41]. However, and consistent with our main hypotheses, initial sensor-level analysis of the MEG data (Fig 2) rather revealed robust attentional modulation during the cue-target interval in the alpha band and during the post-stimulus period in the gamma band, consistent with previous studies [1,6–11]. As well as the beta band, there is also some evidence that gamma-band phase interregional synchronization between frontal and posterior cortex is modulated by direction of attention [6], making this another candidate mechanism for top-down control. Future studies should attempt to further elucidate the precise form these attentional top-down control signals take.
The sources of the modulation of anticipatory alpha and stimulus-induced gamma oscillations were identified in the occipital cortex. The degree to which this attentional modulation was stronger in one hemisphere correlated strongly with hemispheric asymmetry of SLF1 volume. Crucially, however, a region in the superior frontal cortex also showed a similar effect in the gamma band, but with the opposite sign. This means that—whereas greater SLF1 volume in the left hemisphere (versus right) predicted stronger attentional gamma modulation in the left occipital cortex (versus right)—in the superior frontal cortex, greater SLF1 volume predicted weaker ipsilateral gamma modulation as compared to contralateral. Since modulation asymmetry is a measure of interhemispheric difference in modulation, this suggests a coupling of attentional gamma modulation between frontal cortex and contralateral visual cortex. Some evidence of such contralateral connections has been seen in previous TMS studies [18,42]. Furthermore, besides being the hypothesized frontal terminus of SLF1 [21], this frontal region is also adjacent to the human frontal eye field [33], a key node in the dorsal attentional network known to be involved in top-down allocation of attention [43–45]. Notably, one TMS study explicitly demonstrated a link between the disruptive effect of TMS to the right FEF on a visual perception task and properties of the SLF1 [23], suggesting that this white-matter tract indeed serves as the structural basis for communicating signals from FEF to other nodes in the dorsal attentional network.
Whilst we demonstrate a role for a cortico-cortical connection in top-down control of attentional oscillations, it is important to also consider cortico-subcortical connections. A recent nonhuman primate study demonstrated functional and structural connectivity between pulvinar and several visual areas, with the former serving to synchronize neocortical regions during a visuospatial attention task [46]. The cortico-cortical pathway we report on should be considered complementary to the subcortical pathway. The pulvinar may drive local synchrony between visual cortical regions preferentially during attention, whilst the frontal cortex provides top-down control signals that boost or attenuate the amplitude of attentionally relevant oscillations in response to task demands. Delineation of the respective contributions of both cortico-cortical and cortico-subcortical pathways should be the object of further study.
In conclusion, our data demonstrate for the first time (as far as we are aware) evidence for a cortico-cortical pathway providing top-down control of attentional modulations of behaviorally relevant neuronal oscillations in occipital cortex. This provides experimental support for the notion that modulation of visual cortical oscillations—and thus of effective synaptic gain—is the mechanism by which the dorsal attentional network asserts goal-directed attention.
Twenty-eight right-handed subjects (15 males, 13 females, mean age of 24 y and 5 mo) participated in the experiment. All subjects underwent standard screening procedures for MEG and MRI. All experiments were carried out in accordance with the Declaration of Helsinki and following ethical approval by the local ethics board (CMO region Arnhem-Nijmegen, CMO2001/095). One subject elected not to complete the diffusion scanning, meaning diffusion data were unavailable, and for one subject SLF branches could not be reconstructed. Therefore, the analyses were conducted on the remaining 26 datasets.
|
10.1371/journal.pntd.0001918 | Modeling Dynamic Introduction of Chikungunya Virus in the United States | Chikungunya is a mosquito-borne viral infection of humans that previously was confined to regions in central Africa. However, during this century, the virus has shown surprising potential for geographic expansion as it invaded other countries including more temperate regions. With no vaccine and no specific treatment, the main control strategy for Chikungunya remains preventive control of mosquito populations. In consideration for the risk of Chikungunya introduction to the US, we developed a model for disease introduction based on virus introduction by one individual. Our study combines a climate-based mosquito population dynamics stochastic model with an epidemiological model to identify temporal windows that have epidemic risk. We ran this model with temperature data from different locations to study the geographic sensitivity of epidemic potential. We found that in locations with marked seasonal variation in temperature there also was a season of epidemic risk matching the period of the year in which mosquito populations survive and grow. In these locations controlling mosquito population sizes might be an efficient strategy. But, in other locations where the temperature supports mosquito development all year the epidemic risk is high and (practically) constant. In these locations, mosquito population control alone might not be an efficient disease control strategy and other approaches should be implemented to complement it. Our results strongly suggest that, in the event of an introduction and establishment of Chikungunya in the US, endemic and epidemic regions would emerge initially, primarily defined by environmental factors controlling annual mosquito population cycles. These regions should be identified to plan different intervention measures. In addition, reducing vector: human ratios can lower the probability and magnitude of outbreaks for regions with strong seasonal temperature patterns. This is the first model to consider Chikungunya risk in the US and can be applied to other vector borne diseases.
| Chikungunya fever is a mosquito-borne viral infection showing a surprising potential for geographic expansion. Similar to other tropical infectious diseases having no vaccine and no specific treatment, the main control strategy for Chikungunya remains reduction of mosquito population size. We developed a model for disease introduction that combines a climate based mosquito population dynamics stochastic model with an epidemiological model in order to identify temporal windows during which disease introduction through one exposed individual might compromise the health status of the entire human population. We ran this model with temperature data from different locations showing the geographic sensitivity of this risk. The identification of temporal windows with epidemic risk at different spatial locations is key to guiding mosquito population control campaigns. Locations with marked seasonal variation also have a season with high epidemic risk matching the period in which mosquito populations survive and grow, therefore controlling mosquito population sizes might be an optimal strategy in those areas. However, locations with other temperature patterns may need additional control strategies to avoid epidemics. To our knowledge, this is the first model to explore Chikungunya introduction in the USA. Our modeling approach can be used for other vector borne diseases and can be expanded to compare the outcome with different control strategies.
| Chikungunya fever (CHIKF) is a mosquito-borne viral infection first isolated in Tanzania in 1953 [1], [2]. CHIKF is caused by Chikungunya virus (CHIKV), an alphavirus with different variants endemic to countries in Africa and Southeast Asia [3], [4]. Illness caused by CHIKV is usually diagnosed based on symptoms, and often confused with dengue given some overlapping symptomology [5]. One symptom specific for Chikungunya is a debilitating and prolonged joint pain, affecting the peripheral small joints [6], that appears in conjunction with other nonspecific symptoms including fever, severe joint pain, muscle pain, headache, nausea, fatigue and occasionally rash [7]–[11]. CHIKF-related mortality is rare, but can occur, often in patients with other health conditions [1], [9], [11], [12]. There is no specific treatment for the disease; consequently, treatment is focused on symptomatic care and mosquito vector control. No vaccines are currently available for prevention of CHIKV infection, although vaccine candidates currently are under investigation [13].
The onset of the symptoms occurs after an intrinsic incubation period in the human host of approximately 4 days post infection [1], [8]–[11], and viremia in infective individuals usually persists for a period of approximately 7 days [1], [3], [14]. During this period, mosquitoes may be infected with CHIKV when feeding on viremic hosts. After the acute stage of infection, severe joint pain may persist for long periods in affected individuals. Some people show mild to no overt signs of illness. Seroprevalence studies have demonstrated that 25% of infected individuals have mild symptoms or were asymptomatic [3].
Laboratory studies have demonstrated that CHIKV disseminates to the salivary glands in competent mosquitoes quickly, within 2 days (range 1–14 days) post-infection [15]. Once infectious, mosquito vectors are thought to remain infectious for their lifetime. Prior to 2000, Aedes aegypti was the most important vector of CHIKV [6] , with Ae. albopictus considered a secondary vector [16]. Within the last decade several epidemics of CHIKF were reported. In 2005–2006, a severe epidemic occurred in Réunion Island [3], followed shortly after by epidemics in India [3], Southeast Asia [17] and other Indian Ocean islands [1], [18]. Sampling during the Reunion Island epidemic provided evidence for the role of Ae. albopictus as the main vector [19]–[21]. Sequencing of the envelope protein of the Reunion Island CHIKV isolates (CHIKV 226OPY1) showed that the outbreak was caused by a new variant of the virus with a single adaptive mutation. This single amino acid change from Alanine to Valine in the E1 glycoprotein at position 226 [22] increased infection and dissemination in Ae. albopictus [23]. Another epidemic of CHIKV OPY1 genotype occurred in Italy in 2007 [24], [25]. Epidemiological studies strongly implicate introduction of the virus from India by a traveler [21], [26]. This unexpected outbreak is a striking example of disease introduction in an area recently colonized by Ae. albopictus [15], [21], [27]. Moreover, it highlights the fact that CHIKV outbreaks can originate from just one infective individual even in temperate areas with seasonal transmission of arboviruses [25], [19].
The Asian tiger mosquito, Ae. albopictus, is an invasive urban mosquito native to East Asia [28]. It is a diurnally active species and thought to have a broader host range than Ae. aegypti, although in some regions it can be highly anthropophagic when human hosts are readily available [29]. In the past couple of decades this species has invaded many countries through the transport of goods, especially used tires, and increasing international travel [30]–[32]. Native to tropical regions of Asia, Ae. albopictus has successfully adapted to cooler climates within the 10°C isotherm [33]. Thus, eggs from strains in temperate regions are moderately tolerant to cold and can even tolerate short durations of freezing temperatures [34], [35] . Female Ae. albopictus lay eggs in human-made and natural containers just above the waterline. Reported flight range of this species is typically less than 200 m [36].
Several laboratory studies on Ae. albopictus vector competence for the CHIKV LR 226OPY1 epidemic strain have now been conducted reporting a range of dissemination rates from 26–100% in various geographic strains of the vector [20], [37]–[40]. A recent laboratory study, using salivary gland infection as a proxy for transmission, demonstrated transmission rates from above 67% for Galveston, TX strain, Florida strain, and a New Jersey/New York metropolitan strain (Harrington, Sanchez-Vargas and Olson, unpublished data).
Concerns for the role of Ae. albopictus as an active disease vector have been raised since its introduction and in the USA [41], [42]. Since introduction, Ae. albopictus has become established in 26 states primarily in Southeast, gulf coast and mid-Atlantic regions. The species is currently expanding its range through New Jersey and into New York State [43], [44]. Given the establishment of Ae. albopictus in these regions, travel related introductions of several arboviruses suggest a potential increase in epidemic risk for the USA. High numbers of CHIKF cases are periodically reported in US travelers [45], [46]. However, local CHIKV outbreaks have not been detected in the US to date, presumably because of the asynchrony between the arrival of the exposed individuals and the abundance of the vectors [45].
In this study, we explicitly evaluated the risk of epidemic events by simulating the introduction of Chikungunya virus into three naïve US populations. Assuming established mosquito populations in each area, we introduced one exposed individual to evaluate the epidemic potential size of an outbreak, taking into account the population dynamics of the vector and its susceptibility to temperature regimes. We predicted low epidemic risk for disease introduction during periods of low vector abundance and high epidemic risk for certain critical periods that show increasing, or high, vector abundance. These results provide valuable additional information not only for early warning systems but also for the implementation of intervention strategies with the goal of reducing vector populations or human risk of exposure.
To study the dynamics of the introduction of CHIKV in an immunologically naive population we constructed a model with demographic stochasticity for mosquitoes and humans (Figure 1, see Material-S1 for model equations). Using a classical approach, the human host population was divided into susceptible (S), exposed (E), symptomatic infective (IS), asymptomatic infective (IA) and recovered (R) classes. Analogously, the adult mosquito population was divided into susceptible (S), exposed (E) and infected (I) classes. In addition, we considered the immature stages of mosquito population, including mosquito eggs (G), larvae and pupae (L) and eggs undergoing diapause (D). Vital mosquito rates in this model were temperature dependent (Table S1 and Figure 2). Density-dependent effects were added to both mosquito and human populations. It is worthwhile to note that although the net effect of having density-dependent terms in the model is to avoid uncontrolled population growth, they represent a broad range of factors from larval overcrowding effects to human behavior.
The functional forms of the temperature forcing on the parameters for the dynamic of the vector population are presented in Figure 2 (The mathematical forms are presented in Material S1). Two temperature thresholds (TsD and TeD) were used to determine the diapause state. Eggs entered diapause (i.e., arrested development) when temperature was below TsD, and eggs do not undergo diapause for temperatures above TeD. The proportion of eggs undergoing (avoiding) the diapause state linearly decreased (increased) with increasing temperature for environmental conditions between TsD and TeD. Although the determination of diapause periods follows a complex combination of factors — including temperature and photoperiod— temperature was used as a proxy for such combination in here. Dependence on temperature of both egg survival and development time was fitted to experimental data (Harrington, unpublished data) and reports from the literature [47], [48]. Similarly, experimental and reported data were used for the fitting of temperature influence on larvae survival and development time [Harrington, unpublished data, [47], [48]. Adult longevity dependence on temperature was fitted to experimental data using the general assumption that longevity declines linearly with temperatures under 10°C [49]. Although there is no experimental support reported in the literature, CHIKV extrinsic incubation period (EIP) was assumed to be reduced with increasing temperature (up to 32°C) as with other arboviruses such as DENV [50]. The reported minimum extrinsic incubation period for CHIKV in several studies is 2 days [15]. Hence, we modeled EIP as a linear function with a minimum duration of 1.5 days at 32°C and a maximum duration of 4 days at 10°C [15], [51]. For the purposes of the current model we assumed transmission rates based on early experimental work [52].
Daily temperatures were calculated by applying a spline interpolation to the monthly mean temperature data of the last decade obtained from the Intergovernmental Panel on Climate Change (http://www.ipcc-data.org) (Figure S1). The model was run using temperature data from different locations to evaluate variability on epidemic risk with temperature patterns. Here, we present the results for three major US ports of entry that encompass a wide seasonal variation in temperature: New York, Atlanta, and Miami.
Population sizes and carrying capacity parameters for human populations were estimated using the city size data reported in the last census (http://www.census.gov). Mosquito population carrying capacity was estimated assuming a maximum number of vectors per host. We ran independent simulations for the three ports of entry changing this ratio (we present here the results for 0.5, 1 and 3 mosquitoes per human).
The model was run for five years. Initial population sizes in the model were selected according to the expected equilibrium values. During the first year of simulation there was no disease present in the model and therefore both human and mosquito populations drifted to their respective equilibria. CHIKV was introduced into the model during the second year of simulation via one exposed individual, and the simulation was run until the end of the fifth year. We calculated the final number of infective individuals, the number of infected at the outbreak peak, and the time to reach the outbreak peak from the day of introduction for each one of 1000 Monte Carlo simulations. We ran the simulations systematically varying the day of introduction of the disease from January 1st to December 31st, which allowed us to express the outbreak probability as a function of the introduction day (Figure 3). Here, the probability of outbreak was defined as the frequency of cases where the chain of infection was functional (i.e., the number of infected individuals was bigger than 1). In addition, we calculated the risk of an outbreak as the mean of the final epidemic size (summation of all infected individuals) over the average susceptible population size (Figure 4). These simulations were replicated not only varying the ratio of mosquitoes to humans (considering values of 0.5, 1 and 3 for that ratio) but also reducing the mosquito feeding pattern for human blood from 100% to 25%.
Figure 3 displays the probability of outbreak for the different locations as function of the day of introduction and 0.5 mosquitoes to human ratio for each location, and a 100% human meal preference (See Material S1 for results with other vector to host ratios and human blood meal patterns). The probability of an outbreak (defined as at least one successful transmission event to a human) for New York shows a peak around 38% for a CHIKV introduction in August, and is over 30% during the interval from August 6th to September 11th. In addition, there is a significant probability of outbreak after an introduction on June 15th and up to December. The probability of having an outbreak late in November is very small and it is a consequence of using mean monthly temperature data as a basis for the temperature patterns.
Outbreaks also were seasonal for Atlanta, with no significant probability of outbreaks after introductions between January 12th and April 9th. Moreover, in Atlanta, the probability of outbreak was greater than 30% for a longer period, extending from June 6th to September 26th, with peak values similar to those in New York. In contrast, for Miami chances of a CHIKV outbreak were significant after an introduction at any time during the year.
Our model only demonstrated the occurrence of at least one successful transmission event, however, the maximum prevalence reached for those outbreaks is likely to be a more important parameter (Figure 4). Consequently, we explored the peak infection rate with our model and found that it varies significantly between locations and also with the ratio of vectors to hosts and human feeding patterns (see Tables S2 and S3 for results with other vector to host ratios and meal preferences). When we set the ratio of mosquitoes to one human at 0.5 and human feeding rates at 100%, peak infection in New York was very small (0.0002%), peaking at the beginning of the high probability outbreak period. A similar pattern, but with higher prevalence values for CHIK in humans (0.1381%) was observed for Atlanta. In Miami, however a high mean prevalence for CHIK in humans (25.0187%) was observed throughout the year.
Additionally, we calculated the number of days from pathogen introduction until the peak prevalence (Table 1). These calculations reveal that (in general), when mosquito feeding preference is set to be only from humans, the time to peak prevalence is longer than when feeding preferences are broad. Thus, when human blood feeding is low, the epidemic peak usually occurs shortly after introduction because the chain of transmission could be easily interrupted. However, when human blood feeding is set to 100%, the epidemic peaks occurs within 20 days for cities in cooler climates (i.e. New York) and approximately one to three months for warmer locations. In these warmer locations several secondary cases are expected to follow the index case, and the stochastic interruptions of the chain of infection can only slow the development of the epidemic instead of stopping it.
Since the occurrence of the CHIKF outbreak in Italy in 2007, the risk of similar outbreaks in the United States and other temperate countries has become a public health concern [42]. Our models, based on the introduction of one exposed individual, show that the probability of an outbreak in any of the three chosen locations varies by geographic location. As expected, for areas like the New York metropolitan region and Atlanta, that display a strong seasonality in temperature and therefore in mosquito abundance, this risk is bounded to the summer time and low prevalence levels are predicted. In locations with temperature patterns similar to those in Miami, which allow for year-round mosquito activity, the risk for outbreak is not bounded seasonally. In addition, predicted prevalence levels are higher for Miami because usually the chain of infection is not completely interrupted. These models also show that the proportion of people affected in an outbreak is reduced dramatically with increasing latitude (Figure 4). Furthermore, the models suggest the replenishment of susceptible individuals is not enough to create endemic foci, however this result could change when using more realistic models.
Our model outputs display higher sensitivity to parameters controlling the proportion of blood meals from humans than the vector: host ratio. Nevertheless, the increased ratio of mosquitoes to humans led to a two-fold increase in the probability of outbreak at all locations. This result highlights the importance of vector control to reduce both the risk of outbreaks and the proportion of infected individuals. It is important to note that, in this modeling approach, both parameters may be interpreted as proxies for a reduction of human exposure to mosquitoes. Hence, our results confirm the relevance of public campaigns advising residents to control mosquitoes at home and take precautions to avoid mosquito exposure to reduce disease outbreaks.
The time between CHIKV introduction and peak outbreaks revealed that, for locations with temperature patterns similar to those of Miami where mosquito populations may not undergo diapause, CHIKV infections might circulate at low levels for several months until reaching dramatic proportions. Early detection of cases in these regions will be important to reduce the magnitude of an outbreak. However, in locations such as New York and Atlanta, a critical temporal window for interventions could be identified and intervention during such periods may be enough to significantly reduce the probability of an outbreak.
It is clear that these predictive models are highly sensitive to temperature patterns that govern mosquito population dynamics, and could be improved by using non-averaged temperature data (i.e. sampling from the distribution of temperatures), and including other environmental factors such as rainfall and photoperiod that can have a significant influence on vector populations. In addition, reduction of individual exposure (only modeled as a reduction in human feeding patterns here) should be considered in order to have more accurate predictions. This modeling approach highlights the fact that a better understanding of epidemiological dynamics will require further studies on both biological and non-biological processes. Especially important will be: (1) further studies on diapause, abundance and feeding biology of Ae. albopictus, (2) the inclusion of multiple disease introduction events, either simultaneously of temporally spread, and (3) a better understanding of the evolution and plasticity of both pathogen and vector.
Our results strongly suggest that, in the event of an introduction and establishment of CHIKV in the United States, endemic and epidemic regions would emerge initially, mainly defined by environmental factors controlling annual mosquito population cycles. These regions should be identified in order to plan different intervention measures. In addition, reducing mosquito population sizes (and, consequently, reducing vector: human ratios) can lower the probability and magnitude of outbreaks mainly for regions with strongly marked seasonal temperature patterns.
Typical control strategies for vector borne diseases are: (1) reduction of vector population, (2) reduction of host exposure to infectious mosquito bites, and (3) isolation of infective hosts. This model also allows for evaluation of the effects of changes in the mosquito feeding patterns. Simulation results suggest that a reduction of vector population and human exposure could be very effective for a reduction of both the risk of an outbreak and the population at risk.
The results presented here simulating significant CHIK outbreaks in the US were based on a conservative approach of one exposed individual introduced to a region [45]. Given CHIKV infections in returning US travelers [45], [46] and the low numbers of infected individuals needed to spark an outbreak, we conclude that US health systems should be vigilant.
|
10.1371/journal.pgen.0030185 | Genetic Variation and Population Structure in Native Americans | We examined genetic diversity and population structure in the American landmass using 678 autosomal microsatellite markers genotyped in 422 individuals representing 24 Native American populations sampled from North, Central, and South America. These data were analyzed jointly with similar data available in 54 other indigenous populations worldwide, including an additional five Native American groups. The Native American populations have lower genetic diversity and greater differentiation than populations from other continental regions. We observe gradients both of decreasing genetic diversity as a function of geographic distance from the Bering Strait and of decreasing genetic similarity to Siberians—signals of the southward dispersal of human populations from the northwestern tip of the Americas. We also observe evidence of: (1) a higher level of diversity and lower level of population structure in western South America compared to eastern South America, (2) a relative lack of differentiation between Mesoamerican and Andean populations, (3) a scenario in which coastal routes were easier for migrating peoples to traverse in comparison with inland routes, and (4) a partial agreement on a local scale between genetic similarity and the linguistic classification of populations. These findings offer new insights into the process of population dispersal and differentiation during the peopling of the Americas.
| Studies of genetic variation have the potential to provide information about the initial peopling of the Americas and the more recent history of Native American populations. To investigate genetic diversity and population relationships in the Americas, we analyzed genetic variation at 678 genome-wide markers genotyped in 29 Native American populations. Comparing Native Americans to Siberian populations, both genetic diversity and similarity to Siberians decrease with geographic distance from the Bering Strait. The widespread distribution of a particular allele private to the Americas supports a view that much of Native American genetic ancestry may derive from a single wave of migration. The pattern of genetic diversity across populations suggests that coastal routes might have been important during ancient migrations of Native American populations. These and other observations from our study will be useful alongside archaeological, geological, and linguistic data for piecing together a more detailed description of the settlement history of the Americas.
| Patterns of genetic diversity and population structure in human populations constitute an important foundation for many areas of research in human genetics. Most noticeably, they provide an invaluable source of data for inferences about human evolutionary history [1–3]. In addition, the distribution of genetic variation informs the design and interpretation of studies that search for genes that confer an increased susceptibility to disease [4–6].
Recent genomic studies have produced detailed genome-wide descriptions of genetic diversity and population structure for a wide variety of human populations, both at the global level [7–19] and for individual geographic regions, including East Asia [20], Europe [21,22], and India [23]. Here we report the first such analysis of indigenous populations from the American landmass, using 678 microsatellites genotyped in 530 individuals from 29 Native American populations. The study is designed to investigate several questions about genetic variation in Native Americans: what records of the original colonization from Siberia are retained in Native American genetic variation? What geographic routes were taken in the Americas by migrating peoples? What is the genetic structure of Native American populations? To what extent does genetic differentiation among populations parallel the differentiation of Native American languages? In addressing these questions, our analyses identify several surprising features of genetic variation and population history in the Americas.
We collected genome-wide microsatellite genotype data for 751 autosomal markers in 422 individuals from 24 Native American populations spanning ten countries and seven linguistic “stocks” (Tables S1 and S2). We also collected data on 14 individuals from a Siberian population, Tundra Nentsi. To enable comparisons with data previously reported in the worldwide collection of populations represented by the Human Genome Diversity Project–Centre d'Etude du Polymorphisme Humain (HGDP–CEPH) cell line panel [7,11,13], data analysis was restricted to 678 loci typed across all populations (see Methods). The combined dataset contains genotypes for 1,484 individuals from 78 populations, including 29 Native American groups and two Siberian groups (Figure 1).
We compared levels of genetic diversity across geographic regions worldwide (Table 1). A serial founding African-origin model of human evolution [10,11]—in which each successive human migration involved only a subset of the genetic variation available at its source location, and in which the Bering Strait formed the only entry point to the American landmass—predicts reduced genetic diversity in Native Americans compared to other populations, as well as a north-to-south decline in genetic diversity among Native American populations. Indeed, Native Americans were found to have lower genetic diversity, as measured by heterozygosity, than was seen in populations from other continents (Table 1). Additionally, applying a sample size-corrected measure of the number of distinct alleles in a population [24,25], Native Americans had fewer distinct alleles per locus compared to populations in other geographic regions (Figure 2A). Among Native American populations, the highest heterozygosities were observed in the more northerly populations, and the lowest values were seen in South American populations (Table 2). The lowest heterozygosities of any populations worldwide occurred in isolated Amazonian and eastern South American populations, such as Surui and Ache. More generally, heterozygosity was reduced in eastern populations from South America compared to western populations (Table 1, p = 0.02, Wilcoxon rank sum test). Eastern South American populations also had fewer distinct alleles per locus than populations elsewhere in the Americas (Figure 2B).
Assuming a single source for a collection of populations, the serial founding model predicts a linear decline of genetic diversity with geographic distance from the source location [11,26]. Such a pattern is observed at the worldwide level, as a linear reduction of heterozygosity is seen with increasing distance from Africa, where distance to Native American populations is measured via a waypoint near the Bering Strait (Figure 3A). To investigate the source location for Native Americans, we considered only the Native American data and allowed the source to vary, measuring the correlation of heterozygosity with distance from putative points of origin. Consistent with the founding from across the Bering Strait, the correlation of heterozygosity with geographic distance from a hypothesized source location had the most strongly negative values (r = −0.436) when the source for Native Americans was placed in the northernmost part of the American landmass (Figure 3B). The smallest value for the correlation coefficient was seen at 55.6°N 98.8°W, in central Canada, but as a result of relatively sparse sampling in North America, all correlations in the quartile with the smallest values, plotted in the darkest shade in Figure 3B, were within a narrow range (−0.436 to −0.424).
One way to examine the support for particular colonization routes within the American landmass is to determine if a closer relationship between heterozygosity and geography is observed when “effective” geographic distances are computed along these routes, rather than along shortest-distance paths. Using PATHMATRIX [27] to take the precise locations of continental boundaries into consideration in effective geographic distance calculations (see Methods)—rather than using a waypoint approach [11] to measure distance—does not substantially alter the correlation of heterozygosity with distance from the Bering Strait (r = −0.430, 1:1 coastal/inland cost ratio in Figure 4A). However, when coastlines are treated as preferred routes of migration in comparison with inland routes, the percent of variance in heterozygosity explained by effective distance increases to 34% (r = −0.585 for a coastal/inland cost ratio of 1:10 in Figure 4A). In contrast, all scenarios tested that had coastal/inland cost ratios greater than 1 explain a smaller proportion of the variance in heterozygosity than do the scenarios with coastal/inland cost ratio of 1 or less.
The preferred routes in the optimal scenario of a 1:10 coastal/inland cost ratio include a path to the Ache, Guarani, and Kaingang populations that travels around northern South America (Figure 4B). With these three populations excluded, the role of coastlines is almost unchanged (Figure S1), and a 1:10 ratio continues to explain the largest fraction of variation in heterozygosity (r = −0.595). Applying a reduced cost only to the Pacific coast, a preference is still seen for ratios slightly less than 1 compared to ratios greater than 1, and the scenario producing the closest fit is a 1:2 ratio (Figure S2). A stronger preference for a Pacific coastal route was observed excluding from the computations the Chipewyan, Cree, and Ojibwa populations, three groups that follow an Arctic route in Figure 4B, or excluding Ache, Guarani, and Kaingang in addition to Chipewyan, Cree, and Ojibwa (Figure S3). We did not find a closer fit of heterozygosity and effective distance assuming a reduced cost for travel along major rivers, and indeed we observed that a higher cost for riverine routes was preferred (Figure S3).
To investigate population structure at the worldwide level, we used unsupervised model-based clustering as implemented in the STRUCTURE program [28,29]. Using STRUCTURE, we applied a mixture model that allows for allele frequency correlation across a set of K genetic clusters, with respect to which individual membership coefficients are estimated (see Methods).
As has been observed previously [7,9,13,16,23], cluster analysis with worldwide populations identifies a major genetic cluster corresponding to Native Americans (Figure 5), indicating an excess similarity of individual genomes within the Americas compared to genomes in other regions. Inclusion of the Native American data collected here did not substantially alter the clusters identified in previous analyses. When the genotypes were analyzed using a model with five clusters, the clusters corresponded to Sub-Saharan Africa, Eurasia west of the Himalayas, Asia east of the Himalayas, Oceania, and the Americas. For a model with six clusters, the sixth cluster corresponded mainly to the isolated Ache and Surui populations from South America. Almost no genetic membership from the cluster containing Africans and a relatively small amount of membership from the cluster containing Europeans were detected in the Native Americans, indicating that with relatively few exceptions, the samples examined here represent populations that have experienced little recent European and African admixture.
To search for signals of similarity to Siberians in the Native American populations, we used a supervised cluster analysis [28,29] in which Native Americans were distributed over five clusters (Figure 6). Four of these clusters were forced to correspond to Africans, Europeans, East Asians excluding Siberians, and Siberians (Tundra Nentsi and Yakut), and the fifth cluster was not associated with any particular group a priori. Most Native American individuals were seen to have majority membership in this fifth cluster, and considering their estimated membership in the remaining clusters, Native Americans were genetically most similar to Siberians. A noticeable north-to-south gradient of decreasing similarity to Siberians was observed, as can be seen in the declining membership in the red cluster from left to right in Figure 6. Genetic similarity to Siberia is greatest for the Chipewyan population from northern Canada and for the more southerly Cree and Ojibwa populations. Detectable Siberian similarity is visible to a greater extent in Mesoamerican and Andean populations than in the populations from eastern South America.
The level of population structure observed among Native Americans, as determined using FST [30], was 0.081, exceeding that of other geographic regions (Table 1). Comparing regions within the Americas, the highest FST value was observed in eastern South America, with intermediate values occurring in western South America and Central America and with the smallest value occurring in North America (Table 1). These results are compatible with the lower overall level of Native American genetic variation, particularly in eastern South America, as the mathematical connection between heterozygosity and FST predicts that low heterozygosities will tend to produce higher FST values [11,31–33].
Applying unsupervised model-based clustering [28,29] to the Native Americans, considerable population substructure is detectable (Figure 7). For a model with two clusters, one cluster corresponds largely to the northernmost populations, while the other corresponds to populations from eastern South America; the remaining populations are partitioned between these two clusters, with greater membership of the more northerly populations in the “northern” cluster. As the number of clusters is increased, the least genetically variable groups form distinctive clusters (for example, the Ache, Karitiana, and Surui populations). However, variation exists across replicates in the nature of the partitioning, and to illustrate the range of solutions observed, Figure 7 summarizes each clustering solution that was seen in at least 12% of replicate analyses for each K from two to nine. These summaries indicate that the main clustering solutions with a given K “refine” the partitions observed with K − 1 clusters, in the sense that each of the K clusters is either identical to, or is a subset of, one of the K − 1 clusters. A likely explanation for the multimodality is the presence of several population subgroups that are roughly equally likely to form individual clusters. For small K, not enough slots are available, and only when K is sufficiently large is each of these groups able to occupy its own cluster.
For K = 7, a relatively stable clustering solution is observed, appearing in 44 of 100 replicates (compared to seven of 100 for the next most frequently observed solution). This clustering solution has distinctive clusters for three of the smallest and least genetically variable groups in the sample—Karitiana and Surui from Brazil, and Ache from Paraguay. Two separate samples from the Amazonian Ticuna group of Colombia form the basis for a cluster, as does the Pima group from Mexico. The remaining two clusters include one centered on the North American groups and one centered on the Chibchan–Paezan language stock from Central and South America. The cluster containing Chibchan–Paezan populations—the only cluster at K = 7 that corresponds well to a major language stock—separates into two subclusters when K is increased to nine. Despite the large geographic distance between Mesoamerica and the Andes, Mesoamerican populations (Mixtec, Zapotec, Mixe, and Maya from Mexico and Kaqchikel from Guatemala) and Andean populations (Inga from Colombia, Quechua from Peru, and Aymara and Huilliche from Chile) have similar estimated membership across clusters when K = 7, and together with five additional populations (Zenu, Wayuu, and Piapoco from Colombia, and Kaingang and Guarani from Brazil), they comprise a single cluster when K = 9.
We compared the classification of the populations into linguistic “stocks” [34,35] (Table S2) with their genetic relationships as inferred on a neighbor-joining tree constructed from Nei genetic distances [36] between pairs of populations (Figure 8). As the use of a single-family grouping (Amerind) of all languages not belonging to the Na–Dene or Eskimo–Aleutian families is controversial [37], we focused our analysis on the taxonomically lower level of linguistic stocks.
In the neighbor-joining tree (Figure 8), a reasonably well-supported cluster (86%) includes all non-Andean South American populations, together with the Andean-speaking Inga population from southern Colombia. Within this South American cluster, strong support exists for separate clustering of Chibchan–Paezan (97%) and Equatorial–Tucanoan (96%) speakers (except for the inclusion of the Equatorial–Tucanoan Wayuu population with its Chibchan–Paezan geographic neighbors, and the inclusion of Kaingang, the single Ge–Pano–Carib population, with its Equatorial–Tucanoan geographic neighbors). Within the Chibchan–Paezan and Equatorial–Tucanoan subclusters several subgroups have strong support, including Embera and Waunana (96%), Arhuaco and Kogi (100%), Cabecar and Guaymi (100%), and the two Ticuna groups (100%). When the tree-based clustering is repeated with alternate genetic distance measures, despite the high Mantel correlation coefficients [38] between distance matrices (0.98, 0.98, and 0.99 for comparisons of the Nei and Reynolds matrices, the Nei and chord matrices, and the Reynolds and chord matrices, respectively), higher-level groupings tend to differ slightly or to have reduced bootstrap support (Figures S4 and S5). However, local groupings such as Cabecar and Guaymi, Arhuaco and Kogi, Aymara and Quechua, and Ticuna (Arara) and Ticuna (Tarapaca) continue to be supported (100%). This observation of strongly supported genetic relationships for geographically proximate linguistically similar groups coupled with smaller support at the scale of major linguistic groupings is also seen in Native American mitochondrial data [39].
To more quantitatively test the correspondence of genetic and linguistic variation in the Americas, we computed the Mantel correlation of genetic and linguistic distances (Table 3). Nei's Da distance [36] was used for the genetic computations, and linguistic distances were measured along a discrete scale (see Methods). Considering all of the Native American populations and treating all linguistic stocks as equidistant (Table S3), the Mantel correlation of Nei genetic distance with linguistic distance is small (r = 0.04). The correlation is also small when using between-stock linguistic distance measures (Tables S4–S11) that make use of shared etymologies identified by Greenberg [34]. For two ways of computing linguistic distance, using the Dice and Jaccard indices (see Methods), respectively, the correlations are r = −0.01 and r = −0.02. When the effects of geography are controlled, or when stocks are excluded from the computation individually, the partial correlations of linguistic and genetic distance [40] remain low.
A potential explanation for the low correlation coefficients—suggested by the apparent genetic and linguistic correspondence in the neighbor-joining tree for closely related groups—is that sizeable correlation between genetic and linguistic distance may exist only below a certain level of linguistic distance. Considering genetic and linguistic differentiation only for pairs of populations within linguistic stocks, the correlation of genetic distance and linguistic distance increases (r = 0.53). The partial correlation of genetic distance and linguistic distance remains fairly high when the effect of geographic distance is controlled (r = 0.40), although 11% of random matrix permutations produce higher values (Table 3).
By excluding language stocks from the computation individually, it is possible to investigate the extent to which individual linguistic stocks are responsible for the within-stock correlation of genetic and linguistic distance. When the Equatorial–Tucanoan stock is excluded, the correlation increases to 0.68, and the partial correlation controlling for geographic distance increases to 0.66. Excluding the Andean stock, however, both the correlation and the partial correlation decrease (to 0.46 and 0.26, respectively). Excluding any of the three other stocks for which more than one population is represented (Northern Amerind, Central Amerind, Chibchan–Paezan) does not lead to a sizeable change in either the correlation coefficient (0.54, 0.51, 0.55) or the partial correlation coefficient (0.40, 0.39, 0.40).
Considering alleles found only in one major geographic region worldwide, Native Americans have the fewest private alleles (Figure 9A). Private alleles, which lie at the extreme ends of the allele size range more often than expected by chance (p < 0.023), usually have low frequencies in the geographic region where they are found (≤13%). Within the Americas, counting alleles private to one of four subregions, northern populations have the most and eastern South American populations have the fewest private alleles, with western South American populations having slightly more than Central American populations (Figure 9B).
Despite this general lack of high-frequency private alleles, especially in Native Americans, we observed that the only common (>13%) regionally private variant in the worldwide dataset was a Native American private allele. This allele, corresponding to a length of 275 base pairs at locus D9S1120, was found at a frequency of 36.4% in the full Native American sample, and was absent from the other 49 world populations. Allele 275 is the smallest variant observed at the locus and it is present in each of the 29 Native American populations—at frequencies ranging from 11.1% in Ticuna (Tarapaca) to 97.1% in Surui (Figure 10). This allele has now been observed in every Native American population in which the locus has been investigated [41,42], and it has only been seen elsewhere in two populations at the far eastern edge of Siberia [42].
Because of the likely submergence of key archaeological sites along the Pacific coast, the relative absence of a written record, and the comparatively recent time scale of the initial colonization, population-genetic approaches provide a particularly important source of data for the study of Native American population history [43–52]. In this article, building upon recent investigations that have increased the size of Native American genetic datasets beyond classical marker, Y-chromosomal, mitochondrial, and single-gene studies [7,11,13,16,41,53–65], we have examined genome-wide patterns of variation in a dataset that—in terms of total genotypes—represents the largest continent-wide Native American population-genetic study performed to date. Our results have implications for a variety of topics in the demographic history of Native Americans, including (1) the process by which the American landmass was originally populated, (2) the routes taken by the founders during and subsequent to the migration, and (3) the extent to which genes and languages have traveled together during the diversification of Native American populations. We discuss these issues in sequence.
The lower level of genetic diversity observed in the Americas compared to other continental regions is compatible with a reduction in population size associated with a geographically discrete founding, representing one of the most recent in a series of major bottlenecks during human expansions outward from Africa [11]. Gradients of genetic diversity (Figure 3) and decreasing similarity to Siberians (Figure 6) also point to extant Native Americans as the descendants of a colonization process initiated from the northwestern part of the American landmass. An alternative possibility that could produce a genetic diversity gradient—namely, a north-to-south gradient of recent admixture from high-diversity European populations—can be eliminated as a possible explanation given that (1) European admixture is not strongly correlated with distance from the Bering Strait (r = −0.135), (2) inclusion of a European admixture covariate in the regression of heterozygosity on distance from the Bering Strait is not supported (p = 0.37) and only slightly increases the fit of the regression model (R2 = 0.208 compared to R2 = 0.182), and (3) the regression of heterozygosity on distance from the Bering Strait does not change substantially when the most highly admixed populations are excluded from the analysis (Table S12). The genetic diversity and population structure gradients—which are generally compatible with principal component maps of allele frequencies at small numbers of classical markers [1,66] and with some analyses of mitochondrial, X-chromosomal, and Y-chromosomal data [67,68]—are more clearly visible in our study of a larger number of loci.
Although gradients of genetic diversity and Siberian similarity constitute major features of the pattern of Native American variation when considering all of the loci together, one important aspect of Native American variation—the distribution of a private allele at locus D9S1120—deviates from the genome-wide pattern and does not show a north-to-south frequency gradient. The geographic distribution of this allele is similar to the distributions of certain mitochondrial and Y-chromosomal variants that are also ubiquitous in the Americas, but that are absent elsewhere or that are found outside the Americas only in extreme northeast Siberia [69–74]. Such distributions are most easily explained by the spatial diffusion of initially rare variants during the colonization of the continent, rather than by continent-wide natural selection or by an origin considerably later than the colonization [42,75,76]. The restricted distribution in Asia of D9S1120 allele 275 and similar Y-chromosomal and mitochondrial variants suggests one of several explanations [42]: the ancestral population that migrated to the Americas may have already acquired a degree of genetic differentiation from other Asian populations [77], descendants of the original Native American founders are no longer present elsewhere in Asia, or these descendants have not yet been genotyped at loci that carry apparently private Native American variants.
The genomic continent-wide patterns observed here can be explained most parsimoniously by a single main colonization event, as proposed by some interpretations of archaeological, mitochondrial, and Y-chromosomal data [67,74,77–83]. In this view, at each step in the migration, a subset of the population splitting off from a parental group moves deeper into the Americas, taking with it a subset of the genetic variation present in the parental population. This scenario would be expected to produce a set of low-diversity populations with distinctive patterns of variation at the far terminus of the migration, such as those we and others [84] observe in the Ache and Surui populations. It can also explain the gradient of Siberian similarity, and the continent-wide distribution of D9S1120 allele 275. Alternatively, similar patterns could result from gene flow across the Bering Strait in the last few thousand years, together with continual interactions between neighbors on both sides of the Bering Strait [47]. It is also possible to envision a series of prehistoric migrations, possibly from the same source population, with the more recent descendants gradually diffusing into pre-existing Native American populations.
Largely on the basis of archaeological data, a classical model for the colonization of the Americas posits that humans entered the region towards the end of the Wisconsin glaciation (∼11,000 y ago) via a mid-continental ice-free corridor between the Cordilleran and Laurentide glaciers [78,79]. According to this model, migration southwards would have followed a pattern with a front of advance at approximately the same latitude across North America.
It is interesting to consider the patterns of genetic structure observed here within the context of the emphasis placed recently on the Pacific coast as an alternative to the inland ice-free corridor route of population dispersal in the Americas [79,85–87]. The late timing of the rapid inland colonization model has been put into some doubt by the discovery of early archaeological sites that predate by thousands of years the most recent deglaciation of North America [88]. In addition, recent geological evidence indicates that ice-free areas west of the Cordilleran ice sheet may have existed as early as ∼14,000 y ago [79], suggesting the possibility of an early coastal migration. Within South America, the coastal colonization model suggests an early southward migration along the western side of the Andes and is consistent with an interpretation that modern speakers of Andean languages may represent descendants of the first occupiers of the region [1]. Recent computer simulations also suggest that a coastal colonization model may more easily explain observed patterns of classical marker and mitochondrial DNA diversity [89].
Several observations from our data are compatible with the proposal of a coastal colonization route. The stronger correlation of genetic diversity with geographic distance when higher coastal mobility is taken into account (Figure 4) supports a possible role for population dispersals along the coast (note, however, that the difference in the tree structure induced by the optimal route in Figure 4 and the tree in Figure 8 suggests that alternative routes might be preferred if more aspects of the genetic data were incorporated into the coastal analysis). Consistent with observations of recent migration paths of certain Amazonian populations [43], we did not find support for migrations along major rivers. Finally, the relative genetic similarity of Andean populations to populations from Mesoamerica (Figure 7) is also compatible with an early Pacific coastal colonization. Under this view, the east-to-west difference in genetic diversity in South America, a pattern also observed with mitochondrial and Y-chromosomal markers [90–92] (including the extremely low diversity in the Ache [93,94] and Surui [94] populations), could reflect an initial colonization of western South America followed by subsampling of western populations to form the eastern populations.
An alternative interpretation of the Mesoamerican and Andean similarity is that this pattern is recent in origin. In this case, the reduced diversity and increased population structure in eastern South America may reflect a deep divergence between western and eastern populations, so that their different levels of differentiation could result from different levels of gene flow and genetic drift in western and eastern South America. The genetic similarity among Andean populations, and their relative similarity to the populations sampled from Mesoamerica, would perhaps then reflect recent gene flow along the coast.
Similar to results seen in some mitochondrial studies [95–97], Central American and South American populations from the Chibchan–Paezan language stock had slightly reduced heterozygosity compared to neighboring populations. Interestingly, the Cabecar and Guaymi populations from lower Central America (Costa Rica and Panama) were robustly placed at the tips of a northwest South American Chibchan–Paezan cluster in the tree of Figure 8. One explanation of this observation is that these populations may be of South American origin, as the ancestral group for the cluster could have been a South American population, most of whose descendants remain in South America. Alternatively, the large cluster containing the Chibchan–Paezan and Equatorial–Tucanoan populations could be the result of a colonization of South America separate from the colonization by the Andean populations—with the founder population possibly speaking a language from which modern Chibchan–Paezan languages have descended [98]. In this view, Guaymi and Cabecar are the only sampled Central American populations descended from the ancestors of this second migration.
At a qualitative level, the topology of the tree of Figure 8 shows some correspondence between genetic distance and linguistic stock assignment. High bootstrap values are seen for population clusters corresponding mainly to speakers of Chibchan–Paezan and Equatorial–Tucanoan languages and, to a lesser extent, Central Amerind languages. Although the high bootstrap values support previous qualitative comparisons that have suggested a considerable degree of relationship between genetic and linguistic distances [1], quantitative analyses based on matrix correlation coefficients for genetic and linguistic distances have been somewhat more equivocal [39,99–101]. Indeed, the correlation of genetic and linguistic similarity considering all populations in our dataset is quite small (Table 3). Considering only pairs of populations from within major language stocks, however, the correlation increases. Although several populations that do not group in the neighbor-joining tree with their linguistic neighbors appear most genetically similar to their geographic neighbors, the correlation remains moderate when geographic distance is controlled. The within-stock correlations are in most cases not unusually high when applying permutation tests, but are perhaps suggestive that at the local scale, dissimilarities in languages either play a partial role in producing genetic barriers or otherwise co-occur with factors that impede gene flow. The lack of a more general correlation may be due to such factors as deviations from a tree-like history for genetic evolution or for linguistic evolution, or to uncertainties in the linguistic classification [39].
In a genomic study of a relatively large number of Native American populations, our work provides support to a variety of hypotheses about fundamental aspects of Native American demographic history. In particular, we find genetic evidence that supports a single main colonization event from Siberia, a coastal colonization route, and a divergence process that may have been facilitated at the local scale partly by differences between languages. As genomic data proliferate, more formal genetic tests of these hypotheses, together with accumulating evidence from fields such as archaeology [78,79,102], geology [103], and linguistics [104–106], will surely result in a more detailed picture of the settlement by and differentiation of indigenous human populations in the American landmass.
A total of 436 individuals from 24 Native American populations and one Siberian population were included in this study, in addition to data on 1,048 individuals from 53 worldwide populations represented in the HGDP–CEPH human genome diversity cell line panel [107,108]. Alternate names for the Native American populations, together with sample sizes and approximate geographic coordinates, are given in Table S1.
Populations from the HGDP–CEPH panel were classified into geographic regions as in Rosenberg et al. (2002) [7], and the Tundra Nentsi population from Siberia was classified as East Asian. In analyses subdivided by geographic region within the Americas, we grouped the populations as North American (Chipewyan, Cree, Ojibwa), Central American (Cabecar, Guaymi, Kaqchikel, Maya, Mixe, Mixtec, Pima, Zapotec), western South American (Arhuaco, Aymara, Embera, Huilliche, Inga, Kogi, Quechua, Waunana, Wayuu, Zenu), and eastern South American (Ache, Guarani, Kaingang, Karitiana, Piapoco, Surui, Ticuna [Arara], and Ticuna [Tarapaca]). The populations from Mexico, which except Pima were from the southern part of Mexico, were considered as part of the Central American group. Populations were placed linguistically using the classification of Ruhlen [35]. Although disagreement exists about linguistic classifications in the Americas, there is greater agreement at the level of linguistic stocks and at lower levels in the linguistic classification hierarchy, on which we focus.
Each of the newly sampled individuals was genotyped by the Mammalian Genotyping Service for 751 microsatellites spread across all 22 autosomes. The microsatellite markers were drawn from Marshfield Screening Sets 16 and 54 (http://research.marshfieldclinic.org/genetics/). Considering all individuals, we checked each pair of markers to determine if genotypes at one member of the marker pair were identical to those at the other member of the pair, up to a constant of translation. This procedure identified one pair of duplicated markers—MFD600 and MFD601—and MFD600 was discarded from the analysis.
Among the 750 remaining microsatellites that were genotyped in the new samples, 693 had previously been genotyped in the HGDP–CEPH diversity panel [7,11,13]. For some of these loci, there was a change in primer length or position between the two studies, or a systematic change occurred in the algorithm by which allele size was determined from raw genotyping products—or both. In cases where the primers changed, allele sizes from the new dataset were adjusted by the appropriate length in order to align its list of allele sizes with the earlier list for the HGDP–CEPH dataset.
To identify systematic changes between datasets, for each locus the allele sizes of one dataset were translated by a constant and the G test statistic of independence between allele frequencies and dataset (older HGDP–CEPH dataset versus newly genotyped dataset) was then computed [23]. Considering all possible constants for translation of allele sizes, the one that minimized the G statistic was determined. In implementing the G test, two groups of comparisons were performed. In the first group of comparisons, the constant of translation was determined by comparing 80 Jewish individuals genotyped simultaneously with the Native Americans to all 255 individuals from Europe and the Middle East in the HGDP–CEPH H1048 dataset [109], excluding Mozabites. The second group of comparisons involved 346 Native American individuals from Central and South America in this newer dataset (all 336 sampled Central and South Americans excluding Ache, and ten additional individuals who were later excluded) and 63 Native American individuals from the Maya, Pima, and Piapoco populations in the older H1048 dataset (the Piapoco population is described as “Colombian” in previous analyses of these data). The constants expected based on the two G tests—labeled S1 for the comparison of the Jewish populations to European and Middle Eastern populations and S2 for the Native American comparison—were then compared with the constant of translation expected from consideration of three additional sources of information available for the two datasets: the genotypes of a Mammalian Genotyping Service size standard (S3), a code letter provided by the Mammalian Genotyping Service indicating the nature of the change in primers (S4), and the locations of the primers themselves in the human genome sequence (S5).
Among the 693 markers, 687 had the same optimal constant of translation (that is, the constant that minimizes the G statistic) in the two different sets of population comparisons (S1 = S2). The remaining six markers with different optimal constants of translation in the two G tests were compared with the value expected from the locations of the old and new primers in the human genome (S5). In all six cases, the optimal constant for the comparison of the Jewish and European/Middle Eastern datasets agreed with the value based on the primer locations (S1 = S5). As real population differences between datasets are more likely in Native Americans due to the larger overall level of genetic differentiation in the Americas, we used the constant obtained based on the Jewish and European/Middle Eastern comparison (S1) for allele size calibration.
Of the remaining 687 markers, 638 had an optimal constant of translation that agreed with the value expected based on the code letter provided by the Mammalian Genotyping Service (S1 = S2 = S4). Thus, there were 49 markers for which the code letter was either uninformative or produced a constant of translation that disagreed with S1 and S2. For 35 of these markers, the constant of translation based on the size standard (S3) agreed with S1 and S2. For eight of the remaining 14 markers, the constant of translation based on the primer sequences (S5) agreed with S1 and S2. The six markers with disagreements (AAT263P, ATT070, D15S128, D6S1021, D7S817, and TTTAT002Z), having S1 ≠ S5, were then discarded. For the remaining 687 markers that were not discarded, 685 had G < 48 in both G tests, while the other two markers (D14S587 and D15S822) had G > 91 in the Jewish versus European/Middle Eastern comparison. These two extreme outliers, which also had the highest G values for the Native American comparison, were then excluded (Figure S6).
To further eliminate loci with extreme genotyping errors, we performed Hardy-Weinberg tests [110] within individual populations for the 685 remaining markers. This analysis, performed using PowerMarker [111], used only the 44 populations in which all 685 markers were polymorphic. We calculated the fraction of populations with a significant p-value (<0.05) for the Hardy-Weinberg test (Figure S7). Two markers (GAAA1C11 and GATA88F08P) were extreme outliers, with more than 43% of populations producing p < 0.05. For the remaining markers, the proportion of tests significant at p < 0.05 varied from 0 to 35% without any clear outliers, and with most markers having less than 10% of tests significant at p < 0.05. Excluding the two Hardy-Weinberg outliers, 683 markers remained. Five additional markers (AGAT120, AGAT142P, D14S592, GATA135G01, and TTTA033) were excluded due to missing data: for each of these markers there was at least one population in which all genotypes were missing. Thus, 678 loci remained for the combined analysis with the HGDP–CEPH panel.
After the elimination of problematic markers, ten individuals who had potentially been mislabeled were discarded. Seven of these were admixed individuals from Guatemala who, through a clerical error, had been incorporated in the data cleaning phase of the study as members of the Kaqchikel population. The other three were individuals who, on the basis of elevated allele sharing, were inferred to be siblings, but who were classified as belonging to two different populations (Wayuu and Zenu). The final dataset, combining the HGDP–CEPH data and the new data, contained 1484 individuals and 678 markers, with a missing data rate of 4.0%. Each marker had some data present in all populations, with a minimum 88.7% genotypes per marker and 50.1% genotypes per individual. Of the 1,484 individuals, 1,419 had a missing data rate of less than 10%.
Identification of pairs of close relatives was performed using identity-by-state allele sharing combined with likelihood inference as implemented in Relpair [112,113]. A critical value of 100 was used in the likelihood analysis, and the genotyping error rate was set at 0.008. In each population, Relpair was applied using count estimates of allele frequencies in that population. Identification of recommended panels with no first-degree relatives and with no first- or second-degree relatives followed the procedure of Rosenberg [109], except that when an arbitrary decision was required about which individual in a relative pair should be excluded, the individual with more missing data was discarded. Beginning from the 436 newly sampled individuals (termed panel N436), this analysis produced a panel of 379 individuals with no first-degree relatives and a panel of 354 individuals with no first- or second-degree relatives. These panels are termed N379 and N354. Details on the properties of these panels can be found in Tables S13–S26, and plots of allele sharing are shown in Figures S8–S13.
Geographic coordinates for the newly sampled populations are specified in Table S1, and coordinates for the other populations were taken from Rosenberg et al. [13]. For the production of Figure 3B, distances between populations were computed using great circle routes [13], with obligatory waypoints as specified by Ramachandran et al. [11]. Routes to South America required an additional waypoint in Panama at 8.967°N 79.533°W. The computation of Figure 3B excluded the waypoint used by Ramachandran et al. [11] at Prince Rupert, and did not use the Panama waypoint when the origin was placed on a Caribbean island. Geographic distances from East Africa (Figure 3A) were computed using an origin at Addis Ababa [11].
Compared to the waypoint-based geographic distances, effective distances incorporate more detailed information on the effects of landscape components. They are computed as least-cost paths on the basis of a spatial cost map that incorporates these landscape components. For example, a coastal/inland ratio of 1:10 means that it is ten times more costly to go through land than through coastline. The effective distance between two points is computed as the sum of costs (so-called “least-cost distance”) along the least-cost path connecting the points. Because the relative costs of landscape components are somewhat arbitrary, several combinations were tested. We used PATHMATRIX [27] to compute least-cost distances based on a “uniform” cost over the continent (that is, when the boundaries of continental landmasses are the only spatial constraint, so that the coastal/inland cost ratio is 1:1), as well as using the following coastal/inland relative cost combinations: 1:2, 1:5, 1:10, 1:20, 1:30, 1:40, 1:50, 1:100, 1:200, 1:300, 1:400, and 1:500. Inverse cost combinations were also tested (2:1, 5:1, 10:1, 20:1, 30:1, 40:1, 50:1, 100:1, 200:1, 300:1, 400:1, 500:1). We also considered scenarios where the cost differed only for the Pacific coast instead of for all coasts, and where it differed not along coasts, but along major rivers.
Least-cost paths were computed on a Lambert azimuthal equal-area projection of the American landmass (central meridian 80°W, reference latitude 10°N) divided into a grid of 100 km2 square cells. For each cost scheme, we computed a Pearson correlation between heterozygosity and effective distance from the Bering Strait, as specified by the Anadyr waypoint [11] at 64°N 177°E, and we obtained its significance by using the t-distribution transformation [114].
For each population, expected heterozygosity was computed for each locus using an unbiased estimator [115], and the average across loci was taken as the population estimate. Heterozygosity was calculated for pooled collections of populations, and average heterozygosity across populations was obtained within individual geographic regions. Computations of FST were performed using Equation 5.3 of Weir [30], with confidence intervals obtained using 1,000 bootstrap resamples across loci.
To assess whether private alleles lie more often at the ends of the allele size range, for a given allele frequency cutoff, c, all private alleles with frequency at least c in their region of occurrence were obtained. Under the null hypothesis that all alleles are equally likely to be private, the number of private alleles expected to be at one of the two ends of the allele size range was obtained as the sum over the private alleles of 2/ki, where ki denotes the number of distinct alleles worldwide at the locus that produced private allele i. A difference from the value expected was evaluated using a chi-square goodness-of-fit test with one degree of freedom. Considering this test for all possible cutoffs c below 0.06 (above which only seven private alleles were observed), the most conservative p-value was 0.0228, although most values of c produced considerably more stringent p-values (Figure S14). In depicting allele frequencies at tetranucleotide locus D9S1120 (Figure 10), five of 2,914 observations not differing from the remaining alleles by a multiple of four are grouped with the nearest allele sizes (in one case where the allele was halfway between steps, it was assigned to the larger allele).
Analysis of population structure was performed using STRUCTURE [28,29]. Replicate runs of STRUCTURE used a burn-in period of 20,000 iterations followed by 10,000 iterations from which estimates were obtained. All runs were based on the admixture model, in which each individual is assumed to have ancestry in multiple genetic clusters, using the F model of correlation in allele frequencies across clusters. Graphs of STRUCTURE results were produced using DISTRUCT [116].
Worldwide population structure. Using the full worldwide data, ten replicate unsupervised STRUCTURE runs were performed for each value of the number of clusters K from one to 20. For each pair of runs with a given K, the symmetric similarity coefficient [117] (SSC) was computed as a measure of the similarity of the outcomes of the two population structure estimates. Using the Greedy algorithm of CLUMPP [117], distinct modes among the ten runs with a given K were then identified by finding sets of runs so that each pair in a set had SSC ≥ 0.9. The average was then taken of the estimated cluster membership coefficients for all runs with the same clustering mode. Of the ten runs, the number of runs that exhibited the mode shown was ten for K = 2 and K = 4, nine for K = 3 (with the tenth run grouping Africans and East Asians rather than Europeans and other Asians), five for K = 5 (with the remaining runs subdividing various combinations among Karitiana, Surui, and Ache, rather than separating the two populations from Oceania), and six for K = 6 (with the remaining runs subdividing the Native Americans into three clusters rather than separating the two populations from Oceania).
Supervised clustering. Using STRUCTURE, individuals from Europe, Sub-Saharan Africa, East Asia (excluding Siberia), and Siberia were forced into separate clusters, and supervised analysis of the Native American data was performed with K = 5 clusters. Ten replicates were performed, each of which yielded the same clustering mode, and the average membership coefficients across these replicates are displayed in Figure 6.
Native Americans. Using the Native Americans only, 100 replicate unsupervised STRUCTURE runs were performed for each value of K from one to 15 clusters. The settings for the runs were the same as in the worldwide analysis, and modes were identified in a similar manner. For K ≤ 9, average membership coefficients for the most frequently observed mode at each K are displayed in Figure 7. For each value of K, the figure presents the average membership estimates across all replicates that produced the most frequently occurring solution. Because of the high level of multimodality for K ≥ 3, no single mode provides a complete representation of the STRUCTURE results with a given K. Using CLUMPP [117], we identified all modes appearing at least 12 times in 100 replicates, using the SSC ≥ 0.9 criterion. Computations of SSC were based on the best alignment of the 100 replicate analyses obtained using the LargeKGreedy algorithm of CLUMPP with 1,000 (2 ≤ K ≤ 11) or 200 (12 ≤ K ≤ 15) random input sequences.
For 2 ≤ K ≤ 9, using the criterion SSC ≥ 0.9, the relation “in the same mode” had the property of being transitive, so that if runs (R1,R2) were in the same mode and runs (R2,R3) were in the same mode, then runs (R1,R3) were also in the same mode. For K ≥ 10, with the criterion SSC ≥ 0.9, “in the same mode” was not always transitive. While other cutoffs c could sometimes be identified so that “in the same mode” was transitive when the criterion SSC > c was applied, for K ≥ 10 there was no clear plateau in the cumulative probability distribution of SSC values across pairs of runs (Figure S15). Such plateaus, which are observed for 2 ≤ K ≤ 9, represent a large gap between SSC values for pairs of runs truly in the same mode (high SSC) and pairs of runs not in the same mode (lower SSC). The fact that for K ≥ 10 the probability is high that a randomly chosen pair of runs has SSC < 0.9 is also indicative of considerable multimodality across replicates.
Considering the modes with successive numbers of clusters, we identified all sets of modes with K+1 clusters that “refined” modes with K clusters. A mode with K+1 clusters is a refinement of a mode with K clusters if the mode with K+1 clusters consists of K−1 of the clusters in the K-cluster mode together with two clusters obtained by splitting the Kth cluster into two subgroups. More generally, a mode with K > L clusters refines a mode with L clusters if each cluster in the K-cluster mode is either the same as or a subdivision of a cluster in the L-cluster mode. As an example, in Figure 7, the mode depicted for K = 7 is a refinement of all modes depicted for smaller values of K.
For the Native American data, we performed a separate analysis using TESS [118,119], a genetic clustering program that estimates a preferred value of the number of clusters K less than or equal to a prespecified maximum value Kmax. If the estimated K equals Kmax, then the choice of Kmax is insufficiently large. Using the TESS admixture model with burn-in period of length 10,000 followed by 20,000 iterations from which estimates were obtained, we performed 200 runs of TESS with Kmax = 10, 20 each for ten values of a spatial autocorrelation parameter Ψ at intervals of 0.2 from 0.2 to 2. Of these 200 replicates, 183 supported an inference of K = 6, 7, 8, or 9, and only one supported an inference of K = 10. This suggests that the most important components of population structure are apparent with K < 10.
An unrooted neighbor-joining [120] population tree was constructed for the Native American and Siberian populations based on the Da distance of Nei et al. [36], which was found to perform comparatively well in estimation of population trees from microsatellite allele frequency data [121]. To visualize the tree, the root was placed between the Siberian and Native American populations. Confidence values were obtained from 1000 bootstrap resamples across loci. The computation of bootstrap distances was performed using PowerMarker [111], and the consensus tree was obtained and plotted using MEGA3 [122]. For comparison, trees based on Reynolds [123] and chord distances [124] were obtained analogously. Genetic distance matrices based on the Nei, Reynolds, and chord distances are shown in Tables S27–S29.
We used a discretized scale to measure linguistic distance [125,126]. Two populations from different language stocks or “groups” (Table S2) were scored as having distance 4, and within stocks, two populations had distance of 1, 2, or 3 depending on the level at which their languages diverged (Table S3). For some computations, we devised discretized measures of linguistic distance between stocks on the basis of shared and unshared etymologies tabulated in Table C.1 of Greenberg [34] (Tables S4 and S5). Using these etymologies, we computed the Dice (simple matching) and Jaccard indices of dissimilarity between stocks [127] (Tables S6 and S7), which we then converted into discretized between-stock distances (Tables S8 and S9).
For comparison with linguistic distances, Da genetic distances were used (Table S27), and the Mantel correlation coefficients [38] between pairs of distance matrices (among genetic, geographic, and linguistic) were obtained, with significance assessed using 10,000 permutations of rows and columns. Waypoint-based distances (Table S30) were used for the geographic computations. For computations within linguistic stocks, the correlation and significance level were computed as in tests involving the full matrix, except that all entries between language stocks were omitted from the evaluation of the correlation coefficient. Partial correlations of genetic and linguistic distance controlling for geographic distance were also obtained [128], with geographic distance calculated using the waypoint approach as above.
As the inclusion of relatives has the potential to influence various types of population-genetic analysis, we compared some of our results based on the full collection of 1,484 individuals to results based on 1,306 individuals—the H952 set from the HGDP–CEPH diversity panel [109] together with the N354 set from the newly genotyped individuals. The inclusion of relatives does not lead to a bias in allele frequency estimates (that is, E[p̂i] still equals pi), but it does inflate Var[p̂i]. The estimator Ĥ of heterozygosity is
, where n is sample size, the sum proceeds over alleles, and p̂i is the estimated frequency of allele i. Expanding the expression for the expectation E[Ĥ], it can be observed that the coefficient for the Var[p̂i] term is negative. Thus, inclusion of relatives is expected to reduce the estimate of heterozygosity through an increase in Var[p̂i].
The population heterozygosities based on the full and reduced datasets are plotted in Figure S16. The correlation coefficient between population heterozygosities based on the reduced and full datasets was 0.997; as expected, however, heterozygosity was systematically higher in the reduced set (mean difference of 0.0033 across populations; p< 0.001, Wilcoxon signed rank test). Given the greater proportion of individuals excluded when relatives were removed from N436 (18.8%) compared to H1048 (9.2%), the difference in heterozygosities between full and reduced datasets is greater in the 25 newly sampled populations (mean difference of 0.0052; p < 0.001) compared to the 53 HGDP–CEPH populations (mean difference of 0.0024; p < 0.001).
Despite the detectable effect of the removal of relatives on heterozygosity, the systematic nature of this small effect was such that very little difference was observed on the relationship of heterozygosity with distance from the Bering Strait (Figure S17). A number of other analyses, including the analyses of linguistic correlations and numbers of private and distinct alleles, also produced nearly identical inferences when relatives were excluded (Figures S18–S20 and Tables S31–S33), two exceptions being a noticeable decrease in population differentiation (Table S31) and a shift in the position of several populations in the neighbor-joining tree (Figure S19). Via the connection between heterozygosity and differentiation [11,31–33], the decrease in differentiation is a consequence of the increase in heterozygosity upon exclusion of relatives. In the case of the tree, despite a Mantel correlation of 0.99 between genetic distance matrices including and excluding relatives (Tables S27 and S33), the Cree, Huilliche, Maya, Ojibwa, Wayuu and Zenu populations shifted positions slightly, and the Kaqchikel population moved nearer to its geographic neighbors. Although the population groupings were generally quite similar, several bootstrap values decreased, magnifying the effect of the slight decrease in population differentiation. |
10.1371/journal.ppat.1005547 | PPARγ Is Activated during Congenital Cytomegalovirus Infection and Inhibits Neuronogenesis from Human Neural Stem Cells | Congenital infection by human cytomegalovirus (HCMV) is a leading cause of permanent sequelae of the central nervous system, including sensorineural deafness, cerebral palsies or devastating neurodevelopmental abnormalities (0.1% of all births). To gain insight on the impact of HCMV on neuronal development, we used both neural stem cells from human embryonic stem cells (NSC) and brain sections from infected fetuses and investigated the outcomes of infection on Peroxisome Proliferator-Activated Receptor gamma (PPARγ), a transcription factor critical in the developing brain. We observed that HCMV infection dramatically impaired the rate of neuronogenesis and strongly increased PPARγ levels and activity. Consistent with these findings, levels of 9-hydroxyoctadecadienoic acid (9-HODE), a known PPARγ agonist, were significantly increased in infected NSCs. Likewise, exposure of uninfected NSCs to 9-HODE recapitulated the effect of infection on PPARγ activity. It also increased the rate of cells expressing the IE antigen in HCMV-infected NSCs. Further, we demonstrated that (1) pharmacological activation of ectopically expressed PPARγ was sufficient to induce impaired neuronogenesis of uninfected NSCs, (2) treatment of uninfected NSCs with 9-HODE impaired NSC differentiation and (3) treatment of HCMV-infected NSCs with the PPARγ inhibitor T0070907 restored a normal rate of differentiation. The role of PPARγ in the disease phenotype was strongly supported by the immunodetection of nuclear PPARγ in brain germinative zones of congenitally infected fetuses (N = 20), but not in control samples. Altogether, our findings reveal a key role for PPARγ in neurogenesis and in the pathophysiology of HCMV congenital infection. They also pave the way to the identification of PPARγ gene targets in the infected brain.
| Congenital infection by human cytomegalovirus (HCMV) might result in permanent neurological sequelae, including sensorineural deafness, cerebral palsies or devastating neurodevelopmental abnormalities. Infants with such sequelae represent about 0.1% of all live births (>5500 per year in the USA). Given the considerable health and societal burden, a better insight on disease pathogenesis is urgently needed to design new therapeutic or prognostic tools. Here, we studied the impact of HCMV on neuronal development, using human neural progenitors (NSC) as a disease model. In particular, we investigated the outcome of infection on Peroxisome Proliferator-Activated Receptor gamma (PPARγ, a key protein in the regulation of metabolism, inflammation and cell differentiation. We disclosed that HCMV infection strongly increases levels and activity of PPARγ in NSCs. In vitro experiments showed that PPARγ activity inhibits the differentiation of NSCs into neurons. We also found increased PPARγ expression in brains of in utero infected fetuses, but not in controls, suggesting that PPARγ is a key effector of HCMV infection also in vivo. Our study provides new insights on the pathogenesis of HCMV infection and paves the way to the discovery of PPARγ-related molecules secreted in the infected brain.
| Congenital infection by human cytomegalovirus (HCMV) is a leading cause of permanent abnormalities of the central nervous system [1]. About 1% of newborns are congenitally infected with HCMV each year in the USA, as a result of either primary infection of a seronegative mother, or reinfection / viral reactivation in a seropositive mother during pregnancy. Ten percent of congenitally infected newborns are symptomatic at birth, and most of them (60–90%) display neurological sequelae [2]. Further, 10 to 15% of congenitally infected newborns that are asymptomatic at birth show neurological disorder with onset later in infancy [2]. The most severely affected fetuses or newborns show brain development abnormalities such as microcephaly, lissencephaly or polymicrogyria [2–4]. The most frequent permanent sequelae include mental and/or psychomotor disabilities, sensorineural hearing or vision loss, and/or spastic cerebral palsies. Overall, patients with permanent sequelae represent up to 0.1–0.2% of all live births (>5500 per year in the USA). The direct annual care costs for patients are estimated at $1-$2 billion in the USA [5]. No vaccine or reliable prognosis tools are available to date, except for ultrasound examination of macroscopic brain abnormalities. Considering the dramatic health and societal burden of congenital HCMV infection, it is clear that a better insight on its pathogenesis is urgently needed to provide new therapeutic and prognostic tools.
Human cytomegalovirus (HCMV) is a beta herpes virus that infects and replicates in a broad spectrum of organs and cell types. Infection of neural progenitor cells (NPCs) in the developing brain is thought to be a primary cause of the neurological sequelae due to HCMV congenital infection. Consistent with this hypothesis, studies using mouse brain slices or neurospheres reported that murine cytomegalovirus (MCMV) preferentially infected NPCs in the developing brain [6, 7]. Further studies by others and us showed that mouse or human NPCs obtained from neonatal autopsy tissues were permissive to HCMV infection in vivo or ex vivo [8–11]. These reports, however, revealed considerable diversity in the phenotype of NPCs following HCMV infection. Indeed, HCMV infection of neural progenitors was found to (i) inhibit self-renewal and proliferation, along with the induction of apoptosis [11], (ii) inhibit astrocyte differentiation [12], (iii) result in premature and abnormal differentiation [8], (iv) reduce the number of proliferating CD24-expressing NPCs [10]. Whatsoever, common to all studies was the observation that HCMV infection impaired the differentiation of NPC into neurons. Accordingly, two recent studies showed defective neuronal differentiation of neural stem cells generated from human induced pluripotent stem (iPS) cells upon in vitro HCMV infection [13, 14]. Despite these advances, the specific cellular and molecular mechanisms underlying the impaired neuronogenesis consecutive to HCMV infection still remain elusive. Given that a number of studies have established that peroxisome proliferator-activated receptor γ (PPARγ) is critical for proper brain development (reviewed in [15]), we reasoned that PPARγ may be involved in the impact of HCMV infection on neural progenitor cells. PPARγ is a ligand-dependent transcription factor, member of the nuclear receptor superfamily, which plays key roles in regulating cellular function and tissue homeostasis [16, 17]. Natural PPARγ ligands include 15-deoxy-∆12,14 prostaglandin (PG) J2 (15d-PGJ2), 15-hydroxyeicosatetraenoic acid (15-HETE), 9- or 13-hydroxyoctadecadienoic acid (9/13-HODE), all derived from oxidization cascades of poly-unsaturated fatty acids [16].
Here, we describe a new model of infection based on highly neuronogenic human neural stem cells (NSCs) derived from embryonic stem (ES) cells [18]. With this model, we examined the role of PPARγ in the neuropathophysiology of HCMV congenital infection. We further extended our observations to a collection of autopsy samples from HCMV-infected fetuses.
NSCs derived from embryonic stem cells showed self-renewal and continuous growth in defined conditions without the need of generating neurospheres. They expressed the multipotency marker SOX2 and the marker Nestin, and showed ability to differentiate into neurons positive for the markers HUC/D and βIII tubulin (Fig 1). An in-depth phenotypical characterization of NSC has been published elsewhere [19].
We first assessed the permissivity of NSC cultures to HCMV infection. Immunofluorescence analysis revealed that only few cells displayed a clear nuclear staining to HCMV Immediate Early antigen (IE) 24 h post infection (pi). At 48h pi, approximatively 5% of cells showed IE positive immunostaining (MOI 1 or 10) (Fig 2A). Thereafter, an increasing number of cells immunoreactive to IE were observed over time after HCMV infection, with up to 30% of IE-positive cells by 16 days pi (Fig 2B). Together, these results show that NSC cultures become progressively more permissive to HCMV infection overtime although the reason for this delayed kinetics is presently unknown. Our results are, however, consistent with previous reports showing that human neural progenitor cultures contain only 23% of IE-positive cells seven days after infection by the HCMV laboratory strain Towne when infected at a MOI of 1 [11]. All cells, including the IE-positive cells, remained immunoreactive to SOX2, suggesting that infection did not cause detectable changes in the stem cell status of NSCs (S1 Fig). As a control, no cell showed staining to IE when the inoculum had been previously UV-irradiated.
The 86-kDa form of IE, which is required for HCMV replication, was detected by western blot analysis as soon as 4 days pi (Fig 2B). Likewise, the early and late antigens UL44 and pp28 were immuno-detected from 8 days pi (Fig 2C). Electron microscopy revealed morphologically mature HCMV particles in the cytoplasm and pericellular space of infected NSCs, along with dense bodies and immature particles (Fig 2D). Titration of viral particles present in the medium of infected NSCs was performed using recipient MRC5 cells and fluorescence unit forming assay, which confirmed the presence of infectious HCMV particles (up to 4.104/ml infectious particles released the 8th day pi when the MOI was 10) (Fig 2E).
We next investigated whether infection altered the differentiation of NSCs into neurons. To initiate the differentiation of NSCs, it is critical to detach cells and to re-install them at a lower density on a fresh support, in the presence of increased concentration of laminin. In our preliminary experiments, we observed that the majority of infected cells were lost during these steps. As a result, differentiation of NSC cultures had to be initiated before any infection. Using this procedure, we observed that differentiating cultures of HCMV-infected NSCs displayed a dramatically decreased number of cells immunoreactive to βIII tubulin when compared to uninfected cultures (Fig 3A). Consistent with this observation, the overall level of βIII tubulin was strongly reduced in infected cultures (Fig 3B). To better appreciate the HCMV-triggered blockade of neuronal differentiation, we set up an automated procedure to screen populations of differentiating NSC grown in a 96-well plate format, based on the nuclear markers HUC/D (neurons) and SOX2 (NSCs). This analysis confirmed that the absolute and relative numbers of generated neurons decreased strongly and significantly as soon as 4 days pi in infected cultures (Fig 3C). At this stage, uninfected NSC cultures yielded 52% of neurons, whereas HCMV-infected NSCs generated 29% or 10% of neurons when infected, respectively, at MOI of 1 or 10 (p<0.008). At day 16 pi, uninfected NSC cultures yielded nearly 68% of neurons, while HCMV-infected NSCs generated 39% or 20% of neurons when infected, respectively, at MOI of 1 or 10 (p<0.01). Labeling for Ki67 antigen revealed no change in the proportion of dividing cells in the infected populations as compared to their uninfected counterparts, indicating that there was no concomitant increase in proliferative NSCs. Finally, we analyzed the cell death rate in infected or uninfected cultures. In control cultures, 44% of cells underwent developmental cell death within the two first days after onset of differentiation (Fig 3D). This event is classically observed during NSCs differentiation. Interestingly, HCMV infection appeared to limit this wave of developmental cell death, since the death rate was strongly and significantly decreased at day 2 pi in infected cell populations (22%; p<0.0032). This was confirmed by analyzing caspase 3 activation 48 h after infection, which revealed a significant decrease in the number of apoptotic cells among infected populations as compared to their uninfected counterparts (9% vs. 20%; p<0.0043, Mann-Whitney test)(Fig 3D).
Uninfected NSCs displayed expression of only minute amounts of PPARγ, as shown by immunofluorescence (Fig 4A, top row) and western blot (Fig 4B) analyses. In contrast, immunofluorescence, western blot and quantitative mRNA analyses revealed high levels of PPARγ mRNA and protein in HCMV-infected NSCs (Fig 4A–4C). Interestingly, cells with positive PPARγ staining were much more numerous than IE-positive cells. The PPARγ staining was nuclear, suggesting that the receptor was in its active form (Fig 4A). This finding prompted us to investigate whether infection enhanced PPARγ transactivating activity. Activated PPARγ binds to cognate DNA sequences termed PPAR responsive elements (PPRE). We thus performed luciferase assays using a PPRE-containing, PPARγ-responsive, luciferase reporter plasmid (pGL4-PPRE-luc), and the corresponding control plasmid (pGL4). Stimulation of uninfected NSCs with rosiglitazone resulted in a small and non-significant increase in PPRE-luc activity (Fig 4D). Infection by live HCMV increased pGL4-driven luciferase activity (ten fold) suggesting a generalized enhanced transcriptional activity in infected NSCs (Fig 4D). More importantly, however, PPAR-specific luciferase activity as assessed by transfection with the pGL4-PPRE-luc reporter plasmid was strongly and significantly increased in infected NSCs (> 43 fold, p<0.0019) (Fig 4D). Incubation of infected NSCs with the specific PPARγ antagonist T0070907 induced a significant decrease of luciferase activity from PPRE-luc (p<0.0026) (Fig 4D), but no change in luciferase activity from the control plasmid pGL4-luc, further demonstrating specific increased PPARγ activity in infected NSCs. We next performed chromatin immunoprecipitation (ChIP) assays to examine the ability of PPARγ, or of whatever dimer containing it, to bind physically to cognate genetic sequences in infected NSCs. We used as a probe a genomic segment located in the 5’ promoter region of the DLK1 gene, which binds PPARγ [20], and performed ChIP experiment with two different antibodies against PPARγ (Fig 4E). ChIP revealed a significant increase (> 2 fold, p<0.05) in the level of occupancy of the DLK1 gene segment by PPARγ in infected NSCs. Last, Oil red O staining showed that infection was associated with the accumulation of lipid droplets in the cytoplasm of host NSCs, indicative of enhanced lipid metabolism and thus of PPARγ activity [21] (Fig 4F).
Since a large majority of NSCs showed increased PPARγ levels in infected cultures, even though they did not show IE expression, we explored the possibility that infected NSCs could exert a positive effect on PPARγ expression in the surrounding cells. To investigate whether infected NSCs could release soluble mediators able to trigger PPARγ expression, we purified the supernatants from infected or uninfected NSC cultures, 5 days post infection. We next treated uninfected NSCs with these supernatants (after an ultracentifugation step to eliminate viral particles) and we analyzed PPARγ expression two days after this medium change. NSCs treated with supernatants prepared from uninfected cells did not show any increase in PPARγ levels (Fig 5A). In contrast, NSCs treated with supernatants prepared from infected NSC cultures displayed markedly increased PPARγ levels, similar to HCMV-infected NSCs. Importantly, almost all cells of the monolayer appeared to be sensitive to exposure to the supernatant prepared from infected NSCs. As expected, no IE-positive cells were detected in the cultures treated with supernatants from infected NSCs, indicating efficient removal of virus particles during the ultracentrifugation step. These results thus show that infected NSCs release soluble mediators that contribute to increase PPARγ levels in uninfected bystander cells.
Because known PPARγ activators are polyunsaturated fatty acids [16], we next investigated the hypothetical role of lipids released by the infected NSCs. The total lipid fractions from the supernatants of infected or uninfected NSC cultures were purified by chromatography using C18 columns, followed by desiccation of the extract and solubilization in DMSO. Importantly, such a procedure is incompatible with virion survival, thereby preventing any effect due to virus carry-over. Next we added these lipid extracts to the culture medium of NSCs, at a final concentration of 0.1% (v/v). NSCs exposed to lipids purified from the supernatants of infected cultures displayed a strong increase in PPARγ levels, in sharp contrast to NSCs exposed to lipids purified from control uninfected culture supernatants (Fig 5B).
Our findings thus show that infected NSCs release soluble mediators able to activate PPARγ expression in uninfected NSCs, similar to direct infection per se, and that lipid components contribute to this bystander effect.
We next investigated which ligands accounted for PPARγ activation during HCMV infection of NSC. Natural PPARγ ligands include 15-deoxy-∆12,14 prostaglandin (PG) J2 (15d-PGJ2), 15-hydroxyeicosatetraenoic acid (15-HETE), 9- or 13-hydroxyoctadecadienoic acid (9/13-HODE) [16]. The precursor of 15d-PGJ2, PGD2, as well as 15-HETE and 9/13-HODE are generated by oxidization of arachidonic acid (AA) or linoleic acid (LA) by cyclooxygenase (COX) or 5/15- lipoxygenase (LOX). Release of AA or LA from membrane glycerophospholipids is catalytically driven by calcium-dependent phospholipase A2 (cPLA2) activity [22]. Interestingly, during HCMV virion assembly, cellular cPLA2 is packaged into the viral particle and remains within the tegument of the virions, as an “onboarded” cell-derived cPLA2 which is required for infectivity [23]. It has been shown that this cell-derived cPLA2 can be inhibited by treatment of the viral inoculum by the specific cPLA2 inhibitor methyl arachidonyl fluorophosphonate (MAFP) before infection [23, 24]. Accordingly, we observed that treatment of the inoculum by 50 μM MAFP abolished IE expression in NSCs (Fig 6A). This suggested a role for this virion-packaged cPLA2 in the biosynthesis of possible PPARγ activators derived from polyunsaturated fatty acids (PUFA). To test this hypothesis, we measured levels of candidate PUFA-derived PPARγ agonists in control NSCs, HCMV-infected NSCs and, as a control, NSCs infected by MAFP-inactivated HCMV. We used a novel, rapid and sensitive method based on high performance liquid chromatography coupled to tandem mass spectrometry (LC-MS/MS) [25] using lysates and conditioned culture media collected at 24 h pi (Fig 6B). Candidate PPARγ agonists were 9/13-HODE, 15-HETE, and 15d-PGJ2. We also investigated the amounts of 5/8/12-HETE, although they have not been formally identified as PPARγ agonists to date. No significant changes were detected in amounts of 15d-PGJ2, 13-HODE or 5/8/12/15-HETE in infected NSCs or the corresponding culture supernatants (Fig 6B). In contrast, levels of 9-HODE were significantly increased in lysates from HCMV-infected NSCs (> 2.4 fold; p<0.029) (Fig 6B). In uninfected control cell lysates, 9-HODE amounts were detected at 1440.3 pg (4.9 pmol) per mg of total cellular protein, whereas they rose to 3539.8 pg (12 pmol) per mg of protein in infected NSCs. Treatment of HCMV particles with MAFP prior to infection abolished this increase in 9-HODE amounts (Fig 6B), indicating that active virion-packaged cPLA2 is needed for efficient 9-HODE biosynthesis. Importantly, only low amounts of 9-HODE were found in the conditioned culture supernatants (approximately 0.1 μg/ml, i.e., 0.3 nM) and no difference was observed in supernatants between infected and control NSCs. We assume that this was likely due to poor stability of 9-HODE in the serum-free medium and/or high cell permeability to 9-HODE resulting in poor abundance in the medium. In any event, our results suggest that HCMV triggers 9-HODE biosynthesis, at least at early stages of infection. Next, we investigated the outcomes of 9-HODE on PPARγ activity in NSCs, independently from the infectious context. We first carried out immunofluorescence analysis using NSCs stimulated during 24h by a range of 9-HODE concentrations. This analysis revealed a dose-dependent increase in levels of PPARγ staining in NSCs in response to 9-HODE exposure (Fig 6C). We also evidenced the nuclear translocation of PPARγ in cells stimulated with 9-HODE at concentrations greater than 0.5 μg/ml (Fig 6C). We further investigated the effect of 9-HODE on PPARγ activity by using the more sensitive luciferase reporter assay, which showed significantly increased PPARγ activity in NSCs stimulated by 9-HODE from 0.1 μg/ml (p< 0.0022) (Fig 6D). Altogether, our results indicate that 9-HODE efficiently activates PPARγ in NSCs, even outside of the infectious context.
Since PPARγ enhances IE1/2 gene transcription and HCMV replication [26], these results prompted us to investigate the impact of 9-HODE on HCMV replication in NSCs. To this aim, we infected NSCs at a MOI of 10 in the presence of 9-HODE in the medium. Immunofluorescence analysis revealed that treatment with 9-HODE resulted in significantly increased amounts of cells immunoreactive to an antibody specific to IE, at concentrations from 0.5 μg/ml (p< 0.0043, Mann-Whitney test) (Fig 7A and 7B). Lastly, we carried out an HCMV titration assay. Culture supernatants from infected NSCs stimulated or not by 9-HODE were harvested at 5, 6 and 7 days pi, and were added to the culture medium of MRC5 fibroblasts. The day after, immunofluorescence analysis was performed to assess the number of cells immunoreactive to IE. Titration assay showed that MRC5 cultures incubated with supernatants from NSCs infected in the presence of 9-HODE contained significantly greater number of IE-positive cells (Fig 7C) (p<0.01, Wilcoxon test). Together, these findings show that HCMV replication is more efficient when stimulated by 9-HODE, consistent with the fact that HCMV uses PPARγ for its replication [26].
We next investigated whether increased PPARγ activity could play a causative role in the defective neuronogenic differentiation of infected NSCs (Fig 8). First, we generated stably transduced NSCs that constitutively and strongly expressed either mouse PPARγ (NSC-Pg) or eGFP (NSC-GFP) (Fig 8A, top). Next, we carried out in vitro neuronogenesis assays to investigate the level of differentiation of such recombinant NSCs, stimulated or not by the PPARγ activator rosiglitazone. After seven days of differentiation, we observed that unstimulated NSC-GFP and NSC-Pg both yielded a relatively lower number of neurons (<20%) as compared to wild type NSC cultures, probably because of higher cell passaging, transduction and/or selection (Fig 8A, bottom). No significant difference was observed in the number of neurons generated in NSC-GFP cultures stimulated by the PPARγ activator rosiglitazone as compared to unstimulated NSC-GFP (Fig 8A, bottom). In contrast, a significantly decreased number of neurons was generated by NSC-Pg upon rosiglitazone stimulation (p<0.0006, Man-Whitney test) (Fig 8A, bottom). We next used wild type NSCs to investigate the impact of the activation of endogenous PPARγ in on neuronal differentiation in vitro. We used 9-HODE as the activator because, unlike rosiglitazone, it is able to increase both expression and activity of endogenous PPARγ in NSCs (Fig 6). We examined the number of HUC/D positive (HUC/D+) neurons generated from NSCs grown in the presence of 9-HODE at 0.1 μg/ml or 0.5 μg/ml, or in the presence of the vehicle (ethanol), after seven days of differentiation. 9-HODE stimulation at concentrations greater than 0.5 μg/ml resulted in strong cytotoxicity within this time period. In the absence of 9-HODE, NSC differentiation yielded 38% HUC/D positive neurons (Fig 8B). Significantly lower rates of differentiation were found in NSCs cultured in the presence of 9-HODE at a concentration of 0.1 μg/ml (31%; p<0.017) or 0.5 μg/ml (26%; p<0.002) (Fig 8B). Together, these findings establish that activated PPARγ is sufficient to impair neuronal differentiation of NSCs, even without infection. Lastly, we investigated whether pharmacological inhibition of PPARγ could improve neuronal differentiation in vitro from HCMV-infected NSCs. We carried out neuronogenesis assays with NSC infected by HCMV at a MOI of 1, in the presence of T0070907, a PPARγ-specific inhibitor. Because of the cytotoxicity of T0070907 on differentiating NSCs, we had to culture them on glass coverslips for no longer than 5 days, before examination of a randomly-selected set of optical fields (N = 12) and statistical analysis. When T0070907 was added to the culture medium 3 h pi, it strongly and significantly limited HCMV infection. Indeed, immunofluorescence analysis showed that T0070907-treated NSC populations contained two times less IE immunoreactive cells per field than untreated cultures, at 4 days pi (p<0.008, Mann-Whitney test) (Fig 8C, left). Indeed, untreated NSC cultures contained an average of 11% IE-positive cells (with a 5% confidence interval of 2.9%), whereas the T0070907-treated NSC populations contained an average of 5% of IE-positive cells (with a 5% confidence interval of 1.8%), consistent with our previous observations. This result, consistent with previous studies [26], also confirmed that PPARγ was efficiently inhibited by T0070907-treatment. As expected, infected NSCs showed defective differentiation and generated almost two times less HUC/D + neurons than uninfected cells (p<0.0014) (Fig 8C, right). Untreated NSC cultures contained from 20% to 40% of neurons per field among the total cell population. Also, uninfected NSCs differentiating in the presence of T0070907 showed no significant change in the abundance of neurons generated as compared to the untreated controls (Fig 8C). In contrast, infected NSCs differentiating in the presence of T0070907 yielded the same number of neurons as compared to the uninfected controls (Fig 8C). These results suggest that treatment by the PPARγ inhibitor T0070907 can reverse the effects of infection on neuronogenic differentiation in vitro, either directly through PPARγ inhibition, or indirectly through inhibition of viral replication resulting from PPARγ inhibition.
To assess the pathophysiological relevance of our experiments using NSCs, we next investigated the expression of PPARγ in fetal brain samples from aborted fetuses with congenital HCMV infection (N = 20) or from control subjects (N = 4). The clinical and pathological features are summarized in Table 1.
Gestational ages ranged from 23 weeks to 28 weeks, for cases and controls, so that all case samples could be compared with gestational age-matched controls. We first explored the level of infection in each sample by determining the total number of HCMV-positive cells in each slide, using an antibody specific to IE (Table 1, S2 Fig). No correlation could be established between the number of infectious foci and either gestational age, gender, or severity of the phenotype. Immunohistological analysis of PPARγ expression revealed PPARγ immunoreactive cells in the cell-dense, periventricular, brain germinative zone (BGZ) of all HCMV cases (Fig 9A–9E), but in none of the controls (Fig 8H and 8I). PPARγ staining was nuclear in the majority of cells (Fig 9A–9E), suggesting the presence of the active form of the receptor. We were also able to detect IE-positive cells surrounded by PPARγ positive cells (Fig 9F), supporting the hypothesis that viral replication enhances PPARγ expression both in host and neighboring cells. Isolated islets of PPARγ-positive cells were also detected in discrete lesional areas in the BGZ (Fig 9G). PPARγ-positive cells were also detected in the ependyma of HCMV subjects (Fig 9D and 9E).
To assess the relative abundance of cells expressing PPARγ, we counted, for each case, the number of PPARγ positive nuclei in a series (n = 6) of optical fields within the BGZ (Fig 10). The field size was approximatively 10 mm2. The number of nuclei in the fields ranged from 453 to 1592, with an average value of 971.3 and a 5% confidence interval of 48.7. Cases showed individual variability in the abundance of PPARγ expressing cells, ranging from 1.82% to 20.28%, with a mean value of 5.25%. No correlation with gestational age was apparent.
White matter was negative in all infected and control samples (Fig 11). PPARγ is physiologically expressed in vascular cells and is critical in vascular biology [27]. Accordingly, we observed that endothelial cells in brain vessels were positive to PPARγ in all infected and control samples (Fig 11). Together, these findings disclose that PPARγ expression is triggered specifically in the brain germinative areas of cases with congenital HCMV infection.
The main result of our study is the identification of PPARγ activation as a molecular determinant of the pathology induced by HCMV infection in neural precursors, in vitro and presumably in vivo. Our findings unambiguously demonstrate that HCMV infection causes increased PPARγ levels and activity, increased biosynthesis of 9-HODE, impaired neuronogenesis and enhanced viral replication in NSCs (Fig 12).
We here showed that HCMV infection of NSCs associates with increased levels of PPARγ mRNA and protein, although the changes in levels of transcripts appeared much greater than that of the protein. It is now well documented that protein levels are not systematically proportional to mRNA levels [28–30]. The differences in mRNA and protein levels likely arise from the fact that Q-PCR relates to the steady-state levels of transcripts, but does not take into account mRNA stability, protein stability, or translation efficiency.
Enhanced activity of PPARγ appears to be a common feature of HCMV infection, both in NSCs and cytotrophoblasts [26]. In contrast, PPARγ activity is decreased in mouse lung tissues infected by H1N1 influenza A virus [31], whereas it increases in macrophages infected by Mycobacterium tuberculosis or Mycobacterium bovis [32]. The observation that 9-HODE stimulation of uninfected NSCs is sufficient to increase PPARγ levels agrees with previous studies which revealed that 9-HODE selectively increases PPARγ gene expression in human U937 monocytic cells [33] or mesangial cells [34]. We have recently reported that 13-HODE and 15-HETE were the PPARγ agonists secreted by cytotrophoblasts and placenta explants infected by HCMV. In this case, however, no 9-HODE changes were detected [24]. This suggests tissue specificity in the response to HCMV infection with regard to fatty acids metabolism. Strikingly, 9-HODE, 13-HODE and 15-HETE all arise from oxidization of linoleic (9/13-HODE) or arachidonic (15-HETE) acids by lipoxygenase 15 [16]. This route of biosynthesis of PPARγ activating lipids is alternate to that previously described in human foreskin fibroblasts, where cyclooxygenase 2 activity catalyzes the biosynthesis of 15d-PGJ2 from arachidonic acid [35]. Some NSCs showed activated PPARγ and no IE detectable expression in infected monolayers (Fig 4A). We showed that infected cells exert a bystander effect on PPARγ expression in uninfected cells through soluble mediators. A significantly increased production of the highly membrane-permeant 9-HODE in infected cells could explain such a bystander effect, although we were unable to detect increased amounts of 9-HODE in the supernatants. One likely hypothesis is that 9-HODE could be present at detectable amounts only in live infected cells and that it would be released upon lysis of infected cells.
Abnormal PPARγ activity is likely to have multiple outcomes in infected cells and appears as an efficient strategy for HCMV to target simultaneously a number of important functions in the host NSC. PPARγ is required for IE gene expression and efficient HCMV replication in the host cell [26]. It exerts neuroprotective and anti-inflammatory effects, and regulates the oxidative pathway [36–39]. In particular, PPARγ is able to trans-repress the activity of NF-κB, AP-1 and STAT-1 as a response to their activation in the infectious context, resulting in negative modulation of production of the inflammatory mediators iNOS, TNFα, and IL-6 [37]. We report defective neuronal differentiation of NSCs infected by HCMV, in agreement with previous investigations [8, 10, 11, 13, 14]. Previous studies reported that PPARγ agonists either inhibited [40, 41], promoted [42, 43] or had no effect [44] on neuronal differentiation of uninfected rat or mouse neural progenitors. Our findings, most particularly our assays with singled-out expression of PPARγ in NSCs or using 9-HODE or T0070907, demonstrate unambiguously that PPARγ activity inhibits neuronogenesis in NSCs. To rule out the possibility of a receptor-independent effect of T0070907, it would be necessary to knockdown PPARγ expression in NSCs. Unfortunately, we failed to obtain any viable PPARγ knockdown using specific siRNA or shRNA vectors. Future experiments using CRISPR/CAS-9 system to invalidate PPARγ expression in NSCs may help elucidate the role of PPARγ in neurogenic differentiation.
In the present study, NSCs were generated through early neuroepithelial differentiation of human ES cells by using inhibitors of the TGFβ superfamily and a defined medium containing N2 and B27 supplements [19]. This method allows for efficient neural commitment and avoids possible confounding factors such as donor variability, batch-dependent components and feeder cell conditioned media. NSCs displayed a cortical phenotype and no immunoreactivity to non cortical markers [19], which is likely critical since HCMV infection targets cortical areas of the developing brain [7]. A recent study investigated HCMV replication in primitive prerosette NSCs (pNSCs), which represent a very primitive neural developmental stage [45]. That study revealed that viral replication depends on the differentiation status of the target cells. In that study, neither expression of HCMV early antigens or viral spreading could be evidenced from pNSCs or derived progenitors. We assume that the NSCs used in our work represent a later differentiation stage than the progenitors used in that previous study. pNSCs do not express the radial glia progenitor cell marker GFAP and can be differentiated to S100β+ NPCs following treatment with FGF2 [45]. This may explain why our NSCs readily supported viral replication, exhibited strong expression of early and late antigens, showed assembled viral particles, and allowed for efficient spreading. Further studies are required to decipher why the initial stages of infection were delayed, with only a minor part of the NSC population becoming IE-positive during the two first days. The reasons for the delayed initial kinetics of infection are unclear, but may be related to either changes in the surrounding cells resulting in increased permissivity, or by possible changes in the virus cycle. Notably, it is presently unclear whether the number of morphologically mature virions detected in infected NSCs by transmission electron microscopy fully correlates with infectivity (Fig 2).
The involvement of PPARγ in the pathogenesis of congenital HCMV infection is strongly supported by its pattern of expression in HCMV-infected human brain samples. Indeed, all samples from HCMV subjects, and none from control cases, showed PPARγ immunostaining in the brain germinative zone (BGZ). Moreover, BGZ appeared to be the only brain area with detectable abnormal PPARγ expression in infected subjects, unlike blood vessels or white matter. So far, only the presence of inclusion bodies [2], but not that of any specific protein, has been reported in brain sections from cases congenitally infected by HCMV. Our findings are consistent with studies in the mouse which identified the germinative subventricular zone as the most sensitive site to infection by murine cytomegalovirus (MCMV) [7, 46]. The critical role of PPARγ in neurodevelopmental regulation requires fine spatiotemporal tuning of expression and activity [47]. Therefore, we assume that asynchronously increased PPARγ activity could be deleterious to neurogenesis during HCMV congenital infection. Notably, we observed that the number of PPARγ expressing cells in brains slices from an infected case was always similar or greater than the number of IE positive cells. This is consistent with the possibility of a bystander effect from infected cells to uninfected cells during congenital infection, such as that observed in vitro with NSCs.
In conclusion, NSCs turned out to be an invaluable tool for modeling functional correlates of HCMV infection, and this cell platform may probably be extended to other viral pathologies of the central nervous system. Our findings shed a new light on the pathophysiological bases of the neurological outcomes of congenital HCMV infection and on the role PPARγ in neural stem cell and developing brain.
Neural stem cells from human embryonic stem cells were used in the frame of a project approved by the French authorities (Agence de la Biomedecine, authorization number SASB0920178S). Collection of brain histological samples was performed in the frame of a project coordinated by Necker Hospital, AP-HP (Assistance Publique-Hôpitaux de Paris). The study was performed in accordance with the French ethical guidelines and was approved by the French authorities (Agence de la Biomedecine, authorization number PFS-15009). Written informed consent was obtained from all study participants prior to sample collection. All samples were anonymized before processing.
NSCs were generated and grown in vitro as detailed elsewhere [18]. Briefly, neural tube-like structures containing neuro-epithelial cells were obtained in vitro from human embryonic stem (ES) cells by using two inhibitors of the TGFβ superfamily (SB431542 and Noggin). Manually isolated colonies were expanded in the presence of EGF, FGF2 and BDNF to obtain NSCs. NSCs were seeded at 100 000 cells/cm2 and maintained in growth medium (basal medium) consisting of DMEM/F12/Neurobasal medium (Life Technologies, Grand Island, NY, USA) mixed at a ratio of 1/1/2 (v/v/v) in the presence of N2 and B27 supplements (Life Technologies), 10 ng/ml FGF2, 10 ng/ml EGF, 20 ng/ml BDNF (all from Peprotech, Rocky Hill, NJ, USA). Culture supports were coated first by PBS containing 0.05% poly-ornithine (PO) (Sigma, Saint-Louis, MO, USA) then by PBS containing mouse laminin (1 μg/cm2)(Roche, Basel, Switzerland). NSC cultures were checked for the absence of mycoplasma (Plasmotest, Invivogen, Toulouse, France). Neuronogenic differentiation was induced by removal of FGF2 and EGF and addition of laminin (2 μg/cm2) into the medium (differentiation medium) of low passage (<12) NSCs seeded at 50 000 cells/cm2. Culture medium was renewed every two days.
The human immortalized fibroblast line MRC5 (ATCC CCL171, Manassas, VA, USA) was cultured in DMEM containing 10% bovine calf serum (Life technologies).
We used the clinical VHL/E HCMV strain (a gift from C. Sinzger, Tubingen, Germany), at low passage (<8) of amplification in MRC5 cells, and laboratory-adapted AD169 HCMV strain (ATCC VR538). Virus stocks were collected from infected MRC5 fibroblasts when cytopathic effects were >90%. Supernatants were clarified of cell debris by centrifugation at 1,500 × g for 10 min, ultracentrifuged at 100,000 × g for 30 min at 4°C, harvested in NSC basal medium, and stored at −70°C until use. Virus titers were determined upon infection of MRC5 cells by serial dilutions of the inoculum followed by immunofluorescence analysis to count the number of nuclei immunoreactive to HCMV Immediate Early antigen (IE) 24 h post infection (pi)(fluorescence unit forming assay). UV irradiation of HCMV particles was performed for 30 min in a closed propylene tube with a Spectrolyne irradiator (EF-140/F) fitted with a BLE-2537S bulb [254 nm] (Spectronics corporation, Westbury, NY). In these conditions, the theoretical radiant energy density is 36 J/cm2. After such a treatment, irradiated HCMV virions were still able to infect cells 30 min pi, as checked by immunostaining with an antibody specific to the tegument protein pp65 (Virusys corporation, Taneytown, MD), but the viral genome could not be expressed 24 h pi, as checked by the absence of immunoreactivity to the HCMV Immediate Early antigen (IE).
Antibodies specific to PPARγ were H-100 (Santa Cruz Biotechnology, Dallas, TX, USA) and A3409A (Abcam, Cambridge, UK). We used primary antibodies specific to HCMV IE (Argene, Verniolle, France), PPARα (H-74, Santa Cruz Biotechnology), PPARβ (H-98, Santa Cruz Biotechnology), RXRα, β and γ (DN197, Santa Cruz Biotechnology), SOX2 (D6D12, Cell Signaling Technology, Beverly, MA, USA), HUC/D (16A11, Life Technologies), Nestin (10c2, Millipore, Billerica, MA, USA), Class III beta-tubulin (βIII tubulin) (TU-20 or ab18207, both from Abcam), cleaved caspase-3 (Cell Signaling Technology), Ki67 (KiS5, Millipore). Secondary antibodies against rabbit or mouse immunoglobulins were conjugated with Alexa-488 or -555 or -647 fluorophores (Life Technologies). No staining was detectable when cells were incubated with the secondary antibodies alone, or with primary and secondary antibodies for which species of origin did not match.
PPARβ synthetic activator was rosiglitazone (1 μM) (Sigma) and a stimulation time of 2 h was used. PPARβ specific inhibitor was T0070907 (10 nM) (Sigma). Optimal concentrations of rosiglitazone and T0070907 were determined from initial dose-effect experiments using a luciferase PPAR reporter plasmid (as detailed below). Control experiments with rosiglitazone were performed with the vehicle, DMSO. Synthetic 9-HODE was purchased (Cayman, Ann Harbor, MI). Control experiments with synthetic 9-HODE were performed with its vehicle, ethanol.
Cell-derived, virion-packaged cPLA2 was inactivated as described elsewhere [24]. Viral suspensions were incubated in a volume of 1 ml for 1 h at room temperature in the presence of 50 μM methyl arachidonyl fluorophosphonate (MAFP; Sigma), ultracentrifuged at 100,000 × g for 30 min at 4°C, washed twice with phosphate-buffered saline (PBS; Life Technologies) and diluted in culture medium. The working concentration of MAFP (50 μM) was determined by immunofluorescence analysis using NSCs infected by HCMV particles treated by various doses of MAFP, using an antibody specific to IE. Cell viability was checked by 4′6-diamino-2-phenylindole (DAPI) staining. Control viral suspensions were processed identically after incubation in the presence of the vehicle (DMSO) instead of MAFP. When NSCs were infected with MAFP-treated virus, control uninfected NSCs were cultured in the presence of MAFP at a concentration equivalent to that which would have been obtained without the HCMV particles washes (50 nM).
NSC cultures infected by HCMV at a MOI of 10 were fixed in 2% glutaraldehyde in 0.1 M Sorensen phosphate buffer (pH 7.4) for 4 h at 4°C, 6 days post infection. After an overnight wash in 0.2 M phosphate buffer, the cells were post-fixed for 1 h at room temperature with 1% osmium tetroxide in 250 mM saccharose and 0.05 M phosphate buffer, and stained overnight in 2% uranyle acetate. The samples were then dehydrated in a series of graded ethanol solutions and embedded in an Epon-Araldite resin (Embed 812-Araldite 502, Electron Microscopy Sciences, Hatfield, PA). Finally, the cells were sliced into 70-nm thick sections and mounted on 200-mesh collodion-coated copper grids prior to staining with 3% uranyle acetate in 50% ethanol and Reynold’s lead citrate. Examinations were carried out on a transmission Hitachi HU12A electron microscope at an accelerating voltage of 75 kV.
NSC cultures were seeded in 10 cm2 dishes at a density of 100, 000 cells/cm2. NSCs were infected (MOI 10) or stimulated 16 h after plating, lysed in RIPA buffer containing 50 mM Tris-HCL, pH 7.6; 150 mM NaCl; 0.1% sodium deoxycholate, 0.1% sodium dodecylsulfate, 0.1% NP-40, 1 mM EDTA, and a protease inhibitor cocktail (all from Sigma). Lysates were subjected to SDS-PAGE with 4 to 12% Tris-Tricine gels (Life Technologies). Proteins were blotted onto nitrocellulose membranes (GE-Healthcare, Pittsburgh, PA, USA) using a semi-dry transfer device (Biorad, Hercules, CA, USA). Western blot was performed using Tris-Buffered-Saline (TBS) containing 0.1% Tween-20 as the wash buffer, TBS containing 5% non-fat dry milk and 3% BSA as the blocking buffer, and primary or horseradish peroxidase-conjugated secondary antibodies diluted in blocking buffer. Detection was carried out by using a chemoluminescence kit (Sigma). Analysis was performed with a Chemidoc system (Bio-Rad, Hercules, CA) using conditions when signals were not saturating.
Cells were cultured on coverslips coated with PO and laminin, fixed in 100% methanol for 15 min at -20°C (PPAR staining), or in 4% formaldehyde for 20 min at 4°C, and permeabilized in 0.3% Triton X-100 for 15 min at room temperature (other stainings). Blocking buffer was PBS containing 5% fetal calf serum. Primary antibodies were diluted in blocking buffer and applied overnight at 4°C. Secondary antibodies were diluted in blocking buffer and applied for 1 h at room temperature. The cells were washed three times with blocking buffer, then washed three times with PBS, then counterstained with 1 μg/ml DAPI (Sigma), washed again three times with PBS and visualized on an DM4000B inverted fluorescence microscope (Leica, Solms, Germany). Image processing was performed using ImageJ software [48].
For neuronogenesis assays, NSCs were seeded in 0.35 cm2 coated culture wells at a density of 15,000 cells per well in differentiation medium and were infected 24 h later. Next, immunofluorescence was carried out at different time points pi, using double staining with the antibodies specific to HUC/D and SOX2, and counterstaining with DAPI. The culture plates were analyzed with an automated microscopy device (Cellomics) to count the number of nuclei positive for SOX2 or HUC/D. DAPI staining was used as the primary mask. Cell death assays were performed similarly, using the reagent Image-it dead (Life Technologies).
For neuronogenesis assay in the presence of T0070907, NSCs were installed in differentiation medium on 15 mm2 coated glass coverslips in 2.3 cm2 wells (150,000 cells/well). The day after, NSCs were infected or not by HCMV at a MOI of 1. T0070907 (10 nM) was added to the medium 3 h after infection. The medium was renewed everyday. At day 4 pi, immunofluorescence analysis was performed as described above using antibodies specific to HUC/D, SOX2, or IE, and DAPI as a counterstain. Twelve optical fields of each coverslip were visualized on a DM4000B inverted fluorescence microscope, and analyzed using ImageJ. The number of cells immunoreactive to SOX2, HUC/D or IE antibodies was counted manually to exclude dead cells and to resolve cell clusters. Three independent experiments were performed.
We used a firefly luciferase (Luc) reporter plasmid based on a pGL4 backbone (Promega, Madison, WI, USA) containing three PPAR responsive elements (PPREs) [26] upstream of the herpes simplex thymidine kinase promoter (pGL4-PPRE-luc). For normalization, we used a promoter-less renilla luciferase normalization plasmid (pRL-null, Promega). NSCs were seeded in 96-well plates at a density of 25,000 cells per well. Transfection of both the reporter and normalization plasmids was performed 16 h after seeding using Genejuice transfection reagent (Millipore), according to the manufacturer's instructions. Cells were infected (MOI 10) or treated with rosiglitazone (1 μM) or T0070907 (10 nM) during the 24 h following transfection. Last, cell lysis was performed using Cell Culture Lysis Reagent (Promega). Luciferase activity was quantified using a Centro luminometer (Berthold). All assays were done in triplicate and the experiment was repeated twice.
LC-MS/MS was performed as detailed elsewhere [26], using HPLC grade methanol, methyl formate, and acetonitrile (Sigma–Aldrich). Deuterium-labeled lipoxin A4 (LxA4-d5), leukotriene B4 (LTB4-d4) and 5- hydroxyeicosatetraenoic acid (5-HETE-d8) (Cayman Chemicals) were mixed at a concentration of 400 ng/ml in methanol and used as the internal standard (IS) solution. In all experiments, NSCs from 10 cm2 culture wells were harvested 6 h pi in 0.2 ml of PBS, transferred to lysing matrix A (MP Biomedicals) and supplemented with 5 μl of IS solution.
Cells were lysed using a spin homogeneizer (Fastprep, MP Biomedicals) with 2 cycles of 20 sec at 5,000 rpm. 10 μl of the lysed cell suspension were added to 200 μl of 0.1 M NaOH for subsequent protein quantification using a Bradford assay (BioRad). The remaining of the lysate was supplemented with 200 μl methanol, vigorously shaken, and centrifuged for 15 min at 1,000 x g at 4°C. The supernatants were collected and stored at -80°C until lipid extraction. Lipid amounts from cell lysates were expressed in pg per mg of protein in the lysate. Culture supernatants were collected 6 h pi, supplemented with 300 μl of ice-cold methanol and 5 μl of IS solution, clarified by a centrifugation at 1,000 x g for 15 min, and stored at -80°C until lipid extraction. Lipid amounts from supernatants were expressed as pg/ml.
Lipid preparation from all samples was carried out through solid-phase extraction using hydrophobic polystyrene-divinylbenzene resin in dedicated 96-well plates (Chromabond multi96 HR-X 50 mg; Macherey-Nagel). After conditioning of the plate with methanol and sample loading, the plates were washed twice with H2O/MeOH (90/10, v/v) and dried under aspiration for 15 min. Lipids were eluted with methanol (2 ml), dried under nitrogen, dissolved again in methanol (10 μl) and transferred to liquid chromatography tubes before LC–MS/MS analysis.
LC-MS/MS analysis was performed using an UHPLC system (LC1290 Infinity, Agilent) coupled to a 6460 triple quadrupole mass spectrophotometer (Agilent Technologies) fitted with an electro-spray ionization interface. Separation was done at 40°C on a Zorbax SB-C18 column (2.1 mm–50 mm–1.8 μm) (Agilent Technologies). The compositions of mobile phases A and B were water, acetonitrile (ACN) and formic acid (FA) (75/25/0.1) and ACN, FA (100/0.1), respectively. Compounds were separated with a linear gradient from 0 to 85% B in 8.5 min and then to 100% B at 9 min. Isocratic elution continued for 1 min at 100% B, then 100% A was reached at 10.2 min and maintained to 11 min. The flow rate was 0.35 ml/min. The autosampler was set at 5°C and the injection volume was 5 μl. Source conditions were as follows: negative ESI mode, source temperature = 325°C, nebulizer gas (nitrogen) flow rate = 10 l/min, sheath gas (nitrogen) flow rate = 12 l/min, sheath gas temperature = 400°C and spray voltage = −3500 V. Data were acquired in MRM mode. For each compound, the best conditions of separation and quantification were defined: retention time in minutes (RT), specific Q1/Q3 transition (T) fragmentor (F) and collision energy (CE). Peak detection, integration and quantitative analyses were performed using Mass Hunter Quantitative analysis software (Agilent Technologies). At least three independent experiments were performed, each in triplicate wells.
We used pWPXL-GFP (Addgene #12257), a lentiviral vector backbone allowing for stable expression of enhanced Green Fluorescent Protein (eGFP) driven by the human EF1α gene promoter. A MluI- XbaI fragment containing the wild type mouse Pparg2 cDNA (1626 bp) was excised from a modified pSV Sport PPARγ2 plasmid (Addgene #8862) [49] and substituted to the eGFP cDNA into pWPXL-GFP restricted by MluI and SpeI (plasmid pWPXL-Pg). A puromycin resistance cassette containing the gene Pac under the control of the human ubiquitin promoter was excised by AscI digestion of the plasmid pSF-CMV-Ub-Puro-SV40 Ori SbfI (Oxford genetics), blunted, and inserted into the blunted KpnI site of pWPXL-Pg and pWPXL-GFP, generating the plasmids plenti-Pg and plenti-GFP.
Lentiviral vectors were generated by transfection of the plasmids pMD2G (Addgene #12259), pCMVR8.74 (Addgene #22036), and plenti-GFP or plenti-Pg into HEK293 cells using calcium phosphate, as recommended by the supplier (Clontech). Lentiviral particles were collected at 24 h and 48 h post transfection, ultracentrifuged at 50,000 x g for 120 min at 16°C, resuspended in NSC basal medium, and stored at −70°C until use. Recombinant NSCs were generated by transducing cultures by the lentiviral vectors, followed by continuous selection by 1 μg/ml puromycin. Ectopic expression of eGFP or PPARγ was checked by immunofluorescence, western blot and oil red O staining.
Uninfected or infected (MOI 10) cells were incubated for 2 h in growth medium in the presence of oleic acid conjugated with BSA (10 μg/ml), fixed and permeabilized in methanol for 2 min at −20°C and incubated for 10 min with 0.3% Oil Red O diluted in 60% isopropanol. After 30 s of incubation, cells were washed in water and nuclei were counterstained with hematoxylin.
RNA was extracted by using dedicated columns (Qiagen), and 1 μg was reverse transcribed (RT) with Superscript III (Invitrogen), according to the supplier’s recommendations. All quantitative RT-PCR (Q-PCR) assays were based on a SyBr-green based PCR mixture (Roche) using a LC480 system (Roche). All primers pairs were designed using the Primer3 software (http://frodo.wi.mit.edu/) [50] and characterized by real-time amplification of a series of cDNA dilutions to determine linearity range and primer efficiency. Primer sequences are available upon request. All Q-PCR amplifications were done in triplicate and the experiments were performed at least twice. Q-PCRs were carried out according to the MIQE guidelines [51]. Reference gene was GAPDH, as identified by Genorm analysis.
Chromatin immunoprecipitation (ChIP) was carried out using the Magnetic ChIP kit (Pierce) following the supplier’s recommendations. NSCs were seeded at a density of 100,000 cells / mm2 onto 60-mm plates containing one 12-mm glass coverslip, infected after 16 h at a MOI of 10 (AD169 strain), and fixed 48 h post infection. Infection was controlled by immunofluorescence analysis of the cells on the coverslip, using an antibody specific to IE. Sonication of chromatin was performed using a Vibracell device (Bioblock Scientific) and checked by agarose gel electrophoresis. Chromatin was immunoprecipitated with 10 μg of specific antibody or 10 μg of unspecific mouse immunoglobulins. One tenth of the immunoprecipitated DNA samples and 5 ng of input DNA samples were subjected to Q-PCR for normalization. Primers specific to DLK1 were described elsewhere [20].
Brain tissue biopsies were collected from 20 human fetuses aborted electively because of HCMV congenital infection and from 4 controls aborted for non-infectious diseases. Immuno-histopathological brain analysis of control and HCMV subjects was performed on 8 μm sections from paraffin blocks using standard methods, IE antibody, E8 PPARγ antibody, and Mayer’s hematoxilin counterstain, with a Dako Autostainer automated device (Dako, Glostrup, denmark). Slides were scanned with a Panoramic 250 system (3D Histech, Budapest, Hungary) and analyzed with the Panoramic viewer software (3D Histech). For each patient, 6 optical fields within the brain germinative zone were analyzed. The total number of nuclei in each field was determined using the Fast Morphology plug-in of ImageJ software, with a threshold size of 50 square pixels. When required, cell clusters were resolved manually. The number of nuclei with positive PPARγ staining in each field was determined manually to exclude endothelial cells when present.
Statistical analyses were performed with the StatEL plugin (Adscience) for Excel (Microsoft, Redmond, WA) or GraphPad Prism (GraphPad Software, San Diego, CA), using Kruskall-Wallis test unless indicated. Error bars show 5% confidence intervals (CI).
PPARγ gene: PPARG, Ensembl ID: ENSG00000132170
Nestin gene: NES, Ensembl ID: ENSG00000132688
SOX2 gene: SOX2, Ensembl ID: ENSG00000181449
HUC/D gene: Ensembl ID: ELAVL4, ENSG00000162374
β3-Tubulin gene: TUBB3, Ensembl ID: ENSG00000258947
DLK1 gene: DLK1, Ensembl ID: ENSG00000185559
|
10.1371/journal.pgen.0030034 | A Caenorhabditis elegans Wild Type Defies the Temperature–Size Rule Owing to a Single Nucleotide Polymorphism in tra-3 | Ectotherms rely for their body heat on surrounding temperatures. A key question in biology is why most ectotherms mature at a larger size at lower temperatures, a phenomenon known as the temperature–size rule. Since temperature affects virtually all processes in a living organism, current theories to explain this phenomenon are diverse and complex and assert often from opposing assumptions. Although widely studied, the molecular genetic control of the temperature–size rule is unknown. We found that the Caenorhabditis elegans wild-type N2 complied with the temperature–size rule, whereas wild-type CB4856 defied it. Using a candidate gene approach based on an N2 × CB4856 recombinant inbred panel in combination with mutant analysis, complementation, and transgenic studies, we show that a single nucleotide polymorphism in tra-3 leads to mutation F96L in the encoded calpain-like protease. This mutation attenuates the ability of CB4856 to grow larger at low temperature. Homology modelling predicts that F96L reduces TRA-3 activity by destabilizing the DII-A domain. The data show that size adaptation of ectotherms to temperature changes may be less complex than previously thought because a subtle wild-type polymorphism modulates the temperature responsiveness of body size. These findings provide a novel step toward the molecular understanding of the temperature–size rule, which has puzzled biologists for decades.
| Biologists are fascinated by variation in body size, which is hardly surprising, considering that the range of body sizes spans orders of magnitude from bacteria to blue whales. Even within species, body sizes can vary dramatically. This intraspecies variation is intriguing because it suggests strong associations between body size and environment. Already in 1847, Bergmann noticed that mammals tend to be larger in colder environments. More recently similar relationships were found for ectotherms, which rely for their body heat on the temperature of their surroundings, where more than 85% of the species studied grew larger at lower temperatures. This phenomenon, dubbed the temperature–size rule, has caused a renewed interest to understand how temperature affects body size. Yet the control of the temperature–size rule remains enigmatic, and the hypotheses proposed have been inconclusive. In this paper the authors show that a single nucleic acid change in one gene is required for regulation of the temperature–size rule in the nematode C. elegans. Using protein modelling they also show that this subtle change in DNA decreases the function of the encoded protein. The data suggest that temperature adaptation can be simple and far less complex than previously thought.
| For many decades biologists have been intrigued by the relation between body size and temperature. It was discovered that ectotherms—animals that maintain their body temperature by absorbing heat from the surrounding environment such as fish and all invertebrates—reproduce later at a larger size when reared at lower temperatures [1–3]. This phenomenon is known as the temperature–size rule, and nearly 90% of ectothermic species studied so far follow this rule [4]. The magnitude of this phenomenon is illustrated by Azevedo et al. [5] who found a 12% increase in wing and thorax size in Drosophila melanogaster when grown at relatively low temperatures. In the case of the nematode C. elegans (strain Bristol N2), an environmental temperature of 10 °C resulted in adults that were ~33% larger than those grown at 25 °C [6].
About 99.9% of all species are ectothermic, and the temperature–size rule is observed in bacteria, protists, plants, and animals, making it one of the most widespread phenomena in ecology. From the perspective of life-history evolution it is not well understood why growing bigger at lower temperatures is beneficial for organisms. Because this thermal plasticity of body size is taxonomically widespread, the reasons are probably diverse and may vary among groups of organisms. It has been suggested that a large body size is advantageous, because it compensates for delayed reproduction by yielding more offspring [7]. Other explanations may be that a larger body size at maturity enables individuals to produce larger offspring or to provide better parental care [2].
Since body size and temperature are the two most important variables affecting fitness [8,9], many experimental and theoretical attempts have been made to explain the mechanism underlying the temperature–size rule. Essentially, an increase in body size can be achieved by increasing cell number, cell size, or by both. Various studies point at the second (cell size) and the third option (cell size and number) as being the most likely explanation for the observed increase in body size at lower temperatures (Drosophila spp. [10–12], yellow dung fly [13], and the nematode C. elegans [6]).
Next to these empirical observations, various models have been proposed that are based on a combination of changes in cell size and number. Biophysical models show that the temperature–size rule is the result of unequal effects of temperature on cell growth and cell division [14]. When the effect of temperature on the rate of division is greater than its effect on the rate of cell growth, the model predicts that a low temperature should lead to a larger body size. Recently, a physiological model was proposed by Atkinson et al. [15], which assumes that temperature induced changes in cell size and number depend on the optimisation of oxygen supply at different temperatures. Yet, these empirical and theoretical findings give little insight into the molecular genetic control of body size at lower temperatures.
Unravelling the molecular mechanism underlying the temperature–size rule is hampered by the fact that temperature affects nearly all biochemical processes in a cell, and in theory growing bigger at lower temperatures may have numerous causes. However, low temperatures also have been shown to induce a number of specific physiological and genetic responses in ectotherms [16]. In D. melanogaster gene expression analysis revealed a senescence marker smp-30 to be induced by low temperature [17]. Van ‘t Land et al. [18] reported the association of the gene Hsr-omega with low temperatures in D. melanogaster. Next to these specific gene responses, an early indicator of low temperature is a transient elevation of the cytosolic calcium concentration [Ca2+]i. Higher cytosolic calcium levels occur not only in response to a rapid cooling but also to more gradual reductions in temperature, and it is a widespread phenomenon observed in plants [19,20] and ectothermic animals [21–24].
Here we aimed to identify and characterize genes underlying the temperature–size rule in a model ectotherm, the nematode C. elegans. C. elegans is a suitable model for studying the molecular control of temperature–body size responses because of its completely sequenced genome, isomorphic growth, and cell constancy, and because nematode life-history traits are easy to observe [25]. We found that wild-type Bristol N2 (designated as N2) grew bigger at lower temperatures and thus complied with the temperature–size rule, whereas wild-type CB4856 (designated as CB) defied the rule. The natural variation in body size response to temperature between CB and N2 was caused by a single mutation F96L in a calpain-like protease TRA-3 encoded by tra-3. Homology modelling predicts that F96L is likely to reduce the ability of TRA-3 to bind calcium.
We studied the thermal reaction norm for body size (TRB), which is a plot of body size at maturity versus temperature, and defined compliance with the temperature–size rule if body size is significantly and negatively related to temperature. To assess differences in the TRB between the two wild-type strains we measured body size at 12 °C and 24 °C. Body-size measurements were taken from Gutteling et al. [26]. We found a marked difference in TRB between the two wild types. The body size of wild-type N2 exhibited a significant negative relationship with temperature, i.e., N2 grew larger at low temperature (F = 3.49; p = 0.02). In contrast, CB defied the temperature–size rule because body size was not significantly affected by temperature (F = 0.8; p = 0.47) (Figure 1). The results for N2 are in agreement with previous findings where increased body size was found in C. elegans N2 hermaphrodites as well as males at lower temperatures [27,6]. To further study the genetic control of the TRB, we first developed an N2 × CB recombinant inbred panel and performed a quantitative trait locus (QTL) analysis for detecting genomic regions associated with the TRB.
By selfing the CB × N2 F1 offspring for 20 generations, we obtained a segregating population of recombinant inbred lines (RILs), which were also exposed to 12 °C and 24 °C. We found large differences in TRB slopes among the RILs (Figure 1). As generally observed in recombinant inbred crosses between divergent strains, the mean trait values for many of the RILs exceeded the mean value for either parental strain. Apparently the differences between the N2 and CB phenotypes (the slope of the TRB) capture a great deal of genetic variation. This was evident in the variation exhibited in the RILs for the TRB slope. Such transgressive segregation has been reported for many organisms and indicates that alleles at different loci act in the same direction, and when combined these alleles will result in phenotypes more extreme than either parent [28].
In general RILs matured at 12 °C at a bigger size than at 24 °C, which is in accordance with the temperature–size rule (see Atkinson [4] for an overview). We found strong genetic variation among RILs for body size across the two temperatures (F = 40.1; p < 0.001). We then sought to determine which loci were associated with the TRB by genotyping the RILs and performing a QTL mapping study using the recombinant inbred panel.
For the QTL analysis we used a dense single nucleotide polymorphism (SNP) map. A full description of the genetic architecture of the RILs can be found in [29]. In summary, the overall average distance between two SNP markers was 835 kb or 2.38 cM. The overall average chromosomal coverage was 96% if measured in bp or 95% if measured in cM. Compared to the Wormbase F2-derived genetic maps (http://www.wormbase.org, release WS106), the genetic maps showed on average an ample 2-fold expansion. This is common for RILs bred by self-fertilization or sib-mating and can be explained by the multiple rounds of meiosis undergone [30].
Figure 2 shows the detected QTLs associated with the slope of the TRB. Two QTLs on Chromosome IV were associated with a negative effect on the slope of the TRB and were linked to CB alleles. The distal QTL at Chromosome IV showed pleiotropy or linkage for body size at 24 °C (additive effect of 4%). We aimed to identify the gene(s) controlling the QTL at Chromosome IV with a peak at marker pkP4095 at 12 cM, because this QTL was uniquely associated with TRB (hence we named it the TRB-locus) and not with body size itself at 12 °C or 24 °C. This locus had a relatively large additive effect of 34% of the total standard deviation and explained 11% of the among-RIL variance. Introgression of a CB segment spanning the TRB-locus into an N2 background confirmed the QTL analysis. Phenotyping of NIL WN17–9 carrying an ~6-cM region of the TRB locus revealed no significant body-size difference between low and high temperature (Figure 3). Three other QTLs on Chromosome III increased the slope and each of these QTLs was linked to N2 alleles and showed a pleiotropic or close linkage effect for body size at 12 °C [26].
The 2.5-cM genome segment covered by the confidence interval (CI) of the TRB locus harbours a number of mutationally mapped genes of which only one (dpy-4) [31] is known to affect body size. To identify promising candidate genes, we reasoned as follows. Previous studies have shown that body size in C. elegans is controlled by genes that affect cell size and not cell number [32,33]. It is also known that this is one of the main mechanisms, next to cell number, by which ectotherms grow bigger in colder environments [7]. Furthermore, we sought to identify and characterize genes that encode a calcium-activated protein because [Ca2+]i is a key signal of low temperature. Lower temperatures lead to an increase of [Ca2+]i [21–24].
Given these two facts (increased [Ca2+]i and cell size) we searched for genes that are activated by [Ca2+]i and that play a role in increased cell size. Among the few genes with known function in the TRB locus, the most likely candidate gene was tra-3. TRA-3 has a high homology with mammalian calpains [34], which are known to be activated by [Ca2+]i and have been reported to regulate cell size during oncosis (cell swelling) [35]. dpy-4 is not known to be activated by [Ca2+]i [31]. We therefore selected tra-3 as a candidate gene that might explain the difference in temperature responsiveness between N2 and CB. The gene tra-3 seems to be important for the TRB slope because a significant difference was found between tra-3 allelic variants (using the linked marker pkP4095) and the TRB slope (t-test, p = 0.03). RILs with the N2 allele had a larger slope than RILs with a CB allele.
To investigate the hypothesis that tra-3 controlled the TRB, we first sequenced this region in CB. One SNP was found within the coding region where phenylanaline-96 in N2 was mutated into leucine-96 in CB. To see whether other tra-3 mutants displayed the same phenotype as observed in CB, we selected two homozygous artificial allelic mutants in an N2 background, tra-3(e1107) carrying a nonsense mutation [34] and tra-3(e2333). We also sequenced tra-3(e2333) in the ORF ± 1 kb and found a nonsense mutation at nucleotide position 1,779 (G to A) of the spliced tra-3 transcript. This resulted in a premature stop (W to stop) at position 593 of the TRA-3 protein. Both mutants were phenotyped for body size at 12 °C and 24 °C and compared to the wild-type N2. Like CB, body size was not affected by temperature in both mutants (Figure 4). The N2 phenotype was rescued by the fully suppressed mutant tra-3(e1107)sup-24(st354)IV, which promotes translational readthrough of the tra-3(e1107) mutation (Figure 4).
We then tested whether a larger body size could also be obtained by mimicking a low temperature environment through an artificial increase of [Ca2+]i at 24 °C. Although TRA-3 does not have a specific EF calcium-binding site in C. elegans, a well-conserved region has been shown to bind calcium [36,37]. We used thapsigargin (TG) to increase [Ca2+]i [38,39]. We found a clear dose–response relationship between TG and body size, showing that N2 grew larger at 24 °C at increased levels of TG (Figure 4). A significant increase in size was found at 0.015 μM TG compared to a positive control that included the solvent dimethyl sulfoxide (DMSO). Calpain activity was required for the TG-induced body-size enlargement because treatment with 0.015 μM TG did not result in a larger body size in homozygous tra-3(1107) mutants (Figure 4). These results indicate that calcium activation of TRA-3 may be controlling body size at different temperatures.
In addition to the F96L mutation, the observed phenotypic differences could be due to differential expression of tra-3. Therefore, we performed quantitative RT-PCR experiments on cDNA obtained from N2 and CB at 12 °C and 24 °C. It was found that expression was slightly enhanced at 24 °C in both wild types. There was no significant difference in tra-3 expression across temperatures between N2 and CB (results not shown). Based on these findings we hypothesised that observed TRB differences between N2 and CB were the result of a polymorphism in tra-3.
To further investigate the role of tra-3 in the wild-type CB, we performed a complementation analysis by crossing the near-isogenic line (NIL WN17–9) with tra-3(e1107). Heterozygous F1 from a cross between NIL WN17–9 and N2 revealed the recessive nature of the CB–TRB allele (Figure 3). The body size for the e(1107)/+ F1 offspring exhibited increased size at 12 °C indicating that tra-3(e1107) was recessive (Figure 3). Complementation analysis in which NIL WN17–9 was crossed with tra-3(1107) showed no differences in body size of F1 between high and low temperature (Figure 3). These results show that tra-3 is required for regulating body size in response to changing environmental temperatures and that an SNP in tra-3 is able to reduce this ability. We did not attempt to perform a complementation test between NIL WN17–9 and tra-3(e2333) because of the dominant nature of tra-3(e2333) over other tra-3 mutants. Homozygous tra-3(e1107) worms show partial masculinisation whereas homozygous tra-3(e2333) animals are wild-type hermaphrodites. Heteroalleles of these two mutants are also wild-type hermaphrodites indicating a dominance effect of e2333 over e1107 [40].
We next asked whether the N2 version of the tra-3 gene could transform CB to have a larger body size at low temperature. Therefore we carried out a transgenic assay in which tra-3 from N2 was transferred to the CB background. We exposed independently derived strains of CB(gfp) (control strains) and CB(gfp and tra-3(+)) to 12 °C and 24 °C. Figure 5 shows that the N2 phenotype was rescued in CB(gfp and tra-3(+)) because it grew 24% larger at the low temperature. CB(gfp) retained the CB phenotype because it did not grow larger at low temperature.
We next sought to determine whether F96L could lead to a diminished activity of TRA-3 in CB by conducting homology modelling of the 3D structure of TRA-3. The TRA-3 protein consists of four domains (I–III and T), where domain II is the protease catalytic site, and domain T does not have a critical calcium-binding function [34,41] but may be important for protein folding. Although TRA-3 does not have a specific EF calcium-binding site in C. elegans, a well-conserved region spanning the boundaries of domain II and III has been shown to bind calcium [36]. In addition, Moldoveanu et al. [37] reported on non-EF calcium-binding sites in domain II between position 62–74 (Ca-1). Homology modelling shows that F96 is located at the beginning of a short helix, H6, contiguous in space to the loop hosting Ca-1. In “open” configuration, corresponding to the absence of calcium, the distance between the α-carbons of F96 and E68, G69, and A70 reduces to 8–10 Å, as compared to ~10–14 Å corresponding to the “closed” configuration. In addition, the side chain of F96 is oriented toward the Ca-1 loop making their atoms to come frequently in van der Waals contact (<3.0 Å) (Figure 6). As the length of a leucine side chain is ~1.5 Å smaller than that of a phenylalanine, F96L will introduce a void in this region. Therefore, F96L can make a small but important difference by increasing the conformational space that the “opened” Ca-1 loop can sample during its dynamics. As the number of configurations increases this might reduce the probability to find the loop in its “closed” configuration and consequently reduce the ability for calcium binding.
The genetic control of the C. elegans body size has been intensively studied. Mutants such as sma-2, 3, 4, and daf-4 have a small body size and are defective in the TGF-β signalling pathway, which underlies body growth and development [42]. The lon mutants have been found to grow longer but not larger in volume [32,43]. It was shown that egl-4 mutants, defective in a gene encoding a cGMP-dependent protein kinase, have a much larger body size than N2 [32].
Here it is shown that TRA-3 has a prominent role in regulating the thermal plasticity of body size in C. elegans. Homology modelling shows that the F96L mutation in CB4856 attenuates the ability to grow bigger at lower temperatures by destabilizing the calcium-binding site in TRA-3. These data indicate that calcium signalling in response to temperature changes may lead to the activation of TRA-3. This mechanism to control the temperature–size rule is supported by various reports on the elevation of the free cytosolic calcium concentration in response to lower temperatures. Increase of cytosolic calcium levels in response to a gradual reduction of temperature is widely observed in plants [19,20] and ectothermic animals.
Many studies in other organisms have shown the importance of calpains in oncosis showing calpain-mediated cell swelling and disruption of plasma membrane permeability followed by cell death [35]. In C. elegans calcium-activated TRA-3 is known to be involved in the sex determination pathway by activating TRA-2A, a membrane protein that indirectly activates the zinc finger transcriptional regulator TRA-1A by binding and inhibiting a masculinising protein FEM-3 [44]. Current insights are insufficient to link aforementioned findings and to infer a putative pathway by which calcium activation of TRA-3 results in larger cell sizes in C. elegans.
Many different theories have been proposed to unravel the underlying mechanism of the temperature–size rule [2,7,45,46]. Our results partly fit the theory by Van der Have et al. [14] who suggested that the temperature–size rule is regulated by two distinct processes underlying temperature effects on body size: growth rate (which is the biomass increase per time unit) and differentiation rate (which is the reciprocal of development time). Their model presupposes that the temperature–size rule depends on a wide range of alleles differing in sensitivity to temperature. Our results show that a polymorphism in a single gene may attenuate the TRB in C. elegans.
CB was originally isolated in Hawaii while N2 originates from the UK. Whether the F96L mutation in CB reflects adaptive change or a fortuitous event is unknown. Both parental strains have been isolated decades ago and kept in the laboratory ever since, and additional field research is needed to establish whether this polymorphism and/or others in tra-3 are typical for strains isolated from tropical regions.
Our results do not provide insight into how natural selection modifies the temperature–size rule, yet they provide the basis for a more mechanistic understanding of the evolutionary outcomes. Like C. elegans the increased body size at lower temperatures in flatworms, Drosophila spp., and protists [47–49] is caused primarily by increased cell size. Because tra-3 shows a high homology with other ectothermic calpains [34,37], our findings may imply a possible role of calpain in the control of the temperature–size rule in other organisms as well.
We have presented genetic and structural evidence that an SNP in the gene tra-3 encoding a calpain-like protease is required for the regulation of the temperature–size rule in wild-type C. elegans. First, we found that the wild-type N2 complied with the temperature–size rule, whereas wild-type CB4856 defied it, and demonstrated that the genetic variation in the temperature–size response mapped to a single QTL on Chromosome IV harbouring tra-3. Second, we showed similar expression levels in tra-3 between the two wild types. Third, transgenic CB carrying an N2 allele of tra-3 complied with the temperature–size rule. Fourth, we found that F96L in TRA-3 attenuates the ability of wild-type CB4856 to grow larger at low temperatures. Finally, we showed that, based on homology modelling, the CB4856 mutation decreased the calcium-binding activity of TRA-3 rendering it less active. Because TRA-3 shows a high homology with other ectothermic calpains, our findings imply a possible role of tra-3 in the control of the temperature–size rule in other organisms as well. Together our data show that the response of a quantitative trait to temperature changes can be simple and far less complex than previously thought.
Both N2 and CB parental strains were homozygous. Strains were grown in 9-cm petri dishes at 15 °C or 20 °C on standard nematode growth medium with Escherichia coli strain OP50 as food source [50] and transferred to new dishes by a chunk of agar once a week. RILs were constructed according to [29].
NIL WN17–9 was constructed by crossing a single L4 hermaphrodite of RIL WN17 with five males generated from N2 on a 6-cm petri dish. The proportion of males in the offspring was approximately 0.5 indicating a successful cross. Subsequently, 12 crosses were set up, each with the use of single L4 hermaphroditic offspring of the former cross and five males derived from N2. Backcrossing procedure was continued with two L4 hermaphroditic offspring per successful cross. After described three generations backcrossing, six L4 hermaphrodites were picked from each successful cross and placed individually on a 3-cm petri dish to self. Selfing was continued for ten generations for each of the lines. All derived lines were subsequently genotyped at seven marker positions (including marker pkP4095) distributed equally over the fragments that were identified in RIL WN17 to be of CB origin. A total of five lines that appeared to have N2 alleles in all genotyped positions except for the marker pkP4095 were used for detailed genotyping. These lines were genotyped at all remaining marker positions. The results for one of the genotyped lines (NIL WN17–9) showed at all genotyped markers N2 alleles except for CB allele at marker pkP4095, and three neighbouring marker positions at Chromosome IV indicating a single ~6-cM DNA fragment of CB origin introgressed into N2 background. Genotyping was according to Li et al. [29].
Prior to an experiment, all lines (80 RILs and two parental) were synchronised at room temperature by transferring five adult nematodes to fresh 6-cm petri dishes and allowing them to lay eggs for 3–4 h, after which the nematodes were removed. Eggs were allowed to develop at 20 °C, and three days later synchronisation was repeated to get double-synchronised lines.
Measurements for parental and RIL body size at maturation were taken from Gutteling et al. [26]. Maturation was defined as the moment that the first few eggs are laid and can be easily observed. Because of this, body size at maturity can be precisely measured. For each RIL, three replicate experiments were performed using double-synchronised lines as a start. In each replicate, four adult nematodes per RIL were transferred to a fresh 6-cm dish, allowed to reproduce at room temperature for 2–4 h (average 2.5 h), and removed. Dishes were then stored at 12 °C and 24 °C climate chambers (Elbanton, http://www.elbanton.nl). Temperature was recorded with Tinytag Transit temperature loggers (Gemini Data Loggers, http://www.geminidataloggers.com). After 1 d (24 °C) or 4 d (12 °C), 12 juvenile nematodes were transferred at room temperature to individual dishes (3 cm diameter). Dishes were randomised and put back at the appropriate temperature. After 38 h (24 °C) or 145 h (12 °C) dishes were scanned at room temperature at regular intervals (1.5 h for 24 °C and 4 h for 12 °C) for the presence of eggs. If one or more eggs were observed, time and number of eggs were noted and the dish was put at −20 °C to prevent further development; a pilot study (unpublished data) showed that freezing did not affect body size.
Dishes were defrosted and nematodes were transferred to new dishes with NGM-agar. Digital pictures were taken with a CoolSnap camera (Roper Scientific Photometrics, http://www.photomet.com). Each nematode was measured automatically with Image Pro Express 4.0 (Media Cybernetics, http://www.mediacy.com). Using a measurement ocular we calibrated 10.000 pixels3 as 753.516 μm3. We assumed a rod-like shape of a worm where volume Vszm = π·(D/2)2·L = (1/4)·π·A2/L where D is diameter, L is length, and A = L·D. Because perimeter P = 2L + 2D ~ 2L we get:
Area (A, pixels2) and perimeter (P, pixels) of each worm were measured digitally. In subsequent analyses Vszm was used as input value for body size [26]. We studied the TRB, which is a plot of body size versus temperature, and used the slope of the reaction norm as a mapping trait.
For mutant phenotyping the following strains were used for body-size measurements at 24 °C and 12 °C: wild-type Bristol N2 and CB4856 isolate, tra-3(e2333), tra-3(e1107)/dpy-4(e116)IV, and tra-3(e1107)sup-24(st354)IV. The tra-3(e1107)/dpy-4(e116)IV heterozygotes segregate dpy-4 homozygotes, heterozygotes, and tra-3(e1107) homozygote hermaphrodites, which due to maternal effects are phenotypically wild type and segregate pseudomales [51]. We measured body size in these homozygote pseudomales, as well as the homozygote and heterozygote hermaphrodites. Body size was larger only in the hermaphrodites at 12 °C. Body size in the pseudomales was measured after the characteristic male tail [52] was completely formed. Experiments were performed on agar dishes (3 cm diameter) as described above. Samples were not frozen, but body size was measured directly when worms started laying eggs.
Crosses with the mutants and NIL WN17–9 were conducted by transferring J2 stage worms on small agar dishes (3 cm diameter) with three to five males. The worms were allowed to mate at 24 °C after which the females were transferred to new plates thus allowing them to lay eggs for 3–4 h. Mating was considered to be successful if the ratio of males:hermaphrodites was approximately 1:1 in the F1 hybrids. After this period females were removed and eggs allowed to develop at subsequent high or low temperature. When reproduction started body size was measured as described above.
TG (Sigma, http://www.sigmaaldrich.com) was applied to agar plates dissolved in DMSO. Different concentrations were added in a volume of 200 μl to petri dishes (3 cm diameter) each containing 2 ml of agar (end concentration in the agar: 0.004, 0.0075, and 0.015 μM) and seeded with E. coli. A positive control was included containing 200 μl of DMSO. After 24 h eggs were transferred to each dish and allowed to hatch. The size at maturity was recorded as described above.
The number of replicate worms measured for their body size were at 24 °C (italics) and 12 °C (bold): tra-3(e1107) 24, 16; tra-3(e1107) 24 °C DMSO 10; tra-3(e1107) 24 °C TG, 10; tra-3(e1107)sup-24(st354) 11, 13; tra-3(e2333) 19, 10; +/+ DMSO control 5; +/+ 0.004 μM TG 6; +/+ 0.0075 μM TG 6; +/+ 0.015 μM TG 6; NIL/+ 11, 15 16, 11; e1107/+ 8, 9, 8, 8, 8, 9; e1107/NIL 10, 16, 10, 18, 6, 4; NIL/NIL 22, 31.
Populations of N2 and CB were bleached (0.5 M NaOH, 1% hypochlorite) to collect synchronized eggs, which were then inoculated into fresh dishes. For each wild-type strain, four replicate dishes of synchronized eggs were kept in each of the two temperatures until maturity was reached. The nematodes were then collected and frozen in liquid nitrogen. Three independent samples were used for each strain and temperature. For each sample, individuals were synchronized and RNA was extracted using the Trizol method. RNA was subsequently purified (with genomic DNA digestion step) with the RNeasy Micro kit from Qiagen (http://www.qiagen.com). RNA concentration and quality were measured with Nano Drop (http://www.nanodrop.com). From each sample 2 μl of RNA were used to obtain cDNA using Superscript II reverse transcriptase from Invitrogen (http://www.invitrogen.com) and oligo d(t) primers from Genisphere (http://www.genisphere.com). cDNA was diluted 20× and used for RT-PCR with iQ Sybr Green Supermix from Bio-Rad in 20 μl reactions (http://www.bio-rad.com). Standard curves for each sample were generated by serial dilutions of the cDNA to select for primer efficiencies of 90%–110% and correlation coefficients greater than 0.99. We selected two reference genes (rps-20 and rpl-3) using geNorm on the basis of Vandesompele et al. [53]. All primers were designed with Beacon Designer avoiding secondary structures and cross homology. RT-PCR runs were done with MyIQ from Bio-Rad, and expression levels were calculated with the Bio-Rad Gene Expression Macro version 1.1 using the selected reference genes for normalization. Expression levels are presented relative to the lowest expression of the gene. At least two independent experiments were carried out for each gene.
Transgenic worm strains containing tra-3(+) from the Bristol N2 wild-type strain in the CB background were obtained from the Umeå Worm/Fly Transgenic Facility (http://www3.umu.se/utcf/index_eng.html). Standard microinjection methods were used [54]. A DNA fragment spanning the entire tra-3 locus and containing the endogenous tra-3 promoter was injected at a concentration of 25 μg/ml. The coinjection marker was pCC [55], a plasmid containing gfp under the control of the unc-122 promoter, which is active in coelomocytes. pCC was injected at a concentration of 50 μg/ml. Body size was measured as described above for five independently derived strains of CB(gfp) (control strains) and CB(gfp and tra-3[+]).
For RIL analysis a randomised block design was used (three blocks per RIL). Statistical analyses were performed in SAS. All data were found to be normally distributed according to the Box-Cox method. Comparison between treatments was tested with one-way ANOVA using PROC MIXED. In case of crossing experiments, replicate crossings were performed, and the data were analysed with a nested design where each cross was nested within temperature (cross[temperature]). In PROC MIXED we defined cross(temperature) as a random factor. The number of replicates was optimal to obtain the mean to be within the 95% CI. ANOVA was performed to study the effect of temperature, RIL, block, and interactions on body size. QTL mapping was used to identify the genomic regions (Wormbase release WS100) controlling various life-history traits. Composite interval mapping was used to identify responsible QTL because it is statistically a well-established and powerful tool; it has a better resolution of QTL peaks compared to interval mapping and is able to control for a number of background markers [56]. QTL analyses were performed with the software package QTL Cartographer version 2.0 [57] using forward regression, a maximum of five background parameters, and the default window size of 10 cM. The experiment-wise likelihood-ratio threshold significance level was determined by computing 1,000 permutations of each trait [58] as implemented by QTL Cartographer. These permutations can account for non-normality in marker distributions and trait values. A peak in the likelihood ratio LR was taken to indicate a significant QTL if LR > 10. Composite interval mapping is sensitive to the number of background markers included in the analysis. The relatively low number of five background markers was used because too many background markers can over-parameterise the model. However, in order to assess whether detected loci that were close to one another also suggested one QTL, we examined the inclusion of ten background markers. The results show the significant QTL based on this ten-marker correction. CIs for QTL were taken based on a 1-LOD support interval corresponding to 95% CI [59].
Template identification was performed with 3D-PSSM [60]. Results show that the first part of TRA-3 sequence corresponding to domains I–III best matches rat calpains M and μ with E-values of 0.0142 and 0.0332 respectively, corresponding to over 95% fold recognition confidence. By contrast, the last part corresponding to the T domain matches the C2 domain fold with a best hit to protein kinase C alpha C2, E-value: 0.583, corresponding to 90% confidence.
Restricting to only the TRA-3 DII domain where the F96L mutation occurs (Figure S1), best templates were searched for the “open” and “closed” configurations—corresponding to calcium free and bound states respectively.
For the “open” configuration the best match is with human M-calpain (pdb code—1kfu) with an E-value of 0.000686, 36.5% identity, and 67.5% similarity. However this calpain was not crystallized in its “closed” configuration as well, and further search for templates was needed to model this state.
Structural analysis of the existing M and μ calpains crystallized in “closed” state (1tl9, 1mdw, 1kxr, 1tlo, 1zcm, 2ary) showed that all of these are practically identical from a structural point of view, with main chain rms deviations of only 0.731–1.124 Å. Consequently all are equally good templates for TRA-3 DII and the closest sequence homologue can be used. This was found to be the rat μ-calpain (pdb code – 1tl9, identity: 38.6%; similarity: 67.5%, but an E-value of only 0.0553).
In building the models, target and template sequences were first aligned using MULTALIN [61]. This alignment was further optimised manually in several steps by incorporating information on secondary structure, accessibility, contacts, and functionality of important residues. Secondary structure profile of the target was raised by a consensus based on the top five prediction methods according to CASP6 (Critical Assessment of Structure Prediction Methods 6): JPRED [62], HNN [63], SSPRO [64], PROF [65], NNPREDICT [66]. The alignment was further refined by moving the gaps to correct for unfavourable exposures in the 3D model.
The 3D models were then built by coordinate transfer in the sequence conserved regions. Loops with insertions or deletions were generated ab initio, then subjected to multiple rounds of conformational search by simulated annealing and local minimization. Packing of long insertions was investigated using Modeller [67] by generating large number of loop conformers and subjecting them to statistical analysis. Simulated annealing was then used to optimise the top contenders followed by extensive rounds of energy minimisation. In the end, the entire model was subjected to repeated rounds of minimization to relieve unfavourable contacts.
Model building, refinement, and analysis were performed using the Accelrys programs: Insight II, Discover, Homology, Modeller, Charmm, Cdiscover, and the free-ware 8v1 version of Modeller on an Silicon Graphics, Octane 2 station. |
10.1371/journal.ppat.1000355 | Characterization of the Interferon-Producing Cell in Mice Infected with Listeria monocytogenes | Production of type I interferons (IFN-I, mainly IFNα and IFNβ) is a hallmark of innate immune responses to all classes of pathogens. When viral infection spreads to lymphoid organs, the majority of systemic IFN-I is produced by a specialized “interferon-producing cell” (IPC) that has been shown to belong to the lineage of plasmacytoid dendritic cells (pDC). It is unclear whether production of systemic IFN-I is generally attributable to pDC irrespective of the nature of the infecting pathogen. We have addressed this question by studying infections of mice with the intracellular bacterium Listeria monocytogenes. Protective innate immunity against this pathogen is weakened by IFN-I activity. In mice infected with L. monocytogenes, systemic IFN-I was amplified via IFN-β, the IFN-I receptor (IFNAR), and transcription factor interferon regulatory factor 7 (IRF7), a molecular circuitry usually characteristic of non-pDC producers. Synthesis of serum IFN-I did not require TLR9. In contrast, in vitro–differentiated pDC infected with L. monocytogenes needed TLR9 to transcribe IFN-I mRNA. Consistent with the assumption that pDC are not the producers of systemic IFN-I, conditional ablation of the IFN-I receptor in mice showed that most systemic IFN-I is produced by myeloid cells. Furthermore, results obtained with FACS-purified splenic cell populations from infected mice confirmed the assumption that a cell type with surface antigens characteristic of macrophages and not of pDC is responsible for bulk IFN-I synthesis. The amount of IFN-I produced in the investigated mouse lines was inversely correlated to the resistance to lethal infection. Based on these data, we propose that the engagement of pDC, the mode of IFN-I mobilization, as well as the shaping of the antimicrobial innate immune response by IFN-I differ between intracellular pathogens.
| Type I Interferons (IFN-I) are cytokines produced by the innate immune system immediately after intrusion of a pathogen. To produce large quantities of IFN-I once an infection is starting to spread throughout the body, the innate immune system employs a specialized “interferon-producing cell” (IPC). In the case of viral infections, IFN-I protect the host organism from rapid viral replication and spread. Conversely, organisms that cannot produce IFN-I are exquisitely sensitive to viral infections. Intriguingly, the opposite has been reported for the pathogen Listeria monocytogenes. Like virus, this bacterium replicates within cells of the host organism and stimulates IFN-I synthesis. Unlike virus, however, IFN-I sensitize the infected host to lethal pathology resulting from L. monocytogenes infection. In this article, we show that all tested molecules contributing to IFN-I production in Listeria-infected mice are responsible for a corresponding increase in mortality. We address the question of which cell type is responsible for producing vast quantities of IFN-I that can be measured in the serum of mice infected with Listeria. We show that these are not IPC, but rather macrophages, cells specialized to ingest and kill bacteria. We conclude that the engagement of cells for IFN-I production and also the effect of IFN-I on innate immunity is determined by the tropism and lifestyle of a particular pathogen.
| Type I interferons (IFN-I) comprise a family of around 20 members that bind a common receptor, the type I IFN receptor (IFNAR) [1]. Immunologically most relevant are IFNβ, with only one member in humans and mice, and the IFNα family with more than ten members. Pattern recognition receptors (PRR) for all classes of microbes are able to stimulate transcription of the IFN-I genes, establishing IFN-I production as a hallmark of innate immune responses to a vast number of viral and nonviral pathogens [2]–[8]. A common property of these receptors is to stimulate the activation of one or more transcription factors of the interferon regulatory factor (IRF) family, most importantly IRF3 and IRF7 [9].
Cells producing IFN-I employ different molecular strategies to induce synthesis of the cytokines. A feed-forward amplification mode is characterized by IRF3-dependent IFNβ production that predominates the early phase of infection. In a subsequent phase signaling of the IFNβ through the IFNAR stimulates activation of the transcription factor ISGF3 consisting of a STAT1/STAT2 heterodimer in conjunction with IRF9 [10]. This transcriptional complex activates the IRF7 gene promoter and initiates expression of the IRF7 mRNA. Finally, IRF7 is instrumental in transcribing multiple IFNα genes in addition to IFNβ [11]. An alternative way of stimulating IFN-I was described for virus-infected or toll-like receptor (TLR) ligand-stimulated plasmacytoid dendritic cells and, under certain conditions, myeloid dendritic cells (pDC and mDC, respectively) [12]–[14]. In these cells, the IFN-I genes are targeted by a pathway originating from endosomal TLRs (TLR 7 and TLR 9) that assemble a signalosome including MyD88 and IRF7 together with other IRFs such as IRF1 or IRF5 [15]–[20]. In pDC the TLR signal is rapidly relayed to IRF7, which is available for immediate stimulation of all IFN-I promoters. Thus, pDC produce vast quantities of IFN-I in response to infection with viruses and are generally referred to as interferon-producing cells (IPC).
Listeria monocytogenes is a Gram-positive, food-borne bacterial pathogen [21]. Major sites of Listeria replication during systemic infection of mammals are liver and spleen. Infection is exacerbated by the activity of IFN-I, shown convincingly by the increased resistance of IFNAR or IRF3-deficient mice to lethal infection [22]–[24]. Several explanations for this IFN-I effect have been provided, including decreases in TNF production or an enhancement of macrophage death [22]–[25]. Moreover, IFN-I were shown to sensitize cytotoxic T lymphocytes to the lytic action of LLO [24]. Enhanced lymphocyte killing and IL-10-dependent suppression of innate immunity as a result of phagocytic uptake of apoptotic cells increase the severity of infection-borne pathological effects [26].
L. monocytogenes is endowed with the ability to infect the cytoplasm of cells in the host organism [27]. This ability results from the endo/phagosome-disrupting activity of the major L. monocytogenes virulence factor, the Listeriolysin O (LLO), a bacterial cholesterol-dependent hemolysin [28]. A common property of L. monocytogenes and viruses is that recognition and signaling occur through both membrane-associated and cytosolic PRR [3]. In contrast to viruses, the cytosolic L. monocytogenes receptors mediating IFN-I induction have not been identified. In L. monocytogenes-infected macrophages, IFN-I synthesis occurs via feed-forward amplification [29]. The exact pathway and cell type responsible for IFN-I production during infection of mice with L. monocytogenes have not been examined. Considering the fact that IFN-I enhance the adverse effects of the early, innate immune response to L. monocytogenes [22]–[24], it is of major importance to understand their mobilization and mode of action.
In this study we sought to identify the essential IFN-I-producing cells during infection with L. monocytogenes. We used sorted splenic cell populations and multiple mouse models with defects in the IFN-I production machinery or with tissue-specific disruption of the ifnar1 gene to decipher the pathways and cells required for both IFNβ and IFNα production. Our results show that in contrast to viral infection models, pDC are not the source of IFN-I in response to L. monocytogenes. Rather, the cytokines are produced via feed-forward amplification by a splenic cell with cell surface markers characteristic of macrophages. We also show for the first time that IFNβ and IRF7 make strong contributions to the harmful action of IFN-I in L. monocytogenes infection because the lower levels of IFN-I production in these mouse lines were correlated with decreased susceptibility to lethal infection.
Knock-out mice were used to study pathways contributing to increase of IFNβ and IFNα in the serum of mice infected with L. monocytogenes. We established 24 h post intraperitoneal (i.p.) injection of bacteria as the point of maximum IFN-I induction in preliminary experiments (data not shown); therefore, IFNβ and IFNα were measured 24 h after infection with L. monocytogenes in subsequent experiments. Serum levels of IFNβ were generally at the detection limit of the ELISA and differences between wt and the IRF3−/− or IRF7−/− genotypes could not be reliably determined. By contrast, statistically significant differences were observed comparing wt and IFNAR1−/− animals (Figure 1A). This finding most likely reflects the importance of the IFNAR for IFNβ clearance. Increased accumulation of the cytokine in the blood of IFNAR−/− mice is therefore not in contradiction to subsequent findings showing that a fraction of IFNβ production occurs via feed-forward amplification through the IFNAR (e.g. Figures 2 and 3). Serum of infected wt mice contained higher levels of IFNα than IFNβ. IFNα species were reduced in absence of IFNβ, IRF3 or IRF7 (Figure 1B). Serum IFNα was virtually absent in mice lacking IFNAR1. Collectively the data show that the levels of serum IFNα critically depend on ‘early’ IFNβ and on IFN-I signaling, hence feed-forward amplification. Bacterial loads in liver and spleen after 24 h of infection were virtually identical between wt and IFNAR−/− mice (Figure S1). Therefore, differences in IFN-I production did not result from lower numbers of infecting bacteria.
The spleen is a major target organ of L. monocytogenes, irrespective of the route of infection by intragastric, intraperitoneal or intravenous application [21],[30]. Therefore, we tested both IFN-I production and response in spleens following infection with L. monocytogenes (Figure 2A and 2B, and Figure S2A). Production of both IFNβ and IFNα mRNA, assayed by RT-PCR (Figure 2A and 2B), was observed in wt animals. IFNβ mRNA synthesis required both IRF3 and IRF7 and, surprisingly, the IFNAR. Splenic IFNα production was reduced in absence of IRF3 or IFNβ and even more affected by the lack of IRF7 or the IFNAR. A representative analysis of four mice/genotype is shown in Figure 2A, whereas Figure 2B summarizes all experiments to show how many mice of each genotype produced IFN-I and account for individual responses to Listeria infection.
STAT2 is a subunit of the ISGF3 transcription factor complex and its tyrosine phosphorylation is a hallmark of IFN-I-treated cells [10]. Phosphorylated STAT2 was absent in IFNAR1−/− spleens, reduced slightly in absence of IFNβ and IRF3, and reduced strongly in absence of IRF7 (Figure S2A). STAT1 tyrosine phosphorylation occurs in response to both IFN-I and IFNγ [10]. Likewise, the STAT target gene Gbp2 (Figure 2A) responds to both IFN types. Both STAT1 tyrosine phosphorylation and Gbp2 expression occurred at levels similar to wt in all genotypes including IFNAR1−/− (Figure S2A). In agreement with this, IFNγ mRNA levels were equal to wildtype in all investigated genotypes (data not shown). Together the data show that the molecular requirements for splenic IFN-I production are very similar to those found for serum IFNα production. In both cases signaling through IFNAR and upregulation of IRF7 expression are of critical importance and indicate feed-forward amplification.
In bone-marrow-derived macrophages (BMM), L. monocytogenes stimulates IFNβ synthesis through the IRF3 pathway independently of TLR9 and MyD88 [29],[31]. Requirements for IFN-I production were further investigated in BMM and mDC infected with L. monocytogenes. IRF7 or IFNAR-deficiency had no effect on BMM IFNβ production (Figure 2C). Unlike IFNβ, IFNα production by infected BMM was dependent on IRF3, IRF7, IFNβ and the IFNAR (Figure 2D). In case of IRF7-deficiency, residual IFN-I still caused unimpaired tyrosine phosphorylation of STAT2 (Figure S2B). In mDC, surprisingly, both IFNβ and IFNα production strictly required IRF3, IRF7 and IFNAR1 (Figure 2E and 2F).
To investigate the contribution of IFNβ and IRF7 to the IFN-I-dependent increase in mortality, we monitored the survival of animals infected with L. monocytogenes. In accordance with previous findings [22]–[24], IFNAR deficiency caused a strong resistance to lethal infection, particularly during the innate phase of the anti-Listeria immune response (up to day 6, Figure 4A). By comparison, the increase caused by IFNβ deficiency was less pronounced. Lack of IRF7 had a greater impact on survival than absence of IRF3. This difference became smaller if survival was monitored beyond the period of the innate immune response (up to day 10). Lower pathogen burdens in liver and spleens from IRF7 or IFNβ-deficient mice reflected the increase in survival on day 3 (Figure 4B). Similar findings have been reported for IFNAR1 and IRF3-deficient mice [22]–[24]. Since differences in the bacterial load between wt animals and those with defects in IFN-I synthesis and/or response are not present 24 h after infection (Figure S1) the inhibitory effect of IFN-I on bacterial clearance must develop between day 1 and 3 post infection. Together, the data emphasize the importance of the early IFNAR, IFNβ and IRF7-mediated amplification of IFN-I production for adverse IFN-I action during the early, innate immune response.
To examine the cell type(s) important for IFN-I production and response during L. monocytogenes infection, we made use of tissue-specific IFNAR ablation. The contribution of myeloid cells to IFN-I effects was tested by mating mice with a floxed ifnar1 allele [32] with LysM-Cre mice [33]. Using this technology, a high degree of floxed allele conversion is obtained in macrophages and neutrophils, whereas it is inefficient in splenic DC [33]. In accordance with these observations, IFNAR1 expression in IFNARfl/fl-LysM-Cre mice was strongly reduced in CD11b+ cells that include macrophages, but similar to wt on CD11c+ cells that include myeloid and plasmacytoid DC (Figure S3). Absence of the IFNAR in myeloid cells caused a highly significant increase in the survival of Listeria-infected mice during the innate phase of the immune response up to day 6 (Figure 3A). In this period we observed 100% survival in mice with complete IFNAR1 deletion and 80% survival in mice with myeloid IFNAR1 deletion. At later times the survival curves diverged further, most likely due to a contribution of non-myeloid cells to the lethal outcome of infection. The increase in survival compared to wt mice could result from an important role of myeloid cells in IFN-I production. Alternatively, the myeloid response to IFN-I might directly contribute to the cytokines' detrimental action. To distinguish between these possibilities, serum IFN–I was compared in wt mice and mice lacking the IFNAR either on all or only on myeloid cells. This experiment revealed a major contribution of myeloid cells to serum IFNα by IFNAR-dependent feed-forward amplification (Figure 3B). Differences in the low level of serum IFNβ were not statistically significant between wt mice and those disrupted for the myeloid ifnar1 gene (data not shown).
Analysis of splenic IFN-I production and response in mice lacking the myeloid IFNAR showed that IFNα mRNA was completely absent (Figure 3C). Residual IFNβ mRNA was present, but the induction was reduced and more transient (compare 24 h and 48 h time points). No difference between genotypes was observed for the induction of Gbp2.
The data collected so far suggest that the majority of IFN-I is produced by a splenic cell type that employs an IFNβ/IFNAR/IRF7 pathway for feed-forward amplification of IFNα synthesis. This profile does not match IFN-I production by pDC where expression of IFN-I genes is regulated by a pathway originating from endosomal TLRs [34]. To further exclude the relevance of this pathway for IFN-I production during Listeria infection, TLR9-deficient mice were analyzed. The data summarized in Figure 5 show that neither serum IFNβ or IFNα (Figure 5A and 5B), nor splenic IFN-I mRNA synthesis or response were reduced in the absence of TLR9 (Figure 5C, Figure S4A). By contrast, production of IFN-I mRNA in bone-marrow-derived pDCs exposed to L. monocytogenes showed strong (IFNβ) or absolute (IFNα) dependence on TLR9 and/or MyD88 (Figure 5D–5F). Both IFNβ and IFNα mRNA synthesis were strongly affected by IRF7 deficiency and, surprisingly, also the absence of IRF3 (Figure 5G and 5H). Unlike pDC, mDC produced IFNβ independently of the TLR9 pathway, resembling macrophages in this regard (Figure S4B). Consistent with IFN-I production (Figure 5A–5C), the absence of TLR9 did not increase survival following infection with L. monocytogenes (Figure 5I). Rather, TLR9 deficiency slightly enhanced the lethality of L. monocytogenes infection, suggesting that the TLR9 pathway plays a protective role.
To further rule out pDC as major producers of IFN-I during Listeria infection of mice and to confirm a prominent role of myeloid cells, we purified cell populations from L. monocytogenes infected spleens by FACS and measured IFN-I production by RT-PCR. Figure 6 clearly shows that cells with a phenotypic profile of pDC (PDCA1+B220+CD11cdimCD11b−) fail to produce IFN-I in response to L. monocytogenes infection. The CD11chi population (either CD11b+PDCA1−B220− or CD11b−PDCA1−B220−) including typical mDCs did not express detectable IFNα mRNA, whereas these cells produced minor amounts of IFNβ mRNA. In contrast, CD11b+ cells not expressing CD11c, PDCA1 or B220, a profile characterizing macrophages, expressed high levels of both IFNα and IFNβ mRNA. Macrophages, therefore, are most likely the predominant IFN-I producers in Listeria-infected mice.
Using mice and cells ablated for genes involved in IFN-I synthesis, we established a molecular profile of the cell type(s) producing IFN-I upon infection with L. monocytogenes and correlated IFN-I production in the spleen and serum with the survival of infected hosts. Clear results emerging from our studies are that I) the TLR9/MyD88 pathway, a hallmark of IFN-I synthesis by pDC, does not significantly contribute to splenic (local) or systemic IFN-I production; II) consistently, feed-forward amplification requiring signaling through the IFNAR is of great importance and III) the main IFN-I producers are cells with characteristics typical of splenic macrophages. Finally and IV), IFNβ and IRF7 contribute to the enhancement of lethal infection by IFN-I. Reduction in IFN-I production that resulted from IRF3, IRF7, IFNβ or IFNAR deficiency by and large reflected the increase in resistance to lethal infection. Examination of pathogen organ loads revealed that the bacterial growth-promoting, hence lethality-enhancing activity of IFN-I develops between day 1 and 3 post infection.
Our studies focused on the early, innate phase of infection up to about day 6. IFN-I may additionally influence the ensuing adaptive immune response and, as suggested by recent studies, this effect may not be primarily adverse as in the innate response [35]–[37]. Diverse IFN-I effects on adaptive and innate immunity most likely explain that the relative increases in resistance to infection caused by the knock-out mice used in our study differ to some degree when analyzed at day 6 or day 10 after infection (Figure 4). Furthermore, the survival curve of mice with myeloid cell-specific IFNAR1 ablation suggests that cells outside this cell compartment influence the survival of Listeria infection particularly after the innate phase of the immune response.
Two recent studies closely examining the innate response to Listeria in infected spleens revealed a highly complex process requiring CD11c+ DC for the movement of Listeria to the white pulp and to initiate a concerted response involving macrophages, TIP-DC, NK cells and neutrophils [38],[39]. Our study shows that the TLR9 pathway was not engaged by L. monocytogenes for IFN-I synthesis in mice. By contrast, data with pDC-enriched bone marrow cultures show that Listeria can present a TLR9 ligand and stimulate this cell type for IFN-I production. This finding strongly implies that pDC do not contribute to IFN-I production during L. monocytogenes infection of mice, in spite of high bacterial loads in lymphoid organs that harbor this cell type. Although TLR9/MyD88-independent IFN-I synthesis has also been observed in mice infected with viruses [40]–[42], several reports suggest that the engagement of the pDC/TLR system for IFN-I production occurs when viruses gain access to lymphoid tissue either directly or because the pathogen breaks the initial barrier of infection formed by macrophages and/or epithelial cells [34], [41]–[43]. For example, lung infection with Newcastle disease virus led to IFNα production by alveolar macrophages, whereas systemic infection engaged both mDC and pDC for IFNα production [41]. Splenic mDC are reportedly major IFN-I producers during infection with Adenovirus, underscoring pathogen specificity in the choice of IPC and defying a general role of pDCs as producers of systemic IFN-I upon viral infection [44]. Under steady state conditions, pDC are mainly found in peripheral blood and lymphoid organs [13]. A recent report by Tam and Wick demonstrated expansion and activation of cells phenotypically resembling pDC in lymphoid tissue infected with L. monocytogenes [45]. The cells expressed activation antigens, but, consistent with our results, did not produce cytokines. Together, the data suggest that the ability of L. monocytogenes to stimulate pDC expansion occurs without direct contact and may be caused by cytokines released from infected cells. Unlike virus, L. monocytogenes appears to be unable to efficiently gain access to pDC even at lethal doses of infection when lymphoid organs are heavily infected. Direct contact may be prevented by an efficient containment of the pathogen in cells that are more readily infected than pDC. Owing to their phagocytic potential, splenic macrophages activated by NK cell-derived IFNγ, might clear Listeria before they reach pDC. It will be interesting to determine whether the choice of IPC determines the IFN-I effect on infection and whether macrophages as IFN-I producers are related to fatal outcome as in the case of L. monocytogenes. In this regard it will be important to examine IPCs in murine infection models such as S. pneumoniae, where the host benefits from the production of IFN-I [46].
IFNα production in BMM via feed-forward amplification required IRF3-dependent IFNβ production, signaling through IFNAR1 and IRF7. This resembled the molecular mechanism observed for splenic or systemic IFNα production after in vivo infection. To confirm the assumption that cells different from pDC are important IFNα producers in vivo we used IFNARfl/fl-LysM-Cre mice and demonstrated that IFNα synthesis critically depends on the presence of IFNAR1 on myeloid cells. Consistently, cells expressing surface markers characteristic of macrophages express both IFNβ and IFNα mRNA when purified from infected spleens. By contrast, this was not observed for the cell populations that include mDC and pDC. Thus, macrophages are the most likely IFNα producer cells after L. monocytogenes infection in vivo. An additional contribution of neutrophils cannot be ruled out entirely. However, little is known about the presence and abundance of this cell type in the spleen early after Listeria infection and it is similarly unclear whether neutrophils produce IFN-I in response to infection.
Whereas IRF3 was also critically involved in IFNβ production by bone marrow-derived macrophages, the IFNAR or IRF7 were not required. This contrasts the more important role of IRF7 for IFNβ synthesis in the L. monocytogenes infected spleen. Unlike bone marrow macrophages, much of splenic IFNβ resulted from an IFNAR/IRF7-dependent amplification loop because IFNAR−/− spleens expressed strongly reduced amounts of IFNβ mRNA. This may reflect a contribution of splenic mDC to IFNβ production. This cell type required IRF3 as well as IFNAR/IRF7 for IFNβ synthesis also when grown from mouse bone marrow. We assume that mDC produce low amounts of IFNβ immediately after Listeria infection in an IRF3-dependent manner. This initiates the feed-forward amplification loop causing synthesis of the majority of both IFNβ and of IFNα. This model, while consistent with the data, fails to provide a satisfactory explanation for the obvious difference between bone marrow-derived macrophages and mDC regarding the significance of IFNAR1/IRF7 for IFNβ synthesis. Possibly the much larger rate of cytoplasmic infection of macrophages drives IRF3-dependent IFNβ synthesis much more efficiently. A deeper understanding of cell type differences in the mode of IFN-I production is needed to provide ultimate clarity regarding this point. The assumption that splenic mDC participate in IFNβ production via IFNAR and IRF7 in infected mice is supported by the presence of mRNA for the cytokine in CD11c+/CD11b+ cells purified from infected mice. Nonetheless, a larger amount of IFNβ mRNA was detected in purified splenic macrophages, suggesting a major contribution of these cells to total IFNβ production. At the same time, a large fraction of splenic IFNβ requires feed-forward amplification and, judging from our findings with BMM, this appears to contrast the assumption of macrophages being major producers. A simple explanation for this discrepancy might be an intrinsic difference between splenic macrophages and BMM regarding feed-forward amplification of IFNβ production. Alternatively, the IFNAR/IRF7 requirement of splenic macrophages may derive from the fact that feed-forward amplification results from paracrine, not autocrine IFN-I priming. Under this assumption IFN-I is released by infected macrophages within or outside the spleen to prime uninfected splenic macrophages. Signaling through the IFNAR will result in IRF7 expression and the transcription factor is available for both IFNβ and IFNα production once primed cells become infected by Listeria. By contrast BMM are synchronously infected in culture and will immediately produce IFNβ in absence of IRF7 expression, employing exclusively IRF3 instead.
In summary our results provide answers to the question how the IFN-I system is deployed by an intracellular bacterial pathogen. Both, the observations made with cultured cells and our studies in mice, emphasize that molecular mechanisms governing IFN-I synthesis and response strongly depend on the pathogen as well as the host cell type. Deciphering the importance of these differences for various routes of infection and for diverse types of pathogens remains a challenging task for future research.
All animal experiments were discussed and approved by the University of Veterinary Medicine, Vienna institutional ethics committee and carried out in accordance with protocols approved by the Austrian law (GZ 680 205/67-BrGt/2003).
Listeria monocytogenes LO28 [47] were grown in brain heart infusion (BHI) (Difco) broth. Concentrations of bacteria were determined by measurement at OD600 and confirmed by plating serial dilutions onto BHI or Oxford agar (Merck) plates.
All mice were on a C57BL/6 background and housed under specific pathogen-free conditions according to FELASA guidelines [48].
Raw264.7 macrophages were cultured in DMEM (Gibco, Invitrogen) supplemented with 10% FCS (Gibco, Invitrogen). Bone marrow was isolated from femurs of 6–8 week old mice. For differentiation of BMM, cells were grown in DMEM (Gibco, Invitrogen) in the presence of 10% FCS (Gibco, Invitrogen) and L-cell derived CSF-1 as described [49]. The cultures contained >99% F4/80+ cells. mDCs were obtained by culture of bone marrow in DMEM (Gibco, Invitrogen) supplemented with 10% FCS (Gibco, Invitrogen) and X-6310 derived GM-CSF as described [50]. mDC cultures contained virtually no F4/80+ cells and the purity of CD11c+/CD11b+ cells was between 60 and 70%. pDCs were obtained by culture of bone marrow in DMEM (Gibco, Invitrogen) supplemented with 50 ng/ml of Flt-3-L for 6 days. The purity of CD11cdim/B220+ pDCs was between 60 and 70%. pDC cultures contained no F4/80+ cells.
Raw264.7 macrophages, primary BMM, mDCs and pDCs were infected with L. monocytogenes (derived from overnight culture) at a multiplicity of infection (MOI) of 10 and incubated for 60 min at 37°C in a humidified CO2 atmosphere. Extracellular bacteria were subsequently killed by exchanging medium to gentamicin-containing medium (final concentration 50 µg/ml). After another 60 min, medium was changed to medium containing 10 µg/ml gentamicin.
Listeria monocytogenes was grown in BHI broth (Difco) to late logarithmic phase (OD600 0.8), pelleted and resuspended in BHI containing 20% glycerol. Aliquots were shock-frozen and stored at −80°C. The concentration of L. monocytogenes aliquots was quantified by plating serial dilutions onto Oxford agar plates (Merck). For infection, the bacteria were thawn on ice, washed with PBS for three times and diluted in PBS (endotoxin-free, Sigma) to the appropriate concentrations. 500 µl of the bacterial suspension were injected into the peritoneum of 10- to 12-week-old mice. For the experiments shown, survival of infected mice was monitored for 10 days. For determination of the bacterial load, mice were killed at the indicated time points after infection and spleens and livers were homogenized in PBS. Serial dilutions of homogenates were plated on Oxford agar plates and colonies were counted after growth at 37°C for 24–36 h. For RNA and protein extraction from spleen and for sorting of splenic cells, mice were killed after 24 h of infection and spleens were isolated.
Mice were infected intraperitoneally with L. monocytogenes or injected with PBS (Sigma). Sera were collected after 24 h of infection and levels of IFNβ and IFNα were determined by ELISA according to the manufacturer's instructions (PBL InterferonSource).
Spleens were digested in 1 mg/ml collagenase (Roche) for 10 min and homogenized using a cell strainer. The cell suspension was labelled with the following antibodies: CD11b−FITC (Becton Dickinson), CD11c-APC (Becton Dickinson), B220-PE-Cy7 (Beckton Dickinson) and PDCA1-PE-Cy5 (Miltenyi Biotec). Four different cell populations were sorted by FACS: PCDA1+B220+CD11cdim (pDC), CD11b+CD11c−B220−PDCA1− (macrophages), CD11b+CD11c+B220−PDCA1− and CD11b−CD11c+B220−PDCA1− (myeloid dendritic cells).
RNA extraction from cells, reverse transcription and real-time PCR was performed as already described [29]. Data obtained by real-time PCR were analysed using SPSS and a Student's t-Test. Significant values are indicated by: n.s. not significant p>0.05, * p≤0.05, ** p≤0.01, *** p≤0.001.
Primers for real-time PCR were as described, except for IFNα: Probe: panIFNα: 5′-(6-Fam) AG+AA+GAA+A+C+AC+AG+CC (BHQ1)-3′ (+indicating LNA (Locked Nucleic Acid) nucleotides; Proligo); panIFNα-for: 5′-CCACAGGATCACTGTGTACCTGAGA-3′; panIFNα-rev: 5′-CTGATCACCTCCCAGGCACAG-3′
Spleens were isolated from mice infected with L. monocytogenes or injected with PBS for 24 h, shock-frozen in liquid nitrogen and stored at −80°C. For RNA extraction, 20 mg of frozen tissue was homogenized in lysis buffer RA1 (NucleoSpin RNAII kit, Macherey-Nagel) with the Precellys 24 homogenizer (Peqlab) at 6000 rpm for 30 sec. Subsequent isolation of total RNA and cDNA synthesis was performed as described [29]. Primers used for PCR were as follows: panIFNα-for-5′-ATGGCTAG(A/G)CTCTGTGCTTTCCT-3′; panIFNα-rev-5′-AGGGCTCTCCAGA(T/C)TTCTGCTCTG-3′; IFNβ-for 5′-CATCAACTATAAGCAGCTCCA-3′;
IFNβ-rev 5′-TTCAAGTGGAGAGCAGTTGAG-3′; Gbp2-for-5′-TGCTAAACTTCGGGAACAGG-3′; Gbp2-rev-5′-GAGCTTGGCAGAGAGGTTTG-3′;
L32-for 5′-ATTAAGCGAAACTGGCGGAAACCC-3′; L32-rev 5′- TTTCTTCGCTGCGTAGCCTGG-3′.
Supplementary Materials and Methods are presented in Text S1.
|
10.1371/journal.pntd.0004908 | Compounds Derived from the Bhutanese Daisy, Ajania nubigena, Demonstrate Dual Anthelmintic Activity against Schistosoma mansoni and Trichuris muris | Whipworms and blood flukes combined infect almost one billion people in developing countries. Only a handful of anthelmintic drugs are currently available to treat these infections effectively; there is therefore an urgent need for new generations of anthelmintic compounds. Medicinal plants have presented as a viable source of new parasiticides. Ajania nubigena, the Bhutanese daisy, has been used in Bhutanese traditional medicine for treating various diseases and our previous studies revealed that small molecules from this plant have antimalarial properties. Encouraged by these findings, we screened four major compounds isolated from A. nubigena for their anthelmintic properties.
Here we studied four major compounds derived from A. nubigena for their anthelmintic properties against the nematode whipworm Trichuris muris and the platyhelminth blood fluke Schistosoma mansoni using the xWORM assay technique. Of four compounds tested, two compounds—luteolin (3) and (3R,6R)-linalool oxide acetate (1)—showed dual anthelmintic activity against S. mansoni (IC50 range = 5.8–36.9 μg/mL) and T. muris (IC50 range = 9.7–20.4 μg/mL). Using scanning electron microscopy, we determined luteolin as the most efficacious compound against both parasites and additionally was found effective against the schistosomula, the infective stage of S. mansoni (IC50 = 13.3 μg/mL). Luteolin induced tegumental damage to S. mansoni and affected the cuticle, bacillary bands and bacillary glands of T. muris. Our in vivo assessment of luteolin (3) against T. muris infection at a single oral dosing of 100 mg/kg, despite being significantly (27.6%) better than the untreated control group, was markedly weaker than mebendazole (93.1%) in reducing the worm burden in mice.
Among the four compounds tested, luteolin demonstrated the best broad-spectrum activity against two different helminths—T. muris and S. mansoni—and was effective against juvenile schistosomes, the stage that is refractory to the current gold standard drug, praziquantel. Medicinal chemistry optimisation including cytotoxicity analysis, analogue development and structure-activity relationship studies are warranted and could lead to the identification of more potent chemical entities for the control of parasitic helminths of humans and animals.
| Schistosomiasis and trichuriasis affects millions of people worldwide and are caused by blood flukes and whipworms, respectively. Only a handful of anthelmintic drugs exist to treat these infections and the pipeline for the next generation of anthelmintic drugs is sparse, precipitating the need for new drug development. In this context, medicinal plants present a viable source of novel anthelmintic compounds. This inspired us to study the selected naturally occurring compounds derived from a Bhutanese daisy medicinal plant, Ajania nubigena for their anthelmintic activities. Here, using the xWORM motility assay, we demonstrate that two compounds, luteolin (3) and (3R,6R)-linalool oxide acetate (1), display significant broad-spectrum anthelmintic activity against two of the most important genera of human helminth parasites, the nematode whipworm and the platyhelminth blood fluke. Luteolin exhibited the best activities with IC50 values of 5.8 μg/mL against schistosomes and 9.7 μg/mL against whipworms. Using scanning electron microscopy we showed that luteolin damages the tegument of blood flukes and induces abnormalities in the bacillary bands/glands and cuticles of whipworms. Intriguingly, our previous study showed that luteolin (3) was effective against multi-drug resistant Plasmodium falciparum malaria. Due to its broad-spectrum anti-parasitic activities, luteolin (3) is a desirable drug lead scaffold, which could be used for developing effective compounds to control and treat numerous tropical diseases.
| The World Health Organization (WHO) recognises 17 different ‘neglected tropical diseases’ (NTDs) that affect more than 1.4 billion people in 149 countries [1]. Helminth infections caused by roundworms (nematodes) and flatworms (platyhelminths) comprise the largest group of NTDs [2]. Whipworms (Nematoda) cause trichuriasis and infect about 800 million people worldwide, second among the nematodes only to Ascaris infection [3]. The schistosome blood flukes (Platyhelminthes) cause schistosomiasis, a disease that afflicts more than 240 million individuals and kills hundreds of thousands each year [4].
A variety of approaches have been employed to combat these infections including education, vector control, sanitation and hygiene, behavioural change and mass drug administration (MDA) programs. Various in vitro and animal model studies have highlighted the repurposing of existing drugs and discovery and development efforts for new drugs [2, 5] but all things considered, the pipeline for the next generation of anthelmintic drugs is sparse. Indeed, a systematic assessment of databases of drug regulatory authorities and the WHO, as well as clinical trial registries, revealed that no new antiparasitic drugs have been approved during the last decade [6]. There are only a handful of anthelmintic drugs on the market, some of which have unwanted side effects or achieve poor cure rates due to primary drug resistance developing in the parasites [7–9].
For example, praziquantel, which is the sole frontline drug used in the mass treatment of schistosomiasis, is efficacious but has many disadvantages: a) it is ineffective against juvenile stages of the parasite, b) reduced efficacy has been reported in field studies [10], c) there is a strong possibility that praziquantel resistance could appear if sufficient selection pressure is applied and mass drug administration is continued [11], and d) its active (S)-enantiomer and inactive (R)-enantiomer components remain inseparable in the production process, rendering bulky tablets that discourages patients from taking the right doses or the complete dosing regimen, which could trigger the development of drug resistance [12]. Until new arsenals of safe and effective drugs and/or vaccines are made available, helminth infections will continue to affect the world’s most impoverished populations, causing significant morbidity and mortality worldwide.
While new drugs can be developed synthetically, natural products—especially the medicinal plants—have been an important pool of antiparasitic drugs. Quinine and artemisinin discovered from medicinal plants continue to save the lives of millions of people worldwide. As such, the notion of therapeutics derived from medicinal plants has re-surfaced [13]. Crude extracts and compounds of plant origin have been demonstrated to possess broad biological activities in in vitro and ex vivo assays and animal models of parasitic infections [14–19]. Edwards et al. [20] showed that 7-keto-sempervirol isolated from the boxthorn from which goji berries are harvested, Lycium chinense, was effective against Schistosoma mansoni and Fasciola hepatica. A compound that displays such broad anti-parasitic activity against various life stages of multiple parasites is highly desirable. Extracts of the Bhutanese medicinal plant from the flowering daisy family, Ajania nubigena (Syn. Tanacetum nubigenum DC.) have been previously shown to possess broad biological activities including antiparasitic effects against Plasmodium falciparum and antimicrobial properties [21]. It is locally known as m.khan-d.kar and has been used in Bhutanese traditional medicine (derived from Tibetan scholarly medicine) for thousands of years as incense and for treating an array of conditions and infections including wounds, bleeding and swelling [21]. Although this plant is not specifically indicated for treating intestinal worms, the decoction of its closely related species, Tanacetum parthenium L. (feverfew) and Tanacetum dolichophyllum Kitam has been traditionally used by the Ladakhis Amchis (medical system derived from Tibetan medicine and similar to Bhutanese traditional medicine) [22] and Costa Ricans healers [23] against intestinal worms. These plants have reserves of highly aromatic essentials oils that have evolved to aid in plant protection and competition against plant parasites and herbivorous insects. Chemically, these plants contain similar chemotypes including sesquiterpenes and flavonoids [23]. Encouraged by these lead information, we have investigated the anthelmintic properties of four compounds isolated from the Bhutanese A. nubigena against two of the most important genera of human helminth parasites, the nematode whipworm (Trichuris) and the platyhelminth blood fluke (Schistosoma). To monitor worm viability we used xWORM, a technique that monitors helminth motility in real time using xCELLigence [24–25]. The advantage of using xWORM over other methods is that it enables high-throughput screening of a large number of compounds in a fully automated, label-free manner.
The aerial part of wild Ajania nubigena was collected from alpine mountains (altitude range of 3600–4800 meters above sea level) of Lingzhi, Bhutan in August 2009. The collected plant material was air-dried and a herbarium specimen with voucher number 73 was deposited at the herbarium collection section of Menjong Sorig Pharmaceuticals, Ministry of Health, Bhutan. The air-dried plant material (2 kg) was chopped into small pieces and was repeatedly extracted with methanol (AR/HPLC grade, 3 L over 48 h). The extract was filtered and then concentrated using a Buchi rotary evaporator to generate a crude methanol (MeOH) extract (58.2 g). The isolation technique was described previously [21]. MeOH extract was dissolved in MeOH:H2O (200 mL, 1:9) and then fractionated with hexane followed by ethyl acetate to obtain the hexane extract (28.0 g) and the ethyl acetate extract (12.5 g), respectively. Subsequently, essential oil (EO) extraction was performed using hydro-distillation (60°C). One kg of dried plant material yielded 7 mL of pale green EO. The crude MeOH extract and EO were subjected to extensive natural products isolation processes. Flash column chromatography packed with Merck Kieselgel 60 PF254 and pre-coated silica plates (0.2 mm silica thickness, Merck) were used for repeated separation and purification of compounds. UV light (short wavelength of 254 nm, long wavelength of 366 nm) and ceric ammonium molybdate (CAM) were used for visualization and detection of compounds on Thin Layer Chromatography (TLC) plates. Eight compounds were isolated and characterised in total from the MeOH and EO extracts using Infrared (IR) Spectroscopy, Mass Spectrometry (ESI-MS, HR-EI-MS), Gas Chromatography Mass Spectrometry (GCMS), and Nuclear Magnetic Resonance (NMR-1H, 13C, gCOSY, gNOESY, TOCSY, gHSQC and gHMBC) [21]. In this study, we have selected four major compounds whose structures are produced in Fig 1: (3R,6R)-linalool oxide acetate (1), (E)-spiroether (2), luteolin (3) and luteolin-7-O-β-D-glucopyranoside (4).
The stock solutions of the four test compounds were prepared at the concentration of 100 mg/mL in DMSO and then subsequently diluted them with respective tissue culture media to make 10x solutions. 20 μl of 10× drug concentration was added to 180 μl of media containing helminths in the E-plate wells. Control worms were cultured in the presence of DMSO equivalent to that used for the highest drug concentration; this group was used to determine 100% motility. For schistosomula drug assays, stock solutions were diluted in culture media with two-fold dilutions with in-well concentrations of (2–1000 μg/mL).
S. mansoni cercariae were shed from infected Biomphalaria glabrata snails (Biomedical Research Institute, MD, USA) by exposure to light at 26°C for 2 hours and used to infect 12–14 week old male BALB/c mice (120 cercariae/mouse) by abdominal penetration [26]. Adult flukes were perfused from the mesenteries 7 weeks post-infection and then transferred to Basch medium for culturing as previously reported [27]. To monitor the effects of drugs on S. mansoni schistosomula, shedding of cercariae from snails and subsequent in vitro transformation to schistosomula was performed as described by Peak et al [28].
Genetically susceptible mice (STAT6-1-) were orally infected with T. muris eggs (200 μL volume containing ~ 200 eggs) and sacrificed after 4 weeks. Adult worms were harvested from the caecum, washed with PBS/2x antibiotic/antimycotic (AA) and resuspended in 100 μl of RPMI containing 10% foetal calf serum and AA (culture medium) then transferred to E-plates for motility assessment using the xWORM assay.
The trematocidal effects of test compounds against S. mansoni were evaluated using an xCELLigence SP system (ACEA Biosciences) as described by us [25]. Adult flukes (1 fluke per well) were placed in triplicate for each compound into 96 well E-plates (ACEA Biosciences) containing 180 μl of culture medium and cultured overnight at 37°C with 5% CO2 to obtain a baseline motility reading. Test compounds were added to E-plates and motility was monitored for 12–40 hr. The inter-well spaces of E-plates were filled with 100 μL culture media. All experiments were carried out as per manufacturer’s instructions with 15 sec read intervals using the real time cell assay (RTCA) software (ACEA Biosciences) as described previously [24–25].
Similarly, the nematocidal effects of test compounds against T. muris were assessed using the same xCELLigence SP system as described above. We determined the optimal culture duration and worm concentration to maximize the signal to noise ratio using the xWORM technique for the first time with T. muris. Different numbers of adult T. muris (2, 4 and 8) were added to individual E-plate wells and motility was monitored overnight. Four worms of mixed gender per well in a final volume of 200 μl of culture medium was determined to be optimal for this study. The E-plates containing worms were treated with prepared concentrations of the test compounds and were monitored using the xCELLigence SP system for 12–40 hr. Inter-well spaces were filled with 100 μL of culture medium or PBS to prevent evaporation. Each set of conditions was monitored in triplicate.
The 96 well plates containing culture media were loaded with schistosomula (100 μL volume containing ~ 100 schistosomula) in triplicate and treated with the test compounds at various in-well concentrations of 2–1000 μg/mL. Plates were cultured at 37°C with 5% CO2 for 12–40 hr and were finally stained with trypan blue solution to assess final viability after treatment. The stained schistosomula were observed by light microscopy, and live and dead flukes in each well were counted manually and 50% inhibitory concentration (IC50) values were obtained.
The IC50 values of test compounds were determined based on the motility index for adult worms as described by us [25]. Briefly, motility index was calculated as the standard deviation (SD) over 800 data points (i.e. 4 readings per min for 200 min) of the cell index (CI) difference from the rolling average over 20 data points (10 proceeding and preceding CI values—5 min total). One hundred percent motility was determined from the average motility index of the untreated wells, while 0% motility was determined from a media only well (no worms present). The motility index averaged over 100 data points (25 min) was converted to percentage motility and this figure was used in GraphPad Prism 6.0 to calculate dose response curves. We used a log (test compound concentration) vs normalised response (100%–0%) formula, with variable slope when data were sufficient or set -1 hill slope when data was limited, and automatic removal of outliers (with default ROUT coefficient used: Q = 1.0%). IC50 values for each dose concentration were calculated at 1 hr, 6 hr, and 12 hr post-treatment of the worms with the test compounds. Compounds with IC50 values of higher than 100 μg/mL were considered ineffective in this study.
Statistical analyses were undertaken using GraphPad Prism 6.0. When data were sufficient to use the variable slope analysis, the Hill Slope and the Log IC50 value were together compared for significant differences using an extra sum-of squares F-test. For the in vivo mouse experiments, 1-way ANOVA with Holm-Sidak’s multiple comparisons test was used for determining significance p-values.
Worms treated with the test compounds were prepared for scanning electron microcopy (SEM) as follows: a) fixed in 3% gluteraldehyde in Sorensen’s buffer overnight, b) dehydrated for 15 min in a graded ethanol series (50%, 60%, 70%, 80%, 90%, 100%), mixture of ethanol and hexamethyldisilizane (HMDS) (1:1 ratio) and then finally with pure HMDS (100%), c) the dehydrated worms were covered and left overnight in a fume hood to allow the HMDS to evaporate. Completely dried worms were placed on an aluminum stub (at least three worms from each treatment regimen), sputtered with gold and visualized using a JEOL JSM scanning electron microscope operating at 10 kV. Each worm on a stub was scanned from head to tail to determine if the compounds had altered its gross morphology. Digital image acquisition was performed on the affected region of the worms using Semaphore software.
Four to five week-old STAT6-1- mice were grouped (each group with 9 mice) as: solvent control, positive control and luteolin (3). Each mouse was orally infected with 200 μl of PBS containing approximately 200 live embyronated eggs of T. muris. These mice were housed for 4 weeks with constant access to water and pelleted food. After 4 weeks, luteolin (3) and the positive control drug (mebendazole) prepared in 1% DMSO/PBS were orally administered at a single dose of 100 mg/kg (9 mice in total for each group) as per the protocol [29–30]. Five days after one dose of oral treatment the mice were sacrificed, worms were harvested from the caecum, and counted manually using light microscopy. The recorded numbers of worms were averaged to find the percentage reduction in worm burden for each group of mice.
The permit to collect medicinal plants from the park management areas around Lingzhi, Bhutan was obtained from the Department of Forest, Ministry of Agriculture and Forestry in Bhutan. The material transfer agreement and approval was sought from the National Biodiversity Centre of Bhutan. MeOH extracts of the plants were transported to Australia with prior approval from the Bhutan Agriculture and Food Regulatory Authority, University of Wollongong and sample inspections by Australian Quarantine & Inspection Service in 2010. The James Cook University (JCU) animal ethics committee approved all experimental work involving animals (Ethic approval number A2213). Mice infected with S. mansoni and T. muris were raised in cages in the JCU animal facility for 4–7 weeks in compliance with the Australian Code of Practice for the Care and Use of Animals for Scientific Purposes, 7th edition, 2007 and the Queensland Animal Care and Protection Act 2001. Mice were kept under normal conditions at regulated temperature (22°C) and lighting (12 hr light/dark cycle) with free access to pelleted food and water. All reasonable efforts were made to minimise the suffering of the mice.
The isolation and characterization of compounds from the A. nubigena plant were performed as reported previously by us [21]. Out of eight compounds isolated from this plant, several were reported to have antimalarial, antibacterial, antifungal and cytotoxicity activities [21]. Encouraged by the antimalarial activities, four major compounds (Fig 1): (3R,6R)-linalool oxide acetate (1), (E)-spiroether (2), luteolin (3) and luteolin-7-O-β-D-glucopyranoside (4), were tested for broad-spectrum anthelmintic activities against two distinct phyla of helminths, the platyhelminth trematode blood fluke S. mansoni and the nematode whipworm T. muris.
Of the four compounds tested, (3R,6R)-linalool oxide acetate (1), luteolin (3) and luteolin-7-O-β-D-glucopyranoside (4) showed anti-schistosome dose-dependent anti-schistosome effects (Fig 2A). At the highest concentration tested (1 mg/mL) all three compounds killed flukes within 1–12 hr. Lower drug concentrations, however, took longer to kill flukes reflected by higher motility index values (Fig 2A).
When the IC50 values of each compound were averaged or combined (calculated for the dose concentrations of 0.1–1000 μg/mL) for each time point (1 hr, 6 hr and 12 hr), luteolin (3) and luteolin-7-O-β-D-glucopyranoside (4) appeared to be fast acting on worms as their IC50 values did not change significantly between the initial and final 12 hr dosing time points (Fig 2B). On the other hand, (3R,6R)-linalool oxide acetate (1) was slow-acting, showing two-fold decreases in IC50 at each 6 hr time point. Of the compounds assessed, luteolin (3) exhibited significantly better trematocidal activity at both 6 hr and 12 hr time points with IC50 values of 4.6 μg/mL and 5.8 μg/mL, respectively.
The compounds that were active against adult S. mansoni were also tested against the intra-mammalian larval stage, the schistosomulum. Monitoring of schistosomula survival in the presence of different drug concentration using Trypan blue exclusion showed that luteolin (3) started to show lethal effects at the lowest dilution of 3.91 μg/mL and achieved 98–100% killing at the dilution of 31.3 μg/mL. (3R,6R)-linalool oxide acetate (1) however started to show lethal effects at a concentration of 125 μg/mL and only achieved about 43% killing at the maximum dose tested of 250 μg/mL. Schistosomula treated with 1% DMSO alone (solvent control) had 100% survival as measured by Trypan blue exclusion. A dose response curve of schistosomula survival after treatment with luteolin (3) revealed an IC50 of 13.3 μg/ml (Fig 3B).
Luteolin (3) demonstrated the best anthelmintic activity when motility was assessed using xWORM. The effect of this compound and praziquantel on adult fluke morphology at 4–20 μg/mL concentration, with particular emphasis on the tegument, was assessed by SEM. Adult flukes treated with praziquantel adopted a tightly coiled appearance due to contraction. Both male and female flukes treated with luteolin (3) were contracted and coiled compared to control flukes cultured in 1% DMSO in media, but not as tightly coiled as praziquantel-treated parasites (Fig 4A–4F).
Observation of fluke teguments by SEM under high magnification revealed that the tegument of DMSO-treated flukes displayed numerous tubercles, with well-formed spines in the males (Fig 5A) and clearly defined surface grooves with sensory papillae in females (Fig 5B), and oral suckers with clearly defined pits containing sharp spines (Fig 5C). Flukes treated with praziquantel (Fig 5D–5F) and luteolin (3) (Fig 5G–5I) exhibited severe morphological alterations of the tegument.
At the lowest concentration tested of 4 μg/mL we observed clear tegumental damage induced by luteolin (3) (Fig 5G–5I), similar to that induced by praziquantel (Fig 5D–5F). While male flukes (Fig 5G) suffered partial loss of pits and their encased spines, female flukes (Fig 5E) showed surface erosion, constriction of grooves, bursting of small sensory papillae and formation of holes on the tegument. At higher concentrations of 20–1000 μg/mL, luteolin (3) and praziquantel completely destroyed the body surfaces and exhibited erosion of tubercles, oral and ventral suckers (S1 Fig).
Prior to testing the nematocidal effects of (3R,6R)-linalool oxide acetate (1), (E)-spiroether (2), luteolin (3) and luteolin-7-O-β-D-glucopyranoside (4), we standardized the culturing conditions of adult T. muris for the xWORM assay. E-plate wells containing four adult worms (both males and females) yielded optimal motility signals (S2 Fig), and was the condition selected for subsequent anthelmintic screening of the test compounds. (3R,6R)-linalool oxide acetate (1) and luteolin (3) showed the best anti-Trichuris activity with IC50 values of 20.4 μg/mL and 9.7 μg/mL, respectively, calculated on cell motility index at the 12 hr time point (Fig 6A).
The IC50 values of the four compounds tested here were obtained using xWORM and calculated at 1 hr, 6 hr and 12 hr time points (Fig 6B). Luteolin (3) was the most efficacious drug in terms of reduced motility of T. muris, exhibiting the lowest or equally low IC50 values at all time points and a final 12 hr value of 9.7 μg/mL.
Based on the efficacy of luteolin (3) at reducing Trichuris motility, we examined the morphological changes in the cuticle induced by this compound using the SEM protocols described by Stepek et al [30] and Tritten et al [31] specific to T. muris. Live adult T. muris (mixed sexes) were treated for 48 hr with luteolin (3) or mebendazole at dose concentrations of 200–1000 μg/mL. Morphological changes were observed towards the anterior end of the worms in the form of partially damaged bacillary band/glands and adjacent cuticle. The bacillary band is a specialized row of longitudinal cells of some nematodes consisting of glandular and non-glandular cells. These bands host the glands. Worms treated with DMSO/media (solvent control) alone had a moderately coiled appearance with a smooth cuticle displaying knitted parallel segmental joins and minimal shrinkage of bands (Fig 7A–7C) in comparison to the groups treated with mebendazole (Fig 7D–7F) and luteolin (3) (Fig 7G–7I). At higher magnification we observed that the luteolin-treated worms exhibited blister-like formations on the surface of the cuticle, swelling and loosening of cuticle seams/grooves near the bacillary glands (Fig 7I). These morphological changes were similar to that of the mebendazole-treated worms (Fig 7F).
Based on the significant in vitro nematocidal activity demonstrated by luteolin (3), this compound was further assessed for its anti-Trichuris effect in vivo using a mouse model of T. muris infection [29–31]. Four weeks post-infection with T. muris eggs, mice were administered with a single 100 mg/kg oral dose of luteolin (3). When mice were sacrificed 5 days later we observed that a single treatment of this compound resulted in a 27.6% reduction in worm burdens (249 worms); luteolin (3)–treated mice = 651 worms; DMSO-treated mice = 900 worms; P = 0.0087) (Fig 8). The positive control drug, mebendazole, reduced worm burdens by 93.1% (mebendazole treated mice = 50 worms). The activity of luteolin (3) against T. muris, despite being significantly better than the untreated control group, was markedly weaker than mebendazole in the mouse model. Mice treated with luteolin (3) did not show signs of toxicity at any of the concentrations tested.
Globally, helminth infections caused by roundworms (nematodes) and flatworms (platyhelminths) comprise the largest group of NTDs [2]. Schistosomiasis (caused by the platyhelminth blood flukes) and trichuriasis (caused by the nematode whipworms) affect about 240 million and 800 million people, respectively [3–4]. While schistosomiasis is transmitted through infected snails and water, trichuriasis is transmitted through soil and is therefore referred to as a soil-transmitted helminth infections (STHI). The STHI are among the most common infections worldwide and affect the poorest and most deprived communities where sanitation is poor. Sole reliance on praziquantel for schistosomiasis, and only a very small number of drugs (some with poor cure rates) for STHI, has precipitated the need for new anthelmintic drugs to treat parasites that infect both humans and animals [7]. In this context, medicinal plants present a viable source of novel anthelmintic compounds. Indeed, many anti-parasitic drugs including quinine, chloroquine, artemisinin and atovaquone were originally purified from plants. Our initial study, involving the crude CHCl3 extract of A. nubigena (syn. Tanacetum nubigenum DC.) and its compounds luteolin and luteolin-7-O-β-D-glucopyranoside, showed significant antiparasitic activities against the multidrug resistant K1CB1 strain and chloroquine sensitive TM4/8.2 strain of Plasmodium falciparum [16, 21]. Interestingly, this plant and its close relatives have been used in the ethnomedicines for treating arrays of disorders including wound, bleeding and worm infection [16, 21–23].
In this study, we have demonstrated that compounds linalool oxide acetate (1) and luteolin (3) had significant trematocidal activity against S. mansoni and nematocidal activity against T. muris. These compounds are simple small secondary plant metabolites belonging to the terpenes and flavonoids. Luteolin (3) was the most active compound against both parasites with IC50 values of 5.8 μg/mL against S. mansoni and 9.7 μg/mL against T. muris calculated at the 12 hr time point. It also effectively killed schistosomula with an IC50 value of 13.3 μg/mL. Intriguingly, this same compound has been shown to have significant anti-malarial activity against P. falciparum TM4/8.2 (chloroquine-antifolate sensitive strain) and K1CB1 (multidrug resistant strain) [21]. New anti-parasitic drugs require excellent safety and therapeutic profiles, should exhibit broad spectrum activity against different types of infections, and also display significant activity against different developmental stages of parasites. Current anthelmintic drugs are generally effective at treating single stages of target helminths. For example, praziquantel is effective against adult stage schistosomes but not schistosomula/cercariae. Therefore, finding a broad-spectrum drug that could treat multiple diseases or multiple life stages is desirable when treating large populations in resource-poor settings. Luteolin (3) met these criteria in that it has anti-malarial [21], anti-fluke and anti-whipworm properties. In addition, this compound was effective in killing schistosomula, the stage of S. mansoni that is unaffected by praziquantel.
While our findings do not specifically address the mechanism of action of these anthelmintic compounds, we showed that luteolin (3) is capable of damaging the outer surface membranes of the parasites–the fluke tegument and the nematode cuticle and associated glandular structures, and worms adopted a coiled state of paralysis. Previous studies on S. mansoni have demonstrated that the tegument plays essential roles in many processes at the host-parasite interface, and numerous molecular pathways that are represented at the host-parasite boundary are anti-parasitic drug targets [5, 32]. SEM has been used to demonstrate the mechanisms of many anti-parasitic drugs, including oxamniquine, praziquantel, mefloquine, mebendazole and artemisinin [20, 32–39]. Schistosomes treated with these drugs displayed vacuolization or bubble-like-lesions, surface erosion, destruction of tubercles and tissues, loss of sensory papillae, and pore formation leading to death. SEM has also been used to reveal cuticular damage in nematodes, particularly for the benzimidazole class of drugs [40–41]. Stepek et al [30] and Tritten et al [31] first used the plant extracts and nitazoxanide to demonstrate the mechanisms of the antiparasitic action against T. muris. These same morphological changes were observed in adult S. mansoni and T. muris that were treated with luteolin (3), suggesting that the mechanism of action is similar to at least some of the existing anthelmintic drugs. For T. muris, luteolin (3) and mebendazole mainly affected the anterior bacillary band and surrounding glands. The anterior ventrolateral face of the worm contain the glandular pores, bacillary band sensory cells and glands, and the stichosome cells which helps in the formation of the syncytial feeding site in the host, and plays an important role in the excretion of digestive enzymes, pre-digestion and nutrient uptake in Trichuris [42–45].
Luteolin (3), with a LogP value of 2.6, meets the Lipinski rule of 5 criterion for drug-likeness [46]. Generally, compounds with LogP values in the range of 2–3 are more likely to diffuse/permeate through the cell membrane of an organism, and therefore enabling them to interact with target receptors. There was no structural similarity between our active compounds and the currently used anthelmintic drugs, which suggest that compounds that damage the worm surface do not necessarily have similar structural scaffolds. It should be noted that our in vivo assessment of luteolin (3) against T. muris infection at a single oral dosing of 100 mg/kg, despite being significantly (27.6%) better than the untreated control group, was markedly weaker than mebendazole (93.1%) in reducing the worm burden in mice. Moreover, the mice showed no signs of ill health, suggesting that luteolin (3) is not overtly toxic and allows future studies to explore the efficacy of multiple treatments. We did not assess in vivo efficacy of luteolin (3) against S. mansoni in the mouse model. Future work will entail synthesis of luteolin and its derivatised compounds in a thorough in vivo assessment of efficacy in mouse models of schistosomiasis and trichuriasis, as well as other soil-transmitted helminth infections.
|
10.1371/journal.pntd.0000371 | Taenia solium Cysticercosis Hotspots Surrounding Tapeworm Carriers: Clustering on Human Seroprevalence but Not on Seizures | Neurocysticercosis accounts for 30%–50% of all late-onset epilepsy in endemic countries. We assessed the clustering patterns of Taenia solium human cysticercosis seropositivity and seizures around tapeworm carriers in seven rural communities in Peru.
The presence of T. solium–specific antibodies was defined as one or more positive bands in the enzyme-linked immunoelectrotransfer blot (EITB). Neurocysticercosis-related seizures cases were diagnosed clinically and had positive neuroimaging or EITB.
Eleven tapeworm carriers were identified by stool microscopy. The seroprevalence of human cysticercosis was 24% (196/803). Seroprevalence was 21% >50 m from a carrier and increased to 32% at 1–50 m (p = 0.047), and from that distance seroprevalence had another significant increase to 64% at the homes of carriers (p = 0.004). Seizure prevalence was 3.0% (25/837) but there were no differences between any pair of distance ranges (p = 0.629, Wald test 2 degrees of freedom).
We observed a significant human cysticercosis seroprevalence gradient surrounding current tapeworm carriers, although cysticercosis-related seizures did not cluster around carriers. Due to differences in the timing of the two outcomes, seroprevalence may reflect recent T. solium exposure more accurately than seizure frequency.
| Cysticercosis is a parasitic disease caused by the tapeworm Taenia solium, common in areas with limited sanitation or with migration from these populations. The adult parasite is hosted in the human intestine and releases large numbers of eggs with the feces. Human beings sometimes ingest eggs due to poor hygiene, and then eggs sometimes lodge on the brain and after a few years can cause intense headaches and seizures. During a study in seven rural communities in Peru, individuals exposed to T. solium eggs were often tightly clustered at the homes or immediate surrounding of the carriers of the adult parasite. However, no aggregation of cases of seizures was found near carriers. It appears that seizures do not cluster around carriers because several years pass between exposure to T. solium eggs and the onset of seizures. During these years the adult parasite has probably died or people had moved within or even outside their communities. Therefore, only a partial understanding of the epidemiology of cysticercosis is gained by studying seizures cases.
| Taenia solium cysticercosis is a parasitic disease endemic in developing countries where pigs are raised in close contact with human feces. Humans are the only definitive host and harbor the adult tapeworm. Taeniasis occurs after ingestion of improperly cooked pork and tapeworm carriers disseminate eggs in their feces. Cysticercosis is the infection with the larvae or cyst, and both people and pigs can become infected by fecal-oral contamination [1]. In humans, cysts often locate in the central nervous system (CNS) causing neurocysticercosis (NCC). Seizures are NCC's main clinical feature, although manifestations can range from asymptomatic, mild headaches and seizures to death [2]. Neurocysticercosis imposes a heavy financial burden to cases and their families, and treatment costs and productivity losses account on average for 53% of an annual minimum wage salary in the first year of treatment [3].
While it is well known that harboring a tapeworm or living with a carrier are factors associated with increased cysticercosis risk and disease burden [4],[5], it is not known if cysticercosis risk remains elevated outside the carrier's home. Two reports suggested that cysticercosis cases aggregate in neighboring households [6],[7], although they did not determine the magnitude and origin of clustering. In contrast, an epidemiological study published in 1992 actually did not detect clustering in human cysticercosis seroprevalence [8]. Recent work from our group demonstrated a substantial increase in swine cysticercosis seropositivity 50 and 200 m around tapeworm carriers and more extended seropositivity distance gradients [9]. However, given the different exposure patterns of pigs and humans, it remains unclear whether these findings can be extrapolated to human cysticercosis.
Quantifying the aggregation of human cysticercosis around carriers is a potentially relevant public health issue, because focalized control interventions could be developed to target such cysticercosis hotspots. We evaluated this hypothesis in northern Peru with the same methods previously used to assess clustering of swine cysticercosis. We determined if human cysticercosis also clustered around tapeworm carriers, comparing the clustering patterns of two key epidemiological parameters of human cysticercosis: seroprevalence and NCC-related seizures.
We analyzed data collected during the baseline assessments of a longitudinal study that evaluated control measures for T. solium cysticercosis. The study was conducted in seven rural, poor villages of the district of Matapalo, Tumbes, in Peru's northern coast near the Ecuadorian border. The study protocol was approved by the institutional review boards of the Universidad Peruana Cayetano Heredia, the Johns Hopkins Bloomberg School of Public Health and the Centers for Disease Control and Prevention. All study participants provided informed consent and assent if legal minors, using a single consent form was used for all study procedures. Prospective participants were told that they could refuse to participate in specific procedures. The socioeconomic characteristics of the area have been described elsewhere [9],[10]. We evaluated if there were increased rates of human cysticercosis seroprevalence and a lifetime history of NCC-related epileptic seizures surrounding tapeworm carriers.
These procedures have been described in detail previously [9]. Briefly, T. solium taeniasis mass treatment took place between November–December 1999, together with a population census during which household coordinates were recorded with sub-meter accuracy using global positioning system (GPS) hand-held receivers. Eligible, consenting participants received a single dose of niclosamide (Pharmamed, Malta). Children under five and pregnant women were not treated, and treatment coverage was 95% among eligible residents. Stool samples were requested from all consenting residents before and after treatment regardless of age, pregnancy or treatment status. The study provided all the materials and instructions necessary to avoid self-contamination, and the stool collection rate among treated residents was 88%. Stool specimens were treated with an ether sedimentation technique and examined for Taenia sp. eggs by microscopy [11].
During the census blood samples were taken by finger prick and stored on filter paper [12]–[14] to determine the presence of cysticercosis-specific antibodies with the serum enzyme-linked immuno-electrotransfer blot assay (EITB). Reactions to at least one of the seven diagnostic bands on EITB are considered positive. The EITB has 100% specificity, and its sensitivity is 98% in individuals with ≥2 viable lesions [15],[16] and approximately 60%–70% in patients with only one degenerating cyst or calcified lesions only [17].
A validated questionnaire [18],[19] was applied to identify individuals with a history of epileptic seizures. Two different neurologists interviewed consecutively subjects selected by the questionnaire and confirmed or rejected the diagnosis following the criteria of the International League Against Epilepsy [20],[21]. Individuals with a confirmed diagnosis of epileptic seizures were offered a brain computed tomography (CT) scan to detect the presence of NCC-compatible lesions.
The swine serosurvey was conducted between November 1999 and January 2000, capturing pigs 2 months old and older excluding pregnant sows. A 6–8 ml serum sample was obtained by vena cava puncture and tested for cysticercosis-specific antibodies using the EITB assay.
Two main outcomes were analyzed: 1) percent human cysticercosis seroprevalence and 2) lifetime prevalence of cysticercosis-related seizures, operationally defined for our analysis as seizure cases with either positive EITB or NCC-compatible lesions on the CT scan. Both outcomes were calculated at the individual level during univariate analyses and then aggregated by household during multiple regression analysis. The main covariate was the distance to the location of the nearest confirmed tapeworm carrier, calculated by assigning each subject the GPS coordinates of their household. Distances were calculated in meters using equator equivalences of latitude and longitude respectively [22]. The same procedures were used in our previous work [9].
We assessed the association between human cysticercosis and the distance to the nearest tapeworm carrier by separately estimating seroprevalence and lifetime seizure prevalence distance gradients using three different approaches. First, we described the shape of the distance gradient using piecewise cubic splines. Splines with 2–7 sections were initially defined with equal numbers of positive individuals in each section, and the best-fitting spline was then chosen. Second, we evaluated if seroprevalence or lifetime seizure prevalence increased exponentially near tapeworm carriers, testing if they were linearly associated with the base 2 logarithm of the distance [−log2(distance+1)]. Finally, seroprevalence and seizure prevalence rates were estimated for specific distance ranges. An additional analysis by distance ranges was conducted to evaluate the probability of finding tapeworm carriers in the proximity (50 m) of high seroprevalence households.
The human cysticercosis seroprevalence and lifetime seizure prevalence distance gradients were estimated using binomial family, logarithmic link function generalized linear models [23] to calculate prevalence ratios (PR). The relative change in prevalence rates associated to the distance to the nearest tapeworm carrier was estimated by:
Statistical significance was determined using the Wald test and the Akaike's Information Criterion (AIC) was used to assess model fit. One degree of freedom (DF) Wald test are presented unless stated otherwise.
The association of the study outcomes with sociodemographic and swine farming covariates was also assessed. Neighborhood population density was assessed by the number of households in a 100 m radius. In-house crowding was measured with the ratio of household members per bedroom. Numeric variables such as neighborhood density, crowding, pigs owned and swine cysticercosis seroprevalence were categorized in tertiles. All these covariates were included in nested, sequential regression models to estimate adjusted seroprevalence and seizure prevalence gradients. Standard errors were scaled using the square root of the Pearson's χ2 dispersion parameter to account for the correlation in outcomes of individuals from the same household.
For further validation, two secondary outcomes were analyzed at the univariate level only: 1) seroprevalence defined by ≥3 positive EITB bands, as stronger EITB positivity is associated with greater parasitic burden and more severe infection [24],[25], and 2) lifetime overall seizure prevalence, either NCC or non-NCC related. All statistical analyses were performed using Stata 9.2 (Stata Corporation, US) and all confidence intervals (CI) were calculated at the 95% level. Maps were prepared with ArcMap 9.0 (ESRI, US).
The baseline census registered 1004 people and 237 families. We excluded 106 subjects from the analyses: 12 due to incomplete GPS measurements, 54 because they did not reside in the study area, and 40 temporary residents who stayed <2 nights per week in the area. Therefore, our analysis included 898 individuals from 212 families.
The coverage of the serological survey was 89.4% (803/898) and the overall seroprevalence of cysticercosis was 24.4% (196/803), while 10.1% of all individuals had three or more positive EITB bands. Eleven tapeworm carriers were found in nine households during the stool survey, for a prevalence of taeniasis of 1.2%. The household seroprevalence of human cysticercosis and location of tapeworm carriers is presented in Figure 1.
The neurological survey reached 858 residents (95.6%) and 131 subjects reported a history of epileptic seizures. Two neurologists evaluated 111 of the 131 positive respondents and a random sample of negative responders who reported headaches or other neurological symptoms, confirming a lifetime history of seizures in 42 individuals (42/838, 5.0%). Only one tapeworm carrier reported a history of seizures, but was not confirmed during the neurological evaluation.
Out of 42 confirmed seizures cases, 23 were classified as NCC-related seizures because of NCC-compatible lesions on brain CT (15 cases), positive serology with negative CT (8 cases) or positive serology but without CT (2 cases). Using the operational definition proposed for our analysis, the prevalence of NCC-related seizures was 3.0% (25/837) after excluding one seronegative seizure case who did not presented for CT scan.
Cubic splines showed increased human cysticercosis seroprevalence at the carriers' homes and in their immediate proximity (Figure 2, top). On average, seroprevalence had a 12% relative increase each time distance to the carrier halved (Table 1, 95% CI: 8%, 17%, p<0.001, Wald test). A statistically significant seroprevalence gradient was observed across distance ranges. Seroprevalence was 21% (148/693) >50 m from a carrier and increased to 32% (23/71) at 2–50 m (Prevalence ratio [PR] = 1.52, 95% CI: 1.01–2.29, Wald test p = 0.047), and from 1–50 m there was an additional increase to 64% (25/39) at the carriers' home (PR = 1.98, 95% CI: 1.25–3.14, Wald test p = 0.004). Human seroprevalence was marginally higher among carriers themselves compared to their household contacts (9/11 = 82% vs 16/28 = 57%, PR = 1.43, 95% CI: 0.93–2.21, Wald test p = 0.107). Distance to the nearest carrier fitted seroprevalence better using splines (AIC:2.31) compared to distance ranges (AIC:2.33) and logarithmic transformation (AIC:2.36).
In contrast to seroprevalence, there was no indication in any analyses of a trend in the frequency of NCC-related seizures associated to the distance to the nearest tapeworm carrier (Figure 2, bottom). Only two of the 25 cases of NCC-related seizure lived within 50 m of a carrier, and splines did not show any association between the distance to the nearest carrier and seizures, regardless of the number of curve segments fitted. No significant associations were observed in the analyses by the logarithm of distance (Wald test p = 0.639) and distance ranges (Wald test p = 0.629, 2 DF), respectively. No systematic overall trend was observed across distance ranges either (Table 2).
In the analysis of secondary outcomes, seroprevalence defined by the presence of three or more positive EITB bands increased 13% each time distance to the nearest carrier halved (95% CI: 6%, 22%, Wald test p = 0.002). The seroprevalence distance gradient, however, was observed only within 50 m of carriers. There were no differences between >50 m and 1–50 m from a carrier (62/693 = 8.9% vs 5/71 = 7.0%, Wald test p = 0.637), but seroprevalence increased from 7.0% in 1–50 m to 14/39 = 35.9% at the carriers' home (PR = 4.01, 95% CI: 2.32–6.93, Wald test p<0.001). Additionally, carriers had significantly higher seroprevalence than their household contacts (8/11 = 72.7% vs 6/28 = 21.4%, PR = 3.39, 95% CI: 1.48–7.80, Wald test p = 0.004). The prevalence of all seizures, in contrast, was not associated to the distance to the nearest carrier, neither as the base 2 logarithm of the distance nor when analyzed by distance ranges (Wald test p = 0.842 and p = 0.604 [2 DF], respectively).
Human cysticercosis seroprevalence was not associated with age, gender, village, number of pigs owned, crowding and availability of latrines. Families that owned pigs had a marginally lower seroprevalence, and human population density and seroprevalence in pigs were associated with significant differences too. In the only village where some households had sewage (Matapalo), human cysticercosis seroprevalence was positively associated with the presence of sewage installations. None of these variables, however, were significantly associated with human seroprevalence after adjusting for distance to the carrier and harboring a tapeworm. Human cysticercosis seroprevalence remained increasing by 10% as distance to the nearest carrier decreased by half (95% CI: 5%, 15%, Wald test p<0.001) after adjusting for having a tapeworm.
Despite a clear association between distance to the tapeworm carrier and human cysticercosis seroprevalence, the chance of finding a carrier 50 m around a high seroprevalence (≥80%) household never exceeded 31% in any of the models.
Our findings demonstrate a significant gradient in human T. solium-specific antibodies around tapeworm carriers, increasing from 21% at >50 m from tapeworm carriers to 64% at the homes of carriers themselves. Seropositivity hotspots and distance gradients around carriers have also been observed in swine cysticercosis [9], and gradients in T. ovis and T. hydatigena parasitic burden were also observed around canine carriers [26]–[28]. NCC-related seizures, however, did not show any clustering patterns surrounding carriers. Current or previous epileptic crises may not be an efficient, useful indicator to identify active cysticercosis transmission foci, and seroprevalence may reflect recent exposure more accurately than seizure frequency.
After a newly infected tapeworm carrier starts disseminating viable T. solium eggs and a seroprevalence gradient is formed around the carrier, several factors may prevent observing a carrier-centered seizure cluster later on. First, exposure does not always culminate in cyst implantation in the central nervous system and only a fraction of NCC cases will eventually present seizures [2]. Second, most tapeworms probably die before the 3–5 year NCC pre-patent period [29],[30], which leads to the relatively low taeniasis rates found in most symptomatic NCC patients and their relatives [5],[31], and suggests that NCC-related seizure cases found in the community are unlikely to coexist with the tapeworms that infected them. Third, seizure gradients around carriers will be confounded by the migration of carriers and epileptic individuals, other NCC-seizure cases infected by tapeworms outside the community and non-NCC related seizure cases. Therefore, the absence of seizure gradients around carriers may result because NCC-related seizure cases were probably infected mainly by previous generations of tapeworm carriers instead of currently present carriers. This means that when an NCC case first begins seizing, the tapeworm that infected him/her is at least three years old and most likely dead already. Alternatively, either the carrier or the NCC case may have moved farther from the case. Thus, the identification of risk factors for NCC-related seizures is probably clouded if covariates such as proximity to carriers are measured when symptoms onset or later instead of closer to the exposure period.
Close proximity to a T. solium tapeworm carrier is the main risk factor for swine cysticercosis seroprevalence [1], and we observed a strong association between seroprevalence and both harboring a tapeworm as well as the distance to the nearest carrier, even after multiple-regression adjustment. Other factors associated to seroprevalence in univariate analyses were not significant in regression analyses. The precarious latrines and sewage connections present in some villages [32] apparently failed to provide a protective effect, similarly to previous reports where the poor latrine design and maintenance probably did not reduce environmental contamination [9],[10]. Besides the proximity to a tapeworm carrier, most villagers of this uniformly rural area probably had similar chances of exposure to T. solium eggs.
Previous research demonstrated that the number of seropositive EITB bands is correlated to the number of NCC-related lesions observed in CT [24] and we found increased seropositive with three or more positive EITB bands at the homes of tapeworm carriers among other household members. Close, in-house contact with carriers may explain the stronger EITB responses observed, probably related to higher infective doses and increased parasitic burden [24]. However, increased seroprevalence with one or more positive EITB bands was found in subjects living up to 50 m from carriers. A significant fraction of EITB seropositive humans and pigs are known to serorevert [9],[33],[34], and in pigs this was more frequent in animals with 1–2 EITB bands [9]. Therefore, it is likely that most of the excess human seropositivity 50 m from carriers corresponds to subjects with transient immunity. Only a small fraction of the seroprevalent cases may actually develop cysts and could present seizures later, consistently with the rates of NCC-related lesions found in community settings [10].
Both swine and human cysticercosis seroprevalence presented significant gradients surrounding T. solium tapeworm carriers, although with certain differences. Higher human seroprevalence was observed at the carrier's home compared to other households <50 m of a tapeworm, but no such difference in swine seroprevalence was observed. Increased human cysticercosis seroprevalence was observed within 50 m from carriers, while swine cysticercosis remained increased up to 51–500 m [9]. Swine coprophagia and free-range pig rearing probably explain the extended swine seroprevalence gradient. Human transmission by fecal-oral contamination, in comparison, is probably a less-efficient exposure mechanism, concentrating increased exposure mostly in the immediate proximity of carriers.
A limitation of this study is the absence of application of a sensitive coproantigen ELISA for human taeniasis [35], a test that can detect twice as many tapeworm carriers compared to our Ritchie formol ether test [11]. Missed carriers reduced the statistical power of the analyses and probably introduced exposure misclassification, decreasing the chances of finding a true association between seizures and distance to the nearest T. solium carrier. However, the same eleven tapeworms detected were enough to show highly statistically significant cysticercosis gradients both in swine seroprevalence and seroincidence in a previous study [9] as well as in human seroprevalence in this study. The lack of a seizure gradient probably was not due to low statistical power, because with similar number of detected carriers and cases allowing demonstrating a distance seroincidence gradient in pigs but no seizure rate trends in humans.
The observed seroprevalence gradients provide a first estimate of the size of T. solium cysticercosis seropositivity hotspots around carriers at 200 m and 50 m for swine and humans, respectively. In comparison, increased T. hydatigena infectivity was found 80 m from carriers [26] and up to 175 m for T. ovis [27]. Ecological associations suggested transmission over larger distances [27],[28], although without compelling evidence. Despite the limited comparability of findings due to differences between species, hosts, transmission mechanisms, research methods and context, the available data appears to suggest that tapeworm carriers spread contamination over a few hundred meters at best, with greater impact within smaller (50–80 m) radius.
Our results showed that human cysticercosis seroprevalence clustered in the immediate surroundings of T. solium tapeworm carriers but seizure frequency did not. The absence of seizure clusters suggests that the location of seizure cases does not correlate with recent transmission hotspots and probably has little use for prevention. Also, these results suggest that risk factor for seizures should attempt to address the period when cases were exposed instead of the time when symptoms began years later. Further insight into other epidemiological parameters such as tapeworm lifespan, duration of environmental contamination, and egg dispersion mechanisms should enhance neurocysticercosis control in the underdeveloped settings where this disease is one of the main causes of acquired epilepsy.
|
10.1371/journal.pntd.0007672 | Opportunistic pathogens and large microbial diversity detected in source-to-distribution drinking water of three remote communities in Northern Australia | In the wet-dry tropics of Northern Australia, drinking water in remote communities is mostly sourced from bores accessing groundwater. Many aquifers contain naturally high levels of iron and some are shallow with surface water intrusion in the wet season. Therefore, environmental bacteria such as iron-cycling bacteria promoting biofilm formation in pipes or opportunistic pathogens can occur in these waters. An opportunistic pathogen endemic to northern Australia and Southeast Asia and emerging worldwide is Burkholderia pseudomallei. It causes the frequently fatal disease melioidosis in humans and animals. As we know very little about the microbial composition of drinking water in remote communities, this study aimed to provide a first snapshot of the microbiota and occurrence of opportunistic pathogens in bulk water and biofilms from the source and through the distribution system of three remote water supplies with varying iron levels. Using 16s-rRNA gene sequencing, we found that the geochemistry of the groundwater had a substantial impact on the untreated microbiota. Different iron-cycling bacteria reflected differences in redox status and nutrients. We cultured and sequenced B. pseudomallei from bores with elevated iron and from a multi-species biofilm which also contained iron-oxidizing Gallionella, nitrifying Nitrospira and amoebae. Gallionella are increasingly used in iron-removal filters in water supplies and more research is needed to examine these interactions. Similar to other opportunistic pathogens, B. pseudomallei occurred in water with low organic carbon levels and with low heterotrophic microbial growth. No B. pseudomallei were detected in treated water; however, abundant DNA of another opportunistic pathogen group, non-tuberculous mycobacteria was recovered from treated parts of one supply. Results from this study will inform future studies to ultimately improve management guidelines for water supplies in the wet-dry tropics.
| Water providers in the wet-dry tropics of Northern Australia face additional challenges to keep drinking water microbiologically safe. The source water is often rich in iron-cycling bacteria leading to excessive biofilm formation in pipes and it can also contain the emerging opportunistic pathogen Burkholderia pseudomallei causing the severe disease melioidosis in humans and animals. We know very little about the ecology of microbes in remote community water supplies, so to start to fill this gap we assessed the microbial composition from the source to the distribution of three remote water supplies. We not only found that the geochemistry of the source water had a substantial impact on the composition of the iron-cycling bacteria but B. pseudomallei was cultured from source water with low organic carbon but elevated iron levels and from a multi-species biofilm linked to iron bacteria. No B. pseudomallei were detected in treated water; however, abundant DNA of another opportunistic pathogen group, non-tuberculous mycobacteria, was recovered from treated parts of one water supply. This work lays the foundation for future studies to ultimately improve management guidelines for water supplies in the wet-dry tropics.
| Water providers in the wet-dry tropics of Northern Australia face significant challenges to keep drinking water safe and free of opportunistic pathogens. One such opportunistic pathogen is Burkholderia pseudomallei, an environmental saprophytic bacterium and causative agent of the severe disease melioidosis affecting humans and animals [1, 2]. People most at risk are those suffering from diabetes, chronic lung or renal disease or hazardous alcohol use [3]. Until recently, melioidosis was thought to mainly affect people in Northern Australia and Southeast Asia where B. pseudomallei is endemic. However, a recent modelling study predicted 165,000 annual melioidosis cases worldwide of whom 89,000 were estimated to die [1]. B. pseudomallei is a natural component of the soil and surface water microbiota in rural Darwin, Northern Territory in northern Australia and 30% of tested unchlorinated residential water wells (bores) were positive for the bacteria [4, 5]. B. pseudomallei has been isolated from aerator sprays and tank sludge from water treatment plants ([6]; own observation) and melioidosis cases and deaths due to contaminated drinking water have been documented in Northern Australia and Thailand [6–10]. These supplies were either not chlorinated or the disinfection process was interrupted. B. pseudomallei is successfully contained by free chlorine levels of 0.5 to 1 mg/L, although in laboratory experiments, some strains were more chlorine tolerant [11].
Groundwater in many areas of Northern Australia contains naturally high levels of iron and it is unclear to what degree this promotes B. pseudomallei survival. B. pseudomallei has a redundant system of siderophores allowing it to acquire non-bioavailable ferric iron [12, 13]; a positive association between B. pseudomallei and total iron levels was found in bore water [14] while the association was negative in soil with high iron levels [15, 16] suggesting a unimodal rather than linear relationship across the range of iron levels encountered in the environment.
Water with high iron levels attracts naturally occurring iron bacteria which metabolize the iron and contribute to pipe corrosion and reduced bore yield. While some bacteria such as Geothrix fermentens or Shewanella sp. reduce iron in anoxic groundwater using organic carbon as electron donor, in niches with low oxygen iron oxidizers such as Gallionella ferruginea thrive and facilitate the production of abundant ferric oxide precipitates which block pipes, and contribute to biofilm formation reducing disinfection efficiency [17–19]. Most biofilms consist of a complex mix of bacterial taxa and can also be associated with fungi, viruses or protozoa [18]. Although biofilms are a known reservoir for opportunistic pathogens such as nontuberculous Mycobacteria, Legionella pneumophila or Pseudomonas aeruginosa [20], we still do not know to what degree B. pseudomallei colonizes multi-species biofilms in water pipes.
Water supplies of remote communities mainly depend on chlorination as disinfection treatment and are vulnerable to exposure to opportunistic pathogens in the event of a chlorination breakdown or if pathogens are chlorine-resistant. Indigenous people in remote communities often have higher rates of chronic diseases such as diabetes and thus, are more at risk of infection if exposed to opportunistic pathogens [21]. A multiple barrier approach to improve water quality is needed [22]. However, without full knowledge of what microbes occur in the source and distribution water, it can be difficult to design and apply barriers suitable for northern Australia.
This scoping study aimed to provide a first snapshot of the microbiota in bulk water and biofilms from the source and through the distribution system in three water supplies of remote communities; one supply with naturally high iron levels, one with medium and one with low levels. There were three study objectives: A) the detection and culture of opportunistic pathogens with a focus on endemic B. pseudomallei; B) the detection of taxa with known iron bacteria using 16s rRNA gene amplicon sequencing for microbial profiling; and C) the characterisation of the bacterial and archaeal microbiota and its association with nutrients and site characteristics. We hypothesized that water treatment would have the largest impact on the microbiota followed by the origin of the source water. We also hypothesized that water supplies fed from unconfined shallow aquifers would contain more bacteria also occurring in soil including B. pseudomallei. Results from this work will inform and guide future studies to ultimately improve management guidelines suitable for Northern Australia to minimize microbial risk in the drinking water distribution network.
Water and biofilms from the drinking water distribution system (DWDS) were sampled from three remote Indigenous communities in the Top End of the Northern Territory (NT), Australia. The Top End has a tropical savannah climate, with a distinct dry and wet season and average annual rainfall of 1,727 mm between Oct and March (www.bom.gov.au). The “HighFe” or HF community had a water supply with high iron (Fe) levels in the source water with median 0.80 mg/L total iron levels. This was above the aesthetic guideline value of 0.30 mg/L of the Australian drinking water guidelines [22]. The “MidFe” or MF community had a water supply with medium iron (Fe) levels with median 0.25 mg/L total iron and the “LowFe” or LF community had low iron levels with a median 0.05 mg/L total iron. All three communities had reported melioidosis cases in the past (1994–2017: HF 3 cases (incidence rate IR 4.1 cases/1,000 population), MF 11 cases (IR 9.9) and LF 4 cases (IR 2.6)). It is not known where these patients acquired the melioidosis bacteria.
Samples were collected in the late wet season i.e. in March 2017 for two of the three communities (HF and LF) while the third community (MF) was sampled in May 2017 as soon as waters receded sufficiently to allow access to the water bore fields. For each community, samples were collected from five points along the DWDS of which three were unchlorinated (bores and tanks) and two from the chlorinated reticulation system.
Samples were collected from five sites from each of three water supplies. One litre of water was collected in duplicate for subsequent DNA extraction. An additional 500 mL were collected for B. pseudomallei culture (Menzies School of Health Research), 200 mL in duplicate in 200 mL sodium thiosulphate dosed bottles for subsequent faecal indicator, heterotroph and amoebae culture, 100 mL into acid-washed 125 mL bottles for elemental analysis and 100 mL of in situ filtered water (using 0.45 micron filters) into acid-washed 125 mL bottles for nutrient analysis. All bores that were sampled were in operation for >6 hours and bores were purged for five minutes prior to water collection. The surface of biofilms in the bore head, pipes, tanks, and water meter walls was collected in duplicates using sterile swabs (Interpath, Australia). All samples were kept on ice on the sampling day except water and biofilms for subsequent amoebae and B. pseudomallei culture which were kept at room temperature and protected from sunlight. A total of 60 water and biofilm samples were collected in duplicates from 15 water collection points of three water supplies.
A YSI meter (www.ysi.com) was used to measure various physicochemical factors in water namely pH, salinity, temperature, turbidity and dissolved oxygen (DO) content. A colorimeter was used to measure free chlorine levels of the chlorinated water. A redox meter calibrated with Zobell’s solution (YSI) measured the oxidation redox potential (ORP)–redox measurements were conducted for all samples on the same day of collection upon return to the laboratory.
E. coli, coliforms, P. aeruginosa, heterotrophs and free-living amoebae were cultured at the NATA accredited NT Government Dept. of Primary Industry and Resources laboratory and the Australian Water Quality Centre (AWQC) after overnight shipment of samples on ice (room temperature for amoebae). Culture of E. coli and coliforms was based on the Most Probable Number (MPN) method and Colilert-18 Defined Substrate Technology (DST) (AS/NZS 4276.21–2005) while culture of P. aeruginosa was by membrane filtration. Heterotrophic Colony Count was by pour plate method with incubation for 44 h at 36 C (AS 4276.3.1–2007). Culture for B. pseudomallei and near-neighbour Burkholderia was conducted at Menzies School of Health Research. Culture from 500 mL of water was based on membrane filtration (0.22 micron filters) followed by culture in Ashdown broth and agar as previously described [4]. Similarly, biofilm swabs were incubated in Ashdown broth followed by plating on Ashdown agar.
DNA extraction of six B. pseudomallei isolates was as previously described [23] and the genomes were sequenced on a Illumina HiSeq2500 platform (Illumina, Inc., San Diego, CA) at the Australian Genome Research Facility (AGRF).
Orthologous core single nucleotide polymorphism (SNP) variants were identified among 89 B. pseudomallei genomes from the Northern Territory using the default settings of SPANDx v3.2 [24] and the closed Australian B. pseudomallei genome MSHR1153 [25] as reference (N50 4,032,226 bp; 2 contigs; size 7,312,903 bp). A maximum parsimony phylogenetic tree was generated in PAUP* 4.0.b5 [26] based on 174,905 SNPs and rooted using MSHR668. Multi-locus sequence types (MLST) were assigned in silico using the BIGSdb tool which is accessible on the B. pseudomallei MLST website (http://pubmlst.org/bpseudomallei/). The following geographical and virulence genetic markers were extracted in silico using the Basic Local Alignment Search Tool (BLAST) [27] following previously published methods [28]: LPS A (wbil to apaH in K96243 [GenBank ref: NC_006350]), LPS B (BUC_3392 to apaH in B. pseudomallei 579 [GenBank ref: NZ_ACCE01000003]), LPS B2 (BURP840_LPSb01 to BURP840_LPSb21 in B. pseudomallei MSHR840 [GenBank ref: GU574442]), BTFC (lafU in B. pseudomallei MSHR668 [GenBank ref: NC_006350]), YLF (BPSS0124 in B. pseudomallei K96243 [GenBank ref: CP009545.1]), bimABm (BURPS668_A2118 in B. pseudomallei MSHR668 [GenBank ref: NZ_CP009545]), bimABp (BPSS1492 in B. pseudomallei K96243 [GenBank ref: NC_006350]) and fhaB3 (BPSS2053 in B. pseudomallei K96243 [GenBank ref: NC_006350]).
Elements (total Fe Mn Mo Mg K Ca S Ni Cu Zn) were measured at the Environmental Chemistry & Microbiology Unit (ECMU) (CDU, Darwin, Australia) by ICP-MS (AGILENT 7700ce, www.agilent.com)[29]. Dissolved nutrient analysis (TDN, NOx, TDP and DOC) of the filtered water was conducted at the laboratory of Queensland Health (www.health.qld.gov.au).
Within 24h of collection, water samples (1 L) were filtered (0.45 micron filters, Sartorius) and frozen until processed. DNA was extracted from filters and swabs using the FastDNA soil kit (MPBio, Australia) following the manufacturers’ instructions. Bacterial load was measured using a SYBR-based qPCR assay targeting the 16s rRNA gene with PCR primers 331-f and 797-r [30] and using the QuantiTect SYBR Green qPCR mix (Qiagen, Australia) resulting in a qPCR efficiency of 90%. The delta Ct method was used for relative quantification and a positive control was included in each run for inter-run comparisons. Five DNA extraction negative controls on filters (#3) and swabs (#2) with no water or biofilm added were also processed. The DNA was sent to the Australian Centre for Ecogenomics (ACE, https://ecogenomic.org/) for 16s rRNA gene amplicon sequencing.
Sixteen-s rRNA gene amplification and Illumina MiSeq sequencing was conducted at ACE using the Earth Microbiome Project 16s rRNA V4 515FB-806RB universal primers (FWD:GTGYCAGCMGCCGCGGTAA; REV:GGACTACNVGGGTWTCTAAT) targeting bacteria and archaea (accessed June 2017: http://press.igsb.anl.gov/earthmicrobiome/protocols-and-standards/16s/). These primers were extensively validated to minimize bias towards or against taxonomic groups; however, remaining preferential amplification of certain taxa cannot be excluded [31–33]. Sequences were processed to sequence variants (SVs) by ACE with the following pipeline. The software Trimmomatic was used for sequence quality trimming removing poor quality sequences with a sliding window of 4 bases and an average base quality above 15. All reads were hard trimmed to 250 bases, and any with less excluded. Reads were processed to SVs using the QIIME-2 workflow with default parameters and the DADA-2 algorithm [34, 35]. Taxonomic assignment of SVs was through BLAST+ using the reference database SILVA (www.arb-silva.de).
15,590 SVs were further processed using the R library Phyloseq. Due to the low biomass of many samples, special care was taken to exclude potential contaminant SVs such as from lab reagents [36, 37]. Seventeen SVs were excluded which consistently occurred in all five negative controls. A further five SVs were excluded which occurred in minimum two negative controls and showed a significant negative Spearman’s rank correlation with the bacterial DNA abundance based on 16s qPCR results (P<0.05). The R package “decontam” was used to exclude a further 93 SVs based on their occurrence in negative controls (prevalence method). Thus, a total of 104 SVs were excluded due to contamination concerns. A further 9,787 SVs (63%) were excluded as these only occurred in one sample (with duplicate water and biofilm samples collected per site). As a last step, 38 SVs were excluded as these were not assigned to either Bacteria or Archaea.
Nineteen of 60 samples (32%) were excluded due to low sequence counts (< 5,000 sequences) (7 chlorinated water, 6 unchlorinated water, 4 chlorinated biofilms, 2 unchlorinated biofilms). Negative control samples had sequence counts ranging between 334 and 1,840 sequences. A hierarchical cluster analysis was performed in Primer-E v7 (Plymouth, UK) to test whether samples clustered with negative controls. This was only the case for samples with low sequence counts which were excluded from any downstream analyses. The final dataset contained 5,411 different SVs and 41 samples. All remaining samples were rarefied to the lowest common sequence number per sample (5,259 sequences). Rarefaction curves indicated that with this cut-off, the richness of all chlorinated samples plateaued i.e. was reached; however, four of 16 unchlorinated biofilm samples (25%) and two of 12 unchlorinated water samples (17%) only reached approximately 70–80% of their SV richness (S1 Fig). To examine Mycobacteria counts and richness across primarily chlorinated samples, a lower rarefaction threshold of 2,000 sequences was adopted which allowed the inclusion of more chlorinated samples (22 of 24 chlorinated samples) while their full SV richness was reached at this threshold, which was still higher than the sequencing depth of all negative controls (S1 Fig).
The weighted UniFrac and Bray Curtis distance matrices were trialled on the rarefied and square-root transformed SV data. Non-metric multidimensional scaling ordinations (nMDS) on the weighted UniFrac matrices showed a high stress value (>0.2) and therefore, the Bray Curtis dissimilarity matrix on square-root transformed SV’s was chosen for subsequent analysis in Primer-E v7. PERMANOVA was used to test whether the bacterial composition differed between communities, sample type (water vs biofilm) or chlorination status. Sample sites along the DWDS were included as random factor nested in community and chlorination status. A distance-based test for homogeneity of multivariate dispersions (PermDISP) was conducted to test for differences in data dispersion between sample groups. A canonical analysis of principal coordinates (CAP) was performed to assess the predictive ability of the microbiota for sample type and chlorination status. A distance linear model and distance-based redundancy analysis (dbRDA) were performed to associate abiotic factors (water physicochemical factors, nutrients and metals) with the bacterial composition. Model selection was based on the lowest AICc and a combination of forward and backward step elimination. A negative binomial model on the non-rarefied data was applied in Phyloseq (DESeq2 library in R)[38] to identify bacterial taxa whose abundance significantly differed between sample groups. The False Discovery Rate (FDR) method was used to account for multiple testing. To compare the occurrence of SVs across sample groups, the R package labdsv was used and those SVs considered which occurred at least twice in the sample group. The R package Vennerable was used to generate Venn diagrams. Multiple regressions on the log transformed bacterial DNA abundance and negative binomial models on taxa sequence counts and SV richness (based on rarefied dataset) were performed in Stata-14 (www.stata.com) with standard errors clustered for sites and model residual diagnostics conducted. A result was considered significant if P<0.05 unless otherwise stated.
Sixteen-s rRNA gene sequencing data were submitted to the European Nucleotide Archive (PRJEB29497, ERR2882159 to ERR2882214). Accession numbers of B. pseudomallei whole genome sequencing data are in S2 Table.
Water in the HF community had the highest levels of various metals, nutrients and salts (0.21–0.26 ppt) (Fig 1A–1C, S1 Table) and a neutral to slightly alkaline pH (6.9 to 7.8). The MF community water had lower iron levels of 0.03 to 0.78 mg/L, generally lower metal and nutrient levels and was more acidic (pH 4.8 to 5.3). Water in the LF community was the least buffered with the lowest metal and nutrient levels of the three water supplies and also the most acidic with a pH of below 5 for all five tested water samples (S1 Table). For all three communities, total Fe levels were strongly correlated with total Mn levels (Spearman’s rho 0.91, P<0.001). The DO and redox levels reflected the water origin with oxygen-deprived groundwater sampled from the bores showing the lowest DO and ORP levels. Redox levels overall were lowest at the HF community indicating a more reducing environment which was also reflected by the more neutral pH and higher DOC levels.
In this study we analysed a snapshot of the microbiota in bulk water and biofilms in the source and distribution system of three remote communities in Northern Australia. Changes in the microbiota were associated most with changes in redox levels and dissolved oxygen followed by various metals and nutrients such as TDN or DOC. These parameters not only differed along the water treatment train but also between the water supplies. Indeed, the geochemistry of the groundwater varied considerably between the three water supplies which was also reflected by significant differences in the source water microbiota. Remarkably, only 2% of SVs in untreated source samples occurred in all three water supplies as opposed to 23% SVs shared between the supplies for the treated parts. Significant microbial differences in source water of water supplies driven by geochemical differences have previously been reported [41].
The water supply of community LF was fed by the deepest aquifer, had the newest constructed bores and the water contained the least amount of nutrients and was the most acidic. There were fewer and less diverse bacteria in the water and while biofilm growth in the pipes was minimal, this was the only water supply with a significantly higher richness in the untreated biofilms compared to untreated bulk water. The scarce biofilms of all three tested LF bores were rich in chemo-organotrophic Acidobacteriaceae. Bacteria from these taxa typically have a growth optimum at lower pH and have adapted strategies to grow in low-carbon environments [41–43].
There were significantly more metals, bacteria and archaea in the bore water of the MF water supply. The MF bore heads were covered in thick loose iron flocs and biofilms and contained abundant Gallionella bacteria [44–46]. Ideal conditions for Gallionella growth have been reported to be at a neutral to slightly acidic pH, with a redox potential of 200–320 mV [47] matching conditions in the MF water supply. The bore field of the MF community is often inundated during the wet season during which contamination with surface water is possible. Two of the three tested bores were fed by a shallow aquifer and water from these bores grew heterotrophic bacteria. In contrast, the third bore accessed the Kombolgie sandstone aquifer at 20 meters depth, only had scarce heterotrophic growth but its water and biofilm samples were positive for B. pseudomallei. It is not known whether B. pseudomallei indeed occurred in the deeper aquifer. B. pseudomallei is a facultative anaerobe and hardy bacterium able to survive even in distilled water [16, 48]. Alternatively, the bore could have been contaminated with surface water during the wet season although the scarce heterotrophic growth did not support the notion of a recent contamination. Previous research has shown that B. pseudomallei inhabits shallow unconfined aquifers [49–51] and B. pseudomallei has been found more often in residential bores with hard water, acidic pH, increased iron levels and turbid water also containing coliforms [14]. These are indicators for surface water influx or water from shallow seasonal inter-flow aquifers. More research is needed to establish the potential occurrence of B. pseudomallei also in deeper aquifers which would be more difficult to manage by water providers.
Free-living Hartmannella amoebae were also recovered from the B. pseudomallei positive biofilm. Similar to other opportunistic pathogens, B. pseudomallei is able to survive within amoebae as shown in laboratory experiments [52, 53]. Survival within amoebae increases the pathogens’ resistance to chlorination [53].
Community HF was built in a coastal swamp area with shallow unconfined aquifers. Accordingly, the groundwater was buffered with the highest levels of various nutrients and metals. Untreated samples had the largest microbial richness and the pipes were covered in biofilms and iron deposits of a firm and scaly nature. Water from HF had a lower redox potential indicating a more reducing environment and organic carbon levels were high. Consequently, despite the high iron levels no Gallionella were recovered but instead dissimilatory iron reducers or nitrate-dependent anaerobic iron oxidizers like Geobacter, Azospirillum or Ferrovibrio [54]. The sulphur oxidizing Thiobacillus or Thiothrix were also detected. Bacteria of these genera cause biogenic sulphuric acid corrosion of concrete and they produce sulphates used by sulphate reducing bacteria such as Desulfovibrio or Desulfobulbus, both of which were also found at HF and less so at MF [45]. Sulphate reducing bacteria are involved in anaerobic corrosion or pitting of iron or steel by producing hydrogen sulphide and promoting anaerobic iron oxidation [55]. The untreated tank and rising main of the HF water supply showed abundant microbial life which flourished in the warm nutrient-rich water with high heterotroph counts, coliforms and Hartmannella and Naegleria lovaniensis amoebae feeding on the bacteria. It was of interest that these samples were also rich in Sphingomonadaceae bacteria which have been identified as an abundant member of the intra-amoebal microbiota in drinking water [56]. They have also been described in biofilms of chlorinated parts of water supplies and may be a reservoir of antibiotic resistant genes [57]. Water and biofilms of the tank also grew P. aeruginosa, an opportunistic pathogen primarily known for its pathogenicity in nosocomial settings and potential spread of antibiotic resistant genes in water distribution systems [58, 59]. In contrast to P. aeruginosa, there were no B. pseudomallei detected in the tank. Instead, B. pseudomallei was cultured from the shallow HF bore. Similar to the B. pseudomallei positive bore at MF, this bore only had scarce heterotrophic growth. Heterotrophic microbes require organic carbon for growth and HPC are routinely used by water providers to monitor the integrity of the supply and to indicate surface water contamination or presence of biofilms [22]. In this study, increased HPC did not match the presence of B. pseudomallei.
Genome analysis of the B. pseudomallei isolates revealed the presence of the YLF gene cluster and fhaB3 gene in isolates from the MF bore. The YLF cluster is more common in B. pseudomallei from Southeast Asia and remote parts of the Northern Territory [60, 61] while fhaB3 has been associated with B. pseudomallei positive blood culture as opposed to localized skin lesions [61]. LPS type B was found in B. pseudomallei from the HF supply together with the bimA-Bm gene. Both these genetic markers are more common in B. pseudomalllei from remote NT and bimA-Bm is also more widespread in Southeast Asia [62]. The bimA-Bm gene has been associated with neurological disease [61]. A phylogenetic tree with the water supply and other NT isolates showed no closely related B. pseudomallei isolates of clinical origin.
Nitrifying Nitrospiraceae were abundant in the untreated biofilms. Their production of nitrates provides a source of nutrients increasing biofilm mass [63]. Most nitrifiers identified in this study belonged to the genus Nitrospira common in drinking water with a preference for low nutrient or low nitrite environments [41, 64]. It was of interest that nitrate producing Nitrospiraceae were associated with B. pseudomallei positive samples. B. pseudomallei is a denitrifier under anaerobic conditions and in one study, B. pseudomallei load increased in sand upon nitrate treatment while in another study, B. pseudomallei was associated with soil containing elevated total nitrogen [65, 66]. More research is needed to further explore this potential commensal relationship.
Chlorination successfully contained B. pseudomallei and P. aeruginosa and reduced nuisance organisms. Similar to other studies, water treatment had the largest impact on the microbiota [67, 68]. The largest reduction in bacterial richness was observed for the MF water supply. Water disinfection of this water supply also included UV treatment apart from chlorine gas. Gammaproteobacteria were more abundant in chlorinated samples across all water supplies and members of this taxa are more resilient to higher chlorine levels and oxidative stress compared to Alpha- and Betaproteobacteria [69, 70]. One chlorinated site of the MF water supply had abundant DNA of several sequence variants of another group of opportunistic pathogens, called non-tuberculous mycobacteria. Further investigations are needed to establish whether these were from viable bacteria. Environmental mycobacteria are known to persist in water supplies and can cause disease in immunocompromised people or people with chronic lung disease [71].
Due to the low biomass of many samples in this study, the inclusion of several negative controls proved crucial. Various sequence variants in chlorinated samples were also detected in negative controls such as those of Ralstonia or Pseudomonas. This made it difficult to differentiate between hardy bacteria persisting in various environments including chlorinated water or mere contaminants of laboratory reagents and DNA extraction kits [36, 72]. As outlined in the methods, utmost care was taken in excluding samples with low sequence numbers and/or similarity to microbial fingerprints of negative controls and excluding potential contaminant sequence variants. Subsequent studies will use larger water sample volumes and filters with smaller pore size to increase biomass and ensure capturing microbes of all sizes [70].
Overall, there were no significant differences in the microbiota between bulk water and biofilms; this was particularly the case for the turbid water of the MF supply with a high level of suspended solids. Swabs were used to collect biofilms which primarily captured the top layer of biofilms or microbes associated with suspended solids and loose deposits as opposed to other studies which scraped the biofilm off pipes or grew them on coupons inserted into pipes [73]. Nevertheless, we found untreated biofilms to be more heterogeneous than planktonic microbiota with a distinct microbial fingerprint for each water supply. Sequence variants of various nitrifying families were more common in untreated biofilms compared to untreated bulk water as previously reported [74]. Once the water was treated, the microbiota indeed differed between water and biofilms and the proportion of SVs unique to biofilms also increased while the proportion of SVs shared between the sample types decreased. This matches previous reports of an increase in differences between sample types upon water treatment [67].
In summary, we found that the geochemistry of the source water had a substantial impact on the untreated microbiota with largely different microbial communities in untreated parts of the three water supplies. Accordingly, a multiple barrier approach to improve water quality would have to account for the heterogeneous nature of the microbiota in different water supplies across Northern Australia. We detected three opportunistic pathogen groups; namely non-tuberculous mycobacteria, P. aeruginosa and B. pseudomallei. In contrast to our working hypothesis, B. pseudomallei was cultured from a bore accessing a deeper aquifer and future investigations across seasons will determine whether B. pseudomallei indeed occurs in deeper confined aquifers or is mainly linked to surface or shallow aquifer water intrusions during the wet season, with the latter easier to manage for a water provider. Similar to other opportunistic pathogens in water supplies [20], B. pseudomallei was cultured from bulk water with low organic carbon and scarce heterotrophic growth. This matches its ability to thrive under nutritionally poor conditions [16, 48] but also indicates that HPC routinely used by water providers to monitor the supply integrity is a poor indicator for B. pseudomallei presence. We also detected B. pseudomallei in a multi-species biofilm linked to iron bacteria. Further research is needed to examine these interactions as Gallionella is increasingly used in biological iron-removal filters. This study provided a first snapshot of the microbiota in a selection of remote water supplies informing future studies to ultimately improve management guidelines for water supplies in the wet-dry tropics.
|
10.1371/journal.pcbi.1000821 | Beauty Is in the Eye of the Beholder: Proteins Can Recognize Binding Sites of Homologous Proteins in More than One Way | Understanding the mechanisms of protein–protein interaction is a fundamental problem with many practical applications. The fact that different proteins can bind similar partners suggests that convergently evolved binding interfaces are reused in different complexes. A set of protein complexes composed of non-homologous domains interacting with homologous partners at equivalent binding sites was collected in 2006, offering an opportunity to investigate this point. We considered 433 pairs of protein–protein complexes from the ABAC database (AB and AC binary protein complexes sharing a homologous partner A) and analyzed the extent of physico-chemical similarity at the atomic and residue level at the protein–protein interface. Homologous partners of the complexes were superimposed using Multiprot, and similar atoms at the interface were quantified using a five class grouping scheme and a distance cut-off. We found that the number of interfacial atoms with similar properties is systematically lower in the non-homologous proteins than in the homologous ones. We assessed the significance of the similarity by bootstrapping the atomic properties at the interfaces. We found that the similarity of binding sites is very significant between homologous proteins, as expected, but generally insignificant between the non-homologous proteins that bind to homologous partners. Furthermore, evolutionarily conserved residues are not colocalized within the binding sites of non-homologous proteins. We could only identify a limited number of cases of structural mimicry at the interface, suggesting that this property is less generic than previously thought. Our results support the hypothesis that different proteins can interact with similar partners using alternate strategies, but do not support convergent evolution.
| Interaction between proteins is a fundamental process, generic to most biological pathways. The increasing number of protein–protein complexes with atomic data should help us to understand the major factors that guide protein interactions. In particular, a number of examples are available of similar proteins that interact with proteins that are very different in terms of structure and function. An intuitive hypothesis to explain the ability of these different proteins to recognize the same partner is that they display the same local region for interaction, in other words, they imitate the same binding site. Here, we quantify the similarity between these putatively mimicking binding sites. We show that it is not statistically significant. We confirm this observation on the small sets of evolutionarily conserved residues. Our results suggest that different proteins that bind the same protein do not imitate binding sites, but probably target specific locations or residues at the binding site.
| Protein-protein interaction is the basis of numerous biological functions, such as immune response, supra-molecular assembly, enzymatic reactions, and many more. Understanding the way proteins interact is thus a fundamental challenge. The collection of all protein-protein interactions, the interactome, is also of great importance for drug discovery [1]. Given their variety and often transient nature, the number of protein-protein complexes for which crystallographic structures are available is very limited compared to the number of individual protein structures in the Protein Data Bank [2]. But even with this limited amount of data, the observation of available complexes has helped to decipher some rules for protein-protein interactions. Among the properties playing a role in this process, hydrophobicity was suggested as a major factor by Chothia and Janin in their pioneering work [3]. Other characteristics that are important for interaction, or that can be used to describe binding sites, include size, shape complementarity, residue propensity and packing density [4]–[6]. Sequence conservation is also widely acknowledged as an important feature of protein-protein recognition [7], [8]. Additional studies have further refined the picture. For example, binding sites are organized as a core of buried residues, surrounded by a rim of accessible residues, with distinct amino-acid composition and evolutionary conservation patterns [9], [10]. Nicola and Vakser found that the binding site is, on average, closer to the center of mass of the protein compared to other surface residues [11]. Different types of complexes (e.g. homo- or hetero-dimers, transient or permanent) display different properties [8], [12], [13].
A notable element to understand the mechanism of protein-protein interaction is the existence of hot spots, residues that make major contributions to the binding energy, see for example [14] for a review. In their landmark paper, Bogan and Thorn showed that hot spots are localized at the center of interfaces, and surrounded by a ring of energetically unimportant residues, that protect them from the solvent [15]. This is called the O-ring theory, and has been recently refined by Li and Liu [16].
Several groups have addressed the question of the evolutionary conservation of protein-protein binding sites and binding modes. At first found to be insignificant [17], the conservation of interface residues has since been shown to be more pronounced in biological interfaces than in crystallographic ones or over the rest of the protein surface [18], [19]. This change of viewpoint probably comes from the increase of available data, as well as the variety of computational approaches developed to quantify conservation, and also the fact that some proteins have multiple interfaces [20]. The link between evolutionary conservation and hot spots is unclear: overall difference in conservation between hot spot and non hot spot residues is marginal [21], [22]; conservation used in combination with other features has been found to improve hot spot prediction in [21] but not in [22]. From a more macroscopic point of view, complexes that share more than 35% identity commonly share similar structures and interaction modes [23]. The localization of a binding site on a protein is preserved within SCOP families, but not necessarily at the super-family level [24], [25].
Another important notion we want to introduce here is the existence of promiscuous proteins. Promiscuity, also called multi-functionality or moonlighting, denotes the ability of one protein to perform distinct functions, see reviews [26], [27]. A recent review reveals that promiscuity is not as rare as previously thought [28]. Examples notably include transcription regulatory proteins that can act as transcription coactivators or enzymes [29]. More generally, a promiscuous protein can interact with different partners. These multi-partner proteins have been the subject of dedicated studies. For example, Keskin et al have shown that multi-partner protein interfaces have original properties: they are smaller and less packed than other interfaces [30]. A recent survey of proteins with multi-binding protein interfaces involving 97 pairs of complexes from 49 protein families revealed that multi-binding interfaces are not more conserved than other interface sites [31]. The energetic determinants of multi-partner proteins have also been addressed: interactions involving specific binding sites display higher affinities than those of promiscuous binding sites [32]. In an earlier work, Humphris and Kortemme employed a computational design procedure to optimize the binding site of 20 multi-specific proteins, so that they maintained interactions with all their known partners (multi-constraint protocol) or with each partner separately (single-constraint protocol) [33]. For half of the tested cases, they obtained different results using the single and the multi-constraint protocol, suggesting that promiscuous binding sites are optimized for multi-specificity in such a way that each partner prefers its own set of residues on the binding site. A recent analysis using state-of-the-art computational methods applied on calmodulin, whose structure is available in complex with 16 different targets, confirmed this hypothesis [34]. These analyzes focused on the common, promiscuous binding sites, but not on the binding sites of the multiple partners.
The fact that a promiscuous protein can bind to different partners using the same binding site is puzzling, but also of outstanding interest to further understand the mechanisms of protein-protein interactions. Does this observation imply that radically different proteins possess similar binding sites in order to recognize a single promiscuous protein? At first sight, it might seem hopeless to look for similar binding sites on non-homologous proteins that differ in structure, function and ancestry. However, the literature is rich in examples of approaches employing - or searching for - such local similarities between unrelated proteins. This is the case for at least three distinct targets: catalytic sites, ligand binding sites and protein-protein binding sites. In the case of catalytic sites, the well-known example of the catalytic triad pattern, found in diverse serine proteases, has motivated a number of developments [35]–[40]. Concerning ligand binding sites, their generic nature among unrelated proteins has lead to the development of many comparison approaches [41]–[50]. Lastly, for protein-protein interactions, the similarity between proteins with very different folds has been investigated in several studies. An important corpus of work on this problem comes from Nussinov and colleagues. Using geometric hashing, they created clusters of similar interfaces based on the C geometry [51] and found clusters with similar interfaces despite different overall structures, as well as clusters where only one side of the interface was conserved [52], [53]. Shulman-Peleg et al. subsequently developed the I2I-SiteEngine software, dedicated to structural alignment of protein-protein interfaces, based on the similarity of their physico-chemical properties and shapes [54], [55]. These observations have been applied to the prediction of protein-protein interactions, with the development of the PRISM database [56], [57], and to structural alignment of protein-protein interfaces, with the MAPPIS web server [49]. Other groups have also investigated this question. Zhu et al. proposed the Galinter method, based on the representation of interfaces by vectors representing van der Waals interactions and hydrogen bonds between protein chains, allowing binding site comparison using graph algorithms [58]. Very recently, Konc et al. have proposed ProBis, a graph-based method for binding site prediction [59]. Convergent evolution thus seems to exist also for protein-protein interactions [60], [61].
In this paper, we analyze a set of protein-protein complexes involving homologous proteins in interaction with different partners. These examples come from an analysis of PDB complexes in terms of SCOP domains, and are stored in the ABAC database [61]. Truly speaking, these complexes do not illustrate promiscuity, since they involve homologous (same SCOP family) rather than identical proteins. We therefore term this promiscuous binding at the family level. Our goal is to understand how unrelated proteins can bind to similar targets. In particular, we looked for similar atoms or groups of atoms at the interface of different proteins that bind similar partners and assessed the significance of the similarity between interfaces using a bootstrap procedure. We also considered evolutionarily conserved residues, as they probably play a dominant role in the binding. Our results support the hypothesis that different partners often interact with a single partner using alternate strategies, and do not point to convergent evolution.
The overall methodology used to assess the similarity at protein-protein interfaces is summarized in Figure 1 detailed in the Materials and Methods section.
The ABAC pairs are classified into five categories on the basis of the quality of the superimposition between the two complexes, as illustrated in Figure 2. The first two categories, O and M (see Figures 2A and B) represent ideal cases to study promiscuous binding at the family level, with A/A′ domains having very similar structures which are easily superimposed. These two categories, encompassing 299 ABAC pairs, will be privileged in analyzing the similarity of binding sites, since the interfaces of A/A′ domains are well superimposed and the subsequent analysis of B/C binding sites is thus expected to be less noisy. Furthermore, the M category has the interesting particularity of exemplifying interface mimicry: domains B and C, although they have different global folds, display strikingly similar structures at the interface. It should be noted that among the 53 ABAC pairs in the M category, only 3 different SCOP families of A/A′ domains are represented, see Table 1. Eukaryotic proteases (family 50514) are seen in 49 pairs, subtilisin-likes (family 52744) in three pairs, and interleukin 8-like chemokines (family 54118) in one pair. Pairs of the category M are thus largely dominated by eukaryotic proteases complexed with various inhibitors, which, as shown in Figure 2 B, display a protruding/interwound geometry, with the B/C mimicry interfaces embedded in the A/A′ domain. This introduces a significant bias in interface size, with more residues involved in the interface on the A/A′ side than on the B/C side, see Figure 3 and Table 3 in Text S1. The three other categories, E, I and S (see Figures 2 C, D and E), illustrate three degrees of difficulty in A/A′ superimposition, with, respectively, alternate conformations in the binding site, residue insertion/deletion in the binding site, and overall poor structural similarity, which might alter the analysis of interface similarity.
In the rest of the paper, we present a quantitative analysis of similarity at protein-protein interfaces in ABAC pairs, and then evaluate its significance against a random model. We also survey the similarity of interfaces in terms of evolutionarily conserved residues.
We first compute the number of similar elements - atoms, pseudo-atoms or residues - in each partner of the protein complexes after structural superimposition of the common partners A and A′. Domains A and A′ are from homologous domains from the same SCOP family. Consequently, we expect a good level of similarity between them. However, since such similarity results from divergence from a common ancestor and fold conservation, it does not necessarily imply that the similar elements are key determinants for the protein-protein interaction. Domains B and C are from different SCOP superfamilies. They thus have very different structures, but a common ability to bind to the same, or, at least, a similar partner. Similar elements between B and C could thus be a sign of evolutionary convergence to a given binding motif, or indicate which functional groups are essential for the binding.
Figure 3 presents the number of superimposed and similar elements at the interface in the 433 pairs of complexes, and the ratio of similarity, with different interface representations (separate Figures for each category are given in Figures 4 to 8 in Text S1). For each ABAC pair, the number of superimposed and similar elements is computed separately for each domain, and we compare the statistics on the homologous sides (A and A′) versus the non-homologous sides (B and C) of each complex. Each ABAC pair is thus represented by two points: one for complex AB and one for complex A′C. We previously checked that the sizes of the binding sites on A/A′ and B/C sides are roughly similar (see Figure 3 and Table 3 in Text S1), which is true, except for complexes of the M category, due to their protruding/interwound geometry as illustrated in Figure 2.
As expected, there is a positive correlation between the number of superimposed elements - defining the size of the overlap - on the A/A′ domains versus B/C domains, see Figures 3A, D and G, resulting from geometrical considerations. The number of superimposed elements is almost always lower on the B/C side than on the A/A′ side, for every interface representation. This is due to the fact that the structural superimposition is guided by domains A and A′, which favours better overlap on the A/A′ side, as illustrated in Figure 4. This bias introduced by the superimposition results in a mean ratio of overlap sizes equal to 1.3–1.8, depending on the interface representation: for 100 elements superimposed on the B/C side, there is an average of 130 to 180 elements on the A/A′ side (statistics for each pair category are presented in Table 4 in Text S1). Because of this effect alone, the number of similar elements on B/C sides is expected to be lower than on the A/A′ sides. It can be seen, in Figures 3B, E and H, that the number of similar elements on the B/C side is effectively lower. The mean numbers of similar elements for the five categories are given in Table 2. The mean ratio is around 2 for all-atom and coarse-grain representations and 3 for residues: there is, on average only one similar residue on the B/C side for 3 residues on the A/A′ side. Interestingly, the correlation between the similarity ratios, i.e., number of similar elements normalized by the number of superimposed elements (see Figures 3C, F and I) is lower. For example, the Pearson correlation coefficient between the numbers of similar atoms (see Figure 3B) is equal to 0.8, versus 0.4 between the corresponding similarity ratios (see Figure 3C). In other words, a greater similarity between A/A′ interfaces does not automatically correspond to a greater similarity between B/C interfaces. It thus seems that the low level of similarity in B/C domains is not only the result of the superimposition bias, but reflects a real sparsity of common binding determinants in different proteins that bind to similar partners. Indeed, some ABAC pairs with very similar common domains can exhibit very low similarity on the B/C sides. As an example, when complex 1m4u_BA (human bone morphogenetic protein-7 complexed with noggin) is compared with complex 1nys_DC (human activin A complexed with rat activin receptor) 11 out of 16 superimposed residues are similar for the A/A′ domain, and only 2 residues out of 9 for the B/C domain. Similar binding sites can thus bind two proteins that present a very restricted set of similar residues. To go further with this analysis, we computed similarity P-values as explained in the Materials and Methods section.
Similarity P-values, computed using a bootstrap procedure, are presented as histograms in Figure 5 for the ABAC pairs of category O.
A P-value equal to x% means that in x% of the randomly sampled interfaces, the number of similar elements is greater or equal to the number of similar elements in the real interface. Consequently, a high P-values indicates that the similarity has a high probability to occur by chance. Inversely, a very low P-value means that the similarity is significantly higher than expected with a random model. A value of 5% is classically used as the significance cut-off.
It is clear from Figures 5A and 5B that the distribution of similarity P-values is very different between A/A′ and B/C sides, with a bias toward low P-values on the A/A′ sides, and high P-values on the B/C sides. For A/A′ interfaces, we intuitively expect low P-values, indicating a significant similarity, since A and A′ domains belong to the same SCOP family and share a common ancestor. This is the case, see Figure 5A. What is less expected, is that the P-values for the B/C sides are rather high, indicating that the similarity between binding sites of the B and C domains is, most of the time, insignificant, see Figure 5B.
We note that the all-atom model (see Figure 5A) can however result in high P-values for A/A′ domains. This can be due to the background model used for bootstrapping, in which the atom type labels are randomly re-distributed among atom positions. In an all-atom representation, atoms of the same type appear as clusters, simply because they are part of the same amino acid. Such a random model is thus not optimal, because it neglects this aspect. Furthermore, with a distance cut-off equal to 3 Å to detect similar superimposed points, several atoms can be matched by the same point after superimposition. The result is an artificially high number of random similar points, and consequently, high P-values. Another source of error, with a probable significant impact, is the inherent sensitivity of the all-atom model to side chain flexibility. Since the same side chain, upon binding to multiple partners, might undergo different conformational changes, the all-atom model might under-estimate the real level of similarity. For these reasons we considered coarse-grain and C representations only in the following analysis.
As shown in Figures 5C and D, the coarse-grain representation overcomes the high P-value artifact on the A/A′ side. On the B/C histogram, however, a number of complexes still display high P-values, meaning that the similarity level is not significant compared to random. This holds true using a C representation, see Figure 5E and F. We obtained similar results for other categories of ABAC pairs (Figures 4 to 7 in Text S1), although with more noisy results (less significant P-values on the A/A′ side) for the E, I and S categories, as expected due to the difficulty of the structural alignments for these categories.
We next considered the restricted set of evolutionarily conserved residues detected using the ConSurf database (as explained in the Materials and Methods section) and analyzed the interface similarity in this light. More precisely, we repeated the same analysis as for the C representation, but instead of considering five classes of residues, we labelled the residues by their conservation status, i.e., conserved or non-conserved. Then, we considered only the conserved residues at the interface, to see if they are co-localized with conserved residues after domain superimposition. As before, we computed separately the number of conserved residues superimposed on the A/A′ interfaces and the B/C interfaces, and the corresponding P-values. The P-value histograms follow the same trend as for binding site similarity: low P-values on the A/A′ side, but not on the B/C side, see Figures 8 and 9 in Text S1. Note that a considerable number of protein domains have no superimposed conserved residues in their binding sites, limiting the P-value analysis to a more restricted data set.
As shown in Figure 6, interfaces are only partially overlapping after structural superimposition of A/A′ domains. We thus cannot exclude that some residues located outside the overlap play dominant roles in the binding. The correlation between the fraction of similar atoms and the fraction of atoms that are overlapping is weak but positive (see Figure 21 in Text S1). The fact that binding sites with a small fraction of similar atoms tend to have a small fraction of binding site overlap (meaning that a significant proportion of the binding site is excluded from the comparison) suggests that key binding determinants could indeed be missed.
In the same way that there is a limited number of protein folds, it is tempting to speculate that there is a limited number of protein-protein binding interfaces [62]. Since protein structures are made of recurrent local conformations, i.e., -helices and -strands, protein-protein interfaces might be made by the assembly of recurrent binding modules. The present study was motivated by the search for such modules. Indeed, the fact that unrelated, dissimilar proteins are able to bind similar, homologous proteins suggests that common binding strategies might be re-used by different proteins. It is logical to look for generic binding modules in the promiscuous binding sites thus formed.
We were not however able to confirm this hypothesis. Starting from a discrete physico-chemical model, in which interfaces are described by points - be they atoms, pseudo-atoms or residues - belonging to five different classes, we found that, in most of the cases, the similarity between different proteins that bind to homologous partners is not greater than random (but the similarity between the homologous partners is significant, suggesting that the random model is appropriate). It thus seems that protein interfaces with no detectable similarity can nevertheless bind similar partners.
We should temper this result by noting that the energetic contribution of interfacial residues is uneven; some hot spot residues make major contribution, while other residues are unimportant. Unfortunately, energetic information - requiring extensive mutation analysis - is not available for our full data set, we thus approached this particularity in an indirect way. Although evolutionary conservation is a poor discriminant of hot spots [21], [22], it has been shown to improve the prediction when used in combination with other features [21]. Conserved residues do not translate into hot spots but might contain some information. We thus considered conserved residues at protein-protein interfaces, and assessed their co-localization in our complex pairs. This time, the criteria was not to know if superimposed residues are from the same physico-chemical class, but to know if they are both conserved during evolution, independently of their class. The rationale was to restrict the analysis to the subsets of conserved residues. The co-localization of conserved residues in different proteins that bind homologous partners was found to be largely insignificant. Further studies using in silico hot spot prediction methods could bring additional information.
Altogether, our results suggest the following picture for promiscuous protein-protein binding: similar, homologous proteins present binding sites with great similarity, via which they interact with diverse, dissimilar proteins. The binding interfaces of these dissimilar proteins exhibit different atomic/residue patterns, and their conserved residues are not co-localized. It thus suggests that different proteins use their own set of atoms/residues to perform the recognition, as illustrated in Figure 7A. There is also the possibility that atom groups interacting specifically with a single partner could play a dominant role, i.e., different partners use residues or group of residues that are outside the overlap between the two binding sites, see Figure 7B. The mechanism illustrated in Figure 7A is in agreement with the elegant work of Humphris and Kortemme, who have shown that multi-specific binding can be achieved by different mechanisms [33]. Using computational design to “optimize” the interfaces of promiscuous proteins, they observed two distinct patterns: (i) for half of the tested case, all partners shared key interactions; (ii) for the other half, each binding partner preferred its own set of wild-type residues in the common binding site. Some experimental studies of promiscuous proteins support this second pattern. For example, TRAF3 (Tumor Necrosis Factor Receptor-associated Factor) is able to bind two targets, CD40 and Lymphotoxin- receptor, at the same interface, although they present motifs with distinct sequence and structure motifs for the binding [63]. Another example of promiscuous protein is protein kinase A, which is able to bind to different proteins using the same binding site. Entropy calculations suggest that the binding site of protein kinase A provides alternative contact points for the partner side chains [64]. In a recent study of BirA, a protein able to form a homodimer as well as heterodimer using the same binding site, hot spot residues were identified specifically for the homodimerization, but not for the heterodimerization [65]. This suggest that each complex forms using its own preferred and distinct interactions. This has also been observed for protein/ligand complexes. For example, different non-peptidic haptens have been shown to bind to the same site of an antibody, by forming different hydrogen bonds, dependent upon their particular chemistry and the availability of complementary antibody residues [66].
A last point to discuss is the existence of structural mimicry at interface. Protein mimicry is an intuitive concept, that has been successfully used in rational design [67]. Examples of protein interface mimicry - present in our data set - include several chymotrypsin inhibitors with various global folds (49 ABAC pairs), the viral protein M3 that mimics the binding site of chemokines for homodimerization (1 ABAC pair), and different subtilase inhibitors (3 ABAC pairs). Surprisingly, the similarity P-value analysis of these 53 pairs revealed that the physico-chemical similarity of the mimicking binding sites is not significant. However, their structural similarity is obvious, see Figure 2. This might indicate that the shape - not taken into account by our atomic or residue-based representations - is an important determinant for interface mimicry. Indeed, local surface comparison has been successfully used to retrieve chymotrypsin inhibitors [68].
The present study focused on promiscuous binding at the family level. The goal was to find the key determinants that allow unrelated proteins to bind to homologous partners. Our main conclusions are summarized below.
We were not able to find evidence of convergent evolution. Our results support the hypothesis that promiscuous binding is rather achieved by alternative binding strategies for different partners.
We exploited the data from the ABAC database (http://scoppi.biotec.tu-dresden.de/abac/) that contains protein-protein complexes organized in pairs [61]. As illustrated in Figure 4, ABAC pairs are formed by homologous proteins, A and A′, in interaction with non-homologous proteins B and C at equivalent binding sites. The SCOP classification [69] was used to ensure that A and A′ belong to the same family and B and C to different super-families. SCOP families gather proteins that have a clear evolutionary origin, measured by a sequence identity greater than 30%, or lower sequence identity, but very similar structure or function. At the superfamily level, proteins display low sequence identity, but structures and functions suggest that they are evolutionarily related. Proteins classified in different superfamilies are unrelated. Pairs with equivalent binding sites were selected after a two-stage procedure involving an assessment of interface residue overlap on A and A′ sequences and spatial overlap between A/B and A′/C interfaces measured by the angle between the center of mass of A/A′, and the center of mass of the interfacial region of B and C [61].
PDB files of protein-protein complexes were retrieved from the PQS database [70]. Starting from a non-redundant list of ABAC pairs with only one instance per SCOP family combination, we selected pairs that fulfilled two criteria: (i) the two partners are from different chains, i.e., we do not consider intra-chain interactions, (ii) SCOP domains spanning several protein chains involved in the binding site are excluded from the analysis for computational simplicity. We also removed complexes with missing atomic coordinates at the binding site, and pairs with very low overlap between the binding sites resulting in no superimposed atoms on the B/C side. Details concerning the minimum overlap size in the data set are given in Table 5 in Text S1. The final data set comprises 433 ABAC pairs. These 433 pairs were further classified into 5 categories, based on a visual assessment of the quality of the superimposition between A and A′ domains, particularly at the interfaces:
For the category M, the geometry of the main chain of B and C domains in the binding site was taken into account. Globally, O and M categories correspond to smaller rmsd between A and A′ domains, and smaller irmsd (rmsd between interfacial residues) compared to category S; categories E and I are intermediate; and categories overlap in terms of rmsd values, see Figure 2 in Text S1. Note that rmsd and irmsd are average values of structural deviation, hence they only reflect global tendencies; furthermore, they depend on the extent of the structural alignments. Also, irmsd computation does not take into account insertion of residues, because they are unaligned. Structural mimicry of B and C domains cannot be detected using rmsd, since domains B and C are unrelated and hence not superimposable. The classification thus ultimately results from a careful visual examination that takes into account all these parameters.
Our data set is non-redundant in the sense that every SCOP family combination is unique. However, the ABAC pairs are not independent, since the same SCOP family can be shared by several pairs. For example, the SCOP family 49504 (Plastocyanin/azurin-like) is shared by the A/A′ domains of two ABAC pairs:
Overall, 68 SCOP families are present in only one ABAC pair if we consider their A/A′ domains, and the most abundant family - family 52592, G proteins - is represented in 130 pairs. This probably indicates both the capacity of some particular families for promiscuous binding at the family level, but may also reflect the bias of structures deposited in the PDB toward proteins with biomedical interest. The number of distinct SCOP families, for A/A′ domains and B/C domains are reported in Table 1, for each category of ABAC pairs. It can be seen that the number of different SCOP families in A/A′ domains is 105 for the full data set. This apparent redundancy is not a limitation in our context, since we consider the similarity between pairs of complexes. In particular, considering ABAC pairs with unique SCOP domain combinations is enough to explore how different B/C domains interact with similar A/A′ domains.
Interfacial atoms were detected by applying a cut-off of 5 Å between heavy atoms from interacting chains, as in the SCOPPI database [71], [72]. Residues were considered to be part of the binding site if they had at least one interfacial atom.
Atoms were classified into five groups adapted from those proposed by Mintseris and Weng [73] (see Figure 1 in Text S1). These groups were determined by an optimization procedure, so as to maximize the mutual information of the pairwise matrix of atomic contacts at protein-protein interfaces. Although they have been determined by statistical optimization, they are in excellent agreement with biochemical criteria and roughly make the distinction between positively charged/negatively charged/polar/non-polar and hydrophobic groups of atoms.
As in [61], homologous partners of the ABAC pairs, i.e., domains A and A′, were superimposed using Multiprot [74]. After structural superimposition, interfacial atoms from A (resp. B) were considered as superimposed if there was an interfacial atom from A′ (resp. C) less than Å away, and similar if both atoms were from the same group. Cutoff was set to 3 Å, as in [61]. Note that this cut-off is used to compute the number of similar atoms between two binding sites after superimposition, and should not be confused with the cut-off equal to 5 Å that is used to detect atoms that are part of the interface.
Binding site similarity was also quantified on a per-residue basis, by representing each residue by its C. In addition, we considered an intermediate coarse-grain model introduced by Zacharias [75], in which residues - except GLY - are modeled by two or three pseudo-atoms: the C, and one side-chain pseudo-atom (residues ALA, SER, THR, VAL, LEU, ILE, ASN, ASP and CYS) or two side-chain pseudo-atoms (residues PHE, MET, PRO, TRP, HIS, TYR, GLN, GLU, LYS, ARG). Residues and pseudo-atoms were clustered into five groups, deduced from the atom groups (see Tables 1 and 2 in Text S1).
In order to take into account the fact that residues are described by a reduced number of points using these simplified representations, the cut-off to detect similar points after complex superimposition was empirically set to 4 Å for the C and the coarse-grain representations.
The significance of the similarity between binding sites was assessed by bootstrapping. The principle is to generate random binding sites by randomly re-assigning the atom types in the overlapping interfaces. The advantage of this re-sampling is that the sizes of the compared objects are preserved. The procedure was repeated 500 times in order to obtain the distribution of the number of similar atoms (or pseudo-atoms or residues) between two binding sites that can be expected with a random model. The extent of the observed similarity could then be assessed by computing the corresponding P-value, , where and denote respectively the number of similar atoms obtained between random binding sites, and observed between real binding sites. For each ABAC pair, we thus computed four P-values: one for each of A, A′, B and C binding sites.
Evolutionarily conserved residues were detected using the ConSurf database [76]. This database contains pre-calculated conservation scores, obtained after multiple alignment of homologous sequences using an empirical Bayesian algorithm [77]. For each residue of a protein, a normalized conservation score is assigned. Residues with normalized scores lower than -1 were considered as evolutionarily conserved. In some cases, when the number of homologous sequences is too low, the conservation scores were not available. In such cases, all residues were considered as unconserved.
During the comparison of binding sites, 131 comparisons out of 433 involved a binding site with no conserved residues when considering A/A′ domains, and 178 out of 433 when considering B/C domains. The analysis of evolutionarily conserved residues is thus inherently based on a smaller data set.
|
10.1371/journal.pcbi.1004798 | Extracting Behaviorally Relevant Traits from Natural Stimuli: Benefits of Combinatorial Representations at the Accessory Olfactory Bulb | For many animals, chemosensation is essential for guiding social behavior. However, because multiple factors can modulate levels of individual chemical cues, deriving information about other individuals via natural chemical stimuli involves considerable challenges. How social information is extracted despite these sources of variability is poorly understood. The vomeronasal system provides an excellent opportunity to study this topic due to its role in detecting socially relevant traits. Here, we focus on two such traits: a female mouse’s strain and reproductive state. In particular, we measure stimulus-induced neuronal activity in the accessory olfactory bulb (AOB) in response to various dilutions of urine, vaginal secretions, and saliva, from estrus and non-estrus female mice from two different strains. We first show that all tested secretions provide information about a female’s receptivity and genotype. Next, we investigate how these traits can be decoded from neuronal activity despite multiple sources of variability. We show that individual neurons are limited in their capacity to allow trait classification across multiple sources of variability. However, simple linear classifiers sampling neuronal activity from small neuronal ensembles can provide a substantial improvement over that attained with individual units. Furthermore, we show that some traits are more efficiently detected than others, and that particular secretions may be optimized for conveying information about specific traits. Across all tested stimulus sources, discrimination between strains is more accurate than discrimination of receptivity, and detection of receptivity is more accurate with vaginal secretions than with urine. Our findings highlight the challenges of chemosensory processing of natural stimuli, and suggest that downstream readout stages decode multiple behaviorally relevant traits by sampling information from distinct but overlapping populations of AOB neurons.
| Across the animal kingdom, chemical senses play a central role in guiding social behaviors by conveying information about particular behaviorally relevant traits. However, decoding these traits from profiles of chemical cues is challenging since cue levels are modulated by multiple factors. Here, we investigate how the mouse vomeronasal system, a chemosensory system important for processing social information, detects behaviorally relevant traits from natural stimuli. We focus on detection of a female’s genetic background (a model for individuality) and estrus-state (a measure of sexual receptivity) by neurons in the first vomeronasal brain relay, the accessory olfactory bulb (AOB). We show that information about both genetic background and receptivity can be obtained from various stimulus sources: urine, vaginal secretions, and saliva. Importantly, while individual AOB neurons can only provide limited decoding ability of these traits, simple networks sampling AOB neuronal ensembles provide considerable improvement. Our analyses highlight an overlooked challenge associated with chemosensory processing and suggest how it can be overcome by downstream neurons that read information from multiple AOB neurons.
| Social animals are extremely adept at extracting information about conspecifics and many species rely on chemosensory cues to achieve this goal [1–3]. Yet, deriving information about specific traits pertaining to other individuals can be complicated by several factors [4]. One factor is physical variability, for example due to stimulus source dilution. Namely, if a given trait is associated with a particular level of some compound, dilution or concentration of the stimulus source could confound correct trait detection [4, 5]. Another factor is the influence of multiple ecologically relevant traits on the levels of any one type of molecule. For example, analysis of mouse urinary compound levels as a function of genetic background and reproductive state indicates that both factors modulate the levels of all tested compounds. In some cases, even the direction of change as a function of reproductive state varies with the genetic background [6]. A third factor involves the stimulus source identity. Many animals, including rodents [7, 8], investigate multiple body regions of conspecifics and thus inevitably sample different secretions. Compound concentrations likely differ across secretions and interpretation of their content must thus account for the secretion sampled. Given these sources of variability, reliable detection of any trait becomes a significant computational challenge.
Here, we explore the neural representation of genetic background (a model for identity) and female estrus-stage (a measure of receptivity), two traits that are critical for a male mouse to guide reproductive behavior [2, 9, 10]. We study neuronal responses at the level of the accessory olfactory bulb (AOB), the first brain stage receiving vomeronasal inputs [11]. The vomeronasal system (VNS) is ideal for this study as its role in the detection of both of these traits is well established [9, 10, 12–16]. In particular, the two donor strains used here (BALB/c and C57BL/6), are clearly distinguishable by mice in the context of the vomeronasal-system mediated pregnancy block effect (The Bruce effect) [17, 18]. Furthermore, physiological recordings reveal that responses of vomeronasal sensory neurons, and of AOB neurons, are modulated by both the reproductive state [10, 12, 19–22], and the strain [7, 19, 23–25].
An important element in this study is the parallel investigation of multiple chemosensory stimulus sources: urine, saliva and vaginal secretions. While urine has been extensively studied as a chemosensory stimulus source in rodents [2, 22, 25–27], saliva and vaginal secretions have received much less attention. Here we show for the first time that neurons in the AOB respond to all these secretions in a strain and reproductive-state specific manner. However, we find a major difference between sensitivity to a certain trait and the ability to reliably detect it in the presence of multiple sources of variability. Thus, while many individual neurons can provide some information about these traits, any one neuron in isolation is generally insufficient to provide invariant information about them. We show, however, that integration of information across multiple neurons considerably improves trait detection, and that the detection does not rely on a small number of specialist neurons. Our study highlights the complexities of extracting socially relevant information from chemosensory cues, yet also suggests that relatively simple networks can overcome these challenges.
We initially tested whether AOB responses to urine, saliva and vaginal secretions can convey the strain and the state of the stimulus donor. Recordings were performed in anesthetized BALB/c males (BC), using multi-electrode arrays. The electrodes were targeted to the AOB external cell layer (Fig 1A) which contains the cell bodies of AOB projection neurons, known as AOB mitral/tufted cells [11]. Stimuli were collected from estrus and non-estrus adult BC and C57BL/6 (C57) female mice (Fig 1B and 1C). Fig 2 shows examples of twelve different units, each of which is modulated by the stimulus donor’s reproductive state (left), or its strain (right), for one of the three secretions.
While clearly selective to the females’ reproductive-state or strain, the examples shown in Fig 2 are limited to a single pairwise comparison, in which only one trait is varied with all other factors remaining identical. To investigate how invariant representations of strain and state can be attained, we assume the position of a decoder with access to AOB neuronal activity. Unless specified otherwise, we use a simple linear classifier—a perceptron [28]—whose goal is to discriminate among trait values. As inputs, the classifier accepts single-trial responses, defined as averaged firing rate changes calculated during the 40 s period following stimulus presentation. The rate change is defined with respect to the 30 s period preceding vomeronasal stimulation. Fig 1D shows one such single trial (highlighted by the red rasters), used to derive a single value characterizing the neuron’s response (indicated by bar on right side). The broad 40 s time-window was chosen to account for the slow and temporally variable responses of AOB neurons (see Fig 2). A classifier generally receives inputs from multiple neurons (Fig 1D), with single-trial responses sampled independently from each of the neurons. For each classification, we train the classifier with one set of single-trial responses (training set, see Materials and Methods), obtaining a set of weights and a bias term. Following training, classifier performance is tested on a different set of single-trial responses (test set). We first train the classifier with all units in the dataset, and then sequentially remove the unit assigned with the smallest absolute weight. Unless indicated otherwise, classification results denote an average over 10 training cycles. Repeated training cycles were conducted to account for randomness in the training and testing steps.
Vaginal secretions are likely to play an important chemosensory role during anogenital investigation of females [3, 29]. However, it is not known what type of behaviorally relevant information can be derived from this stimulus by the VNS. We thus first focused on decoding reproductive state from vaginal secretions. Our dataset includes 92 AOB units (38 single-units, 54 multi-units, 8 sites from 5 mice, Fig 1C set 1, see S1 Table) that responded to at least one of the 12 stimuli at the 0.05 significance level. The mode, median and mean number of significant responses per unit is 1, 3, and 4.0, respectively (See S2 Fig, which also shows that the fraction of responsive units increases with stimulus concentration). Recording site locations of the 92 responding units spanned the entire anterior-posterior aspect of the AOB, indicating that AOB neurons receiving both basal and apical vomeronasal sensory inputs [11, 30] were sampled (S1 Fig).
The normalized response profiles of all units included in this analysis are shown in Fig 3A. We began with the simplest discrimination, involving two reproductive states, while all other stimulus properties (dilution and strain) remain identical. In our dataset, with two strains and three dilutions, this amounts to six simple discriminations (indicated by the black lines in Fig 3B). The classification performance as a function of the number of units, averaged across all 6 pairwise comparisons, is shown in Fig 3C (mean classifier, solid black lines). The traces reveal that even with individual units, all six discriminations can be made with a very high success rate.
We next asked if a classifier trained on one dilution performs well on stimuli at a different dilution. For example, how does a classifier trained to distinguish estrus from non-estrus C57 vaginal secretions at 3x dilution perform on a 1x dilution of the same stimuli? The gray double-headed arrows in Fig 3B indicate the twelve possible tests of generalization. The arrows are double-headed to indicate that two reciprocal tests of generalization can be made with each pair of dilutions. The resulting classifications, whose average is shown by the gray line in Fig 3C, reveal that generalization across dilutions is only slightly better than chance. In this context, it is interesting to note that different urine concentrations were shown to convey different signals, via activation of distinct populations of vomeronasal sensory neurons [22]. See S3A–S3F Fig, for a more detailed description of classification results, and S4 Fig for a detailed examination of generalization across specific dilutions.
Do these observation imply that a classifier cannot perform well across different dilutions of a stimulus? To answer this, we next trained a classifier to discriminate estrus state across all vaginal secretion dilutions (separately for each strain), as illustrated by the orange lines in Fig 3D. Fig 3E shows that robust classification is indeed possible with a simple linear classifier, even across a range of dilutions. Yet, multiple neurons are needed to achieve maximal performance. Using the neuron removal approach, we observe that roughly 10 units allow classification with a success rate of about 95%. These analyses indicate that classification across dilutions requires explicit training and that polling multiple units provides a substantial benefit.
We next address the effect of changing one trait on the ability to discriminate another. Specifically, we test how the two dilution-invariant classifiers, each of which was trained on one strain, perform on stimuli from the other strain. These comparisons are indicated by the double-headed arrow in Fig 3D. The averaged classification performance is shown in gray in Fig 3E. Regardless of the number of units used for classification, generalization across strains yields near chance performance, revealing that the effect of one trait on another is indeed significant. Finally, we trained a classifier to discriminate reproductive state across both dilutions and strains (Fig 3F). Although the classifier’s performance plateaued at a lower level, and required more units than the less general classifiers, it clearly yielded above-chance performance (Fig 3G).
The unit removal approach often underestimates the performance of the best individual unit. We therefore also include this value, for comparison, in Fig 3C, 3E and 3G as a broken line (when multiple classifiers are present, the line represents their average). The thick solid line in each of these panels shows the maximal perceptron performance. This value represents the best classification performance over the 10 repeated training and classification cycles, with all units considered. Thus, comparison of the thick solid line with the dotted line in each panel shows the benefit of decoding with multiple units over that possible with any individual unit. While this advantage is minor for the simple classifications (Fig 3C), it is substantial for the more general classifications (Fig 3E and 3G). See S11 Fig for additional analyses showing the limited decoding capacity of individual units.
Our analysis thus far was based on one stimulus source (vaginal secretions) and one type of discrimination (reproductive state). We next expand the analysis to another discrimination (strain) and to another secretion (urine). The comparison across discriminations and secretions reveals some general principles but also some notable differences.
The urine dataset comprises 51 units (25 single units, 26 multi-units) that responded to at least one of the 12 urine stimuli (Fig 1C, set 2). Detailed descriptions of the number of units per session and recording site locations, spanning the entire anterior-posterior aspect of the AOB, are given in S1 Table and S1 Fig. For the 51 responding units, the mode, median and mean number of significant responses per unit was 1, 3, and 3.4, respectively (see S2 Fig, which also shows that the number of responsive units increases with stimulus concentration).
Complete descriptions for VS and urine classifications are shown in S3A–S3L Fig. Fig 4 summarizes the classification performance to facilitate comparison across the various cases. To fairly compare the VS and urine datasets (the latter containing fewer units), we consider performance with 50 units. The similarities are highlighted by noting that for both secretions and for both traits, discriminations involving more sources of variability result in poorer performance (panels 4A-D). Comparison of the left-side panels (4A, C) and the right-side panels (4B, D), reveals that for both secretions, strain discriminations are more successful than state discriminations. Comparing the upper and lower panels, particularly for the more difficult, general classifications (red bars), shows that while strain discriminations are marginally better with urine stimuli (4C vs. 4A), state discriminations are considerably more successful with vaginal secretions (4B vs. 4D). Thus, in the context of vomeronasal chemosensation, different secretions seem optimized for conveying distinct socially relevant traits.
Many mammals, and mice in particular, investigate each other in a manner that inevitably samples multiple stimulus sources [7]. From a decoder’s perspective, the challenge is that levels of different cues, and their dependence on particular traits, could vary across secretions. This consideration is particularly relevant in the context of slow vomeronasal sampling, which can lead to mixing of stimuli from different sources within the vomeronasal organ [31, 32].
In the foregoing analyses, each classifier was trained and tested on a single stimulus source. To investigate how classification criteria based on one secretion apply to others, we recorded responses of AOB neurons to urine, vaginal secretions, and saliva (Fig 1C, set 3). The dataset contains 164 units (101 single-units, 63 multi-units) that responded to at least one of the 12 stimuli. Detailed descriptions of the number of units recorded per session and recording site locations, spanning the entire anterior-posterior aspect, are shown in S1 Table and S1 Fig. For the 164 responding units, the mode, median and mean number of significant responses across units was 1, 2, and 2.6, respectively (see also S2 Fig). Note that in this stimulus set, due to the small volumes of salivary samples, stimuli from four to six females were pooled (see Materials and Methods). This has the potential effect of reducing variability and to some extent simplifying the classification problem.
We first trained classifiers for each secretion separately. We began with reproductive-state classifiers that generalize across strains, and strain classifiers that generalize across reproductive states (indicated by the orange lines in Fig 5A and 5C). In these experiments, each stimulus is presented at one dilution, and thus dilution variability does not play a factor. The results of these classifications, shown by the orange lines in Fig 5B and 5D, reveal that in addition to urine and vaginal secretions, salivary cues can also provide information about strain and reproductive-state. Complete classification results for this dataset are given in S3 Fig, panels M-R.
Next, we applied classifiers trained on one secretion to the other two (gray double-headed arrows in 5A, C). The analysis shows that application to other secretions results in chance level performance for state discrimination (Fig 5B, gray traces) and slightly above-chance performance for strain discriminations (Fig 5D). These results suggest that distinct chemical cues provide information about behaviorally relevant traits across the different stimulus sources. The modest above-chance generalization performance observed with strain classifications across different secretions indicates that some units do respond to strain dependent cues that are common across secretions.
Finally, we trained classifiers across secretions, to investigate whether a single classifier (red lines in Fig 5A and 5C) can detect particular traits regardless of the stimulus source. Using the unit removal procedure, we observe that about 15 AOB units suffice for approximately 85% correct classification. Best performance with all units, or one unit, is 90% and 70%, respectively. Performance on strain discriminations is higher, with 10 units sufficient to classify strain with a ~95% success rate. Best performance with all units, or one unit, is 98% and 86%, respectively. As with the previous classifications, polling multiple units provides a considerable advantage over individual units. Notably, the ability of one individual unit to provide 86% correct strain discriminations suggests that some neurons are sensitive to strain dependent cues that are common to the three secretions.
To gain insight about how social information is decoded from AOB population activity, we investigated the influence (classifier weights) assigned to each of the units. Fig 6A and 6B show the vaginal secretion response dataset (data shown in Fig 3A) with the rows sorted according to the values assigned by the general reproductive-state and strain classifiers, respectively. As expected, this ordering shows that the linear classifier assigns more weight to units with response profiles that reflect the detected traits. Yet, while unit reordering highlights each of the two traits, it does not reveal if the observed patterns are represented more than expected by chance. To address this, we defined an index that quantifies the impact of the strain and state dimensions on the dataset (see Materials and Methods). Comparison of the index to its bootstrapped distribution under the null hypothesis, reveals that while the strain dimension is significantly represented (Fig 6D), the reproductive-state dimension is not (Fig 6C). The same is true for responses to urine (strain p-value = 0.0035, reproductive-state p-value = 0.55). This observation indicates why strain discriminations are achieved with higher success than state discriminations.
Fig 6E and 6F show the units ordered according to the weights assigned by the reproductive-state classifiers, specifically for the BC and C57 strains (i.e. classifiers indicated by orange lines in Fig 3D). Inspection of the columns representing stimuli not used for training the classifier (i.e. the six rightmost columns in Fig 6E and the six leftmost columns in Fig 6F) reveals that reproductive state selective responses for one strain are not necessarily associated with similar selectivity for the other. Indeed, examination of matrices in Fig 6A and 6B indicates that even units assigned with the highest absolute weights are influenced by the strain and the dilution. This explains why generalization of reproductive state classifiers across strains is not efficient unless training is done explicitly with stimuli from both strains. The same observations apply with regard to urine stimuli and for strain discriminations across both states, yet the confounding effect of strain on reproductive state discrimination is more dominant than vice versa. As shown in S4 Fig, similar considerations apply to the difficulty of generalizing across concentrations. Generally, we find no systematic relationship between the number of units used and the ability of the classifier to generalize across other instances (S3 Fig). The ability of a classifier to generalize largely depends on the response profiles of the individual units that contribute to it. Specifically, generalization across particular dimensions will be successful if the influential units happen to display invariant responses along these dimensions. To ensure generalization along any given dimension, classifiers must be trained with stimuli that vary along that particular dimension.
Across all secretions and traits, our analysis has shown that consideration of multiple units provides better classification performance than is available with any one unit. This is expected, as the entire population of units also includes the “best” unit. This last consideration raises the possibility that the success of classifiers with many units heavily depends on a small number of key units. To address this possibility, we revisited the classification analysis, but instead of removing the unit with the least effect (i.e. smallest absolute weight) as done above we instead removed at each stage, the unit with the highest absolute weight. For each classifier, the process was repeated until the classifier’s performance decreased below that possible with the best individual unit (Fig 7). The average number of units (across 10 repeated training cycles) that can be removed without impairing classification below the best one-unit performance are: 27.5 (VS state classification), 29.6 (VS strain), 17.1 (urine state), 15.7 (urine strain), 50.5 (across secretions, state) and 11.5 (across secretions, strain). In five out of six cases, about one third of the most influential units can be removed while still maintaining performance above that possible with the best individual unit.
S5 Fig illustrates the results of classification with units that were simultaneously recorded in a single session. As expected, classifier performance is substantially reduced in comparison to that obtained with all units available (S3 Fig). Nevertheless, the main conclusions of our foregoing analyses are reproduced even in smaller samples collected in individual sessions.
S6–S8 Figs show an analysis of the correlations among the response profiles of the most influential units for each of the general classifiers. The analysis shows that while there are some prominent response patterns among the influential units, the responses of most unit pairs exhibit low correlations. Specifically, of all pairwise correlations among the top 20 units within each classifier, the large majority are smaller than 0.5. (% correlations below 0.5: VS state: 86.3%; VS strain: 83.7%; urine state: 89.5%; urine strain: 86.3%; across secretions state: 88.4%; across secretions strain: 93.7%). This analysis thus indicates that to achieve optimal performance, classifiers tend to sample units with a variety of response profiles.
Finally, application of Principal Component Analysis (PCA) to our dataset revealed only weak correlations between the dominant Principal Components and the dimensions analyzed here (S9 Fig). This result is consistent with the idea that neurons in our sample do not represent a homogenous population with a one stereotypic response profile.
Our choice of the perceptron model was motivated by its ability to indicate the contribution of individual units to the classification. Furthermore, in principle, it can be easily implemented by neurons receiving direct excitatory and (indirect) inhibitory input from AOB neurons. However, we are not suggesting that the actual downstream readout of AOB activity involves a perceptron-like classifier. Indeed, reliable trait detection may be improved by multiple processing stages realizing non-linear computations. To illustrate this, we also calculated the ability of a more powerful non-linear classifier, a support vector machine (SVM), to discriminate among traits. Each panel in Fig 7 compares the best performance obtained with an SVM to that of the best perceptron classifier. The comparison clearly shows that SVMs provide significantly improved performance, particularly for the more challenging classifications. In one case (Fig 7D), perceptron performance is slightly better. This appears surprising because an SVM classifier can implement any decision rule that a perceptron can. The explanation for the reduced performance observed in this case is that the SVM likely over-fits the training dataset.
Comparison of SVM and perceptron performance as a function of the number of units (S10 Fig) shows that the SVM advantage is more prominent when more neurons are used. For small ensembles, the differences between the two classifiers are marginal. Overall, this analysis highlights once more the advantage of polling multiple units, and shows that non-linear readout of neuronal activity can yield improved classification performance.
The observation that individual neurons can be modulated by more than one trait implies that each could contribute to the detection of multiple traits. To test whether this is the case, we compared the weights assigned by the reproductive-state and the strain classifiers. Fig 6G shows that across the population of units, the weights for state and strain classifiers for the VS dataset (Fig 6A and 6B) are clearly not correlated. More generally, over 10 classification cycles, the average correlation coefficient between weights for the two distinct classifiers is 0.02±0.03 (mean±SD). For comparison, the average correlation coefficient over all 10 repeated training cycles of the state classifier is 0.99±0.003. Similar results were observed with the other datasets (S2 Table). These results suggest a model according to which individual units participate in multiple distinct and overlapping networks, each associated with classifying a particular trait (Fig 8B).
In this study, we focused on the mouse AOB to study how socially relevant information can be reliably decoded from natural stimuli. In principle, detection of any trait via chemosensation requires a reliable readout of chemical cues that are correlated with the trait. Fig 8A illustrates a very simple (hypothetical) scenario related to the traits addressed here, in which the level of one compound is indicative of estrus state. The example illustrates how factors such as dilution, other traits, or the stimulus source, could modulate compound levels in a manner that confounds discrimination [33]. Even under this simple scenario, trait detection is not a trivial challenge. While there are known cases in which individual molecules elicit well-defined behavioral responses [34–37], encoding even a relatively robust trait such as sex, seems to involve more than a single molecule [38]. More generally, behaviorally relevant information is conveyed by levels of multiple compounds and the relationships between them [2, 4, 20, 25]. These non-trivial relationships between particular traits and levels of chemical cues imply that trait representation in the AOB can be complex.
In one scenario, akin to a labelled line model, the activity of one particular “expert” AOB neuron could provide an unambiguous report of the presence of a particular trait. Under the opposite scenario, the activity of many, potentially all, AOB neurons must be sampled. To provide a definitive answer to this question one would have to test responses of all neurons to an enormous set of chemosensory stimuli varying along multiple dimensions, including: various traits of the stimulus donor (e.g. genetic makeup, sex, age, physiological status), the stimulus source-secretion, and physical properties (e.g. dilution, freshness). In attempt to record many neurons with high temporal resolution and fidelity, we used multisite extracellular recording probes in the AOB. Because of technical limitations on practical dataset sizes, we could only explore a limited stimulus subspace in each single experiment (Fig 1C). Specifically, we tested the responses to urine, saliva, and vaginal secretions of females from different strains and reproductive states. We have shown, for the first time, that AOB neurons have the capacity to respond to all these stimulus sources in a strain and reproductive-state specific manner. Furthermore, compared to urine, vaginal secretions are optimized to convey reproductive state information. Linking these results with behavioral analyses, we note that in hamsters, flank marks are used to detect individuality, whereas estrus state is sensed via vaginal secretions [8]. Adding to previous studies [8, 39, 40], our findings stress the importance of attending to diverse secretions in the context of mammalian chemosensory communication.
Based on the responses to these stimuli, we studied how information about a female mouse’s strain and reproductive state can be decoded from the activity of AOB neurons, despite these sources of variability. While some individual units do allow reliable trait discrimination within our (limited) stimulus set, most are significantly restricted in their ability to provide reliable information across multiple sources of variability. However, a significant gain in classification performance is obtained by considering multiple units. In this respect, our work is in good agreement with a recent study, which demonstrated that deriving sex and strain information from urine also requires consideration of multiple neurons [23].
Our analyses indicate that generalized discriminations require explicit training with stimuli spanning multiple dimensions of variability. The challenge is similar to that of object recognition. For example, viewing an unfamiliar face once as a static image is generally not sufficient to allow identification across a range of contexts (e.g. different moods or ages of the person, as well as viewing angles and lighting conditions). The ability to reliably identify a face under those contexts requires accumulated exposure that provides sampling under diverse conditions. Here, we observed that strain discriminations generally required fewer units and were performed with higher success than reproductive-state discriminations. Indeed, our previous analysis has shown that in the AOB, sex is represented more prominently than strain [24]. Generally, prominently represented traits may be more immune to variability, explaining why the confounding effect of strain on state discrimination is larger than the reciprocal effect. Similarly, the confounding effect of stimulus dilution depends on its magnitude relative to that associated with different trait values [4], suggesting why dilution had a minor confounding effect on sex and strain discriminations in the AOB [23].
One may argue that “expert” neurons for trait detection do exist for all behaviorally relevant discriminations, but that our limited sample simply failed to identify them. This possibility cannot be entirely ruled out. However, even if such “experts” were found for each discrimination considered here, there is no guarantee that they would also provide reliable classification across further sources of variability. Our present study suggests an alternative scenario, in which discrimination is achieved by polling a limited number of distinct units, none of which must be an “expert”. The various response profiles associated with AOB neurons provide a rich substrate to realize a large number of discriminations, including of traits not strongly represented by neuronal activity. This was exactly the case for reproductive state discriminations in our experiments (Fig 6C). More generally, when traits are strongly represented in the dataset, a small number of units may suffice to yield reliable discriminations (Fig 8Bii), and complex decoding mechanisms might not be required. Thus, instead of a dichotomy between labelled lines and combinatorial codes, we suggest that decoding distinct traits may require different population sizes, depending in part on how prominently these traits are represented by individual neurons.
While the perceptron model provides a clear analogy to a downstream neuron, we do not claim that it actually represents the algorithm used by the brain. Indeed, more powerful non-linear classifiers provide better classification (Fig 7), especially when a larger neuronal population is considered (S10 Fig). Note also that we quantified neuronal responses as the mean firing rate changes within a long time window, a choice motivated by the relatively long and variable response patterns of AOB units [41]. A more refined examination of individual units’ rate modulations and spike timing relative to stimulus delivery might lead to improved decoding ability. In particular, temporal response profiles may play an important role during bouts of natural investigation, when the vomeronasal pump is likely activated repeatedly [7, 31]. The dynamics of active sensing could thus substantially affect summation of responses, and hence the ability to discriminate between, or to generalize across, stimuli. Finally, we note that examination of the relationships between individual units’ spike times, and of correlated rate modulations among different units could also lead to improved discriminations. With all these considerations in mind, understanding how AOB activity is indeed read by specific downstream stages remains an important question for future research.
For recordings, adult sexually unexperienced BALB/C (BC) male mice were purchased from Harlan Laboratories (Israel). All experiments were performed in compliance with the Hebrew University Animal Care and Use Committee. Stimuli were collected from adult sexually unexperienced female mice of the BC and C57BL/6 (C57) strains (Harlan Laboratories, Israel).
For urine collection, mice were gently held over a plastic sheet until they urinated. The urine was transferred to a plastic tube with a micropipette and then flash-frozen in liquid nitrogen and subsequently stored in -80°C. Vaginal secretions were collected by flushing the vaginal region with 30 μl of ringer’s solution. 20 μl were immediately frozen and stored at -80°C for stimulus presentation. The remaining volume was smeared on a glass slide for determination of estrus state. For saliva collection, isoproterenol hydrochloride (0.2mg/100g) and pilocarpine (0.05mg/100g) were injected I.P. to increase salivation [42]. Saliva was then collected from the oral cavity using a micropipette and immediately frozen in liquid nitrogen and stored at -80°C. All stimulus dilutions were made with ringer’s solution. Stimulus collection was performed 3–5 times per week, usually after 14:00 (non-reversed light cycle, light on: 7:00–19:00). The estrus stage was determined by examining vaginal secretions smeared on glass slides, dried, and stained with cresyl violet. Slides were examined under a light microscope and the estrus cycle stage determined by cellular morphology [43, 44]. Stages were classified as either estrus/proestrus (designated as estrus), or meta- or diestrus (designated as non-estrus). For the urine and vaginal secretion datasets (sets 1 and 2 in Fig 1C), stimuli for a given strain were obtained from one individual (with stimuli collected during different stages of the cycle). Dilutions used were 9x (L, low), 3x (M, medium), 1x (H, high) for vaginal secretions, and 300x (L), 100x (M) and 33x (H), for urine. For comparison of the three different secretions (set 3 in Fig 1C), stimuli comprised a mixture from 4–6 females. In most cases, stimuli for each session were collected from the same females. For comparison of different stimulus sources, we used undiluted samples, as they are thus sampled during natural investigation. With vaginal secretions, this was not possible since they were collected via flushing. Responses to undiluted urine were clearly stimulus specific, and thus did not represent non-specific activation due to urinary potassium S12A–S12D Fig. Unlike in the MOS [45], where airflow alone can induce activity changes, in the VNS, delivery of ringer’s solution alone does not elicit a response [46]. As examples, S12E and S12F Fig shows recordings of two units, demonstrating a null response to Ringers solution. Because the number of stimuli in our dataset was a limiting factor, we did not use an explicit negative control stimulus here.
Experimental procedures were described previously [24], and are reproduced briefly here, with the differences noted. Mice were anesthetized with 100mg/kg ketamine and 10mg/kg xylazine, a tracheotomy was performed with a polyethylene tube to allow breathing during flushing, and a cuff electrode was placed around the sympathetic nerve trunk with the carotid serving as a scaffold. Incisions were closed with Vetbond (3M) glue and the mouse was placed in a custom-built stereotaxic apparatus where anesthesia was maintained throughout the entire experiment with 0.5–1% isoflurane in oxygen. A craniotomy was made immediately rostral to the rhinal sinus, the dura was removed around the penetration site, and electrodes were advanced into the AOB at an angle of ~30° with an electronic micromanipulator (MP-285; Sutter instruments, Novato, CA). All recordings were made with 32 channel probes with 8 channels on each of 4 shanks (NeuroNexus Technologies, Ann Arbor, Michigan). Before recordings, electrodes were dipped in fluorescent dye (DiI, Invitrogen, Carlsbad, CA) to allow subsequent confirmation of electrode placement within the AOB external cell layer, which contains the mitral-tufted cells [11]. In each session, stimuli were typically presented 5 times in a pseudorandom order. In a minority of sessions, only 4 repeats were possible. Mean number of repeats across all experiments: 4.8. In each presentation, 2 μl of stimulus was applied directly into the nostril. After a delay of 20 s, a square-wave stimulation train (duration: 1.6 s, current: ±120 μA, frequency: 30 Hz), was delivered through the sympathetic nerve cuff electrode to induce VNO pumping and stimulus entry to the VNO lumen. Following a second delay of 40 s, the nasal cavity and VNO were flushed with 1–2 ml of ringer’s solution which flowed from the nostril, into the nasal cavity, and sucked out from the nasopalatine duct via a solenoid-controlled suction tube. The cleansing procedure was 50 s long and included sympathetic trunk stimulation to facilitate stimulus elimination from the VNO lumen. Neuronal data was sampled at 25 kHz using an RZ2 processor, PZ2 preamplifier, and two RA16CH head-stage amplifiers (TDT, Alachua, FL). Signals were band-pass filtered (300–5000 Hz) and custom MATLAB (Mathworks, Natick, MA) programs were used to extract spike waveforms. Spikes were sorted automatically according to their projections on two principal components on 8 channels of each shank using KlustaKwik [47] and then manually verified and adjusted using the Klusters program [48]. Spike clusters were evaluated by consideration of their spike shapes, projections on principal component space (calculated for each session individually) and autocorrelation functions. A spike cluster was defined as a single unit if it had a distinct spike shape and was fully separated from both the origin (noise) and other clusters along at least one principal component projection, and if its inter-spike interval histogram demonstrated a clear trough around time 0 (of at least 10 ms). Clusters comprising more than one single unit were designated as multi-units. Thus, using the present definitions, multi-units could represent the activity of as few as 2 units, or more. Throughout this manuscript, we used both single and multi-unit activity with the aim of increasing the probability of finding individual units conveying robust information. Note that one of our key conclusions is that response profiles of individual units are not sufficiently general, and this cannot be a result of confounding multiple units together. Comparison of classifier weights assigned to single and multi-units (see Fig 6G) revealed no systematic relationship between unit type and the magnitude of classifier weight.
All data analyses and visualizations were performed using either custom or standard MATLAB code. The response of a unit to a given stimulus was defined as the average firing rate change over a 40 s window following sympathetic nerve stimulation (change measured compared to the 30 s period preceding VNO activation). Response significance of a given unit to a given stimulus was determined by comparing its spiking rate distribution following all repeats to the baseline firing frequency distribution during the 10 s period prior to stimulus application. Response significance (of a particular unit to a given stimulus) was determined by a non-parametric ANOVA comparing the set of post-stimulation rates to the set of preceding baseline rates (preceding rates were pooled across all stimuli).
|
10.1371/journal.pcbi.1006946 | TASmania: A bacterial Toxin-Antitoxin Systems database | Bacterial Toxin-Antitoxin systems (TAS) are involved in key biological functions including plasmid maintenance, defense against phages, persistence and virulence. They are found in nearly all phyla and classified into 6 different types based on the mode of inactivation of the toxin, with the type II TAS being the best characterized so far. We have herein developed a new in silico discovery pipeline named TASmania, which mines the >41K assemblies of the EnsemblBacteria database for known and uncharacterized protein components of type I to IV TAS loci. Our pipeline annotates the proteins based on a list of curated HMMs, which leads to >2.106 loci candidates, including orphan toxins and antitoxins, and organises the candidates in pseudo-operon structures in order to identify new TAS candidates based on a guilt-by-association strategy. In addition, we classify the two-component TAS with an unsupervised method on top of the pseudo-operon (pop) gene structures, leading to 1567 “popTA” models offering a more robust classification of the TAs families. These results give valuable clues in understanding the toxin/antitoxin modular structures and the TAS phylum specificities. Preliminary in vivo work confirmed six putative new hits in Mycobacterium tuberculosis as promising candidates. The TASmania database is available on the following server https://shiny.bioinformatics.unibe.ch/apps/tasmania/.
| TASmania offers an extensive annotation of TA loci in a very large database of bacterial genomes, which represents a resource of crucial importance for the microbiology community. TASmania supports i) the discovery of new TA families; ii) the design of a robust experimental strategy by taking into account potential interferences in trans; iii) the comparative analysis between TA loci content, phylogeny and/or phenotypes (pathogenicity, persistence, stress resistance, associated host types) by providing a vast repertoire of annotated assemblies. Our database contains TA annotations of a given strain not only mapped to its core genome but also to its plasmids, whenever applicable.
| Toxin-antitoxin systems (TAS) were originally known for their involvement in a process known as post-segregational killing (PSK), a plasmid maintenance mechanism based on the differential decay of the products of two plasmid-encoded genes: a toxin gene and its antagonistic antitoxin [1–3]. The current model for TA activation is that under normal growth conditions, the antitoxin efficiently counteracts the toxin negative effects. Yet, under certain stress situations the toxin is released, thus leading to a transient metabolic shutdown and growth arrest. TAS can be acquired from mobile genetic elements such as plasmids or phages, and are also present in core genomes [4]. The ability to be transferred both vertically and horizontally renders any phylogenetic analysis difficult and little is known about the distribution of the TAS among phylum. The work by Wood and his group with artificial toxin derived from endogenous antitoxins (and vice-et-versa) highlights the plasticity of ubiquitous TAS and the complexity of their origins [5]. Since the discovery of the PSK, the growing list of TAS related studies has led to a list of more complex (and sometimes controversial) roles for TAS. To name a few, TAS are involved in cell suicide following a phage abortive infection [6] or nutritional stress [7], in regulating biofilm dynamics [8] and in bacterial persistence [9–11]. Some studies even show that chromosomal TAS can counteract PSK [12].
All TAS toxins are proteins that target a variety of essential biological processes (e.g., membrane integrity, translation, replication) and they are divided in groups based on the nature and mechanism of action of the cognate antitoxin [13]. Currently there are six types of TAS described in the literature. In the type I family, an ncRNA antitoxin (generally in antisense of the toxin gene) inhibits the translation of the toxin mRNA. Typical examples of type I TAS are the hok/sok systems [3]. Type II TAS, which constitute the most commonly studied family, are composed of an antitoxin protein that binds directly to the toxin protein and inhibits its activity. Some toxins target DNA replication [14], or affect the cell membrane integrity by phosphorylating peptidoglycan precursors [15], while others have acetyltransferase activity [16,17], or are kinases that target the translation elongation factor EF-Tu [18,19]. Yet, many type II toxins are ribonucleases that i) cleave target mRNAs in a ribosome-dependent manner [20] or ii) cleave free mRNA [21], and they can also target non coding RNA [22,23]. Type III is a more recent addition, with ToxN/ToxI as a reference member [6] and more families added later by the pioneering work from Salmond’s group [24]. The type III toxin is a nuclease that cleaves a broad range of mRNA and RNA, while the antitoxin is a small non-coding RNA that binds directly to the toxin protein, thus inhibiting its action. In type IV there is no direct interaction between the toxin and antitoxin components. Here the antitoxin counteracts the toxin by competing with its targets, like cytoskeleton proteins [25]. Type V currently has so far only a single member, the GhoT/GhoS system [26], in which the antitoxin itself is an endoribonuclease protein that targets the toxin mRNA [27]. Type VI TAS are grouped TA systems that involve a third partner. This partner promotes the toxin decay in trans [28] or the antitoxin stability in cis [29].
The ubiquity of the TAS and the diversity of their functions open question about their potential interactions in trans. Numerous publications suggest that it may be between noncognates from same families [12,30–32] or between noncognates from different TAS types [33,34]. On the other hand, other data suggest isolated TA units [35]. The Laub group used co-evolution study of protein-protein interactions to show that paralogous ParD/ParE pairs are highly specific in their operon cognates [36]. Nevertheless, their model of promiscuous intermediates still leaves room for interactions in trans. Finally, most of the TAS studies focus on the canonical TAS that are usually found in a configuration with the antitoxin gene being upstream of the toxin gene, with few TAS families presenting a reversed order [4,37]. Alternative structures have been mentioned by van Melderen and her group, which highlights the existence of orphan TA loci [38]. So far, TAS screening approaches usually skip the multigene TA systems, despite known tripartite TAS [29,39–41] and TAS modules inserted within operons [7,42].
Validated and predicted TAS are collected in the TADB2 database [43]. TADB2 focuses mainly on type II TAS that were mined from the literature (N = 105 TA loci) and from previous published screens (N = 6088 TA loci) extracted from 870 bacteria and archaea genomes. The 6088 TA loci were predicted using Blastp on 126 genomes [37] or PSI-Blast searches with validated literature datasets [44]. A few of them were additionally combined with known operon structure obtained from STRING [45]. TADB2 also includes a search tool called TAFinder (http://202.120.12.133/TAfinder/index.php) combining homologous search and operon structure module filters [43]. TAFinder uses Blastp searches with the TADB2 dataset and HMM searches with 108 Toxin HMMs and 201 Antitoxin HMMs to select the TA loci. These loci are then filtered using protein size (by default >30aa and <300aa) and intergenic distance (by default from -20nt to +150nt). TADB2 and TAFinder are very stringent in their criteria to minimize false positives.
Our primary goal is to provide the microbiology community with a largely extended database of the type I to type IV (and potentially type V to VI as side hits) toxin and antitoxin loci. We also propose an objective annotation of the TA independently of the cognate components. With the current nomenclature based on the identification of the toxin cognate, the antitoxin would “inherit” the toxin family name. This can be misleading and ignores the modularity of TA cognates. Instead, our method allows the discovery of unexpected combinations of toxin and antitoxin families. We include a “guilt-by-association” approach in our pipeline, similarly to methods developed by others [38,44]. The large dataset of genomes enables us to apply phylogenetic comparisons.
The EnsemblBacteria database (Rel. 33 Nov. 2016) contains N = 41'610 genomic assemblies that correspond to N = 23'921 unique taxonomic identifiers (taxonomy ids), indicating a high degree of redundancy in the assemblies. At least one hit was found for N = 40'993 assemblies present at least one hit with the TASmania HMM scan, of which N = 22'950 correspond to unique taxonomy ids. A closer look at the taxonomy ids shows that 40% of the genomic assemblies belong to the Proteobacteria phylum and 34% to the Firmicutes phylum, these two groups making up three quarters of the database (S1 Fig). The Actinobacteria and Bacteroidetes phyla represent 12% and 3% of the assemblies, respectively. The remaining 11% of the assemblies correspond to N = 72 other phyla and/or unclassified bacteria.
TASmania is based on the pipeline summarized in Fig 1. Briefly, the strategy relies on TA HMM profiles built from an initial set of proteins annotated with TA InterPro (IPR) (S1 Table). This critical initial set is a known limitation affecting other methods like TADB2 or TAFinder and might lead to missing families. From the protein clustering we obtain N = 369 toxin HMM profiles (with at least 10 unique protein sequences) and N = 305 antitoxin HMM profiles (with at least 10 unique protein sequences). From the theoretical N = 369*305 = 112’545 possible combinations in canonical AT/TA operons, we only observe N = 2’600 HMM profile combinations.
We combine the HMM profiles into larger HMM clusters by similarity. This allows to decrease the number of toxin HMM profiles (N = 369) and antitoxin HMM profiles (N = 305) combinations to plot. When using clustered HMM profiles (N = 152 clusters for toxin HMM profiles and N = 130 clusters for antitoxin HMM profiles), we go from theoretical N = 152*130 = 19’760 combinations to only N = 1’567 observed pairs. Thus, grouping the HMM profiles into clusters allows a decrease of ∼40% in the number of combinations and reduces potential redundancy of certain HMM profiles. We always keep the link between HMM profiles and their clusters. We call each of these clusters TASMANIA.T1 to TASMANIA.T152 (T1 to T152) for the toxins, and TASMANIA.A1 to TASMANIA.A130 (A1 to A130) for the antitoxins. We enhance the value of the putative TA hits by structuring the loci into pseudo-operons and including phylogenetics information. A given combination of two clusters within pseudo-operon is dubbed “popTA”. Finally, for reverse-compatibility with the current TA nomenclature, we also include a nearest Pfam annotation for a given HMM profile and cluster (S2 Table). More details are given in Materials and Methods section.
After scanning EnsemblBacteria with the HMM profiles, we obtain N = 1'155'070 putative toxin gene hits, corresponding to N = 228'074 unique toxin protein sequences; and N = 1'283'761 putative antitoxin genes hits, corresponding to N = 270'733 unique antitoxin protein sequences. In total, the putative toxin or antitoxin hits correspond to N = 2'298'903 unique pseudo-operons containing TA modules (including redundant ones). A phylogenetic analysis of the TA hits distribution shows that Cyanobacteria are very TA-rich and are the most common phylum in the top 200 most TA-enriched genomes (S2 Fig). Our method does not use a protein length filtering, thus allowing for discovery. The protein length distribution of the putative toxin and antitoxin hits confirms previous results [46], as shown in Fig 2. We can see that the absence of length thresholding allows the discovery of more putative TAs (right tail of the distributions). When focusing on the canonical—i.e., the two-gene T->A or A->T modules—the protein length distribution mimicks the previously published data by narrowing the proteins length into the 30–210 residues window used by [46]. This effect is most probably due to the bias of annotation favouring AT/TA modules. However, as can be seen in green on Fig 2, some toxin and antitoxins of the canonical AT/TA modules exceed the 210 aa limit from [46] and 300 aa from [43].
The distribution of the pseudo-operon structures of the HMM scan hits in Fig 3A i) indicates that TAS can be multi-cistronic organisation, not uniquely bi-cistronic.; ii) confirms that the A->T module type is more common than the T->A type and iii) shows the existence of many “orphan” hits, i.e., a toxin or antitoxin gene as single-gene pseudo-operon. These hits could be either true orphaned T’s or A’s, and/or false positives and/or could be due to the mis-annotation of the operons and/or potentially type I or type III toxins as we cannot detect the ncRNA with our current method. The prevalence of the A->T type is highlighted when comparing only canonical two-genes structures (Fig 3B).
We compared TASmania putative TAS hits with the ones proposed by TAfinder. Since we cannot download the entire datasets from this webtool, we used a few reference model strains as a proof of principle: Mycobacterium tuberculosis H37Rv (M.tuberculosis), Mycobacterium smegmatis MC2155 (M.smegmatis), Caulobacter crescentus CB15 (C.crescentus) and Staphylococcus aureus NCTC8325 (S.aureus). The putative hits were manually downloaded from these websites and compared against TASmania hits (Fig 4). These data show that TASmania covers most of TAfinder hits and gives many other putative candidates (Fig 4 and S3 Table).
Looking closely at the TAfinder hits missed by TASmania, the module Rv2653c/Rv2654c in M.tuberculosis H37Rv seems to encode prophage proteins, with no IPR annotation, hence their absence from TASmania (S4 Table). This module could be a real TAS and if this hypothesis happens to be confirmed experimentally, they will be added to TASmania profiles. The remaining TAfinder hits missed by TASmania fall into the transcriptional regulators (e.g., ArsR, LysR, TetR, MarR), transposases and uncharacterized proteins categories. It is difficult to evaluate if these loci are true TA missed by TASmania or false positives from TAfinder.
Although it is technically not possible to assess the overall rate of false positives in the TASmania-specific hits, the in vivo analysis performed on some TASmania candidates shows promising results. We investigated whether some of the putative TA systems of M.tuberculosis identified by TASmania were indeed bona fide new TA systems. We selected 11 putative TA systems that are not found by TAfinder or TADB2 and asked whether expression of their putative toxins could affect growth of the closely related M.smegmatis strain MC2155. Putative toxin encoding genes were cloned into the pLAM12 vector under the control of an acetamide inducible promoter, transformed into MC2155 and incubated for 3 days at 37°C on kanamycin agar plates without or with 0.2% acetamide inducer. Under these conditions we found that six out of eleven putative toxins affected M.smegmatis growth, with four of them exhibiting a robust toxicity, namely Rv0078A, Rv0366c, Rv2016 and Rv2514c, and two only inducing a slow growth phenotype, namely Rv0207c and Rv0269c (Fig 5). These results suggest that these six genes could encode toxins of new or uncharacterized TA systems in M.tuberculosis, thus further extending the long list of TA in this bacterium [47].
In order to investigate whether these toxic genes are part of bona fide TA systems, the six corresponding TA operons composed of the putative toxin encoding genes and of the putative cognate antitoxin genes were cloned in pLAM12 vector, transformed in MC2155 and their effect on bacterial growth was monitored as in Fig 6. Note that 4 out of these 6 putative TA systems are in antitoxin first, toxin second (AT orientation), and the last two in toxin first, antitoxin second gene organization (TA orientation) (Fig 6).
We found that in all cases bacterial growth could be rescued by the presence of the putative antitoxin genes in all cases, although to various levels (Fig 6). Rv0078B/Rv0078A (A->T) and Rv2515c/Rv2514c (A->T) operons both support the in silico prediction of putative TAS: the high toxicity of the putative toxin expressed alone is inhibited by the co-expression of the putative cognate antitoxin. Rv0078B/Rv0078A (A->T) is a very interesting case. Remarkably, although Rv0078B acts as an antitoxin and rescues the toxicity of Rv0078A, TASmania HMM scan flags Rv0078B as a putative toxin from the cluster T52 (nearest Pfam SymE_toxin type I). Rv0078A is also flagged as a toxin via its IPR annotation (IPR014942 AbiEii toxin type IV). This unexpected predicted “TT” pair could be the signature of a new family of TAS, with Rv0078B being a potential example of a TAS cognate that “switched” function [4]. T52 hits like Rv0078B are found in diverse pseudo-operons structures, although T52 should in theory be a toxin of type I and therefore rather appears in pseudo-operons looking like orphans (“T”). M.tuberculosis presents only a single pseudo-operon with T52 hit, while it is absent from M.smegmatis and appears in N = 34 different loci in Thalassomonas actiniarum. In the latter, T52 hits are all orphan toxins, suggesting that, in this species at least, T52 looks more like a classical SymE-like toxin type I (the antitoxin cognate being a ncRNA, it cannot be annotated currently by TASmania).
On the other hand, Rv0208c/Rv0207c and Rv0269c/Rv0268c are both putative TAS operons with the toxin exhibiting a weak toxicity when expressed in M.smegmatis. This could be due to various reasons, including missing/divergent M.tuberculosis toxin targets in M.smegmatis, potential cross-interactions in trans with the cognate antitoxins of other similar TAS, a poorly expressed toxin in M.smegmatis, a non-essential toxin target or a target not required under the growth conditions tested. Rv0269c/Rv0268c is a TAS in T->A conformation, with the antitoxin Rv0268c annotated as a A24 (nearest Pfam family PhdYeFM_antitox), while Rv0269c is proposed as a guilt-by-association toxin. In M.tuberculosis, only Rv0268c is found as a A24 hit, but many other loci (N = 12) belong to PhdYeFM_antitox clusters (A24, A9, A27, A81, A94, A100). Rv0269c/Rv0268c is interesting since it is in a T->A configuration, which is unusual for the PhdYeFM antitoxin. Homologies suggest that Rv0269c is related to proteins with a DNA polymerase/primase/ligase domain. Therefore Rv0269c/Rv0268c is a puzzling pair worth deeper investigation. Whether these two systems are bona fide TA pairs remains to be investigated.
Rv0367c/Rv0366c (A->T) is a putative TA couple where both loci are hit by TASmania HMM profiles belonging to the A123 (nearest Pfam ParD_like) and T70 (nearest Pfam Zeta_toxin) clusters, respectively. The combination A123.T70 (nearest Pfam ParD_like.Zeta_toxin) could represent a new TAS family, since the canonical zeta toxin is described in the literature as the cognate of epsilon antitoxin. In the TASmania database, most of T70 clusters hits appear as paired with A49 and A123 clusters (both with nearest Pfam ParD_like).
Finally, in the case of Rv2016 (T144 nearest Pfam HicA_toxin), which is highly toxic when expressed in M.smegmatis, we could also detect an effective but very limited suppression of toxicity in the presence of the putative antitoxin gene Rv2017 (A32 nearest Pfam HTH_3). Whether this is due to the genetic organization with the toxin and/or to the lack of a chaperone partner is unknown [48]. All together, these experimental validations of TASmania in silico predictions show how our database can be a very powerful tool in discovering unexpected TAS families.
For clarity and reproducibility, we focus on the two-genes modules to study the toxin and antitoxin clusters co-occurrence within the pseudo-operons, i.e., popTAs. In order to minimize bias introduced by the overrepresentation of certain phylogenetic groups over others (see S1 Fig), we apply a correction to cluster counts with the weight of each phylum in the database. Out of the theoretical N = 152*130 = 19'760 possible combinations, we find N = 1'522 popTAs, independently of their T->A or A->T orientation; and N = 1'567 popTAs if the orientation is taken into account.
The popTA features highlight the potential issues that the TA annotations can produce. In the current way toxins and antitoxins are annotated, namely by giving priority to the toxin for naming the antitoxin, many inconsistencies are created. For example in M.tuberculosis, several antitoxins are annotated as a “VapB” while the TASmania HMM profiles hitting these antitoxins belong to diverse Pfam families like PhdYeFM, ribbon-helix-helix (RHH), CopG or MazE (Table 3).Therefore, we here propose a more objective and systematic annotation of the toxins and antitoxins based on cluster identifiers, rather than misleading functional names inferred from cis-occurrence.
The guilt-by-association approach [38,44] allows the discovery of previously undescribed protein families. This strategy relies on the non-targeted cognate loci of TASmania hits in two-genes operons—“xT”, “Tx”, “Ax” and “xA”. For convenience we focus on xT/Tx starting by collecting and pooling the protein sequences corresponding to the “x” cognates of toxins HMM hits in TASmania. These x cognates are loci that do not have any previous IPR annotation corresponding to known TAS families, nor are they picked up by any of HMM profiles. But they have a toxin as direct neighbour gene, identified by TASmania HMM profile(s) and/or direct IPR annotation. As a proof of principle, we screen all the “x” genes having as neighbour a toxin T cognate, in two-genes pseudo-operons “xT” and “Tx” (we dub these two types of pairs as “popTx”, independently of the orientation). We obtain N = 24’377 unique protein sequences that could potentially belong to new uncharacterized antitoxins. We build and cluster the HMM profiles using the same procedure as for TASmania (see Methods below). These putative new antitoxin families are summarized in Table 4.
Many x antitoxins are annotated as nearest to Pfam HTH_3 (A*1 and A*8) and RHH_1 (A*27) features, for instance in the following pairing types: HigB_toxin.HTH3, HipA_C.HTH_3, HTH_3.HipA_C, ParE_toxin.HTH_3, RelE.HTH_3 and RHH_1.ParE_toxin. These HTH_3 and RHH_1 Pfam annotations are too general to directly infer functional clues for these putative new antitoxin families but they are good candidates to discover new antitoxin families. Each of the different popTx groups derived from these HTH_3 and RHH_1 combinations would require further characterization based on cognates alignments and structural analyses for example. Some other interesting x antitoxins are the ones with nearest Pfam annotations of Colicin_Pyocin (A*190, as in Colicin_Pyocin.YafQ_toxin—.A*190.T4), VraX (A*371, as in VraX.PemK_toxin—A*371_T143, specific to Staphylococcus), Glyoxalase (A*77, as in YafQ_toxin.Glyoxalase—T32.A*77), Antirestrict (A*237, as in Antirestrict.CbtA_toxin—A*237.T3) and Response_reg (T5, as in Cpta_toxin.Response_reg—T5.A*2). VraX (IPR035374) and Glyoxalase (IPR004360) are both involved in antibiotics resistance pathways. The VraX-like putative antitoxins seem to originally be derived from a phage protein. Intriguingly, the VraX.PemK pair is not found in the reference Staphylococcus aureus subsp. aureus NCTC 8325 while it is present in other S.aureus strains (S4 Fig). Colicin_Pyocin and Response_reg families could potentially give some clues in the evolution of the TAS. The Colicin_Pyocin (IPR000290) family contains the immunity proteins and/namely members of the effector-immunity system, which is a two-component genetic system (TCS) similar to the TAS but where both cognates are secreted in order to protect the bacteria itself and its clonemates [50]. Response_reg (IPR001789) belongs to another two-component genetic system called “two-component signal transduction system”, which also presents similarities with the TAS. Previous publications have already suggested potential interplay and/or homology between different TCS [51,52]. Finally, annotations from other x antitoxins indicate that many more popTx could be promising candidates: Ap_endonuc_2 (as in AP_endonuc_2.ParE_toxin) and Phage_integrase (as in CcdB.Phage_integrase, Phage_integrase.PemK_toxin or Phage_integrase.Zeta_toxin). These two candidates highlight the link between the TAS and the phages. More investigation will be needed to confirm these candidates as functional new antitoxin families.
We believe that the strength of TASmania is its discovery-oriented feature. Although this may lead to unwanted false positives, it also allows for the identification of candidate TAS in species previously described as not containing any TAS loci. Typically, the Prochlorococcus marinus and Mycoplasma are good examples to show the advantage of TASmania. Indeed, while no hit is predicted using TAfinder, TASmania shows that various Mycoplasma assemblies harbour putative type II (e.g., the TA pair D500_0109/D500_0110 in Mycoplasma feriruminatoris, which corresponds to a Pfam YafQ/RelB-like pair) and type IV (e.g., MAGb_3900/MAGb_3910, an AbiEii/AbiEi_4-like pair in Mycoplasma agalactiae 14628) hits. In addition, TASmania identifies several putative TAS (including many orphan loci) in various Prochlorococcus marinus assemblies, which would need further investigation before validation as type II, and also some less clear TAS types like P9303_20011/P9303_20021 pair in Prochlorococcus marinus str mit 9303 (similar genes also in other related assemblies) that correspond to a PIN/Clp-like pair. Intriguingly the next neighbour gene P9303_20031 is also a Clp protease. Overall, TASmania data indicate that even species previously considered as “TAS-free” in the literature might actually contain TAS loci, but whether these are expressed in vivo and are biologically functional would require to be investigated in further experimental analysis.
By avoiding any assumption in the TA protein length and the type of operon—TASmania includes orphan TA loci and TAS hits from multigene pseudo-operons—our database opens up to new TAS families and possible networks. In parallel, we use our large database to apply a meaningful analysis of the biology of the TAS by looking at their organisation in pairs. Our results highlight the modularity of the TA cognates and the issues raised by the conventional misleading family annotations of the TAS. Currently TASmania has three main limitations: i) due to our discovery approach, we suspect that the false positive rate might be high, but it is difficult to assess ii) the downside of automated clustering methods in general iii) the absence of the phage genomes (but prophages and plasmids are included). One should also note that TASmania can contain putative type V and type VI as “side hits”, although these were not mined for purposely. These hits correspond to T or A mined from type I-IV HMM profiles, but due to the modularity, plasticity and the rapid evolvability of the TAS [4,5,38], they can be found in type V-VI. Beside the discovery of uncharacterized TAS missed by alternative sources, TASmania can provide valuable help in the experimental design step. Indeed, the frequent presence of multiple TAS within same genomes, including orphan loci, raise the issue of potential (positive and/or negative) interference in trans. By providing an in silico updated map of putative TAS, TASmania offers the possibility to consider a maximum of potential interferences of TAS in trans when designing an experiment, and to compare this with other strains of interest. Ideally, RNA-seq data should be combined with the TAS in silico annotation in order to get an accurate landscape of TAS. TASmania is very powerful thanks to its large number of assemblies (>41K), which has never been proposed so far. Some of TASmania’s potential applications are phylogenetics and phenotypic comparisons of different isolates. For instance, TASmania can help in making comparative studies by more accurately mapping putative TA loci in E.coli strains with various pathogenicity [53], or in Endozoicomonas sequenced strains from different ecosystems [54], highlighting how this could link to the associated hosts (our own unpublished data).
TASmania is a new resource for the discovery of toxin-antitoxin in known bacterial genomes. Even though it is based on existing protein domain descriptions, its flexibility allows for the uncovering of potential new combination of pairs and totally new families of toxins and/or antitoxins using a guilt-by-association strategy. The experimental validation in vivo of several predicted TAS confirms the potential of this resource for the identification of TAS.
The global strategy is to build an updated list of toxin and antitoxin HMM profiles and scan a local version of the EnsemblBacteria database (N>41K assemblies) with thoses HMM profiles. To achieve this, we have downloaded EnsemblBacteria (release 33, November 2016) [55], updated its InterPro (IPR) (version Nov 2016) [56] annotation and applied a pseudo-operon annotation with arbitrary definition where a maximal intergenic distance of 100 bp is applied, as shown in Fig 1.
In parallel, we perform HMM profiles comparison in order to reduce the number of profiles, using the Profile Comparer program PRC (v1.5.6) [59]. Combining the PRC results with the NetworkAnalyzer [60] in CytoScape (3.5) [61] network analysis, we select the first PRC E-value of 10−12 where the number of connected components (CC) (i.e clusters of HMM profiles) is reaching the plateau. For clarity and continuity with previous TAS annotations found in the literature, each TASMANIA cluster identifier is given the nearest corresponding Pfam family names (release 31.0) to which the TAS scientific community is used to. The “nearest” Pfam annotation is performed as follows: using the PRC program for profile-profile comparison (default settings), each TASmania HMM profile is scanned against Pfam database. The best Pfam profile match for each TASmania HMM profile (i.e., the lowest E-value) is selected and the identifier of this Pfam annotation is used as the Pfam equivalent of the given TASmania HMM profile. On top of the HMM profile annotation, the TASmania clusters are also attributed a Pfam annotation. For each TASmania cluster we attribute the common profile Pfam annotation when there is no ambiguity. In cases of heterogeneity (more than one Pfam annotation per cluster), the Pfam match with the smallest E-value is selected. But in all cases, the individual Pfam annotation of each TASmania HMM profile is kept and shown in S2 Table for methodology coherence. We used the word”nearest” to emphasize the potential issues of such equivalences.
The final TASmania database contains: i) the putative hits from the HMM scan; ii) the genes annotated with a reference TAS IPR and that were filtered out due to the small size of their proteins clusters (less than 10 unique sequences) when building the HMM profiles; iii) the guilt-by-association “x” cognates (see S5 Fig). We also add an extra annotation of the putative TAS hits by analysing the cis-occurrence—within a same pseudo-operon—of toxins and antitoxins clusters: we call these T<->A clusters associations “popTA” groups. To construct these popTAs we first define the pseudo-operon structures using a relaxed model containing one, two or more genes. Our pseudo-operon model is simply based on an arbitrary intergenic distance -100nt < = D < = +100nt between adjacent genes oriented in the same direction (strand), keeping in mind that there is no "one-size-fits-all" D value. We selected the arbitrary value of 100nt based on some previous studies of intergenic distances distributions [62]. The pipeline is summarized in the Fig 1. The popTA sequences comparisons in Fig 9 are done with ClustalO, the MSA plots with Jalview (2.9.0b2) [63] and the HMM profiles of the MSA are plotted with Skylign [64].
Plasmid constructs. Plasmid pLAM12 [65] has been described elsewhere. The eleven putative new toxins identified by TASmania were PCR amplified using primers from S7 Table and cloned in pLAM12 under the control of an acetamide inducible promoter. Cloning was performed using appropriate restriction enzymes or by In-Fusion methodology (Clontech), as indicated in S7 Table. Constructs were sequence verified using primers pLAM-For 5’- ACCCTCCACCGGCCGCGCTC and pLAM-Rev 5’- TGGCAGTCGATCGTACGCTA. For toxins that affected M.smegmatis growth, their respective toxin-antitoxin operons (six in total) were then PCR amplified and cloned into pLAM12, using appropriate primers from S7 Table.
In vivo growth assay. The pLAM12-based constructs were first electroporated in Strain M.smegmatis MC2155 (strain ATCC 700084). Following 3 h incubation at 37°C in LB medium + tween 80 (0,05%), 1/100 of the transformants were directly plated on LB agar supplemented with kanamycin (20 μg/ml) and acetamide (0,2%). Plates were incubated 3 days at 37°C.
Note that Rv0229c only showed a tiny but reproducible effect on M.smegmatis growth when expressed alone (Fig 5). Therefore, we decided to test it within the context of its operon as well. The effect of the putative antitoxin Rv0230c was hardly detectable (S6 Fig), indicating that Rv0229c/Rv0230c may not be a functional TA system when expressed in M.smegmatis.
Similar to the popTAs analysis performed on the canonical TA/AT hits previously, we pool all the “x” protein sequences, cluster them with MMseqs2, make an MSA of each cluster, build an HMM profile for each protein cluster, and compare and cluster the HMM profiles (N = 805) with PRC and Cytoscape. We dub these putative antitoxin HMM clusters as TASMANIA.A*n (A*n) (N = 536 at E-value = 10−5). After Pfam annotation of these putative antitoxin clusters, we perform a semi-automated curation to discover new antitoxin families. One criterion of selection we applied is that the nearest Pfam annotation of the “x” antitoxin should not belong to known antitoxin families (e.g., ParD_antitox, CcdA, CbeA_antitoxin, MazE_antitoxin, PhdYeFM_antitox, CopG_antitoxin, AbiEi, VAPB_antitox). We then go further in stringency by selecting only pairs whose T toxin cognate had an HMM E-value below 10−20 and we thus obtain N = 222 xT/Tx combinations. We find that 27 popTx contain putative new antitoxin protein families worth investigating, since they are conserved up to high stringency.
|
10.1371/journal.pgen.1004943 | DNA Damage Response Factors from Diverse Pathways, Including DNA Crosslink Repair, Mediate Alternative End Joining | Alternative end joining (Alt-EJ) chromosomal break repair involves bypassing classical non-homologous end joining (c-NHEJ), and such repair causes mutations often with microhomology at the repair junction. Since the mediators of Alt-EJ are not well understood, we have sought to identify DNA damage response (DDR) factors important for this repair event. Using chromosomal break reporter assays, we surveyed an RNAi library targeting known DDR factors for siRNAs that cause a specific decrease in Alt-EJ, relative to an EJ event that is a composite of Alt-EJ and c-NHEJ (Distal-EJ between two tandem breaks). From this analysis, we identified several DDR factors that are specifically important for Alt-EJ relative to Distal-EJ. While these factors are from diverse pathways, we also found that most of them also promote homologous recombination (HR), including factors important for DNA crosslink repair, such as the Fanconi Anemia factor, FANCA. Since bypass of c-NHEJ is likely important for both Alt-EJ and HR, we disrupted the c-NHEJ factor Ku70 in Fanca-deficient mouse cells and found that Ku70 loss significantly diminishes the influence of Fanca on Alt-EJ. In contrast, an inhibitor of poly ADP-ribose polymerase (PARP) causes a decrease in Alt-EJ that is enhanced by Ku70 loss. Additionally, the helicase/nuclease DNA2 appears to have distinct effects from FANCA and PARP on both Alt-EJ, as well as end resection. Finally, we found that the proteasome inhibitor Bortezomib, a cancer therapeutic that has been shown to disrupt FANC signaling, causes a significant reduction in both Alt-EJ and HR, relative to Distal-EJ, as well as a substantial loss of end resection. We suggest that several distinct DDR functions are important for Alt-EJ, which include promoting bypass of c-NHEJ and end resection.
| Alternative EJ (Alt-EJ) is a chromosomal double strand break (DSB) repair pathway that often uses short stretches of homology (microhomology) to bridge the break during repair. Alt-EJ involves bypass of the classical non-homologous end joining (c-NHEJ) pathway, and hence may be important for DSBs that are not readily repaired by c-NHEJ, such as DSBs requiring extensive end processing prior to ligation. Since the factors that mediate Alt-EJ are unclear, we identified DNA damage response factors that differentially promote Alt-EJ relative to an EJ event that is a composite of c-NHEJ and Alt-EJ. Several of these factors promote other repair events that are enhanced by loss of c-NHEJ, namely homologous recombination (HR), including DNA crosslink repair factors, such as FANCA. We then investigated distinctions among individual factors. For instance, we found that loss of c-NHEJ appears to diminish the influence of FANCA on Alt-EJ, but enhances the effect of PARP inhibition. Furthermore, we find that FANCA and DNA2 differentially affect another aspect of the DNA damage response, namely end resection. Based on these findings, we suggest that several aspects of the DNA damage response are important for Alt-EJ.
| End joining (EJ) repair of chromosomal breaks is important for cellular resistance to clastogens, and for antibody maturation that is induced by programmed double-strand breaks (DSBs) [1]. However, EJ can be prone to cause loss of genetic information, as it does not require the use of extensive homology or a template for repair. Loss of genetic information can include insertions or deletions, point mutations, and/or formation of gross chromosomal rearrangements. Such gross chromosomal rearrangements are associated with cancer and inherited diseases, and can often show evidence of short stretches of homology (microhomology) at the rearrangement junctions [2–4]. Defining the factors that influence the frequency of these different EJ outcomes provides insight into the processes that ensure genome maintenance.
Repair via EJ can be classified into two major types: classical non-homologous EJ (c-NHEJ) and alternative-EJ (Alt-EJ) [5–8]. C-NHEJ events are mediated by a set of factors important for antibody maturation, including the DSB end binding factor Ku (Ku70/80 heterodimer), the kinase DNA-PKcs, and the XRCC4/Ligase 4 complex [1]. These c-NHEJ factors are also important for radioresistance, yet in their absence, chromosomal EJ remains relatively proficient, but repair junctions show increased frequencies of insertions and deletions, as well as greater evidence of microhomology usage [9–11]. The term Alt-EJ refers to such repair events that are independent of c-NHEJ factors [9–13]. While Alt-EJ events often show microhomology at the repair junction, microhomolgy is not absolutely essential for Alt-EJ [9–13]. Furthermore, c-NHEJ can also use microhomolgy during repair [14].
The increase in Alt-EJ caused by loss of c-NHEJ is a feature shared with homologous recombination (HR). Namely, at least two types of HR are more frequent in the absence of c-NHEJ factors: the conservative homology-directed repair (HDR) pathway that is mediated by the strand invasion factor RAD51, as well as the non-conservative single-stranded annealing (SSA) pathway [15, 16]. Since loss of c-NHEJ causes a substantial increase in the frequency of HDR, SSA, and Alt-EJ, bypass of c-NHEJ is likely an important step of these repair events. Although, such c-NHEJ bypass may not be an absolute requirement, since c-NHEJ is not necessarily the default pathway of DSB repair in all circumstances, such as distinct chromatin and cell cycle phase contexts. Nevertheless, c-NHEJ bypass likely enables DSB end resection that generates 3’ ssDNA. Accordingly, factors important for end resection, including CtIP and the Mre11-complex, mediate HDR, SSA, and Alt-EJ [12, 17–23]. The role of these factors is conserved in S. cerevisiae, in that the CtIP ortholog (SAE2) and MRE11 are important for Alt-EJ/MMEJ in this organism [24]. In contrast, factors that are implicated in promoting extensive end resection (e.g. BLM and EXO1) appear to favor HR over Alt-EJ [18, 25]. Conversely, factors that inhibit end resection, such as 53BP1 and RIF1, suppress HDR, SSA, and Alt-EJ [25–28].
Such inhibition of c-NHEJ to enable chromosomal break end resection for HR and/or Alt-EJ has implications for genome maintenance beyond DSB repair outcome. Namely, chromosomal breaks that occur at DNA replication forks can be one-ended DSBs, which if used during c-NHEJ can result in gross chromosomal rearrangements. In this context, bypassing c-NHEJ likely favors DSB end resection to support restart of replication forks, which could be mediated either by extensive stretches of homology, or by microhomology [18, 29, 30]. As another example, repair of DSB ends that are not readily ligated, such as those blocked by DNA or protein crosslinks [31], may also be facilitated by c-NHEJ bypass. Indeed, inhibition of c-NHEJ has been posited as a key function of some factors important for resistance to such DNA damage. For example, the requirement of the HR factor BRCA1 for cellular resistance to DNA damage caused by chemical inhibitors of poly ADP-ribose polymerase (PARP) can be partially rescued by loss of either the c-NHEJ factor Ku80, or the end resection inhibition factor 53BP1 [32, 33]. As another example, the role of Fanconi Anemia (FANC) factors for cellular resistance to DNA crosslinking agents has been shown to be suppressed by disrupting c-NHEJ in some experimental systems [34, 35], although not in others [33].
Apart from inhibiting c-NHEJ to favor end resection, other DNA damage response (DDR) factors have been implicated in distinct steps of Alt-EJ, including PARP. The influence of PARP on Alt-EJ has been examined with genetic disruption of PARP-1, as well as with catalytic inhibitors directed against PARP-1, which also likely act on multiple members of the PARP superfamily of proteins [36]. Specifically, inhibition of PARP causes reduced EJ frequency and radiosensitivity in Ku-deficient cells, and has been shown to reduce the frequency of DSB-induced chromosomal translocations [8, 37–39]. The role of PARP during EJ is unclear, but could include DSB end bridging and/or ligase recruitment [8, 37–39]. Apart from PARP, translesion polymerases have also been implicated in promoting insertion formation during Alt-EJ [5–8].
We have sought to identify other DDR factors that mediate Alt-EJ. Using a set of chromosomal break reporters, we screened an RNAi library targeting known DDR genes for siRNAs that cause a specific decrease in Alt-EJ, relative to an EJ event that is a combination of Alt-EJ and c-NHEJ (Distal-EJ between two tandem DSBs). From this analysis, we identified several DDR factors from diverse pathways that mediate Alt-EJ. Many of these factors also promote HR, including factors important for cellular resistance to DNA crosslinks, such as FANCA. Since bypassing c-NHEJ is important for both HR and Alt-EJ, we tested whether the role of Fanca during Alt-EJ in mouse cells is affected by c-NHEJ, and found that loss of Ku70 diminishes the influence of Fanca on Alt-EJ. In contrast, PARP-inhibition causes a reduction in Alt-EJ that is enhanced by Ku70 loss. Along these lines, we also found that the influence of the helicase/nuclease DNA2 during Alt-EJ and end resection is distinct from FANCA and PARP. Finally, we examined the effect of the cancer therapeutic Bortezomib on chromosomal break repair, since it has been identified as an inhibitor of FANC signaling [40, 41], and found that this small molecule also disrupts both HR and Alt-EJ, as well as end resection. We suggest that multiple distinct aspects of the DNA damage response are important for Alt-EJ, including FANCA mediated bypass of c-NHEJ.
To identify mediators of Alt-EJ, we sought to determine the influence of known DDR factors on a previously described Alt-EJ reporter, EJ2-GFP [12], integrated into human osteosarcoma cells (U2OS) [42]. In this reporter, the reading frame of a GFP cassette with an N-terminal tag is disrupted by an I-SceI site followed by stop codons in all frames, which is flanked by 8 nucleotides of microhomology (Fig. 1A). Alt-EJ of an I-SceI generated chromosomal DSB that deletes the stop codons, which occurs predominantly via the use of 8 nucleotides of flanking microhomology to bridge the DSB, restores the GFP cassette and causes a 35 nucleotide deletion [12]. Notably, since this predominant Alt-EJ event shows microhomology at the repair junction, and also is enhanced by loss of c-NHEJ factors [12], the terms MMEJ and Alt-EJ are both applicable to this repair event, and hence are largely interchangeable in this context. We examined a set of DDR factors on this Alt-EJ repair event using an RNAi library targeting 238 genes with pools of 4 siRNAs per gene, by pretreating cells with each siRNA pool prior to transient transfection with the I-SceI expression vector, and subsequent analysis of the frequency of GFP+ cells using flow cytometry.
We also performed a secondary screen, since siRNAs that cause a reduced frequency of Alt-EJ using the EJ2-GFP reporter could reflect not just a reduction in this particular repair outcome, but also an overall loss of DSB repair proficiency. For this secondary screen, we examined a distinct reporter for total EJ: EJ5-GFP [12, 42]. In this reporter, a GFP cassette is separated from its promoter by a marker gene that is flanked by two tandem I-SceI sites, such that EJ repair that uses distal ends of the two I-SceI-induced DSBs restores GFP expression (Distal-EJ, Fig. 1A). These Distal-EJ events reflect a composite of c-NHEJ and Alt-EJ, because the GFP cassette can be restored by diverse EJ repair junctions (e.g. I-SceI restoration, deletions, or insertions) [12, 21, 43], and hence can occur by multiple EJ mechanisms. Consistent with this notion, GFP restoration with this reporter is not dependent on c-NHEJ factors (e.g. Ku70 or Xrcc4) [12, 21, 43]. Thus, examining both the Alt-EJ (EJ2-GFP) and the Distal-EJ (EJ5-GFP) reporters provides a means to distinguish between specific effects on Alt-EJ, versus relatively nonspecific effects on DSB repair that are reflected in changes to the frequency of total EJ.
Accordingly, to identify factors that specifically affect Alt-EJ, we sought to identify siRNAs that cause not only a significant decrease in the frequency of Alt-EJ (EJ2-GFP) relative to non-targeting siRNA (siCTRL), but also cause a significantly greater fold-decrease in Alt-EJ versus Distal-EJ (EJ5-GFP). As an example of this approach, depletion of the end resection factor CtIP has been previously shown to cause both a significant decrease in Alt-EJ, as well as a greater decrease in Alt-EJ versus Distal-EJ [12], which is consistent with the conclusions of other studies on CtIP and EJ [18, 44]. We confirmed this result with an individual siRNA targeting CtIP (siRNA CtIP-1) and the CtIP siRNA pool from the library (Fig. 1B, P≤0.0008). Thus, we included siCtIP-1 as a positive control for our screen to identify siRNA pools that cause a decrease in Alt-EJ versus Distal-EJ.
To perform the screen, we determined the fold change caused by each siRNA pool on the frequency of Alt-EJ and Distal-EJ (N = 2), relative to parallel siCTRL treatments. We then calculated the ratio of this fold change on Alt-EJ versus Distal-EJ, and performed additional repeats of several siRNA pools that appeared to cause the greatest effects on the Alt-EJ/Distal-EJ ratio. We then ranked the siRNA pools according to this Alt-EJ/Distal-EJ ratio to complete the screen (S1 and S2 Tables, S1A Fig.). Although we did identify a few genes causing a relative increase in Alt-EJ, we focused on the siRNAs showing a low Alt-EJ/Distal-EJ ratio in order to identify mediators of Alt-EJ. For this, we examined individual siRNAs for the 15 genes showing the lowest Alt-EJ/Distal-EJ ratio. As well, since one of these genes was FAAP24, which is required for recruitment of FANCA to chromatin damaged by the crosslinking agent mitomycin-C [45], and since FANCA ranked in the top 10% of the screen, we also examined individual siRNAs for FANCA. From this analysis of 16 genes, we found that 13 showed at least two individual siRNAs that caused a significantly greater decrease in Alt-EJ compared to Distal-EJ: FANCA, FAAP24, NTHL1, UNG, SOD1, RUVBL2, TIP60/KAT5, RAD23B, GEN1, DNA2, MSH6, PRP19/PSO4, and POLA1 (Fig. 1C, P≤0.02, Alt-EJ/Distal-EJ ratio of this data in S1B Fig.). To validate the siRNAs from these Alt-EJ mediators, we confirmed depletion of the target mRNA using quantitative RT-PCR (qRT-PCR, Fig. 1D, S1 Table).
We then sought to identify commonalities among these 13 mediators of Alt-EJ by also examining their effect on HR. For this, we tested the siRNA pools in U2OS cells with reporters for HDR and SSA (Fig. 2A). DR-GFP is a reporter for HDR in that GFP expression is restored via repair that uses a homologous template (iGFP) for RAD51-dependent gene conversion [16, 46]. In contrast, SA-GFP is a reporter for SSA, in that GFP expression is restored by a repair event that uses two flanking homologous repeats to bridge the DSB, causing a deletion between the repeats [16]. For SA-GFP, GFP restoration could also occur by certain gene conversion events: long-tract gene conversion that is resolved by EJ (LTGC-EJ), or with crossing over. The relatively low frequency of these events, combined with the finding that RAD51 disruption causes an increase in GFP+ cells with this reporter, indicates that the restoration of GFP predominantly occurs via SSA [16]. However, we note that deficiencies in some DDR factors (e.g. BRCA1 and CtIP) have been shown to cause an increase in LTGC [47], such that it is important to also consider the possibility that the relative contribution of SSA versus LTGC-EJ could be altered by depletion of certain DDR factors. As our positive control for this analysis, we again used CtIP depletion that has been shown to cause a significant reduction in both HDR and SSA, compared to Distal-EJ [12], which we have confirmed (siRNA CtIP-1, Fig. 2B).
From this analysis, 6/13 factors showed a similar pattern to CtIP, in that RNAi of these factors caused a greater decrease in Alt-EJ, HDR, and SSA, each compared to Distal-EJ (P≤0.004): FANCA, FAAP24, TIP60/KAT5, PRP19/PSO4, RAD23B, POLA1 (Fig. 2B). An additional four factors showed a significant decrease in either HDR or SSA relative to Distal-EJ (P≤0.003): DNA2 (SSA only), SOD1 (HDR only), RUVBL2 (HDR only), and MSH6 (SSA only) (Fig. 2B). The remaining three showed a specific effect only on Alt-EJ (P<0.0001): UNG, NTHL1 and GEN1 (Fig. 2B).
Since Alt-EJ and HR appear more proficient in the S/G2 phases of the cell cycle [18, 48], we next considered that the above 13 Alt-EJ mediators might affect cell cycle phase distribution. In particular, we wanted to test whether depletion of these factors via siRNA causes a substantial shift to G1 phase cells, which could possibly cause a reduction in Alt-EJ and/or HR. Thus, following siRNA treatment (siCTRL, siRNA CtIP-1, and siRNA pools of the above 13 Alt-EJ mediators), we labeled S-phase cells using a pulse of bromodeoxyuridine (BrdU) incorporation, performed co-staining with the DNA dye propidium iodide (PI), and analyzed cells by flow cytometry. From this experiment, we found that siRNAs targeting a few of the genes caused a significant increase in G2 phase cells (Fig. 2C, FAAP24, DNA2, SOD1, and NTHL1, P<0.02), whereas only the siRNAs targeting UNG caused a significant, but modest, increase in G1 phase cells (Fig. 2C, UNG 56%, siCTRL 45%, P<0.005). These findings indicate that, apart from possibly UNG, the reduction in Alt-EJ and HR caused by depletion of the 13 genes described above cannot be readily attributed to an increase in G1 phase cells.
Since depleting both FANCA and FAAP24 caused a reduction in Alt-EJ and HR, and furthermore since we perform additional analysis below with FANCA, we considered that other FANC factors may also have a similar influence on Alt-EJ and/or HR. Namely, while siRNAs targeting other FANC genes did not cause the greatest decrease in Alt-EJ versus Distal-EJ relative to other genes in the library (S2 Table), we nevertheless considered the possibility that targeting these FANC genes may cause a statistical difference among different repair outcomes. For this, we examined the siRNA pools in the library that target several FANC genes, using each of the reporter assays described above. To begin with, we evaluated FANCD2, and for comparison, we included the FANCA siRNA pool and two individual FANCA siRNAs. From these experiments, we found that each of these siRNAs (siFANCA-P, siFANCA-3, siFANCA-4, siFANCD2-P) depleted the target protein as detected by immunoblot (Fig. 3A). Furthermore, each of these siRNAs showed a similar effect as the FANCA siRNA pool (siFANCA-P) on the reporter assays, in that Alt-EJ, HDR, and SSA were each reduced to a greater extent than Distal-EJ (Fig. 3B). To extend this analysis, we then examined each of the FANC core complex and associated factors represented in the library (FANCC, E, F, G, I, L, and M) [49]. We confirmed that each of these siRNA pools caused depletion of the target mRNA (S1 Table, Fig. 3A). From the reporter assay analysis, we found that siRNAs targeting FANCC, E, F, and M showed a similar pattern on repair as depletion of FAAP24, FANCA, and FANCD2 (i.e. a significant decrease in Alt-EJ, HDR, and SSA, relative to the effect on Distal-EJ, Figs 2B, 3B, 3C, P<0.01 for Alt-EJ and SSA, P<0.04 for HDR). In contrast, while siRNAs targeting FANCG, L, and I caused a significant decrease in SSA relative to Distal-EJ, only FANCG caused a relative decrease in Alt-EJ, and only FANCL caused a relative decrease in HDR (Fig. 3C, P<0.04). Although, the modest effects of siRNAs targeting FANCG, L, and I on repair outcome could reflect incomplete disruption of these factors, which is an inherent limitation of siRNA experiments. In any case, these results indicate that depletion of several FANC factors (C, D2, E, F, and M) show a similar pattern as depletion of FANCA and FAAP24 in causing a reduction in Alt-EJ and HR relative to Distal-EJ.
The above findings indicate that FANCA is important for Alt-EJ, HDR, and SSA, which are all repair events that involve bypass of c-NHEJ, in that these events are elevated in the absence of c-NHEJ factors, such as Ku70 and Xrcc4 [10, 12, 15, 21, 50]. Accordingly, these findings support the notion that the FANC pathway may be important for c-NHEJ bypass, which has been suggested by studies showing that the DNA crosslink sensitivity of FANC-deficient cells can be rescued by c-NHEJ disruption [34, 35]. However, this genetic interaction does not appear to be consistent among all experimental systems [33, 49]. Thus, to examine the genetic relationship of FANC and c-NHEJ, we sought to test how loss of c-NHEJ affects both Alt-EJ and DNA crosslinking sensitivity in FANC-deficient cells. For this, we used mouse embryonic stem (mES) cells, in which Ku is not essential for viability, in comparison to human cells that require Ku for survival, likely due to its role in telomere maintenance [51].
Specifically, we examined the influence of Fanca on Alt-EJ and DNA crosslink sensitivity in mES cells, both in the presence and absence of Ku70. For this, we integrated the EJ2-GFP and EJ5-GFP reporters into a previously described Fanca−/− mES cell line [51–53], which we confirmed involves deletion of exons 37 to 39, causing a frame-shift at position N1202 (Fig. 4A). Then, we transfected these cell lines with an expression vector for I-SceI, along with either an expression vector for Fanca, or the associated empty vector (EV). By immunoblotting with an anti-human FANCA antibody, we detected Fanca in cells transfected with the Fanca expression vector, although this antibody was unable to detect endogenous Fanca in WT cells (Fig. 4A). In comparison to parallel EV transfections of Fanca−/− cells, we found that expression of Fanca caused a significant increase in Alt-EJ (1.8 fold, P<0.0001), but not Distal-EJ (Fig. 4A), which is consistent with our above findings that FANCA promotes Alt-EJ in U2OS cells. To examine the effect of Ku70 on Alt-EJ in Fanca−/− cells we introduced frame-shift mutations in both alleles of Ku70 in the Fanca−/− EJ2-GFP cells using CAS9-mediated genome engineering [54], which we confirmed caused a loss of the Ku70 protein (Fig. 4B). With these cells, we expressed I-SceI, along with either a complementation vector for Fanca, Ku70, or EV (expression confirmed by immunoblotting, Fig. 4B). From this analysis, we found that loss of Ku70 in Fanca−/− cells caused an increase in Alt-EJ (Fig. 4C, 9-fold, P<0.0001), which was substantially reduced with transient expression of Ku70 (Fig. 4C, 3.7-fold, P<0.0001). The finding that Alt-EJ in Fanca−/−Ku70−/− cells was not completely reduced to the level of Fanca−/− cells by transient expression of Ku70 likely reflects limitations of this approach to precisely mimic endogenous expression, which is consistent with prior findings with transient Ku70 expression [16]. Using this pair of Fanca−/−Ku70−/− and Fanca−/− cells, we next examined the effect of transient Fanca expression on Alt-EJ. From these experiments, we found that the fold effect of Fanca expression on Alt-EJ is significantly reduced in Fanca−/−Ku70−/− cells, compared to Fanca−/− cells (Fig. 4C, 1.2-fold and 1.7 fold, respectively, P<0.0001). These results indicate that loss of Ku70 can diminish the influence of Fanca on Alt-EJ.
To further examine the genetic interaction between Fanca and Ku70 in response to DNA damage, we tested the effect of clastogen exposure on clonogenic survivial of the above described mES cell lines, as well as a previously described Ku70−/− cell line [55], and a WT line. Specifically, we exposed these cells to different concentrations of the DNA crosslinking agent cisplatin, or doses of ionizing radiation (IR), and compared colony formation to untreated cells (Fig. 4D). From this analysis, we found that Fanca−/− cells show substantial hypersensitivity to cisplatin, as compared to both WT and Ku70−/− (Fig. 4D, P≤0.006). Furthermore, while Ku70−/− cells were not obviously more sensitive to cisplatin than WT, the Fanca−/−Ku70−/− cells showed even greater cisplatin sensitivity than the Fanca−/−cells (Fig. 4D, P≤0.01). This latter finding is consistent with a previous study of Fancd2−/− and Fancd2−/−Ku80−/− mouse cells [33], but not in experiments from other systems [34, 35, 49]. In contrast to cisplatin response, both Ku70−/− and Fanca−/−Ku70−/− showed IR hypersensitivity, compared to both WT and Fanca−/− cells (Fig. 4D). As well, Fanca−/− cells did not show hypersensitivity to IR, and indeed were modestly IR resistant compared to WT (Fig. 4D, P<0.0001 at 3 Gy IR). Thus, Ku70 appears important for IR resistance regardless of Fanca, but is important for cisplatin resistance specifically in Fanca−/− cells. These findings indicate that loss of Ku70 can diminish the role of Fanca in promoting Alt-EJ, but not DNA crosslink resistance.
In contrast to the above findings with FANCA, PARP function during DSB repair has been suggested to be independent of c-NHEJ. Namely, chemical PARP inhibitors cause radiosensitivity and reduced EJ frequency, in a manner enhanced by Ku-deficiency [8, 37–39]. Thus, we considered whether PARP inhibition might have a distinct influence on DSB repair versus FANCA. We first examined the effect of the PARP inhibitor (PARPi) Olaparib, on each of the U2OS reporter cell lines described above (Figs 1A and 2A). Inhibition of PARP activity via Olaparib was confirmed by immunoblotting analysis of PARylated protein (Fig. 5A). Furthermore, we found that Olaparib treatment did not obviously affect cell cycle phase distribution (Fig. 5A). From the reporter assay analysis, we found that Olaparib treatment caused a significantly greater decrease in Alt-EJ compared to Distal-EJ (Fig. 5A, 1.7 fold and 1.3 fold, respectively, P<0.0001). In contrast, the effect of Olaparib on HDR was the same as Distal-EJ, and SSA was affected to a lesser degree than either (Fig. 5A). Thus, in contrast to FANCA and many other DDR factors that cause a similar effect on Alt-EJ and HR, the PARPi Olaparib appears to cause a relatively specific decrease in Alt-EJ.
This finding raised the possibility that the influence of FANCA versus Olaparib on Alt-EJ may be distinct, which we first tested co-treatment experiments with Olaparib and siFANCA in U2OS cells. From these experiments, we found that Olaparib and siFANCA each caused a significant decrease in Alt-EJ compared to Distal-EJ as shown above, and that the effect of combined treatment was approximately additive (Fig. 5B). We then tested whether the effect of Olaparib on Alt-EJ might be affected by Ku70 loss, using the mES cell lines described above. As with the U2OS experiments, inhibition of PARP activity in mES cells with Olaparib was confirmed by immunoblot analysis of PARylated protein (Fig. 5C). We found that Olaparib treatment caused a greater decrease on Alt-EJ in Ku70−/− cells compared to WT (Fig. 5C, 1.6-fold, P<0.001), and in Fanca−/−Ku70−/− cells compared to Fanca−/− cells (Fig. 5C, 2.4-fold, P<0.0001). Olaparib also appeared to cause a modestly greater decrease on Alt-EJ in Fanca−/− cells, compared to WT, although this effect was not statistically significant. In any case, these results indicate that the effect of PARP inhibition on Alt-EJ is enhanced by Ku70 loss, whereas the effect of Fanca disruption on Alt-EJ is diminished by Ku70 loss (Fig. 4). Accordingly, we suggest that Fanca disruption and Olaparib treatment have distinct effects on Alt-EJ.
Among the genes besides FANCA that we identified as mediating Alt-EJ, we sought to further examine DNA2, since the siRNA pool targeting DNA2 caused the greatest decrease in Alt-EJ/Distal-EJ ratio in the screen (S2 Table), and because this factor has been implicated in end resection [56]. Furthermore, DNA2 has been shown to co-immunoprecipitate with FANCD2 [57], and appears to influence the requirement for the FANC pathway in the response to DNA crosslinking agents [58]. To begin with, we examined the effect of expressing DNA2 on Alt-EJ in U2OS cells, using stable integration of an expression cassette for 3xFlag-tagged DNA2 with silent mutations for resistance to an individual DNA2 siRNA (siDNA2–4), which we compared to cells transduced with an EV (Fig. 6A, expression confirmed by Flag immunoblot). We then treated these cells with siDNA2–4 or siCTRL, and examined the frequency of Alt-EJ after I-SceI expression. In cells treated with siDNA2–4, we found that those expressing 3xFlag-DNA2 showed a significantly greater frequency of Alt-EJ compared to EV cells (Fig. 6A, 2-fold, P<0.0001). While Alt-EJ was not restored to siCTRL levels by 3xFlag-DNA2 expression, this finding is consistent with a previous report showing partial complementation of DNA2 function during genome maintenance using a similar experimental system [59]. We then tested whether the influence of DNA2 on Alt-EJ is distinct from that of PARPi and/or FANCA, using co-treatment experiments in the U2OS reporter cells. From these experiments, we found that siDNA2–4 treatment caused an approximately additive decrease in Alt-EJ with Olaparib treatment (Fig. 6B), as well as with siFANCA-3 or siFANCA-4 treatment (co-depletion of DNA2 and FANCA mRNA confirmed by RT-PCR, Fig. 6C, D).
The above findings that depletion of DNA2 causes an approximately additive decrease in Alt-EJ with either FANCA depletion or Olaparib treatment indicates that the influence of DNA2 on Alt-EJ may be distinct from FANCA or PARP inhibition. Accordingly, we sought to examine whether these factors may also have distinct effects on other aspects of the DNA damage response, specifically end resection. Namely, we considered that depletion of DNA2, but not FANCA or PARP inhibition, may cause a decrease in end resection. We base this hypothesis on findings in S. cerevisiae, as well as mammalian cells, that DNA2 mediates end resection [56, 57, 60]. To test this, we employed a recently described flow cytometry-based assay for end resection, which uses a ssDNA binding protein (RPA, specifically RPA32) as a marker of ssDNA formation [61]. Specifically, this assay measures chromatin-bound (i.e. detergent-resistant) ssDNA binding protein (RPA32), which is detected by RPA32 immunostaining, and combined with counterstaining with the DNA dye DAPI. This assay is designed to measure end resection at DNA replication forks that are blocked by the topoisomerase I poison camptothecin (CPT), and such end resection was shown to be dependent on CtIP [61]. A limitation of this assay is that end resection at blocked replication forks induced by CPT treatment is not necessarily equivalent to end resection at DSBs. Nevertheless, this assay provides a quantitative measure of end resection. To begin with, we confirmed previous findings that CPT treatment of U2OS cells significantly induced a population of cells with chromatin-bound RPA staining, which was abolished in cells treated with an siRNA targeting CtIP (Fig. 6E, F, S2A, B Fig.). We next examined cells treated with siRNAs targeting DNA2, as well as each of the other 13 Alt-EJ mediators (siRNA pools), and cells treated with Olaparib. We found that DNA2 depleted cells showed a shift towards G2 phase cells, which was consistent with the cell cycle analysis described above (Fig. 2C). Without CPT treatment, these G2 phase cells showed a significant increase in RPA signal, compared to siCTRL treated cells without CPT treatment (Fig. 6E, F, P<0.0001). Such RPA staining may reflect defects in processing replication intermediates in DNA2 deficient cells [59, 62]. After CPT treatment, the frequency of cells with RPA signal was significantly lower for DNA2 depleted cells compared to siCTRL treated cells (6E, F, P<0.0002), although they were not completely abolished as with CtIP depletion. In contrast, cells that were depleted of the other 12 Alt-EJ mediators we described above (including FANCA), or cells treated with Olaparib, did not show an obvious decrease in the frequency of cells with CPT-induced RPA signal, compared to siCTRL treated cells (Fig. 6E, F, S2B Fig.). These findings indicate that DNA2 depletion, but not FANCA depletion or PARP-inhibtion, causes a reduction in end resection by this assay. These findings support the notion that DNA2 has a distinct influence from FANCA and PARP-inhibition during Alt-EJ and end resection.
Since our above analysis indicates that disrupting the FANC signaling pathway can cause a reduction in both HR and Alt-EJ, we sought to test whether a small molecule inhibitor of this pathway could have a similar effect on repair. For this, we examined the small molecule 26S proteasome inhibitor Bortezomib [40], which has been shown to disrupt FANC signaling, inhibit HDR, and cause DNA crosslink sensitization [41, 63, 64]. Indeed, Bortezomib has been proposed as a possible therapeutic crosslink sensitizer for FANC proficient tumors [41, 63, 64], particularly since it is already used as a cancer therapeutic for multiple myeloma [40]. Thus, given these prior studies demonstrating Bortezomib as an inhibitor of the FANC pathway, and because of the clinical relevance of this small molecule as a cancer therapeutic, we sought to examine the influence of Bortezomib on the DDR and DSB repair outcomes. First, we tested whether Bortezomib treatment affects FANC signaling by examining FANCD2 accumulation into ionizing radiation induced foci (IRIF, DSB induction confirmed by γH2AX staining, S3 Fig.). We found that Bortezomib treatment caused a marked reduction in the frequency of FANCD2 IRIF, which is similar to the effects of FANCA depletion (siFANCA-3) (Fig. 7A, 5-fold and 3.5-fold respectively), and is consistent with previous studies [41]. Next, we tested the effect of Bortezomib treatment on the DSB reporter assays in U2OS cells, using a concentration that was sublethal over the course of the experiment (17 nM). From this experiment, we found that Bortezomib treatment caused a significantly greater decrease in Alt-EJ, HDR, and SSA, compared to Distal-EJ (Fig. 7B, P<0.004). Thus, Bortezomib treatment causes a similar pattern on DSB repair as depletion of FANCA and several other FANC factors (Fig. 3).
However, since Bortezomib inhibits the FANC pathway as a proteasome inhibitor [41], we considered that the effects of Bortezomib on Alt-EJ may not necessarily be dependent on the FANC pathway, and furthermore that Bortezomib may disrupt other parts of the DDR. To begin with, we examined the effect of Bortezomib treatment on Alt-EJ in cells that had been treated with siRNAs targeting FANCA, DNA2 and FANCD2, and found Bortezomib treatment caused a reduction in Alt-EJ in each of these instances (Fig. 7C). Thus, the reduction in Alt-EJ via Bortezomib treatment is not dependent on the FANC pathway. Accordingly, we next examined whether Bortezomib treatment may affect other aspects of the DDR. For this, we treated cells with Bortezomib with the same conditions as the FANCD2 IRIF experiment (Fig. 7A), and examined cell cycle distribution and end resection, using the flow cytometry assays described above (Figs 2C, 6E, respectively). From this analysis, we found that Bortezomib treatment did not obviously affect cell cycle phase distribution or BrdU incorporation, but caused a striking loss of end resection (Fig. 7D, cells with RPA staining after co-treatment of CPT and Bortezomib was substantially reduced compared to CPT treatment alone, P<0.0001). Thus, Bortezomib treatment not only disrupts FANC signaling, but also inhibits end resection, HR, and Alt-EJ.
We have sought to characterize factors that influence Alt-EJ, since many cancer-associated chromosomal rearrangements show hallmarks of Alt-EJ (e.g. microhomology) [2–4], and since this DSB repair pathway likely affects cellular responses to clastogenic therapeutics. From a targeted RNAi screen of known DDR factors, we identified 13 genes that mediate Alt-EJ, and found that the majority of these also mediate HR (HDR and/or SSA), as shown previously for the end resection factor CtIP [12, 18]. While some of these factors had not been clearly identified as HR mediators (FAAP24, RAD23B, POLA1, SOD1), the other factors have been shown to influence HR. The TIP60 chromatin remodeling complex, which includes TIP60/KAT5 and RUVBL2, has been shown to promote HDR [65], and may function to acetylate nucleosomes proximal to DSBs to block recruitment of 53BP1 [66], which is a factor that inhibits HDR, SSA, and Alt-EJ [25–28]. The nuclease/helicase DNA2 is a major end resection factor in S. cerevisae [56], and has been shown to promote HR and end resection in mammalian cells [57, 60], which we have confirmed here. The influence of the mismatch repair factor MSH6 on HR is somewhat complex, as mismatch repair has both pro and anti-HR activities [21, 67], although our finding here that MSH6 promotes SSA is consistent with findings in S. cerevisae [68]. Factors important for the FANC pathway, including FANCA, have been shown to promote HDR and SSA [52, 69]. Finally, the ubiquitin ligase PRP19/PSO4 has been shown in a recent study to promote HDR [70].
This finding that Alt-EJ and HR share several common mediators supports a model whereby Alt-EJ may often be important for completing repair of aborted HR initiation events. Namely, DDR factors that mediate both HR and Alt-EJ may promote initiation of HR under circumstances where HR is not readily finished (i.e. absence of the sister chromatid), which would then rely on Alt-EJ to guard against chromosome loss. An important initiation step of HR or Alt-EJ is likely bypass of c-NHEJ, since c-NHEJ-deficient cells (e.g. Ku70 and XRCC4 deficient) show substantially elevated levels of HDR, SSA, and Alt-EJ [9–13]. Accordingly, initiation of HR or Alt-EJ may be important under conditions whereby c-NHEJ is not feasible, such as DSB ends that are blocked by DNA crosslinks [31] or base damage, or for one-ended DSBs that arise during DNA replication. Consistent with this model, we have found that numerous factors important for DNA crosslink resistance (i.e. FAAP24, several FANC factors, and PRP19/PSO4) [49, 71, 72], appear to mediate HDR, SSA, and Alt-EJ.
Along these lines, a role for the FANC pathway in bypassing c-NHEJ was suggested by studies showing that the DNA crosslink sensitivity of FANC-deficient cells can be rescued by loss of c-NHEJ. Specifically, such suppression of DNA crosslink hypersensitivity was shown for FANCC-deficient chicken DT40 cells with deletion of Ku70 [34], Fancd2-deficient C. elegans with loss of Ligase IV, and FANCA or FANCD2-deficient human cells with disruption of DNA-PKcs [35]. Additional evidence that FANCA may influence EJ includes reduced DNA ligation in cell-free extracts from FANCA-deficient human cells [73], as well as altered class switch recombination junctions (greater intra-switch recombination) in Fanca−/− mice [74]. Our findings support the notion that the FANC pathway may inhibit c-NHEJ, in that FANCA promotes several repair events that are enhanced by loss of c-NHEJ (Alt-EJ, SSA, and HDR), and that the influence of Fanca on Alt-EJ in mES cells is diminished by loss of Ku70. The mechanism of such c-NHEJ inhibition is unclear, which in part reflects our relatively limited understanding of how c-NHEJ factors may be blocked from associating with DSBs and/or evicted from DSBs, the latter of which may include targeted degradation of c-NHEJ factors [75]. Distinct from these findings with Alt-EJ, we found that loss of Ku70 did not rescue DNA crosslink (cisplatin) sensitivity of Fanca-deficient cells, and indeed caused hypersensitivity, which is consistent with previous findings with Fancd2−/−Ku80−/− mouse cells [33]. Apart from differences in the specific FANC and c-NHEJ-deficiencies tested in different studies, the distinct cellular toxicity to DNA crosslinking agents might also reflect the diverse mechanisms of cell death caused by such clastogens (e.g. apoptosis, necrosis, and mitotic catastrophe) [76], which could be affected by cell type and experimental conditions. In summary, we suggest that the FANC pathway plays a role in bypass of c-NHEJ that is important for DSB repair and cellular response to DNA crosslinks, and hence this function could be a therapeutic target for tumor sensitization to these clastogens. However, the FANC pathway is likely important for additional aspects of the cellular response to DNA crosslinks, such that loss of the c-NHEJ pathway is not sufficient in all systems to rescue DNA crosslinking hypersensitivity.
Apart from inhibition of c-NHEJ per se, the initiation of end resection is also likely an important step of Alt-EJ, since we found that the factors CtIP and DNA2 appear to mediate Alt-EJ, as well as end resection at replication forks blocked by CPT. Although, the influence of DNA2 on this assay is distinct from CtIP, in that CPT-induced RPA staining was abolished with CtIP depletion, but not DNA2 depletion. Furthermore, DNA2 depleted cells without CPT treatment showed elevated levels of chromatin-bound RPA in G2 phase cells, as compared to siCTRL treated cells, which may reflect ssDNA caused by defects in processing DNA replication intermediates [59, 62]. Along these lines, the end resection associated with blocked DNA replication that is measured by this assay (i.e. chromatin-bound RPA induced by CPT) may not necessarily be equivalent to end resection that mediates Alt-EJ, or even HR of DSBs. For instance, relatively extensive end resection may be required to generate substantial RPA staining, but such end resection may not be required for Alt-EJ. Furthermore, c-NHEJ bypass to facilitate Alt-EJ or HR of DSBs may not be rate limiting for end resection at replication forks that are blocked by CPT. Nevertheless, the role of DNA2 during Alt-EJ and end resection appears distinct from that of FANCA, since FANCA depletion did not obviously cause a loss of end resection, and furthermore since depletion of FANCA and DNA2 caused an approximately additive decrease in Alt-EJ.
Similar to FANCA, the remaining Alt-EJ mediators that we examined did not cause an obvious effect on end resection, at least by the assay system employed in our study, which indicates that some of these factors may play distinct roles during Alt-EJ. Certainly, it is conceivable that such factors are important for bypass of c-NHEJ, as we have suggested for FANCA, but of course they could also influence other aspects of this pathway that is currently poorly understood, such as short-range 5’ DSB end resection that does not cause RPA recruitment, annealing of microhomology, DNA synthesis that is primed by the annealed microhomology, ssDNA tail processing, and/or ligation. For example POLA1, the catalytic subunit of Pol α, could conceivably initiate microhomology-mediated DNA synthesis to facilitate DSB end synapsis, which is consistent with the function of Pol α to synthesize short DNA primers during initiation of DNA replication [77]. In related findings from S. cerevisiae, repair synthesis primed by microhomology has been proposed to be mediated by another replicative polymerase (Pol δ), based on findings with the non-essential subunit POL32 [78]. Considering another step of Alt-EJ, the GEN1 nuclease could conceivably promote short-range 5’ end resection, based on its 5’ flap cleavage activity [79, 80]. Additionally, we found that Alt-EJ is mediated by the UNG and NTHL1 DNA glycosylases [81, 82]. The role of these particular factors during Alt-EJ is unclear, although notably UNG is also important for a programmed EJ event, class switch recombination, which may involve non-catalytic functions of UNG [81]. Another possible role for DNA glycosylases during Alt-EJ could be cleavage of nucleotide base damage near DSBs, since such processed DSB ends may be more prone to Alt-EJ.
In addition to these factors, we also find that inhibiting PARP activity causes a specific decrease in Alt-EJ relative to HR, indicating that PARP inhibition does not affect all repair events that are inhibited by c-NHEJ. Indeed, we found that the effect of PARPi on Alt-EJ is enhanced by loss of c-NHEJ (e.g. Ku70-deficiency), which is consistent with other reports showing that PARPi causes radiosensitization and end-joining defects that are amplified by Ku-deficiency [37–39]. Furthermore, PARPi effects on Alt-EJ were approximately additive with loss of DNA2 and FANCA, indicating that PARPi affects a distinct step during repair. These findings support a c-NHEJ-independent influence of PARPi on Alt-EJ, which could include direct effects on end bridging and/or ligation [8, 37–39]. With respect to mechanism, we note that PARPi treatment is not equivalent to genetic loss of PARP-1 and PARP-2, and may function to block repair by trapping PARP complexes on DNA damage [25]. In summary, we suggest that several DDR factors function to mediate both HR and Alt-EJ, which may be important for cellular response to DNA damage not readily repaired by c-NHEJ, but that Alt-EJ has multiple mechanistic requirements that are distinct from HR.
Finally, we also examined the effect of the proteosome inhibitor Bortezomib on DSB repair, since this small molecule has been identified as an inhibitor of FANC signaling and HDR, and hence has been proposed as a DNA crosslink sensitizer for FANC proficient tumors [41, 63, 64]. Our findings support this concept, since we confirmed that Bortezomib inhibits FANC signaling. However, we also found that this small molecule substantially inhibits end resection, and furthermore causes a specific decrease in Alt-EJ and HR, relative to Distal-EJ. We suggest that the influence of Bortezomib on the DDR, and hence on DNA crosslink sensitivity, may not be limited to disrupting FANC signaling, but could include blocking Alt-EJ, HR, and/or end resection. In general, we suggest that blocking the initiation of not only HR, but also Alt-EJ, should be considered in developing therapeutics for DNA crosslink sensitization.
Establishment and culturing of U2OS reporter cell lines were described previously [42]. The WT mES cell line was acquired from ATCC (J1 strain), and the Fanca−/− cell line was generously provided by Drs. Maria Jasin and Koji Nakanishi [53, 69], and the Ku70−/− EJ2-GFP reporter cell line was described previously [12, 55]. The Fanca RT-PCR genotyping primers were 5’gtgtggtcggtggatgagat and 5’aacagctgaggctcctggta. The mES cells were cultured and used for integration of the EJ2-GFP and EJ5-GFP reporters at the pim1 locus as described [12]. The Fanca−/−Ku70−/− EJ2-GFP cell line was generated using the expression plasmid pX330 [54] (generously deposited by Dr. Feng Zhang, Addgene 42230) for the CAS9 nuclease and gRNA with the following sequence introduced at the 5’ end of the gRNA: 5’gccatgggggtcgtcttcat, which targets a site in exon 5 of Ku70. This CAS9/Ku70-gRNA plasmid (0.5 μg) was co-transfected with a dsRED expression plasmid (0.1 μg, Clontech) into the Fanca−/− EJ2-GFP cell line using Lipofectamine 2000 (Invitrogen, 1.8 μl, total transfection volume 0.6 ml), and subsequently dsRED+ cells were sorted by fluorescence activated cell sorting, and seeded to isolate individual colonies. The genomic DNA from individual colonies was examined by PCR amplification (Platinum Hi-Fi Supermix, Invitrogen), using Kuex4UP 5’agatttggacaacccaggtc and Kuex5DN 5’gaggtcgctggctttggt, and Bbs1 digestion analysis to identify clones with mutations in the Bbs1 recognition site. Sequencing analysis of Bbs1 resistant PCR products identified a heterozygous clone with a frameshift mutation, which was re-transfected with dsRED and CAS9 plasmids and analyzed as above to identify a clone with loss of the Bbs1 site in the second allele that also caused a frameshift mutation, as determined by subcloning of the PCR product for sequencing analysis.
The library of pools of 4 siRNAs targeting 238 DDR factors is based on a commercially designed library (Dharmacon, G-006005). The sequence of each siRNA shown in Figs 1 and 3 are listed in S1 Table. The pCAGGS-BSKX empty vector and expression vectors for Ku70, I-SceI (pCBASce), and GFP (pCAGGS-NZE-GFP) were described previously [12]. The Fanca expression vector was generated by inserting the coding sequence from IMAGE clone 9087616 (Open Biosystems) into pCAGGS-BSKX. The 3xFlag-DNA2 vector was derived from pBABE-3xFlag-DNA2 (generously deposited by Dr. Sheila Stewart, Addgene 31955) [59] by introducing silent mutations for resistance to siDNA2–4 (5’TGATATcGAcACtCCtcTA), and subcloning into pMX-IRES-neo (Cell Biolabs). These pMX vectors were transfected with packaging plasmids (Addgene 8454 and 8449) [83] into 293T/17 cells (ATCC), and media from these transfected cells was filtered (0.45 μm) and used to treat the U2OS EJ2-GFP cell line, and integrants were subsequently selected using G418 (0.5 mg/ml).
Similar siRNA transfections were used for both DSB reporter assays and IRIF analysis: 0.5 × 105 U2OS cells were plated in 0.5 ml antibiotic-free media on a 24 well plate with 5 pmol of each siRNA incubated with 1.8 μl RNAiMAX (Invitrogen), and cultured overnight (20 hrs). For the reporter assays, following the overnight RNAi treatment, cells were transfected with 0.5 μg of the I-SceI expression vector (pCBASce) using 1.8 μl Lipofectamine 2000 in 0.6 ml antibiotic-free media. The transfection media was removed after 3 hrs and replaced with antibiotic media. For mES cells, 0.3 × 105 cells were plated 20 hours prior to transfection with 0.4 μg I-SceI and 0.2 μg of a second vector (expression vector for Fanca, Ku70, or EV). Transient GFP expression transfections used 0.4 μg of pCAGGS-NZE-GFP in place of pCBASce. For drug treatment, Olaparib (AZD2281, Selleck Chemicals, S1060), Bortezomib (Santa Cruz Biotech, sc-217785), or vehicle (DMSO) was added immediately after removing the transfection complexes (4 hrs for mES, 3 hrs for U2OS) for a continued treatment for the rest of the experiment. Three days after the plasmid transfections, GFP+ frequencies were determined by flow cytometery using a CyAn ADP Analyzer (Beckman Coulter, Inc.), as described [42]. The GFP+ frequencies shown relative to a control sample were calculated by dividing the GFP+ frequency for each transfection by the mean value for the control samples treated in parallel (siCTRL, EV, and/or DMSO treated). Each repair value is the mean of at least three independent transfections, error bars reflect the standard deviation, and statistics were performed with the unpaired t-test. Error bars denote the standard deviation from the mean.
For IRIF analysis, following overnight siRNA treatment (20 hrs), cells were plated onto chamber slides and cultured for a second day before IR treatment. For experiments without siRNA, cells were plated directly on chamber slides, and the next day were pre-treated with 100 nM Bortezomib or vehicle (DMSO) for 4 hrs prior to IR. Slides were treated with 10 Gy of IR (Gammacell 3000), allowed to recover for 4 hr (with Bortezomib or vehicle when appropriate), fixed with 4% paraformaldehyde and treated with 0.1 M glycine and 0.5% triton-X 100 prior to probing with antibodies against γH2AX (Abcam, ab18311) and FANCD2 (Abcam, ab2187), followed by secondary antibodies (Invitrogen/Life Technologies, A-11036 and A11029), and with DAPI using Vectashield Mounting Medium (Vector Laboratories). Images were acquired using a BX-50 (Olympus) microscope with Image-Pro software. At least 100 cells that showed γH2AX foci were scored for >20 FANCD2 IRIF from at least three independent treatments per condition. Statistics were performed as for the reporter assays.
For qRT-PCR analysis to examine RNAi depletion of target mRNAs, total RNA was isolated 2 days after siRNA transfection (Qiagen RNAeasy Plus) and reverse transcribed with MMLV-RT (Promega). The RT reactions were amplified with primers for the target mRNA (primers in S1 Table) and actin (5’actgggacgacatggagaag and 5’aggaaggaaggctggaagag) using SYBR Select Master Mix (Life Technologies) and quantified on a ViiA 7 Real Time PCR System (Life Technologies). Fold depletion for each siRNA treatment was determined as 2ΔΔCt, for which the cycle threshold (Ct) value for the target mRNA was subtracted by Ct value for actin (mean of duplicate amplifications from the same RT reaction) to calculate the ΔCt value, which was then subtracted from the corresponding ΔCt from siCTRL treated cells to calculate ΔΔCt.
Cells were lysed with NETN (20 mM Tris pH 8, 100 mM NaCl, 1 mM EDTA, 0.5% IGEPAL, 1.25 mM DTT and Roche Protease Inhibitor) or RIPA (Sigma R0278 with 1.25 mM DTT and Roche Protease Inhibitor Cocktail, Roche PhosStop for U2OS, for PAR analysis), and using several freeze/thaw cycles or sonication (QSonica Q800RS ultrasonic horn). Blots of these extracts were probed with antibodies against CtIP (Santa Cruz Biotech, sc-5970), FANCA (Bethyl Laboratories, Inc., A301–980A to detect human FANCA, and Santa Cruz Biotech, sc-23612 to detect expression of mouse Fanca), FANCD2 (Abcam, ab2187), Ku70 (Santa Cruz Biotech, sc-1487), PAR (Trevigen, 4335-MC-100), Flag (Sigma, A8592), actin (Sigma, A2066), and HRP-conjugated secondary antibodies (Santa Cruz Biotech, sc-2004 and sc-2005). ECL reagent (Amersham Biosciences) was used to visualize HRP signals.
For cisplatin sensitivity, mES cells were plated one day prior to continual exposure to 0.12 μM or 0.25 μM cisplatin (Pfizer NOC-0069-0081-01) for one week to form colonies. For IR sensitivity, mES cells were treated in suspension with 1 Gy or 3 Gy IR (Gammacell 3000) and plated for one week to form colonies. Colonies were fixed (10% Acetic Acid, 10% Methanol), stained with 1% crystal violet, and counted under the miscroscope (10X). To calculate fraction clonogenic survival, the number of colonies per well were normalized to the number of cells plated, and this colony forming value for each treated well was divided by the mean value of parallel untreated plates. Each clonogenic survival value represents the mean of at least six independent treatments, and error bars denote the standard deviation.
Cells were treated with siRNAs as for the reporter assays, but scaled up to one well of a 6 well plate, and cultured for two days prior to either end resection and cell cycle analysis. For drug treatments, 100 nM bortezomib and 5 μM Olaparib were added prior to performing the assays (4 hr and 20 hr, respectively). For the end resection assay, cells were treated with 1μM camptothecin for 1 hr prior to harvesting for RPA and DAPI staining, as described [61]. Briefly, cell pellets were detergent extracted in 100 μl of 0.2% Triton X-100 in PBS on ice for 7 minutes, and washed with 1 ml of BSA-PBS (0.1% BSA in PBS). Next, cells were fixed with 100 μl BD Cytofix/Cytoperm buffer (BD Biosciences) for 15 min, washed with BSA-PBS, and incubated in 50 μl of BD Perm/Wash buffer (BD Biosciences) with 1:200 RPA antibody (RPA2 9H8 Abcam ab2175) for 1 hour. Cells were washed with BSA-PBS and re-suspended in 50 μl BD Perm/Wash buffer with 1:200 secondary antibody (goat anti mouse Alexa Fluor 488, Life Technologies A11029) for 30 min. Finally, cells were washed with BSA-PBS and re-suspended in 0.3 ml PBS with 0.02% sodium azide, 250 μg/ml RNase A (Sigma R4642) and 2 μg/ml DAPI (Sigma D8417) for 30 min at 37°C. For cell cycle analysis, cells were incubated with 10 mM bromodeoxyuridine (BrdU, Sigma B5002) for 30 min prior to harvesting, fixed in 70% ethanol, stained with FITC-conjugated anti-BrdU antibody (BD Biosciences, 51–33284X), and incubated with propidium iodide (PI, Sigma P4170) and RNase A (Sigma R4642). Stained cells for both assays were analyzed on a CyAn ADP Flow Cytometer.
|
10.1371/journal.pgen.1006220 | Mutation of Growth Arrest Specific 8 Reveals a Role in Motile Cilia Function and Human Disease | Ciliopathies are genetic disorders arising from dysfunction of microtubule-based cellular appendages called cilia. Different cilia types possess distinct stereotypic microtubule doublet arrangements with non-motile or ‘primary’ cilia having a 9+0 and motile cilia have a 9+2 array of microtubule doublets. Primary cilia are critical sensory and signaling centers needed for normal mammalian development. Defects in their structure/function result in a spectrum of clinical and developmental pathologies including abnormal neural tube and limb patterning. Altered patterning phenotypes in the limb and neural tube are due to perturbations in the hedgehog (Hh) signaling pathway. Motile cilia are important in fluid movement and defects in motility result in chronic respiratory infections, altered left-right asymmetry, and infertility. These features are the hallmarks of Primary Ciliary Dyskinesia (PCD, OMIM 244400). While mutations in several genes are associated with PCD in patients and animal models, the genetic lesion in many cases is unknown. We assessed the in vivo functions of Growth Arrest Specific 8 (GAS8). GAS8 shares strong sequence similarity with the Chlamydomonas Nexin-Dynein Regulatory Complex (NDRC) protein 4 (DRC4) where it is needed for proper flagella motility. In mammalian cells, the GAS8 protein localizes not only to the microtubule axoneme of motile cilia, but also to the base of non-motile cilia. Gas8 was recently implicated in the Hh signaling pathway as a regulator of Smoothened trafficking into the cilium. Here, we generate the first mouse with a Gas8 mutation and show that it causes severe PCD phenotypes; however, there were no overt Hh pathway phenotypes. In addition, we identified two human patients with missense variants in Gas8. Rescue experiments in Chlamydomonas revealed a subtle defect in swim velocity compared to controls. Further experiments using CRISPR/Cas9 homology driven repair (HDR) to generate one of these human missense variants in mice demonstrated that this allele is likely pathogenic.
| Growth-Arrest Specific 8 (Gas8) is implicated in dual roles at both the primary cilium to regulate hedgehog signaling and in motile cilia to coordinate cilia movement. To investigate these roles in vivo, we created a Gas8 genetrap mutant mouse. Though no overt primary cilia phenotypes were evident in the Gas8 genetrap mutant mice, there were severe motility defects and the mice presented with Primary Ciliary Dyskinesia (PCD) like symptoms including situs inversus and hydrocephalus. We also identified two potential disease causing GAS8 missense variants (A391V and E199K) in humans. Utilizing CRISPR/Cas9 we generated a mouse to mimic the A391V allele. When we crossed the Gas8AV mutants with the Gas8GT mutant, the compound Gas8GT/AV heterozygous animals developed mild hydrocephalus. Rescue experiments using Chlamydomonas with mutations in the Gas8 homolog revealed only a modest decrease in swim velocity raising the possibility that the E199K allele is not pathogenic.
| Primary cilia are solitary and immotile cellular appendages that serve as signaling hubs for pathways such as Hedgehog (Hh) during development [1]. Motile cilia initiate and maintain fluid flow and are critical in the brain for cerebral spinal fluid flow and are necessary for mucus transport in the lungs [2]. During development, motile cilia are responsible for initiating flow at the embryonic node which is critical for setting up left-right asymmetry in the mammalian body [3–5].
While all cilia have common core components such as tubulin and intraflagellar transport proteins, motile cilia possess several accessory structures such as inner dynein arms (IDAs), outer dynein arms (ODAs), radial spokes, and the nexin-dynein regulatory complex (N-DRC). In Chlamydomonas reinhardtii, data indicate that the N-DRC functions to link the A microtubule of one doublet with the B microtubule of the adjacent doublet. It coordinates the activities of the outer and inner dynein arms to regulate flagellar beat frequency and waveform [6,7]. Studies in Chlamydomonas have led to the identification of several N-DRC proteins many of which appear to be conserved in mammals [8,9]. As in Chlamydomonas, mutations in putative mammalian N-DRC proteins CCDC164 (DRC1), CCDC65 (DRC2), and most recently, GAS8 (DRC4) are correlated with defects in ciliary motility [10–13].
The human homolog of Gas8 was originally identified in human breast cancer and referred to as Growth Arrest Specific 11 (GAS11) [14]. This gene shares 56% protein identity to an N-DRC component in Chlamydomonas known as DRC4, the protein product of the paralyzed flagella 2 (PF2) gene [15]. Loss of PF2 (DRC4) in Chlamydomonas leads to loss of IDAs and the majority of the N-DRC (N-DRC proteins DRC3-7) visible by transmission electron microscopy (TEM) and results in a slower forward swimming velocity and defective waveform [6,16,17]. The function of Gas8 in the mammalian N-DRC remains poorly understood.
Gas8 localizes to the axoneme of motile cilia and also to the base of primary cilia in vertebrate cells [18]. This led us to question if Gas8 serves as an N-DRC component in motile cilia and whether it has a separate role in non-motile primary cilia. This possibility is supported by data from knockdown studies of Gas8 in NIH3T3 cells showing defects in Hh pathway responses. Expression of truncated versions of Gas8, after knockdown of endogenous Gas8, revealed that the C-terminal region of Gas8 bound to and facilitated the transport of Smoothened into the cilium in response to Hh pathway activation using the Smoothened agonist (SAG) [18,19]. In mammals, cilia are essential for normal regulation of Hh signaling activity with many of the Hh signaling components such as Smoothened, Patched and Gli transcription factors dynamically localizing in primary cilia [20–22].
Primary Ciliary Dyskinesia (PCD, OMIM #244400) is a human disease characterized by abnormal motile cilia. PCD patients exhibit bronchiectasis, infertility, and chronic respiratory infections, and in some cases can present with hydrocephalus. A subset of PCD patients will also have a reversal of their left-right body axis that includes situs inversus totalis which is referred to as Kartagener syndrome [23]. PCD patients often have changes in cilia axonemal ultrastructure that include defects in the inner or outer dynein arms, central complex, radial spokes, and the N-DRC [24–28]. These structural defects alter ciliary beat frequency (CBF), ciliary waveform, and cilia orientation. Recent data indicate that mutations in the putative mammalian N-DRC components CCDC164, CCDC65, and Gas8 correlate with the clinical presentation of PCD. Mutations in these genes lead to dyskinetic cilia with subtle changes in cilia ultrastructure pointing to an importance for these components in ciliary motility. Other proteins such as CCDC39 and CCDC40 are responsible for the assembly and attachment of the IDAs and N-DRC in motile cilia. The absence of these proteins results in severe motility defects [10,11,29–31].
In this study, we investigate a role for Gas8 in both primary and motile cilia in vivo. For this we generated a Gas8 genetrap mutant mouse. Gas8 mutants present with severe hydrocephalus and cilia motility defects on both the ependyma and trachea, as well as a situs inversus phenotype. Given the role for Gas8 in cilia motility and recent data suggesting it is a PCD causing allele, we screened human PCD patients for GAS8 mutations and identified two independent missense variants. The potential pathogenicity of these alleles was tested by rescue experiments in Chlamydomonas PF2 mutants and by generating a mouse model for one of the variants. In contrast to the PCD phenotypes, we did not observe Hh associated defects in any of the mutant mice or cell lines derived from them or other phenotypes typically associated with defects in primary cilia function. These results suggest that GAS8 plays a highly conserved role in ciliary motility and mutations in Gas8 are associated with human disease through their impact on motile cilia.
A β-geo cassette containing the β-galactosidase enzyme, a neomycin resistance cassette, an N-terminal splice acceptor and poly-A tail was inserted in intron 7 of the Gas8 mouse allele (Fig 1A, herein referred to as Gas8GT). RT-PCR analysis using primers located before the genetrap insertion indicates that the 5’ end of the transcript is generated (Fig 1B, left). In contrast, primers located 3’ to the insertion failed to detect any Gas8 mRNA (Fig 1B, right).
Western blot analysis shows a product of expected size (57kDa) in wildtype and heterozygous Gas8 mice. This product is absent in homozygous mutants (Fig 1C). Additionally, the Gas8::β-geo fusion protein is detected (approx. 230kDa) in heterozygous and homozygous mutants indicating that the genetrap allele is being transcribed and translated. Loss of Gas8 led to lethality at approximately postnatal day 14 (P14) with few living to P21. All mutants presented with severe hydrocephalus (Fig 1D). Gas8GT mutant mice also presented with situs inversus at a rate of 36% (6 of 16 mutants) in live births based on position of the heart and stomach (Fig 1E and 1F). Both the hydrocephalic and situs inversus phenotypes suggested a defect in the function of motile cilia.
Based on a previous study reporting Gas8 as a positive effector of Hh signaling in mammals, we anticipated Gas8GT mice would present with phenotypes related to Hh signaling defects, especially since this mutation would lack the putative Smo binding domain (amino acids 386–478). However, we did not observe any hedgehog-associated phenotypes in limb patterning or neural tube formation. To further test a role for Gas8 in the Hh pathway, we isolated Mouse Embryonic Fibroblasts (MEFs) from Gas8WT and Gas8GT mutant mice and treated them with 150nM Smoothened agonist (SAG). The MEFs were then immunolabeled for Smo and acetylated tubulin to analyze differences in Smo trafficking into the cilium (Fig 2A). In contrast to the outcome of the knockdown studies, there was no difference between the amounts of Smo present in the cilia of Gas8GT mutants when compared to Gas8WT cilia (Fig 2B). These data indicate that a least in the Gas8GT mutants, Gas8 is not an essential factor involved in regulating Smo cilia trafficking. Similarly, none of the Gas8GT mutants exhibited defects in dorsal ventral patterning of the neural tube typical of altered Hh signaling (Fig 2C).
To investigate the hydrocephalus phenotype, cilia morphology, ultrastructure and motility on ependymal and tracheal cells was assessed. DIC analysis and immunofluorescence staining of trachea indicate motile cilia are present on the epithelium, but the Gas8GT protein fails to localize to these cilia (Fig 3). We counted cilia from trachea TEMs for broken doublet rings and found that about 9% of cilia from Gas8GT mutants showed disorganization of the arrangement of the microtubule doublets (Fig 4A arrowhead and 4E). To analyze ultrastructure within the doublets, we averaged 202 doublets of both genotypes to reduce variability due to random sectioning of the 96nm repeat of the microtubule doublet. We did not observe any major structural differences in the inner or outer dynein arms (Fig 4B). However, high speed video and Fourier transformation analysis revealed that cilia are largely static with only a few moving (Fig 4C and 4D, S1 Movie and S2 Movie). Those cilia that did moved were dyskinetic, resulting in an inability of cilia to propel fluid as seen by tracking of fluorescent beads added to either brain ventricle or trachea preparations (Fig 4F and 4G). Beat frequency of cilia that remained motile in Gas8GT mutants was modestly decreased from 17.0Hz in Gas8WT to 12.7Hz in Gas8GT (Fig 4H). Cilia length is also affected in Gas8GT mice, with Gas8GT motile cilia measuring 0.9μm shorter than Gas8WT motile cilia (Gas8WT 5.3μm and Gas8GT 4.4μm) (Fig 4I). Cilia orientation in Gas8GT tracheas is also more randomized than in Gas8WT controls (Fig 4J). These phenotypes observed in the motile cilia of Gas8GT mutant mice are similar to those observed in PCD patients and animal models.
The phenotypes in the Gas8GT mutants led us to evaluate whether mutations in GAS8 are associated with PCD in humans. We identified two independent missense variants, c.595G>A E199K and c.1172C>T A391V, in human patients through a previously published screen (Fig 5A) [32]. The E119K patient is of Latino decent and presented with heterotaxy. Unaffected parents of the patient are heterozygotes, and an unaffected female sibling is a homozygote. This allele appears at a frequency of 11% in Latino populations (87 homozygotes and 1279 heterozygotes in a total of 11564 alleles sequenced according to ExAC). The prevalence of this allele in the Latino community makes it unlikely to be associated with disease. The A391V patient met the diagnostic criteria for PCD. This allele is infrequent, occurring only 3 times heterozygously and 0 times homozygously in 84864 alleles sequenced according to ExAC. Both variants affect highly conserved regions across multiple species (Fig 5B). We utilized the PolyPhen-2 program to predict the pathogenicity of these alleles. The A391V allele had a PolyPhen-2 score of 0.762 suggesting that it is a potentially damaging mutation while the E199K allele had a score of only 0.082, suggesting that this is a benign mutation. Given the low allele frequency, PolyPhen-2 score, and the confirmation of the PCD diagnosis in the patient carrying the A391V variant, we chose to test potential pathogenicity of this allele in mice.
To further assess the potential pathogenicity of the human allele, we created a mouse harboring the A391V mutation via homology driven repair with CRISPR/Cas9 technology. Sequencing confirmed the presence of the c.1172 C>T mutation resulting in an A391V amino acid change (Fig 6A). We crossed Gas8AV mice onto the Gas8GT background to create compound heterozygous (Gas8GT/AV) mice. To determine the impact on motile cilia and test possible cause of the hydrocephalus, we took brains from 6 week old mice and analyzed cilia beat and the ability of motile cilia to move fluid. While there were no differences in beat frequency, bead flow analysis shows a modest decrease in the ability of Gas8GT/AV cilia to move fluid compared to Gas8GT/WT cilia (Fig 6B and 6C). Compound heterozygotes develop mild hydrocephalus at approximately 10 weeks of age (Fig 6D) but there were no evident laterality defects. While all the Gas8GT/AV mice analyzed (n = 6) display hydrocephalus at this age, the severity ranged from mild (arrowhead) to moderate (arrow). The phenotype in the Gas8GT/AV mice is not as severe as in the Gas8GT mice and hydrocephalus was not present in any (n = 2) of the Gas8GT/WT or (n = 2) of the Gas8WT/AV mice analyzed.
We chose to generate the A391V mouse model because of the PCD symptoms of the patient but given the lack of full PCD symptoms in the E199K patient, we decided to test first whether or not the E199K is pathogenic in Chlamydomonas before proceeding to a potential mammalian model. Alignment of GAS8 and the Chlamydomonas orthologue DRC4 revealed that E199 in GAS8 aligns with D198 in DRC4 (Fig 5B). To better understand the mechanisms underlying these defects, we generated strains expressing the Chlamydomonas equivalent (D198K) of the human E199K alleles in a null mutant background (pf2). Interestingly, transformation with DRC4-DK-GFP rescued the severe motility defects seen in the pf2 null mutant, but measurements of forward swimming velocities revealed a subtle defect in the swimming phenotype of the rescued strains (Fig 6E). Furthermore, the DRC4-DK-GFP protein is assembled at wild-type levels in the flagellar axonemes of Chlamydomonas, as assayed by western blot (S1 Fig). These observations show that the D198K DRC4 mutant protein is properly localized in the axoneme and may not correspond to a pathogenic allele.
Defects involving cilia motility cause severe phenotypes in humans including infertility, hydrocephalus, respiratory defects, and reversal of left-right asymmetry. Much of our understanding about cilia motility has come from studies in organisms such as Chlamydomonas. These studies and how defects in cilia motility cause disease are now being extended into mammalian systems. Recently GAS8 was implicated as a cause for Primary Ciliary Dyskinesia (PCD) as well as a positive effector of Smoothened transport into cilia during Hh pathway activation [12,13,19].
To further evaluate the connection between Gas8 and PCD in mammals, we generated a mouse with a β-geo cassette inserted in intron 7 of the Gas8 gene. Insertion of the β-geo genetrap cassette effectively eliminated the presence of wildtype transcript and protein in mutants as verified by RT-PCR and western blot analysis. Though the Gas8GT mutant allele is translated into a large fusion protein between the N-terminal portion of Gas8 and β-geo, it does not localize to motile cilia. Gas8GT mutant mice present with hydrocephalus starting at postnatal day 5 (P5) that becomes more pronounced as the mice mature and eventually leads to mortality between P14-P21. Development of hydrocephalus is associated with severe impairment of cilia motility on ependymal cells lining the ventricles of the brain.
Previous studies using image average procedures to analyze flagella ultrastructure in Chlamydomonas showed that strains with mutations in PF2/DRC4, the Gas8 homolog, were associated with the loss of the majority of the N-DRC complex along with a subset of the IDAs [6,16,17,33]. In contrast to the Chlamydomonas results, the N-DRC and IDA do not appear to be overtly affected in Gas8 mutant mice based on standard thin-section TEM analysis. However, loss of Gas8 does effect microtubule organization, as indicated by a higher percentage of cilia with disorganized microtubule doublets in Gas8GT mutant mice when compared to Gas8WT mice. Altered cilia microtubules were recently also recently observed in human Gas8 patients [12,13]. Together these data suggest that defects in the mammalian N-DRC may not always be detectable using traditional TEM averaging of cross-sections. The inability to observe ultrastructural defects in human PCD patients could be attributed to having only one N-DRC per 96nm repeat. Future studies using better imaging approaches such as cryo-electron tomography and image averaging of longitudinal sections to assess the human N-DRC will likely continue to reveal structural and functional differences similar to those described for the radial spokes by Lin, et al 2014.
Most Gas8GT mutant cilia failed to move, however those that were observed moving displayed a modest decrease in beat frequency. The most distinguishable phenotype observed in the cilia that moved was a very rigid and short wave pattern. This pattern has also been observed in other cilia motility mutants thought to affect the NDRC [10–12]. These changes in waveform and the lack of overall motility result in the defective fluid flow observed in these mice. Previous data show a complex relationship between planar cell polarity (PCP) and fluid flow in establishing motile cilia orientation [34,35]. Gas8GT cilia show a more random distribution of cilia orientation than their Gas8WT counterparts supporting the necessity of proper fluid flow in establishing cilia orientation.
Variants in Gas8 were recently identified in human PCD patients. These mutations resulted in a similar, albeit not significant, decrease in beat frequency along with an abnormally rigid ciliary waveform [12,13]. This motility phenotype is similar to our observation in the mutant mice. Here we identified an additional independent missense mutation, c.1172C>T A391V, in PCD patients as well as a variant c.595G>A E199K that appears to have minimal effect on cilia motility. The A391V mutations lies in close proximity to the other published mutants, C309*, A334*, and G357* suggesting that this area is critical for GAS8 function. Similarly, the genetrap cassette in the Gas8GT allele was inserted in close proximity (K337) to the A334* mutation. The E199K mutation also affects a highly conserved region within Gas8 that is proposed to be a Microtubule Association Domain (GMAD) [36].
To test pathogenicity of the A391V allele, we used CRISPR/CAS9 homology driven repair (HDR) to generate a mouse line mimicking the human mutation. Mice compound heterozygous for the Gas8GT and Gas8AV mutations develop age dependent, mild hydrocephalus, but did not present with situs defects (n = 6 Gas8GT/AV mice). The phenotype was associated with a reduced ability of ependymal cilia to move fluid. Interestingly, beat frequency was not significantly altered from that of controls, suggesting that the defect lies within a subtle waveform difference or in cilia orientation. These data suggest that the A391V allele is pathogenic though more in-depth analysis of ciliary defects will be necessary to determine the precise mechanism. Data from the Chlamydomonas rescue experiments suggest that the E199K allele may have very subtle effects on motility. The D198K rescued strain in Chlamydomonas showed a small but statistically significant reduction in forward swim velocity of approximately 10 percent. While statistically significant, additional work is needed to determine whether such small changes might impact ciliary motility and have pathogenic consequences in different tissues and different organisms. As this variant is commonly found in Latino populations, it seems more likely that this variant is a benign polymorphism.
Gas8 was previously implicated as a modulator of the Hh pathway. In vitro data indicated that the C-terminal region of Gas8 binds to Smoothened (Smo) and acts at the base of primary cilia as a regulator of Smo entry into the cilium following Hh pathway activation [19]. These data showed that in the absence of Gas8, Smo accumulation in the cilium is abrogated and that it cannot activate the Gli transcription factors and turn on downstream genes. Based on these in vitro findings, we expected to see Hh defects in our mutant mice. However, the Gas8 mutants survive to birth and have normal digit number and patterning as well as normal neural tube dorsal ventral patterning. Furthermore, there were no significant differences in Smo accumulation in cilia between Gas8WT and Gas8GT MEFs after SAG stimulation, suggesting that in this mutant model, Gas8 does not act as a regulator for Smo entry. The role that Gas8 plays at the base of primary cilia remains uncertain; however, we do not see any other pathologies that would suggest there is a defect in primary cilia such as cystic kidney disease.
The data presented here solidify GAS8 as a disease causing gene in humans and elucidate the mechanisms by which loss of Gas8 causes disease. We identified new independent, homozygous missense mutations and used model systems to test the pathogenicity of the alleles. Importantly, these results suggest the A391V allele is pathogenic while the E199K variant is not. Our results demonstrate the importance of testing the potential pathogenicity of human alleles in easily amenable model systems such as Chlamydomonas and further reveal the ease with which CRISPR/Cas9 has now made it possible to conduct similar tests in mouse models.
The Gas8 mutant mouse line was generated using embryonic stem cell line CH0760 (BayGenomics) in which a β-galactosidase neomycin resistance fusion cassette was inserted into intron 7 of Gas8. The insertion site was confirmed by genomic PCR and sequence analysis. PCR primers for genotyping were designed based on the insertion site and are as follows 5’-GGGACAAGCAGATTCTGGTC-3’, 5’-CAGGGTTACACACAGAGAAACC-3’, and 5’-CCGCAAACTCCTATTTCTG-3’. The Gas8GT embryonic stem cells were from the 129P2/OlaHsd genetic background and were injected into C57BL/6 blastocysts using standard procedures. Chimeras were bred with albino C57BL/6 females and germline transmission was confirmed by coat color and subsequent PCR genotyping.
All experimental procedures were approved by the Institutional Animal Care and Use Committee (IACUC) regulations at the University of Alabama at Birmingham under the animal protocol number (130208061).
RNA was isolated from Gas8WT, Gas8GT/WT, and Gas8GT mouse embryonic fibroblasts using Trizol reagent according to the manufacturer’s protocol (cat# 15596–026, Thermo-Fisher Scientific). cDNA was generated using Verso cDNA kit (cat# AB-1453/B, Thermo-Fisher Scientific). 5’ Gas8 RT-PCR was performed using the following primers spanning exons 3 and 4: 5’-GAATCGAAGAATACCACCATC-3’ and 5’-CTGAGAAGATGGCTATGTAG-3’. 3’ Gas8 RT-PCR was performed with primers spanning exons 9 and 10: 5’-CTGGACCCCACAGCATTAAC-3’ and 5’-CTTGATGGTGGTATTCTTCG-3’. Actin control primers: 5’-ATGGGTCAGAAGGACTCCTA-3’ and 5’-GGTGTAAAACGCAGCTCA-3’ were used in all samples.
Animals were anesthetized by a 0.1 ml per 10 g of body weight intraperitoneal injection of 2.5% tribromoethanol (cat# T48402, Sigma–Aldrich), killed by cardiac puncture, and perfused with PBS followed by 4% paraformaldehyde (cat# 19943, Thermo-Fisher Scientific). The brains were further fixed in 4% paraformaldehyde 1h at room temperature followed by successive dehydration through 1 hour alcohol incubations at 30% and 50% and placed finally in 70% overnight. Tissues were further dehydrated through 1 hour alcohol incubations at 80%, 95%, and finally 100%. Tissues were placed in xylenes for 1 hour and then placed in a 50/50 xylenes/paraffin mix for 1 hour at 60°C under vacuum followed by a final paraffin penetration in paraffin at 60°C under vacuum for 1 hour and then paraffin embedded. The brains were sectioned at 10μm and stained with Cresyl Violet stain as previously described [37].
Fresh tracheas were extracted from p21 Gas8WT, Gas8GT/WT, and Gas8GT mice. Samples were submerged in ice cold RIPA (10mM Tris pH7.5, 150mM NaCl, 1%NP-40, 1% sodium deoxycholate, 0.1% SDS) mixed with one cOmplete Protease Inhibitor tablet (cat# 11 836 170 001, Roche Diagnostics) per 10mL at 300uL per 5mg of tissue. Tissues were sonicated 3x for 10 seconds each. After sonication, tissues were placed on a rotary agitator for 2 hours at 4°C and then spun for 20 minutes at 12,000rpm at 4°C. Supernatant was removed and protein levels were assayed using a BioRad DC protein assay kit (cat# 5000111, Bio-Rad). Approximately 20μg per sample was used for SDS-PAGE with a 12% Tris-Glycine gel (cat# 00252562, Thermo—Fisher Scientific). Proteins were transferred overnight to nitrocellulose. The membrane was blocked for 45 minutes in 5% milk in PBS and incubated with primary antibody in 5% milk in PBS with 0.02% Tween-20 overnight at 4°C. Primary rabbit anti-Gas8/DRC4 antibody was used at 1:20000 [6]. Primary mouse anti-GAPDH was used as a loading control at 1:1000 (cat# ab8245, Abcam, Cambridge UK). Blots were washed 5x for 5 minutes each in 0.02% PBS-Tween-20. Secondary antibody in 5% milk in 0.02% PBS-Tween-20 was added and the blot was incubated for 1 hour at room temperature with the following secondary antibodies: anti-mouse IRDye 800CW (cat# 827–08364, LI-COR, Lincoln NE USA) and anti-rabbit IRDye 680RD (cat# 926–68071, LI-COR). Blots were washed 5x for 5 min each in 0.02% PBS-Tween-20 and then dried. Images were taken on a LI-COR Odyssey CLx imaging system (LI-COR).
Mouse embryonic fibroblasts (MEFs) were generated from E14.5 embryos and cultured in DMEM growth medium with High Glucose, 0.05mg/ml Penicillin/Streptomycin, 2mM L-Glutamine, 0.2mM β-mercaptoethanol, and 20% Fetal Bovine Serum (FBS). Prior to immunolabeling, MEFs were cultured in reduced serum medium containing 0.5% FBS for 48 hours to induce primary cilia formation as previously described [38].
Cells were fixed in 4% paraformaldehyde and permeabilized with 0.1% Triton X-100 in PBS with 2% donkey serum, 0.02% sodium azide and 10 mg/ml bovine serum albumin (BSA). Cells were labeled with anti-acetylated α-tubulin, 1:1000 (cat# T-6793, Sigma-Aldrich), anti-SmoN, 1:1000 (gift from Dr. Matthew Scott, Stanford University). Sections from E10.5 neural tubes were immunolabeled with the following antibodies from Developmental Studies Hybridoma Bank (University of Iowa, Iowa City, IA): anti-ShhN 1:1000 (5E1), anti-FoxA2 1:1000 (74.5a5), anti-Mnr2 1:1000 (81.5C10), anti-Pax7 1:1000 (Pax7), and anti-Msx1+2 1:1000 (4G1) as previously described [38]. Trachea sections were labeled with anti-Gas8/DRC4, 1:2000 [6]. All incubations and washes were carried out in PBS with 2% normal donkey serum, 0.02% sodium azide and 1% BSA. Primary antibody incubations were performed for 16–24 hours at 4°C and secondary antibody incubations were performed for 1 hour at room temperature. Secondary antibodies all from Thermo-Fisher Scientific include the following: Alexa Fluor-594 donkey anti-mouse (cat# A21203), Alexa Fluor-488 donkey anti-mouse (cat# A21202), Alexa Fluor-594 donkey anti-rabbit (cat# A21207), and Alexa Fluor-488 donkey anti-rabbit (cat# A21206). Nuclei were visualized by Hoechst nuclear stain. Coverslips were mounted using Immu-Mount (cat# 9990402, Thermo-Fisher Scientific). Fluorescence imaging was performed using a Nikon TE-2000U inverted microscope (Melville, KY) outfitted with a PerkinElmer UltraVIEW ERS 6FE-US spinning disk laser apparatus (Shelton, CT) and a Hamamatsu C9100.
DIC images of p14 trachea prepared for IF were used for length analysis. Images were captured with a 40x objective (Plan-Fluor, 1.3NA). Length was measured manually by drawing a line from the tip of the cilium to the base using Volocity v6.3.
Gas8WT and Gas8GT MEFs were grown to confluency on 0.17mm coverslips and serum starved for 48 hours to induce ciliation. Cells were treated with 150nM Smoothened agonist (SAG) (cat#566660, CALBIOCHEM) for 2 hours in low serum media to induce Hedgehog pathway activation and Smoothened translocation. Cells were fixed and stained as described in the immunofluorescence section and imaged by spinning disk confocal. Amount of Smoothened per cilia volume was measured using Volocity v6.3 software.
Postnatal day 14 (P14 mice were anesthetized and perfused with PBS followed by a perfusion of 2% glutaraldehyde in 0.1M cacodylate buffer pH 7.4. Tracheas were extracted and fixed overnight at 4°C in 2% glutaraldehyde in 0.1M cacodylate buffer pH 7.4. Samples were then washed thoroughly four times for 15 minutes each in 0.1M cacodylate Buffer pH 7.4. A post fix in 1% OsO4 in 0.1M cacodylate buffer pH 7.4 was performed. Samples were washed two times for 10 minutes each in 0.1 M cacodylate pH 7.0. Samples were then prepped in 1% tannic acid in 0.1M cacodylate Buffer pH 7.0; 30 minutes followed by 1% NaSO4 in 0.1M cacodylate Buffer pH 7.0; 5 minutes. Dehydrate the samples in 50%, 75%, and 95% at 4°C for 20 minutes each and finally 100% EtOH for 20 minutes; warm to RT°. Dehydrate samples totally with four washes of 100% EtOH 15 minutes each. Infiltrate the sample with Propylene Oxide for 30 minutes. Mix the EMbed 812 according to instructions from EMS and Infiltrate with 25% Embed in propylene oxide for 30 minutes, 50% for 40 minutes, 75% overnight, 100% for four hours, 100% for 1 hour and harden at 60°C. Samples were sliced at 90nm and imaged on a Phillips CM110 Electron Microscope.
TEM averaging of doublets was performed by isolating individual doublets from cilia and importing the doublets into Photoshop CS5. Individual doublets were aligned to a single template doublet and then averaged and flattened.
TEMs were used to determine cilia orientation. Cilia orientation was determined by measuring the angle of central pairs by drawing a line across the central doublets and measuring the angle relative to the image. Each angle was normalized to the average (or most common angle) after setting the average angle to 0°. The frequency of angles in each image was measured and plotted.
Brains of experimental mice were extracted, sliced in half to expose the ependymal of the lateral ventricles and placed in pre-warmed, pre-oxygenated artificial cerebrospinal fluid (125mM NaCl, 2.5mM KCl, 1.25mM NaH2PO4, 2mM CaCl2, 1mM MgCl2, 25mM NaHCO3, 25mM Glucose, pH 7.35). Brains were placed on a Zeiss Axioskop microscope and imaged with a 5x objective (Plan-Neofluor, 0.15NA) and a 10x objective (Fluor, 0.5NA) using a Photometrics CoolSnap HQ CCD camera at 30fps. Red fluorescent latex beads (cat# L3530-1mL, Sigma-Aldrich) were diluted 1:100 from stock and 10μL of diluted beads was added to the ventricles. Bead tracking analysis was performed using the MTrack2 plugin in FIJI.
Mice trachea were dissected out into fresh PBS and cut lengthwise into strips. Trachea were kept in warm media (DMEM F/12, 20% FBS, and Pen-Strep and allowed to adapt for 20 minutes in an environmental chamber (37°C, 45% relative humidity, and 5% CO2) before imaging with Differential Interference Contrast (DIC). All high speed video was captured at 240fps using a modified Casio Exilim EX-ZR100 attached to a Nikon TE-200 using a 60x water objective (Plan-Apo WI NA = 1.2). Videos saved as quick time files were then extracted into individual frames using VirtualDub 1.10.4 software and all analysis was performed in ImageJ.
Kymographs were created using Metamorph v6.1. A line was drawn from the tip of the cilium to the base and kymographs were made from the results.
To make the D198K mutation in Chlamydomonas, the DRC4-GFP plasmid [6] was used as template for PCR with the primers 5’-CAGTGCTGTGAGCCTGACG and 5’-AAACCAAAGCACCTTGAGCG to generate a 1483bp product that contains the restriction sites BclI and ClaI flanking the desired mutation site. The PCR product was cloned into pGEM-T-Easy (cat# A1360, Promega Corp) to generate the plasmid pf2-Y1-A. This plasmid was further digested with KpnI and SpeI to removed repetitive DNA and subcloned into pBlueScript to generate the plasmid pf2-Y1-B. The D198K mutation was introduced into pf2-Y1-B using the primers 5’-GAAGATGCTGCGAGACaAaATGGAGCTGCGGAGAAAG-3’ and 5’-CTTCTACGACGCTCTGtTtTACCTCGACGCCTCTTTC-3’ and the QuickChange II kit (Agilent Technologies) to generate the plasmid pf2-Y1-C. After sequence verification by Genewiz, the pf2-Y1-C plasmid was digested with KpnI and SpeI and subcloned back into pf2-Y1-A to generate the plasmid pf2-Y1-D. The pf2-Y1-D plasmid was digested with BclI and ClaI to release the 1483 bp fragment now carrying the D198K mutation. This fragment was subcloned back into the original DRC4-GFP plasmid by Genewiz. The completed plasmid, DRC4-D198K-GFP, was linearized with EcoRI for transformation into the pf2-4 strain [6]. Transformants were screened as described above. RT-PCR confirmed that the D198K mutation was expressed in the rescued strains without any other sequence modification. Forward swimming velocity was recorded and measured as previously described [6]. For transformations with the control DRC4-GFP plasmid, rescued colonies were recovered at a frequency of 5–15%
CRISPR/sgRNA target sequences were queried using the MIT CRISPR server. Three sites most proximal to the desired SNP change were selected to test nuclease efficiency. CRISPR1: 5’- CTTCTCCACAGCAGCGTTCA GGG-3’ (reverse strand), CRISPR2: 5’-GGTGCTGGCCGCCTCCAACC TGG-3’ (forward strand), CRISPR3: 5’-GACACAAGCGTTAATGCTGTGGG-3’ (reverse strand). Pronuclear injections were performed with Cas9 mRNA (100 ng/ul), CRISPR3/sgRNA (50 ng/ul) and ssODN (200 ng/ul). Efficiency of nuclease activity was assessed using a blastocyst assay. In brief, injected zygotes were cultured to the blastocyst stage and lysed to obtain genomic DNA. Genomic DNA was used in PCR and the amplicons (215 bp) were resolved by heteroduplex mobility assay (HMA). CRISPR3 was found to be most efficient and was used to generate the SNP edited mouse (C57Bl/6 background). Injected zygotes were cultured to 2-cell stage in KSOM mixed with the NHEJ inhibitor SCR7 at a final concentration of 10 mM. The 2-cell stage embryos were transferred to psuedopregnant recipient female mice, which gave birth to 13 pups. The SNP was introduced with the help of a 154 nt single stranded oligo DNA (ssODN) HDR template. Since the PAM sequence (CCC>Pro) could not be modified without changing the amino acid, multiple silent changes were made in the protospacer (sgRNA binding) sequence (indicated by small letters in the sequence below). These changes were made to eliminate the chances of the sgRNA binding to the repaired allele. The SNP change introduced a restriction enzyme recognition enzyme site (BsmBI/Esp3I) and the silent changes introduced two new restriction enzyme recognition sites (BtgI and HaeII). HaeII sites were used to distinguish the wildtype and the modified alleles. Specific primers were also designed that can preferentially amplify the modified allele. HDR template (ssODN) 5’-GGCCCTGAACGCTGCTGTGGAGAAGAGAGAGGTTCAGTTCAATGAGGTGCTGGCCGTCTCCAACCTGGACCCCACgGCgcTgACGtTgGTGTCCCGCAAACTTGAGGTAGGTGCCCTCCTGTCCTGTGCTGTGGTACGCCTTCTTGGGTGGCAC-3’. After the initial characterization of the F0 litter by PCR, the 215 bp amplicons were cloned, and selected individual clones were subjected to Sanger sequencing. Sequence analysis of the 13 pups revealed that 2 had complete knock-in of the edited/repair sequence, 1 pup had incorporated the silent changes but did not have the desired SNP change, and 1 pup had indels. F0 animals were bred with wildtype C57Bl/6 mice to test germline transmission of the desired alleles. All alleles were successfully transmitted through the germline, and the positive F1 animals were used to create homozygous and compound heterozygous F2 animals.
Cilia length analysis, bead flow tracking, cilia orientation, and cilia beat frequency were tested with Student’s t-test and graphed in Microsoft Excel. Smoothened trafficking assay and Chlamydomonas swim speeds were tested by ANOVA followed by Student’s t-test with a Bonferroni correction and graphed in Microsoft Excel. All error is represented in Standard Error of Means (SEM).
|
10.1371/journal.ppat.1007674 | Biphasic and cardiomyocyte-specific IFIT activity protects cardiomyocytes from enteroviral infection | Viral myocarditis is a serious disease, commonly caused by type B coxsackieviruses (CVB). Here we show that innate immune protection against CVB3 myocarditis requires the IFIT (IFN-induced with tetratricopeptide) locus, which acts in a biphasic manner. Using IFIT locus knockout (IFITKO) cardiomyocytes we show that, in the absence of the IFIT locus, viral replication is dramatically increased, indicating that constitutive IFIT expression suppresses CVB replication in this cell type. IFNβ pre-treatment strongly suppresses CVB3 replication in wild type (wt) cardiomyocytes, but not in IFITKO cardiomyocytes, indicating that other interferon-stimulated genes (ISGs) cannot compensate for the loss of IFITs in this cell type. Thus, in isolated wt cardiomyocytes, the anti-CVB3 activity of IFITs is biphasic, being required for protection both before and after T1IFN signaling. These in vitro findings are replicated in vivo. Using novel IFITKO mice we demonstrate accelerated CVB3 replication in pancreas, liver and heart in the hours following infection. This early increase in virus load in IFITKO animals accelerates the induction of other ISGs in several tissues, enhancing virus clearance from some tissues, indicating that–in contrast to cardiomyocytes–other ISGs can offset the loss of IFITs from those cell types. In contrast, CVB3 persists in IFITKO hearts, and myocarditis occurs. Thus, cardiomyocytes have a specific, biphasic, and near-absolute requirement for IFITs to control CVB infection.
| Viruses can infect the heart, causing inflammation–termed myocarditis–which is a serious, and sometimes fatal, disease. One way to combat the infection is by stimulating our immune system, encouraging it to fight the virus. However, the treatment that is currently used “revs up” many different parts of our immune system, including some that play little or no role in clearing the virus, and this wide-ranging activation increases the risk of potentially-harmful side effects. We want to identify the parts of the immune system that fight virus infections of the heart, so that we can improve the treatment of viral myocarditis by selectively stimulating only those immune responses, thereby retaining the benefit of treatment (i.e., clearing the virus) while reducing its cost (i.e. lowering the risk of harmful side effects). In this paper, we demonstrate that a family of proteins called IFITs play a role in protecting many tissues against these infections, but are particularly important in heart muscle cells, in which they are indispensable. Thus, IFITs represent a possible target for the treatment of viral myocarditis.
| Myocarditis, which can cause serious, and sometimes fatal, complications including heart failure, cardiac arrest, and dilated cardiomyopathy, is commonly caused by infection and, most frequently, by viruses including coxsackievirus B3 (CVB3) [1, 2]. This enterovirus infects mice and humans, replicates to high titers, and causes acute viral myocarditis through two major pathological mechanisms; virus-mediated direct lysis of the infected cells and immune-mediated tissue damage (immunopathology). Limiting virus infection by activating the immune system through type I interferon therapy (T1IFN) has shown promise [3, 4], but comes with an increased risk of immunopathology, because T1IFNs have strong and pleiotropic biological effects. Therefore, it is important to better understand how T1IFNs exert their anti-enteroviral effects, with the aim of retaining their biological benefits while reducing concomitant immunopathology.
T1IFNs (mainly, ~12 subtypes of IFNα and the sole IFNβ in human and mouse) are important innate immune mediators against virus infection. T1IFN production is initiated in a virus-infected cell by the tripping of series of innate immune sensors; the resulting downstream signaling upregulates the transcription of genes encoding T1IFNs and pro-inflammatory cytokines, which are secreted from the infected cell [5]. After secretion, all of the T1IFN proteins signal through a common heterodimeric receptor, the T1IFN receptor (T1IFNR), expressed by the great majority of somatic cells. T1IFN binding to this receptor activates the JAK-STAT pathways, leading to the induction of a large number of interferon-stimulated genes (ISGs), which then exert various effects including innate immune antiviral action and modulation of cytokine production. Mice lacking this receptor rapidly succumb to CVB3 infection [6, 7], as do IFNβ knockout (KO) mice [8], demonstrating the essential role played by T1IFNs in protecting against this virus. We recently generated inducible conditional knockout mice (CMMCM T1IFNRf/f mice) in which the administration of tamoxifen efficiently deleted T1IFNR expression specifically in cardiomyocytes and, using these mice, we revealed the importance of local T1IFN signaling into cardiomyocytes during CVB3 infection. Without such signaling, at ~2–3 days post-infection (p.i.) we observed increased cardiac titers; myocarditis was accelerated, and virus clearance was delayed [7]. These data raised several questions: during CVB3 infection, which ISGs are induced in cardiomyocytes in response to CVB3 infection, which of these ISGs are needed to suppress virus replication, and which ISGs regulate the rapid influx of inflammatory cells into the heart?
We show here that, following CVB3 infection, IFIT (IFN-induced with tetratricopeptide) family genes are highly induced in cardiomyocytes in vivo. The IFITs are a large family comprising six murine (Ifit1, Ifit2, Ifit3, Ifit1b, Ifit1c and Ifit3b) and five human (IFIT1, IFIT2, IFIT3, IFIT5 and IFIT1B) members [9]. These genes exert antiviral responses against various different viral species by binding to both host and viral molecules [9–11], but the role of IFIT locus genes in enterovirus infection and the consequent pathogenesis have not been previously investigated. In this study, we use mice lacking the entire IFIT locus (IFITKO mice), several primary cell types from these mice, and cardiomyocytes modified by CRISPR/Cas9-mediated gene editing. Both of these approaches—in vivo and in vitro–indicate that the IFIT locus acts biphasically, and in a cardiomyocyte-specific manner. During the first phase, constitutive IFIT expression is required for suppressing early CVB3 replication in several tissues and cell types. In the second phase, which follows T1IFN signaling, the upregulation of IFITs is vital for CVB3 clearance from cardiomyocytes, and for the prevention of myocarditis. We conclude that the second phase of IFIT activity is cardiomyocyte-specific because the T1IFN-driven induction of IFITs is expendable in other cell types. In these cells–unlike in cardiomyocytes–other ISGs can provide compensatory anti-CVB3 activity, offsetting the absence of IFITs.
To identify which ISGs are expressed in normal cardiomyocytes during CVB3 infection in vivo, we exploited CMMCM T1IFNRf/f mice [7]. Two weeks after receiving a Tamoxifen injection, Cre- or CMMCM T1IFNRf/f mice were challenged with 500 PFU CVB3 intraperitoneally (i.p.). Two days later, a time point when the cardiac virus titers are still comparable in both groups [7], the animals (and uninfected controls) were sacrificed, and ISG expression in the hearts of the four different data groups (Fig 1A) was determined by PCR array. CVB3 infection induced multiple ISGs in the genetically-intact heart (Fig 1B, left), but this was largely abolished by T1IFNR deficiency in cardiomyocytes (Fig 1B, right), indicating that ISG upregulation in the heart is limited mainly to those cells. By comparing the PCR signals of Cre- and CMMCM hearts, we estimated the extent to which various ISGs were upregulated in cardiomyocytes (Fig 1C). IFNβ mRNA was ~20-fold more abundant in the hearts of Cre- mice, demonstrating that (i) cardiomyocytes are the major source of IFNβ during CVB3 infection and (ii) CVB3 infection triggers abundant IFNβ production by cardiomyocytes only if these cells can receive (or have already received) T1IFN signals. This is consistent with a previous report in which, using HL-1 cells (a murine cardiomyocyte cell line [12]), the authors showed that CVB3 infection did not directly trigger IFNβ production [13]. We have independently confirmed this finding (see S1 Fig). In addition to ISGs that been described previously (Ifnb1, Socs1, Isg15 and Il6) [8, 14–16], we found that several of the IFIT family genes were up-regulated in cardiomyocytes at 2 days p.i. (small arrows, Fig 1C; this PCR array assayed only Ifit1, Ifit2 and Ifit3). As described above, a broad spectrum of cellular functions of individual IFIT family genes, including antiviral effects, has been reported previously [9–11], but little is known about the collective importance of the IFIT locus, and less still about its role during enterovirus infection in vivo. Therefore, the remaining experiments reported herein focused on the role of this gene family in responding to CVB3 infection. To confirm the PCR array results in vitro, and to extend them to IFIT family members not covered in the PCR array, we isolated primary cardiomyocytes from C57BL/6 (B6) mice. CVB3 could efficiently infect, and replicate in, this cell type (Fig 1D). Real-time PCR analysis of CVB3-infected cardiomyocytes showed that, in the absence of exogenous T1IFNs, there is a substantial (> 30 hour) delay in IFIT expression, but by 72 hours p.i. there is robust induction of Ifit1, Ifit2 and Ifit3, and also of one previously-uncharacterized family member, Ifit3b. However, there was little, if any, induction of Ifit1b and Ifit1c at 72 hours p.i. (Fig 1E). Thus, our in vivo and in vitro data indicate that, by 48–72 after CVB3 infection, 3–4 members of IFIT family genes (Ifit1, Ifit2, Ifit3 and Ifit3b) are highly induced in cardiomyocytes.
Next, we infected B6 mice with CVB3 (104 pfu, i.p.). The animals (and uninfected controls) were sacrificed at 2 days after CVB3 infection, and IFIT family gene expression in 11 different tissues was analyzed by real-time RT-PCR (Fig 1F). Similar to cardiomyocytes, at this time point after CVB3 infection, Ifit1, Ifit2, Ifit3 and Ifit3b were induced in most tissues. Up-regulation of the other two family members is more tissue-restricted: Ifit1b is induced almost exclusively in kidney, liver and pancreas, and Ifit1c in liver and pancreas. In uninfected B6 mice, expression of most of the IFIT family mRNAs was low but detectable in most tissues (S2A Fig). We also determined the constitutive expression of the mRNAs in three primary cell types isolated from B6 mice (cardiomyocytes, peritoneal macrophages and cardiac fibroblasts), and found that all were expressed in uninfected cells (S2B Fig). The expression pattern of IFIT mRNAs differed between the two cardiac-derived cell types, reflecting others’ findings that basal levels of ISG expression–including IFIT1 –can be detected in both cell types, and that the expression pattern is cell-type-specific [17]. The relative levels of IFIT1 expression differed between the cited study and our own data; we speculate that this might result from unidentified differences in culture conditions. In contrast to the marked induction of most IFIT mRNAs at 48–72 hours after CVB3 infection in vivo and in vitro (Fig 1B, 1C, 1E and 1F), barely any increase was observed in mouse hearts at 24 hours p.i. (S2C Fig), supporting our observation that there was little increase at 30 hours p.i. (Fig 1E). Taken together, these data indicate that CVB3 infection of cardiomyocytes does not, in itself, drive the rapid and abundant expression of ISGs (including IFITs and T1IFNs); rather, the induction of ISGs depends on the cardiomyocytes having been primed by T1IFN signaling. Basal levels of IFIT2 and IFIT3 proteins were detectable in the liver and heart of naïve B6 mice, and both were markedly up-regulated following the in vivo administration of recombinant IFNβ (S2D Fig, left panels). Low constitutive expression of IFIT2 and IFIT3 proteins also was detectable in primary cardiomyocytes, and was up-regulated after 24 hours of IFNβ treatment (S2D Fig, right panels). Thus, we conclude that: (i) IFITs are constitutively expressed in many tissues / cell types; (ii) the constitutive expression pattern of the various IFIT genes can vary among cell types; (iii) CVB3 infection causes a robust increase in expression of most of the IFIT family genes, but (iv) in cardiomyocytes, this takes at least 30 hours to occur, suggesting that the increase depends upon these cells having received signals by systemic T1IFNs which, perhaps originate from other cell types, e.g. dendritic cells in vivo.
To study the antiviral effect of T1IFNs on CVB3 infection in vitro, we first employed HL-1 cells, which we modified using the CRISPR/Cas9 system [18]. To validate the approach, we began by transfecting HL-1 cells with a vector encoding a single guide RNA targeted to exon 2 of the Ifnar1 gene, which encodes one of the heterodimeric T1IFN receptor proteins (Fig 2A). HL-1 cells did not expand after selection / single cell dilution, preventing us from developing HL-1 clonal lines. Therefore, we relied on bulk edited and selected HL-1 cells, in which T1IFNR protein expression was dramatically decreased (Fig 2B). Effective functional depletion was demonstrated by treating these cells, or their wt counterparts, with IFNβ; Ifit1, Ifit2 and Ifit3 mRNA induction was ablated in Ifnar1-edited HL-1 cells (Fig 2C). Western blotting showed that IFIT proteins were constitutively expressed at similar levels in wt and Ifnar1-edited HL-1 cells, but that IFNβ-driven induction of IFIT family gene products was ablated in the latter (Fig 2D). Finally, exogenous IFNβ pre-treatment reduced the production of infectious CVB3 by ~2,300-fold in WT HL-1 cells but had no suppressive effect in Ifnar1-edited HL-1 cells (Fig 2E). A further conclusion can be drawn from the data in this panel. There is no statistically-significant difference in virus titers between the non-treated WT and Ifnar-1-edited HL-1 cells, indicating that, during the time of infection, the cells did not produce sufficient IFNβ to confer any antiviral effect. This provides additional evidence that CVB3 infection of resting cardiomyocytes does not directly trigger abundant IFNβ production; if it did so, then one would have predicted that this endogenously-synthesized IFNβ would have suppressed viral replication in the WT cells to a level below that observed in Ifnar-1-edited cells, which are unable to respond to the cytokine. These findings confirm the in vitro data in S1 Fig, as well as our in vivo observations made using the CMMCM T1IFNRf/f mice; (i) T1IFN signaling into cardiomyocytes markedly inhibits CVB3 infection [7] and (ii) unless they receive T1IFN signals, cardiomyocytes do not produce abundant T1IFNs following infection by CVB3 (Fig 1). When combined with our demonstration that T1IFNs drive the strong up-regulation of IFIT expression in cardiomyocytes, these data raise the question: do IFIT genes participate in the observed T1IFN-mediated protection of cardiomyocytes against CVB3?
To address this question, we again applied the CRISPR/Cas9-mediated bulk gene-editing approach to HL-1 cells, this time deleting the entire IFIT locus. HL-1 cells were transfected with two CRISPR/Cas9 expression vectors expressing the indicated sgRNAs (Fig 3A) and, after drug selection, we confirmed the deletion of entire IFIT locus on genomic DNA of IFIT locus-edited HL-1 cells by PCR analysis (Fig 3B); the combination of primer P1 + reverse primer did not generate a detectable PCR product on WT DNA, because the two primers are separated by ~100 kbp, but a ~500 bp amplicon was present when DNA from the sgIFITs edited cells was used, indicating that these cells contained IFIT locus KO DNA. However, the bulk-edited and selected population was not 100% pure, because PCR using the reverse primer together with the P2 primer (which is absent from the KO DNA) produced an amplicon. To determine the impact of IFIT locus editing on IFIT protein content, IFIT2 and IFIT3 protein levels in HL-1 cells were determined by western blot without or with prior stimulation with recombinant IFNβ (100 U/ml). As shown in Fig 3C, robust induction of IFIT2 and IFIT3 was observed in WT cells, and this was markedly reduced in the sgIFIT population, although some protein was detected, consistent with the conclusion that the population contains some cells with intact IFIT genes. We then infected both populations of HL-1 cells with CVB3 (multiplicity of infection (MOI) of 1) in the presence or absence of IFNβ pre-treatment (100 U/ml). 72 hours p.i., virus titers in the supernatant of these cells were determined by plaque assays (Fig 3D). When cells were not pre-treated with IFNβ, infectious virus yield in the supernatant of IFIT locus-edited HL-1 cells was higher than in WT HL-1 cells at all time points, and became ~70-fold higher at 72 hours p.i. (p<0.0001) (Fig 3D, left panel). As shown in Fig 3D, right panel, IFNβ pre-treatment of WT HL-1 cells very effectively suppressed the infectious virus yield throughout the 72 hour time course (black symbols), whereas the virus still replicated to high titer in IFNβ-treated IFIT locus-edited HL-1 cells, in which the virus titer was significantly higher as early as 6 hours p.i., and ultimately became ~40,000-fold higher than in IFNβ pre-treated WT HL-1 cells (p = 0.0008). This viral titer data was reflected in analyses of virus RNA (Fig 3E). In the absence of IFNβ treatment, we found ~30-fold higher quantities of virus RNA in IFIT locus-edited HL-1 cells compared to WT cells (p<0.0001), and IFNβ pre-treatment reduced virus RNA in both cell populations, but much more so in the WT (P < 0.01). Parallel findings came from analyses of virus protein accumulation (Fig 3F). In summary, these data indicate that control of CVB3 infection by IFITs in HL-1 cardiomyocytes is biphasic: in the first phase, the constitutive (T1IFN-independent) expression of IFITs helps to constrain viral replication and in the second phase, T1IFN-driven up-regulation of IFIT expression in cardiomyocytes maintains this protection. The minimal antiviral effect of IFNβ on IFIT locus-edited cells (Fig 3D & 3E) indicates that, in cardiomyocytes, other ISGs are unable to confer substantial protection against CVB3.
To corroborate our HL-1 cell findings, we isolated primary cardiomyocytes from B6 mice and from IFIT locus-deleted (IFITKO) mice, which were generated in the laboratory of one of the authors (GCS) and will be described in detail in a separate paper. These IFITKO mice lack all of the IFIT family genes (Ifit1, Ifit2, Ifit3, Ifit1b, Ifit1c and Ifit3b), and the successful deletion is demonstrated by the fact that cardiomyocytes isolated from these IFITKO mice did not express any of the IFIT family genes, even after recombinant IFNβ treatment (S3 Fig). Primary cardiomyocytes from B6 and IFITKO mice were infected with CVB3, then recombinant IFNβ was (or was not) added to the culture media. 72 hours p.i., virus titers in the supernatant of these cells were determined (Fig 3G). These data confirmed the biphasic activities of IFITs in cardiomyocytes. CVB3 titer in the supernatant of primary cardiomyocytes (not treated with IFNβ) was significantly higher in IFITKO than B6 mice, and IFNβ treatment suppressed infection in WT cells, but not in IFITKO cells. Thus, in primary cardiomyocytes as in HL-1 cells, (i) IFIT locus genes appear to constitutively suppress CVB3 replication, (ii) IFN-driven IFIT up-regulation is required to maintain and extend this protection, and (iii) other ISGs, induced by exogenous IFNβ, appear unable to exert independent antiviral effects. We executed similar experiments with peritoneal macrophages, and with cardiac fibroblasts, from B6 and IFITKO mice (Fig 3H & 3I respectively). In both cell types, in the absence of exogenous IFNβ, virus yield was significantly increased in the IFITKO cells, similar to what was observed for cardiomyocytes. However, in contrast to cardiomyocytes, IFNβ treatment strongly inhibited infectious virus production by IFITKO macrophages and cardiac fibroblasts, although a stronger effect was observed for the former cell type. These data indicate that: (i) constitutive IFIT expression plays a role in suppressing CVB3 replication in cardiomyocytes, cardiac fibroblasts and peritoneal macrophages, and (ii) the second, T1IFN-induced, phase of IFIT activity shows cell specificity; if IFITs are absent, other ISGs can confer antiviral protection in macrophages, but not in cardiomyocytes, with cardiac fibroblasts showing an intermediate phenotype. Sherry and colleagues have previously proposed that cardiomyocytes, being non-replenishable cells, may have developed near-unique mechanisms to cope with viral challenges [19], and our observations support and extend this suggestion. Studies are ongoing to identify the precise mechanism(s) by which the IFIT locus mediates its anti-CVB3 activity.
Next, we investigated the in vivo consequence of the loss of the IFIT locus during CVB3 infection. B6 and IFITKO mice were challenged with 104 pfu of CVB3, i.p., and their body weights were monitored daily (Fig 4A). An early, and statistically-significant, loss of body weight was observed in IFITKO mice. Nevertheless, IFITKO mice survived significantly longer than B6 mice (Fig 4B). The CVB3-infected animals were sacrificed at 12 days p.i., and the two strains displayed dramatic macroscopic differences in the small intestine, pancreas, and liver (Fig 4C). The small intestine of CVB3-infected IFITKO mice was swollen and fulfilled with gas, and their pancreata were smaller than those of their B6 counterparts. In contrast, the livers of the IFITKO mice looked grossly normal, while those of B6 mice were pale, and confocal analyses revealed numerous apoptotic hepatocytes (S4 Fig), possibly contributing to the higher mortality in the WT animals (Fig 4B). This is consistent with the recent demonstration, by others, that hepatic disease may contribute significantly to the mortality associated with CVB3 infection [20]; we speculate that, in WT mice, the antiviral activity of IFITs in the liver may contribute to hepatic immunopathology that is detrimental to survival. No obvious macroscopic differences were observed in the hearts, but histological analyses were revealing (see below).
To analyze the in vivo impact of loss of the IFIT locus on CVB3 replication, we measured the virus titer in the pancreas, liver and heart of both groups at different time points over the course of CVB3 infection. At day 1 p.i., there were higher amounts of CVB3 in IFITKO mouse pancreas, liver and heart (Fig 4D to 4F), indicating that–as we showed for several IFITKO cell types in vitro (Fig 3G–3I)–IFIT locus genes are required for restricting very early CVB3 infection in vivo. However, at later time points (days 7, & 11–12 p.i.) CVB3 appeared to be more rapidly cleared from the pancreata and livers of IFITKO mice; at day 7 p.i., titers were ~300-fold lower in the IFITKO pancreata, and ~2 million-fold lower in the IFITKO livers (Fig 4D & 4E). We also found accelerated virus clearance in the feces of IFITKO mice (Fig 4G). These data indicate that, while the IFIT locus is required for immediate virus control in several tissues in vivo, its absence eventually leads to much more effective viral clearance from most of those tissues. The heart appears to be an exception: at day 7 p.i., CVB3 titers were high in the hearts of both mouse strains the small difference observed at this time point (~2-fold higher in the B6 mice) was far less than the differences observed in pancreata and livers and, by day 11–12 p.i., titers in IFITKO pancreata and liver were far below those in WT tissues (Fig 4D & 4E), while the titers in IFITKO hearts were ~20-fold higher than those in B6 mice (Fig 4F). This conclusion was supported by studies in which mice were challenged with a lower dose of CVB3 (103 pfu); IFITKO mice showed 100-fold higher cardiac titers compared to B6 mice (P < 0.05; S5A Fig).
Exogenous IFNβ treatment has been shown to improve CVB3-induced pathogenesis in mouse models [21, 22] and in some human clinical trials [3, 4]. Since our in vitro data (Fig 3D & 3E) had shown us that the IFIT locus was required for most of the T1IFN-mediated inhibition of CVB3 replication in cardiomyocytes, we next examined whether the IFIT locus is required for IFNβ-mediated beneficial effects in vivo. B6 and IFITKO mice were challenged with 104 PFU of CVB3 and, 24 hours p.i., received a single i.p. injection of either PBS or recombinant IFNβ (2 × 104 units). We monitored the body weight loss of these animals, and found that IFNβ treatment protected B6 mice from weight loss and overt signs of disease over a 12 day period of CVB3 infection, but failed to do so in IFITKO mice (Fig 4H). Surviving mice were sacrificed at day 12 p.i., and virus titers were measured in the pancreas, liver, and heart. For the heart, data are shown in Fig 4I; data for pancreata and livers are shown in S6 Fig. Three conclusions may be drawn: (i) in both mouse strains, there was a clear relationship between substantial body weight loss and high cardiac titers; (ii) most of the B6 mice benefited from IFNβ (cleared virus from the heart, and showed minimal weight loss); and (iii) most importantly, for the purpose of our study, no such effects were observed in IFITKO mice (only one of the IFITKO mice had cleared virus from the heart, and that mouse still had significant weight loss). These data are consistent with the near-complete requirement for IFITs in cardiac clearance. Similar conclusions can be drawn from the pancreatic data, but for the liver it is more difficult to interpret any relationships with virus titer, because the great majority of the livers in both strains scored negative at this time point p.i. (S6 Fig).
Next, we sought to determine why virus clearance is accelerated in pancreas and liver, but not in heart, of IFITKO mice at later time points (days 7 & 11–12 p.i.). Since viral load at early time points is higher in IFITKO cells (Fig 3D) and in IFITKO mice (Fig 4D–4F), we reasoned that this might lead to the more rapid induction of innate responses in these mice compared to B6 animals. Strikingly, at day 1 p.i., several chemokine mRNAs were highly induced in the hearts of IFITKO mice, but only modestly so in B6 tissues (S7 Fig). We chose to carry out a more detailed analysis at 48 hours p.i., because we knew that, by this time point, systemic T1IFNs have driven the upregulation of many ISGs in the infected heart (see Fig 1B & 1C). PCR array heat maps (Fig 5A) revealed that, at 48 hour p.i. many ISGs appeared to be expressed at higher levels in IFITKO tissues than in their WT counterparts. Of interest, too, the pattern of ISG induction differed among the three organs analyzed. To more readily visualize the overall differences between the two mouse strains, the data were re-plotted to compare the range and extent of ISG up-regulation in all three organs of both strains (Fig 5B). In all three organs, the highest ISG up-regulation occurred in IFITKO conditions, and the overall range of ISG expression appeared higher in IFITKO pancreas and heart, and almost equivalent in livers from both mouse strains. Thus, we consider it likely that this increased ISG expression may underpin the faster resolution of infection in the pancreas and liver of IFITKO mice (Fig 4D & 4E). This enhanced virus clearance does not occur in the IFITKO heart (Fig 4F) despite there being increased overall ISG expression (Fig 5B), consistent with the notion that, at least in cardiomyocytes, other ISGs cannot functionally compensate for the loss of the IFIT locus.
Taken together, these in vivo data indicate that constitutive IFIT expression plays a key role in restricting CVB3 replication in most/all tissues and, in its absence, virus RNA-driven induction of the T1IFN response is accelerated; this, in turn, leads to the rapid up-regulation of a variety of other ISGs in all tissues, and these ISGs quickly control CVB3 replication in all of the tested tissues, except the heart. Thus, we suggest that the IFIT locus is especially vital for protecting cardiomyocytes, because these cells lack the functional redundancy that the other ISGs can provide in most cell types.
Finally, we investigated the impact of loss of the IFIT locus on the extent of CVB3-induced myocarditis. Mice were infected with 104 pfu CVB3 i.p. and, 12 days later, hearts were harvested, and paraffin sections were stained (Fig 6A). Immune cell infiltration was quite limited in the B6 heart at 12 days p.i., while numerous infiltrating cells, and collagen deposition (light blue), were observed in the hearts of IFITKO mice. These findings were reproducible when mice were challenged at a dose of 103 pfu of CVB3 (S5B Fig). Since we had observed enhanced chemokine expression in the IFITKO hearts soon after infection (S7 Fig), we analyzed infiltration of macrophages by staining heart vibratome sections with an antibody against Iba-1 (Ionized calcium-binding adaptor molecule 1, also known as Aif-1), a protein that is predominantly expressed on cells of the macrophage lineage [23]. At 12 days post-CVB3 infection (104 pfu), Iba-1 signals (Fig 6B, red) were brighter and more frequent in the hearts of CVB3-infected IFITKO mice compared to the WT animals. Our previous work showed that development of myocarditis in the mice lacking T1IFN signaling into cardiomyocytes was not only exacerbated but also accelerated [7]. Likewise, we found more myocardial inflammation in IFITKO than in B6 mice at 7 days p.i. (Fig 6C, yellow arrows indicate immune infiltrates), a time point when the mice of both groups showed comparably high cardiac virus titers (Fig 4F). Real-time RT-PCR analysis was applied to RNA extracted from the hearts of both mouse strains at d7 p.i., to identify genes expressed by immune cells, and revealed a small but significant increase of Cd8 RNA, and a massive increase of Ly6G, a marker of granulocytes and monocytes, in IFITKO hearts (Fig 6D). Therefore, in the absence of the IFIT locus, there is more rapid inflammatory cell infiltration into the heart. Taken together, these data indicate that the IFIT locus plays an important role in limiting CVB3-induced myocarditis.
The present study was aimed at identifying the genes responsible for T1IFN-dependent antiviral protection in the heart. We report: (i) that the IFIT locus plays a central role in controlling CVB3 infection in multiple tissues including the heart; (ii) that it does so in two distinct phases, separated by the onset of T1IFN signaling; (iii) that the first phase, which depends on constitutive IFIT activity, impacts all analyzed tissues; but (iv) that the second, T1IFN-induced, phase of IFIT activity is cell-specific, being almost indispensable in cardiomyocytes, and redundant in other cell types.
Most cell types can respond to T1IFNs, thereby increasing their ability to resist virus challenge. However, prior to being stimulated by T1IFNs, many cells also have a constitutive capacity to withstand virus infection. This has been referred to as “intrinsic antiviral immunity” [24], and here we demonstrate that the IFIT family genes play such a role in protecting many cell types and tissues against CVB3 infection. Our in vitro observations show that, in the absence of T1IFN treatment, CVB3 replicates to higher titers in IFIT-deficient cardiomyocytes (both in HL-1 cells and in primary isolates), peritoneal macrophages, and cardiac fibroblasts (Fig 3G–3I). Others have reported that, specifically in cardiomyocytes, the mitochondrial antiviral signaling (MAVS) pathway is spontaneously activated, resulting in increased basal levels of IFNβ [25], suggesting the possibility that IFNβ might contribute to the intrinsic immunity of cardiomyocytes to CVB3 that we report herein. However, as noted above, CVB3 infection of primary cardiomyocytes does not trigger abundant IFNβ production. Furthermore, it is intriguing to note that some viral proteases have been shown to cleave MAVS protein, potentially nullifying the pathway’s activity; these include the 3C protease of CVB3 [26], and the 3ABC complex of another picornavirus, hepatitis A virus [27]. Moreover, cytokine responses to CVB3 appear to be independent of MAVS, and CVB3 titers are not increased in MAVS-deficient mice [28]. Given that, both before and after IFNβ stimulation, cardiomyocytes display a near-absolute requirement for endogenous IFITs (Fig 3D–3G), we consider it likely that one or more of the proteins in the IFIT family play(s) the key role in conferring both constitutive and inducible anti-CVB3 protection in these cells.
The importance of constitutive IFIT expression was confirmed by in vivo studies. IFITs are constitutively expressed in many tissues and cell types (S2 Fig panels A, B and D, and Fig 2D) and, compared to WT mice, CVB3 titers at 1 day p.i. were markedly higher in multiple tissues of IFITKO mice (Fig 4D–4F). By two days after CVB3 infection, mice have transitioned from the first phase of antiviral immunity (cell-intrinsic resistance) to the second, T1IFN-induced, phase. At this time point, genetically-intact hearts express many ISGs (Fig 1B & 1C), one of which is IFNβ, whose abundance is reduced ~20-fold if cardiomyocytes are unable to respond to T1IFN (Fig 1C). These data suggest that cardiomyocytes are, by far, the major source of IFNβ in the CVB3-infected heart. Moreover, they demonstrate that extensive T1IFN synthesis by cardiomyocytes requires that the cells be able to respond to the cytokine–i.e., wt cardiomyocytes exhibit a positive feedback loop in vivo, leading to the escalation of local T1IFN concentration, with consequent rapid and marked induction of numerous other ISGs, including several IFITs (Fig 1B and 1C). This is consistent with the in vitro and in vivo observations that suggest that there is a 1–2 day delay in IFIT up-regulation, followed by an explosive increase. What are the antiviral consequences of T1IFN signaling into cardiomyocytes? As noted above, we have reported that T1IFNR-deficient cardiomyocytes show delayed clearance of CVB3 in vivo [7], and here we confirm in vitro the importance of T1IFN signaling in cardiomyocytes; IFNβ treatment reduced the yield of infectious virus by ~2,300-fold in wt HL-1 cells, while no such effect was observed using CRISPR/Cas generated T1IFNR-deficient cardiomyocytes (Fig 2E).
Strikingly, this T1IFN-driven suppression of CVB3 infection in cardiomyocytes is almost entirely dependent on the IFIT locus. IFNβ-treatment of wt HL-1 cells or wt primary cardiomyocytes dramatically inhibited CVB3 replication, but equivalent treatment of IFIT-deficient cardiomyocytes had very little effect, suggesting a near-absolute requirement for the IFIT locus in this cell type (Fig 3D to 3G). In contrast, IFNβ-treated IFITKO peritoneal macrophages very efficiently controlled the infection, while IFNβ-treated cardiac fibroblasts showed an intermediate phenotype (Fig 3H & 3I) indicating that, for both of these cell types, other ISGs could wholly or partially restore antiviral resistance. Thus, IFITs act in a biphasic manner following CVB3 infection, and the T1IFN-driven second phase shows cell-type specificity. This mirrors previous reports of neuron-specific antiviral activity of IFIT2 against vesicular stomatitis virus [29, 30]. The biological importance of this T1IFN-driven shift, from intrinsic immunity (phase 1) to ISG-mediated antiviral responses (phase 2) at ~1–2 days p.i. is clearly shown by the prior observations that IFNβKO, T1IFNRKO and CMMCMT1IFNRf/f mice all failed to control CVB3 infection [6–8], demonstrating that intrinsic immunity alone is insufficient. Furthermore, our data suggest that intrinsic immunity–or, at least, the intrinsic immunity conferred by the IFIT locus–can be lost without fatal effects. In IFITKO mice, virus titers were initially higher than in wt animals–perhaps explaining the early and transient weight loss that occurred in these mice (Fig 4A)–but at later time points virus clearance was accelerated in the pancreas, liver and feces (Fig 4D, 4E & 4G). This enhanced virus clearance is probably attributable to the more rapid induction of T1IFNs and ISGs in IFITKO mice (Fig 5), presumably driven by the extremely high viral titers that were present at 1 day p.i. Consistent with our in vitro data, this accelerated CVB3 clearance showed substantial cell (tissue) specificity: it was not observed in the hearts of IFITKO mice which, at 12 days p.i., still contained much higher levels of CVB3 than did WT hearts (Fig 4F & 4I). In addition, cardiac CVB3 titers in IFITKO mice were prolonged even if the mice were treated with exogenous IFNβ (Fig 4I). Taken together, these findings suggest that other ISGs can substitute for the absence of IFITs in many tissues, but not in the heart, because of cardiomyocytes’ requirement for IFITs. The inability of other ISGs to protect cardiomyocytes against CVB3 after ~day 2 p.i., when the T1IFN system has exerted its effects, appears to render these cells a more hospitable environment for the virus.
As mentioned in the Introduction, virus-mediated direct cell lysis and immunopathology are the two major pathological mechanisms of myocarditis. Previous studies in mice in which the viral receptor (CAR) has been deleted in cardiomyocytes found that the hearts of these mice are resistant to infection, and the mice are largely protected against cardiac disease [31, 32], demonstrating that virus replication in cardiomyocytes is a prerequisite for myocarditis. Our data showed that enhanced cardiac virus replication in IFITKO mice was accompanied by accelerated and exacerbated myocarditis (Fig 6 and S5B Fig). However, in contrast to our tissue culture data, which show clearly that IFIT expression in cardiomyocytes is the key factor, these in vivo data in IFITKO mice–which show that IFITs protect against myocarditis–are open to interpretation. It is possible that, in genetically-intact mice, protection against myocarditis is mediated solely by IFITs in cardiomyocytes (paralleling the tissue culture data), but it also is possible that protection is mediated, at least in part, by the early actions of constitutively-expressed IFITs in multiple tissues. This issue can be resolved in the future by generating mice carrying a floxed IFIT locus, and crossing them to CMMCM mice [which are described in ref 7], allowing the inducible deletion of the locus specifically from cardiomyocytes. Whichever mechanism is in play, the primary means by which the IFIT locus prevents viral myocarditis in WT animals is to inhibit CVB3 replication, thereby reducing both virus-mediated direct cell lysis and the immunopathological damage caused by infiltrating cells. In addition, we observed that the increased early viral load in the IFITKO mice is accompanied by a more rapid T1IFN response (Fig 5). Interestingly, in Sendai virus (SeV) -infected Ifit2 KO mice, which show elevated T1IFN induction triggered by uncontrolled virus replication, both SeV infection and abnormal production of T1IFN are required for the virus pathogenesis [33]. Hence, during CVB3 infection of IFITKO mice, the rapid / robust induction of the T1IFN response together with the prolonged virus presence in the heart may synergistically promote the cardiac immune cell infiltration and contribute to the pathogenesis of myocarditis.
In conclusion, we have revealed a key role for the IFIT locus in modulating CVB3 infection. Through analyzing genetically-manipulated cells and mice, we show that the IFIT locus constitutively limits early virus replication in many tissues, and that its subsequent upregulation by the T1IFN response plays an especially-important role in cardiomyocytes, delaying or preventing the development of myocarditis. Future studies to reveal the precise mechanism(s) by which the IFIT locus acts, and to determine the function of each IFIT family gene, may lead to new and improved strategies for treating enterovirus-induced disease.
All animal experiments were approved by The Scripps Research Institute (TSRI) Institutional Animal Care and Use Committee (protocol number 09-0131-3) and were carried out in accordance with the National Institutes of Health’s Guide for the Care and Use of Laboratory Animals.
C57BL/6 mice were purchased from the TSRI rodent breeding colony. Generation of IFITKO mice by the Sen laboratory will be described in detail elsewhere. CMMCM T1IFNRf/f mice were described previously [7].
Mice serum was isolated using K3-EDTA coated microvette tubes (Starstedt, Nümbrecht, DEU) from the blood by centrifuging at 12,000 rpm for 15 min at room temperature, and stored at -20 °C until its use.
The cardiomyocyte cell line HL-1 [12] was obtained from Drs. William C. Claycomb and Ikuo Tsunoda at Louisiana State University (Shreveport, USA). The cells were cultured in Claycomb medium supplemented with 100 μM norepinephrin, 10% fetal bovine serum (FBS) and 4 mM L-glutamine, in 37°C, 5% CO2 atmosphere. The Pierce primary cardiomyocytes isolation kit (Thermo Fisher Scientific, #88281) was used to isolate primary cardiomyocytes and primary cardiac fibroblasts, following the manufacturer’s instructions with some modifications [17]. In brief, day 2 post neonate mice were sacrificed and the hearts were isolated. After mincing each heart into 1–3 mm3 pieces, the tissues were washed twice with HBSS (Hanks-based salt solution), resuspended and incubated in working solution including primary cardiomyocyte isolation enzymes 1 and 2 (components of the kit) at 37°C for 30 minutes. The tissues were washed with HBSS several times and then with DMEM. Cells were plated at a density of 1.25 x 106 per well in six-well plates for 2 hours in order to separate cardiac fibroblasts (rapidly adhering) from cardiomyocytes (still floating at 2 hours after plating). Adherent cells were washed with PBS and cultured for nine days, changing media every 3 days. Then, cell clusters that do not include any beating cells were collected and used as primary cardiac fibroblasts. For cardiomyocyte culture, 24 hours after plating the cells, fresh medium was added, with cardiomyocyte differentiation supplement (another component of the kit). After 7 days’ growth and differentiation with one medium change at day 4, cells were used as primary cardiomyocytes. For peritoneal macrophage isolation, mice were injected with aged 3% thioglycollate medium (SIGMA Aldrich, #T9032) i.p.. Three days later the mice were sacrificed, and peritoneal cells were recovered by lavage and seeded onto a tissue culture plate. The next day, floating cells were removed, and the adherent cells were used as macrophages.
The wild-type CVB3 used in these studies is a plaque purified isolate (designated H3) of the myocarditic Woodruff variant of CVB3. Mice were infected intraperitoneally (i.p) with the indicated dose of CVB3 and their survival and body weight were monitored during the course of infection. At the indicated times p.i., feces were collected, and at the time of sacrifice, pancreata, livers and hearts were isolated, weighed, and homogenized in 1 ml Dulbecco modified Eagle medium (DMEM). Virus titers were assessed using standard plaque assays as previously described [34].
T1IFN-related gene expression in pancreata, livers and hearts was quantified using Mouse Type I interferon response RT2 Profiler PCR Array (PAMM-016Z, SA Biosciences, Frederick, MD). For real-time RT-PCR analysis, RNA was isolated from tissue and cell suspensions using the RNeasy Mini Kit (QIAGEN, # 74104), and 1–2 μg of RNA was reverse transcribed using iScript Reverse Transcription supermix (Bio-rad, #1708841). Real-time PCR was performed using Power SYBR Green PCR mastermix reagent (Applied biosystems, #4367659) with specific primer sets (see S1 Table). All of the values in PCR array analysis and real-time PCR analysis were normalized to the values of Gapdh.
2 x 105 HL-1 cells were seeded onto gelatin/fibronectin-coated plate. 24 hours later, pX459 (ver. 2) encoding either sgIfnar1 or sgIfit1 and sgIfit2 was transfected into HL-1 cells and incubated for further 24 hours. The sequences of the sgIFITs are shown in S1 Table. Then culture medium was changed to the media containing puromycin (3 μg/ml) and Cas9-expressing cells were selected for three days. After drug selection, puromycin was removed from culture media and cells were recovered. These bulk gene-edited cells were used for in vitro studies at early passage numbers.
To determine surface IFNAR1 protein expression, HL-1 cells were analyzed by flow cytometry. WT or Ifnar1-edited HL-1 cells were incubated in trypsin-EDTA at 37°C for 5 min. Then, the reaction was stopped by adding DMEM supplemented with FBS. After washing several times, isolated cells were incubated with PE-conjugated anti-IFNAR1 antibody or PE-conjugated isotype control on ice for 20 min. After washing several times with FACS buffer, IFNAR1 expression on HL-1 cells was analyzed by flow cytometry using an LSR II (BD Bioscience).
Cells were lysed in RIPA buffer (Millipore, #20–188). After centrifugation, cell debris was discarded and protein concentration in the supernatant was determined by Pierce BCA Protein Assay Kit (Thermo Fisher Scientific, #23225). Colorimetry was measured using plate reader, Victor X3 (Perkin Elmer). 5 μg of total cell lysates were mixed with Laemmli Sample Buffer (Bio-Rad, #161–0747) and 10% of 2-Mercaptoethanol (SIGMA Aldrich, #M6250) and used for western blotting. Blotting of the proteins to membrane was performed using Trans-blot Turbo RTA Transfer Kit (Bio-rad, #170–4272) as follows. After developed on SDS gels, proteins were transferred on PVDF membranes (Bio-Rad, a component of the Transfer Kit) by using Trans-blot Turbo system (Bio-Rad). Membranes were then blocked with 1% skim milk for an hour, and overnight with the relevant diluted primary antibodies. Then, membranes were washed three times with Tris buffer Tween 20 (TBST) and incubated with diluted secondary antibodies. One hour later, membranes were washed again three times with TBST, then protein-antibody complexes were visualized by Super Signal ELISA Femto Maximum Sensitivity Substrate (Thermo Scientific, #37074).
Mice were perfused with Dulbecco’s PBS (DPBS), and tissues were harvested and fixed using buffered zinc formalin at room temperature (RT) overnight. For standard histological analyses, tissues were paraffin embedded, and 3-μm sections were cut and stained with hematoxylin-eosin or Masson’s trichrome. Images were captured at 10x magnification with an BZ-X710 inverted microscope (KEYENCE) using BZ-X Viewer software (KEYENCE). For confocal studies, 70 μm sections of liver and heart were cut with a Leica VT 1000S vibratome. Sections were incubated with primary antibody for 1 hr at RT and then at 4°C overnight. After washing, they then were incubated with secondary antibody for 1 hr at RT, washed and then incubated with Phalloidin 488 at 4°C overnight to label F-actin. After incubation, sections were washed, counterstained with Hoechst 33342 and mounted with ProLong Gold Antifade Mountant for confocal microscopy. Confocal images were captured using a Zeiss LSM 710 Laser Confocal Scanning Microscope running Zen 2009 Zeiss software suite. Representative regions within each vibratome section of the tissues were scanned as 8-bit optical sections (1,024 × 1,024 image sizes) and reconstructed for analysis. Exposure and image acquisition settings were identical for all sections.
All data were analyzed using Prism software (GraphPad Prism 8). An unpaired, two-tailed t-test was used to determine statistically significant differences for in vitro experiments. The Mann-Whitney test was used to analyze differences in viral burden. Kaplan-Meier survival curves were analyzed by the log rank test. P values less than 0.05 were considered significant, and are indicated in figures as follows: * 0.05>p>0.01; ** 0.01≥p>0.001; *** 0.001≥p>0.0001; **** p≤0.0001.
|
10.1371/journal.pcbi.1000496 | Global Motions of the Nuclear Pore Complex: Insights from Elastic Network Models | The nuclear pore complex (NPC) is the gate to the nucleus. Recent determination of the configuration of proteins in the yeast NPC at ∼5 nm resolution permits us to study the NPC global dynamics using coarse-grained structural models. We investigate these large-scale motions by using an extended elastic network model (ENM) formalism applied to several coarse-grained representations of the NPC. Two types of collective motions (global modes) are predicted by the ENMs to be intrinsically favored by the NPC architecture: global bending and extension/contraction from circular to elliptical shapes. These motions are shown to be robust against tested variations in the representation of the NPC, and are largely captured by a simple model of a toroid with axially varying mass density. We demonstrate that spoke multiplicity significantly affects the accessible number of symmetric low-energy modes of motion; the NPC-like toroidal structures composed of 8 spokes have access to highly cooperative symmetric motions that are inaccessible to toroids composed of 7 or 9 spokes. The analysis reveals modes of motion that may facilitate macromolecular transport through the NPC, consistent with previous experimental observations.
| The nuclear pore complex (NPC) serves as the sole gateway to the cell nucleus, and its proper functioning is therefore crucial for gene expression and many vital signaling pathways. Although it is typically circular, the overall structure of the NPC has been observed to change in response to the presence of cargo. Recently, the molecular architecture of the yeast NPC, including the shapes and relative positions of its constituent proteins, has been resolved. These new structural data provide us with a first opportunity to construct an accurate dynamical model of a macromolecular machine containing hundreds of proteins. By modeling the NPC as a network of masses connected by springs, we investigate its probable large-scale dynamics. We start from a very coarse model and gradually refine it, observing how the structural details influence the calculated dynamics. We find that the NPC dynamics are quite similar to those of a flexible toroid with an uneven mass distribution, and that the 8-fold symmetry that is universally observed in NPCs enables them to undergo certain collective motions that are inaccessible to structures of other symmetries.
| Any macromolecule entering or exiting the nucleus must pass through a nuclear pore complex (NPC). NPCs are formed by hundreds of proteins organized into cylindrically symmetric pores that act as gateways to the cell nucleus. The NPC has a mass of ∼50 MDa in yeast and up to 125 MDa in vertebrates, comparable to a small organelle. The yeast NPC is a ring of about 100 nm diameter, creating a central pore of ∼30 nm diameter. It contains approximately 450 proteins, termed nucleoporins, arranged into eight similar “spokes” extending from the NPC's central channel to its outer perimeter (Figure 1). Each spoke of the yeast NPC exhibits quasi-twofold symmetry about the lumenal plane of the nuclear envelope (NE), giving the NPC quasi-sixteenfold symmetry and reducing the number of unique proteins in the complex to only about 30. The NPC central channel is coated by several “FG nucleoporins” that characteristically contain multiple structurally disordered phenylalanine (F) and glycine (G) repeats. The FG nucleoporins collectively serve as a selective barrier between the nucleoplasm and the cytoplasm: Small (<5–30 kDa) particles diffuse freely through the NPC channel, but larger particles require the assistance of karyopherin transport factors to pass through the entropic barrier created by the FG nucleoporins. Transport through the NPC is passive, and the rate of material exchange through the NPC can be accounted for by the concentration gradient of the karyopherins and their cargoes [1].
The NPC is somewhat plastic [2] and has been observed to undergo large-scale conformational fluctuations in the form of elongation or dilation under a variety of conditions. Nuclear export of large mRNA molecules seems to be facilitated by dilation of the NPC [3],[4]. Similarly, some modes of viral infection suggest that the NPC channel widens concurrently with import of viral genetic material [5]: Viral infection often includes the nuclear import of intact capsids, as is the case with papovaviruses [6],[7] and hepatitis B virus [8], or import of the intact viral genome, as occurs with adenoviruses [9] and type-1 herpes simplex virus [5]. The NPC also dilates in response to cation concentration [10] and to certain steroids [11],[12], possibly as a mechanism to regulate transport rates [11]. Recent studies with cryo-electron tomography [13],[14] highlight the structural changes of the NPC during transport, indicating that the central channel alters its profile in response to the presence of cargo.
Although conventional models of nuclear transport focus primarily on the local interactions of FG nucleoporins with karyopherins [15]–[21], a number of mechanisms have been proposed to account for the large-scale motions of the NPC [22],[23]. The limiting factor in determining the nature of the global NPC motions has been the resolution of available data. Cryo-EM and electron tomography images of NPCs reveal global features of the complex, while blurring over its finer details. Alternatively, X-ray crystallography captured the structures of several individual nucleoporins to atomic resolution, but it does not provide the relative positions or orientations of the nucleoporins within the assembled NPC. Recently, an integrative approach has been developed to determine the molecular architectures of macromolecular assemblies from diverse data [24]. Combining spatial restraints that account for the NE excluded volume (from EM), nucleoporin excluded volumes (from the protein sequences and ultracentrifugation), protein positions (from immuno-EM), protein contacts (from affinity purification), and the eight-fold and two-fold symmetries of the NPC (from EM), this approach was used to generate a coarse-grained structural model for the yeast NPC, defining the relative positions of its constituent proteins [25]. Although each individual restraint contains little structural information, the concurrent satisfaction of all restraints from independent experiments drastically reduces the degeneracy of the structural solutions and yields pronounced maxima in the localization probabilities of almost all nucleoporins, thus leading to one predominant NPC architecture [25]. This architecture provides us with enough detail to construct a dynamical model of the NPC at the resolution of a single protein, enabling the exploration of the large-scale dynamics of the NPC.
A dynamical model that has enjoyed considerable success in elucidating the machinery of biomolecular assemblies near physiological conditions is the elastic network model (ENM) [26]. In the simplest ENM, each of the constituent bodies in the system is represented as a point mass, or node, and a network is constructed by joining neighboring nodes through elastic edges that supply linear restoring forces to displacements from equilibrium. ENMs are straightforwardly paired with normal mode analysis (NMA) to provide analytical solutions to motions accessible to biomolecular systems under native state conditions. It has been repeatedly shown that the collectives modes predicted by the ENM correlate well with experimentally observed functional changes in structure triggered by substrate binding or activation [27]. This observation supports the idea that biomolecular systems possess intrinsic dynamics that are uniquely defined by their three-dimensional (3D) structures [27], in accord with experimental observations [28]. Further, ENMs are readily extensible to large systems through a variety of coarse-graining techniques, and they have been applied at various degrees of detail to systems ranging in scale from small proteins to supramolecular assemblies such as the ribosome [29],[30] and viral capsids [31],[32]. The robustness of the global modes predicted by the ENMs and their elegant simplicity and scalability lead us to explore the dynamics of the NPC using ENM-based theory and methods. The known agreement between ENM normal modes and molecular fluctuations can be exploited for structural refinement [33],[34]. Starting from an initial structure, one can iteratively construct an ENM and modify the structure until the structure matches the observed low-resolution density data. Such techniques are commonplace in generating atomic coordinates to fit cryo-EM data [35]–[37], and may similarly be used to refine the structure of the NPC.
Here, we examine the dynamics of the NPC at several levels of coarse-graining with a newly developed version of a classical ENM. We begin by modeling the NPC on the coarsest scale possible, i.e., representing the full complex as a flexible toroid. We then increment the level of detail to include non-uniform mass distributions, spokes, and finally the full NPC architecture. We find that the global modes of the NPC are largely captured by the modes of a simple toroid, suggesting that the global dynamics of the NPC does not depend on the details of the NPC architecture. We note that the eight-fold symmetry permits ease of certain cooperative motions that are inaccessible to structures with other rotational symmetries. Two highly robust modes of collective deformation emerge from the analysis, providing insight into the details of NPC motions that are observed with low-resolution tomography [13],[14].
|
10.1371/journal.pgen.1003652 | Disease-Related Growth Factor and Embryonic Signaling Pathways Modulate an Enhancer of TCF21 Expression at the 6q23.2 Coronary Heart Disease Locus | Coronary heart disease (CHD) is the leading cause of mortality in both developed and developing countries worldwide. Genome-wide association studies (GWAS) have now identified 46 independent susceptibility loci for CHD, however, the biological and disease-relevant mechanisms for these associations remain elusive. The large-scale meta-analysis of GWAS recently identified in Caucasians a CHD-associated locus at chromosome 6q23.2, a region containing the transcription factor TCF21 gene. TCF21 (Capsulin/Pod1/Epicardin) is a member of the basic-helix-loop-helix (bHLH) transcription factor family, and regulates cell fate decisions and differentiation in the developing coronary vasculature. Herein, we characterize a cis-regulatory mechanism by which the lead polymorphism rs12190287 disrupts an atypical activator protein 1 (AP-1) element, as demonstrated by allele-specific transcriptional regulation, transcription factor binding, and chromatin organization, leading to altered TCF21 expression. Further, this element is shown to mediate signaling through platelet-derived growth factor receptor beta (PDGFR-β) and Wilms tumor 1 (WT1) pathways. A second disease allele identified in East Asians also appears to disrupt an AP-1-like element. Thus, both disease-related growth factor and embryonic signaling pathways may regulate CHD risk through two independent alleles at TCF21.
| As much as half of the risk of developing coronary heart disease is genetically predetermined. Genome-wide association studies in human populations have now uncovered multiple sites of common genetic variation associated with heart disease. However, the biological mechanisms responsible for linking the disease associations with changes in gene expression are still underexplored. One of these variants occurs within the vascular developmental factor, TCF21, leading to dysregulated gene expression. Using various in silico and molecular approaches, we identify an intricate allele-specific regulatory mechanism underlying altered expression of TCF21. Notably, we observe that two apparently independent risk alleles identified in distinct populations function through a similar regulatory mechanism. Together these data suggest that conserved upstream pathways may organize the complex genetic etiology of coronary heart disease and potentially lead to new treatment opportunities.
| A recent meta-analysis of 14 Genome-wide association studies (GWAS) for CHD, Coronary ARtery DIsease Genome-wide Replication And Meta-analysis (CARDIoGRAM), including 22,233 cases and 64,762 controls in Europeans, elucidated 13 novel susceptibility loci [1]. One of these novel loci includes a variant, rs12190287 at 6q23.2, located within the 3′ untranslated region (3′UTR) of TCF21 [1]. This lead SNP at 6q23.2 had the lowest P value (P<4.6×10−11) of the novel loci in the meta-analysis and was also highly associated in the combined meta-analysis (P<1.1×10−12). rs12190287 was also identified as an expression quantitative trait locus (eQTL) through correlation with increased TCF21 gene expression in both liver and adipose tissue [1], [2]. Importantly, the TCF21 locus was recently replicated in another GWAS for CHD in a Han Chinese population (15,460 cases and 11,472 controls), however a second variant (rs12524865) that is poorly correlated with rs12190287 and located 14 kb upstream of TCF21 was the lead SNP in this racial ethnic group [3].
TCF21 is a member of the basic helix-loop-helix (bHLH) transcription factor (TF) family and regulates cell differentiation and cell fate decisions during development of the coronary vasculature, lung, kidney, and spleen [4], [5]. Tcf21 is expressed in mesodermal cells in the proepicardial organ (PEO) as early as E9.5 in mice, and later in mesenchymal cells forming the pericardial layer [4]. Recent elegant studies employing knockout mice have established a specific role for this factor in the origin of coronary artery smooth muscle cells and cardiac fibroblasts [6], [7]. Loss of Tcf21 expression in mouse results in increased expression of SMC markers in cells on the heart surface consistent with premature SMC differentiation [7], and a dramatic failure of cardiac fibroblast development [6], [7]. These data are most consistent with a role for Tcf21 in a bipotential precursor cell for SMC and cardiac fibroblast lineages, with loss of Tcf21 expression being essential for SMC development, and persistent Tcf21 expression being required for cardiac fibroblast development [6], [7].
In studies described here we examine the function of a regulatory element at the lead variant rs12190287 though allele-specific reporter assays, gel mobility shift assays, and haplotype specific chromatin immunoprecipitation (haploChIP). We further demonstrate allele-specific regulation of TCF21 gene expression, though modulating this regulatory element, via platelet derived growth factor receptor beta (PDGFR-β) and Wilms tumor 1 (WT1) mediated signaling. Lastly, we identify a conserved AP-1 dependent mechanism acting upstream of TCF21 at rs12524865, which was recently associated with CHD in East Asians [3]. Taken together, these studies elucidate both disease-related and embryonic pathways upstream of TCF21, at two independent risk alleles, thus providing further pathophysiological insight into the common heritable risk of CHD.
The CARDIoGRAM meta-analysis in Caucasians identified rs12190287 as the lead CHD association at 6q23.2 (P<4.64×10−11), which was 3 orders of magnitude more significant than other SNPs in this region (Figure 1A) [1]. We then set out to identify potential causal risk-associated mechanisms at 6q23.2 using an integrated workflow (Figure 1B). We examined the overall linkage disequilibrium (LD) plot around the TCF21 locus at 6q23.2 using 1568 individuals of European descent genotyped on the fine-mapping Metabochip array [8] (Illumina), which contains 196,726 polymorphisms [8], [9], of which approximately 280 markers were contained in the 170 kb block between Chr6: 134,171,000–134,341,000. We identified regions of high LD surrounding the lead SNP, rs12190287, which is located in the 3′UTR of the non-coding exon of the long variant 1 of TCF21. Two haplotypes contained the high-risk allele at the lead SNP rs12190287, with one containing 3 out of the 5 additional eSNPs, while the haplotype with frequency 0.366 contained all of the eSNPS in the TCF21 locus (Figure 1C and 1D). We examined the LD plot containing the eSNPs for TCF21 (Figure 1C) and found that none of these variants had r2 values >0.8 with the lead SNP rs12190287, suggesting that if a single variant is responsible for the association observed by CARDIoGRAM, it is most likely to be rs12190287. This LD pattern is also consistent with the observation that rs12190287 was 3 orders of magnitude more significant than other SNPs in the region [1]. We then mapped regulatory chromatin regions surrounding rs12190287 using the ENCylopedia Of DNA Elements (ENCODE) Integrated Regulation data from 7 cell lines (Figure 1E), which demonstrated enrichment of the enhancer mark H3K4me1. Enrichment for histone modifications H3K4me3 (marks promoters), and H3K27ac (marks active regulatory elements) were also observed to a lesser extent. We also found regions of DNaseI hypersensitivity for open chromatin and overlapping RNA-seq peaks for transcriptional activity in the region containing rs12190287. We validated these histone modification and DNaseI ENCODE data with our own ChIP-seq and FAIRE-seq (Formaldehyde-Assisted Isolation of Regulatory Elements) experiments in HCASMC (Figure S1), demonstrating consistent regulation at rs12190287 and relevance to CHD. We also mapped the transcription factor binding sites (TFBS) using ENCODE data, which identified enrichment of an activator protein 1 (AP-1) component, JUND in a human embryonic stem cell line (Figure 1E).
Given that rs12192087 appeared to be the most likely causal variant at 6q23.2, we proceeded with in vitro functional studies to identify the risk-associated mechanisms through this variant. First, we mapped the putative TFBS in silico using various bioinformatics tools, including TRANSFAC, PROMO, MatInspector, and TFSearch (Table 1). Interestingly, multiple AP-1 TFs were predicted to preferentially bind to the major risk C allele, containing the binding motif, TGACTTCA (Figure S2A). Luciferase reporters containing the putative binding site for the risk and protective alleles were then transfected into various cell lines, including HepG2, HEK, and A7r5, as well as primary human coronary artery smooth muscle cells (HCASM), and rat aortic smooth muscle cells (RASM). We observed approximately 150–200 fold increase in activity with the rs12192087-C (C-Luc) reporter relative to the empty vector reporter, and this relative activity of the C-luc reporter was ∼20-fold greater than the rs12190287-G reporter (G-Luc) (Figure 2A). Similar results were observed in primary HCASM and RASM, suggesting that the G allele disrupts TF binding, leading to reduced TCF21 transcription. Given the ubiquitous expression of AP-1 factors in various cell types, it is not surprising that cell-specific activity was not observed. Interestingly, the allele-specific difference in transcription was lost when we mutated the 8-mer binding site to create a classical AP-1 7-mer (closely resembling a TPA element), but not when we mutated to another atypical AP-1 8-mer (Figure 2B). This is consistent with predicted binding of either c-Jun homodimers or c-Jun/ATF heterodimers, rather than classical c-Jun/c-Fos AP-1 complex [10], to confer allele-specific transcriptional regulation. In order to measure relative TF protein binding we performed electrophoretic mobility shift assays (EMSA). We observed binding to both radiolabeled alleles containing a single putative binding site in nuclear extracts from various cell types (Figure 2C). Greater binding to the C risk allele was observed, while competition with excess cold probe was more effective at the G allele. These results are consistent with the reporter assays, suggesting the G allele has weaker transcriptional regulatory activity due to decreased TF binding.
Given that the putative binding site closely resembles an AP-1 or CRE element, we measured the effects of activating protein kinase C (PKC) via phorbol-12-myristate-13-acetate (PMA) or adenylyl cyclase (AC) via forksolin (fsk) on the transcriptional activity at rs12190287 (Figure S2B). Surprisingly, neither PMA nor forskolin altered C or G-Luc reporter activity, while both activated the consensus AP-1 and CRE reporters, respectively. This may indicate the element at rs12190287 can be activated in normal growth media, which contains growth factors upstream of AP-1. Overexpression of constitutively active MEKK (preferentially upstream of AP-1 elements), but not active protein kinase A (PKA) (preferentially upstream of CRE elements) led to greater transactivation of the C allele (Figure 2D). We observed an increase in the bound TF complex upon PMA stimulation, which was greater at the C allele, suggesting binding of an AP-1-like element (Figure 2E). This is further supported by the gel shift observation of a similar higher molecular weight complex bound to C and G alleles, compared to the consensus AP-1 probe (Figure S2C). Interestingly, a second lower molecular weight complex bound to the consensus AP-1 probe was not observed with C and G probes, suggesting some differences in TFs binding to these distinct elements.
We then sought to determine the specific identity of the TFs predicted to bind rs12192087 in vitro. Using allele specific reporters for rs12190287, we measured the regulatory effects of overexpression of AP-1 related factors meeting a predicted in silico binding threshold of >0.85, which included c-Jun, JunD, and ATF3 (Figure 3A). c-JUN overexpression elicited robust activation of the C allele, similar to the activation of consensus AP-1-luc. Less overall activity was observed with the G allele and minor effects were observed upon JUND and ATF3 overexpression. Prior reports also demonstrate that c-Jun predominately activates AP-1 elements in vitro, whereas JunD and ATF3 alone often result in transrepression [11]. We also measured the transcriptional regulation at rs12192087 via loss-of-function experiments. The dominant negative mutant ΔJun (TAM67), which lacks the transactivation domain of c-Jun, resulted in blunted transcriptional activity at C and G alleles (Figure 3B). Similar results were observed with ΔATF3, whereas ΔCREB led to slightly increased activity. Transfection of siRNAs against c-JUN, JUND, and ATF3 also led to reduced transcriptional activity at the C and G alleles, specifically implicating these factors in mediating the activity at rs12190287 (Figure 3C). siRNA-mediated protein knockdown for each AP-1 TF was confirmed by immunoblotting (Figure S3). Using EMSA we also observed super-shifted complexes upon incubation with antibodies against c-Jun and JunD (Figure 3D). Together these results implicate the AP-1 factors c-Jun, JunD, as well as ATF3 in regulating putative enhancer activity at rs12190287.
To ascertain the functional implications of allelic variation at rs12190287, we measured the effects of relevant upstream stimuli on TCF21 expression in HCASMC. Platelet-derived growth factor (PDGF) is a potent growth factor ligand responsible for activation of AP-1-dependent gene expression in SMCs, leading to synthetic phenotypic properties such as increased proliferation, survival, and migration [12]. Further, signaling through PDGFRβ is required for epithelial-mesenchymal transition (EMT) and epicardial fate determination in developing CASMC [13]. Transforming growth factor beta (TGF-β1) is also critically involved in both EMT and adult SMC phenotypic modulation. Interestingly, PDGF-BB treatment resulted in a time-dependent increase in TCF21, whereas transforming growth factor beta (TGF-β1) and PMA led to slightly reduced or unaltered TCF21 levels (Figure 4A). Western blots demonstrated changes in TCF21 protein levels were consistent with changes in TCF21 mRNA levels in response to PMA and PDGF-BB (Figure S4). Concordantly, c-JUN and ATF3 were upregulated by PDGF-BB, while JUND levels were unchanged (Figure 4B). We then assessed the effects of PDGF-BB and TGF-β1 treatment on allele-specific expression (ASE) of TCF21 in HCASMC using TaqMan allelic discrimination (Figure 4C, Figure S5A, B). Using heterozygous CG HCASMC, we observed that PDGF-BB treated samples had greater normalized C/G ratios, which peaked around 6 hours, while TGF-β1 treated samples had lower C/G ratios (Figure 4C). The phasic allelic imbalance observed with PDGF-BB may be partially dependent on activation of AP-1 to regulate TCF21 gene expression.
Transcriptional regulation of gene expression is tightly controlled by the native chromatin architecture in vivo. Therefore, we interrogated allele-specific AP-1 occupancy at rs12190827 using chromatin immunoprecipitation (ChIP) and haplotype specific ChIP (haploChIP). In HCASMC treated with PDGF-BB we observed a significant increase in enrichment at rs12190287 by c-Jun and ATF3 (Figure 5A). JunD enrichment, while significantly above IgG in control samples, was unchanged with PDGF-BB. We then measured allele-specific enrichment in heterozygous HCASMC treated with PDGF-BB, as done previously for haploChIP [14]. Interestingly, c-Jun was predominately enriched at the C allele under control treatment, indicated by greater C/G ratios (Figure 5B). Both c-Jun and ATF3 were more enriched at the C allele upon PDGF-BB treatment, and JunD enrichment was unchanged. Similar observations were made using pyrosequencing-based allelic discrimination (Figure S5C). ChIP products were also amplified at FOSB and MYOG promoters, as AP-1 positive and negative control regions, respectively (Figure 5C, D). We then measured putative enhancer activity at rs12190287 via post-transcriptional histone modification. PDGF-BB treatment led to significantly increased enrichment of H3K4me1 (marks active/poised promoters) and H3K27ac (marks active enhancers) and H3K27me1 (marks active/poised promoters) (Figure 5E). We also observed increased relative enrichment of active histone modifications at the C allele, which was further potentiated with PDGF-BB stimulation in HCASMC (Figure 5F). These data indicate that the AP-1 complex positively regulates the rs12190287 risk allele in the native and active chromatin state.
We also investigated the potential functional effects of non-AP-1 TFs predicted to bind rs12190287 (Table 1). Wilms tumor 1 (WT1) was of particular interest given its known role in CASMC development [15], and evidence in developmental models indicating wt1 directly regulates tcf21 [16]. The G allele resides in a WT1-like binding element [17] (WTE; 5′-GCGTGGGAGT-3′), which was previously implicated in the regulation of the human thromboxane A2 receptor [18]. WT1-D (+KTS amino acid insertion) binds DNA with reduced affinity compared to WT1-B (−KTS) [19], and the ratio of the two alternatively spliced isoforms has been implicated in Frasier syndrome [20]. We observed that expression of WT1-B (−KTS) and WT1-D (+KTS) led to similar transrepression of both C and G alleles at rs12190287 (Figure 6A), consistent with the role of WT1 as a transcriptional repressor. As a tumor suppressor gene, WT1 often represses AP-1 mediated transcription [21], [22] and WT1-B and WT1-D also repressed c-Jun-mediated activation of rs12190287 in vitro (Figure 6B). Similar regulation was observed at both C and G alleles suggesting WT1-mediated regulation may not be the rate-limiting step altered by rs12190287. Consistently, WT1 siRNA mediated knockdown led to increased activity of C and G alleles (Figure 6C), with protein knockdown verified by immunoblotting (Figure S3D). Next, we assessed the expression changes of WT1 upon TGF-β1 or PDGF-BB treatment of HCASMC (Figure 6D). Interestingly, PDGF-BB led to a rapid decline in WT1, which recovered by 24 hours. TGF-β1 led to a slower yet persistent reduction of WT1. We also observed decreased enrichment of WT1 at rs12190287 upon PDGF-BB stimulation, consistent with effects observed at the FOSB promoter (AP-1 positive control region) (Figure 6E). Less reduction was observed at the MYOG promoter (AP-1 negative control region). Surprisingly we observed WT1 preferentially enriched at the C allele, which was increased upon PDGF-BB stimulation (Figure 6F). Given that WT1 negatively regulates transcription at rs12190287 and is downregulated by PDGF-BB, may suggest that the WT1 cofactor preferentially associates with the C risk allele to temporally fine-tune AP-1 mediated transcription upon growth factor stimulation.
TCF21 was one of three Caucasian CAD associated loci that was recently replicated in a Han Chinese population [3]. While association at rs12190287 did not reach genome-wide significance at the TCF21 locus, rs12524865 (Figure 1A) represented the lead SNP in this racial ethnic group and showed consistent association in the discovery and replication stages [3]. We examined the haplotype structure at 6q23.2 in ∼2400 Han Chinese from the HALST study in Taiwan who were also genotyped with the Metabochip, and augmented genotype data in this region through imputation using data from the HapMap II and III Han Chinese (CHB) samples. We identified regions of high LD surrounding rs12190287, however much less LD surrounding rs12524865 compared to the European samples, and we found that rs12524865 is located in a distinct haplotype block from the lead SNP in Europeans, rs12190287 (Figure 7A). Further, the risk haplotype containing rs12524865 and four other TCF21 eSNPs occurs at similar frequency (0.361) in this population (Figure S6A). rs12524865 is in perfect LD with other eSNPs for TCF21 in one haplotype block, including rs1967917 and rs7752775. However, the haplotype block for rs12190287 does not contain other alleles in LD (Figure 7A). We then mapped the putative TFBS at rs12524865 using multiple prediction tools (Table 2). Interestingly, rs12524865 is also located within an AP-1/CREB-like element, TAA[C/A]GTCA, which closely resembles the consensus ATF/CREB binding site, TGACGTCA (Figure S6B). As expected, the C allele (also major, risk allele) is predicted to bind AP-1 and CREB family members, whereas predicted binding is disrupted by the minor, protective allele. Using luciferase reporters containing this putative enhancer we observed robust transcriptional activity with the risk allele, which was absent with the protective allele (Figure 7B). Forskolin but not PMA stimulation potentiated this activity, suggestive of a cAMP-responsive ATF/CREB element (Figure 7B). Dominant negative mutants of CREB, Jun, and ATF3 reduced this activity (Figure S6C). We then measured occupancy of AP-1 and active chromatin histone modifications at rs12524865. Interestingly, we observed increased c-Jun and ATF3 enrichment with PDGF-BB treatment, with much greater enrichment by ATF3 (Figure 7C). Enrichment of active histone modifications, H3K4me1 and H3K27ac also suggest a functionally active chromatin state at rs12524865 (Figure 7D). Together these results implicate both rs12190287 and rs12524865 risk alleles in a shared AP-1-dependent mechanism for regulating TCF21 in HCASMC, thus further defining the genetic risk mechanisms of CHD which have been conserved across racial ethnic groups during evolution (Figure 8).
Therapeutic targeting of traditional CHD risk factors has reduced overall mortality rates, however there are currently no therapies that directly target disease processes of the vessel wall. Recent GWAS have identified 46 independent risk-associated loci for CHD/MI, and 104 independent loci associated at a false discovery rate <5% [8], [23]. Many of the genes identified are implicated in the regulation of SMC plasticity during atherosclerosis, including PDGFD, COL4A1/2, CDKN2B and CDKN2A/p19ARF [24]–[26]. However, the molecular mechanisms and relevant pathways underlying these risk associations are relatively underexplored. Here, using an integrated beyond GWAS strategy (Figure 1B), we reveal the interplay of both developmental and disease-related pathways, which coordinate the regulation of TCF21 at independent CHD susceptibility alleles in humans. Individuals carrying the risk haplotype for rs12190287 and rs12524865 are predicted to have greater TCF21 expression due to increased binding of AP-1 complexes to a cis-regulatory element. Our studies further reveal a potential PDGFRβ-dependent mechanism for the CHD risk association in human coronary artery smooth muscle cells both in vitro and in vivo.
Of the 13 novel loci identified in the initial CARDIoGRAM meta-analysis, TCF21 was particularly attractive as a missing link to CHD. Tcf21 (Pod-1/Capsulin/Epicardin) was initially cloned in our laboratory and two others [4], [27], and is expressed in epicardial progenitor cells that give rise to developing CASMC [4]. Studies of TCF21 function in the adult have been hampered since Tcf21 null mice die postnatally due to pulmonary hypoplasia and respiratory failure [5]. TCF21 has been identified as a candidate tumor suppressor gene and is frequently epigenetically silenced in various human cancers [28], [29]. These studies have implicated loss of TCF21 expression as an early-stage biomarker for increased cancer risk. Based on our findings, we can reason that aberrant upregulation of TCF21 in coronary SMC may increase CHD risk through alteration of the SMC response to injury in the vessel wall.
Identification of cis-regulatory elements altered by disease related variants is critical for post-GWAS functional characterization studies [30]–[32]. Here, we observed that PDGF-BB mediates binding of an AP-1 complex, likely containing c-Jun, and JunD or ATF3 heterodimers to the risk allele at rs12190287, which is preferentially in an active chromatin state. The activator protein-1 (AP-1) family of TFs have been implicated in growth factor-dependent SMC activation following vessel injury [33]. The prototypical basic region-leucine zipper (bZip) protein, c-JUN is expressed in human atherosclerotic plaques and promotes SMC proliferation and neointima formation in vivo [34]. ATF3 is another stress-inducible gene upregulated in many cancers [35] and also in SMCs within injured mouse femoral arteries, to promote SMC migration and ECM synthesis [36]. It has been shown that c-Jun readily forms heterodimers with ATF2 and ATF3, which have distinct DNA binding affinities to CRE and AP-1 elements [37]. Interestingly, genes encoding extracellular matrix (ECM) proteins are often the targets of c-Jun/ATF enhancer elements [10].
A challenge in prioritization of regulatory SNPs is elucidating the biologically relevant upstream pathways driving these associations [38], [39]. Platelet-derived growth factor (PDGF) is a critical growth factor involved in vascular development. It has been shown in mice that PDGFR-β is required for development of mural cells, CASMC and pericytes, involving epithelial-to-mesenchymal transition (EMT) in the epicardium [13], [40]. PDGF-BB is also a potent inducer of the synthetic SMC phenotype, increasing migration, lipid uptake, and ECM synthesis, both in vitro and in vivo during vascular injury and atherosclerosis [12]. A recent GWAS for CHD in Europeans and South Asians identified the PDGFD gene, and this PDGF family member has been shown to have many of the same disease-related actions as related PDGFs [41]. In contrast, TGF-β is a pleiotropic cytokine mostly responsible for maintaining SMC differentiation, through activation of Smad, SRF, and RhoA signaling pathways [42]. Indeed, VSMC differentiation during aortic development likely depends on TGF-β rather than PDGF-BB/PDGFR-β [43]. Our observations that TCF21 was selectively induced by PDGF-BB rather than TGF-β in HCASMC is consistent with the notion that TCF21 inhibits coronary artery SMC differentiation while inducing SMC phenotypic modulation.
Similar to Tcf21, Wilms tumor 1 (Wt1) is expressed in the early proepicardium, epicardium and mesenchyme during development of the heart and other mesoderm-derived tissues [15], [44], [45]. In fact, previous studies in zebrafish have demonstrated that tcf21 expression in the proepicardial organ is dependent on wt1 [46]. Wt1 expression is also induced in the coronary vasculature in regions of ischemia and hypoxia after MI in mice [47]. As a tumor suppressor gene, WT1 was previously shown to repress PMA induced transcription [22]. WT1 binding to the thrombospondin-1 promoter leads to repression upon c-Jun overexpression in ECs [48] and fibroblasts [21]. Here we identify WT1 upstream of TCF21 to repress the enhancer at rs12190287. While in silico analyses predicted preferential binding to the G allele, haploChIP data suggest that WT1 preferentially associates with the C allele. This is consistent with greater c-Jun enrichment at the C allele. The orthogonal regulation of WT1 expression by PDGF-BB compared to TCF21, c-JUN, JUND, and ATF3, also may imply that WT1 acts to spatially and temporally fine-tune TCF21 expression, rather than cause repression.
CHD involving atherosclerosis continues to burden both developed and developing countries, largely due to urbanization and westernization of diet and lifestyle. Compared to developed countries, CHD related deaths are predicted to rise more than 3-fold in China and India, for instance [49]. While most GWAS for CHD have focused on individuals of European ancestry, large-scale studies of non-European populations may allow further understanding of the risk-associated mechanisms driving CHD. A recent meta-analysis of GWAS for CHD in a Han Chinese population (15,460 cases and 11,472 controls) replicated the TCF21 association in Europeans [3]. The combined discovery and replication stages identified a near genome-wide significant association at rs12524865, upstream of TCF21 at 6q23.2 (P = 1.87×10−7). The discovery stage identified primarily rs12524865 (P = 3.40×10−3), although rs12190287 showed a trend and directionality as a reporter for Caucasian cohorts (P = 3.03×10−2). The linkage disequilibrium r2 values between these two eQTLs for TCF21 is 0.62 in Europeans and only 0.18 in Han Chinese, and these variants are found in separate haplotype blocks in Han Chinese. Further, we demonstrated rs12524865 disrupts a binding site for CREB/ATF in silico and measured enrichment for c-Jun and ATF3 at rs12524865 in HCASMC. These studies highlight the value of multi-ethnic post-GWAS validation of causal variants to assess both the functional impact and heritable risk of common variants in complex diseases.
The compelling promise of the new disease associated genes and pathways afforded by GWAS methodology is that they will provide biological insights and targets for the development of new therapeutic approaches, and this is particularly compelling for CHD where there are no therapies directed at the blood vessel wall. TCF21 and its downstream targets provide one such pathway. eQTL data have suggested that the disease-associated major allele shows increased TCF21 expression, and this is consistent with the functional studies described herein, where the major risk C alleles at rs12190287 and rs12524865 confer greater transcriptional activity compared to the minor protective G and A alleles, respectively. The embryonic function of TCF21 in the developing coronary circulation seems most consistent with a role aimed at inhibiting differentiation of SMC progenitors, and thus it is likely that TCF21 function might interfere with the SMC response to vascular injury in the disease setting. It is now generally accepted that vascular SMC provide a stabilization of the atherosclerotic plaque, and thus therapeutic inhibition of the TCF21 pathway would be expected to decrease the risk for coronary events. However, such an approach might also put the patient at increased risk for head and neck and lung cancer, as TCF21 is a potent tumor suppressor gene that is frequently mutated or silenced in cells of these tumors. This situation contrasts with the emerging information related to the risk mechanisms at 9p21.3, where the function of one likely causal gene CDKN2B is associated with decreased risk for vascular disease [25], [50] as well as a broad range of human cancers [51]–[53]. Therapeutic activation of expression of this gene would be expected to decrease risk for both types of disease. While it is still early days for such extrapolations, follow-up of vascular wall GWAS genes is expected to provide insights into disease-related pathways to better inform therapeutic manipulation.
Primary human coronary artery smooth muscle cells (HCASMC) were purchased from three different manufacturers, Lonza, PromoCell and Cell Applications and were cultured in complete smooth muscle basal media (Lonza) according to the manufacturer's instructions. All experiments were performed on HCASMC between passages 4–7, using lots identified as heterozygous at rs12190287 as indicated. Primary rat aortic smooth muscle cells (RASM) were obtained from Dr. Phil Tsao (Stanford University) and cultured in Dulbecco's' Modified Eagle Media (DMEM) low glucose with 10% fetal bovine serum (FBS). The rat aortic smooth muscle cell line, A7r5 was purchased from ATCC and also obtained from Dr. Joe Miano (University of Rochester) and were maintained in DMEM low glucose with 10% FBS. HepG2 and HEK cells were purchased from ATCC and maintained in DMEM low glucose with 10% FBS.
Pre-designed Silencer Select siRNA duplexes against human c-JUN, JUND, ATF3, and WT1 were purchased from Ambion/Life Technologies. At least two individual siRNAs were tested for each. Briefly, HCASMC were plated in 12-well (dual-luciferase assay) or 6-well plates (qPCR) in complete media. Approximately 24 hours after plating, and between 40–60% confluence, cells were transiently transfected with negative control or TF specific siRNAs (50 nM) using RNAiMAX reagent (Life Technologies) according to the manufacturer's instructions. Cells were incubated for 48 hours prior to performing dual-luciferase assays, harvesting total RNA using miRNeasy Mini kit (Qiagen) for TaqMan based qPCR assays, or nuclear protein extraction for Western blotting.
Oligonucleotides containing the putative enhancer elements for rs12190287 C/G and rs12524865 C/A (Table S1) were annealed at 95 degrees for 10 minutes in annealing buffer and allowed to cool to room temperature. Double-stranded DNA fragments were then subcloned into the MCS of the minimal promoter containing pLuc-MCS vector (Agilent). Constructs were validated by Sanger sequencing. Empty vector (pLuc-MCS), rs12190287-C or rs12190287-G and Renilla luciferase constructs were transfected into HCASMC, RASMC, A7r5, HepG2, and HEK using Lipofectamine 2000. Media was changed after 6 hours, and dual luciferase activity was measured after 24 hours using a SpectraMax L luminometer (Molecular Devices). Relative luciferase activity (firefly/Renilla luciferase ratio) is expressed as the fold change of the empty vector control (pLuc-MCS).
Double stranded oligonucleotides for rs12190287-C/G, AP-1, CREB, rs12190287 mixed were obtained by annealing single stranded oligos (Table S1), as previously described [54]. Annealed oligos were then labeled with [γ32P]-ATP (Perkin Elmer) using T4 polynucleotide kinase (NEB) for 30 minutes at room temperature and then purified through Sephadex G-50 Quick Spin columns (Roche). After measuring radioactivity, reactions were assembled with 1× EMSA binding buffer, 1 µg poly-dIdC, 10 µg nuclear extract, 100× unlabeled probe (for competitions), 2 µg polyclonal antibody (for super-shifts), [γ32P]-ATP labeled probe, and incubated at room temperature for 30 min prior to protein separation on a 4% TBE gel. Gels were dried on Whatman paper using a heated vacuum drier and proteins were detected on radiographic film.
Primary human coronary artery smooth muscle cells (HCASMC) were cultured in normal growth media until approximately 75% confluent, then cultured in the absence of serum and supplements for 24 hours, prior to stimulation with 50 ng/ml human recombinant PDGF-BB (R&D Systems), 5 ng/ml human recombinant TGF-β1 (R&D Systems), 100 nM PMA (Sigma) or vehicle for indicated times. Total RNA was prepared using miRNeasy Mini kit (Qiagen) and total cDNA was prepared from 0.5 µg of RNA using the TaqMan High Capacity cDNA synthesis kit (Life Technologies). TaqMan gene expression probes (Table S1) were used to amplify human TCF21, c-JUN, JUND, ATF3, and WT1, which were normalized to human 18S levels.
Nuclear extracts were generated from HCASMC harvested at indicated time points. Protein concentrations were determined using a BCA assay (Pierce) and 50–100 µg nuclear protein for each condition was loaded on a pre-cast NuPAGE 4–12% Bis-Tris polyacrylamide gel (Invitrogen/Life Technologies), with gel run at 150 v for 1 h using MES buffer (Invitrogen/Life Technologies), and transferred to PVDF membrane at 35 v for 1 h. Membranes were blocked in 5% non-fat dry milk in 1× TBST for 1 h and incubated overnight with rabbit polyclonal antibodies against TCF21 (Sigma; 0.25 µg/ml), cJUN (Santa Cruz; 1.0 µg/ml), JUND (Santa Cruz; 1.0 µg/ml), ATF3 (Santa Cruz; 1.0 µg/ml), or WT1 (Santa Cruz; 1.0 µg/ml), followed by incubation in a secondary anti-rabbit HRP-conjugated antibody (Invitrogen/Life Technologies; 0.2 µg/ml) and detection by standard ECL (Pierce). Blots were reprobed with a mouse monoclonal antibody against GAPDH as a loading control.
Chromatin immunoprecipitation (ChIP) was performed according to the Millipore EZ-ChIP protocol with slight modifications. HCASMC were cultured as described above and treated with PDGF-BB or vehicle. Cells were fixed in 1% formaldehyde to cross-link chromatin, followed by quenching with glycine. 2×107 cells per condition were collected, and nuclear lysates were prepared as previously described [55]. Cross-linked chromatin nuclear extracts were sheared into ∼500 bp fragments using a Bioruptor (Diagenode) and clarified via centrifugation. 1×106 nuclei per condition were precleared with 20 ul Protein G Dynabeads (Invitrogen) for 1 hour, followed by incubation with 2 ug Rabbit IgG or anti-c-Jun, JunB, JunD, ATF3, WT1 (Santa Cruz or Active Motif), H3K4Me1, H3K4Me3, H3K27Ac, H3K27Me1 (Diagenode or Abcam) overnight at 4C. Immunoprecipitated chromatin samples were incubated with 20 µl Protein G Dynabeads for 1 hour at 4C to capture the protein-DNA complexes. Complexes were washed and eluted as described. Protein-DNA cross-links were reversed, treated with RNase A and proteinase K and free DNA was purified using Qiagen PCR purification kits. Total enrichment was measured using rs12190287 or rs12524865 specific primers, or a known AP-1 regulatory region, or a negative control region using the primers listed (Table S1). Semi-quantitative PCR was used to verify ChIP products via gel electrophoresis. Quantitative real-time PCR (ViiA 7, Life Technologies) was performed using SYBR Green (Applied Biosystems) assays and fold enrichment was calculated by measuring the delta Ct – delta Ct IgG. Melting curve analysis was also performed for each ChIP primer pair.
1×107 HCASMCs per condition were processed as previously described [56], using anti-H3K4me1 (pAb-037-050, Diagenode), anti-H3K4me3 (pAb-003-050, Diagenode), anti-H3K27me3 (pAb-069-050, Diagenode), or anti-rabbit IgG (X 0903, DAKO). ChIP-seq library generation, cluster formation and next-generation sequencing was performed at the Stanford Functional Genomics Facility, Stanford University, Stanford CA, USA, on an Illumina MiSeq instrument. 36 bp single reads from next-generation sequencing of ChIP libraries were then mapped to the reference genome using Burrows-Wheeler Aligner (BWA). BigWig files were created using the R/Bioconductor environment.
1×107 HCASMCs were mainly processed similar to the ChIP-seq samples. However, instead of preclearing and immunoprecipitation, protein-depleted DNA was extracted from cross-linked nuclear lysates by phenol-chloroform extraction. After DNA precipitation, purification and reverse cross-linking, samples were sequenced and further processed as described above.
Genomic DNA was isolated from >106 HCASMC cultured in complete media for approximately 48 hours, using the Blood and Tissue DNA isolation Kit (Qiagen). 50 ng of gDNA template was amplified using primers flanking rs12190287 to generate 250 bp fragments. Fragments were then sequenced via Sanger sequencing using an internal forward sequencing primer, and genotypes were determined from chromatograms using Sequence Analyzer (Applied Biosystems).
Heterozygous genotypes were determined by Sanger sequencing, and RNA and cDNA prepared as described above. Allele-specific expression of TCF21 at rs12190287 was determined using a pre-designed TaqMan SNP genotyping assay for rs12190287 (Table S1). Calibration of the SNP genotyping assay was determined by mixing 10 ng of HCASMC gDNA or cDNA, homozygous for each allele at the following ratios: 8∶1, 4∶1, 2∶1 1∶1, 1∶2, 1∶4, 1∶8. The Log2 ratio of the VIC/FAM intensity at cycle 40 was then plotted against the Log ratio of the two alleles to generate a linear regression standard curve. The Log ratio of the intensity of the two alleles from cDNA samples was fitted to the standard curve. These values were then normalized to the ratio of gDNA for each allele to obtain the normalized allelic ratio.
Heterozygous genotypes were determined as described above. Briefly, heterozygous HCASMC were cultured for the indicated timepoints, and chromatin cross-linked, sheared and immunoprecipitated as described above. Purified DNA was then amplified using TaqMan SNP genotyping assay probes for rs12190287. The Log2 ratio of VIC/FAM intensity at cycle 40 was then fitted to the standard curve and normalized to gDNA ratio, with the normalized allelic ratio of IgG control enrichment arbitrarily set to 1.
Pyrosequencing assay for rs12190287 was generated using PyroMark Assay Design software (Qiagen). Forward rs12190287 PCR primer: 5′- and biotinylated reverse PCR primer, and forward pyrosequencing primers were synthesized by the Protein And Nucleic acid (PAN) facility (Stanford). Approximately 20 ng ChIP DNA was amplified using forward and reverse pyrosequencing primers under the following conditions: 94 C 4 min, (94 C 30 s, 60 C 30 s, 72 C 45 s) ×45, 72 C 6 min. Pyrosequencing reaction was performed on a PyroMark Q24 according to manufacturer's instructions. Allelic quantitation was obtained automatically from the mean allele frequencies derived from the peak heights using PyroMark Q24 software.
Transcription factor binding site (TFBS) prediction was determined using the following online bioinformatics tools: TRANSFAC (BIOBASE), PROMO, MatInspector, JASPAR, and TFSearch. Sequences from dbSNP for each allele were scanned for TFBS in vertebrates meeting a minimum similarity score of 0.85. Regional association plot was generated from CARDIoGRAM meta-analysis dataset at TCF21 using LocusZoom [57]. Linkage disequilibrium plots and haplotype frequencies were generated from Europeans in the ADVANCE cohort (Stanford) from the CARDIoGRAM consortium, and East Asians from the HALST cohort within the TAICHI consortium. Briefly, genotyping data was extracted for each region of interest using PLINK [58] and transposed files were imported into Haploview [59].
Experiments were performed using at least three independent preparations with individual treatments/conditions performed in triplicate. Data is presented as mean ± standard deviation (SD) of replicates. GraphPad Prism 6.0 was used for statistical analysis. Comparisons between two groups were performed using paired two-tailed t-test. P values <0.05 were considered statistically significant. For multiple comparison testing, two-way analysis of variance (ANOVA) accompanied by Tukey's post-hoc test were used as appropriate.
All samples reported in this study were obtained under written informed consent for participation in the Atherosclerotic Disease, VAscular functioN, and genetiC Epidemiology (ADVANCE) and Healthy Aging Longitudinal Study in Taiwan (HALST) studies with the approval of the Institutional Review Boards of Stanford University and National Health Research Institutes, respectively.
|
10.1371/journal.pcbi.1000155 | The Effect of a ΔK280 Mutation on the Unfolded State of a Microtubule-Binding Repeat in Tau | Tau is a natively unfolded protein that forms intracellular aggregates in the brains of patients with Alzheimer's disease. To decipher the mechanism underlying the formation of tau aggregates, we developed a novel approach for constructing models of natively unfolded proteins. The method, energy-minima mapping and weighting (EMW), samples local energy minima of subsequences within a natively unfolded protein and then constructs ensembles from these energetically favorable conformations that are consistent with a given set of experimental data. A unique feature of the method is that it does not strive to generate a single ensemble that represents the unfolded state. Instead we construct a number of candidate ensembles, each of which agrees with a given set of experimental constraints, and focus our analysis on local structural features that are present in all of the independently generated ensembles. Using EMW we generated ensembles that are consistent with chemical shift measurements obtained on tau constructs. Thirty models were constructed for the second microtubule binding repeat (MTBR2) in wild-type (WT) tau and a ΔK280 mutant, which is found in some forms of frontotemporal dementia. By focusing on structural features that are preserved across all ensembles, we find that the aggregation-initiating sequence, PHF6*, prefers an extended conformation in both the WT and ΔK280 sequences. In addition, we find that residue K280 can adopt a loop/turn conformation in WT MTBR2 and that deletion of this residue, which can adopt nonextended states, leads to an increase in locally extended conformations near the C-terminus of PHF6*. As an increased preference for extended states near the C-terminus of PHF6* may facilitate the propagation of β-structure downstream from PHF6*, these results explain how a deletion at position 280 can promote the formation of tau aggregates.
| Alzheimer's disease pathology is characterized by two types of protein aggregates that are found in the brain—amyloid plaques and neurofibrillary tangles. Several studies suggest that these aggregates also play an active role in the disease process. Thus, an understanding of disease pathogenesis may be facilitated by a detailed characterization of the proteins that comprise these aggregates. Our study aims to model structural characteristics of tau protein, which is found in neurofibrillary tangles. Modeling of tau is particularly difficult because the protein is intrinsically disordered and therefore must be modeled as an ensemble of structurally dissimilar states. We developed a novel modeling approach that incorporates experimental measurements to generate ensembles of conformations that model the unfolded state of tau. By analyzing structural properties in these model ensembles for both normal and disease-associated forms of the protein, we identify structural features that may facilitate tau aggregation.
| Alzheimer's disease (AD) pathology is characterized by extracellular aggregates of Aβ-amyloid (Aβ) and intraneuronal tau aggregates, known as senile plaques and neurofibrillary tangles (NFTs), respectively [1]. Despite much focus on Aβ amyloid in AD research, tau seems to play an important role as well. For example, the number of NFTs and not the number of senile plaques in the neocortex correlates with the severity of dementia in AD patients, and there are data that imply that abnormalities in tau alone may cause neurodegeneration [2]. In light of these observations, a detailed characterization of the structure of tau protein may provide insights into the pathogenesis of AD and other neurodegenerative disorders associated with tau pathology. However, probing the structure of tau is difficult because tau protein is natively unfolded (or intrinsically disordered) in solution. Several studies suggest that tau retains its function after heat or acid-induced denaturation and both CD and X-ray scattering experiments imply that tau does not adopt a well-defined folded structure in solution 3–5. Consequently, obtaining structural and hence functional information on tau is problematic because the direct observation of unfolded states is typically difficult to achieve experimentally.
Initially, unfolded proteins were described as random coils whose properties are derived from Flory's statistical description of chain molecules [6]. For such polymers, the radius of gyration, RG, follows the scaling law RG = R0Nν, where R0 is the radius of gyration of a monomeric subunit (a function of the persistence length), N is the number of subunits in the polymer, and ν is a scaling factor that depends on the solvent characteristics. The most common measure of whether a protein behaves as a random coil is to test whether its radius of gyration follows this scaling law. However, while a structurally disordered molecule can exhibit random coil statistics, the converse is not necessarily true; i.e., random coil statistics do not imply that the structure is completely disordered [7]. Slight structural preferences may exist for some natively unfolded proteins and small changes in the distribution of conformers within an unfolded ensemble may play a role in the normal and pathological functioning of intrinsically disordered systems. A recent study, for example, suggests that inducer-mediated tau polymerization involves an allosterically regulated conformational change [8]. This is consistent with the notion that the formation of tau fibrils is associated with a shift in the conformational distribution of tau such that the unfolded state has a preference for proaggregatory conformations in the presence of an inducer. In light of this, constructing detailed ensembles that model the unfolded ensemble of tau may facilitate the identification of structural properties that promote aggregation.
As full-length tau contains more than 400 amino acids (441 residues for the htau40 isoform [9]) constructing detailed ensembles that model the unfolded state of this protein is a daunting task. Fortunately, tau contains three or four imperfect microtubule-binding repeats (MTBRs) near the C-terminus of the protein, and almost all known mutations of tau that are associated with inherited forms of neurodegenerative diseases are located in MTBR domains or their nearby flanking regions [10]. As these data suggest that MTBRs play an important role in the progression of inherited tauopathies, we first focus on building ensembles that model the structure of individual MTBRs. It is important to note, however, that we do not strive to model the structure of a given MTBR fragment alone in solution. Rather, our goal is to generate ensembles that model the range of conformations that a MTBR can adopt when it is part of full length tau. In the present study we focus on building ensembles for the second MTBR, henceforth referred to as MTBR2. This repeat is of particular interest because it contains both a six amino-acid repeat, PHF6*, which is a minimum interaction motif that can initiate tau aggregation in vitro [11],[12], and the site of the proaggregatory mutation, ΔK280, which is associated with some forms of frontotemporal dementia [13]–[16].
We have developed a method, called energy-minima mapping and weighting (EMW), to construct ensembles that model the unfolded state of proteins. The underlying assumption that forms the basis of this approach is that the unfolded state can be modeled as a set of energetically favorable conformers, where each conformer corresponds to a local energy minimum. The method involves constructing a library of energetically favorable conformations and selecting conformations from this library to form ensembles that are consistent with a given set of experimental data. We use EMW to build ensembles for wild-type (WT) MTBR2 and the corresponding ΔK280 mutant. By comparing data from the two sets of ensembles, we deduce structural preferences in the ΔK280 ensemble that explain its increased propensity to form tau aggregates.
The EMW method begins by constructing sets of energetically favorable conformations for a sequence of amino-acids within a natively unfolded protein (Figure 1). In the case of tau we focus on MTBR2 since this region contains the aggregation-initiating sequence PHF6* as well as the site of a mutation that is associated with increased tau aggregation in vitro [17]. A set of local energy minima is then constructed for this subsequence, hence forming the candidate ensemble (Figure 1). Associated with each structure in this ensemble is a weight, ωi, which corresponds to the probability that the given subsequence adopts the ith conformation in the candidate ensemble. We say that an ensemble is fully specified when the local energy minima and weights are known.
Initial weights for structures in the candidate ensemble are calculated from the relative energies of each structure, as shown in Figure 1. However, as sampling is performed on a relatively small subsequence these weights may not reflect the relative probabilities of different conformations when the subsequence is part of the larger protein. For example, compact states may be preferred over extended states when the subsequence is in isolation but not when part of tau. Therefore, the composition of the ensemble is optimized and the members of the candidate ensemble are reweighted in light of experimental data that is obtained on a larger segment of tau protein. Sampling small subsequences increases the chance that we will observe a relatively large number of accessible states for this system. Using experimental data obtained on a larger region of tau (and not just the subsequence of interest) helps to ensure that the calculated ensemble represents the local structure of the sequence as it appears within full length tau.
A central component of EMW is that we do not strive to construct a single model for the unfolded state. We recognize that the construction of unfolded ensembles that agree with any given set of experimental data is largely an underdetermined problem; hence it is likely that there are a number of different ensembles that are consistent with a given set of experimental data. Consequently, we constructed several ensembles that are all consistent with the experimental measurements and focused our analysis on local structural motifs that are present in all ensembles. For this study, we focused on NMR data that are available for both WT MTBR2 and a ΔK280 mutant. These data were kindly provided by Marco Mukrasch, Daniela Fischer, and Markus Zweckstetter [17],[18].
Using the EMW method, 100 ensembles were constructed for both wild-type (WT) and ΔK280 sequences of MTBR2 (a total of 200 ensembles). Each ensemble was constructed to minimize the difference between calculated 13Cα chemical shifts and the corresponding experimentally determined 13Cα chemical shifts. The number of structures in each ensemble corresponds to the minimal number of structures needed to fit the available chemical shifts. Preliminary calculations found that 15 conformers were needed; i.e., fewer structures resulted in worse fits to the 13Cα chemical shifts and more structures did not significantly improve the quality of fits. We note that other models examining residual structure in the unfolded state have utilized a similar number of representative conformers [19].
Application of EMW yielded ensembles that were in excellent agreement with experimentally determined absolute 13Cα chemical shifts (Figure 2A and 2B). The average RMS error between the calculated 13Cα chemical shifts and the corresponding experimental values was 0.1 ppm—well below the error associated with SHIFTX chemical shift predictions and similar to the error associated with experimental chemical shift measurements on K18 constructs [17],[20]. However, given that measured absolute chemical shifts for the 20 amino acids vary significantly according to the amino-acid type, reasonable correlations to absolute chemical shifts may be achieved by simply predicting amino-acid specific random coil values. Given this, we analyzed the relationship between the chemical shifts, after subtracting out residue-specific random coil chemical shift values; i.e., the secondary chemical shifts. Overall, there is excellent agreement between calculated secondary chemical shifts and the corresponding experimental values for each residue in the sequence (Figure 2C and 2D). These data demonstrate that the calculated models yield agreement with experiment on a per residue basis.
In the next step of our protocol, carbonyl carbon (13CO) chemical shifts were used to test whether the resulting ensembles can predict experimental observations that were not used to construct the model. This helps to ensure that our models are not “overly fit” to the 13Cα chemical shifts. In general, a model that is over-fit to a given set of experimental data can reproduce that data remarkably well but cannot reproduce data that was not used to generate the model. Therefore we consider an ensemble to be validated if new experimental results can be accurately predicted from the ensemble. For both the WT and ΔK280 sequences, each of the 100 ensembles was ranked based on its ability to predict 13CO chemical shifts. Based on these data the thirty best ensembles were chosen for further analysis. The RMS difference between the calculated 13CO chemical shifts and the corresponding experimental values are below 0.9 ppm; i.e., below the error associated with available chemical shift prediction algorithms (Table 1) [20]. To further demonstrate that these thirty ensembles can reproduce additional data not used in the model constructed, we computed the error between calculated amide hydrogen (1HN) chemical shifts and the corresponding experimental values. The resulting values agreed with the experimentally measured ones to within 0.3 ppm (Table 1).
As expected, structures that comprise the WT (Figure 3A) and ΔK280 (Figure 3B) ensembles are heterogeneous in that they sample a wide range of conformations. Since each of the 30 ensembles represents an independent representation of the unfolded state, we searched for local structural motifs that are found in all of the ensembles. More precisely, the existence of a local conformation that is consistently adopted by a given subsequence in MTBR2 suggests that this conformation is needed to reproduce the experimental results. We therefore consider conserved motifs to represent local conformational preferences.
We begin with an assessment of the local conformation of PHF6* in both the WT and ΔK280 ensembles. Since PHF6* in the WT sequence spans residues 275–280, the ΔK280 mutant sequence has a deletion in the six-residue stretch corresponding to PHF6*. However, since residue 281 is also a lysine, the ΔK280 mutant contains an equivalent PHF6* subsequence at its N-terminus (Figure 4). This allows us to directly compare the conformation of PHF6* in both sequences. To identify preserved conformations of PHF6*, we first determined the different types of structures that this subsequence can adopt by clustering structures using only the backbone atoms of PHF6* (Figure 5). The probability that a given cluster occurs in an ensemble is equal to the sum of the weights of structures in that ensemble that contains a motif in the cluster. Preserved structural motifs are defined as clusters that have a nonzero weight in every ensemble (Figure 5); i.e., a preserved motif is found in all ensembles. For comparison, we repeated this procedure for all contiguous six-residue subsequences within MTBR2, yielding a collection of approximately 300 clusters that represent all possible structural motifs in our ensembles that any six-residue sequence in MTBR2 can adopt. Using the criterion outlined above, roughly 5% of these clusters were preserved across all ensembles.
In WT MTBR2, clustering based on the conformation of PHF6* yielded 12 distinct conformations. However, only one of these states was present in all 30 ensembles (Figure 6A and 6B). Similarly, while PHF6* clusters into 11 distinct conformations in the mutant ΔK280 ensembles, only one conformation was preserved (Figure 6C and 6D). In both cases, the preserved conformation of PHF6* is extended and has φ, ψ angles that fall within the broad region of the Ramachandran plot corresponding to β-structure. This observation is consistent with the notion that PHF6* a priori adopts extended conformations that can readily form cross β-structure with other tau monomers [21]. Since the formation of cross β-structure is believed to play an essential role in the formation of protein aggregates, these data are consistent with the notion that PHF6* promotes aggregation by forming β-structure between tau monomers [11],[12].
To explore the effect of the ΔK280 mutation on the local structure of MTBR2, we analyzed the structure of the subsequences 278INKKLD283 and 278IN-KLDL284 in the WT and ΔK280 sequences, respectively. For WT MTBR2, two conformations for 278INKKLD283 were found in all ensembles. The first is a loop/turn that is associated with a change in the direction of the mainchain (Figure 7A and 7B). In this structure residue K280 has φ, ψ angles of approximately −102° and −30°, respectively; i.e., mainchain dihedral angles consistent with an α-helical/turn conformation. The second conformation is more extended, having φ, ψ angles that place its residues within the broad region corresponding to extended β-structure (Figure 7C and 7D). In the mutant sequence, residue K280 is absent and the corresponding sequence, 278IN-KLDL283, has one preserved conformation. The deletion of residue 280, which can adopt an α-helical/turn conformation in the native sequence, leads to a relative increase in results in extended states in this region (Figure 7E and 7F). The deletion, however, also introduces a slight kink in the mainchain of the sequence (Figure 7F).
In a prior work, N–H residual dipolar coupling (RDC) values were measured for residues in the WT K18 construct in polyacrylamide gel [22]. While most residues in MTBR2 have relatively large negative RDC values, S285 has a large positive value [22]. This difference can be explained by either a change in the local alignment tensor at S285, or the presence of α-helical/turn structure at this site [23]–[26]. Accelerated molecular dynamics simulations of WT K18, however, confirm that the sequence 283DLSN286 samples turn conformations with relatively high frequency [22]. In light of these observations, we explored the structure of the six residue segment, 282LDLSNV287, which includes residue S285. This region adopts two conformations that are preserved across all WT ensembles. One of the conformations contains a loop/turn (Figure 8A and 8B) where residue S285 has φ, ψ angles of −63° and −39°, respectively; i.e., near the optimal α-helical values (Figure 8B). The alternate conformation is extended and does not result in a change in the direction of the mainchain (Figure 8C and 8D). However, in the ΔK280 mutant, 282LDLSNV287 has one structure that is preserved across all ensembles (Figure 8E and 8F). In this structure S285 again adopts φ, ψ angles (−95° and −63°, respectively) that are consistent with an α-helical/turn conformation (Figure 8F). These data agree with the RDC data mentioned above and suggest that this region in both the WT and mutant sequences is able to adopt turn-like conformations in solution as well as in a polyacrylamide gel.
Dynamical simulations provide a valuable tool for the analysis of unfolded proteins, providing insights that would be difficult to obtain from experiments alone [27]. A number of simulation methods have been developed to model the unfolded states of proteins and useful insights have been obtained with these techniques. Many of these approaches generate ensembles by directly incorporating experimental constraints into molecular dynamics simulations in order to facilitate conformational sampling. These methods bias molecular trajectories to sample conformers that are consistent with a given set of experimental data. One problematic issue with biased sampling, however, is that it can suffer from over-fitting—a process that may yield a distribution of conformers that does not accurately model the range of structures that comprise the unfolded state [27]. Given this concern, a number of unbiased methods have been developed to generate ensembles for unfolded proteins. These approaches utilize fast algorithms, which do not employ a physical potential energy function, to obtain representative structures of the unfolded state, and in some cases experimental data can then be used to improve the resulting ensembles [28]–[31]. The algorithm ENSEMBLE, for example, adjusts population weights for pregenerated conformers to improve agreement with experimental data in a manner similar to that described here [30].
A unique feature of the present method is that it does not strive to generate a single ensemble that represents the unfolded state. Given that accurate modeling of an unfolded protein is an undetermined problem, it is likely that there are a number of different ensembles that agree with any given set of experimental data. Moreover, given the immense number of potential conformations that an unfolded protein can adopt, this may be true even when a relatively large number of experimental constraints are used to construct the ensemble. Hence our goal was to construct several candidate ensembles, each of which agrees with a given set of experimental constraints, and focus our analysis on local structural features that are preserved across all ensembles. Local structural features that are found in all independent ensembles likely represent motifs that are required to reproduce the experimental data. In other words, given the underdetermined nature of the problem, it is not clear how to determine when one has the “correct” ensemble. However, local structural motifs that consistently appear in all independent ensembles are likely to also be present in the “correct” ensemble. Consequently, we consider locally preserved structural motifs to represent local conformational preferences.
An important consideration in our method is the choice of experimental data that is used to build and validate the constructed ensembles. In principle, EMW can use any set of experimental measurements to optimize and validate model ensembles. Indeed, as more structural information is made available, additional data can and should be used to further refine the set of model ensembles. In this regard, we note that although a number of NMR measurements have been made on native tau constructs, the data available for constructs containing a ΔK280 mutation is relatively limited. In a prior study, nuclear chemical shifts and HSQC spectra were measured for the K18ΔK280 construct, which contains all four MTBRs and the ΔK280 mutation [17]. Data were obtained for both free K18ΔK280 and for K18ΔK280 in the presence of the polyanion heparin and microtubules [17]. However, as we are interested in building structural models for MTBR2 in solutions free of compounds that promote tau self-association (e.g., heparin) and free of proteins known to bind tau, we focused on measurements obtained with the free K18ΔK280 construct. Additionally, as there are a number of existing methods that relate chemical shift measurements to three dimensional protein structures [20], [32]–[34] we considered 13Cα, 13CO, 1HN, and 15N chemical shift measurements; i.e., the only available chemical shifts for K18ΔK280 [17]. Furthermore, established methods for estimating NMR chemical shifts can predict carbon and amide proton chemical shifts with an error of approximately 1 ppm or less, while the error associated with predicting nitrogen chemical shifts is substantially larger (∼2–2.5 ppm) [20], [33]–[35]. Therefore we focused on the 13Cα, 13CO, and 1H chemical shifts for this study because these data represent measurements that can be calculated with the greatest accuracy and that are available for both native tau constructs and the ΔK280 mutant.
It has long been recognized that chemical shifts of a given residue are, in general, largely a function of the local environment of the residue in question [36],[37]. Since we generate ensembles that agree with chemical shifts, a limitation of the results reported here is that we do not explicitly include experimental data that more directly reveal information about non-local interactions. While long range contacts have been identified in some natively unfolded proteins (e.g., [19]), the dimensional scaling characteristics of intrinsically disordered proteins suggests that stable long-range contacts are sparse in these systems [38]. Nevertheless, we suggest that the combination of a physical potential energy function, which can in principle model long range interactions, and experimentally determined chemical shifts can provide insight into the structure of proteins in general. In this regard we note that data are emerging that suggest that backbone chemical shifts, when used in conjunction with a physical energy function, may be sufficient to adequately predict tertiary folds, and consequently stable non-local contacts, for some proteins [39],[40].
Although our work focuses on the structure of the MTBR2 without explicitly including other MTBRs, our findings may also have implications for full length tau. Once a representative set of conformers for MTBR2 is generated, we strive to ensure that the calculated chemical shifts agree with chemical shifts obtained using a construct that contains all MTBRs. This helps to guarantee that the ensemble models the structure of MTBR2 as it appears in full length tau. In short, we are not interested in the structure of MTR2 as it appears alone in solution; instead we hope to deduce structural features of MTBR2 as it appears in full length tau. In addition, as MTBR2 contains an aggregation-initiating sequence that is known promote tau aggregation in vitro as well as the site of a mutation that leads to in increased tau aggregation in vitro and in vivo, studies of both its WT and mutant forms may lead to insights into the mechanism of tau aggregation [12],[15],[41].
The ability to form intermolecular β-sheet conformations appears to be a relatively general property of polypeptide chains that are associated with disorders of protein misfolding and aggregation [42]–[45]. Therefore it is likely that an inherent propensity to form extended conformations, that are consistent with β-structure, will promote aggregation in natively unfolded systems. When EMW is applied to MTBR2, we find that the aggregation-initiating sequence, PHF6*, adopts an extended conformation in both the WT and ΔK280 ensembles, a finding consistent with the observation that these peptides can initiate tau aggregation [11],[12]. Interestingly, in a prior work we demonstrated that a related hexapeptide, PHF6, preferentially adopts an extended state that can facilitate the formation of cross-β-structure between tau monomers [21]. The present study suggests that this property is preserved when aggregation-initiating sequences are part of their corresponding MTBRs. That is, PHF6* a priori adopts extended conformations that can readily form hydrogen-bonded β-structure. Additionally, a recent survey of amyloidogenic proteins suggests that fibrillogenesis for natively unfolded proteins involve the formation of partially folded intermediates that can subsequently go on to form amyloid fibrils [45]. Our findings are consistent with these observations. That is, our results imply that formation of a locally stable, and extended, conformation plays a role in the formation of tau aggregates.
Recently, several studies have attempted to characterize residual structure of MTBRs in tau [17], [18], [22], [46]–[48]. These studies can be roughly divided into two categories: descriptions of ensemble average characteristics based on NMR measurements [17],[18],[22],[46], and NMR solution structures of local regions obtained by adding organic solvents to stabilize a unique fold [47],[48]. Since the presence of organic solvents leads to significant changes in the conformational distribution of states, as evidenced by dramatic changes in the CD spectra [5],[47],[48], the physiologic relevance of these latter results remains unclear. However, early characterizations of MTBRs in nonorganic solvents, found that the PHF6 region likely has a higher propensity for extended, β-strand-like conformations—a finding in accord with our data [18],[46].
Given that both WT and ΔK280 tau contain aggregation-initiating sequences (Figure 4), it is not clear how β-strand propensity in this region explains the difference in aggregation potential between the two sequences. Therefore to deduce structural features of the ΔK280 mutant that explain its proclivity to form aggregates, we analyzed the structure of MTBR2 in the vicinity of the mutation site. Unfolded ensembles of WT MTBR2 contain two conformations at the mutation site that were present in all ensembles—a loop/turn conformation and an extended state. In contrast to the WT MTBR2 ensembles, models of ΔK280 in the same region had one conformation that was present in all ensembles. This state is relatively extended and contains a kink at the site of the deletion. While the slight disruption in the extended state of the mutant may also influence the ability to form hydrogen-bonded cross-β-structure, a loop/turn at the C-terminus of PHF6* constitutes a much greater impediment to the formation of β-structure. Since residue K280 has a relative preference for nonextended states, deletion of this residue leads to increased sampling of extended states downstream from PHF6*. The relative preference for extended structures downstream from PHF6* in the ΔK280 mutant suggests that the ability to propagate β-structure distal to PHF6* can affect the aggregation potential of tau. These observations therefore explain how the deletion of a single residue can change the aggregation potential of tau.
We also find that in both WT and mutant ensembles residue S285 can adopt φ, ψ angles consistent with an α-helical/turn structure. Recent data on the WT sequence are also consistent with these observations as RDC values and molecular dynamics simulations suggest that S285 adopts an α-helical/turn structure. Since those experiments were performed in polyacrylamide gel, our data suggest that this structure also occurs with relatively high frequency in solution. It is also worthwhile to note that although we find that a six-residue region including K280 can adopt a similar loop/turn conformation, the associated RDCs for this region are not associated with a change in sign, like that observed at S285 [27]. Nonetheless, unlike RDC measurements for folded proteins, RDC values for unfolded proteins can be difficult to interpret [49]. This is due, in part, to the fact that prior to the measurement of RDC values, the protein of interest must first be embedded in an alignment medium [26]. This induced steric alignment of unfolded proteins may lead to results that do not fully capture the range of structures that an unfolded protein can adopt in solution. Hence the absence of particular RDC values in polyacrylamide gel (or any other alignment media) does not necessarily imply that a given conformation is not present in solutions containing the unfolded protein of interest.
The formation of tau aggregates is likely a complex process as a number of factors have been shown to influence the formation of tau aggregates in vitro [1]–[3]. Consequently, there may be additional factors that contribute to the increased ability of the ΔK280 mutant to form aggregates; e.g., a ΔK280 mutation leads to an overall decrease in the strength of the intermolecular charge-charge repulsion between tau monomers that self-associate [12]. Nonetheless, our data demonstrate that small changes in the sequence of tau can lead to localized structural changes in the unfolded ensemble that may affect tau's ability to form cross-β-structure. Overall, our data suggest that small sequence-specific changes can promote tau aggregation and that interventions that prevent the propagation of β-structure downstream from aggregation-initiating sequences, may form the basis for therapies that prevent tau aggregation.
The EMW method constructs ensembles for unfolded proteins that are consistent with a given set of experimental data. Our model for an unfolded ensemble consists of structures corresponding to local energy minima and associated probabilities (weights) that are assigned to the different conformations. For this work, the experimental measurement used to optimize and validate the model ensembles are chemical shifts for the second tau microtubule binding repeat [17]. In principle, EMW can be used with any given set of experimental data. In this application we focus on chemical shifts that were available for both the K18 and K18ΔK280 constructs.
The EMW method can be decomposed into three steps (i) conformational sampling, (ii) model optimization, and (iii) ensemble validation. Conformational sampling uses high temperature molecular dynamics (MD) followed by minimization of the resulting structures (i.e., quenched dynamics) to create a library of widely varying conformations representing minima on the potential energy surface. Model optimization is performed to select a subset of these structures and optimize weights that represent the relative prevalence of each structure. Validation is performed by computing additional chemical shifts that not used to construct the ensemble and comparing these data to experimentally measured carbonyl carbon shifts. In what follows we outline each step of the EMW method.
We searched for conformations of six-residue subsequences that are present in every ensemble. Six residues was a natural characteristic size for a local region of interest, as it is the length of PHF6*. To this end, all structures in each ensemble of either WT or ΔK280 MTBR2 were clustered using a matrix consisting of the pairwise RMSD backbone deviation of the each contiguous six-residue segment. Structures were clustered using MATLAB (Mathworks) such that the maximum RMSD between two structures in a cluster was 2.5 Å. A range of maximum RMSD values (1–6 Å) were examined empirically, and it was found that a cutoff of 2.5 Å was sufficient to prevent similar conformations from being divided into separate clusters, while also ensuring that clusters included a relatively homogeneous set of conformations. The probability that a given cluster occurs in an ensemble is equal to the sum of the weights of all structures that contain that motif. Preserved local structural motifs were found by identifying clusters where the total weight of its structures was non-zero across all ensembles.
Structures for each cluster were visualized in VMD. To facilitate visualization of the overall conformation associated with a cluster, an average structure for each cluster was generated after 5,000 steps of steepest descent minimization to remove bad contacts (only the 6 residues were minimized). Visual inspection verified that the energy minimized structures did not differ significantly from their un-minimized counterparts. All molecular structures were made with VMD [59].
|
10.1371/journal.pntd.0003280 | Post-translational Modification of LipL32 during Leptospira interrogans Infection | Leptospirosis, a re-emerging disease of global importance caused by pathogenic Leptospira spp., is considered the world's most widespread zoonotic disease. Rats serve as asymptomatic carriers of pathogenic Leptospira and are critical for disease spread. In such reservoir hosts, leptospires colonize the kidney, are shed in the urine, persist in fresh water and gain access to a new mammalian host through breaches in the skin.
Previous studies have provided evidence for post-translational modification (PTM) of leptospiral proteins. In the current study, we used proteomic analyses to determine the presence of PTMs on the highly abundant leptospiral protein, LipL32, from rat urine-isolated L. interrogans serovar Copenhageni compared to in vitro-grown organisms. We observed either acetylation or tri-methylation of lysine residues within multiple LipL32 peptides, including peptides corresponding to regions of LipL32 previously identified as epitopes. Intriguingly, the PTMs were unique to the LipL32 peptides originating from in vivo relative to in vitro grown leptospires. The identity of each modified lysine residue was confirmed by fragmentation pattern analysis of the peptide mass spectra. A synthetic peptide containing an identified tri-methylated lysine, which corresponds to a previously identified LipL32 epitope, demonstrated significantly reduced immunoreactivity with serum collected from leptospirosis patients compared to the peptide version lacking the tri-methylation. Further, a subset of the identified PTMs are in close proximity to the established calcium-binding and putative collagen-binding sites that have been identified within LipL32.
The exclusive detection of PTMs on lysine residues within LipL32 from in vivo-isolated L. interrogans implies that infection-generated modification of leptospiral proteins may have a biologically relevant function during the course of infection. Although definitive determination of the role of these PTMs must await further investigations, the reduced immune recognition of a modified LipL32 epitope suggests the intriguing possibility that LipL32 modification represents a novel mechanism of immune evasion within Leptospira.
| Leptospirosis, caused by pathogenic Leptospira spp., constitutes an increasing global public health threat. Humans are accidental hosts, and acquire the disease primarily from contact with water sources that have been contaminated with urine from infected animals. Rats are asymptomatic carriers of infection and are critical for disease transmission to humans, particularly in urban slum environments. In this study, investigation of Leptospira directly isolated from the urine of infected rats showed acetylation or tri-methylation of the highly abundant leptospiral lipoprotein, LipL32. In comparison, Leptospira grown in culture did not result in any LipL32 lysine modifications. A synthetic peptide derived from LipL32 that incorporated a tri-methylated lysine modification exhibited less reactivity with serum from leptospirosis patients compared to an unmodified version of the peptide, suggesting LipL32 modifications may alter protein recognition by the immune response. This study reports, for the first time, modification of a Leptospira protein during infection, and suggests these modifications may have a functional consequence that contributes to bacterial persistence during infection.
| Pathogenic Leptospira spp. are the causative agents of leptospirosis, which is considered to be the world's most widespread zoonotic disease [1]–[4]. Recent data shows the incidence of leptospirosis is increasing, with outbreaks frequently occurring within urban slum settings [5]–[8], an environment in which approximately 31.6% of the world's total urban population resides [9]. Therefore, leptospirosis represents a significant public health threat in these communities.
Mammalian hosts that are chronically infected with Leptospira contain the spirochetes within the renal tubules of the kidney. Dissemination of pathogenic Leptospira occurs largely via urine excreted from these infected hosts, with bacterial acquisition by new hosts occurring through cuts/abrasions or mucous membranes as a consequence of direct exposure to infectious urine or urine-contaminated water [1]. One of the most significant reservoir hosts of Leptospira is the brown rat, Rattus norvegicus [10], which is critical for the spread of leptospirosis among urban slum settings. Rats can shed Leptospira in urine at high densities (104 to 107 cells/ml) over a prolonged period of time due to persistent renal colonization, and without apparent detrimental effects on the health of the rat [11], [12]. Within humans, clinical symptoms associated with leptospirosis range from mild illness including fever, chills, headache and myalgia to serious sequelae including hepatic, renal or pulmonary disorders. Severe cases of leptospirosis, which frequently display renal involvement, are reported to be diagnosed in 5–10% of patients, with mortality rates among these patients estimated at 5–40% [13], [14]. Asymptomatic renal colonization and excretion of Leptospira within the urine of patients for weeks and, in rare instances, months have been reported [2], [15], [16].
Although the underlying mechanisms by which pathogenic Leptospira spp. are able to persist within mammalian hosts are incompletely understood, we have put forward the hypothesis that post-translational incorporation of methyl groups in leptospiral outer membrane proteins (OMPs) may enhance bacterial survival by modifying proteins and their respective functional interactions within the host. In support of this we have previously reported the addition of O-methyl esters to selected glutamic acid residues within the leptospiral protein OmpL32 [17], and Cao et al. have detected mono-, di- and tri-methylations on proteins from L. interrogans serovar Lai, including the major leptospiral lipoprotein LipL32 [18]. Also in support of this hypothesis is the observation in the Gram-negative bacterium Rickettsia prowazekii that methylation of OmpB alters the immunogenicity of this outer membrane protein [19]. Further, OmpB proteins from the virulent Rickettsia Breinl and Evir strains have been shown to be more extensively methylated, and more protective, than OmpB from the avirulent Madrid E strain [20]–[24].
In the current study we investigate the presence of PTMs, including methylations, occurring within the major leptospiral lipoprotein LipL32. Since LipL32 is found only within pathogenic Leptospira serovars and constitutes the most abundant protein within the cell (approximately 38,000 copies per cell) [25], the incorporation of PTMs within LipL32 could have significant implications for the currently elusive function of this protein. Herein we analyzed the extent of LipL32 PTMs observed in leptospires directly isolated from a rat chronic infection model in comparison with in vitro-cultured Leptospira. Through these analyses we show that PTMs, composed of either tri-methylations or acetylations, occur solely on lysine residues from in vivo-derived LipL32. Four PTMs occurred in regions of LipL32 previously identified as B cell epitopes, and one PTM decreased epitope recognition by human leptospirosis serum. This study provides the first in vivo proteomic investigation of post-translational modifications of a major leptospiral lipoprotein and provides insight into the potential functional consequences of such modifications.
Samples were obtained from patients with confirmed leptospirosis [5] and healthy control individuals [7] who were residents of the city of Salvador, Brazil and provided written informed consent. The study protocol was approved by the institutional review boards of Yale University and Oswaldo Cruz Foundation. Leptospires shed in rat urine were collected using a protocol approved by the University College Dublin Animal Research Ethics committee, approval P-42-05, and licensed by the Department of Health and Children, Ireland, license number B100/3682.
Leptospira interrogans serovar Copenhageni strain RJ16441, a human clinical isolate [26] was passaged through guinea pigs to maintain virulence as previously described [27]. In vitro cultivated Leptospira (IVCL) samples were cultured at 30°C in Ellinghausen-McCullough-Johnson-Harris liquid medium (EMJH; Becton Dickinson, Oxford, England) [28], [29] supplemented with 6% rabbit serum (Sigma-Aldrich, Arklow, Ireland). Bacteria were enumerated by dark-field microscopy, harvested by centrifugation at 12,000× g for 10 min at 4°C when cultures reached a density of 1×108 leptospires/ml, and bacterial pellets were washed three times with 10 mM Tris pH 7.4/1 mM EDTA and stored at −20°C. Rat urine isolated Leptospira (RUIL) samples were prepared as previously described [12]. Briefly, 6 six-week old male Rattus norvegicus Wistar strain animals (150–190 g, Charles River Laboratories, UK) were infected by intraperitoneal injection with 5×107 low passage in vitro cultivated Leptospira. Upon detection of infection, rats were placed into metabolic cages and urine was collected twice weekly for six weeks. Urine samples were centrifuged at 12,000× g for 10 min at 4°C and pellets were washed and stored at −20°C as outlined above.
IVCL pellets (8×108 cells) and RUIL pellets (a total of 12 pellets collected from six rats) were solubilized by overnight incubation at room temperature in a total volume of either 80 µL (IVCL) or 240 µL (RUIL) rehydration buffer (7M urea, 2M thiourea and 1% ASB-14). The protein concentration of the IVCL and RUIL samples was determined using the RC/DC protein assay kit (Bio-Rad, Hertfordshire, UK). A total of 100 µg of protein from the IVCL samples and 400 µg of protein from the RUIL samples was used to rehydrate 18 cm 4–7pH IPG strips overnight, and proteins were separated by two dimensional gel electrophoresis (2DGE) and stained with SYPRO Ruby as previously described [30], [31]. A higher amount of protein was used in the RUIL samples to compensate for the presence of rat urine proteins. Protein spots corresponding to LipL32 were identified by comparison with published data [32], [33] and the same isoform of LipL32 was excised from each gel (as determined by alignment of gel images) for analysis by LC-MS/MS.
Excised LipL32 spots were digested with trypsin and analyzed using a QSTAR Pulsar I Hybrid Quadrupole-TOF liquid chromatography-electrospray ionization-tandem mass spectrometer (LC-MS/MS) as previously described [17].
MS/MS data were searched by centroiding with Analyst QS 1.1 Mascot script 1.6 b24 (MDS Sciex) to create the Mascot generic format file (.mgf) for data base searching. The.mgf files were searched against the L. interrogans proteome in UniProtKB annotation (http://www.uniprot.org/uniprot/?query=leptospira+interrogans+serovar+copenhageni&sort=score&format=*) that was maintained in-house at the University of Victoria-Genome BC Proteomics Centre. Search parameters used were as follows: fixed modification was set to carbamidomethyl (C) and variable modifications were set to methyl (D/E/K), tri-methyl (K), acetyl (K), peptide mass tolerance +0.3 Da, fragment mass tolerance +0.15 Da and trypsin enzyme specificity with 2 missed cleavages. Peptides with ion scores at or above the significance threshold (set at 28) were included in the analysis. Post-translational modifications comprising tri-methylations and acetylations were confirmed via manual de novo sequencing using the user input sequence for a given peptide spectrum. To be classified as a confirmed PTM and to identify the precise location of the PTM within the peptide, at least one of the following requirements needed to be achieved: (1) the presence of a neutral loss ion (MH+-59) unique for a tri-methylated lysine [34], [35]; (2) the presence of spectral ions encompassing the modified residue and the residue immediately preceding the modified residue, with confirmation that the former contained the PTM and the latter lacked the PTM; or (3) the presence of spectral ions derived from the residue immediately preceding the modified residue and the residue immediately following the modified residue, with confirmation that the former lacked the PTM, the latter included the PTM and did not constitute a residue that could be modified by a tri-methyl or acetyl group, and cumulative modifications could not result in a Δ m/z of 42 daltons.
Two LipL32 peptides corresponding to previously identified LipL32 immunodominant epitopes [36] that were identified through MS/MS analyses as containing confirmed PTMs (see requirements outlined above) were synthesized by either GenScript (Piscataway, NJ, USA) or ChinaPeptides (Shanghai, China). For each peptide, unmodified and tri-methylated versions were synthesized (Table 1). Ninety six-well MaxiSorp plates (Nalge-Nunc, Rochester, NY, USA) were coated overnight at 37°C in triplicate with 4 µg per well of either unmodified or modified LipL32 peptides or, as a negative control, an unrelated peptide from the Treponema pallidum Tp0751 protein sequence [37]. Wells were blocked with 4% skim milk powder/PBS pH 7.0 for 1.5 hours at room temperature and washed four times with PBS pH 7.0 containing 0.05% Tween-20 (PBST). Wells were incubated for 1.5 hours at room temperature with either a 1∶100 dilution of pooled serum prepared in blocking buffer collected from patients with laboratory-confirmed leptospirosis (n = 15) [5] or, as a positive control, a 1∶5000 dilution of polyclonal rabbit anti-LipL32 serum [38] prepared in blocking buffer. After washing 6 times with PBST, wells were incubated for 1.5 hours at room temperature with a 1∶3000 dilution of either goat anti-human IgG (Fab specific)-peroxidase or goat anti-rabbit IgG (whole molecule) F(ab′)2 fragment-peroxidase (both purchased from Sigma-Aldrich, Oakville, Ontario, Canada). Wells were washed six times with PBST and developed at room temperature with the TMB peroxidase substrate system (Kirkegaard & Perry Laboratories, Gaithersburg, MD, USA). Plates were read at 600 nm with a Synergy HT plate reader (BioTek, Winooski, VT, USA), and statistical analyses were performed using the Student's two-tailed t-test.
Annotation of the identified lysine modifications within the solved Ca2+-bound LipL32 structure [39] was accomplished using the PyMOL Molecular Graphics System, Schrödinger, LLC., available at http://www.pymol.org/pymol.
In the current study we investigated the presence of PTMs, with a focus upon methylations, occurring in LipL32 from L. interrogans serovar Copenhageni cells grown under in vivo and in vitro growth conditions. Total protein was prepared from leptospires shed in the urine of rats housed in metabolic cages (rat urine-isolated Leptospira; RUIL) and from leptospires cultured in EMJH medium (in vitro-cultivated Leptospira; IVCL), and each of the whole cell proteomes was independently subjected to two-dimensional gel electrophoresis (Figure 1). Two protein spots corresponding to LipL32 were excised from each of the RUIL and IVCL samples; to allow for a direct comparison, the same LipL32 isoform was selected from each gel. The protein spots were digested with trypsin and analyzed by LC-MS/MS to verify the identity of the excised proteins. The resultant spectrometry data was subjected to manual analysis to confirm detected PTM position assignments. A list of peptides identified by LC-MS/MS, including peptides containing confirmed PTMs, is shown in Table 2, and the obtained MS/MS data and spectra for peptides containing confirmed PTMs can be found in the Supplementary Data Files (Table S1 and Figure S1, respectively).
High protein coverage was obtained for the two LipL32 protein spots excised from each of the IVCL and RUIL samples (designated as IVCL1, IVCL2, RUIL1 and RUIL2; see Table 2). Combined, the detected peptides covered 45% (IVCL1 and RUIL1) and 47% (IVCL2 and RUIL2) of the LipL32 sequence. Analysis of PTMs observed within the RUIL samples identified eight peptides containing modifications, whereas no modifications were detected in the IVCL samples. All modifications were confirmed by manual analysis of the mass spectra and were exclusively observed on lysine residues. The observation that five out of eight of the modified lysine residues were detected in peptides containing a missed trypsin cleavage (Table S1) is noteworthy, for the presence of lysine modifications is known to reduce the cleavage efficiency of trypsin by masking the enzyme substrate site [34]. Thus, the number of lysine modifications detected within LipL32 in this study may be under-representative of the actual number present within this protein.
For five of the eight modified peptides that were detected, the nature of the PTMs could only be narrowed down to either tri-methylation (Δ m/z 42.0471) or acetylation (Δ m/z 42.0106). Due to the highly similar m/z ratios obtained for tri-methylation and acetylation, differentiation of these two modifications via mass spectrometry relies upon detection of the neutral loss ion (MH+-59) which is unique for a tri-methylated lysine [34], [35]. In our analyses, three peptides contained MH+-59 neutral loss ions (LDDDDDGDDTYKEER from RUIL1, SFDDLKNIDTK from RUIL2, and QAIAAEESLKK from RUIL2; Table 2). For each peptide the presence of the neutral loss ion on a fragment containing the modified lysine, combined with the presence of spectral ions surrounding the modified lysine, enabled precise residue assignment of the tri-methyl group addition to residues K152, K178 and K246 (Figure S1). Of note, a modified version of the LDDDDDGDDTYKEER peptide was also detected in the RUIL2 sample. Although the spectra obtained for this peptide prevented assignment of a precise location for the modification, it could be narrowed down to the region of the peptide encompassing the sequence TYKE (residues 150–153; Figure S1). For seven out of eight of the modified peptides detected within RUIL1 and RUIL2 the corresponding unmodified peptide was also detected, indicating that the observed modifications were not ubiquitous within the LipL32 spots selected for analysis. No further PTMs were detected within either the IVCL or RUIL samples.
Analysis of the modified LipL32 peptides detected within the RUIL samples highlighted four peptides that overlap with regions previously identified through epitope mapping studies as antigenic regions of LipL32 [36]. Within the reported LipL32 epitope spanning residues 132–158 of the mature protein (designated as P1; AAKAKPVQKLDDDDDGDDTYKEERHNK) we detected tri-methylation of residue K152. For the LipL32 epitope encompassing residues 162–185 of the mature protein (designated as P2; LTRIKIPNPPKSFDDLKNIDTKKL) we detected modifications on K166, K172 and K178, with the latter constituting a confirmed tri-methylation. In order to investigate if the modifications altered immune recognition of these epitopes, synthetic peptides corresponding to the two previously identified epitopes were prepared; specifically, unmodified and K152 tri-methylated versions of P1 were synthesized, while for P2 an unmodified and a representative modified peptide containing a tri-methylated K178 residue were synthesized (Table 1). When tested in an ELISA assay (Figure 2), pooled sera obtained from patients with laboratory-confirmed leptospirosis reacted strongly to the unmodified version of P1, while reactivity against the K152 tri-methylated peptide version was significantly decreased (p<0.0001). As a control, polyclonal antiserum raised against recombinant LipL32 was tested for reactivity against the unmodified and modified versions of P1; interestingly, no difference in reactivity was observed.
In our study minimal reactivity of both patient sera and rabbit serum was observed against the unmodified and modified versions of P2, a result that differs from the reactivity observed by Lottersberger et al. against P2 using hyperimmune rabbit serum [36]. Assessment of the coated P1/P2 peptide concentrations demonstrated no difference in coating efficiency between the peptides, indicating this result was not due to inadequate coating of P2 within the ELISA wells. Due to the minimal reactivity observed against the unmodified version of P2, the effect of modification of P2 on immune recognition of this epitope could not be assessed.
The structural position of the eight modified lysines detected within the LipL32 sample originating from the rat urine-isolated leptospires was determined through examination of the Ca2+-bound LipL32 crystal structure [39]. For consistency and ease of interpretation, we have used the numbering system C1-K253 which corresponds to the mature LipL32 protein (lacking the 19 residue cleaved SPII signal sequence) set forward by Vivian et al. [40] and adopted by Tung et al. [39]. As shown in Figure 3A, the observed modifications were localized to the β2 strand (K29), the loop between the β8 and β9 strands (K152), the β9 strand (K166), the loop between the β9 strand and the α3 helix (K172), the α3 helix (K178), the loop between the β10 and β11 strands (K199) and the α4 helix (K245 and K246). All eight modified lysines localized to surface accessible regions of the structure (Figure 3B). Of note, the modified residues K152 and K199 are located in immediate proximity to the LipL32 Ca2+-binding site, which comprises the negatively charged surface formed by residues D142–D149 (encompassing a portion of the β8 strand and the β8β9 loop) and the Ca2+ ion-coordinating residues in the α1β7 loop (D113 and T114) and the β8β9 loop (D145, D146 and Y159) [39]. Lysine 199 is also in close proximity to the reported collagen-binding site encompassing residues L53, V54, Y62, W115, R117, Y151 and Y198 [40] (Figure 3B). In fact, K199 is positioned on a flexible loop between strands 10 and 11 that bridges the Ca2+ and reported collagen-binding sites of LipL32, suggesting that modification of K199 may play a role in regulating protein architecture and function.
Two previous proteomic studies have established the presence of post-translationally modified proteins within L. interrogans. The leptospiral protein OmpL32 (corresponding to LIC11848) was shown to be differentially methylated on glutamic acid residues from L. interrogans grown under in vitro conditions [17]. Additionally, a global PTM analysis of in vitro-cultured L. interrogans serovar Lai detected multiple PTMs on a broad range of proteins, including 46 proteins with 54 lysine acetylation sites, 104 proteins with 135 glutamic acid/glutamine methylation sites, and 58 proteins with 64 lysine/arginine methylation sites. One of the lysine-modified proteins detected within the latter study was LipL32, which was shown to be devoid of lysine acetylations but to contain mono-, di-, and tri-methyl groups on residue K152 [18]. In the study presented herein, we similarly detected modification of residue K152, as well as seven other lysine residues, within LipL32 from in vivo-grown leptospires. Of the modified lysines that were detected an unambiguous assignment of tri-methylation could be made for only residues K152, K178 and K246. By extrapolation from these results it is plausible that the remaining five detected PTMs constitute similar lysine tri-methylations, however definitive assignment could not be made from the spectra obtained in this study.
The prior proteomic studies outlined above establish that post-translational protein modification is a phenomenon which occurs within in vitro-cultured L. interrogans. Our study extends this observation and establishes that, at least for modification of lysine residues within the major leptospiral protein LipL32, this phenomenon also occurs under biologically relevant in vivo infection conditions. Of particular interest is the fact that in our study modification of LipL32 lysine residues was restricted to in vivo-grown organisms, while in the study performed by Cao and colleagues LipL32 lysine modification was observed at K152 within in vitro-cultured Leptospira [18]. These divergent results may stem from the use of different strains in the two in vitro studies (Lai in the Cao et al. study, Copenhageni in our study). Alternatively, the difference may have arisen due to the higher in vitro culture density attained for L. interrogans in the Cao et al. study, which reached approximately 6.6×108 bacteria/ml compared to 1×108 bacteria/ml in our study. The higher density may be more representative of that achieved during kidney colonization by Leptospira and thus may indicate the process of protein modification is dependent, at least partially, upon elevated bacterial density. Such a phenomenon has recently been described for Pseudomonas aeruginosa with respect to the PrmC methyltransferase, a protein that mediates methylation of key virulence factors and whose activity is controlled by quorum sensing [41].
In this study we observed a correlation between lysine modifications occurring within LipL32 and regions of the protein that have been previously identified as immunogenic. Seven out of eight of the identified lysine modifications occurred in regions identified by Hauk and colleagues as being reactive with patient sera [42]. Of particular interest, six of the modifications occurred in the region of the mature protein encompassing amino acids 166–253 that was shown by Hauk et al. to be most strongly reactive with acute phase sera [42]. Further, four out of eight of the identified lysine modifications occurred in regions previously identified as epitopes via peptide mapping studies [36]. Our subsequent demonstration that modification of K152 within one of these epitopes significantly decreased the reactivity of leptospirosis patient serum compared to the unmodified epitope version suggests the possibility that lysine tri-methylation within LipL32 may decrease immune recognition of this protein during infection. Supporting this theory is our finding that decreased reactivity was solely observed using patient sera and not rabbit serum raised against recombinant LipL32, suggesting the decreased immune recognition of the modified LipL32 epitope is a phenomenon observed only under conditions of infection. Additionally, the modifications present within Leptospira isolated from rats may differ from those arising during infection in humans, a possibility that would further explain the decreased immune recognition of patient sera against a rat-derived modified LipL32 peptide.
An alternative, or additional, proposed function for LipL32 lysine modification comes from determination of the location of the modified residues within the solved LipL32 structure. LipL32 has been shown to possess a calcium-binding cluster that comprises residues D113, T114, D145, D146, and Y159, with the latter three residues being localized to the β8β9 loop within the LipL32 structure [39]. The β8β9 loop undergoes a significant conformational change upon calcium binding, a shift that has been suggested to account for the enhanced stability observed in the LipL32 calcium-bound state [39], [43]. The modification observed within our study at K152 similarly lies within the β8β9 loop. The relatively large size and charge modification of a tri-methyl group at K152 and either an acetyl or tri-methyl modification at K199, combined with the close proximity to the Ca2+-coordinating residues D145, D146 and Y159, has the potential to affect Ca2+ binding. This may affect the stability of LipL32 and/or impact LipL32 functions that are linked to calcium binding, such as the putative immune system modulatory function that has been ascribed to the calcium-bound state of LipL32 (promotion of interaction with the Toll-like receptor 2 to induce an inflammatory response) [44]. Whether methylation could influence binding of LipL32 to extracellular matrix components, which has been shown by Hauk et al. to be independent of calcium binding [45], or other currently unknown LipL32 functions remains to be seen.
Although originally identified as a surface-exposed outer membrane lipoprotein, a recent study using multiple, independent methods has suggested that LipL32 has limited surface exposure [46]. If discrete sub-surface and surface-exposed LipL32 subpopulations exist, PTMs incorporated into the latter may contribute to immune evasion, as hypothesized in the current study. Interestingly, we detected both unmodified and modified versions of the identified LipL32 peptides from in vivo-isolated organisms, indicating that subpopulations of modified/unmodified LipL32 proteins are present during the course of an infection. These could correspond to distinct subpopulations comprised of surface-exposed/modified and sub-surface/unmodified LipL32 molecules. Additionally, since the RUIL samples analyzed in this study constitute a pool collected over a six week time frame, PTMs could have arisen over time in response to immune pressure, with a higher proportion of modified LipL32 being present at later time points. Both scenarios suggest LipL32 modification could play a role in establishment of persistent infection. However, a study by Murray and co-workers showed no reduction in the ability of an L. interrogans lipL32 transposon mutant to colonize rats [47]. Although this result is seemingly at odds with our proposed link between LipL32 modification, immune evasion and leptospiral persistence, the ability of the lipL32 mutant to persist beyond 15 days post-infection was not assessed, thus leaving open the possibility that wild-type and LipL32-deficient Leptospira may exhibit differences in the capacity to establish long-term infection within rats.
The discovery of LipL32 lysine modification in organisms shed in the urine has potential implications for understanding how leptospires persist for extended periods of time in the renal tubules. Such persistence is responsible for the ability of leptospires to maintain high rates of shedding, transmission, and infection among reservoir host populations. An unanswered paradox in leptospiral biology is why leptospires exhibit high levels of LipL32 expression in the renal tubules [38] despite a robust LipL32 immune response during infection [48], [49]. The data presented here provide a partial explanation for this paradox by showing that serum from infected patients recognized a methylated LipL32 peptide far less well than the unmethylated form. This result indicates that lysine methylation sufficiently alters peptide epitopes as to prevent antibody binding. Since PTMs have also been shown to alter immunologic processing and presentation of peptides in the context of MHC class I [50], [51], the cell-mediated arm of the immune system may similarly exhibit an altered response to modified compared to unmodified LipL32. Of interest in this regard, an IFN-γ-stimulating T cell epitope has recently been detected within LipL32 and has been shown to reside between amino acids 1 and 181 [52], a region that corresponds to five of the eight modified lysines detected in this study (K29, K152, K166, K172, K178). The possibility exists that LipL32 modification decreases both humoral and cell-mediated immune recognition of the protein, thus preventing effective bacterial clearance and promoting leptospiral persistence. The importance of prevention of renal tubule colonization with respect to protection from leptospiral infection is exemplified by the study performed by Seixas et al. In this study, hamsters immunized with Mycobacterium bovis bacillus Calmette-Guérin (BCG)-expressing LipL32 were more likely to survive Leptospira challenge than hamsters immunized with BCG alone, and enhanced survival corresponded with an absence of renal tubule colonization [53]. Collectively, these studies suggest that an increased understanding of how PTMs affect the humoral and cell-mediated immune responses to LipL32 and, relatedly, whether modification of LipL32 contributes to bacterial persistence, may be key to development of vaccines that are able to generate sterilizing immunity.
In summary, this study reports the detection of multiple modified lysine residues within the major leptospiral protein LipL32 isolated from leptospires grown under in vivo, but not in vitro, conditions and provides insight into the potential functional and immunological consequences of these modifications. In addition this study provides a potential molecular mechanism for, and link between, establishment of chronic Leptospira infection in rats and acute Leptospira infection in humans. Of importance to assess through future studies are the temporal occurrence, prevalence, and biological relevance of LipL32 lysine modifications at both the cellular and community levels, as well as whether PTMs are detected on other leptospiral OMPs during the course of infection. These investigations may uncover the elusive function of this highly abundant leptospiral protein and may reveal a novel mechanism of bacterial immune evasion.
|
10.1371/journal.pcbi.1004720 | Recovery of Dynamics and Function in Spiking Neural Networks with Closed-Loop Control | There is a growing interest in developing novel brain stimulation methods to control disease-related aberrant neural activity and to address basic neuroscience questions. Conventional methods for manipulating brain activity rely on open-loop approaches that usually lead to excessive stimulation and, crucially, do not restore the original computations performed by the network. Thus, they are often accompanied by undesired side-effects. Here, we introduce delayed feedback control (DFC), a conceptually simple but effective method, to control pathological oscillations in spiking neural networks (SNNs). Using mathematical analysis and numerical simulations we show that DFC can restore a wide range of aberrant network dynamics either by suppressing or enhancing synchronous irregular activity. Importantly, DFC, besides steering the system back to a healthy state, also recovers the computations performed by the underlying network. Finally, using our theory we identify the role of single neuron and synapse properties in determining the stability of the closed-loop system.
| Brain stimulation is being used to ease symptoms in several neurological disorders in cases where pharmacological treatment is not effective (anymore). The most common way for stimulation so far has been to apply a fixed, predetermined stimulus irrespective of the actual state of the brain or the condition of the patient. Recently, alternative strategies such as event-triggered stimulation protocols have attracted the interest of researchers. In these protocols the state of the affected brain area is continuously monitored, but the stimulus is only applied if certain criteria are met. Here we go one step further and present a truly closed-loop stimulation protocol. That is, a stimulus is being continuously provided and the magnitude of the stimulus depends, at any point in time, on the ongoing neural activity dynamics of the affected brain area. This results not only in suppression of the pathological activity, but also in a partial recovery of the transfer function of the activity dynamics. Thus, the ability of the lesioned brain area to carry out relevant computations is restored up to a point as well.
| Open-loop brain stimulation has emerged as a common tool to restore aberrant neuronal activity. The most successful example is the application of high-frequency deep-brain-stimulation (DBS) used to ameliorate motor symptoms in Parkinson’s disease (PD). However, even in this case the stimulation induces side-effects such as gait imbalance, cognitive impairment, speech impairment, depression etc. [1]. The main cause of these side-effects is likely to be the constant stimulation, but additional explanations are plausible, e.g. the inability of open-loop stimulation to recover the original computations carried out by the impaired brain area. Thus, there is a clear need for more sophisticated brain stimulation schemes [2–4].
Moreover, to exploit the full potential of external brain stimulation as a research and therapeutic tool it is important to obtain theoretical insights that can guide the design of novel stimulation protocols. The goal for these new stimulation methods should ideally be twofold: to alter the dynamical state of the brain activity in a desired manner and to recover the computations performed by the network. Here, we demonstrate that DFC, a method with its origin in chaos control [5, 6], can achieve these objectives.
To show that DFC is effective in altering the global activity state, we focus on its ability to switch the network state between synchronous-irregular (SI) oscillatory and asynchronous-irregular (AI) non-oscillatory activity. This choice is motivated by the fact that several brain diseases are manifested as a transformation of the AI state to persistent SI oscillations, e.g. in PD [7] and in certain forms of epilepsy [8], or as the inability of the network to generate transient SI activity, e.g. in schizophrenia [9]. To demonstrate that DFC facilitates the recovery of certain types of computations, we also illustrate how a network under DFC can effectively process and route rate as well as temporally coded signals. Thus, DFC not only steers the system to a more physiological activity regime, but it also recovers to a considerable degree the coding abilities of the network as they were present before the onset of the pathology.
Previous theoretical models of closed-loop stimulation are not suitable to study the control of SI oscillations because the dynamics that arise in networks of phase oscillators [10–12], in networks of Hodgkin-Huxley neurons [13, 14] and in Wilson-Cowan type firing rates models [15] are qualitatively different from the SI oscillations [16, 17]. In addition, the physiologically plausible SI oscillations are known to be robust to both noise and heterogeneities [18–20] and, therefore, require a more differentiated control approach.
Our control strategy is applicable to any network with arbitrary connectivity that undergoes a Hopf bifurcation and it is useful but not critical to know the parameters a priori. In fact, we show how adaptive tuning methods can be used to estimate the control parameters if the precise network parameters are not known. The DFC based stimulation method proposed here can be in principle applied in all animal models where the use of optogenetic tools allows for direct modulation of the membrane potential. In human patients, where optogenetics is currently not an option, manipulation of the membrane potential can be achieved indirectly via subthreshold electrical stimulation. Importantly, the method is not restricted to stimulation of deep structures, but could be applied non-invasively to modulate activity in cortical layers as well [21]. Finally, the theoretical insights we provide into the mechanisms of feedback control in SNNs could also explain the recent success of event-driven stimulation schemes [22–24].
Excitation and inhibition (EI) in balanced random SNNs cause asynchronous irregular (AI) and non-oscillatory population activity. This state resembles the ongoing activity in the healthy state [17]. Changes in the EI balance, caused by altered inputs and/or changes in the recurrent synaptic strengths, can result in two qualitatively different types of oscillations. The synchronous-regular (SR) oscillations arise when the mean input to the individual neurons exceeds their spiking threshold, resulting in high firing rate and high frequency regular oscillations [16, 17]. By contrast, the synchronous-irregular (SI) oscillations arise because of strong synaptic coupling and increased variance of the total input to the neurons. Importantly, the emergence of the SI oscillations is accompanied by a change in the network transfer function and its ability to represent stimulus-related activity. Persistent SI oscillations often are signature of brain diseases, e.g. in PD [7] and epilepsy [8]. The altered network transfer function and the robustness of the oscillations to noise and neuronal heterogeneities pose a serious challenge for stimulation-based therapeutic approaches. In the following we show that DFC is able to both quench SI oscillations and to recover the original network transfer function.
In order to provide an explanatory account of the exact mechanism by which DFC operates we provide specific examples of neuronal networks with known parameters that can generate SI activity. To this end, we used mean-field theory to describe the dynamics of homogeneously and randomly connected recurrent SNNs. The mathematical analysis was corroborated with appropriate numerical simulations of sparsely connected SNNs. We considered networks with only inhibitory and both excitatory and inhibitory neurons.
While our goal is to reveal the mechanisms by which DFC controls SI activity in excitatory-inhibitory SNNs, it is more instructive to first demonstrate the concept in a simple, purely inhibitory SNN. In the AI state the population average of the firing rate is constant in time r(t) = r0. Therefore, the mean recurrent input that each neuron receives is also constant:
I r e c ( t ) = C · J / C · ( s ⋆ r ) [ t ] = C · J / C · ∫ s ( τ ) r ( t - τ ) d τ = J · r 0 · ∫ s ( τ ) d τ
where C is the average in-degree, J/C the synaptic coupling strength and s(t) the postsynaptic current. In such a network, the emergence of SI oscillations can be investigated by analyzing the stability of the network firing rate in the AI state [18, 19]. A small perturbation in the steady-state firing rate
r ( t ) = r 0 + R e [ r ^ 1 ( λ ) e λ t ]
where eλt is an eigenmode of the network dynamics with complex eigenvalue λ, leads to a perturbation in the recurrent input
I ( t ) = I 0 + R e [ I ^ 1 ( λ ) e λ t ]
with I ^ 1 ( λ ) = - J S ( λ ) r ^ 1, where J is the synaptic coupling strength and S is the synaptic response function. In a recurrent network both perturbations have to be consistent, that is
r ^ 1 ( λ ) = R ( λ ) I ^ 1 ( λ )
where R is the neuron response function. This results in the self-consistency equation:
J · R ( λ ) · S ( λ ) = 1 (1)
In a purely inhibitory network J is negative, but here the negative sign of J has been absorbed in the phase S(λ). We can then compute the eigenvalue spectrum, that is the roots λ that satisfy Eq (1). When the eigenvalues have a positive real part, the AI state is unstable and the SNN settles in the SI state. Note that due to the synaptic delays the spectrum is infinite. However, in time-delay systems of the retarded type that we are considering here, the total number of unstable eigenvalues is always finite [25]. Increasing J shifts the spectrum towards more positive values on the real axis. For a critical value Jcr a complex pair of eigenvalues crosses the imaginary axis and the system becomes unstable through a supercritical Hopf bifurcation [18] (Fig 1). In the following we consider a SNN in which J > Jcr, thus resulting in the emergence of SI oscillations. We aim at designing a controller that can alter the global activity state from SI to AI by placing the unstable eigenvalues back to the left half-plane (Fig 1e).
For the implementation of our DFC controller we made two assumptions. First, we chose the instantaneous population rate to be the output state of the system. This state needs to be continuously monitored by the controller. Based on this state, the controller evaluates an appropriate control signal. The second assumption is that the control signal can be applied via current injection directly to the somata of the neurons. Note that the control signal is identical for all neurons in the network. Thus, the synchrony in the network activity was not decorrelated because each neuron received different input. To include the contribution of DFC the self-consistency Eq (1) needs to be modified:
J R ( λ ) S ( λ ) e - λ · d - K R ( λ ) M ( λ ) e - λ · d c = 1 (2)
where K is the control gain, dc the control delay and M the control kernel. The roots of the above equation (see methods) yield the range of parameters K, dc that move the unstable eigenvalues back to the left-half plane (Fig 1e), which results in a switch of activity from SI to AI.
To verify the analytical solution we simulated an Erdös-Rényi type inhibitory recurrent network of N = 10,000 sparsely connected leaky-integrate-and-fire (LIF) neurons. The neurons were connected with a connection probability ϵ = 0.1. Switching on the DFC with parameters estimated from Eq (2), almost immediately results in suppression of oscillations and in a network state that resembles the AI regime (Fig 2a–2d). The suppression of stochastic oscillations is evident both in the spiking activity of single neurons (Fig 2a) and in the population activity of the network (Fig 2b). The spike count variability and the irregularity of single neuron interspike intervals, estimated by the Fano Factor (FF) and the coefficient of variation (CV) respectively, confirm that under DFC the firing of individual neurons in the network follows Poisson statistics (AI: FF = 1.04, CV = 1.01, DFC: FF = 1.02, CV = 0.99). Moreover, the oscillation index that captures the degree of oscillatory activity (see methods) is in both conditions comparable (AI: PT = 1.47, DFC: PT = 1.45) and significantly smaller than in the SI state (PT = 3). The change in the network spiking activity is also observed in the subthreshold membrane potential of individual neurons (Fig 2c and 2d).
Next we demonstrate the applicability of DFC in changing the SI state in recurrent networks of excitatory and inhibitory neurons. To this end we simulated a SNN composed of 8000 excitatory and 2000 inhibitory neurons with Erdös-Rényi type connectivity and a connection probability of ϵ = 0.1. As in the I-I network, we used the mean-field approach to derive appropriate values for the parameters to attain an SI state. The self-consistency equation for the coupled EI-network under DFC is given by
[ ( J E I · S E I ( λ ) e - λ · d E I - K · M ( λ ) e - λ · d c ) · R E ( λ ) ] · J I E · R I E ( λ ) · S I E ( λ ) · e - λ · d I E = 1 (3)
where Jij, dij is the synaptic coupling strength and delay from population j to population i and RE (RI) is the neuron response function of excitatory (inhibitory) neurons. Note that we ignored the recurrent couplings within E and I populations, because we were interested only in the oscillations created by the EI-loop. We implemented DFC by recording the activity of neurons in the inhibitory population while stimulating excitatory neurons. Here again K is the control gain, dc the control delay and M the control kernel. Switching on the controller yielded a near instantaneous transition in the network activity from SI to AI (Fig 2e–2h). In this case the original physiological state that we wanted to recover was characterized by slightly less irregular firing of the individual neurons. Nevertheless, DFC successfully steered the network to a regime with statistics comparable to the AI activity (DFC: FFE = 0.85, CVE = 0.91, AI: FFE = 0.83, CVE = 1.03).
In a coupled network with more than one population additional possibilities for recording and stimulating neurons exist. For instance, we could both record and stimulate the excitatory population (see below “stability and robustness of control domains”). Our results, however, do not depend on the exact identity of the recorded and stimulated neurons. The oscillation frequency in the SI state of the network was in the beta range, f ≈ 25–30Hz, which is characteristic for PD [7]. This suggests that if PD-associated beta oscillations are caused by a strong coupling between STN and GPe [26, 27], then our DFC approach could be used to suppress these beta band oscillations. The results presented here are general and the same approach can be applied to suppress oscillations in other frequency bands as well as long as the oscillations underlie a Hopf bifurcation.
To determine the range of values that led to stable control we fixed the control kernel M, using a box function of width 1 ms, and parametrized the system by the control gain K and delay dc. For each pair of values we simulated the SNN and computed the oscillation index (Fig 3a and 3c). The (K, dc)-plane shows that a stable control domain exists at 7 ms. That is, an effective control delay of dc,eff = 7 ms yields the maximum stability for the resulting AI state. The semi-analytical results from mean-field theory (red contour lines derived from Eqs 2 and 3 for the I-I network and E-I network, respectively) are in good agreement with the numerical simulations (blue shades). The only discrepancy occurs when the difference between synaptic and control coupling is small. In such a scenario it is more difficult to maintain constant rates of the stimulated population and the system may become effectively excitatory leading to rate instabilities [28]. Moreover, fluctuations in the mean input that are ignored in our mean-field approach could also become more important. The analysis of these fluctuations is beyond the scope of this work and will be addressed in a future study.
Despite the fact that one stable control domain exists, a compensation mechanism to maintain constant firing rates is required to achieve stable control. In real-life applications a detailed fine-tuning may not always be possible. Therefore, we modified our control protocol and introduced an additional delay term dc2, thus, effectively feeding into the controller the difference between two time-delayed versions of the population activity. For such differential DFC scheme the control signal is given by (see methods):
I C ( t ) = K · M ( t ) ⋆ ( v ( t - d c 1 ) - v ( t - d c 2 ) )
Differential control has been previously used to control unstable periodic orbits [5, 29] and to suppress synchrony in networks with discrete-time neuron models [30]. With the differential control we accounted for the fact that recording neural activity and injecting a control current into the neurons introduces a finite time-delay. Therefore, we used a small but non-zero value for dc2, i.e. dc2 = 1 ms, which is close to the overall closed-loop delay introduced by current technologies [31, 32]. A crucial advantage of differential DFC is that no additional rate compensation is required, because the mean contribution of the control signal vanishes
l i m T → ∞ 1 T ∫ 0 T ( v ( t - d c 1 ) - v ( t - d c 2 ) ) d t = 0
Moving in the control parameter space (K, dc), therefore, did not affect the firing rates of the neurons. This was reflected in the near perfect overlap of theoretical predictions and numerical simulations of the SNN (Fig 3b). In addition, differential DFC introduced two positive effects on the stability of the control domains: (i) The first control domain was expanded, which amounts to an increase in the robustness in the parameter variation. That is, small deviations from the estimated values of the gain and the delay would not be critical for the stability of the AI state achieved by differential control. (ii) A new stable control domain appeared at t = 23 ms. Thus, with differential control there is an increase of the range of parameters that lead to stability.
DFC also enhanced the robustness of the system to external disturbances, e.g. undesired signals at the controller output, measurement noise etc. This becomes evident when we consider the distance Bcr of the complex eigenvalues λi from the imaginary axis for the main stable control domain at t = 7 ms. A more robust closed-loop system is reflected in higher values of Bcr. Differential and direct control yielded B d i f f c r = m a x ( R e ( λ i ) ) = - 257 and B d i r e c t c r = m a x ( R e ( λ i ) ) = - 224, respectively, clearly revealing a more robust system with differential DFC.
Both direct and differential control were effective in a E-I network as well. The location of the stable control domains depended on the exact implementation (Fig 3c–3e). When the activity of the inhibitory population was monitored while stimulating the excitatory population, the main stable control domain appeared at t = 7 ms (Fig 3c). This location is identical with the purely inhibitory network and reflects the overall delay of the I-E path (I-I loop) in the E-I (I-I) network. Indeed for both the I-E path and I-I loop the effective delay is deff = 7 ms (see methods). By contrast, when the excitatory population was both recorded and stimulated then the location of the domains shifted to around t = 14 ms reflecting the larger overall delay in the E-I-E loop. (Fig 3d and 3e). Note that in this case the stable control domain for direct control was smaller. The reason is that the size of the stable control domains shrinks for larger delays.
In both the I-I and E-I SNNs we applied an identical control signal to all stimulated neurons. That is, we did not disrupt oscillations and decorrelated network activity by injecting different currents to each of the neurons. This is in contrast with a widespread assumption that common input always tends to increase correlations in neural activity [33]. The results from the application of DFC reveal that common input can both increase or decrease correlations in SNNs. It is the timing and the amplitude of the common input that determines the direction in which correlations are affected.
It is important to point out that injection of a control signal is not equivalent to the application of additive noise to the system. To demonstrate this we simulated an I-I network and injected Gaussian noise with the same mean and variance as the control signal to all neurons. This stimulation approach failed to suppress SI oscillations (Fig 4a and 4b) indicating that the temporal structure of the control signal is crucial for successful control. Increasing further the noise intensity, e.g. by a factor of ten, eventually resulted in desynchronization of the activity and in quenching of oscillations (Fig 4d and 4e). However, with such strong external noise the network dynamics is predominantly influenced by the input rather than the recurrent activity. This condition is disastrous from a computational point of view, because any information processing taking place within the stimulated brain region would be severely impaired. To illustrate this we recorded the subthreshold dynamics of ten randomly selected neurons in the network (Fig 4f). The huge fluctuations in the membrane potential under the influence of strong external noise are rather pathological. By contrast, the fluctuations in the case of DFC are comparable to those in the physiological AI regime.
The detrimental effect of strong external noise became even more apparent when we studied the response of the network to incoming stimuli. We examined two scenarios. First, we tested how a series of incoming pulse packets composed of randomly distributed spikes are processed by the SNN. We evaluated the network response by the area under the curve (AUC, see methods) for each of the following network states: AI, SI quenched DFC and SI quenched by noise stimulation. A high AUC value reflects better separability of two conditions. It is evident that the AUC in the AI state and in the DFC condition is close to unity indicating that both conditions are comparable in terms of stimulus separability (Fig 5a and 5b). By contrast, when the SI oscillations were quenched by the injection of strong external noise the AUC dropped significantly. That is, DFC, in contrast to strong noise stimulation, does not impair the ability of the network to detect incoming stimuli.
Next we tested how DFC affects temporal aspects of the network response. To this end, we provided external correlated inputs to all stimulated neurons and measured the spike train similarity in the network response. We computed the spike distance D that captures the time-resolved degree of synchrony between individual spike-trains ([34], see methods). Again DFC did not impair the temporal processing as indicated by a clear separation of the two clusters during baseline DB and stimulation DS (Fig 5c). For external noise, however, the two distributions of values strongly overlapped, showing that aspects of temporal processing as measured by pairwise synchrony are clearly compromised when the SI state is disrupted by open-loop noise injection. These results suggest that processing of incoming signals either locally or by downstream areas is feasible in a DFC scheme, but not in an external noise scheme.
The above two results clearly demonstrate that DFC has multiple advantages compared to the open-loop noisy stimulation. DFC does not only suppress SI activity steering the network to an AI regime, it also facilitates the recovery of the network’s ability to process stimulus related information. From its design it is evident that DFC effectively counteracts the increase in coupling strength, which is one of the main causes for the emergence of SI activity. Indeed, the goal of the DFC design was to move the poles of the system at, or close, to their original positions. Ideally, the stimulation kernel M would match the synaptic kernel S with dc = d and the amplitude of the control gain K would be tuned to match the pathological increase of the coupling strength ΔJ. If this were the case, DFC would completely eliminate the effects on the mean recurrent input. This is evident if we consider the modulation to a perturbation in the average input to a neuron
I ( λ ) = ( J + Δ J ) · R ( λ ) · S ( λ ) · e - λ d - K · R ( λ ) · M ( λ ) · e - λ d c = K = Δ J , M = S ( J + Δ J ) · R ( λ ) · S ( λ ) · e - λ d - Δ J · R ( λ ) · S ( λ ) · e - λ d = J · R ( λ ) · S ( λ ) · e - λ d
That is, under DFC the effects of ΔJ are not visible in the perturbed current term. In practical applications, of course, a perfect match between the control parameters (K,dc,M) with the synaptic values is not feasible, because the exact shape of the synaptic kernels are not known a priori and have to be estimated. Nevertheless, within a certain reasonable range of parameters (see also section “stable control domains”), DFC still places the eigenvalues close to the initial position they had before the onset of pathology. Therefore, as we showed above, aspects of both rate and temporal coding that the network may be performing are recovered.
The understanding of the exact mechanisms by which DFC suppressed SI activity allowed us to precisely investigate how the neuron and synapse response function R and S respectively influence the stability of the closed-loop system. To this end, we again used a mean-field approximation, which explicitly incorporates the expressions for R and S. In general, the neuron response R depends on the specific neuron model as well as on the external input. Here, we did not change the neuron model, but altered the external Gaussian white noise input by using different values for the mean and variance (μ, σ2). We then assessed the stability of the system. It is apparent that for a given pair of coupling and control parameters (J, d) and (K, dc), respectively, the system becomes unstable as we move in the two dimensional parameter-space spanned by the mean and variance (Fig 6a). For meaningful comparison we used (μ, σ2)-combinations that yield constant rates. In the ideal case where M(λ) = S(λ) and dc = d Eq (2) becomes:
J · R ( λ ) · S ( λ ) · e - λ d - K · R ( λ ) · M ( λ ) · e - λ d c = 1 ( J - K ) · R ( λ ) · S ( λ ) · e - λ d = 1 ( J - K ) · G S · R n ( λ ) · S ( λ ) = 1 (4)
where Gs is the slope (or the static gain) of the ‘f-I curve’ at the operating point and Rn the normalized neuron response (see methods). The critical effective coupling is then given by
L c r ( λ ) = ( J - K ) ( λ ) = 1 G s | R n ( λ ) · S ( λ ) |
As we move along the constant output firing rate lines both Gs and |Rn(λ)| increase (S1b and S1b Fig) leading to a decrease of Lcr. The changes in Gs are significantly larger than those in |Rn(λ)|, implying that the static gain is the dominant factor that affects stability. The changes in |S(λ)| are negligible (S1c Fig). This is expected, because the frequency range we are interested in is much smaller than the cut-off frequency of the synaptic filter ω < ω3db. Thus, when the system operates in a dynamic regime in which single neuron responses have a higher gain the control domains shrink and the range of K values that stabilizes the system decreases.
Next, we investigated the interaction between the synaptic S(λ) and the control kernel M(λ). The amplitude responses for different kernels do not vary significantly (S2 Fig). Therefore, the important factor that influences stability is the phase difference or, alternatively, the difference Δd between the effective delays of the synaptic deff and the coupling kernel dc,eff. An optimal result is achieved if this difference vanishes (see methods) i.e. when
Δ d = d e f f - d c , e f f = 0
This point is illustrated for the case where dc = d+1ms (Fig 6c). These results show that DFC does not strongly depend on the shape but rather on the effective delay of the kernel.
Interestingly, the same control strategy can be used to induce or enhance rather than to suppress oscillations. Choosing appropriate control parameters to increase the effective coupling, i.e. selecting K to have the same sign as J (see methods), results in SI activity (Fig 7). This may be helpful for the treatment of symptoms in several pathological conditions that are characterized by impaired oscillations, e.g. gamma power decrease in schizophrenia [35]. Thus, DFC is a generic control approach and the control parameters (K,dc,M) can be tuned to quench or to enhance oscillatory activity, depending on the nature of aberrant activity.
Open-loop stimulation has been the main non-pharmacological approach to control the symptoms in a wide range of pathological conditions. It has been successful to some degree, but it often introduces clinical side-effects [36]. Moreover, it inherits the drawbacks from open-loop systems: (i) The stimulation profile is predetermined and is not adjusted to the clinically observed short-term fluctuations in the patients’ symptoms [37]. In addition, stimulation is continuously applied even though it may not be always necessary. (ii) The stimulation does not adapt to long-term changes of the system, e.g. structural alterations due to the progression of the disease. (iii) The operating point cannot be altered to deal with perturbations, caused, for instance, by a drift of the electrode lead [38]. (iv) External disturbances due to transient undesired signals are not being suppressed. In contrast to open-loop control, closed-loop control can by design deal with all these situations. For this reason there is a growing interest in investigating feedback-control both experimentally [2, 22–24] and theoretically [39]. The goal of the experimental work has been to demonstrate that closed-loop control is indeed effective, whereas the theoretical studies aimed at providing a deeper understanding of the underlying conditions and mechanisms.
Here, we provide a theory for DFC, a conceptually simple but powerful form of control [5, 6], applied to the control of stochastic SI oscillations in SNN. These oscillations are generic, they occur in many brain areas and in multiple conditions [40] and they emerge via a supercritical Hopf bifurcation [16]. Therefore, the control objective was specific: to counteract this bifurcation. We provide a mean-field approximation to estimate the DFC parameters and confirm the analytical predictions in numerical simulations in purely inhibitory and in coupled excitatory-inhibitory SNNs. DFC is effective in networks with arbitrary connectivity as long as the temporal instabilities, that is uniform oscillations, dominate the dynamics.
We used two control approaches, direct and differential control, and demonstrated that both schemes are effective in suppressing oscillations. Consistent with previous findings [5, 30], our results reveal that differential control has two main advantages over direct control. First, the control domains are enlarged, which renders the selection of control parameters an easier task. Larger control domains imply increased robustness of the system both to perturbations in the parameters and to disturbances. This means that neither small deviations from the nominal values of K, dc1, dc2 nor external signals compromise its stability. Second, in differential control the stimulation signal vanishes which translates to decreased power consumption. In clinical settings this is a highly desirable property and is, in fact, a basic requirement of any neuroprosthetic device.
The key advantage of the approach we presented here is that the system under control is being steered back towards its primary operating point (Fig 8). That is, DFC effectively decreases the synaptic coupling strength and, therefore, it counteracts the causes that originally led to the instability. This is obviously true only for the first-order statistics, because DFC does not counteract changes in the variance of the input that a random neuron in the network receives. Nevertheless, this is sufficient for the network to recover basic processing abilities both for rate and temporal coding schemes. Alternative approaches that rely on increased external noise are able to suppress oscillations [41], but they do not allow the network to perform any meaningful computations. We think that a similar explanation is valid also for the traditional open-loop DBS. The exact mechanisms of this type of DBS are still debated [42], but one of the reasons for the induced side-effects may be compromised information processing.
DFC suppresses oscillations in SNN not by decorrelating individual neurons, but rather by applying a common signal to all neurons that counteracts the mean input they receive from the network. Besides the many advantages described earlier, the utility of DFC lies in the fact that it is a very general control strategy, which does not depend qualitatively on lower level properties such as the specific coupling kernels of the connections. Neither does it depend qualitatively on the exact neuronal type (S3 Fig). Thus, DFC can in principle deal with more complex scenarios, e.g. heterogeneities in network and in neuronal properties. Nonetheless, the exact shapes of the neural and synaptic response functions do affect the system quantitatively and do modify the stability landscape. Thus, it is essential to have a good understanding of their precise contribution.
The results we presented provide a clear picture about how exactly the static gain of the neuron affects the size of the stable control domains. We also showed that the width rather than the shape of the control kernel affects the stability boundary. Further, we explained how the coupling strength and delay influence the overall stability landscape. Besides their theoretical value, these insights have a direct implication for the design of neuroprosthetic devices and are, therefore, of immediate practical and clinical relevance.
It is also important to address certain limitations in our approach. First, we assumed that we can directly affect the neuron membrane potential. This is currently only possible in animal models, where optogenetic tools are applied not only for superthreshold excitation/inhibition but also for subthreshold modulation [32]. In humans only electrical stimulation is realistic at the moment, therefore incorporation of volume conduction models [43] to describe the effects of electrical fields on individual neurons would be required [44]. Second, we used the population average of single neuron firing as our observable. Again, in more realistic settings the population rate has to be reconstructed from the available recordings, e.g. from multi-unit activity. This could be done, for instance, via the use of Kalman filters [45]. Third, our theoretical analysis was based on a mean-field approximation that ignores fluctuations in the input. Analytical and numerical results were largely in a good agreement, but additional work is necessary to specifically deal with the fluctuations in the activity. Last, we had access to the relevant parameters required for the tuning of the controller. This was intended, because our goal was not to simply demonstrate that the method is effective in large networks of spiking neurons, but also to provide an explanation in terms of mean-field theory of the precise mechanism. In real applications these parameters have to be estimated online from the recorded activity. The delay could be inferred from the frequency of the oscillatory activity. Inference of the coupling strength is less straightforward, but may still be feasible. Alternatively, once the delay is estimated, methods of adaptive tuning could be used to retrieve also the optimal control gain (S4 Fig). Tuning the controller is in general a difficult problem, even for open-loop DBS, and additional research in this direction is required.
Few studies have addressed the problem of suppressing oscillations in neural activity (see [39] for a detailed review). They are based (i) on population dynamics [46][15], (ii) on detailed single neuron descriptions [13], (iii) on simplified but computationally efficient models (e.g. Rulkov maps [47]), (iv) or on combinations thereof [14]. These approaches have their merits, but they come with limitations: (i) the parameters cannot be directly mapped to experimental measurable quantities (ii) it is not clear if the results scale to large networks of neurons.
The approach that we presented here is a trade-off between biophysical realism and analytical tractability. We used the LIF model, which captures single-neuron dynamics to a sufficient degree, while at the same time allows computationally efficient simulations of large networks. We applied DFC that was originally proposed in the context of chaotic systems as a method to control unstable periodic orbits [5]. DFC has been also used to control dynamics in extended media [48] and has been applied in different contexts as well [49]. It was later used to control coherence [50] and to suppress synchronous activity in networks in which the neurons themselves act as oscillators [10–12].
Here, we did not use simplified population dynamics or phase oscillators. Instead, we used spiking neurons that fire irregularly and are nevertheless able to generate oscillations. We used a single proportional control term to provide a proof of principle of the method and to be able to delineate the control domains semi-analytically. Alternative approaches, e.g. linear PI/PID control [4] or non-linear control schemes [11] are also possible and may improve performance, however, their theoretical analysis is less straightforward. We also inserted realistic descriptions of synaptic dynamics and, therefore, were able to explicitly study their contribution to stability. This allowed us to design an appropriate control kernel, which resulted in increased control domains. In addition, by using a mean-field theory that explicitly incorporates the synaptic and neuronal response functions we could study their contribution in a systematic way. The neuronal response function enabled us to investigate the influence of external and recurrent inputs and to relate them to experimentally measurable quantities. Indeed, as we showed above, the statistics of the mean field for activity states with very similar firing rate profiles may be significantly different affecting stability. Therefore, feasible measurements of the population activity can be directly used to characterize the operating point of the network and to fine-tune the control parameters to achieve the desired results.
Finally, the results presented here provide us with an understanding of the recent success of event-triggered control strategies. Event-triggered control can be placed between open-loop and continuous closed-loop control (e.g. DFC). Open-loop provides constant magnitude stimulation independent of the ongoing activity. Event-triggered approaches provide also constant magnitude stimulation, but only if a certain event occurs, for instance the power of beta oscillations crosses a certain threshold. Thus, the overall stimulation time and, therefore, the undesired stimulation side-effects are reduced. In DFC the stimulation side-effects are likely to be further reduced, because the stimulation amplitude is continuously adjusted to the ongoing activity. That is, no excessive stimulation is applied. This is particularly true for differential DFC, in which the stimulation amplitude vanishes over time.
We used DFC, a relatively simple form of control that includes only a proportional gain term, because it is still possible to analytically study the stability of the closed-loop control system. More sophisticated control strategies could further increase the performance of the system. They come, however, at the price of increasing the number of control parameters that have to be estimated and of increasing complexity precluding a formal proof of stability. Our approach spans multiple levels of analysis of neuronal dynamics, enabling an understanding of how the control stimulus interacts with both low-level synaptic and high-level properties of the population activity to influence stability. At the same time the complexity of the controller is low enough to be of practical relevance. Thus, here we have provided a general conceptual framework for future studies that address both theoretical and practical aspects of closed-loop control in neuronal systems.
We simulate networks of N LIF neurons randomly connected with a probability of ϵ = 0.1. Thus each neuron receives on average C = ϵN connections from other neurons in the network. For the purely inhibitory network we use N = NI and for the coupled excitatory-inhibitory case N = NE+NI. The subthreshold dynamics of a neuron i in the network is given by
τ m d v i ( t ) d t = ( v r e s t - v ( t ) ) + R m · I i , r e c ( t ) + R m · I i , e x t ( t ) (5)
where Rm is membrane resistance, τm is the membrane time constant and vrest is the resting potential. The recurrent input term
I i , r e c ( t ) = - ∑ j = 1 N J i j c i j ∑ k s ( t - t j k - d i j ) (6)
describes the total synaptic current arriving at the soma due to presynaptic spikes. cij are elements of the binary connectivity matrix. Each presynaptic spike causes a stereotypical postsynaptic current s(t) modeled as an α-function [51]
s ( t ) = t τ s e 1 - t t s H ( t ) (7)
where τs is the synaptic time constant and H(t) the Heaviside function.
The double sum in Eq 6 runs over all firing times t j k of all presynaptic neurons connected to neuron i. For all connections in the network we use the same synaptic coupling strength Jij = J/C, where C is the average in-degree and dij = d the transmission delay. The external input
I i , e x t ( t ) = μ + σ τ m η i ( t ) (8)
contains a mean term μ and a fluctuating term resulting from the Gaussian white noise ηi(t) that is uncorrelated from neuron to neuron with <ηi(t)> = 0 and <ηi(t)ηi(t′)> = δ(t−t′).
In the stable asynchronous state the population average of the firing rate is constant in time, r(t) = r0. The mean recurrent input that each neuron receives is therefore also constant and given by
μ r e c ( t ) = ⟨ J / C · ∑ c i j ∫ s ( τ ) ∑ k δ ( t j k - τ - d ) d τ ⟩ = J · r 0 · ∫ s ( τ ) d τ = μ r e c
similarly the variance of the recurrent input is
σ r e c 2 ( t ) = V a r [ J / C · ∑ c i j ∫ s ( τ ) ∑ k δ ( t j k - τ - d ) d τ ] = J 2 / C · r 0 · ∫ s 2 ( τ ) d τ = σ r e c 2
We study the stability of the asynchronous state following a linear perturbation approach [18]. A small oscillatory modulation of the stationary firing rate r(t) = r0+r1 e−λt with v1 ≪ 1 and λ = x+jω where ω is the modulation frequency leads to corresponding oscillation of the synaptic current
I 1 = - J · r 1 · e · τ s ( 1 + λ · τ s ) 2 e - λ d (9)
The firing rate in response to an oscillatory input is given by
r 1 = I 1 · r 0 σ ( 1 + λ τ m ) ∂ U ∂ y ( y t , λ ) - ∂ U ∂ y ( y r , λ ) U ( y t , λ ) - U ( y r , λ ) (10)
The function U is given in terms of combinations of hypergeometric functions
U ( y , λ ) = e y 2 Γ ( 1 + λ · τ m 2 ) F 1 - λ · τ m 2 , 1 2 , - y 2 + e y 2 Γ ( λ · τ m 2 ) F 1 - λ · τ m 2 , 3 2 , - y 2
In a recurrent network the modulation of the firing rate and the modulation of the synaptic input must be consistent. Combining Eqs (9) and (10) we get
1 = - J · r 0 · e · τ s e - λ d σ ( 1 + λ τ m ) ( 1 + λ τ s ) 2 ∂ U ∂ y ( y t , λ ) - ∂ U ∂ y ( y r , λ ) U ( y t , λ ) - U ( y r , λ )
which we write as
1 = J · R ( λ ) · S ( λ ) · e - λ d (11)
where the terms
R ( λ ) = 1 σ ( 1 + λ τ m ) ∂ U ∂ y ( y t , λ ) - ∂ U ∂ y ( y r , λ ) U ( y t , λ ) - U ( y r , λ )
and
S ( λ ) = e · τ s ( 1 + λ · τ s ) 2
describe the neuronal and synaptic response functions respectively. The negative sign of J is absorbed in the phase of S(λ).
The critical coupling values at which modes have marginal stability with frequency ωi can then simply be computed by
J i = 1 R ( ω i ) · S ( ω i )
The smallest value Jcr = min{Ji} is the critical coupling at which the first complex pair of eigenvalues crosses the imaginary axis and the system becomes unstable. In the case of the inhibitory network for μ = 14 mV and σ = 6 mV we have Jcr ≈ 115 mV. In the simulations we used for the coupling between two neurons i and j, Jij = 0.2mV thus the total coupling is J = C ⋅ Jij = 1000 ⋅ 0.2 mV = 200 mV >Jcr (Fig 2a–2d).
In the simulations we implement DFC by recording and stimulating all neurons in the network. The subthreshold dynamics of a neuron i with DFC is given by
τ m d v i ( t ) d t = ( v r e s t - v ( t ) ) + R m · I i , r e c ( t ) + R m · I i , e x t ( t ) + R m · I C ( t ) (12)
where IC(t) is the control input. Note that IC(t) is identical for all neurons in the network given by
I C ( t ) = K · m ( t ) ⋆ ( v ( t - d c ) ) (13)
where v(t) is the instantaneous population activity at time t and ⋆ denotes the convolution operation ( f ⋆ g ) ( t ) = ∫ - ∞ ∞ f ( t - τ ) g ( τ ) d τ . We used as control kernel m(t) a box function
m ( t ) = H ( t - a ) - H ( t - b )
where H(t) is the Heaviside function
H ( t ) = 0 , t < 0 m s 1 , 0 ≤ t ≤ 1 m s
Thus the control input IC(t) was updated in steps of 1ms.
|
10.1371/journal.pcbi.1006635 | Modelling the mechanics of exploration in larval Drosophila | The Drosophila larva executes a stereotypical exploratory routine that appears to consist of stochastic alternation between straight peristaltic crawling and reorientation events through lateral bending. We present a model of larval mechanics for axial and transverse motion over a planar substrate, and use it to develop a simple, reflexive neuromuscular model from physical principles. The mechanical model represents the midline of the larva as a set of point masses which interact with each other via damped translational and torsional springs, and with the environment via sliding friction forces. The neuromuscular model consists of: 1. segmentally localised reflexes that amplify axial compression in order to counteract frictive energy losses, and 2. long-range mutual inhibition between reflexes in distant segments, enabling overall motion of the model larva relative to its substrate. In the absence of damping and driving, the mechanical model produces axial travelling waves, lateral oscillations, and unpredictable, chaotic deformations. The neuromuscular model counteracts friction to recover these motion patterns, giving rise to forward and backward peristalsis in addition to turning. Our model produces spontaneous exploration, even though the nervous system has no intrinsic pattern generating or decision making ability, and neither senses nor drives bending motions. Ultimately, our model suggests a novel view of larval exploration as a deterministic superdiffusion process which is mechanistically grounded in the chaotic mechanics of the body. We discuss how this may provide new interpretations for existing observations at the level of tissue-scale activity patterns and neural circuitry, and provide some experimental predictions that would test the extent to which the mechanisms we present translate to the real larva.
| We investigate the relationship between brain, body and environment in the exploratory behaviour of fruitfly larva. A larva crawls forward by propagating a wave of compression through its segmented body, and changes its crawling direction by bending to one side or the other. We show first that a purely mechanical model of the larva’s body can produce travelling compression waves, sideways bending, and unpredictable, chaotic motions. For this body to locomote through its environment, it is necessary to add a neuromuscular system to counteract the loss of energy due to friction, and to limit the simultaneous compression of segments. These simple additions allow our model larva to generate life-like forward and backward crawling as well as spontaneous turns, which occur without any direct sensing or control of reorientation. The unpredictability inherent in the larva’s physics causes the model to explore its environment, despite the lack of any neural mechanism for rhythm generation or for deciding when to switch from crawling to turning. Our model thus demonstrates how understanding body mechanics can generate and simplify neurobiological hypotheses as to how behaviour arises.
| Exploratory search is a fundamental biological behaviour, observed in most phyla. It has consequently become a focus of investigation in a number of model species, such as larval Drosophila, in which neurogenetic methods can provide novel insights into the underlying mechanisms. However, appropriate consideration of biomechanics is needed to understand the control problem that the animal’s nervous system needs to solve.
When placed on a planar substrate (typically, an agar-coated petri dish), the Drosophila larva executes a stereotypical exploratory routine [1] which appears to consist of a series of straight runs punctuated by reorientation events [2]. Straight runs are produced by laterally symmetric peristaltic compression waves, which propagate along the larval body in the same direction as overall motion (i.e. posterior-anterior waves carry the larva forwards relative to the substrate, anterior-posterior waves carry the larva backwards) [3]. Reorientation is brought about by laterally asymmetric compression and expansion of the most anterior body segments of the larva, which causes the body axis of the larva to bend [2].
Peristaltic crawling and reorientation are commonly thought to constitute discrete behavioural states, driven by distinct motor programs [2]. In exploration, it is assumed, alternation between these states occurs stochastically, allowing the larva to search its environment through an unbiased random walk [1, 4–6]. The state transitions or direction and magnitude of turns can be biased by sensory input to produce taxis behaviours [4, 5, 7–13]. The neural circuits involved in producing the larval exploratory routine potentially lie within the ventral nerve cord (VNC), since silencing the synaptic communication within the brain and subesophageal ganglia (SOG) does not prevent substrate exploration [1]. Electrophysiological and optogenetic observations of fictive locomotion patterns within the isolated VNC [14, 15] support the prevailing hypothesis that the exploratory routine is primarily a result of a centrally generated motor pattern. As such, much recent work has focused on identifying and characterising the cells and circuits within the larval VNC [16–32]. However, behaviour rarely arises entirely from central mechanisms; sensory feedback and biomechanics often play a key role [33–35] including the potential introduction of stochasticity. Indeed, thermogenetic silencing of somatosensory feedback in the larva leads to severely retarded peristalsis [36] or complete paralysis [37, 38].
In line with the ethological distinctions drawn between runs and turns, computational modelling of the mechanisms underlying larval behaviour has so far focused on either peristaltic crawling or turning. An initial model based on neural populations described a possible circuit architecture and dynamics underlying the fictive peristaltic waves observed in the isolated ventral nerve cord [39]. A subsequent model described the production of peristaltic waves through interaction of sensory feedback with biomechanics, in the absence of any centrally generated motor output [40], in a manner similar to earlier models of wave propagation via purely sensory mechanisms in C. elegans [41, 42]. This model produced only forward locomotion as it incorporated strongly asymmetric substrate interaction. Recently, a model combining biomechanics, sensory feedback, and central pattern generation reproduced many features of real larval peristalsis [43]. However, this model only aimed to explain forward locomotion, and accordingly contained explicit symmetry-breaking elements in the form of posterior-anterior excitatory couplings between adjacent segments of the VNC, and posterior-anterior projections from proprioceptive sensory neurons in one segment into the next segment of the VNC. No biomechanical models of turning in the larva have yet been published, but the sensory control of reorientation behaviour has been explored in more abstract models [4, 5, 8, 11–13, 44]. No current model accounts for both peristalsis and reorientation behaviours, and no current model of peristalsis can account for both forward and backward locomotion without appealing to additional neural mechanisms.
Here we present a model of unbiased substrate exploration in the Drosophila larva that captures forward and backward peristalsis as well as reorientation behaviours. We provide a deterministic mathematical description of body mechanics coupled to a simple, reflexive nervous system. In contrast to previous models, our nervous system has no intrinsic pattern-generating ability [39, 43, 44], and does not explicitly encode discrete behavioural states or include any stochasticity [4, 5, 8, 11–13]. Nevertheless, the model is capable of producing apparently random “sequences” of crawling and reorientation behaviours, and is able to effectively explore in a two-dimensional space. We argue that the core of this behaviour lies in the chaotic mechanical dynamics of the body, which result from an energetic coupling of axial (“peristaltic”) and transverse (“turning”) motions.
Our choice not to explicitly model navigational decision-making and central pattern generation circuits is motivated by our desire to illustrate the powerful insights offered by focusing upon the mechanics of the body with which the nervous system interacts. The model neuromuscular system we have constructed is based upon simple physical arguments, yet ultimately bears a striking resemblance to known features of the larval nervous system. By starting from the mechanics of the body, and not assuming the existence of particular neural circuits, we are able to provide a new explanatory framework within which to re-interpret existing neurophysiological observations, including observations of central pattern generation within the larval VNC, and also provide unique predictions for future neurophysiological experiments.
In what follows, we first outline the key components and assumptions of our model of body mechanics. We then follow simple arguments to guide the construction of a neuromuscular model capable of producing power flow into the body, and motion of the body’s centre of mass relative to the substrate. Crucially, the neuromuscular model neither senses nor drives transverse motions. In analysing the behaviour of our model, we begin by focusing on the small-amplitude, energy-conservative behaviour of the body in the absence of frictive and driving forces. In this case, the motion of the body is quasiperiodic and decomposes into a set of energetically isolated axial travelling waves and transverse standing waves. Reintroducing friction and driving forces, we demonstrate the emergence of a pair of limit cycles corresponding to forward and backward peristaltic locomotion, with no differentiation of the neural activity for the two states. We then shift focus to the behaviour of the model at large amplitudes. In this case the axial and transverse motions of the body are energetically coupled, and the conservative motion becomes chaotic. The energetic coupling allows our neuromuscular model to indirectly drive transverse motion, producing chaotic body deformations capable of driving substrate exploration. Analysis of our model supports a view of larval exploration as an (anomalous) diffusion process grounded in the deterministic chaotic mechanics of the body.
To explore larval crawling and turning behaviours, we choose to describe the motion of the larval body axis (midline) in a plane parallel to the substrate (Fig 1, S1 Fig). The larval body is capable of more diverse motions including lifting/rearing [21], rolling [45], digging [46], self-righting / balancing, and denticle folding which we have recently observed to occur during peristalsis (S1 Video). However, while exploring flat surfaces, the larva displays fairly little out-of-plane motion (neither translation perpendicular to the substrate nor torsion around the body axis) and only small radial deformations [47]. Furthermore, the majority of ethological characterisations of larval exploration treat the animal as if it were executing purely planar motion [4, 6, 8–13, 48]. A planar model is thus a reasonable abstraction for the issues addressed in this paper, i.e., the generation of peristalsis, bending, and substrate exploration.
The segmented anatomy of the Drosophila larva allows us to focus our description of the midline to a set of N = 12 points in the cuticle, located at the boundaries between body segments and at the head and tail extremities. We assign each point an identical mass, and measure its position and velocity relative to a two dimensional cartesian coordinate frame fixed in the substrate (the laboratory or lab frame). We therefore have NDOF = 2N = 24 mechanical degrees of freedom. We note that our assumption of a uniform mass distribution along the midline is somewhat inaccurate, since thoracic segments are smaller than abdominal segments. However, simulations with non-uniform mass distribution give results which are qualitatively close to those presented here.
We assume that the larval body stores elastic energy in both axial compression/expansion and transverse bending, due to the presence of elastic proteins in the soft cuticle. We assume that energy is lost during motion due to viscous friction within the larva’s tissues and sliding friction between the body and the substrate. Sliding friction also allows shape changes (deformations) of the body to cause motion of the larva as a whole relative to the substrate (centre of mass motion).
Since the mechanical response of the larva’s tissues is yet to be experimentally determined, we assume a linear viscoelastic model. This is equivalent to placing linear (Hookean) translational and torsional springs in parallel with linear (Newtonian) dampers between the masses in the model, as shown in Fig 1, or to taking quadratic approximations to the elastic potential energy and viscous power loss (as in S1 Appendix). We note that the accuracy of the approximation may decrease for large deformations, in which nonlinear viscoelastic effects may become important.
As with larval tissue mechanics, there has been little experimental investigation of the forces acting between the larva and its environment. We therefore assume a simple anisotropic Coulomb sliding friction model, in which the magnitude of friction is independent of the speed of motion, but may in principle depend upon the direction of travel. This anisotropy could be thought of as representing the biased alignment of the larva’s denticle bands, or directional differences in vertical lifting or denticle folding motions which are not captured by our planar model. A mathematical formulation of our sliding friction model is given in S1 Appendix.
In addition to power losses due to friction, we also allow power flow due to muscle activation. For the sake of simplicity, we choose to allow only laterally symmetric muscle tensions. In this case, the musculature cannot directly cause bending of the midline, and can only explicitly drive axial motions. We will see later that even indirect driving of bending motion can lead to surprisingly complex behaviour, due to energetic coupling of axial and transverse degrees of freedom.
The choice to neglect asymmetric muscle tensions is made in order to simplify our model and provide a clearer illustration of the potential role of body mechanics in generating complex larval behaviour. We note that there is only one way for muscle activations to be symmetric—if we were to allow asymmetry we would have to specify exactly what form that asymmetry should take, and we have little empirical or theoretical grounds on which to do so. Nevertheless, there are some interesting cases which may be considered in passing—the presence of a constant torque about the model’s segment boundaries should cause a shift in the equilibrium posture towards a resting curved shape. The presence of torques which are a linear function of the local body bending angle or local angular velocity should shift the effective transverse stiffness or viscosity of the body S5 Appendix. In this sense the model presented here could be considered to already include the effect of asymmetric muscle tensions, they have simply been incorporated into the passive stiffness and viscosity of the body. We have recently developed an extension of the model presented here which uses a similar local reflex to modulate the body’s effective transverse viscosity in proportion to a stimulus input, allowing the model to exhibit taxis behaviour [49].
Finally, we model the internal coelomic fluid of the larva. Given the extremely small speed of the fluid motion compared to any reasonable approximation to the speed of sound in larval coelomic fluid, we can safely approximate the fluid flow as incompressible [50]. This would ordinarily require that the volume contained within the larval cuticle remain constant. However, since we are modelling only the motion of the midline and neglecting radial deformations, we constrain the total length of the larva to remain constant. We note that this constraint is not entirely accurate to the larva, as the total length of the animal has been observed to vary during locomotion [47]. Nevertheless, for the sake of simplicity we will continue with this constraint in place, noting that this approximation has been used with success in previous work focused on peristalsis [40, 43], and that there is experimental support for kinematic coupling via the internal fluid of the larva [3]. We note that we satisfy the incompressibility condition only approximately in some sections (Model behaviour—Conservative chaos, Dissipative chaotic deformations, and Deterministic exploration), by introducing an additional potential energy associated with the constraint, which produces an energetic barrier preventing large changes in the total length of the body (see S1 Appendix for details of this approximation along with specifics of the mathematical formulation of our mechanical model).
Note that in the absence of transverse bending, the total length constraint causes the head and tail extremities of the larva to become mechanically coupled and move in unison [40, 43]. The axial mechanics thus has periodic boundary conditions, and the most anterior (T1) and posterior (A8) segments of the larva may be considered adjacent. This means, for instance, that a compression wave travelling from tail to head will cause motion of the tail on termination at the head, thus initiating a new compression wave. This view also allows us to reason about what should happen if we relax the total length constraint. In particular, if we were to replace the direct coupling of head and tail by a viscoelastic coupling, representing the capacity for storage and dissipation of energy within the internal fluid or in radial expansion of the cuticle, the axial mechanics would still have periodic boundary conditions but would now have a step change in mechanical impedance. Waves hitting such “sudden” impedance boundaries in their transmission media will generally be partially transmitted (i.e. passing directly from head to tail in the larva) and partially reflected (i.e. changing direction and moving backwards from the head extremity), providing one possible cause of transitions between forward and backward locomotion in the animal. As will be seen, however, these transitions may occur even in the absence of an impedance discontinuity, and we will continue with the total length constraint in place in order to simplify our model.
Let us now consider how we should use muscle activity to produce locomotion. There are two basic requirements. First, we must have power flow into the body from the musculature, so that the effects of friction may be overcome and the larva will not tend towards its equilibrium configuration. Second, we must be able to produce a net force on the centre of mass of the larva, so that it can accelerate as a whole relative to the lab frame. Note that in this section, we motivate the neural circuits in the model from this purely functional point of view, but will present relevant biological evidence in the discussion.
To satisfy the first criterion, let us examine the flow of power into the body due to the action of the musculature
P = - ∑ i = 1 N - 1 b i MF i q ˙ i (1)
Here, qi describes the change in length of the i’th body segment away from its equilibrium length, q ˙ i is the rate of expansion of the i’th body segment, bi is a (positive) gain parameter, MFi is a (positive) dimensionless control variable representing muscle activation, and the product biMFi is the total axial tension across the i’th body segment. From this expression, it is clear that if we produce muscle tensions (MFi > 0) only while segments are shortening (q ˙ i < 0), we will always have positive power flow into the body (P > 0). This is a mathematical statement of the requirement for the larva’s muscles to function as motors during locomotion, rather than as springs, brakes, or struts [33].
A simple way to fulfil this condition is to introduce a segmentally localised reflex circuit (Fig 2, [40]). We place a single sensory neuron in each segment which activates when that segment is compressing (q ˙ i < 0). Each sensory neuron then projects an excitatory connection onto a local motor neuron, which in turn projects to a muscle fibre within the same segment. Assuming for now that there are no other influences on the motor neurons, so that sensory activation implies local motor neuron activation, segmental shortening will produce an immediate muscle tension serving to amplify compression of the segment and thus counteract frictive energy losses.
Let us now consider the second criterion for peristaltic locomotion. Assuming all segment boundaries are of equal mass, the force on the centre of mass of the larva is proportional to the sum of the forces acting on each segment boundary, i.e.
F C O M ∝ ∑ i = 1 N - 1 F s e g m e n t (2)
Newton’s third law tells us that any forces of interaction between segment boundaries (i.e. viscoelastic and muscle forces) must be of equal magnitude and opposite direction, so that they cancel in this summation and we are left only with contributions arising from substrate interaction. If the motion of the body is such that some number nf of segments move forward at a given time, against a frictional force −μf, while nb segments remain anchored or move backward, experiencing a frictional force μb, then the summation becomes
F C O M ∝ n b μ b - n f μ f (3)
In the limiting case of isotropic (direction-independent) substrate interaction we have μb = μf, and this expression tells us that the centre of mass will accelerate in the forward direction only when there are less segments moving forward than are moving backward or anchored to the substrate. Similarly, moving a small number of segments backward while the others remain anchored will result in backward acceleration of the centre of mass. Therefore, if the animal is to move relative to its substrate, it must ensure that only a limited number of its segments move in the overall direction of travel at a given time (indeed, this matches observations of the real larva [3, 22]). A more lengthy exposition of this requirement on limbless crawling behaviours can be found in [51].
We fulfil the requirement for a small number of moving segments by introducing mutually inhibitory interactions between the segmentally localised reflex circuits (Fig 2). We add a single inhibitory interneuron within each segment. When the sensory neuron within the local reflex activates, it excites this interneuron, which then strongly inhibits the motor neurons and inhibitory interneurons in non-adjacent segments, effectively turning off the local reflexes in distant neighbours. Adjacent segments do not inhibit each other in our model, allowing reflex activity to track mechanical disturbances as they propagate from one segment to the next. We comment on the plausibility of this feature of our model, given the experimental observation of nearest-neighbour inhibitory connections in the larval ventral nerve cord [28], in the discussion. Similarly, the head and tail segments do not inhibit each other, which permits peristaltic waves to be (mechanically) reinitiated at one extremity as they terminate at the other (see Discussion at the end of the previous subsection). This effectively introduces a ring-like topology into the neural model, matching our model of axial mechanics which couples head and tail motion through the total length constraint [40].
We now have a neuromuscular model consisting of four cell types repeated in each segment—sensory neurons, inhibitory interneurons, motor neurons, and muscle fibres. For the sake of simplicity we model all neurons as having a binary activation state governed by the algebraic relation
V i = { 1 ∑ j w j V j > θ i 0 otherwise (4)
where Vi is the activation of the i’th cell, θi is its activation threshold, Vj is the activation of the j’th presynaptic cell, and wj is the associated synaptic weight. Numerical values for the weights and thresholds used in our model are given in S1 Table, supplemental. Note that the muscle tension over a segment either vanishes (when the muscle fibre is in the inactive state) or has fixed magnitude bi (when the muscle fibre is activated by local sensory feedback). For this reason we refer to bi as the reflex gain.
Our choice to neglect neural dynamics is based on the large difference in timescales between the neural and mechanical dynamics. Typical motor neuron spiking occurs with a timescale on the order of 10−3 seconds. Spiking is observed to be significantly “averaged out” by the graded (non-spiking) muscle fibre responses, and respond on the order of ∼10−1 seconds to prolonged motor neuron spiking [52, 53]. During locomotion, segmental compressions are driven by several longitudinal muscle fibres activating simultaneously [3, 14, 29] in response to largely independent motor neuron populations [54, 55] which are unlikely to spike with identical timing. This spatial integration should further “mask” the effects of neural dynamics. Note that the lack of neural dynamics in our model immediately rules out central pattern generation. However, this does not prevent our model from producing complex, larva-like behaviour, and we consider how our model could account for observations of central pattern generation in the discussion.
To summarise, the neural model we have constructed can be seen as consisting of two parts, a segmentally repeating local reflex and a mutual inhibition circuit acting between non-adjacent reflexes. The local reflex is constructed so that muscles will act as motors, amplifying segmental compressions and counteracting friction. The mutual inhibition circuit couples distant reflexes to allow only localised amplification. By limiting the number of moving segments, this should ensure that the model larva can produce a net force on its centre of mass.
One of the advantages of grounding our model of larval exploration in the body’s physics is that we are now able to apply powerful analytical results from classical mechanics in order to understand the model’s behaviour. In this section we attempt to elucidate the naturally preferred motions of the larva by focusing our attention on the conservative mechanics of the body while neglecting friction forces, which would cause all motion to stop, and driving forces, which might impose a particular pattern of motion.
In this case, the general character of motion is specified by the Liouville-Arnold integrability theorem. This theorem asks us to look for a set of conserved quantities associated with a mechanical system, which remain unchanged as the system moves (energy, momentum, and angular momentum are examples of some commonly conserved quantities). If we can find a number of these quantities equal to the number of mechanical degrees of freedom in our model, then the theorem tells us that the motion of the body is integrable—it can be expressed analytically, and must be either periodic or quasiperiodic. If there are not enough conserved quantities, then the system is said to be nonintegrable. In this case the motion is much more complicated and will be chaotic for some initial conditions. These chaotic motions do not permit analytical expression and must be approximated through simulation.
In this section, we explicitly seek a case for which there is a “full set” of conserved quantities (we provide only major results here, for detailed derivations see S2 Appendix). We begin by restricting ourselves to considering only small deformations of the larval midline, in the case where all segments are of identical axial stiffness ka, transverse stiffness kt, mass m, and length l. Under these assumptions, the total mechanical energy of the body may be written
H ( x , y , p x , p y ) = 1 2 [ p x T p x + ω a 2 x T D 2 x ] + 1 2 [ p y T p y + ω t 2 y T D 4 y ] (5)
where x and y are vectors giving the displacement of each segment boundary along the body axis and perpendicular to the body axis, respectively, px and py give the translational momentum associated with each direction, D2 and D4 are difference matrices arising from a Taylor series expansion of our model’s potential energy (see S2 Appendix), and ω a = k a / m and ω t = k t / m l 2 are characteristic axial and transverse frequency scales. By making a linear change of coordinates {x, y, px, py} → {X, Y, pX, pY} to the eigenbasis of D2 and D4 (see S2 Appendix) this simplifies to
H ( X , Y , p X , p Y ) = ∑ i = 1 N - 1 1 2 [ p X , i 2 + ω a 2 λ a , i X i 2 ] + ∑ i = 1 N 1 2 [ p Y , i 2 + ω t 2 λ t , i Y i 2 ] (6)
where λa,i and λt,i are eigenvalues associated with the coordinate transformation. This expression is a sum of component energies, each of which is independently conserved. The Liouville-Arnold theorem immediately tells us that the motion of the body must be (quasi)periodic in the case of conservative small deformations. Indeed, the energy associated with each of the new coordinates Xi, Yi is in the form of a harmonic oscillator, and each coordinate executes pure sinusoidal oscillations. By transforming back to the original coordinates x, y we obtain a set of collective motions (modes) of the body which are energetically isolated and have a sinusoidal time dependence, corresponding to axial and transverse standing waves. We will refer to the Xi, Yi as modal coordinates since they describe the time dependence of each of the collective motions.
Each transverse standing wave corresponds to a periodic lateral oscillation of the body, with a unique frequency given by ω t , i = ω t λ t , i. We determined these frequencies numerically, along with the spatial components of the lowest frequency standing waves (Fig 3A). These can be seen to match the eigenmaggot shapes extracted from observations of unbiased larval behaviour [56].
The axial standing waves correspond to oscillating patterns of segmental compression and expansion. While each transverse standing wave had its own unique frequency of oscillation, the axial standing waves come in pairs with identical frequency but different spatial components—each member of the pair corresponds to a different spatial pattern of segmental compression/expansion spread across the body, but these patterns oscillate in time with the same frequency. We were able to analytically determine the frequency of oscillation of the i’th pair of axial standing waves to be
ω a , i = ω a λ a , i = 2 ω a | sin ( π i N - 1 ) | , i ∈ [ 0 , N / 2 - 1 ] (7)
The spatial components of the axial standing waves could also be determined analytically
x k , i = 1 N - 1 cos ( 2 π i k N - 1 ) , or x k , i = 1 N - 1 sin ( 2 π i k N - 1 ) , i ∈ [ 0 , N / 2 - 1 ] (8)
Where xk,i is the displacement of the k’th segment boundary for the i’th pair of standing waves. We plot the axial frequencies ωa,i and spatial components xk,i in Fig 3B.
The fact that the axial oscillation frequencies come in identical pairs allows us to combine the axial standing waves with a ±90° relative phase shift to form pairs of forward and backward travelling wave solutions (see S2 Appendix for the full derivation)
x k , i ( t ) = cos ( ω a , i t ± 2 π i k N - 1 ) , i ∈ [ 0 , N / 2 - 1 ] (9)
where xk,i(t) gives the displacement of the k’th segment boundary as a function of time for the i’th pair of travelling waves. The choice of a plus or minus sign corresponds to the choice between forward or backward wave propagation. These solutions correspond to propagating waves of segmental compression and expansion similar to those seen during larval peristalsis. We plot the lowest frequency pair of axial travelling wave solutions in Fig 3C, and directly visualise the synthesis of travelling wave solutions from standing wave solutions in S2 Video.
To summarise, in this section we have shown that for the case of conservative, small oscillations, the motion of the body may be decomposed into a combination of transverse standing waves and axial travelling waves. This is of clear relevance to understanding the behaviour of the larva, which moves across its substrate by means of axial peristaltic waves while reorienting using lateral oscillations. Indeed, the transverse modes of oscillation that we have derived here match principal components of bending computed from actual larval behaviour [56]. Our results can be interpreted as providing a physical basis for these observations—the principal components extracted from real larval data correspond to a “natural” coordinate basis that is grounded in the animal’s mechanics. Furthermore, the proportion of postural variance explained by each principal component of the experimental data decreases with increasing modal frequency in our model (and thus increasing energy). We can therefore help to explain the observed ordering of principal components, as this corresponds to the larva “preferring” to occupy low-frequency, low-energy modes during most of its behaviour. We comment further on this observation in S3 Appendix in the context of axial modes. We will now focus on the small-amplitude motion of the body in the presence of energy dissipation due to friction and driving forces.
Reintroducing friction will clearly lead the motions described above to eventually terminate due to energy dissipation, unless opposed by transfer of power. In a previous section (Models—Neuromuscular system, see also S1 Appendix), we introduced a neuromuscular system to produce power flow into the body, but as described, it can only directly transfer power into the axial degrees of freedom. In the small deformation model we have just analysed, the axial and transverse degrees of freedom are energetically decoupled. It follows that transverse friction is unopposed and any transverse motion must eventually terminate in the case of small deformations. In this section we will therefore focus only on the axial degrees of freedom, which correspond to the peristaltic locomotion of the larva.
In Fig 4, we show the effect of coupling our neuromuscular model to the axial mechanics. For small reflex gains, the power flow into the body from the musculature is too low to effectively counteract frictive losses and the larva tends towards its passive equilibrium state, in which there is complete absence of motion. However, increasing reflex gain past a certain critical value leads to the emergence of long-term behaviours in which the larva remains in motion, away from its passive equilibrium. These motions correspond to forward and backward locomotion, driven by forward and backward propagating compression waves (see below), as predicted from our earlier description of the conservative motions of the body, and as observed in the real larva [3]. The qualitative changes in behaviour that occur as reflex gain is varied are depicted in Fig 4A, where we have measured the long-term centre of mass momentum of the larva, along with the long-term relative phase of the lowest frequency modal coordinates.
The exact value of reflex gain required for onset of locomotion depends on the particular mechanical parameters used in our model (see S2 Table for parameters used in Fig 4). In principle, this bifurcation point of the dynamics should be amenable to analytical investigation by studying the stability of the linearised model dynamics around the passive equilibrium state [57, 58]. In practise, however, the presence of hard nonlinearities in the sliding friction model makes such an approximation inaccurate.
For low reflex gains the centre of mass momentum tends to 0 as the body comes to a stop and enters a passive equilibrium state. The relative phase of the low frequency modal coordinates tends to either 0 or 180 degrees, which also corresponds to a loss of momentum. For larger values of reflex gain, the total momentum is either positive, zero, or negative. Positive and negative total momentum represent forward and backward locomotion, respectively, while zero momentum corresponds to two unstable cases which we discuss below. The relative phase of the lowest frequency modal coordinates tends to ±90° at high reflex gains, corresponding to the presence of forward- or backward-propagating compression waves (see previous section). As in the larva [1, 3], forward-propagating waves drive forward locomotion while backward-propagating waves drive backward locomotion (Fig 4B).
We believe that these behaviours arise from the production of a pair of limit cycle attractors in the system’s phase space, which we visualise in Fig 4C by projecting the system state onto the lowest frequency pair of axial modes, and plotting the associated modal coordinates along with the centre of mass momentum. Since wave motion implies that pairs of modal coordinates should perform pure sinusoidal oscillations with equal amplitude and frequency, and a ±90° relative phase shift (see previous section and S2 Appendix), the travelling wave trajectories of the system become circles in this coordinate system (see discussion of Lissajous figures, [59]). Forward and backward locomotion can then be distinguished by the momentum of the centre of mass.
In this model, the speed of forward and backward locomotion are equal for a fixed value of reflex gain, while in the real larva the speeds are known to differ [15]. We comment on some possible explanations for this difference in the discussion.
In S2 Fig we show the neural state of the model larva during forwards and backwards locomotion. As expected given our previous exposition, we observe waves of activity in the nervous system which track the mechanical waves propagating through the body. Our sensory neurons also show a second, brief period of activation following propagation of the mechanical wave caused by a slight compression that occurs as segments return to equilibrium. This activity is “cancelled out” by the mutual inhibition circuit, so that motor neurons do not exhibit a secondary burst of activity.
The larva has zero long-term total momentum in the presence of large reflex gain in only two cases, both of which are highly unstable. First, if we start the larva so that it is already in its passive equilibrium state, so that no relative motion of segment boundaries occurs, it is obvious that there will be no activation of local reflexes and the larva will not spontaneously move out of equilibrium. The second case corresponds to a pure axial standing wave. In this case the larva is divided into two regions by nodal points where the axial displacement is zero. During the first half-cycle of the standing wave, one region will experience compression while the other experiences expansion, and in the second half-cycle these roles will reverse. The neural circuit we have constructed can amplify compression during both half-cycles since they are separated by a configuration in which no compression or expansion occurs, and this allows the entire nervous system to become inactive and “reset”. Since these behaviours are extremely unstable and require very specific initial conditions to be observed, we have not visualised them here.
While the mutually inhibitory connections in our model are not required for the propagation of locomotor waves, which can be maintained entirely by local reflexes [40], these connections do greatly enhance stability. In the absence of the mutual inhibition circuit, small mechanical disturbances in any stationary body segments can be amplified, giving rise to multiple compression waves which travel through the body simultaneously. If this instability produces an equal number of forward and backward moving segments then overall motion of the larva relative to the substrate will stop, in line with the argument presented earlier. We have also observed that roughly symmetrical substrate interaction is required to produce both forward and backward locomotion in our model. If friction is too strongly anisotropic, then locomotion can only occur in one direction regardless of the direction of wave propagation.
It is worth noting that the axial model presented in this section does display discrete behavioural states. However, there are no explicit decisions regarding which behavioural states to enter, since the particular neural states occupied during forwards and backwards locomotion are indistinguishable. Rather, both the apparent decision and the eventual direction of travel can only be understood by examining the dynamics and mechanical state of the body.
Having successfully produced peristaltic locomotion using our model, we will now turn our attention to the larva’s turning behaviours. As before, we will start from physical principles. In a previous section (Results—Conservative axial compression waves and transverse oscillations) we showed that, for the case of conservative small oscillations, transverse motions of the body were energetically decoupled from axial motions, and could be decomposed into a set of periodic standing waves. We will first extend our previous analysis to the case of energy-conservative, large amplitude motions in the absence of damping and driving; and then in the following section consider the impact of energy dissipation and the neuromuscular system on transverse motion.
To keep our presentation simple and allow visualisation of model trajectories, we will focus on a reduced number of the mechanical degrees of freedom. Namely, we will examine the bending angle ϕ and axial stretch q of the head segment (Fig 5A). We introduce an amplitude parameter ϵ by making the substitutions q → ϵq and ϕ → ϵϕ, so that the total mechanical energy of the head may be written in nondimensional form as (see S4 Appendix)
H = 1 2 [ p q 2 + 1 ( 1 + ϵ q ) 2 p ϕ 2 + q 2 + λ 2 ϕ 2 ] (10)
where pq, pϕ are the radial and angular momentum of the head mass, and we have scaled the time axis of the model so that the natural frequency of axial oscillation is unity. The parameter λ = ωt/ωa = kt/kal2 is the ratio of transverse and axial frequencies.
In the case of small oscillations, i.e. ϵ → 0, the mechanical energy reduces to the simpler expression
H = 1 2 [ p q 2 + q 2 ] + 1 2 [ p ϕ 2 + λ 2 ϕ 2 ] (11)
which is clearly a sum of independent axial and transverse energies. These energies are individually conserved, so that the Liouville-Arnold theorem applies, and the motion of the head is integrable and (quasi)periodic. This is in clear agreement with our earlier results. For the more general case of large amplitude motion (ϵ > 0), we may have in principle only a single conserved quantity—the total energy of the system. Indeed, it should be clear from the presence of a “mixed” term in the mechanical energy (Eq 10) that the axial and transverse motions are energetically coupled at large amplitudes, so that the individual energies associated with each motion are no longer independently conserved. Given that we have only one conserved quantity for a two degree of freedom system, we can no longer rely on the Liouville-Arnold theorem to prove (quasi)periodicity of the motion, and must accept that the system’s behaviour may be chaotic.
To investigate this possibility further, we first note that conservation of energy implies that flow within the four dimensional phase space must be constrained to lie on the energy surface given implicitly by the relation H(q, ϕ, pq, pϕ) = E. Therefore, given a particular value E for the total energy, the system dynamics becomes three dimensional. This allows us to visualise the behaviour of the system by plotting the points at which trajectories intersect a two-dimensional Poincare section [57, 58]. We define our Poincare section by the condition that the angular momentum vanishes pϕ = 0 (equivalently, angular velocity vanishes dϕ/dt = 0), and plot successive crossings of the section as points in the q, ϕ plane. This allows us to intuitively interpret points in the Poincare section as configurations of the head at successive turning points (extrema) in the transverse motion (Fig 5B).
In what follows, we set the total energy to be E = 1 2 so that when ϵ = 1 we can in principle obtain complete compression of the head (q = −1). We choose to set λ = e 6 ≈ 0 . 45, giving an irrational frequency ratio. This loosely matches observations of the real larva in which the frequency of transverse oscillations is approximately half that of axial oscillations but does not satisfy an exact (rational) resonance relationship [44]. The results we obtain with these parameters do not differ much from results for other energies or other frequency ratios, including resonant relationships.
Poincare plots for the cases ϵ → 0 and ϵ ∈ { 1 3 , 2 3 , 1 } are shown in Fig 6. When ϵ → 0 (Fig 6A), conservation of transverse energy implies that the turning points of the transverse motion must remain constant. The fact that the frequency ratio λ is irrational implies that the overall motion is quasiperiodic, and the values of q obtained at successive transverse turning points should not repeat. In accordance with these observations, the Poincare section for ϵ → 0 consists of a series of verticle lines (Fig 6Ai). For ϵ = 1 3 the Poincare plot becomes distorted, but the majority of trajectories still trace out one-dimensional curves in the section (Fig 6Bi), which is indicative of persistent quasiperiodic behaviour. At ϵ = 2 3 the Poincare plot appears qualitatively different. There is now a large region of what appears to be “noise”, indicating that the configuration of the head at successive transverse turning points has become unpredictable. This is a clear signature of deterministic chaos. There are, however, several regions of the section indicative of (quasi)periodic behaviour. These appear as one-dimensional curves or discrete points in the Poincare section (Fig 7Ai). At ϵ = 1, the region of the Poincare plot occupied by chaos has expanded, although there still appear to be some regions corresponding to (quasi)periodic behaviour (Fig 7Bi). These results qualitatively agree with the Kolmogorov-Arnold-Moser theorem [59], which tells us that quasiperiodic behaviour should persist under small nonintegrable (chaotic) perturbations of an integrable Hamiltonian, and that the region of phase space corresponding to chaotic behaviour should grow with the perturbation size (in our case, the perturbation size corresponds to the amplitude of motion ϵ). However, our model as presented here does not formally meet the requirements of this theorem (see S4 Appendix).
Analysis by Poincare section provides an invaluable method to determine the character of overall system behaviour by direct visualisation of trajectories, but cannot be applied to higher dimensional systems. This is problematic, since we would like to assess the existence of chaos beyond our reduced model of the larva’s head. We therefore deployed a series of other methods which are possibly less reliable than the method of Poincare section but can be applied equally well to higher dimensional systems. These included estimation of the maximal Lyapunov characteristic exponent (MLCE) for the dynamics along with calculation of the power spectrum and autocorrelation of internal variables [57, 58, 60]. The MLCE can be thought of as quantifying the rate of separation of nearby trajectories, or, equivalently, the rate at which information is generated by the system being analysed [61]. A positive MLCE is generally considered a good indicator of chaotic behaviour. The power spectrum of a periodic or quasiperiodic process should consist of a “clean” set of discriminable peaks, whereas that of a chaotic process should appear “noisy” and contain power across a wide range of frequencies. Meanwhile, the autocorrelation of a periodic or quasiperiodic process should show a strong oscillatory component with an envelope that decays linearly with time, while that of a chaotic process should show a much quicker decay, similar to a coloured noise process. In Fig 6 we plot these measures at each value of ϵ, for a trajectory starting with initial conditions indicated on the corresponding Poincare plot by a filled black circle. These measures confirm increasingly chaotic behaviour as the amplitude ϵ increases, in agreement with our Poincare analysis. In Fig 8 we show a solution including all degrees of freedom in our conservative mechanical model (i.e. not just those of the head). The trajectory of individual segments relative to the substrate appears qualitatively irregular, while the indirect measures we introduced above (MLCE, power spectrum, autocorrelation) are all indicative of deterministic chaotic behaviour.
We will now reintroduce dissipative effects into our model of larval motion in the plane. We previously saw that conservative mechanics predicted axial travelling waves and transverse oscillations. These were lost in the presence of friction, but the axial travelling waves could be recovered with the addition of a neuromuscular system designed to selectively counteract frictive effects. We have now seen that conservative mechanics predicts chaotic planar motion. Although our neuromuscular model transfers power only into the axial degrees of freedom, we recall from the previous section that axial and transverse motions are energetically coupled at large amplitudes. We therefore tentatively expect that we may be able to recover the complete chaotic planar motion without requiring any additional mechanism for direct neuromuscular power transfer into tranverse motion.
To investigate whether our dissipative planar model shows chaotic behaviour, we ran n = 1000 simulations starting from almost identical initial conditions (euclidean distance between initial mechanical state vectors < 10−7, with no initial neural activity) and identical parameters (see S5 Table). We can indeed observe that the simulated larva perform peristalsis with interspersed bending of the body (turns), and that the fully deterministic system produces apparently random turning such that the simulations rapidly diverge (S3 Video). Since most working definitions of chaos require strictly bounded dynamics, we here restrict our analysis to the coordinates describing deformation of the body (segmental stretches and bending angles), ignoring motions of, or overall rotations about, the centre of mass (i.e., the trajectory through space of the body, which we will analyse in the following section).
Qualitatively, the deformations of the large amplitude dissipative model appear irregular (Fig 9A). However, there are persistent features reminiscent of the ordered small-amplitude behaviour described in previous sections. In particular, there are clear axial travelling waves and lateral oscillations. However, whereas forward- and backward-propagating axial waves previously corresponded to stable limit behaviours, the large amplitude system appears to go through occasional “transitions” between these behaviours. In addition, apparently spontaneous large bends appear occasionally, suggesting that the neuromuscular system is effectively driving transverse motion.
The irregularity of the axial motion is reflected in the pattern of sensory neuron activation (S3 Fig). However, the mutual inhibitory interactions in our model act to filter this input, allowing only a small window of excitability within the central nervous system. As a result, interneuron and motor neuron activity appears fairly ordered, tracking and reinforcing axial compression waves.
We used four measures to assess whether our qualitative observation of irregular motion betrays the existence of deterministic chaos. First, we analysed the power spectrum of individual cooordinates (Fig 9B). The power spectra of all degrees of freedom showed consistent harmonic peaks along with a strong “noisy” non-harmonic component, a pattern consistent with chaos and incommensurate with (quasi)periodicity (Fig 9B shows data for head bending angle and stretch Next, we computed the autocorrelation of the same degrees of freedom. The autocorrelations of all degrees of freedom showed a periodic pattern with a peak at 0 seconds time lag followed by a rapid decay (Fig 9C). This is characteristic of oscillatory chaotic behaviour, and the exponential loss of information regarding initial conditions that chaotic systems display. We then estimated the correlation dimension (Dc) of the limit set of our model’s dynamics. Note that we did not apply this measure to the conservative models in the previous section since the Liouville theorem rules out attracting limit sets for these systems. The distribution of correlation dimension estimates for our dissipative system across all 1000 trials is shown in Fig 9D. Estimates were clustered around ∼ 3.5 (median dimension 3.46), with 93% of estimates lying in the range [3–4]. These results are suggestive of a limit set with fractal dimension, which is a signature of “strange” chaotic attractors. Finally, we computed an estimate of the maximal Lyapunov characteristic exponent (MLCE). As in the previous section, we estimated the MLCE for our system to be positive (∼ 13textrmbitss−1, Fig 9D), a very strong indicator of chaotic behaviour. All of these results point to the presence of a chaotic dynamical regime in our dissipative large amplitude model.
As the coupled biomechanical and neuromuscular system produces both forward and backward peristalsis and lateral oscillations, each simulated larva will trace out a 2D trajectory over time. As shown in Fig 10A, the chaotic deformations characterised in the previous section caused the larvae to disperse across their two-dimensional substrate, following paths reminiscent of the real animal’s exploratory behaviour.
To characterise the trajectories of our model, we first investigated them at a global level, based on the centre of mass (COM) trajectory of each simulated larva, computing the tortuosity and fractal dimension of the paths (Fig 10B) [62]. We defined our tortuosity measure as
T = 1 - D L (12)
where D is the net displacement of the COM between initial and final times, and L is the total distance travelled by the COM along its path. Note that if the COM travels in a straight line between initial and final times we will have D = L so that T = 0. In the limit L → ∞, corresponding to the COM taking an extremely long path between its initial and final states, we have D L → 0 so that T → 1. We calculated the mean tortuosity of our COM trajectories to be 〈T〉 = 0.43, with a variance of 〈(T − 〈T〉〉)2 = 0.05. The lowest (highest) tortuosity observed was T = 0.05 (T = 0.95).
We estimated the fractal dimension Db of the COM trajectories using a box-counting algorithm. The minimum expected dimension Db = 1 would correspond to curvilinear paths (e.g. straight line or circular paths), while the maximum expected dimension of Db = 2 corresponds to plane-filling paths (e.g. brownian motion). We calculated the mean dimension of the COM trajectories to be 〈Db〉 = 1.37, with variance 〈(Db − 〈Db〉)2〉 = 0.01. The lowest (highest) path dimension observed was Db = 1.17 (Db = 1.95). We have plotted the tortuosity and fractal dimension of every path in Fig 10B. These results show that the trajectories of the model differed markedly from straight lines (tortuosity T > 0 and box-counting dimension Db > 1), and displayed a good ability to cover the planar substrate (box-counting dimension 1 < Db < 2). We also note that our COM trajectories display the power-law relationship between angular speed and curvature reported by [63], with a scaling exponent (β ≈ 0.8) falling within the range reported for freely exploring larvae (S4 Fig).
We next investigated the rate at which the simulated larvae explored their environment. To do this, we calculated the mean-squared displacement (MSD) of the COM over time (Fig 10C). This is a standard measure used to characterise diffusion processes, and is defined as
⟨ d 2 ⟩ ( t ) = 1 n ∑ i = 1 n ( R i ( t ) - R i ( 0 ) ) 2 (13)
where Ri(t) is the position of the i’th larva’s COM at time t and n = 1000 is the number of trials being averaged over. We observed an initial transient, lasting on the order of 10 seconds, during which the MSD grew as ∼ t2 (blue line, S5 Fig), after which growth slowed and tended to ∼ t (linear fit for t > 80 seconds shown by red line, Fig 10C and S5 Fig, r2 = 0.99, diffusion constant D = 144segs2s−1). The initial transient was not due to our particular initial conditions, since it remained even after discarding > 60s of initial data. These results therefore tell us that, although on long timescales our model appears to execute standard Fick diffusion or a Brownian random walk (linear growth of MSD), on short timescales the model’s behaviour is superdiffusive (approximately quadratic growth of MSD) [64, 65]. This is in good agreement with observations of the real larva [6, 48]. The superdiffusive behaviour of the larva was previously explained in terms of a persistent random walk [6], in which the larva’s current and previous headings are highly correlated during straight runs so that the animal follows an approximately ballistic trajectory on short timescales. We believe that persistence effects arise in our model due to the finite time required for the deterministic chaotic dynamics to destroy information regarding initial conditions.
We next calculated some other standard measures found in the larva literature. To do so, we built a two-segment representation of each simulated larva by drawing vectors from the tail extremity to the A2-A3 segment boundary (the tail vector, T), and from the A2-A3 boundary to the head extremity (the head vector, H). We then defined the body bend, θ, to be the angle between tail and head vectors, θ = atan(Hy/Hx) − atan(Ty/Tx). The head angular velocity ν was computed as the cross-product of the head vector and the head extremity’s translational velocity r ˙ head measured relative to the lab frame, ν = H × r ˙ head, while the tail speed v was taken to be the magnitude of the tail extremity’s translational velocity r ˙ tail measured relative to the lab frame, v = r ˙ tail · r ˙ tail. The tail speed and head angular velocity both show a strong oscillatory component, which can be seen in the time and frequency domains (S6 Fig). The power spectra of these variables contains considerable “noise” over a wide spread of frequencies, in accordance with the results of the previous section. The distribution of tail speeds for our model is bimodal, similar to that of the real larva [44]. The body bend angle was observed to be symmetrically distributed (Fig 10D), with roughly zero mean (〈θ〉 = 0.005), small variance (〈(θ − 〈θ〉)2〉 = 0.13), slight positive skew (SK(θ) = 0.23), and high excess kurtosis KU(θ) = 7.9. The kurtosis of our data precludes a good fit by the von Mises distribution (maximum likelihood estimate shown by red line in Fig 10D). Our data appears to be better fitted by a wrapped Cauchy distribution, though this overestimates the central tendency of our data (maximum likelihood estimate shown by blue line in Fig 10D). The high excess kurtosis of the body bend distribution gives a quantitative measure of the rare large bending events mentioned at the beginning of the previous section, and qualitatively matches experimentally observed distributions of real larval bends [44, 66]. Our model also reproduces the observed overall speed of larval locomotion (median model speed = 0.26 body lengths s−1, real larval range ∼ 0.1 − −0.35 body lengths s−1), the turn rate (median model turn rate = 2.08 turns min−1, real larval range ∼ 0 − −4.5 turn min−1, threshold body bend for turn classification = 30deg to match relevant literature), and the relative distance gained during free locomotion (median model distance gained = 0.14 body lengths s−1, real larval range ∼ 0 − −0.2), with our results being more consistent with observations of third instar than first instar larvae [66].
Finally, we computed a run-length distribution by setting a threshold body bend angle θturn = 20° (as in [13]) and calculating the length of time between successive crossings of this threshold. The distribution we obtained appears approximately linear on a log-linear plot (Fig 10E, linear fit r2 = 0.99 with slope λ = −0.075), and is better fit by an exponential than a power law distribution (maximum likelihood estimates, log likelihood ratio = 5281, p < 0.01). Together with our observation of asymptotic linear growth of MSD, the exponential distribution of run lengths suggests that the model can be considered to be effectively memoryless on long timescales [65]. This again agrees with the observed rapid loss of information from the system due to its chaotic dynamics, as quantified by the Lyapunov exponent and autocorrelation analysis of the previous section.
Ultimately, the analysis of our model supports a view of the larval exploratory routine as an (anomalous) diffusion process arising from the deterministic chaotic dynamics of the body. The model nervous system functions purely to recover these dynamics from the effects of frictive energy dissipation, and to ensure centre of mass motion, rather than explicitly directing exploration.
The intrinsic capabilities of an organism’s body determine the field of possibilities that neural circuits for behaviour can exploit. Here, by focusing first on the biomechanics of Drosophila larva, we find that its body already contains an inherent exploratory routine. This is demonstrated through a combined biomechanical and neuromuscular model that is the first to be able to generate both forward and backward peristalsis and turning, allowing 2D motion in the plane to be simulated. We show that, in the absence of friction, the body’s conservative mechanics alone supports both axial travelling waves and transverse standing waves. These are energetically coupled at larger amplitudes, such that no driving, sensing, or control of body bend is required for the system to start producing spontaneous coordinated bending motions. Frictional losses can be recovered, to maintain axial waves, by a neuromuscular system consisting of only simple local sensorimotor reflexes and long-range inhibitory interactions. This is sufficient to produce emergent crawling, reversal and turning that resembles larval exploratory behaviour, and which is chaotic in nature. At a population level, we observe a deterministic anomalous diffusion process in which an initial superdiffusive transient evolves towards asymptotic Fickian/Brownian diffusion, matching observations of real larvae [6, 48]. We therefore propose that the role of biomechanical feedback in Drosophila larvae goes beyond the periphery of basic neuromuscular rhythms [40, 43], to provide the essential “higher order” dynamics on which exploratory behaviour is grounded.
Most existing models of larval exploration abstract away from the mechanics underlying the production of runs and reorientations [4–6, 8, 11–13]. The larva is often described as executing a stochastic decision-making process which determines which state (running or turning) should be occupied, and when to initiate a change of behavioural state. In contrast, our model produces the entire exploratory routine without making any decisions (the transverse motion is neither sensed nor driven by the nervous system) nor introducing any stochastic process (neural or otherwise). Similarly, transient “switching” is seen to occur between forward and backward peristalsis even though there is no neural encoding or control of the direction of wave propagation. In other words, the body dynamics generate the basis of a chaotic exploratory routine which only needs to be amplified by the neural circuitry, making the search for underlying stochastic or state switching circuitry superfluous for this behaviour.
The work presented here also stands in contrast to previous models of larval peristalsis [39, 43] and the prevailing hypotheses regarding this phenomenon [15, 67] by eschewing any role for intrinsic neural dynamics. Such stereotyped and rhythmic locomotion is widely assumed to be the signature of a central pattern generator (CPG), that is, a neural circuit that intrinsically generates a rhythmic output, and thus determines a particular mechanical trajectory to be followed by the body [68–70]. However it is recognised that systems vary in the degree to which coordinated behaviour is independent of biomechanical and sensory feedback [70]. Indeed, evidence from studies employing genetic manipulations to disrupt sensory neuron input suggest that proprioceptive feedback is necessary for correct larval locomotive patterns [16, 36–38, 71]; although in some cases coordinated waves of forward and backward peristalsis can be produced, in both intact [16, 36, 71] and isolated VNC preparations [14, 15], these are reported as abnormal with the most evident defects being time-dilation [15, 36, 71] and abnormal frequency in polarity changes [71].
In fact, our intent is not to adjudicate between the roles of intrinsically generated activation sequences vs. biomechanical feedback in this system, but rather to note that we should expect neural circuits of locomotion to adhere to the dynamical modes of the associated body, instead of working against them. Thus it should be unsurprising if these dynamics also exist (potentially in a latent form) in the neural circuitry. For example, a simple modification of the neural circuit presented here could produce instrinsic ‘peristaltic’ waves. Recall that the long-range global inhibition pattern in our model treats head and tail segments as ‘neighbouring’ nodes (see Models—Neuromuscular system). If local constant input or recurrent feedback were added to each segment, the circuit would then resemble a ring attractor [72–74] and a stable activity bump would be formed. Breaking the forward/backward symmetry of the circuit, e.g., by introducing asymmetric nearest-neighbour excitatory connections [75], would cause the activity bump to move along the network, giving rise to intrinsic travelling waves. This would complement any mechanical compression waves travelling through the body, i.e., remain consistent with the principles set out in this paper. Would such a network be a CPG? The answer is unclear. On the one hand, it would show spontaneous rhythmic activity in the absence of sensory input. On the other, sensory feedback would do much more than simply correct deviations from the CPG output or provide a “mission accomplished” signal [36]. Rather, feedback would play a crucial role in orchestrating motor output to ensure power flow into the body, consistently with its dynamical modes.
It is important to note that the emergence of rhythmic peristalsis and spontaneous turns in our model is not strongly dependent on the specific assumptions made in our mechanical abstraction. For example, the observation for small amplitude motions of sinusoidal axial travelling waves, along with transverse standing waves whose shapes match the experimentally observed “eigenmaggots” [56], is a direct result of the second-order Taylor series approximation of the model Hamiltonian (S2 Appendix). The small-amplitude model is thus non-unique, since many different mechanical models could have identical second-order approximations. Similarly, we expect that the deterministic chaotic behaviour derived from our conservative model for large amplitude motions will hold for other models of the larval body, given that it is conjectured that the majority of Hamiltonian systems are nonintegrable. This may also mean that our results can be applied to other animals with body morphologies and mechanics similar to the Drosophila larva.
In our model we constrain the total length of the larva to be constant. This constraint is intended to represent the fact that there is minimal observable radial deformation of the larva’s body during behaviour, yet its body is filled with fluid which should conserve volume. We were further motivated by the experimental observation of “visceral pistoning” [3] in which the head and tail extremities of the larva appear to be mechanically coupled via the coelomic fluid during peristalsis. However, the total length of the real larva is known to change during behaviour [47], and it is therefore important to consider the effect of weakening the length constraint in our model. When restricted to small-amplitude motion, the total length constraint appears as periodic boundary conditions in the axial mechanics, allowing waves of compression to propagate from head to tail and vice versa. In the complete absence of the length constraint, these waves will instead be reflected back from the head and tail extremities, leading to alternating forward and backward waves. Alternatively, replacing the constraint with a simple linear viscoelastic model to represent energy storage and dissipation within the internal fluid and in radial cuticle deformation leads to the presence of a new mechanical impedance between the head and tail. It is well known that sudden impedance changes in wave transmission media lead to simultaneous reflection and transmission of waves—in our model, this means that some amount of the axial compression wave will be transmitted between head and tail while some will be reflected. Since our neural model cannot sustain two peristaltic waves concurrently due to the presence of mutual inhibition between distant segments, this causes occasional “switching” between forward and backward peristalsis. If the extent of radial deformations is under neural control in the larva, this could provide a potential route for control or biasing of transitions between forward and backward peristalsis.
As a consequence of exploiting body mechanics, our model explains a wider range of behaviour than previous models, using a simpler nervous system. The properties included in the neuromuscular circuitry were derived from basic physical considerations, i.e., what was necessary and sufficient to produce exploration, rather than from known neuroanatomy or neurophysiology. However, it is useful to now examine what insights and predictions regarding this circuitry can be derived from our model.
Firstly, we consider the connections between segments. Unlike the model from [43], we did not require assymmetric connections to obtain forward (or backward) waves as these (and spontaneous switching between them) arise inherently in the mechanics. Rather, obtaining centre of mass motion of the entire body required the “ring attractor” layout of mutual inhibition between distant segments described above. The model thus predicts that motor output should be strongly inhibited (by signalling from other segments) the majority of the time, so that motor neurons only activate as the (mechanical) peristaltic wave passes through the corresponding body segment. This is in contrast to previous models which appealed only to local, nearest-neighbour inhibitory connections [39, 43].
What might be the neural substrate for the proposed inhibition? There are two currently known intersegmental inhibitory pathways in the larva. GVLI premotor inhibitory neurons synapse onto motor neurons within the same segment but extend their dendritic fields several segments further anterior along the VNC. Accordingly, the GVLIs inhibit motor neurons at a late phase during the local motor cycle [22]. Our model predicts that there should be a larger set of GVLI-like neurons within each segment, with dendritic fields tiling distant segments. Although in our model the mutual inhibition is (for simplicity) arranged to act on all non-adjacent segments, we would in practice expect that active compression is actually spread across more segments [3, 22] to transfer power to the body more efficiently (S3 Appendix), and this should be reflected in the inhibitory connection pattern. The second inhibitory pathway involves GDL inhibitory interneurons, which receive input from the excitatory premotor neuron A27h in the nearest posterior segment, and synapse onto A27h within the same segment while simultaneously disinhibiting premotor inhibitory neurons in distant segments [28]. Thus, GDL effectively produces both local and long-range inhibition of motor output. However, GDL receives axo-axonic connections from vdaA and vdaC mechanosensory cells within the same segment, so local inhibition is likely gated by sensory input. This would match our model, in which sensory activation within a segment should be sufficient to produce motor output when one of the neighbouring segments is active. We thus predict that simultaneous experimental suppression of GDL, GVLI, and all other long-range inhibition in the VNC should allow the propagation of several, concurrent locomotor waves in response to mechanical input.
Secondly, within a segment, our model highlights the importance of the timing of neuromuscular forces relative to body motion. Specifically, during locomotion, the larva’s muscles should act primarily as motors rather than as springs, brakes, or struts (see [33] for a discussion of these differences), and thus should activate in phase with the segmental stretch rate. This hypothesis could be tested by performing work-loop experiments, for which we predict the existence of a counterclockwise cycle in a plot of muscle force (potentially measurable by calcium imaging) over segment length during locomotion.
Can our model’s requirement that neurons sensing stretch-rate provide a direct excitatory connection to motor neurons, within the same segment, be mapped to identified pathways in the larva? One possible monosynaptic implementation of such a link are the dda mechanosensory cells which have been observed to make synapses onto aCC and RP2 motor neurons [23]. However, synapse counts show high variability both within and across individuals, so it seems unlikely to be a core component of the locomotor circuitry. A more promising candidate is the excitatory premotor interneuron A27h, which receives input from vpda and vdaC and sends bilaterally symmetric outputs to aCC and RP5 [28]. It is known that A27h activation is sufficient to activate downstream motor neurons, but it remains unknown whether proprioceptive sensory input is sufficient to activate A27h. Additionally, we hypothesise that A02 (PMSI) interneurons [20], which have been recently shown to form an inhibitory sensory-motor feedback pathway between dbd mechanosensory cells and motor neurons [27], could play a role in filtering this signal to obtain the necessary stretch-rate activation independently of stretch. General models of mechanotransduction suggest that larval mechanosensory cells may be sensitive to both rate of stretch as well as absolute stretch, depending upon the mechanical properties of the sensory dendrites and the biophysics of the relevant mechanosensitive ion channels [76]. If PMSIs have a slow-activating, integrator dynamics that encodes stretch, while A27h activate quickly in response to proprioceptive sensory input to encode stretch and stretch-rate, the combined input to motor neurons would be only stretch-rate dependent excitation, as our model requires. This could explain the observation that optogenetic disturbance of PMSIs [20] slows the timescale of peristaltic waves, as the inclusion of absolute stretch in this feedback loop would produce muscle forces that not only counteract friction but also decrease the effective stiffness of the cuticle, slowing peristalsis (see S5 Appendix).
It is clear the real larval nervous system exhibits many complexities not reflected in our model, and likewise that the real larva performs many more behaviours than exploration. These include appropriate (directed) reactions to sensory stimuli such as stopping, withdrawal and reverse in response to touch stimuli [38]; differences in the speed of forward and backward locomotion [15]; and modulation of the frequency and direction of (large) turns in response to sensory gradients such as odour, heat or light [4, 8–10, 12, 13, 77–82] to produce positive or negative taxis. In a previous model of taxis [44] we have shown that by a continuous coupling of the amplitude of a regular lateral oscillation to the experienced change in stimulus strength in a gradient, a larva-like response to gradients can emerge, again without requiring active switching between states. In the current model, this could be effected by incorporating direct neuromuscular driving of bending degrees of freedom, since the real larva can likely use asymmetric activation of its lateralised muscles to produce active bending torques to influence the transverse motion. Alternatively, the degree of bend could be influenced indirectly by altering the stiffness and viscosity of segments (as explored in our upcoming paper [49]), or their frictional interaction with the substrate. We note that the effective viscoelasticity of body segments can be neurally controlled by local reflex arcs (see S4 Appendix and [40]). Indeed, this could partially explain the experimental observation of increased bending on perturbation of a contralateral segmental reflex mediated by Eve+ interneurons [24]. The muscle activation caused by this reflex should produce bending torques which are proportional to current bend or bending rate, thus effectively modulating transverse stiffness or viscosity, respectively. Notably, in the taxis model of [44], it is not required that the descending signal that alters turn amplitude is lateralised, but rather that it has the right temporal coordination, which itself is naturally created by the interaction of body and environment.
Backward locomotion is observed to be slower than forward locomotion in the real larva [15], yet in our model both behaviours are of equal speed for a fixed value of reflex gain. We believe that this is due to the preservation of mechanical symmetry between forward and backward motion in our model. The real larva likely experiences asymmetric substrate interaction forces. For instance, this could be due to the exact coordination of denticle folding/lifting during forward and backward locomotion (S1 Video) or due to the geometry of the larva’s denticle bands, which display a degree of anisotropy [83]. Alternatively, there may be asymmetries within the larva’s neural circuitry responsible for this difference. Indeed, there do appear to be neurons in the larval VNC which are only active specifically during forwards or backwards locomotion, and these may be functionally asymmetric [28].
The model presented in this paper does occasionally produce stops (cessation of peristalsis) during exploration, but this only occurs in concert with a large body bend (this stored transverse energy can subsequently and spontaneously restart the peristalsis); whereas in larva slowing, stopping and resumption of peristalsis (or transition from a stop to a large bend) can occur while the body is relatively straight [2, 10]. As for ‘directed’ turning, this suggests that additional neural control might be needed to terminate or initiate movement in response to sensory stimuli. It is worth noting that our model predicts that peristalsis can be restarted by almost any small disturbance of the physical equilibrium provided the local feedback gain is high enough; similarly, lowering the gain means that energy losses due to friction are not compensated and the animal will stop. In general, we have found that altering assumptions about the sliding friction forces by which the model interacts with the substrate can often have unexpected and subtle effects on the motion produced, thus it would be interesting to further explore the functions provided by segmental lifting [3, 84], folding of the denticle bands (S1 Video), and extrusion of the mouth-hooks [3, 85] during locomotion. Indeed, detailed experimental characterisation of the substrate interaction forces experienced by the larva would be a major advance in understanding how the animal behaves. Inspiration for approaches to this problem could perhaps be taken from the literature on C. elegans substrate interaction (see for instance [42, 86–90], though this list is not exhaustive). In the more extreme case, larva are capable of burrowing through a soft substrate, and it is clear that a complex interaction of forces, mechanics, sensing and neural control must be involved that go well beyond the scope considered here.
|
10.1371/journal.pntd.0003705 | Tsetse Fly (G.f. fuscipes) Distribution in the Lake Victoria Basin of Uganda | Tsetse flies transmit trypanosomes, the causative agent of human and animal African trypanosomiasis. The tsetse vector is extensively distributed across sub-Saharan Africa. Trypanosomiasis maintenance is determined by the interrelationship of three elements: vertebrate host, parasite and the vector responsible for transmission. Mapping the distribution and abundance of tsetse flies assists in predicting trypanosomiasis distributions and developing rational strategies for disease and vector control. Given scarce resources to carry out regular full scale field tsetse surveys to up-date existing tsetse maps, there is a need to devise inexpensive means for regularly obtaining dependable area-wide tsetse data to guide control activities. In this study we used spatial epidemiological modelling techniques (logistic regression) involving 5000 field-based tsetse-data (G. f. fuscipes) points over an area of 40,000 km2, with satellite-derived environmental surrogates composed of precipitation, temperature, land cover, normalised difference vegetation index (NDVI) and elevation at the sub-national level. We used these extensive tsetse data to analyse the relationships between presence of tsetse (G. f. fuscipes) and environmental variables. The strength of the results was enhanced through the application of a spatial autologistic regression model (SARM). Using the SARM we showed that the probability of tsetse presence increased with proportion of forest cover and riverine vegetation. The key outputs are a predictive tsetse distribution map for the Lake Victoria basin of Uganda and an improved understanding of the association between tsetse presence and environmental variables. The predicted spatial distribution of tsetse in the Lake Victoria basin of Uganda will provide significant new information to assist with the spatial targeting of tsetse and trypanosomiasis control.
| Trypanosomiasis is a vector-borne disease transmitted to both humans and animals by the tsetse fly. The tsetse vector is distributed across sub-Saharan Africa. Trypanosomiasis maintenance is determined by the interrelationship of three elements: vertebrate host, parasite and the vector responsible for transmission. Mapping the distribution and abundance of tsetse flies assists in predicting trypanosomiasis distributions and developing rational strategies for disease and vector control. This study makes available dependable tsetse fly distribution data (maps) for use by decision makers. The approach makes use of modelling techniques involving limited field-sampled tsetse data points distributed across an area of approximately 40,000km2 within the Lake Victoria basin of Uganda. Precipitation, temperature, landcover, normalised difference vegetation index (NDVI, a measure of the amount of green vegetation) and elevation data were used as environmental covariates. We used logistic regression to analyse the relationships between presence of tsetse and the environmental covariates. The results indicated that tsetse are more likely to be present in areas with a greater proportion of riverine vegetation and forest cover. The key outputs are a predicted tsetse distribution map for the Lake Victoria basin of Uganda and an increased understanding of the association between tsetse presence and environmental variables. This will provide a vital resource for the planning and spatial targeting of future tsetse control activities.
| Tsetse flies are responsible for the transmission of human African trypanosomiasis (HAT), also known as sleeping sickness and its animal form (nagana). Trypanosomiasis occurs in 38 sub-Saharan African countries with an average of 15,000 human cases reported annually (period 2000–2012 [1]), and 70 million people at risk of contracting the infection [2]. Uganda reports approximately 500 cases of sleeping sickness annually [1], and it is the only country reporting the presence of both forms of HAT: the gambiense form in the north-west and the rhodesiense form in the south-east and, more recently, in the centre of the country [3, 4]. Animal trypanosomosis presents major constraints to livestock production among many livestock keeping communities in Africa. The disease is widely reported in Uganda [5], and the removal of African animal trypanosomiasis (AAT) could generate direct economic benefits in the region of 400 million US$ in a 20-year period [6].
Glossina fuscipes fuscipes is known to be present in several parts of Uganda, with its geographical extent stretching from Lake Victoria’s shores through central Uganda up to the West Nile region. In addition, G. f. fuscipes is assumed to be present around Lakes Albert, Edward and George in western Uganda. The islands of Kalangala and Buvuma located within Lake Victoria have also been identified as having G. f. fuscipes [7].
The major drivers of tsetse fly habitation are generally known to be temperature, humidity, rainfall, vegetation and presence of host animals [8, 9, 10]. This implies that tsetse flies are found in ecologically suitable habitats, represented through a set of conditioning environmental variables. Such variables determine: feeding behaviour; infection rates; fly movements; fly density; species-diversity; and fly reproduction [10]. Therefore, spatial information on such environmental variables can be helpful in predicting the relative distribution of tsetse flies in an area.
Tsetse distribution maps are crucial in the control and management of human and animal trypanosomiasis in affected areas [11, 12]. Accurate maps should ideally be based on high precision fly data derived from field investigations. In the absence of such data, tsetse distribution maps may be constructed using partial district-level entomological reports, existing publications, sector reports and modelled environmental covariates. Given scarce resources to carry out regular field tsetse surveys, there is a need to devise inexpensive means for periodically obtaining reliable large area and high precision tsetse information across target areas. A potential solution is provided by spatial statistical modelling (e.g., spatial regression analysis) using tsetse presence or abundance data acquired from field survey and fine spatial resolution satellite-generated environmental variables.
Regression is a statistical tool used to quantify the association between an outcome measure and predictor variables [13]. Logistic regression, in particular, is commonly used to explain or predict a binary variable response using a set of predictor variables or covariates [14]. This approach has been used in the predictive mapping of various vectors and associated vector-borne diseases including malaria and Rift valley fever, with broad applications in environmental disease risk [15]. The use of GIS and temporal Fourier-processed surrogates for vegetation, temperature and rainfall derived from satellite sensor data in predicting tsetse distributions has been investigated with significant utility [16]. Further use of GIS and remote sensing in attempting to explain tsetse vector distributions is described in Rogers et al.[17, 18] and Wint et al. [19, 20, 21].
Wint and Rogers [19, 21], at a spatial resolution of 5 km, predicted tsetse presence at the continental level using logistic regression, targeting 23 tsetse sub-species from the three major species groups (Fusca, Palpalis and Morsitan). The process involved fitting statistical regression models between tsetse data and remotely sensed predictor variables. The tsetse data used were derived from the Ford and Katondo tsetse maps [22, 23], through systematic extraction of 12,000 points across the entire continent. Predictor variables included; NDVI, surface temperature, middle-infrared reflectance, vapour pressure deficit and surface rainfall [19, 21].
Wint [20], in an effort to provide more accurate tsetse maps, derived sub-continental tsetse fly distribution maps at a spatial resolution of 1 km for East Africa (Uganda) and selected parts of some countries in West Africa. This approach made use of; (i) modified Ford & Katondo presence/absence maps, (ii) 5 km-continental tsetse predictions in 2000, (iii) 17,000 data points extracted for East Africa and satellite-derived data. According to these maps, Uganda is approximately 80% tsetse infested. Although an improvement from the Wint continental version [19, 21], these sub-continental tsetse distribution maps are associated with low precision. The lack of up-to-date field data on tsetse is a key concern, while the absence of land cover data as a predictor, which is known to be important in determining tsetse distributions, is another.
In Uganda, there is a need to produce dependable and up-to-date tsetse distribution information, preferably at sub-national level, to support decision-making and improved planning of tsetse control interventions. Relatively few studies have used recently gathered data from traps. The purpose of this study was to quantify the relationships between tsetse presence/absence and external factors in the study area and also to predict the spatial distribution of G. f. fuscipes in the Lake Victoria basin of Uganda.
The study area is predominantly a lake basin stretching for approximately 50 to 100 km from the Lake Victoria shoreline in Uganda. This region is characterized by high annual rainfall (1000–1500 mm) with two distinct rainfall peaks in April and November. Tsetse data were obtained from a systematic entomological survey conducted from May to June, 2010, to ascertain tsetse presence and abundance. Biconical traps [24, 25] were used to capture tsetse flies during the survey. Five thousand geo-referenced tsetse trap sites were spread uniformly over a ground area of approximately 40,000km2 within the target region [26]. Trapping at each site lasted 72 hours and was conducted by teams led by district entomologists. Single collection was made at the end of this 72 hour period. The parameters recorded in the entomological survey sheet included: trap code, latitude, longitude, altitude, vegetation type around the trap site, start date and time, end date and time, species trapped, number of females, males and flies of un-identified sex, and number of other biting insects. Data were collated and entered into a database. These tsetse data were used as the dependent variable in the regression modelling, while all other variables were used as independent variables.
Several covariates (Table 1) were used in the analysis, based on an understanding of the factors important for tsetse reproduction and survival [11, 27, 28]. These included; (i) land cover, (ii) temperature, (iii) normalised difference vegetation index (NDVI), (iv) elevation, and (v) rainfall. The land cover data were extracted from the fine spatial resolution, multi-purpose land cover dataset GlobCover for 2009 [29]. This global land cover series is described by a legend of 22 core land cover categories in total. The region under study contained only 19 of the 22 classes presented. Land cover variables used in the analysis were estimated through the creation of buffers of 1000 m (catchment) around each entomological tsetse survey point. Within each buffer, area percentages of the different land cover types were computed and used as the set of land cover predictor variables. NDVI, as a measure of vegetation cover or biomass production, was derived from the National Oceanic and Atmospheric Administration (NOAA) Global Inventory Monitoring and Modelling Studies group (GIMMS) dataset [29]. The temperature and precipitation data used were obtained as interpolated raster data at a spatial resolution of 30 arc-seconds from the WorldClim—Global Climate Data facility [29]. Elevation data were obtained from the Shuttle Radar Topography Mission (SRTM).
Tsetse survey count data were transformed to a binary variable representing tsetse fly presence or absence (0, 1). Presence of tsetse flies was represented by a “1” while absence was represented by “0”. Preliminary visualisation of the geographical distribution of tsetse presence was carried out using the ArcMap10 GIS software (ESRI, Redlands). Exploratory analysis was performed as a means to check for outliers, and aspects of homogeneity, normality and collinearity within the predictor variables.
A forward step-wise approach was applied to select the final multivariate logistic regression model. Covariates were added one after the other cumulatively and were retained if they retained statistical significance (p < 0.05). Estimated multivariate regression model coefficients were compared with those obtained at the univariate analysis stage to ascertain the consistency of final covariates in influencing the outcome variable.
A residual variogram was constructed to assess the presence of spatial autocorrelation in the model residuals. Autologistic regression was applied to account for the residual spatial autocorrelation [30, 31, 32, 33, 34, 35]. This process involves the introduction of a new explanatory variable (autocovariate). Autologistic regression involving several covariates is determined using the formula;
lnπ1−πi=α+βs(yi)+∑kykxki+εi
1
Where;
ɑ is the model intercept
β is the coefficient that relates to the autocovariate
s(yi) is the autocovariate and is a function that summarises the y-values in the neighbourhood of i. It is calculated from the observed data only once and used throughout.
γk are the coefficients relating to the k different environmental covariates
xki are the k different environmental covariates at location i.Ԑi is the error
The spatial autocorrelation was quantified by the Global Morans’s I index as extending up to a distance of 20 km [31, 32]. Thus, a spatial range of 20 km was used for the calculation of the autocovariate.
Receiver operating characteristic (ROC) curves were generated to evaluate model performance based on suggested cut-off points. Sensitivity and specificity were used to assess the predictive ability of the model. The area under the ROC curve (AUC) was calculated to provide an assessment of how accurately the model can classify the study area into tsetse presence and absence [13, 36]. Spatial prediction was carried out using the final multivariate model parameters, along with spatially continuous covariate datasets, to enable visualisation of predicted probability of occurrence for both the sampled and unsampled locations. The unsampled locations were represented on a regular grid and the predictions were used to produce continuous surface maps. The probabilities were derived from the regression equation in which the linear predictor was transformed using the logit function into a value between 0 and 1. Values close to ‘0’ represent a high probability of tsetse absence while ‘1’ represents a high probability of tsetse presence. All analyses were performed using the software R, version: Rstudio2011, with additional packages; geoR, gstat, MASS and spdep.
A map of tsetse abundance based on the tsetse sampling points is presented in Fig 1. These data indicate spatially heterogeneous distributions, with high tsetse abundance particularly in the Kalangala islands, along the river Nile, and in the south eastern regions of the study area.
In an initial univariate logistic regression stage, 44% of the land cover variables had a statistically significant association with tsetse presence-absence (p<0.05). Covariates; cropland, forest, riverine vegetation, woody vegetation, NDVI, elevation, temperature and rainfall were all positively correlated (p<0.05, Odds ratio (OR)>1), while savannah vegetation, herbaceous vegetation and built-up area were negatively correlated (p<0.05, OR<1) with tsetse presence.
Seven covariates were included in the multivariate logistic regression model. The significant covariates were; rainfall, elevation, temperature, cropland, savannah vegetation, forest, and riverine vegetation. Parameter estimates are given in Table 2. The presence of tsetse flies was negatively correlated with savannah vegetation, and positively correlated with the remainder of the model covariates. However, the covariates cropland, riverine vegetation, elevation and rainfall presented only very small positive associations, with wide confidence intervals.
The map of residuals and the residual variogram based on the multivariate logistic regression model revealed the existence of residual spatial autocorrelation. This situation is a problem as it violates the assumption of independence of residuals and can result in biased parameter estimates, leading to inflation of significance. Since non-spatial models fail to account for the autocorrelation effect, there was a need to apply a spatial model: in this case, autologistic regression [37, 38].
Autologistic regression was applied based on the seven significant variables obtained from the multivariate logistic model together with the computed autocovariate. The resultant statistics are presented in Table 3 and the residual variogram from the autologistic model is shown in Fig 2. The residual variograms for the two models were compared. The autologistic regression model reduced the spatial autocorrelation in the residuals compared to the multivariate logistic model.
In the autologistic model, forest (p<0.05, OR = 1.105) and riverine vegetation (p<0.05, OR = 1.008) were positively correlated with tsetse presence. Savannah vegetation (p<0.05, OR = 0.993) and elevation (p<0.05, OR = 0.997) were negatively correlated. These three land cover classes and elevation are, thus, considered to be important determinants of tsetse presence and absence in the study area. Cropland, temperature and rainfall failed to retain their significant association with tsetse presence (p>0.05) after accounting for spatial autocorrelation.
The Pearson X2 test parameter and Deviance parameter were evaluated as measures of goodness-of-fit. These measures were statistically non-significant (Pearson X2 = 4654, p = 0.196 (i.e p>0.05) and Deviance = 4890), indicating that the model fits the data appropriately and, therefore, could be used to predict probabilities of tsetse presence across the study area.
Model evaluation was conducted to assess prediction accuracy. The area under the curve (AUC) was computed as 72.7%, indicating adequate predictive ability. The plot of sensitivity and false positives (1-specificity) against expected probabilities (Fig 3) indicates a probability cut-off point of 0.28, leading to a sensitivity and specificity of 53%. This is the threshold value for the prediction of tsetse presence where both sensitivity and specificity are maximised, and can be used to classify areas as containing tsetse or not [39]. At a probability cutoff of 0.5, the sensitivity is 10% while specificity is 90% (Fig 3). This implies that at this cutoff approximately 90% of all the true positive cases (tsetse presence) will be missed. As the threshold increases, the sensitivity decreases and the specificity increases.
Fig 4 shows the predicted probability of tsetse presence across the study area, based on the multivariate logistic regression model (non-spatial model), while Fig 5 shows the predicted probability of tsetse presence across the study area, based on the autologistic regression model (spatial model). The two models identify areas of scaled potential tsetse fly risk with estimated probabilities of tsetse presence ranging from 0 to 1. The outcome reflects the presence of a clear tsetse infestation corridor in the Eastern part of the study area. High probability of tsetse occurrence (predicted probability of occurrence > 75%) was predicted in the eastern sections of the study area close to the Kenya-Uganda border (Bugiri, Busia, Tororo Kaliro, Kamuli and Pallisa districts) as well as on islands located in Lake Victoria. Low probability of tsetse occurrence (below 20%) was predicted in the western and north-western parts of the Lake Victoria basin.
The primary objective of the study was to develop a predictive model that can reliably inform decision-makers about the spatial distribution of G.f.fuscipes in the target area of Uganda, based on entomological survey results and a set of environmental covariates. The research was intended to provide high precision, up-to-date, sub-national tsetse maps to guide control interventions. The tsetse presence and absence data (dependent variable) represent one of the most comprehensive tsetse datasets collected over such a large area and are fully geo-referenced.
At the univariate investigation stage, tsetse presence was found to be significantly (positively) associated with eight variables (cropland, woody vegetation, forest, riverine vegetation, NDVI, elevation, temperature and rainfall (P<0.05, OR>1)). Rainfed cropland, savannah vegetation and herbaceous vegetation demonstrated a negative association (p<0.05, OR<1). Temperature demonstrated the largest correlation with the outcome variable (p<0.05, OR = 2.61).
The multivariate logistic regression model established that the presence of tsetse was positively associated with temperature, elevation, rainfall and proportion of forest cover, riverine vegetation and cropland. Savannah vegetation was negatively correlated with the outcome. Temperature remained highly influential in determining tsetse presence in the multivariate model (p<0.05, OR = 2.63). Tsetse flies are very sensitive to environmental changes and ecological instability, and are found in ecologically suitable habitats which have the necessary temperature, humidity and vegetation cover [40]. G.f.fuscipes, as a riverine species of the palpalis group, thrives in zones with high humidity [41].
After accounting for spatial autocorrelation, the covariates temperature, rainfall, and cropland lost their statistical significance in influencing tsetse presence or absence. The use of a spatial autologistic regression enabled the detection of key environmental variables that are highly influential in positively determining tsetse presence and these were; forests and riverine vegetation. Savannah vegetation (p<0.05, OR = 0.993) and elevation (p<0.05, OR = 0.997) retained their negative association with tsetse presence. The discussion below is based entirely on the results from the spatial regression model (autologistic regression).
Tsetse presence was positively correlated with forest cover. These correlations are consistent with the known aspects of the fly’s ecology [10]. Tsetse (G.f.fuscipes) thrives in environmental conditions where the vegetation is not too dense such as to enable them to fly easily and spot the feeding host readily. In addition, tsetse presence was positively correlated with riverine landcover. G. f. fuscipes is ecologically considered a riverine species and is commonly found in zones of high humidity offered by the interaction between forest vegetation and water bodies. It is important to make use of data with a fine spatial resolution, especially when considering drainage systems, to enable the identification of small rivers and streams that may support riverine vegetation. The spatial resolution of the land cover data used may not have been detailed enough to enable small rivers to be detected.
Tsetse presence was negatively correlated with savannah vegetation. Such vegetation can be categorised as “humanised” or “disturbed” landscapes, and tsetse flies usually avoid disturbed habitats [40]. Additionally, the low humidity in savannah landscapes (due to less water and vegetation cover) is less suitable habitat for riverine tsetse flies. Tsetse presence was also negatively correlated with elevation. Such association has been detected in previous research [42]. Generally, elevation may influence the micro-climatic conditions or landcover variations of an area. However, the entire study area had limited height variation (1034 to 1412 m asl) and the model, thus, illustrates the lack of an altitudinal control on tsetse presence within this particular study area, as evidenced by the odds ratio which was close to unity. Water courses are located at lower elevations. Thus the altitude effect is bound to be influenced by proximity to existing waters courses.
Tsetse presence was not correlated with cropland (p>0.05). This association could be linked to its characteristic of being a completely humanised landscape. There is a tendency for tsetse flies to avoid such environments [40] due to removal of vegetative cover ideal for tsetse survival. However, following habitat degradation, G. f. fuscipes can take refuge in remnant tree cover (thickets), which may explain the presence of tsetse in cultivated fields during the survey [43]. Tsetse presence was not significantly correlated with rainfall. The entire study area had monthly total rainfall ranging from 34 to 339 mm with no significant spatial variation across the study area. Therefore, it is unlikely that precipitation would influence tsetse distributions.
Tsetse presence was not significantly correlated with temperature (p>0.05). Tsetse flies thrive in areas with mean annual temperatures between 19 and 30°C [44]. Temperatures below 19°C slow down tsetse activity and general physiology [44], while extreme low temperatures (below 15°C) increase fly mortality [45]. Tsetse are severely affected by high temperature conditions and once exposed to a temperature of more than 36°C tsetse will have a survival capacity of close to zero [46]. From the training data, the lowest temperature for the study area was recorded as 13.5°C, the mean temperature was 27°C and the maximum temperature was 29.7°C. Temperature variation was by about 4°C at most sites across the region. These temperature ranges were within the acceptable envelope for the fly and therefore had no specific consequences on fly availability in the study area.
Fitting the autologistic regression model permitted us to assess the influence of spatial autocorrelation on the probability of tsetse presence. The parameters for temperature, rainfall, and cropland appeared less important (not statistically significant) after accounting for the effect of forest cover, riverine vegetation, elevation and spatial dependence in the observations. The predictive outputs from the autologistic regression model are considered to be more reliable than those from the initial logistic regression model, as they account for spatial autocorrelation in the data by incorporating information from neighbouring locations. The autocovariate term captured part of the spatial pattern in the data observations, thus, providing a more robust estimation of the covariate effects after accounting for the spatial dependence in the observations. The autologistic regression predictive outputs should be considered to be the most important in terms of future planning of interventions. However, it should be noted that the predictive models did not predict tsetse presence along the river Nile and this may be due to the spatial resolution of the covariate data used (1 km) not allowing accurate representation of relatively small areas of suitable habitat.
Other methodological approaches can be used to deal with spatially autocorrelated data, such as model-based geostatistics [15], although the fitting of these models and subsequent spatial predictions are very demanding computationally. Future research will refine the spatial models presented in this paper using these computationally intensive methods. The present analysis provides much needed empirical data on tsetse distributions in south east Uganda, along with spatially continuous predicted outputs which will provide significant benefits for the planning of future interventions.
Several tsetse sub-species have long been associated with the Lake Victoria basin. The location-specific entomological data gathered for this study provide further evidence of the extensive distribution of tsetse in the area. Using logistic and autologistic regression models coupled with extensive field survey entomological data and a set of environmental covariates, a tsetse distribution map for the lake basin was constructed. These regression models enabled the identification of the important environmental variables determining tsetse presence across the study area. Notably, the final model identified forests and riverine vegetation (positive) and savannah vegetation and elevation (negative) as the key covariates associated with tsetse presence in the study area. Knowledge of the influential factors and availability of detailed sub-national tsetse distribution maps offers a platform for making meaningful decisions when planning tsetse control interventions. The findings are based on data from Uganda, but the approach is certainly of much broader interest and application.
|
10.1371/journal.pntd.0000952 | The Case for Reactive Mass Oral Cholera Vaccinations | The outbreak of cholera in Zimbabwe intensified interest in the control and prevention of cholera. While there is agreement that safe water, sanitation, and personal hygiene are ideal for the long term control of cholera, there is controversy about the role of newer approaches such as oral cholera vaccines (OCVs). In October 2009 the Strategic Advisory Group of Experts advised the World Health Organization to consider reactive vaccination campaigns in response to large cholera outbreaks. To evaluate the potential benefit of this pivotal change in WHO policy, we used existing data from cholera outbreaks to simulate the number of cholera cases preventable by reactive mass vaccination.
Datasets of cholera outbreaks from three sites with varying cholera endemicity—Zimbabwe, Kolkata (India), and Zanzibar (Tanzania)—were analysed to estimate the number of cholera cases preventable under differing response times, vaccine coverage, and vaccine doses.
The large cholera outbreak in Zimbabwe started in mid August 2008 and by July 2009, 98,591 cholera cases had been reported with 4,288 deaths attributed to cholera. If a rapid response had taken place and half of the population had been vaccinated once the first 400 cases had occurred, as many as 34,900 (40%) cholera cases and 1,695 deaths (40%) could have been prevented. In the sites with endemic cholera, Kolkata and Zanzibar, a significant number of cases could have been prevented but the impact would have been less dramatic. A brisk response is required for outbreaks with the majority of cases occurring during the early weeks. Even a delayed response can save a substantial number of cases and deaths in long, drawn-out outbreaks. If circumstances prevent a rapid response there are good reasons to roll out cholera mass vaccination campaigns well into the outbreak. Once a substantial proportion of a population is vaccinated, outbreaks in subsequent years may be reduced if not prevented. A single dose vaccine would be of advantage in short, small outbreaks.
We show that reactive vaccine use can prevent cholera cases and is a rational response to cholera outbreaks in endemic and non-endemic settings. In large and long outbreaks a reactive vaccination with a two-dose vaccine can prevent a substantial proportion of cases. To make mass vaccination campaigns successful, it would be essential to agree when to implement reactive vaccination campaigns and to have a dynamic and determined response team that is familiar with the logistic challenges on standby. Most importantly, the decision makers in donor and recipient countries have to be convinced of the benefit of reactive cholera vaccinations.
| Cholera outbreaks have had catastrophic impact on societies for centuries. Despite more than half a century of advocacy for safe water, sanitation and hygiene, approximately 100,000 cholera cases and 5,000 deaths were reported in Zimbabwe between August 2008 and by July 2009. Safe and effective oral cholera vaccines have been licensed and used by affluent tourists for more than a decade to prevent cholera. We asked whether oral cholera vaccines could be used to protect high risk populations at a time of cholera. We calculated how many cholera cases could have been prevented if mass cholera vaccinations would have been implemented in reaction to past cholera outbreaks. We estimate that determined, well organized mass vaccination campaigns could have prevented 34,900 (40%) cholera cases and 1,695 deaths (40%) in Zimbabwe. In the sites with endemic cholera, Kolkata and Zanzibar, a significant number of cases could have been prevented but the impact would have been less dramatic. The barriers which currently prevent the implementation of mass vaccinations, including but not only the cost to purchase the vaccine, seem insurmountable. A concerted effort of donors and key decision makers will be needed to offer better protection to populations at risk.
| In October 2009, the World Health Organization's (WHO) Strategic Advisory Group of Experts (SAGE) on immunization made the pivotal recommendation that oral cholera vaccination should be considered as a reactive strategy in areas with ongoing outbreaks. This is in addition to the continuing recommendation that oral cholera vaccines be used in areas where the disease is endemic and should be considered in areas at risk for outbreaks in conjunction with other prevention and control strategies [1]. Previously, the WHO did not recommend oral cholera vaccination once an outbreak had started due to “the time required to reach protective efficacy and the high cost and heavy logistics associated with its use” [2]. This reluctance has since changed because of the emergence of large and prolonged outbreaks, particularly in sub-Saharan Africa [3]. The large cholera outbreak in Zimbabwe is the latest of these catastrophes [4]. By 2008, 179,323 (94%) of the reported 190,130 cholera cases and 5,074 (99%) of 5,143 cholera deaths reported to the WHO occurred in Africa [5].
There are two oral cholera vaccines (OCVs) available. Dukoral is internationally licensed and prequalified by the WHO for purchase by United Nations agencies and consists of inactivated Vibrio cholerae O1 whole cells combined with the B subunit of the cholera toxin (BS-WC). A large-scale field trial in Bangladesh in the 1980's showed that the BS-WC vaccine is safe and conferred high-grade (∼85%) protective efficacy (PE) during the first 6 months after vaccination, decreasing to ∼60% during the following 18 months and much lower ∼20% in the 3rd year following vaccination [6], [7]. More recently the BS-WC vaccine has been evaluated in Beira, a cholera endemic region of Mozambique, which demonstrated the feasibility and effectiveness of vaccination (PE∼80% during the first year after vaccination) under actual public health conditions in a setting in sub-Saharan Africa [8]. Two more mass vaccination campaigns confirmed the feasibility of this approach in the complex emergency settings of Darfur, Sudan, and in Aceh, Indonesia [9]. A further WHO sponsored evaluation of the BS-WC vaccine is currently under way in Zanzibar. Mass vaccination campaigns with this vaccine in other cholera endemic sites in sub-Saharan Africa are currently under discussion. The BS-WC is administered with a buffer solution in two doses with an interval of at least seven days. Protection is conferred 7 to 10 days after the second dose. The price of this vaccine which is produced in Sweden (approximately USD $18–30/dose on the commercial market) has been a major barrier to wider use. Recently a similar, much less expensive killed OCV has been licensed in Vietnam and subsequently in India [10], [11] and is undergoing WHO prequalification. This second vaccine (WC only), licensed in India as Shanchol, consists of WC without a B subunit and does not require the co-administration of a buffer solution. A double-blind, placebo-controlled trial showed that this vaccine (given in two doses with a minimum inter-dose interval of 14 days) is safe and efficacious, providing nearly 70% protection against clinically significant cholera for at least 2 years after vaccination [12]. Preparations for a trial of a single-dose Shanchol regimen are underway.
The SAGE also recommended that the impact of oral cholera vaccination in halting outbreaks should be documented. In the absence of data from reactive vaccination campaigns we used existing data from cholera outbreaks occurring in endemic and non-endemic settings in Asia and Africa to compare actual cholera outbreaks with simulated outbreaks during which a mass vaccination takes place. To add realism to our simulation we have varied the response times, and made estimates for single- and two-dose vaccines.
Three sites in three countries Zimbabwe, Kolkata (India), and Zanzibar (Tanzania) were selected based on a) availability of reliable data and b) absence of interventions such as vaccinations. The sites vary in cholera endemicity (Table 1). In Kolkata cholera is endemic and seasonal cholera outbreaks can be predicted. Cholera is also endemic in Zanzibar, however the location of cholera clusters tends to shift between years and is not easily predictable. In Zimbabwe cholera is not yet endemic. Large cholera outbreaks occurred in 1999 and 2002 [5]. Between 2002 and 2008 no cholera outbreaks were reported. In Zanzibar and Zimbabwe the large majority of cholera cases were clinically diagnosed. In Kolkata all cholera cases were laboratory-confirmed. A detailed description of the Kolkata site in has been published [13].
Zimbabwe is a republic located in the southern part of Africa. The population of the country is approximately 13,4 million. The cholera outbreak in 2008/9 is one of the largest outbreaks ever recorded. The outbreak started in August 2008, lasted for 49 weeks and affected all 10 of the country's provinces. The data used in our study comes from the daily cholera updates posted by United Nations Office for the Coordination of Humanitarian Affairs (OCHA), Zimbabwe [14]. The large majority of reported cases were based on clinical diagnosis. The daily updates were entered in an excel spreadsheet and analysed. Since daily updates were only posted from November 2008 onwards we extrapolated the epidemic curve backwards between August, the date the first cases were reported, and November 2008 when daily reporting started. The country-wide attack rate during 49 weeks of the 2008/9 outbreak was 7.4 per 1000 population.
The site in India consists of legally registered urban slum areas (bustees) within the administrative wards 29 and 30 in the city of Kolkata [13]. The area has a high population density and residents do not have sufficient water supply or sanitary facilities. A baseline census of the study population was done in early 2003 and enumerated 57,099 individuals. Study site residents of all ages were under surveillance for diarrhoea treated at any of the five project health clinics set up in the field and the city's infectious diseases and children's hospitals. Rectal swabs in Cary-Blair media were brought on the same day from the project health outposts to the study laboratory at the National Institute of Cholera and Enteric Diseases. Vibrio cholerae were isolated and identified using standard methods. The annualized cholera incidence during the surveillance period from May 2003 to April 2005 was 1.6 per 1000 population.
In Zanzibar, the site consists of two islands Unguja and Pemba, 40km and 60km off mainland Tanzania, with a total population of 1,182,804 in 2008, calculated from 2002 national census data. In Zanzibar the first cholera cases were detected in 1978. For this study the seven outbreaks occurring in Unguja and Pemba during the decade 1997 to 2007 were reviewed. In Unguja, cholera patients tend to reside in densely populated urban areas with water supply from communal and private taps. In contrast, in Pemba outbreaks were mainly reported in rural areas, with shallow wells as the primary water source. Routine surveillance reports completed by the national surveillance system were reviewed. The majority of cases are clinically diagnosed with the initial cases laboratory confirmed using standard methods, at Mnazi Mmoja Hospital in Unguja, and the Public Health Laboratory Ivo de Canieri in Pemba. Between 1997 and 2007, 5640 cholera cases were reported in Zanzibar, assuming a population of 1,037,183 during this period, the annualised incidence was 0.5 per 1000 population over the decade.
Outbreaks are defined as temporally- and geographically-clustered cholera cases, separated by the absence of cases for a minimum of 6 months. We assume that the vaccine confers 85% protection during the first 6 months following vaccination decreasing to 60% up to the end of year 2 and 20% protection in the 3rd year following vaccination. We define vaccination as the ingestion of the OCV and immunisation as the protective biologic response following vaccination (Table 2). Administration of two vaccine doses either at 7, 14 or 28–42 day intervals results in similar immune responses after the second dose [15]. Protection against cholera can be expected to start one week after the primary immunization series.
We assume the availability of a cholera vaccine through a rotating stockpile to prevent the expiration of doses. The overall response time is divided into the following components, the outbreak is recognized, an agreement to send vaccine is reached, the vaccine shipment arrives, vaccinations start, the administration of the first dose is completed, delay between first and second dose, second dose starts, vaccinations are completed, and finally the participants are immunized (Table 2 and 3). There is currently no agreement on a threshold to trigger cholera vaccination campaigns. We assume that in the endemic settings such as Kolkata or Zanzibar, the time from the report of the initial cholera cases to the recognition that an outbreak is occurring could take between 24 hours to 6 weeks, the number of cases reported before an outbreak was recognised under these assumptions is displayed in Table 4. In contrast the outbreak in Zimbabwe was in a non-endemic setting and orders of magnitude bigger. We arbitrarily set the threshold which should have triggered a vaccination campaign in Zimbabwe at 400 cases. The period required for the stockpile administrators to come to an agreement to implement a mass oral cholera vaccination and ship the vaccine could also take between 24 hours and 6 weeks. The vaccine shipment via air courier could take between 2 and 6 weeks depending on the urgency as well as potential delays clearing customs. Setting up the posts and starting vaccination will take 1 to 4 weeks and completion of the first dose will take 2 to 4 weeks. An interval of at least 7 days is required between the two doses. Starting the administration of the second dose will require up to 7 days, assuming staff and materials are on stand-by following the first dose. Completion of the second dose will take 2 to 4 weeks. The time required for vaccinated individuals to mount an immune response is 7 days. In total, the minimum time required to immunize the community is about 10 weeks, a delayed response would be 21 weeks and in the worst case scenario, a maximum response time could take as long as 33 weeks (Table 2).
The use of a hypothetical single-dose vaccine with similar protective efficacy and duration of protection as the licensed two-dose vaccine will reduce the response period, as the minimal delay between first and second dose and the time to complete the second dose will no longer apply. The use of single-dose vaccine would reduce the rapid response time in our simulations from 10 weeks to 6 weeks (a 40% reduction), the delayed response time from 21 to 16 weeks (a 24% reduction) and finally a maximum response time from 33 to 27 weeks, (an 18% reduction, Table 3).
In compliance with Good Clinical Practice guidelines all information that could reveal the identity of study participants was removed prior to analysis. None of the data available to the investigators of the submitted study contained information which could potentially reveal the identity of the participants. The data from Zimbabwe was public domain data so ethics approval was is not necessary. The data from Kolkata was collected in preparation of large cholera vaccine trials. It was considered sufficient and appropriate by the local investigators to obtain verbal, informed consent since the participation in the study consisted of taking a medical history, physical examination, testing of stool specimens. All procedures included in the study participation are part of good routine management of diarrhea patients. No experimental procedures of any kind were conducted on study participants. Besides the local (ethics committee of the National Institute of Cholera and Enteric Diseases) and the national ethics review committee (the Health Ministry Screening Committee of India), the study was approved by the WHO Secretariat Committee on Research Involving Human Subjects and the International Vaccine Institute Institutional Review Board. The data from Zanzibar was collected in preparation for a large cholera vaccine effectiveness study. Data was summarised routine surveillance data without any patient identifiers and so individual patient consent was unobtainable and unnecessary. Permission to use the data was granted by the Ministry of Health and Social Welfare, Zanzibar, who provided the data for the study.
We created graphs of each outbreak showing number of cholera cases by week. The number of prevented cases (PC) was calculated as the product of the number of reported cases (RC), protective efficacy (PE), and percent of the population participating in mass vaccination campaigns (Vaccine coverage, VC) or PC = RC×PE×VC. PE was set at 85% during the first six months after vaccination, 60% after 6 months and 20% in the third year based on previously published data [6], [7]. We assumed that between 50% and 75% of the target population will participate in mass vaccination campaigns. VC was therefore set at 50% and repeated at 75%. The number of prevented cases was subtracted from the reported number of cases to model the outbreak curves after mass vaccinations. The number of prevented cases was calculated for the varying response times shown in Table 2 for a two-dose vaccine and in Table 3 for a single-dose vaccine.
In Zimbabwe the first cholera cases were reported in August 2008 [5]. The outbreak reached its peak in the last week of January 2009 and had subsided by the beginning of July 2009. By July 2009, 98,591 cholera cases had been reported with 4,288 deaths attributed to cholera. The overall case fatality rate (CFR) was 4% with no significant decrease in CFR throughout the outbreak. If a stockpile of cholera vaccines would have been available mass vaccinations could have been implemented once a critical number of cholera cases had been diagnosed. The outbreak affected the whole of Zimbabwe hence a nationwide vaccination campaign would have been required. We calculated the reduction in cases that a two-dose vaccine would have achieved under different conditions (Figs 1a to c and Table 5). Had a rapid response taken place after the initial 400 cases had been reported, we estimate that as many as 34,900 of 98,591 (40%) cholera cases and 1,695 of 4,288 (40%) deaths could have been prevented by a mass OCV campaign with 50% coverage (Fig 1a and Table 5). Delayed and maximum time responses would have resulted in fewer cases prevented (Figs 1b and c and Table 5).
In Kolkata three outbreaks in 2003, 2004, and 2005 were reviewed. During the 2003 outbreak, 53 cases were detected. The number of cholera cases that could have been prevented by reactive mass vaccination with a coverage of 50%, was between 19 (36%) with a rapid response time to as low as 7 (13%) with a delayed response time (Fig 2a and b and Table 5). With a higher participation rate of 75%, which is realistic in this setting, the number of cholera cases avoided increased to 29 (54%) with a rapid response time (Table 5). There would be no reduction of cases associated with a maximum response time of 33 weeks, as the 2003 outbreak lasted only 30 weeks (Figure 1c). During the 2004 outbreak, 136 cases were detected. The number of cholera cases that could have been prevented by reactive mass vaccination with a coverage of 50%, was between 18 (13%) with a rapid response, 14 (10%) with a delayed response and 3 (3%) with the maximum response time (Fig 2d and e and Table 4). The number of cholera cases which could have been prevented in the 2005 season is shown in Table 5, the number of cases reported before an outbreak is recognised is shown in Table 4.
In Unguja, Zanzibar four outbreaks were reviewed. From the 52nd week of 1997 to the 34th week of 1998 a total of 452 cases were detected (0.85 cholera cases/1000 population). A reactive island-wide mass vaccination campaign with 50% vaccine coverage would have prevented 108 (24%) cholera cases with a rapid response and 6 (1%) cases with the maximum response time. With a 75% vaccine coverage, the number of cholera cases prevented would have been 162 (36%) with a rapid response time to 8 (2%), with the maximum response time. To initiate a rapid response an outbreak would be recognised after 24 hours, when 14 cases had been reported, in a maximum response time an outbreak would have been recognised after 6 weeks after a total of 164 cases had been reported (Table 4). The number of cholera cases which could have been prevented in three subsequent outbreaks is shown in Table 5. In Pemba three outbreaks were reviewed. A total of 119 cholera cases were detected (0.32 cholera cases/1000 population) from the 50th week of 2002 to the 12th week of 2003. A reactive island-wide mass vaccination campaign with 50% vaccine coverage would have prevented 5 (4%) cholera cases with a rapid response time. Delayed and maximum response times did not prevent any cholera cases. The following year the number of preventable cases increased to 253 (29%) and 124 (14%) respectively for rapid and maximum response times. The number of cases that could be avoided during the following years and under different assumptions is shown in Table 5, and the number of cases reported before recognising an outbreak shown in Table 4. Figure 3a to 3f shows the number of preventable cases in Unguja and Pemba during the 2006–2007 outbreak had reactive vaccination been employed under different assumptions.
While the protective efficacy is waning in the years following immunisation there remains ∼60% PE during the following 18 months and ∼20% PE in the 3rd year following vaccination. More cases are prevented in years following maximum and delayed responses vaccinations as there is no delay in protection. In endemic settings a reactive vaccination campaign using a the two-dose OCV in the first outbreak with 50% coverage in Zanzibar, would prevent 137–168 cases in Unguja and 184–189 cases in Pemba during the three years following a rapid response and maximum response vaccination campaigns, respectively (Table 6). In Kolkata 43–55 cases would be prevented during the three years following a rapid response and maximum response vaccinations, respectively.
If a hypothetical single-dose vaccine with similar characteristics as the two-dose vaccine would have been available for a rapid mass vaccination in Zimbabwe, 41,059 cases could have been prevented (42%) and 1,748 deaths (41%; Table 7). In three other sites a single-dose vaccine in a rapid response vaccination could have prevented a total of 1,768 cases (29%), reducing outbreak size by 3% compared to a two-dose vaccine.
We found that the number of cholera cases prevented by reactive mass vaccination campaigns depends on the size and shape of the outbreak curve. Reactive mass vaccinations can be expected to be most effective in large, long-lasting outbreaks which are most likely to occur in populations with no past exposure, i.e. where cholera appears de novo or returns after a long period of absence. In the presence of a sufficiently large, susceptible population an outbreak can continue for months. The outbreak in Zimbabwe petered out approximately 11 months after the first cases were reported. It has been estimated previously that cholera outbreaks in a refugee camp last 20 weeks hence reactive vaccination was unlikely to be cost–effective [16]. The large and long cholera outbreak in Zimbabwe demonstrated a different dynamic and lasted more than twice as long as average outbreaks in refugee camps. A reactive cholera vaccination campaign would have prevented significant numbers of cholera cases and deaths. In contrast during shorter outbreaks in the endemic settings of Kolkata and Zanzibar, where the majority of cases occur early on, only an immediate, brisk response will prevent a substantial number of cases. A second important finding was the minimal advantage of a single-dose in reducing the number cholera cases compared to a two-dose vaccine regimen. The advantage of single dose vaccines was most pronounced in small short outbreaks and less important in larger and longer outbreaks. A single dose vaccine would increase coverage as there would be no drop out between the first and the second dose.
Public health experts have considered establishing a cholera vaccine stockpile similar to the existing yellow fever and the meningococcal vaccine stockpiles [17], [18]. A concerted effort to distribute a hypothetical cholera vaccine stockpile could have potentially prevented more than a third of the cholera cases and deaths in Zimbabwe 2008–9. However there is currently no cholera vaccine stockpile in existence which the international aid community could have used for this purpose. Secondly there is a consensus opinion that the political situation in Zimbabwe at the time of the outbreak would have prevented mass vaccination campaigns. Thirdly not all provinces were initially affected by the cholera outbreak. Early strategically targeted mass vaccination campaigns potentially could have prevented the spread of cholera which ultimately affected all provinces. Important lessons could and should be learned from this disaster for future cholera outbreaks.
A major challenge for the administration of a cholera vaccine stockpile will be a consensus on the number of cholera cases which represent the threshold to trigger mass vaccination campaigns. The threshold should discriminate between sporadic cholera cases and an outbreak. This threshold may have to be calibrated for the cholera endemicity and the dynamics of outbreaks. The number of cholera cases may have a different significance in a cholera endemic area compared to an area where cholera hasn't been detected for several years. A steady increase in the daily number of reported cases is more worrying than stable or a declining numbers of daily cholera cases. In the absence of a consensus of a threshold to start cholera mass vaccinations we have assumed for the purpose of this paper that it may take between 24 hours and 6 weeks until a threshold is reached which triggers a mass vaccination.
We have underestimated the benefit of vaccinations as we could not include the added benefit of herd immunity which is likely to further reduce the number of cholera cases and deaths. Widespread administration of cholera vaccines protects the vaccinated individual as well as the unvaccinated community members with indirect protection proportional to vaccine coverage [19]. Longini and co-workers have used modelled data to show that cholera transmission could be controlled in endemic areas with 50% coverage with OCVs [20]. However the available evidence for herd immunity is based on data collected in a randomised controlled trial conducted in Bangladesh in the 1980s where cholera was endemic. We have currently no data on the added protection conferred by herd immunity following reactive vaccinations in cholera outbreaks. It seems likely that herd immunity will have an additive effect in reactive vaccinations. There is hope that a vaccine coverage of 50% or more may reduce the basic reproductive number below equity and abort cholera outbreaks altogether. However without empiric evidence it is highly speculative trying to quantify the added protection conferred by herd immunity. Estimating such added protection conferred through herd immunity would be highly informative for future mass vaccination campaigns.
Another limitation of this study was the selection of culture confirmed cases from Kolkata. It seems likely that cholera cases were not captured because they did not present to the treatment centre or had false negative microbiology results. The true number of cases is most likely higher. For our calculations, we used a minimum inter-dose interval of 7 days which is recommended for the BS-WC vaccine, whereas efficacy data is available only for an inter-dose interval of 14 days for the WC-only vaccine [12]. Due to the expected dropout between the first and second dose it may be necessary to vaccinate 10 to 20% more people to achieve a coverage of 50% or 75% with two complete doses. However the 14 day interval may not be essential as immunogenicity data suggests protection starting even after a single dose [21], [22]. Finally our database may not be representative of other cholera outbreaks. However in the absence of surveillance, other available epidemic curves are likely to be incomplete, especially missing the cases at the start of the outbreak. Furthermore the Kolkata data comes from a restricted population which may have underestimated outbreak duration compared to Zimbabwe and Zanzibar, where data were collected from the whole population. Clearly our models do not provide precise predictions of the number of cases prevented but they serve to highlight questions and point to areas where further research is needed.
As it is currently impossible to predict the size and shape of an outbreak curve, our data suggest that time is of the essence, a brisk response will provide most benefit. If circumstances prevent a rapid response there remain good reasons to roll out cholera mass vaccination campaigns in response to an outbreak report. Even if the vaccination does not impact greatly the current outbreak, once a substantial proportion of a population is vaccinated, outbreaks in subsequent years may be reduced if not prevented. This benefit should be taken into consideration when deciding on reactive cholera vaccinations. Furthermore containment of the disease in one area may prevent the spread to other susceptible populations. Because outbreaks are heterogeneous a large number of outbreaks would have to be randomized to assess the impact of reactive vaccinations. Hence a meaningful assessment of the impact of reactive mass vaccinations may not be feasible. Modelling the cost effectiveness of reactive mass vaccinations may well provide further evidence of the benefit of this intervention.
Our findings support the SAGE recommendations to include reactive mass vaccination campaigns in the interventions to manage cholera outbreaks. An explosive cholera outbreak in Haiti at the end of 2010, where a population of 10.1 million people has no prior immunity has added poignancy to our report [23]. There is an urgent need for financial mechanisms to establish and maintain a stockpile of cholera vaccines as well as a determined and dynamic team to administer such a stockpile. Perhaps most importantly decision makers in affected countries have to become aware of the benefit of reactive vaccination campaigns and actively promote their use.
|
10.1371/journal.pntd.0000404 | Successful Interruption of Transmission of Onchocerca volvulus in the Escuintla-Guatemala Focus, Guatemala | Elimination of onchocerciasis (river blindness) through mass administration of ivermectin in the six countries in Latin America where it is endemic is considered feasible due to the relatively small size and geographic isolation of endemic foci. We evaluated whether transmission of onchocerciasis has been interrupted in the endemic focus of Escuintla-Guatemala in Guatemala, based on World Health Organization criteria for the certification of elimination of onchocerciasis.
We conducted evaluations of ocular morbidity and past exposure to Onchocerca volvulus in the human population, while potential vectors (Simulium ochraceum) were captured and tested for O. volvulus DNA; all of the evaluations were carried out in potentially endemic communities (PEC; those with a history of actual or suspected transmission or those currently under semiannual mass treatment with ivermectin) within the focus. The prevalence of microfilariae in the anterior segment of the eye in 329 individuals (≥7 years old, resident in the PEC for at least 5 years) was 0% (one-sided 95% confidence interval [CI] 0–0.9%). The prevalence of antibodies to a recombinant O. volvulus antigen (Ov-16) in 6,432 school children (aged 6 to 12 years old) was 0% (one-sided 95% IC 0–0.05%). Out of a total of 14,099 S. ochraceum tested for O. volvulus DNA, none was positive (95% CI 0–0.01%). The seasonal transmission potential was, therefore, 0 infective stage larvae per person per season.
Based on these evaluations, transmission of onchocerciasis in the Escuintla-Guatemala focus has been successfully interrupted. Although this is the second onchocerciasis focus in Latin America to have demonstrated interruption of transmission, it is the first focus with a well-documented history of intense transmission to have eliminated O. volvulus.
| Brought to the Americas from Africa by the slave trade, onchocerciasis is present in six countries in Latin America. The disease is caused by a round worm and is transmitted to humans by the bite of an infected black fly. Once in a human, the adult worms produce larvae that circulate through the body, causing itching or even blindness. Ivermectin, a drug that kills the larvae, is delivered by public health authorities in countries where the disease is present. If the larvae are killed, then the disease cannot be transmitted to more people. People living in the Escuintla-Guatemala focus, a region in Guatemala where the disease was common, have been taking ivermectin for many years. The Ministry of Health of Guatemala believes that onchocerciasis is no longer being transmitted in the area. To prove that there is no more transmission of the disease, the authors examined the eyes of residents of the area to see if they could find any evidence of the worms. They also conducted analyses of blood in school children to see if they had ever been exposed to the worm, and they caught thousands of black flies and tested them to see if they were infected. These evaluations found no evidence of transmission of the disease in the Escuintla-Guatemala focus. As a result, local public health authorities can stop giving ivermectin and invest their human resources in other important diseases.
| Onchocerciasis (river blindness) is caused by a filarial nematode transmitted by black flies of the genus Simulium [1]. The disease may be mild (dermatitis) or severe (visual impairment and blindness) and is caused by the human immune response to microfilariae (mf) released by female adult worms as they move across subcutaneous tissue and spread throughout the body. Humans are the only known reservoir [2].
Onchocerciasis occurs throughout much of East and West Africa and Yemen, and was brought to the Americas through the slave trade [3]. It is now endemic to 6 countries in Latin America (Brazil, Colombia, Ecuador, Guatemala, Mexico and Venezuela). Foci of transmission in the Americas are relatively small and geographically delimited compared to areas of transmission in Africa [4]. In part due to the geographical isolation of foci, the goal of the Onchocerciasis Elimination Program of the Americas (OEPA) is both to eliminate ocular morbidity throughout the region, and to permanently interrupt transmission where possible [1],[5].
Control and eventual regional elimination of transmission is considered feasible due to the efficacy of ivermectin (Mectizan®, donated by Merck&Co, Inc.) as a microfilaricide, when used twice per year [6],[7]. While ivermectin used in this manner prevents transmission of infections, it does not kill adult worms [8] although it may reduce their fecundity and lifespan [9]. OEPA, along with its ministry of health counterparts, supports mass treatment with ivermectin twice per year, with the goal of reaching 85% of eligible individuals (those ≥5 years of age, ≥90 cm of height and ≥15 kg of weight; excluded are pregnant women and individuals with severe disease) living in endemic areas. Recent reanalysis of information on the effectiveness of ivermectin delivered in this strategy has suggested that six and a half years (13 treatment rounds) of such coverage can be sufficient to interrupt transmission [7].
Guatemala, with a population eligible for treatment of 175,881 (to receive 351,762 treatments) in 2006 [5] accounts for 38.5% of the endemic population eligible for treatment in Latin America. Guatemala has four endemic foci: Santa Rosa (Department of Santa Rosa), Huehuetenango (Department of Huehuetenango), Escuintla-Guatemala (Departments of Escuintla and Guatemala) and the Central Endemic Zone (Departments of Suchitepéquez, Sololá and Chimaltenango; Figure 1) [10]. The Guatemalan Ministry of Public Health and Social Welfare (MSPAS, in its Spanish acronym) has been delivering ivermectin to endemic communities, through mass drug administration (MDA), since 1988 [8] and has reached 85% of the eligible population at risk twice per year in all foci since 2001 (Figure 2) [5].
Beginning in 2004, the MSPAS, in partnership with OEPA, the US Centers for Disease Control and Prevention (CDC) and the Universidad del Valle de Guatemala (UVG), began evaluating three of the four endemic foci in the country to determine whether transmission had been interrupted in these areas and if semiannual treatment could be suspended. Criteria for making these determinations are based on World Health Organization (WHO) guidelines presented in the 2001 document “Certification of Elimination of Human Onchocerciasis: Criteria and Procedures” [11] as adapted for field conditions by Lindblade et al. [12] and the OEPA steering committee (the Program Coordinating Committee-PCC) [1]. In summary, the criteria to be applied in areas with historically documented onchocerciasis transmission include: 1) demonstration of a prevalence of mf in the anterior segment (MfAS) of the eye (anterior chamber and cornea) to be less than 1%; 2) a cumulative incidence of O. volvulus infection of less than 0.1% in school-age children; and 3) a prevalence of infection in vectors of less than 0.05% [11],[13]. Based on these criteria, Lindblade et al. demonstrated that transmission had been interrupted in the Santa Rosa focus; [12] subsequently, the PCC recommended to the Minister of Public Health of Guatemala that treatment be suspended in this focus. That recommendation was accepted and the Santa Rosa focus is currently under a three year post treatment surveillance phase to monitor for transmission recrudescence [5].
In this report, we evaluate the current status of transmission of O. volvulus in the Escuintla-Guatemala focus based on the adapted WHO criteria.
In 2007, the Escuintla-Guatemala focus consisted of 49,616 individuals at risk, with 45,224 eligible for treatment, divided among 103 communities in the department of Escuintla (14.30°N, 90.79°W), and 14 communities in the department of Guatemala (14.62°N, 90.53°W). Historically, this focus included areas with intense transmission: between 1979–1982, the community mf prevalence ranged from 8 to 38%. [14] A larval control effort from 1979–1989 significantly reduced both biting density and community mf prevalence.
To ensure that all areas with current or past evidence of onchocerciasis transmission were included in this evaluation, all communities with at least one of the following characteristics were identified using historical data, including unpublished reports from the MSPAS and published articles: a) past evidence of onchocerciasis transmission (nodules or mf in at least one community resident); b) suspicion of past transmission suggested by a documented survey, which may not have found evidence of transmission; or c) currently under semiannual ivermectin treatment by the MSPAS. A total of 155 communities satisfied at least one of these criteria) (Figure 3). These potentially endemic communities (PEC) served as the sampling frame for all evaluations of the status of transmission of onchocerciasis (Figure 3).
Because ocular morbidity is more likely to be found where onchocerciasis transmission is most intense, [15] we evaluated ocular lesions associated with onchocerciasis only in communities that historically had the highest rates of transmission. PEC with >0% nodule prevalence in the last 3 MSPAS surveys (1989, 1990 and 1991) and elevation >800 m were considered in order to maximize the possibility of finding O. volvulus related morbidity. A total of 16 communities satisfied these criteria. Two of these communities were dropped before the evaluation began because they had less than 5 inhabitants, and one was not included because it effectively serves as a bedroom community for Guatemala City, making it very difficult to locate potential participants in their homes.
The calculation of minimum sample size was based on estimating a population prevalence of MfAS of the eye of less than 1%. Finding 0 positive individuals out of 300 examined will allow a one-sided 95% confidence interval (CI) to exclude 1%. Given an estimated non-response rate of 10%, the total sample size required was 330.
All houses of the selected communities were mapped and the residents were censused using a pre-programmed hand-held personal digital assistant (PDA) with a global positioning system (GPS) attached. Eligible residents (those who were ≥7 years old and who had resided in the community for at least the last 5 years) were identified and recorded. As the ophthalmologist is capable of evaluating up to 90 individuals per day, we stratified communities into those with <90 eligible residents and those with ≥90 eligible residents. In the small communities (<90 residents), all eligible individuals (N∼266) were invited to participate. In the larger communities (≥90 residents), a PDA-based algorithm was applied in the field to randomly select 12% of the households and their members for inclusion in the evaluation (N∼223).
An ophthalmologist (OO) with extensive experience conducting evaluations of onchocerciasis-related eye disease performed the ophthalmologic evaluations. Visual acuity was measured with a Snellen chart using standard methods. Ocular examinations were conducted with a split-lamp in a darkened area after the patients were asked to sit with their head between their legs for 5 minutes [16]. MfAS were noted as live/coiled or dead/straightened. Data were entered in the PDA and later downloaded to a database for subsequent analysis.
We estimated the cumulative incidence of O. volvulus infection by measuring the prevalence of antibodies (IgG4) to a recombinant antigen of O. volvulus, OV-16, [17] in a stratified sample of school children 6–12 years old. We chose to stratify the sample into urban and rural schools because it was possible that levels of transmission would differ significantly between industrialized urban areas and rural communities located close to black fly breeding sites. The urban areas were taken to be the 2 large cities in the focus (San Vicente Pacaya and Palín), and the remainder of the schools located outside these cities were considered to be rural.
Information about schools and the number of children aged 6 through 12 who were attending was obtained with the help of the MSPAS and the Ministry of Education. Based on these figures, we estimated 4,674 eligible children in the urban schools and 9,815 in the rural schools. Schools were ordered at random within each stratum and then selected until the target sample size had been reached. The selected schools were visited and meetings were held with directors, teachers and parents to explain the evaluation. Teachers were asked to prepare a list of all enrolled children for the day of the evaluation.
Based on the WHO certification for elimination criteria (cumulative incidence <0.1%) and considering antibody prevalence equivalent to the cumulative incidence rate, 3,000 children were required in each stratum to calculate a one-sided 95% CI that excluded 0.1% when no seropositives were encountered. Given an expected 30% non-response rate, our target sample size was 4,286 in each stratum.
The methods used to collect finger-prick blood samples and data on residency from children participating in the evaluation have been described previously [12]. Briefly, each participant provided 80–120 uL of blood by standard sterile finger prick procedures. Whatman filter paper No. 2 was used to collect the blood directly after the finger prick. Children who didn't attend school on the appointed day were traced to their homes and asked to participate. Blood samples were processed within two months of collection using a standard ELISA [12].
Simulium flies were collected from November 2005 to April 2006 (peak biting season) in seven PEC and tested for infectivity in order to calculate the Seasonal Transmission Potential (infected stage larvae per person per season, STP). Collection sites were selected through a rapid assessment of PEC to find those that satisfied the following criteria: a) high densities of S. ochraceum; b) presence of appropriate collection sites that capture areas where residents are most likely to be exposed to vectors (i.e. casco, near a house and cafetal, near coffee plantations); and c) willingness of the owner (in the case of fincas [plantations]) or residents to participate.
We used similar methods described by Lindblade et al. [12]. Two teams of two people (collector and paid attractant, a male resident of the finca ≥18 years old) rotated between two collection sites (cafetal and casco) in each PEC. The paid attractants were given ivermectin before starting collections and had finger-prick blood samples taken on filter paper at the beginning and at the end of the evaluation to evaluate exposure to O. volvulus. Collections started at 8:00 AM and ended at 5:00 PM taking 10 minute breaks at the end of every hour and a 1 hour break at noon. Each community was sampled two days per month.
At the laboratory at the UVG, the heads and thoraces of S. ochraceum were separated from their bodies and up to 50 flies were pooled per tube, maintaining separate months and communities. The bodies were analyzed first using a standard polymerase chain reaction (PCR) assay to detect O. volvulus DNA [18]. Positives were confirmed by a second PCR. If a positive was confirmed, all the fly heads of that community were tested.
The one-sided 95% confidence intervals for the prevalence of MfAS and antibodies to Ov16 were calculated using the SAS (version 9.0, SAS Institute, Cary NC) FREQ procedure with the EXACT statement, BINOMIAL option and an alpha level of 0.10. The Poolscreen 2.0 program was used to calculate the proportion of infective flies based on the number of positive pools [19]. Biting rates and STP were calculated according to standard methods [12].
All protocols received appropriate review and approval by the CDC (Atlanta, GA), the ethics committee of the UVG (Guatemala City, Guatemala), and the MSPAS (Guatemala City, Guatemala). All participants ≥18 years of age or the parents or guardians of children <18 years old read or had read to them an informed consent form and then were asked to sign or mark with their finger to indicate their consent to participate. Children aged 7 to <18 years old were read or had read to them an assent form and asked to sign or mark with their fingerprint to indicate their willingness to participate. Paid attractants also read or had read to them a consent form and indicated their willingness to participate with their signature or fingerprint.
Of the 13 communities selected for the evaluation, one could not be reached due to road and weather conditions, and the only family in a second community could not be found on the day of the evaluation. We evaluated 329 (73.1%) of the 450 eligible residents selected for inclusion in the evaluation. Of the total evaluated, 55% were women and 36% were 7–15 years old. Blindness due to onchocerciasis was not observed in any of the patients and 306 (93%) individuals evaluated had their visual acuity measured in the range of 20/20–20/70. No MfAS were found; the prevalence of MfAS was therefore 0, with a one-sided 95% CI of 0–0.9%.
In the urban area, we registered 4,674 enrolled children in 24 local schools, and 3,130 (67%) participated in the evaluation. Due to an insufficient blood sample, the results for seven children could not be determined. Out of the 3,123 samples analyzed, there was 1 positive (a 9 year old male living in an urban area) for antibodies against O. volvulus. A second blood sample was requested and also tested positive. The sample was sent for additional testing at an experienced onchocerciasis laboratory in Mexico (Instituto Politécnico Nacional, Reynosa, Mexico) against recombinant antigens Ov10, Ov11 and Ov16; the sample did not test positive for any of these antigens and we therefore concluded that it was a false positive and have recorded it as a negative result. In the rural area, we registered 4,614 enrolled children in 34 schools, and 3,316 (72%) participated in the evaluation. A total of seven samples again had to be discarded due to insufficient sample. None of the 3,309 samples tested were positive for OV-16. Therefore, the prevalence of antibodies to Ov16 in the Escuintla-Guatemala focus was 0, and the one sided 95% IC was 0–0.05%.
A total of 28,423 Simulium flies were caught in 1,320 hours of sampling, from November 2006 through April 2007. None of the human attractants was found to be seropostitive for OV-16 either before or after the evaluation. Of the flies collected, 17,336 (61%) were S. ochraceum and 11,087 (39%) were S. metallicum (not considered to be a vector of onchocerciasis when community mf prevalence is low [20]). The highest biting densities were measured in November and December. A total of 14,099 S. ochraceum in 303 pools were tested for O. volvulus DNA by PCR; all pools were negative, thus, prevalence was 0% and the 95% CI was 0–0.01%. The geometric mean biting rate for S. ochraceum was 11.0 bites/person/hour while the arithmetic mean daily biting rate flies was 177 bites/person/day. As the proportion of infective flies was 0, the STP was also 0. To calculate the maximum potential STP, we used the upper end of the 95% CI of the proportion infective and the geometric mean biting rate to calculate the maximum STP, assuming that each infective fly would have 1 infective-stage larva. The calculated maximum potential STP was 1.0 infective stage larvae transmitted per person per season.
Our findings, based on the ophthalmologic, entomologic and serologic evaluations adapted from the WHO guidelines for certification of elimination, indicate that transmission of O. volvulus has been successfully interrupted in the Escuintla-Guatemala focus. Our studies in this formerly endemic area demonstrated that the prevalence of mf in the anterior segment of the eye was less than 1%, evidence of active or prior infection (or exposure) as measured by antibodies to a recombinant O. volvulus antigen (Ov-16) in school children was less than 0.1%, and O. volvulus DNA in vectors was under 0.05% (with a STP of 0 infective stage larvae per person per season).
O. volvulus transmission in the Escuintla-Guatemala focus was extensively documented from 1979 to 1984 by the Guatemala-Japan Cooperative Project on Onchocerciasis Research and Control, which conducted a large-scale larval elimination program in the area around the town of San Vicente Pacaya in the Department of Escuintla [21]. Several communities in that area had a prevalence of mf in the skin of 8% to 38% as recently as 1982 [14]. Nevertheless, community mf prevalence dropped from an average of 26% to 7% during the years of the larval elimination program [14]. Larval control efforts ceased in 1989, and a MSPAS survey in1991 found community mf rates of 3% (MSPAS, unpublished data; Figure 4). No surveys had been conducted in Escuintla until the current report. However, the MSPAS provided semiannual ivermectin treatments in the focus, reaching more than 85% of the eligible population at risk twice per year from 2002 to 2007 (Figure 2).
Santa Rosa was the first focus in the Americas to demonstrate interruption of onchocerciasis transmission, but there is evidence that transmission was extremely low to nonexistent prior to ivermectin distribution [5]. In contrast, levels of transmission in the Escuintla-Guatemala focus were historically higher than the Santa Rosa focus and transmission was well documented until at least the early 1980s. The larviciding efforts in the San Vicente Pacaya area from 1983–1989 were responsible for a significant decline in vector biting rates and, subsequently, community mf prevalence. The successful interruption of transmission after 12–13 rounds of MDA with ivermectin in San Vicente Pacaya may be at least partially due to the reduction in mf prevalence resulting from the larviciding campaign. While other areas of the Escuintla-Guatemala focus experienced a decline in mf prevalence without vector control, larviciding may be considered a complementary strategy to mass drug administration in areas of intense transmission.
Although the reported specificity of the Ov-16 ELISA test is 90%, [17] our laboratory has now tested over 9,964 samples from endemic areas with only 1 false positive, a specificity of 99.99%. However, distinguishing false from true positives is challenging. We undertook extensive interviews of the family of the child initially found positive to rule out travel to other endemic areas or potential exposure to vectors during his daily activities. We tested other family members, including a grandfather who reported a nodule that was extirpated in the past, and none was found positive. After a second blood sample from the same child tested positive, we sent the samples for testing in another laboratory against additional antigens, where the child's samples were negative in all external tests. We, therefore, feel confident reporting this finding as a false positive.
The data presented in this report were extensively reviewed by OEPA and the PCC. Based on the results, the PCC recommended to the Minister of Health of Guatemala that ivermectin treatments be suspended in the Escuintla-Guatemala focus in 2008. The recommendation was accepted, and, as in Santa Rosa, three years of surveillance for recrudescence has now begun during which a final set of evaluations to ensure that transmission has been completely eliminated will be undertaken [5].
Currently we are conducting a similar series of evaluations in the focus of Huehuetenango along the border with Chiapas, Mexico, to determine whether transmission has been interrupted there. Results from these studies are expected mid-2008. As of the writing of this article, transmission of O. volvulus continues in the Central Endemic zone of Guatemala [5].
|
10.1371/journal.pcbi.1004460 | Computational Identification of Mechanistic Factors That Determine the Timing and Intensity of the Inflammatory Response | Timely resolution of inflammation is critical for the restoration of homeostasis in injured or infected tissue. Chronic inflammation is often characterized by a persistent increase in the concentrations of inflammatory cells and molecular mediators, whose distinct amount and timing characteristics offer an opportunity to identify effective therapeutic regulatory targets. Here, we used our recently developed computational model of local inflammation to identify potential targets for molecular interventions and to investigate the effects of individual and combined inhibition of such targets. This was accomplished via the development and application of computational strategies involving the simulation and analysis of thousands of inflammatory scenarios. We found that modulation of macrophage influx and efflux is an effective potential strategy to regulate the amount of inflammatory cells and molecular mediators in both normal and chronic inflammatory scenarios. We identified three molecular mediators − tumor necrosis factor-α (TNF-α), transforming growth factor-β (TGF-β), and the chemokine CXCL8 − as potential molecular targets whose individual or combined inhibition may robustly regulate both the amount and timing properties of the kinetic trajectories for neutrophils and macrophages in chronic inflammation. Modulation of macrophage flux, as well as of the abundance of TNF-α, TGF-β, and CXCL8, may improve the resolution of chronic inflammation.
| A recent approach to quantitatively characterize the timing and intensity of the inflammatory response relies on the use of four quantities termed inflammation indices. The values of the inflammation indices may reflect the differences between normal and pathological inflammation, and may be used to gauge the effects of therapeutic interventions aimed to control inflammation. Yet, the specific inflammatory mechanisms that can be targeted to selectively control these indices remain unknown. Here, we developed and applied a computational strategy to identify potential target mechanisms to regulate such indices. We used our recently developed model of local inflammation to simulate thousands of inflammatory scenarios. We then subjected the corresponding inflammation index values to sensitivity and correlation analysis. We found that the inflammation indices may be significantly influenced by the macrophage influx and efflux rates, as well as by the degradation rates of three specific molecular mediators. These results suggested that the indices can be effectively regulated by individual or combined inhibition of those molecular mediators, which we confirmed by computational experiments. Taken together, our results highlight possible targets of therapeutic intervention that can be used to control both the timing and the intensity of the inflammatory response.
| Prolonged inflammation is a recognized contributor to a multitude of pathological conditions, including cardiovascular, metabolic, and neurodegenerative diseases, as well as chronic injuries [1, 2]. Timely resolution of inflammation is essential for tissue homeostasis. Inflammation resolution, previously believed to be a passive, self-regulatory process, is now known to be actively modulated by several different classes of endogenous molecular mediators, such as anti-inflammatory cytokines and growth factors [interleukin-10 (IL-10) and transforming growth factor-β (TGF-β)], oxygenated lipid mediators (lipoxins, resolvins, protectins, and maresins), and protease inhibitors [2–5]. Ongoing research efforts, including pharmacological animal model research and clinical trials, are focused on novel pro-resolution therapies for a variety of inflammatory conditions [6, 7]. There is a clear, documented need for new approaches to develop resolution-centric therapeutic interventions [2] to supplement or replace the currently used anti-inflammatory therapies, which are only modestly effective [8–10].
The concentrations of molecular and cellular components of the inflammatory process are typically characterized by single-peak temporal trajectories reflecting a distinct period of activation followed by resolution [2, 5, 11, 12] (Fig 1). The quantitative properties of these trajectories vary as a result of differences in inflammatory conditions and scenarios. Recently, four quantitative indices [namely, peak height (Ψmax), activation time (Tact), resolution interval (Ri), and resolution plateau (Rp)] (Fig 1) were introduced as informative measures to analyze the quantitative patterns characterizing temporal inflammatory trajectories [1, 13]. For a given molecular species or cell type, the Ψmax is defined as the maximum value of the corresponding temporal trajectory. Tact is the time after inflammation initiation required for the temporal trajectory to reach the Ψmax level. Ri is the time difference between Tact and the time it takes to reach 50% of Ψmax. Rp is the trajectory level at the end of the considered time period, expressed as a percentage of the Ψmax value; it therefore reflects residual inflammation. Based on the aspects of the temporal trajectories that they represent, the indices can be divided into amount indices (Ψmax and Rp) and timing indices (Tact and Ri) [14, 15]. Additionally, the area under the curve (AUC) of a kinetic trajectory has been recently introduced as a metric to quantitatively assess cumulative tissue damage under various inflammatory scenarios [16].
Experimental studies have demonstrated the utility of the inflammation indices in distinguishing between the time courses for normal and pathological inflammation, as well as in establishing the quantitative efficacy of external interventions [1, 3, 17]. For example, in a mouse peritonitis model, it was shown that the macrophage Ψmax, Tact, Ri, and the neutrophil Tact were noticeably higher (specifically, higher by 2-fold, 3-fold, 2-fold, and 12-fold, respectively) in the chronic inflammatory scenario compared to the acute inflammatory scenario [17]. In the same study, several drugs (specifically, ibuprofen, resolvin E1, a prostaglandin D2 receptor 1 agonist, dexamethasone, rolipram, and azithromycin) were evaluated for their inflammation resolution efficacy by quantifying their regulation of these inflammation indices in the chronic inflammation scenario. Another peritonitis mouse model study investigated the efficacy of drug-filled nanoparticles in inflammation resolution. The drug efficacy was established based on its ability to reduce the neutrophil Ψmax and Tact by ~1.5-fold [3]. Yet, systematic experimental characterization of the patterns and mutual relationships for the inflammation indices is challenging due to the immense diversity of possible inflammatory scenarios. This diversity hampers our understanding of the global control of the indices by specific molecular mechanisms, and our understanding of the modulation of the indices by therapeutic interventions. Computational modeling offers a possibility to complement experimental investigations and address this complex problem using an integrated approach.
Computational modeling can be used to “screen” thousands of inflammatory scenarios in a systematic way and thereby guide the generation of focused, mechanistic, experimentally testable hypotheses [11, 16, 18–21]. Previous works have demonstrated the utility of mathematical models in the study of inflammation in specific disease scenarios and in the identification of crucial inflammatory mechanisms [11, 12, 16, 20–32]. For example, computational models have been used to predict the efficacy of cytokine regulation therapies in chronic inflammatory diseases [24, 27, 33]. However, quantitative characterization of the inflammatory response of different cell types and molecular mediators in terms of indices, as well as a rigorous analysis of their mechanistic determinants, have not been carried out in previous modeling efforts. In the present study, we used our recently developed quantitative kinetic model of acute and chronic inflammation in wounds to generate hypotheses regarding mechanistic control of the inflammation indices [11]. Our model could capture the behavior of inflammatory cell type and molecular mediator kinetics in acute and chronic inflammation initiated by both infection and injury [17, 34, 35]. In our simulations, the chronic inflammatory scenarios were characterized by higher concentrations (Ψmax) and delayed resolution timing (Ri) for key inflammatory components in comparison with acute inflammatory scenarios. Using this model, we identified essential inflammation-driving mechanisms (specifically, macrophage influx and efflux rates) and informative indicators (i.e., IL-6, TGF-β, and PDGF) of chronic inflammation. Our findings regarding the mechanistic regulation of inflammation by macrophage fluxes are supported by experimental studies where wound macrophage levels (regulated by macrophage fluxes) controlled the timing of wound healing [36–39] and the quality of wound scarring [40]. Furthermore, our modeling predictions characterizing IL-6 as an informative indicator of pathological inflammation are consistent with a recent clinical wound study [41]. Here, we use this model to elucidate the functional relationships between specific molecular/cellular processes and inflammation indices during normal and pathological inflammation.
The goal of our study was to identify specific mechanistic determinants that can be targeted to modulate the index values during abnormal (delayed) inflammation and drive them toward a desired outcome. We used two complementary analyses (namely, sensitivity and correlation analysis) and identified such targetable mechanisms for regulating the inflammatory indices in the model. Furthermore, we wanted to test the effectiveness of cytokine inhibition as an intervention strategy to regulate the inflammation indices. For this purpose, we extended the model to represent cytokine inhibition kinetics for three cytokines (chosen based on our predictions regarding targetable mechanisms). We used this extended model to study the efficacy of individual/combined cytokine inhibition in the regulation of the inflammation indices for neutrophils and macrophages.
Our modeling results indicate that, for the majority of the model output variables representing inflammatory cell types and molecular mediators, the amount indices Ψmax and Rp were robustly regulated by the macrophage influx and efflux rate, respectively. In contrast, for the timing indices (i.e., Tact and Ri), such a robust functional dependence on single model parameters was not detected in the sensitivity analysis. Yet, the timing indices for the inflammatory components were strongly correlated with the platelet degradation rate. Moreover, strong correlations between the timing indices and certain mechanistic processes existed, but only under specific inflammatory situations representing chronic inflammation. Our inflammatory mediator inhibition modeling suggested that during an abnormal (delayed) inflammatory response, TNF-α and TGF-β inhibition strongly shifted the macrophage Ψmax and Rp indices toward inflammation resolution, whereas CXCL8 inhibition regulated the neutrophil Ri and could nearly restore this index to its normal (i.e., acute-inflammation) value. Notably, combined inhibition of TNF-α and CXCL8 resulted in improved restoration of normal neutrophil dynamics compared with the inhibition of these two targets acting independently. Comparisons with available experimental data provided validation for our TNF-α inhibitor modeling results.
To investigate the regulation of the inflammation indices via changes in the model parameters, we performed a local sensitivity analysis for the model in 10,000 inflammatory scenarios (simulations), as described in the Materials and Methods Section. We regarded the regulation of an inflammation index (computed for a given model output variable) by a specific parameter as robust, if the sensitivity sij (Eq 1, defined in the Materials and Methods Section) of the index with respect to this parameter was the highest (or second highest, third highest, etc.) across all parameters in the majority of the 10,000 simulations. In the figures and tables, for the sake of brevity, we identify each of the model’s 69 parameters using its assigned number in the model preceded by the prefix “P#.” We use these designations when referring to the figures and tables in the text and, in addition, provide descriptive names for the parameters (see Table 1 for a full list of the model parameters).
The amount indices, Ψmax and Rp, for the majority of the model output variables were robustly regulated by the parameters representing the rate of macrophage influx (P#7) into the inflamed area and the rate of their efflux (P#10), respectively. Table 2 shows the first, second, and third most influential (based on the ranking of the corresponding sensitivities) model parameters for the inflammation indices for six specific model output variables. These variables (namely, TNF-α, IL-1β, IL-6, IL-10, total neutrophils, and total macrophages) are the ones that typically demonstrate abnormal characteristics during pathological inflammation [17, 34, 36, 42]. By definition, Ψmax and Rp characterize the inflammation intensity at its peak and during its resolution (Fig 1). Therefore, these results are consistent with, and complement, our earlier findings identifying macrophage influx (P#7) and efflux (P#10) rates as the main regulators of kinetic trajectories during the initial and final phases of inflammation, respectively [11]. In addition to the well-pronounced regulation by macrophage flux rates, our local sensitivity analysis identified the TNF-α degradation rate (P#19) and the neutrophil influx rate (P#3) as robust regulators of the Ψmax for the TNF-α and the neutrophil (Ntot) model output variables, respectively (Table 2).
In contrast to the robust regulation of the amount indices, no model variables had their timing indices (i.e., Tact and Ri) robustly regulated by a model parameter. Indeed, the largest sensitivity values for Tact and Ri corresponded to different parameters depending on the specific simulation (i.e., on the specific model parameter set). For all of the16 variables, the model parameter effecting the strongest Tact regulation in the largest fraction of the 10,000 inflammatory scenarios was the platelet degradation rate (P#1) (Table 2). Yet, this largest fraction (which we call the “robustness fraction”) was too small (~18–46%) to consider this regulation robust. For the Ri, the parameter effecting the strongest regulation of this index in the largest fraction of scenarios was macrophage efflux rate (P#10) for 10 out of the 16 model output variables. However, the corresponding robustness fraction for it was even smaller (~15–35%). These results suggest that robust control of the timing indices may require a sophisticated strategy involving simultaneous modulation of multiple inflammatory mechanisms.
To gain additional insights into the regulation of the inflammation indices, we calculated the correlation coefficients (CCs) and the associated p-values between the inflammation indices of each model output variable and each model parameter (see Materials and Methods). The CC values ranged between −1 to +1 reflecting a negative or a positive index-parameter correlation, respectively. The signs of the CCs for the four indices are shown in S2 Fig. Absolute index–parameter correlations above 0.5 were considered strong [43, 44]. Among the correlations identified as strong, only correlations with p ≤ 0.05 were considered statistically significant.
For the amount indices, we detected a strong positive correlation between inflammation peak height (i.e., Ψmax) and macrophage influx rate (P#7) for 10 model output variables, including macrophages, IL-6, and IL-10 (Fig 2a and 2b). Furthermore, we found a strong positive correlation between the Ψmax for neutrophils (Ntot) and neutrophil influx rate (P#3) (Fig 2a and 2b), which was consistent with our sensitivity analysis results (Table 2). This result suggests that the inflammation indices for the neutrophil inflammatory response trajectory may depend on the degradation rates of CXCL8 (P#31) and TGF-β (P#13), which are the two main neutrophil chemoattractants in our model [11]. For the other amount index, i.e., Rp, only one parameter [namely, macrophage efflux rate (P#10)] showed a strong (negative) correlation with a large number (namely, 13) of model output variables (Fig 2c and 2d). These findings were consistent with our sensitivity analysis results regarding the Ψmax and Rp regulation by macrophage influx and efflux rates (Table 2).
For the timing indices, the correlation analysis highlighted a strong negative correlation between TGF-β degradation rate (P#13) and the Tact of four model variables, including neutrophils (Ntot) and IL-6 (Fig 3a and 3b). Moreover, we detected a strong negative correlation between the platelet degradation rate (P#1) and inflammation activation time (i.e., the Tact index) for 9 model variables (Fig 3a and 3b). However, this relationship simply demonstrates the connection between Tact and the strength and persistence of inflammation-initiating stimuli, reflected in our model by the presence of platelets in the wound. Additionally, the Ri for the majority of the model variables, including macrophages and IL-6 (results not shown), exhibited a strong negative correlation with macrophage efflux rate (P#10), which could be expected based on our sensitivity analysis results (Table 2). Yet, our correlation analysis did not detect any other strong correlations for the timing indices.
The limited number of detected strong correlations prompted us to hypothesize that a larger number of strong correlations between the timing indices and the model parameters could be detected in a smaller set of simulations reflecting specific conditions, such as chronic inflammation. To test this, we divided the 10,000 simulations into two subsets, “acute” and “chronic” (see Materials and Methods). Then, for only the “chronic” subset, we performed the correlation analysis between each of the timing indices (i.e., Tact and Ri) for the 16 model output variables and the 69 model parameters. As hypothesized, in this analysis many model parameters emerged as strongly correlated with Tact (18 parameters) and Ri (9 parameters), for different model variables (Fig 3c and 3d, respectively). In subsequent analyses, we specifically focused on TGF-β and CXCL8 degradation rates (P#13 and P#31, respectively), because they were strongly correlated with several key model outputs and those correlations were statistically significant. In summary, the correlation analysis provided us with candidate mechanisms for directly regulating both of the amount indices (via neutrophil influx rate and the macrophage flux rates) and the timing indices (via the TGF-β and CXCL8 degradation rates) of the model output variables.
The goal of model parameter randomization in our 10,000 simulations was to account for possible biological variability in inflammation scenarios [11]. While wider variation ranges for the parameters may allow for a fuller representation of this variability, such random, simultaneous, uncorrelated variations may also introduce excess noise that could mask the biologically relevant patterns we want to detect. An informative analysis should therefore utilize parameter ranges sufficiently wide to represent variation and yet narrow enough to define the vicinity of our carefully chosen default parameter set, which had been derived directly from experimental data and represented the expected “typical” injury-triggered inflammation scenario [11]. To assess the impact of large parameter deviations, we performed the sensitivity and correlation analysis for an additional set of 40,000 simulations. In these simulations, the parameters were randomly and uniformly sampled from a 9-fold range (i.e., 3-fold down and 3-fold up) around the default parameter values. In the sensitivity analysis, the introduction of these larger parameter deviations reduced by 5–30% (results not shown) the robustness of the Ψmax regulation by macrophage influx rate for the six different outputs shown in Table 2. Similarly, the robustness of the Rp regulation by macrophage efflux rate was reduced by 1–14% (results not shown) in comparison with the results shown in Table 2. However, in the correlation analysis, the major trends [i.e., the strong correlation between the Ψmax and macrophage influx rate and the strong correlation between the Rp and macrophage efflux rate] were preserved, while other strong correlations (Figs 2 and 3) were not detected in the case of large parameter deviations (S3 Fig). Thus, it appears that some of the detected biological patterns are particularly strong in the fewfold vicinity of our default parameter set, which reflects their dependence on specific type of inflammatory scenario.
The results of our sensitivity and correlation analyses (Table 2 and Figs 2 and 3) suggested that specific inflammation indices of the model variables can be considerably impacted by the following five model parameters: macrophage influx and efflux rates and the TNF-α, CXCL8, and TGF-β degradation rates. We therefore wanted to investigate whether modulation of these parameters during chronic inflammation could result in (at least, partial) restoration of the normal (i.e., acute-inflammation) values of the respective inflammation indices. We addressed this question for all of the five parameters except macrophage influx rate, because we used its modulation to induce chronic inflammation in the model (see caption for Fig 4). Using the same protocol as implemented in our previous modeling study (see Figure 5 in [11]), we simulated a chronic inflammatory scenario (Fig 4a and 4b, red line). Here, we specifically focused on the kinetic trajectories for neutrophils and macrophages, because their kinetic behavior is well studied and is known to be disrupted during delayed (or chronic) inflammation [34, 45].
To investigate the regulation of neutrophil and macrophage inflammatory trajectories, we repeated the chronic inflammation simulations upon modifying the values of the parameters mentioned above, as follows. In the chronic inflammation simulation for neutrophil regulation, we implemented two distinct parameter modification strategies. In one strategy, we increased the CXCL8 degradation rate, which caused the Ri for the neutrophils to decrease (Fig 4a, solid black line and green values in the table) compared to the chronic scenario with no modification (Fig 4a, red line). In the other strategy, we increased the TGF-β degradation rate, which caused both Tact and Ψmax for the neutrophils to decrease (Fig 4a, dashed black line, green values in the table) compared to the chronic scenario with no modification (Fig 4a, red line).
Similarly, for macrophage regulation, we introduced two distinct parameter modification strategies into the chronic inflammation simulation. In the first strategy, we increased the TNF-α degradation rate, which caused the Ψmax for the macrophage variable to decrease (Fig 4b, solid black line and green values in the table) compared to the chronic scenario with no modification (Fig 4b, red line). In the second strategy, we increased the macrophage efflux rate, which caused the Tact, Ri, and Rp for the macrophage variable to decrease (Fig 4b, dashed black line and green values in the table) compared to the chronic scenario with no modification (Fig 4b, red line). The observed decrease in the inflammation index values suggested that the considered parameter modifications can partially restore acute inflammatory kinetics, i.e., can bring the inflammation index values for a chronic inflammatory scenario closer to those for an acute inflammatory scenario.
The analyses described above were supplemented with modulation of other parameters. This was done to illustrate that not all model parameters exhibiting strong correlations with certain model variables could effectively regulate their indices. For example, the IL-10 production rate parameter (P#26) was strongly and negatively correlated with the macrophage Tact (Fig 3c), and the parameter for TNF-α inhibition by TGF-β (P#55) had a strong and negative correlation with the neutrophil (Ntot) Ri (Fig 3d). However, increasing those two parameters did not result in a significant modulation of the respective neutrophil or macrophage indices (Fig 4a and 4b, dotted lines). The model parameters that effected weak inflammatory index regulation or exhibited low-confidence correlations with the inflammatory indices were not selected for further analysis.
Because the TGF-β, CXCL8, and TNF-α degradation rates represent inherent biochemical properties of the respective molecular mediators, they cannot be easily changed in vivo. We therefore wanted to test whether a mechanistically distinct process, such as cytokine inhibition, could be used to obtain functionally similar outcomes. For this purpose, we extended our model [11] to include the kinetics of inhibitors for TGF-β, CXCL8, and TNF-α and performed inhibition kinetics simulations for chronic inflammatory scenarios (Figs 5 and 6 and S1). We derived the values for the association (kon) and dissociation (koff) rate constants (Eq 2, defined in the Materials and Methods Section) for each mediator inhibitor from literature data (S1 Table). In our simulations, CXCL8 inhibition primarily regulated inflammation timing (specifically, the Ri index for the neutrophil variable; Figs 5a and 6a, black lines), whereas TNF-α and TGF-β inhibition primarily regulated inflammation intensity (specifically, the Ψmax and Rp indices for both neutrophil and macrophage variables) (Figs 5b and 5e and 6b and 6e; S1a and S1d Figs, black lines).
For each inhibitor, we performed simulations for three different inhibitor concentrations. We used inhibitor concentrations of 10 nM, 100 nM, and 500 nM for both the TNF-α and CXCL8 inhibitors. For the TGF-β inhibitor, we used the concentrations of 1 nM, 20 nM, and 200 nM. We found that the inhibitors for different targets are most effective within concentration ranges that can be vastly different (Figs 5 and S1). Indeed, the CXCL8 and TNF-α inhibitors were most effective in restoring (at least, partially) their target indices when the inhibitors were added at concentrations ≥100 nM (Fig 5a, 5b and 5e), while TGF-β inhibition was effective for inhibitor concentrations in the range ~1–10 nM (S1 Fig). These results attest to the efficacy of the simple strategy involving the inhibitor addition at time 0.
To validate our modeling predictions regarding TNF-α inhibition, we compared our simulations with the experimental data generated for H1N1 virus-induced lung inflammation regulated by the TNF-α inhibitor etanercept [46]. These data characterized three scenarios: 1) control (i.e., no infection and, therefore, no inflammation), 2) inflammation with infection without added TNF-α inhibitor, and 3) inflammation with infection and with added 200 nM TNF-α inhibitor. From these data, we calculated the ratios of the total neutrophil and macrophage concentrations for the inflammation scenario without the TNF-α inhibitor to the corresponding concentrations for the inflammation scenario with added TNF-α inhibitor. We compared these ratios with the respective ratios calculated from our model simulation of injury-induced chronic inflammation. There was a reasonable agreement between the simulation-derived and experiment-derived ratios (Table 3). Note that, in our analysis, we did not attempt to model the nonzero neutrophil and macrophage concentrations detected for the experimental control scenarios, because in our model, which represents extravascular space, the concentrations of neutrophils and macrophages in the absence of inflammation are zero. In contrast, the control experiments measured the cell concentrations in entire lung lobes, which included cells present in the vasculature, resulting in nonzero neutrophil and macrophage concentrations even in the absence of inflammation. We chose this particular experimental study for the validation because TNF-α is a key pro-inflammatory cytokine that is secreted by inflammatory cells in most inflammatory scenarios, and whose expression and regulatory functions are largely independent of the specific inflammation-inducing stimuli (e.g., viral load, LPS [47], and wounding [39]). Thus, despite the difference in the inflammation-inducing stimulus between the experimental study and our modeling study (viral loading in lungs vs. injury, respectively), the reasonable agreement of the neutrophil and macrophage kinetics between the two studies suggests that our model adequately captured TNF-α inhibition.
To test whether the timing of inhibitor administration can impact the index restoration outcomes, we performed simulations in which the inhibitors were added at three different time points (i.e., 24, 48, and 72 h) during chronic inflammation. We chose 24 h as the earliest intervention time point because neutrophils are the first blood leukocytes to arrive at the inflammation site, and 24 h approximately corresponds to the neutrophil peak time for acute inflammatory response [35, 48]. Macrophages peak at 48 h, which motivated our choice of the second intervention time. For all the simulations, the added CXCL8 and TNF-α inhibitor concentrations equaled 200 nM (selected based on the observed effective range >100 nM, Fig 5a, 5b and 5e). In the case of CXCL8, the inhibitor added at 24 h was more effective in reducing the Ri for the neutrophil variable than the inhibitor added at other time points (Fig 6a, dotted black line). For TNF-α, the inhibitor added at 24 h provided a more complete restoration of the Ψmax for both neutrophils and macrophages to its acute-inflammation value than the TNF-α inhibitor added at later time points (Fig 6b and 6e, dotted black lines).
A nearly identical effect on the neutrophil Ri and macrophage Ψmax was observed when the CXCL8 and TNF-α inhibitors, respectively, were added at 48 h (Fig 6a and 6e, dashed black lines nearly overlapped the dotted black lines). Mediator addition at 72 h showed the least degree of restoration in the neutrophil Ri and macrophage Ψmax among the three inhibitor addition times analyzed (Fig 6a and 6e, black solid lines). However, this observed effect of mediator inhibition on the inflammation indices was not monotonic. Specifically, when the CXCL8 and TNF-α inhibitors were added at 36 h (results not shown), the degree of restoration in the neutrophil Ri and macrophage Ψmax values, respectively, was less than that when the mediator inhibitors were added at the 48 h time point. Experimental data from cytokine inhibition studies [49, 50] support the possibility of this type of non-monotonic behavior, which may be due to the complex nonlinear functional dependencies at work in the system. In summary, the CXCL8 and TNF-α inhibitors were characterized by similar optimal-efficiency concentration ranges and preferred administration timing regimens, but distinct preferentially regulated inflammation indices. Among the three time points considered, mediator inhibition was most effective when introduced during peak neutrophil response (i.e., at 24 h). For TGF-β, however, the inhibition outcomes were more complicated (see S1 Fig and S1 Text).
Because the neutrophil is the primary microbicidal cell type that produces powerful cytotoxic molecules that can potentially destroy the surrounding healthy tissue, (at least partial) restoration of the “normal” (i.e., acute-inflammation) timing and intensity of the neutrophil surge is essential for the resolution of chronic inflammation [51]. Based on the preferential regulatory action of individual CXCL8 and TNF-α inhibitors on the total neutrophil Ri and Ψmax, respectively (Fig 5a and 5b), we hypothesized that simultaneous inhibition of these two mediators in chronic inflammation might induce simultaneous restoration of the normal values for both of these indices. We tested this hypothesis by simulating the effect of adding both CXCL8 and TNF-α inhibitors at time 0. Each of the two inhibitors was added at a concentration of 200 nM. As hypothesized, the combined inhibition resulted in simultaneous restoration of both the Ri and Ψmax of the total neutrophil trajectory in a chronic inflammation simulation (Fig 5c, solid black line). Similarly to the action of the CXCL8 inhibitor (Fig 6a), the combined inhibition of CXCL8 and TNF-α was most effective at 24 h (Fig 6c and 6f, dotted black lines). These results suggest that “cocktails” of therapeutic agents with different targets may provide strategies of improved efficacy for simultaneous control of both timing and intensity of inflammation after wounding.
Timely resolution of inflammation following an injury, infection, or disease is essential for the maintenance of healthy tissue [2, 52]. Despite ongoing research and development efforts, current anti-inflammatory therapies are only modestly effective and have significant negative side effects [9, 53, 54]. This work was motivated by the need to identify molecular mechanisms that could serve as targets for intervention strategies intended to modulate chronic inflammatory responses. Chronic inflammation may have distinct phenotypic manifestations depending on the inflammatory condition, e.g., increased apoptotic neutrophil levels in diabetic ulcers [36], or increased levels and prolonged presence of classically activated macrophages in ischemic wounds [45], or prolonged oscillations in the levels of inflammatory cells and cytokines [55]. Here, we focused on chronic inflammation characterized by heightened levels and/or delayed resolution timing for kinetic trajectories describing accumulation and depletion of the inflammatory cell types and molecular mediators in comparison with acute inflammation [2]. In our simulations (both acute and chronic), the kinetic trajectories of all the inflammatory components return to their baseline levels following a peak triggered by inflammation initiation. We used our computational model of injury-initiated local inflammatory response [11] to identify the mechanistic factors regulating the four inflammation indices, i.e., Ψmax, Rp, Tact, and Ri (Fig 1), which were recently introduced as quantitative markers of the levels and timing of the inflammation time course [1, 3, 13, 17].
Our sensitivity analysis elucidated the effects of parameter changes on the inflammation indices. For the majority of the model output variables, the amount indices (i.e., Ψmax and Rp) were robustly regulated by macrophage influx (P#7) and efflux (P#10) rates, respectively, in the 10,000 performed simulations (Table 2). Because macrophages are major cytokine-producing cells at the site of inflammation [39], identification of the macrophage flux rates as critical inflammation intensity control mechanisms is, perhaps, not surprising. However, the kinetics of each cytokine are additionally affected by a large number of factors, such as its individual production and degradation rates, feedback parameters, cell phenotype conversion rates, etc. In view of this diversity, identification of the macrophage flux rates as the only robust modulators for almost all model variables is an unexpected result. Interestingly, these results are supported by experimental studies involving direct macrophage manipulation [36, 37, 39, 40]. For example, elevated macrophage abundance was shown to cause wound fibrosis [40] and delayed wound healing [38]. Moreover, reducing the levels of functional macrophages in wounds can severely delay the wound healing time in mouse models [36, 37, 39].
The macrophage flux rates, being strong regulators of the intensity inflammation indices, (Table 2 and Fig 2), are in fact the primary targets of certain recently emerged pro-resolution molecular mediators, such as lipoxins and resolvins [4]. By perturbing the recruitment mechanisms of neutrophils and macrophages in our model (Fig 4), we were able to reproduce the experimentally observed effects of exogenous delivery of resolvins and lipoxins in models of murine peritonitis (Fig 4, black lines and green values in the tables), e.g., a reduction in the total neutrophil Ψmax and Tact (see Table II in [1]), and a reduction in the total macrophage Ψmax (see Figure 5 in [17]). While these experiments addressed only a handful of specific inflammatory scenarios, the robustness of the inflammation index regulation by the macrophage flux parameters, detected in our sensitivity analysis (Table 2), attests to the general nature of this regulation type. It should be noted, however, that the macrophage influx rate parameter (P#7) in our model is a lumped parameter representing macrophage chemotaxis stimulated by any combination of four distinct macrophage chemoattractants (specifically, TGF-β, PDGF, TNF-α, and MIP-1α) (see Table 1 in [11]). Likewise, the neutrophil influx rate parameter (P#3), which robustly regulates the total neutrophil Ψmax (Table 2), characterizes the chemotaxis of active neutrophils stimulated by any combination of TGF-β and CXCL8. Thus, the sensitivity analysis alone [which identified P#7 and P#3 as influential parameters (Table 2)] was not sufficient for obtaining deeper insights into the specific mechanistic factors regulating the macrophage and neutrophil influx rates. This was accomplished in our parameter/output correlation analysis.
The correlation analysis explored the effects of simultaneous, random parameter variations within specified limits. The analysis identified a number of functional associations between model parameters and inflammation indices (Figs 2 and 3). Consistently with the sensitivity analysis results, macrophage influx and efflux rates were strongly correlated with Ψmax and Rp of model outputs (Fig 2). However, in contrast to the sensitivity analysis results, where the timing indices (i.e., Tact and Ri) did not display robust regulation by any of the kinetic model parameters across the 10,000 simulations, a strong negative correlation was observed between the platelet degradation rate and the Tact of the majority of output variables (Fig 3a and 3b). This correlation may simply reflect that the presence of platelets in the wound area is the only inflammation-initiating stimulus in our model. While other mediator-releasing cells (e.g., endothelial cells, resident macrophages, mast cells, etc. [56]) may contribute to the inflammatory response, this simplifying modeling assumption is in accord with the prominent roles played by platelets in initiating wound healing [57, 58]. Of the many platelet-secreted molecular mediators facilitating inflammation initiation [57], we only modeled the most potent neutrophil and macrophage chemoattractant, i.e., TGF-β [59, 60]. The results of this choice are evident from the strong correlations between the TGF-β degradation rate and key inflammatory components, such as total neutrophils, pro-inflammatory macrophages, and IL-6 (Fig 3b).
Overall, Ri was the least sensitive of all the inflammation indices and had a limited response to the parameter variations applied, which was consistent with its experimentally detected insensitivity [1]. Notably, for a subset of simulated scenarios reflecting chronic inflammation, certain parameters (e.g., TGF-β and CXCL8 degradation rates) exhibited a strong correlation with the Tact and Ri values for many model output variables, including key outputs such as total neutrophils, TNF-α, and IL-6 (Fig 3c and 3d). Thus, our correlation analysis identified relationships between specific inflammation indices and the specific molecular components regulating the macrophage (TGF-β, Fig 3c) and neutrophil (TGF-β and CXCL8, Fig 3c and 3d) influx.
We extended our model to describe inflammatory mediator inhibition kinetics in order to mechanistically represent the regulatory effects of inhibiting TNF-α, CXCL8, and TGF-β (Figs 5 and 6; S1 Fig, black lines). Mediator inhibition could provide a pharmacologically feasible strategy to mimic the functional effects of the modulation of the neutrophil and macrophage flux mechanisms, as well as of the modulation of TNF-α, CXCL8, and TGF-β degradation rates, which we identified as robust regulators of specific inflammatory indices. In our simulations, individual and combined mediator inhibition demonstrated the potential for significantly improved restoration of acute-scenario kinetics for specific cell types in certain injury-initiated chronic inflammatory scenarios (Figs 5 and 6). Interestingly, the molecular mediators TGF-β and CXCL8, which were identified as strong regulators of the timing indices Tact and Ri (Fig 3), successfully regulated the amount indices Ψmax and Rp of neutrophils and macrophages (Figs 5 and S1). This suggests that CXCL8 and TGF-β may be the primary contributors to the neutrophil and macrophage abundance in the wound, respectively. Yet, the detected neutrophil Ψmax regulation by TNF-α inhibition (Fig 5) was surprising because, unlike CXCL8, TNF-α is not a chemoattractant for neutrophils. TNF-α has been reported to be pro-apoptotic to neutrophils in specific concentration ranges [61]. It is conceivable that the non-linear feedback effects present in our model allowed us to capture this indirect effect of the neutrophil Ψmax regulation by TNF-α.
The detected difference in regulation robustness between the timing and amount inflammation indices appears to be an intrinsic property of the inflammation regulation system. Interestingly, a regulation dichotomy between timing and amount properties of the output kinetics has been detected for other systems in molecular biology. For example, in bacterial signal transduction systems, the amount characteristics of the response curves are primarily defined by the numerical values of the systems’ kinetic parameters, whereas the response timing properties are determined largely by the systems’ architecture and are less sensitive to parameter variation [14]. Therefore, a lack of robust patterns characterizing the control of response timing via kinetic parameter modulation might be a common property of biological control systems.
Traditional anti-inflammatory therapies, such as nonsteroidal anti-inflammatory drugs targeting the production of prostaglandins [10], have focused on the inhibition of pro-inflammatory pathways. More recently, administration of pro-resolution molecular agents, such as lipoxins and resolvins, has shown promise in promoting inflammation resolution and is now a topic of active research [4, 62]. Direct inhibition of inflammatory mediators is another therapeutic strategy that has seen moderate success for specific inflammatory conditions. For example, TNF-α inhibition and IL-1β inhibition have been successful as therapies for rheumatoid arthritis (RA) [8, 9] and chronic inflammatory pathologies associated with cancer [63], respectively. Previously published reports suggest that single mediator inhibition strategies are useful predominantly in chronic inflammation scenarios driven by specific mediators and/or when only one inflammation system component is affected, e.g., in the neutrophil overload condition [36]. Our results are consistent with these findings and suggest that the observed therapeutic effects can be explained by preferential regulation of a given inflammation index by a specific mediator. Indeed, TGF-β and TNF-α inhibition regulate the total neutrophil and total macrophage Tact and Ψmax (Figs 4a and 4b and 5b and 5e and 6b and 6e and S1a and S1d), while CXCL8 inhibition generally regulates the total neutrophil Ri (Figs 5a and 6a).
The generality of our approach suggests that TNF-α inhibition could potentially be used (alone or in combination with other mediator inhibitors) for a variety of inflammatory conditions characterized by elevated neutrophil and macrophage levels, such as traumatic injuries [64, 65]. TGF-β regulation has largely been studied for its role in fibrosis during the later stages of wound healing [66, 67]. Our results show that TGF-β inhibition might have a role in promoting the resolution of the crucial inflammatory stage of wound healing. In fact, recent clinical trials have begun evaluating the efficacy of TGF-β inhibition for various inflammatory pathologies, such as cancer [68]. CXCL8 inhibition has until now been speculated to have a role in inflammatory airway diseases, such as chronic obstructive pulmonary disease [69] and severe asthma [53]. Our results indicate a potential role for CXCL8 inhibition in timely neutrophil resolution during injury-initiated inflammation that warrants further experimental investigation. For pathological situations affecting several components of the inflammation process, the use of combination mediator therapy has been recently proposed and is being tested in clinical trials for rheumatoid arthritis [54]. Encouragingly, our simulations showed that a combined intervention targeting both TNF-α and CXCL8 in a chronic inflammatory scenario reduced both Ψmax and Ri for the total neutrophil concentration and thereby restored the near-“normal” (in comparison with the acute inflammatory scenario) total neutrophil kinetic trajectory (Figs 5c and 6c). Overall, our modeling results suggest that individual and combined mediator inhibition approaches may be applicable during trauma-induced chronic inflammation, in addition to the inflammatory conditions for which they are currently indicated or being tested. Our model may be used to test the efficacy of different intervention strategies, involving combined administration of inhibitors for TNF-α, TGF-β, and CXCL8, and to possibly determine the intervention timing. The use of robust control methodologies with our model might facilitate the design of improved parameter selection strategies to simulate chronic inflammation and to optimize drug administration regimens [70].
The main limitations of our approach arise from the computational nature of our study and from the assumptions made during the computational model development [11]. First, in our sensitivity analysis, we only examined the effects of parameter variations in the vicinity of the default parameter set. Larger (e.g., several‐fold) parameter changes, which could represent a higher degree of dysregulation or severity of inflammation, were not examined. Nevertheless, local sensitivity analysis is an established research methodology that allowed us to generate model results consistent with experimental data (Table 2) [11, 19, 71–74]. While this methodology is not designed to investigate simultaneous variations in several parameters, simultaneous parameter variation was a part of our correlation analysis strategy. Moreover, the effects of simultaneous parameter variations can be understood using global sensitivity analysis, as was recently done for a model of acute inflammation [16]. Second, our model contained a simplified representation of the interaction between inhibitors and their targets. Specifically, we did not include the individual degradation/removal kinetics for the inhibitors, which reflects the assumption that the inhibitors are present at the inflammation site at nearly constant total concentrations. While this assumption may be only an approximation to the in vivo situation, it allowed us to focus our inhibition analysis specifically on the inhibitor-target interactions (rather than potentially complex inhibitor pharmacokinetics). Besides, an increase in the number of modeled components and their interactions would have caused an increase in the number of unknowns, thereby introducing additional uncertainty, which we wanted to avoid at this stage of the inhibitor modeling. Finally, our model did not explicitly represent the action of the recently emerged lipid mediators of inflammation resolution, such as lipoxins and resolvins [1, 17]. Yet, our modeling approach allowed us to implicitly represent their effects by modifying their suggested target mechanisms (see Figure 8 and Table II in [1]).
With the recognition of inflammation as the key contributor to several pathologies, new approaches are needed to identify mechanistic regulators of pathological inflammation. A promising methodology is based on the analysis of the inflammation indices that quantitatively characterize the shapes of inflammation trajectories. Our study shows the applicability of systems biology approaches to identify mechanistic regulators of the inflammation indices. Moreover, computational models can provide a non-invasive and cost-effective framework for testing the efficacy of potential therapeutic strategy aimed to improve inflammation resolution. Computational analyses employing robust control strategies can guide the development of focused, hypotheses-driven experimental and clinical studies by reducing the ambiguity in the timing and strength of therapeutic interventions.
To simulate the inflammatory response, we used our recently developed quantitative model of acute and chronic local inflammation in a wound [11]. The model reflects initiation of inflammation by the platelets present at the site of injury; the platelets release TGF-β, whose gradient attracts inflammatory cells that interact and release soluble inflammatory mediators. The model’s variables describe the kinetics of 5 types of inflammatory cells (namely, active and apoptotic neutrophils, pro- and anti-inflammatory macrophages, and platelets) and 11 molecular mediators [namely, TNF-α, interleukin(IL)-1β, IL-6, IL-12, IL-10, CXCL8, TGF-β, platelet derived growth factor (PDGF), macrophage inflammatory protein-1α and 2 (MIP-1α, MIP-2), and interferon gamma-induced protein 10 (IP-10)]. We chose to model these components because they are widely regarded as essential cell types and molecular mediators involved in an innate immune response to injury [48, 75]. In our analysis, we did not include inflammatory molecular mediators produced by cells of adaptive immunity (e.g., IL-4 and IFN-γ) that are reported to affect innate immune components, such as macrophages [40]. For modeling tractability, we reduced the wide spectrum of possible wound macrophage phenotypes [34, 40] to just two: pro-inflammatory (similar to the “classically activated,” or M1, phenotype induced in vitro by IFN-γ and bacterial LPS) and anti-inflammatory (similar to the “alternatively activated,” or M2, phenotype induced in vitro by IL-4 and IL-13) phenotypes. The essential mechanisms (i.e., chemotaxis and phenotype conversion of inflammatory cells, cellular apoptosis, and molecular mediator production/degradation and positive/negative feedback effects) that govern the kinetics of the inflammatory response are represented via 69 model parameters (Table 1). The model describes extracellular signaling between different cell types and cytokines. Intracellular processes, such as transcription, translation, and export of proteins, are not represented mechanistically and are instead described implicitly via the model’s rate parameters.
Our model reflects experimentally observed acute and chronic inflammatory response kinetics initiated by injury or infection (see Figures 3 and 5 in [11]). While the kinetic trajectories of the model’s output variables in our simulations had qualitatively similar shapes during both acute and chronic inflammation, the chronic inflammatory scenarios were characterized by higher concentrations (Ψmax) and delayed resolution timing (Ri) for neutrophils, macrophages, and pro-inflammatory mediators (e.g., TNF-α, IL-1β, and IL-6) in comparison with acute inflammatory scenarios, which is consistent with experimental reports of chronic inflammation [17, 34, 36, 45]. The model is a coupled system of 15 ordinary differential equations and one delay differential equation (DDE). The DDE in the model is used to describe the chemotaxis of pro-inflammatory macrophages. Indeed, there exists a ~12 h delay between the arrival of the macrophage precursors (i.e., monocytes) at the wound site and their differentiation into pro-inflammatory macrophages [12], which is accounted for in the model by using the DDE. Each of our simulations reflected a 20-day period after inflammation initiation. We performed all computations in the software suite MATLAB R2012a (MathWorks, Natick, MA) and solved the model equations using the MATLAB solver DDE23 with default tolerance levels. The MATLAB files used to simulate the results reported in this article and a document providing details on how run the code are provided as S1 Code and S1 Text, respectively.
For each model output variable representing a molecular species or cell type (with two exceptions), we calculated the four inflammation indices defined in the Introduction (Fig 1). We did not calculate the indices for the model variables representing the platelet and TGF-β concentrations, because their kinetic trajectories do not have the same characteristic single-peak shape as the trajectories for other model variables. We additionally considered, and computed the inflammation indices for, two variables representing the total concentrations of neutrophils (Ntot) and macrophages (Mtot). We computed these variables directly from the model output variables representing two distinct neutrophil and two distinct macrophage phenotypes. We thus analyzed 16 model variables in total. In the article text, we sometimes refer to the total neutrophil and total macrophage model variables simply as neutrophils and macrophages, respectively. First, we calculated temporal trajectories and the inflammation indices for the model’s default parameter set, which represents an acute inflammatory scenario [11]. Then, applying a previously described Latin hypercube sampling approach [71], we generated 10,000 random 69-parameter sets, in which each individual parameter was sampled independently from an interval permitting up to twofold deviations (up or down) from the parameter’s default value. We used the MATLAB function LHSDESIGN to perform this sampling. The parameter sets were intended to represent the natural variability in the inflammatory scenarios occurring under different circumstances or in different individuals. In the existing literature, there is no consensus regarding the criteria for selecting the parameter randomization sample size to ensure sufficient coverage of the possible kinetic scenarios. It is recommended that this sample size be increased until no significant further changes in the main analysis results are detected [76]. Following this strategy, we arrived at the parameter randomization sample size equal to 10,000 parameter sets. For the considered model variables, we calculated the inflammation indices from the temporal trajectories generated for each of these 10,000 parameter sets. We refer to these trajectories as 10,000 simulations.
We calculated logarithmic local sensitivities, sij, for each inflammation index of every model output variable with respect to every model parameter according to the standard definition (see, e.g., [11, 73]):
sij=∂logXi/∂logpj=(dXi/Xi)/(dpj/pj),
(1)
where Xi is a given inflammation index (Tact, Ψmax, Ri, or Rp) for the model’s ith variable (of the 16 output variables) and pj is the model’s jth parameter (of the model’s 69 parameters). To obtain numeric approximations of the derivatives in Eq 1, each parameter was individually perturbed by ±1% of its value, and the derivative was approximated using the second-order central finite difference formula. We restricted our attention to local sensitivities because local sensitivity analysis is a powerful tool that has been successfully used to identify critical mechanisms and intervention points in a variety of biological systems [11, 19, 72–74]. We performed a local sensitivity analysis for the default parameter set, as well as for each of the 10,000 simulations with random parameter sets. For each considered parameter set, each model variable, and each inflammation index, we sorted the absolute sensitivity values in descending order to determine the top three most influential parameters for that variable’s index.
Using the MATLAB function CORR, we calculated Spearman’s rank correlation coefficients (CCs) between each of the 69 model parameters and each of the four inflammation indices, for each of the 16 model variables. For these calculations, we used the model parameter values and the inflammation index values from the 10,000 simulations with randomized parameters described above. Moreover, based on the calculated Ri values for the neutrophil and macrophage variables, we introduced a criterion for dividing the 10,000 simulations into two subsets. One of the subsets contained simulations representing acute inflammation, and the other one contained simulations representing chronic inflammation. To create these subsets, we first calculated 20,000 ratios by dividing the Ri values for neutrophils and macrophages in each of the 10,000 simulations by the respective Ri values calculated using the default parameter set. Second, we used a cutoff value (equal to 2) to identify the simulations (out of 10,000) for which both the neutrophil and macrophage ratios were above the cutoff. The criterion for choosing the cutoff value was based on experimental studies, in which the Ri was increased by ~2-fold in abnormal inflammatory scenarios [12, 17]. The simulations separated based on this criterion were regarded as reflecting chronic inflammation. We then re-calculated the CCs between the model parameters and model output timing indices (i.e., Tact and Ri) using only the chronic inflammation simulations. This was done to identify any significant correlations that may exist for a specific chronic inflammation condition and were not detected when all 10,000 inflammation scenarios (simulations) were considered. A description of the algorithm for the separation of the simulation subsets is provided in S1 Text.
We modeled inflammatory mediator inhibition by adding two new differential equations to the original model for each modeled inhibitor. The equations represented the volumetric concentration of the inhibited mediator and its respective mediator-inhibitor complexes. We used mass action kinetics to model the individual reactions between the inhibitors and their target mediators according to the following reaction scheme (which reflects inflammatory mediator sequestration that prevents the mediator’s participation in normal signaling):
I + C ←koff→ kon IC
(2)
where C denotes a mediator, I represents an inhibitor, and IC denotes the mediator–inhibitor complex. For any given inhibitor, kon and koff denote the values for the association and dissociation rate constants, respectively. Using our model extended in this way, we performed simulations to predict the effects of mediator inhibition at different inhibition concentrations and the addition of inhibitors at different time points after inflammation initiation. In these analyses, the inhibitors were added only in chronic inflammation simulations, and only one mediator was inhibited in each simulation, unless stated otherwise. A detailed description of the computational implementation of inflammatory mediator inhibition is provided in the Supplemental Material (S1 Text).
|
10.1371/journal.pcbi.1003478 | Computational Design of the Affinity and Specificity of a Therapeutic T Cell Receptor | T cell receptors (TCRs) are key to antigen-specific immunity and are increasingly being explored as therapeutics, most visibly in cancer immunotherapy. As TCRs typically possess only low-to-moderate affinity for their peptide/MHC (pMHC) ligands, there is a recognized need to develop affinity-enhanced TCR variants. Previous in vitro engineering efforts have yielded remarkable improvements in TCR affinity, yet concerns exist about the maintenance of peptide specificity and the biological impacts of ultra-high affinity. As opposed to in vitro engineering, computational design can directly address these issues, in theory permitting the rational control of peptide specificity together with relatively controlled increments in affinity. Here we explored the efficacy of computational design with the clinically relevant TCR DMF5, which recognizes nonameric and decameric epitopes from the melanoma-associated Melan-A/MART-1 protein presented by the class I MHC HLA-A2. We tested multiple mutations selected by flexible and rigid modeling protocols, assessed impacts on affinity and specificity, and utilized the data to examine and improve algorithmic performance. We identified multiple mutations that improved binding affinity, and characterized the structure, affinity, and binding kinetics of a previously reported double mutant that exhibits an impressive 400-fold affinity improvement for the decameric pMHC ligand without detectable binding to non-cognate ligands. The structure of this high affinity mutant indicated very little conformational consequences and emphasized the high fidelity of our modeling procedure. Overall, our work showcases the capability of computational design to generate TCRs with improved pMHC affinities while explicitly accounting for peptide specificity, as well as its potential for generating TCRs with customized antigen targeting capabilities.
| T cell receptors (TCRs) play a major role in immunity, recognizing peptide antigens presented by major histocompatibility complex proteins. Due to their capacity to target intracellularly produced proteins and initiate cell killing, there is significant interest developing TCR-based therapeutic strategies, particularly towards cancer. A concern with TCRs is their weak-to-moderate affinities, which limits therapeutic potential. While in vitro evolution has been used to enhance TCR affinity, with sometimes spectacular results, these techniques can reduce peptide specificity and offer little control over affinity enhancements. Here we explored the use of structure-based computational design to enhance TCR affinity, which in principle can permit control over both specificity and affinity gains. We examined a clinically relevant TCR recently used in melanoma immunotherapy, identifying and characterizing mutations which enhanced affinity with no detectable impacts on binding specificity. We solved a crystal structure of our highest affinity designed TCR in complex with antigen, which indicated high accuracy of the structural modeling during the design process, and we critically evaluated several design protocols and functions to further improve design success. These results provide valuable insights into the use of computational design for TCRs. Lastly, the enhanced affinity variants identified may be of potential clinical benefit.
| T cell receptors (TCRs) are key elements of adaptive immunity, as they specifically recognize antigenic peptides bound to MHC proteins (pMHCs) on cell surfaces and are responsible for initiating immune responses against targeted cells. The TCR-pMHC interaction is of considerable importance in health and disease, notably in transplantation, autoimmunity, and is a target for development of vaccines and therapeutics for infectious disease and cancer [1]–[3]. For example, the adoptive transfer of genetically engineered T cells, whereby tumor-specific TCRs are transduced into T cells and then infused into the patient, is being explored as a means for cancer immunotherapy. Clinical trials of such genetically engineered T cells have shown promise in the treatment metastatic melanoma [4]–[6] and synovial cell carcinoma [7], leading to durable tumor regression and long-term survival in patients.
The observations that TCRs have relatively weak affinities towards pMHC (typically 1–300 µM; ∼1000-fold lower than mature antibody/antigen interactions) and that pMHC affinities are correlated to some extent with in vivo potency [8] have led to a number of efforts to engineer TCRs with enhanced binding affinity. These efforts include in vitro selection [9]–[13] as well as computational structure-based design [14]–[16], resulting in up to 1,000,000-fold improvements in affinity. However, a major concern in enhancing TCR affinity is maintenance of peptide specificity. As TCRs recognize peptides presented by MHC proteins, yet invariably form contacts to both peptide and MHC [17], enhancements to TCR affinity risk dangerous cross-reactivity if affinity-enhancing substitutions preferentially target the MHC protein. Such “off-target” interactions can be challenging to predict from peptide sequence and are a major concern for high affinity TCRs [18]. Indeed, the unanticipated cross-reactivity of a high affinity TCR resulted in serious consequences and deaths in a recent clinical trial [19]. Additionally, significant enhancements in antigen-specific affinity may be detrimental for T cell activity, as there is evidence of a TCR “threshold affinity” above which T cell responsiveness is attenuated [20], [21]. Thus, careful control of affinity and specificity is crucial in the development of enhanced TCRs for therapeutic purposes.
The αβ TCR DMF5 was originally isolated from tumor infiltrating lymphocytes present in a patient with metastatic melanoma [22]. DMF5 recognizes the 27–35 nonameric and 26–35 decameric peptide epitopes from the MART-1 melanoma antigen presented by the class I MHC protein HLA-A*0201 (HLA-A2), and was the second TCR to be used in clinical trials of genetically engineered T cells [5]. Without knowledge of structure or affinity, Robbins and colleagues previously examined a series of point mutations in DMF5, generating variants that resulted in improved antigen-specific responses yet also showed evidence of reduced specificity, underscoring the need for incorporating structural information in the design process [23]. More recently, the DMF5 TCR has been crystallized by our laboratory in complex with both the MART-1 nonameric epitope (AAGIGILTV; referred to as AAG) as well as the anchor-modified decameric epitope (ELAGIGILTV; referred to as ELA), both bound to HLA-A2 [24]. The structures show that despite the significant difference in peptide conformation between the ELA/HLA-A2 and AAG/HLA-A2 ligands [24], DMF5 engages them with an identical binding mode. These structures along with associated affinity measurements provide an ideal opportunity to explore the applicability of computational structure-based design for rationally enhancing a clinically relevant TCR, while simultaneously exploring the impact on peptide specificity.
Utilizing a refined algorithm initially developed for our redesign of the A6 TCR [14], we applied structure-based design to the DMF5 TCR, generating variants and characterizing mutants with affinity enhancements of up to 400-fold toward ELA/HLA-A2. Highlighting the ability of structure-based design to directly target regions of interest within protein interfaces, and in contrast with results seen with in vitro selection, the strongest affinity enhancement was achieved with only two previously identified amino acid substitutions [25] that directly interact with the peptide. Importantly, the highest affinity variant showed no detectable recognition of unrelated peptides presented by HLA-A2. We determined the crystallographic structure of this variant bound to ELA/HLA-A2, permitting a detailed analysis of the accuracy of the various structural modeling protocols, and together with the affinity measurements, a quantitative assessment of scoring functions and terms. Further, by purposely disrupting interactions with the ELA peptide, we were able to shift TCR specificity away from the ELA peptide toward the AAG peptide, albeit with more modest efficacy. Altogether, these results highlight the promise of structure-based design for TCR engineering, and provide a rich dataset for further improvements in design strategies, including the broadening of efforts to other TCR-pMHC systems. Lastly, given the ongoing use of the DMF5 TCR in efforts to develop immunological therapies for melanoma (e.g., [26]), the high affinity DMF5 variants identified here may have future clinical applicability.
We used the ZAFFI and Rosetta software tools [14], [27] to predict the affinity changes of DMF5 mutants for ELA/HLA-A2 or AAG/HLA-A2, simulating all point mutations for each DMF5 residue within 5.5 Å of the pMHC ligand in the tertiary structures. In total, we examined 589 substitutions of 31 DMF5 residues within each complex, which were then ranked based on predicted TCR-pMHC affinity. Twelve computationally designed mutations were chosen for experimental testing. To help maintain peptide specificity, with the exception of two αR27 mutants, we only chose mutants that were predicted to contact the peptides. The αR27 mutants were selected to compare with our previously designed substitutions at the corresponding position in the A6 TCR [14], which shares the germline α chain gene (TRAV 12-2) and some MHC contacts with DMF5. We performed mutagenesis using soluble DMF5 gene constructs, expressed and purified the mutant proteins, and measured their binding affinities toward ELA/HLA-A2 and AAG/HLA-A2 via surface plasmon resonance (Figure 1). The mutations and their measured affinities for ELA/HLA-A2 and AAG/HLA-A2 are given in Table 1, organized by the method through which they were selected: Affinity, Specificity, or Proline, as discussed in detail below. In addition, Table 1 includes four mutations, listed under “Test”, that we selected for measurement based on manual inspection of the TCR-pMHC structures.
Mutations in the Affinity category were chosen on the basis of predicted enhancement in affinity towards both ELA/HLA-A2 and AAG/HLA-A2. Three of the six mutations in this category had significantly improved affinities: αD26W, αD26Y, and βL98W. The two αD26 mutants had the highest measured binding affinities among all tested mutants (up to 40-fold improvement for αD26W towards AAG/HLA-A2), while the βL98W mutant had a 3-fold affinity improvement for both ELA/HLA-A2 and AAG/HLA-A2.
Mutations in the Specificity category were chosen on the basis of predicted differential affinity towards ELA/HLA-A2 and AAG/HLA-A2. These were predicted to contact a portion of the interface that varies between the two peptides, where the alanine at the N-terminus of the nonamer is replaced by a larger glutamate residue in the decamer (Figure S1). Several mutations at TCR position αG28 were chosen that would potentially destabilize the interaction with ELA/HLA-A2 via steric hindrance while favoring AAG/HLA-A2. Of the specificity-altering substitutions, all shifted specificity toward the AAG nonamer as predicted, albeit the shifts were relatively modest (up to a 5-fold shift; Table 1 and Figure 2).
Based on work with the A6 TCR [28], as well as the observation of proline CDR mutants in high affinity TCR selection experiments [9], [11], we tested three proline mutations that were predicted to stabilize CDR loops in the bound conformation while not negatively impacting contacts with the pMHC (Proline category in Table 1). None of these proline substitutions showed a significant improvement in affinity, indicating that while potentially reducing the entropic cost for binding, the magnitudes of any such improvements were not substantial enough to yield a net increase in binding free energy, possibly because these loops appear relatively rigid in the unbound DMF5 TCR [29]. Moreover, given the >1 kcal/mol loss in binding free energy with both ELA/HLA-A2 and AAG/HLA-A2, the αG28P substitution may have directly or indirectly impacted pMHC contacts, consistent with its relatively buried position in the pMHC interface.
Combining the affinity-enhancing αD26Y and βL98W mutations (this double mutant is referred to as YW) yielded a substantial improvement towards ELA/HLA-A2. This high affinity double mutant was previously described in a brief report, with a preliminary affinity measurement yielding an approximate 200-fold enhancement [25]. Here, however, we measured a 400-fold improvement (from 9.5 µM to 24 nM). The difference is attributable to our use of a kinetic titration binding assay in this case (Figure 1b), which is more accurate at quantifying binding in the nanomolar range or higher, as it permits analyses of high affinity binders without requiring surface regeneration [30]. The on and off rates of the YW mutant towards ELA/HLA-A2 determined from the kinetic titration were 1.7×106 M−1 s−1 and 0.05 s−1, respectively. The dissociation rate of wild type DMF5 from ELA/HLA-A2 was too fast to accurately measure [29], indicating that the combined mutations result in a slower TCR off rate, as seen with the majority of affinity-enhanced TCRs [31].
The combined YW mutations were somewhat nonadditive (−3.5 kcal/mol enhancement versus −2.5 kcal/mol assuming additivity), suggesting a modest degree of communication between the CDR1α and CDR3β loops; the same degree of cooperativity was also observed for the αD26W/βL98W (WW) mutant binding ELA/HLA-A2 (Table 1). Nonadditivity within TCR binding interfaces has been observed previously [32], [33], and could be attributable to structural or dynamic effects of mutations on neighboring loops. The YW variant also showed a smaller but still considerable 30-fold enhancement towards AAG/HLA-A2. The reduced affinity enhancement is likely attributable to the lack of the N-terminal glutamate in the AAG peptide as discussed below.
Given its dramatic affinity improvement toward both ELA/HLA-A2 and AAG/HLA-A2, we next asked whether the high affinity YW variant could recognize targets other than the MART-1 nonamer and decamer. No binding was detectable towards HLA-A2 presenting the Tax or gp100 peptides, even at concentrations more than 25-fold higher than those used to characterize binding to wild type DMF5 (Figure S2). The Tax and gp100 peptides have markedly different sequences from ELA or AAG (Tax: LLFGYPVYV; gp100: IMDQVPFSV), yet the conformations of HLA-A2 are identical in the four peptide/HLA-A2 crystal structures [24], [34], [35]. The lack of detectable binding of the high affinity DMF5 YW variant towards the other peptides thus suggests that we may have improved its specificity towards the MART-1 peptides, and at the minimum demonstrates that our design has avoided peptide-independent targeting of HLA-A2.
To quantify the performance of the design methods that we used to generate candidate mutations, ZAFFI and Rosetta, we compared predicted versus measured affinities towards ELA/HLA-A2 and AAG/HLA-A2 for each of the point mutations that were experimentally characterized (excluding the αY50A and αG94T mutants, for which binding was too weak to measure). Mutants were scored with or without structural minimization (referred to as Min and NoMin respectively), as shown in Figure 3 (with scores in Table 2). For both the Rosetta and ZAFFI scoring functions, the NoMin simulations yielded higher agreement with experimental data (Figure 3a–b), with the Rosetta scoring function achieving an impressive 0.72 correlation with measured ΔΔGs (excluding four outlier points correctly predicted to have poor affinities). Except for the proline mutant αG28P, the Rosetta NoMin protocol made no other false positive predictions, and its top four predictions (αD26Y and αD26W for the two pMHCs) had the highest measured affinities among all predicted point mutations (βL98W was also correctly ranked highly, particularly for AAG). This predictive success is notable as the majority (8 out of 14) of these mutants involved glycine and proline, which are often overlooked during in silico studies due to difficulties predicting backbone-related effects [27].
The ZAFFI NoMin protocol gave a correlation of 0.59 with measured data (again excluding several true negative outlier points due to predicted steric hindrance). Though it previously outperformed Rosetta in scoring A6 TCR mutants [14], and correctly gave favorable scores for the DMF5 αD26 mutants, ZAFFI made several false positive DMF5 predictions for both AAG/HLA-A2 and ELA/HLA-A2, possibly due to its parameterization on a more limited dataset than Rosetta and the distinct biophysical properties of the A6 and DMF5 interfaces. This led us to evaluate and reparameterize the terms in the ZAFFI function using a larger set of energy terms and mutants, as described further below.
Both minimization-based protocols (ZAFFI Min and Rosetta Min; Figure 3c–d), while displaying positive correlations with the experimental results, were lower in their predictive success than the NoMin protocols. However, ZAFFI Min scored the αD26 mutants favorably, and correctly identified βL98W (for AAG) as within the score cutoff for predicted binding improvement (≤−0.6; for ELA, βL98W was near this cutoff). Overall though, false positive predictions for ZAFFI and Rosetta led to relatively weak correlations, suggesting that minimization may have led to incorrect structures in some cases. We additionally tested other minimization protocols as well as more extensive side chain packing (Table S1), each of which gave lower correlations with measured energies than the relatively restrictive NoMin protocol.
To examine the structural basis of the 400-fold binding affinity improvement and compare with the models generated during the design process, we crystallized and determined the structure of the DMF5 YW mutant bound to ELA/HLA-A2 at 2.56 Å resolution (Figure 4, with crystallographic data in Table S2). Clear electron density was observed for the TCR-pMHC interface, and the positions of the mutated amino acids were unambiguous as indicated by an unbiased, iterative-build OMIT map [36] (Figure S3). As with other structurally characterized TCRs engineered for high pMHC affinity [12], [13], [37]–[39], the docking orientation was conserved when compared to the wild-type complex, with a TCR-pMHC crossing angle of 32°, versus 33° for the wild-type. Essentially no perturbations of the interface CDR loops or peptide were observed (0.34 Å backbone atom RMSD for TCR and pMHC residues within 10 Å of the binding interface), indicating that our relatively conservative design strategy of selecting point substitutions against a fixed pMHC structure did not substantially alter the interface or proximal side chains (Figure 4b–e). This tight structural conservation of the binding loops and target pMHC residues is in contrast to some high affinity TCRs generated by in vitro selection where moderate (1G4 designs c5c1, c48c50, c58c61, c58c62) [12],[37] or pronounced (2C designs m6, m13, m67, and Mel5 design α24β17) [13], [38], [39] perturbations of CDR loops were exhibited, along with adjacent CDR loop remodeling [39] and addition of a synergistic ion adduct in the interface [12]. In the recently described structure of the c134 TCR [39] which is an in vitro selected variant of the A6 TCR with nearly 1000-fold improved affinity for Tax/HLA-A2, the mutant CDR3β loop retained largely the same backbone structure as the wild-type loop, yet it led to a shifted footprint of the α chain over the pMHC.
As anticipated from our modeling, both the tyrosine and tryptophan mutant side chains directly contact the MART-1 peptide in the αD26Y/βL98W structure, and make more extensive peptide contacts than their wild-type counterparts (Table S3). These mutations led to a 5% increase in buried solvent accessible surface area for the pMHC, from 1059 Å2 to 1113 Å2. Unexpectedly, as explicit water molecules were not used in our structural modeling or scoring, a water-mediated hydrogen bond to the peptide was introduced between the mutant residue αY26 and the side chain of the N-terminal glutamate of the peptide, in addition to a direct hydrogen bond between side chains (Figure 4c). This polar network may explain the superior affinity of αY26 versus αW26 for ELA/HLA-A2, despite the fact that they were predicted to be similar (ZAFFI) or αW26 was preferred (Rosetta; Table 2). In contrast, αD26W binds more strongly than αD26Y to AAG/HLA-A2, which lacks the N-terminal peptide glutamate and hydrogen bonding capability at that side chain. In light of the water-mediated contacts observed in the mutant crystal structure, we re-ran simulations using explicit water molecules from the wild-type and mutant structures, but no improvement in correlation was observed (Table S1).
The crystal structure of the YW variant bound to ELA/HLA-A2 allowed us to evaluate the performance of several structural modeling protocols. After least squares fitting of the backbone of the TCR and pMHC interface residues to the crystal structure, we compared positions of the modeled side chains to those in the crystal structure (Figure 5, with RMSDs in Table 3). In addition to the NoMin and Min methods, we evaluated models generated using two intermediate minimization methods: MinSC (minimizing interface side chains only) and MinBB (minimizing interface backbone atoms). Finally, we re-modeled the engineered side chains in the context of the mutant crystal structure (NoMinMut) to determine whether accurately positioned backbone and neighboring side chain atoms could improve modeling results.
For modeling the side chain of αY26, all protocols performed well in predicting the general orientation of the Tyr side chain, with NoMin outperforming the other protocols (RMSD = 1.06 Å). Though generally accurate, all models exhibited a rotation in the aromatic ring and a slight shift in the OH group with respect to the crystal structure. As these errors were possibly due to the absence of explicit waters in the modeling omitting the water mediated hydrogen bonding observed in the YW/ELA/HLA-A2 crystal structure, we re-ran the NoMinMut simulation with water molecules from that structure, but found little improvement in RMSD (0.98 Å, versus 1.13 Å without water molecules).
The predicted side chain conformations for the mutant βL98W were more variable than αY26, including a flip of the aromatic rings in the Min and MinBB models, leading to relatively high RMSDs (>2 Å) relative to the experimentally determined structure for this residue. The structure modeled without minimization had a sub-optimal positioning of the Trp side chain (tilted away from the pMHC) (Figure 5), which improved substantially (from 1.52 Å to 0.89 Å RMSD) when modeled in the context of the backbone and side chains from the mutant crystal structure. This indicates that Rosetta's packing protocol is sensitive to small structural perturbations and accurate modeling of backbone and neighboring side chains can lead to improved predictions.
In light of the lower accuracy of the ZAFFI scoring function on the measured DMF5 point mutants (Figure 3) than for the A6 TCR, we performed a systematic evaluation of scoring functions to better predict DMF5 affinities while still maintaining accuracy with the set of A6 mutants. We included several statistical potentials in addition to the energetic and knowledge-based terms from the original ZAFFI study [14]. Given that minimization yielded false positive results for both ZAFFI and Rosetta functions (Figure 3) and that unminimized structures more closely matched the YW-ELA/HLA-A2 crystal structure, we used unminimized models for this analysis. In addition to correlation with measured ΔΔGs, we evaluated scoring functions using receiver operating characteristic area under the curve (AUC) in order to judge discrimination of binding improvement without penalizing true negative or true positive outliers.
We identified a scoring function (referred to as ZAFFI 1.1) with a higher correlation (0.74) than ZAFFI (0.59) and Rosetta (0.72) for the set of DMF5 point mutants (excluding the four αG28 outlier mutants), and high AUC values for both DMF5 and A6 mutants (Table 4 and Figure S4). Correlation P-values are included in Table 4 for all functions, highlighting significant predictive performance of ZAFFI 1.1 (p<0.001) for both sets of data. ZAFFI 1.1 includes six terms: van der Waals attractive and repulsive components, desolvation, intra-residue clash, hydrogen bonding and Coulombic electrostatics. While its correlation with A6 TCR data (0.65) was not as high as the original ZAFFI function (0.77), both the correlation and AUC are considerably higher than Rosetta on that set of data. Although a few outlier points persisted, including αG28P in the AAG/HLA-A2 interface, the overall success of this function demonstrates that a relatively simple scoring function and packing scheme can be used to model a large proportion of energetic changes in three designed TCR-pMHC interfaces. To examine the performance of this function in the context of other protein-protein interactions, we applied it to two large sets of interface point mutants (285 mutants each) of two proteins designed de novo to target influenza hemagglutinin (Table S4), recently used in a collaborative effort to evaluate protein design algorithms as part of the protein docking experiment CAPRI [40]. We found that ZAFFI 1.1 (with NoMin packing) performed similarly to the other tested functions for scoring the HB36 mutants (r = 0.36; p = 2.1×10−10), while for HB80 mutants it outperformed all other functions (r = 0.5; p<2.2×10−16), with a Kendall tau rank correlation (0.38) higher than we achieved in the CAPRI experiment using a ZAFFI-related function (0.31), where our Kendall correlation surpassed all other groups [40].
Structure-based design of TCRs provides a means to improve upon low wild-type affinities for pMHC while maintaining, improving, or altering specificities for desired targeting capabilities. While some studies have determined the fine specificities of designed TCRs using biophysical [39], [41] and cell-based [23] methods, here we demonstrated that point substitutions selected using structure-based methods can be used to efficiently engineer pMHC specificity and affinity. We then utilized structural modeling and x-ray crystallography to gain atomic-level insights into these substitutions. We achieved higher affinity improvements than previously reported in structure-based TCR design, with just two point substitutions resulting in an approximately 400-fold affinity improvement, versus 150-fold for four combined point mutants of the BC1 TCR selected using molecular mechanics [16], and 100-fold for four combined point mutants of the A6 TCR selected using ZAFFI [14].
Despite the structural plasticity commonly observed in TCR-pMHC interfaces [42]–[45], our computational modeling and crystal structure indicate that carefully selected point substitutions can improve pMHC affinity and modulate peptide specificity without grossly perturbing the interface structure. We note though that a broad extension this approach to other TCRs of interest will likely entail further refinement of the energy function based on measured data, in addition to improvements in high-resolution modeling of TCR-pMHC complexes [46]. Large-scale datasets of mutant binding affinities, including the CAPRI data we utilized to assess our design functions [40], can provide possible training sets for re-weighting terms and derivation of energy-based statistical potentials that would add discriminating power and predictive breadth to the ZAFFI function. Additionally, our analysis of the YW-ELA/HLA-A2 structure indicates that there is room for improving structural modeling of mutant residues, with modeling of fine structural effects and bound water molecules representing two avenues for further development.
The modulation of nonamer versus decamer specificity by many point mutants of the DMF5 TCR highlights the sensitive nature of TCR-antigen recognition, as well as the potential to fine-tune TCR recognition properties via structure-based design. We achieved a shift in specificity toward the nonameric MART-1 peptide via mutation of αG28 residues that were predicted to clash with the decameric E1 residue but would be accommodated in the cleft near the nonameric A1, similar in concept to the “knob-in-to-hole” designs utilized to alter binding specificity in other protein-protein interfaces [47]. The clash with the decamer was overestimated using the NoMin modeling methods (which had the greatest overall predictive success), thus leading to lower than anticipated specificity shifts; better modeling of clashes through judicious use of minimization (avoiding false positive predictions as we observed) could potentially reduce such errors. In contrast, we found an increase in specificity (>4-fold) toward the decameric peptide with the DMF5 double mutants YW and WW, resulting from the cooperativity of these mutants in the presence of the decamer. This peptide-dependent cooperative effect is a previously undescribed mechanism for shifting TCR specificity. As the structure of the YW/ELA/HLA-A2 complex did not suggest any major alterations in the binding interface compared to the wild-type complex, this effect may be dynamic in nature. As recently reported, the Mel5 TCR mutant α24β17, which targets ELA/HLA-A2 with a 30,000-fold affinity improvement over wild-type, was found to retain peptide specificity, albeit towards alanine substituted ELA variants rather than between the ELA and AAG decameric/nonameric peptides [13]. In this case specificity was mediated through subtle solvent interactions. By modeling solvent and dynamic effects, as well as exploring explicit specificity design methods, such as multi-state design [48], greater control of TCR specificity could be achieved via rational engineering.
Three of the α chain mutants we tested were previously examined in the A6 TCR (αD26W, αG28I, and αG28L) [14], whose CDR1α and CDR2α loops are identical to DMF5 due to the common use of the TRAV12-2 gene. αD26W improved pMHC affinity significantly for both TCRs, though to varying extents. On the other hand, the αG28 mutants improved the affinity of A6 modestly (∼2-fold) but resulted in no change or weakened affinity with DMF5. This behavior likely follows from the positions of the mutations, as the αG28 mutants are predicted to make extensive contacts with the varying N-terminus of the peptide, while αD26W would primarily target the same HLA-A2 site to improve affinities for all three pMHCs. However, both αD26 mutants of DMF5 still exhibited a measurable peptide dependence with ΔΔG, compared with, for instance, βL98W which had identical effects in the context of both MART-1 peptides. Data from more mutants and positions, as well as other TCR-pMHC systems, such as the Mel5 TCR which shares the TRAV12-2 gene with DMF5 and A6 and also targets ELA/HLA-A2 with a similar docking mode [49], would help to further delineate the extent of any conserved effects of affinity-enhancing or destabilizing mutants. Indeed, the structure of the high affinity α24β17 Mel5 TCR mutant in complex with ELA/HLA-A2 [13] features a large hydrophobic substitution at position αD26 (Phe), which closely matches the αD26Y conformation and the pMHC binding site in the YW/ELA/HLA-A2 structure (Figure S5), although as Mel5 α24β17 contained 18 additional substitutions, the energetic effect of αD26F alone is unclear. A more detailed study of the impact of affinity-enhancing mutations in germline CDRs would help to further probe TCR germline binding permissiveness suggested by a recent double mutant cycle deconstruction of the interface with the A6 TCR [50].
In conclusion, we have shown that rational, computational-based design offers the potential to simultaneously alter the efficacy and antigen targeting of a therapeutic TCR, potentially enabling the development of improved TCRs for adoptive cell therapy [51] or biotherapeutics [52] customized to bind antigens presented by tumors or virally infected cells from individual patients. Given the ongoing use of the DMF5 TCR in clinical trials for cancer immunotherapy, the higher-affinity YW variant of DMF5 generated here may also be of potential clinical benefit.
As with our previous study designing the A6 TCR [14], we used the “interface” mode of Rosetta 2.0.2 [27] to model point mutations of the DMF5 TCR. Command line options were specified to include extra chi1, chi2, and chi3 rotamers (“-extrachi_cutoff 1 -ex1 -ex2 -ex3”). Only the mutant side chain was repacked (the default behavior of this mode) while the protein backbone from the wild-type structure was retained. Rosetta predicted mutant structures as well as ΔΔGs, and the structures were then re-scored by our energetic scoring function ZAFFI to generate its own set of predicted ΔΔG scores. The ZAFFI filter, parameterized using the A6 TCR data and designed to remove false positive predictions that destabilized native electrostatic contacts, was not used in this study, given that our focus was evaluation and development of binding energy prediction functions, and the new system and protocols being explored would require tuning of the parameters of this filter. However, the filter function was used to corroborate avoidance of mutations in some cases (such as hydrophobic mutants of αQ30) where key hydrogen bonds would likely be disrupted.
To generate predictions of point mutants using side chain and/or backbone minimization we used Rosetta 2.3, a more recent version of this program that includes minimization functionality in its interface mutagenesis mode. Minimization was specified using the command line flags (“-min_interface -int_bb -int_chi”) to perform minimization of interface backbone and side chain atoms in the wild type and mutant structures (“Min” protocol), while just “-int_chi” or “-int_bb” was used to perform only side chain or backbone minimization (“MinChi”, “MinBB”). Point mutant simulations with explicit water molecules taken from the input structure were also performed using Rosetta 2.3, using the command line flag: “-read_hetero_h2o”.
We analyzed residue backbone conformations in the bound and unbound DMF5 TCR structures using a Ramachandran plot analysis server [53] (http://zlab.bu.edu/rama/). DMF5 CDR positions with favorable backbone conformations for proline (as well as favorable pre-proline conformations for the preceding residue), in addition to either improved or maintained pMHC affinity predicted for the proline mutant by at least one prediction method, were selected for experimental mutation to proline.
Expression and refolding of soluble constructs of DMF5 TCRs and HLA-A2 were performed as previously described [29], [54]. In brief, the TCR α- and β-chains, the HLA-A2 heavy chain, and β2-microglobulin (β2m) were generated in Escherichia coli as inclusion bodies, which were isolated and denatured in 8 M urea. TCR α- and β-chains were diluted in TCR refolding buffer (50 mM Tris (pH 8), 2 mM EDTA, 2.5 M urea, 9.6 mM cysteamine, 5.5 mM cystamine, 0.2 mM PMSF) at a 1∶1 ratio. HLA-A2 and β2m were diluted in MHC refolding buffer (100 mM Tris (pH 8), 2 mM EDTA, 400 mM L-arginine, 6.3 mM cysteamine, 3.7 mM cystamine, 0.2 mM PMSF) at a 1∶1 ratio in the presence of excess peptide. TCR and pMHC complexes were incubated for 24 h at 4°C. Afterward, complexes were desalted by dialysis at 4°C and room temperature respectively, then purified by anion exchange followed by size-exclusion chromatography. Refolded protein absorptions at 280 nm were measured spectroscopically and concentrations determined with appropriate extinction coefficients. Mutations in the DMF5 α- and β-chains were generated by PCR mutagenesis and confirmed by sequencing. Peptides and plasmids were commercially synthesized and purified (Genscript).
Surface plasmon resonance experiments were performed with a Biacore 3000 instrument using CM5 sensor chips. In all experiments, TCR was immobilized to the sensor chip via standard amine coupling and pMHC complex was injected as analyte. All samples were thoroughly dialyzed in HBS-EP buffer (20 mM HEPES (pH 7.4), 150 mM NaCl, 0.005% Nonidet P-20), then degassed for at least 15 minutes prior to use. Steady-state experiments were performed with TCRs coupled onto the sensor chip at 1000–1500 response units. Injections of pMHC spanned a concentration range of 0.5–150 µM at flow rates of 5 µl/min at 25°C. Multiple data sets were globally fit using a 1∶1 Langmuir binding model utilizing BIAevaluation 4.1. Kinetic titration experiments were performed with TCRs coupled at approximately 500 response units. A series of five ELA titrations, spanning 10–160 nM and 20–320 nM at 2-fold increase per titration, were flowed over YW and WW respectively. Flow rates of 30 µl/min were used at 25°C. Data were fit with a 1∶1 association model with drift using BIAevaluation [30].
Crystals of the DMF5 YW-ELA/HLA-A2 complexes were grown from 12% PEG 3350, 0.25 M MgCl2 buffered with 0.1 M HEPES (pH 8.0) at 25°C. Crystallization was performed using sitting drop/vapor diffusion. For cryoprotection, crystals were transferred into 20% glycerol/80% mother liquor for 30 s and immediately frozen in liquid nitrogen. Diffraction data were collected at the 22ID (SER-CAT) beamlines at the Advanced Photon Source, Argonne National Laboratories. Data reduction was performed with HKL2000. The ternary complexes were solved by molecular replacement using PHENIX and Protein Data Bank (PDB) entry 3QDG as the reference model [29]. Rigid body refinement, followed by translation/libration/screw (TLS) refinement and multiple steps of restrained refinement were performed. TLS groups were automatically chosen by phenix.refine. Once defined, TLS parameters were included in all subsequent steps of the refinement. Anisotropic and bulk solvent corrections were taken into account throughout refinement. After TLS refinement, it was possible to unambiguously trace the position of peptides and TCR CDR loops in all structures against σA-weighted 2Fo-Fc maps. Evaluation of models and fitting to maps were performed using COOT [55]. The template structure check in WHATIF [56] and MolProbity [57] was used to evaluate the structures during and after refinement. Atomic positioning was verified with an iterative-build OMIT map calculated in PHENIX [36]. Structures were visualized using PyMOL [58]. Analysis of hydrogen bonds was performed with HBPlus [59], using hydrogen-acceptor maximum distance of 2.7 Å and a donor-acceptor maximum distance of 3.6 Å. Solvent accessible surface areas were measured in Discovery Studio (Accelrys Inc.) using a probe radius of 1.4 Å. The structure has been deposited with the Protein Data Bank (PDB ID 4L3E).
ROC AUC analysis was performed using the CROC package [60]. Multi-linear regression to determine weighting of terms was performed as described previously, using 760 measured point mutants from four enzyme-inhibitor complexes [14]. However, we used van der Waals attractive and repulsive terms from Rosetta [27] rather than the corresponding terms from ZRANK [61], as the former led to some improvement in performance across the tested systems. As with the original ZAFFI training, we removed mutants with high clash during training (van der Waals repulsive score >580, corresponding to 48 mutants removed out of 760). We included a number of statistical potential terms for evaluation that were recently tested for binding affinity prediction [62], though none led to substantial improvements in predictive performance in this context. The terms and weights for the retrained energy function (ZAFFI 1.1) are:
van der Waals attractive: 0.57
van der Waals repulsive: 0.0045
solvation: 0.58
hydrogen bonding: 1.2
intra-residue repulsion: 0.026
electrostatics: 0.03
Solvation, hydrogen bonding, and intra-residue repulsion terms were obtained from Rosetta (along with the van der Waals terms as noted above), while the electrostatics term is the long-range Coulombic electrostatics energy from ZRANK [61].
All correlations (with the exception of the Kendall tau rank correlations reported in Table S4) are Pearson correlations. P-values for correlations were calculated using the program R (www.r-project.org).
|
10.1371/journal.pgen.1003148 | Pre-Disposition and Epigenetics Govern Variation in Bacterial Survival upon Stress | Bacteria suffer various stresses in their unpredictable environment. In response, clonal populations may exhibit cell-to-cell variation, hypothetically to maximize their survival. The origins, propagation, and consequences of this variability remain poorly understood. Variability persists through cell division events, yet detailed lineage information for individual stress-response phenotypes is scarce. This work combines time-lapse microscopy and microfluidics to uniformly manipulate the environmental changes experienced by clonal bacteria. We quantify the growth rates and RpoH-driven heat-shock responses of individual Escherichia coli within their lineage context, stressed by low streptomycin concentrations. We observe an increased variation in phenotypes, as different as survival from death, that can be traced to asymmetric division events occurring prior to stress induction. Epigenetic inheritance contributes to the propagation of the observed phenotypic variation, resulting in three-fold increase of the RpoH-driven expression autocorrelation time following stress induction. We propose that the increased permeability of streptomycin-stressed cells serves as a positive feedback loop underlying this epigenetic effect. Our results suggest that stochasticity, pre-disposition, and epigenetic effects are at the source of stress-induced variability. Unlike in a bet-hedging strategy, we observe that cells with a higher investment in maintenance, measured as the basal RpoH transcriptional activity prior to antibiotic treatment, are more likely to give rise to stressed, frail progeny.
| Individual organisms of identical genetic background, living in a homogeneous constant environment, may nonetheless exhibit observable differences dubbed phenotypic plasticity or variability. When such a population is challenged with an unforeseen stress, the disparity among individuals may increase, yielding different strategies in response. This work addresses the occurrence and propagation of phenotypic variation as it affects bacterial survival in response to mild antibiotic treatments. We recorded images of single bacterial cells as they divide prior to and during exposure to a sub-lethal level of streptomycin, a ribosome-targeted antibiotic. We found that individual differences increase upon stress to the extent that cells may either die or survive the treatment. Differentiation events were traced back prior to exposure. We suggest that a positive feedback loop, governed by increased membrane permeability, underlies the transient cell memory observed. Cells with relatively high basal stress-response levels prior to stress are not primed for better survival, but are rather more likely to succumb to antibiotic treatment. As pathogens commonly encounter sub-lethal doses of antibiotics, their survival may be better understood in light of this study.
| Microbial phenotypic heterogeneity, defined as variability of a given trait in a genetically identical population in a homogeneous environment, has been repeatedly observed [1], [2]. It is manifest, for example, in the broad distributions of individual gene expression levels recorded in studies of both prokaryotic and eukaryotic cells [3], [4]. Stress conditions may induce further differentiation of clonal cells, in agreement with the observed higher variability of stress response genes' expression in comparison with other gene families [5]. At the extreme, initial stochastic variability is funneled into bistable states via positive feedback mechanisms that persist through generations [6], [7]. Recent evidence suggests that fate decisions can be partly made even before cells experience an environmental change [8], [9], [10]. A cell's ultimate fate depends on its historical state, indicating that phenotypic variability is shaped by pre-disposition factors [11].
It has been proposed that population heterogeneity increases fitness in unpredictable environments [12], [13]. This may work as a kind of bet-hedging [7], [14], [15], allowing a given genotype to express multiple phenotypes of differing viability. One phenotype may be better adapted to the current environment while others are prepared for future environmental changes under which they may gain higher fitness. On the other hand, heterogeneous populations may simply undergo performance-based selection, in which fitter cells always perform better despite an environmental change.
Stress-responsive genes show greater expression variability than genes from other classes [5], suggesting the hypothesis that variability arises as an anti-stress adaptation evolutionary strategy. Among the stresses that bacteria face, antibiotics are prominent and widespread [16]. Yet the consequences of low-grade antibiotic stress are rather poorly understood. Our interest here is to characterize the dynamic process of stress-induced phenotypic heterogeneity. Specifically, we address the following questions: Will sub-inhibitory antibiotic concentrations further amplify phenotypic variation to the extent of producing persistant and sensitive sub-populations? Are there any predetermining factors that modulate the response? How does this variation propagate through the bacterial lineage?
To this end, we followed the growth of micro-colonies from single Escherichia. coli (E. coli) cells under microfluidic control. We exposed cells to sub-inhibitory concentrations of the aminoglycoside antibiotic streptomycin and tracked their responses at the single-cell level. We find that mild antibiotic treatment results in rapid generation of increased phenotypic variability in terms of stress-induced gene expression, growth rate, survival and death. Stochastic events leading to differentiated outcomes may precede the application of stress, propagating in a more deterministic fashion within the lineage as the stress persists. Counter-intuitively, progenitors that exhibit relatively higher maintenance activity prior to stress are not primed for survival, but are rather more likely to develop frail progeny.
Streptomycin penetrates aerobically growing bacteria and targets the ribosome, causing mistranslation of nascent proteins [17]. These in turn may misfold, resulting in the induction of RpoH-mediated heat-shock-responsive gene expression [18]. We monitored the heat-shock response using a chromosomal transcriptional fusion of the yellow fluorescent protein (YFP) to the RpoH-driven ibpAB promoter [19], [20]. This construct was found to be a highly sensitive reporter (Figure S1). We found streptomycin concentrations (<4 µg/ml), lower than the minimal inhibition concentration (MIC), where significant induction of the heat-shock response can be detected with minimal perturbation to bacterial population growth rate (Figure S2). The survival rate in these conditions, as determined by plating experiments, is 100% (see Materials and Methods).
We followed the outcome of low-dose streptomycin treatment at the single-cell level within its lineage context by time-lapse fluorescence microscopy. This allowed us to determine the extent to which a cell's stress state depends on its ancestors and life history. From a single cell exposed to antibiotics, large variations in fluorescence and growth rate phenotypes were found to propagate through the lineage (Video S1). As can be seen in this typical movie, cells may either survive or die. As early as the first division, the two daughter cells differentiate into sub-lineages: one with higher fluorescence signal, visible inclusion bodies, slower growth and fewer total divisions before the ultimate death of all its descendants. Here ‘death’ is defined as prolonged arrest in cell growth and gradual loss of contrast in phase contrast images. The other sub-lineage grows faster (engulfing the dead cousins), exhibits lower fluorescence and further develops variation in fluorescence signal and growth rate. Periodic ‘switch on’ events, characterized by increased fluorescence and slowed growth, recur within this sub-lineage (Video S1). Thus, in response to stress induction, single cells give rise to progeny of diverse phenotypes. Other examples of stressed 2D colonies can be found in Figure S3.
We further studied the emergence of variability using a microfluidic setup allowing controlled environmental changes while following micro-colony growth with time-lapse microscopy [8]. In this setup, single cells were grown without stress for four generations prior to streptomycin treatment. The micro-colonies were monitored by phase contrast and fluorescence time-lapse microscopy (Video S2, Video S3 as representative examples). The time-series images were analyzed by our custom-made open-source software ‘Cellst’ [21] to segment the cells, quantify their growth rate and fluorescence intensities, and reconstruct their lineage (Materials and Methods).
Under induced stress conditions, the pibpAB-YFP signal was found to negatively correlate with growth rate (Figure 1A). In contrast, in absence of stress, a positive correlation prevails (Figure S4A). Therefore, the promoter fusion is a valid reporter for the protein quality, streptomycin-induced stress response. Notably, when the stress is so severe that cells stop growing, overall promoter activity diminishes. As shown in Figure 1a, the correlation saturates at low growth rates. Single cell growth rates exhibit a bimodal distribution (Figure 1B), with one sub-population identified as death-prone (Figure 1A, data points in red). The commitment to eventual cell death can be traced back as early as one generation (30 minutes, Figure S6) after induction, even though the actual death may take up to 3 generations to occur (Figure 1C, Figure S6). Staining with Propidium iodide (PI), a widely used death marker that fluoresces upon intercalation between DNA bases yet can diffuse only through depolarized cellular membranes, supports our conclusion that growth-arrested cells are indeed killed by continuous antibiotic exposure (Figure S5). While a significant (>5 hours) delay occurs between growth-arrest and PI signal, all growth-arrested cells are eventually marked.
We quantified phenotypic variation as the sub-lineage coefficient of variation (SLCV) and individual coefficient of variation (IDCV) of cellular fluorescence intensity or growth rate across time (Text S1). The IDCV measures the phenotypic heterogeneity among a population of single cells regardless of their lineage relation, while the SLCV quantifies the differences among sub-populations of cells grouped according to common progenitors. For example, a single cell may divide twice to form a four-cell micro-colony. These four cells continue to divide respectively. Under normal conditions, the four subsequent sub-lineages are expected to have similar phenotypes, with relatively small differences. However, if the four sub-lineages show significantly large variation in phenotype, we would conclude that differentiation had occurred in the four-cell micro-colony, leading to significantly different sub-lineages. It follows, as depicted theoretically below, that large SLCV with respect to IDCV, indicates occurrence of differentiation.
Consider a micro-colony originated from a single cell. At time s, the micro-colony reaches Ns cells. At a later time point t>s, each cell from time s has produced ni progeny, whose fluorescence intensity or growth rate are denoted as xik (i = 1∼Ns; k = 1∼ni). Therefore, the total number of cells at time t is Nt = n1+n2…+nNs
The SLCV for a starting point s and end point t is calculated aswhere is the average phenotype among cells within the same sub-lineage and is the average across all (see Text S1 for the precise definitions).
Let IDCV be the coefficient of variation among all individual cells in the micro-colony.where is defined as the overall average of single cell phenotypes xik at given time (Text S1).
We expect that if no differentiation occurs between the sub-lineages (see Text S1 for the derivation):While if differentiation occurs:In support of this statistical model, we performed a mathematical simulation reflecting the lineage dynamics in response to the streptomycin-induced stress. A set of stochastic differential equations were constructed to describe reporter gene expression and cell division. We account for the possible positive feedback between stress level and reporter gene expression. The reporter gene expression, in turn, inversely correlates with the cellular growth rate (Figure 1A). Such feedback and correlation can lead to extended cell memory. Model parameters were set to fit the mean and variance of single cell phenotypes measured from the experimental data. As shown in Figure S7, in agreement with our expectation, the simulation results show that IDCV and SLCV are comparable in the non-stressed condition, while the extended cell memory effect leads to significantly increased SLCV in stress response.
We then calculated the actual SLCV and IDCV curves from the growth rate and fluorescence signal experimental data with different starting points Ns = 4, 8, 16, 32, 64 (eg. 2∼6 generations) under induced and non-induced conditions (Figure 2). The IDCV and SLCV values are similar and stable through 8 generations of micro-colony growth without induction, indicating no differentiation (Figure 2C and 2D). In contrast, when streptomycin is added at the 8–16 cell stage, both values increase (Figure 2A and 2B). The SLCV increases faster than the IDCV, indicating differentiation. The SLCV curves with a starting point prior to induction (Ns = 4, 8) also increase relative to the IDCV, indicating that differentiation potentially occurs among sibling cells even before they encounter the stress condition. This suggests that the stress has revealed a pre-existing difference in physiological states among the non-induced cells. In other words, there may exist pre-disposition factors in non-induced cells that prime the stress-induced differentiation.
We used data randomization to assess the significance of these experimental results. Randomly-chosen cells were switched within the lineage tree as follows: For a micro-colony with final population of N cells, N pairs of cells were chosen for switching to achieve sufficient mixing. Only cells born after stress induction were selected. In order to preserve the time course profile, switching was only allowed between cells of the same generation. As expected, while IDCV remains unchanged, SLCV decreases and is indistinguishable from IDCV (Figure S8). This result highlights the existence of extended memory effect in the original data.
We could exclude a genetic component to the observed variability increase under stress. Identical variability emerged by repeating the above microfluidics experiments with cells from an exponential phase culture, initially stressed (2 hours, 3 ug/ml Streptomycin), washed, and recovered for 4 hours in absence of stress (data not shown). Indeed, mutations would not be expected to reproducibly manifest these phenotypic effects given the rapid emergence of variability by 4–16 cell stage.
In agreement with the SLCV analysis, the detailed view of the induction phenotype within the lineage context reveals significant sub-lineage divergence as well as clustering of stress induction (Figure 3; Video S2). To highlight the existence of pre-disposition factors in single cells, we compared the RpoH-driven stress response and growth rate of the descendants of each sister cell at the tree nodes prior to induction. In most cases, there was a significant difference between the mean fluorescence (T-test, p-value <0.01; circled nodes, Figure 3) and mean growth rate (Figure S10) of the two progeny groups. To assess the significance of this result, we randomly exchanged progeny measurements in the experimentally derived tree. For each pair of sister cells born prior to induction (15 nodes), we generated 500 randomized trees where progeny were randomly re-assigned. In the stark majority of the runs, no significant difference was detected between the descendants of the pre-induction sister cells. At most, fewer than two percent of the runs per node were statistically significant (p value <0.01). This is in contrast with the experimental data (Figure 3) where the majority of these events (12 of 15) are significant indicating a 15!/12!/3!)*0.02∧12 = 2E-18 probability of generating our experimental tree by chance. This suggests that differentiation between progenitor sister cells occurred prior to stress induction.
In search of a marker for pre-disposition, we considered differentiation events occurring within the time-scale of the stressed cellular phenotype memory half-life time (90 minutes, see below, Figure 4). That is, we compared sibling progeny at 90 minutes after induction. We found no global correlation in the comparison of fluorescence intensity, promoter activity, or growth rate between the non-induced progenitor cells and their induced progeny (Figure S9). However, for the specific identified differentiation events (T-test, p-value <0.01, Figures S11, S12, S13, S14), there is a clear bias (p-value of binomial distribution test <0.003 Figures S11, S12, S13, S14) that the more fluorescent sister gives rise to a sub-lineage with more stressed siblings.
The memory or epigenetic effect was further quantified with a gene expression level auto-correlation function. In its simplest form (i.e. stable gene product and constant production rate), this auto-correlation is expected to decrease exponentially with half-life equal to the cellular doubling time ([22]; Text S1 and Figure S19). In case of nonlinear regulation, such as a positive feedback loop, the half-life will be longer than the doubling time. To this end, we calculated the auto-correlation function of the fluorescence signal, representing the RpoH-driven gene expression level. As expected, before induction, the auto-correlation function decreases exponentially with a half-life close to the cells' doubling time (23 Minutes; Figure 4). However, a significant delay of the auto-correlation function decrease is observed after induction (Figure 4). Note that the auto-correlation function half-life increases after induction to as long as three times the cell doubling time (140 minutes, Figure 4 red line). This is indicative of epigenetic effects that last longer than a generation. It is thus likely that nonlinear effects such as positive feedback contribute to the delayed decrease in auto-correlation.
It was previously proposed that streptomycin exposure could induce further streptomycin uptake by damaging the bacterial cytoplasmic (inner) membrane [23]. Such a positive feedback loop could be responsible for the epigenetic effects described above. Upon streptomycin treatment, the cytoplasmic membrane integrity is challenged by mistranslated periplasmic [23] and membrane proteins [18]. However, despite reports of increased secretion of small molecules [24], direct evidence for increased membrane permeability after streptomycin treatment is scarce. If streptomycin (molecular weight MW = 581 g/mol) treatment increases the membrane permeability, it should also increase permeability of other molecules with similar size. Therefore, controlled gene expression by transcriptional inducers such as anhydrotetracycline (ATC, analog of tetracycline, MW = 463 g/mol) should function as indicators of a parallel increase in streptomycin uptake. Similar to streptomycin, tetracycline (MW = 444 g/mol) can penetrate the outer membrane through porins [25] and diffuse across the cytoplasmic membrane. The latter step is rate limiting for both streptomycin [26] and tetracycline, with half-equilibration time of 35±15 minutes [27]. Such slow permeation rates produce detectable variation in the intracellular inducer concentration. Consider cells co-induced by streptomycin and ATC, a positive correlation between the heat-shock reporter and an ATC-inducible reporter is expected if the higher stress level induced by streptomycine leads to higher membrane permeability, with a corresponding influx of ATC molecules.
To test this hypothesis, a tetR-controlled fluorescence reporter was chromosomally integrated in the ibpAB-promoter-driven fluorescence reporter strain. When the strain was co-induced with both ATC and streptomycin, a positive correlation between two reporters was observed (Figure 5). Furthermore, compared to ATC induction alone, the expression level of the tetR reporter is stronger in the presence of streptomycin. The possibility that the positive correlation is due to elevated global protein expression level in higher stressed cells was excluded as no positive correlation was found between prrna promoter (e.g. a constitutive promoter) and pibpAB activity after stress (Figure S15). These results suggest that cells accumulate higher concentration of ATC under streptomycin stress, supporting the hypothesis that streptomycin stress increases the cytoplasmic membrane permeability. Such increased membrane permeability is likely to allow higher uptake of streptomycin as well, closing a positive feedback loop of stress induction which leads to the observed epigenetic effect (Figure 4). As control, we tested two other antibiotics at sub-inhibitory concentrations: Mitomycin C (a DNA cross-linker) and Nalidixic acid (topoisomerase inhibitor) that are not expected to significantly impact translational fidelity. Indeed, while these antibiotics induced the SOS response (judged by characteristic filamentation) they did not induce the ibpAB promoter and did not enhance but rather reduced the ATC induction levels (Figure S16).
We demonstrated that sibling E. coli cells diverge in their response to a sub-inhibitory concentration of streptomycin, to the extent that sub-populations may die and others survive within the same growing micro-colony in a homogeneously defined environment (Figure 1 and Figure S5). Upon induction, phenotypic differentiation events occurred, manifested as a stronger increase of the coefficient of variation among sub-lineages as compared to that of the coefficient of variation among individual cells (Figure 2), leading to significant differences between sister's progeny (Figure 3). Increased phenotypic variation upon stress is coupled with transient epigenetic inheritance that lasts for up to three generations, as opposed to a typical autocorrelation half-life of one generation time in absence of stress (Figure 4). Our results indicate the existence of nonlinear feedbacks that prolong the memory lifetime. The correlated expression of an ATC-induced tetR promoter and a streptomycin-induced ibpAB promoter (Figure 5) agrees with the hypothesis that streptomycin treatment leads to higher cellular membrane permeability, allowing more streptomycin as well as ATC molecules to enter the cell. Such feedback could be triggered by random events such as bursts of membrane damage by nascent mistranslated proteins or the asymmetric segregation of damaging factors during cell division [28]. While some cells are induced earlier and pass on the stressed state to descendants, others stay relatively healthy for a longer time, resulting in sustained diversification of cell fate. Therefore, we argue that positive feedback and stochasticity are responsible for the differentiation and increased variation.
Apart from membrane permeability, there may be other feedback pathways that can affect cell fate. For example, streptomycin may lead to production of ribosomes with lower accuracy, which in turn produce more dysfunctional ribosomes [29]. Or the amount of misfolded protein in the cell could exceed the capacity of the chaperone system, preventing the latter from maintaining protein homeostasis [30]. Recent work presents a revised view of the antibiotic mode of action, showing that apart from targeting a single entity, antibiotics broadly effect on the global metabolism of the cell [31], [32]. It is well established that sub-inhibitory concentration of antibiotics can directly or indirectly interact with different functional modules in the cell [33]. In the presence of stochastic events, such complex response processes are expected to produce diverse phenotypic outcomes. The process described here may play a role in other systems, since stochastic fluctuation and positive feedback are common. Our methodology could be applied to test other stresses, where incurred damage weakens the defense system of cells, leading to further damage accumulation.
The early occurrence of differentiation events (Figure 2 and Figure 3) indicates that the stress condition can reveal differences in cellular physiological state existing prior to induction. Some cells are intrinsically more resistant to stress than others. After induction, the difference is amplified and passed on in the respective sub-lineages, resulting in differentiation and increased variability. Similar pre-disposition phenomena have been reported in other induction-response systems. For example, the probability of lysogeny during phage infection is determined by both the number of infecting phage and the size of the host cell [9]. In the lactose switch, the bacterial growth rate and basal LacI level are highly predictive of switching outcome after induction [8]. In Bacillus subtilis, the decision to form an endospore is made two generations before encountering starvation conditions [10]. Whether these pre-disposition factors are a consequence of natural selection for bet hedging is mostly unclear. In the case of ampicillin ‘persisters’, a sub-population of bacteria transiently enter a dormant state in a non-stressing environment and can thus survive ampicillin treatment that attacks only growing cells [14]. Such persister cells pay a cost to express a phenotype which is less fit in the current environment but more fit for a particular environmental change. This example was interpreted as a bet-hedging strategy anticipating the arrival of future stress conditions. Yet whether it is beneficial to apply a bet-hedging strategy depends on the phenotypic switching rate, the time scale of environmental change and the fitness cost [15], or on rather stochastic events inherent to cellular physiology rather than resulting from a positive evolutionary fitness gain. In our observations, there is no sign of a pre-disposition factor working to hedge phenotypic bets. Higher stress responses prior to induction do not prime our cells for the stress to come. Instead, cells with relatively higher basal RpoH transcriptional activity are more likely to give rise to more stressed progeny (Figure S11, S12, S13, S14). This suggests that under non-stressed conditions, cells with a higher basal stress level may be paying a cost which will not help them to survive the upcoming stress. It is the weaker cells that simply suffer more, while fitter cells prevail, suggesting a simple performance-based selection.
It was recently shown that antibiotic-resistant mutants can emerge rapidly in a structured environment with a gradient of antibiotic concentrations, even from small population of 100 cells [34]. The fact that a single bacteria can generates highly variable progeny at sub-inhibitory antibiotic concentrations may facilitate this process, as it has been shown theoretically that higher variation in cellular growth rate indicates higher selection pressure [35]. Our findings may have clinical relevance as it is common that pathogens encounter sub-lethal doses of antibiotics, due either to disruptions in the prescribed medication regime or limited diffusion through structured niches such as biofilms.
All strains were derived from the wild-type strain E. coli MG1655 [36]. The YFP gene was integrated downstream of the ibpAB promoter with the ibpAB operon [19]. The strain with pibpAB-RFP and ptetR-YFP is from [37]. The strain with prrna-CFP and pibpAB-RFP is from [19].
E. coli were cultured overnight at 37°C in Luria-Bertani (LB) medium (Bacto). The cell culture was diluted and plated on LB-agar plates with different concentrations of streptomycin (0–4 µg/ml). The number of colonies on the plates were counted following overnight incubation at 37°C.
A detailed description of the microfluidic setup can be found in [8]. In short, cells were plated on a thin agarose pad (1.5% agarose in LB medium). The agarose pad was then inverted and laid on a cover slide with cells contacting the glass. A block of crosslinked poly(dimethylsiloxane) (PDMS RTV615, General Electric) with the feeding channel structures is exposed to air plasma (HARRICK PLASMA) and then placed over the agarose pad with the rest of surface area sticking to the cover-slide. LB medium or LB supplemented with 3 µg/ml streptomycin (Sigma) is injected into the feeding channel (2 ml per hour) and diffuses through the agarose pad to feed the cells. With this setup, it is possible to switch the medium on the spot with <1 minute homogenisation time [8].
We controlled for positional effects and no difference in cellular growth rate was found at different locations within a micro-colony (Figure S17 and S18). Additional controls on homogeneous permeability of the agarose layer have already been reported [8].
All the experiments are run at 37°C using a Zeiss automated microscope (Axio Observer Z1, HXP 120, 63× objective) with a temperature-controlled chamber (Live Imaging Services). For each media condition (with or without streptomycin), four single cells were chosen to be followed. Phase contrast photos were taken every 90 seconds while fluorescence photos were taken every 180 seconds (2% lamp energy, 3 second exposure).
Overnight cultures in LB 37°C were diluted 200 fold into fresh LB and agitated at 37°C for 2 hours. 1 µl of cell culture was dropped onto an agarose pad (1.5% agarose in LB medium with or without 3 µg/ml streptomycin or 25 ng/ml ATC). The agarose pad was covered with a cover-slide and the border sealed with nail polish [38], [39].
Phase contrast images were analyzed by customized software “Cellst” [21] for cell segmentation and micro-colony lineage reconstruction. The cell border was then projected onto the corresponding fluorescence image to determine the fluorescence intensity of the cells, defined as the mean grey level (background subtracted) of the pixels inside cell border. The exact location of a cell was set as the pixel coordinate of the centre of mass of the cell area. The length of a cell is measured as the long axis of the cell area.
|
10.1371/journal.ppat.1005994 | Autographa californica Multiple Nucleopolyhedrovirus Ac34 Protein Retains Cellular Actin-Related Protein 2/3 Complex in the Nucleus by Subversion of CRM1-Dependent Nuclear Export | Actin, nucleation-promoting factors (NPFs), and the actin-related protein 2/3 complex (Arp2/3) are key elements of the cellular actin polymerization machinery. With nuclear actin polymerization implicated in ever-expanding biological processes and the discovery of the nuclear import mechanisms of actin and NPFs, determining Arp2/3 nucleo-cytoplasmic shuttling mechanism is important for understanding the function of nuclear actin. A unique feature of alphabaculovirus infection of insect cells is the robust nuclear accumulation of Arp2/3, which induces actin polymerization in the nucleus to assist in virus replication. We found that Ac34, a viral late gene product encoded by the alphabaculovirus Autographa californica multiple nucleopolyhedrovirus (AcMNPV), is involved in Arp2/3 nuclear accumulation during virus infection. Further assays revealed that the subcellular distribution of Arp2/3 under steady-state conditions is controlled by chromosomal maintenance 1 (CRM1)-dependent nuclear export. Upon AcMNPV infection, Ac34 inhibits CRM1 pathway and leads to Arp2/3 retention in the nucleus.
| Actin is one of the most abundant molecules in eukaryotic cells. Actin polymerization is a process that nucleates actin monomers into filamentous structures, and this cellular process is frequently used by viruses to facilitate virus multiplication in host cells. Arp2/3, the central regulator of actin polymerization, is predominantly localized in the cytoplasm under steady-state conditions. Alphabaculoviruses assemble their progeny nucleocapsids in the nucleus of host cells, and this process is heavily dependent on nuclear actin polymerization, which requires the virus to accumulate Arp2/3 in the nucleus. Yet, how baculovirus retains Arp2/3 in the nucleus remained largely unknown. In this study, we found that the distribution of Arp2/3 is dependent on CRM1, a receptor located on the nuclear membrane that mediates the export of a large number of proteins from the nucleus to the cytoplasm. AcMNPV protein Ac34 can inhibit the CRM1 function, and lead to Arp2/3 retention in the nucleus to assist in virus replication.
| Actin polymerization is an evolutionarily conserved biological process in eukaryotic cells. The key elements of cellular actin polymerization machinery include, but are not limited to, actin, nucleation promoting factors (NPFs), and the actin-related protein 2/3 complex (Arp2/3). Arp2/3 was first isolated from Acanthamoeba castellani [1] and consists of seven subunits, including Arp2, Arp3, P40/ARPC1 (P40), P34/ARPC2 (P34), P21/ARPC3 (P21), P20/ARPC4 (P20), and P16/ARPC5 (P16) (Reviewed in [2, 3]). Activated by NPFs, Arp2/3 initiates globular actin (G-actin) polymerization into filamentous actin (F-actin) (Reviewed in [4]). Under steady-state conditions, Arp2/3 and other actin polymerization elements are predominantly localized in the cytoplasm. However, increasing evidence has shown that actin polymerization elements are also present in the nucleus and play important roles ranging from chromatin remodeling to transcription regulation (Reviewed in [5, 6]). The nuclear import mechanisms of actin and N-WASP, one of the best characterized NPFs, were previously determined [7–10], whereas nucleo-cytoplasmic shuttling mechanism of Arp2/3 remains enigmatic.
Intracellular pathogens, such as Listeria monocytogenes [11], Rickettsia spp. [12], vaccinia virus [13], alpha-herpesvirus [14], human immunodeficiency virus [15], and Burkholderia thailandensis [16], frequently use the host actin polymerization machinery to assist in pathogen reproduction (Reviewed in [17–20]). Alphabaculovirus is thus far the smallest pathogen known to profit from the host actin polymerization machinery for their propagation [21–23]. After the host cell entry of the Autographa californica multiple nucleopolyhedrovirus (AcMNPV), one of the best-characterized alphabaculoviruses, cellular Arp2/3 is activated by P78/83, a virus-encoded NPF [23]. In this way, P78/83 induces cytoplasmic actin polymerization to propel nucleocapsid migration towards the nucleus, where viral genome replication, gene transcription, and nucleocapsid assembly occur [21, 24]. However, unlike most pathogens that induce primarily cytoplasmic actin polymerization, AcMNPV also induces nuclear actin polymerization, which is essential for nucleocapsid assembly in the nucleus and for progeny nucleocapsid transport to the nuclear periphery [22, 23, 25–28]. The unique feature of nuclear actin polymerization induced by AcMNPV requires the accumulation of the cytoplasmic actin polymerization machinery, including Arp2/3, in the nucleus [27, 29–31], which makes this virus-infection system ideally suited as a research model for investigating the nucleo-cytoplasmic shuttling mechanism of Arp2/3.
Chromosomal maintenance 1 (CRM1), also known as exportin-1, is a highly versatile transport receptor in eukaryotic cells. In the nucleus, CRM1 binds to its cargo protein, usually harboring a nuclear export sequence (NES) containing a leucine-rich motif LxxxLxxLxL, along with RanGTP, to form a CRM1-cargo-RanGTP complex [32]. This complex interacts with several nucleoporins within the nuclear pore complex (NPC) and migrates across the NPC to the cytoplasm (Reviewed in [33]). After its nuclear export, RanGTP is hydrolyzed to RanGDP, and the complex releases the cargo protein to the cytosol.
In this research, we found that Arp2/3 subcellular distribution is controlled by CRM1-dependent nuclear export under steady-state conditions. AcMNPV infection induced Arp2/3 nuclear retention by inhibiting the CRM1 pathway with a viral late gene product, Ac34. To our knowledge, this is the first study describing the nuclear retention mechanism of Arp2/3 under steady-state and virus-infection conditions. We also provide the first example of a virus specifically blocking the CRM1 nuclear export pathway to promote its replication.
Previously, we and other groups have revealed the nuclear accumulation mechanism of P78/83 and G-actin [29–31], two key elements of the actin polymerization machinery, during AcMNPV infection. To investigate how AcMNPV accumulates Arp2/3, the central regulator of actin polymerization, in the nucleus, we cloned the cDNA sequences of Arp2/3 subunits from Sf9 cells, a commercially available Spodoptera frugiperda cell line commonly used for baculovirus infection (GenBank Accession: KJ187399.1, JQ364941.1, KJ187400.1, GU356595.1, KJ187401.1, KJ187402.1) [34]. Here, P40 was selected to represent Arp2/3 because P40 appeared to be the most abundant protein detected by either Western blot or fluorescence microscopy (Arp2 and P20 were less abundant than P40; Arp2 could only be detected by Western blot; other subunits were barely detected by Western blot or fluorescence microscopy when transiently expressed in Sf9 cells). We prepared plasmid-based expression constructs encoding P40 tagged with a V5 epitope (P40-V5) at its C-terminus or P40 fused to enhanced green fluorescent protein at its N-terminus (EGFP-P40) to monitor the Arp2/3 dynamics during AcMNPV infection.
Cytoplasmic localization was noted by immunofluorescence for P40-V5 for mock infected cells (Fig 1A, left panel), but some nuclear localization was observed for cells infected with AcMNPV carrying an EGFP marker (vAcegfp, diagramed in S1A Fig). As evidenced by cell fraction and Western blot, P40-V5 was present in only the cytoplasmic fraction of mock infected cells, while some P40-V5 was found in the nuclear fraction of vAcegfp infected cells (Fig 1A, right panel). The nuclear and cytoplasmic control proteins, histone and tubulin respectively, were identified in the nuclear and cytoplasmic fractions, respectively, validating the effectiveness of the fractionation (Fig 1A, right panel). Similarly, by fluorescence microscopy, EGFP-P40 localized to the nucleus only in cells infected with AcMNPV expressing polyhedrin (vAcpolh, diagramed in S1A Fig) (Fig 1B). This phenotype is in accordance with the observation described by Goley et al., in which yellow fluorescent protein-tagged P21 (P21-YFP) was observed to accumulate in the nucleus during AcMNPV infection [23].
To test whether EGFP-P40 associates with other Arp2/3 subunits, Arp2-Ha was co-expressed with EGFP or EGFP-P40 in Sf9 cells, respectively. Western blot assay demonstrated that Arp2-Ha (approx. 46 kDa), EGFP (approx. 27 kDa), and EGFP-P40 (approx. 69 kDa) were present in the whole cell lysates (WCL) (Fig 1C, left panel). A co-immunoprecipitation (Co-IP) assay using anti-Ha showed that EGFP-P40, but not EGFP, was pulled down with Arp2-Ha (Fig 1C, left panel), indicating that EGFP-P40 is associated with Arp2-Ha. Similarly, EGFP-P40 is shown to interact with P20-Ha (approx. 21 kDa) (Fig 1C, right panel), implying that EGFP fusion to P40 does not impair the incorporation of P40 into Arp2/3. Taken together, these phenotypes demonstrated that either the C-terminally tagged P40-V5 or the N-terminally tagged EGFP-P40 can be used to monitor Arp2/3 dynamics during AcMNPV infection.
We next investigated which class of viral genes needed to be expressed for P40 nuclear accumulation. Aphidicoline (APH), an inhibitor of DNA synthesis, was used to shut off AcMNPV late gene expression [35]. The dynamic localization of P40 in AcMNPV-infected cells was monitored in the presence or absence of APH. Early during infection (0–12 hpi), P40 resided predominantly in the cytoplasm irrespective of APH treatment (Fig 1D). During the late phase of infection (After 12 hpi), AcMNPV infection resulted in detectable P40 accumulation in the nucleus in the absence of APH, suggesting that viral late gene products may play an important role in P40 nuclear accumulation. When the expression of viral late genes was shut off by APH, P40 failed to accumulate in the nucleus, as demonstrated by both immunofluorescence microscopy and cell fractionation assays (Fig 1D). Together, these data indicated that viral late gene products are responsible for P40 nuclear accumulation.
To identify the viral protein responsible for the nuclear accumulation of P40, AcMNPV ORFs were individually cloned into a pIZ-V5 transient expression vector (Invitrogen). Each individual viral ORF was co-expressed with EGFP-P40 and the subcellular distribution of P40 was determined using fluorescence microscopy. Among the 118 viral ORFs screened (S1 Table), only Ac34, a viral late gene product, appeared to be sufficient to induce P40 nuclear accumulation. Ac34 tagged with mCherry (mC-Ac34) was shown to accumulate EGFP-P40 or P40-V5 in the nucleus when co-expressed in Sf9 cells (Fig 2A and 2B). As a control, we co-expressed P40 with non-fused mCherry (Fig 2A and 2B), resulting in a predominantly cytoplasmic localization of P40. Similar nuclear relocation induced by Ac34 also occurred for P20 (S2A Fig), indicating that Ac34 is sufficient to accumulate Arp2/3 in the nucleus.
To further verify the role of Ac34 in P40 nuclear accumulation during AcMNPV infection, an ac34-knockout bacmid with an EGFP expression cassette (vAc34KOegfp, diagramed in S1B Fig) was constructed [36]. Immunofluorescence microscopy at 48 hours post-transfection (hpt) demonstrated that exogenous P40 (P40-V5) resided in the cytoplasm of vAc34KOegfp-transfected cells, whereas the restoration of ac34 to vAc34KOegfp (vAc34KOac34, diagramed in S1B Fig) could accumulate P40-V5 in the nucleus (Fig 2C). Similar nuclear accumulation also occurred for P20 (S2B Fig), indicating that ac34 is responsible for the Arp2/3 nuclear accumulation induced by AcMNPV.
Previously, we revealed that virus-encoded NPF P78/83, another key element of the nuclear actin polymerization machinery during AcMNPV infection, is relocated to the nucleus by binding to and co-transportation with C42, which harbors a nuclear localization sequence (NLS) [31]. Based on this scenario and the fact that Ac34 is present in the nucleus (Fig 2A and 2B), we were prompted to explore whether P40 nuclear accumulation is also correlated to the presence of Ac34 in the nucleus.
A series of mCherry-fused C-terminal and N-terminal Ac34 truncations (S1C Fig) was prepared to identify the sequence responsible for Ac34 nuclear localization. Fluorescence microscopy demonstrated that the removal of amino acids (aa) 195–215 of Ac34 (mC-Ac341-195) resulted in the cytoplasmic localization of Ac34 (Fig 3), which is in sharp contrast to the full-length Ac34 (mC-Ac34) and all the tested N-terminal Ac34 truncations, which exhibited a predominantly nuclear localization pattern (S3 Fig). This phenotype indicated that the aa 195–215 region plays a major role in determining the presence of Ac34 in the nucleus, although sequence analysis did not show any classic NLS pattern (tandem repeats of lysine and arginine) within this region. Notably, when the C-terminal truncation of Ac34 was extended to aa 75 or further (mC-Ac341-75 and mC-Ac341-55), a diffuse cellular distribution of Ac34 was observed (Fig 3), which could be attributed to free nucleo-cytoplasmic shuttling of the resulting low-molecular-mass polypeptides.
Interestingly, among all the tested Ac34 truncations, only full-length Ac34 could accumulate EGFP-P40 in the nucleus, and the removal of aa 195–215 of Ac34 resulted in a lack of EGFP-P40 nuclear accumulation (Fig 3), thus supporting our hypothesis that P40 nuclear accumulation is dependent on the presence of Ac34 in the nucleus. Similar nuclear accumulation also occurred for P20 (S2A Fig), indicating that the aa 195–215 region is essential for Ac34 to accumulate Arp2/3 in the nucleus.
To verify the role of aa 195–215 of Ac34 in P40 nuclear accumulation during AcMNPV infection, Ac341-195 was used to rescue vAc34KOegfp, generating vAc34KOac34Δ195–215 (diagramed in S1B Fig). When P40 was co-expressed in bacmid-transfected cells, only vAc34KOac34 could induce P40 nuclear accumulation at 48 hpt, in contrast to the cytoplasmic distribution pattern of P40 in vAc34KOegfp and vAc34KOac34Δ195-215-transfected cells (Fig 4A). Similar nuclear accumulation also occurred for P20 (S2B Fig), further confirming that Ac34 is responsible for the Arp2/3 nuclear accumulation induced by AcMNPV, and aa 195–215 are required for the accumulation.
Nuclear actin polymerization requires the nuclear localization of Arp2/3. To explore whether Ac34 is involved in AcMNPV-induced nuclear actin polymerization, Sf9 cells were transfected with vAc34KOegfp, vAc34KOac34, or vAc34KOac34Δ195–215 and stained with phalloidin at 48 hpt to visualize F-actin. Among all the transfected bacmids, only vAc34KOac34 induced typical nuclear actin polymerization, with F-actin accumulating in the nuclear region (Fig 4B). The cells transfected with the other bacmids showed no significant F-actin accumulation in the nucleus (Fig 4B). This phenotype can easily be attributed to the absence of Arp2/3 in the nucleus due to either ac34 knockout (vAc34KOegfp) or the loss of its nuclear localization determinant (vAc34KOac34Δ195–215).
CRM1 is a highly versatile transport receptor that mediates the nuclear export of a large number of proteins. Inhibition of CRM1 results in nuclear retention of NES-bearing protein. Bioinformatics assay (LocNES, http://prodata.swmed.edu/LocNES/) [37] predicted that the P40 C-terminus (aa 360–374), a leucine-rich sequence, is a putative NES. We then explored whether the cytoplasmic distribution of P40 is CRM1-dependent.
P40-V5 was transiently expressed in Sf9 cells. Immunofluorescence microscopy showed that P40 exhibited significant nuclear accumulation after adding leptomycin B (LMB), a specific CRM1 inhibitor (Fig 5A) [38–40]. Removing aa 360–374 of P40 resulted in P40 (P40Δ360-374-V5) accumulation in the nucleus (Fig 5A), implying that the P40 C-terminus functions as a NES to determine the cytoplasmic distribution of P40. To further confirm P40 nuclear accumulation is CRM1-dependent, cellular CRM1 was knocked-down using double-stranded RNA (dsRNA) targeting the 1–1000 nt (ds-crm11-1000) or the 1001–2000 nt (ds-crm11001-2000) of CRM1 mRNA (Genbank accession KT208379.1). Western blot assay demonstrated that both dsRNAs significantly down-regulated the endogenous CRM1 level (Fig 5B). Nuclear accumulation of P40-V5 and EGFP-P40 was observed in the ds-crm11-1000 bearing cells, in comparison with the control cells (Fig 5C).
Similar nuclear retention upon LMB treatment (S2A Fig) or CRM1 knockdown (S2B Fig) was also observed in P20-expressing cells, indicating that the presence of Arp2/3 in the cytoplasm is controlled by CRM1-dependent nuclear export.
Given the evidence that the cytoplasmic distribution of Arp2/3 is controlled by CRM1-dependent nuclear export, and AcMNPV infection induces Arp2/3 nuclear accumulation, one of the possible explanations is that AcMNPV infection inhibits cellular CRM1-dependent nuclear export and subsequently leads to Arp2/3 retention in the nucleus.
To evaluate the influence of AcMNPV infection on the CRM1 pathway, a classic NES peptide (LQNKLEELDL) [41] was fused to mCherry (mCherry-NES) and EGFP (EGFP-NES) to construct probes for CRM1-dependent nuclear export.
When mCherry-NES was transiently expressed in Sf9 cells, a predominantly cytoplasmic distribution pattern was observed (Fig 6). Adding LMB to the culture medium resulted in the accumulation of the majority of mCherry-NES in the nucleus (Fig 6), indicating that the nuclear export of mCherry-NES is CRM1-dependent. Similar nuclear retention upon CRM1 knockdown was also observed in EGFP-NES expressing cells (S4 Fig).
When virus stock solution (vAcegfp) was added to the culture medium, mCherry-NES accumulated in the nucleus of infected cells and remained in the cytoplasm of uninfected cells (Fig 6). This differential distribution indicated that AcMNPV causes dysfunctional cellular CRM1-dependent nuclear export. To identify which class of viral genes was responsible for the dysfunction, APH was added to the culture medium after virus infection. All the cells showed cytoplasmic distribution of mCherry-NES (Fig 6), suggesting that AcMNPV late gene products were responsible for the virus-induced dysfunction in the CRM1 pathway.
Because we demonstrated that Ac34 is responsible for virus-induced Arp2/3 nuclear accumulation, and AcMNPV inhibits CRM1-dependent nuclear export, which can lead to Arp2/3 retention in the nucleus, it is highly possible that Ac34 is involved in the dysfunction of the CRM1 pathway induced by AcMNPV.
To test this hypothesis, EGFP-NES was co-expressed with mCherry or mC-Ac34 in Sf9 cells. Fluorescence microscopy showed that EGFP-NES resided in the cytoplasm in the presence of mCherry, whereas it accumulated in the nucleus in the presence of mC-Ac34 or LMB (Fig 7A). This phenotype indicated that Ac34 is sufficient to inhibit CRM1-dependent nuclear export. Removing the Ac34 C-terminus (aa 195–215), which is essential for Ac34 nuclear localization and Arp2/3 nuclear accumulation, also abolished EGFP-NES nuclear retention (Fig 7A).
To validate the role of Ac34 in AcMNPV-induced CRM1 pathway dysfunction, mCherry-NES was co-expressed in bacmid-transfected cells. Fluorescence microscopy showed that mCherry-NES resided in the cytoplasm in vAc34koegfp-transfected cells, whereas the restoration of ac34 (vAc34koac34), but not ac34Δ195–215 (vAc34koac34Δ195–215), could accumulate mCherry-NES in the nucleus (Fig 7B), indicating that Ac34 is involved in the CRM1 pathway dysfunction induced by AcMNPV.
Taken together, this evidence demonstrated that Ac34 induces Arp2/3 nuclear retention by inhibiting CRM1-dependent nuclear export during AcMNPV infection.
The nuclear import mechanisms of key elements of actin polymerization machinery, including actin and N-WASP, have been previously identified [7–10]. However, nucleo-cytoplasmic shuttling mechanism of Arp2/3, the central regulator of actin polymerization, has not been elucidated yet. In this study, a unique virus-infection system was employed to reveal how Arp2/3 is retained in the nucleus, which could shed light on the nucleo-cytoplasmic shuttling mechanism of Arp2/3 under different physiological or pathophysiological conditions.
Viral manipulation of cellular the nucleo-cytoplasmic transport of proteins has been extensively documented in recent years (reviewed in [42]), in particular in cardioviruses and enteroviruses. Cardioviruses use their leader proteins to induce the hyper-phosphorylation of nucleoporins and disrupt the RanGTP gradient [43, 44], thus inducing an efflux of the nuclear proteins required for viral replication and leading to interferon suppression. Enterovirus infection results in cellular protein retention in the cytoplasm via the degradation of nucleoporins mediated by the virus-encoded proteases 2A and 3C [45–47]. Other viruses, such as herpes simplex virus [48], human papillomavirus [49, 50], severe acute respiratory syndrome coronavirus [51], Ebola virus [52], and measles virus [53], employ a variety of methods to interfere with the nucleo-cytoplasmic shuttling of cellular proteins, therefore facilitating viral replication and escape from the host anti-viral immune response.
Unlike most viruses, which primarily induce impaired protein nuclear import or enhance protein nuclear export, our results demonstrated that AcMNPV infection results in impaired protein nuclear export. As a nucleopolyhedrovirus, most AcMNPV replication processes, including viral genome replication, gene transcription, and nucleocapsid assembly, all occur in the nucleus. These processes require a variety of proteins, including, but not limited to, virus-encoded transcription factors, transcriptases, and capsid proteins, as well as some cellular proteins (e.g., actin, Arp2/3), to accumulate in the nucleus. AcMNPV contains 156 predicted ORFs at least 50 aa in length. Aside from a limited number of exceptions, the nuclear import mechanisms of most viral and cellular proteins during AcMNPV infection remain unknown. Currently, at least 7 exportins have been identified in eukaryotic cells [8, 32, 54–58]. Unlike other exportins that only transport highly specialized cargoes (Reviewer in [59]), CRM1 mediates the nuclear export of many NES-bearing proteins, and its dysfunction leads to the nuclear accumulation of these proteins. Based on bioinformatics prediction (NetNES, http://www.cbs.dtu.dk/services/NetNES/) [60], 98 AcMNPV proteins contain putative residues that could serve as a NES (S2 Table). Such a high percentage of viral proteins bearing putative NESs implies that CRM1-dependent nuclear export may determine the subcellular distribution of many viral proteins, and the inhibition of CRM1-dependent nuclear export by Ac34 could possibly play a key role in the AcMNPV-induced nuclear accumulation of proteins. Whether Ac34 also influences other exportins or these exportins also contribute to the virus-induced protein nuclear accumulation remain to be explored.
Ac34 homologues are presented in all sequenced alphabaculoviruses but absent in betabaculoviruses [61]. Alphabaculoviruses and betabaculoviruses behave in significantly different ways. In respect to cytopathology, alphabaculoviruses assemble their nucleocapsid in the nucleus, whereas betabaculoviruses induce nuclear membrane rupture, and nucleocapsid assembly occurs in a combination of the cytoplasm and the nucleoplasm [62]. This cytopathologic difference suggests that unlike alphabaculoviruses, betabaculoviruses do not need to accumulate the cytoplasmic actin polymerization machinery to the nucleus. As a consequence, betabaculoviruses do not need a viral protein or mechanism to induce nuclear accumulation of Arp2/3 (although only P40 and P20 were proved to be retained in the nucleus of AcMNPV-infected cells in this study, both Arp2/3 components behave in a similar way upon virus infection), which is supported by the evidence that Ac34 homologues are absent in the genomes of betabaculoviruses [61].
Nuclear G-actin is required for the transcriptional activity of RNA polymerases [63–65] and the epigenetic activation of chromatin (Reviewed in [5, 66]). Among the three key actin polymerization elements that are accumulated in the nucleus during AcMNPV infection, only G-actin is recruited to the nucleus by early viral gene products [29, 30]. This early nuclear accumulation of G-actin could increase the nuclear G-actin pool and promote the transcription of viral early genes that are transcribed by host RNA polymerase II [61]. Late in infection, P78/83 and Arp2/3 accumulate in the nucleus and induce nuclear actin polymerization that converts G-actin to F-actin. The resulting nuclear G-actin pool depletion could lead to the loss of the transcriptional activity of host RNA polymerases and the epigenetic reprogramming of host chromatin towards transcriptional inhibition, which could contribute to the host gene transcription shutoff that occurs in the late phase of baculovirus infection [67, 68]. Consistent with this, cytochalasin D, a chemical that specifically prevents actin polymerization, behaves as an antagonist of the virus-induced shutdown of host gene expression [69]. In this respect, nuclear actin polymerization induced by baculovirus infection may also participate in the regulation of host/virus gene expression by the modulation of the nuclear G-actin pool, in addition to its role in assisting viral nucleocapsid assembly and transport, which has long been recognized.
In summary, Ac34 subversion of the CRM1-dependent nuclear export during AcMNPV infection suggests that alphabaculoviruses may employ an efficient way by encoding a single protein to accumulate multiple viral and host proteins in the nucleus to assist in virus replication. As a key element of actin polymerization machinery, Arp2/3 is present in both the cytoplasm and the nucleus. Our finding that Arp2/3 nuclear-cytoplasmic shuttling is CRM1-dependent sheds light on how cells manage to control actin polymerization machinery in different cellular compartments to exert different functions.
Sf9 cells from S. frugiperda were cultured in Grace’s medium (Invitrogen) with 5% fetal bovine serum (Invitrogen) and 0.1% Antibiotic-Antimycotic (Invitrogen) at 27°C. Sf9 cells were transfected with the indicated plasmids or bacmids using the Cellfectin II reagent (Invitrogen) following the standard procedures. For infection, the Sf9 cells were incubated with virus stock solution for 1 h at a multiplicity of infection (MOI) of 2. (MOI = 2). The cells were then rinsed twice and then incubated in fresh medium or medium with APH (5 μg/ml) (Sigma). The cells were fixed for further immunofluorescence detection at 6, 12, and 24 hpi. To block CRM1-dependent nuclear export, LMB (0.1 μg/ml) (Beyotime) was added to the culture medium and the cells were incubated for 4 hours before the fluorescence assays.
One hundred fifty-four ORFs of AcMNPV were cloned by polymerase chain reaction (PCR) and inserted into pIZ-V5 (Invitrogen). All the viral ORFs began with ATG and ended without the stop codon to create an in-frame fusion with the V5 epitope. All the constructs were sequenced, and 118 viral ORFs were tested for their impact on the change in P40 subcellular distribution (S1 Table).
All the plasmids used in this research for transient expression were prepared by standard molecular cloning protocols. The indicated genes, gene truncations, and genes with epitope tags were generated by PCR or site-directed mutagenesis (Transgene) and inserted into pIZ-V5/Ha vectors (Invitrogen).
To prepare recombinant bacmids, the Bac-to-Bac system was employed according to Invitrogen’s protocol. In brief, Ac34 expression cassettes controlled by the native ac34 promoter were cloned into pFbdg, a pFastbac-Dual vector (Invitrogen) bearing an EGFP expression cassette controlled by the p10 promoter [31]. The resulting shuttle vectors were then used to transform DH10B E. coli cells harboring the vAc34KO bacmid provided by Cai et al. to generate the transposed bacmid constructs [36].
Maps of the plasmids and bacmids prepared in this research are diagramed in S1 Fig.
Cells were rinsed with ice-cold PBS and lysed with homogenization buffer (10 mM HEPES pH = 7.9, 10 mM KCl, 1.5 mM MgCl2, 0.1 mM EGTA, 0.5 mM DTT, 2 mM PMSF, 1 μg/ml Proteinase Inhibitors (Roche)). The cell membranes were disrupted by passing through a 25G needle 5 times, and the lysates were then spun at 1000×g for 10 min at 4°C. The supernatant containing the crude cytoplasmic fraction was collected in 1.5 ml tubes and spun at 20,817×g for 30 min at 4°C, and the supernatant was collected as the purified cytoplasmic fraction. The nuclear pellet was rinsed in 1 ml homogenization buffer and centrifuged at 1000×g for 10 min at 4°C. The pellet was re-suspended in 100 μl extraction buffer (10 mM HEPES pH = 7.9, 0.4 M NaCl, 1.5 mM MgCl2, 0.1 mM EGTA, 0.5 mM DTT, 2 mM PMSF, 1 μg/ml Proteinase Inhibitors) under gentle shaking for 30 min at 4°C. The suspension was centrifuged at 20,817×g for 30 min at 4°C and the supernatant was collected as the nuclear fraction. The protein concentrations of all samples were determined using Bradford assays (Bio-Rad) and the samples were subjected to Western blot assays. Anti-histone H3 (Sigma) and anti-tubulin (Sigma) diluted to 1:1000 were used to verify the quality of the cytoplasmic and nuclear fractions, respectively. After HRP-conjugated secondary antibody (1:10,000 dilution, Jackson Laboratory) incubation, the blots were developed using an enhanced chemiluminescence kit (Pierce).
Sf9 cells were rinsed with ice-cold PBS and lysed with RIPA buffer (50 mM Tris, pH = 7.5, 1 mM EGTA, 1 mM EDTA, 1% Triton X-100, 150 mM NaCl, 2 mM DTT, 100 μM PMSF, 1 μg/ml Proteinase Inhibitors). The cell lysates were centrifuged at 20,817×g at 4°C for 10 min and the supernatants (WCL) were collected. The protein concentrations of the WCL were determined by Bradford assays and 1500 μg was mixed with 2 μg anti-Ha (Sigma) and Protein G Agarose (Millipore) and incubated at 4°C overnight according to the manufacturer’s protocol. The immunoprecipitated samples were centrifuged and washed three times and subjected to Western blot assays using anti-Ha (1:1000 dilution) and anti-EGFP (1:1000 dilution, Invitrogen).
The immunofluorescence assays were performed as described previously [31]. Briefly, the cells were fixed with 3.7% paraformaldehyde in PBS for 30 min, permeabilized with 0.5% Triton X-100 and blocked in 1% normal goat serum (Boster) in PBS for 30 min on ice. The cells were incubated with anti-V5 (1:500 dilution, Invitrogen) or anti-Ha (1:500 dilution, Sigma) primary antibodies. The secondary antibodies were Alexa Fluor 568- or 488-conjugated anti-mouse and anti-rabbit antibodies (1:500 dilution, Invitrogen). The nuclear DNA was stained with Hoechst 33258 (Beyotime).
For F-actin staining, the cells were transfected with different recombinant bacmids, fixed, and permeabilized as described above and then stained with 0.7 U/ml Alexa Fluor 568-phalloidin (Invitrogen) and Hoechst 33258 for 10 min. The cells were then washed three times with PBS and examined by confocal microscopy using a PerkinElmer UltraVIEW VoX microscope.
The fluorescence quantification data were obtained using Volocity 6.3 software (PerkinElmer) and Student’s T-test was performed to compare the differences between the tested samples.
To knockdown the expression of CRM1, primers encompassing the 1–1000 nt (TAATACGACTCACTATAGGGATGGCAACTTTAGAGCAACA, TAATACGACTCACTATAGGGACTTCAGATATCAGTACAAG) or the 1001–2000 nt (TAATACGACTCACTATAGGGAGAAGAAGTAGAAATTTTTA, TAATACGACTCACTATAGGGTGTCCAAATATATTCTACCC) of S. frugiperda CRM1 mRNA (Genbank accession: KT208379.1) were synthesized and served as gene specific primers to prepare dsRNA by using MEGAscript RNAi kit (Ambion) according to the manufacturer’s protocols. Sf9 cells were transfected with 5 μg dsRNA/105 cells using the Cellfectin II reagent (Invitrogen).
|
10.1371/journal.pgen.1003389 | Molecular Networks of Human Muscle Adaptation to Exercise and Age | Physical activity and molecular ageing presumably interact to precipitate musculoskeletal decline in humans with age. Herein, we have delineated molecular networks for these two major components of sarcopenic risk using multiple independent clinical cohorts. We generated genome-wide transcript profiles from individuals (n = 44) who then undertook 20 weeks of supervised resistance-exercise training (RET). Expectedly, our subjects exhibited a marked range of hypertrophic responses (3% to +28%), and when applying Ingenuity Pathway Analysis (IPA) up-stream analysis to ∼580 genes that co-varied with gain in lean mass, we identified rapamycin (mTOR) signaling associating with growth (P = 1.4×10−30). Paradoxically, those displaying most hypertrophy exhibited an inhibited mTOR activation signature, including the striking down-regulation of 70 rRNAs. Differential analysis found networks mimicking developmental processes (activated all-trans-retinoic acid (ATRA, Z-score = 4.5; P = 6×10−13) and inhibited aryl-hydrocarbon receptor signaling (AhR, Z-score = −2.3; P = 3×10−7)) with RET. Intriguingly, as ATRA and AhR gene-sets were also a feature of endurance exercise training (EET), they appear to represent “generic” physical activity responsive gene-networks. For age, we found that differential gene-expression methods do not produce consistent molecular differences between young versus old individuals. Instead, utilizing two independent cohorts (n = 45 and n = 52), with a continuum of subject ages (18–78 y), the first reproducible set of age-related transcripts in human muscle was identified. This analysis identified ∼500 genes highly enriched in post-transcriptional processes (P = 1×10−6) and with negligible links to the aforementioned generic exercise regulated gene-sets and some overlap with ribosomal genes. The RNA signatures from multiple compounds all targeting serotonin, DNA topoisomerase antagonism, and RXR activation were significantly related to the muscle age-related genes. Finally, a number of specific chromosomal loci, including 1q12 and 13q21, contributed by more than chance to the age-related gene list (P = 0.01–0.005), implying possible epigenetic events. We conclude that human muscle age-related molecular processes appear distinct from the processes regulated by those of physical activity.
| A fundamental challenge for modern medicine is to generate new strategies to cope with the rising proportion of older people within society, as unaddressed it will make many health care systems financially unviable. Ageing impacts both quality of life and longevity through reduced musculoskeletal function. What is unknown in humans is whether the decline with age, referred to as “sarcopenia,” represents a molecular ageing process or whether it is primarily driven by alterations in lifestyle, e.g. reduced physical activity and poor nutrition. Because the details of such interactions will be uniquely human, we aimed to produce the first reproducible global molecular profile of human muscle age, one that could be validated across independent clinical cohorts to ensure its general applicability. We combined this analysis with extensive data on the impact of exercise training on human muscle phenotype to then identify the processes predominately associated with age and not environment. We were able to identify unique gene pathways associated with human muscle growth and age and were able to conclude that human muscle age-related molecular processes appear distinct from the processes directly regulated by those of physical activity.
| Discovery of the biological determinants of muscle mass and functional molecular phenotypes has substantial bearing on the fields of human performance (e.g. hypertrophy, strength or endurance adaptations [1], [2]) and human health (countering muscle atrophy and deconditioning occurring in older age or with conditions such as cancer cachexia [3], [4], respiratory disease or medically enforced immobilization (e.g. hospitalized bed-rest, cast-immobilization [5], [6])). Resistance exercise (RE) training (RET) is an effective intervention to increase muscle mass in many, but not all people, and thus provides an excellent opportunity to study gene-network regulation during muscle hypertrophy and the proposed relationship to muscle aging. Many exogenous factors may influence RET-induced hypertrophy including manipulation of exercise volume [7], intensity [8] and adequate macronutrient availability [9] all of which interact with an individual's genotype to determine muscle growth. Establishing the molecular diagnostics that enable a personalized approach to tackle ageing has great appeal.
To date, the molecular regulation of muscle hypertrophy has mostly focused on aspects of post-genomic signaling, with early work concluding that canonical insulin-like growth factor (IGF-1) signals control muscle hypertrophy though a phosphatidyl-inositol-3 kinase/protein kinase B/mechanistic target of rapamycin (PI3K/AKT/mTORc1) pathway [10], abbreviated to mTOR. This protein complex can control cell growth through two mechanisms; firstly, mTORC1 regulates efficiency of translation through inducing phosphorylation of its substrates, ribosomal protein S6 kinase (S6K1) and 4E binding protein 1 (4EBP1) and, secondly, mTORC1 increases translational capacity through regulating ribosomal RNA (rRNA) production within the nucleolus [11]. There are however conflicting data regarding the importance of mTOR regulation (protein phosphorylation or target gene mRNA responses) or its up-stream regulators, and acute anabolic or chronic growth responses to resistance exercise [12]–[17] reported from the same laboratories, indicating that important biological rather than methodological issues remain to be identified.
More recent evidence indicates that the mechanisms regulating muscle hypertrophy go beyond the canonical IGF-1/PI3K/AKT/mTORc1 pathway. While circulating IGF-1 concentrations do not determine RET-induced hypertrophy in humans [18], hypertrophy has now been shown to potentially occur through both PI3K-AKT [19] and mTOR [20] independent pathways, even in pre-clinical models. Perhaps the most convincing observation in favor of a more divergent regulation of muscle growth, is the fact that disparate exercise modes (e.g. RET vs. endurance exercise training (EET)) can produce similar protein signaling patterns in humans [21]. This suggests that the molecules, so far studied, are pleiotropic and in our opinion probably important for any type of tissue remodeling, regardless of the final physiological phenotype [22]. Therefore, a more innovative approach is needed to define links between molecules and ensuing in vivo physiological adaptations, than can be achieved with targeted western-based molecular profiling.
Exercise training has also been postulated as a key tool to reverse the impact of ageing on human skeletal muscle phenotypes [23], [24]. Yet, while some ‘genomic features’ of ageing have been reported [25], we have noticed that the available global molecular profiles of human muscle [23], [24], [26], [27] do not identify consistent molecular features. Furthermore, our recent work has highlighted that physiological adaptations to exercise, whether that be hypertrophy [28] or aerobic function [29], are highly heterogeneous in humans, implying that exercise may not be able to “reverse” muscle ageing [23] for some people. For example, following >10-wk of supervised EET, ∼20% of subjects show no improvement in aerobic capacity while ∼30% demonstrate no improvement in insulin sensitivity [30], [31]. Similarly, we reported muscle hypertrophy ranging from 0.8 to 6.0 kg [28], while Raue et al reported changes in muscle cross-sectional area (CSA) ranging −1.2 to +10.4 cm2 [24]. Both of these RET studies reported that the outcome of supervised progressive RET did not relate to pre-existing differences in characteristics (i.e., gender, age, pre-existing muscle mass, physical activity levels or dietary habits) indicating that there is not a simple explanation for the heterogeneity of the gains in lean mass [28], [32].
In recent years, we have focused on using the heterogeneous responses to exercise training and OMIC techniques to uncover molecular networks regulated by EET [33][29] or generate molecular predictors of trainability [34], directly in humans. The aim of the present work is to produce the first reproducible molecular signature of human muscle age, and examine how such a profile relates to new and established exercise adaptation gene networks. We generated new gene-chip profiles from muscle samples derived from two independent clinical cohorts, with a continuous range of ages (18–79 y) and which originate from distinct environments (UK and USA) and which were independently processed in the laboratory. We also generated a new data set of paired global RNA-responses to a supervised 20-wk RET program (N = 44), as well as utilizing various sets of published acute-RET and chronic-EET gene-chips (total N = 200) data sets. Finally, Ingenuity's new IPA up-stream analysis tool [35] was used to identify key features, within these novel age and exercise signatures, to provide independent and robust molecular insight into the heterogeneous nature of muscle hypertrophy, and human muscle age.
The paired differential analysis, comparing expression in 38 from 44 subjects before and after RET yielded a dataset of ∼700 regulated genes (Dataset S1) and this gene-list related to a few upstream regulators in IPA (Dataset S2). For logical reasons we used only the 38 subjects that demonstrated a training effect [31]. This list included a 62-gene network (Figure 1A), representing the transcriptional action of all-trans-retinoic acid (Tretinoin) and this signature was highly ‘activated’ following 20-wk RET (Z-score = 4.5 for directional consistency; P-value for transcript overlap (p = 6×10−13)). In contrast, aryl hydrocarbon (AhR) and V-Ki-ras2 Kirsten rat sarcoma viral oncogene homolog (KRAS) signaling (Figure S1A) was inhibited, such that repression of connected genes was removed, and thus target mRNAs up-regulated.
We utilised the differential RNA response to 6-wks of EET [29] to address the question of specificity of this RET molecular profile to conditions of muscle growth. We have previously speculated that a core set of gene-networks will be common to all types of exercise and could represent basic determinants of tissue adaptability [22]. Indeed, we found that several significant upstream transcriptional regulators from the RET profile were also significant significantly regulated by EET (Dataset S1) and the majority of individual genes in the RET RNA signatures were activated in an identical manner between RET and EET (Figure 1B and Figure S1B). Thus while differential expression in the Derby Resistance Exercise Training (DRET) study clearly identifies a number of tissue remodeling related processes, these are not specific to exercise training modality.
We then utilized the dataset from the Trappe laboratory (GSE28422, ‘Study A’ from GEO). We removed the following samples (or the related paired sample) because the following samples appeared to have technical issues; GSM702424, GSM702439, GSM702431, GSM702421, GSM702414, GSM702436 and GSM702405 GSM702393, GSM702399, GSM702411. Unfortunately we were declined access to the phenotype data (individual muscle growth changes) for each subject (Dr S Trappe, personal communication) and thus we were unable to re-analyze their data to capture the growth related responses, as their data-set included a number of ‘non-responders’ for lean mass gains. Nonetheless, we used the acute RE response dataset found in the Trappe study (GSE28422) so that we could explore the overlap between the acute RNA responses and chronic RET adaptation. What we noted was that <2% of the RNA changes that occur with chronic adaptation in the DRET study, were regulated in response to acute resistance exercise.
Of greater physiological importance was our effort to identify genes which link to the magnitude of muscle hypertrophy in humans. Quantitative SAM analysis [3], [36] identified 642 probe-sets and interestingly the majority of genes identified were negatively correlated with gains in lean mass (Dataset S2). The probe-set list was imported into IPA (filtered at a 5% FDR) and 384 genes could be mapped to the IPA database for pathway analysis. The enrichment score generated for the EIF2 canonical pathway enrichment was extremely significant (P-value = 1×10−64) and the combination of these observations indicates that the gene-list was both predictable (valid with extensive literature) and statistically very robust (Figure S2). We identified a number of regulators that could be responsible for regulating the transcriptional signature that correlated to gain in lean mass (Dataset S2).
Figure 2A presents the most striking finding, an active rapamycin signature, equating to inhibition of mTOR signaling [37]. This signature was comprised of genes that almost entirely negatively correlated with lean-mass gain (Z-score = 2.8 for directional consistency; P-value for transcript overlap p = 1.4×10−30). In short, subjects that demonstrated the largest gains in lean tissue mass following 20-wks of RET have suppressed mTOR signaling over the training period, a novel and controversial observation (all raw data was manually checked for consistency of direction). A second major transcriptional regulator was MYC, and the gene-list driving its inclusion (Z-score −5.8 for directional consistency; p-value for transcript overlap p = 4×10−43) overlapped with the rapamycin network, equating to the inhibition of MYC activity. MYC can be an upstream activator of mTOR signaling in cell culture systems [38]. Thus these two robust observations are consistent, and notably the signature evidence is based on entirely independent data.
Close examination of the rapamycin associated gene network (Figure 2A) reveals ∼30 ribosomal RNA genes (a total of 70 genes were in the lean mass gain associated list but not all featured in the data-base of rapamycin regulated genes). The remaining genes (Figure 2A) belonged to metabolic processes and other facets of protein metabolism or signaling. To more easily appreciate the characteristics of those subjects that were found to have the greatest increase in lean tissue mass combined with a reduction in ribosomal gene abundance, we plotted the quartile response in lean mass (Figure 2B). Baseline lean mass could not explain our observation and in fact the four groups had the same mean age, physiological characteristics, while the highest and lowest quartiles for lean mass gain had exactly the same proportion of males and females (Table 1). Figure 2C specifically illustrates a subset of the ribosomal RNA genes, however all other rRNA genes were consistent with this plot. There was an almost universal pattern of down-regulation observed in high responders (shown in purple) while subjects that were unable to increase muscle mass (shown in blue) substantially up-regulated these genes, as would be expected from the qSAM statistical analysis (FDR<5%) in Figure 1A.
Thus, our analysis strategy enabled the discovery of an entirely novel in vivo feature of the mTOR growth pathway, while standard differential RNA expression analysis (pre vs. post sample) could not. We also plotted the relationship between physiological characteristics, protein-phosphorylation during acute net anabolic situations (resistance exercise coupled with feeding) with lean mass gain in these subjects using principal component analysis (PCA). Selected variables were scaled and principal components 1 and 2 are presented. In Figure 3A, it is abundantly clear that none of the metabolic or physiological characteristics shared variance with the main component capturing lean mass gain variation following RET. Likewise, while protein kinase abundance or protein kinase activation status varied in a manner consistent with the literature, none of these acute net anabolic responses were correlated to gains in RET induced lean mass or shared variance with gains in lean mass when studied prior to 20-wk RET (Figure 3B). In short, only the unbiased transcriptomics method was able to identify a biological profile distinguishing between high and low responders for lean mass gain.
Identification of the determinants of skeletal muscle mass has obvious implications for treating age-related sarcopenia. There is no longitudinal molecular analysis of ageing muscle in humans. However using cross-sectional gene-chip data-sets, attempts have been made to identify age-related gene expression changes. For example, Melov et al., reported that differences in gene expression between a cohort of young and old subjects can be removed through a RET program [23]. RET removed some aspect of the ‘inactivity’ related component of the difference between young and old subjects but when we contrasted their age-signature with other publically available muscle age datasets [23], [24], [26], [27] no overlap was apparent.
To investigate this issue further, we utilised the Melov et al., data and the data from Trappe and our lab. We used SAM analysis to compare young with old subjects in each study. For the Trappe study we had 13 young (20–30 y) and 11 old subjects (>80 y). We used baseline samples from DRET, 10 young (20–29 y) versus 16 old subjects (64–75 y), and finally we re-analysed the Melov et al., data [23] which had 26 younger (18–28 y) and 25 older subjects (65–85 y). As can be clearly observed in Figure 4 there was no overlap between the three studies indicating that a reproducible set of ‘ageing’ genes cannot be generated with this statistical or experimental approach. Our re-analysis of the GO analysis (using DAVID) of earlier studies [26], [39], using the appropriate back-ground files [40], also confirmed that mitochondrial RNA changes in ageing cannot be claimed as being a reproducible hallmark of ageing, despite the presumed association with inactivity. Re-analysis of the Melov et al., data did identify a mitochondrial gene expression signature (less significant than the original analysis due to comparison with a more appropriate ontological background) but in that study the older subjects were substantially de-trained and aerobic function was not presented. This gene-set is also known to reflect physical activity [29] and inactivity in humans [41] and thus it shouldn't be attributed to age per se anyway. To consolidate the conclusion that there was no common finding across these three studies, we entered the individual gene-lists in a gene ontology analysis to evaluate if some common pathways were enriched in each list, even if the member genes differed. Only 1 ontological group was common to 2 from the 3-gene lists and it related to mitochondrial processes indicating that even when the older subjects have a lower physical capacity a decline in mitochondrial genes is not always a prominent feature of age-related changes.
Therefore, an alternative approach to identify age-related gene expression profiles in human muscle was required. To achieve this we utilised QSAM, which we have previously applied and validated to some extent for human studies [3]. We applied QSAM to generate a list of muscle transcript levels, which correlated with subject age (20–75 yr), with correction for multiple testing. This allowed us to identify genes that either negatively or positively associated with the subject age. This has not been attempted before because previous studies did not have a sufficiently wide and continuous range of ages to generate such data [23], [24], [26], [39]. However, such an analysis would be of limited value if some of the observations could not be independently reproduced, using a distinct set of clinical samples. To this end, we generated 52 new U133+2 profiles (17–63 yr) from muscle biopsy samples from the HERITAGE Family Study [42] and were able to identify a set of 580 genes or transcripts (Dataset 3) which were correlated with age in both clinical studies and which was enriched in post-transcriptional and chronic disease traits (Figure S3) but not mitochondrial related gene-sets.
We found in IPA that the age-related dataset was consistent with the activation of the PGR (z-score = 2.6 and P-value = 0.001) and RXR (z-score = 2.0 and p-value = 0.0001) proteins and 5-fluorouracil agonism (Z-score = 2.2 and P-value = 0.0005, Figure 5A). Each mediator orchestrated a set of either positive (yellow) or negatively (green) age-correlated genes such that both overlap and direction were similar to the literature-constructed networks. Critically, these networks were not significantly related to the lean-mass associated gene-list (Figure 5B) or differentially regulated by either acute RE (Figure S4A) or chronic endurance exercise (Figure S4B). Thus it is unlikely that these new age-related observations reflect simple confounding factors, such as exercise training or being physically active.
There was also inhibition of certain protein mediators with age (Figure 6A) including c-MYC (z-score = −2.8 and p-value<0.0001) and CLDN7 (z-score = −2.6 and p-value = 0.05). Again, no clear relationship with acute exercise or endurance training was apparent (Figure S5), while a closer association with genes related to gains in lean mass was noted (Figure 6B) with some key exceptions. Notably inhibition of MYC was predicted in both the lean-mass and age-related gene lists (with gene-correlations in the same direction) which we would not expect if muscle ageing was simply the ‘opposite’ of muscle growth or lack of response to exercise training. Furthermore, large differences in gene expression still existed when comparing the age groups and the pre and post-training samples in the Trappe dataset (data not shown). The age-related expression signature was also related to RNA signatures in the Broad Connectivity database, including multiple serotonin antagonists and appears opposite to DNA topoisomerase inhibition (Dataset S4). Finally, we examined whether the age-related genes were over represented at genomic loci using Positional enrichment analysis [43]. Both Chromosome 1 (q12) and 13 (q21.33) had significant hits and the genes associated with those locations can be found in Figure 7A and 7B and the remaining analysis in Dataset S5.
We have carried out unbiased molecular analyses on both new and pre-existing human muscle data-sets (acute and chronic exercise, RET or EET and ageing) from which we have been able to define: (i) a ‘generic’ set of molecular networks that are activated irrespective of exercise-training mode in humans, i.e., RET or EET, (ii) the differential effects of acute exercise versus chronic exercise training on molecular networks, (iii) the molecular networks that are specifically modulated in relation to the extent of human muscle hypertrophy, and (iv) the first reproducible set of age-related transcriptional changes further supporting our view that large human sample sizes, using unbiased molecular profiling techniques [4], [29], [44] is an important strategy for translational medical science.
Muscle hypertrophy is the most recognized adaptation to RET. However, there are numerous other adaptations that occur, to support the biochemical, physical and metabolic requirements of a growing muscle. For example, hypertrophy is associated with activation of muscle satellite cells to support growth [45] while RET stimulates angiogenesis proportional to muscle fiber growth (rather than increasing capillary-to-muscle fiber area as EET [46]). Furthermore, along with accretion of contractile proteins, the extracellular matrix (ECM) undergoes substantial remodeling after RET [47]. Clearly then, successful hypertrophy is the summation of complex intra/extra-muscular cross talk to co-ordinate hypertrophy. As such, we believe that teasing out the molecular networks regulating adaptation in vivo requires charting the relationship between the modulation of molecular factors with that of the physiological outcome(s) [29].
Our initial analysis revealed that activation of a Tretionin (all-trans retinoic acid (ATRA)) and inactivation of aryl hydrocarbon receptor (AhR) are common molecular responses to training, irrespective of exercise mode i.e., RET or EET. ATRA is the active form of vitamin A, which serves as a ligand for two families of widely expressed nuclear receptors; the retinoic acid receptors (RAR) that bind to ATRA and the retinoid X receptors (RXR) that bind to its stereoisomer, alitretionin (9-cis-RA). Although little work exists on ATRA signalling in skeletal muscle, Halevy et al. showed exogenous ATRA promoted myogenic cell differentiation [48] which is allied to the function of ATRA amongst various cell types [49]. Given the post-mitotic properties of myonuclei, this may point to a novel link between ATRA-like signalling and aspects of in vivo satellite cell activities induced by exercise training. In support of this idea, we also observed IGF-1 and IGF-2 up-regulation in both RET and EET, and this is thought to play a role in satellite cell activation and differentiation [50], [51][52]. ATRA-like signalling has also been shown to modulate endothelial cell maturation and angiogenesis in tube formation assays [53] suggesting a role for activation of this network in exercise-induced angiogenesis, while angiogenic factors are also associated with satellite cell activation and differentiation [54]. Indeed, we identified up-regulation of homeobox (HOX) genes (e.g. HOXB3/7), which have roles in vascular remodelling and angiogenesis [55]. Thus this collection of genes is likely contributing to vascular remodelling-induced both by RET [56] and EET [57]. In addition, the turnover of ECM components is activated when smooth muscle cells are exposed to exogenous ATRA, thereby suggesting this pathway is involved in ECM remodelling [58]. Indeed, the (activated) ATRA gene list was highly enriched in collagen genes. Finally, while one cannot rule out that the degree of modulation of ECM (which differed between RET and EET) may influence aerobic adaptation [29], [31] it would appear to be a constituent feature of muscle growth and remodelling per se.
The AhR is a ligand-activated transcription factor known to mediate the negative effects of environmental xenobiotic contaminants such as dioxin (i.e., TCDD; 2,3,7,8-tetrachlorodibenzo-[p]-dioxin). This receptor belongs to the basic-helix-loop-helix (bHLH)/PAS (Period [Per]-AhR nuclear translocator [Arnt]-Single minded [Sim]) family of heterodimeric transcriptional regulators [59]. There have been a number of studies in which physiological clues have been gathered as to the function of AhR. For example, AhR has shown tumour suppressor effects i.e., when receptor levels are down-regulated by siRNA [60] the G1/S transition of the cell cycle and cell proliferation is promoted, suggesting a growth inhibitory role of the receptor. There also exists links between AhR and angiogenesis. Vascular endothelial growth factor (VEGF), a major growth factor that regulates angiogenesis is transcriptionally regulated by hypoxia-inducible factor-1 alpha (HIF-1α) in response to tissue hypoxia and muscle contraction [61]. Hypoxia stabilizes HIF-1α, which forms heterodimers with HIF-1β, also known as the AhR nuclear translocator (ARNT). ARNT can heterodimerize with AhR, minimizing the ability of ARNT to interact with stabilized HIF-1 to induce VEGF production. Finally, matrix remodelling is negatively affected during AhR activation. For example, AhR activation blocks regenerative processes during zebra fish caudal fin regeneration while impairing expression of genes involved in the structure and remodelling of the ECM [62]. Thus, inactivation of AhR-signalling may facilitate ECM and vascular remodelling occurring in response to both RET and EET (Figure 1).
The notion that both ATRA and AhR pathways are potent regulators of cell growth, differentiation, ECM remodelling, vascularization, organogenesis and embryogenesis [63] underlines their key roles in tissue development and homeostasis. Moreover, since adaptation to RET and EET both involve aspects of cellular remodelling, muscle satellite cell activities [64], ECM and vascular remodelling, we suggest that ATRA and AhR molecular programs are playing a previously undefined but central role in regulating “generic” features of exercise adaptation [22]. Finally, since ATRA and AhR gene networks that were regulated during long-term exercise training (Figure 1), were not reflective of those modulated in the hours after a single bout of exercise [24], [33], this casts doubt over ascribing formative purpose of acute exercise gene networks, which more likely represent stress pathways instigated by unfamiliar activities or simply the acute energy crisis in exercised muscle (in agreement with the lack of a striking ontology profile). This may explain why acute mRNA changes do not overlap with the chronic exercise patterns or, in our hands, relate to the networks that associate with the degree of gain in lean mass (see below).
Following the identification of what might be called generic ‘adaptability’ [22] molecular networks, we were interested to see if we could identify any molecular networks that were regulated in proportion to the degree of muscle hypertrophy in individual subjects. The justification for this approach was based on the marked heterogeneity in capacity for muscle growth in humans, with gains ranging from 0% to 22% [24], [28], [65]. In the DRET study, we found a similar range of changes in muscle size (−3% to +28%) and the analyses of the gain-related gene networks yielded striking results. We first discovered that there existed a correlation between capacity for human muscle lean mass gains and activity of both c-MYC and mTORc1 sensitive genes, including a large group of ribosomal RNA (rRNA) genes (∼70/560 total rRNAs). We further demonstrated that the nature of the association was not as one would be expecting from pre-clinical research, but rather there was a reduction in rRNA gene expression when greater muscle hypertrophy was observed (Figure 2B).
We speculate that human high-responders to hypertrophy potentially show superior efficiency in protein synthesis (i.e., protein yielded per RNA) and/or a reduction in proteolytic responses to RET. Nonetheless, it should also be noted that while rRNA expression negatively correlated to gain in lean mass, the highest responders tended to have higher levels of rRNA pre-training for individual rRNAs and hence the absolute abundance was similar post-training. This pattern of response clearly demonstrates that ‘more’ mTOR activation (e.g. more rRNA production) in humans is neither a hallmark nor a necessity for gains in lean mass in vivo. Our observations in humans also contrast with the molecular responses observed during acute synergist ablation induced hypertrophy [66], raising further doubts over the relevance of such pre-clinical models to inform about physiological muscle growth in humans. Regardless, our data demonstrate that high-responders for muscle hypertrophy evoke an “anti-growth” transcriptional response during a period of successful muscle growth with more studies being required in high or low lean-mass responders to investigate this phenomenon unambiguously.
While the unbiased transcriptional profiling yielded novel insights into human muscle growth responses, we also attempted to link training responsiveness to subject baseline physiology and the acute response of phospho-protein (AKT-mTORc1) signaling in response to an anabolic stimulus (RET combined with optimal nutrition [67]). We used correlation analysis to examine the relationship between key factors that have been speculated to be important for human muscle growth (e.g. body composition, fiber-type, metabolic and signaling molecule status). We found no significant shared variance between any of these variables and gains in lean mass, as presented in Figure 3 using PCA. We utilized PCA for visualizing the integration of physiological and molecular data because it enables an over-view in a single plot of how physiological and molecular aspects may vary within the major ‘units’ of variance of a particular dataset rather than plotting multiple individual scatter plots. By doing this, it becomes easier to visualize when paired relationships are behaving as expected, such as total protein and phospho-protein positions within each principal component. Unusual patterns could be used to identify problem areas such as when detection methods (e.g. antibodies) are poorly functioning.
As can be seen in Figure 3A, while principal component 1 was dominated by the variation in lean mass gains, basal lean mass, fat mass or fasting glucose did not vary. These data are in agreement with previous studies in which baseline physiological status was unable to select out high-responder status to hypertrophy [28]. We integrated the acute changes in phospho-protein signaling under the conditions of RET plus nutrition as an index of growth signaling potential in each subject. This seemed a valid approach as muscle hypertrophy is the product of nutrition and exercise-induced muscle protein synthesis [68]. While pre-training acute “anabolic” signals did not share variance with gains in muscle lean mass (Figure 3A), some weak relationships appeared between acute “anabolic” signaling elements following exposure to RET (Figure 3B) and gains in lean mass. Collectively, this underlines that while acute changes in phosphorylation may associate with those of acute remodeling processes (i.e., muscle protein synthesis responses) they are a poor indicator of future leans mass gains. Previously we have found a degree of dissociation between AKT-mTOR signals [69], [70] and muscle protein synthesis, while others have shown that acute synthesis does not, but some signaling molecules do, relate to future gains in lean mass [71]. We contend that using these individual signals as acute proxies for RET muscle growth is not going to be the most sensitive strategy.
There is a long established relationship between canonical pathways related to muscle growth and mechanisms that are associated with extended life-span in model organisms [72]. For example, inhibitors of mTORc1 and PI3K (a gene also known as “age-1”) activity can extend life-span in Caenorhabditis elegans, drosophila and mice [72]. As the number of people living beyond their eighth decade rises, it is expected that skeletal muscle atrophy and dysfunction (sarcopenia and dynapenia respectively) will become an increasing public health challenge [73]. Activation of mTORc1 and muscle IGF-1 signalling is associated with muscle cell growth experimentally while chronic inhibition of mTOR has been predicted to induce muscle frailty in humans [14], [74]. Likewise, reduced S6K1 (ribosomal protein S6 kinase) activity of the multifaceted regulator of cell growth would be thought to impair the retention of muscle mass in humans [75]. Thus there is a clear molecular basis to believe that processes under-pinning ageing, longevity, sarcopenia and muscle growth will be strongly inter-connected.
We found a list of ∼500 genes which track with age in human muscle across two independent cohorts, when our analyses utilised a continuum of ages. This contrasts with the irreproducibility of previous RNA versus muscle age datasets. Using this data we then evaluated which age-related gene networks link to muscle growth signalling or a variety of exercise scenarios [24], [29], [56], [76]. The motivation for this is that these factors cannot be independently controlled for in human studies and thus post hoc considerations are essential. The first clear observation was that the reproducible age-related gene-list and the lean-mass gain related gene-list both had inhibition of MYC as a key transcriptional feature immediately indicating that age and muscle growth are not exact opposites. Further, the key upstream regulators of the age-related gene list (e.g. PGR, RXR, Claudin7) contained a set of genes which were largely unrelated to the RET and EET transcriptomes, or the acute exercise responses [24], [76] for that matter. Claudin-7 activity was inhibited with age (Figure 6) and appears to relate to developmental differentiation and is strongly regulated by androgen signaling and HGF in vitro [77]. According to the IPA publication database the RXR ligand, CD437, induces a transcriptional signature that is consistent with age-related gene correlations, involving a network that is for 2/3rd negatively correlated with increasing age and for 1/3rd positively associated with age (i.e. with age KLF5 and IRS2 expression increases). Note that the ATRA signature, characteristic of increased physical activity in our hands, should operate through the RAR pathway. However, there is some overlap in the down-stream gene activation and thus both age-related changes and exercise activate some common features related to Vitamin A biology. In short, it is abundantly clear that the age-related changes in gene expression are not simply the ‘opposite’ [23] of the profiles seen with exercise or exercise training.
Positional enrichment analysis [43] was used to map the ∼500 age-related genes to chromosomal localisation, as attempts to link DNA variants with ageing have so far been only partly successful suggesting that an alternative approach may provide useful insights. We found several loci that yielded a significant enrichment score with ∼1q12 (e.g. PDE4DIP) and ∼13q21 (e.g. LMO7 (FBXO20)) yielding the most significant scores (Figure 7). PDE4DIP is a binding partner of phosphodiesterase 4D (PDE4D) which partners with Rheb to be a cAMP-specific negative regulator of mTORc1 [78]. When PDE4D binds Rheb it inhibits the ability of Rheb to activate mTORC1 and hence it is plausible that PDE4DIP (also called myomegalin) impacts on this relationship, as an increased interaction between Rheb and mTOR should promote growth. On chromosome 13, another growth and differentiation gene, Lmo7, was identified. Lmo7 impacts on myoblast differentiation, being required to induce Pax3 and MyoD expression [79]. Lmo7 is positively associated with age in muscle, while we have previously identified PAX3 as an up-stream transcriptional regulator of the EET induced transcriptome [29], highlighting further why the age-related transcriptome is not simply the opposite of exercise training. i.e., muscle ageing is not simply inactivity and thus is unlikely to be reversed by only activity (this does not at all contest that muscle function can be substantially retained by physical training in many but not all subjects).
Over the past 2 decades, therapeutic advances for complex chronic diseases have failed to generate all the progress predicted by the emergence of genetic technologies [80]. Part of the reason for this is that the investment in forward genetic pre-clinical models [81] has not yielded the expected validation of drug targets and it is now widely accepted that the clinical approach to chronic-disease management will have to reflect on numerous interactions between environmental and molecular factors [82]. Thus, an alternative approach, whereby identification of disease networks directly observed in clinical populations may have merit and lead to a more rapid translation of basic science [3], [4], [34], [40], [83]. Validation of such observations is dependent on access to sets of independent clinical data sufficiently large to have robust statistical power and diverse enough to be able to generalize the conclusions. It is safe to say that our approach is not at all universally favoured, as there is still a great reliance on so-called ‘validation’ studies involving forward and reverse genetic strategies in mice and cells. We suspect that such validation studies will, in the end prove to be context dependent and no easier to interpret than our in vivo molecular studies.
The singular advantage of our approach is that our data is generated under the precise conditions that would ultimately require therapeutic intervention. However, in the present analysis we focus on components of dynamic function, namely muscle mass, as the study of muscle performance will require determination of a wider range of muscle functional parameters (power, torque, velocity) and larger clinical studies, studies which do not exist at present. We also appreciate that our current observations would benefit greatly from follow-on genetic association studies in humans and by pharmacological or nutritional intervention studies. For example we have shown that a simple relationship between more mTOR activation and muscle growth does not exist in vivo in humans, albeit we are relying in part on the pharmacological activity of rapamycin to support this observation (off target effects may be present) and do not yet have kinetic data to understand the dynamic nature of this new relationship. We also failed to find a link between inter-subject variation in acute phospho-protein anabolic signalling and gains in lean mass. This may reflect the choice of time-point that we profiled the protein responses under. However, it must also be recognised that quantification of protein abundance changes suffers from numerous complications, including compression or exaggeration of dynamic range and challenges with specificity. Time will tell if our systems-biology translational medicine approach [84] exceeds the performance of traditional approaches taken to yield new therapeutic advances for human health.
This manuscript relies on several microarray studies to produce and independently validate molecular associations. We present a new 89-chip analysis (44 paired samples and 1 baseline-only sample) combined with direct comparisons to literature-based microarray data [23], [24]. The Trappe Lab chip data (GSE28422) was processed using identical methods to our new clinical data. Our QC analysis identified a number of chips that should be removed prior to statistical analysis. Inclusion of such data can both increase or decrease detection of differential gene expression (derived from technical and not biological variation). The analysis as a whole represents the most comprehensive examination of human molecular exercise-ageing pathways, and the application of independent chip resources provides a gold-standard level of validation.
Subjects were recruited from an age range of 18 to 75 y. Before beginning the study all subjects were screened using a medical questionnaire, physical examination and resting ECG with exclusions for overt muscle wasting (>2 SD below age norms) [85], metabolic, respiratory/cardiovascular disorders or other major contraindications to a healthy status. All subjects had normal blood chemistry and were normotensive (BP<140/90). All subjects performed routine activities of daily living and recreation but did not participate in moderate to high intensity aerobic exercise and none had participated in RET in the last 24 months. Body composition was measured at screening and following RET by dual energy X-ray absorptiometry (DEXA) (Lunar Prodigy II, GE Medical Systems). Subject positioning on the DEXA bed was optimized to allow the region of interest (ROI) body compartments to be analyzed separately. The upper leg ROI was selected as the area inferior to the lowest visible point of the coccyx to the mid-point of the patella. All subjects gave their written, informed consent to participate after all procedures and risks were explained. This study was approved by the University of Nottingham Medical School Ethics Committee and complied with the Declaration of Helsinki. The clinical data from these subjects were first reported in 2012 [56]. For the purposes of this article a total of 45 from the original 51 subjects were utilised as this represents the total number of gene-chip profiles that passed the appropriate quality control processes (N = 89 U133+2 Affymetrix chips).
The 20-wk fully supervised RET programme was designed to achieve skeletal muscle hypertrophy. Subjects trained three times per week, with each session lasting approximately 60 min. During four weeks of induction training (to ensure adoption and adherence to correct technique) intensity was increased from 40% to 60% 1-RM. For the remaining 16-wk of training intensity was set at 70% 1-RM with multiple sets of 12 repetitions, with two min rest between sets. 1-repetition maximum (1-RM) assessments were made every four weeks to ensure that the intensity of loading was constant. Subjects were excluded from the study for non-compliance, defined as: non-attendance for >6 consecutive sessions, less than 75% attendance, or failure to complete the set exercise regime on >15% attendance. Muscle biopsies (∼150 mg) were taken under fasted-non-exercised (“basal”) and optimal growth conditions (acute exercise-fed conditions, 2.5 h after a single bout of exercise) both before and after chronic-RET from the vastus lateralis muscle under local anaesthesia (2% lidocaine, with the use of a conchotome biopsy forceps, as previously described [56]). Blood was collected in pre-chilled tubes containing Lithium Heparin, plasma was separated by centrifugation and was then stored at −80°C until analyses. Plasma glucose concentration was measured on a clinical chemistry glucose analyser (ILAB 300 Plus). For insulin, blood was collected in pre-chilled tubes containing EDTA, plasma was separated by centrifugation within 30 min of collection and was then stored at −80°C until final analyses. Plasma insulin concentration was determined using high sensitivity insulin ELISA systems (Immunodiagnostic systems limited).
Muscle biopsies (∼20 mg) were homogenized in ice-cold extraction buffer (10 µL.mg−1) containing 50 mM Tris-HCl (pH 7.4), 0.1% Triton X-100, 1 mM EDTA, 1 mM EGTA, 50 mM NaF, 0.5 mM activated sodium orthovanadate (Sigma Aldrich, Poole, UK) and a complete protease inhibitor cocktail tablet (Roche, West Sussex, UK). Homogenates were centrifuged at 10,000 g for 10 min at 4°C. Bradford assays were used to determine supernatant protein concentrations after which samples were standardized to 1 mg.mL−1 in Laemmli loading buffer. Samples were heated at 95°C for 5 min before 15 µg of protein/lane was loaded on to Criterion XT Bis-Tris 12% SDS-PAGE gels (Bio-Rad, Hemel Hempstead, UK) for electrophoresis at 200 V for ∼60 min. Gels were equilibrated in transfer buffer (25 mM Tris, 192 mM glycine, 10% methanol) for 30 min before proteins were electro-blotted on to 0.2 µm PVDF membranes (Bio-Rad) at 100 V for 30 min. After blocking with 5% low-fat milk in TBS-T (Tris-Buffered Saline and 0.1% Tween-20; both Sigma-Aldrich, Poole, UK) for 1 h, membranes were rotated overnight with primary antibody (all AbCam, Cambridge, UK) against our target proteins (AKT, mTOR, p70S6K1, 4EBP1, eEF2) at a concentration of 1∶2000 at 4°C. Membranes were washed (3×5 min) with TBS-T and incubated for 1 h at room temperature with HRP-conjugated anti-rabbit secondary antibody (New England Biolabs, UK), before further washing (3×5 min) with TBS-T and incubation for 5 min with ECL (Immunstar; Bio-Rad). Blots were imaged and quantified by assessing peak density, after ensuring bands were within the linear range of detection using the Chemidoc XRS system (Bio-Rad, Hemel Hempstead, UK). Protein phosphorylation was corrected for loading to actin loading control before the protein signals were subject to PCA to explore relationships between ‘anabolic signals’, specifically in terms of muscle hypertrophy responsiveness.
Total RNA was isolated from muscle biopsies taken before and after (72 h after the final training session) RET by chloroform-phenol based extraction. In brief, paired tissue samples (obtained before and after RET) of ∼20 mg each were processed simultaneously in 1 mL TRIzol (Invitrogen) using a Mini-Beadbeater-8 (Biospec Inc.) for 15 sec on the “homogenize” setting. After 5 min of incubation at room temperature, 200 µL of chloroform (Sigma-Aldrich) was added and samples shaken vigorously by hand. Samples were briefly incubated on ice prior to centrifugation at 12,000 g for 15 min. The supernatant was removed and mixed with isopropanol (Sigma-Aldrich) and spun once more at 12,000 g for 10 min, after 10 min of incubation. After a single washing step with 75% EtOH RNA pellets were re-suspended in 40 µL DEPC-treated water (Ambion) and quantified using a NanoDrop Spectrophotometer (NanoDrop Technologies). RNA purity was assessed using the A260/A280, A260/A230 ratios and stored at −80°C. Samples were put through this process in pairs (pre-post samples) while the order of subject processing was carried out to distribute ‘non-responders’ equally.
Reverse transcription of RNA was carried out using the Affymetrix 3′ IVT express kit. 100 ng of total RNA was reverse transcribed as per manufacturer's protocol, and quantified using a Nanodrop ND-1000 instrument. aRNA was fragmented and labeled as per manufacturers protocol and hybridized to Affymetrix U133+2 arrays (Affymetrix, USA). Arrays were washed, stained and scanned following Affymetrix standard procedures, using an Affymetrix 3000 7G scanner and Affymetrix 450 wash station. A visual inspection of each array was carried out.
Low-level processing of all arrays was undertaken using Bioconductor in R. The Affy package was used to carry out MAS5 based normalization and generate present, marginal and absent (PA) scores. NUSE plots were generated and combined with PCA, outlier samples were identified where both the NUSE plot and PCA was supportive of its exclusion (∼2% of arrays). For baseline correlation analysis, all samples that passed QC were utilized (N = 45). This procedure was applied to the data-set originating from the Trappe laboratory (GSE28422, [24]) and outliers removed from the dataset that failed the QC process, leaving n = 96 for analysis. Pre-exercise training muscle biopsy samples from the HERITAGE family study (N = 50) were also analyzed to yield a second independent data set with a continuous span of age-ranges (see below). The Trappe and HERITAGE datasets therefore represent independent datasets which we utilized, where possible, to validate the pathway analysis of our study. Such confirmation benchmarks results using thousands of data-points and is more desirable that targeted real-time qPCR confirmation (where the gene selection is biased and the sample size inappropriate to make statistical conclusions). Annotation of all CEL files used ‘hgu133plus2cdf_2.9.1.tgz’ while annotation of probe-set lists was then updated using the Ingenuity Pathway Analysis database, as of August 2012.
Our first objective was to identify the gene-networks regulated in proportion to gains in upper leg muscle mass (hypertrophy), the same location as our biopsy sample. Such analysis relies on the established principal that adaptation responses (for the majority of phenotypes) to exercise training in outbred populations is highly variable, typically reflecting genetic and epigenetic variation and in genomic variation. We utilized quantitative SAM analysis [3], [36] to generate a list of genes which vary in a positive and negative manner with changes in DEXA assessed upper leg lean mass. This was applied to PA filtered data and the statistical parameter generated is a q-value (false discovery rate). This provided for the first time a candidate list of gene-changes that may exhibit primary or secondary influence over muscle growth in humans. The gene-list was then subject to IPA based pathway analysis and in particular the Upstream Analysis tool in IPA was utilized. This analysis has similarities to the Molecular Connectivity Database [86] where pre-existing collections of RNA signatures are compared with our lean-mass related gene list, and significant overlaps identified. An overlap P-value is generated based on the degree of overlap between the gene-set within the IPA database (which reflects the RNA molecules changed in response to a ‘mediator’ such as a transcription factor or a drug) and our data set, adjusting for data set sizes using the Fischer's Exact Test. We accepted a stringent P-value of p<0.001 as being significant. A second parameter is the activation “z-score” where the directional change in RNA is compared between the IPA mediator data-set and our lean-mass gain gene list. The z-score informs on whether the drug/protein mediator is likely to be ‘active’ or ‘inhibited’ during gains in lean mass. Thus, if we discovered that an antagonist is ‘inhibited’ in our analysis, this indicates that the drug target is activated. However, in the present study the data-input refers to genes, which positively or negatively correlate with lean mass gains e.g. if we find a “Statin” signature was inhibited, it is interpreted that HMG-CoA reductase regulated genes are negatively correlated with lean mass gain.
The two-step process presented above generates a focused gene-list with a high statistical rigor for true positive associations. This type of analysis also utilizes the full range of physiological response observed, however it assumes that expression of important genes will relate in a linear manner to lean mass gain and thus can not discover all appropriate associations. We then contextualize the statistical findings both in terms of subject characteristics and through comparison of the response of these significant networks with independent gene-array data (e.g. [24]). At this stage we utilized descriptive statistics, plotting the significant network genes as simple expression values relative to the quartile distribution of lean-mass gains to allow for clear discussion of the results. As these plots are based on the z-scores and P-values as above, no further statistical analysis is presented.
Following identification of our primary objectives we then carried out a classic differential expression analysis using SAM. Given that we have established that chronic differential expression patterns, following exercise training, are dependent on the presence of physiological adaptation we removed 6 subjects that demonstrated no gain in lean mass. This yielded a list of differentially expressed genes that could then be compared with the RET gene-list generated from the Trappe laboratory data and our published exercise studies [24]. Secondary analysis, where subject age or baseline lean-mass was related to baseline gene-expression was carried out using quantitative SAM analysis as described above [3], [36]. This allowed us to present comparisons of the RET gene-list with other modes of exercise, such as endurance exercise training or disease [3], [29] and age-related analysis [23], [24]. PCA was utilized to visualize the association between selected physiological and protein expression parameters and training induced changes in muscle lean mass. PCA was implemented in R, using prcomp() command, which calculated a singular value decomposition and plots the selected principal components using the plot command in R. All data was individually transformed to a median value within that data set so that all variables were within a consistent data range. In each case the majority (∼65%) of the total variance was captured by the first two principal components.
Finally, positional gene enrichment analysis (PGE) was used to identify whether the classification genes (or the classifier network genes) were significantly enriched within given chromosomal regions [43]. This analysis is based on the following rules: Rule 1: it contains at least two genes of interest, Rule 2: there is no smaller region containing the same genes of interest, Rule 3: there is no bigger region with more genes of interest and the same genes not of interest, Rule 4: there is no larger encompassing region with a higher percentage of genes of interest, Rule 5: there is no smaller encompassed region with a better P-value, Rule 6: it does not contain any region having less than expected genes of interest. The approach of PGE exhaustively evaluates the over-representation at all chromosomal resolution levels simultaneously.
|
10.1371/journal.pntd.0003688 | Onchocerciasis Transmission in Ghana: Persistence under Different Control Strategies and the Role of the Simuliid Vectors | The World Health Organization (WHO) aims at eliminating onchocerciasis by 2020 in selected African countries. Current control focuses on community-directed treatment with ivermectin (CDTI). In Ghana, persistent transmission has been reported despite long-term control. We present spatial and temporal patterns of onchocerciasis transmission in relation to ivermectin treatment history.
Host-seeking and ovipositing blackflies were collected from seven villages in four regions of Ghana with 3–24 years of CDTI at the time of sampling. A total of 16,443 flies was analysed for infection; 5,812 (35.3%) were dissected for parity (26.9% parous). Heads and thoraces of 12,196 flies were dissected for Onchocerca spp. and DNA from 11,122 abdomens was amplified using Onchocerca primers. A total of 463 larvae (0.03 larvae/fly) from 97 (0.6%) infected and 62 (0.4%) infective flies was recorded; 258 abdomens (2.3%) were positive for Onchocerca DNA. Infections (all were O. volvulus) were more likely to be detected in ovipositing flies. Transmission occurred, mostly in the wet season, at Gyankobaa and Bosomase, with transmission potentials of, respectively, 86 and 422 L3/person/month after 3 and 6 years of CDTI. The numbers of L3/1,000 parous flies at these villages were over 100 times the WHO threshold of one L3/1,000 for transmission control. Vector species influenced transmission parameters. At Asubende, the number of L3/1,000 ovipositing flies (1.4, 95% CI = 0–4) also just exceeded the threshold despite extensive vector control and 24 years of ivermectin distribution, but there were no infective larvae in host-seeking flies.
Despite repeated ivermectin treatment, evidence of O. volvulus transmission was documented in all seven villages and above the WHO threshold in two. Vector species influences transmission through biting and parous rates and vector competence, and should be included in transmission models. Oviposition traps could augment vector collector methods for monitoring and surveillance.
| The World Health Organization (WHO) aims at eliminating onchocerciasis by 2020 in selected African countries. The success of elimination using ivermectin treatment alone will depend on several interacting factors including baseline endemicity, treatment coverage and vector species mix. In Ghana, transmission persists despite prolonged control. We investigated entomological determinants of this persistence. Blackflies were collected from seven villages with 3–24 years of ivermectin treatment. A total of 12,196 flies was dissected, with 463 larvae (all Onchocerca volvulus) in 97 infected and 62 infective flies. Transmission indices in the wet season, at Gyankobaa and Bosomase, amounted to, respectively, 86 and 422 infective larvae/person/month after 3 and 6 years of ivermectin treatment. Infection levels at these villages were over 100 times the WHO threshold of one L3/1,000 parous flies. At Asubende, an infective fly was caught among ovipositing flies in nearby breeding sites, indicating that infection was just over the WHO threshold despite extensive ivermectin and vector control. Spatial and seasonal vector species composition influences the magnitude of transmission indices through variations in biting and parous rates, and vectorial competence and capacity, and should be reflected in transmission models. Oviposition traps could enhance vector collection for transmission monitoring and surveillance.
| The London Declaration on Neglected Tropical Diseases (NTDs) [1] and the World Health Organization’s (WHO) road map to accelerate progress for overcoming the impact of NTDs [2] have set goals for the elimination of human onchocerciasis by 2020 in selected African countries. Based on the results of epidemiological studies conducted in some foci of Mali, Senegal and Nigeria [3,4,5], it has been suggested that 14–17 years of annual (or biannual) ivermectin treatment may lead to local elimination of the infection reservoir in the absence of vector control. The repeatability of these achievements depends, in part, on the initial level of onchocerciasis endemicity, geographical and therapeutic coverage, treatment compliance and frequency, parasite susceptibility to ivermectin, and the intensity and seasonality of transmission, including the species composition of the simuliid vectors [6].
Previous reports assessing the feasibility of onchocerciasis elimination have concluded that although ivermectin mass drug administration (MDA) alone would help to eliminate the public health burden of onchocerciasis, it would not lead to elimination of infection in most foci, with the possible exception of areas of low endemicity [7]. However, more recent and encouraging results in areas of moderate to higher endemicity [3,4,5], have spurred the African Programme for Onchocerciasis Control (APOC) to shift its goals from morbidity control to local elimination of Onchocerca volvulus where possible [8]. Recognising the need to understand the nature and extent of transmission zones, APOC and WHO have emphasized the importance of conducting entomological studies on the determinants and feasibility of elimination [8,9,10]. Current WHO guidelines state that parasite levels within the vector need to be below a threshold of one L3 larva per 1,000 parous flies [11]. However, understanding how this measurement relates to the rate of transmission assessed via the biting rate, the infectious biting rate, the parous rate and the transmission potential, and importantly, how it varies with vector species composition and season, is vital for accurate monitoring and interpretation of this threshold [8].
Ghana was originally a country under the umbrella of the Onchocerciasis Control Programme in West Africa (OCP), which operated between 1974 and 2002, and was initially a vector control programme [12,13]. Vector control activities started in 1975 in the onchocerciasis savannah foci of northern and central Ghana, but the southern forest foci were not part of the programme [9]. When the microfilaricidal drug ivermectin was licensed for human use in 1987 [14,15], Ghana was one of the first countries to commence MDA. In particular, community trials were conducted in the then highly hyperendemic focus of Asubende (initial microfilarial prevalence of 80%) [16], where vector control had taken place but was suspended during the ivermectin distribution pilot study in the late 1980s. When the OCP ceased operations in 2002, the persistence of onchocerciasis at Asubende required this focus to be part of the so-called Special Intervention Zones, which maintained extensive coverage with ivermectin leading to dramatic reductions in infection intensity and prevalence [17].
In 2007, Osei-Atweneboana and co-workers [18] reported on the epidemiological situation in Ghana after the closure of the OCP and observed that despite vector control, and 19 years of annual ivermectin treatment, some communities exhibited high microfilarial prevalence and intensity (measured as the community microfilarial load) [19]. This was subsequently attributed to adult female worms being less responsive to the anti-fecundity effects of multiple treatments with ivermectin in some communities [20], but others pointed out the possibility of programmatic causes such as poor coverage permitting significant residual transmission [21,22,23]. Concerned by these findings, the NTD Programme of the Ghana Health Service initiated biannual ivermectin distribution in some communities in 2009 [6,24]. From 2003, ivermectin distribution was also extended to include endemic areas in Ghana which had not previously been included in the OCP.
Motivated by the need to understand the feasibility of elimination in Ghana, and in particular the entomological determinants of transmission persistence despite prolonged control, we conducted a study on the transmission of onchocerciasis in areas both within and outside the original OCP area. We have already reported on the spatial and temporal distribution of species within the Simulium damnosum complex found at breeding sites in southern Ghana from 1971 to 2011 [25], and on the biting and parous rates of host-seeking females [26]. In this paper, we present the spatial and temporal patterns of infection with Onchocerca spp. larvae of host-seeking and ovipositing flies in communities that have experienced different durations (and frequency) of ivermectin treatment. We relate our findings to the therapeutic coverage recorded in each study village and discuss the potential of fly trapping techniques, not based on the traditional OCP vector collector method, for the monitoring of transmission prior to or after the initiation of post-MDA surveillance.
Ethical clearance was obtained from the Imperial College Research Ethics Committee (ICREC_9_1_7) and the Institutional Review Board of the Noguchi Memorial Institute for Medical Research, University of Ghana (IRB:0001276, 006/08-09). No tissue samples were taken from human subjects; however, local villagers and elders assisted with blackfly collections. Signed informed consent was obtained from all individuals involved after detailed explanations in their local languages about the study. Participating individuals were not at an increased risk of exposure, nor were human samples obtained for diagnosis, therefore, no treatments were offered. However, all participants were receiving ivermectin as part of the national programme following appropriate (annual or biannual) schedules according to the Ghana Health Service strategy [24].
Site selection, geography and key simuliid species are described elsewhere [26], but, in brief, blackfly collection was conducted in seven villages within four regions of Ghana: Asubende (08°01'01.4"N, 00°58'53.8"W) and Agborlekame (08°14'04.0"N, 2°12'23.2"W) in the Brong-Ahafo Region; Asukawkaw Ferry (07°40'55.9"N, 00°22'16.0"E), Dodi Papase (07°43'22.5"N, 00°30'38.3"E) and Pillar 83 (07°42'20.3"N, 00°35'21.5"E) in the Volta Region (Pillar 83 is the village on the Ghanaian side of the river Wawa, which forms the border and is known as the Gban-Houa in Togo, opposite the former OCP catching site of Djodji in Togo); Bosomase (05°10'44.7"N, 01°36'23.1"W) in the Western Region and Gyankobaa (06°20'12.4"N, 01°16'11.3"W) in the Ashanti Region (Fig 1). A pilot study was conducted at Bosomase in January–February 2006 to assess the efficacy of Bellec traps (see below) as a fly collection method, and to test the performance of DNA amplification methods for the determination of blackfly species, infection status and blood meal origin. The main sample collection took place during one wet season, 23rd July–5th September 2009, and two dry seasons, 14th February–28th March 2010 and 30th January–5th March 2011. Villages were visited and samples were collected for up to five consecutive days per site per trip. Not all sites were successfully sampled during each period due to weather conditions and variability in blackfly population abundance.
In Ghana, six main species are known to contribute to the transmission of O. volvulus. These are S. damnosum sensu stricto (s.s.) Vajime and Dunbar; S. sirbanum Vajime and Dunbar; S. sanctipauli Vajime and Dunbar; S. yahense Vajime and Dunbar; the Beffa form of S. soubrense Vajime and Dunbar [32] and S. squamosum (Enderlein) (of which both C and E forms occur) [25]. Morphological identifications, parity status and molecular fly identifications have been described in detail previously [26] and were carried out using standard methods [32,33,34,35,36,37,38,39,40,41]. The colour of the fore-coxae used by some authors [33,34] to separate S. damnosum s.s. from S. sirbanum is unreliable since many individuals of both species with either dark or pale fore-coxae have been noted, especially in the eastern parts of the former OCP, and therefore these two species were not split by definitive identification and are termed S. damnosum s.s. /S. sirbanum. Morphological identifications and parity status of the host-seeking blackflies were performed the day after being caught. Parous females’ abdomens were separated from the head and thorax, which were preserved individually in corresponding wells of two 96-well PCR plates (one for heads plus thoraces, one for abdomens) in absolute ethanol for subsequent molecular analysis. When catch numbers were manageable (up to 300 flies per day), all host-seeking flies were first dissected for parity in the field. When parity of some blackflies was not assessed due to high catch numbers and time constraints (>300 per day), all remaining host-seeking flies were only morphologically identified and their heads and thoraces separated from their abdomens and stored as above. Simulium squamosum shares many morphological traits with other sympatric species, causing difficulties when identifying some adult blackflies [33]. Therefore, DNA from all abdomens was extracted and used for definitive molecular identification of S. squamosum and for Onchocerca spp. infections as described below. Flies caught in Bellec and Monk’s Wood traps were morphologically identified using the same techniques [35,36,37,38,39,40,41], and the heads, thoraces and abdomens separated and stored individually as for the host-seeking flies [26].
The heads and thoraces of all the known parous and unknown parous (physiological age not determined) host-seeking blackflies were dissected for Onchocerca infection. Flies caught in Bellec and Monk’s Wood traps were in the process of ovipositing and hence were not dissected for parity, as their gravid status made parity assessment impossible without counting their ova [42]. Although the flies coming to lay eggs in breeding sites would comprise both nulliparous (laying eggs for the first time) and parous flies (having laid eggs before), it was assumed that they would have all taken at least one blood meal (as S. damnosum s.l. is obligatorily anautogenous [43]) and, therefore, capable of ingesting Onchocerca microfilariae if feeding on infected hosts 2–3 days previously. By the time of oviposition, some of these microfilariae could have migrated out of the abdomen and established in the thorax as L1 larvae. In parous flies, infections picked up 2–3 gonotrophic cycles earlier, could have developed into pre-infective (L2) in the thorax, or infective (L3) larvae, found in heads or thoraces. Therefore, the heads and thoraces of all ovipositing flies were dissected for infection with Onchocerca larvae. Heads and thoraces were soaked in distilled water for one hour, stained with a solution of 7% lactopropionic orcein in distilled water for a further hour [44], and examined in a drop of the staining solution under a dissecting microscope. The numbers, developmental stage (L1, L2, L3), and location within the fly (head or thorax) of any Onchocerca spp. larvae were recorded. Larvae were transferred to steel-frame 0.9μm POL-membrane slides (Microdissect, Leica, Germany) [45] for subsequent individual DNA-based identification of parasite species (such as O. volvulus, O. ochengi, O. ramachandrini, O. dukei, O. denkei and the Siisa-clade of O. ochengi) [46,47,48]. In the field, during the morphological identification and parity dissection, any Onchocerca larvae which emerged were also recorded and transferred to a POL-membrane slide.
Since S. damnosum s.l. is also involved in the transmission of other Onchocerca species [46,49], parasite larvae were identified by molecular methods to ensure that transmission of human onchocerciasis would be accurately recorded. POL-membrane slides with the Onchocerca spp. L1, L2 and/or L3 were placed on a Leica LMD6000 laser dissection microscope, viewed on a computer screen, and any Onchocerca larvae were cut out individually using an ultraviolet laser, with the sample falling into a PCR tube cap below [45]. Larvae were stored in 15μl Qiagen ATL buffer and frozen until DNA extraction. DNA extraction was performed using the QIAamp DNA Micro kit (QIAGEN) following the ‘isolation of genomic DNA from laser-microdissected tissues’ protocol, with DNA eluted into 30μl sterile distilled water. DNA was amplified using general Onchocerca (primer O-150) [47,50] and the O. volvulus specific (C1A1-2) [47] primers and the results run on agarose gels for species identification through presence or absence of the O. volvulus specific amplicon, when the Onchocerca general PCR had been successful. In addition, PCR amplifications were performed using three further pairs of primers 12SOvB and C, 16SOvB and C, and ND5OvA and C amplifying 12S rRNA, 16S rRNA, and ND5 mitochondrial genes respectively [51,52]. PCR clean-up, quantification and sequencing was performed on these 12S, 16S, and ND5 amplicons. Sequences were then individually run through BLAST and Onchocerca species identification scored when successful matches occurred. Sequences were also compared to known sequences of Onchocerca on ClustalW for additional clarification of any species identification. PCR plates contained negative water controls, O. ochengi (adult worm DNA) positive controls, and O. volvulus (microfilarial DNA) positive controls. Presence of Onchocerca (most likely microfilariae or infective larvae) in the abdomens was detected using the same 16S protocol [51] mentioned above for dissected Onchocerca larvae; any positive amplicons were then also sequenced and run through BLAST and ClustalW.
The study communities currently receive community-directed treatment with ivermectin (CDTI) but with varying treatment histories in terms of number of years of MDA and treatment frequency, as well as having experienced a range of historical vector control activities, summarised in Table 1. Community drug distributors were interviewed regarding recent drug administrations in each village, as well as village, regional and national treatment records checked for historical treatments. Dates of historical vector control are indicated in Fig 1 and previously discussed in [26]. Data on yearly therapeutic coverage of ivermectin for each study village for annual or biannual treatment rounds were provided by the Ghana Health Service.
Except where specified as PCR results on the blackfly abdomens, all data presented are from dissections of heads and thoraces only. Data are reported as per fly, per parous fly, per infected fly or infective fly throughout. The proportion infected is taken as the number of flies of each species with any larval stage (L1, L2 or L3) divided by the total number of flies of that species dissected and are presented with 95% exact confidence intervals (95% CI), determined using the method of Clopper-Pearson [53].
Because Onchocerca L3s can migrate from other parts of the body to the head during a blood meal, a fly with L3s in any body part is counted as infective [54,55,56]. (Infective larvae develop in the fly’s thoracic muscles and typically migrate to the head capsule and the fly’s proboscis, but they have also been detected in the halteres and abdomen.) Therefore, the proportion infective is the number of flies of each species with L3 larvae (in head and/or thorax) divided by the total number of flies of that species dissected and is presented with 95% CIs. In addition we also present the number of flies with L3s in the head only for comparison with published literature.
We calculated monthly infective biting rates, which take into account the number of infective flies that (come to) bite a host per month, but not their parasite burden. These were calculated by multiplying the proportion of infective flies (with L3 larvae in head and/or thorax) by the monthly biting (landing) rates as reported elsewhere [26], but summarised in S1 Table. Monthly parous biting rates, the monthly rate at which a host would be bitten by parous flies, have been presented and analysed by species elsewhere [26].
We calculated the arithmetic mean number of L3 per infective fly per species (L3s/infective fly) as the total number of L3 larvae divided by the number of flies which contained any L3 larvae. The monthly transmission potential is the mean number of L3 larvae to which a host is exposed per month. These were calculated by multiplying monthly infective biting rates by the number of L3s/infective fly. We report transmission potentials for given months in the wet and dry seasons, but as we did not collect data throughout the whole year we do not extrapolate these results to annual transmission potentials. As fly survival rates have been shown to affect variations in transmission rates [57,58], we also present the number of L3 larvae per 1,000 parous flies as recommended by the WHO [11]. These values are reported, separately, for parous host-seeking flies and ovipositing flies for each location and season. The mean number of L3s/1,000 parous (or ovipositing) flies was calculated as the total number of L3 larvae divided by the total number of parous (or ovipositing) flies dissected for Onchocerca multiplied by 1,000. We did not assume that the same parity rates determined in samples of host-seeking flies would apply to the ovipositing flies caught near (by light traps) or in breeding sites (by Bellec traps) because a phenomenon of differential dispersal of nulliparous and parous flies along rivers and inland from rivers has been documented in S. damnosum s.l., which varies between the savannah and forest members of the species complex [59].
The transmission indices described above were calculated from flies captured by vector collectors (and therefore relate to human exposure and the potential of transmission from flies to humans) unless stated otherwise. Host-seeking infective flies collected in the cow tents—had they been able to bite cattle and shed their entire L3 larval load—would not have contributed effectively to the transmission of O. volvulus. However, these flies indicate occurrence of active transmission from humans to flies, as they have become infected and survived the incubation period of the parasite. Therefore, these transmission parameters are presented for each host-seeking catching technique. Also, our results indicate that flies that bite cattle may also bite humans (blood meal results to be presented elsewhere) and so, if able to survive further gonotrophic cycles, infected and infective flies attracted to cattle could subsequently feed on humans and transmit their remaining infective larval load as, on average, only 50 to 80% of L3 larvae are shed per bite [55,60]. The proportion infected, proportion infective, the mean number of L3s/infective fly and the number of L3s/1,000 (parous or ovipositing) flies are reported, separately, for host-seeking and ovipositing flies.
Statistical analyses were performed on SPSS version 22 (SPSS, Inc., Chicago, IL, USA) or R [61]. Numbers of infected and infective flies, for all catches, and per species, were compared among villages, seasons and trapping methods using chi-squared (χ2) tests. Ninety five percent CIs for the number of L3/1,000 parous, L3 per 1,000 ovipositing and L3 per infective flies were determined using a percentile bootstrap method [62]. A correlation between the number of years since the start of ivermectin treatment and the proportion of infected and infective flies was tested using Spearman’s Rank correlation coefficient (rS). Variation in infection intensities among different species was compared using Kruskal Wallis and Mann–Whitney U tests. Numbers of infected versus uninfected flies as measured by PCR of the abdomens were compared between catching techniques using the chi-squared (χ2) test. Therapeutic coverage of ivermectin distribution was plotted against time since each village commenced treatment, with a best fit polynomial plotted for each village.
A total of 17,300 S. damnosum s.l. flies was collected, of which 6,142 (35.5%) were caught by vector collectors; 2,207 (12.8%) were trapped in the human-baited tents; 1,567 (9.1%) in the cow-baited tents; 7,212 (41.7%) on Bellec traps—including 3,352 (46.5% of the Bellec total) from the pilot study in Bosomase during the dry season in 2006—and 172 (1%) in Monk’s Wood light traps. A total of 16,478 (95.2%) blackflies was morphologically identified, of which 5,812 (35.3%) were dissected for parity in the field, with 4,247 (73.1%) nullipars and 1,565 (26.9%) parous flies. These nullipars were not further dissected for Onchocerca infection, but pooled samples of the nullipars were used as molecular controls, with no positive Onchocerca results obtained. The heads and thoraces of 12,196 flies (6,918 ovipositing flies, 3,713 host-seeking flies of unknown parity status and 1,565 known parous flies) were stained and dissected for Onchocerca spp. larvae. These, plus the known uninfected nullipars (4,247), totalled 16,443 flies whose infection status was assessed. A total of 97 (0.6%) was infected (with any larval stage) of which 58 (0.4%) were infective (with L3s in head and/or thorax), with 45 flies (0.3%) harbouring L3s in the head (S2 Table).
DNA was extracted and amplified from all 463 larvae of all stages, from the 97 infected flies (on average, 4.8 larvae per infected fly and 0.03 per fly). The PCR product using the ND5 primers was consistently of poor quality and therefore only the 12S and 16S amplicons [51] were used for Onchocerca spp. identification with BLAST and ClustalW. Of all individual L1 to L3 larvae, 76% (352/463) were positively identified as O. volvulus using either 12SOv, 16SOv primers and/or O. volvulus specific (O-150 versus C1A1-2) amplicons in the agarose gels. The remaining 24% were not successfully amplified. No O. ochengi was observed in the field-caught flies, but the positive O. ochengi controls were successfully identified by BLAST and/or ClustalW and did not have O. volvulus specific amplicons in the agarose gels. There were no ambiguous results for the species identification. Of the 111 non-identifiable larvae, 107 (96%) came from flies in which other larvae of the same stage had been successfully identified as O. volvulus.
Blackflies infected with O. volvulus larvae were recorded at Asubende, Asukawkaw Ferry, Bosomase and Gyankobaa, and infective flies (with L3s in head and/or thorax) were recorded at Asubende, Bosomase and Gyankobaa (Tables 1 and S2). No infected or infective flies were observed at Agborlekame, Dodi Papase or Pillar 83 during our study from the heads and thoraces; however, O. volvulus DNA was amplified in flies from all seven villages from the abdomens (see below).
There was no statistically significant difference in the proportion of infected (χ2 = 5.06, d.f. = 3, p = 0.168) and infective (χ2 = 2.79, d.f. = 3, p = 0.425) flies caught at Gyankobaa or Bosomase by the different trapping methods. A higher but, not statistically significant, proportion of infected and infective parous flies were caught in the cow-baited tents (infected = 2.54%, infective = 1.34%) than the other trapping methods (Fig 3), with infected and infective levels of 1.53% and 0.77% in the human-baited tents, 2.30% and 0.73% by the vector collectors and 1.06% and 0.97% in the oviposition traps, respectively. Twenty seven percent of the infected flies were caught in the cow-baited tents, 19% in the human-baited tents, 34% in the vector collector caught flies and 20% by the oviposition traps.
There was no statistically significant difference between the proportion of infected (χ2 = 0.90, d.f. = 1, p = 0.353) or infective (χ2 = 2.09, d.f. = 1, p = 0.148) flies caught by the oviposition and host-seeking methods combined, nor between the two most successful catching techniques, namely the Bellec traps and the vector collector method (infected: χ2 = 3.08, d.f. = 1, p = 0.079; infective: χ2 = 0.30, d.f. = 1, p = 0.584). In contrast, in the abdomens, statistically significantly more flies had O. volvulus infections, as recorded by PCR, in the ovipositing flies than in the host-seeking flies (χ2 = 19.58, d.f. = 1, p<0.001), as well as in just the Bellec-caught flies in comparison with the vector collector-caught flies (χ2 = 8.51, d.f. = 1, p = 0.004).
There was a negative correlation between the number of years since the start of ivermectin treatment and the proportion of infected and infective flies as measured from all those dissected, including the nullipars (Table 1) (infected: rs = –0.717, p = 0.045; infective: rs = –0.654, p = 0.078) and parous flies (infected: rs = –0.700, p = 0.188; infective: rs = –0.667, p = 0.219) (Fig 4), but this reached statistical significance only for the overall proportion infected.
Therapeutic coverage (the proportion of the overall population treated with ivermectin) for all villages was rarely below 60%. Coverage in Asubende, Pillar 83 and Gyankobaa had steadily increased since the beginning of mass treatment implementation, whilst Agborlekame, Dodi Papase, and Bosomase appeared to experience a recent decreasing trend in treatment coverage (Fig 5).
Monthly infective biting rates and monthly transmission potentials calculated from host-seeking flies only were zero in all villages except for Bosomase and Gyankobaa, the villages most recently incorporated into the CDTI programme. These transmission indices were also negative for Asubende, as the only fly identified as infective was an ovipositing blackfly caught using a Bellec trap, rather than a host-seeking fly.
Monthly infective biting rates varied greatly between villages, seasons, catching techniques and vector species (Table 2). In Bosomase, for human and cattle-seeking catching methods, these rates ranged from 0 to 42.2 infective bites/host/month, with higher values in the wet season than in the dry season. In the wet season of 2009, the forest form of S. sanctipauli was the main vector recorded in Bosomase and the only species with infective larvae, whilst in the dry season of 2010, S. yahense was the main vector species harbouring infective larvae (Table 2). At Gyankobaa, only 11 flies were collected in the dry seasons of 2010 and 2011 (S2 Table), all from Bellec traps, but in the wet season of 2009, the infective biting rates ranged from 38.9 infective bites/person/month, caught by vector collectors, to 50.4 infective bites/cow/month, for flies collected in the cow-baited tents (Table 2). Simulium sanctipauli flies harboured infective larval stages across all catching techniques at Gyankobaa indicating that this species was able to pick up infections from humans (although they would later attempt to feed on a non-human host), whereas infective S. damnosum s.s./S. sirbanum were only caught in the man-baited tents or by vector collectors, contributing both to transmission from humans to flies and from flies to humans. However, the overall sample sizes of S. damnosum s.s./S. sirbanum at Gyankobaa from the wet season of 2009 were low, with only 35, 4 and 13 S. damnosum s.s./S. sirbanum caught and dissected for infection from the vector collectors, human-baited and cow-baited tents, respectively.
The number of L3 larvae recorded varied between villages, seasons and catching techniques (S2 Table). The WHO states that a level of less than one L3 per 1,000 parous flies is required to control onchocerciasis transmission [11]. Gyankobaa in the wet season had levels of over 100 L3s per 1,000 parous flies whilst at Bosomase in the dry and wet season these were more than 350 and 250 L3s per 1,000 parous flies respectively (Fig 6A). Both these villages had not been included in the former OCP and were incorporated into the CDTI programme more recently in 2006 for Gyankobaa and 2003 for Bosomase. Asubende was just above this level, with 1.35 L3/1,000 (95% CI: 0–4.0) ovipositing (of which not all would be parous) flies (Fig 6B). This is despite 24 years of ivermectin treatment at the time of sampling, but the infection leading to this result was detected in an ovipositing rather than in a host-seeking fly, with 0 L3/1,000 host-seeking parous flies.
Combining L3 numbers and infective biting rates for the different vector species across trapping techniques for Gyankobaa and Bosomase resulted in transmission potentials ranging from 0 to 422.1 L3/host/month (Table 3). All flies at Asubende were S. damnosum s.s./S. sirbanum, but at Bosomase and Gyankobaa vector composition varied between seasons and catching techniques (S1 Table) [26]. Simulium sanctipauli was the most important vector species at both Bosomase and Gyankobaa in the wet seasons, whilst S. yahense played a more important role in transmission at Bosomase in the dry season of 2010. The importance of vector species at Bosomase in the dry season also differed between catching techniques, with S. sanctipauli having higher transmission potentials by flies caught in the human-baited tents, and S. yahense having higher transmission potentials by flies caught in the cow-baited tents (Table 3).
For all infected flies successfully identified to species (95 out of 97), the arithmetic mean number of O. volvulus larvae per infected fly varied greatly and statistically significantly among species, with S. damnosum s.s./S. sirbanum harbouring 1.33 larvae per infected fly ± 0.33 SE; S. sanctipauli 3.61 ± 0.40 and S. yahense 17.86 ± 4.32 (Kruskal Wallis χ2 = 15.50, d.f. = 2, p<0.001). The mean number of L3s per infective fly also differed statistically significantly among vector species, with S. damnosum s.s./S. sirbanum harbouring 1.33 L3s per infective fly ± 0.33 SE; S. sanctipauli 2.73 ± 0.50 and S. yahense 17.67 ± 9.23 (Kruskal Wallis χ2 = 6.83, d.f. = 2, p = 0.033). These differences were also observed when analysed at the village level, controlling for variations in local transmission levels, with S. yahense having significantly higher infection intensities at Bosomase in the infected flies (Mann Whitney U = 8.50, d.f. = 46, p<0.001). The difference had only borderline significance in the infective flies (U = 0.00, d.f. = 33, p = 0.057), as there was only one infective S. yahense with 17 L3s, despite the large difference between this and the mean in S. sanctipauli of 2.12 ± 0.56 L3/infective fly.
Overall, 258 of the 11,122 (2.3%) abdomens tested for Onchocerca infections were positive. The majority of these (240) were from Gyankobaa or Bosomase; however, there was also one positive result from each of Agborlekame, Dodi Papase and Pillar 83, which had been negative by dissection of heads and thoraces. The number of infected abdomens in vector collector flies was lower (0.8%) than that in the flies caught using all other methods combined (2.3%, χ2 = 58.0, d.f. = 1, p<0.001), suggesting that the infections did not originate from the flies acquiring an infectious blood meal with microfilariae at the point of collection.
As the goals of onchocerciasis control programmes shift from morbidity reduction towards elimination, knowledge of ongoing transmission by local vector species is urgently required [8,9,10]. This enables entomological monitoring of programmes’ progress, and helps to understand the determinants of persistent transmission despite prolonged control interventions. We report O. volvulus transmission, in Ghanaian communities with different treatment and control histories, and its variation according to simuliid species composition, vector trapping technique and season.
Factors influencing the feasibility of achieving elimination with the current ivermectin treatment strategy include baseline levels of endemicity, patterns of treatment coverage and compliance, parasite ivermectin susceptibility, duration and effectiveness of former vector control, seasonality of transmission in relation to ivermectin distribution, parasite immigration in flies or people, vector species mix and their associated vectorial capacity and competence for O. volvulus [63].
We have documented active onchocerciasis transmission, raising questions regarding the potential for CDTI alone to interrupt transmission under the treatment frequency and coverage levels commonly achieved in Africa. We report high monthly infectious biting rates and transmission potentials (measuring transmission from vectors to humans) for the communities most recently incorporated into the CDTI strategy. We also report infections in fly abdomens from all study villages, providing evidence of transmission from humans to flies. These infections were identified molecularly as O. volvulus. Infection levels above the WHO threshold of one L3 larva per 1,000 parous flies were recorded in the villages of Bosomase and Gyankobaa which started receiving treatment, respectively, in 2003 and 2006, i.e. 6 and 3 years prior to our entomological study. The WHO’s value forms part of the criteria for achieving the operational elimination thresholds for treatment cessation and commencement of surveillance [8], which in some West African foci have been reached after 14–17 years of annual (or biannual) ivermectin distribution [3,4]. This threshold was also exceeded in Asubende, which by the time of our study had received 24 years of ivermectin. Clear interpretation of this result is difficult since it is based on one infective fly caught in a Bellec trap, and flies using local breeding sites may originate from afar. However, there is also evidence from other studies that transmission in Asubende is continuing at a rate of >40 L3/person/month in some months (F.D.B. Veriegh, pers. comm.). Similarly, after 15 [64] and 17 [65] years of CDTI in Cameroon, or 20 years in the Central African Republic [66] have not resulted in interruption of transmission. Due to these and similar studies, there is a strong call for introducing more frequent (e.g. biannual) ivermectin treatments (or other strategies) if elimination is to be attained [67].
In regions in North Cameroon, approximately 70–90% of the filarial larvae in S. damnosum s.l. caught biting man were O. ochengi [68,69]. Given that cattle are present in some of our study villages (e.g. Agborlekame (~300 cows) and Asukawkaw Ferry (~500 cows), that S. damnosum s.l. flies feed on a range of blood hosts, and that 20% of the infective flies were caught using cattle-baited tents, we anticipated that we might have identified cattle-borne Onchocerca species such as O. ochengi but we only found O. volvulus. Over three quarters of the larvae had definitive O. volvulus identifications, and 96% of the unidentified larvae were from blackflies which had also contained known O. volvulus (of the same larval stage). No other species were identified and we are therefore confident that all of the Onchocerca larvae originated from flies infected with O. volvulus. This indicates active onchocerciasis transmission from humans to flies (early larval stages or infective flies attempting to feed on cattle) and from flies to humans (infective larvae in flies attempting to feed on humans). During the OCP, transmission potentials had been initially calculated on the assumption that all larvae would be O. volvulus; these ‘crude’ transmission potentials were subsequently corrected when tools for molecular identification of parasite larvae became available revealing that a geographically variable proportion of infective flies harboured non-volvulus Onchocerca spp. of zoonotic origin [70].
In 1980 (pre-ivermectin and pre-vector control), over 75% of the Asubende population were infected with microfilariae, and in 1987, prior to the ivermectin community trials, an infection prevalence of 80% was recorded [16], only slightly higher than that of Agborlekame (both in the Brong-Ahafo region). These communities were highly hyperendemic at baseline. The absence of infective flies observed at Agborlekame may be attributable to our low sample sizes, and/or recent treatment, rather than true lack of transmission. This conjecture is supported by on-going entomological studies (F.B.D. Veriegh pers. comm.) indicating high levels of L3 infections in flies from Agborlekame reaching 68 L3/person/month. This is further supported by our molecular analyses of fly abdomens, which revealed one infected fly in 83 flies analysed. At Asubende, biting rates have returned to pre-vector control levels [26], suggesting ecological conditions propitious for continuing transmission. Asubende has received regular annual treatment since 1987, and bi-annual treatment since 2009, with the most recent treatment round just 2 months before our sample collection. The village had a population of only 88 inhabitants at the time of sampling, and inspection of the community distributor’s notebooks and district records indicated a high therapeutic coverage. Therefore, in addition to the return of high biting rates and the possibility of infective flies migrating into the area [71], the potential for sub-optimal responses to ivermectin, perhaps suggesting decreased drug susceptibility, cannot be ignored. After 20 years of annual ivermectin administration, epidemiological assessments in 19 communities in Ghana, including Asubende, indicated a persistent reservoir of microfilarial infection [18,20].
In contrast, in the three Volta Region villages, transmission was low, despite a shorter history of vector control and ivermectin treatment than in Brong-Ahafo. The lack of infections may be attributable to the success of the OCP vector control strategy, which eliminated the Djodji form of S. sanctipauli [72], one of the S. damnosum complex species with the highest vector competence. Previous studies had shown that the Djodji form of S. sanctipauli carried, on average, three times as many L3 larvae per 1,000 biting flies as S. squamosum [73]. The reduction in biting rates associated with the disappearance of the Djodji form of S. sanctipauli [26] may also explain the reduction in transmission. Ivermectin treatment records also indicate that Pillar 83 had repeated ivermectin treatments in the years from 1993 to 1997 (potentially rapidly reducing levels of transmission in this community at the early stages of ivermectin control), followed by annual CDTI.
At Bosomase and Gyankobaa, which never received vector control and were incorporated into CDTI only recently, high levels of active transmission are still occurring, despite their lower baseline levels of infection intensity and prevalence, and current biannual or annual ivermectin treatment. In Gyankobaa, the most recent round of ivermectin distribution had taken place over a year before our sample collection date, providing ample opportunity for the reappearance of microfilariae in the hosts’ skin and their ensuing transmission [74,75]. In Bosomase, the high infection levels observed in the wet season in August 2009 are probably explicable by the missed annual treatment in that year, highlighting the importance of understanding the programmatic determinants of persistent transmission.
The transmission in the dry season of 2010 at Bosomase is of concern, with flies collected just one month after ivermectin treatment. However, seasonal variations (transmission levels in the 2009 wet season were higher than in the 2010 dry season), and in vector species composition and competence may also play a role in explaining the reported transmission patterns.
In the dry season, monthly transmission potentials were driven by S. yahense, with a higher number of L3s per infective fly than the extant form of S. sanctipauli. In contrast, the higher monthly infective biting rates in the wet season were driven by higher numbers of infective S. sanctipauli flies, despite their lower numbers of L3s per infective fly. Although not as anthropophagic and efficient a vector as the eliminated Djodji form, the forest form of S. sanctipauli has previously been demonstrated to be a highly efficient vector. In an area environmentally similar to, and just north of, Bosomase, a mean of 377 L3 in 1,000 parous flies, and 122 L3 per 1,000 biting flies (with 44% of parous flies infected) were recorded [76]. Even higher values, of 616 L3 per 1,000 parous flies have been reported in other African localities [56]. Overall, we observed lower infection rates than these, potentially due to the high therapeutic coverage of annual CDTI in this community. However, some reductions attributed to CDTI may actually be due to river pollution, lowering fly breeding success and associated transmission, particularly for S. sanctipauli [77], further supporting our previous biting rate findings and potential factors involved [26].
The influence of vector competence on transmission observed in Bosomase was also seen in Gyankobaa, where S. yahense, and to a lesser extent S. squamosum, were responsible for lower monthly transmission potentials due to lower biting rates and parous biting rates. In contrast, the forest form of S. sanctipauli, contributed to high numbers of L3/person/month due to high biting rates. Consequently, although both Bosomase and Gyankobaa have a shorter history of CDTI, the high transmission parameters recorded here for the vector species prevailing in this area must be emphasised. In Gyankobaa, infection levels (numbers of L3/1,000 parous flies) were 129 times as high, and in Bosomase, 291 to 365 times as high, as the WHO threshold. In both localities, the greatest proportion of L3 were found in S. sanctipauli, a species poorly or not at all represented in current transmission models.
Transmission models for African onchocerciasis have been mostly parameterised using S. damnosum s.s./S. sirbanum data [6,63,78,79,80,81,82,83] to reflect transmission dynamics in savannah areas suffering from severe ocular sequelae due to onchocerciasis. Exceptions to these models are the studies by Davies (1993) [84], based on transmission of forest onchocerciasis by S. soubrense B sensu Post; some quantitative analyses on other S. damnosum complex species, including S. leonense and S. squamosum B [85,86], and the recent modelling study of the effect of climate change on onchocerciasis transmission in Ghana and Liberia, including S. soubrense [87]. Our findings highlight that data on vector competence and vectorial capacity for O. volvulus for other important vector species are crucially needed, particularly as regions with diverse and seasonally varying simuliid vector composition strive towards elimination.
Approximately 40% of the flies were caught on Bellec traps, a similar proportion to that caught by the traditional OCP vector collector method, resulting in roughly equal numbers caught by host-independent and host-dependent methods. Light traps performed poorly, despite previous success at trapping S. squamosum [29] and other members of the S. damnosum complex [30] in Ghana. The prevalence of infected and infective flies, assessed by dissection, was similar among our host-dependent and host-independent catching techniques. Bellec-caught flies had higher infection prevalence, measured by DNA analyses of the abdomens, than the vector collector-caught flies. Positive abdomens in ovipositing flies could originate from microfilariae ingested with the blood meal (that did not escape the peritrophic matrix)—indicating transmission from humans to flies, and/or from L3 larvae migrating out of the thorax—indicating potential transmission from flies to humans. These results suggest that using oviposition (Bellec) traps in breeding sites along rivers close to villages, could augment (and perhaps replace) the more labour-intensive methods of human vector collection for monitoring vector infection levels. Large numbers of flies are required by techniques such as pool-screening [88], and with decreasing infection rates, the numbers to power transmission studies seeking to quantify reductions in transmission may need to be even larger [89]. Potential replacements for human landing catches, such as the Esperanza Window Trap, have been developed for S. ochraceum s.l. (the vector in Mexico and Guatemala) [90,91] and evaluated for host-seeking flies in Africa [92]. Oviposition traps have the added advantage that even nulliparous flies could contribute to the quantification of infection in thoraces, as sufficient time between an infected bite and oviposition elapses allowing any potential microfilariae to establish as L1s within the flies. The O. volvulus larvae thus collected could also be tested for ivermectin resistance markers once field probes are developed, helping in the monitoring and evaluation of transmission and of the potential spread of decreased ivermectin efficacy. This will become particularly pertinent with the increasing need for large-scale entomological evaluation of interventions as programmes strive for elimination, which will raise ethical concerns surrounding the widespread use of human landing catches. The host-independent Bellec traps could also be used in wider geographical perimeters during the post-MDA surveillance phase to complement more human exposure-focused methods in sentinel sites.
As vector competence is known to vary between seasons [93], blackfly collection was performed in both wet and dry seasons at five of the seven locations. (Due to incorporation at a later stage in the study of Asubende and Agborlekame, data were only collected during the dry season in these communities.) However, due to low blackfly catches at four of the study locations in one or the other of the seasons, Bosomase was the only location where substantial data were collected during both seasons. Although this reflects a lack of biting or ovipositing blackflies at the sampling times in these localities during these seasons, our results may not reflect true absence of simuliids and of any associated transmission for the whole season. This is particularly highlighted by our inability to detect Onchocerca larvae at Agborlekame, despite recent observations of on-going transmission (F.B.D. Veriegh, pers. comm.). Indeed, when blackfly abdomens were analysed, at least one positive result was obtained for O. volvulus infection in each of the villages assessed, indicating some level of active transmission. A potential limitation of analysing fly abdomens by molecular means is that higher levels of infection in vector collector-caught flies might be expected if any of the vector collectors caught the flies after the start of feeding and were themselves infected with microfilariae. There was no evidence that O. volvulus-positive abdomens were caused by microfilariae from the vector collectors as proportions of infected blackfly abdomens were significantly lower in the vector collector-caught flies than in those obtained by the remaining trapping methods.
|
10.1371/journal.ppat.1000004 | IL-10 from CD4+CD25−Foxp3−CD127− Adaptive Regulatory T Cells Modulates Parasite Clearance and Pathology during Malaria Infection | The outcome of malaria infection is determined, in part, by the balance of pro-inflammatory and regulatory immune responses. Failure to develop an effective pro-inflammatory response can lead to unrestricted parasite replication, whilst failure to regulate this response leads to the development of severe immunopathology. IL-10 and TGF-β are known to be important components of the regulatory response, but the cellular source of these cytokines is still unknown. Here we have examined the role of natural and adaptive regulatory T cells in the control of malaria infection and find that classical CD4+CD25hi (and Foxp3+) regulatory T cells do not significantly influence the outcome of infections with the lethal (17XL) strain of Plasmodium yoelii (PyL). In contrast, we find that adaptive IL-10-producing, CD4+ T cells (which are CD25−, Foxp3−, and CD127− and do not produce Th1, Th2, or Th17 associated cytokines) that are generated during both PyL and non-lethal P. yoelii 17X (PyNL) infections are able to down-regulate pro-inflammatory responses and impede parasite clearance. In summary, we have identified a population of induced Foxp3− regulatory (Tr1) T cells, characterised by production of IL-10 and down regulation of IL-7Rα, that modulates the inflammatory response to malaria.
| Much of the pathology of malaria infection is due to an excessive inflammatory response to the parasite. The regulatory cytokine IL-10 is known to control inflammation during malaria infections and thus protect against immunopathology, but, in so doing, it reduces the effectiveness of other immune mechanisms which remove the parasites. In order to try to dissociate these two effects of IL-10, to allow simultaneous control of infection and avoidance of pathology, we need a better understanding of the processes leading to IL-10 production, the timing of its production, and the cells that produce it. In this study we have found that the major source of IL-10 during malaria (Plasmodium yoelii) infection is adaptive regulatory CD4+ T cells. This population is distinct from natural regulatory T cells and classical effector T cells. IL-10 derived from these adaptive CD4+ T cells prevents hepatic immunopathology but also suppresses the effector T cell response, preventing parasite clearance. Further work is now required to determine how these two key cell types (anti-parasitic effector T cells and IL-10-producing regulatory T cells) are induced, so that vaccines can be designed that will induce optimal numbers of each cell type at appropriate stages of the infection.
| The erythrocytic stage of malaria infection is characterised by the development of strong pro-inflammatory immune responses which, although required to control parasite replication and promote clearance of infected erythrocytes, must be tightly regulated to prevent the immune-mediated pathology which is integral to the development of the severe complications of infection in humans and in a number of well-characterised animal models [1]–[3]. Previous studies have highlighted important roles for IL-10 and TGF-β in regulating the pro-inflammatory response during malaria infection [4]–[11]. Thus, although IL-10−/− and TGF-β-depleted mice are able to control parasite replication during P. chabaudi AS infection as effectively as WT mice, unlike WT mice they develop severe TNF-mediated pathologies which are typically fatal [4], [9]–[11]. Similarly, IL-10 can prevent the onset of cerebral malaria in P. berghei ANKA-infected mice [8].
However, the exact role of IL-10 and TGF-β appears to vary between infections with different malaria species and strains, depending on the timing of cytokine production in relation to disease progression. Thus, production of TGF-β and IL-10 during the first few days of a lethal P. yoelii 17XL (PyL) infection is associated with inhibition of pro-inflammatory responses, rapidly escalating parasitaemia and death [5],[7]. In contrast, mice infected with the non-lethal variant (P. yoelii 17X; PyNL) produce no or only low levels of TGF-β and IL-10 during early acute infection and eventually control their parasitaemia [5]. Blockade of IL-10R signalling in combination with anti-TGF-β treatment restores the type-1 immune response during lethal P. yoelii infection, and a proportion of infected animals are able to control their infections and survive [5]. Moreover, splenocytes from susceptible BALB/c mice, but not resistant DBA/2 mice, infected with PyNL produce IL-10 and TGF-beta during the early acute stage of infection, which is associated with an increase in the proportion of splenic CD25+ CD4 T cells [12]. Taken together, these studies demonstrate a causal role for immunoregulatory cytokines in suppressing parasite clearance mechanisms.
In accordance with these findings, a study by Hisaeda and colleagues indicated that differential activation of natural regulatory T cells (nTreg) may account for the differing virulence of P. yoelii strains, since depletion of CD4+CD25hi T cells (with anti-CD25 antibody) prior to infection converted PyL from a rapidly lethal infection into a resolving infection but had no effect on the course of PyNL infection [13]. Although first identified as cells that limit autoimmune pro-inflammatory responses [14], nTreg (defined by expression of CD4, the transcription factor Foxp3 and high levels of CD25) have since been shown to regulate the immune response in a number infections including Leishmania spp infections, Mycobacterium tuberculosis and helminth infections [15]–[18], mediating their effects either via direct cell contact or by release of cytokines. However, it is now becoming apparent that both adaptive (Foxp3−) regulatory T cell populations and classical T-bet expressing Th1 cells also play crucial immunoregulatory roles during infection and mediate their effects through secretion of IL-10 [19]–[21].
In this study we have examined the generation and function of both nTreg and adaptive IL-10-secreting T cells during malaria infection. We observe equivalent expansion of natural Foxp3+ regulatory T cells during both lethal and non-lethal P. yoelii infections but, using either anti-CD25 treatment or adoptive transfer of purified CD25hi/Foxp3+ nTreg or CD25−/Foxp3− non-Treg T cell populations, we find no role for nTreg during PyL infection. Conversely, we demonstrate that populations of adaptive regulatory CD4+ T cells, that are CD25−, Foxp3− and CD127−, and which do not make IFN-γ, IL-4 or IL-17, develop during both PyL and PyNL infections. These cells inhibit parasite clearance but, importantly, also prevent the development of pathology via production of IL-10. These data are consistent with the notion that whilst endogenous populations of nTreg may be sufficient to prevent immune-mediated pathology during chronic infections which induce rather modest inflammatory responses, such as avirulent leishmania, tuberculosis or helminth infections, rapid induction of distinct populations of adaptive/Th1 CD4+ T cells producing IL-10 may be required to counter the powerful inflammatory signals provided by virulent, rapidly replicating pathogens.
In accordance with previous observations [5],[22], infection of C57BL/6 mice with 104 P. yoelii 17XL (PyL) parasites was associated with a rapid onset of fulminant parasitaemia (approaching 100% by day 7 pi) that was universally fatal (Figure 1A, B). In contrast, infection with 104 P. yoelii 17X (NL) (PyNL) parasites led to a more gradual increase in parasitaemia with peak parasitaemia of approx. 30% on day 14 pi, before the infection eventually resolved. Significant differences in malaria-induced anaemia were also evident between lethal and non-lethal infections, with more rapid onset and increased severity of anaemia occurring in PyL-infected mice compared with PyNL-infected mice (Figure 1C).
We have previously reported that simultaneous neutralisation of TGF-β and blockade of IL-10 signalling allows a proportion of PyL-infected mice to resolve their infections and survive [5], suggesting that active immune regulation/immune suppression occurs during PyL infection that inhibits optimal parasite control. In agreement with these observations, Kobayashi et al [7] have reported that IL-10 is produced very early during PyL (but not during PyNL) infection and Perry et al [23] have reported a switch from IL-12 (at day 3 pi) to IL-10 (at day 17 pi) production by splenic dendritic cells during the course of a non-lethal Py infection. These data are consistent with the hypothesis that protective pro-inflammatory responses develop during the acute phase of PyNL infection that limit parasite numbers, whereas an early anti-inflammatory cytokine response during the acute phase of PyL infection inhibits the development of protective immune responses. As CD4+CD25+ regulatory T cells (nTreg) have been reported to regulate immunity in a number of auto-immune and infectious diseases [14]–[18] and can exert their regulatory role through secretion of IL-10 and/or TGF-β we investigated, using intracellular staining for Foxp3 as well as transgenic Foxp3-GFP reporter mice [24], whether nTreg activation is correlated with the virulence of PyL infection.
CD4+ splenic lymphocytes from uninfected (control) mice, or from PyL- or PyNL- infected mice, were analysed for intracellular Foxp3 expression (Figure 2A) and the numbers of CD4+Foxp3+ cells, the expression levels of Foxp3 and the ratios of CD4+Foxp3+ (nTReg) to CD4+Foxp3− (non-regulatory T cells) were assessed over the first 7 days pi (Figure 2B–D). In accordance with previous observations [25] a significant increase in the numbers of splenic CD4+Foxp3+ nTreg was observed during the first 5 days of PyL infection (Figure 2B) and this was accompanied by increased levels (MFI) of Foxp3 expression (Figure 2C) and a transient increase (on day 3pi) in the nTreg/non-Treg ratio (Figure 2D). However, almost identical changes in nTreg numbers and Foxp3 expression levels were observed in mice infected with PyNL, and there were no significant differences in any nTreg parameter between PyL-infected and PyNL-infected mice at any time up to 7 days pi, after which the PyL-infected mice succumbed to their infections.
Similar results were obtained with Foxp3-GFP reporter mice [24]. Importantly, the course of PyL and PyNL infections were equivalent in Foxp3-GFP mice and C57BL/6 mice (data not shown). A representative plot showing Foxp3-GFP expression in infected and uninfected mice is shown in Figure 2E. Numbers of CD4+GFP+ cells were significantly increased in the spleen on 5 day pi (Figure 2F) and on day 7 pi (data not shown) but did not differ significantly between PyL-infected and PyNL-infected mice. Finally, no significant differences were observed in expression of Foxp3 mRNA in CD4+ T cells purified from spleens of PyL and PyNL-infected mice on days 1, 3, 5 and 7 pi (data not shown).
The similarity of the nTreg response during PyL and PyNL infections suggested that, in our hands, suppression of effector cell responses by nTreg was unlikely to explain the highly virulent nature of PyL infections. However, to formally test the role of nTreg, mice were treated with a cocktail of anti-CD25 antibodies (previously shown to give optimal depletion of CD4+CD25hiFoxp3+cells; 25) 3 days prior to infection with PyL (Figure 3). As previously reported [25], the 7D4 (IgM, anti-CD25) antibody substantially reduced the proportion of splenic CD25+ CD4 cells within 3 days (i.e day of infection) but CD25+ cells recovered to normal levels by day 4 pi (results not shown) and 7D4 treatment had no significant effect on the frequency of CD4+Foxp3+ve cells (results not shown). In contrast, PC61 (IgG anti-CD25) given in combination with 7D4 induced an approximately 50% reduction in the frequency of both CD25+ and Foxp3+ cells that was sustained throughout the 7 day infection period [25]. Nonetheless, neither 7D4 treatment nor combined 7D4+PC61 treatment significantly altered the course of parasitaemia, anaemia or survival of PyL infection in C57BL/6 mice (Figure 3A–C). As these observations contradict those of a similar published study [13] we considered whether some effect of natural T reg might be being masked by the rapidly ascending parasitaemia and early mortality associated with infection with 104 PyL parasites. We therefore repeated the anti-CD25 antibody treatment in C57BL/6 mice infected with either a 10 fold lower dose of PyL parasites (103 PyL) or with 104 PyNL parasites. However, anti-CD25 antibody treatment did not alter the outcome of either of these infections (Figures S1 and S2) suggesting that natural T reg cells do not markedly influence P. yoelii infections in C57/BL6 mice.
It has been reported that regulatory T cell responses are more effective at limiting pro-inflammatory responses in BALB/c mice than in C57BL/6 mice [26]. Therefore, to determine whether mouse strain influences the outcome of anti-CD25 treatment during PyL infection, we repeated the CD25-depletion experiments in BALB/c mice and compared our depletion strategies (single dose of 7D4 or 7D4+PC61 given 3 days prior to infection) with a strategy previously shown to affect PyL infection [13], namely repeated injections of 7D4 antibody on days −3, −1 and 5 relative to PyL infection. Repeated administration of 7D4 did not increase either the duration of CD25+ T cell depletion or the extent of depletion of CD4+Foxp3+ve cells compared to the other treatment regimes (Figure S3). Consistent with this, repeated administration of 7D4 did not alter the course of PyL infection compared with single 7D4 administration or combined 7D4 and PC61 administration (Figure S3), and none of our CD25-depletion regimes had any effect on PyL infection in BALB/c mice (Figure S3).
It is becoming increasingly evident that anti-CD25 antibody treatment is not a specific or robust strategy to examine the importance of natural regulatory T cells during inflammatory episodes [25], [27]–[29]. CD25 expression is not limited to nTreg [24]. Moreover, depending on the precise protocol used, a variable but significant proportion of Foxp3+ cells escape depletion by anti-CD25 antibody. We therefore compared the outcome of PyL infection in RAG−/− mice reconstituted or not with purified naïve CD4+CD25− (putative effector) T cells or a 10∶1 ratio of effector (CD4+CD25−) to nTreg (CD4+CD25+) cells. Furthermore, as nTreg can down-regulate NK cell responses [30], and as NK cells have previously been reported to play a protective role during malaria infection [31]–[33], we adoptively transferred CD4+CD25+ (nTreg) cells in the absence of CD4+CD25− (effector) cells, to determine whether nTreg modulate innate immune responses during malaria infection. The proportion of Foxp3+ cells fell from 10–15% in unsorted CD4+ T cells to 1–2% in the CD25−CD4+ population, whereas CD25+ cells were highly enriched for Foxp3+ cells (70–80%; Figure 3D).
In accordance with our previous studies [22], we found that control (unreconstituted) RAG−/− mice succumbed to PyL infection with the same kinetics as WT mice (compare Figure 3E, F with Figure 1). Furthermore, the course of infection was virtually indistinguishable in RAG−/− mice reconstituted with CD4+CD25−, CD4+CD25+ or a 10∶1 ratio of CD4+CD25−/CD4+CD25+ T cells (Figure 3E, F). Thus, using two independent models of nTreg depletion, we have found no significant role for natural CD4+CD25+Foxp3+ regulatory T cells in suppression of anti-parasitic immunity during PyL infection in either C57BL/6 or BALB/c mice.
Having found no evidence that nTreg influence the outcome of PyL infection we next investigated the possibility that IL-10 producing CD4+ T cells (“adaptive” Treg or Tr1 cells) might be induced during PyL and/or PyNL infection that regulate parasite killing and/or pathology.
Expression of IL-10 mRNA was determined by real time PCR in purified splenic CD4+ T cells obtained on days 1, 3, 5 and 7 post-infection from wild type (WT) C57/BL6 mice and plasma levels of IL-10 were determined by ELISA on days 1, 3, 5 and 7 pi from WT and RAG−/− mice. We find that CD4+ T cells are a significant source of IL-10 by day 5 of both PyL and PyNL infections, since IL-10 mRNA is significantly upregulated in splenic CD4+ cells on days 5 and 7 post-infection compared with cells from uninfected mice (Figure 4A). Furthermore, CD4+ T cells (and potentially B cells) may be the major source of IL-10 during infection since plasma IL-10 does not increase above baseline levels in RAG−/− mice (Figure 4B, C) except on day 3 pi of PyL infection.
To more accurately determine the cellular source of IL-10 during P. yoelii infection, splenocytes from IL-10-GFP reporter mice [21] were examined for expression of GFP and various cell surface markers on selected days after PyL or PyNL infection (Figure 5A–C). In both infections, from day 5 onwards, the vast majority of the IL-10+ cells were CD4+ lymphocytes. At no point during either PyL or PyNL infection did we observe significant IL-10 production by myeloid (CD11b+), lymphoid dendritic cells (CD11c+) or macrophages (F4-80+) (results not shown). IL-10 production by CD19+ B cells was observed, on day 7 post-infection, only during PyL but not PyNL infection (results not shown). Moreover, IL-10 producing non-CD4+ T cells produced only low quantities of IL-10, whereas CD4+ T cells were heterogeneous in their ability to produce IL-10 (Figure 5A).
Since it is not possible to stain for intranuclear Foxp3 without quenching the fluorescence of GFP, IL-10/GFP+ CD4+ T cells were analysed for expression of CD25, CD62L and CD127 and separately analysed for CD25 and Foxp3 (Figure 5B, C). On day 5 post-infection, IL-10+ CD4+ T cells showed very variable expression of CD25 with approx. 60% being CD25−, indicating that they are not a typical nTreg population. As we have previously observed transient upregulation of CD25 on CD4+Foxp3− T cells at this time (Figure 5B and [22]), we considered it likely that at 5 days post-infection the majority of IL-10+ cells were Foxp3−. In confirmation of this, by day 7 post-infection, IL-10+ CD4+ T cells were almost exclusively CD25− indicating that, since the majority of Foxp3+ cells maintain CD25 expression during P. yoelii infection (Figure 5B), CD25−Foxp3− CD4+ T cells are the primary source of IL-10 during both PyL and PyNL infection. Interestingly the frequencies and numbers of IL-10+ CD4+TCR-β+ cells were equivalent in PyL and PyNL infected mice on day 7 post-infection (Figure 5D).
IL-10+ cells were heterogeneous in terms of expression of CD62L suggesting that they comprise of a mixed population of cells in terms of memory/activation status, and despite being Foxp3−, the majority of IL-10+ CD4+ cells were CD127−, suggesting that down-regulation of IL-7Rα may be a useful marker for differentiating adaptive Treg from other antigen-experienced T cells (Figure 5C).
We have shown that CD4+ T cells are the primary source of IL-10 during malaria infection, and that these cells do not express CD25, suggesting that they may not be conventional nTreg cells. Since IL-10 can be produced by various effector CD4+ T cell subsets (including Th1, Th2 and Th17 cells), as well as specialised regulatory populations such as Tr1 [19, 20 34–36], we examined the expression of Th1, Th2 and Th17 lineage-associated cytokines in IL-10-producing (GFP+) and IL-10-GFP− CD4+ T cells purified from IL-10-GFP reporter mice on day 7 of infection. As seen previously (Figure 5), GFP expression was similar in CD4+ T cells isolated from PyL and PyNL infected mice (Figure 6A). As expected, IL-10 mRNA was expressed at much higher levels in GFP+ than in GFP− cells but cells isolated from PyL and PyNL infected animals expressed similar levels of IL-10 mRNA (Figure 6B). Importantly, Foxp3 mRNA was not upregulated in IL-10-GFP+ cells isolated during either PyL or PyNL infection, confirming that the IL-10-producing CD4+ T cells that develop during P. yoelii infection are neither natural nor induced Foxp3+ regulatory T cells. Moreover, GFP+ cells did not express significant amounts of mRNA for IFN-γ, IL-4 or IL-17, thus distinguishing them from classical Th1, Th2 and Th17 cells. Although IL-10-GFP+ cells expressed IL-13 mRNA, levels were comparable to those seen in GFP− cells indicating that IL-10 producing cells did not preferentially co-produce IL-13. Thus, the IL-10-producing CD4+ T cells induced during P. yoelii infection fit the definition [35] of adaptive, Tr1, regulatory T cells.
To determine whether IL-10 production from T cells is functionally important during Py infection, we first compared the course of PyL and PyNL infection in IL-10−/− and WT mice (Figure 7). PyNL infection was significantly attenuated - with significant reductions in parasitaemia and anaemia in IL-10−/− mice compared with WT mice (Figure 7A–D), although the IL-10−/− mice did lose significantly more weight than age-matched WT mice (Figure 7C). Furthermore, approx 30% (6/21 mice) of IL-10−/− (but not WT) mice infected with 104 PyL pRBC were able to control their infections and survived (Figure 7E–H), with parasitaemia declining from a peak of approx 45% on day 6pi. Moreover, IL-10−/− (but, again, not WT) mice given a low dose PyL infection (103 pRBC) were fully able to control parasitaemia and 100% of the mice survived (Figure 7I–L).
Taken together, these data indicate that IL-10 suppresses immune effector mechanisms which would otherwise be able to control low dose PyL infections. Since this IL-10 emanates principally from CD4+ T cells (Figure 5) we hypothesised that IL-10-deficient CD4+ T cells may promote more effective parasite control than WT CD4+ T cells. To test this, purified naïve WT or IL-10−/− CD4+ T cells were adoptively transferred into RAG−/− mice which were then infected with PyNL or PyL parasites.
PyNL-infected RAG−/− mice that had received IL-10−/− CD4+ T cells developed significantly lower parasite burdens than those which had received WT CD4+ T cells (Figure 8A). Although both groups developed similar levels of anaemia, mice that received IL-10−/− T cells lost significantly more weight and succumbed to infection more rapidly than mice that received WT CD4+ T cells (Figure 8B–D). Exacerbation of disease despite improved parasite control in mice receiving IL-10−/− CD4+ T cells was associated with more extensive proliferation of the adoptively transferred T cells (IL-10−/− T cells comprised >30% of total splenic leucocytes compared with <10% for transferred WT cells), higher concentrations of circulating IFN-γ and lower plasma concentrations of IL-10 (data not shown). These data are consistent with the conclusion that recipients of IL-10−/− CD4+ T cells died of immunopathology whilst recipients of WT CD4+ T cells eventually died because they were unable to fully resolve their infections.
By contrast, RAG−/− mice that had received IL-10−/− CD4+ T cells were somewhat better able to control infections with 103 (8E–H) or 104 (8I–L) PyL infections than were mice receiving WT CD4+ T cells; a proportion of mice receiving IL-10−/− T cells were able to control their infections, although failure to fully eliminate parasites eventually led to death from anaemia. Thus, IL-10 derived from CD4+ T cells significantly modulates the outcome of both PyL and PyNL infection.
It has previously been shown that IL-10−/− mice succumb to normally avirulent P. chabaudi chabaudi infections despite comparable - or more effective - control of malaria parasitaemia compared to WT mice [9]. The increased susceptibility of IL-10−/− mice is due to elevated plasma concentrations of IFN-γ and TNF-α [10] and survival of IL-10−/− mice following malaria infection can be enhanced by treatment with anti-TNF-α [10]. Whilst there was no marked difference in mortality between P. yoelii-infected IL-10−/− and WT mice, IL-10−/− mice (and RAG−/− mice reconstituted with IL-10−/− T cells) lost significantly more weight than mice reconstituted with WT T cells during PyNL infection, indicative of more severe morbidity (Figure 7C, 8C).
Histopathological examination of infected animals did not reveal any liver or lung damage 3 days post-infection (data not shown) but revealed significantly more hepatic cellular changes including periportal inflammation, necrosis and bridging necrosis in IL-10−/− mice than in WT mice on days 7 and 14 post-infection (Figure 9A) and this was significantly more severe in PyL-infected than PyNL-infected animals on day 7 post-infection. We also found that by day 25 of PyNL infection, RAG−/− recipients of IL-10−/− CD4+ T cells had developed significantly more severe hepatic periportal inflammation and necrosis (including bridging necrosis) than RAG−/− recipients of WT CD4+ T cells (Figure 9B). Thus, T cell derived IL-10, although negatively regulating parasite killing, is protective during malaria infection by preventing the onset of immunopathology.
It is well established, in a variety of infections, that regulatory cytokines both ameliorate immunopathology and delay pathogen clearance [5], [8], [9]–[11], [37]–[42]. Manipulation of these cytokines by vaccination or immunotherapy, to simultaneously enhance pathogen clearance and limit the associated pathology, requires a better understanding of their cellular sources and mechanisms of induction. Important roles have been demonstrated for both IL-10 [6]–[11] and TGF-β [4],[5] in modulating the outcome of murine malaria infections, and observational data strongly suggests that they play a similar role in human infections [43]–[45]. Recently, endogenous or natural, CD25hi, Foxp3+ CD4+ T cells (nTreg) have been implicated as major regulators of malarial pathology [13],[46] but their mechanisms of action remain undefined. Attempting to elucidate the role of nTreg in murine Plasmodium yoelii infections, we were surprised to find no role for these cells in regulating the outcome of either high dose (104) or lower dose (103) lethal (Py17XL; PyL) or non-lethal (Py17X; PyNL) infection in either C57BL/6 or BALB/c mice. In contrast, we find that adaptive, IL-10-producing CD4+ Tr1 cells (CD25−, Foxp3−, CD127−, IFN-γ−, IL-4− and IL-17−), are generated during both PyL and PyNL infections and are associated with down-regulation of pro-inflammatory responses, moderation of both morbidity and mortality and failure to clear parasites. Crucially, we were able to demonstrate a causal relationship between these various observations by showing that IL-10−/− CD4+ T cells adoptively transferred into RAG−/− mice provided more effective parasite control than did WT CD4+ T cells, but at the cost of more severe pathology.
We conclude that induced Foxp3− regulatory T cells, characterised by down-regulation of CD127/IL-7Rα, modulate the inflammatory response to Plasmodium yoelii malaria by production of IL-10. Although it has been observed in humans [47],[48] and mice [48] that CD127 is down-regulated on endogenous (Foxp3+) regulatory T cells, our data demonstrate – for the first time - that CD127 is also down-regulated on adaptive (Foxp3−) Tr1 cells.
The lack of any effect of anti-CD25 antibody treatment on the course of PyL infection in our experiments contradicts the published data [13],[46]. Despite numerous attempts, using three different depletion protocols - including a protocol identical to that previously reported to ameliorate PyL infection [13], in both C57BL/6 and BALB/c mice we were unable to reproduce the published observations. Anti-CD25 treatment failed to ameliorate PyL infection initiated by a 10 fold lower dose of parasites, discounting the possibility that the virulence of high dose PyL infection masks regulatory activity of nTreg that would otherwise be evident during a less virulent infection. Furthermore, adoptive transfer of CD25+ and/or CD25− CD4+ T cells into RAG−/− mice also failed to reveal any role for CD25+ Foxp3+ cells in this infection. The discrepancy between our findings and those of other labs is reminiscent of the disparate results obtained for P. berghei ANKA infection which report that depletion of CD25+ regulatory cells either facilitates parasite control and prevents the onset of the cerebral pathology infection in C57BL/6 mice [49], or enhances effector T cell responses and increases the severity of brain pathology in normally resistant BALB/c mice [50] or has no effect on cerebral pathology [51]. One explanation for these inconsistent results may be differences in prior exposure to pathogens or commensal organisms between mice in different laboratories. Components of the normal intestinal flora of conventionally housed animals are essential for development of intestinal nTreg [52] and nTreg development is facilitated by the presence of covert infections such as Helicobacter hepaticus [53]. Depletion of nTreg by anti-CD25 treatment in such mice may lead to more profound alteration in the effector / regulatory cell balance than in mice (such as those used in our studies) raised in low-infection environments.
Although, we could show no role for nTreg in acute PyL and PyNL infection, we have shown that the adaptive IL-10-producing regulatory T cells that develop during P. yoelii infection hinder parasite control but simultaneously limit disease severity. In contrast to recent studies describing a role for IL-10 producing Th1 cells in Toxoplasma gondii [20] and Leishmania spp [19],[54],[55] infections, the adaptive IL-10 producing Tregs we describe do not co-express IFN-γ or other effector cytokines and better fit the definition of Tr1 cells. Nevertheless in several virulent protozoal infections (PyL, Toxoplasma gondii, Leishmania major SD and L. donovani), adaptive IL-10-producing CD4+ T cells are required to regulate the fulminant Th1-effector responses that are induced whereas classical (Foxp3+) Treg appear to be sufficient to regulate the effector response to a less virulent (healing) strain of L. major [15]. Given that naïve CD4+ T cells can develop into adaptive Treg after interaction with IL-10-producing dendritic cells expressing low levels of co-stimulatory molecules [56]–[59], it is possible that the induction of Treg during P. yoelii infection is linked to the modulation of macrophage and dendritic cell function that occurs in response to prolonged toll-like receptor signalling [60]. Alternatively, parasite-induced TGF-β [40],[60], IL-6 [44] and/or IL-27 may synergise to promote production of IL-10 by Th1, Th2, Th17 and Tr1 cells [34],[36],[61],[62].
We have not yet definitively identified the mechanism by which IL-10 suppresses parasite clearance but given our recent findings that control of the acute phase of P. yoelii parasitaemia is critically dependent on macrophages [22], it is likely that T cell-derived IL-10 acts directly on macrophages to inhibit their anti-parasitic mechanisms. It is also possible that, as in mycobacterial infections, adaptive Treg induce an autocrine signalling loop in which macrophages both secrete and respond to IL-10 with consequent down regulation of effector function and pathology via a STAT-3 –dependent pathway [63]–[66].
In summary, we have demonstrated that adaptive, but not natural, regulatory T cells control parasite numbers during PyL and PyNL infections whilst also limiting the onset of immunopathology. These cells are characterised by lack of expression of CD25 and Foxp3, down-regulation of CD127 and production of IL-10 but not IFN-γ, IL-4 or IL-17. Taken together with our data highlighting the importance of macrophages in the control of malaria infection [22], these findings identify an important pathway of adaptive, T cell- mediated control of innate immune responses. Further studies are required to identify the pathways leading to induction of this important regulatory cell population.
C57BL/6, Foxp3-GFP (F2: 129/C57BL/6; from A. Rudensky, University of Washington, 24), C57BL/6 RAG-1−/−, C57BL/6 IL-10−/− and BALB/c mice were bred in-house or purchased from Harlan and maintained under barrier conditions at LSHTM. IL-10-GFP reporter mice [21] were maintained under barrier conditions at the National Institutes of Health. Cryopreserved Plasmodium yoelii 17X (non lethal; PyNL) and P. yoelii 17XL (lethal; PyL) parasites were passaged once through mice before being used in experimental animals.
Unless stated otherwise, male or female mice, 7–9 weeks of age, were infected intraveneously with 1 × 103 or 1 × 104 parasitised red blood cells (pRBC). Parasitaemia was determined daily by examination of Giemsa-stained thin smears of tail blood for the first seven days of infection and every second day thereafter. On every second day, mice were weighed and RBCs were counted using an automated haemoanalyser (Beckman Coulter). Plasma was stored (at −20°C) for cytokine quantification. On selected days post-infection, mice were sacrificed and spleens were removed. Single spleen cell suspensions were prepared by homogenisation through a 70 µm cell strainer (BD Biosciences) and live cells enumerated by trypan blue exclusion.
CD4+ T cells were positively selected using anti-mouse CD4-conjugated midiMACS beads (Miltenyi Biotec) according to the manufacturer's instructions and the purity of eluted cells was checked by flow cytometry. In some experiments, the CD4+ cells were labelled with anti-mouse CD4 (GK1.5: Rat IgG2b: E-bioscience) and anti-mouse CD25 (PC61: Rat IgG1: E-bioscience) fluorochrome-labelled antibodies and sorted, using a BD FACSVantage SE, into CD4+CD25+ and CD4+CD25− populations. In separate experiments CD4+ T cells, isolated from IL-10-GFP reporter mice [21] on day 7 of infection, were labelled with anti-mouse CD4 (GK1.5: Rat IgG2b: E-bioscience) and IL-10 producing (GFP+) and non-IL-10 producing (GFP−) CD4+ T cells were purified by flow cytometric cell sorting.
IL-10, Foxp3, IFN-γ, IL-4, IL-13 and IL-17A mRNA were quantified by Taqman (ABI, Warrington, UK). RNA was extracted (RNAeasy) and DNAse1 treated prior to cDNA synthesis. cDNA expression for each sample was standardised using the housekeeping gene GAPDH. Cycling conditions were: initialisation 2 min at 50°C and 10 min at 95°C followed by 40 cycles of 15 sec at 95°C and 1 min at 60°C.
Primer sequences:
IL-10 Forward ATGCTGCCTGCTCTTACTGACTG
Reverse CCCAAGTAACCCTTAAAGTCCTGC,
Foxp3 Forward CACCTATGCCACCCTTATCC,
Reverse CGAACATGCGAGTAAACCAA
IFN-γ Forward AGA GCC AGA TTA TCT CTT TCT ACC TCA G
Reverse CCT TTT TCG CCT TGC TGT TG
IL-4 Forward ACG AGG TCA CAG GAG AAG GGA
Reverse AGC CCT ACA GAC GAG CTC ACT C
IL-13 Forward CCTCTGACCCTTAAGGAGCTTAT
Reverse CGTTGCACAGGGGAGTCT
IL-17A Forward TGTGAAGGTCAACCTCAAAGTC
Reverse AGGGATATCTATCAGGGTCTTCATT
Rat anti-mouse IL-10 (JES5-2A5; Rat IgG1; Mabtech, Sweden) or rat anti-mouse interferon (IFN)-γ (AN-18; Rat IgG1; eBioscience) antibodies were used as capture antibodies, diluted in 0.5 M Tris-HCl, pH 8.9 buffer. Biotinylated rat anti-mouse IL-10 MAb (JES5-16E3; Rat IgG2b; Mabtech) or rat anti-mouse IFN-γ MAb (R4-6A2; Rat IgG1; Mabtech) were used as detecting antibodies and were visualised using streptavidin-alkaline phosphatase (eBioscience) and p-nitrophenyl phosphate (Sigma Aldrich, UK). Absorbance was read at 405 nm using a MRX TC II microplate reader (Dynex Technologies Ltd,. UK)
For flow cytometric analysis, cells were surface stained with anti-mouse CD4 (RM4-5; Rat IgG2a; BD Biosciences), anti-mouse CD25 (PC61; Rat IgG1; eBioscience), anti-mouse CD69 (H1.2F3; Armenian Hamster IgG; eBioscience), anti-mouse CD62L (MEL-14; Rat IgG2a; eBioscience) or anti-mouse CD127 (A7R34; Rat IgG2a; eBioscience). Intracellular Foxp3 staining using anti-mouse Foxp3 (FJK-16s; Rat IgG2a; eBioscience) was performed by permeabilising cells with 0.1% Saponin/PBS. Cells were concurrently incubated with anti-mouse CD16/32 (Fc block) when staining with all conjugated antibodies. Flow cytometric acquisition was performed using a FACSCalibur (BD Immunocytometry Systems, USA) and all analysis was performed using Flowjo software (Treestar Inc., OR, USA)
Liver and lung tissues were fixed in 10% Formalin-saline. Fixed tissues were paraffin embedded and stained with haematoxylin and eosin. Slides were microscopically examined at 20X magnification.
Statistical significance was determined using Student's T test, unless otherwise stated, with P<0.05 taken as indicating a significant difference. |
10.1371/journal.ppat.1000373 | Synthetic Double-Stranded RNAs Are Adjuvants for the Induction of T Helper 1 and Humoral Immune Responses to Human Papillomavirus in Rhesus Macaques | Toll-like receptor (TLR) ligands are being considered as adjuvants for the induction of antigen-specific immune responses, as in the design of vaccines. Polyriboinosinic-polyribocytoidylic acid (poly I:C), a synthetic double-stranded RNA (dsRNA), is recognized by TLR3 and other intracellular receptors. Poly ICLC is a poly I:C analogue, which has been stabilized against the serum nucleases that are present in the plasma of primates. Poly I:C12U, another analogue, is less toxic but also less stable in vivo than poly I:C, and TLR3 is essential for its recognition. To study the effects of these compounds on the induction of protein-specific immune responses in an animal model relevant to humans, rhesus macaques were immunized subcutaneously (s.c.) with keyhole limpet hemocyanin (KLH) or human papillomavirus (HPV)16 capsomeres with or without dsRNA or a control adjuvant, the TLR9 ligand CpG-C. All dsRNA compounds served as adjuvants for KLH-specific cellular immune responses, with the highest proliferative responses being observed with 2 mg/animal poly ICLC (p = 0.002) or 6 mg/animal poly I:C12U (p = 0.001) when compared with immunization with KLH alone. Notably, poly ICLC—but not CpG-C given at the same dose—also helped to induce HPV16-specific Th1 immune responses while both adjuvants supported the induction of strong anti-HPV16 L1 antibody responses as determined by ELISA and neutralization assay. In contrast, control animals injected with HPV16 capsomeres alone did not develop substantial HPV16-specific immune responses. Injection of dsRNA led to increased numbers of cells producing the T cell–activating chemokines CXCL9 and CXCL10 as detected by in situ hybridization in draining lymph nodes 18 hours after injections, and to increased serum levels of CXCL10 (p = 0.01). This was paralleled by the reduced production of the homeostatic T cell–attracting chemokine CCL21. Thus, synthetic dsRNAs induce an innate chemokine response and act as adjuvants for virus-specific Th1 and humoral immune responses in nonhuman primates.
| Novel adjuvants that facilitate the induction of strong cellular immunity could be of help in the design of vaccine strategies to combat infections such as HIV or tuberculosis. Our immune cells possess archaic receptors recognizing structures of infectious pathogens, and the interaction of these receptors with their ligands results in an activation of the immune system. Here we exploited synthetic forms of one of these ligands, i.e., dsRNA, to define an adjuvant for the induction of cellular immune responses in primates. We injected model and viral proteins together with three different forms of dsRNA subcutaneously (s.c.) in rhesus macaques, and all compounds served as adjuvants for the induction of cellular immunity without the incidence of major side effects. These adjuvant effects depended on the adjuvant dose and coincided with profound alterations in the chemokine production in the draining lymph nodes. dsRNA also helped to induce cellular and humoral immune responses against capsomeres of low immunogenicity derived from the human papillomavirus 16, the causative agent in about 50% of all cases of cervical cancer worldwide. Therefore, formulations involving synthetic dsRNA are promising candidates for development of novel vaccines.
| Effective vaccines against infections caused by intracellular pathogens including HIV infection, malaria, or tuberculosis most likely will need to induce strong cellular and humoral immune responses [1]. Current vaccine strategies under development are based on prime-boost immunizations, such as vaccination with plasmid DNA followed by booster injections with replication-incompetent viral vectors (e.g., adenoviruses or poxviruses), with both DNA and viruses encoding immunogenic proteins of the pathogen [2]. There is concern that these strategies may be insufficiently immunogenic and protective, so alternative vaccine approaches are under development [3],[4]. While protein based vaccines allow the delivery of large amounts of immunogenic vaccine antigens, particularly when targeted to antigen presenting dendritic cells (DCs) [5], these vaccines require the identification of appropriate adjuvants [6], which may act by differentiating the DCs to elicit strong immunity [7]–[10]. Monkeys are being used as an animal model to develop AIDS vaccines and are likely to be a valuable preclinical model to identify adjuvants and understand their mode of action.
Currently the most widely used adjuvant is aluminum hydroxide. It predominantly induces Th2 immune responses [11], and as such may be inappropriate for HIV or tuberculosis vaccines or for immune therapy of tumors related to infection by human papillomaviruses (HPV). Ligands for pathogen recognition receptors, e.g., Toll-like receptor (TLR) ligands, can stimulate cells of the innate and adaptive immune systems and have therefore been proposed as promising adjuvant candidates [12],[13]. We have previously studied the effects of TLR9 ligands, i.e., CpG-A and CpG-B, on the induction of protein-specific immune responses in nonhuman primates. However, we did not observe strong CD4+ T cell-mediated immune responses as indicated by T cell proliferative assays [14]. This may in part be due to the lack of TLR9 expression in myeloid primate DCs [15], which can be valuable for the priming of naïve T cells and the induction of cellular immune responses [16],[17]. In this study, we have focused on synthetic double stranded RNA (dsRNA) compounds as adjuvants. They can be recognized by both TLR3 [18] and the melanoma differentiation-associated gene-5 (MDA-5) [19], pattern recognition receptors that are expressed by many cell types and are involved in anti-viral immune responses [20].
In mice, polyriboinosinic-polyribocytoidylic acid (poly I:C) has long been known as a strong IFN-α inducer and provides anti-viral and adjuvant activity [21],[22]. Poly I:C also works as a mucosal adjuvant for the induction of humoral and cell-mediated immune responses [23]–[25]. MDA-5 is important for the IFN response induced by poly I:C [26],[27].
In primates, poly I:C is a less effective IFN-α inducer, most likely due to nucleases, which reduce the biostability of poly I:C and are reported to be more prevalent in the serum of primates than rodents [28]. A complex of poly I:C with poly-L-lysine and carboxymethylcellulose (poly ICLC), however, is five to 10 times more resistant to hydrolysis by RNAse in primate serum than the parent poly I:C and induces significant levels of interferon in monkeys under conditions in which poly I:C itself induces no interferon [29],[30]. Poly ICLC possesses anti-viral activity against a variety of viruses in monkeys [31]–[33] and chimpanzees [34], and also inhibits malaria infection of macaques [35]. Furthermore, it has shown potent adjuvant activity on the induction of humoral immune responses in the nonhuman primate models of Venezuelan equine encephalomyelitis virus and swine influenza virus [36],[37]. In humans, dose-dependently, mild to moderate side effects of poly ICLC were observed in a number of phase I and II studies conducted in children and adults [38]–[45]. Another synthetic dsRNA, poly I:C12U (Ampligen), supports the induction of broad antiviral immune responses in mice [46],[47], shows low toxicity in humans [48], and should therefore also be considered as an adjuvant in human vaccine trials. To date, no studies have been reported on the potential of synthetic dsRNA to augment cellular immunity in primates.
We therefore have performed studies in rhesus macaques to address the impact of dsRNA on the induction of protein-specific immune responses. As a prelude to studies with protein based vaccines, we selected keyhole limpet hemocyanin (KLH). In contrast to a previous study where TLR7/8 and TLR9 ligands have been used as adjuvants for cellular immunity in rhesus macaques [49], we injected the dsRNA plus KLH in aqueous solution without additional emulsification in water-in-oil adjuvants, such as Montanide, to minimize the risk of undesired side-effects at the site of injection. To confirm that the adjuvant effect of dsRNA is also manifest in the context of the injection of viral proteins, we injected some animals with the major capsid protein (L1) of HPV16 with or without poly ICLC. HPV16 is the major carcinogenic genotype of HPV in most countries and involved in about 50% of the cases of cervical cancer worldwide [50]. Recently, prophylactic vaccines against HPV16 have been marketed that consist of L1 virus-like particles (VLPs) and induce neutralizing antibodies that efficiently protect against persistent HPV infection and premalignant cervical lesions [51]. However, therapeutic vaccines for the use in individuals who are already infected will need to induce cellular immunity, most likely against the E6/E7 antigens of HPV. Subunits of VLPs (pentameric capsomeres) have potential advantages over VLPs, i.e., higher stability and reduced production costs but their immunogenicity has not yet been evaluated in nonhuman primates.
To monitor also the innate response to dsRNAs, we concentrated on the rapid innate production of CXCL9 (MIG) and CXCL10 (IP-10) chemokines, which are induced by dsRNA [52] as well as CCL21 (SLC), which attracts naïve T lymphocytes and DCs [53]. Here we show that dsRNAs act as adjuvants for the induction of innate and adaptive cellular and humoral immunity in nonhuman primates.
Poly ICLC has adjuvant activity on the induction of humoral immunity at doses as low as 0.1 mg/kg [37]. Since we assumed that higher doses might be required for the induction of cellular immune responses, we immunized rhesus macaques subcutaneously (s.c.) with KLH and either poly ICLC (0.5 mg/kg body weight; 6 animals), poly I:C (0.5 mg/kg; 4 animals), or without adjuvant (4 animals). To monitor the development of T cell immunity, we cultured peripheral blood mononuclear cells (PBMC) with or without KLH and determined whether immunization resulted in T cell proliferative responses to the administered antigen by 3H thymidine incorporation assays. Peak stimulation indices (SI; KLH-induced proliferation divided by proliferation in medium alone) were significantly higher (p = 0.040) in animals injected with poly ICLC (week 0, 1.93±1.38; peak, 23.00±8.02) compared with KLH alone (week 0, 3.28±1.55; peak, 8.97±7.47), individual maximum proliferative responses are shown in Table S1. The kinetics of the responses are shown in the Figure S1 and reveal significantly stronger proliferative responses in poly I:C co-injected animals than in controls six weeks post injection (p = 0.03). Thus, KLH-dependent proliferation of PBMCs was induced with poly ICLC or poly I:C by the s.c. route.
The poly I:C analogue poly I:C12U requires TLR3 to be active in vivo [54],[55] and shows little toxicity in humans [48]. Like poly I:C and in contrast to poly ICLC, it is not stabilized against primate serum nucleases. We therefore compared the effectiveness of poly I:C12U to poly ICLC in a separate study, using fixed standardized doses per animal rather than adjustment to body weight. To study the effects of dsRNA on cellular immune responses in more detail, we used a carboxyfluorescein diacetate succinimidyl ester (CFSE) dilution assay, which allows the separate evaluation of CD4+ and CD8+ T cells (Figure 1A). KLH at 200 µg/animal was administered to five animals each either alone or together with 2 or 6 mg poly I:C12U or 2 mg poly ICLC per animal (i.e., 0.27 to 0.44, 1.00 to 1.13, and 0.33 to 0.43 mg/kg, respectively). The dose of 2 mg per animal has been used in previous studies on TLR agonists as vaccine adjuvants in monkeys [49] and thus facilitates comparison between studies of different adjuvant compounds. KLH-specific CD4+ T-cell proliferation at week 2 after immunization was significantly higher when KLH was given with either 6 mg poly I:C12U (p = 0.001) or 2 mg poly ICLC (p = 0.002) whereas at that time point no significant difference to KLH alone was observed after the injection of 2 mg poly I:C12U (p = 0.16; Figure 1B). The effect of immunization with poly I:C12U or poly ICLC on proliferative responses was sustained over 6 weeks, and there was a significant difference also for the 2 mg poly I:C12U group over KLH alone at this time point (p = 0.013; Figure 1B). Thus, all three synthetic dsRNA compounds that we tested could serve as adjuvants for the induction of protein-specific T-cell proliferation in primates.
To examine possible dose-dependent adjuvant effects of poly ICLC, we compared in a prime-boost experiment the effects of 0.5 mg/kg body weight with those of 0.1 mg/kg body weight, which is still sufficient for the induction of humoral immune responses [37]. After the first immunization increased KLH-specific CD3+CD4+ T cell proliferative responses were seen in both animals immunized with 0.5 mg poly ICLC/kg, and CD3+CD4− T cells (as a surrogate for CD8+ T cells) were expanded to a similar extent (Fig. S2A). CD4+ and CD4− T cell-proliferative responses were less pronounced after the primary immunization together with the lower 0.1 mg/kg dose of poly ICLC (Figure S2B). Booster immunization at week 14 enhanced the proliferative CD4+ T cell responses in the animal 13404 immunized with 0.5 mg poly ICLC/kg (Figure S2A) and in the animal 13406 receiving the lower dose of poly ICLC (Figure S2B). Therefore, 0.5 mg/kg of poly ICLC might be more active as an adjuvant for cellular immunity than lower doses.
To confirm that dsRNA analogues also serve as adjuvants in the context of a clinically relevant viral antigen, we injected s.c. six animals each with a low dose of HPV16 L1 capsomeres (10 µg) with or without 2 mg of poly ICLC. Another six animals were injected with 2 mg of the TLR9 ligand CpG-C (ODN 2396), which supports the induction of protein-specific cellular immune responses in monkeys when injected in a water-in-oil emulsion [49]. We selected the L1 pentamers rather than the complete virus-like particles (VLP; 360 molecules of L1), since capsomeres are promising candidates for 2nd generation vaccines but their immunogenicity in nonhuman primates has not yet been evaluated. The capsomeres were obtained by expression of a modified L1 protein in baculovirus-infected insect cells [56]. In the immune assays, we re-stimulated PBMCs with HPV16 VLPs and used mouse norovirus VLPs (A447) generated in the same expression system as a negative control antigen. All animals were boosted with a second injection of antigen +/− adjuvant eight weeks later.
Numbers of IFN-γ secreting cells in the peripheral blood were determined by ELISPOT assay. At week 2, we detected increased numbers of HPV-specific, IFN-γ secreting cells in PBMCs from 4 out of 6 animals (Figure 2A) injected with antigen plus poly ICLC, and the responses waned in all animals by week four. Two weeks after the booster injection (at week 10 after the first injections), however, all six animals injected with antigen together with poly ICLC had developed HPV-specific, IFN-γ secreting cells, which also were maintained two weeks later (week 12) and still present in 3 animals at week 19. In contrast, IFN-γ secreting cells were detectable at elevated numbers in only one of the CpG-injected monkeys (animal 13928) four weeks after the first injection and this response could not be boosted by the second injection. None of the control animals showed substantial numbers of HPV-specific IFN-γ secreting cells, neither following the first nor the booster injection (Figure 2A). The background responses against A447 might be induced by contaminating protein fractions derived from the expression system, in which both antigens, i.e., HSV16 L1 capsomeres and A447, had been generated.
When we assessed T cell proliferation in CFSE assays, we found significantly enhanced HPV-specific CD4+ T-cell proliferative responses in the poly ICLC-injected monkeys four weeks after the second application of antigen (p = 0.008) (Figure 2B). Figure 2B depicts the proliferation of CD3+CD8− cells, re-stimulated for the last 6 h of the assay with peptide pools 1–4. Similar results (p = 0.012, week 12) were obtained with cells re-stimulated for the final 6 h with pools 5–8 (data not shown). At week 19, proliferative responses did not differ significantly in poly ICLC-injected and control animals.
To further characterize the Th cell responses, we determined the concentrations of IFN-γ, IL-4, and IL-17 in supernatants collected from re-stimulated PBMCs 2 d after setting up the assays. We used ELISAs for the detection of monkey cytokines or, in the case of IL-17 an ELISA for the detection of the human protein but known to cross-react with monkey IL-17 [57]. Following the booster injection at week 8, we detected significantly more IFN-γ in the supernatants of cells collected from poly ICLC-injected animals than in those of cells from control animals and these responses were sustained until week 19 (Figure 2C). In contrast, we were unable to detect IL-4 or IL-17 in the supernatants from assays set up with PBMCs from either group of animals. Thus, poly ICLC supports the induction of HPV-specific Th1 immune responses, i.e., CD4+ T cell proliferative responses and IFN-γ secretion.
We also determined the humoral immune responses induced by the injection of HPV16 L1 capsomeres with or without adjuvants. Injection of poly ICLC or CpG-C resulted in up to 1000fold increased titers of binding antibodies (measured by ELISA) compared with control animals (Figure 3A; p<0.01 for both adjuvants for weeks 4, 8, 10, and 12), and at weeks 4, 8, and 10, poly ICLC also induced higher titers than an equal dose of CpG-C (p<0.05 for week 4, p<0.01 for weeks 8 and 10). The individual antibody titers of all animals are shown for all points in time in the Table S2. In addition, we performed neutralization assays using the serum samples collected 12 weeks after first immunization and HPV16 pseudovirions as targets. Sera of the animals from both adjuvant groups showed considerable neutralizing activity while samples from the control animals were not able to neutralize the activity of the pseudovirions in our assay (Figure 3B). Poly ICLC injected animals showed stronger responses than monkeys that had received CpG-C (p = 0.03 for serum dilutions of 1∶1000). There was a good correlation between ELISA and neutralization titers in the sera of the individual animals (Figure S3). Therefore, while CpG-C mainly affects the induction of antibodies, poly ICLC acts as adjuvant for both humoral and cellular immunity.
Since we have previously observed that poly I:C activates monkey DCs [58], immunohistochemistry was performed to determine the number and activation status of DCs present in lymph nodes taken prior to immunization and at 18 h after injection of poly ICLC. The numbers of phenotypically immature (CD1a+) and mature (CD83+ or CD208+) DCs varied between animals but did not show a clear decrease or increase after immunization (Figure S4).
Draining inguinal lymph nodes were also analyzed for CXCL10, CXCL9, and CCL21 by in-situ hybridization and immunohistochemistry. In comparison to control lymph nodes removed before immunizations, elevated expression of CXCL10 (Figure 4A and 4B) and CXCL9 (Figure 4C and 4D) was detected in the T cell areas of draining lymph nodes at 18 hours after immunization. Chemokine mRNA expression correlated with protein expression detected by immunohistochemistry (insets in Figure 4). Expression of CCL21 mRNA (Figure 5A) and protein (Figure 5B and 5C) 18 hours after immunization was markedly decreased in draining lymph nodes compared with control lymph nodes obtained before immunization. Thus, the innate response to dsRNA is detectable in lymph node cells.
The administration of poly ICLC or poly I:C together with KLH led to a significant increase of serum levels of CXCL10 (Figure 6A; p = 0.001 for both compounds at 18 or 24 h). Furthermore, 48 h after immunization, serum levels of CXCL10 were significantly higher in poly ICLC- than in poly I:C-injected monkeys (p = 0.027). Like poly I:C at 0.5 mg/kg, poly I:C12U or lower doses of poly ICLC (0.1 mg/kg) induced increased CXCL10 levels, which were less sustained (Figure 6B and 6C).
We detected increased CXCL9 serum levels in animals injected with poly I:C or poly ICLC (0.5 mg/kg), and there were minor changes in CXCL9 concentrations in monkeys receiving KLH alone (Figure 6D). No changes of CXCL9 serum concentrations were observed when 0.1 mg/kg poly ICLC were administered (data not shown). At 6, 24, or 48 h after infection, no significant differences in serum levels of IFN-α, IFN-γ, TNF, IL-12p40, and CCL3 (MIP-1α) were observed between groups receiving KLH alone or together with dsRNA (data not shown). We were not able to detect considerable serum concentrations of IFN-α at any point in time including 1 h post injection.
Since immunohistochemistry and in-situ hybridization revealed that CXCL10 was mainly produced in the T cell-areas of the draining lymph nodes (Figure 4), we considered DCs as a potential source for this chemokine in vivo. Unfortunately, double-labeling with DC identifying mAbs was not possible on formalin-fixed specimens. We therefore tested whether dsRNA may directly induce CXCL10 secretion by highly purified rhesus macaque DCs in vitro. When monocyte-derived monkey DCs were incubated with poly ICLC at two different concentrations (50 and 200 µg/ml), significantly elevated CXCL10 concentrations were detectable 48 h later in the cell culture supernatants (p = 0.002 compared to un-stimulated controls), and both doses of of poly ICLC induced comparable levels of CXCL10 (Figure 7). Thus, primate DCs produce CXCL10 upon stimulation with synthetic dsRNA, making DCs one of the candidate sources of CXCL10 observed in the draining lymph nodes.
This study shows that s.c. injection of synthetic dsRNA, i.e., poly I:C, poly ICLC, or poly I:C12U supports the induction of cellular immune responses to protein antigens in nonhuman primates. These responses could also be boosted by a second injection of antigen together with dsRNA. We observed antigen-specific T cell proliferation of CD3+CD4+ and CD3+CD4− T cells. High but nontoxic doses (toxicity starts in M. mulatta at i.v. doses >2 mg/kg, i.m. or s.c. injections are better tolerated than i.v. injections; unpublished observations) of poly ICLC (0.5 mg/kg or 2 mg/animal) might be more potent than lower doses (≤0.1 mg/kg). Using HPV16 capsomeres at low doses (10 µg/animal) as a relevant viral antigen with low immunogenicity, we also showed that poly ICLC, but not CpG-C (which supported the induction of humoral responses, however), supports the induction of HPV16-specific Th1 responses. The lack of effect of CpG-C in our system compared to other studies where the same compound helped to elicit cellular immunity in nonhuman primates is most likely due to the fact that we injected the antigens in PBS, while others injected CpG-C and antigens in the synthetic water-in-oil emulsion, Montanide [49]. Amongst the three different formulations of synthetic dsRNA, poly ICLC appears to possess the most potent adjuvant activity on the induction of cellular immune responses. Subsequent studies will show whether it will help to induce protective immune responses against other pathogens, e.g., SIV.
Both adjuvants supported the induction of humoral immune responses, including neutralizing antibodies. Therefore, subsequent in vivo studies should compare poly ICLC with the adjuvants currently used in vaccine formulations, e.g., alum, and investigate whether its co-application might allow fewer injections than required today for the currently licensed vaccine formulations.
In order to understand the activity of dsRNA, we examined the innate response since this includes events that can improve the function of antigen presenting DCs and T cells. Surprisingly, we did not detect the expected increase of serum IFN-α shortly after injection of poly I:C or poly ICLC. This might be due to the s.c. route of injection. While i.v. injections of poly ICLC give rise to high serum interferon levels [30], the s.c. application of dsRNA may lead to a more protracted release from the site of injection and a delayed bioavailability. In mice, type I interferon induced by poly I:C has been shown to be essential for its adjuvant effect on humoral immunity and isotype switching [59], and it also seems essential for TLR3-mediated cross-priming of CD8+ T cells [60]–[62]. Likewise, type I interferon is critical for the CD8+ T cell expansion induced by TLR agonists in combination with CD40 [63]. Poly I:C and poly ICLC induce proliferation of CD8+ T cells, both have been shown to be effective as an adjuvant for the induction of specific CD8+ T cell responses in mice [64]–[66], and this effect partially depends on NK cells [67]. Thus, poly I:C, and most likely also poly ICLC, support the induction of CD8+ T cell responses, and the KLH-specific responses expressed by CD3+CD4− T cells observed by us might reflect true CD8 responses.
In contrast to our inability to detect IFN-α in the serum in response to dsRNA, we did detect enhanced levels of CXCL10. These were sustained over 48 hours in animals injected with 0.5 mg/kg poly ICLC but decreased more rapidly in monkeys following injection of lower concentrations of poly ICLC, 0.5 mg/kg poly I:C, or a comparable dose (2 mg/animal) of poly I:C12U. This may reflect the reduced biostability of the nonstabilized poly I:C and poly I:C12U compared with that of poly ICLC as described before [29],[30]. CXCL10 is known for its activity to attract effector Th1 cells through interaction with its receptor CXCR3 at sites for the expression of Th1 immune responses [68], e.g., rejection of allografts or the inflammatory response upon mycobacterial infection [69],[70]. CXCL10 is also required for resistance to protozoan or viral pathogens [71],[72]. Studies in mice revealed additionally that CXCL10 is secreted early (e.g., earlier than CXCL9, which we did not detect at the same levels in the serum as CXCL10) [73], and stimulates T cell proliferation [74]. In fact, CXCL10-deficient mice have impaired T cell responses following primary immunization with exogeneous protein antigen indicating a role for CXCL10 in effector T cell generation [75]. Since CXCR3 also is induced early in CD4 T lymphocyte differentiation [76], the literature suggests an enhancing role for CXCL10 in both the expression and induction of Th1 immune responses. Notably, while we have previously detected an increase in serum CXCL10 after injection of CpG-A or CpG-B [14], these concentrations were around ten-fold lower than in animals injected with 0.5 mg/kg poly ICLC. Since both the two forms of CpGs and low doses of poly ICLC had only marginal adjuvant effects on the induction of cellular immunity, high and sustained serum levels of CXCL10 after injection of an adjuvant seem to be indicative of its ability to support the induction of cellular immune responses. Interestingly, the Th2-adjuvant alum considerably inhibits TLR-induced production of CXCL10 [77].
Expression of CXCL9 and CXCL10 was primarily in the T cell areas of the draining lymph node. Thus, DCs should be considered as a potential source of these chemokines, since they are abundant in this area of the lymph node. We show that monocyte-derived DCs produce CXCL10 upon activation with dsRNA, which suggests a direct role of these cells in the production of the pro-inflammatory chemokines and induction of cellular immune responses in our system. Monkey DCs express TLR3 (manuscript in preparation) and can be activated by poly I:C [58], so pattern recognition receptors on DCs likely contribute to the observed adjuvant effects of dsRNA for CD4+ T-cell proliferation. Synthetic dsRNA, however, may also target and activate other TLR3+ or TLR3− leukocyte subsets. In vitro, it activates human NK cells [78],[79], γ/δ TCR+ T cells [80], CD8+ α/β TCR+ T cells [81], and also monocytes/macrophages, which are TLR3− [82],[83]. These cells (or the corresponding cells in lymphoid tissues) could contribute to its adjuvant activity, e.g., through the secretion of pro-inflammatory cytokines and notably type I and II interferons. While it remains to be determined whether dsRNA can promote survival of primate CD4+ T cells as recently shown for murine cells [84], analyses of human blood leukocytes shortly after poly ICLC injection revealed increased percentages of CD4+ T cells, but also effects on the activity of NK cells and the frequency of HLA-DR+ cells [85]. Nevertheless, cells other than leukocytes including keratinocytes and neurons also can produce type I interferons and other pro-inflammatory cytokines upon stimulation with poly I:C [86],[87].
After injection of dsRNA, we observed a down-regulation of the homeostatic chemokine CCL21, which attracts CCR7+ cells, such as DCs and naïve T cells, to lymph nodes. In agreement with the advuvant effect of poly ICLC on the induction of HPV-specific Th1 immune responses shown in the present study, this process has recently been described for the early phase of the induction of Th1 but not Th2 immune responses in mice and is controlled by the production of IFN-γ [88]. This is mirrored by our findings using HPV16 capsomeres as viral protein antigen with relevance to the human system. Animals injected with HPV together with poly ICLC developed Th1 immune responses characterized by antigen-specific T cell proliferation and IFN-γ secretion in the absence of detectable IL-4 or IL-17 production.
In conclusion, dsRNA compounds induce the innate production of CXCL10 in the draining lymph nodes and high CXCL10 concentrations in the serum early after injection, and these compounds are effective adjuvants for the induction of adaptive pathogen-specific T cell and humoral immune responses.
Healthy young adult male and female rhesus macaques (Macaca mulatta) housed at the German Primate Center (Deutsches Primatenzentrum, Göttingen, Germany) were used. The animals were antibody negative for simian T-lymphotropic virus type 1, simian D-type retrovirus, and simian immunodeficiency virus. All animal care operations were in compliance with the guidelines of the German Primate Center and approved by the local authorities. For immunizations and collection of blood samples animals were sedated with ketamine. For the removal of lymph nodes, a deeper anesthesia consisting of a mixture of xylazine, atropine, and ketamine was used. 200 µg endotoxin-free KLH (Calbiochem, San Diego, CA, USA) or 10 µg HPV 16 capsomeres alone or in combination with either poly I:C (Invivogen/Cayla, Toulouse, France), poly ICLC (Hiltonol, Oncovir, Washington, D.C.), poly I:C12U (Ampligen, Celldex Therapeutics, Bloomsbury, NJ, USA), or CpG-C (ODN 2396, generously provided by Coley Pharmaceutical Group, Wellesley, MA, USA) were administered bilaterally s.c. at doses indicated in the text at volumes between 1.0 and 2.0 ml, partially diluted in PBS, by injecting close to the inguinal lymph nodes. All animals remained well following the application of dsRNA plus antigen and no local signs of inflammation apart from transient lymph node swellings and mils hyperemia were observed at sites of injection.
Blood samples were drawn at 0, 1 or 6, 18 or 24, and 48 h after injections for measurements of serum cytokine and chemokine concentrations. To determine humoral and cellular immune responses additional blood samples were drawn at points in time indicated. Axillary lymph nodes were removed before the immunizations, and 18 h after the injections one draining lymph node from each immunized animal was removed. Lymph nodes were divided in two parts. One part was fixed in 4% neutral-buffered formalin overnight and embedded in paraffin. The other halves were embedded in tissue-freezing medium (Leica, Nussloch, Germany), snap-frozen in liquid nitrogen, and stored at −70°C until use.
HPV16 L1 capsomeres were produced using recombinant baculoviruses containing the mutated L1 (L1_2xCysM: C175A+C428A) as described previously [56]. In short, High Five insect cells (Invitrogen, Germany) were infected with recombinant baculoviruses and harvested by centrifugation. Proteins were extracted by sonification from cell pellets resupended in 20 ml of extraction buffer (5 mM MgCl2, 5 mM CaCl2, 1 M NaCl, 0.01% Triton ×100, 20 mM Hepes pH 7.4 and 1 mM PMSF). The cleared lysate was loaded on a two-step gradient consisting of (30% w/v) sucrose and CsCl (58% w/v), followed by a centrifugation at 96,500 g at 10°C for 3 h in a SW32 rotor (Beckman Ultracentrifuge). The interphase between the sucrose and CsCl and the complete CsCl layer were centrifuged again in Quickseal tubes (Beckman, USA) for 16–18 h at 20°C at 184,000 g in a Sorval TFT 65.13 rotor. Fractions of 1 ml fractions were collected and the L1 containing determined by antigen-capture ELISA and western blot analysis and the structure of the particles was characterized by electron microscopy [89]. The control antigen (mouse norovirus VP1 VLPs) were generated by the identical protocol. The VP1 clone was kindly provided by W. Nicklas, DKFZ Heidelberg.
Standard proliferation assays were set up with 1×105 PBMCs/well in 96-well round-bottom trays (Nunc) with KLH (100 µg/ml) in cell culture medium consisting of RPMI 1640, supplemented with 2 mM L-glutamine, penicillin (100 U/ml)-streptomycin (100 µg/ml), 10 mM HEPES (all GIBCO, Invitrogen, Karlsruhe, Germany), 50 µM 2-mercaptoethanol (Sigma), and 10% heat-inactivated FCS (Biochrom, Berlin, Germany). Controls included PBMCs in medium alone and PBMCs stimulated with 5 ng/ml staphylococcal enterotoxin B (SEB; Alexis Corp., Lausen, Switzerland). All conditions were set up in triplicates and cultures were incubated at 37°C and 5% CO2. Supernatants were harvested on day 2 and frozen at −80°C for analyses of cytokine concentrations. 3H-thymidine (1 µCi/well, NEN, Perkin Elmer, Boston, MA, USA) was added to the wells on day 3 (for SEB and medium alone) or day 5 (KLH and medium alone). Cells were harvested 24 h later onto glass fibre filter mats (ICN Biomedicals, Aurora, OH, USA), and incorporated 3H-thymidine was measured in a liquid scintillation counter. To facilitate the comparison of proliferative responses, SIs were calculated by dividing the mean counts per minute (cpm) of triplicates of antigen-containing wells by the mean cpm of triplicate wells with unstimulated PBMCs.
Additionally, CFSE (Invitrogen/Molecular Probes, Karlsruhe, Germany) assays were used to determine proliferation. PBMCs at 1×107 cells/ml were stained with 0.25 µM CFSE in pre-warmed PBS for 15 min at 37°C, washed in medium, incubated in pre-warmed medium for another 30 min, and washed again. The cells were then adjusted to 1×106 cells/ml and cultured in medium with or without SEB or KLH as described above or and incubated for 6 to 7 days. Alternatively, cells were incubated at 1.25 µg/ml with HPV16 VLPs or an unrelated control antigen, i.e., mouse norovirus VLPs similarly produced as the HPV antigen (A447), at the same dose. After 7 days cells were harvested and washed in PBS/5% FCS/0.05% sodium azide, stained with anti-CD3 PerCP- and anti-CD4 APC-conjugated mAbs, washed, and fixed. T cell proliferation was assessed as the percentage of CFSElow cells, gating on live CD3+CD4+ or CD3+CD4− cells (Figure 1A). Alternatively, cells were re-stimulated with eight pools of HPV16-specific, 15mer peptides (124 peptides, pool 1–4 with 16 peptides each, pool 5–8 with 15 peptides each), 2 µg/ml SEB, or medium alone in the presence of 1 µg/ml co-stimulatory mAbs CD28 and CD49d (BD Pharmingen) for 6 h, and Brefeldin A (Sigma) was added at a final concentration of 10 µg/ml for the last 4.5 h. Cells were then washed in PBS/5% FCS/0.05% sodium azide, stained with anti-CD3 PerCP- and anti-CD8 APC-conjugated mAbs, washed, fixed with 4% paraformaldehyde, and stained with PE-conjugated mAbs against IFN-γ after cell permeabilization with 0.5% saponin in PBS/5% FCS/0.05% sodium azide. T cell proliferation was assessed as the percentage of CFSElow cells, gating on live CD3+CD8+ or CD3+CD8− cells, and IFN-γ secretion was measured as the percentage of PE-stained, CFSElow cells in the gated cell populations.
ELISPOT assays were preformed using commercially available reagents (Mabtech AB, Hamburg, Germany) as previously described [90]. Briefly, PBMCs were resuspended in culture medium and seeded at 1×105 cells/well in 96-well plates (MAIP S4510, Millipore, Schwalbach, Germany), which had been coated with 1 µg/well of anti-human IFN-γ monoclonal antibody overnight at 4°C. For antigen stimulation, HPV16 L1 protein or control antigen (A447) was added at 1.25 µg/ml to the wells in triplicates. Positive and negative controls consisted of cells stimulated by SEB (1 µg/ml, Sigma) and cells kept in medium alone. After 20 h of incubation at 37°C in 5% CO2, cells were removed and biotinylated anti-human IFN-γ detector antibody was added (0.1 µg/well), followed by the addition of streptavidin-alkaline phosphatase conjugate at 1∶1000 in PBS/0.1% FBS. Spots were developed with NBT/BCIP solution (25 µg NBT and 15 µg BCIP in 0.1 M Tris–HCl pH 9.5 per well) for 30 min, the wells were washed with distilled water and air-dried, and spots were counted using a BIOSYS2000 ELISPOT reader. The counts were extrapolated to 106 PBMCs. Average spot numbers of background responses (to A447) plus twice the standard deviation were considered positive responses.
The presence of L1-specific IgG antibodies in plasma samples from immunized monkeys was determined by VLP-ELISA as described earlier [56]. Briefly, 96-well plastic plates were coated overnight at 4°C with VLP produced in baculovirus infected High Five insect cells and purified according to a previously published method [89]. After washing with PBS-T, plates were blocked with MPBS-T (5% skim milk in PBS- 0–05% Tween) for 1 hr at 37°C. Prediluted sera (in two-fold dilutions starting from 1∶50 to 1∶819,200) were added, and plates were incubated for 1 hr at 37°C. After washing, plates were incubated for 1 hr at 37°C with 1∶2000 diluted HRP-coupled antihuman IgG F(ab')2 secondary antibody (Dianova, Germany) in MPBS-T, TMB (3,3′,5,5 -tetramethylbenzidine) substrate solution (Sigma, Germany) was used as substrate. OD was measured in an ELISA reader at 450 nm after 10 min and 30 min incubation at room temperature. Nonspecific binding was determined by using the same dilutions on plates coated with extracts of High Five cells infected with wt baculovirus. IgG titers were expressed as the reciprocal of the highest dilution giving an absorbance above the cut off value (the average of the negative controls plus three times standard deviation).
Pseudovirions were prepared by transfecting 293TT cells (cultivated in DMEM containing 50 µg of hygromycin/ml) with a plasmid coding for the humanized HPV16 L1 and L2 genes, together with a plasmid containing the gene for secretable alkaline phosphatase (SEAP) under the control of the CMV promoter. For pseudovirion extraction, cells were harvested 3–4 days later by trypsination, washed once with PBS and resuspended in 1 ml PBS containing 1 mM CaCl2 and 5.6 mM MgCl2 per 5×107 cells and lysed by 50 µl Brij58 (Sigma) in the presence of Benzonase (250 U/ml) for 5 min on ice. The cellular lysate was centrifuged after the addition of NaCl to a final concentration of 710 mM, and the cleared supernatant containing the pseudovirions was used for infection of 293TT cells. For this purpose, pseudovirions were diluted 1∶5000 in DMEM and preincubated with the sera (dilution from 1∶50 to 1∶100,000) for 15 min at room temperature. Pseudovirions were then added to the cells, followed by incubation at 37°C for 5 days. Detection of SEAP activity in cell culture supernatant was measured by using a commercial assay (Roche, Mannheim, Germany) according to the manufacturer's recommendations.
Chemokine and cytokine concentrations in serum or plasma samples and cell culture supernatants were measured using ELISA kits for human CXCL10, CXCL9 (both R&D Systems, Wiesbaden, Germany), CCL3 (Antigenix America, Huntington, NY, USA), IFN-α (PBL, Brunswick, USA) [14], IL-17 (eBioscience, NatuTec, Frankfurt/Main, Germany) [57], and human TNF as well as monkey IFN-γ, IL-4, and IL-12p40 (all U-Cytech, Utrecht, The Netherlands).
Cryostat sections were cut, fixed in acetone for 30 min and incubated with monoclonal antibodies against human CD1a (dilution: 1∶100; Medac, Hamburg, Germany), CD83 (dilution: 1∶100) or CD208 (1∶70; both Immunotech, Hamburg Germany). Antibody binding was visualized by the alkaline phosphatase anti-alkaline phosphatase method using New Fuchsin as chromogen. The sections were counterstained with hemalaun and mounted. The numbers of DCs were quantified with a Zeiss AxioImager M1 microscope (Carl Zeiss, Jena, Germany). Using a 40× objective, a standard area was set (unit area). Ten non-overlapping unit areas were selected. The positive cells were counted using AxioVision (Release 4.6) software (Zeiss). The values were averaged to represent the numbers of positive cells per unit area. Due to inadequate immunohistochemical staining the draining lymph node from animal number 13408 was omitted from the examination.
Immunohistochemistry on paraffin sections was performed as previously described [92]. The following antibodies diluted in antibody diluent (S3022, DAKO, Glostrup, Denmark) were used: mouse anti-CXCL10 (MAB266, R&D Systems, 1 µg/ml), goat anti-CXCL9 (AF392, R&D Systems, 1 µg/ml), and goat anti-CCL21 (AF366, R&D Systems, 1 µg/ml). After over night incubation, sections were washed and incubated with rabbit anti-mouse (E0413, DAKO) or rabbit anti-goat (E0466, DAKO,) biotinylated antibodies followed by streptavidin-alkalyne phosphatase complex (K0391, DAKO), following the manufacturer's instructions. Positive cells were detected using New Fuchsin (K0698, DAKO) as substrate, and tissue sections counterstained with Meyer's Haematoxylin (1.09249, Merck, Zug, Switzerland).
35S-labeled sense and antisense CXCL9, CXCL10, and CCL21 mRNA probes, 411 bp in length corresponding to position 26 to 437 of the CXCL9 sequence (NM_002416), 372 bp corresponding to position 28 to 400 of the CXCL10 sequence (NM_001565), and 367 bp corresponding to position 27 to 394 of the CCL21 sequence (NM_002989), respectively, were generated by in vitro transcription (Roche Molecular Biochemicals, Indianapolis, IN). Tissue sections were dewaxed, rehydrated in graded ethanol solutions, and subjected to in situ hybridization, according to a previously described method [93]. Finally, the sections were dipped in photo emulsion NTB-2 (Kodak, Rochester, NY) and exposed in complete darkness for 2 to 4 weeks at 4°C. Development and fixation were performed according to the instructions provided by Kodak, and counterstaining was done with haematoxylin.
Rhesus macaque monocyte-derived DCs were generated from heparinized peripheral blood as previously described [57]. CD14+ monocytes were magnetically separated (Miltenyi Biotec, Bergisch-Gladbach, Germany) and cultured at 1.5–2×106 cells/3 ml in RPMI 1640, supplemented with 5% human AB serum (PAN Biotech, Aidenbach, Germany), human rGM-CSF (1000 U/ml, sargramostim, Leukine, Berlex, Richmond, CA, USA), human rIL-4 (100 U/ml, R&D Systems, Wiesbaden-Nordenstadt, Germany, and L-glutamine, 2-mercaptoethanol, HEPES, and penicillin-streptomycin as described under T cell assays. At day 6, DCs at 1×105/well were stimulated for 48 h with 50 or 200 µg/ml poly ICLC in 96-well round bottom plates. Supernatants were harvested for analysis of cytokine and chemokine secretion.
Data are expressed as means±standard error of the mean (SEM), standard deviation (SD), or median, where appropriate. Statistical significance of differences was determined by Student's t-test or Mann Whitney U-test. Differences were considered statistically significant for p<0.05.
|
10.1371/journal.pgen.1003074 | Recessive Mutations in SPTBN2 Implicate β-III Spectrin in Both Cognitive and Motor Development | β-III spectrin is present in the brain and is known to be important in the function of the cerebellum. Heterozygous mutations in SPTBN2, the gene encoding β-III spectrin, cause Spinocerebellar Ataxia Type 5 (SCA5), an adult-onset, slowly progressive, autosomal-dominant pure cerebellar ataxia. SCA5 is sometimes known as “Lincoln ataxia,” because the largest known family is descended from relatives of the United States President Abraham Lincoln. Using targeted capture and next-generation sequencing, we identified a homozygous stop codon in SPTBN2 in a consanguineous family in which childhood developmental ataxia co-segregates with cognitive impairment. The cognitive impairment could result from mutations in a second gene, but further analysis using whole-genome sequencing combined with SNP array analysis did not reveal any evidence of other mutations. We also examined a mouse knockout of β-III spectrin in which ataxia and progressive degeneration of cerebellar Purkinje cells has been previously reported and found morphological abnormalities in neurons from prefrontal cortex and deficits in object recognition tasks, consistent with the human cognitive phenotype. These data provide the first evidence that β-III spectrin plays an important role in cortical brain development and cognition, in addition to its function in the cerebellum; and we conclude that cognitive impairment is an integral part of this novel recessive ataxic syndrome, Spectrin-associated Autosomal Recessive Cerebellar Ataxia type 1 (SPARCA1). In addition, the identification of SPARCA1 and normal heterozygous carriers of the stop codon in SPTBN2 provides insights into the mechanism of molecular dominance in SCA5 and demonstrates that the cell-specific repertoire of spectrin subunits underlies a novel group of disorders, the neuronal spectrinopathies, which includes SCA5, SPARCA1, and a form of West syndrome.
| β-III spectrin is present in the brain and is known to be important in the function of the cerebellum. Mutations in β-III spectrin cause spinocerebellar ataxia type 5 (SCA5), sometimes called Lincoln ataxia because it was first described in the relatives of United States President Abraham Lincoln. This is generally an adult-onset progressive cerebellar disorder. Recessive mutations have not previously been described in any of the brain spectrins. We identified a homozygous mutation in SPTBN2, which causes a more severe disorder than SCA5, with a developmental cerebellar ataxia, which is present from childhood; in addition there is marked cognitive impairment. We call this novel condition SPARCA1 (Spectrin-associated Autosomal Recessive Cerebellar Ataxia type 1). This condition could be caused by two separate gene mutations; but we show, using a combination of genome-wide mapping, whole-genome sequencing, and detailed behavioural and neuropathological analysis of a β-III spectrin mouse knockout, that both the ataxia and cognitive impairment are caused by the recessive mutations in β-III spectrin. SPARCA1 is one of a family of neuronal spectrinopathies and illustrates the importance of spectrins in brain development and function.
| Spectrins are a diverse family of membrane scaffold proteins. They were originally found in erythrocytes where mutations result in various haemolytic anemias [1], [2]. Spectrins have been identified in the brain [3] but until recently little was known of the effects in humans of brain spectrin mutations. In 2006, heterozygous mutations of the brain spectrin gene SPTBN2, encoding β-III spectrin, were found to cause Spinocerebellar Ataxia Type 5 (SCA5) [4]. SCA5 is an autosomal dominant, slowly progressive, adult onset, pure cerebellar ataxia, which was first identified in a large family who are the descendents of relatives of the US President Abraham Lincoln; SCA5 is therefore sometimes referred to as “Lincoln ataxia” [5], [6], [7]. Two other SCA5 families have been described in the literature, one from France and one from Germany [8], [9].
β-III spectrin is a 2,390 amino acid protein comprising an N terminal domain containing the actin/ARP1 binding site, 17 spectrin repeats, (the latter containing regions which bind the glutamate transporter EAAT4 [10], and ankyrin [11]), and a C terminal domain of uncertain function. β-III spectrin forms antiparallel tetrameric heterodimers with α-II spectrin, encoded by SPTAN1. The tetrameric self-association probably requires the presence of the C terminal β spectrin repeats, B16 and B17, and the N terminal α spectrin repeats, A0 and A1, with absence of these regions highly likely to impair the formation of a functional tetramer [12]. Three heterozygous dominant mutations in SPTBN2 have been reported to cause SCA5: in the US (Lincoln) family a 13 amino acid in-frame deletion (E532_M544del) in the third spectrin repeat, in the French family a small complex in-frame deletion-insertion (L629_R634delinsW), also in the third spectrin repeat, and in the German family a missense mutation (L253P), in the N terminal domain. The mechanism of action of these mutations is not immediately obvious and could be explained by haploinsufficiency, in which the mutant allele is inactive and the normal stoichiometry for tetramer formation is lost, a dominant negative effect which suppresses wild type (wt) function, or a gain of function effect. Several lines of evidence have suggested that a dominant negative effect in SCA5 is most likely. Using targeted gene disruption of mouse β-III spectrin, Perkins et al, reported that homozygous knockout mice (β-III spectrin −/−) had cerebellar ataxia, a progressive loss of cerebellar Purkinje cells and an associated decrease in the Purkinje cell specific glutamate transporter EAAT4 [13]. The β-III spectrin −/− mutant mice lack all full-length β-III spectrin but do express, at a low level, a form of β-III spectrin (∼250 KDa) that lacks most of the actin-binding domain encoded by exons 2–6. The heterozygous mice (β-III spectrin +/−) were reported to be normal. Further work has shown that the L253P (German) missense mutation has a dominant negative effect on wild type function by preventing protein trafficking from the Golgi apparatus [14]. There is evidence also that de novo in-frame mutations in SPTAN1 encoding α-II spectrin have dominant negative effects, causing a form of West Syndrome (infantile epilepsy with developmental delay) [15]. However, although experimental data has strongly suggested that small in-frame mutations or missense mutations in α-II or β-III spectrins have a dominant negative effect, no recessive mutations in spectrins have been found, and such data would lend further strong support for this hypothesis.
Here we report the first description of recessive mutations in SPTBN2 in which there is a severe developmental childhood ataxia but also significant cognitive impairment. The homozygous stop codon c.1881C>A (p.C627X), was identified in three affected individuals from a consanguineous family using targeted capture and next generation sequencing and both the ataxia and cognitive impairment co-segregate with the mutation. However, since more than one mutation can co-segregate, particularly in consanguineous families, we considered whether a second recessive mutation, either homozygous or compound heterozygous, could account for the cognitive impairment. We investigated this using a combination of SNP array analysis and whole genome sequencing, but found no evidence of a second mutation.
We also investigated β-III spectrin −/− knockout mice [13] for supportive evidence that the cognitive impairment in the human subjects is caused by loss of β-III spectrin. We examined the mouse model for morphological abnormalities of neurons in brain regions (other than cerebellum), which are thought to be involved in memory function including prefrontal cortical (PFC) layers, the caudate putamen/striatum and hippocampus (HPC). Finally we tested the mice using object recognition tasks, which have been shown to correlate with function of the PFC and HPC [16], [17]. The morphological and behavioural abnormalities found in the knockout mice provide further evidence that the cognitive impairment in our human subjects is an integral part of this novel recessive disorder which we have called SPARCA1 (“Spectrin-associated Autosomal Recessive Cerebellar Ataxia type 1”). We suggest that this represents one of a novel group of disorders, the neuronal spectrinopathies, which demonstrate that the cell-specific functional repertoire of spectrin subunits are involved in brain development including the cortex, in addition to cerebellar development and function.
The three affected individuals are from a UK family of Pakistani origin with complex consanguinity (see Figure 1A), but no other family history of neurological disorders. The clinical phenotype in the 3 individuals is identical (Table 1). V1 was referred at the age of 13 months with motor delay; she was extremely floppy and was unable to crawl. She sat at 10 months, crawled at 18 months and was pulling to stand at 20 months. She walked with a walker by the age of 5 and started to walk with support at age 7. She was noted to have language delay and at age 5 was just starting to join words together. Global developmental delay was subsequently noted, she was educated at a special school and now attends a college for adults with special educational needs. On examination there are abnormal eye movements with a convergent squint, hypometric saccades, jerky pursuit movements, and an incomplete range of movement particularly in the horizontal plane. There is obvious dysmetria and dysdiadochokinesia of the limbs and gait ataxia with inability to tandem walk without falling. Limb tone is normal, reflexes are normal and plantars flexor and there is no evidence of any sensory abnormality. Rombergs sign is normal. Neuropsychological assessment reveals significant global cognitive impairment with all IQ scales falling at the second percentile or below, and with Full Scale IQ scores falling in the learning disabled range (Table 1). A brain CT scan at age 2 did not show any abnormality, but a recent MRI brain reveals significant cerebellar atrophy (Figure 2A). V2 is the younger sibling of V1. She was noted to have developmental delay in early childhood and also did not start to walk until age 7. On examination, she has an identical clinical phenotype to that of her sister except for occasional beats of nystagmus on eye examination. She attends a school for children with learning disabilities and a recent assessment (at age 16) shows functioning in English and Mathematics at the level of an average 5–7 year old in the UK requiring special educational support. Formal cognitive assessment also showed very similar impairments to V1 with scores on all IQ scales falling at the second percentile or below, and with Full Scale IQ scores falling in the learning disabled range (Table 1). The difference between Verbal and Performance IQ for each individual was not statistically significant (p = 0.15). MRI imaging in V2 at age 6 revealed cerebellar atrophy and this was found to have progressed over time (Figure 2Bi and Bii). V3 is the first cousin of V1 and V2. He was noted to have poor head control and balance in early childhood. Clinical examination is identical to his cousins and also shows an identical developmental profile in that he has just started to walk with assistance at the age of 7. He also has an identical eye movement disorder, a convergent squint, dysmetria and dysdiadochokinesia. He is hypotonic with normal reflexes downgoing plantars and no evidence of a sensory neuropathy. He attends a mainstream school but requires full time one to one support. Cognitive assessment of V3 also showed significant global cognitive impairment (Table 1). The slightly higher IQ scores in V3 results from a floor effect in the normative data rather than a significant difference in cognitive ability from his older cousins. In this age cohort the lowest attainable scores are VIQ = 62, PIQ = 73 and FSIQ = 63 and therefore V3 falls in the same learning disabled range as his cousins. Brain imaging of V3 showed a normal cerebellum at age 5, but mild hypoplasia of the posterior corpus callosum (Figure 2C). The normal appearance of the cerebellum in V3 at an early age is not unexpected as both his cousins imaging shows progression with time. Neurological examination of both sets of parents was entirely normal, with no evidence of ataxia. The father of V1 and V2 works as a bus driver, having left school at age 16 with 5 GCSEs (General Certificates of Secondary Education) and the father of V3 works in a warehouse and has a similar educational background. Formal psychometric testing in the father of V1 and V2 showed IQ indices falling in the low average range consistent with his educational attainment. The father of V3 was not available for testing but has very similar attainment levels to his brother. Formal assessment of the mothers could not be performed since neither speak English, but interview of the family did not reveal any evidence of learning disability. There is no history of the siblings or grandparents of the affected individuals having any cognitive or neurological abnormalities.
We initially performed targeted capture of >100 known ataxia genes (including SPTBN2) in a group of children with unexplained ataxia including patient V3, followed by next generation sequencing. In V3 we identified only one mutation, a homozygous stop codon p. C627X (c.1881C>A), located in the third spectrin repeat in SPTBN2 and used Sanger sequencing to confirm that all three affected patients in the family had the same mutation whereas the neurologically normal parents of V3, were shown to be heterozygous for the mutation (Figure 1B). Since mutations in β-III spectrin are associated with cerebellar degeneration in SCA5, the newly identified mutation was considered likely to explain the ataxia, although of a developmental type with a much earlier onset. However, since more than one mutation can co-segregate, particularly in consanguineous families, we went on to consider the contribution of the mutation in SPTBN2 to the observed cognitive impairment. We therefore used SNP array analysis and whole genome sequencing to search for any evidence of a second mutation.
To investigate whether a second homozygous mutation segregated with the cognitive impairment, all 3 affected individuals (V1, V2 and V3) and the unaffected parents of V3 (IV3 and IV4) were genotyped to identify regions of homozygosity (ROH) shared by V1, V2 and V3 and not present in either IV3 or IV4. This analysis identified 20 shared homozygous segments on autosomes totalling 17.1 Mb (Table 2). SPTBN2, on chromosome 11, was located in the largest ROH shared by V1, V2 and V3 and not present in either IV3 or IV4 (Figure 3).
Whole genome sequencing of patient V2 was performed on the Illumina HiSeq2000 as 100 bp paired end reads, using v3 clustering and sequencing chemistry. After duplicate reads removal, the mean coverage across the genome was 25.6× with 90.4% of bases covered at 15× or more. The mean coverage over the 17.1 Mb ROH identified by SNP analysis was 25.9× with 93.4% of bases covered at 15× or more. Variant calling was performed as detailed in the Materials and Methods.
We firstly based our data analysis on an autosomal recessive disease model, caused by one or more rare homozygous mutations and focused on homozygous variants occurring in the shared ROH identified by SNP array analysis, filtering them out if they were:
These filtering steps identified 68 candidate variants, subdivided into functional classes (Table 3). Only 2 exonic variants were found: a synonymous variant, NPHP1 L551L on chr2 which is not predicted to be pathogenic and is not located near a splice site, and the stop codon C627X in SPTBN2 on chr11 (Table 2 and Table 3). Of the remaining variants, 21 were intergenic and also considered unlikely to be disease related, and 4 variants were in untranslated regions (5′ UTR) or in non-coding RNAs and all were in positions which scored poorly with PhyloP and GERP. In addition, none of the associated genes (UBIAD1, LINC00116, LOC100130987) appear to be relevant for this disorder. The other 41 were in intronic and upstream regions but based on evolutionary conservation and available information in databases (eg HGMD [18]) we found no evidence of potential involvement in the disease. The only likely pathogenic variant is the stop codon in SPTBN2.
We also considered a model of recessive inheritance with compound heterozygous mutations segregating with the ataxia and/or cognitive impairment. Our criteria were that all 3 affecteds must have two different variants in the same gene and where this occurred the variants should be in trans (ie each parent is a carrier). We identified all potential compound heterozygous coding variants present in the WGS data for individual V2. In total there were variants fulfilling our criteria at 13 different loci but in only 1 case were both variants present in all 3 affecteds and further analysis revealed that in this instance both variants were also in the father of V3 (ie were in cis). Furthermore, none of the variants identified are known to be associated with ataxia or cognitive impairment and the majority of genes had data suggesting an alternative function (such as taste or fertility), nor were there any likely candidates based on pathogenicity bioinformatic prediction programs (Table S2).
The phenotype of our patients suggested that β-III spectrin is involved in cognitive development, in addition to being essential for motor functions. We therefore utilised β-III spectrin knockout mice which have progressive cerebellar degeneration and lack any full length β-III spectrin [13], to further investigate the role of β-III spectrin in other brain regions. Our previous work revealed that β-III spectrin is required for the correct dendritic development of Purkinje cells [19], [20] and therefore we initially examined dendritic organisation in other brain regions by immunostaining sagittal sections from the brains of 6-week-old wild-type and β-III spectrin knockout animals for microtubule associated protein 2 (MAP2), a dendritic marker. This revealed irregular reactivity throughout the PFC layers and within the caudate putamen/striatum of knockout animals when compared to WT mice but no obvious difference in the HPC (Figure 4A). However no difference was observed between WT and β-III spectrin knockout animals when the cortex and striatum were immunostained for tau or myelin basic protein (MBP) indicating that there was no change to axonal structure (Figure S1).
The PFC in humans is believed to be important for complex cognitive tasks, and given there is evidence of a close association between this area and the neocerebellum, as well as high expression levels of β-III spectrin in mouse [10] we further investigated the prefrontal cortical region in β-III spectrin knockout animals. There was no difference in the thickness of individual prefrontal cortical layers (data not shown) but the morphology of individual pyramidal neurons in β-III spectrin knockout animals was found to be altered. Morphometric analysis of dye-injected pyramidal neurons from layer 2/3 showed basal dendrites in 8-week-old β-III spectrin knockout mice were significantly thinner distally compared to wild type cells (Figure 4B–4D). Moreover, the basal dendrites of knockout mice tapered more rapidly than those of wild types, being significantly reduced in thickness between 20 and 30 µm from the soma, whereas wild type dendrites showed no significant narrowing until 90 µm from the soma. However, no difference in spine density was observed between genotypes in either dye injected (Figure 4D: +/+, 2.8±0.6, n = 8; −/−, 3.2±0.2 spine/µm3, n = 7; p = 0.56) or Golgi-impregnated (Figure 4E: +/+, 12.4±1.7, n = 4; −/−, 13.7±1.3 spine/10 µm, n = 6; p = 0.56) pyramidal neurons. Only small sections of apical dendrites could be reconstructed from the serial stacks of dye-injected cells. Nevertheless, quantification of the short regions imaged, when normalized to length analysed, indicated reduced apical dendritic volumes, and hence thinner apical dendrites in β-III spectrin knockout animals (+/+, 4.3±0.47; −/−, 2.5±0.36 µm3/µm, n = 6 for each genotype; p = 0.011).
Since patient V3 shows mild hypoplasia of the posterior corpus callosum we examined this brain structure in 8-week old β-III spectrin knockout animals to determine if the morphological defect in the human subject could be a consequence of β-III spectrin loss or is unlinked to the homozygous stop codon c.1881C>A (p.C627X) mutation in SPTBN2. No signs of posterior hypoplasia were observed in sagittal sections stained either with cresyl violet (Figure 5A) or an anti-tau antibody (Figure 5B). Similarly width of corpus callosum, measured from coronal sections immunostained for MBP (Figure 5C), was no different between WT and knockout animals (+/+, 469.7±46.6; −/−, 480.6±41.3 µm, N = 3 for each genotype; p = 0.28).
Four object recognition memory tasks (two- and four- novel object preference, object-in-place and object location; Figure 6A–6D) were carried out to assess whether β-III spectrin knockout animals displayed any cognitive deficits. No impairment in the two novel object recognition task (“object identity”) was observed in β-III spectrin knockout animals compared with wild type animals (Figure 6A); however knockout animals performed worse in the four novel object recognition task (Figure 6B). Knockout animals were also worse at discriminating between rearranged and non-rearranged objects in the object-in-place task compared with litter mate controls, shown by their failure to spend more time exploring the two objects in different locations compared with the two objects that had not moved (“object displacement”) (Figure 6C). However, there was no significant difference in performance for the object location task (Figure 6D). The poorer performance in the four-novel object recognition task for knockout animals was not a consequence of less exploration in the 5 minute sample phase as in fact they explored more than wild type animals (+/+, 64.9±6.7; −/−, 88.7±4.8 sec; p = 0.018). Similarly for the object-in-place task although there was no significant difference between genotypes there was a trend for greater exploration in knockout animals (+/+, 42±3.6; −/−, 62.2±8.7 sec; p = 0.054).
The integrated evidence from clinical, genetic and neuropsychological analysis in humans and behavioural and morphological analysis in a mouse model demonstrate that we have identified a novel recessive disorder, SPARCA1, associated with mutations in β-III spectrin. The 3 human subjects with a premature stop codon and the mouse knockout all have very early onset cerebellar ataxia, indicating a developmental role for β-III spectrin. The human and mouse knockout phenotype also show that β-III spectrin is involved in cognitive development and function. The human subjects have global cognitive impairment in the mild/moderate range. The specific brain structures and connections associated with this impairment are not yet known and further detailed neuropsychological testing will be required. However, we have shown that in the mouse knockout there are morphological abnormalities especially thinning of dendrites in PFC neurons, similar to that previously reported for Purkinje neurons [19], but with no obvious changes in various regions of HPC (CA1, CA3 and dentate gyrus), and the behavioural tests in the mouse are consistent with this. Based on published lesion studies, deficits in the object-in-place task but not the object location task would indicate defects in the PFC not HPC, since PFC is believed to mediate memory for object location (Òobject displacementÓ), whereas HPC integrates information as to object identity and the temporal order of object presentation with HPC lesioned animals being impaired on object location task [16], [17], [21]. However, further to the above discussion, there is also increasing recognition that the cerebellum itself has a direct role in cognition [22] and it is possible that some of the phenotype results directly from cerebellar abnormalities. Further investigation should also allow a detailed analysis of which specific brain regions mediate mild/moderate cognitive impairment in humans.
The data demonstrate that our β-III spectrin knockout mouse [13] is an excellent model for the novel recessive disorder we have identified and will allow further molecular analysis of β-III spectrin, in addition to the morphological and behavioural analysis. β-III spectrin is known to be expressed widely throughout the brain, kidney, liver and testes and to be associated with the Golgi and other cytoplasmic vesicles [23], but the mechanisms by which mutations lead to impaired brain development are unknown. The premature stop codon C627X identified in our family is predicted to result in truncation of β-III spectrin near the end of the 3rd spectrin repeat (Figure 7). This truncated protein would be unable to form tetramers with α-II spectrin, nor be able to bind to EAAT4 or ankyrin, but it is possible that there is nonsense mediated decay and loss of the entire protein. Since SPTBN2 is expressed at only very low levels in peripheral blood, further in vitro expression studies will be required to determine this. However, it is most likely that β-III spectrin is absent in the brain of the human subjects and this has resulted in neuronal dysfunction in widespread brain regions, notably cerebellum and prefrontal cortex. Future studies will investigate other brain regions such as striatum and perirhinal cortex as well.
Our findings also provide insights into the mechanism of molecular dominance in SCA5: the heterozygous carrier parents of the C627X stop codon in the SPARCA1 family are neurologically normal despite carrying a stop codon which in the homozygous state is a recessive loss of function mutation. Therefore haploinsufficiency is highly unlikely to be the mechanism underlying SCA5 and this lends considerable weight to the body of experimental evidence suggesting that SCA5 results from a dominant negative effect, possibly by interfering with normal binding to ARP1 [13], [14], [24].
One difference between the human and mouse model is that the mouse shows progressive motor deficits in addition to progressive Purkinje cell loss whereas there is no evidence of clinical progression in the patients at the moment despite one of our subjects having progressive cerebellar atrophy on imaging. This lack of clinical progression and discordance between the clinical and imaging findings could suggest that there is significant plasticity within the human cerebellum, although we cannot exclude the possibility that slow clinical progression will occur with time.
The phenotypic spectrum of neuronal spectrinopathies now appears to be very wide. In SCA5, the ataxia is generally a pure adult-onset ataxia whereas recessive mutations in SPTBN2 cause SPARCA1, a more severe childhood ataxia with cognitive impairment. In West Syndrome, associated with SPTAN1 mutations, the patients have epilepsy, profound developmental delay and in addition have shortening of the corpus callosum and cerebellar vermis atrophy. Only one of our patients, V3, had shortening of the corpus callosum and it is tempting to speculate that this additional feature may be part of the SPARCA1 phenotype, although there are no signs of hypoplasia in the β-III spectrin knockout mice. It also may be that this feature is caused by another gene mutation or a genetic modifier and to clarify this additional cases will need to be identified. Overall, our data suggest that region specific expression of spectrin subunits is important in prenatal brain development and further work is required to define their temporal and spatial contribution.
Our data also suggest the possible and testable hypothesis that the phenotype in neuronal spectrinopathies relates in part to the total amount of functional spectrin tetramers: in SCA5, all α-II/β-II tetramers are normal and functional but α-II/β-III tetramers will contain mutant β-III spectrin which likely have a dominant negative action and may not be fully functional; in SPARCA1, a recessive disorder, there is complete loss of the tetramerisation site of β-III spectrin so there will be normal α-II/β-II tetramers but no functional α-II/β-III tetramers, whereas the heterozygotes who are effectively “haploinsufficient” have enough α-II/β-III tetramer to be clinically normal; in West Syndrome, caused by in-frame dominant SPTAN1 mutations [15], the majority of both α-II/β-II and α-II/β-III tetramers are abnormal resulting in the most severe of the disorders to be described so far (Figure S2). This model would suggest that homozygous loss of function α-II spectrin mutations might be more severe or lethal and a very recent report of an α-II knockout mouse supports this and it will be important to identify the equivalent human disorder [25]. There may be other disorders associated with human disease: dominant negative or recessive mutations in β-II and proteins interacting with brain spectrins may also have similar phenotypes. For example, a mouse knockout model of Ankyrin G, was reported to cause Purkinje cell degeneration [26] but a human phenotype has not yet been found. In addition, seizures are described in SPTAN1 mutations [15] and another β-III spectrin knockout [24] and it will be important to search for spectrin mutations in epilepsy patients.
In conclusion, the identification of recessive mutations in β-III spectrin provides evidence that the cell-specific repertoire of spectrin subunits underlies a novel group of disorders, the neuronal spectrinopathies, including SCA5, a dominant form of West Syndrome and SPARCA1. It is likely that other human disorders are caused by mutations in neuronal spectrins and searches for these are in progress. We also demonstrate the power of analysing complex phenotypes in consanguineous families by using whole genome sequencing, which was critical in establishing that both the ataxia and the cognitive impairment were caused by the same mutation and illustrate how the use of genome sequencing, even in single human families, can help provide mechanistic insights into disease.
Our institutional ethics committee approved the study on human participants and specific consent was obtained to include whole genome analysis. All procedures involving analysis of mutant mice were carried out according to the United Kingdom Animals (Scientific Procedures) Act (1986) and other Home Office regulations under specific pathogen-free conditions.
The exonic sequences of 129 genes known or suspected to be associated with ataxia were selected for targeted capture (Table S1) and 120-mer baits with 2X tiling designed using the Agilent eArray design tool. The total size of the targeted region amounted to 605.8 kb. Multiplex sequencing was performed on the Illumina GAII with 51 bp paired-end reads. A total of 5,046,154 reads were generated for patient V3 and aligned to the human reference genome (GRCh37/hg19) with STAMPY [27] About 60% of the reads mapped to the target region, providing a mean depth coverage of 218.4× with 89.8% of target bases covered at 30× or more. Single nucleotide variants (SNVs) and indels were called respectively with SAMTOOLS [28] and DINDEL [29]. Variants were annotated with respect to gene and transcripts using the Ensembl database (release 62, Apr 2011 [30]) by means of the associated Variant Effect Predictor tool.
Results were confirmed using Sanger Dideoxy Sequencing with the following primers across exon 14 of SPTBN2: Forward: CTACCTCTGCTGCACGACCT; Reverse: AGGGAGGGAAGTCCAAGAGA. Genomic DNA was amplified with Taq Polymerase (Roche) and PCR products were used as templates for sequencing with BigDye Terminator reagents (Life Technologies) on a 3730xl DNA Sequencing Analyzer (Life Technologies). The sequence traces were aligned to the gene-specific reference sequence (NCBI build 37) with Sequencher 4.10.1 (Gene Codes).
Genotyping was performed using the Illumina HumanCytoSNP-12v1 BeadChip, containing nearly 300,000 genetic markers. Hybridization to the chip was performed according to manufacturer's protocols found on registration at http://www.illumina.com/support/array/array_kits/humancyto-snp-12_v2-1_dna_analysis_kit/documentation.ilmn. In brief, patient DNA was denatured, amplified and enzymatically fragmented and then hybridized onto CytoSNP-12 BeadChips by rocking in an Illumina hybridization oven at 48°C for 16–24 hrs. The BeadChips were washed according to the Illumina Inc. protocol and the hybridized DNA detected by primer extension with labelled nucleotides followed by detection using fluorescent antibodies. The data were processed using Illumina's GenomeStudioV2009.2.
As SNP coordinates in the chip were reported with respect to human genome build 36, we downloaded the corresponding coordinates for build 37 from the website http://www.well.ox.ac.uk/~wrayner/strand/, cross-checking them using the USCS Genome Browser liftOver utility (http://genome.ucsc.edu/cgi-bin/hgLiftOver) and the dbSNP database (Build 135 [31]). We filtered out ∼18,000 markers which could not be mapped unambiguously to build 37 of the human genome. We further excluded SNPS with missing calls in one or more samples, thus reducing the number of markers to 271,208.
PLINK v1.07 (http://pngu.mgh.harvard.edu/purcell/plink/ [32]) was used to identify regions of homozygosity (ROH) shared by V1, V2 and V3 and not present in either IV3 or IV4. For V1, V2 and V3, we applied relaxed parameters in order to include all potential ROH, resulting in potential false positives but minimizing false negatives. We defined a homozygous region as a run of (at least) 50 homozygous SNPS spanning more than 500 kb, allowing for some heterozygous calls within it. Shared ROH were identified from overlapping and allele matching segments. Further details of the algorithm are provided on the PLINKwebsite. We used the options: –homozyg –homozyg-group –homozyg-window-kb 500 –homozyg-window-snp 50 –homozyg-snp 50 –homozyg-kb 500. All other parameters were left at default values. ROH were then identified in IV3 and IV4. In this case very stringent criteria were applied to confidently include only true ROH and avoid false positives. We defined a homozygous region as an uninterrupted run of (at least) 500 homozygous SNP's spanning more than 5 Mb. In IV3 we identified 8 ROH on autosomes totalling 78 Mb (the largest ROH was 18.4 Mb); in IV4 we identified 2 large ROH on chromosome 11 present also in V1, V2 and V3 (Table 2 and Figure 3). These regions were excluded in the search for pathogenic variants as both IV3 and IV4 are unaffected. As a result, the search was restricted to 20 regions totalling 17.1 Mb, among which the ROH harbouring SPTBN2 was the largest.
Screening of cognitive function was undertaken using the Wechsler Abbreviated Intelligence Scale (WASI).
For immunostaining and histological analysis brains from wild type and β-III spectrin knockout animals were removed and immersion-fixed with either 1 or 4% paraformaldehyde in 0.1 M sodium phosphate buffer, pH 7.4 overnight at 4°C and cryoprotected in 0.1 M sodium phosphate buffer (pH 7.4) containing 30% sucrose. Tissue was embedded in OCT then 16 µm-thick sections cut and mounted onto poly-L-lysine coated slides. Primary antibodies used were mouse anti-MAP2 (Sigma), rabbit anti-tau (DAKO) and rat anti-myelin basic protein (AbD Serotec). Secondary antibodies were cyanine 3 (Cy3)-conjugated goat anti-mouse IgG (Jackson laboratories), fluorescein isothiocyanate (FITC)-conjugated goat anti-rabbit IgG (Cappel) and Alexa Fluor 488 –conjugated donkey anti-rat (Invitrogen). For Golgi impregnation brains were removed and immersion-fixed with 4% paraformaldehyde in 0.1 M sodium phosphate buffer, pH 7.4 overnight at 4°C and processed as described previously [39]. For cell filling animals were deeply anesthetized with isofluorane and sacrificed by transcardial perfusion with 4% paraformaldehyde in 0.1 mM phosphate buffer, pH 7.4. Brains were dissected and postfixed in 1% paraformaldehyde overnight at 4°C. Coronal sections were cut (250 µm-thick) and individual neurons in layer 2/3 of the prefrontal cortex were visualized with a 20× immersion objective and injected with 0.2 mM Lucifer Yellow (Sigma) and 0.02 mM Alexa FluorAR 568 hydrazide (Invitrogen). Slices were post-fixed and 4% paraformaldehyde overnight at 4°C and wet-mounted with Vectashield onto 0.13 mm thick borosilicate glass and neurons imaged using the Alexa 568 dye. All images were captured using a Zeiss inverted LSM510 confocal scanning laser microscope and serial stacks used for three-dimensional reconstruction of dendritic arbors using NeuronStudio software (CNIC).
Animals were handled for 1 week and then habituated to the arena (40 cm×40 cm×40 cm) for 5 d before testing. All tests involved a 5 min sample phase followed by a 5 min test phase after a delay of 5 min. Exploratory behaviour was recorded via a WebCam positioned above the testing arena and two researchers blind to genotype scored the investigation of each sample using ANY-maze software (Stoelting). As described previously [16], [21] for the novel object preference tasks one object from the sample phase was replaced with a novel object in the test phase; the object-in-place task comprised switching the location of two familiar objects in the test phase; and for the object location task position of one familiar object was changed (Figure 6A–6D). Duplicate copies of familiar objects were used in the test phases to remove any chance of olfactory cues being present. Discrimination ratios were calculated as the time spent exploring the novel or location switched object(s) divided by the total time spent exploring all objects.
Statistical analysis was performed using Student's t-test, two sample assuming unequal variance, apart from analysis of filled pyramidal cells where a two-way ANOVA was used.
|
10.1371/journal.pbio.2001164 | The mismatch repair and meiotic recombination endonuclease Mlh1-Mlh3 is activated by polymer formation and can cleave DNA substrates in trans | Crossing over between homologs is initiated in meiotic prophase by the formation of DNA double-strand breaks that occur throughout the genome. In the major interference-responsive crossover pathway in baker’s yeast, these breaks are resected to form 3' single-strand tails that participate in a homology search, ultimately forming double Holliday junctions (dHJs) that primarily include both homologs. These dHJs are resolved by endonuclease activity to form exclusively crossovers, which are critical for proper homolog segregation in Meiosis I. Recent genetic, biochemical, and molecular studies in yeast are consistent with the hypothesis of Mlh1-Mlh3 DNA mismatch repair complex acting as the major endonuclease activity that resolves dHJs into crossovers. However, the mechanism by which the Mlh1-Mlh3 endonuclease is activated is unknown. Here, we provide evidence that Mlh1-Mlh3 does not behave like a structure-specific endonuclease but forms polymers required to generate nicks in DNA. This conclusion is supported by DNA binding studies performed with different-sized substrates that contain or lack polymerization barriers and endonuclease assays performed with varying ratios of endonuclease-deficient and endonuclease-proficient Mlh1-Mlh3. In addition, Mlh1-Mlh3 can generate religatable double-strand breaks and form an active nucleoprotein complex that can nick DNA substrates in trans. Together these observations argue that Mlh1-Mlh3 may not act like a canonical, RuvC-like Holliday junction resolvase and support a novel model in which Mlh1-Mlh3 is loaded onto DNA to form an activated polymer that cleaves DNA.
| In sexually reproducing organisms, crossing over between homologous chromosomes in meiosis creates physical linkages required to segregate chromosomes into haploid gametes. In baker’s yeast, which utilizes meiotic recombination pathways conserved in mice and humans, the majority of meiotic crossovers are initiated through the formation of a branched DNA intermediate, which is stabilized by the Msh4-Msh5 complex. This DNA intermediate is further processed to form a structure (a double Holliday junction), which requires the endonuclease activity of the Mlh1-Mlh3 DNA mismatch repair factor to be resolved exclusively into a crossover product. Current meiotic recombination models invoke the use of structure-specific enzymes that symmetrically cleave single Holliday junctions. In this study, we provide evidence that the yeast Mlh1-Mlh3 complex is unlikely to act as a structure-specific enzyme. Furthermore, we showed that Mlh1-Mlh3’s endonuclease activity is dependent upon its ability to form a polymer on DNA and suggest that it is capable of cleaving DNA that is captured in an active complex. Together, our biochemical observations support a novel model involving regulated polymerization of Mlh1-Mlh3 for its cleavage function, potentially in meiotic crossovers or in mismatch repair.
| Mismatch repair (MMR) acts during DNA replication to remove DNA polymerase misincorporations such as single-nucleotide base–base mismatches and insertion or deletion loops that result from polymerase slippage events. In Escherichia coli, DNA mismatches are initially recognized by MutS, which recruits MutL in an ATP-dependent reaction to activate MutH. MutH functions as a latent endonuclease that nicks the newly replicated and undermethylated DNA strand up to several kilobases away from the mismatch. The resultant nick, 5′ or 3′ to the mismatch, directs excision and resynthesis steps that remove the misincorporated DNA [1].
MMR is highly conserved across organisms. In baker’s yeast, mice, and humans, MutS homologs (MSHs) form heterodimers Msh2-Msh6 and Msh2-Msh3, which recognize and bind with overlapping specificities to base–base mismatches and insertion or deletion loops. These MSH complexes primarily interact with the MutL homolog (MLH) complex Mlh1-Pms1 (MLH1-PMS2 in humans), which nicks the newly replicated strand to promote excision and resynthesis in steps coordinated with a yet-to-be-understood strand discrimination mechanism [1]. A minor MMR pathway has been identified in baker’s yeast in which Mlh1-Mlh3 interacts with Mlh1-Pms1 to repair a subset of DNA mismatches that are recognized by Msh2-Msh3 [2–4].
A subset of MSH and MLH protein complexes act during yeast meiosis to promote the major class of crossovers (COs) that form between homologs. In Saccharomyces cerevisiae, meiotic recombination is initiated by the formation of ~200 programmed double-strand DNA breaks (DSBs) that appear throughout the genome. During prophase I, DSBs are resected to form 3′ single-strand tails, which are directed to invade and pair with the unbroken homolog. About half of these invasions are stabilized by a family of ZMM proteins (Zip1-4, Mer3, and Msh4-Msh5) to form single-end invasion (SEI) intermediates. The SEIs are further processed through synthesis and ligation steps to form double Holliday junction (dHJ) intermediates, which are subsequently resolved exclusively into COs [5–7]. Invasion intermediates that are not protected by ZMM proteins can also mature into dHJs but are not sensitive to interference and are resolved into both COs and noncrossovers (NCOs). Work in Sordaria, mouse spermatocytes, and human oocytes suggested that Msh4 foci (presumably bound to Msh5) form in leptotene at all DSB sites [8–11]. As meiosis progresses, the number of Msh4 foci decrease. When the cell enters pachytene, Msh4 foci were found to colocalize with Mlh1 foci at sites that ultimately form COs [12].
How does the Mlh1-Mlh3 endonuclease act to resolve dHJs to form crossovers? Genetic and molecular work performed in baker’s yeast showed that dHJs stabilized by ZMM proteins are resolved through the actions of factors that include the Mlh1-Mlh3 endonuclease, Sgs1-Top3-Rmi1, and the exonuclease-independent functions of Exo1 [13,14]. Resolution of dHJs in this pathway is biased to cut the two Holliday junctions (HJs) in opposite orientations to produce exclusively COs [15–23]. Interestingly, the HJ resolvases Mus81-Mms4, Slx1-Slx4, and Yen1 can compensate for the absence of Mlh1-Mlh3, but dHJs are processed through an interference-nonresponsive pathway that randomly resolves dHJs into COs and NCOs. These structure-selective HJ resolvases, which can bind to and cleave synthetic HJs and other branched structures at specific sites in vitro, act independently of the ZMM proteins and in the presence of Mlh3 account for only a minority of crossover events [18] (reviewed in [24]). Mus81-Mms4 and Yen1 are regulated by phosphorylation in a cell cycle–dependent manner [25].
Previously, we showed that yeast Mlh1-Mlh3 is an endonuclease whose activity is stimulated on plasmid substrates by yeast mismatch repair factor Msh2-Msh3 but not the replication processivity clamp/clamp-loader (PCNA/RFC), which stimulates Mlh1-Pms1 [26,27]. Mlh1-Mlh3 displays binding preference for oligonucleotide substrates containing mismatches and branched DNA structures, particularly HJs [26,27], but does not display endonuclease activity on these oligonucleotide substrates, even in the presence of Msh2-Msh3. This behavior is significantly different from the biochemical activities of the well-characterized HJ resolvases Mus81-Mms4, Slx1-Slx4, and Yen1, which cleave such substrates (reviewed in [7,24]), but is consistent with activities seen for the MMR endonuclease Mlh1-Pms1 in yeast or MLH1-PMS2 in humans ([28,29]; see below). In vivo analysis of MMR factors in S. cerevisiae suggests that multiple Mlh1-Pms1 complexes bind to a DNA mismatch [30]. In vitro, yeast Mlh1-Pms1 exhibits high affinity and cooperative binding to large DNA substrates [31], and single-molecule studies showed that yeast Mlh1-Mlh3 and Mlh1-Pms1 have similar diffusion properties on duplex DNA [32,33]. These observations indicate that DNA-binding properties are conserved among MLH family members; in support of this idea, work in E. coli suggested that multiple MutL proteins interact with a MutS-bound mismatch substrate [34].
Because MLH proteins display similar DNA binding properties that appear critical for their roles in MMR, we hypothesize that such properties are utilized by Mlh1-Mlh3 in meiosis to resolve dHJs into COs. Here we show that Mlh1-Mlh3 displays an endonuclease activity on large duplex DNA substrates; DNA binding, wild type and mutant protein complex mixing experiments, and electron microscopy assays indicate that Mlh1-Mlh3 polymer formation is required for this activity. We also observe that Mlh1-Mlh3 is capable of making DSBs in a concerted manner that are religatable. Finally, we observe that Mlh1-Mlh3 is capable of cleaving DNA substrates in trans. These data support a novel model in which a polymer of Mlh1-Mlh3, positioned and directed by other meiotic recombination proteins, may cleave recombination intermediates to form COs. These properties distinguish Mlh1-Mlh3 from structure-selective endonucleases found in other dHJ cleavage pathways.
Mlh1-Mlh3 is active on plasmid substrates in the presence of divalent magnesium or manganese, with manganese giving slightly greater nicking activity ([26,27]; Fig 1A). We did not observe endonuclease activity with either metal using purified Mlh1-mlh3D523N, a complex bearing a mutation that disrupts a conserved residue seen in the MutL family endonucleases ([19,26,27]; Fig 1A).
Curiously, Mlh1-Mlh3 shows preferential binding (Fig 1B and 1C) to a variety of branched DNA structures such as HJs and insertion or deletion loop mismatch DNA substrates but does not cleave them ([26,27]). In addition, when a cruciform was incorporated into a plasmid substrate to mimic a HJ intermediate, specific nicking at the cruciform was not observed [27]. These observations suggested to us that Mlh1-Mlh3 does not display structure-specific nuclease activities on HJ substrates, as was seen for RuvC, Yen1, and Mus81-Mms4. To further test this, we performed the three experiments described below.
First, we observed that incorporating a +8 loop mismatch into a plasmid substrate had an inhibitory effect on Mlh1-Mlh3 endonuclease activity and did not localize its endonuclease activity to sites near the mismatch (see details in the next section). Second, we performed an oligonucleotide competition endonuclease assay in which Mlh1-Mlh3 was incubated with a circular plasmid substrate in the presence of homoduplex, HJ, and +8 loop–containing oligonucleotides. As shown in Fig 1B and 1C, HJ and +8 loop–containing oligonucleotide substrates acted as very strong competitors relative to homoduplex DNA, which was an ineffective competing substrate. The differences in competition efficiency among the oligonucleotide substrates were much greater than Mlh1-Mlh3’s DNA-binding affinities for these substrates—the Kd values are in the nanomolar range for all substrates [26,27]. Finally, we tested whether Mlh1-Mlh3 activity is sensitive to phosphorylation. In meiosis, Mus81-Mms4 and Yen1 endonuclease activities are regulated by their phosphorylation state. Phosphorylation of Mus81-Mms4 by Cdc5 has been shown to hyperactivate the endonuclease, while Yen1 is inhibited by phosphorylation [25]. Based on Mlh1-Mlh3’s identity as an MMR factor, its activity is not expected to be sensitive to phosphorylation; there is no evidence in the literature that MLH endonuclease activity is regulated by phosphorylation. Nevertheless, we treated Mlh1-Mlh3 with either lambda protein phosphatase or CDK1-cyclinB, which has a recognition motif in Mlh3. We then compared the protein’s endonuclease activity to controls not treated with a phosphatase or kinase (S1 Fig). We did not observe an effect on endonuclease activity when Mlh1-Mlh3 was treated with a phosphatase or kinase, suggesting that it is unlikely to be sensitive to these modifications. Together, our data argue against Mlh1-Mlh3 acting according to archetypes set by canonical HJ resolvases, and HJs and mismatched substrates are unlikely to be the preferred in vivo substrates for the Mlh1-Mlh3 endonuclease.
Previously we showed in electrophoretic mobility shift assays (EMSA) that a specific Mlh1-Mlh3-DNA gel shift could only be detected in a narrow range of Mlh1-Mlh3 concentrations that are suboptimal for nuclease activity. At higher concentrations, a more robust nuclease activity was seen, but in EMSA assays performed at these concentrations, Mlh1-Mlh3-DNA complexes do not enter the gel [26,27]. This observation and the lack of endonuclease activity on oligonucleotide substrates suggested to us that multiple Mlh1-Mlh3 heterodimers are required to activate its endonuclease function. The more thoroughly studied yeast Mlh1-Pms1 complex was shown to bind larger DNA substrates with higher affinity than smaller DNA substrates [31]. Hall et al. hypothesized that this behavior is suggestive of cooperative binding, which necessitates the presence of multiple protein complexes [31]. They supported this biochemical observation using atomic force microscopy, where they observed long tracts of protein bound to plasmid DNA. Consistent with the in vitro data, multiple (~15) Mlh1-Pms1 molecules are recruited to mismatch sites in vivo [30]. These observations encouraged us to test whether disruptions to DNA substrates inhibit Mlh1-Mlh3’s endonuclease activity.
We tested if Mlh1-Mlh3 would display a similar endonuclease activity on a 2.7 kb linear duplex identical in size and sequence to a supercoiled circular substrate (Fig 2A and 2B and S2A and S2B Fig). While we observed significant activity on the circular substrate, we were unable to observe nicking on the linear substrate using either Mg2+ or Mn2+ as a metal cofactor. To determine if the lack of endonuclease activity on the linear substrate was caused by Mlh1-Mlh3 sliding off of the ends of the substrate before it could nick, we biotinylated the 3′ ends of the linearized plasmid and attached streptavidin to block the ends. Attaching streptavidin to the ends of the linear DNA did not restore Mlh1-Mlh3 nicking (Fig 2A and 2B). The preferential nicking of circular double-stranded DNA (dsDNA) was not due to the negatively supercoiled topology, because both relaxed and negatively supercoiled plasmids were nicked with similar efficiencies (S3 Fig).
To better understand Mlh1-Mlh3’s substrate preferences, we tested Mlh1-Mlh3 endonuclease activity on 1.4 to 15 kb circular dsDNA substrates (Fig 2C and 2D, S2C, S2D and S6 Figs; data for 1.4 kb circular substrate are in S5B Fig). Although Mlh1-Mlh3 nicked all circular substrates tested, activity was directly proportional to plasmid size in the presence either Mg2+ or Mn2+ cofactor. Optimum endonuclease activity for Mlh1-Mlh3 was seen at an ~100:1 ratio of Mlh1-Mlh3 to 2.7 kb plasmid. This finding suggested that multiple Mlh1-Mlh3 molecules interact on the DNA substrate to activate endonuclease activity. It is also consistent with the finding that a large molar excess of Mlh1-Mlh3 is required to nick even the smallest DNA substrate tested (~1,400 bp), which is at least 30-fold larger than the size of DNA required to form a Mlh1-Mlh3-duplex DNA complex in gel-shift assays [27].
Interestingly, Mlh1-Mlh3 can cleave both circular and linear DNA substrates that are at least 7 kb in size (Figs 2E, 2F, 3E and 4C). This result correlates with the data presented in Fig 2C and 2D and shows that Mlh1-Mlh3 has higher DNA cleavage activity on larger substrates but less overall activity on linearized DNA. One way to explain this is that on circular DNA, Mlh1-Mlh3 can bind to any initial site and form a polymer. On linear DNA, however, some sites are closer to an end compared to others. If Mlh1-Mlh3 initially nucleates near the end of linear DNA, the polymer length sufficient for robust cleavage may not be achieved. This could explain why Mlh1-Mlh3 can nick a 12 kb linear substrate but not a 2.7 kb linearized substrate, as the majority of random nucleation sites within a 12-kb-long linearized plasmid will be located away from a DNA end. These data and the fact that Mlh1-Mlh3 can nick 1.4 kb and 2.7 kb circular substrates provide clues for the critical length of an Mlh1-Mlh3 polymer needed for nuclease activity (see Discussion).
The above observations suggested that disruptions in DNA that interfere with polymer formation would inhibit Mlh1-Mlh3 endonuclease activity. To test this, we synthesized a 7.2 kb circular substrate containing a single biotin moiety [35]. We titrated Mlh1-Mlh3 into reactions containing either homoduplex 7.2 kb circular DNA or 7.2 kb circular DNA with a biotin with or without bound streptavidin. All substrates were prepared by primer extension in an identical manner. In these experiments, we found that the biotin alone was sufficient to inhibit the reaction by a small amount, but the addition of bound streptavidin inhibited nicking efficiency ~1.5-fold (Fig 3A and 3B). These data suggest that a relatively small perturbation in a large DNA sequence can inhibit Mlh1-Mlh3 activity. As indicated above, we observed a similar result when an eight-nucleotide loop was incorporated into the substrate (Fig 3C and 3D). In addition, we found that the loop insertion inhibited Mlh1-Mlh3 to a greater extent when present at the center of a linear substrate compared to an end, which is consistent with Mlh1-Mlh3 polymer formation being an important requirement for activating its endonuclease activity (Fig 3E).
We next tested whether Msh2-Msh3, which preferentially binds to loop mismatches and stimulates Mlh1-Mlh3 endonuclease activity on duplex DNA [26], would overcome the inhibitory effect that the loop mismatch has on Mlh1-Mlh3 activity (S4C Fig). We also tested whether Mlh1-Mlh3 nicking on the +8 loop circular substrate was localized to the mismatch. This was accomplished by incorporating a radiolabel into the circular substrate ten base pairs away from the mismatch either on the same or opposite strand as the mismatch (S4B Fig). Under conditions when we observed Mlh1-Mlh3 stimulation by Msh2-Msh3, we first measured the total amount of nicking on the substrate in the presence and absence of Msh2-Msh3 (S4C Fig, left). We then excised from a denaturing polyacrylamide gel, a 2.3 kb fragment centrally encompassing the mismatch and the radiolabel (S4C Fig, right). On a homoduplex substrate, Msh2-Msh3 increased the amount of nicked product from 15% to 36%. In reactions containing only Mlh1-Mlh3 and the loop substrate, we observed 10% nicking of the substrate. The addition of Msh2-Msh3 increased this level to 21%. If the inhibition of cleavage of the +8 loop substrate was accompanied by shifting the distribution of nicks to positions near the mismatch, we should observe either the appearance of a specific low-molecular-weight band or a loss in a radioactive signal of the 2.3 kb fragment in the denaturing gel corresponding to the amount of substrate nicked in the agarose gel where we detect nicking on the entire substrate. Instead, we observed no significant loss of signal when the radiolabel was on either strand relative to the mismatch, similar to what was observed on the homoduplex substrate. These results indicate that the nicking observed on the +8 loop substrate was random (see Discussion).
Hall et al. observed using atomic force microscopy that yeast Mlh1-Pms1 forms long protein tracts on DNA and that these tracts associate with two regions of dsDNA in a single plasmid that do not appear to require homology [31]. They suggested that yeast Mlh1-Pms1 has higher binding affinity for larger DNA molecules and additionally shows a preference for circular over equivalent linearized DNA substrate because of the propensity of larger and circular molecules for having two regions of dsDNA in close proximity as a result of greater substrate flexibility in solution. It is possible that this is a common property of MLH proteins that is also utilized by Mlh1-Mlh3 (see Discussion). We observed similar properties for human MLH1-PMS2 (S5 Fig).
To directly observe Mlh1-Mlh3’s DNA-binding properties on large DNA substrates, we linearized and radiolabeled 2.7 to 12 kb plasmid DNAs and performed nitrocellulose filter binding assays in the presence of increasing amounts of Mlh1-Mlh3. In these assays, a single protein bound to DNA is sufficient to retain DNA on the nitrocellulose filter and be scored as bound protein. For this reason, the shape of the resulting binding curve can be used only as a rough measure of a protein’s binding properties. This analysis was performed previously to characterize the DNA-binding properties of Mlh1-Pms1 [31]. They observed sigmoidal binding curves, suggesting that Mlh1-Pms1 binds cooperatively to DNA (see [36]). We show here that Mlh1-Mlh3 displayed a modest preference for binding to larger over smaller DNA substrates with Kd values of ~40, 65, and 100 nM for 12, 7, and 2.7 kb linear substrates, respectively (Fig 4A). For the 7 kb substrate, when the amount of potassium chloride (KCl) was increased, the protein’s binding curve was sigmoidal, which suggests the presence of multiple Mlh1-Mlh3 molecules (Fig 4B).
To directly visualize whether multiple Mlh1-Mlh3 molecules bound to DNA are critical for Mlh1-Mlh3 nicking activity, we performed negative staining electron microscopy experiments on samples containing the following mixtures: circular DNA alone (Fig 5A); 300 nM Mlh1-Mlh3 alone (Fig 5B); 30 nM Mlh1-Mlh3 in the presence of DNA, which confers suboptimal nicking activity (Figs 2A, 2B and 5C); and 300 nM Mlh1-Mlh3 in the presence of DNA, which confers robust endonuclease activity on circular DNA (Figs 1A, 2B and 5D). By analyzing multiple grids and multiple locations on these grids, we observed the following: (1) Circular DNA alone displays different degrees of supercoiling but no higher-order structures (Fig 5A), and Mlh1-Mlh3 in the absence of DNA did not show any distinct structures (Fig 5B). (2) At 30 nM Mlh1-Mlh3, small protein–DNA clusters are observed that range from loosely (Fig 5C, white arrow) to tightly condensed (Fig 5C, black arrow). Naked DNA was also observed in this sample but with low frequency. The fact that under limiting Mlh1-Mlh3 concentrations, both partially and tightly condensed protein–DNA clusters were observed supports the intrinsic propensity of Mlh1-Mlh3 to multimerize on DNA. (3) At 300 nM Mlh1-Mlh3, we rarely observed loosely packed clusters or naked DNA. We found that most of the DNA was in tightly packed clusters (Fig 5D). The fact that these conditions resulted in the highest endonuclease activity suggests that the condensed clusters are not inactive coated molecules, but rather active Mlh1-Mlh3 complexes optimally capable to cleave DNA.
Our biochemical analysis suggested that large DNA molecules can accommodate multiple Mlh1-Mlh3 complexes, and interactions among these dimers can activate the endonuclease activity. We hypothesize that polymer formation licenses the Mlh1-Mlh3 cluster to introduce a nick. Our data also suggest that polymer formation can be disrupted by modifications to DNA (Fig 3). Such activities appear distinct from those seen for established HJ resolvases, which recognize and symmetrically cleave branched DNA structures (reviewed in [7]).
To further test the idea that interactions between Mlh1-Mlh3 dimers activate endonuclease activity, we performed mixing experiments in which the endonuclease dead Mlh1-mlh3D523N complex was added to reactions containing suboptimal concentrations of Mlh1-Mlh3. Mlh1-mlh3D523N, although deficient for endonuclease activity, is an intact heterodimer and retains its DNA binding properties [26,27]. Consistent with this, the mlh3D523N allele confers a mlh3Δ-like phenotype for meiotic crossing over [19]. These experiments were performed in the linear range for Mlh1-Mlh3 endonuclease activity. As shown in Fig 6, Mlh1-mlh3D523N addition (up to 75% of the MLH complex present) increased Mlh1-Mlh3 endonuclease activity to levels similar to exclusively wild-type complex addition, providing further support for polymer formation being critical for Mlh1-Mlh3 endonuclease function.
The Mus81-Mms4 structure-selective endonuclease is active on HJ substrates containing a nick directly adjacent to a branch point (reviewed in [24]). This discontinuity provides flexibility in the substrate that allows the bound heterodimer to undergo a conformational change to position the branch point in the active site of the complex [37,38]. Mus81-Mms4 then introduces a nick at a discrete point opposite the preexisting nick. Being directed by a preexisting nick in a DNA substrate is a hallmark of Mlh1-Pms1 in partially reconstituted mismatch repair reactions. In MMR, Mlh1-Pms1 introduces a nick on the same DNA strand that contains the preexisting nick, although the position of the nick can be several hundred base pairs away [28,29].
Mlh1-Mlh3 is able to convert closed circular substrate into nicked open circular and linear product (Figs 2C and 3C, for example). The appearance of the linear product suggests that one nick may encourage additional nicking to a site opposite the first nick. We observed that incubating Mlh1-Mlh3 with 12 kb plasmids (pEAE99, pEAO202, pEAE324) resulted in the exclusive formation of linear product with no nicked intermediate, an activity that was also seen for even larger (14 kb-pEAE107, 15 kb-pEAM58) substrates (Figs 2C and 7A, S6 Fig). In time course experiments performed with pEAE99, we did not observe nicked product at any time point assayed (Fig 7B, compare lanes 3–7 to background nicked circle observed in lane 2, DNaseI-treated plasmid in lane 1, and HindIII-linearized pEAE99 in lane 8). Circular DNA was directly converted into a linear product, suggesting that the two nicks on this large DNA substrate were rapidly made through a concerted mechanism. Consistent with a mechanism that creates concerted nicks to form a double-strand break, the linear fragments created by Mlh1-Mlh3 endonuclease activity could be religated to form closed circular and nicked products, as well as material that failed to enter the well of the gel, much like the religation behavior of linearized product that contains HindIII overhangs (Fig 7E). When Mlh1-Mlh3 linear products were blunt-ended by T4 DNA polymerase, the amount of closed circular product increased, and the religation appeared to behave more similarly to a blunt-end linear fragment. Curiously, when linearized, pEAE99 was used as a substrate, DSBs were not detected in native agarose gels (Fig 7D), but single-strand nicks were detected in denaturing gels (Fig 2E). We also observed concerted nicks on the 2.7 kb circular substrate; linear product was seen at time points at which nicked product was still being produced, and this was seen in both relaxed and supercoiled plasmid (Fig 7C, lanes 4–6, for example; S3 Fig). Together, these data suggest that polymer formation and possibly close-range DNA interactions are critical for Mlh1-Mlh3 to make DSBs through a concerted nicking mechanism (see below).
Human MLH1-PMS2 also converted a 12 kb circular substrate to linear product but did not do so on smaller plasmid substrates (S5A and S5G Fig). To better understand this observation, we tested MLH1-PMS2 activity on a 2.7 kb closed circular substrate and a substrate prenicked using a commercially available restriction endonuclease [26]. We did not observe the appearance of a linear product using either substrate, despite robust nicking of the closed circular substrate (S5G Fig). Nevertheless, because human MLH1-PMS2 can generate linear product on very large circular substrate, these data suggest that the ability to form a DSB is a general property of the MLH complexes and that Mlh1-Mlh3 does this efficiently.
Previously, we observed Mlh1-Mlh3–dependent conversion of a prenicked substrate to linear DNA using a substrate containing four preexisting nicks [26]. In our reactions, all DNA can be accounted for as either uncut circular, linear, or nicked circular product, indicating that Mlh1-Mlh3 does not introduce an abundance of nicks into the substrate. This suggests that the linear product is not formed upon frequent nicking at random positions. Supporting this, we observed the conversion of nicked into linear DNA under conditions in which only ~50% of closed circular DNA was nicked (Fig 8A lane 2, and in [26]). Once formed, linear product of this size does not support endonuclease activity (Fig 2A). The endonuclease-deficient Mlh1-mlh3D523N was unable to make DSBs, indicating that this is an intrinsic property of the Mlh1-Mlh3 endonuclease (Fig 8B). These observations indicate that the nicks did not occur at a large number of sites on the substrate that would produce linear DNA when two nicks are located close to each other by chance. Rather, our data show that the endonuclease cleavage sites on nicked substrates were not distributed randomly.
We entertained three explanations for the above observations: (1) Preexisting nicks are preferential loading sites for Mlh1-Mlh3 polymer formation and direct Mlh1-Mlh3 to nick directly opposite the preexisting nick. This proposed mechanism is similar to that seen for the Mus81-Mms4 resolvase (reviewed in [24]). (2) Preexisting nicks have no impact on Mlh1-Mlh3’s ability to form a DSB, but the complex introduces two nicks, making a DSB at a site independent of the position(s) of preexisting nicks. This idea is suggested by our finding that linear product is formed to similar extents by using closed plasmid and prenicked plasmid as substrates [26] and by the time course experiments presented in Fig 7B and 7C. (3) In the absence of a preexisting nick, an Mlh1-Mlh3 polymer can introduce a nick that is then used by the active polymer to introduce an additional nick on the opposing strand. In this model, a preexisting nick serves as a preferential site for an Mlh1-Mlh3–generated nick on the opposite strand but does not serve as a preferential loading site. Such a model argues that Mlh1-Mlh3’s endonuclease acts distinctly from Mus81-Mms4 and is supported by the finding that the amount of linear product is similar in reactions containing closed circular or prenicked substrates [26].
To differentiate between the above mechanisms and address whether preexisting nicks are used as preferential substrates for Mlh1-Mlh3, we generated nicked substrates containing one, four, or ten preexisting nicks in the pUC18 plasmid. If preexisting nicks act as preferential loading sites for Mlh1-Mlh3, we would expect an increase in linear product as the number of preexisting nicks increases, and if each nick is used as a recognition site, we would expect to see smaller linear fragments or perhaps a smear on the agarose gel. The amount of linear product was comparable for closed circular substrate or circular substrate with any of the number of preexisting nicks tested (Fig 8A). These data imply that formation of the DSB is likely independent of discontinuities in the DNA substrate and that nicks are not preferential loading sites for an Mlh1-Mlh3 polymer (favoring models 2 and 3).
If model 3 is correct, then we should be able to map a DSB to a specific location in the plasmid. On closed circular DNA, Mlh1-Mlh3 generated nicks are not made at sequence-specific locations or near any known secondary structure on the plasmid. To determine where Mlh1-Mlh3–generated nicks are located relative to a preexisting nick, we used a substrate that contains one nick introduced by a restriction-nicking endonuclease. We incubated this substrate with Mlh1-Mlh3, isolated the linear product, and then annealed primers either complementary to the strand with the restriction nick (primer A) or complementary to the opposing strand (primer B) approximately 60 nt away from the restriction nicking site. A primer extension assay was then performed. In this assay, primer extension terminates at the site of a nick, allowing us to determine if Mlh1-Mlh3 is nicking immediately opposite the preexisting nick (Fig 8C, top). Because of the low amount of linear product generated from this substrate (Fig 8A), we surmised that the preexisting nick does not preferentially act as a loading site for Mlh1-Mlh3. If, however, Mlh1-Mlh3 loaded near the preexisting nick by chance uses a single preexisting nick to direct a nick to a specific site in the opposite strand, we expected to detect a discrete band that would be visible when generated in the primer extension assay and analyzed by denaturing polyacrylamide gel electrophoresis (PAGE). If the preexisting nick only directs nicking to the region but not to a specific nucleotide, we would observe a smear on the gel. If Mlh1-Mlh3 does not use the preexisting nick as a preferred site and the DSB occurs completely independently of the preexisting nick, we will observe extension of the primer to a site much farther than 60 bp away from the preexisting nick on the 2.7 kb substrate.
As a control, we performed the primer extension assay with substrate generated by using a restriction enzyme that linearizes the plasmid by cutting at the site of the preexisting nick and at a site 3 bp away (Fig 8C, top). For the linear control, when primer A was extended, a 60 nt product formed (Fig 8C, bottom, lane 3). When primer B was extended, a 63-nt nucleotide product formed (Fig 8C, bottom, lane 5). For the linear product generated by the Mlh1-Mlh3 reaction, when primer A was extended, a 60-nt product formed (Fig 8C, bottom, lane 8). This corresponds to the site of the preexisting nick. When we extended primer B, we observed a single band at approximately the same intensity as the unextended primer, migrating at ~60 nt (Fig 8C, bottom, compare lane 10 to 9 for intensity and lane 10 to 11 and 12 for size).
The above data indicate that if the Mlh1-Mlh3 complex is loaded near a preexisting nick, it uses the nick to direct endonuclease activity to a specific site precisely opposite this initial precursor nick when present in the substrate. Again, because of the low yield of this linear product, we do not believe that this is a preferred loading site for the polymer. Together, these data suggest that an active Mlh1-Mlh3 polymer can introduce a DSB made by concerted nicks in a mechanism distinct from Mus81-Mms4. If loaded by chance near a preexisting nick, however, the nick can be used as a landmark for Mlh1-Mlh3 to introduce a nick on the opposite strand. With no preexisting nick, nicking is otherwise random, as observed by analysis with denaturing gels (Figs 2A, 2E and 3E). If there were hotspots for nicking on covalently closed substrate, discrete bands would be observed in these gels as opposed to a smear. These observations also provide a hint for how Mlh1-Mlh3 could be directed by a loading factor to specifically cleave recombination intermediates such as dHJs (see Discussion).
The finding that Mlh1-Mlh3 can create concerted DSBs on large circular but not equivalent linear substrates where it made nicks, and that it did not make nicks on smaller linear substrates, suggested that close-range interactions and/or synapsis between DNA molecules could license Mlh1-Mlh3 polymer to display endonuclease activity “in trans.” To test this idea, we performed reactions in which a closed circular 7.2 kb substrate, which is nicked by Mlh1-Mlh3, and a 2.7 kb linear substrate, which is not nicked, were incubated together with Mlh1-Mlh3. As shown in Fig 9, the 2.7 kb linear substrate was nicked only when incubated in the presence of the 7.2 kb closed circular substrate. This observation is consistent with an Mlh1-Mlh3-DNA complex being able to interact with DNA substrates in trans (see Discussion).
Little is known about the mechanism by which Mlh1-Mlh3 acts to resolve meiotic recombination intermediates to form COs. This has been a challenging effort because Mlh1-Mlh3 has little in common with the well-characterized structure-selective endonucleases (e.g., Mus81-Mms4, Slx1-Slx4, and Yen1), both in terms of homology and intrinsic behavior in vitro. The absence of a biochemical paradigm provided by other structure-selective endonucleases makes it difficult to model Mlh1-Mlh3’s role in HJ resolution. Mlh1-Mlh3 likely relies on other protein factors, including Msh4-Msh5, to recruit and coordinate endonuclease activity and collaborates with other factors to spatially and temporarily coordinate the resolution of dHJs into COs in a mechanism that is distinct from the previously established archetypes set by structure-selective endonucleases.
Here, we showed that Mlh1-Mlh3 polymer formation is a requirement for its in vitro endonuclease activity (Fig 10A and 10B), and that Mlh1-Mlh3 is capable of making DSBs through a concerted nicking mechanism (Fig 10B). Strikingly, we observed that the Mlh1-Mlh3 polymer cleaves long duplex DNA but not small DNA substrates containing loop and branched structures. These observations are reminiscent of the involvement of multiple Mlh1-Pms1 molecules during MMR [30] and the finding that Mlh1-Pms1 does not nick mismatched DNA directly at the site of the mismatch [28,29]. For this reason, we cannot exclude the possibility that such activities may be more critical for mismatch correction than crossing over. The presence of distinct Mlh1 and Mlh3 foci at CO sites in vivo also supports a necessity for multiple Mlh1-Mlh3 heterodimers being required for resolution, since a large number of molecules are needed for foci to be visible. The requirement for polymerization for nuclease activation provides a regulatory mechanism to precisely control nuclease activity. Such a model also predicts that the CO endpoints that involve Mlh1-Mlh3 may be different than those seen by resolution with structure-specific endonucleases that make precise incisions, as DNA cleavage may occur at locations away from the junctions analogous to MMR mechanisms. DNA cleavage followed by branch migration is sufficient to result in CO products (Fig 10C).
Taken together, our data argue strongly against Mlh1-Mlh3 acting on HJs using a Mus81-Mms4–like mechanism. Mus81-Mms4 and other structure-selective endonucleases recognize branched molecules and cleave at precise, discrete locations at the branch points to resolve them. Based on molar ratios and substrate requirements in vitro, there are no polymerization requirements for these canonical resolvases (reviewed in [24]). Mus81-Mms4 and Yen1’s endonuclease activities are also sensitive to their phosphorylation state [25]. Our Mlh1-Mlh3 expressed in Sf9 cells was unaffected by phosphatase treatment, which may suggest that Mlh1-Mlh3 is not regulated by phosphorylation, although we cannot rule out an effect of phosphorylation by cognate factors during meiosis in vivo or other posttranslational modifications such as SUMOylation. It should be noted that protein components involved in formation of the synaptonemal complex and stabilization of Msh4-Msh5 have been found to be SUMOylation enzymes (reviewed in [39]) [40–44]. It is possible that such a modification could have an as yet unknown regulatory function on Mlh1-Mlh3, though posttranslational modifications have yet to be implicated in the activity of other MLH family complexes.
Our data indicate that Mlh1-Mlh3 is unlikely to act directly on Holliday junctions and mismatched substrates. We showed that incorporating a +8 loop mismatch into a plasmid substrate has an inhibitory effect on endonuclease activity and does not localize it, and cruciforms present in plasmids are not substrates for Mlh1-Mlh3 cleavage [27]. These observations suggest that HJs and mismatched substrates are not likely to be the preferred in vivo substrates for the Mlh1-Mlh3 endonuclease. Previously, we showed that Mlh1-Mlh3 has a binding preference for oligonucleotide substrates containing a +8 loop mismatch and a HJ over homoduplex DNA but that the binding affinity differences in buffers used in our endonuclease assay were modest [26,27]. However, in the oligonucleotide competition endonuclease assay presented here (Fig 1), HJ and +8 loop–containing oligonucleotide substrates acted as very strong competitors relative to homoduplex DNA, which was an ineffective competing substrate compared to plasmid DNA. Although having preferences for certain structures, Mlh1-Mlh3 is an overall efficient DNA-binding complex [26,27]. Mismatch recognition factor Msh2-Msh3, while preferring 8-nt loop mismatches, does not bind mismatch DNA particularly tightly (Kd ~50 nM for loop mismatch; Kd ~200 nM for homoduplex) [45]. In experiments where an 8-nt loop mismatch was incorporated into a 7.2 kb plasmid substrate, it should be noted that the portions of the plasmid substrate that do not contain the mismatch may serve as competitor DNA (S4 Fig). Since Mlh1-Mlh3 binds DNA at least as tightly as Msh2-Msh3, it follows that Mlh1-Mlh3 may be competing with Msh2-Msh3 for the loop mismatch as well as the remaining duplex DNA, which is a possible explanation for why we do not observe the ability of Msh2-Msh3 to direct Mlh1-Mlh3 nicking to sites near the mismatch.
We hypothesize that Mlh1-Mlh3 forms a polymer on DNA to activate its endonuclease activity (Fig 10A). We suggest that there is a critical polymer length (on the order of 1 kb; S5B Fig) needed to activate the endonuclease activity that is achieved more readily on a circular substrate where all initiating sites are equivalent. It is important to note that one nick is sufficient to convert closed circular substrate to nicked product, and because nicking occurs in random locations on the plasmid, quantifying the number and location of nicks is not feasible. This is evidenced by the lack of a discrete product in denaturing agarose gels (Figs 2A, 2E and 3E). It should also be noted that our experiments suggest Mlh1-Mlh3 makes only a few nicks per substrate molecule. The entirety of the DNA is accounted for after the reaction as either uncut closed circular substrate, nicked circular product, or linear product. If an abundance of nicks were generated, we would observe loss of DNA density in the native agarose gels. Analysis by alkaline agarose gel suggests that nicks are being made in both strands of the duplex because we observe loss of DNA density in these experiments when high concentrations of Mlh1-Mlh3 are added to the reaction.
Our observation that a 2.7 kb circular plasmid can be nicked but the linearized form cannot (Fig 2A and 2B) indicates that the Mlh1-Mlh3 polymer is directional (Fig 10B). Without directionality, one would expect a polymer to form on the linear substrate in different orientations and thus be able to nick the substrate. Although we do observe binding to this linearized substrate, it is insufficient to promote nuclease activity. Together, these data suggest that a unidirectional polymer is required for endonuclease activation.
Why does Mlh1-Mlh3 make DSBs on large circular plasmids but nick smaller ones? Hall et al. showed through atomic force microscopy that yeast Mlh1-Pms1 forms polymers on plasmid DNA and can simultaneously interact with two regions of the substrate [31]. One way to explain this observation is that a single MLH heterodimer simultaneously interacts with adjacent regions of dsDNA. Alternatively, polymers of MLH proteins located at two different locations on a plasmid DNA interact with each other (Fig 10B, panel 2). An observation similar to that made for Mlh1-Pms1 has been made for type IB DNA topoisomerases (TopIBs) [46,47]. Using both atomic force microscopy and electron microscopy, TopIB, in the presence high protein concentrations relative to DNA, forms a polymer on DNA and brings distant, noncontiguous regions within the same circular DNA molecule into close proximity (intramolecular synapsis). When a linear substrate was used, regions of DNA not necessarily within the same molecule were synapsed (intermolecular synapsis). Analysis of a cocrystal structure of TopIB in complex with DNA showed that TopIB has two distinct DNA-binding regions and that a single TopIB polymer can simultaneously interact with two regions of dsDNA [48], in contrast to a model in which two polymers, each interacting with DNA, form a complex. It will be valuable to perform analogous experiments with the MLH proteins to determine if they undergo an intramolecular synapsis mechanism similar to that seen for TopIB.
We observed a substantial overlap in biochemical properties of Mlh1-Mlh3 and MutLα, both of which act in MMR [32]. Similar to Mlh1-Pms1, Mlh1-Mlh3 may associate with two nearby regions of dsDNA, possibly through a synapsis-type mechanism in which two nucleoprotein complexes interact (Fig 10B, panel 2). Larger substrates are more likely to accommodate multiple Mlh1-Mlh3 polymers of sufficient length to activate nicking. These regions of DNA coated with Mlh1-Mlh3 may then be able to interact with one another on these large substrates. Such properties could thus account for the exclusive conversion of closed circular substrates into linear products for plasmids greater than 12 kb, perhaps by creating a nucleoprotein conformation that restricts nicking activity to one of the two interacting duplexes. Our observation that a small linear substrate can be nicked only when incubated in the presence of a larger closed circular substrate is consistent with the idea that Mlh1-Mlh3 can associate with nearby regions of dsDNA through a synapsis-type mechanism. In this case, the nucleoprotein conformation permits nicking activity to an interacting duplex in trans.
Our data indicate that Mlh1-Mlh3 is not creating a large number of nicks on a given DNA molecule, suggesting that an activated polymer may introduce as few as one nick per substrate. Our data do not give a clear understanding, however, of where within an active polymer a nick is created. In our assays, Mlh1-Mlh3 polymers bind and nick dsDNA nonspecifically, so mapping a nick to a unique heterodimer within a polymer is not feasible. One possibility is that an interior heterodimer within the polymer enjoys a high degree of stabilization and that there is a critical polymer length to optimize this stabilization, which triggers nicking. Introducing interruptions, such as insertion loops, interferes with achieving optimal stabilization. A second possibility is that the first Mlh1-Mlh3 molecule that binds DNA to initiate stable polymer formation introduces a nick. Mlh1-Mlh3 has a binding preference for branched DNA substrates and DNA with insertion loops. Incorporating features such as insertion loops into circular DNA inhibits Mlh1-Mlh3 activity, however. One interpretation of these data is that an Mlh1-Mlh3 polymer nucleates from the insertion loop because of it being a preferred binding site, but such binding distorts the conformation of the initial Mlh1-Mlh3 heterodimer so that it is inhibited from nicking. Our data in Fig 6 argue against this model, however. When wild-type and endonuclease-inactive proteins are mixed together at up to a 1:3 stoichiometric ratio, one would expect that if the initiating heterodimer were responsible for nicking, the presence of Mlh1-mlh3D523N would be inhibitory. We therefore favor a model in which an Mlh1-Mlh3 heterodimer interior to the polymer is activated and introduces nicks. The addition of other factors, such as Msh4-Msh5 or Exo1, may regulate the extent of polymerization that is required to induce Mlh1-Mlh3 nicking.
How are dHJs resolved into COs in the Msh4-Msh5/Mlh1-Mlh3 pathway? Resolvases, such as Mus81-Mms4 and Yen1, which cleave dHJs in an Msh4-Msh5–independent pathway, produce both COs and NCOs, suggesting that they cleave each HJ in an independent manner. Such a mechanism does not necessitate a precise orientation of the endonuclease and only requires a nuclease to recognize and cleave DNA within a junction. In the Msh4-Msh5/Mlh1-Mlh3 pathway, only COs form as the result of dHJ resolution. Since Mlh1-Mlh3 is the major nuclease activity required in this pathway, either it is “presented” with an inherently asymmetric substrate or it must be organized in different orientations at each junction to facilitate asymmetric cleavage. The finding that in MMR the human MLH1-PMS2 endonuclease activity is strand-specifically activated by PCNA suggests that meiotic factors could regulate Mlh1-Mlh3 endonuclease specificity in an analogous way [49].
In meiosis, the two HJs present in a dHJ are not created simultaneously, and identical proteins may not occupy them prior to resolution. In the leptotene stage of meiotic prophase I, each DSB is resected to form two 3′ single-strand overhangs, one of which invades the homologous chromosome. The invading end is extended and stabilized by ZMM proteins, including Msh4-Msh5, in zygotene, creating the first HJ in the dHJ intermediate. In early pachytene, the newly synthesized invading strand can reanneal to the other side of the DSB in a process involving RPA, Rad52, Rad54, and potentially Zip3 to create the second HJ [16,50–58]. This process suggests how an asymmetry could be created between the two junctions to exclusively resolve dHJs into COs. It also implies a coordinated assembly of the dHJ substrate and that the junctions may be at least partially protected from Mus81-Mms4, which acts at the same time as Mlh1-Mlh3 during Meiosis I [25]. The complex assembly of the substrate and the in vivo requirements for other factors in creating and stabilizing the substrate strongly support the proposal that other protein factors, such as Msh4-Msh5 or other ZMM proteins, are present in precise positions prior to Mlh1-Mlh3’s activities and that these factors are likely critical for recruiting and orienting Mlh1-Mlh3. Msh4-Msh5 is explicitly implicated to function upstream of Mlh1-Mlh3 by meiotic crossover control experiments in which a mms4Δ mlh1Δ baker’s yeast strain exhibited a 13- to 15-fold decrease in crossing over compared to the wild type, but a mms4Δ mlh1Δ msh5Δ triple mutant only showed a 5-fold decrease, similar to the decrease observed in a mms4Δ msh5Δ double mutant [20]. We hypothesize, based on relative DNA-binding affinities between MLH proteins and MSH proteins, that once recruited, the Mlh1-Mlh3 polymer displaces Msh4-Msh5 [27,31,45] (Fig 10A and 10C). Such an idea is consistent with work in Sordaria in which Msh4 foci were observed to diminish between early and midpachytene, a time frame in which Mlh1-Mlh3 is believed to be recruited [8]. It is also supported by an experiment analogous to those shown in Fig 2 in which hydrolytically inactive EcoRI(E111Q) was tested as a roadblock for activation of the Mlh1-Mlh3 endonuclease. Although EcoRI(E111Q) bound to DNA, it did not have an inhibitory effect on Mlh1-Mlh3 activity (S7 Fig).
In addition to the ZMM and second-end capture factors, a candidate for an in vivo specificity factor is the nuclease-independent activity of Exo1, which was shown by Zakharyevich et al. to be required in conjunction with Mlh1-Mlh3 and Sgs1-Top3-Rmi1 in dHJ resolution [13,18]. It is known that Mlh1 physically interacts with Exo1 and that such an interaction may serve the purpose of stabilizing and thus activating Mlh1-Mlh3 on dHJ substrates in addition to interactions that are likely critical for MMR [59,60]. It is likely that in order to model HJ resolution by Mlh1-Mlh3 in vitro, these additional factors will need to be present and positioned appropriately. It is also likely that interactions with these other factors stabilize Mlh1-Mlh3 in such a way that a smaller amount of Mlh1-Mlh3 is required for activation than we can demonstrate in vitro in the absence of other factors. This is supported by the stimulation of Mlh1-Mlh3 by MMR factor Msh2-Msh3, in which the amount of nicked product observed with 100 nM Mlh1-Mlh3 can be observed with 25 nM Mlh1-Mlh3 in the presence of Msh2-Msh3 (S4 Fig; [26]).
Can we reconcile the Mlh1-Mlh3 enzymatic activities observed here with asymmetric resolution of dHJs? Endonuclease activity being contingent upon Mlh1-Mlh3 polymer formation would regulate the endonuclease activity in vivo and prevent promiscuous nicking of DNA substrates. In vivo, it is unlikely that there would be sufficient naked DNA to accommodate a polymer that can nick DNA at nonspecific sites as we observe in our in vitro assays. It would also restrain the double-strand break activity that we observe on very large substrates. During dHJ resolution, there are likely to be constraints imposed upon the endonuclease by both the substrate and other protein factors that tightly regulate the endonuclease activity and promote polymer formation and activation at specific sites. We also observe that disrupting the continuity of dsDNA has an inhibitory effect on endonuclease activity. This suggests that pure junction recognition, followed by polymer formation and activation, is unlikely to take place in vivo. It is likely that a very specific substrate and positioning of other factors, including Msh4-Msh5, provide a unique substrate for Mlh1-Mlh3 to act on (Fig 10C). In the absence of a fully reconstituted system, we provide early models for how dHJ resolution occurs in vivo. Below, we put forward possible models for dHJ resolution that exploit the biochemical properties presented here.
To our knowledge, there is no direct evidence that dHJs are covalently closed in vivo. If they are not, nicks would likely be present in dHJ structures at sites of DNA synthesis termination (Fig 10D, top). In such a scenario, an endonuclease recruited to a dHJ could nick the strand opposite to the preexisting nick present in an unligated HJ. Such a model has been previously suggested for a generic HJ resolvase activity [61]. One drawback of this model is that Mlh1-Mlh3 does not appear to recognize and cleave HJs like Mus81-Mms4. If DNA synthesis terminates within a junction, the resultant substrate is a nicked junction, which is recognized and cleaved efficiently by Mus81-Mms4. Thus, collaboration with a structure-selective endonuclease may be required (Fig 10D, right). If DNA synthesis terminates such that the nicks are some distance away from the junctions, an Mlh1-Mlh3 polymer could form and be directed to nick specifically upon encountering the unligated duplex (Fig 10D, left). Since DSBs are neither desired nor required for CO formation, this would necessitate an Mlh1-Mlh3 polymer to nick the duplex that does not contain the preexisting nick (nicking in trans). The ability of Mlh1-Mlh3 to nick in trans is supported by our observations in Fig 9. Such a model may also couple Mlh1-Mlh3 polymer directionality with DNA polarity, which we have not explicitly observed in vitro.
Modeling of recombination intermediates suggested that dHJs can isomerize between two different geometries (type I and II; [62]). Type I represents the archetypal depiction of a double Holliday junction (Fig 10E). Type II is an alternate configuration of a dHJ, which, like type I, involves a twisting or isomerization of one of the two HJs from a common intermediate ([62]; Fig 10E). The type II configuration provides an attractive model to integrate the Mlh1-Mlh3 biochemical activities described here; for resolution of this geometry into crossover products, cleavage can take place symmetrically by using an Mlh1-Mlh3 polymerization and nicking mechanism (Fig 10F; see below). A challenge in explaining how such geometries are relevant to recombination is why a type II configuration would be favored, as well as the steric issues associated with isomerization, which could include chromosome arm rotations.
Based on a molecular analysis of zip1 mutants in yeast, Storlazzi et al. provide an elegant argument for the maintenance of a type II configuration [58]. They described an early role for Zip1 at the transition from resected DSBs to strand invasion steps: specifically, COs were significantly decreased in zip1 mutants. Based on a mechanical stress model developed to explain CO patterning, they suggested that Zip1 acts in a CO differentiation pathway to preisomerize dHJs into type II junctions committed to a CO pathway. Such a model is attractive because it provides a way to deal with the steric issues outlined above. Consistent with this idea is a recent analysis in Sordaria of Hei10, a Zip3 family protein that has been implicated as a structure-based signaling molecule acting in several pathways, including meiotic recombination [42]. DeMuyt et al. observed, using three-dimensional structured illumination microscopy, that ~50% of Hei10 foci localize to sites of twists in the synaptonemal complex (SC) [42]. The authors argue that “this structural distortion points to structural and/or geometric interplay between the DNA events of CO formation and the SC.” This analysis may also be consistent with work performed by Oke et al. in observing recombination products in yeast, wherein the authors suggest that Zip3 is required for the conversion of dHJs into exclusively CO products [51]. To this point, if we invoke the presence of a type II double Holliday junction substrate in CO formation, one can imagine Mlh1-Mlh3 being recruited and forming a polymer to cut at both junctions (Fig 10F). Perhaps between the two junctions there is inadequate DNA length to accommodate the critical length of an Mlh1-Mlh3 polymer.
In budding yeast, the two junctions of a dHJ have been estimated to be either a few hundred base pairs away from each other [63,64] or converged to point junctions [64]. The requirement for a helicase/topoisomerase (Sgs1-Top3-Rmi1) in the Msh4-Msh5/Mlh1-Mlh3 pathway also suggests that junctions may converge or nearly converge, creating a structure that Mlh1-Mlh3 is capable of cleaving [21,23]. Such a mechanism would also require other factors to constrain Mlh1-Mlh3 polymer formation so that nicking can be localized. It should be noted, however, that nicking does not need to explicitly take place at a HJ to yield a CO product. Nicks could be produced away from the junction. A subsequent branch migration step would then be sufficient for resolution.
Our data in Fig 9 support the idea that Mlh1-Mlh3 can interact with two DNA molecules simultaneously and that these interactions can stimulate the endonuclease to act on a substrate that it does not act on in the absence of the second substrate. These data in conjunction with observations for Mlh1-Pms1 presented by Hall et al. [31] support a synapsis-type mechanism. Such an activity could be used in meiosis to permit nicking to interacting duplex regions in trans. Orientation of Mlh1-Mlh3 by other factors could aid in directing the polymer to nick so that CO products are generated.
Mlh1-Mlh3 is a member of the MLH family of MMR proteins and plays a minor role in repairing mismatches recognized by Msh2-Msh3 [3,4]. Models for MMR in eukaryotes suggest that mismatches are recognized by MSH factors, which recruit MLH factors to specifically cleave the newly replicated strand. The endonuclease activity of Mlh1-Pms1 (MLH1-PMS2 in humans) is activated through interactions with the DNA replication processivity clamp PCNA [28,29]. Pluciennik et al. propose that asymmetric loading of PCNA is responsible for directing the Mlh1-Pms1 endonuclease to cleave the mismatch-containing strand [49]. Together, these observations support the idea that MLH endonuclease activities are directed by other factors and are consistent with the biochemical studies of Mlh1-Mlh3 presented here. Interestingly, work by the Crouse laboratory suggests that Mlh1-Mlh3 acts in conjunction with Mlh1-Pms1 during MMR [4]. Mlh1-Mlh3 appears to lack residues present in other MutL proteins that appear critical for activation by the DNA replication processivity clamp [65]; consistent with this, Mlh1-Mlh3 endonuclease is not activated by RFC and PCNA in vitro [26,27]. These observations suggest that during MMR, Mlh1-Mlh3 is recruited and activated by Msh2-Msh3 but must retain an intimate association with Mlh1-Pms1, which is presumably oriented by PCNA, to coordinate MLH endonuclease activities with strand-specific repair.
Sequences of all oligonucleotides used in this study are listed in S1 Table.
For experiments in Fig 1, homoduplex and +8 loop substrates were constructed from AO3142 and either AO3144 (homoduplex) or AO3143 (+8) as described [26,45]. HJ substrate X26 [66], with a 26-bp homologous core allowing branch migration, was constructed from AO3147 (X26-1), AO3148 (X26-2), AO3149 (X26-3), and AO3150 (X26-4). Annealed substrates were purified by HR S-300 spin columns (GE Healthcare).
Yeast wild-type Mlh1-Mlh3 and Mlh1-mlh3D523N were purified from baculovirus-infected Sf9 insect cells as previously described [26]. Human MLH1-PMS2 was a gift from Peggy Hsieh’s laboratory and was purified as previously described [67,68]. Yeast Mlh1-Pms1, Msh2-Msh3, RFC, and PCNA were purified from yeast as described previously [69–72].
For experiments in S1 Fig, either 200 units of lambda protein phosphatase (NEB) or 50 units of CDK1-cyclinB (NEB) was combined with purified Mlh1-Mlh3 (1.2 μM final concentration) in the supplied buffer in a 15 μL reaction. The reaction was allowed to proceed according to the manufacturer’s instructions. For CDK1-cyclinB, we performed a control with Mlh1-Mlh3 and Ɣ 32-ATP and with Ɣ 32-ATP alone, followed by loading onto a Bio-Rad P-30 spin column. After centrifugation, Mlh1-Mlh3 protein elutes from the column, but Ɣ 32-ATP is retained. The specific activity of the probe was 30 CPM/fmol. Approximately 16,000 fmol of Mlh1-Mlh3 was reacted with ~20,000 fmol of Ɣ 32-ATP (conditions reported in the Materials and methods; except for the assayed complex, no radioactivity was used, and a larger excess of ATP was included). For the reaction that did not contain Mlh1-Mlh3, the elutant did not contain radioactive signal above background (as measured by a Geiger counter). For the reaction that contained Mlh1-Mlh3, the elutant was radioactive (~500,000 CPM, which is equivalent to ~16,000 fmol of material given the specific activity of the probe), indicating that a majority of the protein was radiolabeled if only one site per complex was phosphorylated.
With respect to the lambda protein phosphatase experiments, we treated the radiolabeled Mlh1-Mlh3 above with lambda phosphatase. After the spin column step, the eluant no longer gave a radioactive signal above background, as measured using a Geiger counter. We also performed a spectroscopic control using p-nitrophenyl phosphate using the same conditions used to treat Mlh1-Mlh3. Those conditions converted 100% of the p-nitrophenyl phosphate to p-nitrophenol, which is detected at 405 nm and has an extinction coefficient of 18,000 M-1cm-1. We also ran an SDS-PAGE on Mlh1-Mlh3 samples treated with lambda protein phosphatase and did not observe a mobility shift, which suggests that the complex is at least not hyperphosphorylated.
The synthesized 7.2 kb circular DNA substrates in Fig 3 and S4 Fig were generated essentially by using the protocol described by Baerenfaller et al. [35]. HPLC-purified primers were obtained from IDT. Homoduplex 7.2 kb circular substrate was generated from AO3266 and biotinylated 7.2 kb circular substrate was generated from BIO_M13mp18. +8 loop–containing 7.2 kb circular substrate was generated from AO3267 (with loop mismatch disrupting BmrI restriction site). Primers were phosphorylated on their 5′ ends by T4 polynucleotide kinase (NEB) in a 50 μL reaction in the supplied buffer containing 6 μM primer, 1 mM ATP, and 10 units enzyme at 37°C for 60 min. Kinase was inactivated for 20 min at 65°C. Unincorporated nucleotide was removed using a P-30 spin column (Bio-Rad) according to the manufacturer’s instructions. Phosphorylated primer (1 μM) was annealed to 250 nM M13mp18 ssDNA template (Affymetrix) in a 25 μL reaction in a buffer containing 50 mM Tris-HCl (pH 7.5), 10 mM MgCl2, and 10 mM DTT by incubating for 6 min at 85°C then cooling to room temperature overnight. Annealed primers were extended and closed circular substrates were generated by incubation in a 50 μL reaction containing 96 nM primer template with 1 mM dNTPs, 30 units T4 DNA polymerase (NEB), and 1,000 units T4 DNA ligase (NEB) in a buffer containing 50 mM Tris-HCl (pH 7.5), 100 μg/mL BSA, 10 mM MgCl2, 10 mM DTT, and 1 mM ATP. Reactions were incubated at 37°C for 60 min, followed by enzyme inactivation at 70°C for 20 min. Closed circular DNA was isolated from ssDNA and open circle DNA by resolution in a 0.8% agarose gel and extraction using a QIAGEN gel extraction kit. The presence of the +8 loop mismatch was verified by sequencing and lack of BmrI restriction enzyme digestion. Radiolabeled circular substrate used in S4 Fig was prepared by an identical method to the above with the exception that primers AO3266, AO3267, or AO3346 (for generating substrate in which the radiolabel is on the strand opposing that containing the mismatch) were phosphorylated by incubation with Ɣ32-ATP (Perkin Elmer) prior to annealing.
For experiments in Fig 2C, S2 Fig, Fig 7, and S5 Fig, pUC18 2.7 kb and pBR322 4.4 kb closed circular plasmids were purchased from Invitrogen. The 7 (pEAE399), 12 (pEAE99, pEAO202, pEAE324), 14 (pEAE107), and 15 (pEAM58) kb plasmids were amplified and mini-prepped from DH5α-competent cells by standard methods. Prenicked circular substrates used in Fig 8 were generated as previously described using Nt.BspQI, Nt.BstNBI, or Nt.AlwI purchased from New England Biolabs [26].
The 2.7, 7 kb, and 12 kb linear DNA substrates were generated by digesting pUC18, pEAE399, or pEAE99 with HindIII (NEB) according to the manufacturer’s instructions. Reactions were incubated at 37°C for 60 min, followed by enzyme inactivation at 80°C for 20 min. Linearized fragment was isolated via resolution by agarose gel and gel extraction (QIAGEN). Biotinylated pUC18 used in Fig 3 was prepared by incubating the HindIII fragment with 1 mM dATP, dCTP, biotin-11-dGTP (Perkin Elmer), and dTTP and 10 units of Bsu DNA polymerase large fragment (NEB) in the supplied buffer at 37°C for 60 min, followed by enzyme inactivation at 75°C for 20 min. Excess nucleotide was removed using a P-30 spin column (Bio-Rad) according to the manufacturer’s instructions. Streptavidin was bound to biotinylated DNA by incubating 50 nM DNA with 1 μM streptavidin (NEB) at room temperature in the endonuclease reaction buffer for 15 min immediately prior to use. The presence of streptavidin was confirmed by gel shift.
Substrates used in Fig 3E were generated identically to those used in Fig 3C to created circularized versions. The homoduplex or +8 loop mismatch–containing substrate was then linearized with the indicated restriction enzyme (NEB) according to the manufacturer’s instructions. Linearized DNA was then gel isolated to be used in endonuclease assays.
For radiolabeled linear substrates used in filter binding reactions, pUC18, pEAE399, or pEAE99 HindIII fragments were extended by Bsu DNA polymerase large fragment using α 32-dATP by a method analogous to that used to biotinylate linear DNA fragment above.
Endonuclease reactions were performed as previously described [26] in endonuclease buffer: 20 mM HEPES- KOH (pH 7.5), 20 mM KCl, 0.2 mg/mL BSA, 1% glycerol, and 1 mM MgCl2 unless otherwise indicated. Reactions were stopped by the addition of a stop mix solution containing final concentrations of 0.1% SDS, 14 mM EDTA, and 0.1 mg/mL ProteinaseK (NEB). Products were resolved by 1% agarose gel containing 0.1 μg/mL ethidium bromide, which results in covalently closed circular DNA isoforms migrating similarly to supercoiled DNA. Gels were run in 1x TAE (Tris-acetate-EDTA; 40 mM Tris base, 20 mM acetic acid, 1 mM EDTA) at 100 V for 40 min or by denaturing agarose gel where indicated. For experiments with circular DNA substrate analyzed by denaturing agarose gel, circular DNA was linearized with HindIII for 60 min after incubation with Mlh1-Mlh3 but prior to the addition of stop mix. Linear substrate was treated identically. Denaturing agarose gels consist of 1% (w/v) agarose, 30 mM NaCl, 2 mM EDTA (pH 7.5) run in a buffer containing 30 mM NaOH and 2 mM EDTA. Prior to sample loading, reactions were diluted in five volumes of buffer containing 180 mM NaOH, 6 mM EDTA, 20% glycerol, 0.1% xylene cyanol, and 0.1% bromophenol blue, heated for 5 min at 70°C, then cooled for 3 min on ice. Gels were run at 50 V for ~3 h. After running, alkaline agarose gels were neutralized in 0.5 M Tris-HCl (pH 7.5) for 30 min and stained with 0.5 μg/mL ethidium bromide for ~2 h. To generate a marker for closed circular, linear, and nicked 12 kb plasmid, pEAE99 was treated for 2 min with 0.1 μL of a 1:1,000 dilution of DNaseI (NEB) at 37°C. Quantifications were performed using GelEval (FrogDance Software, v1.37), and negative control reactions were used for background subtractions.
pUC18 plasmid was either nicked with Nt.BspQI or linearized using SapI (NEB) according to the manufacturer’s instructions with heat inactivation and gel isolation of the nicked or linear product. Nicked plasmid was then used as an endonuclease substrate in a reaction performed with 24 replicates containing 300 nM Mlh1-Mlh3. After stopping the reactions, replicates were combined and the DNA was ethanol precipitated by standard methods. The concentrated DNA sample was then resolved by 0.8% agarose gel, and the linear endonuclease product was gel isolated. This product or control linearized plasmid (5 nM) was annealed to the radiolabeled primer (50 nM) either complementary to the strand with the Nt.BspQI nicking site (primer A [AO3516]) or to the primer complementary to the opposing strand (primer B [AO3518]). Template annealed to the radiolabeled primer was then isolated using S-300 HR spin columns (GE) and incubated with 2.5 mM dNTPs and 2,000 U of T4 DNA polymerase (NEB). Primer extension products were analyzed by 15%, 8 M urea PAGE and phosphorimaging. Radiolabeled oligonucleotide AO3535 was used as a marker for migration of a 60-mer primer extension product.
DNA filter binding assays were performed essentially as described [73]. Briefly, 20 μL reactions containing 15 μM total nucleotide were combined with increasing amounts of protein in a reaction containing 20 mM Tris-HCl (pH 7.5), 0.01 mM EDTA, 2 mM MgCl2, 40 μg/mL BSA, and 0.1 mM DTT. Upon the addition of Mlh1-Mlh3, reactions were incubated for 10 min at 30°C. The reaction was then filtered through KOH-treated nitrocellulose filters using a Hoefer FH225V filtration device for approximately 1 min. Filters were subsequently analyzed by scintillation counting.
Mixtures of Mlh1-Mlh3 alone or in combination with either circular or linear DNA were applied at the concentrations stated in the Results onto EM grids freshly coated with a continuous layer of amorphous carbon. Grids were floated on a 5 μL drop of the diluted assembly reaction for 2 min immediately after a glow discharge treatment of 5 mA for 15 s. Excess of sample was blotted with filter paper and the grids were stained with 1% uranyl acetate for 1 min. Grids were loaded in a room-temperature holder and introduced into a FEI Tecnai F20 electron microscope operated at 200 kV and equipped with a Gatan K2 Summit direct detector camera. This detector was used in counting movie mode with five electrons per pixel per second for 15-s exposures and 0.5 s per frame. This method produced movies consisting of 30 frames with an exposure rate of ~1 e-/Å2. Movies were collected with a defocus range of –1 to –2.5 microns and a nominal magnification of 11,500x, which produced images with a calibrated pixel size of 3.15Å. The 30 frames in each move were aligned using the program alignframesleastsquares_list [74] and averaged into one single micrograph with the shiftframes_list program [74]. These programs are available from (https://sites.google.com/site/rubinsteingroup/home). These programs perform whole frame alignment of the movies using previously published motion correction algorithms [75]. A denoising filter using Photoshop was applied to the entire image to obtain the figures shown.
For experiments in S4 Fig, circular substrate with and without a mismatch was synthesized as described above except with the inclusion of a radioactive phosphate on the 5′-end of the primer used for synthesis. When Msh2-Msh3 was present, the protein was preincubated with the substrate at 30°C for 10 min, after which 20 nM Mlh1-Mlh3 (final concentration) was added to reactions. The reaction was then incubated, stopped, and deproteinated as described above. The reaction was then either resolved by agarose gel to determine the total amount of nicking, or the DNA was digested with BsaHI and BsrGI at 37°C for 60 min. The digested sample was then resolved by 4%, 8 M urea PAGE and phosphorimaged.
The closed circular form of pEAE99 was initially gel isolated and treated with ScaI, HindIII, or used as a substrate in an Mlh1-Mlh3 endonuclease reaction containing 300 nM Mlh1-Mlh3 in quintuplicate. The linearized product was then gel isolated from these reactions and replicates were combined. Where + T4 polymerase is indicated, in a 10 μL reaction, ~1 μg of linearized DNA was combined with 100 μM dNTPs in NEB buffer 2.1. One unit of T4 DNA polymerase (NEB) was added to the reaction. The reaction was incubated for 15 min at 12°C followed by a heat inactivation step at 75°C for 20 min. An 8 μL portion of the reaction was then added to a 10 μL reaction containing the supplied buffer and 400 units of T4 DNA ligase (NEB) where indicated. The reaction was incubated at room temperature for 2 h followed by heat inactivation per the manufacturer’s instructions.
For all plotted data, individual data points from each trial as well as means and standard deviations are found in S1 Data.
|
10.1371/journal.pbio.1000533 | The DNA Methylome of Human Peripheral Blood Mononuclear Cells | DNA methylation plays an important role in biological processes in human health and disease. Recent technological advances allow unbiased whole-genome DNA methylation (methylome) analysis to be carried out on human cells. Using whole-genome bisulfite sequencing at 24.7-fold coverage (12.3-fold per strand), we report a comprehensive (92.62%) methylome and analysis of the unique sequences in human peripheral blood mononuclear cells (PBMC) from the same Asian individual whose genome was deciphered in the YH project. PBMC constitute an important source for clinical blood tests world-wide. We found that 68.4% of CpG sites and <0.2% of non-CpG sites were methylated, demonstrating that non-CpG cytosine methylation is minor in human PBMC. Analysis of the PBMC methylome revealed a rich epigenomic landscape for 20 distinct genomic features, including regulatory, protein-coding, non-coding, RNA-coding, and repeat sequences. Integration of our methylome data with the YH genome sequence enabled a first comprehensive assessment of allele-specific methylation (ASM) between the two haploid methylomes of any individual and allowed the identification of 599 haploid differentially methylated regions (hDMRs) covering 287 genes. Of these, 76 genes had hDMRs within 2 kb of their transcriptional start sites of which >80% displayed allele-specific expression (ASE). These data demonstrate that ASM is a recurrent phenomenon and is highly correlated with ASE in human PBMCs. Together with recently reported similar studies, our study provides a comprehensive resource for future epigenomic research and confirms new sequencing technology as a paradigm for large-scale epigenomics studies.
| Epigenetic modifications such as addition of methyl groups to cytosine in DNA play a role in regulating gene expression. To better understand these processes, knowledge of the methylation status of all cytosine bases in the genome (the methylome) is required. DNA methylation can differ between the two gene copies (alleles) in each cell. Such allele-specific methylation (ASM) can be due to parental origin of the alleles (imprinting), X chromosome inactivation in females, and other as yet unknown mechanisms. This may significantly alter the expression profile arising from different allele combinations in different individuals. Using advanced sequencing technology, we have determined the methylome of human peripheral blood mononuclear cells (PBMC). Importantly, the PBMC were obtained from the same male Han Chinese individual whose complete genome had previously been determined. This allowed us, for the first time, to study genome-wide differences in ASM. Our analysis shows that ASM in PBMC is higher than can be accounted for by regions known to undergo parent-of-origin imprinting and frequently (>80%) correlates with allele-specific expression (ASE) of the corresponding gene. In addition, our data reveal a rich landscape of epigenomic variation for 20 genomic features, including regulatory, coding, and non-coding sequences, and provide a valuable resource for future studies. Our work further establishes whole-genome sequencing as an efficient method for methylome analysis.
| DNA methylation plays a vital role in genome dynamics. In the human genome, it predominantly occurs at cytosine guanine dinucleotide (CpG) sites in somatic cells [1] and at non-CpG cytosines in embryonic stem cells [2] and perhaps other cells as well. DNA methylation at any of these sites can vary and thus affect many biological processes that impact on human health and disease [3]. Therefore, detailed knowledge of the of DNA methylation status of all cytosines (the methylome) is paramount for understanding the mechanisms and functions underlying DNA methylation.
The emergence of the next-generation sequencing of bisulfite converted DNA represents an important advance in the field of DNA methylation analysis [4]–[6]. This technology has enabled human methylome analysis to advance from single chromosomes [7] to low (100 bp) resolution whole genomes [8] to single-base resolution whole genomes using bisulfite sequencing [2],[9]. For a comprehensive description of methylome analysis methods, please refer to the recent review by P. Laird [10].
Using whole-genome bisulfite sequencing, we here report the methylome analysis of peripheral blood mononuclear cells (PBMC) from an anonymous male Han Chinese individual (YanHuang) whose genome was determined in the first Asian genome project, henceforth referred to as YH [11]. This approach allowed us to analyse approximately 20 million CpG sites of this clinically important human methylome for genomic landscape, allele-specific methylation (ASM), and allele-specific expression (ASE) in primary cells in a single individual.
The methylome reported and analyzed here was generated from the same sample of peripheral blood mononuclear cells (PBMCs) from a consented donor whose genome was deciphered in the YH project [11]. The nuclear DNA was extracted and subjected to unbiased, whole-genome bisulfite sequencing (BS-seq) using the Illumina Genome Analyzer (Table S1a) [5],[12]. In total, we generated 103.5 Gbp of paired-end sequence data. Of these, 70.4 Gbp (68%) were successfully aligned to either strand of the YH genome [11] with an average mismatch rate of 1.3% (Table S1b), resulting in an average sequencing depth of 12.3-fold per DNA strand or a 24.7-fold overall depth. Of the 18,962,679 CpGs present in the unique haploid part (2.21 Gb) of the YH reference genome sequence, approximately 99.86% were covered by at least one unambiguously mapped read of quality score >14 on either strand, and 92.62% were unambiguously covered on both strands (Figure S1 shows the cumulative distribution of sequencing depth; see Methods for details). Based on the 24.7-fold overall coverage, we estimated that about 88.1% of CpGs were covered on both alleles, but only 6.2% of CpGs could be definitively defined due to the limited number of nearby SNPs. We therefore only used these 6.2% (or 1.17 million) CpG sites for our allele-specific methylation analysis. Based on alignment to in silico converted non-CpG cytosines, the bisulfite conversion rate was determined to be at least 99.8% even assuming all non-CpG methylcytosines are due to conversion failure, ensuring reliable ascertainment of CpG methylcytosines at a false positive rate of <0.5%. All five libraries (Table S1a) showed similar conversion rates (99.7% to 99.9%), and a linear correlation was observed in methylation levels estimated from different libraries (Figure S2). This demonstrates high consistency between technical replicates. We also performed conventional bisulfite Sanger sequencing in randomly selected regions and found that 100% (50 of 50 tested CpG sites) showed a consistent methylation level (p>0.01 in chi-square test; Table S8). The rate of unconverted non-CpG cytosines is a combination of incomplete conversion and authentic non-CpG methylation, which indicates very low methylation levels (<0.2%) of non-CpG cytosines in PBMC. We also used the methylation ascertainment method based on binomial test and false discovery rate constraint that was applied by Lister et al. [2] to distinguish putative non-CpG methylation sites from incomplete bisulfite conversion and found a comparable (<0.2%) rate of non-CpG methylation in human PBMC. Non-CpG methylation roughly followed an exponential distribution where only a few (<1e−5) cytosines had methylation levels of >80% (Figure S3). We used these findings to exclude non-CpG methylation from subsequent analyses and estimate the overall specificity of identified methylcytosines in the PBMC methylome presented here to be 99.5%. We also used computing simulation to estimate the false negative rate of methylation site discovery. Assuming the methylation levels of CpG cytosines are similar between hESC [2] and PBMC, we estimate that about 13% of methylated CpG sites would be missed, of which a majority would be hypomethylation (<20%) sites. This indicates the PBMC methylome has a sensitivity to detect most methylated CpG sites.
We carried out a global analysis of the PBMC methylome and found the overall CpG methylation level to be 68.4%, which is lower than in H1 human embryonic stem cells (ESC) [2] but is still considered to be relatively high. Next, we determined the methylation distribution (Figure 1a) and showed it was less bimodal (9.27% hypo (<20%) methylated, 28.81% hyper (>80%) methylated) than has been previously observed (27.4% and 42.4%, respectively) [7], reflecting less bias of the whole-genome approach used here. Chromosome-specific effects could be excluded based on a separate analysis (Figure S4) of the three chromosomes analysed in the previous study [7]. Most notable was that the methylation distribution was not significantly affected by our depth threshold (where 4-fold was the lowest depth; Figure S10). In support of the conclusions drawn here, these data were consistent with the previous observation showing the CpG methylation level to peak at >70% in the human ESC methylome using the same bisulfite sequencing technology [2]. The whole genome CpG density showed a negative correlation with previously observed methylation levels [2],[6], while a major decrease was observed when CpG density rose from 10 to 15 per 200 bp windows.
We next performed a comprehensive analysis of the PBMC methylome for an additional 20 distinct genomic features (Figure 1b–u). Although some of these features have been analysed before [2],[6],[7],[13] (reviewed in [13]), our analyses provided additional information as well as a more global assessment of some of these components. For example, with respect to protein-coding genes, our data enabled the integration of multiple features into higher-order structures, such as canonical DNA methylation profiles across the entire transcriptional units of expressed and silent genes (Figure 2). Up to 13% difference (p<1e−42) in methylation between highly expressed and silent genes (as determined by digital gene expression profiling (DGEP) of the same sample) are clearly visible, as are two discrete switchover zones, one upstream of the TSS and one in intron 1 that demarcates the transition from hypo- to hypermethylation in the inverse relationship between promoter and gene-body methylation and expression. Also evident is a distinct elevation in methylation level at internal exons with clear demarcation of intron/exon boundaries.
To define genes as expressed or silent, we grouped them according to their DGEP tags, allowing correlation to be assessed between averaged levels of DNA methylation and gene expression. However, other factors than DNA methylation can of course affect expression levels, and future analysis of samples from different tissues should help to address this issue. Within these limitations, we observed a clear trend for DNA methylation levels of expressed genes to decrease at TSS and to increase at gene bodies. This is consistent with results reported from bisulfite sequencing of human ESC [2].
For non-coding RNA genes, we found that different gene families had very different methylation profiles (Figure 1h–j). For instance, tRNA genes had a significantly (p<1e−343) lower methylation level than rRNA genes and the genome average. We further conducted a comprehensive analysis of repeat elements, which is a particular strength of our unbiased whole-genome BS-seq approach (Figure 1k–t). Here we found elements that were still active, such as long terminal repeats (LTRs), LINE/L1, and SINE/Alu, and had significantly higher methylation levels than genome average (p<1e−100), displaying hypermethylation even at high CpG density (>12 CpGs in 200 bp). For instance, we found that methylation levels in Alu elements negatively correlated with evolutionary sequence divergence (Figure S5) and thus negatively correlated with retrotransposon mobility [14]. Loss of methylation in such transposable elements is known to be associated with tumorigenesis [15],[16], and the above observations are consistent with DNA methylation playing a role in controlling retrotransposon mobility by lowering their activities and thereby stabilizing the genome.
In methylome studies, CpG islands are a special genomic feature of great interest (Figure 1u). To investigate these, we performed a canonical analysis of CpG islands and found CpG density and methylation levels displayed a mirrored pattern (Figure S11). CpG islands are CpG-rich and generally hypomethylated, and the shores [17] showed gradual transition of CpG density and methylation levels between the CpG islands and genome average.
Next, we examined the correlation of methylation level of any two nearby CpGs and the relationship between spatial distance (from one CpG to another) and strength of this correlation. Gaining knowledge of genomic regions or features that are highly correlated in methylation status is advantageous for developing efficient designs for genome-wide association studies by enabling the selection of tag CpGs, analogous to tag SNPs [18]. As has been previously observed [7], co-methylation deteriorates over distance and becomes nearly undetectable at distances >1,000 bp (Figure S6a). The co-methylation observed here was not affected by the underlying CpG density (Figure S12).
Analysis of CpG cytosines that had the same distance between them showed that higher methylation levels correlate when they were located on the same strand than on opposite strands (p<6e−7; Figure S6a). This is presumably due to temporary hemi-methylation as a result of post-replication lag in methylation maintenance in proliferating cells. Co-methylation is also markedly different between different genomic features (Figure S6b–t). For example, the correlation was significantly (p<1e−30) higher in gene- than in repeat-associated features. Using Fourier transformation, we also tested the methylation correlation for patterns and found a significant (p<1e−4) peak in periodicity of approximately 170 bp (Figure S7). A similar (CHG) methylation pattern was observed in Arabidopsis [4], which the researchers suggested was due to a nucleosome positioning effect on co-methylation. However, no significant periodicity of smaller motifs was observed in our data.
We compared the PBMC methylome to that of fetal lung fibroblast cells (IMR90) [2],[8] to assess potential tissue-specific differentially methylated regions (tDMR). In total, 240,856>200 bp independent regions (range 200–3.5 kbp; median size 500 kbp; see Methods for more details) that had significant differences in methylation level (>2-fold change, at least in one tissue is not hypomethylated (<20%) and Fisher test p value <1e−2) were identified as candidate tDMRs. Of these, 6,197 were located in the 2 kb flanking sequences of transcription start sites (TSSs) of 6,415 genes. GO classification showed that genes associated with PBMC-specific, hypomethylated tDMR candidates (and confirmed to be expressed according to DGEP analysis and/or the GEO database [19]) were significantly (p<1e−4) overrepresented in categories that related to DNA damage checkpoint (Tables S2, S3).
We examined the PBMC methylome to assess allele-specific methylation (ASM) in the context of genomic imprinting [20] and allele-specific expression (ASE) [21]. Integration of our methylome data with the YH haploid genome sequences [11] enabled us to determine ASM for 1.17 million CpG sites (see above), providing an unprecedented opportunity to identify a first and comprehensive set of haploid differentially methylated regions (hDMR) in any human cell type. Using a conservative threshold (≥5 CpGs with at least 2-fold methylation difference and p value <0.001 in Fisher test), we identified 599 hDMRs (mean size of 312 bp), which accounted for 0.61% of all CpGs with biallelic methylation information or 0.33% of 181,599 regions with bi-allelic sequence information and ≥5 CpGs in their 300 bp flanking sequences (Table S4, see Methods for details). For each of the hDMRs, we randomly selected genomic CpGs with same sequencing depths to that of the hDMRs and subjected them to 10,000 bootstrap iterations to determine how many times the randomly selected CpGs would show differential methylation as defined above. The simulation indicated that 4.17% of these hDMRs were stochastic (showing hDMR signals in >5% of the simulations). As there are approximately 28 million CpG sites in the human genome and ASM could be ascertained for 1.14 million sites, we extrapolated the total number of hDMRs in the YH methylome to be approximately 10,000. This rate, however, is likely to be an overestimation because: (1) 300 bp windows with <5 CpGs may not have enough statistical power to distinguish ASM and (2) CpGs in such lower-CpG-density regions are generally hypermethylated and statistically less likely (likelihood ratio <0.133 compared to regions with ≥5 CpGs in 300 bp windows based on 100,000 simulations on the PBMC methylome) to qualify as ASM according to the conservative threshold described above (p<0.001, 2-fold methylation level change). Nonetheless, if none of the 300 bp windows with <5 CpGs were to display ASM, and those with ≥5 CpGs were to have the same rate of displaying ASM as observed in flanking regions with bi-allelic sequence information, the lower limit of the total number of hDMRs in the YH methylome would still be expected to be 5,000. Thus, we estimate that 0.3%–0.6% of the YH genome are subject to ASM. Annotation analysis revealed that some of these hDMRs were associated with 287 genes (see Table S5 for full list). Figure 3 shows an example of such an association (with the gene FANK1), which displays ASM. FANK1 is a testis-specific gene and has been proposed to play a role in the transition from the diploid to the haploid phase during spermatogenesis [22]. In addition, we investigated the distribution of hDMRs within the YH genome. This analysis revealed a significant (p<1e−343) tendency for the hDMRs to cluster, particularly when in proximity to telomeres or centromeres, which are both hallmarks of imprinting.
To assess the potential of the hDMRs to denote known or novel imprinted loci, we tested the 599 identified hDMRs for correlation with known imprinted loci [23]. First, we analysed the known genomic imprinted space (defined by 40 loci in 15 chromosomal regions [23]) and identified 17 overlaps (Figure 4), including with well-known imprinted loci such as IGF2, H19, KCNQ1, GNAS, and others (Figure S8). Reciprocal analysis of known imprinted loci for which bi-allelic information was available showed 87.8% ASM, indicating that most of the ASM regions can be identified by bisulfite sequencing and that the major limiting factor is a lack of SNPs to differentiate the two alleles. We therefore estimate that most of the hDMRs are not attributable to imprinting but to other mechanisms such as sequence-dependent ASM [24].
Finally, we analysed the possible involvement of ASM (defined by presence of hDMRs) in epigenetically driven allele-specific expression (ASE) (reviewed in [21]). For this, we randomly selected 6 of the 76 genes that had one or more hDMR(s) within 2 kb of their TSS and measured their expression by TA clone sequencing. Five of the six genes (83%) showed a >1.5-fold difference in expression level between the two alleles (Table S6), confirming the inverse relationship between promoter ASM and ASE. As ASM is definitive for 6.2% of the genome and 76 genes had hDMRs within 2 kb flanking sequence, we estimated that 600 to 1,200 genes (3%–6%) display ASM, which indicates that up to a quarter of the 20% of human genes that have been reported to display ASE [25] may be driven by ASM.
To determine possible biological functions of the 76 genes displaying ASM, we carried out gene ontology (GO) analysis. Our results showed that these hDMR-containing genes are significantly (p<1e−4) overrepresented in function categories related to cell division and differentiation (see Table S7 for full list of significant GO categories), such as “negative regulation of S phase of mitotic cell cycle,” “mitotic metaphase/anaphase transition,” and “negative regulation of lymphocyte proliferation.” This functional enrichment pattern was also supported by hDMR-containing non-coding RNA genes hY3 and hY5 that were reported to be essential in DNA replication [26], which is an integral part of cell division.
In this study, we have generated and analysed the two haploid methylomes of human peripheral blood mononuclear cells (PBMC) from an individual whose genome was previously sequenced. This allowed, for the first time, for assessment of the level of ASM within a human methylome and extends recent studies analysing variation between different human methylomes [2],[9]. Compared to what was observed in embryonic stem cells in these studies, non-CpG methylation in human PBMC was negligible.
Our results show that ASM is more frequent than can be accounted for by known imprinted loci [23] and correlates very well with ASE for genes displaying ASM in their promoter regions. To further quantify this observation, additional methylomes will be required to increase the number of parental polymorphisms at imprinted regions. Nonetheless, our work provides a first proof-of-concept for the importance of including ASM in methylome analyses.
In addition, our data revealed a rich landscape of distinct epigenomic features for regulatory, coding, and non-coding sequences. Exons, for instance, were clearly discernable from introns by elevated methylation levels, demarcated by sharp intron-exon boundaries. This finding confirms and extends a recent observation that exons can be defined by epigenetic marks such as nucleosome positioning [13],[27]. The nature of our whole-genome approach enabled us to also analyse features that have previously been difficult to assess [28] such as repeat elements that constitute about 50% of the human genome [29]. Mobility of Alu repeat elements, for instance, was found to negatively correlate with their methylation levels, emphasizing the critical role of DNA methylation in genome stability.
In conclusion, we have reported the first comprehensive methylome analysis at single base-pair resolution for human blood cells with relevance to basic and clinical research. Our results demonstrate this methylome to be rich in biological information, compatible for integration with functional data, and we expected it to form a lasting resource as part of the International Human Epigenome Project [30].
The PBMC methylome data have been deposited into the NCBI Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE17972). In addition, the PBMC methylome and other data are available at the YH genome database (http://yh.genomics.org.cn).
The YH genome was downloaded from YH database (http://yh.genomics.org.cn). Gene and repeat annotations were downloaded from the UCSC database (http://genome.ucsc.edu/). The NCBI reference genes with prefix “NM” were mapped to the reference genome using BLAT by UCSC. Hits with >90% identity were retained for further analysis and only one transcript was retained for each gene. Known imprinted genes were extracted from the content of [23], respectively.
Peripheral blood was obtained from the same individual as in the YH project, and mononuclear cells were separated through Ficoll-Paque (GE Heatlthcare) gradient centrifugation. The total DNA was prepared by proteinase K/phenol extraction, and RNA was extracted from mononuclear cells with RNeasy Mini Kit (Qiagen) following the manufacturer's instructions. The DNA was fragmented by sonication using a Bioruptor (Diagenode, Belgium) to a mean size of approximately 250 bp, followed by the blunt-ending, dA addition to 3′-end and, finally, adaptor addition (in this case of methylated adaptors to protect from bisulfite conversion), essentially according to the manufacturer's instructions. The bisulfite conversion of the adaptor-added DNA was carried out as previously described [31]. Raw GA sequencing data were processed by Illumina Pipeline v1.3.1.
Validation of the methylated state of selected candidate loci was performed by sequencing of multiple T-cloned PCR fragments from the bisulfite converted DNA. The bisulfite treated DNA was amplified by 18 PCR cycles and used for Solexa sequencing.
The reads generated by Illumina sequencing were aligned to the YH genome [11]. As DNA methylation has strand specificity, separate alignments of 6 Gbp in combined length were generated for the Watson and Crick strands of the YH genome. All cytosines in the 6 Gbp target sequence (“original form”) were replaced in silico by thymines (“alignment form”) to allow alignment after bisulfite conversion. In addition, the original forms of the reads were also transformed to cope with BS-treatment nucleotide conversion in the alignment process using the following criteria: (1) observed cytosines on the forward read of each read pair were in silico replaced by thymines, and (2) observed guanines on the reverse read of each read pair were in silico replaced by adenosines. We then mapped the “alignment form” reads to the “alignment form” target sequence using SOAPaligner [32]. Every hit with a single placement with minimum number of mismatches and a clear strand assignment was defined as an unambiguous alignment and was used in methylcytosine ascertainment. Ambiguously aligned reads were only used to estimate the approximate copy number of the local region. Local copy number of a genomic location was calculated by averaging the hit counts of all reads that cover a certain genomic location. Genomic bases with a copy number larger than 1.5 were not used to call methylcytosines and not used in any subsequent analysis to avoid errors caused by misalignment. In total, 2.21 Gbp (77.5% of the whole genome excluding N's) were of local copy number <1.5, which we defined as the “unique” part of genome that contained all cytosines analyzed in this study.
For methylcytosine identification, we transformed each aligned read and the two strands of the YH genome back to their original forms to build an alignment between the original forms. In the unique part of genome, cytosines that were covered by cytosines from reads on the same strand or guanines from those on the opposite strand (hereafter, referred to as ascertainment bases) were called as potentially methylated sites. To exclude spurious ascertainment bases that were caused by sequencing errors, we filtered out all bases with quality scores lower than 14. Increasing the quality threshold further did not change the non-CpG methylation rate. The false positive rate of methylcytosine identification was calculated as: where r is the conversion rate (proportion of non-CpG cytosines with Q14 ascertainment bases), NCpG is the total number of CpG cytosines, and NmCpG is the total number of ascertained methylated CpG cytosines. As non-CpG methylation may occur, though at a very low level, the false positive rate is an overestimation.
Sequencing errors could affect the ascertainment of methylation; therefore, we used the non-CpG methylation level as an indicator of errors. Overall methylation level of non-CpG sites becomes stable when quality >14 (Figure S9), which means the estimate is reliable above such a threshold. To eliminate the effect of low quality bases when estimating the methylation level of a specific genomic CpG cytosine, we divided the number of ascertainment bases by the number of total Q14-covering bases of that genomic location. To estimate the methylation level of a single base accurately, we only used CpG cytosines with a per-strand depth of more than 4 in the analysis of distribution of single CpG methylation levels, as for Figure 1 in the main text. Distribution of methylation level on CpG sites with 4× to 10× coverage, between which there was at minimum a consistent 10× coverage, indicated that the depth requirement was reasonable at even 4× coverage and could provide 5 different results (Figure S10).
For estimating the methylation level in a specific region, we divided the number of all ascertaining bases in the region by the number of all Q14 bases covering CpG cytosines in that region.
Putative tDMRs were identified by comparison of the PBMC and fetal lung fibroblast cell (IMR90) [2] methylomes using windows that contained at least 5 CpG sites with a 2-fold change in methylation level and Fisher test p value <1e−20. In addition, we require that both tissues should not be hypomethylated in tDMR discovery. Two nearby tDMRs would be considered interdependent and joined into one continuous tDMR if the genomic region from the start of an upstream tDMR to the end of a downstream tDMR also had 2-fold methylation level differences between sperm and PBMC with a p value <1e−20. Otherwise, the two tDMRs were viewed as independent. After iteratively merging interdependent tDMRs, the final dataset of tDMRs was made up of those that were independent from each other.
We checked single-end and paired-end reads that were aligned across heterozygotes identified from the YH genome [11] to assign them to specific alleles. We calculated the methylation level of CpGs in SNP-allele containing reads that were assigned to an allele, and the number of methylcytosines and cytosines in the reads from each allele were subjected to Fisher test. Regions with at least 5 genomic CpGs, 2-fold change in methylation level, and a p value <0.001 were defined as hDMRs. Two hDMRs were joined if the phasing relationship could be validated by haplotype analysis of the corresponding YH sequence data or by reads spanning two heterozygotes.
|
10.1371/journal.pntd.0001050 | Testing the Efficacy of a Multi-Component DNA-Prime/DNA-Boost Vaccine
against Trypanosoma cruzi Infection in Dogs | Trypanosoma cruzi, the etiologic agent of Chagas Disease, is
a major vector borne health problem in Latin America and an emerging
infectious disease in the United States.
We tested the efficacy of a multi-component DNA-prime/DNA-boost vaccine
(TcVac1) against experimental T. cruzi infection in a
canine model. Dogs were immunized with antigen-encoding plasmids and
cytokine adjuvants, and two weeks after the last immunization, challenged
with T. cruzi trypomastigotes. We measured antibody
responses by ELISA and haemagglutination assay, parasitemia and infectivity
to triatomines by xenodiagnosis, and performed electrocardiography and
histology to assess myocardial damage and tissue pathology.
Vaccination with TcVac1 elicited parasite-and antigen-specific IgM and IgG
(IgG2>IgG1) responses. Upon challenge infection, TcVac1-vaccinated dogs,
as compared to non-vaccinated controls dogs, responded to T.
cruzi with a rapid expansion of antibody response, moderately
enhanced CD8+ T cell proliferation and IFN-γ production,
and suppression of phagocytes’ activity evidenced by decreased
myeloperoxidase and nitrite levels. Subsequently, vaccinated dogs controlled
the acute parasitemia by day 37 pi (44 dpi in non-vaccinated dogs), and
exhibited a moderate decline in infectivity to triatomines. TcVac1-immunized
dogs did not control the myocardial parasite burden and electrocardiographic
and histopatholgic cardiac alterations that are the hallmarks of acute
Chagas disease. During the chronic stage, TcVac1-vaccinated dogs exhibited a
moderate decline in cardiac alterations determined by EKG and
anatomo-/histo-pathological analysis while
chronically-infected/non-vaccinated dogs continued to exhibit severe EKG
alterations.
Overall, these results demonstrated that TcVac1 provided a partial resistance
to T. cruzi infection and Chagas disease, and provide an
impetus to improve the vaccination strategy against Chagas disease.
| Immunization of dogs with DNA-prime/DNA-boost vaccine (TcVac1) enhanced the
Trypanosoma cruzi-specific type 1 antibody and
CD8+ T cell responses that resulted in an early control of
acute parasitemia and a moderate decline in pathological symptoms during chronic
phase. Further improvement of vaccine-induced immunity would be required to
achieve clinical and epidemiological benefits and prevent transmission of
parasites from vaccinated/infected dogs to triatomines.
| American trypanosomiasis (Chagas disease) is a disease of humans and caused by the
protozoan Trypanosoma cruzi of the family trypanosomatidae [1]. Approximately
30–40% of the infected patients develop a chronic debilitating illness
of the cardiac system, characterized by clinically irreversible and progressive
tissue destruction, and myocardial hypertrophy, eventually leading to heart failure
and the patient’s death [2], [3].
Several investigators have shown the potential utility of T. cruzi
surface antigens as vaccine candidates in murine experimental models [4], [5](reviewed in
[6], [7]). We have
shown the protective efficacy of amastigote surface proteins ASP-1 and ASP-2, and
trypomastigote surface antigen TSA-1 as DNA vaccines in mice [8]. Vaccination with ASP-2 provided
maximal immunity to T. cruzi infection in mice that was further
enhanced by co-delivery of cytokine adjuvants [8]. In recent studies, we have
identified additional potential vaccine candidates by computational screening of
T. cruzi sequence database [9]. Of these, TcG1-TcG8 were
phylogenetically conserved in clinically important strains of T.
cruzi and expressed in the infective and intracellular stages of the
parasite [9]. When
delivered as a DNA vaccine in mice, TcG1, TcG2 and TcG4 elicited a significant
trypanolytic antibody response and Th1 cytokine (IFN-γ) response, a property
associated with immune control of T. cruzi
[10]. These novel
vaccine candidates, thus, increased the pool of protective vaccine candidates
against T. cruzi.
In this study, we proceeded to examine the prophylactic and transmission-blocking
efficacy of the multi-component vaccine constituted of TcG1, TcG2 and TcG4 in dogs.
We chose dogs for our studies because dogs provide an excellent model for studying
the human disease [11]–[13]. Experimentally and naturally infected young dogs
(2–3 months) elicit reproducible and comparable acute infection associated
with increase in blood parasitemia, IgG and IgM antibodies [14], [15] and T cell response
[16]. The
presence of myocarditis with a moderate or small number of parasitized cells and
extensive and frequent focal necrosis turns the disease in dogs similar to acute
Chagas disease in humans [17], [18]. A few infected dogs
(10–20%; 5% humans) develop severe acute myocarditis and may die
of cardiac arrest. More than 80% of dogs recover from acute parasitemia as
parasites become undetectable in the blood, and remnant mild myocardial changes with
scattered microscopic foci of fibrosis and lymphocytic infiltration [19] present a
picture similar to that of human infections [20], [21]. At 12–18 months
post-infection (pi), ∼50% of infected dogs exhibit symptomatic chronic
cardiac disease (20% develop severe myocarditis), associated with progressive
cardiomegaly; arrhythmia, including RBBB with left anterior hemiblock [22], [23]; diffused
myocarditis with focal and interstitial fibrosis, and self-perpetuating myofibril
destruction - a picture reminiscent of the chronic form of Chagas disease observed
in humans. Thus, testing the vaccine efficacy in dogs would provide a strong basis
for developing a human vaccine against T. cruzi and Chagas
disease.
Further, dogs are an important reservoir host for domestic transmission of T.
cruzi. The prevalence rate of T. cruzi infection in
dogs may reach up to 84% in endemic areas (e.g. rural Argentina, Chiapas,
Mexico), determined by serological procedures and xenodiagnosis [24], [25]. Dogs are also
the most frequent blood meal source for the domestic triatomines, i.e., T.
barberi and T. pallidipennis in Mexico [15] and
T. infestans in Argentina [25], [26]. Likewise, a high prevalence
of seropositive dogs and infected triatomines is routinely noted in rural and urban
development in the southern US [27], [28], and suggested to maintain T. cruzi
transmission in the human habitat. Triatomines are several times more likely to take
their blood meal from dogs than from humans [26]. The ratio of dog blood meals
to human blood meals in the engorged guts of triatomines is estimated to be
2.3–2.6 times higher than the ratio of the number of dogs to the number of
humans in a household [29]. Thus, the probability of infecting an insect in one
blood meal from dogs is estimated to be 200-times higher as compared to the
probability from adult humans [25]. These studies demonstrate that dogs are an important
host blood source for domiciliary triatomines, and the risk of T.
cruzi infection in humans is increased by the presence of infected
dogs. Strategies that can limit T. cruzi infection in domestic
reservoir host may, thus, prove to be effective in interrupting the parasite
transmission to the vector, and consequently, to the human host.
We immunized dogs with DNA-prime/DNA-boost vaccine (TcVac1). We examined the efficacy
of TcVac1 in eliciting antigen-and parasite-specific antibody and T cell immunity,
and determined if vaccination with TcVac1 modulated the host immune response towards
protective type 1 upon T. cruzi infection. We also examined the
efficacy of TcVac1 in controlling acute parasitemia, blocking the parasite
transmission to triatomines, and preventing clinical severity of chronic
disease.
Twelve mongrel dogs (6 males and 6 females, 3–4 months old) were acquired
locally and kept at the animal facility at the UAEM Research Center until they
were included in the experiment, at eight months of age (8–12 kg body
weight). Dogs were confirmed free of T. cruzi infection by
microscopic examination of blood smears and serological evaluation of
anti-T. cruzi antibodies using an indirect
haemagglutination assay (IHA) and enzyme-linked immunosorbent assay (ELISA)
[15], [18]. Before inclusion
in experimental studies, dogs were treated with anti-helminthes and vaccines
against regional infectious diseases (Canine distemper, Parvovirus infection,
Canine hepatitis, Leptospirosis, and Rabies). All dogs received water ad
libitum, and commercial dog food fed twice a day according to their
age and development requirements. Experimental protocols were conducted under
the technical specifications for the production, care, and use of lab animals
from the Norma Official Mexicana (NOM-062-ZOO-1999), and the Council for
International Organizations of Medical Science [30], [31]. The research protocols
were approved by the Laboratory Animal Care Committee at the Universidad
Nacional Autonoma de Mexico.
TcVac1 vaccine was constituted of antigen-encoding plasmids (pCDNA3.TcG1,
pCDNA3.TcG2 and pCDNA3.TcG4) and IL-12- and GMCSF-expression plasmids, described
previously [8],
[32]. The
eukaryotic expression plasmids encoding dog cytokines (IL-12 and GM-CSF) were a
kind gift from Dr. Peter Melby [33]. All recombinant
plasmids were transformed into E. coli DH5-α competent
cells, grown in L-broth containing 100 µg/ml ampicillin, and purified by
anion exchange chromatography using the Qiagen maxi prep kit (Qiagen,
Chatsworth, CA) according to the manufacturer’s specifications.
Trypomastigotes of T. cruzi (SylvioX10/4) were maintained and
propagated by continuous in vitro passage in C2C12 cells.
Dogs (n = 6/group, 3 males and 3 females) were
intramuscularly immunized with TcVac1 (200 µg each plasmid DNA/dog),
delivered four-times at 2-week intervals. Dogs vaccinated with empty vector
(pcDNA3 only) were used as controls. Two-weeks after the last immunization, dogs
were challenged with culture-derived T. cruzi SylvioX10/4
(3.5×103 trypomastigotes/kg body weight, i.p.). The
selected dose of the parasites was sufficient to produce acute parasitemia
within 1–2 weeks of inoculation, and symptomatic clinical disease within
6–8 weeks post-infection [18]. Dogs were
observed daily for general physical condition, at weekly intervals for clinical
condition, and at 2-week intervals for cardiac function, monitored by
electrocardiography (EKG). Sera samples were obtained before each immunization
and at two-week intervals thereafter. After challenge infection, in addition to
sera samples, blood samples for parasitemia diagnostics were collected beginning
day 5 pi, on alternate days up to 50 dpi and at two-week intervals
thereafter.
We measured blood parasitemia using hemacytometer counts of 5 µl blood
mixed with equal volume of ACK red blood cell lysis buffer. Xenodiagnostic
analysis was performed as described [25], [34], [35]. Briefly,
stage 4 naive triatomine (T. pallidipenis) nymphs (6 per dog)
were fed on vaccinated and control dogs on day 30 and day 60 pi. Fecal samples
were collected from triatomines at day 60 after feeding, and analyzed by light
microscopy to detect epimastigote and metacyclic trypomastigotes. At least 10
microscopic fields were analyzed for each fecal sample, and triatomines were
considered T. cruzi positive when >1 parasites were
detected.
The cDNAs for TcG1, TcG2 and TcG4 were cloned
in pET-22b plasmid (Novagen) such that the encoded proteins would be expressed
in-frame with a C-terminal His-tag. All cloned sequences were confirmed by
restriction digestion and sequencing at the Recombinant DNA Core Facility at
UTMB. For the purification of recombinant proteins, plasmids were transformed in
BL21 (DE3) pLysS competent cells, and recombinant proteins purified using the
polyhistidine fusion peptide-metal chelation chromatography system
(Novagen).
Blood samples were obtained by venopuncture of the cephalic vein, and immediately
processed to separate sera, using standard methods [15], [18].
Sera samples (1∶50–1∶100 dilution) were analyzed for IgM and
IgG by using the Chagas diagnostic kits for ELISA (Laboratorio-Lemos SRL, Buenos
Aires, Argentina). The horseradish peroxidase (HRP)-labeled anti-human-IgG in
ELISA kit was replaced with HRP-conjugated goat-anti-dog IgM- or IgG-specific
secondary antibody (Bethyl Laboratories) [15], [18].
In some experiments, instead of T. cruzi lysate, plates were
coated with recombinant antigens (TcG1, TcG2 or TcG4, 10-µg protein/ml) to
capture the antigen-specific antibodies. Sera samples from chronically infected
dogs with confirmed T. cruzi infection and from healthy
domestic dogs were used as positive and negative controls, respectively (cut off
value: ELISA, mean OD450nm from negative dogs ±2 SD; IHA,
positive titer at ≥1∶16 serum dilution).
To identify the antibody sub-types (IgG1 and IgG2), plates were coated with
T. cruzi antigen, and, then sequentially incubated at room
temperature with sera samples (1∶50 dilution) for 2 h, biotin-conjugated
goat anti-dog Ig subtypes (IgG1 and IgG2) for 2 h, and streptavidin-horseradish
peroxidase conjugate for 30 min. All antibodies and conjugates were from Bethyl
Laboratories, and used at a 1∶3000 dilution in PBST-0.5% NFDM
(100-µl/well). Color was developed with 100-µl/well Sure Blue TMB
substrate (Kirkegaard & Perry Labs), reaction was stopped with 2N sulfuric
acid, and antibody response was monitored at 450 nm using a SpectraMax M5
microplate reader.
Splenic level of CD4+ and CD8+ cell population in
immunized/challenged dogs was determined by flow cytometry. Briefly, splenocytes
were suspended in PBS (1×106 cells/100 µl) and incubated
for 30 min with FITC-conjugated anti-CD4 and PE-conjugated anti-CD8 antibodies
(1∶50 dilution, from ABD Serotec). Following incubation, cells were fixed
with 2% paraformaldehyde, washed and re-suspended in 500 µl PBS,
and analyzed on a FACScan apparatus (BD Biosciences). Cells stained with PE- and
FITC- conjugated rat IgGs (isotype matched) were used as negative controls. Flow
data were analyzed by Cell Quest software (BD Biosciences).
Cytokine levels in sera of vaccinated dogs were measured by sandwich ELISA.
Briefly, 96 well plates were coated overnight with anti-IFN-γ or anti-IL-10
antibodies (500-ng/ml in PBS), washed with PBS/0.05% Tween-20 (PBST), and
incubated for 2 h with 1% BSA. Plates were then sequentially incubated at
room temperature with sera samples (50-µl/well) for 2 h, biotinylated
anti-dog IFN-γ antibody (0.5-µg/ml) or anti-dog IL-10 antibody
(2-µg/ml) for 2 h, and streptavidin conjugated horse radish peroxidase
(1∶3000 dilution) for 45 min. All antibodies were from R&D systems.
Colorimetric reaction was performed as above. Cytokine concentrations were
calculated using a standard curve derived using recombinant IFN-γ or IL-10
(1–4000 pg/ml).
MPO activity was determined by a dianisidine-H2O2 method
[36],
modified for 96-well plates [37]. Briefly, plasma samples (10-µg protein) were
added in triplicate to 0.53 mM o-dianisidine dihydrochloride
(Sigma) and 0.15 mM H2O2 in 50 mM potassium phosphate
buffer (pH 6.0). After incubation for 5 min at room temperature, the reaction
was stopped with 30% sodium azide and the change in absorbance was
measured at 460 nm (ε = 11,300 M−1
cm−1). Results were expressed as units of MPO/mg protein,
whereby one unit of MPO was defined as the amount of enzyme degrading one n mol
H2O2 per min at 25°C.
The nitrite/nitrate content, indicative of inducible nitrite oxide synthase
(iNOS) activity, was monitored by the Greiss reagent assay, as described [37]. In 96-well
plates, reduced plasma samples (10 µg protein) were mixed with 100
µl Greiss reagent, consisting of 1% sulfanilamide in 5%
phosphoric acid and 0.1% N-(1-napthyl)ethylenediamine
dihydrochloride (1∶1, v/v), and incubated for 10 min. The change in
absorbance was monitored at 545 nm (standard curve, 0–200 n mol sodium
nitrite).
Changes in cardiac rhythm and conduction in all dogs was monitored before
inclusion in the study, and after challenge infection, at 2-week intervals up to
8-weeks and at monthly intervals thereafter. We used electrocardiograph (Stylus,
EK-8, USA) setting at 120 V, 60 Hertz, 20 amps, and 25 Watt in all experiments.
Six leads of the electrocardiogram were considered at 25 mm/sec at 1-mV,
standardized to 1 cm for the present study.
Necropsy was performed the day animals died due to infection or after
humanitarian sacrifice at day 60 (acute phase) and day 365 (chronic phase) post
challenge infection. Dogs were sedated with xylazine (1–3 mg/kg body
weight) and then euthanized according to the Mexican Norma Official Mexicana
[30],
[31],
using protocols approved by the Laboratory Animal Care Committee at the
Universidad Nacional Autonoma de Mexico. A macroscopic and microscopic analysis
of affected organs was performed. Postmortem studies were conducted using
standard protocols with emphasis on macroscopic findings related to Chagas
disease in heart tissue [15]. For histological
analysis, tissue samples were fixed in 10% buffered formalin for 24 h,
dehydrated in absolute ethanol, cleared in xylene, and embedded in paraffin.
Tissue sections (5-µm thick) were stained with hematoxylin-eosin, and
evaluated by light microscopy at 100× and 400× [15], [18].
Tissues were scored 0 to 4 in blind studies, according to the extent of
inflammation and tissue damage from normal to total wall involvement [38].
Data are expressed as means ± SD, and derived from duplicate experiments
(n≥6 animals/group/experiment) with at least duplicate observations per
sample. Results were analyzed for significant differences using ANOVA procedures
and Student’s t-tests. The level of significance was accepted at
*p<0.05 (vaccinated versus non-vaccinated).
The development of an antibody response induced by TcVac1 was determined by an ELISA.
All dogs were seronegative before vaccination was initiated. The T.
cruzi-specific IgM and IgG antibody response was detectable in sera
(1∶50 dilution) of vaccinated dogs after the first immunization, and
moderately increased upon delivery of booster vaccine doses (Fig. 1A&B). The level of antigen-specific
antibody response was detected in the order of TcG1>TcG2>TcG4, and was
additive in nature (Fig. 1E).
The vaccine-induced antibody response was predominantly of the Th1 type with
IgG2/IgG1 ratios >1 (Fig.
1C&D). Control dogs immunized with plasmid vector alone exhibited no
parasite- and antigen-specific antibody response (Fig. 1).
After challenge infection with T. cruzi, sera samples were analyzed
at 2-week intervals (1∶100 dilution) (Fig. 2). Non-vaccinated/infected dogs exhibited a
slight increase in parasite-specific IgM levels (Fig. 2A). All dogs, irrespective of vaccination
status, responded to T. cruzi infection by a gradual increase in
anti-parasite IgG levels (Fig.
2B). The TcVac1-immunized dogs exhibited a faster increase in IgG
antibody response to T. cruzi infection (Fig. 2B, p<0.05) as compared
to that detected in non-vaccinated/infected dogs. Likewise, the vaccine-induced
dominance of IgG2 antibodies (compared to IgG1 subtype) was significantly expanded
after infection (Fig. 2C&D).
Together, the results presented in Fig.
1 and Fig. 2
suggested that vaccination of dogs with TcVac1 skewed the antibody response towards
Th1 type that was further expanded upon challenge infection with T.
cruzi.
Next, we measured vaccine’s efficacy in activation of phagocytic (neutrophils
and macrophages) response to T. cruzi by evaluating the plasma
level of MPO activity and nitrite contents (Fig. 3). Vaccinated and non-vaccinated dogs
exhibited no detectable level of MPO activity before challenge infection (Fig. 3A). After exposure to
T. cruzi, all dogs responded by a rise in MPO activity. During
day 15–30 pi, non-vaccinated/infected and vaccinated/infected dogs exhibited a
2-fold and 25% increase in MPO activity in response to T.
cruzi infection (Fig.
3A). After day 30 pi, all dogs exhibited a similar decline in circulatory
MPO activity. Likewise, the nitrite levels, indicative of iNOS activation and NO
production, were increased by 2-fold in non-vaccinated/infected dogs at 30 dpi,
while vaccinated/infected dogs exhibited ∼22% increase in plasma nitrite
contents (Fig. 3B). These data
suggested that immunization with TcVac1 suppressed the T.
cruzi-mediated activation of phagocytes evidenced by decreased plasma
levels of MPO and nitrite in vaccinated/acutely-infected dogs as compared to
non-vaccinated controls.
A predominance of CD8+ T cells and type 1 cytokines (IFN-γ) is
shown to be essential for control of T. cruzi infection [32]. All dogs,
irrespective of vaccination regimen, responded to T. cruzi
infection by a strong increase in parasite-specific lymphocyte activation (Fig. 4). The vaccinated/infected
dogs exhibited a moderately stronger CD8+T cell response as compared
to the non-vaccinated/infected dogs that was maintained during acute infection and
chronic disease phase (Fig. 4A).
The circulatory cytokine levels (IFN-γ and IL-10) were below detection limit
before and after immunization with TcVac1. The sera level of IL-10 remained
undetectable after challenge infection with T. cruzi in all dogs.
In comparison, all dogs responded to infection by a rapid increase in circulatory
IFN-γ level that was significantly higher in vaccinated/infected dogs as
compared to that noted in non-vaccinated/infected dogs (Fig. 4B). These results indicated that
TcVac1-immunized dogs were moderately better than non-vaccinated dogs in responding
to T. cruzi infection by elicitation of higher level of type 1
biased CD8+T cell response.
Detectable parasitemia that peaked during day 30–35 pi was noted in all dogs
(Fig. 5A). Dogs vaccinated
with TcVac1 exhibited an early rise in parasitemia that was controlled by day 37 pi.
In comparison, non-vaccinated/infected dogs exhibited a slight delay in peak
parasitemia; however, blood parasitemia persisted beyond day 37 pi. No signs of
clinical illness were apparent in vaccinated/infected and non-vaccinated/infected
dogs during the physical exam, yet 33% of the TcVac1-vaccinated/infected dogs
succumbed during 40–42 dpi (Fig.
5B).
Xenodiagnostic studies were performed to determine if dog’s infectivity to
triatomines is altered by vaccination. Triatomines were fed on dogs during acute
phase (30 dpi) and after control of acute parasitemia (60 dpi), and feces were
analyzed 30 days post-feeding for the detection of parasites by microscopy. In
agreement with the peak parasitemia, all triatomines fed at day 30 pi on vaccinated
and non-vaccinated dogs became T. cruzi positive. Of the 36
triatomines fed on each group of dogs at day 60 pi, 47% (17 out of 36)
insects fed on TcVac1-vaccinated/infected dogs and 30% (11 out of 36) insects
fed on non-vaccinated/infected dogs died during the incubation period. Of those
surviving, we detected T. cruzi in feces of 52.63% (10/19)
and 84.6% (21/25) of the insects fed on vaccinated/infected and
non-vaccinate/infected dogs, respectively. These results indicated that TcVac1 was
not effective in preventing infection or early rise in acute parasitemia, and was
moderately effective in reducing the time-course of parasitemia and dogs’
infectivity to triatomines after day 37 post-infection.
Normal electrocardiographic readings were noted in all dogs included in the study,
before and after immunization. After challenge infection, vaccinated/infected and
non-vaccinated/infected dogs exhibited no cardiac alterations up to 30 dpi. By day
60 pi, 67% of non-vaccinated dogs displayed electrocardiographic alterations,
including reduced P-R interval, reduced R wave voltage, axis rotated to the right,
S-T segment line elevation from the isoelectric line (>0.2 mV), long QT segment,
J wave elevation, and sinus tachycardia that were diagnostic of myocarditis,
pericarditis, and high degree of myocardiocyte necrosis. Two vaccinated/infected
dogs died by day 42 pi due to high electrical conductance problems and arrhythmia.
Of remaining, 50% of the vaccinated/infected dogs exhibited at 60 dpi no
electrocardiographic alterations, and other 50% exhibited a moderate level of
EKG abnormalities including low voltage complex, a positive deviation of S-T with
elevation of J-wave, left axel rotation, and tachycardia that were diagnostic of
ventricular dilation, myocarditis, and arrhythmia. At one-year post-challenge
infection, electrocardiographic analysis revealed a spectrum of cardiac dysfunction
in chronically infected dogs. Among the unvaccinated/chronically-infected dogs,
66% exhibited major electrical conduction problems (right axel rotation and
LBBB), and 33% exhibited lateral re-polarization problems. Among the
vaccinated/chronically-infected dogs, 33% exhibited electrical conduction
problems (right axel rotation and LBBB) similar to that noted in
unvaccinated/infected dogs, 33% showed minor axel rotation problems, and
33% dogs exhibited normal EKG. On a scale of 0 (normal) to 10 (severe EKG
alterations), 66% of non-vaccinated/chronically infected dogs were graded as
10 and 33% as normal (zero EKG alterations). In comparison, 33% of
vaccinated/chronically infected dogs were graded normal (0), 33% moderate
(score: 5), and 33% with severe electrical conduction problems (score: 10) at
one year post-infection.
Next, we evaluated the pathology of the heart in dogs. Anatomo-pathological analysis
of the heart, performed at day 60 pi, showed dilated cardiomyopathy (bi-ventricular
dilation) and focal and diffused myocarditis in vaccinated dogs as well as in dogs
injected with vector alone (Fig.
6). Irrespective of vaccination status, some animals exhibited whitish
zones and rounded edges in the spleen, dilation of esophagus, and pinkish ampoules
at the cecum. At one-year post-challenge infection, all dogs had round shaped
hearts. Sixty six percent of non-vaccinated/chronically infected dogs exhibited
severe right ventricle dilation. In comparison, 66% of
vaccinated/chronically-infected dogs exhibited moderate level of right ventricle
dilation. Epicardial hemorrhages were seen in 66% dogs from the control group
and 33% dogs in the TcVac1 group.
Histopathology studies on day 60 pi demonstrated some differences between two groups
(Fig. 7). In epicardium of
dogs vaccinated with TcVac1, non-suppurative moderate to severe myocarditis with
focal or zonal mononuclear and polymorphonuclear inflammatory infiltrate associated
with presence of amastigotes nests and severe active necrosis was generally noted.
In non-vaccinated/infected dogs, histopathological findings were similar to that
noted in vaccinated dogs with the exception that inflammatory infiltrate tended to
be mainly constituted of mononuclear rather than polymorphonuclear cell type (Fig. 7). In the myocardium,
vaccinated and non-vaccinated dogs exhibited multiple coagulative necrosis foci with
mononuclear infiltration and some necrotic areas with polymorphonuclear and
neutrophil infiltration. The diffused multi-focal and zonal mononuclear inflammatory
infiltrate and hemorrhagic areas in myocardium appeared to be larger in
vaccinated/infected dogs (Fig. 7C,
C2) as compared to non-vaccinated/infected dogs (Fig. 7B, B2). Vaccinated/infected dogs also
showed abundant cellular detritus in myocardium. Mononuclear infiltration was
moderate in ventricles and septum of all dogs. Folding of myocardial fibers in right
ventricle was observed in both vaccinated/acutely-infected and
non-vaccinated/acutely-infected dogs. Mural multifocal endocarditis with mononuclear
infiltration was also noted in dogs from the two groups. Abundant amastigote nests
(range 18–21 per microscopic field (mf)) were found in each region (right and
left ventricles, and septum) of the heart of TcVac1-vaccinated/acutely-infected
dogs. Non-vaccinated/infected dogs exhibited, in general, lesser number of parasite
foci in equivalent studied areas. Statistical analysis showed that overall, the
number of myocardial necrotic foci, lymphocyte infiltration foci and number of
amastigote nests were more abundant in TcVac1-vaccinated dogs than was observed in
non-vaccinated/infected dogs at 60 dpi (p<0.05) (Fig. 7).
At one-year post-challenge infection, severe myocardial inflammation persisted in
66% of the non-vaccinated/infected dogs while remaining 33% exhibited
slight inflammatory infiltrate in the heart. In comparison, 66% of
vaccinated/chronically infected dogs exhibited moderate level of myocardial
inflammatory infiltrate. Slight to moderate presence of connective tissue was
apparent in all chronically infected dogs; however, it was more evident in
TcVac1-vaccinated/chronic dogs. Folding of myocardial fibers and vacuolization of
Purkinje fibers was observed in 33% of vaccinated/chronically infected
dogs.
The objective of the present study was to test the efficacy of a multi-component DNA
vaccine (TcVac1) in dogs. The antigenic candidates included in TcVac1 were
identified by computational analysis of T. cruzi sequence database
and selected because they were conserved among several clinically relevant
T. cruzi strains, expressed in infective and intracellular
stages of T. cruzi
[9], and
recognized by the antibody and T cell response in infected mice [32] and humans
(unpublished data). When delivered as a DNA vaccine in mice, TcG1, TcG2 and TcG4
elicited trypanolytic antibody response and Th1 cytokines (e.g. IFN-γ) [9] that resulted in
significant protection from acute infection and chronic disease severity. We
utilized IL-12 and GM-CSF expression plasmids as adjuvants as these cytokines induce
type 1 B and T cell responses [33], and shown to significantly enhance the protective
immunity elicited by the vaccine candidates in mice [8], [39] and dogs [33]. To the best of our
knowledge, this is the first report testing the prophylactic and
transmission-blocking efficacy of DNA vaccine against T. cruzi in
dogs.
Immunization of dogs with TcVac1 resulted in elicitation of antigen-specific and
parasite-specific antibody response that was dominated by IgG2 subtype. The delivery
of booster doses of vaccine resulted in no significant increase in antibody response
that could be explained, at least partially, by the fact that DNA delivery system,
used in this study, is known to drive a low level of antigen expression. Several
investigators have reported that needle delivery of DNA vaccines in muscle induce
low immune response in large animals and humans, even when 1000-fold higher doses of
DNA than those proved to be effective in rodents were given [40], [41]. Other DNA vaccine delivery
systems such as gene gun (biolistic gun) [42], [43], adenovirus or vaccinia virus
delivery vectors [44], replicating attenuated strains of intracellular
microorganisms, such as Salmonella
[45] have shown
promising results in eliciting antigen expression. Additionally, heterologous
prime/boost approaches are noted to be more effective in eliciting stronger,
long-term immunity against intracellular pathogens [46], [47], to be tested in future
studies.
Despite low vaccine-induced antibody response, vaccinated dogs, upon challenge
infection with T. cruzi, exhibited an early expansion in antibody
response. The Ig (G+M) response during acute phase was of higher magnitude in
vaccinated/infected dogs than that observed in dogs injected with vector alone. The
IgG response in vaccinated/infected dogs was primarily of the Th1 type with
IgG2/IgG1 ratios being >1, known to provide protection from acute infection in
dogs [48], [49]. The higher
level and rapid expansion of antibody titers indicates that TcVac1 primed the B cell
response that was expanded upon exposure to T. cruzi. Previously,
we have shown that TcG1-, TcG2- and TcG4-specific antibodies,
elicited in vaccinated mice, were lytic in nature, and efficiently killed
trypomastigotes in a complement-dependent manner [9]. In this study, our observation
of a shorter detectable parasitemic period of 37 days in vaccinated/infected dogs
than that noted in control dogs (44 days) indicate that antibody response primed by
vaccination with TcVac1 was lytic in nature and contributed to a control of blood
parasitemia. Yet, immunization with TcVac1 failed to prevent peak parasitemia. This
was likely because other components of immune system, i.e., innate response
constituted by phagocytes and type 1 biased CD8+ T cell response,
were not strongly primed by vaccine or expanded upon challenge infection in
vaccinated dogs. It is well documented that phagocytes, through activation of NADPH
oxidase, MPO, and iNOS activities and production of cytotoxic reactive oxygen and
nitrogen species, play an important role in control of T. cruzi
[37], [50]-[52]. Numerous
studies have also demonstrated that an efficient control of acute parasitemia
requires concerted activities of Th1 helper cells, and cytotoxic
CD8+ T lymphocytes (CTLs) (reviewed in [53], [54]). Vaccination with TcVac1
resulted in a suppression of phagocytes’ response to challenge infection as
was evidenced by decreased activation of MPO and iNOS activity in
vaccinated/infected dogs as compared to that noted in non-vaccinated/acutely
infected dogs. Equally, immunization with TcVac1 resulted in a significant but only
moderately better expansion of CD8+ T cells and IFN-γ levels
upon challenge infection when compared to that noted in non-vaccinated/infected
dogs. Consequently, it was not surprising to find no significant decline in
infectivity of vaccinated dogs to triatomines during the acute period of infection.
All triatomines, fed on vaccinated/infected or non-vaccinated/infected dogs when
dogs were exhibiting peak parasitemia (day 30 pi), and analyzed at day 60
post-feeding, were infected evidenced by fecal presence of T.
cruzi. However, at 60 dpi, vaccinated dogs exhibited a better control of
parasitemia and moderately reduced infectivity to triatomines (52.63% versus
84.6% infected). The mathematical modeling of transmission dynamics [55] and other
studies using insecticide-treated dog collars [56], [57] indicate that a decline in
infectivity to <20% would be required to block vectorial transmission of
T. cruzi to humans. Thus, we surmise that current formulation
of TcVac1, though provided a decline in dogs’ infectivity to triatomines after
peak parasitemia, would not be effective in blocking the transmission cycle, and
further improvement in vaccination strategy is required.
Infection of dogs with SylvioX10/4 strain of T. cruzi produced
reproducible acute phase and chronic pathology as we have previously reported [18],
causing sudden death in some of the infected dogs, and cardiomyopathic changes in
most of the infected animals during acute stage. EKG alterations were found in more
than half of the acutely infected dogs and ranged from electrical conduction
problems, ventricular dilatations, pericarditis, myocarditis, high lateral necrosis,
and arrhythmia. Most of these changes were validated by necropsy and histopathology
findings, thus, confirming that this T. cruzi strain is highly
pathogenic in dogs. Despite a similar or higher infiltration of inflammatory
infiltrate in the heart, vaccinated dogs exhibited a significantly higher number of
amastigote nests (P<0.05) in cardiac tissue than was observed in non-vaccinated
control dogs (Fig. 7). Others
have reported a direct correlation between in vitro infectivity and
blood parasitism kinetics with heart parasitism intensity during long-term infection
of Beagle dogs [58], [59]. Because of high inflammatory infiltrate and tissue
parasite burden, two of the vaccinated dogs exhibited myocarditis and died suddenly
due to arrhythmia.
Chronically infected/vaccinated dogs were better equipped in controlling the disease
symptoms. EKG findings demonstrated mild-to-moderate cardiac alterations in animals
given TcVac1 vaccine while severe EKG alterations persisted in dogs injected with
vector alone. These findings were supported by anatomo-pathological analysis
performed at one-year post-challenge infection. Anatomo-pathological lesions and
epicardial hemorrhages were fewer and moderate in TcVac1-vaccinated/chronically
infected dogs as compared to non-vaccinated/chronic dogs that exhibited severe right
ventricle dilation and extensive epicardial hemorrhages. These findings were
observed despite no decline in inflammatory infiltrate in the heart in chronically
infected dogs. These data indicate that TcVac1-induced immunity was at least
partially effective in controlling the clinical progression of cardiac disease
severity in chagasic dogs.
Summarizing, in this study, we tested a multi-component DNA vaccine against
T. cruzi infection in dogs. Our data showed that TcVac1 geared
a modest parasite- and antigen-specific type 1 antibody and CD8+ T
cell response that was effective in providing an early control of acute parasitemia
and moderately decreased the infectivity of dogs to triatomines. However, tissue
parasite burden was not controlled in vaccinated dogs, likely due to suppression of
phagocytic cell response, evidenced by decreased myeloperoxidase and nitrite (iNOS)
levels in immunized dogs. Despite this, vaccinated dogs exhibited a moderate decline
in cardiac alterations determined by EKG and anatomo-/histo-pathological analysis
during chronic stage of disease development. Overall, our data demonstrated that
TcVac1-elicited immunity provided a partial protection from chronic Chagas disease
and provided an impetus to further improve the vaccination strategy against Chagas
disease.
|
10.1371/journal.pgen.1004604 | Foxf Genes Integrate Tbx5 and Hedgehog Pathways in the Second Heart Field for Cardiac Septation | The Second Heart Field (SHF) has been implicated in several forms of congenital heart disease (CHD), including atrioventricular septal defects (AVSDs). Identifying the SHF gene regulatory networks required for atrioventricular septation is therefore an essential goal for understanding the molecular basis of AVSDs. We defined a SHF Hedgehog-dependent gene regulatory network using whole genome transcriptional profiling and GLI-chromatin interaction studies. The Forkhead box transcription factors Foxf1a and Foxf2 were identified as SHF Hedgehog targets. Compound haploinsufficiency for Foxf1a and Foxf2 caused atrioventricular septal defects, demonstrating the biological relevance of this regulatory network. We identified a Foxf1a cis-regulatory element that bound the Hedgehog transcriptional regulators GLI1 and GLI3 and the T-box transcription factor TBX5 in vivo. GLI1 and TBX5 synergistically activated transcription from this cis-regulatory element in vitro. This enhancer drove reproducible expression in vivo in the posterior SHF, the only region where Gli1 and Tbx5 expression overlaps. Our findings implicate Foxf genes in atrioventricular septation, describe the molecular underpinnings of the genetic interaction between Hedgehog signaling and Tbx5, and establish a molecular model for the selection of the SHF gene regulatory network for cardiac septation.
| Atrioventricular septal defects (AVSDs) are a common severe class of congenital heart defects. Recent work demonstrates that events in the second heart field (SHF) progenitors, rather than in the heart, drive atrioventricular (AV) septation. Our laboratory has shown that both Hedgehog signaling and the T-box transcription factor, Tbx5, are required in the SHF for AV septation. To understand the molecular underpinnings of the AV septation process we investigated SHF Hedgehog-dependent gene regulatory networks. Transcriptional profiling and chromatin interaction assays identified the Forkhead box transcription factors Foxf1a and Foxf2 as SHF Hedgehog targets. Compound haploinsufficiency for Foxf1a and Foxf2 caused AVSDs in mice, demonstrating the biological relevance of this pathway. We identified a cis-regulatory element at Foxf1a that bound TBX5 and Hedgehog transcriptional regulators GLI1 and GLI3 in-vivo. Furthermore, TBX5 and Gli1 co-activate transcription of the identified cis-regulatory element in-vitro. The enhancer is expressed primarily in the pSHF in-vivo, where Tbx5 and Gli1 expression overlap. Our findings implicate Foxf1a and Foxf2 in AV septation and establish Tbx5 and Hedgehog signaling upstream of Foxf genes in a gene regulatory network for cardiac septation.
| Cardiac septation, the morphogenetic process that transitions the looped heart tube into the multi-chambered heart observed in mammals, is complex and often goes awry in Congenital Heart Disease (CHD). Atrioventricular septation is the crucial process that separates the common atrioventricular canal into right and left compartments. Atrioventricular septal defects (AVSDs) are a common severe form of CHD. A novel paradigm for the developmental ontogeny of the atrioventricular septum has recently emerged [1]–[6]. This work describes atrioventricular septation as a process driven by molecular events in second heart field (SHF) cardiac progenitors rather than in the heart itself [1]–[6]. The identification of extracardiac lineages that generate the atrial and atrioventricular septum implies that the search for gene regulatory networks germane to cardiac septation should occur in SHF cardiac progenitors not in the heart itself.
Hedgehog signaling is an essential developmental pathway conserved from flies to man [7], [8]. Mutations in key Hedgehog pathway genes, including ligands such as Sonic hedgehog (Shh; 20423) and downstream signaling cascade member Smoothened (Smo; 319757) cause significant cardiac defects including complete atrioventricular septal defects [9], [10]. Tissue specific knockout of Hedgehog signaling in the SHF recapitulates atrioventricular septal defects [4], [5] and genetic inducible fate mapping showed that the atrial/atrioventricular septum is derived from Hedgehog-receiving SHF cardiac progenitors [5]. These observations laid the groundwork for identifying the Hedgehog-dependent SHF gene regulatory networks essential for atrial septation.
Cardiogenic transcription factor genes Tbx5 (21388), Nkx2.5 (18091) and GATA4 (14463) have been implicated in human atrial septation [11]–[14]. These transcription factors form a complex and can co-activate gene expression [12], [15]–[17]. Tbx5 has been shown to be required in multiple contexts during cardiac development and adult function in mice. Tbx5 is required in the SHF for atrioventricular septation [6], [15], in embryonic cardiomyocytes for proliferation [18], in adult myocardium for contractile function [19], and in the adult cardiac conduction system for cardiac rhythm control [20]. Tbx5 target genes differ significantly between these distinct cellular and temporal contexts [6], [21]. Yet the Tbx5-responsive cis-regulatory elements specific to these cellular contexts and the molecular cues that establish context dependent selectivity remain unknown.
We previously described genetic interactions between Tbx5 and Hedgehog signaling in the SHF for atrioventricular septation in mice [6]. Mice haploinsufficient for both Tbx5 and the obligate Hedgehog signaling receptor gene Smo express AVSDs more frequently than mice haploinsufficient for either gene alone [6]. Furthermore, constitutive Hedgehog signaling in Tbx5-mutant SHF progenitors can rescue atrioventricular septation [6]. These studies predict that Hedgehog-dependent and Tbx5-dependent gene regulatory networks share vital, yet undescribed overlap in the SHF that is necessary for atrioventricular septation.
In this study we attempted to define Hedgehog-dependent SHF gene regulatory networks and identify the molecular basis of the genetic interaction between Hedgehog signaling and Tbx5. We characterized the Hedgehog-dependent SHF gene regulatory networks by in vivo whole genome transcriptional profiling and GLI-chromatin interaction studies. We found that Foxf1a (15227) and Foxf2 (14238) are downstream of Hedgehog signaling in the SHF. Mice haploinsufficient for both Foxf1a and Foxf2 compound heterozygotes have atrial septal defects, demonstrating the biological relevance of these Hedgehog targets. GLI3T (14634) binding data identified a candidate cis-regulatory element upstream of Foxf1a that contained an adjacent Tbx5 binding site. This enhancer binds to GLI1 (14632), GLI3 and TBX5 in the SHF in vivo. In vitro and in vivo analysis demonstrated that this cis-regulatory element integrates Hedgehog signaling with Tbx5 activity and provides strong specific activity in the posterior SHF. This work identifies a novel role for Foxf transcription factors at the intersection of Tbx5 and hedgehog signaling in atrioventricular septation and describes a SHF gene regulatory network for cardiac morphogenesis.
Progenitor cells for the atrial and atrioventricular septum require Shh signaling in the posterior SHF (pSHF) between embryonic day 8 and embryonic day 10 (E8–E10) to migrate into the heart to form the atrial septum between E9–E11 [4], [5]. To identify the Hedgehog-dependent gene regulatory networks required for this process, we compared transcriptional profiling of the posterior SHF from wild-type and Shh (MGI: 1932461) null embryos at E9.5 to identify differentially expressed transcripts. We isolated the pSHF by microdissection including the dorsal mesenchymal protrusion and closely associated surrounding ventral lateral plate mesenchyme. Our dissection included the attached foregut, but excluded the heart, dorsal lateral plate mesenchyme and neural tube (Figure 1A). RNA was isolated and known Hedgehog-dependent transcripts were evaluated by RT-PCR to verify genotyping prior to whole genome transcriptional profiling. Shh, Ptch1 (19206) and Gli1 all demonstrated significantly reduced expression (p>0.05) in the Shh null samples compared to wild-type micro-dissected samples (Figure 1B). Specifically, Shh was reduced more than 90%, while Ptch1 and Gli1 were each reduced approximately 50%, consistent with significantly reduced Hedgehog signaling in the mutant samples and confirming the genotypic fidelity of the isolated samples.
Transcriptional profiling of pSHF samples was performed on Agilent Mouse Whole Genome Arrays. Using a significance threshold with a multi-test adjusted p-value (Q-value) <0.005 and absolute fold change larger than 2, comparing Shh−/− mutant mouse embryos (n = 4) with wild-type embryos (n = 3) identified a differentially expressed 560-gene signature (Table S1). Gene Ontology (GO) enrichment analysis of differentially expressed genes captured known processes disrupted in Hedgehog pathway mutants, such as pattern specification and organ morphogenesis (Figure 1C) [22]. To further identify the best candidates for an experimental validation, 65 genes were computationally evaluated according to more stringent criteria by three statistical tests (non-parameter Wilcox-tested theoretical p<0.15, empirical t-tested FDR<0.1, and absolute fold change>3, Figure S1) on the same data sets. From the Shh down-regulated candidates, we chose 21 targets and validated significant misexpression of 13 by qPCR (p<2e-16, Fisher's Exact test, FET) (Figure 1D). Eight others did not meet criteria for statistically significant misexpression primarily due to large expression variation, possibly related to the presence of non-SHF tissue isolated by our dissection process.
To define loci directly downstream of Hedgehog signaling, we analyzed genome-wide chromosomal binding locations of the Hedgehog transcriptional regulator Gli3 in the embryonic SHF by chromatin immunoprecipitation with deep sequencing (ChIP-seq). We performed ChIP using a Cre-inducible flag-tagged Gli3T expression line (RosaGli3TFlag c/c MGI: 3828280) [23] combined with the SHF Cre driver Mef2c-AHF-Cre [24] (MGI: 3639735). The SHF tissue from 50 Mef2cAHF-Cre+; RosaGli3TFlag/+ embryos was micro-dissected and immunoprecipitated using an anti-FlagM2 antibody (Sigma). To verify enrichment of Gli3T bound sequences by immunoprecipitation prior to sequencing, we tested a previously identified Gli3T peak upstream of Ptch1 (Chromosome 13, nucleotides 63577408–63579384, mm9), a known Gli3T-bound cis-regulatory element in the limb [23]. This sequence was 13.7-fold enriched in the SHF IP fraction by ChIP-PCR. We proceeded to sequence the IP library and apply Model-based Analysis for ChIP-Seq (MACS) [25]. We identified 1316 Gli3-bound peaks by comparing 68 million sequence tags in IP to 21 million sequence tags in input (tag size = 36 bps, effective genome size = 2e+9, band width = 200, 2<model fold<200, and p-value cutoff = 1e-05) [25]. From these peaks, we analyzed the distribution of the signal around the peak center and identified a typical distribution, confirming successful sequencing (Figure 2A). The predominant GLI3T peak location from the binding data was intergenic and a considerable distance from the transcriptional start sites. We therefore considered the possibility that genes and up to 100 kbp in both directions from intergenic peaks may fall under control of GLI-mediated cis-regulatory elements, given that enhancers often reside thousands of base pairs away from their target of regulation and act independently of their orientation [26], [27]. We therefore annotated GLI3T-bound regions to all transcription start sites within 100 kbp and to the nearest TSS if it resided outside the 100 kbp window [28], [29]. This consideration resulted in mapping the 1316 peaks to 3296 neighbor genes (Table S2). The enrichment between GLI3T-bound and Shh-dependent genes was significant among approximately 22,000 mouse genes (FET p<0.01, Figure 2B).
To define the direct Hedgehog-dependent SHF gene regulatory networks, we intersected the SHF Shh-dependent transcriptional profile signature with the SHF Gli3T chromatin contact results to define candidate Hedgehog-dependent Gli-target genes. This dataset intersection comprised 119 peaks annotated to 112 genes (Figure 2B, Table S3). The enrichment between Gli3T-bound and Shh-dependent genes was significant among ∼22k mouse genes (FET p = 0.003, odds ration = 1.4, Figure 2B). The 119 Shh-dependent Gli3T-bound sites contained significant enrichment of the de novo and known Gli3-binding motif, as derived by ChIP-Chip (CGTGGGTGGTCC) [23] and by computational implication (TRANSFAC database; Figure 2C, bottom panel) [30], [31] at a high degree of significance (p≤1e-10; Figures 2C, top panel).
A significant enrichment of transcription factors was observed in SHF Hedgehog target genes. Transcription factor activity and DNA binding were the two most significant gene-sets over-represented among the 112 Shh-dependent Gli3-bound genes. We directly analyzed our gene set for overrepresentation of transcription factors by searching TRANSFAC version 2013.1 [31] and identified 26 TFs among the 112 unique genes with significant Gli3T-bound peaks (Figure 2D, Table S4), representing a significant enrichment (p = 0.0001, odds ratio = 2.7, Fishers exact test). Specifically, Shh transcriptional profiling and GLI3T chromatin interaction data both identified an enrichment of FOX gene family members, encoding Forkhead transcription factors, identifying FOX genes as potential SHF Hedgehog targets (Figure S2). The set of 112 Shh-dependent Gli3T bound genes included four Fox transcription factors, Foxb1 (64290), Foxc1 (17300), Foxd1 (15229) and Foxf1a, representing a significant enrichment (Figure 2E, p = 0.0001, odds ratio = 18.4).
We investigated the hypothesis that Foxf1a and Foxf2 expression was downstream of Hedgehog signaling in cardiac development. Shh-dependent expression of both genes in the SHF was confirmed by qPCR: Foxf1a expression was reduced by 50% (p = 0.05) and Foxf2 was reduced by 80% in the SHF of Shh−/− versus wild-type controls (p = 0.01) (Figure 1D). In-situ hybridization to evaluate the patterning of expression showed that Foxf1a and Foxf2 were both expressed in the posterior SHF, but not in the heart, in wild-type embryos at E9.5, with Foxf1a expression extending more ventrally than Foxf2 to include the DMP (Figures 3A, A′, E, E′). Mesenchymal expression of both Foxf1a and Foxf2 demonstrated a severe decrement in shh−/− mutant embryos (Figures 3B, B′, F, F′).
In a search for common targets between Tbx5 and Hedgehog signaling in the SHF, we tested whether Foxf1a and/or Foxf2 SHF expression was Tbx5-dependent. We performed in situ hybridization for Foxf1a and Foxf2 in Tbx5+/− heterozygous mutant embryos (MGI: 2387850), which demonstrate 40% penetrance of AVSDs [15]. Foxf1a but not Foxf2 expression demonstrated significant reduction in Tbx5 heterozygotes at E9.5. In Tbx5+/− embryos, Foxf1a expression was specifically decreased in the posterior SHF (Figure 3C, C′, D, D′, arrow) in the area of expression overlap with Tbx5 expression [6]. In regions where Foxf1a expression does not overlap with Tbx5 expression, such as the anterior SHF, Foxf1a expression appeared normal (Figure 3D, D′). Foxf2 expression in Tbx5+/− embryos appeared unaltered compared to wild-type embryos (Figure 3G, G′, H, H′). Taken together, this analysis demonstrates that posterior SHF Foxf1a expression was Shh- and Tbx5-dependent whereas Foxf2 pSHF expression was Shh-dependent alone.
We hypothesized that Foxf1a and Foxf2 were required in a dosage sensitive manner for atrioventricular septation. We analyzed the cardiac anatomy of embryos from an intercross between Foxf1a+/− and Foxf2+/− at E14.5, when cardiac septation is normally complete. Foxf1a+/−; Foxf2+/− double-heterozygote embryos all exhibited atrioventricular septal defects (Figure 4D, D′ asterisk; p = 0.03). Primum-type atrial septal defects, characterized by absence of the dorsal mesenchymal protrusion, were observed in each case (Figure 4D, D′). Additionally, Foxf1a+/−; Foxf2+/− double-heterozygotes displayed larger than normal mesenchymal caps covering the primary atrial septum (Figure 4D′ arrow), an observation in keeping with the known redundant requirement for Foxf1a and Foxf2 in limiting mesenchymal growth in other contexts [32]. Atrial septal defects were never observed in Foxf1a+/− (Figure 4B, B′) or Foxf2+/− (Figure 4C, C′) single-heterozygotes or wildtype control littermate embryos (Figure 4A, A′). We concluded that Foxf1a and Foxf2 are redundantly required for atrioventricular septation.
We hypothesized that Foxf1a may represent a direct downstream target of Hedgehog signaling and/or Tbx5 in the SHF. We identified Foxf1a as a candidate direct target based on unbiased interrogation of GLI3T and TBX5 transcription factor chromatin interaction and transcriptional profiling data sets. We intersected our SHF GLI3T ChIP data set (Figure 2B) with a published TBX5 ChIP-seq data set generated from HL-1 cardiomyocytes [33] to define regions with potential co-occupancy of both transcription factors. The intersection of the ChIP-seq datasets identified a single overlapping interaction peak for Gli3T (in the SHF (Figure 2B)) and TBX5 (in HL-1 cardiomyocytes) [33] located approximately 90 kb upstream of the Foxf1a transcription start site (Figure 5A and Figure S3). The Foxf1a transcriptional start site is the closest protein-coding gene to the described peak. The transcriptional start site for a non-coding RNA is located approximately 1.3 kbp upstream of Foxf1a, oriented in the opposite direction [34]. Closer interrogation of the sequence underlying the interaction domains revealed a conserved canonical T-box binding site (AGGTGTGG; chr 8, nucleotides 123,517,714–721, NCBI137/mm9 assembly) and a conserved canonical Gli binding site (GGACCACCCAGC; chr 8, nucleotides 123,517,754–762, NCBI137/mm9 assembly) within 30 base pairs of one another (Figure 5A). We evaluated the sequence information content for these sites from our SHF Gli3 ChIP-seq experiment and found close agreement with published binding sites for Gli3 [23], [33]. This chromatin interaction data in combination with the Tbx5 and Hedgehog signaling-dependent Foxf1a SHF expression (Figure 3) identified this conserved region (mouse chromosome 8, nucleotides 123,517,714–762) as a candidate Foxf1a cis-regulatory element.
We evaluated the binding of TBX5 and the Hedgehog transcriptional regulators GLI1 and GLI3 to the candidate cis-regulatory element at Foxf1a in vivo in the SHF. We evaluated TBX5 binding in vivo by performing ChIP using an anti-TBX5 antibody on the micro-dissected wildtype SHF at E10.5 and observed 35-fold enrichment of the cis-regulatory element in the TBX5-immunoprecipitation fraction compared to the input fraction by qPCR (Figure 5B). We evaluated GLI1 and GLI3T binding in vivo by performing ChIP on the micro-dissected SHF of mice carrying either a Cre-activated flag-tagged Gli3 (RosaGli3TFlag c/c) or Gli1 allele (RosaGli1Flag c/c MGI: 4460761) in concert with the Nkx2.5-Cre (MGI 2654594), broadly expressed cardiac tissues and progenitors. We performed ChIP using an anti-flag antibody on the SHF from R26R-Gli3-flagNkx2.5-Cre/+ or R26R-Gli1-flagNkx2.5-Cre/+ embryos at E10.5 and observed, respectively, 6.8-fold and 7.1-fold enrichment of the Foxf1a cis-regulatory element in the GLI1- and GLI3T-overexpressing embryos over the input control by qPCR (Figure 5B). We also evaluated two genomic loci between our identified binding site and the Foxf1a transcription start site to determine whether nonspecific pulldown occurred in our ChIP experiments. These loci were not significantly enriched in the IP'd DNA (Figure S4) These results validate in vivo SHF binding of TBX5, GLI1, and GLI3 to the candidate cis-regulatory element at Foxf1a.
We hypothesized that the conserved, adjacent, and functional in vivo Gli and Tbx5 binding sites may integrate Tbx5 and Hedgehog activity as a component of a downstream gene regulatory network. We evaluated the activity of TBX5 and GLI1 on the candidate Foxf1a enhancer in vitro. The conserved element was cloned into a pGL4.23 vector containing a minimal promoter driving luciferase as a transcriptional readout and was transfected into HEK293T cells along with expression vectors for Gli1 and/or Tbx5. Co-transfection with the expression vector for Gli1, a Hedgehog-responsive transcriptional activator, provided a 91.9-fold induction of luciferase activity (p = 0.0017). Co-transfection with the expression vector for Tbx5 alone provided a 3.9-fold increase of luciferase activity (p = 0.039). Co-transfection with both Gli1 and Tbx5 expression constructs provided a 171.6-fold increase in luciferase activity (p = 0.00091), demonstrating synergistic activity between these transcriptional co-activators (Figure 5C).
We assessed the requirement of TBX5 and GLI binding sites for transcriptional activation of the enhancer. To assess the requirement of TBX5-dependant transcriptional activation of the enhancer on TBX5 binding sites, a TBX5-mutant enhancer-luciferase construct with the 7 base pair core of 3 canonical TBX binding sites was generated by site-directed mutagenesis. This TBX5-mutant construct eliminated transcriptional activation by TBX5 alone (p = 0.04) and limited transcriptional activation by TBX5 and GLI1 together (p = 0.006) (Figure 5C). A GLI-mutant enhancer-luciferase construct was also constructed with the 8 base pair core of 3 canonical binding sites altered by site-directed mutagenesis (see materials and methods). This GLI-mutant construct profoundly diminished transcriptional activation by GLI1 alone (p = 0.001) (Figure 5C). Interestingly, transcriptional activation by TBX5 and GLI1 on the GLI-mutant enhancer construct was only modestly abrogated luciferase compared to the activity of GLI1 and TBX5 on the wild-type enhancer (p = 0.003).
We hypothesized that the cis-regulatory element at Foxf1a may integrate Hedgehog signaling and Tbx5 activity as a SHF-specific enhancer in vivo. We cloned the Foxf1a genomic region into an Hsp68-LacZ expression construct, whose minimal promoter affords no intrinsic in vivo expression activity [35]. We evaluated the enhancer activity of the Foxf1a genomic fragment by evaluating LacZ expression in transient transgenic mouse embryos at E9.5. The posterior SHF demonstrated strong lacZ expression and was the only anatomic region demonstrating consistent and robust expression in the 8 transgenic embryos genetically positive for LacZ (Figure 5D). Interestingly, the SHF region with the most consistent and robust expression was the area of overlap between Hedgehog signaling and Tbx5 expression [6], including the early dorsal mesenchymal protrusion and surrounding mesenchyme of the posterior SHF (less frequent and intense expression was also observed in other anatomic locations that receive Hedgehog signaling outside of the Tbx5 expression domain, including the anterior SHF (5/8), anterior lateral plate mesoderm (5/8) and somites (2/8) (Figure 5D). These observations, in concert with the in vitro analysis suggested that Hedgehog and Tbx5 act synergistically to provide strong reproducible transcriptional activation of this enhancer in the posterior SHF.
Identification of the gene regulatory networks required for atrioventricular septation will be the basis for a mechanistic understanding of AVSDs, a common severe form of CHD. We investigated Hedgehog-dependent networks and implicated Foxf genes for the first time in vertebrate heart development. We examined the overlap between Hedgehog pathways and Tbx5, both known to be integral in the pSHF for atrioventricular septation. Using transcription factor-chromatin interaction data, we identified a cis-regulatory element at Foxf1a that bound both TBX5 and the Hedgehog pathway transcriptional activator GLI1. In vitro analysis of TBX5 and GLI activity on the cis-regulatory element at Foxf1a proved predictive of in vivo biology: this enhancer exhibited strong transcriptional activation only in the pSHF region where Tbx5 expression and Hedgehog signaling intersect. This region is the location of atrial septum progenitors [5], where both Hedgehog signaling and Tbx5 are required for atrioventricular septation (Figure 6) [4]–[6]. These observations provide molecular detail for the genetic interaction between Tbx5 and Hedgehog signaling in atrioventricular septation.
Our observations identified a requirement for the Forkhead-box transcription factors Foxf1a and Foxf2 in heart development. Compound haploinsufficiency for both Foxf1a and Foxf2 caused an atrial septal defect of the primum type, an atrioventricular septal defect characterized by absence of the dorsal mesenchymal protrusion. Foxf1a and Foxf2 were expressed selectively in the SHF, not in the heart (Figure 3). The requirement for Foxf genes in atrioventricular septation (Figure 4) provided further support for a model of atrioventricular septation as driven by molecular events in SHF cardiac progenitors as opposed to in the heart itself. We found that Foxf1a and Foxf2 are required downstream of Hedgehog signaling in atrioventricular septation, adding cardiac development to the previously described Hedgehog-dependent role for Foxf genes in murine gut development [32], [36]–[37]. Atrioventricular septal defects are also observed in Shh-null mutant embryos [10]. Because Foxf1a and Foxf2 expression were each decreased in the SHF by more than 50% in shh−/− null embryos (Figure 1D), Foxf1a+/−; Foxf2+/− double heterozygote embryos provided a reasonable developmental facsimile of their diminished expression levels in shh−/− embryos. The observation that Foxf1a and Foxf2 compound haploinsufficiency resulted in AVSDs is therefore consistent with the supposition that Foxf1a and Foxf2 are essential components of the Hedgehog-dependent SHF gene regulatory network.
Foxf genes have also been previously implicated in cardiac specification in the ascidian Ciona intestinalis [38], [39]. In ascidians, the single Foxf orthologue lies at the center of a pathway regulating numerous migration-related cellular processes, such as polarity, migration and membrane protrusion in trunk ventral cardiac progenitor cells [38], [39]. Ciona trunk ventral cells with disrupted Foxf activity fail to migrate properly, but still differentiate into cardiac tissue at an improper location. Interestingly, removing Hedgehog signaling from the mouse SHF causes a migration failure of SHF progenitors [4], [5]. Like the Ciona trunk cells without Foxf, SHF cells without Hedgehog responsiveness differentiate into cardiomyocytes, but their altered migration causes AVSDs [5]. Future efforts will determine whether the requirement for Foxf genes in cardiac progenitor migration is a conserved feature of mammalian cardiac development.
Genetic interaction and rescue experiments investigating the requirement for Hedgehog signaling and Tbx5 in atrioventricular septation were consistent with Tbx5 acting either in parallel or upstream of Hedgehog signaling in atrioventricular septation [6]. Our interrogation of these pathways on a cis-regulatory element at Foxf1a provides molecular detail for their interaction. We observed that TBX5 and GLI1, the Hedgehog-dependent transcriptional activator, synergistically activated the cis-regulatory element in vitro (Figure 5C) predicting strong activation of expression in areas of overlap between Tbx5 expression and Hedgehog signaling. This prediction held in vivo, where transcriptional activity of the enhancer was strong and reproducible only in the posterior SHF region, where Tbx5 expression and Hedgehog signaling overlap (Figures 5D and 6). This work is consistent and a model describing a SHF-specific gene regulatory network driven by GLI1 and TBX5 and essential for atrioventricular septation (Figure 6). This model provides specific predictions for the logic underlying enhancer choice in the SHF with ramifications for understanding the molecular and biochemical basis of atrioventricular septation and clinical AVSDs.
Mouse experiments were completed according to a protocol reviewed and approved by the Institutional Animal Care and Use Committee of the University of Chicago, in compliance with the USA Public Health Service Policy on Humane Care and Use of Laboratory Animals.
The Shh− line was obtained from the Jackson laboratory. The Tbx5+/− mice have been previously reported [15]. Foxf1+/− and Foxf2+/− mouse lines were generated in the Kalinichenko lab (Cincinnati Children's Hospital Medical Center) by breeding Foxf1afl/fl and Foxf2fl/fl mice with EIIA-Cre transgenic mice (Jackson Lab). Mef2c-AHF-Cre [24], ROSA26-Gli1 [40] and ROSA26-Gli3T [23] were reported previously.
For ChIP, transcriptional profiles and in-situ hybridizations, embryos were dissected in nuclease-free PBS on ice. For SHF microdissection procedures, head tissues anterior to the heart were removed, as were tail tissues posterior to the heart. Portions of these tissues were retained for genotyping if necessary. Neural tube tissues were also removed. The SHF mesenchyme was bisected into anterior and posterior portions when necessary, and then removed from the cardiac tissue (Figure 1A).
Shh+/+ and Shh−/− embryos were dissected as described above at E9.5. SHF tissues from these embryos were pooled to isolate sufficient amount of RNA for synthesis of labeled cRNA. Transcriptional profiles were performed using Agilent Mouse Whole Genome Arrays mgug4122a.
Microdissected SHF tissues were grouped into pools of approximately 50. Tissues were briefly fixed in 1.8% formaldehyde, then washed and homogenized. Sonication was performed with a Misonix 4000 sonicator until the sheared chromatin was approximately 100–300 bp in length. Input control samples were reserved prior to overnight immunoprecipitation with the appropriate antibody bound to magnetic Dynabeads (Invitrogen). Beads were precipitated and washed, the chromatin was eluted, de-crosslinked and purified using a PCR cleanup kit (Qiagen). To determine fold enrichment, qPCR was performed using input controls compared with DNA bound to immunoprecipitated proteins, using primers specific to the site of interest as well as primers to two sites not expected to be enriched.
ChIP-seq and microarray data were deposited in the Gene Expression Omnibus (GEO) database with a super accession number GSE44756.
In-situ hybridization was performed as in Moorman et al. [45] with slight modifications. Specifically, after post-hybridization washes with 50% formamide/2X SSC, specimens were incubated for 30 minutes at 37 degrees C in 20 ug/ml RNase A to remove unbound probe and reduce nonspecific staining. All in-situ hybridization experiments were performed on a minimum of three control and three experimental embryos.
Expression vectors for Gli1 and Gli3T were obtained from the Vokes lab. Tbx5 was cloned into the pCDNA 3.1 expression construct [20] Foxf1a fragment was cloned into the pGL4.23 vector (Promega). Expression and reporter vectors were transfected into HEK293T cells using FuGENE (Promega). Cells were cultured for 48 hours after transfection, then lysed and assayed using the Dual-Luciferase Reporter Assay system (Promega).
The Foxf1a enhancer and minimal promoter used in the luciferase assays were subcloned from the pENTR vector into the Hsp68-LacZ vector [35] using the Gateway system (Invitrogen). The resulting construct was digested with NotI enzyme to remove the pBlueScript backbone, gel-purified, injected into fertilized mouse eggs at the University of Chicago Transgenics Core Facility and implanted into female mice. Embryos were harvested at E9.5 and stained as described previously [5].
|
10.1371/journal.pntd.0006959 | The role of fibroblast growth factor signalling in Echinococcus multilocularis development and host-parasite interaction | Alveolar echinococcosis (AE) is a lethal zoonosis caused by the metacestode larva of the tapeworm Echinococcus multilocularis. The infection is characterized by tumour-like growth of the metacestode within the host liver, leading to extensive fibrosis and organ-failure. The molecular mechanisms of parasite organ tropism towards the liver and influences of liver cytokines and hormones on parasite development are little studied to date.
We show that the E. multilocularis larval stage expresses three members of the fibroblast growth factor (FGF) receptor family with homology to human FGF receptors. Using the Xenopus expression system we demonstrate that all three Echinococcus FGF receptors are activated in response to human acidic and basic FGF, which are present in the liver. In all three cases, activation could be prevented by addition of the tyrosine kinase (TK) inhibitor BIBF 1120, which is used to treat human cancer. At physiological concentrations, acidic and basic FGF significantly stimulated the formation of metacestode vesicles from parasite stem cells in vitro and supported metacestode growth. Furthermore, the parasite’s mitogen activated protein kinase signalling system was stimulated upon addition of human FGF. The survival of metacestode vesicles and parasite stem cells were drastically affected in vitro in the presence of BIBF 1120.
Our data indicate that mammalian FGF, which is present in the liver and upregulated during fibrosis, supports the establishment of the Echinococcus metacestode during AE by acting on an evolutionarily conserved parasite FGF signalling system. These data are valuable for understanding molecular mechanisms of organ tropism and host-parasite interaction in AE. Furthermore, our data indicate that the parasite’s FGF signalling systems are promising targets for the development of novel drugs against AE.
| To ensure proper communication between their different cell populations, animals rely on secreted hormones and cytokines that act on receptors of target cells. Most of the respective cytokines, such as FGFs, evolved over 500 million years ago and are present in similar form in all animals, including parasitic worms. The authors of this study show that the metacestode larva of the tapeworm E. multilocularis, which grows like a malignant tumor within the host liver, expresses molecules with homology to FGF receptors from mammals. The authors show that human FGF, which is abundantly present in the liver, stimulates metacestode development and that all parasite FGF receptors are activated by human FGF, despite 500 million years of evolutionary distance between both systems. This indicates that cells of the Echinococcus metacestode can directly communicate with cells of the mammalian host using evolutionarily conserved signaling molecules. This mode of host-pathogen interaction is unique for helminths and does not occur between mammals and single-celled pathogens such as protozoans or bacteria. The authors finally demonstrate that BIBF 1120, a drug used to treat human cancer, targets the Echinococcus FGF receptors and leads to parasite death. This opens new ways for the development of anti-parasitic drugs.
| The flatworm parasite E. multilocularis (fox-tapeworm) is the causative agent of alveolar echinococcosis (AE), one of the most dangerous zoonoses of the Northern hemisphere. Infections of intermediate hosts (rodents, humans) are initiated by oral uptake of infectious eggs which contain the parasite’s oncosphere larval stage [1,2]. After hatching in the host intestine and penetration of the intestinal wall the oncosphere gains access to the liver, where it undergoes a metamorphotic transition towards the metacestode larval stage [3]. The E. multilocularis metacestode consists of numerous vesicles which grow infiltratively, like a malignant tumor, into the liver tissue [1–3]. Due to the unrestricted growth of the metacestode, blood vessels and bile ducts of the liver tissue of the intermediate host are obstructed, eventually leading to organ failure [1]. Another hallmark of AE is extensive liver fibrosis which can lead to a complete disappearance of the liver parenchyma, and which most probably involves the activation of hepatic stellate cells during chronic infection [4,5]. Surgical removal of the parasite tissue, the only possible cure, is not feasible in the majority of patients leaving benzimidazole-based chemotherapy as the only treatment option. However, benzimidazoles act parasitostatic only and have to be given for prolonged periods of time (often life-long), underscoring the need for novel treatment options against AE [1].
We previously established that E. multilocularis development and larval growth is exclusively driven by a population of somatic stem cells, the germinative cells, which are the only mitotically active cells of the parasite and which give rise to all differentiated cells [6]. Using in vitro cultivation systems for metacestode vesicles and germinative cells [7–10], we also demonstrated that host insulin fosters parasite development by acting on evolutionarily conserved receptor kinases of the insulin receptor family that are expressed by the metacestode [11]. Evidence has also been obtained that host epidermal growth factor (EGF) stimulates Echinococcus germinative cell proliferation, most probably by acting on parasite receptor tyrosine kinases (RTK) of the EGF receptor family [12,13]. These studies indicate that the interaction of host-derived hormones and cytokines with corresponding receptors of evolutionarily conserved signalling pathways that are expressed by the parasite may play an important role in AE host-parasite interaction. Although the E. multilocularis genome project already indicated that the parasite expresses receptor TK of the fibroblast growth factor (FGF) receptor family in addition to insulin- and EGF-receptors [14], no studies concerning the effects of host FGF and their possible interaction with parasite FGF receptors have been carried out to date.
FGFs are an ancient group of polypeptide cytokines that are present in diploblastic animals, in deuterostomes and, among protostomes, only in ecdysozoa (with some distantly related members in lophotrochozoa)[15,16]. Humans express 22 different FGFs of which several (FGF11 –FGF14) are not secreted and act independently of FGF receptors in an intracrine modus only [15]. The remaining FGFs act in a paracrine fashion and are typically released via N-terminal signal peptides. Notable exceptions are the prototypic FGF1 (acidic FGF) and FGF2 (basic FGF) which are ubiquitously expressed in human tissues, are the most active members of the FGF family, and are released in a signal peptide-independent manner [15]. FGFs have a key role in metazoan embryonic development and, in adults, are typically involved in regeneration processes (angiogenesis, wound healing, liver regeneration, regeneration of nervous tissue)[15]. In the liver, particularly FGF1 but also FGF2 are present as proteins in significant amounts [17], are crucially involved in tissue regeneration upon damage [18,19], and are also upregulated and released during fibrosis [20]. Secreted FGFs act through surface receptor TK of the FGF receptor family, of which four isoforms, Fgfr1-Fgfr4, are expressed by humans [15]. The mammalian FGF receptors comprise an extracellular ligand-binding domain made up of three immunoglobulin (Ig)-like domains, a transmembrane domain, and a split intracellular kinase domain. FGF binding to the cognate FGF receptors typically results in receptor dimerization, transphosphorylation and subsequent activation of downstream signalling pathways such as the Ras-Raf-MAPK (mitogen-activated protein kinase) cascade or the PI3K/Akt pathway [15].
FGF signalling pathways have, in part, already been studied in flatworms. In the free-living planarian species Dugesia japonica, two members of the FGFR TK are expressed of which DjFGFR1 exhibits three immunoglobulin-like domains in the extracellular region whereas DjFGFR2 only contains two such domains [21]. Both receptors are expressed by X-ray sensitive planarian stem cells (neoblasts) and in cephalic ganglia and an important role of these FGFRs in planarian brain formation has been suggested [21,22]. Furthermore, similar FGF receptors were also detected in stem cells of the planarian Schmidtea mediterranea [23]. In the genome of the flatworm parasite species Schistosoma mansoni, two FGFR-encoding genes were identified of which fgfrA codes for a predicted protein with two extracellular immunoglobulin domains and a split TK domain whereas the fgfrB gene product only comprises one immunoglobulin domain in the extracellular region [24]. Expression of fgfrA and fgfrB in neoblast-like somatic stem cells has been shown and evidence was obtained for an important role of both receptors in schistosome stem cell maintenance [25–27]. Hahnel et al. [24] also demonstrated that both receptors are enzymatically active, are expressed in the gonads of schistosomes, and are upregulated following pairing, indicating a role in parasite fertility. Interestingly, these authors also showed that treatment of adult schistosomes with FGFR inhibitors leads to a reduction of somatic neoblast-like stem cells in both genders [24].
In the present work we provide a detailed analysis of three FGFRs in the cestode E. multilocularis and show that the expression patterns of these receptors differ from those in planaria and schistosomes. We also demonstrate that all three Echinococcus FGFRs are activated in response to human FGFs and that host FGF stimulates parasite development in vitro. Finally, we also show that inhibition of FGF signalling in Echinococcus larvae drastically reduces parasite development and survival.
FGF stimulation and inhibitor experiments were performed with the natural E. multilocularis isolate H95 [14]. Whole mount in situ hybridization was carried out using isolate GH09 which, in contrast to H95, is still capable of producing brood capsules and protoscoleces in vitro [14]. All isolates were continuously passaged in mongolian jirds (Meriones unguiculatus) as previously described [9]. The generation of metacestode vesicles and axenic cultivation of mature vesicles was performed essentially as previously described [7,9] with media changes usually every three days. Primary cell cultures were isolated from mature vesicles of isolate H95 and propagated in vitro essentially as previously described [8–10] with media changes every three days unless indicated otherwise. For FGF stimulation assays, 10 nM or 100 nM of recombinant human acidic FGF (FGF1) or basic FGF (FGF2)(both from ImmunoTools GmbH, Friesoythe, Germany) were freshly added to parasite cultures during medium changes. In the case of primary cells, cultivation was usually performed in cMEM medium which is host hepatocyte-conditioned DMEM (prepared as described in [10]). For inhibitor studies, specific concentrations of BIBF 1120 (Selleck Chemicals LLC, Houston, TX, USA) were added to parasite cultures as indicated and as negative control DMSO (0.1%) was used. The formation of mature metacestode vesicles from primary cells and measurement of metacestode vesicles size was performed essentially as previously described [11].
RNA isolation from in vitro cultivated axenic metacestode vesicles (isolate H95), protoscoleces (isolate GH09), and primary cells (H95, GH09) was performed using a Trizol (5Prime, Hamburg, Germany)-based method as previously described [11]. For reverse transcription, 2 μg total RNA was used and cDNA synthesis was performed using oligonucleotide CD3-RT (5’-ATC TCT TGA AAG GAT CCT GCA GGT26 V-3’). PCR products were cloned using the PCR cloning Kit (QIAGEN, Hilden, Germany) or the TOPO XL cloning Kit (invitrogen), and sequenced employing an ABI prism 377 DNA sequencer (Perkin-Elmer).
The full-length emfr1 cDNA was cloned using as starting material the partial sequence of a cDNA of the closely related cestode E. granulosus, which encoded a FGFR-like TK domain but which lacked the coding regions for transmembrane and extracellular parts [28]. Using primers directed against the E. granulosus sequence (5’-CTA CGC GTG CGT TTT CTG ATG-3’for first PCR; 5’-CCC TCT GAT CCA ACC TAC GAG-3’for nested PCR), the 3’ end of the corresponding E. multilocularis cDNA was subsequently PCR amplified from a metacestode (isolate H95) cDNA preparation using primers CD3 and CD3nest as previously described [29]. 5’-RACE was performed using the SMART RACE cDNA amplification kit (Clontech) according to the manufacturer’s instructions using primers 5’-ACC GTA TTT GGG TTG TGG TCG-3’ (first PCR) and 5’-GAA CAG GCA GAT CGG CAG-3’ (touchdown PCR) as previously described [30]. The presence of an in frame TAA stop codon 110 bp upstream of the emfr1 ATG start codon indicated that the correct 5’ end had been identified. In a final step, the entire emfr1 cDNA was PCR amplified from metacestode cDNA using primers 5’-GAC ACA TCT CCT TGG CCG-3’ and 5’-GCG AGT TGA TAC TTT ATG AGA G-3’ and cloned using the TOPO XL PCR cloning kit (Invitrogen). The sequence is available in the GenBankTM, EMBL, and DDJB databases under the accession number LT599044.
For emfr2 cloning we first identified by BLAST analyses on the published E. multilocularis genome sequence [14] a reading frame encoding a FGFR-like TKD annotated as EmuJ_000196200. Transcriptome analyses [14] and 5’_RACE experiments, however, indicated that there is actually read-through transcription between gene models EmuJ_000196300 and EmuJ_000196200. We thus designed primers 5’-ATG TGT CTC CGA GCT CTC TG-3’, binding to the 5’ end regions of gene model EmuJ_000196300, and primer 5’-TTA CTC GCT CGA TCG TGG GG-3’, binding to the reading frame 3’ end of gene model EmuJ_000196200, to PCR amplify the entire reading frame from metacestode cDNA. The resulting PCR fragment was subsequently cloned using the TOPO XL cloning kit (Invitrogen) and fully sequenced. The sequence is available in the GenBankTM, EMBL, and DDJB databases under the accession number LT599045.
For emfr3 cloning and sequencing we used primers directed against the CDS 5’ end (5’-ATG GCA CCT AAG GTT GTG TCA GGA-3’) and 3’ end (5’-GCA GAT GAG TAA GAA ACC CTC-3’) of gene model EmuJ_000893600 [14] for direct PCR amplification of the reading frame from metacestode cDNA. The resulting PCR fragment was subsequently cloned using the TOPO XL cloning kit (Invitrogen) and sequenced. The sequence is available in the GenBankTM, EMBL, and DDJB databases under the accession number LT599046.
Proliferation of E. multilocularis metacestode vesicles and primary cells was assessed by a bromodesoxyuridine (BrdU)-based method. Axenically cultivated metacestode vesicles (2–4 mm in diameter) were manually picked and incubated in 12-well plates (Greiner BioOne, Kremsmünster, Germany; 8 vesicles per well) in DMEM medium without serum for 2 days. Freshly isolated primary cells were plated on 12-well plates and grown for 2 days under axenic conditions in conditioned DMEM (cMEM) medium with serum [8]. BrdU (SigmaAldrich, taufkirchen, Germany) as well as recombinant human FGF1 and FGF2 were added at 1mM (BrdU) and 100 nM or 10 nM (FGF1, FGF2) final concentrations as indicated. Cultures were incubated for 48 h at 37°C under 5% CO2 for metacestode vesicles or under nitrogen atmosphere [7,8] in the case of primary cells. Samples were analysed in duplicates in three independent experiments. As controls, metacestode vesicles or primary cells were incubated in either DMEM without serum or conditioned DMEM, without addition of FGFs.
Primary cells and metacestode vesicles were then isolated for genomic DNA analysis. In detail, vesicles and primary cells were first washed with 1xPBS, pelleted, and subsequently transferred to lysis buffer (100 mM NaCl, 10 mM Tris-HCl, pH 8.0; 50 mM EDTA, pH 8.0, 0,5% SDS) supplemented with 20 μg/ml RNAse A and 0,1 mg/ml proteinase K. Samples were then incubated at 50°C for 4 h under constant shaking for complete lysis. DNA was isolated by two rounds of phenol/chlorophorm extraction (1 vol of phenol/chlorophorm/isoamylalcohol 25:24:1). DNA was then precipitated with 2 vol of 96% ethanol and 0,1 vol of LiCl (pH 4,5) after overnight incubation at -20°C and centrifugation at 20.000 rcf for 30 min at 4°C and washed with 70% ethanol. The pellet was then air dried for 15 min an resuspended in 1 x TE buffer (10 mM Tris, 1 mM EDTA, pH 8,0).
The DNA was then prepared for coating onto a 96-well plate (96 well optical bottom plates, Nunc, Langenselbold, Germany). To this end, 5 μg of DNA were combined with 1 vol of Reacti-Bind DNA Coating solution (Pierce Biotechnology, Rockford, IL, USA) and mixed for 10 min. The DNA mix was then added to the microplates in duplicates and incubated overnight at room temperature with gentle agitation. The TE/Reacti-Bind DNA coating solution mix served as a negative control. Unbound DNA was removed by washing three times with 1xPBS. After blocking with 5% nonfat dry milk in 1xPBS for 1 h at room temperature and extensive washing with 1xPBS, 100 μl of anti-BrdU-POD (Cell Proliferation ELISA, BrdU; Roche Applied Science, Mannheim, Germany) was added and incubated for 90 min at room temperature. After incubation, microplates were washed three times with 1xPBS buffer before substrate solution (Cell Proliferation ELISA, BrdU; Roche Applied Science, Mannheim, Germany) was added and the wells were incubated for 60 min. Stop-solution (25 μl of 1 M H2SO4) was added and absorbance of the samples was measured using an ELISA reader at 450 nm.
Coding sequences of FGF receptors from E. multilocularis were amplified by RT-PCR and cloned into the vectors pDrive (Qiagen) or pJET1.2 (Thermo Fisher). In the case of emfr2, the full length coding sequence was amplified using primers 5’-ATG TGT CTC CGA GCT CTC TG-3’(forward) and 5’-TTA CTC GCT CGA TCG TGG GG-3’ (reverse), whereas partial coding sequences were amplified for emfr1 (using forward primer 5’-GCA GTG GGC GTC TTC TTT CAC-3’ and reverse primer 5’-GTA AAT GTG GGC CGA CAC TCA G-3’) and for emfr3 (using forward primer 5’-TTG CCC AGT CAT CCG CTA CAA G-3’ and reverse primer 5’-GCA AGC GGT CAT GAG GCT GTA G-3’). The recombinant plasmids were used for in vitro transcription of digoxigenin-labelled RNA probes as previously described [6]. These probes were used for fluorescent WMISH of in vitro cultured E. multilocularis larvae as described in [6]. Control WMISH experiments using the corresponding sense probes were always negative.
For expression in the Xenopus system, the emfr1, emfr2, and emfr3 coding sequences without predicted signal peptide information were cloned into the pSecTag2/Hygro expression system (ThermoFisher Scientific, Germany) leading to an in frame fusion of the Igk leader sequence (provided by the vector system) and the E. multilocularis FGF receptor sequences under control of the T7 promoter. Capped messenger RNAs (cRNA) encoding EmFR1, EmFR2 and EmFR3 were then synthesized in vitro using the T7 mMessage mMachine Kit (Ambion, USA). Microinjection of EmFGFR cRNAs (60 ng in 60 μl) was performed in stage VI Xenopus laevis oocytes according to the procedure previously described [31]. Following 48h of receptor expression, human FGF1 or FGF2 (R & D systems, UK) were added to the extracellular medium at the final concentration of 10 nM. cRNA of Pleurodeles FGFR1 identified as homologous to human receptor [32] was a gift of Shi D.L. (CNRS UMR 722, Paris VI) and was used as a positive control.
In some experiments, BIBF1120 (stock solution 10mM in DMSO, Selleck Chemicals LLC) was added (0.1 to 20 μM final concentration) 1 h before the addition of 10 nM FGF1 or FGF2 on EmFR1, EmFR2, EmFR3 and Pleurodeles FGFR1 expressing oocytes. Following 15 h of FGF1 or FGF2 stimulation, oocytes were analyzed for their state of progression in the cell cycle. The detection of a white spot at the animal pole of the oocyte attested to G2/M transition and GVBD. Non-injected oocytes treated with progesterone (10 μM) were used as positive controls of GVBD. For each assay, sets of 20–30 oocytes removed from 3 different animals were used.
Dead kinase (TK-) receptors were obtained by site-directed mutagenesis of the EmFR1, EmFR2 and EmFR3 constructs. The active DFG sites present in EmFR1 (D442FG), EmFR2 (D647FG) and EmFR3 (D701FG) were replaced by an inactivating motif (DNA) as described in [31].
For western blot analysis, oocytes were lysed in buffer A (50 mM Hepes pH 7.4, 500 mM NaCl, 0.05% SDS, 5 mM MgCl2, 1 mg ml−1 bovine serum albumin, 10 μg ml−1 leupeptin, 10 μg ml−1 aprotinin, 10 μg ml−1 soybean trypsin inhibitor, 10 μg ml−1 benzamidine, 1 mM PMSF, 1 mM sodium vanadate) and centrifuged at 4°C for 15 min at 10,000 g. Membrane pellets were resuspended and incubated for 15 min at 4°C in buffer A supplemented with 1% Triton X-100 and then centrifuged under the same conditions. Supernatants were analyzed by SDS-PAGE. Proteins were transferred to a Hybond ECL membrane (Amersham Biosciences, France). Membranes were incubated with anti-myc (1/50 000, Invitrogen France) or anti-PTyr (1/8000, BD Biosciences, France) antibodies and secondary anti-mouse antibodies (1/50 000, Biorad, France). Signals were detected by the ECL advance Western blotting detection kit (Amersham Biosciences, France)
Axenically cultivated metacestode vesicles of about 0.5 cm in diameter were incubated in DMEM medium with or without 10% FCS for 4 days. Vesicles cultivated without FCS were subsequently incubated with 10 nM FGF1 (aFGF) or 10 nM FGF2 (bFGF) for 30 sec, 60 sec or 60 min. Immediately after stimulation, vesicles were harvested, cut by a scalpel to remove cyst fluid and then subjected to protein isolation as described previously [33]. Isolated protein lysates were then separated on a 12% acrylamide gel and analysed by Western blotting using a polyclonal anti-Erk1/Erk2 antibody (ThermoFisher Scientific; #61–7400), recognizing Erk-like MAP kinases in phosphorylated and non-phosphorylated form, as well as a polyclonal antibody against phospho-Erk1/Erk2 (ThermoFisher Scientific; #44-680G), specifically directed against the double-phosphorylated (activated) form of Erk1/Erk2 (Thr-185, Tyr-187). We had previously shown that these antibodies also recognize the Erk-like MAP kinase EmMPK1 of E. multilocularis in phosphorylated and non-phosphorylated form [33]. As secondary antibody, a peroxidase-conjugated anti-mouse IgG antibody (Dianova, Hamburg, Germany) was used. In inhibitor experiments, axenically cultivated metacestode vesicles were incubated with either 5 μM or 10 μM BIBF 1120 for 30 min and then processed essentially as described above.
qRT-PCR experiments have been performed as previously described by del Puerto et al. [34] with several modifications. Briefly, RNA samples were prepared with TRI-reagent and the DirectzolTM RNA Mini Prep Kit (Zymo Research, USA). Prior to RNA-isolation, in vitro cultured metacestode vesicles were deprived of stem cells by treatment with hydroxyurea (HU) and the Polo-like kinase inhibitor BI 2536 essentially as previously described by Koziol et al. [6] and Schubert et al. [35], respectively. Reverse transcription was performed with Omniscript RT Kit (Qiagen, Germany) from 500 ng of RNA using oligo dT23 as primers according to the manufacturer’s instructions. qRT-PCRs were then performed using primers Forward F_emfr1_qPCR (5’-CCG TAT GAA GGG AAA TGG TCG TGT T-3’) and Reverse R_emfr1_qPCR (5’-TGG TGA ATC GCC AAG GCT GAA A-3’) for emfr1, Forward F_emfr2_qPCR (5’-GGG AAT TTC CAA GGT CAT CAG GGA C-3’) and Reverse R_emfr2_qPCR (5’-ATC GTG GGG GCA CAA CAT AAT TGC-3’) for emfr2, as well as Forward F_emfr3_qPCR (5’-GTC TAC CTT GAG GAA ATT GCT GTG GTC-3’) and Reverse R_emfr3_qPCR (5’-CGT GAG GAA TGA CGC AGG C-3’) for emfr3. As a control gene for normalization, the constitutively expressed gene elp was used with primers as described previously [34]. qRT-PCR conditions and primer amplification efficiency control experiments were performed essentially as previously described [34]. The relative expression of emfr genes in each sample was estimated using the delt-delta-Ct method [36], normalising with elp expression levels using the StepOne software (Thermo Fisher). All experiments were carried out as technical triplicates.
Amino acid comparisons were performed using BLAST on the nr-aa and swissprot database collections available under (https://www.genome.jp/). Genomic analyses and BLAST searches against the E. multilocularis genome [14] were done using resources at (https://parasite.wormbase.org/index.html). CLUSTAL W alignments were generated using MegAlign software (DNASTAR Version 12.0.0) applying the BLOSUM62 matrix. Domain predictions were carried out using the simple modular architecture research tool (SMART) available under (http://smart.embl-heidelberg.de/) as well as PROSITE scans available under (https://prosite.expasy.org/scanprosite/). Two-tailed, unpaired student’s T-tests were performed for statistical analyses (GraphPad Prism, version 4). Error bars represent standard error of the mean. Differences were considered significant for p-values below 0.05 (indicated by *).
All experiments were carried out in accordance with European and German regulations on the protection of animals (Tierschutzgesetz). Ethical approval of the study was obtained from the local ethics committee of the government of Lower Franconia (permit no. 55.2 DMS 2532-2-354).
By cDNA library screening and mining of the available E. multilocularis genome sequence we identified a total of three Echinococcus genes encoding members of the FGFR family of RTK. A partial cDNA for a gene encoding a TK with homology to FGFRs was previously cloned for E. granulosus [28] and by RT-PCR amplification of metacestode cDNA as well as 5’-RACE, the entire cDNA of the E. multilocularis ortholog, designated emfr1 (E. multilocularis fibroblast growth factor receptor 1), was subsequently cloned. As shown in Fig 1, the encoded protein, EmFR1, contained an N-terminal export directing signal peptide, followed by one single Ig-like domain, a transmembrane region, and an intracellular TK domain (Figs 1 and S1). In the recently released E. multilocularis genome information [14], this gene was correctly predicted on the basis of genome and transcriptome data (E. multilocularis gene designation: EmuJ_000833200). In the upstream genomic regions of emfr1, no information encoding potential Ig-like domains was identified which, together with the presence of a signal peptide sequence upstream of the single Ig-like domain, indicated that EmFR1 indeed contained only one single Ig-like domain. Amino acid sequence alignments indicated that the kinase domain of EmFR1 contains all residues critical for enzymatic activity at the corresponding positions (S2 Fig) and SWISS-PROT database searches revealed highest similarity between the EmFR1 kinase domain and that of human FGFR4 (42% identical aa; 59% similar aa).
A second gene encoding a TK with significant homology to known FGFRs was identified on the available E. multilocularis genome sequence [14] under the annotation EmuJ_000196200. The amino acid sequence of the predicted protein only contained an intracellular TK domain, a transmembrane region, and one extracellular Ig-like domain, but no putative signal peptide. We therefore carried out 5’-RACE analyses on a cDNA preparation deriving from protoscolex RNA and identified the remaining 5’ portion of the cDNA, which contained one additional Ig-like domain and a predicted signal peptide. In the genome annotation, these remaining parts were wrongly annotated as a separate gene under the designation EmuJ_000196300. Hence, the second FGFR encoding gene of E. multilocularis, emfr2, encoding the protein EmFR2, actually comprises the gene models EmuJ_000196300 and EmuJ_000196200 of the genome sequence. EmFR2 thus comprises a signal peptide, two extracellular Ig-like domains, a transmembrane region, and an intracellular TKD (Figs 1 and S1). The TKD contained all residues critical for TK activity (S2 Fig) and, in SWISS-PROT BLASTP analyses, showed highest similarity to two FGF receptor kinases of the flatworm Dugesia japonica (45% identical, 65% similar residues), to the S. mansoni receptor FGFRB (55%, 68%) and to human FGFR3 (48%, 62%).
The third FGF receptor encoding gene of E. multilocularis, emfr3, was identified under the designation EmuJ_000893600 and was originally listed as an ortholog of the tyrosine protein kinase Fes:Fps [14]. However, unlike the Fes:Fps kinase which contains FCH and SH2 domains, the EmuJ_000893600 gene product, EmFR3, comprised an N-terminal signal peptide, two extracellular Ig-like domains, a transmembrane region, and an intracellular TKD (Figs 1 and S1), in which 22 of 30 highly conserved residues of TK are present at the corresponding position (S2 Fig). Furthermore in SWISS-PROT BLASTP analyses the EmFR3 TKD displayed highest similarity to several vertebrate FGF receptors and to human FGFR2 (32%, 47%). We thus concluded that EmuJ_000893600 actually encoded a third Echinococcus FGF receptor TK.
Apart from emfr1, emfr2, and emfr3, no further genes were identified in the E. multilocularis genome which displayed clear homology to known FGFR TKDs and which contained characteristic IG domains in the extracellular portions. In vertebrates, structural homology has been described between the TKDs of the receptor families of FGF receptors, the vascular endothelial growth factor (VEGF) receptors, and the platelet-derived growth factor (PDGF) receptors, which also contain varying number of Ig domains in the extracellular parts [37]. Furthermore, VEGF receptor-like molecules have also been described in invertebrates such as Hydra [37]. We therefore carried out additional BLASTP searches on the E. multilocularis genome using human VEGF- and PDGF-receptors as queries, but only obtained significant hits with the above mentioned TKDs of EmFR1, EmFR2, and EmFR3. These data indicated that members of the VEGF- and PDGF-receptor families are absent in Echinococcus, as has already been described for the closely related schistosomes [24].
Genes encoding canonical FGF ligands have so far neither been identified in genome projects of free-living flatworms [38], nor in those of trematodes [39] or cestodes [14]. In vertebrates [15] as well as several invertebrate phyla [40–42], however, canonical FGF ligands are clearly expressed. To investigate the situation more closely, we carried out BLASTP and TBLASTN analyses against the predicted proteins and E. multilocularis contig information, respectively, using FGF ligand sequences of insect, nematode, and cnidarian origin as queries. The product of only one E. multilocularis gene, EmuJ_000840500 (annotated as ‘conserved hypothetical protein’) showed certain similarity to these FGF ligands and according to SMART protein domain analyses could contain a FGF-ligand domain between amino acids 166 and 258, although this prediction was clearly below the prediction threshold and of low probability (E-value: 817). No export directing signal peptide was predicted for the EmuJ_000840500 protein, as would be typical for FGF ligands. Furthermore, although EmuJ_000840500 had clear orthologs in the cestodes Taenia solium (TsM_000953800) and Hymenolepis microstoma (HmN_000558500), none of these gene models had any prediction of an FGF-ligand domain in SMART analyses (nor predicted signal peptides). We thus considered it highly unlikely that EmuJ_000840500 encodes a so far not identified flatworm FGF-ligand.
Taken together, our analyses indicated that E. multilocularis contains genomic information for three members of the FGFR family of RTK with either one (EmFR1) or two (EmFR2, EmFR3) extracellular Ig-like domains, of which one, EmFR3, showed higher divergence within the TKD as it contained only 22 of otherwise 30 highly conserved amino acid residues of TK. On the other hand, no members of the VEGF- and PDGF-receptor families are encoded by the E. multilocularis genome, nor does it contain genes coding for canonical FGF-ligands.
To investigate gene expression patterns of the Echinococcus FGF receptors in parasite larvae, we first inspected Illumina transcriptome data for parasite primary cells after 2 and 11 days of culture (PC2d, PC11d, respectively), metacestode vesicles without or with brood capsules (MC-, MC+, respectively), as well as protoscoleces before or after activation by low pH/pepsin treatment (PS-, PS+, respectively), that had be produced during the E. multilocularis genome project [14]. According to these data, emfr1 was moderately expressed in primary cells, metacestode vesicles, and protoscoleces (S3 Fig). Likewise, emfr2 was expressed in all stages, but very lowly in primary cells, somewhat more in metacestode vesicles, and highest in protoscoleces. emfr3, on the other hand, was low to moderately expressed in primary cells, low in metacestode vesicles, and highest in protoscoleces (S3 Fig). Since primary cell preparations are characterized by a much higher content of germinative (stem) cells than metacestode vesicles [6], these expression patterns could indicate that emfr3 is stem cell-specifically expressed. We therefore carried out quantitative RT-PCR experiments on cDNA preparations from in vitro-cultivated metacestode vesicles (MV) versus metacestode vesicles after treatment with hudroxyurea (MV-HU) or the Polo-like kinase inhibitor BI 2536 (MV-BI), in which the stem cell population had been selectively depleted [6,40]. While emfr1 expression levels were equal in MV when compared to MV-HU and MV-BI, both emfr2 and emfr3 transcripts were significantly reduced in MV-HU and MV-BI (S4 Fig). This indicated that emfr1 does not have a typical stem cell-specific expression pattern whereas emfr2 and emfr3 could be expressed in the parasite’s stem cells or their immediate progeny (since HU- and BI 2536-treatment has to be carried out for at least one week [6,34]).
To clarify the situation we carried out whole-mount in situ hybridization experiments on metacestode vesicles according to recently established protocols [6,43,44]. In these experiments, proliferating parasite stem cells were labeled by incorporation of the nucleotide analog EdU, which was combined with detection of gene transcripts by using fluorescently labeled probes. According to these experiments, emfr1 was expressed at low levels throughout the germinal layer and the signal was somewhat higher during the development of protoscoleces (Fig 2), maybe due to higher cell density in the protoscolex. The intensity of the signal was heterogenous, but no clear pattern could be discerned and it was too low to clearly determine the percentage of positive cells.
emfr2, on the other hand, was specifically expressed in a population of small sized cells, which comprised 1.7% to 6.3% of all cells in the germinal layer (Fig 3A). None of these emfr2+ cells were also EdU+, indicating that they were post-mitotic [6]. During initial protoscolex development, emfr2+ cells accumulated in the peripheral-most layer of cells, as well as in the anterior-most region (which will form the rostellum), and in three longitudinal bands of cells in the interior of the protoscolex buds (Fig 3B). Again, practically none of the emfr2 labelled cells was EdU+ (less than 1% of emfr2+ cells were EdU+; Fig 3C). Later during development, some emfr2+ cells were found in the protoscolex body, but most accumulated in the developing rostellum and the suckers (Fig 3D). Importantly, emfr2 expression was restricted to the sucker cup, where cells are differentiating into em-tpm-hmw+ muscle cells [6], and not the sucker base, where EdU+ germinative cells accumulate [6] (Fig 3D). Taken together, these data indicate that emfr2 is expressed in post-mitotic cells, of which many are likely to be differentiating or differentiated muscle cells.
Finally, emfr3 was expressed in very few cells in the germinal layer (less than 1% of all cells, although the number is difficult to estimate since they were absent in most random microscopy fields) (Fig 4A). emfr3+ cells accumulated in small numbers during brood capsule and protoscolex development (Fig 4B). The emfr3+ cells had a large nucleus and nucleolus and, thus, had the typical morphology of germinative cells (S5 Fig). At the final stages of protoscolex development, few emfr3+ cells were present in the body region and some signals were also present in the rostellar region (S5 Fig). In the developing protoscolex and in the germinal layer, emfr3+ EdU+ double positive cells were found (Fig 4). These data indicated that emfr3 is expressed in a very small number of proliferating cells with the typical morphology of germinative cells.
In summary, the three E. multilocularis FGF receptor genes showed very different expression patterns in metacestode and protoscolex larval stages. While emfr1 was lowly expressed in cells that are dispersed throughout the germinal layer, emfr2 displayed an expression pattern indicative of differentiating/differentiated muscle cells. emfr3, on the other hand, appeared to be expressed in a minor sub-population of germinative cells.
Since E. multilocularis larvae do not express canonical FGF ligands (see above), but usually develop in an environment in which FGF1 and FGF2 are abundant [17], we next tested whether host-derived FGF ligands can stimulate parasite development in vitro. To this end, we employed two different cultivation systems which we had previously established [7–10]. In the axenic metacestode vesicle cultivation system, mature metacestode vesicles of the parasite are cultivated in the absence of host cells under reducing culture conditions (e.g. low oxygen)[7,9]. In the primary cell cultivation system [8,10], axenically cultivated metacestode vesicles are digested to set up cell cultures which are highly enriched in parasite germinative (stem) cells (~ 80% [6]), but which also contain certain amounts of differentiated cells such as muscle or nerve cells. In the primary cell cultivation system, mature metacestode vesicles are typically formed within 2–3 weeks, which is critically dependent on proliferation and differentiation of the germinative cell population [6] in a manner highly reminiscent of the oncosphere-metacestode transition [3].
As shown in Fig 5A, the exogenous addition of 10 nM FGF1 to mature metacestode vesicles already resulted in an elevated incorporation of BrdU, indicating enhanced proliferation of parasite stem cells, which was even more pronounced in the presence of 100 nM FGF1. In the case of FGF2, addition of 100 nM resulted in enhanced BrdU incorporation in a statistically significant manner (Fig 5A). Likewise, metacestode vesicles cultured for 4 weeks in the presence of 10 nM FGF1 or FGF2 displayed a considerably larger volume (about two-fold) when compared to non-FGF-stimulated vesicles (Fig 5B).
In the primary cell cultivation system, 100 nM concentrations of host ligands had to be added to observe statistically significant effects. Again, the incorporation of BrdU by primary cell cultures was stimulated in the presence of host-derived FGF ligands (Fig 5C), as was the formation of mature metacestode vesicles from primary cell cultures (Fig 5D).
Taken together, these results indicated that host-derived FGF ligands, and in particular FGF1, can stimulate cell proliferation and development of E. multilocularis primary cell cultures and mature metacestode vesicles.
Having shown that host-derived FGF ligands can stimulate parasite proliferation and development in vitro, we were interested whether these effects might be mediated by one or all three of the parasite’s FGF receptors which are expressed by the metacestode larval stage. To this end, we first made use of the Xenopus oocyte expression system in which the activity of heterologously expressed protein kinases can be measured by germinal vesicle breakdown (GVBD). This system has previously been used to measure the activities of the TKD of schistosome FGF receptors [24], as well as the host-EGF (epidermal growth factor) dependent activation of a schistosome member of the EGF receptor family [31]. We, thus, expressed Pleurodeles FGFR1 (as a positive control), which is highly similar to human FGFR1 [32], and all three parasite FGF receptors in Xenopus oocytes, which were then stimulated by the addition of 10 nM FGF1 or FGF2. As negative controls, we also expressed kinase-dead versions of all Echinococcus FGF receptors in Xenopus oocytes.
As can be deduced from Table 1, control (non-stimulated) oocytes were negative for GVBD, whereas progesterone-stimulated oocytes displayed 100% vesicle breakdown. Expression of FGFR1 did not yield GVBD but, after stimulation with 10 nM FGF1, 100% of oocytes underwent GVBD, indicating stimulation of the Pleurodeles receptor by FGF1 (as expected). Upon expression of any of the parasite receptors in Xenopus oocytes, no GVBD was observed when no ligand was added. The addition of 10 nM FGF1 to these receptors, however, resulted in 100% GVBD for EmFR1 and EmFR2, as well as to 80% GVBD in the case of EmFR3. In the case of human FGF2 (10 nM), 90% GVBD was observed for EmFR1 and EmFR3, and 85% for EmFR2. No GVBD was observed upon addition of 10 nM FGF1 or FGF2 when the kinase-dead versions of the parasite receptors were expressed (S6 Fig). These data clearly indicated that all three parasite receptors were responsive to host derived FGF ligands (albeit to somewhat different extent) and that the kinase activity of the parasite receptors was essential to transmit the signal. We also measured the phosphorylation state of the parasite FGF receptors upon addition of exogenous FGF1 and FGF2 (10 nM each) to Xenopus oocytes and obtained significantly induced levels of tyrosine phosphorylation in all three cases (S6 Fig).
Taken together, these data indicated that all three parasite FGF receptors were activated by binding of host-derived FGF1 and FGF2, which was followed by auto-phosphorylation of the intracellular kinase domain and downstream transmission of the signal to the Xenopus oocyte signaling systems which induce GVBD.
The small molecule compound BIBF 1120 (also known as Nintedanib or Vargatef) is a well-studied and highly selective, ATP-competitive inhibitor of mammalian members of the FGF-, VEGF-, and PDGF-receptor families with very limited affinity to other RTK [45,46]. As a possible agent to selectively inhibit FGF RTK activities in the parasite, we measured the effects of BIBF 1120 on EmFR1, EmFR2, and EmFR3 upon expression in the Xenopus oocyte system. As can be deduced from Table 1, a concentration of 1 μM of exogenously added BIBF 1120 already diminished the activity (after stimulation with 10 nM FGF1) of FGFR1 to 40%, and led to a complete block of kinase activity upon addition of 10 μM BIBF 1120. In the case of EmFR1 and EmFR2, 1 μM BIBF1120 also led to a marked decrease of receptor kinase activity, although to a somewhat lower extent than in the case of FGFR1. In the presence of 10 μM BIBF 1120, on the other hand, the activities of EmFR1 and EmFR2 were completely blocked. In the case of EmFR3, exogenous addition of 1 μM BIBF 1120 had only slight effects on GVBD, whereas 10 μM BIBF 1120 reduced the activity to less than 50% and 20 μM BIBF 1120 completely blocked TK-dependent GVBD (Table 1). Upon addition of 20 μM BIBF 1120, TK activity of all receptors tested was completely inhibited.
Taken together, these data indicated that all three Echinococcus FGF RTK were affected by BIBF 1120, although in all three cases higher concentrations of the inhibitor were necessary to completely block TK activity when compared to FGFR1. BIBF 1120 treatment had the lowest effects on the activity of EmFR3.
We next tested the effects of different concentrations of BIBF 1120 on parasite development and survival in the primary cell and metacestode vesicle culture systems. As shown in Fig 6A, the addition of 1–10 μM BIBF 1120 had clear concentration-dependent effects on mature metacestode vesicle survival which, after cultivation for 18 days, led to about 20% surviving vesicles in the presence of 1 μM BIBF 1120, 10% surviving vesicles in the presence of 5 μM BIBF 1120, and no survival when 10 μM BIBF 1120 was applied. To test whether the metacestode vesicles were indeed no longer capable of parasite tissue regeneration, we set up primary cell cultures from microscopically intact vesicles which had been treated with 5 μM BIBF 1120 for 5 days (90% intact vesicles) and let the cultures recover in medium without inhibitor. In these cultures, however, we never observed the formation of mature vesicles, indicating that either the parasite’s stem cell population, and/or other cell types necessary for parasite development in the primary cell culture system, were severely damaged after BIBF 1120 treatment.
We then also tested the effects of BIBF 1120 on fresh primary cell cultures from previously untreated metacestode vesicles. As shown in Fig 6B, a concentration of 1 μM BIBF 1120 had no effect on the formation of metacestode vesicles from these cultures, whereas vesicle formation was completely blocked in the presence of 5 μM or 10 μM BIBF 1120.
Altogether, these results clearly indicated detrimental effects of BIBF 1120 on parasite development already at concentrations as low as 5 μM. Since the parasite does not express known alternative targets for BIBF 1120, such as VEGF- or PDGFR-receptors, we deduced that these effects are due to the inhibition of one or more of the parasite’s FGF RTK.
One of the major downstream targets of FGF signaling in other organisms is the Erk-like MAPK cascade, a complete module of which we had previously identified in E. multilocularis [33,47,48]. In particular, we had previously shown that the phosphorylation status of the parasite’s Erk-like MAP kinase, EmMPK1, can be measured by using antibodies against the phosphorylated form of the human Erk-kinase [33]. To investigate whether exogenously added host FGF can affect the E. multilocularis Erk-like MAPK cascade module, we first incubated mature metacestode vesicles for 4 days in serum-free medium (which has no effect on vesicle integrity [29]) and then stimulated these vesicles for 30 sec, 60 sec, and 60 min with 10 nM FGF1 and FGF2. As shown in Fig 7A, FGF1 treatment had a clear effect of EmMPK1 phosphorylation already after 30 sec of exposure. In the case of FGF2, the effect was still measurable, but clearly less pronounced than in the case of FGF1 (Fig 7A). We then also measured the effects of BIBF 1120 treatment on EmMPK1 phosphorylation. To this end, metacestode vesicles were incubated in hepatocyte-conditioned medium and were then subjected to BIBF 1120 treatment (5 μM, 10 μM) for 30 min. As shown in Fig 7B, this led to diminished phosphorylation of EmMPK1 when 10 μM BIBF 1120 was used.
Taken together, these data indicated that, like in mammals and other invertebrates, the E. multilocularis Erk-like MAPK cascade can be activated through FGF signaling, initiated by exogenously added, host-derived FGF ligands.
An important difference between parasitic helminths and all other infectious agents (excluding viruses) is that these organisms are evolutionarily relatively closely related to their vertebrate or invertebrate hosts with which they share an ancestor that has lived around 500 to 600 million years ago. Since all metazoans share evolutionarily conserved signalling systems for cell-cell communication, this opens the possibility for host-parasite cross-communication involving evolutionarily conserved cytokines of one partner (e.g. the host) and cognate receptors of the other (e.g. the parasite), which would be of particular relevance for systemic helminths that infect host organs. Several previous studies indicated that this type of host-pathogen interaction is indeed important in helminth infections. In the E. multilocularis system, which develops in close association with host liver tissue, we previously demonstrated that host insulin stimulates parasite development and growth via acting on evolutionarily conserved insulin-receptor TK that are expressed by the parasite [11]. A similar type of cytokine receptor interaction appears to involve epidermal growth factor (EGF)-like cytokines and cognate parasite receptors, of which three (EmER, EmERb, EmERc) are expressed by E. multilocularis larvae [3,12]. As recently shown by Cheng et al. [13], host-derived EGF is able to stimulate germinative cell proliferation in in vitro cultivated parasite larvae and can stimulate at least one of the parasite EGF-receptors, EmER, when heterologously expressed in Xenopus oocytes. Although the parasite itself expresses several EGF-like molecules [14], which likely act on its EGF receptors, these data indicate that host-EGF could act as an additional stimulus, particularly in response to liver tissue damage as it is inflicted upon entry of the parasite into the host liver [4]. Apart from insulin- and EGF-signalling systems, host-parasite interactions in larval cestode infections might also involve the family of transforming growth factor (TGF)-β/bone morphogenetic protein (BMP)-family of cytokines since host TGF-β has very recently been shown to stimulate larval growth of the cestode Taenia crassiceps in vitro and was found to interact with parasite TGF-β receptors [49]. In a similar way, human BMP2 was reported by us to interact with an E. multilocularis BMP receptor [50], although no direct effects of host BMP on parasite development were yet observed. Like in cestodes, the occurrence of insulin- and EGF-receptor TK as well as TGF-β/BMP serine/threonine kinases with the capability of interacting with respective human hormones/cytokines was reported for schistosomes [31,51–55], and stimulatory effects of host EGF on the development of schistosome snail stages were observed [31]. In the present work, we extend the list of respective host-parasite cross-communication systems to the FGF-family of host cytokines and cognate FGF RTK which, to our knowledge, have never before been addressed in this context.
We herein clearly show that mammalian FGF1 and FGF2 stimulate the development of E. multilcularis metacestode vesicles from cultivated parasite primary cells, which are highly enriched in germinative (stem) cells [6]. Furthermore, we show that human FGF also stimulates proliferation and growth of mature metacestode vesicles. Both FGF1 and FGF2 are abundantly present in mammalian liver tissue where they are mostly released upon liver cell damage and during regeneration processes [17–20]. Although the precise amounts of host FGF in periparasitic lesions of E. multilocularis infected mice has not been measured to date, it is very likely that the early establishing metacestode is encountering considerable amounts of these cytokines since extensive damage to liver tissue is observed not only in chronically infected mice but also in early stages of the infection [4]. The early infectious stage is critical in the establishment of the parasite since the invading oncosophere is not yet producing the laminated layer (LL), an important structure that protects the parasite from direct actions of immune cells [56]. The laminated layer surrounds mature metacestode vesicles (established after 1–2 weeks after invasion) in the chronic phase of the disease. The stimulation of parasite development from stem cells towards mature metacestode vesicles by host FGF, as observed in our cultures, could thus help the parasite to abridge this critical phase and to successfully establish within the liver. Chronic AE is characterized by extensive liver fibrosis, particularly in the peri-parasitic area [57] and is thought to be mediated by hepatic stellate cells [5], which greatly upregulate FGF release during liver regeneration and fibrosis induction [20]. Hence, not only in the initial phase of parasite establishment, but also in the chronic phase of AE, the E. multilocularis metacestode should be in contact with elevated levels of host FGF that can continuously support parasite growth and proliferation.
Using the Xenopus expression system we clearly showed that all three identified FGF receptors of E. multilocularis are functionally active kinases that are capable of inducing GVBD when properly stimulated. We also demonstrated that all three Echincoccus kinases are responsive to human FGF1 and FGF2, albeit to somewhat different extent. While 10 nM FGF1 fully stimulated both EmFR1 and EmFR2, EmFR3 was less activated than the other receptors by both FGFs. This does not necessarily indicate, however, that human FGFs bind less well to EmFR3 than to the other two receptors. Instead, EmFR3 might be activated to a lesser extent since two tyrosine residues, Y653 and Y654, which in human FGFR1 are necessary for full activation of the receptor [58], are conserved in EmFR1 and EmFR2 but absent in EmFR3 (S2 Fig). In any case, our data clearly show that particularly human FGF1, but also human FGF2, are capable of activating the parasite receptors. Since the parasite apparently does not produce intrinsic FGF ligands, the only canonical FGFs it encounters during liver invasion are host derived. It is, thus, logical to assume that the effects of FGF1 and FGF2 on parasite growth and development are mediated by one, two or all three Echinococcus FGF receptors. In line with this assumption is the stimulation of the parasite’s Erk-like MAPK cascade module upon exogenous addition of host FGF to metacestode vesicles. In mammals, the two prominent downstream signalling pathways of FGF receptors are the Ras-Raf-MAPK cascade and the PI3K/Akt pathway, while two others are STAT signalling and the phospholipase γ (PLCγ) pathway [15]. The STAT signalling pathway is absent in Echinococcus [14] and PLCγ involves binding to human FGF receptors at tyrosine residues that are not conserved in the Echinococcus receptors. We had previously demonstrated that the PI3K/Akt pathway exists in Echinococcus [11] but we could not measure differential phosphorylation of EmAKT in response to exogenous FGF or after metacestode vesicle treatment with BIBF 1120, indicating that in contrast to human cells, Echinococcus mainly involves the MAPK cascade pathway for downstream FGF signalling. In the related planaria, an important role of the Erk-like MAPK cascade in the differentiation of stem cells into multiple cell lineages has been reported [59]. In this system, inhibition of the Erk-like MAPK cascade resulted in reduced differentiation of neoblasts which retained their ability to proliferate and produced new neoblasts [59]. In a recent study on E. multilocularis stem cell dynamics, on the other hand, inhibition of the Erk-like MAPK cascade resulted in strongly diminished cell proliferation whereas activation of Erk signalling by exogenous addition of EGF stimulated stem cell self-renewal and clonal expansion, at least under conditions of low stem cell densities [13]. Furthermore, these authors showed that stimulating EGF signalling can activate quiescent stem cells in the protoscolex to re-enter the cell cycle [13]. Unfortunately, in the Echinococcus system no markers are known so far to distinguish between self-renewing stem cells from transit amplifying cells. It will thus in future experiments be interesting to study whether host FGF stimulates Echinococcus stem cell renewal, including activation of quiescent stem cells, or the differentiation of transit amplifying stem cells towards differentiated cell lineages. Furthermore, since apparently both FGF and EGF signalling in Echinococcus converge on the Erk-like MAPK cascade pathway, it will be worthwhile to study differential or additive effects of both stimuli on E. multilocularis stem cell dynamics.
Concerning cellular expression patterns we detected significant differences between the Echincoccus FGF receptors and those of the related schistosomes and planaria. In planaria, the expression of FGF receptors is a hallmark of neoblast stem cells and also occurs in cephalic ganglions [21,23]. In schistosomes, one of the two FGF receptors, fgfrA, is a marker for a prominent subset of somatic stem cells [25–27] and, like fgfrB, is also expressed in the reproductive organs [24]. In Echinococcus, we found only one receptor, EmFR3, which is expressed in germinative cells, but only in a very tiny subpopulation. EmFR2 was also only expressed in a few cells which were, however, clearly post-mitotic and most probably represented muscle cells. Likewise, the third Echinococcus FGF receptor, EmFR1, was expressed throughout the parasite’s metacestode tissue without specific association with the germinative cells. In qRT-PCR analyses on metacestode ´vesicles which were specifically deprived of stem cells after treatment with hydroxyurea [6] or the Polo-like kinase inhibitor BI 2536 [35] we also never observed diminished levels of emfr1 expression (S4 Fig), which further supports that the gene is not exclusively expressed in germinative cells. Taken together, these data indicate that the close association of FGF receptor expression with stem cells as observed in planaria and schistosomes is at least highly modified in the Echinococcus system, in which at least EmFR1 and EmFR2 are clearly expressed in post-mitotic cells. This adds to stem cell-specific gene expression differences which we had previously observed between planaria and cestodes e.g. concerning argonaute or histone deacetylase-orthologs [6], and also indicates clear differences between stem cell–specific gene expression patterns in the parasitic flatworm lineages of trematodes and cestodes.
We observed clear inhibitory effects of the TK inhibitor BIBF 1120 on the enzymatic activity of all three Echinococcus FGF receptors upon heterologous expression in the Xenopus system and also demonstrated that this compound can profoundly affect parasite survival and development in vitro. Since in the Xenopus system concentrations of 20 μM BIBF 1120 fully inhibited all three Echinococcus FGF receptors we cannot clearly state whether the detrimental effects on parasite development were due to the specific inhibition of one of the three Echinococcus FGF receptors, or on combined activities against all three enzymes. However, based on the fact that EmFR3 is only expressed in less than 1% of the cells of the metacestode, that emfr3 expressing cells only accumulate during the formation of protoscoleces, and that EmFR3 showed lowest levels of inhibition upon expression in Xenopus cells, we do not consider this receptor as a prominent candidate for mediating the BIBF 1120 effects on primary cells and the metacestode. We also consider it unlikely that the inhibition of EmFR2 had produced these effects since the respective gene is only expressed in a small subset (2–5%) of all metacestode cells. Based on the expression of emfr2 in muscle cells or muscle cell progenitors, however, we cannot completely rule out that EmFR2 inhibition might have led to the depletion of specific parasite cells that are necessary to form a stem cell niche for the parasite’s germinative cells. At least in planaria it has already been shown that muscle cells provide important positional information on the stem cell population [60] and our recent studies on the wnt signalling pathway in Echinococcus clearly demonstrated that this is also the case for cestodes [44]. However, based on the fact that emfr1 is expressed throughout the metacestode and that in both primary cells and metacestode vesicles emfr1 is the highest expressed FGF receptor gene we propose that most of the effects of BIBF 1120 on Echinococcus development are mediated by EmFR1 inhibition. On the cellular level, BIBF 1120 could either affect emfr1-expressing germinative cells or differentiated cells of the germinal layer or both. Since metacestode vesicles treated with stem cell-specific drugs such as hydroxyurea or BI 2536 still remain intact for 2–3 weeks [6,35], but BIBF 1120-treated vesicles already loose structural integrity after 5 days (Fig 6), we consider it highly likely that EmFR1 inhibition largely damages germinal layer cells which form an important niche for the germinative cells to survive and proliferate.
BIBF 1120 has originally been described as a multi-kinase inhibitor with specificity against members of the VEGF-, FGF-, and PDGF-families, but also with activity against members of the Src-family of non-receptor TK [45]. Our genomic analyses revealed that, similar to schistosomes [24], no genes encoding VEGF- or PDGF-receptors are present on the E. multilocularis genome. So far, no Src-like kinase has been biochemically characterized in cestodes. In the related trematode Schistosoma mansoni, however, at least one clear Src family kinase, SmTK3, has been described [61] and a close ortholog, EmuJ_000598200, is encoded on the E. multilocularis genome and, at least to a certain level, expressed in primary cells and metacestode vesicles. We thus cannot completely exclude that some of the effects of BIBF 1120 on these cell systems were due to an inhibition of the Echinococcus SmTK3 ortholog. On the other hand, we never observed clear effects of Src-kinase specific inhibitors such as PP2 or herbimycin A on Echinococcus development in vitro which indicates that the observed effects of BIBF 1120 on Echinococcus larval development are rather due to the inhibition of other kinases, most probably EmFR1 and/or EmFR2. For the development of novel chemotherapeutics against AE, EmFR1 and EmFR2 would thus be highly interesting target molecules although, of course, BIBF 1120 as originally developed against human FGF receptors showed somewhat higher activities against FGFR1 in the Xenopus expression system than against the parasite FGF receptors. Nevertheless, and using the activity assays developed in this work, it should be possible to identify compounds which are related to BIBF 1120 but which show higher affinities against the parasite enzymes than against human FGF receptors.
In the present work we provide clear evidence that human FGF ligands are capable of activating evolutionarily conserved TK of the FGF receptor family that are expressed by the larval stage of E. multilocularis and that the parasite’s Erk-like MAPK cascade is stimulated upon exogenous addition of human FGFs to metacestode vesicles. We also showed that human FGF1 and FGF2 are stimulating the development of metacestode vesicles from parasite primary cell cultures and that they accelerate metacestode vesicle proliferation and growth in vitro. Since FGF1 and FGF2 are expressed in considerable amounts within the host liver, the primary target organ for the establishment of the E. multilocularis metacestode, and since FGF ligands are also constantly produced during liver regeneration and fibrosis, which are consequences of parasite growth within the intermediate host, we consider the observed in vitro effects on parasite FGF signalling and metacestode development also of high relevance in vivo. Liver-specific activities of host FGF could thus support the development of metacestode vesicles from stem cells that are delivered to the liver by the oncosphere larva, and could constantly stimulate asexual proliferation of the metacestode during an infection. We finally showed that at least one compound that inhibits the activities of mammalian FGF receptors, BIBF 1120, also inhibits the parasite orthologs, leads to metacestode inactivation, and prevents parasite development of stem cell-containing primary cell cultures. This opens new ways for the development of anti-Echinococcus drugs using the parasite FGF receptors as target molecules.
|
10.1371/journal.pgen.1007226 | RNA helicase, DDX27 regulates skeletal muscle growth and regeneration by modulation of translational processes | Gene expression in a tissue-specific context depends on the combined efforts of epigenetic, transcriptional and post-transcriptional processes that lead to the production of specific proteins that are important determinants of cellular identity. Ribosomes are a central component of the protein biosynthesis machinery in cells; however, their regulatory roles in the translational control of gene expression in skeletal muscle remain to be defined. In a genetic screen to identify critical regulators of myogenesis, we identified a DEAD-Box RNA helicase, DDX27, that is required for skeletal muscle growth and regeneration. We demonstrate that DDX27 regulates ribosomal RNA (rRNA) maturation, and thereby the ribosome biogenesis and the translation of specific transcripts during myogenesis. These findings provide insight into the translational regulation of gene expression in myogenesis and suggest novel functions for ribosomes in regulating gene expression in skeletal muscles.
| Inherited skeletal muscle diseases are the most common form of genetic disorders with primary abnormalities in the structure and function of skeletal muscle resulting in the impaired locomotion in affected patients. A major hindrance to the development of effective therapies is a lack of understanding of biological processes that promote skeletal muscle growth. By performing a forward genetic screen in zebrafish we have identified mutation in a RNA helicase that leads to perturbations of ribosomal biogenesis pathway and impairs skeletal muscle growth and regeneration. Therefore, our studies have identified novel ribosome-based disease processes that may be therapeutic modulated to restore muscle function in skeletal muscle diseases.
| Ribosome biogenesis is fundamental to all life forms and is the primary determinant of translational capacity of the cell. Ribosome biogenesis is a complex process that involves transcription, modification and processing of ribosomal RNA, production of ribosomal proteins and auxiliary factors and coordinated assembly of ribonucleoprotein complexes to produce functional ribosomes [1]. While previously considered a “house keeping” constitutive process, recent studies have shown that ribosome biogenesis is regulated differently between cells and can be modulated in a cell type-specific manner [2]. These differences are required to generate ribosomes of different heterogeneities and functionalities that contribute to the translational control of gene regulation by selecting mRNA subsets to be translated under specific growth conditions potentially by identifying specific recognition elements in the mRNA [3–5]. Protein synthesis is the end stage of the gene regulation hierarchy and despite the identification of translational regulators of specific genes, a systematic identification of translational regulatory processes critical for tissue specific control of gene expression is still lacking. Moreover, upstream regulatory factors/processes that regulate ribosome heterogeneity in a cellular and organ-specific context in vertebrates still remains to be identified.
Skeletal muscle is a contractile, post-mitotic tissue that accounts for 30–50% of body mass. Skeletal muscle growth and repair is a highly coordinated process that is dependent on the proliferative expansion and differentiation of muscle stem cells that originate from the muscle precursor cells at the end of embryogenesis [6–8]. Activated muscle stem cells through symmetric/ asymmetric divisions give rise to daughter cells that maintain the muscle stem cell population (satellite cells) and committed myogenic progenitor cells (MPC) that fuse to form the differentiated skeletal muscle [6, 9, 10]. Defects in processes regulating muscle stem cells, myoblast fusion and differentiation therefore, constitute pathological pathways affecting the muscle growth and regeneration [11–17]. Although epigenetic and transcriptional controls of myogenesis have been studied extensively, the importance of translational regulation of these processes in skeletal muscle function is less defined. Transcriptional and translational analysis of myoblasts has shown that differential mRNA translation controls protein expression of specific subset of genes during myogenesis. Recent studies have revealed ribosomal changes during skeletal muscle growth and atrophy. An increase in ribosome biogenesis is often observed during skeletal muscle hypertrophy [18, 19]. Ribosomal perturbations on the other hand are associated with skeletal muscle growth and diseases [20–23]. Deletion of ribosomal protein genes S6k1 and Rps6 in mice results in smaller myofibers and reduced muscle function. In several muscular dystrophy models and atrophied muscles, a reduction in ribosome number and/or activity is observed [22, 24, 25]. In spite of these dynamic changes observed in ribosomes during different growth conditions, our understanding of the mechanism(s) that regulate ribosomal biogenesis in skeletal muscle and eventually control the translational landscape during muscle growth is still poor.
In a forward genetic screen to identify critical regulators of myogenesis in vivo, we recently identified a RNA helicase gene, ddx27, that controls skeletal muscle growth and regeneration in zebrafish [26, 27]. We further show that DDX27 is required for rRNA maturation and ribosome biogenesis in skeletal muscles. Strikingly, DDX27-deficient myoblasts exhibit impaired translation of mRNA transcripts that have been shown to control proliferation as well as differentiation of muscle progenitors during myogenesis. These studies highlight the specific role of upstream ribosome biogenesis processes in regulating tissue specific gene expression during myogenesis.
Zebrafish and human exhibit similar skeletal muscle structure and the molecular regulatory hierarchy of myogenesis is conserved between zebrafish and mammals [9, 28–30]. Therefore, to identify regulators of skeletal muscle growth and disease in vivo, we performed an ENU mutagenesis screen in zebrafish [26]. Analysis of skeletal muscle of 4–5 dpf (days post fertilization) larvae of one mutant identified in the screen, Osoi (Japanese for “slow”), displayed highly reduced birefringence in polarized light microscopy in comparison to the control, indicative of skeletal muscle defects (Fig 1A). Positional mapping and sequencing of this mutation identified a 20 bp deletion in exon 18 of the ddx27 gene (DEAD-box containing RNA helicase 27) (Figs 1B, S1A and S1B). qRT-PCR analysis showed significantly lower levels of ddx27 transcripts reflecting probable nonsense-mediated decay (S1C Fig). Overexpression of human DDX27 resulted in a rescue of skeletal muscle defects in mutant embryos, demonstrating functional evolutionary conservation and confirming that mutation in ddx27 are causal for the osoi phenotype (Fig 1C). Evaluation of a large number of mutant embryos showed that homozygotes die by 6–7 dpf. qRT-PCR analysis of mouse Ddx27 mRNA demonstrated Ddx27 expression in several tissue types (S1D Fig). Whole-mount immunofluorescence of wild-type zebrafish (4 dpf) detected Ddx27 expression in Pax7 labeled muscle progenitor cells in skeletal muscle (Fig 1D).
To identify DDX27 expression domains in mammalian skeletal muscle, immunofluorescence was performed on single myofibers isolated from extensor digitorum longus (EDL) muscles of wild-type mice (Fig 1E). Immunofluorescence on freshly isolated EDL myofibers (Day 0) showed that Ddx27 is expressed in Pax7 positive satellite cells. EDL myofibers were cultured for 3 days, which results in activation (day1), proliferation and differentiation of satellite cells (day 2). Ddx27 expression remained high in activated satellite cells (day1). Ddx27 expression was observed in proliferating Pax7 positive nuclei as well as MyoD expressing nuclei. Subsequently, after culture of myofibers for 1–2 days, satellite cells typically undergo cell division. After 3 days in culture, proliferating satellite cells as well as proliferating myoblasts (MyoD positive) exhibited Ddx27 expression. To determine Ddx27 expression during muscle differentiation, western blot was performed on proliferating and differentiating C2C12 myoblasts (Fig 1F). Ddx27 expression was detected in proliferating myoblasts and declined as cells are committed towards differentiation (100% confluence, day 0). The downregulation of ddx27 corresponded to an increase in MyoG and myosin heavy chain expression (MF20). These data show that Ddx27 expression is high in proliferating satellite cells and myoblasts and is reduced in myoblasts as they are committed towards the terminal differentiation into myotubes.
Finally, analysis of subnuclear localization by immunofluorescence revealed that DDX27 is co-localized with UBF (fibrillar component) and Fibrillarin (dense fibrillary component) in nucleoli of human myoblasts (Fig 1G). No co-localization of ddx27 was detected with B23, labeling the granular component of nucleoli. The nucleolus is the primary site of ribosome biogenesis, therefore, this sub-nucleolar organization of DDX27 suggests a functional requirement of DDX27 in rDNA transcription and/or pre-rRNA processing steps in skeletal muscles.
To investigate the role of RNA helicase Ddx27 in skeletal muscle homeostasis, we analyzed the structure and function of mutant zebrafish embryos (2 dpf) and larvae (4–5 dpf. Analysis of skeletal muscle histology at the end of primary myogenesis in 2 dpf embryos showed no visible differences between control and mutant skeletal muscles (S2A Fig). Quantification of fiber cross section area (CSA) also showed no significant differences in control and mutant muscles at 2 dpf (S2B Fig). The nuclear content of the myofibers was analyzed as fiber CSA per nuclei that was similar in control and mutant myofibers at 2 dpf, however showed a significant decrease in mutants at 4 dpf (S2C Fig). This suggests that skeletal muscle development is normal during embryogenesis (0–2 dpf) in ddx27 fish. During development, Ddx27 expression is observed during embryogenesis and persists during larval stages in zebrafish (www.zfin.com). However, normal skeletal muscle growth during embryogenesis could be due to a functional redundancy with other family members, many of which have overlapping expression patterns with Ddx27. Homozygous ddx27 mutant fish (5 dpf) showed reduced fiber diameter (Fig 2A and 2B arrow, inset) and central nuclei as observed in several muscle diseases (Fig 2A and 2B, arrowhead) [31]. Electron microscopy of longitudinal sections of ddx27 mutant myofibers also showed disorganized myofibrillar structures in ddx27 mutant fish (Fig 2C–2E, arrow). The mutant myofibers displayed smaller Z-lines and disorganized actin-myosin assemblies in comparison to controls suggesting differentiation defects of these muscles. Immunofluorescence of cultured myofibers from control and mutant zebrafish further showed a reduction and disorganization of skeletal muscle differentiation markers at 4 dpf, (Actn2/3 and Ryr1) validating our earlier observation that absence of Ddx27 results in a defective differentiation of skeletal muscles (S2B Fig). The centralized mutant nuclei were round in shape with highly enlarged nucleoli in comparison to controls (Fig 2E, arrowhead). Cross-sections of muscles showed a reduction in myofiber diameter (52 ± 14%) in ddx27 mutant fish in comparison to controls (Fig 2F–2H). Skeletal muscles of mutant fish also displayed whorled membrane structures associated with nuclei (Fig 2D, arrow). These whorled membrane structures are often observed in skeletal muscle of congenital myopathy patients, however their origin and biological significance remains unknown. Ddx27 belongs to a family of highly conserved RNA helicases in vertebrates. Therefore, to evaluate if the function of Ddx27 in skeletal muscle is conserved in vertebrates, myogenic differentiation of control and Ddx27 knockout C2C12 myoblasts was analyzed. C2C12 myoblasts exhibited reduced proliferation and impaired differentiation, that failed to form mature myofibers (S2F and S2G Fig). These results show that Ddx27 deficiency affects skeletal muscle growth during zebrafish larval stages leading in to skeletal muscle hypotrophy and DDX27 functions in growth and differentiation of skeletal muscle are conserved among vertebrates.
To identify molecular events leading to growth defects in ddx27 mutant zebrafish, qRT-PCR was performed during embryonic (2 dpf) and larval (4 dpf) stages. qRT-PCR revealed no significant expression changes in muscle progenitor cell markers, pax3 and pax7 during the embryonic stage (Fig 2I). In zebrafish, the PAX7 orthologue is duplicated in two copies; pax7a and pax7b. pax7a expressing cells act to initiate myofiber formation post-injury, whereas pax7a/b expressing cells are required for myofiber growth [32]. Significant down regulation of both pax7a and b was observed during the larval stage in ddx27 mutant zebrafish. Concurrently, with a downregulation of pax3 and pax7a/b, we also observed a high expression of myoD, myf5 and desmin in mutant muscles suggesting a premature activation of the myogenic program in Ddx27 deficiency (Fig 2I). However, this precocious activation of the myogenic program appears to be unable to progress to a proper differentiated state, as documented by low levels of late differentiation markers myog and mylz leading to disorganized sarcomeres.
ddx27 zebrafish larvae exhibit significantly slower swimming than wild-type controls at 5 dpf (91 ± 29 mm/10 min for ddx27 vs.16823 ± 214 mm/10 min for controls). To quantify the functional deficits in skeletal muscles of homozygous ddx27 mutant larvae (5 dpf), peak twitch and tetanic forces of individual zebrafish skeletal muscle preparations were measured following electrical stimulation. Substantially depressed twitch and tetanic forces as well as the slower rise and fall of tension were observed in ddx27 mutants (Fig 3A). Absolute tetanic force was significantly less for ddx27 larvae compared to controls (mean ± SD: 0.28 ± 0.18 vs. 1.43 ± 0.27 mN; p < 0.0001) (Fig 3B). These force deficits in ddx27 persisted in the force measurements normalized for cross-sectional areas (Fig 3C: 11 ± 8 vs. 48 ± 6 kPa; p < 0.0001), suggesting that the depressed tetanic force of the ddx27 mutant preparations is primarily due to intrinsic skeletal muscle deficits. Mutant ddx27 preparations also showed significant reductions in mean absolute twitch force (Fig 3D: 0.16 ± 0.12 vs. 1.11 ± 0.23 mN; p < 0.0001) and twitch force normalized to the CSA of individual larvae (Fig 3E: 7 ± 5 vs. 37 ± 4 kPa; p < 0.0001). Interestingly, twitch force was depressed to a relatively greater extent than tetanic force as revealed by the significantly reduced twitch to tetanic force ratios of the ddx27 preparations (Fig 3F: 0.57 ± 0.15 vs. 0.78 ± 0.06 kPa; p = 0.0006). In addition to these differences in force magnitude, the kinetics of force development and relaxation were also affected in mutant skeletal muscles. The ddx27 larvae displayed a significantly slower rise in the maximal rate of twitch tension development, +dP/dt (Fig 3G: 1.06 ± 0.71 vs. 8.00 ± 1.30 kPa/ms; p < 0.0001), and a significantly slower maximal rate of twitch tension relaxation, -dP/dt (Fig 3H: -0.47 ± 0.24 vs. -2.55 ± 0.35 kPa/ms; p < 0.0001). This disproportionate reduction in twitch vs. tetanic force and the slowing of twitch kinetics indicate a reduction in the quality of the contractile performance of the ddx27 larvae that are consistent with reduced motility of mutant muscles.
To understand if the decreased Pax7 expression observed in ddx27 mutant is due to reduced Pax7 expression in MPCs or is caused by reduced number of MPCs, we performed whole mount immunofluorescence with Pax7 antibody during embryonic and larval stages. In zebrafish, Pax7 labels proliferative cells in dermomyotome, quiescent muscle stem cells and myoblasts that are required for muscle growth and regeneration [8, 33]. Quantification of Pax7 positive cells by whole mount immunofluorescence showed no differences in control and mutant muscles during embryogenesis (2 dpf). However, a significant reduction in Pax7 expressing nuclei (40 ± 8.2%) was observed in larval mutant skeletal muscle compared with control fish (4 dpf) (Fig 4A). This implies that reduced pax7 expression observed in ddx27 mutant fish is a consequence of reduced number of Pax7 cells in these fish. This decrease in Pax7 cells in mutants could either be due to a defect in proliferation of MPC population or an increased apoptosis in Ddx27 deficiency. To study the proliferative potential of Ddx27-deficient MPC, we performed 5-ethynyl-2’-deoxyuridine (EdU) incorporation assay and Pax7 immunofluorescence in control and mutant fish. The proportion of nuclei that were double labeled with Pax7+Edu+ in total Pax7 labeled population, decreased significantly to ~14.1% in ddx27 skeletal muscle in comparison to the control (~22.0%). This indicates that the proliferation of Pax7 cells is significantly reduced in ddx27 mutant fish (Fig 4B). We also investigated if Ddx27 deficiency leads to increased apoptosis of Pax7 MPC population. Whole mount TUNEL staining and western blot analysis with caspase 3 antibody at 3 and 4 dpf in control and mutant fish did not reveal any significant increase in apoptotic changes in mutant muscles compared to controls at these stages suggesting that decreased proliferation associated with premature differentiation rather than enhanced cell death underlie reduced MPC population observed in mutant muscles (S3 Fig).
A stem cell niche equivalent to mammalian satellite cell system exists in zebrafish that involves migration and asymmetric division and/or proliferation of Pax7 expressing muscle progenitor cells to repair muscle upon injury [8, 34, 35]. Moreover, processes and timing of skeletal muscle repair in larval zebrafish are highly similar to adult mammalian skeletal muscle. As Pax7 MPC population also contributes to skeletal muscle repair in zebrafish, we next investigated if Ddx27 deficiency affects muscle repair in mutant fish. To induce muscle injury, cardiotoxin was injected into the epiaxial myotome of somites of larval fish at 3 dpf and skeletal muscles were analyzed at 5 dpf as described previously [9]. Whole mount immunofluorescence showed that cardiotoxin administration resulted in accumulation of a pool of Pax7 expressing cells at the site of injury in control fish which exhibited efficient myofiber repair post injury as seen by newly formed lighter stained actin myofibers (Fig 4C). In contrast, injured Ddx27-deficient zebrafish muscles exhibited numerous degenerating myofibers and no accumulation of Pax7 expressing cells or muscle repair was observed post injury suggesting an impaired regeneration (Fig 4C). These findings suggest that Ddx27 plays a pivotal role in skeletal muscle regeneration. The major steps in skeletal muscle repair in zebrafish involve proliferation of MPC population, migration to the injury site and fusion with the damaged myofibers or form new myofibers [9]. ddx27 mutant fish exhibit reduced MPCs proliferation as well as differentiation defects to form mature myofibers suggesting that either or both of these processes could underlie the repair defects observed in Ddx27 deficiency in skeletal muscle.
Ddx27 is localized in the nucleolus that is primarily the site of ribosome biogenesis. Therefore, we evaluated nucleolar structure and functions to understand the impact of Ddx27 on these processes in skeletal muscle. Immunofluorescence with different nucleolar markers showed changes in localization of the fibrillary component marker Ubf (labeling rRNA transcription sites) from small punctate foci to larger condensed areas, suggesting a perturbation in active transcription sites in the nucleolus. Similarly, fibrillarin-enriched dense fibrillary component areas of early rRNA-processing regions were also disrupted and merged, forming larger, more condensed structures. Lastly, granular component of the nucleolus which is the site of late rRNA processing was also perturbed in the ddx27 mutant. B23, a nucleolar granular component marker exhibited altered localization to nucleoplasm in mutant nucleoli in comparison to the nucleolar restricted expression controls suggesting a structural disruption of fibrillary and dense fibrillary component potentially disrupts the organization of granular compartment in mutant nuclei (Fig 5A). These results suggest that Ddx27 deficiency disrupts sites of rRNA synthesis, processing and early ribosomal assembly.
To investigate the effect of Ddx27 on rRNA synthesis, we performed in situ rRNA transcription analysis in zebrafish skeletal muscles (5 dpf) by 5-EU labeling of newly synthesized rRNA. Following 5-EU labeling, immunofluorescence was performed to visualize MPCs (Pax7 positive cells) or myonuclei (Actn2/3 labeling). Analysis of 5-EU signal in MPC population revealed a significant reduction in rRNA transcripts in Pax7 positive cells in ddx27 fish (30 ± 8%). Interestingly, a significant decrease in rRNA synthesis was also observed in myonuclei in control myofibers. Further, examination of mutant myonuclei revealed a reduction in rRNA synthesis in comparison to control myonuclei. Together, these results suggest that Ddx27 deficiency impairs rRNA synthesis directly in MPCs. As these MPCs differentiation to form myofibers, rRNA synthesis defect persists in myonuclei contributing to impaired ribosome biogenesis and reduced muscle function in a non-autonomous manner in myofibers (Fig 5B).
Next, to evaluate the role of Ddx27 on pre-rRNA processing, the pre-rRNA maturation pattern was evaluated in the skeletal muscle of ddx27 mutant zebrafish. Skeletal muscles (containing MPCs and myonuclei) were dissected from control and mutant larval fish (5 dpf) and total RNA was isolated. Northern blotting was performed with different probes that were specific to various precursor rRNAs representing different pre-rRNA processing steps (Fig 5C and 5D). ddx27 mutant skeletal muscle displayed significant accumulation of long pre-rRNAs (precursors A and B), corresponding to 47S and 43S pre-rRNAs in the human rRNA processing pathway. Precursors C (corresponding to human 32S pre-rRNA) also accumulated, while precursors to the 18S rRNA (precursors D, E) decreased. This accumulation of early precursors and of precursors C suggests an early defect in the rRNA maturation process and indicative of delayed cleavages in the 5’ETS and 3’ETS. These results are in accordance with recent data showing that depletion of DDX27 in human cells leads to the release of an extended form of the 47S primary transcript [36]. In addition, our data reveal an accumulation of 41S pre-rRNAs and a concomitant decrease of 30S pre-rRNAs, which are indicative of an impaired cleavage at site 2 (Fig 5C).
To study the impact of rRNA maturation defects on ribosomes, we performed ribosomal profiling in control and mutant zebrafish muscle that revealed a significant decrease in free 60S large ribosomal subunits. A reduction of mature 80S monosomes as well as polysomes was also observed in mutant muscles (Fig 5E). Together these studies show that ddx27 expression is required for the formation of mature rRNA species and thereby for biogenesis of functional polysomes in skeletal muscle.
We next sought to investigate if the ribosomal deficits due to Ddx27 deficiency affect translation of global processes or a specific mRNA repertoire in skeletal muscle. To identify mRNA repertoire exhibiting perturbed translation in Ddx27 deficiency, polysome profiling and subsequent RNA sequencing was performed in Ddx27 knockout C2C12 myoblasts that exhibit proliferation and differentiation defects similar to zebrafish. Also, Ddx27 is highly expressed in proliferating C2C12 myoblasts suggesting that mRNA species identified from polysomal profiling will potentially be a direct consequence of Ddx27 deficiency. Polysomes (actively translating ribosomes) were purified from control and knockout Ddx27 C2C12 myoblasts grown in the proliferation media. RNA-sequencing of total and polysome bound mRNA transcripts revealed that 124 transcripts that showed increased and 300 which showed decreased association with polysomes were common in both control and mutant. 1057 mRNA transcripts were exclusively enriched in control polysomes whereas Ddx27 deficient polysomes showed an enrichment of 286 mRNA transcripts (Fig 6). Data analysis (S4 and S5 Figs) showed that control polysomes associated transcripts exhibited an enrichment in mRNAs encoding ribosomal, RNA polymerase, RNA degradation and splicing pathways suggesting a high requirement of these RNA metabolic processes during muscle cell growth. On the other hand, in DDX27 deficiency an enrichment of apoptotic and inflammatory pathway genes was observed suggesting that absence of DDX27 activates the atrophic processes in muscle. Interestingly, mutant fish did not exhibit any visible increase in apoptosis. The enrichment of apoptotic transcripts in mutants could potentially be due to an initiation of end stage changes as mutants die by 5 dpf. In addition, mRNAs required for protein biosynthesis (amino acid biosynthesis, aminoacyl-tRNA biosynthesis) showed significantly lower enrichment in mutant polysomes. These result suggest that DDX27 is crucial for the translation of mRNAs that are necessary for generating building blocks for active biosynthesis of proteins and suppression of transcripts associated with atrophic processes during muscle growth.
Interestingly, a number of signal transduction pathways associated with skeletal muscle growth and diseases are also perturbed in DDX27 deficiency. Mutant polysomes exhibited a reduced association with Fgfr1 transcripts. FGF-signaling pathway is crucial for skeletal muscle growth and muscle specific ablation of Fgfr1 impairs proliferation of muscle satellite cells [37]. Additionally, an increase in mRNAs encoding members of MAP Kinase pathway that is associated with precocious differentiation was observed in ddx27 mutant zebrafish muscles [38]. Lastly, we identified several novel mRNAs that were highly enriched in control polysomes but decreased in DDX27-deficienct polysomes (e.g. Ctsw, Hddc2, Tagln, Wfdc1 Sprr2h, and Vmn2r78) (Fig 6C). Many of these genes are involved in the maintenance of proliferative state of human ES and ipS cell however, their role/s in myogenesis are not known [39]. To confirmed the validity of our polysomal profiling data and expression of these transcripts in skeletal muscle we analyzed the expression of HDDC2 in control and Ddx27 mutant myoblasts. Immunofluorescence and Western blotting revealed a significant downregulation of HDDC2 protein in Ddx27 mutant myoblasts suggesting that HDDC2 may be contributing to DDX27 mediated muscle stem cell defects in skeletal muscle (Fig 6D). A reverse analysis also identified the downregulation of 4 novel transcripts in control myoblasts that were enriched in Ddx27 polysomes, including A330074K22Rik, Fhdc1, Gpr153, and Ston2 and future studies will be able to identify their functional roles in skeletal muscle. In sum, polysome profiling revealed that DDX27 is required for the translation of mRNAs regulating RNA metabolism and signaling pathway that are crucial for muscle satellite cell proliferation and differentiation. In addition, we identified novel genes that are required for cellular proliferation and future studies on in vivo function of these genes in skeletal muscle may help to understand novel processes regulating muscle growth.
In this study, we hypothesized that in vivo identification of novel factors regulating myogenesis should help to elucidate the molecular mechanisms that regulate myogenic processes. Indeed, in this work we found that an RNA helicase, Ddx27, regulates skeletal muscle growth and regeneration by controlling ribosome biogenesis and translational processes. Our investigation into mechanisms of Ddx27 function in skeletal muscle growth leads to three major conclusions. First, Ddx27 is required for the skeletal muscle growth by regulating MPCs proliferation and differentiation in to mature myofibers in zebrafish. Second, Ddx27 is critical for skeletal muscle repair of injury in larval fish. Third, DDX27 mediated processes contributing to skeletal muscle growth and repair are primarily regulated at the level of protein translation by ribosomes.
DDX27 belongs to the DEAD-box family of RNA helicases which represent a large protein family with 43 members that catalyze the ATP-dependent unwinding of double stranded RNA and variously functions in remodeling structures of RNA or RNA/protein complexes, dissociating RNA/protein complexes, or RNA annealing [40]. Studies in yeast and cellular models have shown that several DDX family members regulate different steps of rRNA processing; however, in vivo functions of these RNA helicases in vertebrates are still mostly unknown. While this plethora of RNA helicases implies the potential for functional redundancy, it also raises the attractive possibility that RNA helicases might perform a generic, unifying function in neuromuscular system by regulating RNA metabolism. DDX5/p68 RNA helicase promotes the assembly of proteins required for transcription initiation complex and chromatin remodeling during skeletal muscle differentiation [41]. Overexpression of DDX5 also restores the skeletal muscle function in a mouse model of myotonic dystrophy [42]. RNA helicases (DDX1 and DDX3) play significant roles in muscle diseases by interacting with muscle specific transcription factors or with disease causing genes [43, 44]. We further demonstrate that DDX27 is a nucleolar protein that is highly expressed in muscle progenitor cells. Although nucleolar proteins often have ubiquitous localization, high expression of DDX27 in satellite cells and myoblasts suggests a specialized role for this protein in controlling MPC-regulated processes in skeletal muscle.
Ddx27 deficiency impairs skeletal muscle growth and regeneration in larval Ddx27-deficient zebrafish. The expression of pax7 RNA as well as number of Pax7 expressing MPC were found to be significantly reduced in skeletal muscles of ddx27 fish. This reduced number of Pax7 positive MPC population could be either due to a direct role of Ddx27 in regulating proliferation, premature differentiation or increased cell death. Our studies demonstrate that Ddx27 deficiency leads to a decrease in proliferation of the Pax7-positive muscle progenitor cell population in mutant skeletal muscles by accumulation of cells at the G1 stage of cell cycle (S2H Fig). Considering Ddx27 is expressed in murine Pax7 positive cells and proliferating myoblasts (Fig 1), this reduced proliferation of MPC in zebrafish is likely due to a cell autonomous defect. Moreover, no significant increase in apoptosis was observed in ddx27 mutant muscles suggesting that the reduction in MPC is not a consequence of an increased cell death in mutants. During myoblast differentiation to myotubes, downregulation of Ddx27 expression is associated with an upregulation of MyoD, MyoG and subsequently, myosin heavy chain expression. These data suggest that Ddx27 expression may be required for maintaining the proliferative state of muscle progenitor cells in an autonomous manner whereas preventing their differentiation to mature muscles under normal conditions in a non-autonomous manner. This is supported by the observation that an absence of Ddx27 results in upregulation of myoD and myf5 in ddx27 mutants implicating a precocious differentiation of mutant myoblasts. Previous studies have shown that overexpression of MyoD is sufficient to induce myogenic differentiation suggesting a similar mechanism may be contributing to premature differentiation of Ddx27 deficient MPCs [45, 46].
A lack of muscle repair in mutant larval fish further demonstrated that Ddx27 is also crucial for skeletal muscle regeneration. Skeletal muscle regeneration is a complex process involving migration and proliferation of muscle progenitor cells to the injury site followed by differentiation to form new myofibers or repair the existing damaged myofibers. In vivo imaging studies have shown that the process of muscle repair in larval zebrafish is highly similar to that in adult mammalian muscle. In zebrafish, Pax7-marked muscle progenitor cells migrate to the injury site, divide and undergo terminal differentiation and regenerate muscle fibers [9, 32]. Our studies demonstrate an absence of Pax7 expressing MPCs at the injury site in ddx27 skeletal muscle. This could be either due to a defect in the proliferation of Pax7 muscle progenitor cells as observed by reduced Edu labeling of Pax7 positive nuclei in ddx27 mutant fish or due to the inability of muscle progenitor cells to migrate to the site of injury and/or fuse with damage fibers. Considering we observed reduced MPCs proliferation as well as a lack of proper myofiber differentiation in Ddx27 deficiency, either or a combination of both of these processes could be contributing to skeletal muscle repair defects in mutant fish. Follow up studies with transgenic reporter lines will help us investigate the contribution of different cell lineages and processes contributing to impaired skeletal muscle regeneration in Ddx27-deficient skeletal muscle.
The nucleolus is a prominent organelle, central to gene expression, in which ribosome synthesis is initiated. Alterations in nucleolar structure are indicators of changes in cellular growth and proliferation, cell cycle regulation and senescence; therefore, identification of mechanisms that guide the nucleolar structure-function relationship in vivo is needed. Our studies show that Ddx27 is highly expressed in MPCs that regulate skeletal muscle growth and repair and Ddx27 deficiency results in rRNA synthesis defects in the MPC population. A number of recent studies have shown that nucleoli are actively involved in stem cell maintenance [47, 48]. Actively proliferating stem cells and progenitor cells possess large nucleoli that change to smaller foci during differentiation. These changes in nucleolar structure are associated with reduced rDNA transcription and ribosome biogenesis during differentiation. Depletion of a number of nucleolar proteins results in reduced cell proliferation, abnormal cell cycle and enhanced differentiation, demonstrating that proper nucleolar function is required for self-renewal of ES cells [49, 50]. During differentiation of mesenchymal progenitors in to osteoblasts, myoblasts or adipocytes, phenotypic regulatory factors critical for each lineage (e.g. Runx2, MyoD, Mgn or C/EBP) suppress rDNA transcription suggesting regulation of rRNA synthesis by cell fate-determining factors is a broadly used mechanism for coordinating cell growth with lineage progression [51]. Notably, we identified that lack of Ddx27 also results in rRNA synthesis defects in myonuclei. As mutant MPCs form mature muscles, rRNA defects are also persisted in differentiated muscle fibers. There are 43 DDX proteins in vertebrates, several of which are also expressed in myonuclei (DDX1, 3 and 5). Therefore, rRNA synthesis defects in ddx27 mutant myonuclei imply highly specialized roles of Ddx27 in skeletal muscle that are not compensated by other DDX family members. Collectively, these studies suggest that Ddx27 deficiency results in impaired proliferation and differentiation of MPCs leading to defective myofibers. This lack of Ddx27 in MPCs leads to a direct ribosomal defects in MPCs and have an indirect effect on myofibers derived from these MPC population.
Molecularly, we established that Ddx27 regulates rRNA maturation and ribosome biogenesis. DDX27 is highly conserved in evolution, and mutation of the yeast ddx27 ortholog, drs1p, results in 25S rRNA maturation and 60S ribosome subunit biogenesis defects [52]. In zebrafish and mammalian cells, we identified an accumulation of primary transcripts and long-pre-rRNAs reflecting early rRNA maturation defects resulting in a decrease in functional ribosomes. Previous studies have shown that ribosome numbers are increased during skeletal muscle hypertrophy and reduced in atrophied muscles or skeletal muscle diseases [20–22]. However, the regulators of these processes remain to be identified. Our studies demonstrate that DDX27 is required for normal muscle growth by controlling rRNA maturation and ribosomal biogenesis. A number of recent studies have shown that ribosomes of different heterogeneities and functionalities exist and contribute to the translational control of gene regulation by selecting mRNA subsets to be translated under specific growth conditions [5, 53]. For example, analysis of ribosome populations in mouse embryonic stem cells revealed that RPL10 enriched ribosomes preferentially regulate mRNAs controlling cellular growth whereas RPL10 depleted ribosomes exhibited increased binding to mRNA pools regulating stress responses and cell death [54]. Therefore, reduced proliferation of MPCs and terminal differentiation into mature myofibers in zebrafish could potentially be due defects in ribosome biogenesis affecting either the global translation rates or translation of specific subsets of mRNA required for proliferation and differentiation of MPC population. Polysomal profiling identified reduced association of mRNAs in FGF and MAPK signaling pathways that are known to regulate different stages of skeletal muscle growth in both cell autonomous and non-autonomous manners. These signaling pathways are associated with activation and myogenic commitment of muscle stem cells and thus could be contributing to the skeletal muscle defects observed in Ddx27 deficiency in zebrafish [37, 38]. Polysomal profiling also revealed an enrichment of transcripts regulating RNA metabolism pathways in control polysomes that was lacking in the mutant muscles. Defects in RNA metabolism underlie disease pathophysiology in a number of neuromuscular diseases [55]. Therefore, identification of critical regulators of RNA based processes that contribute towards myogenesis is essential in order to understand the molecular basis of pathological changes in disease conditions. Interestingly, many pathways altered in mutant myoblast are also crucial for myofiber hypertrophy. Identification of these different processes in control and mutant muscles signifies an additional regulatory layer of translational regulation controlled by DDX27 that fine tunes the crucial processes associated with skeletal muscle growth. Interestingly, most of the translational changes observed in both control and mutant skeletal muscle are associated with genes that were previously not known to play significant roles in myogenesis. In particular, dysregulation of genes associated with cellular pluripotency and membrane remodeling suggests novel roles for these factors in skeletal muscle biology. Future work targeting these genes will help to illuminate additional processes that are crucial for myogenesis. This work provides new insight into nucleolar function in skeletal muscle growth, and opens a new avenue to explore the specific roles of nucleolar proteins and ribosome biogenesis in normal and disease muscle.
Fish were bred and maintained using standard methods as described (56). All procedures were approved by the Boston Children’s Hospital Animal Care and Use Committee. Wild-type embryos were obtained from Oregon AB line and were staged by hours (h) or days (d) post fertilization at 28.5°C. Zebrafish embryonic (0–2 days post fertilization) and larval stages (3–5 dpf) have been defined as described previously [56].
Cardiotoxin-induced muscle regeneration studies in zebrafish were performed following previously published protocols [57]. Briefly, control and ddx27 mutant larvae (3 dpf) were anesthetized and immobilized by embedding into 3% low melting agarose. Cardiotoxin (10mM, 1 μl) was injected into dorsal somite muscles, and fish (8–10, in 4 independent experiments) were analyzed at 5 dpf by immunofluorescence analysis. Muscle degeneration and regeneration in mice was performed as described previously [58].
Functional experiments were performed as previously described [59]. Briefly, fish were studied in a bicarbonate buffer of the following composition: (in mM) 117.2 NaCl, 4.7 KCl, 1.2 MgCl2, 1.2 KH2PO4, 2.5 CaCl2, 25.2 NaHCO3, 11.1 glucose (Dou et al., 2008). 4–5 dpf larvae were anesthetized in fish buffer containing 0.02% tricaine and decapitated. The head tissue was used for genotyping. The larval body was transferred to a small chamber containing fish buffer equilibrated with 95% O2, 5% CO2 and maintained at 25°C. The larval body was attached to an isometric force transducer (Aurora Scientific, Aurora, Ontario, CAN, model 403A) and position motor (Aurora Scientific model 308B) using a 10–0 monofilament tie placed at the gastrointestinal opening and another tie attached several myotomes proximal from the tip of the tail. Twitches (200 μs pulse duration) and tetani (300 Hz) were elicited using supramaximal current delivered to platinum electrodes flanking the preparation. All data were collected at the optimal preparation length (Lo) for tetanic force. At the conclusion of the experiment, images of the preparations width and depth at Lo were obtained by carefully rotating the preparation about the gastrointestinal opening attachment point. Each image was analyzed by ImageJ using an internal length calibration. Preparation cross-sectional area (CSA) was calculated from width and depth measurements assuming the preparations cross-section was elliptical. Forces were calculated as active force, i.e. peak force minus the unstimulated baseline force, and are presented in absolute terms (valid because all larvae were attached at a consistent anatomical landmark, the gastrointestinal opening) as well as normalized to preparation CSA. The maximal rate of twitch tension development was determined as the maximal derivative of the force by time response between the onset of contraction and peak force. Likewise, the maximal rate of twitch tension relaxation was calculated as the first derivative of the force by time relaxation response, ranging from peak force until force had declined to approximately baseline. Statistical differences in the ddx27 (n = 12) and control (n = 10) group means were evaluated by a two-sample t-test.
Whole mount immunofluorescence in zebrafish was performed as described previously [60]. For immunofluorescence studies in zebrafish myofiber culture, previously published protocol was followed and 30–40 myofibers were analyzed in each condition [61]. Paraformaldehyde (4%) or methanol (100%) were used as fixative for different antibodies. EDL culture and immunofluorescence was performed using previously published studies [62]. Primary antibodies used in this study were: α-actinin (1:100; Sigma, A7811), RYR1 (1:100; Sigma, R-129), Fibrillarin (1:50; Santa Cruz, sc-25397), B23 (1:50; Santa Cruz, sc-5564), UBF (Sigma, 1:50, HPA006385), DDX27 (1:50, Santa Cruz, sc-81074), Pax7 (1:20; Developmental Studies Hybridoma Bank), Mef2 (1:20, Santa Cruz, sc-17785). For phalloidin staining, paraformaldehyde fixed embryos/larvae were incubated with phalloidin (1:40, Thermo Fisher Scientific, A12379) (and with primary antibody; for double immunofluorescence), overnight at 4°C followed by incubation with secondary antibody (if primary antibody was used). Nuclear staining was done using DAPI (Biolegend, 422801). Secondary antibodies (Thermo Fisher Scientific) were used between 1:100–1:250 dilutions.
Mouse C2c12 cells were cultured in growth medium consisting of DMEM supplemented with 20% fetal bovine serum. To induce differentiation, growth medium was replaced with the differentiation medium consisting of DMEM supplemented with 2% horse serum. Cells were maintained in the differentiation medium for 5 days. Western blot analysis was performed with Ddx27(1:100, Santa Cruz, sc-81074), MyoD (1:200, Santa Cruz, sc-760), MyoG (1:200, DHSB, F5D) and MF20 (1:50, DHSB).
rRNA transcription was detected using the Click-iT RNA Alexa Fluor imaging kit (C10329, Invitrogen) as described previously [63]. Briefly, to detect the synthesis of rRNAs in the nucleoplasm, control or ddx27 knockout zebrafish (10–12) or cultured myofibers (5 dpf) were treated with 1μg/mL actinomycin for 20 minutes and then incubated for 2 hours with 1mM 5-ethyl uridine (5-EU) in the presence or absence of actinomycin, fixed with paraformaldehyde (4%) and incubated with Click-iT reaction cocktail for 1 hour. This was followed by immunofluorescence with Pax7 in zebrafish (1:20, DHSB) to detect MPCs and α-actinin (1:250, Sigma: A7811) in cultured fish myofibers to label myonuclei. DNA was counterstained with DAPI. For the analysis of rRNA levels, the average 5-EU signal intensity in the nucleoplasm was measured from 20 nuclei per fish (10–12 fish in each group) and three independent repeats using the ImageJ program. The level of rRNA was determined as the 5-EU dye signal level in actinomycin-treated samples subtracted from that of untreated samples, thus representing mature rRNAs in control and mutant fish or myofibers as described previously [63].
In order to analyze the precursors to the 28S and 18S rRNAs, total RNA samples were isolated from skeletal muscle of control and mutant larvae (n = 100) and separated on a 1.2% agarose gel containing 1.2% formaldehyde and 1× Tri/Tri buffer (30 mM triethanolamine, 30 mM tricine, pH 7.9). RNAs were transferred to Hybond N+ nylon membrane (GE Healthcare, Orsay, France) and cross-linked under UV light. Membrane hybridization with radiolabeled oligonucleotide probes was performed as described (Preti, 2013). Signals were acquired with a Typhoon Trio PhosphorImager and quantified using the MultiGauge software. The human probes were: 5’-TTTACTTCCTCTAGATAGTCAAGTTCGACC-3’ (18S), 5’-CCTCGCCCTCCGGGCTCCGTTAATGATC-3’ (5’ITS1), a mixture of 5'-CTGCGAGGGAACCCCCAGCCGCGCA-3' (ITS2-1) and 5'-GCGCGACGGCGGACGACACCGCGGCGTC-3' (ITS2-2), 5’-CCCGTTCCCTTGGCTGTGGTTTCGCTAGATA-3’ (28S), 5’-GCACGCGCGCGCGGACAAACCCTTG-3’ (28S-3’ETS). The zebrafish probes were: 5’-GAGGGAGGCGCGTCGACCTTCGCTGGGC-3’ (3’ETS), 5’- CAGCTTTGCAACCATACTCCCCCCGGAAC-3’ (18S), 5’-GAGATCCCCTCTCGAACCCGTAATGAT-3’ (ITS1), 5’-GAGCGCTGGCCTCGGAGATCGCTGGGTCGC-3’ (ITS2), 5’-CCTCTCGTACTGAGCAGGATTACTATTGC-3’ (28S). For ribosome profile analysis zebrafish (100 larvae) were treated with cycloheximide (Sigma, 100ug/ML) for 10 minutes at room temperature. Subsequently, skeletal muscle was dissected and flash frozen in liquid nitrogen in lysis buffer (10mM Tris-Cl, pH 7.4, 5mM MgCl2, 100mM KCl, 1% TritonX-100). To purify ribosomal fractions, cell lysate (0.75 ml) was layered on 10–50% sucrose gradient and centrifuged at 36,000 rpm using SW-41 Ti rotor, 2 hours at 4°C. The fractions were collected and were analyzed at 254nm.
Polysome ribosome fractions were prepared from control and Ddx27 C2C12 myoblasts. Equal number of control and Ddx27 knockout C2C12 myoblasts were plated in 10 cm dishes in the proliferation media and collected after 24 hours for polysomal analysis. Polysomes were isolated as described above for zebrafish muscle and fractions were pooled together, treated with proteinase K and RNA was isolated using acid phenol-chloroform extraction and ethanol precipitation. As Ddx27 myoblasts exhibit a reduced proliferation, equal amounts of proteins were used to fractionate polysomes from control and KO myoblasts. Deep sequencing libraries were generated and sequenced as described [64]. Ribosomal profiling was repeated in triplicate and principal component analysis was performed to identify variation between samples. The Spearman R2 was > 0.9 for all replicates (except one sample) that was subsequently removed from follow up analysis. Splice-Aware alignment program STAR was used to map sequencing reads to Mus musculus (mm10 build). R package “edgeR” was employed to identify differential gene expression calls from these sequence reads. Gene expression was considered to be up-regulated if log2FC> +1 or downregulated if the log2FC< -1 (FC = fold change of average CPM) with respect to the condition being compared at a false discovery rate <0.05. Functional ontological classification of different gene lists was performed by DAVID.
To analyze the proliferating cells in C2C12 cell cultures and zebrafish embryos, Edu labeling was performed. For zebrafish, 20–25 embryos (48 hrs) were placed in 1mL of 500μM EdU/10% DMSO in E3 media and incubated on ice for 2 h. Embryos were transferred back to the incubator at 28.5°C and samples were collected at desired time intervals and fixed in methanol (-20°C, 20 mins). Fish were permeabilized with 1% tritonX-100/PBS for 1 hr and Click-iT reaction cocktail (Thermo Fisher Scientific) was added and incubated in dark for 1 hr at room temperature. Immunofluorescence was performed with Pax7 antibody (1:10 DHSB) and analyzed (10–12 larvae) by confocal microscopy. Proliferation was analyzed by calculating the proportion of Pax+Edu+ double positive cells in total Pax7 labeled population in control and mutant fish. For analyzing proliferating C2C12 cells, equal number of control and mutant cells (to 40–50% confluency) were plated for 12–14 hours. 2X EdU solution was added to the cells and incubated for 2 hrs at 37°C. After incubation, cells were permeabilized with 0.5% triton-100 and labeled with Click-iT reaction cocktail as described for zebrafish. Immunofluorescence was performed with Pax7 antibody (DHSB) and nuclei were stained with DAPI. Apoptosis was performed on 3-4dpf zebrafish larvae by in situ cell detection kit (Roche) or western blot analysis using caspase 3 antibody (ab13847, Abcam).
Zebrafish swimming behavior was quantified by an infra-red tracking activity monitoring system (DanioVision, Noldus, Leesburg, VA, USA). Control or ddx27 mutant larvae were placed individually into each well of a 24 well plate in dark for 10 minutes. The activity of these larvae was recorded during a follow-up light exposure of 20 minutes. Four independent blind trials were performed and mean velocity, total distance and cumulative duration of movement were recorded. Reported values reflect an average of 30–35 control or mutant larval fish.
Quantification of myofiber size, area, western blots and northern blots were performed using the ImageJ program. Data were statistically analyzed by parametric Student t-test (two tailed) and were considered significant when P<0.05. All data analyses were performed using XLSTAT software.
Additional detailed experimental details are provided in the supplemental material.
|
10.1371/journal.ppat.1006809 | STAT3 expression by myeloid cells is detrimental for the T- cell-mediated control of infection with Mycobacterium tuberculosis | STAT3 is a master regulator of the immune responses. Here we show that M. tuberculosis-infected stat3fl/fl lysm cre mice, defective for STAT3 in myeloid cells, contained lower bacterial load in lungs and spleens, reduced granuloma extension but higher levels of pulmonary neutrophils. STAT3-deficient macrophages showed no improved control of intracellular mycobacterial growth. Instead, protection associated to elevated ability of stat3fl/fl lysm cre antigen-presenting cells (APCs) to release IL-6 and IL-23 and to stimulate IL-17 secretion by mycobacteria-specific T cells. The increased IL-17 secretion accounted for the improved control of infection since neutralization of IL-17 receptor A in stat3fl/fl lysm cre mice hampered bacterial control. APCs lacking SOCS3, which inhibits STAT3 activation via several cytokine receptors, were poor inducers of priming and of the IL-17 production by mycobacteria-specific T cells. In agreement, socs3fl/fl cd11c cre mice deficient of SOCS3 in DCs showed increased susceptibility to M. tuberculosis infection. While STAT3 in APCs hampered IL-17 responses, STAT3 in mycobacteria-specific T cells was critical for IL-17 secretion, while SOCS3 in T cells impeded IL-17 secretion. Altogether, STAT3 signalling in myeloid cells is deleterious in the control of infection with M. tuberculosis.
| We studied the role of STAT3, a major regulator of immunity, in the control of the infection with M. tuberculosis. Stat3fl/fl lysm cre mice, deficient in STAT3 in myeloid cells, showed lower bacterial levels in organs and reduced extension of lung granulomas after infection with M. tuberculosis. STAT3-deficient APCs stimulated with innate receptor agonists released high levels of IL-6 and IL-23, and promoted IL-17 production by mycobacteria-specific CD4+ T cells. Increased IL-17 levels accounted for the increased resistance to M. tuberculosis of the STAT3-deficient mice. Instead, stat3fl/fl lysm cre macrophages showed no improved control of mycobacterial growth. SOCS3 is a negative regulator of STAT3 activation. The ability of socs3fl/fl lysm cre APCs to secrete IL-6 and IL-23 and to stimulate IL-17 production by antigen-specific T cells was reduced. In agreement, mice lacking SOCS3 in DCs showed increased susceptibility to M. tuberculosis infection. Different to a role in myeloid cells, STAT3 expression by mycobacteria-specific T cells was required for IL-17 secretion while SOCS3 in T cells hampered IL-17 production. Therefore, despite STAT3 expression in T cells is required for Th17 differentiation, STAT3 in APCs hampers secretion of Th17 promoting cytokines and the secretion of IL-17 by mycobacteria-specific T cells and reduces the resistance of mice to infection with M. tuberculosis.
| Tuberculosis (TB), caused by infection with Mycobacterium tuberculosis, remains a leading public health problem worldwide. TB causes 9 million new cases and 1.5 million deaths each year [1]. However, host factors determining the outcome of infection are not completely understood.
A host counters mycobacterial infections primarily via TH1 immune responses that involve cellular effector mechanisms such as macrophage activation [2, 3]. IL-12 secreted by APCs is crucial for the differentiation and maintenance of IFN-γ-secreting antigen-specific TH1 cells [4, 5] and both IL-12 and IFN-γ mediate mycobacterial control in mice and man [6–9].
The transcription factor STAT3 is a central regulator of immunity, mediating inflammatory but also anti-inflammatory responses [10, 11]. The functions of STAT3 are pleiotropic. STAT3 is activated by phosphorylation in response to cytokines of the IFN-receptor family (such as IL-10) and by some members of the IL-2 receptor family that uses the common γ chain receptor or after stimulation of several receptor tyrosine kinases (EGF, CSF-1, and PDGF). Additionally, STAT3 is activated by the common signal transducing molecule gp130 utilized by the IL-6 receptor family [12], and in response to G-CSF and leptin as their receptors are homologous to gp130.
STAT3 is critical for defense against bacterial and fungal infections. Low IL-17 secreting T-cell proportions were reported in patients bearing STAT3 mutations. These patients were prone to chronic candidiasis and staphylococcal diseases [13]. Chronic candidiasis is frequently present in patients deficient in IL-17 receptor A [14]. STAT3 deficient patients may also display impaired immunity against chronic viral infections [15, 16].
In mice, knockout of STAT3 is lethal, so in vivo studies on STAT3 functions have been performed using conditional knock out mice. Stat3fl/fl lysm cre mice, deficient in STAT3 in myeloid cells, display enhanced susceptibility to endotoxic shock and develop chronic enterocolitis with age [17]. The phenotype of these animals is similar to IL-10-/- mice, including increased expression of TNF and other inflammatory cytokines, since IL-10 suppresses induction of TNF-α via STAT3 [18]. Recently, STAT3 was shown to favour intracellular growth of M. tuberculosis in human macrophages [19]. Moreover, the presence of pSTAT3+ monocytes associated with the progression of the disease in M. tuberculosis infected non-human primates [20].
We have previously analysed the role of SOCS3, a molecule that inhibits STAT3 activation after triggering of several cytokine and growth factor receptors, and found that mice devoid in SOCS3 in myeloid or lymphoid cells showed increased susceptibility to M. tuberculosis [21].
The role of STAT3 during infection with M. tuberculosis in vivo is still unknown. We here examine the role of STAT3 in M. tuberculosis by using stat3fl/fl lysm cre mice. We highlight that STAT3 expression in APCs inhibits TH17 associated responses resulting in an increased susceptibility to infection with M. tuberculosis.
First, the role of STAT3 expression in myeloid cells in the control of infection with M. tuberculosis was examined using stat3fl/fl lysm cre mice. Lungs and spleens from stat3fl/fl lysm cre mice after 4 and 8 weeks of infection showed significantly lower M. tuberculosis burden than stat3fl/fl littermates (Fig 1A and 1B). A smaller area of the lung parenchyma of stat3fl/fl lysm cre mice was occupied by granulomas when compared to control lungs 4 but not at 8 weeks after infection (Fig 1C).
The density of granulocytes in the lung parenchyma was determined either by H&E staining of sections (Fig 1D) or by labelling of CD11b+CD11c-Ly6CdimLy6G+ neutrophils (Fig 1E and 1F) in lung suspensions from stat3fl/fl lysm cre and stat3fl/fl mice 3 and 4 weeks after M. tuberculosis infection. The neutrophil density (Fig 1D–1F) and the levels of neutrophil myeloperoxidase (mpo) and elastase (elane) mRNAs (Fig 1G and 1H) were also higher in lungs from stat3fl/fl lysm cre at mice 4 and 8 but not at 14 weeks after infection with M. tuberculosis- compared to controls (Fig 1D–1H and S1A–S1C Fig).
Activated STAT3 hampers TNF expression [22, 23]. Lungs from stat3fl/fl lysm cre mice infected with M. tuberculosis (Fig 2A) as well as BMM infected with M. tuberculosis or BCG (Fig 2B–2D) showed higher TNF protein and mRNA levels than controls. Since TNF has been shown to mediate M. tuberculosis control in macrophages [24], we speculated that stat3fl/fl lysm cre macrophages could display a better control of intracellular mycobacteria.
A higher frequency of stat3fl/fl lysm cre BMMs were infected when measured 4 h after co-incubation with the M. tuberculosis, although the number of bacteria per infected cell was similar (Fig 2E–2G). Three days after infection M. tuberculosis infected mutant and WT BMM showed similar numbers of infected cells and bacteria per total or infected cell (Fig 2E–2G). Stat3fl/fl lysm cre BMM showed no improved control of M. tuberculosis or BCG growth in vitro 6 days after infection as measured by CFU in lysates (Fig 2H and 2I). Altogether, we observed no indication of an improved bacterial growth control or reduced bacterial uptake in stat3fl/fl lysm cre BMM.
Several cytokines controlled by STAT3 are potent regulators of the expression of co-stimulatory molecules on APCs. Therefore, we studied whether STAT3 played a role in regulation of T cell priming. As expected, the density of co-stimulatory molecules CD80 and CD86 as well as MHC-II levels increased on BMDCs after mycobacterial stimulation. The expression of MHCII, CD80 and CD86 in either control or stat3fl/fl lysm cre BMDCs before or after mycobacterial stimulation was similar (Fig 3A–3C). To investigate if the expression of STAT3 by myeloid cells could modulate T cell priming during infection with M. tuberculosis T cell receptor transgenic T cells specific for the immunodominant mycobacterial Ag85B240-254 peptide (p25-tg) cells were inoculated i.v. into stat3fl/fl lysm cre or stat3fl/fl mice 17 days after infection with M. tuberculosis (Fig 3D). Three days after transfer, the expression of CD69 (which increases after T cell receptor triggering) and CD62L (the L-selectin ligand that hampers T cells to traffic to the periphery) was measured on p25-tg T cells and host T cells from the mediastinal lymph nodes (MLN). The expression of CD69 was increased and the expression of CD62L was reduced in p25-tg T cells from MLN of infected mice when compared to uninfected control mice. Similar levels of the CD69 and CD62L were expressed by p25-tg or host T cells from stat3/fl lysm cre or stat3fl/fl infected mice (Fig 3E and 3F).
SOCS3 inhibits STAT3 activation by different cytokine receptors, e.g. those of the IL-6 receptor family [10]. In accordance with the results obtained with socs3fl/fl lysm cre mice [21], socs3fl/fl cd11c cre mice showed higher bacterial levels in lungs and spleens after infection with M. tuberculosis than control animals (Fig 3G and 3H). The cd11c cre transgene has been shown to be expressed in the majority of conventional and plasmacytoid DCs [25]
Mycobacteria-stimulated BMDC from socs3fl/fl lysm cre showed lower levels of MHCII, CD80 and CD86 than control cells (Fig 3I–3L). When T cell priming in vivo was studied by transfering p25-tg naïve T cells into M. tuberculosis infected animals, the density of CD62L on donor p25-tg T cells and on host MLN T cells was lower in WT mice as compared to socs3fl/fl lysm cre mice recipients (Fig 3M and 3O). The p25-tg T cells in the MLN of socs3fl/fl lysm cre mice also showed lower surface density of CD69 as compared to those from WT-infected mice (Fig 3N and 3P). However, the expression of CD69 in host MLN T cells from M. tuberculosis-infected WT and socs3fl/fl lysm cre mice was similar (Fig 3P).
Thus, deficiency of SOCS3 but not STAT3 in APCs regulates T cell priming during M. tuberculosis infection in vivo.
IFN-γ is required for protection against M. tuberculosis [2, 3]. STAT3 has been shown to inhibit the transcription of IL-12, a potent inducer of IFN-γ secretion by T cells [23]. We then analysed if the increased resistance to M. tuberculosis of stat3fl/fl lysm cre mice is associated with higher IFN-γ secretion by T cells. Higher levels of IL-12p40 (the α-chain of IL-12 and IL-23) in supernatants and il12p40 mRNA in cell lysates of BCG-infected stat3fl/fl lysm cre BMM or BMDC compared to controls were measured (Fig 4A and 4B). The il12p35 mRNA coding for the β-chain of the IL-12 heterodimer was also expressed in higher amounts by M. tuberculosis- or BCG-stimulated stat3fl/fl lysm cre BMDC compared to controls (Fig 4C).
Thus, whether STAT3-deficient APCs are better stimulators of IFN-γ secretion by mycobacteria-specific T cells than WT APCs was investigated. To test this hypothesis, p25-tg T cells were incubated with either BCG- or M. tuberculosis-infected stat3fl/fl lysm cre or stat3fl/fl BMDCs and the IFN-γ titers in the supernatants measured. IFN-γ levels in supernatants were elevated compared to those incubated with WT APCs (Fig 4D and 4E). Supernatants from cultures of p25-tg T cells incubated with heat-killed BCG-stimulated stat3fl/fl BMDC or BMM also contained higher levels of IFN-γ than those using control APCs (Fig 4F), indicating that infection is not required for such responses. In line with this, IFN-γ levels were higher in supernatants from p25-tg T cells co-incubated with stat3fl/fl lysm cre BMDC or BMM stimulated with oligopeptide p25 (amino acids 240–254) from Ag85b, a major immunodominant H2b epitope [26] recognized by the p25-tg T cells, in presence of LPS (Fig 4G).
Confirming previous results [21, 27], socs3fl/fl lysm cre BMDC showed diminished IL-12 secretion after mycobacterial stimulation (Fig 4H). Furthermore, IFN-γ secretion by p25-tg T-cells incubated with mycobacteria-infected socs3fl/fl lysm cre or socs3 fl/fl cd11 cre BMDCs was reduced (Fig 4I and 4J).
In line with these results, mycobacteria-infected il12p40-/- BMDCs showed reduced ability to trigger IFN-γ secretion by p25-tg T cells than controls (Fig 4K). Moreover, the addition of rec IL-12 restored the capacity of socs3fl/fl lysm cre BMDC to stimulate IFN-γ secretion by p25-tg T cells (Fig 4L).
Cells derived from gp130F/F mice, harbouring a mutation that ablates SOCS3 binding to the gp130, show exaggerated gp130-mediated STAT3 responses [28]. Mycobacteria-infected gp130F/F BMDCs also showed a reduced ability to stimulate IFN-γ secretion p25-tg T cells compared to WT cells (Fig 4M).
The frequency of IFN-γ-secreting mycobacteria-specific T cells in lung cell suspensions from stat3fl/fl lysm cre and stat3fl/fl mice 4 and 8 weeks after infection with M. tuberculosis was similar. The frequencies of lymphoid cell populations (S2A Fig) and of PPD- and PMA/ ionomycin -stimulated IFN-γ secreting CD4+ or CD8+ cells (Fig 4N and 4O and S2B and S2C Fig) from lungs stat3fl/fl lysm cre and stat3fl/fl mice 4 and 8 weeks were also similar. In addition, levels of ifng, and the IFN-γ-regulated inos and cxcl9 transcripts were increased in lungs after infection as compared to uninfected controls, but the titers of these transcripts in lungs from stat3fl/fl lysm cre and stat3fl/fl-infected mice were comparable (Fig 4P–4R).
The neutrophil density and the levels of neutrophil transcripts were enhanced in the lungs of M. tuberculosis-infected stat3fl/fl lysm cre as compared to control mice (Fig 1D–1H). IL-17 has been shown to stimulate granulopoiesis via G-CSF production and to induce the expression of CXC chemokines involved in granulocyte recruitment [29]. Thus, we investigated whether the increased neutrophil levels in lungs from M. tuberculosis-infected stat3fl/fl lysm cre was associated with augmented TH17 responses. The frequency of IL-17-secreting, PPD-stimulated CD4+ T cells from lungs from stat3fl/fl lysm cre mice 4 and 8 weeks after infection with M. tuberculosis were elevated when compared to stat3fl/fl controls (Fig 5A–5C). Instead, the frequency of γδ T cells in lungs and the frequency of IL-17 secreting pulmonary γδ+ T cells from WT or stat3fl/fl lysm cre infected mice was similar (S3A–S3C Fig).
In addition, levels of il17a and il22, transcripts that code for TH17 cytokines were higher in lungs from stat3fl/fl lysm cre mice than in those from littermate controls when measured at 4 and 8 weeks after infection (Fig 5D and 5E). Higher levels of il17 mRNA were also observed in stat3fl/fllysm cre mice 14 weeks after infection with M. tuberculosis, while the increase of Il22 mRNA did not reach statistical significance (S4A and S4B Fig). CXCL5 is a neutrophil chemotactic protein stimulated by IL-17 [30]. The level of cxcl5 mRNA was increased in lungs from M. tuberculosis-infected stat3fl/fl lysm cre mice (Fig 5F).
Substantial in vivo data support the notion that IL-6 and IL-23 are required at different stages of TH17-cell differentiation [31, 32]. Levels of il6 and il23 mRNA were elevated in the lungs of M. tuberculosis-infected stat3fl/fl lysm cre mice when compared to levels in lungs from WT mice (Fig 5G and 5H).
TH17 cells show a high degree of developmental flexibility, and when exposed to IL-12 or IL-23, they can rapidly acquire effector functions that are normally associated with TH1 responses such as IFN-γ production [33]. These IFN-γ and IL-17 secreting cells were shown to be pathogenic in murine models of autoimmune diseases, and were also associated with murine colitis and human IBDs [34]. The majority of PPD or PMA/ ionomycin-stimulated IL-17 secreting CD4+ T cells in lungs from M. tuberculosis-infected stat3fl/fl lysm cre or stat3fl/fl mice were not IFN-γ co-producers (Fig 5I–5K). We next studied whether IL-17 played a role in the increased control of infection of stat3fl/fl lysm cre mice. For these experiments, mice were treated with neutralizing IL-17RA mab (M751) before and during infection with M. tuberculosis. Similar bacterial levels were found in lungs and spleens from stat3fl/fl lysm cre and stat3fl/fl animals treated with anti-IL17RA mAb. As expected, stat3fl/fl lysm cre mice from untreated mice showed reduced bacterial numbers in lungs and spleens than those from stat3fl/fl controls (Fig 5L and 5M).
The levels of mpo mRNA was measured in lungs from anti-IL-17RA to control for the IL17RA neutralization. As expected, levels of mpo mRNA were increased in lungs from M. tuberculosis infected stat3fl/fl lysm cre mice as compared to WT. In contrast, levels of mpo mRNA in lungs from infected or anti-IL17RA treated stat3fl/fl lysm cre and stat3fl/fl mice was similar. Lower titters of mpo mRNA were found in stat3fl/fl lysm cre infected mice treated with anti-IL-17RA compared to untreated infected controls, while mpo mRNA levels in anti-IL17RA treated or untreated infeceted stat3fl/fl mice were similar (Fig 5N).
Hence, increased M. tuberculosis control during infection of stat3fl/fl lysm cre mice is dependent on IL-17, but IL-17 neutralization did not increase the susceptibility to M. tuberculosis of mice with normal STAT3 function.
The role by which STAT3 in activated APCs regulates IL-17 secretion by specific T cells was then studied. The mRNA levels of Il6 and il23p19 were both increased in BMM and BMDCs co-incubated with mycobacteria in vitro (Fig 6A and 6B and S5A Fig). An increased accumulation of il6 and il23p19 mRNA was also observed after stimulation with the TLR agonists LPS, CpG or Pam3K of stat3fl/fl lysm cre as compared to stat3fl/fl BMM at 6 and 24 h after stimulation (Fig 6C and 6D and S5B Fig). This indicates that STAT3-mediated inhibition of the expression of IL-6 and IL-23 is not restricted to mycobacterial infection or stimulation with mycobacterial molecules. Supernatants from cultures of mycobacteria-infected stat3fl/f lysm cre BMM or BMDC co-incubated with naïve p25-tg T cells contained higher titers of IL-17 than those using stat3fl/fl controls (Fig 6E–6G). IL-17 levels were also higher in supernatants from p25-tg T cells co-incubated with stat3fl/fl lysm cre BMDC stimulated with either live or heat-killed BCG, M. tuberculosis or peptide 25 from Ag85b in presence of LPS (Fig 6F–6H).
Whether SOCS3 in mycobacteria-infected APCs also regulated IL-17 secretion by antigen-specific T cells was then measured. We found that mycobacteria-infected socs3fl/fl lysm cre BMM contained lower levels of il6 and il23p19 mRNA than their WT counterparts (Fig 6I and 6J). Moreover, p25-tg T cells incubated with socs3fl/fl lysm cre BMM stimulated either with BCG, M. tuberculosis or the cognate p25 peptide secreted lower levels of IL-17 than those stimulated by WT BMDCs (Fig 6K–6M). Similarly, supernatants from co-cultures of p25-tg T cells with mycobacteria-infected gp130F/F BMDC contained higher levels of IL-17 than those using wild type BMDCs (Fig 6N).
We then asked whether STAT3 deficiency regulated the levels of socs3 mRNA transcripts in mycobacteria-infected macrophages. M. tuberculosis-infected stat3fl/fl lysm cre and stat3fl/fl BMMs showed similar levels of socs3 mRNA (Fig 6O).
Different to the inhibitory role of STAT3 in myeloid cells we here showed, STAT3 expression in T cells has been indicated to be required for TH17 cell differentiation in vitro and in vivo [31]. SOCS3, via hyperactivation of STAT3, has been shown to increase IL-17 secretion [35]. In order to compare the role of STAT3 and SOCS3 in T cells and APCs in the regulation of cytokine secretion by T cells, stat3fl/fl lck cre p25-tg and socs3 fl/f lck cre p25-tg mice were generated. The culture supernatants of stat3fl/fl lck cre p25-tg T cells stimulated with BCG-infected or Ag85b peptide-pulsed BMDCs showed low or undetectable levels of IL-17 as compared to controls (lck cre p25-tg T cells) (Fig 7A). Instead IL-17 levels in supernatants from socs3fl/fl lck cre p25-tg T cells co-incubated with BCG or peptide loaded BMDCs were higher than controls (Fig 7C).
The titters of IFN-γ in the supernatants of stat3 fl/fl p25-tg T cells incubated with BCG- (but not with Ag85b peptide-) stimulated BMDCs were higher than those incubated with control T cells (Fig 7B). IFN-γ levels in supernatants from mycobacteria or peptide pulsed BMDCs incubated with socs3fl/fl lck cre p25-tg were instead lower than those using control p25-tg T cells (Fig 7D).
Thus, STAT3 in APCs and T cells, has a dissimilar ability to regulate IL-17 secretion by ag-specific T cells, while SOCS3 in APCs and T cells promote T cell mediated IFN-γ secretion.
We report here that stat3fl/f lysm cre mice show reduced M. tuberculosis load in lungs and spleens, indicating that STAT3 expression in myeloid cells is detrimental for the control of infection with M. tuberculosis. Despite reduced area of the lung occupied by granuloma the area with inflammation was not reduced and the numbers of infiltrating pulmonary neutrophils were elevated in stat3fl/fllysm cre mice. Neutrophil accumulation late during infection have been associated with susceptibility to M. tuberculosis, whereas early after infection neutrophils play a protective role and contribute to early priming of T cells in the draining lymph node [36–39].
Although mice lacking STAT3 expression in bone marrow progenitors display peripheral neutrophilia under resting conditions [40], the pathways involved in neutrophil mobilization and response to chemokines during inflammation have been shown to be STAT3 dependent [41, 42]. Thus, we consider it unlikely that the increase in pulmonary neutrophils observed during infection occurs as a direct consequence of STAT3 deficiency in these cells. Rather, increased levels of IL-17 and IL-22, cytokines that stimulate the expression of neutrophil recruiting chemokines [29], might contribute to the accumulation of granulocytes in lungs from M. tuberculosis-infected stat3fl/fl lysm cre mice. In agreement with this hypothesis, the frequency of IL-17 secreting mycobacteria-specific CD4+ T cells, but not of γδ+ T cells, were elevated in the lungs from stat3fl/fl lysm cre mice compared to controls.
STAT3 expression in APCs proved to be a major regulator of the expression of cytokines that control T cell differentiation. This was shown in STAT3- and in SOCS3-deficient APC, which display higher levels of activated STAT3 when stimulated with mycobacteria or other innate receptor agonists [21, 43]. Thus, while mycobacteria-infected STAT3 deficient APCs showed an improved ability to trigger IL-17 secretion by antigen-specific T cells, the opposite was observed using socs3fl/fl lysm cre or gp130F/F BMDCs or macrophages as APCs.
We observed an increased IFN-γ secretion by antigen-specific T cells incubated with STAT3-deficient, mycobacteria-infected APCs. However, IFN-γ responses were not increased in vivo. Whether this is due to the present of cytokines that might stimulate IL-17 responses while antagonizing TH1 cells (such as for example TGF-β) remains to be explored. Whereas in vitro cultures used provide a proper tool to gain mechanistic insights, the diversity of populations, the tissue localization and the balance between the host immune responses and mycobacteria in the chronic infection might account for differences observed between in vitro responses and the control of infection in mice.
The increased resistance to M. tuberculosis in stat3fl/fl lysm cre mice is mirrored by data showing that mice with SOCS3 deficiency in myeloid cells display reduced resistance to TB and toxoplasmosis [21, 44]. Here we show that socs3fl/fl cd11c cre mice. CD11c cre in these mice has been shown to be expressed in ca 90% of splenic DCs, compared with <10% of lymphocytes and <1% of myeloid cells such as granulocytes [25]. Thus, these animals in which SOCS3 is deleted in DCs but not in inflammatory macrophages and neutrophils [25] are also more susceptible to infection with M. tuberculosis. This supports that the major role of STAT3 and SOCS3 in myeloid cells in the control of infection with M. tuberculosis is not due to an altered ability of SOCS3 or STAT3-deficient macrophages to control the growth of the intracellular mycobacteria in vitro, as shown here and ref [21].
To our knowledge, this is the first report showing that STAT3 deficiency in myeloid cells promotes IL-17 secretion by antigen-specific T cells in vitro and in vivo. Such a role was related to the increased secretion of TH17 inducing IL-6 and IL-23 by STAT3-deficient APCs. The increased expression of IL-6 and IL-23 in stat3fl/fl lysm cre APCs was not restricted to the infection with attenuated or virulent mycobacteria since, it was observed after incubating mutant APCs with different TLR agonists or bacterial lysates, confirming previous data [45]. The opposite effect was observed using socs3fl/fl lysm cre BMM, which were poor inducers of IL-17 secretion by mycobacteria-specific T cells. In relation to this, DC that secrete IL-12p40 (required for T cell differentiation into Th17 or Th1) in the lymph nodes of mycobacteria infected mice are primarily uninfected [46].
Since IL-6 can be produced by various hematopoietic and non-hematopoietic cells, we suggest that APCs are relevant cellular sources of IL-6 for the differentiation of IL-17 secreting cells during infection. Moreover, our data using gp130F/F BMDCs indicate that DC-derived IL-6 acts in an autocrine/ paracrine manner on DCs to regulate their ability to stimulate IL-17 secretion by T cells. A role for gp130/ IL6/ STAT3 pathway in susceptibility to M. tuberculosis has been previously determined. The high susceptibility of gp130F/F mice to infection with M. tuberculosis was not observed in gp130F/Fil6-/- or gp130F/F stat3+/- mice [21].
Our observations on stat3fl/fl lysm cre mice are reminiscent of those seen in il10-/- or anti-IL-10R mAb treated mice that resulted in enhanced lung TH1 and TH17 responses after BCG vaccination [47]. Depletion of IL-10 resulted in elevated protection to M. tuberculosis in some studies but not others [47–50]. However different to our model, IL-10 might not only impair the functions of APCs but is also secreted by T cells and has a direct inhibitory effect on TH1 or TH17 cells [51].
We showed that the improved M. tuberculosis control in stat3fl/fl lysm cre is IL-17-mediated, since administration of neutralization anti-IL-17RA antibodies abrogated differences in bacterial burden between mutant and control mice. IL-17 might contribute to long term protection, control of infection after vaccination or control of hypervirulent strains of M. tuberculosis [52–55]. IL-17 has been suggested to induce of protective TH1 responses against mycobacterial infection [52, 56]. IL-17 has been also shown to mediate CXCL13 induction in the lung, a chemokine that contributes to the localization of pro-inflammatory cytokine-producing CXCR5+ T cells within lymphoid structures, promoting those macrophage activation and mycobacterial control [53, 57].
However, other studies have indicated that IL-17 is dispensable after primary infection with M. tuberculosis [58]. In line with the later observations, we observed similar M. tuberculosis load in lungs or spleens of WT mice treated or not with anti-IL-17RA.
Levels of MHCII and CD80 and CD86 were lower on socs3fl/fl lysm cre BMDCs after mycobacterial stimulation confirming previous findings showing reduced MHCII and co-stimulatory molecules after LPS stimulation of SOCS3-deficient BMDCs [59]. Furthermore, the activation of mycobacteria specific p25-tg T cells was also diminished in MLN from M. tuberculosis-infected socs3fl/fl lysm cre mice as compared to controls. Of importance, p25-tg T cell proliferation was not detectable in the MLN of mice infected with an Ag85b deficient strain of M. tuberculosis indicating the specificity of p25-tg T cell priming [60]. The Ag85b KO M. tuberculosis strain grew in the lungs and disseminated to the MLN at a rate equivalent to that of wild-type bacteria. Instead, stat3fl/fl lysm cre and control BMDCs expressed similar levels of MHC-II and co-stimulatory molecules after mycobacterial stimulation in vitro and similar levels of activated antigen-specific T cells in vivo. Different to these results, STAT3 deficient APCs have been shown increased MHCII levels after IL-6 stimulation [61].
Finally, the role of STAT3 in T cells in regulation of antigen-specific IFN-γ and IL-17 T cell responses was investigated. Contrary to the role of STAT3 in APCs, IL-17 secretion was hampered in mycobacteria-specific STAT3-deficient T cells. STAT3 is required for the responses to both IL-6, IL-21 and IL-23 and for the expression of RORγt by T cells [62]. Instead, SOCS3-deficient antigen-specific T cells secreted higher IL-17 levels as previously reported in other systems [35], while IFN-γ responses were inhibited. Thus, while the role of STAT3 in T cells in the control of M. tuberculosis remains to be studied, these results illustrate the pleiotropic effect of STAT3 in regulation of infection-induced immune responses in different cell types.
In summary, we here showed using SOCS3- and STAT3-deficient mice that STAT3 in myeloid cells is detrimental for the control of infection with M. tuberculosis. Surprisingly, this occurs via impairing secretion of IL-17 by antigen-specific T cells (Fig 7E).
The animals were housed and handled at the Dept. of Microbiology, Tumor and Cell Biology and the Astrid Fagreus Laboratory, Karolinska Institute, Stockholm, according to directives and guidelines of the Swedish Board of Agriculture, the Swedish Animal Protection Agency, and the Karolinska Institute (djurskyddslagen 1988:534; djurskyddsförordningen 1988:539; djurskyddsmyndigheten DFS 2004:4). The study was performed under approval of the Stockholm North Ethical Committee on Animal Experiments permit number N397/13 and N487/11. Animals were housed under specific pathogen-free conditions.
Mice containing loxP-flanked stat3 and socs3 alleles have been described before [43]. For a myeloid-specific deletion these were bred with transgenic lysm cre mice [63]. Socs3fl/fl mice were also bred with cd11c cre transgenic animals. Stat3fl/fl or socs3fl/fl littermates negative for cre expression were used as controls for all experiments. Gp130F/F mice with a homozygous substitution of tyrosine (Y)757 to phenylalanine (F) within the common IL-6 family receptor gp130 abrogating the SOCS3 binding site have been described before [64]. Transgenic T cell receptor p25-tg mice with a T-cell receptor specific for peptide 25 (aa 240–254) of mycobacterial Ag85B on H2b haplotype were used [65]. p25-tg rag2-/- mice expressing ECFP were generated by crossing p25-tg with rag1-/- mice [65] with ECFP mice on a rag2-/- background (kindly provided by Dr. Ronald Germain, NIAID, NIH). The ECFP expression co-localized with Vβ11 used by p25tg T cells [65]. Socs3fl/fl lck cre and stat3fl/fl lck cre mice deficient in SOCS3 and STAT3 in T cells were crossed with p25-tg mice to generate p25-tg socs3fl/fl lck cre and p25-tg stat3fl/fl lck cre mice. P25-tg lck cre mice were also obtained and used as controls.
BCG Montreal and M. tuberculosis Harlingen were grown in Middlebrook 7H9 (Difco, Detroit, MI) supplemented with albumin, dextrose, catalase and, for BCG cultures, 50 μg/ ml hygromycin (Sigma, St. Louis, MO). Mice were infected with 250 M. tuberculosis Harlingen strain by aerosol using a nose-only exposure unit (In-tox Products, Uppsala, Sweden)[66].
Bacteria were quantified on Middlebrook 7H11 agar containing 10% enrichment of oleic acid, albumin, dextrose, catalase, 5 μg of amphotericin B per ml and 8 μg/ ml polymyxin B grown for 3 weeks at 37°C.
Bone marrow was extracted from tibia and femurs of mice and resuspended in DMEM containing glucose and supplemented with 10% FCS and 30% L929 cell-conditioned medium (as a source of macrophage-colony stimulating factor). Bone marrow cells were passed through a 70 μm cell strainer, plated and incubated for 6 days at 37°C, 5% CO2. Bone marrow-derived macrophage (BMM) cultures were then washed vigorously to remove non-adherent cells, trypsinized, counted and cultured for one day at 37°C in 24, 12 or 6 well plates. We have previously shown that these BMM are F4/80+, CD14+ and Mac-3+ [67].
In order to quantify intracellular M. tuberculosis uptake and growth, BMM cells were plated on glass slides at 2.105 cells per well in 24 well plates, incubated for 4 h with M. tuberculosis (MOI 2) and washed with PBS for 3 times to remove the extracellular bacteria before either fixation or replacing the medium. Three days after infection cells were washed with PBS, fixed with 2% PFA and stained with phalloidin to label F-actin (Life technologies, 1:100), DAPI (1:500) and auramine-rhodamine T to label mycobacteria (BD). Micrographs from infected macrophages (400X) were obtained and a total of at least 1000 BMM from 3 independent cultures and categorized as infected or uninfected. The intracellular M. tuberculosis were enumerated. BMM harboring 5 or more bacteria were considered as containing 5. In some cultures, mycobacterial CFU from BMM 6 days after infection were determined.
Mouse bone marrow-derived dendritic cells (BMDC) were differentiated as previously described [68]. Briefly, bone marrow was extracted from tibia and femurs and cell suspensions cultured in RPMI-1640 medium containing 10% FCS and 2 ng/ ml GM-CSF (Peprotech, Rocky Hill, NJ). Fresh medium and cytokine were replaced after 3 days. After six days of culture, loosely adherent cells were harvested and seeded in concentrations for infection. Harvested cells were further selected for CD11c expression with magnetic beads (Miltenyi Biotech) before seeding.
BMDC or BMM were stimulated with either live or heat killed BCG, M. tuberculosis or Ag85b peptide in presence of LPS for 6 h. Then, cells were washed and co-incubated with p25-tg CD4+ lymph node transgenic T cells from rag2-/- p25-Tg mice (at a ratio of 4:1 BMDC). The cultures were further incubated for 24–48 hs at 37C 5% CO2. At these time points the concentration of IFN-γ and IL-17 in the supernatants was measured by ELISA.
Transcripts were quantified by real time PCR as previously described[66]. Hprt was used as a control gene to calculate the ΔCt values for individual samples. The relative amount of cytokine/ hprt transcripts was calculated using the 2-(ΔΔCt) method. These values were then used to calculate the relative expression of cytokine mRNA in uninfected and infected cells and tissues.
Lungs were perfused with PBS through the heart before removal from mice. Lungs were mechanically minced into small pieces and digested with 3 mg/ ml Collagenase D and 30 μg/ ml DNase I for 1 h at 37°C, and single-cell suspensions prepared by filtering lung tissue through 70-μm nylon cell strainers. To further remove impurities cells were loaded in 40/ 70% Percoll gradient in PBS and centrifuged 30 min room temperature. The cells at the interphase were collected and washed. Single spleen cell suspensions were obtained by mechanical disruption, lysis of erythrocytes and straining over a 70-μm nylon mesh. Lung, lymph node and spleen cells were stained for CD3, CD4, CD8, γδ-TCR, CD62L, CD69, CD44, CD11b, CD11c, Ly6C and Ly6G (all eBioscience) and fixed before acquisition.
For determination of IFN-γ and IL-17-producing cells, lung cells were incubated with PPD or with 50 ng/ml phorbol myristate acetate (PMA) and 2 μg/ml ionomycin (Sigma) for 6 or 18 h at 37 oC. Brefeldin (10 μg/ ml) was added to the cultures the last 4 h of stimulation. Cells were then stained with cell population-specific antibodies, and live/ dead staining, fixed, permeabilized using leukocyte permeabilization reagent IntraPrep™ (Immunotech, Marseille, France) and further stained with anti-IL-17a or anti-IFN-γ (eBioscience).
Data were acquired in a CyAn™ ADP (Beckman Coulter) or an LSRII Flow cytometry and analyzed with FlowJo software (Tree star Inc., Ashland, OR).
Formalin fixed left lungs of mice experimentally inoculated with M. tuberculosis were blocked on paraffin. From each lung sample 4 sections were obtained, one longitudinal along the long axis of the lobe and 3 across/transversal of the remaining piece of lung.
The blocks were processed and sections were stained with haematoxylin-eosin. All sections were interpreted by the same pathologist (D. G-W.) and scored semi-quantitatively, blinded to the variables of the experiment.
The following features were scored:
The Mann Whitney test for the bacterial CFU load in vivo and of the ICS analysis. For each experiment, 8–10 control and 8–10 mutant mice were infected. We performed separated experiments for 4 and 8 weeks post infection. One of two independent experiments showing similar results is shown.
The analysis of cytokine secretion or mRNA, histopathological scores and frequencies was done using the Student’s t test for unpaired samples. All in vitro experiments were performed at least twice. A two-way ANOVA was used to compare the differences in IL-17 secretion between genotypes, as well as between cells that co-secrete IFN-γ or not.
|
10.1371/journal.pntd.0000294 | Hyaluronidase of Bloodsucking Insects and Its Enhancing Effect on Leishmania Infection in Mice | Salivary hyaluronidases have been described in a few bloodsucking arthropods. However, very little is known about the presence of this enzyme in various bloodsucking insects and no data are available on its effect on transmitted microorganisms. Here, we studied hyaluronidase activity in thirteen bloodsucking insects belonging to four different orders. In addition, we assessed the effect of hyaluronidase coinoculation on the outcome of Leishmania major infection in BALB/c mice.
High hyaluronidase activity was detected in several Diptera tested, namely deer fly Chrysops viduatus, blackflies Odagmia ornata and Eusimilium latipes, mosquito Culex quinquefasciatus, biting midge Culicoides kibunensis and sand fly Phlebotomus papatasi. Lower activity was detected in cat flea Ctenocephalides felis. No activity was found in kissing bug Rhodnius prolixus, mosquitoes Anopheles stephensi and Aedes aegypti, tse-tse fly Glossina fuscipes, stable fly Stomoxys calcitrans and human louse Pediculus humanus. Hyaluronidases of different insects vary substantially in their molecular weight, the structure of the molecule and the sensitivity to reducing conditions or sodium dodecyl sulphate. Hyaluronidase exacerbates skin lesions caused by Leishmania major; more severe lesions developed in mice where L. major promastigotes were coinjected with hyaluronidase.
High hyaluronidase activities seem to be essential for insects with pool-feeding mode, where they facilitate the enlargement of the feeding lesion and serve as a spreading factor for other pharmacologically active compounds present in saliva. As this enzyme is present in all Phlebotomus and Lutzomyia species studied to date, it seems to be one of the factors responsible for enhancing activity present in sand fly saliva. We propose that salivary hyaluronidase may facilitate the spread of other vector-borne microorganisms, especially those transmitted by insects with high hyaluronidase activity, namely blackflies (Simuliidae), biting midges (Ceratopogonidae) and horse flies (Tabanidae).
| Hyaluronidases are enzymes degrading the extracellular matrix of vertebrates. Bloodsucking insects use them to cleave the skin of the host, enlarge the feeding lesion and acquire the blood meal. In addition, resulting fragments of extracellular matrix modulate local immune response of the host, which may positively affect transmission of vector-borne diseases, including leishmaniasis. Leishmaniases are diseases with a wide spectrum of clinical forms, from a relatively mild cutaneous affection to life-threatening visceral disease. Their causative agents, protozoans of the genus Leishmania, are transmitted by phlebotomine sand flies. Sand fly saliva was described to enhance Leishmania infection, but the information about molecules responsible for this exacerbating effect is still very limited. In the present work we demonstrated hyaluronidase activity in salivary glands of various Diptera and in fleas. In addition, we showed that hyaluronidase exacerbates Leishmania lesions in mice and propose that salivary hyaluronidase may facilitate the spread of other vector-borne microorganisms.
| Hyaluronidases are a family of enzymes that degrade hyaluronan (HA) and several other glycosaminoglycan constituents of the extracellular matrix of vertebrates (for review see [1]). In insects, hyaluronidases are well-known from venoms of Hymenoptera and represent clinically important allergens of honey-bees, wasps and hornets [2]–[4]. Hyaluronidases were found also in cDNA libraries of salivary glands (sialomes) of various bloodsucking insects [5]–[8] and the enzyme activity was found in saliva of three groups of Diptera, namely sand flies, blackflies, and horse flies [9],[10]. Salivary hyaluronidases of parasitic insects may have diverse effects on the host. They play an important role in blood meal acquisition; by degrading HA abundant in host skin, hyaluronidases increase tissue permeability for other salivary components that serve as antihaemostatic, vasodilatory or anti-inflammatory agents [5],[9]. This is why hyaluronidases are frequently called “spreading factors” [11]. The enzyme activity facilitates the enlargement of the feeding lesion and the insect acquires the blood meal more rapidly. In addition, HA fragments were shown to have immunomodulatory properties; they affect maturation and migration of dendritic cells, induction of iNOS and chemokine secretion by macrophages and proliferation of activated T cells (reviewed in [12]). As blood sucking insects represent the most important vectors of infectious diseases, local immunomodulation of the vertebrate host may positively enhance the infection.
Leishmaniasis is one of the most prevalent vector-borne diseases. It is initiated by the intradermal inoculation of Leishmania promastigotes during the bite of an infected sand fly (Diptera: Phlebotominae). As shown first by Titus and Ribeiro [13] saliva of the sand fly vector exacerbates the initial phase of Leishmania infections in terms of parasite burden and size of the cutaneous lesion. Sand fly saliva was described to contain an array of pharmacologically active compounds affecting host hemostasis and immune mechanisms (reviewed in [14],[15]) but the information about molecules responsible for the exacerbating effect is still very limited. Morris et al. [16] showed that maxadilan, a well-known vasodilator of the New World vector Lutzomyia longipalpis, exacerbates Leishmania infection to the same degree as whole saliva. Maxadilan inhibits splenocyte proliferation induced in vitro and delayed type hypersensitivity in mice [17] and it also has several inhibitory effects on macrophages and monocytes that would support Leishmania survival in the host [18]. However, this important peptide was not found in Old World vectors of genus Phlebotomus (www.ncbi.nih.gov), including P. papatasi where exacerbating effect of saliva was repeatedly demonstrated [19],[20]. The vasodilatory activity of P. papatasi was instead ascribed to adenosine and AMP present in saliva of this sand fly [21].
In the present work, we studied hyaluronidase activity in bloodsucking insects of four different orders. In addition, we assessed the effect of hyaluronidase coinoculation on the outcome of Leishmania major skin lesions and spreading into draining lymph nodes.
Samples used are summarized in Table 1. The insects originated from laboratory colonies or were collected in the wild. Salivary glands were dissected out in Tris buffer (20 mM Tris, 150 mM NaCl, pH 7.8) and stored in batches (usually 20 glands in 20 µl of Tris buffer) at −70°C. Where dissection of salivary glands was not feasible, whole bodies (Ctenocephalides flea, Culicoides midge) or the thoracic parts containing salivary glands (Pediculus louse) were used at protein concentration 20 µg/µl. Salivary gland extracts (SGE) or body extracts (BE) were obtained by disruption of tissue by three freeze-thaw cycles in liquid nitrogen, homogenization and centrifugation at 12,000 g for 5 min. Protein concentration was determined by Bradford assay using bovine serum albumin in Tris buffer as a standard.
Enzyme activity was detected by the dot method on 10% polyacrylamide gels with copolymerized hyaluronic acid (HA, potassium salt, from human umbilical cord, ICN Pharmaceutical, CA). Gels were prepared using 0.1 M acetate, pH 5.5, containing 0.1 M NaCl, 0.05% Tween-20 and 0.002% HA. This method was previously proved as sensitive and reproducible [10]. Preliminary experiment with selected salivary extracts revealed that Phlebotomus papatasi and Culex pipiens samples were positive at pH 4.5, 5.5, 6.5 and 7.5 while Aedes aegypti, Anopheles stephensi and Glossina fuscipes samples were consistently negative (Fig. S1). Therefore pH 5.5 was chosen for this assay as this is about the pH optimum known for salivary hyaluronidases of various Diptera [9],[10]. Insect samples (2 µl volume) were dotted on the gel and sheep testicular hyaluronidase, (Sigma, 1 µg in 1 µl) was used as a control. Incubation was carried out for 24 hrs at 37°C in a moist chamber. The gels were then washed in water, soaked in 50% formamide for 30 min and stained in Stains-all (Sigma) solution (100 µg/ml in 50% formamide) for 24 hrs in the dark. After a rinse in distilled water the gels were scanned and photographed. To determine whether the enzyme activity was specific for cleaving HA, we tested positive samples also with another component of extracellular matrix, chondroitin sulfate. The method was performed as described above, only HA was replaced by 0.002% chondroitin sulfate (Sigma).
Electrophoresis (SDS PAGE) was carried out on 10% slab gels (0.75 mm thick) using Mini-Protean II apparatus (Biorad) and constant voltage 150 V. Substrate gels were copolymerized with 0.002% HA. As the hyaluronidase activities and band patterns varied among insects, different loads were used per lane in order to obtain bands of equal intensity. Following electrophoresis, gels were rinsed 2×20 min in 0.1 M Tris, pH 7.8, 20 min in 0.1 M acetate buffer, pH 5.5 (both with 1% Triton X-100 to wash out SDS) and then incubated in 0.1 M acetate buffer (without detergent) for 120 min at 37°C. After rinsing in water the gels were stained with Stains-all as described above. Hyaluronidase activity was visible as a pink band on a dark blue background.
Experiments on mice were done in accordance with Czech Act No. 246/1992 and approved by IACUC of the Fac. Sci., Charles University in Prague. A mouse ear infection model [19] was used to assess the effect of hyaluronidase coinoculation on the outcome of Leishmania infection. Leishmania major clone LV561 (MHOM/IL/67/LRC-L137 Jericho II) was cultured on blood agar from defibrinated rabbit blood, supplemented with 50 µg/ml gentamicin. Female BALB/c mice (Charles River Deutschland, Sulzfeld, Germany) were used at the age of 8 weeks. Ether-anaesthetized mice were inoculated in the ear dermis with 104 or 105 L. major stationary-phase promastigotes (subculture 1) in 5 µl sterile saline. The inoculum also contained bovine testicular hyaluronidase (Sigma) in an amount equivalent to 2 or 10 “optimal salivary glands” of Phlebotomus papatasi [10], i.e. 0.4 and 2.0 relative turbidity reducing units, respectively. Bovine testicular hyaluronidase belongs to the same enzyme class as the sand fly salivary hyaluronidases [10] and shares sequence homology with the enzyme of L. longipalpis [5]. Control animals were inoculated with parasites in sterile saline only. Sixty mice (10 for each of six groups) were used for Q-PCR and another 48 (8 for each group) for lesion monitoring. The size of skin lesions was measured weekly using a Vernier caliper gauge. Lesions were monitored for 6 weeks post infection: the area was calculated from two perpendicular measurements as an ellipse area, and its appearance (degree of ulceration) was assessed using an arbitrary scale from 1 to 5 (1 - low induration, 2 - high induration, 3 - small ulcer, 4 - large ulcer, 5 - perforated ear pinna). Independently in both parasite doses (104 and 105), the significance of the hyaluronidase effect was tested using nonparametric Kruskal-Wallis ANOVA and post hoc comparisons of mean ranks using Statistica 7 routines [22]. The tests were performed separately for weeks 3, 4, 5, and 6 post-infection; the size of a lesion was calculated as its area weighted by the degree of ulceration.
Mice were sacrificed 24 hrs post inoculation (p.i.) as the preliminary experiment revealed that lymph nodes of mice dissected 24 hours p.i. gave more consistent results than those dissected 48 hours p.i. (Fig. S2). Parasite numbers in draining retromaxillar lymph nodes were determined by quantitative PCR (Q-PCR) as described earlier [23]. Briefly, dissected lymph nodes were stored in 10 µl saline at −70°C. Total DNA was isolated from homogenised samples using High Pure PCR Template Preparation Kit (Roche); kinetoplast DNA was targeted using primers described elsewhere [24]. The relative effectiveness of three hyaluronidase doses (equivalent to 0, 2, and 10 P. papatasi salivary glands) with both infection doses (104 and 105 parasites) was evaluated by analysis of variance (Statistica v. 7.1, factorial and one-way ANOVA).
The dot method on gels with copolymerized HA and chondroitin sulfate was used to study the presence of hyaluronidase activity and its substrate specificity. The highest hydrolysis of HA was observed in SGE of deer fly Chrysops viduatus. Pronounced hydrolysis was found in SGEs of blackflies Odagmia ornata and Eusimulium latipes, mosquito Culex quinquefasciatus, sand fly Phlebotomus papatasi and whole body extract of biting midge Culicoides kibunensis (syn. C. cubitalis). Lower activity was detected in BE of cat flea Ctenocephalides felis (Fig. 1). On the other hand, no detectable hydrolysis of HA occurred in SGEs of kissing bug Rhodnius prolixus, mosquitoes Anopheles stephensi and Aedes aegypti, tse-tse fly Glossina fuscipes, stable fly Stomoxys calcitrans and in thoracic extracts of human louse Pediculus humanus (Fig. 1). Positive samples were then tested also for chondroitin sulfate hydrolysis (Fig. 2). High activity was observed in Culex quinquefasciatus and Culicoides kibunensis, in other samples the hydrolysis of chondroitin sulfate was either moderate (Chrysops viduatus) or low (Phlebotomus papatasi, Ctenocephalides felis) (Fig. 2); clearly, HA is the preferred substrate for the enzymes of these three insects.
Seven samples positive in the dot method were analyzed by zymography to reveal the apparent molecular weight (MW) of hyaluronidases. The MW of the enzymes differed among various insects (Figs. 3 and 4). Under nonreducing conditions hyaluronidases were detected as major diffuse bands (Fig. 3). The SGE activity in Phlebotomus papatasi had a MW about 70 kDa while those in both blackfly species tested, Eusimulium latipes, and Odagmia ornata, about 40 kDa. In BE of Culicoides kibunensis, the major band of about 35 kDa was accompanied with a minor one of 70 kDa, supposedly a dimer. Chrysops viduatus SGE revealed one major band with estimated MW of 50 kDa. In BE of flea Ctenocephalides felis, three enzyme bands were detected, the most prominent one of about 52 kDa (Fig. 3). Under reducing conditions, SDS PAGE revealed sharper enzyme bands allowing more precise assignment of corresponding MW (Fig. 4). In sand fly P. papatasi, both blackfly species and deer fly Chrysops viduatus, hyaluronidase activity was observed within the same MW ranges as under nonreducing conditions (70, 40 kDa, and 50 kDa, respectively). In Culicoides kibunensis and Ctenocephalides felis hyaluronidase activity was not detectable under reducing conditions (Fig. 4).
No hyaluronidase activity was detected in Culex quinquefasciatus SGE under either zymography conditions used, reducing and nonreducing. An additional experiment was performed to explain the contradictory results from the dot method and zymography; SGE of C. quinquefasciatus was dotted on the gel with copolymerized HA with and without the presence of SDS. Hydrolysis was observed only in the sample without SDS (Fig. 5).
Next we examined whether hyaluronidase altered the course of Leishmania major infection in BALB/c mice. We used intradermal inoculation into the ear and the disease burden was calculated from weekly measuring the lesion size. As shown in Fig. 6, mice coinjected with parasites and hyaluronidase developed bigger lesions. In all groups of mice, the onset of lesion development was at three weeks p.i. Thereafter, the lesions grew faster in coinoculated groups. The experiment was terminated six weeks post infection when, in some animals, ulcerating lesion spread over the majority of ear pinna. In mice inoculated by higher parasite numbers (105), both hyaluronidase treatments produced similar effects (Fig. 6A). In mice with an inoculation dose one order of magnitude lower (104), the effect of hyaluronidase was concentration-dependent: lesions were bigger in mice coinoculated with hyaluronidase activity equivalent of 10 P. papatasi salivary glands than in those coinoculated with equivalent of 2 glands (Fig. 6B). In both parasite numbers (104 and 105) over all considered weeks (3 to 6) post-inoculation, Kruskal- Wallis ANOVA showed significant differences among hyaluronidase treatments (p always≤0.025), with only one exception in week 3 of 104 parasites treatment (p = 0.23). Consequently, the post-hoc comparison of treatments tests confirmed the significant difference between controls (no hyaluronidase) and corresponding inoculated hyaluronidase doses (2 or 10 glands equivalents). We also tested the difference between the 2 and 10 gland equivalent doses: however, despite the common trends apparent in Fig. 6 indicating that there may be a systematic difference between 2 and 10 gland equivalents doses, the post-hoc comparison of treatments test did not prove it in any case but in week 5 of the 104 parasites treatment.
We also examined whether hyaluronidase affected Leishmania major load in draining lymph nodes of BALB/c mice one day p.i. Using Q-PCR, no significant differences were observed among control and experimental groups of mice at both parasite doses (104 or 105 L. major) tested (F(2, 54) = 0.043; p = 0.96) (Fig. S3).
Parasitic insects utilize two strategies for finding blood: solenophagy (or vessel feeding) and telmophagy (or pool feeding). In solenophagic approach, the feeding fascicle cannulates a blood vessel, while in the pool-feeding mode the mouth part stylets slash through the skin, and the insect sips blood that oozes out from the hemorrhage. In our experiments, pronounced hyaluronidase activity was found in black flies, biting midges, sand flies and deer flies. All these insects belong to parasitic Diptera with pool-feeding mode of blood meal acquisition. The activity was detected also in cat flea (Ctenocephalides felis, Siphonaptera) and in Culex quinquefasciatus mosquito (Diptera). Although these two species belong to different insect orders, they are both vessel feeders. In contrast, no activity was detected in other vessel-feeding insects: human lice, kissing bugs, Anopheles and Aedes mosquitoes, tsetse flies, and stable flies.
Hyaluronidase activity was previously detected in the saliva of various sand fly species [9],[10] as well as in the saliva of the black fly Simulium vittatum [9] and horse fly Tabanus yao [8]. Sequences predicted to code for hyaluronidases were found in the salivary transcriptomes of the mosquito Culex quinquefasciatus [6] and the biting midge Culicoides sonorensis [7]. Herein, we demonstrated that Culex quinquefasciatus and Culicoides kibunensis possess hyaluronidase activity and, in parallel experiments, we detected hyaluronidase activity in saliva of two other species of biting midges Culicoides sonorensis and C. nubeculosus (Volfova et al., unpublished). Therefore, we showed that in biting midges and in Culex quinquefasciatus, the transcripts coding for putative hyaluronidases are translated into functional enzymes.
To determine whether the enzyme activity was specific for cleaving HA, we also tested another component of mammalian extracellular matrix, chondroitin sulfate. All hyaluronidase-positive samples tested cleaved chondroitin sulfate, which would indicate that insect hyaluronidases fall into the same class as mammalian hyaluronidases (E.C. 3.2.1.35 according to IUBMB Enzyme Nomenclature) [25]. Indeed, sequence analysis of transcripts putatively coding for hyaluronidase enzymes reveals their homology to mammalian enzymes [6],[7],[9]. While HA was found as the preferred substrate for most samples tested, very high hydrolysis of chondroitin sulfate was found in Culex quinquefasciatus SGE. This mosquito species differs from other samples tested also in other aspects. In zymography assay, salivary hyaluronidase of Culex quinquefasciatus was irreversibly sensitive to denaturation effect of SDS while enzymes of other insects tested refolded and regained activity after removal of the denaturating agent. Further work is needed to understand the differences in the molecular structure and substrate specificity of hyaluronidases from Culex mosquitoes versus other bloodsucking insects. In addition, in other mosquitoes studied, Anopheles darlingi [26], funestus [27] and gambiae [28] and Aedes aegypti [29] and albopictus [30], neither hyaluronidase activity nor hyaluronidase gene was found in salivary transcriptomes. Sequences homologous to hyaluronidase were, however, found by genome sequencing in Anopheles gambiae [31] and Aedes aegypti [32].
As revealed by zymography, hyaluronidases of different insect species tested vary substantially in MW and the structure of the molecule. Putative oligomers were seen in Culicoides kibunensis. Oligomeric forms have been found frequently among mammalian hyaluronidases. In sand flies, oligomers or dimers were found in Lutzomyia longipalpis, Phlebotomus papatasi, and P. sergenti [10]. Multiple bands observed by zymography in Ctenocephalides felis whole body extract could, however, represent multiple hyaluronidase enzymes.
In Culicoides and Ctenocephalides, reducing conditions affected the stability of the enzymes; 2-mercaptoethanol inhibited hyaluronidase activity. In Culicoides midges, the sensitivity of hyaluronidase activity to reducing conditions was confirmed by experiments with pure saliva of laboratory bred Culicoides sonorensis and C. nubeculosus (Volfova et al., unpublished). This implies that reduction-sensitive residues are either important for the function of the active site of the enzyme, or steric relations in the molecule. On the other hand, hyaluronidases of other insects tested, namely Phlebotomus papatasi, Eusimulium latipes, Odagmia ornata, and Chrysops viduatus, remained active under reducing conditions. Addition of 2-mercaptoethanol did not result in differences in the apparent MW, suggesting that the enzymes consist of a single polypeptide chain. These results correspond with previous observations on sand flies; sand fly hyaluronidases strikingly differed in structure and sensitivity to reducing conditions, even among various species of the genus Phlebotomus [10].
We showed that hyaluronidase is a common constituent of saliva of bloodsucking insects. It seems to be essential for insects with pool-feeding mode, where it facilitates the enlargement of the feeding lesion, serving as a spreading factor for other pharmacologically active compounds present in saliva. Very little is known, however, about the possible role of salivary gland hyaluronidase in allergic reactions which occur in some patients after repeated bites of bloodsucking Diptera. In Hymenoptera, venom hyaluronidase is largely responsible for the cross-reactivity of venoms with sera of allergic patients [4]. In several patients, coexistent anaphylaxis to Hymenoptera sting and Diptera bite was described [33] and hyaluronidase is a candidate allergen responsible for this type of crossreactions. In experiments of Sabbah et al. [34],[35], IgE of allergic patients recognized shared proteins within MW range 44–50 kDa between wasp venom and total extracts of mosquito and horse fly. Unfortunately, these interesting data are difficult to assess given the incomplete identitification of the mosquito and horse fly species tested.
A mouse ear infection model was used to assess the effect of hyaluronidase coinoculation on the outcome of Leishmania major infection. The activity of sand fly enzyme was mimicked by commercially available bovine hyaluronidase. More severe lesions developed in mice where L. major promastigotes were coinjected with hyaluronidase. Even the lower dose of the enzyme corresponding with the activity produced by 1–2 sand fly females resulted in significant differences against the control mice where parasites alone were injected. It would be worth testing if differences observed in lesion size are mainly due to number of parasites or to inflammatory response to coinoculated hyaluronidase. In contrast, there was neither more rapid onset of lesions, nor faster dissemination of Leishmania in the lymph node. Parasite numbers in draining lymph nodes collected 24 and 48 hrs p.i. were similar in all experimental groups. Although hyaluronidase activity exacerbated Leishmania lesions in the skin, it did not support its visceralization. However, we can not exclude the possibility that consequences of hyaluronidase for parasite visceralization are not immediate and thus could not be detected in the present study.
The way by which hyaluronidase enhances the establishment of Leishmania is unknown, but we suggest that it is due to HA fragments generated by hyaluronidase activity in the host skin. HA occurs in two main forms: the high MW (HMW) polymers and the low MW (LMW) fragments. HMW HA is a common component of vertebrate extracellular matrix. LMW HA fragments are generated under inflammatory conditions by endogenous or bacterial hyaluronidases [36], or non-enzymatically by free radicals [37]. HA fragments have diverse immunomodulatory properties; they affect DC maturation, T cell proliferation, cytokine, and chemokine synthesis by lymphocytes and macrophages (reviewed in [12]). Thus, following injury or infection, HA fragments have been implicated as both endogenous and exogenous triggers of repair and/or defense mechanisms [38],[39] and might truly represent a “danger signal” [40]. Leishmania parasites, however, may profit from the local increase of HA fragments. Specifically, endothelial cells were shown to respond to LMW HA by IL-8 production [38] that results in neutrophil recruitment. As neutrophil granulocytes were indicated as Trojan horses enabling Leishmania silent entry into macrophages [41] their accumulation at the site of sand fly bite might promote infection establishment.
In conclusion, we demonstrated that hyaluronidase promotes Leishmania establishment in murine skin. As this enzyme is present in all Phlebotomus and Lutzomyia species studied to date [10] it seems to be one of the factors responsible for enhancing activity present in saliva of the New-World as well as the Old-World sand flies. We propose that hyaluronidase, in concert with other insect-derived molecules, may facilitate the spread of other vector-borne diseases, especially those transmitted by vectors with high hyaluronidase activity in saliva, namely blackflies, biting midges, deer flies and horse flies.
|
10.1371/journal.pntd.0005785 | An economic evaluation of vector control in the age of a dengue vaccine | Dengue is a rapidly emerging vector-borne Neglected Tropical Disease, with a 30-fold increase in the number of cases reported since 1960. The economic cost of the illness is measured in the billions of dollars annually. Environmental change and unplanned urbanization are conspiring to raise the health and economic cost even further beyond the reach of health systems and households. The health-sector response has depended in large part on control of the Aedes aegypti and Ae. albopictus (mosquito) vectors. The cost-effectiveness of the first-ever dengue vaccine remains to be evaluated in the field. In this paper, we examine how it might affect the cost-effectiveness of sustained vector control.
We employ a dynamic Markov model of the effects of vector control on dengue in both vectors and humans over a 15-year period, in six countries: Brazil, Columbia, Malaysia, Mexico, the Philippines, and Thailand. We evaluate the cost (direct medical costs and control programme costs) and cost-effectiveness of sustained vector control, outbreak response and/or medical case management, in the presence of a (hypothetical) highly targeted and low cost immunization strategy using a (non-hypothetical) medium-efficacy vaccine.
Sustained vector control using existing technologies would cost little more than outbreak response, given the associated costs of medical case management. If sustained use of existing or upcoming technologies (of similar price) reduce vector populations by 70–90%, the cost per disability-adjusted life year averted is 2013 US$ 679–1331 (best estimates) relative to no intervention. Sustained vector control could be highly cost-effective even with less effective technologies (50–70% reduction in vector populations) and in the presence of a highly targeted and low cost immunization strategy using a medium-efficacy vaccine.
Economic evaluation of the first-ever dengue vaccine is ongoing. However, even under very optimistic assumptions about a highly targeted and low cost immunization strategy, our results suggest that sustained vector control will continue to play an important role in mitigating the impact of environmental change and urbanization on human health. If additional benefits for the control of other Aedes borne diseases, such as Chikungunya, yellow fever and Zika fever are taken into account, the investment case is even stronger. High-burden endemic countries should proceed to map populations to be covered by sustained vector control.
| Transmitted by the Aedes mosquito, dengue affects more than 100 countries and is rapidly emerging as the leading vector-borne disease. There has been a 30-fold increase in the number of cases reported since 1960. The cost of the illness to the health system and to society at large is estimated at several billions of dollars annually. The health sector response has depended in large part on controlling mosquito populations during outbreaks. Recently, the first-ever dengue vaccine received regulatory approval for use in several countries. However, its roll-out and long-term impact still needs to be evaluated in the field. In this paper, we examine how the introduction of this vaccine might alter the investment case for sustained effort to control mosquitoes. To our knowledge, this is the first economic evaluation of mosquito control in the era of the dengue vaccine. We model the cost and effects of mosquito control in Brazil, Columbia, Malaysia, Mexico, the Philippines, and Thailand. We evaluate the cost-effectiveness of mosquito control in the presence of a vaccine that does not offer full protection to all individuals. Our results suggest that sustained mosquito control will continue to be cost-effective, even if roll-out of the current vaccine is highly targeted and low-cost. These results support current global policies and strategies for the prevention and control of dengue.
| Dengue is a rapidly emerging disease endemic in more than 100 countries, with evidence of transmission reported in 128 countries [1]. Today, hundreds of thousands of severe dengue cases arise every year, including about 20 000 deaths. The economic cost of the illness in the Americas and South-East Asia is already measured in the billions of dollars annually, including costs such as work and school days lost [2][3][4]. In the Western Pacific, between half and two-thirds of affected households have incurred debt as a result of the care they received [5][6]. Environmental change and unplanned urbanization are conspiring to raise the cost of dengue infection further beyond the reach of health systems and households. In 2014, Southern China suffered the worst outbreak of dengue fever in more than two decades; Japan saw autochthonous transmission in its first outbreak of the disease since 1945 [7,8].
In the absence of a fully effective vaccine or any treatment, dengue control has depended solely on the control of the Aedes aegypti and Aedes albopictus vectors. Current strategies include personal protection or biological, chemical, and environmental measures. In 2006, the second edition of the Disease Control Priorities Project (DCP2) put the cost per disability-adjusted life year (DALY) averted by vector control at US$ 1992–3139 (presumably in 2005 US$). Since then, the dengue economics literature suggests lower cost-effectiveness ratios ranging from 2005 US$ 227 (2013 US$ 344) per DALY averted by larval control in Cambodia to 2009 US$ 615–1267 (2013 US$ 802–1652) per DALY averted by adult mosquito control in Brazil [9][10].
A recent systematic review concluded that results were not easily comparable due to differences in methodological assumptions, and that combined control strategies remained largely unexplored [11].
The authors note that “there is growing interest in combining vector control with vaccination once a dengue vaccine becomes widely available, which recognizes that one intervention is insufficient to effectively reduce the burden of disease.” They cite results from studies with malaria and lymphatic filariasis that support the impact of simultaneously targeting vectors and pathogens.
Prior to 2015, the absence of a dengue vaccine did not preclude efforts to model the conditions under which such an immunization strategy might be cost-effective [12][13][14][15]. Prospects for a viable immunization strategy have since improved, though not without complications [16]. In December 2015, the first-ever dengue vaccine, known as chimeric yellow fever virus-dengue virus tetravalent dengue vaccine or CYD-TDV (Dengvaxia), was approved for use in the Philippines. A strategy consisting of one year of catch-up vaccinations targeting children 9–15 years of age, followed by regular vaccination of 9-year-old children, may be cost-effective at costs up to $72 from a health-care perspective and up to $78 from a societal perspective [17]. However, a review of the results of eight independent modelling groups concluded that “the potential risks of vaccination in areas with limited exposure to dengue as well as the local costs and benefits of routine vaccination are important considerations” [18].
It has been argued that, even in the era of a vaccine, the health sector response to dengue is expected to continue to depend in large part on vector surveillance and control [19,20]. In this paper, we undertake a comprehensive review and synthesis of the evidence on the cost and cost-effectiveness of dengue control interventions from the health system perspective. We appraise the cost-effectiveness of sustained vector control (including outbreak response) in the presence of a (hypothetical) highly targeted and low-cost immunization strategy using a (non-hypothetical) medium-efficacy vaccine. We consider also outbreak response and/or medical case management alone. We model a zero-cost or null scenario in which no intervention at all is implemented (not even medical case management).
We generate cost and cost-effectiveness estimates for six middle-income countries: Brazil, Columbia, Malaysia, Mexico, the Philippines, and Thailand. These countries were the focus of recent systematic reviews of epidemiological trends and make up an estimated 15% of the global burden of dengue [21]. Our approach is deliberately conservative. As the focus of the study is assessing the cost-effectiveness of vector control in the age of a dengue vaccine, a conservative approach means making generous assumptions about the efficacy and cost of the vaccine and ungenerous assumptions about the efficacy and cost of vector control.
A dynamic compartmental model using Markov chains was built using the open source software R. The model builds on that of Luz et al. (2011) [9]. We introduce outbreaks, urbanization, and climate change and update the model within a probabilistic sensitivity analysis (PSA) framework. We generate 1000 simulations for each of 780 weekly cycles (15 years), representing the years 2016–2030, following a model burn-in period of 500 cycles. Best estimates and 95% uncertainty intervals (UIs) are obtained by the mean and the 2.5th and the 97.5th percentiles across all iterations. For cost-effectiveness, we report uncertainty using cost-effectiveness acceptability curves and a Net Monetary Benefit (NMB) approach.
The model runs two probabilistic Markov chains in parallel–one each for vector and human populations. The model is depicted in Supporting Information S1 Fig. In every cycle, the probability of a vector being infected with the dengue virus, thus moving from susceptible (SV) to infected (IV), is dependent on the number of infected humans in the previous cycle. Likewise the probability of a susceptible human (SH) being infected is dependent on the number of infected vectors. However, susceptible humans are separated into two categories, denoted SH1 and SH2, representing susceptibility to first and second infections respectively, or DH1 and DH2.
New-borns enter the model in state NBH. New-borns have maternal immunity and therefore remain immune from dengue infection for a period of time. As maternal immunity wanes, new-borns enter the susceptible population (SH1). When a human recovers from their first infection (DH1), they enter a state of recovery from first infection (RH1). In state RH1, humans are assumed to have cross-immunity and cannot contract dengue, after which they become susceptible to a second infection (SH2) from a different dengue serotype. If a human contracts dengue for a second time (DH2)–that is, contracts a different dengue serotype–they may progress to severe dengue (SDH). In SDH, there is an elevated probability of mortality (Death). If an individual recovers from DH2 or SDH they are assumed to no longer be susceptible (RH2).
The baseline model therefore allows for the effect of herd immunity (and loss thereof). As individuals move through the states, contracting dengue, the proportion of individuals that are still susceptible decreases and the population as a whole builds up herd immunity. If new individuals entering the model are less likely to contract dengue for reasons other than acquired immunity (for example, the introduction of sustained vector control, as described below), herd immunity wanes over time, at the rate of population replacement, and the probability of transmission increases.
We model symptomatic or apparent cases only. The number of cases of asymptomatic or unapparent cases is potentially much larger [22]. Several studies have demonstrated high levels of asymptomatic infection in endemic countries [23][24], leading to more severe primary symptomatic infections in those individuals [25]. Recent evidence suggest that asymptomatic viraemic individuals can infect Aedes aegypti [26], although there is no evidence as to whether these are then able to infect humans. We have not estimated the effect of asymptomatic infection in our models due to the limited literature available.
The (biting) female vector population is assumed to be proportional to the human population. The average number of female vectors per host is 1.0–6.0, with a sine function of period length 2π and a standard deviation of 20% [27]. In addition to seasonal variation, there is also a probability in every cycle of switching to an outbreak model of the vector population. In defining dengue outbreaks, previous studies have used moving means of the number of human cases and standard deviations of those means, but these differ between countries and even at subnational level [28].
We model outbreaks as a 100–200% increase in the vector population over a period of 2–9 weeks [29]. The increase is relative to the baseline vector population of any given intervention. This choice allows us to more directly model the effect of outbreak response. The Markov model is structured such that the increase in vector population results, through a higher probability of infective bites, in a higher number of human cases. During the burn-in period, the probability of outbreak is set such that frequency of outbreak is once every 3–5 years [30]. This range is consistent with the country reviews, although somewhat conservative with regard to Malaysia [31].
Urbanization is reflected in part by increases in the new-born and susceptible human populations, given by the crude birth rate and the urban population growth rate minus the crude birth rate, respectively. We model the combined effect of both unplanned urbanization and climate change after the burn-in period by allowing for the frequency of outbreak to drop to as low as once every 2 years (that is, frequency follows a triangular distribution with 2 and 5 as minimum and maximum, respectively, and 3 as the most likely value). Alternatively, we could have modelled climate change as an increase in the average number of female vectors per host and/or standard deviation of the sine function. However, this alternative would suggest predictable rather than unpredictable variation over time in the number of vectors and, by extension, a different type of outbreak response than we have defined below.
A list of the baseline model parameters and distributions of the probabilities underpinning the model are listed (with sources) in Table 1. Additional parameters for the intervention models are listed in Table 2, and described here. Whenever possible, we used country-specific parameter values; in practice, these were mainly available for population at risk estimates, unit cost estimates, and basic health statistics such as population, birth rates, death rates (adult and child), life expectancy, and urban growth rates. Of note, we used country-specific probabilities of death among severe dengue cases receiving care. These were based on country-reported data, and varied considerably: from a low of 0.65% in Mexico to a high of 7% in Brazil [32].
In the presence of medical management, the fatality rate of severe dengue cases is assumed to be 0.7–7% (range of best estimates from the six countries). The fatality rate is the same regardless of whether there is vector control or immunization. In the absence of any medical case management, however, the fatality rate for severe dengue is assumed to be 5–20% [40][41].
Recent reviews reveal that many existing vector control technologies have not been robustly evaluated for impact on reducing human dengue cases [57,58]. By technology, we refer to any combination of strategies and intervention tools that have been costed in the economics literature; these include biological (e.g. copepods), chemical (e.g. insecticides) and environmental (e.g. screens) interventions. To our knowledge, only one study has measured the epidemiological effect of such vector control technologies. A randomized controlled trial (RCT) of community mobilization for dengue prevention demonstrated a lower risk of infection with dengue virus in children (relative risk reduction 29.5%, 95% confidence interval 3.8% to 55.3%) and fewer reports of dengue illness (24.7%, 1.8% to 51.2%) [59].
Several trials have, however, evaluated entomological impact. The above-mentioned RCT found a 51.7% (36.2% to 76.1%) reduction in the number of pupae per person (pupae found/number of residents). An earlier systematic review and meta-analysis found that the most effective method (integrated vector control) resulted in decreases of 67–88% (best estimates) in the following indices: the number of containers with larvae per 100 houses (Breteau index), the percentage of water containers positive for larvae or pupae (Container index) and the percentage of houses with water containers containing larvae or pupae (House index) [60].
The value of larval indices has been challenged, however; pupal indices may be more valuable given lower pupal mortality and higher correlation with adult densities [61]. All evidence considered, the systematic review concluded that “dengue vector control is effective in reducing vector populations, particularly when interventions use a community-based, integrated approach, which is tailored to local eco-epidemiological and sociocultural settings and combined with educational programmes to increase knowledge and understanding of best practice”.
We opted to model the effect of vector control as a reduction of the vector population (or, more generally, a reduction of the vector population capable of transmitting the virus). Entomological effects may vary significantly between settings, not least because the choice of vector control technology is country- if not community-specific [62]. We therefore considered two broad categories of vector control technology: a medium-efficacy technology that reduces vector populations by 50–70%; and a high-efficacy technology that reduces vector populations by 70–90%. We acknowledge that such reductions in adult vectors may not have been conclusively demonstrated for long-term use of existing technologies.
These above reductions are applied to periods of outbreak also, such that sustained vector control limits the increase in vector populations during the initial phases of the outbreak, before the outbreak response is deployed. We define sustained vector control as vector control activities undertaken routinely throughout the year, usually monthly, regardless of changes in the number of vectors or human cases. Outbreak response activities, in comparison, are undertaken only in response to spikes in the number of vectors or human cases and only for as long as those spikes last. During outbreaks, increases in vector populations still occur in the presence of vector control, but from a lower baseline level.
In the baseline model of outbreak, the vector population remains high for 2–9 weeks. The effect of outbreak response is modelled by a switch of the vector population to pre-outbreak levels after a lag of 1–2 weeks–the minimum time assumed to be required to detect the increase in vector populations and deploy the outbreak response. Since we are assuming that any sustained vector control programme includes also an outbreak response component, this same effect is modelled for both sustained vector control and outbreak response alone. This assumption is optimistic with regard to outbreak response and therefore conservative with regard to demonstrating the cost-effectiveness of sustained vector control relative to outbreak response.
In the absence of sustained control, vector populations return to baseline values, whereas in the presence of sustained control, vector populations return to a level determined by the vector control technology (e.g. 70–90% below baseline values, in the case of the high efficacy technology). We assume that when sustained vector control and outbreak response are introduced, it is the susceptible (rather than infected) vector population that is immediately reduced. The number of infected vectors decreases more gradually (over their 28-day life span). This assumption is pessimistic with regard to vector control and therefore conservative with regard to demonstrating the cost-effectiveness of vector control versus immunization.
The first-ever dengue vaccine has now been licensed for use in persons aged 9–45 by countries in Asia and Latin America and is under regulatory review by others. Phase III clinical trials of CYD-TDV (Dengvaxia) have measured efficacy over 25 months from the first dose [63][64]. Pooled results (age ≥ 9) suggest efficacy of 38% (3–63%) among seronegatives and 78% (65–86%) among seropositives [16]. Unfortunately, these same trials reveal high rates of hospitalization among those who were vaccinated when seronegative. This result suggests that the immunization strategy should target the seropositive population [55].
In our model, we consider an optimistic scenario in which the immunization strategy is perfectly effective in targeting the susceptible seropositive population (further described below). The effect of immunization is modelled as a probability among individuals susceptible to a second infection (SH2) of moving directly to the state of recovery from a second infection (RH2). The same percentage reductions are applied to both outbreak and non-outbreak cycles because outbreaks are modelled as changes in vector population rather than in serotype distribution.
The entire at risk human population is targeted for vector control. The broadest definition of the at-risk human population includes the urban population of the six countries. We also considered the subset living in urban slum areas. Poor urban communities typically have environmental characteristics that facilitate Aedes spp. breeding, including presence of refuse deposits and containers for water storage [65][66]. We also extracted from DengueMap a list of all locations from which alerts of dengue cases or deaths had been issued in 2013 [67]. We obtained the latitude and longitude coordinates for these locations using the geocode program of the R package mapproj. We then merged this dataset with the g-econ database [33].
We counted populations living within gecon-coded areas satisfying at least one of the following two conditions: 1) population density of more than 250 per km2 (non-rural settings); AND 2) average minimum temperature of no less than 5 degrees Celsius AND an average maximum temperature of no more than 36 degrees Celsius; AND 3) gross domestic product (GDP) per capita (2005 purchasing power parity) of less than US$ 10 000 (excludes areas with a level of development equivalent to a high income country); OR occurrence of a dengue alert within the 1-degree latitude by 1-degree longitude cell. These cut-offs were based on studies identified in a recent systematic review of dengue risk mapping models [68].
In PSA, the susceptible human population was represented by a triangular distribution using g-econ, urban slum and urban populations as the most likely, minimum and maximum values, respectively.
Mathematical models suggest that, to achieve significant reduction in the disease burden, immunization is most effective if it includes only individuals that have been already exposed to at least one dengue virus [69]. Some countries have decided to roll out immunization by targeting specific age groups, such as all 9–14 year olds in the first year and then all (new) 9 year olds in the second year onwards. Targeting strategies may be adapted to local settings. In theory, age or other individual characteristics could be selected to either minimize the number of seronegative people or maximize the number of seropositive people vaccinated.
We assume optimistically that the susceptible seropositive population is so effectively targeted that 0% of the people that receive the vaccine are seronegative (100% are seropositive). In other words, we assume that targeting strategies are perfectly effective in avoiding vaccine-enhanced disease in vaccinated seronegative people. Furthermore, we assume that 70–80% of the population that is seropositive (and still susceptible to second infection) is contained in the population of people that are targeted over a period of 52 weeks. Of these, 80–90% are (again, optimistically) assumed compliant with vaccination (comparable to yellow fever vaccination coverage among children in Brazil).
It bears emphasizing here that the objective of this paper is to assess the cost-effectiveness of vector control in the presence of a (hypothetical) highly targeted and low-cost vaccination strategy using a (non-hypothetical) medium-efficacy vaccine; it is not to assess the cost-effectiveness of the medium efficacy vaccine itself.
The cost of medical management of dengue cases is based on the utilization of general health services only, or the “hotel cost” of hospital bed days and ambulatory visits excluding any laboratory tests or drugs. The hospitalization rate, duration of hospitalization, and number of ambulatory visits are provided in Table 1. We assume a hospitalization rate of 14% for non-severe dengue and 100% for severe dengue (references are provided in Table 1). Hospitalized non-severe cases are hospitalized in primary hospitals, and severe cases are hospitalized in specialist hospitals. Unit costs (best estimates and standard errors) were obtained using data and methods from WHO-CHOICE [70]. These unit cost estimates are summarized in Table 3. All symptomatic cases were assumed to receive medical management in all scenarios, regardless of whether there was sustained vector control, outbreak response and/or immunization.
We conducted a search of the literature on the cost of sustained vector control interventions and identified eight studies with primary data, from 12 countries, considering different biological, chemical, and environmental measures. A subsequent systematic review revealed no additional studies [11]. We extracted data on costs as well as populations or households covered. Costs were converted to per capita terms and inflated to 2013 US$ using GDP deflators. These unit costs were then modelled in a multivariate log-log regression on population covered and GDP per capita. More than 50% of the variation in unit cost between the studies was explained by these two variables alone, driven by a strongly negative relationship with population. In addition to economies of scale, it is likely that higher cost biological and environmental measures were only implemented at smaller scale, in targeted communities. A plot of the data and regression model results are available in Supporting Information S2 Fig and Supporting Information S1 Table.
We calculated means and standard errors for the predictions for the six countries considered in this study, using their urban populations and GDP per capita. For each country, 1000 values of (log) unit cost were drawn from a normal distribution, and then exponentiated. These unit cost estimates are summarized in Table 3. The best estimates are in the range of about 2013 US$ 0.04–0.05 per person per month–similar to the cost of larviciding and/or adulticiding programs described in studies from Brazil, Cambodia, Venezuela, and Thailand. The confidence intervals are wide, however, with highs of up to 2013 US$ 0.06–0.09 allowing for smaller scale biological and environmental measures. These unit costs are similar to those described in studies from Guatemala, Kenya, Mexico, Myanmar, and the Philippines, published prior to 2013. No costing studies have been undertaken for technologies under development using Release of Insects with Dominant Lethality (RIDL) or Wolbachia symbiont infection of vectors.
The cost of outbreak response was taken from a study from Panama [53]. That study found that the cost of vector control during an outbreak, including larviciding and adulticiding, was about 2005 US$ 0.035 per person per week. The Panama costs are conservative relative to a more comprehensive costing of outbreak response in Cuba, including health education and/or replacement of defective water tanks [71][72]. Similarly, the Panama costs are conservative relative to environmental and/or chemical interventions triggered by case reports in Malaysia and Thailand [73][74]. We considered adjustments for our six countries using the proportion of labour in costs (71%) and GDP per capita relative to that of Panama in the year of the study. Minimum, most likely, and maximum values were obtained considering adjustments for GDP per capita only, or GDP per capita and the labour proportion, or no adjustment at all.
We assumed that a sustained vector control programme would implement extraordinary outbreak response interventions in the midst of an outbreak; therefore, during an outbreak, the total cost of sustained vector control includes the cost of both the sustained vector control and outbreak response interventions. To the cost of both sustained vector control and outbreak response alone we also added the cost of sustained surveillance, based on a study from Brazil, at a cost of 2013 US$ 0.014 per person per week. Again we generated country-specific minimum, most likely, and maximum values using the proportion of labour in costs (39%) and relative GDP per capita [56].
Since the cost of a future vaccine is unknown, the cost per person vaccinated was assumed to be about 2013 US$ 20. This unit cost is assumed to include the cost of screening for seropositivity, or whatever the cost of the perfectly effective targeting strategy, as well as administration of the vaccine itself. It is purposefully optimistic. An earlier study considered a range of US$ 10–300 per unit for the cost of vaccine production alone [12]. In most countries, US$ 20 is less than the cost of the 1–3 outpatient visits that would be required for administration alone (Table 3). It is also considerably lower than the maximum of the median willingness-to-pay results (US$ 70) from Vietnam, Thailand, and Colombia [75].
The number of cases of symptomatic dengue in the baseline model (medical case management alone) is depicted in Fig 1, for each of the six countries. The waves represent seasonal variation. Outbreaks are not visible in the best estimate or uncertainty intervals obtained from the 1000 simulations but are visible in individual simulations, only one of which is depicted for illustration. We compare these baseline model results to recent published estimates to ensure that we are not overstating the potential effects of intervention in terms of cases or DALYs averted [22]. For most countries our estimates in year one are towards the lower bound of those published estimates, and for all countries the uncertainty intervals overlap. Our estimates trend slightly upward over time. Note that the published estimates take into account current (but variable) efforts to control dengue whereas the estimates presented in Fig 1 are for our baseline model (medical case management alone).
The effect of sustained vector control on the total number of dengue cases (including severe cases) is depicted over time in Fig 2. Different sustained vector control technologies (medium and high efficacy) are considered, and compared to the baseline (medical case management only). Vector control technologies of low efficacy (<50% reduction in vector populations) have limited longer-term impact on transmission, and are therefore not depicted.
Initially, medium- and high-efficacy vector control technologies result in a significant decrease in the number of cases. The combination of vector control and acquired immunity push the basic reproduction number to below one. In time, the number of susceptible people waxes and herd immunity wanes. With high-efficacy vector control, the period of low transmission lasts from four to seven years, varying between countries. Differences between countries are explained by differences in the country-specific parameters, namely birth rates, death rates, and urban growth rates, which together determine the rate at which susceptible people are introduced into the model.
The number of cases of severe dengue follows a similar pattern, but with a somewhat longer-term effect of high efficacy vector control, as can be verified in Supporting Information S3 Fig.
The effect of highly targeted immunization using a medium-efficacy vaccine on the number of dengue cases is depicted over time in Fig 3, alone and in combination with sustained vector control (high efficacy). The total number of dengue cases is largely unaffected by immunization alone, because it is targeted at seropositives only and the number of severe dengue cases is small relative to the total. Only the combination of sustained vector control and immunization maintains the number of cases at very low levels, for between four and nine years depending on the country. Again, differences between countries can be explained by differences in the country-specific epidemiological parameters.
The number of cases of severe dengue, however, exhibits a different trend, as can be seen in Supporting Information S4 Fig. Here, because of our optimistic assumptions about how effectively an immunization strategy could target the seropositive population, we have a large and sustained effect of immunization, alone and in combination with sustained vector control (high efficacy). This figure should be interpreted as confirmation that our assumptions about the immunization strategy have been optimistic.
Table 4 summarizes the average annual number of DALYs in the period 2015–2030 under different intervention scenarios, compared to WHO Global Health Estimates for the year 2012. Our model suggests that medical case management alone would result in an average of 22.3–232.3 thousand DALYs per year (range of best estimates across the six countries). The uncertainty intervals on these estimates lie just above WHO estimates for the year 2012, with the exception of the Philippines (for which our uncertainty interval overlaps with the WHO estimate). Life expectancy is lower in the Philippines (69 years), than for any of the other countries (74–78 years). Note that WHO Global Health Estimates reflect variable levels of intervention across countries; they are also based on highest observed life expectancy globally.
Outbreak response has a negligible impact on the average burden, given that outbreaks occur relatively infrequently on average and that the response is triggered only with a lag after the increase in the vector population. The introduction of sustained vector control (medium efficacy) reduces the burden to an average of 18.5–167.0 thousand DALYs per year. Sustained vector control (high efficacy) reduces the burden to an average of 12.0–89.6 thousand DALYs per year.
Average annual costs over the 15-year period are reported in Table 5. At current prices of existing technologies, the total cost of sustained vector control (using a high efficacy technology) including outbreak response and medical case management (2013 $US 58.0–377.6 million, range of best estimates across the six countries) is comparable to what would have to be spent on outbreak response and medical case management (2013 $US 57.7–368.7 million). This result is driven by differences in the cost of treating the cases that would not be averted by outbreak response, given a minimum one week lag between the increase in vector populations and deployment of the outbreak response.
In Table 5, the best estimates of cost are expressed also as a percentage of government health expenditure (GHE) in 2013. Sustained vector control (high efficacy) including medical case management would cost 0.4–1.2% of GHE in five of the six countries. In the Philippines, where GHE in 2013 was much lower in per capita terms than for the other five countries, it would cost 4.0% of GHE.
Average cost-effective ratios (ACERs) and incremental cost-effectiveness ratios (ICERs) are presented for each country in Tables 6 and 7, considering high- and medium-efficacy vector control technologies, respectively.
If sustained vector control is effective in reducing mosquito populations by 70–90% (high-efficacy, Table 6), the average cost per DALY averted would be US$ 679–1331 (range of best estimates across the six countries) relative to the null scenario. The combination of high-efficacy sustained vector control and a highly targeted and low-cost immunization strategy using a medium-efficacy vaccine dominates all other interventions except immunization alone. However, immunization alone averts far fewer DALYs. In five of the six countries, the combination of vector control and immunization is very cost-effective, with an ICER well below one times GDP per capita. In the Philippines, it is cost-effective at a threshold just above one times GDP per capita. Recall that the Philippines is the country with the lowest life expectancy among the six; it is also the country with lowest GDP per capita, and therefore the country with the (presumed) lowest willingness-to-pay (WTP).
If vector control is effective in reducing mosquito populations by only 50–70% (medium-efficacy, Table 7), the average cost per DALY averted would be US$ 808–1907 (range of best estimates across the six countries) relative to the null scenario. Again, the combination of medium-efficacy sustained vector control and a highly targeted and low-cost immunization strategy using a medium-efficacy vaccine dominates all other interventions except immunization alone. However, again, immunization alone averts far fewer DALYs. In four of the six countries, the combination of medium-efficacy vector control and immunization is very cost-effective (the ICER is below one times GDP per capita). In Mexico and the Philippines it is cost-effective at thresholds between two and three times GDP per capita. Mexico is the country with the lowest reported death rate among severe dengue cases receiving medical management.
Uncertainty around cost-effectiveness is reflected in the cost-effectiveness acceptability curves of Figs 4 and 5. The combination of high-efficacy sustained vector control and a highly targeted and low-cost immunization strategy using a medium efficacy vaccine (Fig 4) has the highest probability of being most cost-effective at WTP thresholds as low as one quarter of GDP per capita (per DALY averted). At a WTP threshold of one times GDP per capita, the probability that this is the most cost-effective strategy exceeds 85% in four of the six countries.
We estimated the cost and cost-effectiveness of sustained vector control in six high-burden countries. Our model suggests that, at current prices, the cost of sustained vector control and medical case management is comparable to what would otherwise have to be spent on outbreak response and medical case management. In December 2015, in its decision to approve Dengvaxia, Mexico reported that it was spending about 2.5% of its health budget on medical treatment alone [77]. Many countries have abandoned or lack effective surveillance systems to respond rapidly enough to outbreaks to make much of a dent in the number of cases requiring medical management [78].
Our results on cost are nonetheless conservative relative to an earlier review and qualitative synthesis of the evidence from single settings, which found that the cost of outbreak response exceeds that of sustained vector control [79]. We considered only direct medical costs and control programme costs–we did not consider direct non-medical costs faced by patients during their care (e.g. food, transportation) nor any productivity losses (time spent away from work) of the patients or their caretakers.
Our estimates of the cost per DALY averted by sustained vector control (related to doing nothing) are lower than that of the DCP2, but higher than those from studies of single settings. In those studies, cost-effectiveness ratios ranged from 2005 US$ 227 (2013 US$ 344) per DALY averted by larval control in Cambodia to 2009 US$ 615–1267 (2013 US$ 802–1652) per DALY averted by adult mosquito control in Brazil [9][10]. Taking a broader societal perspective including productivity losses, the Cambodia programme cost only 2005 US$ 37 (2013 US$ 56) per DALY averted. Again, we did not consider these productivity losses–their measurement remains controversial, although few would argue that they are zero.
We have not (nor to our knowledge has any earlier study) considered the additional benefits of vector control targeted at dengue for the prevention of other diseases, such as Chikungunya, yellow fever and Zika fever, transmitted by the same vectors and (for all but yellow fever) lacking effective vaccines and specific treatment. These additional benefits of sustained vector control would increase its cost-effectiveness. Even with conservative assumptions, our results suggest that sustained vector control can be highly cost-effective.
Our model suggests that the introduction of highly targeted and low-cost immunization strategy using a medium-efficacy vaccine does not alter the conclusion that sustained vector control can be highly cost-effective. On the contrary, vector control may complement a medium-efficacy vaccine, or a vaccine that is highly effective but against only secondary infections or only one of the four dengue serotype, or whose production is highly constrained. In these instances, sustained vector control may compensate to some extent for lower levels of coverage by immunization.
There are as yet no costing studies for dengue vaccine delivery, nor even a known price for the vaccine itself [77][80][81]. Indeed, there are still many unknowns about immunization for dengue. Overall, our assumptions for immunization can be considered conservative from the perspective of the cost-effectiveness of sustained vector control. That is, we have been optimistic in our assumptions about immunization with regard to both effects and costs, in order to have the most conservative estimate of the cost-effectiveness of vector control in the presence of immunization. We have been particularly optimistic in assuming that the targeting strategies for immunization are perfectly effective in avoiding seronegative people. This analysis should not, therefore, be used to draw conclusions about the cost-effectiveness of the current vaccine. When new information or new vaccines become available, these model parameters should be adapted.
We accounted for many of the known sources of uncertainty within a probabilistic framework. In spite of considerable uncertainty, sustained vector control emerged with a high probability of cost-effectiveness. Some sources of uncertainty, however, could not be accounted for and remain as more serious limitations to our study.
First, there is remaining uncertainty around particular parameter values. The serotype immunity variable (serotype composition) is assumed to be constant over time. The probability that dengue fever following a second infection develops into severe dengue is based on clinical paediatric data from four hospitals [39]. While the probability of death among severe dengue cases receiving medical management was based on country-reported data, the considerable disparity across countries raises questions. The availability of better data for these parameters would improve the precision of our model.
Second, there is remaining uncertainty about the structure of the model itself, namely with respect to outbreaks, which are modelled as increases in the vector population. Vector indices have been correlated to increases in dengue cases during outbreaks, but the strength of evidence is limited by a lack of well-controlled studies [82][83]. Other studies have considered spatial, meteorological, epidemiological, and entomological factors, and dengue serotype [84]. Although this study could benefit from a more sophisticated outbreak model, uncertainty remains in the choice of variables and availability of data.
Another limitation of the model structure is that we have not modelled heterogeneity in transmission. The probability of contracting dengue assumes an even distribution of vectors to humans throughout the susceptible population. However, a study from Armenia and Colombia found that 95% of Ae aegytpi pupae were concentrated in only 5% of houses [85]. More detailed risk mapping could improve our analysis and lead to better targeting and enhanced cost-effectiveness of vector control.
Third, we have made no assumptions about the relative costs of the different vector control technologies–in fact, we have assumed the same unit cost for both medium- and high-efficacy technologies. Obviously, this analysis does not help us to choose amongst available vector control technologies. Further research is needed to understand the drivers of cost and effect across current and future vector control technologies. Indeed, it is unclear whether reductions in excess of 70–90% can be sustained using existing technologies; more research on the long-term impact of both existing and upcoming technologies is needed.
Nonetheless, the result on the cost-effectiveness of high-efficacy vector control is fairly robust to higher unit costs. Comparing ICERs to WTP, we find that in most settings, the cost of sustained vector control using high-efficacy technologies could be considerably higher than that assumed in our model (about US$ 0.05 per capita per month) before it would no longer be considered cost-effective. At a WTP threshold of three times GDP per capita, the cost could be as much as 2.9–13.9 times higher without affecting our conclusions for any of the settings considered. At a WTP threshold of one times GDP per capita, the cost could be as much as 2.2–4.6 times higher without affecting our conclusions for Brazil, Colombia, Malaysia, and Thailand.
Fourth, our model does not reflect uncertainty around the impact of resistance to insecticides in a scenario of sustained vector control. The data on this are limited, as a result of poor routine resistance monitoring [86]. New vector control technologies, such as genetic modification or symbiont infection of vectors, once developed, may help in mitigating the risk of resistance [87]. Our results are favourable to sustained vector control in general (not to any one technology in particular) and should not discourage the development of new technologies with demonstrated efficacy and safety. Indeed, our results apply to any technology that reduces the number of vectors able to transmit disease, rather than the number of vectors per se.
Finally, our results may not be generalizable to other settings, for which we do not have good epidemiological or cost data. Low-income countries, especially those in Africa, are arguably more vulnerable and less prepared for the effects of unplanned urbanization and climate change, including the spread of vector-borne diseases [88]. These same countries have the poorest quality data on dengue, starting with the frequent misclassification of dengue as malaria. A regional study of the cost and cost-effectiveness of sustained vector control in low-income countries of Africa is needed.
This paper focussed on the cost-effectiveness of sustained vector control in six middle-income endemic countries representing 15% of the estimated global burden of dengue. We have shown that sustained vector control will be cost-effective in most of these countries if it succeeds in reducing mosquito populations by more than 50%. Importantly, we show that the introduction of a highly targeted and low-cost immunization strategy using a medium-efficacy vaccine does not weaken the investment case for sustained vector control. Middle-income endemic countries should proceed with mapping the populations to be covered by sustained vector control.
|
10.1371/journal.pcbi.0030119 | Protein–Protein Interaction Hotspots Carved into Sequences | Protein–protein interactions, a key to almost any biological process, are mediated by molecular mechanisms that are not entirely clear. The study of these mechanisms often focuses on all residues at protein–protein interfaces. However, only a small subset of all interface residues is actually essential for recognition or binding. Commonly referred to as “hotspots,” these essential residues are defined as residues that impede protein–protein interactions if mutated. While no in silico tool identifies hotspots in unbound chains, numerous prediction methods were designed to identify all the residues in a protein that are likely to be a part of protein–protein interfaces. These methods typically identify successfully only a small fraction of all interface residues. Here, we analyzed the hypothesis that the two subsets correspond (i.e., that in silico methods may predict few residues because they preferentially predict hotspots). We demonstrate that this is indeed the case and that we can therefore predict directly from the sequence of a single protein which residues are interaction hotspots (without knowledge of the interaction partner). Our results suggested that most protein complexes are stabilized by similar basic principles. The ability to accurately and efficiently identify hotspots from sequence enables the annotation and analysis of protein–protein interaction hotspots in entire organisms and thus may benefit function prediction and drug development. The server for prediction is available at http://www.rostlab.org/services/isis.
| Interactions between proteins underlie all biological processes. Hence, to fully understand or to control biological processes we need to unravel the principles of protein interactions. The quest for these principles has focused predominantly on the entire interfaces between two interacting proteins. However, it has been shown that only few of the interface residues are essential for the recognition and binding to other proteins. The identification of these residues, commonly referred to as binding “hotspots,” is a first step toward understanding the function of proteins and studying their interactions. Experimentally, hotspots could be identified by mutating single residues—an expensive and laborious procedure that is not applicable on a large scale. Here, we show that it is possible to identify protein interaction hotspots computationally on a large scale based on the amino acid sequence of a single protein, without requiring the knowledge of its interaction partner. Our results suggest that most protein complexes are stabilized by similar basic principles. The ability to accurately and efficiently identify hotspots from sequence enables the annotation and analysis of protein–protein interaction hotspots in an entire organism and thus may benefit function prediction and drug development.
| Interactions of proteins are at the heart of almost every biological process. Thus, the understanding of biological mechanisms requires the knowledge of protein–protein interactions and the molecular principles that underlie them. Large-scale studies unravel networks of protein–protein interactions in cells and identify interacting pairs of proteins [1–5]. However, to fully understand these interactions, and to manipulate them, we need to identify the residues that account for binding of the proteins and stabilizing the complexes. It has been postulated that only very few of the residues in protein–protein interfaces are absolutely essential for the interaction (in a typical 1,200- to 2,000-Å2 interface, less than 5% of interface residues contribute more than 2 kcal/mol to binding. In small interfaces, this can mean as few as one amino acid on each protein) [6]. These residues may be instrumental in understanding the interaction and could be desired drug targets [7].
The ability to predict hotspots on a large scale may assist in identifying, analyzing, and comparing binding sites for drugs. Given a detailed 3-D structure of a complex, the residues crucial for binding are often identifiable. The Hendrickson lab, for instance, identified the most essential binding residues from their 3-D structure of HIV glycoprotein (gp120) and CD4 receptor [8]. Unfortunately, 3-D structures are available for less than 1% of all known pairs of interacting proteins. In the absence of 3-D structures, the most conclusive way to probe the importance of particular residues for interaction is to experimentally mutate them, typically to alanine, and measure the effect of this substitution on the interaction [9,10]. Many experiments have demonstrated that most interface residues could be mutated without affecting the affinity of the protein to its partners [11,12]. Those few residues that, upon mutation, change the affinity are often assumed to be the most essential for the interaction and are deemed “hotspots” [6]. The limited overlap between interface residues and hotspots is demonstrated in Figure 1, which depicts the complex of the human growth hormone and its receptor [13]. In the bound state (Figure 1A), a large patch on the surface of the receptor is buried in the interface. There are 31 residues on the receptor that are in physical contact with a hormone (Figure 1B). However, mutation experiments indicate that only six of these residues are energetically crucial for the interaction (Figure 1B).
The ways to identify hotspots have been subject to theoretical debates. It has been pointed out that given the structural and physicochemical complexity of proteins, the physicochemical features of a protein are not a simple sum of the features of its individual residues [14]. Therefore, single mutations may not always convey accurate assessments of the contribution of a residue to the interaction [15,16]. The theoretical validity of this argument notwithstanding, alanine scans have become the most widely used tool for identifying binding sites. While single mutations may not be tantamount to isolating the contribution of a single residue to the interaction, they are still considered a good approximation. Here, we adopt the following operational definition: if a mutation of a residue in a protein–protein interface changes the binding energy of the protein to its binding partner substantially (ΔΔG > 2.5 kcal/mol), then this residue is a hotspot residue.
To the best of our knowledge, there is currently no method that was designed to identify hotspots from sequence. However, many methods attempt to use sequence or structure to identify which residues are located in the interface between proteins [17–32]. Many of the methods that identify residues in protein–protein interfaces reach impressive levels of positive accuracy (residues correctly predicted to be in protein–protein interfaces as a fraction of all residues predicted to be in protein–protein interfaces; often also referred to as selectivity, or precision; Equation 1). However, their coverage (residues correctly predicted in interfaces as percentage of observed interface residues; often also referred to as sensitivity, or recall; Equation 2) remains fairly low. In other words, although these methods attempt to identify all interface residues (all the residues that are colored blue or red in Figure 1B), they capture only a small fraction of them (e.g., only the green residues in Figure 1C). We hypothesized that the reason for the low coverage of many prediction methods might be that the residues that are missed are more similar to the general population of surface residues than to the essential residues (i.e., they are inconsequential for the interaction). Therefore, a machine-learning algorithm trained on all protein–protein interface residues may learn to disregard the non-hotspot residues as noise, and identify only hotspot residues as the signal to be learned.
To test this hypothesis, we applied ISIS, a prediction method developed for the prediction of all interface residues [28], to the task of predicting only hotspots. ISIS was never trained on hotspots (Methods). Instead, we trained on all interface residues found in Protein Data Bank (PDB) complexes (i.e., all interface residues were labeled “positive,” and all other residues were labeled “negative”). The features on which ISIS was trained included the sequence environment of each residue (four residues on each side), the evolutionary profile of all nine residues in that window, the predicted solvent accessibility of the residue and the solvent accessibility of its immediate sequence environment (one residue on each side), the predicted secondary structure state of the residue and its immediate sequence environment (one residue on each side), and a conservation score for each residue. Like several other methods mentioned above, ISIS predicts residues in protein–protein interfaces very accurately (∼90% accuracy). However, at this high level of accuracy, ISIS identifies fewer than 5% of the residues that were experimentally mapped to the interface.
The novelty here is that we applied a generic interface-prediction method to the specific task of identifying only the residues that are crucial for stabilizing the interactions (i.e., the hotspots). The results demonstrated a surprising overlap between two principally unrelated datasets, namely on the one hand the subset of residues that was identified by experimental alanine mutations as hotspots, and on the other hand the subset of residues predicted by ISIS to be protein–protein interface residues. We obtained a large dataset of hotspots that were determined experimentally through alanine scans (Methods) and assessed the performance of ISIS on these hotspots. The results confirmed our hypothesis that the residues predicted by the machine-learning method are, in fact, the hotspots. Analysis of the results indicated that accurate predictions of hotspots required the combination of sequence features, evolutionary information, and predicted structural features; all this information was generated from the amino acid sequence, suggesting that the commonalities of hotspots have been imprinted clearly onto amino acid sequences in the course of evolution.
Using 296 point mutations from 30 proteins, we compared the residues predicted by ISIS with the ones experimentally identified to be hotspots (Methods). We first analyzed the results for two representative examples. Then, we assessed the performance in predicting hotspots based on the analysis of the entire dataset of 296 mutations. Note that although the 3-D structures for most of these proteins were experimentally known, ISIS predicted interface residues from sequence alone. At no stage of the predictions did we use the experimentally determined structure. The only way in which we used 3-D information was to visualize our results, as we mapped the predictions to the experimentally determined structure (Figure 2).
One of the most comprehensive alanine scans of all the complexes with known 3-D structures is that between the CD4 receptor and the HIV glycoprotein gp120. This interaction involves backbone interactions, mainly on the gp120 side. However, we focused our analysis on the human CD4 receptor. Ashkenazi et al. [33] sequentially mutated many residues in the V1 domain of the CD4 receptor and studied the effect of each substitution on the binding affinity between CD4 and the HIV gp120 protein. Using a set of specific antibodies, they also assessed which mutation had no effect on the structure. They identified 25 positions within a stretch of 94 residues on CD4 that upon substitution changed the affinity of CD4 substantially, without strongly altering the conformation of the protein. Within the same 94-residue segment (Figure 2A), we predicted 30 residues as interface residues; 19 of these were found experimentally to have a strong effect on binding. Of the six residues that ISIS missed, four were next to predicted interface residues. Five of the predictions that were not confirmed experimentally were residues that were not mutated in the study. Our method uses predicted structural features (solvent accessibility and secondary structure). Hence, its performance depends to some extent on the accuracy of these predictions. If we have a 3-D structure of the unbound chain, we can improve accuracy and coverage by using the experimental rather than the predicted features. For example, when we used the unbound structure of CD4 as input for ISIS, we found a few additional residues that were not identified from sequence alone. The two residues that scored highest (i.e., about which we were most confident that they participate in binding) were Arg59 and Phe43. The high-resolution structure of the complex between gp120 and CD4 complex [8] revealed two residues as the most important contacts between these two proteins: Arg59 and Phe43.
For a variety of reasons, membrane proteins are a particularly popular target for alanine scans. One such alanine scan is available for the shaker voltage-gated K+ channel [34]. Within a region of 29 consecutive residues that have been scanned, eight have a significant effect on the affinity of the channel to its inhibitors agitoxin2 and charybdotoxin. We used this region as input to our method, ignoring any available structural information, and predicted 13 residues (Figure 2B). Seven of the eight residues that were found experimentally were predicted by ISIS; the only residue that was missed is buried in the structure and hence is likely to affect the interaction indirectly through a conformational change. Of the six residues in our prediction that did not coincide with the residues implicated as important by the alanine scanning, five coincided with positions that were found to have significant although less dramatic effects on binding [34].
Within our set of alanine scans, almost all binding residues predicted by ISIS were found experimentally to have significant effect on binding (Figure 2C). Furthermore, more than 90% of the negative predictions (predicted not to be involved in protein–protein interactions) were confirmed experimentally to have no effect on the energy of binding. These results were particularly surprising in light of the fact that ISIS never explicitly evaluated any energetic parameters. Using different confidence thresholds (i.e., picking a different point on the curve in Figure 2C), it is possible to increase accuracy (true positives/all positives) at the expense of coverage (true positives/predicted positives). Note that the results for the two examples (Figure 2A and 2B) discussed in detail are similar to the performance of ISIS on the entire dataset of 296 mutations.
We used ISIS to represent methods that predict interface residues at high accuracy and low coverage. The results suggested that the system of neural networks that underlies ISIS learned to identify the hotspots, despite the fact that they were only a small subset of the samples that were labeled as interaction residues. The system effectively disregarded most of the residues observed in interface (i.e., the pupil [neural network] clearly ignored the teacher [labeled data]). We found that the residues ignored were mostly non-hotspot residues. These results indicated that the biophysical common denominators of hotspots are so pronounced that the neural networks could identify them without specific labeling in the training phase.
What are these features that are common to hotspots? Unfortunately, we cannot simply list a few rules or features to describe these commonalities. The neural networks identified a set of complex nonlinear correlations between the input features we used and hotspot residues. It is impossible to translate the subtle and complex dependencies that were identified by the neural networks into simple explanations, or a set of rules, in English. However, it is possible to infer which features are more or less relevant. To that end, we trained several systems using different combinations of input features. Neural networks that were trained only on the sequence environment of interface residues performed only slightly (although significantly) better than random (unpublished data). Adding evolutionary information significantly improved performance on both interface residues and on hotspots. This result was somewhat surprising given that the conservation of predicted hotspots was only marginally different from that of all other residues (Figure 3). Conversely, predicted non-hotspot residues were only marginally less conserved than the background. In other words, although the overall difference in conservation was marginal, the addition of this information to the neural network input substantially improved performance. Apparently, the neural networks have learned to distinguish between conservation that is indicative of hotspots, and conservation that is not. Strikingly, they did so without being trained on hotspots. This underscores why linear combinations of input features did not suffice and why the extraction of singly important commonalities would at best be misleading.
The analysis of the contribution of each feature suggested that successful predictions of hotspots required the combination of all features. However, even when some of these features were not available, ISIS still could provide accurate predictions (e.g., 15% of the proteins found less than ten homologues in today's databases). For these proteins, the success in predicting hotspots was lower, but still significantly higher than random (at 70% positive accuracy, >10% of the experimentally determined hotspots were identified compared with about 70%/20% for all proteins; Figure 2).
We did not benchmark the ability of prediction methods other than ISIS to predict hotspots. The main reason was that no existing method (including ISIS) was designed to predict hotspots. The ability of ISIS to identify hotspots is an unintended consequence of the power of neural networks. Therefore, when comparing ISIS with other methods, one should remember that this comparison does not benchmark these methods in the task for which they were originally developed. Still, the question remains of whether or not any method designed to predict interface residues could predict hotspots at levels of accuracy as high as the ones we reported for ISIS. To address this question, we applied a few representative interface prediction methods to the task of predicting hotspots. In particular, we chose methods that rely on a different input feature. Analysis of the results indicated that methods that did not rely on a combination of physicochemical features, evolutionary conservation, and structural features failed to identify hotspots.
We applied several prediction methods that were designed to identify interface residues to the task of predicting hotspots. To eschew obfuscation: our aim was not to benchmark methods not designed to identify hotspots. Instead, we applied these methods to narrow down the features needed to successfully predict hotspots.
The evolutionary trace (ET) method [35] correlates evolutionary importance of residues with their importance for function. We used ET to represent the approach that relies predominantly on evolutionary conservation. Gallet et al. [22] have attempted to predict interaction sites from simple biophysical features; the method computes the hydrophobic moment [36] around each residue based on its sequence environment to determine whether this residue could be a binding site. ProMate [26] extracts its input from the 3-D structure of an unbound protein; we used it to represent methods that rely on experimentally determined 3-D structures. We also included another method that predicts interfaces exclusively using amino acid information (and no aspects of predicted structure or evolutionary profiles) [29]. We arbitrarily chose the operating point at which the coverage of hotspots was 15% (Methods) and checked the accuracy of each method for this coverage (Figure 4). ISIS and ProMate, the two methods that were most successful, use physicochemical features, evolution, and structural features. ISIS is the only sequence-based method, and the structural features it uses are based on predictions. ProMate, which relies on the 3-D structure, performed even better. The conclusion of this analysis is that no single feature suffices to characterize hotspots. Rather, it takes a complex combination of the aforementioned features that defines a residue as a hotspot.
It is apparent that the neural networks identified some common denominators between hotspots that distinguish them from other interface residues. This question is hard to address given our current gold standard (namely the dataset of experimental alanine scans). The number of features we use for the prediction (189) is greater than the number of positive data points in our set of alanine scans. To determine to what extent each input feature differentiates between hotspots and other interface residues, we need a substantially larger dataset of hotspots and non-hotspot residues. This could be achieved if we assume that ISIS indeed identifies hotspots. Thus, by running ISIS on a large dataset of interface residues, we can create a large dataset of predicted hotspots and a large dataset of interface residues that are predicted not to be hotspots. Then, we can use these large datasets to analyze the characteristics of hotspots versus the characteristics of other interface residues. We did this using the large dataset of interface residues that was used as a test set for training ISIS. On this dataset we compared the residues that were classified by ISIS as positive (i.e., hotspots) with those that are annotated experimentally as interface residues but are classified by ISIS as negatives. Table 1 is based on the multiple sequence alignment of each protein in this dataset. For each interface residue, it shows the average occupancy of its position by each type of amino acid. We also present the average occupancy of each residue in the alignment for experimentally determined hotspots (through alanine scan). These values are presented in parentheses, as the data that underlie them are sparse (only 100 positions). Note that for some amino acids there are significant differences between hotspot and non-hotspot interface residues, while for others there are no substantial differences. Table 1 also presents the p-value for the difference based on a t-test. Note, for example, the 400% overrepresentation of arginine in predicted hotspots (and the extremely low p-value) with reference to other interface residues. However, the percentages of lysine are virtually the same for both categories. Thus, it is not simple considerations of hydrophobicity that characterize hotspots. Four aliphatic residues are depleted in hotspots (A, V, I, and L), while amide side chains are overrepresented (N and Q). However, the role of aromatics is unclear since tyrosine is enriched in hotspots, phenylalanine is depleted, and tryptophan has similar propensities across the interface. The experimental values (shown in parentheses) are very close to the values obtained for the predicted hotspots, supporting our assumption that ISIS identifies hotspots. However, the limited amount of experimental data limits our ability to elaborate on this comparison. We also compared the conservation and the structural features of both groups. As shown in Figure 3, there were hardly any differences in conservation. However, the most striking differences were found between structural features (Table 2). The secondary structure state of 39% of the non-hotspot interface residues was a loop. In the predicted hotspots, on the other hand, 57% of the residues were in a loop state. In both categories, the rest of the residues were divided roughly equally between helices and strands. Again, there is a striking agreement between the properties of predicted hotspots and the properties of experimental hotspots, despite the fact that ISIS was trained on all interface residues. Predicted hotspots were also much more accessible to solvent than other interface residues.
Several studies suggested that hotspots have certain structural characteristics that differentiate them from other residues [37,38]. The Baker lab has shown that given a 3-D structure of a protein complex, it is possible to predict the results of alanine scans specifically and accurately [39,40]. This indicates that alanine scans indeed capture some genuine physicochemical commonalities of interaction hotspots that could be identified by a general method that is applicable to all protein complexes. The in silico alanine scanning is based on analysis of the 3-D structure of the interface between two proteins. Thus, it requires a high-resolution structure of the protein complex, while ISIS needs only sequence of a single chain regardless of its binding partner. On the other hand, in silico alanine scanning produces numerical prediction of the ΔΔG, While ISIS produces a binary prediction (hotspot/non-hotspot). We compared our predictions to those of the in silico alanine scanning by translating their numerical predictions to binary ones according to cutoffs defined above. Of 55 experimental mutations with ΔΔG > 2.5, in silico alanine scanning identified 36 (66%) residues as hotspots. At this coverage, ISIS reached accuracy of about 60% while the in silico alanine scanning reached accuracy of greater than 75%. Scaled to an accuracy of 80%, ISIS identified 18 of these mutations (33%). Thus, for similar levels of positive accuracy, the coverage of ISIS is roughly half that of the in silico alanine scanning. Obviously, when structures of the complex are available, the in silico alanine scan is a powerful tool for identifying hotspots. However, when only the sequence is available, ISIS can provide accurate predictions for a substantial fraction of the hotspots. Our results indicate that some hotspots can be predicted accurately not only without relaying the 3-D structure of the complex but even without the 3-D structure of the unbound proteins. Furthermore, our predictions did not require knowledge of the binding partner. Analyzing a single protein using ISIS typically requires a few minutes. Thus, ISIS may allow large-scale analysis of hotspots at a relatively small CPU cost.
We used the ASEdb database of experimental alanine scans [12], which lists residues that were mutated to alanine and the effect (in terms of ΔΔG) this mutation had on the interaction between two proteins. We checked the correlation between the predictions and the residues that were shown experimentally to substantially affect the affinity of the proteins in a complex to each other. In order to reduce the number of cases in which the effect of the mutation on binding was not due to a change in the interface (e.g., the cases in which the mutation destabilized the structure), we considered only exposed residues in proteins of known structure. Thus our test set included 80 protein chains with hundreds of experimental substitutions. From among these, we analyzed the mutations that substantially changed the binding energy (ΔΔG > 2.5 kcal/mol), and those that had no effect (ΔΔG = 0). Altogether, we attempted to predict the experimental effect of 296 substitutions. The predictions were performed using ISIS [28]. ISIS can take as input either sequence or the coordinate of 3-D structure of unbound chains (the results are more accurate when using known 3-D structures). However, for all values reported here, we ran ISIS from sequence alone.
The accuracy and coverage of ISIS were measured using ratios derived from TP (true positives), defined as the number of residues predicted by ISIS (below) to be in a protein–protein interface and observed to be in a hotspot (i.e., was found to have an extreme effect on binding; ΔΔG > 2.5kcal/mol); FP (false positives), defined as the number of residues predicted in protein–protein interfaces, were found however, upon mutation, to have no effect on binding (ΔΔG = 0); and FN (false negatives; i.e., the number of residues predicted not to be in a protein–protein interface that were observed to have a strong effect on binding [ΔΔG > 2.5 kcal/mol]). We used the following equations:
ISIS is a knowledge-based method we developed to identify interface residues from sequence [28]. It is based on a system of neural networks and uses as input the sequence environment of each residue, its evolutionary profile (the frequency of each type of amino acid in a given position of the alignment), and its predicted secondary structure and accessibility to the solvent. In particular, when a sequence is submitted as a query, ISIS runs PSI-BLAST [41], generates a multiple sequence alignment, and produces an evolutionary profile for each residue. These data are then sent to PROF [42], a system of neural networks that predicts the secondary structure state and the solvent accessibility of each residue. Finally, the sequence environment, the evolutionary profile, and the predicted structural features serve as input to another neural network, which annotates each residue as interface or noninterface. ISIS was trained on a nonredundant version of all transient protein–protein interfaces [27] in the PDB. (The 3-D structures were used only to identify the residues spatially in the interface. No experimental 3-D information was used for training.)
We trained standard feed-forward neural networks with back-propagation and momentum terms on windows of nine consecutive residues. A window was defined as positive if the central residue had any atom that was within 6 Å of any atom in a different protein. This yielded a set with 59,559 positive samples. We trained on two-thirds of the data and tested it on the remaining one-third.
Next, we filtered the raw network predictions. Our analysis of protein interfaces at the sequence level suggested that most interacting residues have other interacting residues in their sequence neighborhood. Therefore, we eliminated predictions with fewer than seven raw predictions within ten adjacent residues (five on either side).
To obtain the expected coverage and accuracy at random, we reshuffled the predictions in the following way: each protein was represented by two strings of the same length, one representing its sequence and the other representing the predictions (“P” for an interacting residue, “–” for a noninteracting residue). Then, we split the prediction string in half and assigned the predictions of the first half of the sequence to the second and vice versa. This process accounted for any size effect that could be caused by the number of predictions and for any effect caused by the heterogeneous distribution of contacting residues along the sequence. Furthermore, it enabled us to find a specific expectation for each scaling of the prediction. We generated different random models for different values of the receiver operating characteristic (ROC)–like curve (Figure 2C). Our background model captured how random our predictions were rather than how well we could predict interface residues at random.
ISIS was developed on a dataset of 1,134 chains in 333 complexes that contained 59,559 residue contacts. In the assessment of ISIS, no sequence that was used for training had any significant similarity for any of the sequences that were used for testing. That is, no protein in the test set could have been modeled by any protein in the development sets by homology-based predictions [43,44].
We chose methods that represent the variety of approaches for predicting interaction sites. ProMate [26] is a structure-based method that extracts features from an unbound chain and uses them to predict the binding site. We also chose three sequence-based methods: a sequence-only method [28], an evolutionary-based method (ET [35,45]), and a biophysics-based method (hydrophobic moment [22]). The first two were available as servers for public use. The hydrophobic moment was not publicly available; thus, we implemented it for the purpose of this analysis. We chose an operating point of coverage equal to 15%, which was the highest coverage reached by the hydrophobic moment tool.
We used the dataset of interface residues that was used to test ISIS originally [28]. In this dataset there are more than 20,000 interface residues, 2,182 of which were classified by ISIS as positive. Attempting to zoom in on the differences between hotspots and other interface residues, we compared the features of these 2,182 residues with the features of the residues that were classified as negative. The results of the comparison for amino acids are presented in Table 1, and are based on the evolutionary profile we used for prediction. For each interface residue, we used a multiple sequence alignment to check how often each residue is present in this position. We performed the same analysis for all the positions that were found experimentally, by alanine scanning, to be hotspots. Table 1 shows the average percentage occupancy of each amino acid in all positively predicted positions in all negatively predicted interface residues. |
10.1371/journal.pgen.1004055 | The Candidate Splicing Factor Sfswap Regulates Growth and Patterning of Inner Ear Sensory Organs | The Notch signaling pathway is thought to regulate multiple stages of inner ear development. Mutations in the Notch signaling pathway cause disruptions in the number and arrangement of hair cells and supporting cells in sensory regions of the ear. In this study we identify an insertional mutation in the mouse Sfswap gene, a putative splicing factor, that results in mice with vestibular and cochlear defects that are consistent with disrupted Notch signaling. Homozygous Sfswap mutants display hyperactivity and circling behavior consistent with vestibular defects, and significantly impaired hearing. The cochlea of newborn Sfswap mutant mice shows a significant reduction in outer hair cells and supporting cells and ectopic inner hair cells. This phenotype most closely resembles that seen in hypomorphic alleles of the Notch ligand Jagged1 (Jag1). We show that Jag1; Sfswap compound mutants have inner ear defects that are more severe than expected from simple additive effects of the single mutants, indicating a genetic interaction between Sfswap and Jag1. In addition, expression of genes involved in Notch signaling in the inner ear are reduced in Sfswap mutants. There is increased interest in how splicing affects inner ear development and function. Our work is one of the first studies to suggest that a putative splicing factor has specific effects on Notch signaling pathway members and inner ear development.
| The organ of Corti is a sensory structure in the cochlea that mediates our sense of hearing. It consists of one row of inner hair cells and three rows of outer hair cells, together with an array of neighboring supporting cells. The precise arrangement of these different cell types is regulated very tightly by a number of signaling pathways during embryonic development, and mutations in genes that regulate this pattern often lead to deafness. We have generated a mouse mutant containing a lentiviral insertion in a gene encoding a putative RNA splicing factor called Sfswap. Homozygous mutant mice have hearing and balance defects, and have an abnormal arrangement of hair cells in their cochlea. These defects are consistent with defects in the Notch signaling pathway. We show that Sfswap mutants interact genetically with a mutation in Jagged1, which encodes a Notch ligand. We show that expression of some genes involved in Notch signaling is disrupted in Sfswap mutant mice. Our work is one of the first studies to show that a putative splicing factor has specific effects on Notch signaling pathway members and on inner ear development.
| The organ of Corti is an excellent system to study mechanisms of cell patterning due to its highly organized array of sensory cells. It contains one row of inner hair cells, three rows of outer hair cells and several classes of specialized supporting cells, including pillar and Deiters' cells. The signals responsible for this intricate and fine-grained cellular pattern are beginning to be understood, and include the Notch signaling pathway. The Notch1 receptor is expressed in supporting cells, while the Notch ligands Jagged2 (Jag2), Delta1 and Delta3 are expressed in hair cells after they differentiate from prosensory precursors [1], [2], [3], [4]. Supernumerary inner and outer hair cells are generated at the expense of supporting cells in the absence of Notch1, Jag2 or Delta1 [5], [6], [7], [8], [9]. In addition, mutations in members of the Hes and Hey family of downstream Notch effectors also cause an increase in hair cell numbers at the expense of supporting cells, with mutations of multiple Hes/Hey family members causing progressively more severe phenotypes [10], [11]. These studies suggest that lateral inhibition mediated by Notch signaling acts to regulate and maintain the correct proportion of hair cells and supporting cells in inner ear sensory organs.
The Notch ligand, Jagged1 (Jag1) is expressed in all sensory organs of the inner ear prior to the onset of hair cell differentiation [2], [3]. In the developing mouse cochlea, Jag1 is expressed broadly at first, and then becomes excluded from the prosensory domain and restricted to Kölliker's organ by E13.5 [12]. As prosensory progenitors in the cochlea differentiate into hair cells and supporting cells, Jag1 is down-regulated from Kölliker's organ and is expressed with Notch1 in supporting cells [2], [3]. Although several hypotheses have been proposed for the mechanism of Jag1 function in the developing cochlea, the precise role of this gene is still poorly understood. Conditional inactivation of Jag1 in the developing inner ear leads to a severely disrupted organ of Corti [6], [13]. Sensory cells are entirely absent from the basal region of the Jag1 conditional mutant cochlea, whereas two rows of inner hair cells but no outer hair cells are observed in the apical region of the cochlea [6], [13]. Jag1 mutant heterozygotes generated by ENU mutagenesis show a milder phenotype; they lack some cells in the third row of outer hair cells and display ectopic inner hair cells [14], [15].
As part of a study to determine whether self-inactivating (SIN) lentiviruses can be used for efficient insertional mutagenesis in transgenic mice, we used a tyrosinase-expressing lentiviral vector to infect pre-implantation albino (FVB/N) mouse embryos by subzonal injection. Tyrosinase expression rescues albinism and provides a visible, dosage-sensitive, reporter for different integration sites. Transgenic founder (F0) mice were bred to establish families with single lentiviral integration sites and the mice were then inbred and assayed for evidence of insertional mutations. In one family (OVE2267B), homozygous mice displayed a robust circling behavior, suggesting inner ear defects. The lentiviral integration site in this family was mapped to the Sfswap gene. Sfswap was originally identified in Drosophila as a suppressor of the transposon-induced white-apricot mutation [16]. Sfswap encodes an RS-domain containing (SR-Like) protein that is a putative splicing factor. RS-domain containing proteins are known to regulate many aspects of RNA processing, including splicing, transcript elongation, transcript stability, nuclear export and miRNA cleavage as well as genome stability, (reviewed in [17]). In Drosophila, Sfswap regulates splicing of several genes, including Sfswap itself [18], [19], [20]. In vitro evidence suggests that Sfswap is involved in RNA processing in mammals as well by promoting fully spliced transcripts [21], [22], [23]. It is unclear whether Sfswap regulates other aspects of RNA processing, however some evidence in Drosophila suggests Sfswap may influence transcript stability [18], [24]. Our Sfswap mouse mutants have hearing loss, circling behavior, and show cochlear defects that are remarkably similar to those seen in Jag1 hypomorphic mutants - they show reduced numbers of outer hair cells and their associated supporting cells and increased numbers of inner hair cells. Compound mutants of Sfswap and Jag1 have a more pronounced cochlear phenotype and have truncations of their semicircular canals, suggestive of a genetic interaction. Moreover, we show that levels of expression of a number of genes involved in Notch signaling in the inner ear, such as Hey1, Neurl1, Numb, and MamlD1, are also affected in the Sfswap mutants. Our results suggest that Sfswap is necessary for the proper development and patterning of sensory structures of the inner ear and shows a genetic interaction with Jagged1 which may be mediated by a reduction in several genes involved in Notch signaling.
We conducted a random insertional mutagenesis study using a tyrosinase-tagged lentiviral vector (Figure 1A) to infect pre-implantation albino (FVB/N) mouse embryos. Lentiviral infection provides a number of distinct advantages for insertional mutagenesis. Lentiviral integration sites are scattered throughout the genome, are single copy, and are well-defined since integration is catalyzed by the lentiviral integrase. Since the injection needle is not inserted into the pronucleus of the embryos, the genomic DNA is not mechanically damaged and the yield of transgenic newborns is much higher. The tyrosinase minigene rescues albinism and provides a dosage-dependent, visible, reporter gene. Greater than 85% of the newborn mice from infected embryos were pigmented, verifying efficient transgenesis and effective expression of the reporter gene (data not shown). F0 mice were bred to FVB/N partners to generate F1 offspring and the F1 mice were again bred to albino mice to establish families with a single lentiviral insertion site. Mice were then inbred and assayed for evidence of insertional mutations. Eighty unique mutant phenotypes were identified (data not shown). This manuscript describes the characterization of the insertional mutation in family OVE2267B. Using inverse PCR, the integration site in this family was amplified, sequenced, and shown to be located in the fourth intron of Sfswap, 115 bases 5′ of exon 5. The mutation was labeled SfswapTg(Tyr)2267BOve (MGI ID: 5287267), and will be referred to throughout this study as SfswapTg or Tg. The tyrosinase minigene allows for identification of genotype by coat color when transgenics are maintained in an albino background [25]. Homozygote Tg mice are more darkly pigmented than heterozygotes and non-transgenic mice are albino. We back-crossed SfswapTg mice onto the FVB/N background for more than 10 generations. Tyrosinase expression and the inner ear phenotypes in homozygous mutants correlated 100% with the insertion in Sfswap as assayed by PCR. Additionally, Southern blots showed a single lentiviral integration site (data not shown). Northern blot analysis of Sfswap revealed that the wild-type Sfswap transcript was significantly reduced in SfswapTg/Tg mice (Figure 1B). In its place we observed an accumulation of Sfswap RNA migrating at an abnormal size (greater than 10 Kb). This RNA is most likely unspliced or incompletely spliced Sfswap mRNA, suggesting a hypomorphic allele. Reverse transcriptase PCR (RT-PCR), followed by sequencing, reveals that some of the Sfswap mRNA produced in SfswapTg/Tg mice includes sequences from the lentiviral insert and exclusion of exons surrounding the insert (data not shown). These abnormal RNAs are not found in mice heterozygous for the transgene, suggesting that SfswapTg is likely to be a recessive allele. Previous studies have demonstrated that Sfswap can regulate the splicing of its own transcript [18], [21], [22]. Disruption of Sfswap function by the mutation may explain why the Sfswap transcript is aberrantly spliced in homozygotes but not heterozygotes.
We initially identified SfswapTg/Tg mutants in our screen because they displayed a significant circling behavior (Figure 1C, Video S1). In a 30-minute open field assay, Tg/Tg mice circle almost five times more than wild-type or heterozygous littermates. SfswapTg/Tg mice are also almost twice as active as littermates (Figure 1D). Since circling behavior is often associated with vestibular dysfunction, we tested SfswapTg/Tg mice for balance defects. We found that Tg/Tg mice curl and grasp their feet in a tail hang assay, circle and tumble in a forced swim test, and have a delay in righting behavior (data not shown), all indicative of vestibular defects [26]. SfswapTg/Tg mutants on the FVB/N background are about 23% smaller than wild-type (WT) littermates at 8 weeks (WT = 29+/−1.11 g, Tg/Tg = 22.28+/−0.99 g p = 0.001), and this size difference is detectable at birth (WT = 1.39+/−0.045 g, Tg/Tg = 1.22+/−0.056 g, p = 0.008). Tg/Tg animals mate at very low frequency in the FVB background, but when crossed to a C57Bl/6 background, the circling behavior ceases and homozygous mice are able to mate more successfully.
Balance defects are often associated with hearing deficits in mice. To test if Sfswap mutants also have hearing defects we measured auditory-evoked brainstem responses (ABR) and found that Tg/Tg mice have an average 18 dB increase in threshold to elicit an auditory response, indicating a moderate degree of hearing loss (Figure 1E). Furthermore, ABR waveforms to auditory stimuli presented at intensities above threshold in SfswapTg/Tg mice have a qualitatively normal shape but smaller peak-to-peak amplitudes (Figure S1). This suggests that the auditory pathway is intact in Sfswap mice, but that fewer neurons are being stimulated to suprathreshold stimuli compared to wild-type mice. We next performed distortion product otoacoustic emissions assays (DPOAE) to identify if the hearing loss is at the level of hair cells. Tg/Tg mutants display an average 15 dB increase in DPOAE thresholds (Figure 1F), suggesting that outer hair cell defects may contribute to the hearing deficits [27]. To test how hearing loss affects behavior, we tested auditory startle responses. In this assay, mice are first acclimated to 70 dB white noise and their movement in response to subsequent varying sound pressures is recorded. As previously reported [28], the magnitude of the startle response increases rapidly between 100–120 dB in wild type mice (Figure 1G). We found that the startle responses in Tg/Tg mice are significantly attenuated compared to wild type. The maximum response we observe in Tg/Tg mice at 118 dB resembles those seen in wild type at sound pressures between 102–106 dB. This reduction in the threshold required to elicit a startle response resembles the increases we observed in ABR and DPOAE thresholds.
RNA in situ hybridization reveals that Sfswap is expressed in the inner ear and brain at E10.5 (Figure 2A). Sfswap is expressed broadly at low levels throughout the E10.5 embryo. Sfswap is expressed broadly and uniformly in the cochlea and surrounding mesenchyme as early as E13.5 in wild-type embryos (Figure 2B, C). Expression of Sfswap persists through birth and is maintained broadly in the cochlea, spiral ganglion, cristae, utricle, and saccule, but is reduced in the surrounding tissues (Figure 2D, E, F, G).
To study the effects of the SfswapTg mutation on inner ear morphology, we performed paint fills of the inner ear at E13.5–16.5. We found no defects in semicircular canal structure, and all components of the inner ear appear to be present. However, the cochlea is shorter, and the organs of the vestibular labyrinths are reduced in size (Figure 3A). Measurement of the flat-mounted cochlear preparations reveals a 38% reduction in the length of the cochlea (WT = 5058.5+/−201.45 µm Tg/Tg = 3135.9+/−221.92 µm p = 5×10−6). We stained surface preparations of newborn Tg/Tg cochleas with fluorescently-labeled phalloidin to detect actin in stereocilia and found regions of the mutant cochlea are missing the third row of outer hair cells and regions that have ectopic inner hair cells (Figure 3B, C). We quantified the number of hair cells per 200 µm and found there are significantly fewer third row outer hair cells throughout the length of the cochlea and significantly more inner hair cells at the apex (Figure 3C, Table 1). Although inner hair cells are locally increased, we found a significant decrease in total numbers of both inner and outer hair cells (37.5% and 50% respectively; Figure 3D).
To determine if there are also defects in supporting cells in SfswapTg/Tg cochleas, we examined expression of the transcription factor Prox1 on whole mount and sectioned inner ears to reveal pillar cells and Deiters' cells (Figure 4A, B) [29]. We found that in regions of the cochlea where outer hair cells are missing, a row of Prox1+ supporting cells are typically missing as well. Based on morphology and location, the missing supporting cell is likely a Deiters' cell or the outer pillar cell. In addition, we examined expression of the low affinity NGF receptor p75 that marks pillar cell apical processes (Figure 4C) and β-tectorin, which marks pillar cells and the greater epithelial ridge (Figure 4D). We found that in the mid-regions of the mutant cochlea individual pillar cell pairs are often absent, and replaced with an ectopic cell that is either a hair cell or an undifferentiated cell. β-tectorin is also absent in some regions of the mutant cochlea in a location corresponding to pillar cells, suggesting that the supporting cell loss is due at least in part to missing pillar cells. We also found that the hair cell marker parvalbumin is occasionally ectopically expressed in the Hensen's cell region of Tg/Tg mutants (Figure 4B). These ectopic parvalbumin-positive cells did not express other hair cell markers such as Myosin VI or show hair bundles with phalloidin staining, suggesting these ectopic cells are not bona fide hair cells, but rather represent mis-expression of at least one hair cell marker.
To determine if the cochlear defects we observed in Tg/Tg mice were due to gross abnormalities in the formation of the prosensory domain which gives rise to the organ of Corti, we examined mutant and wild-type animals for expression of the prosensory domain markers Sox2, and p27kip1 by antibody staining, and Hey2 by in situ hybridization. We used Jag1 antibodies and an in situ probe for Bmp4 to reveal the greater epithelial ridge and outer sulcus, respectively, which form a boundary with the prosensory domain. We did not observe any significant differences in the size of the prosensory domain, nor of the regions of the cochlea that border the prosensory domain in SfswapTg/Tg (Figure S2).
As described above, SfswapTg/Tg mice were initially identified as exhibiting circling behavior indicative of a vestibular defect. To analyze cellular defects that might cause this behavior, we stained flat mounts of the cristae and maculae of newborn mice with fluorescently labeled phalloidin to detect actin in stereociliary bundles, and measured the area of the sensory structures. Anterior and horizontal canal cristae remained attached to the utricle during dissection so that they could be unambiguously identified based on location. All semicircular canal cristae and the maculae of the utricle and saccule are smaller in Tg/Tg mutants than in wild-type mice (Figure 5A, B, C). Strikingly, the saccule is reduced to 25% of wild-type size. In addition, the anterior semicircular canal cristae exhibited smaller or absent eminentia cruciata in 85% of ears examined. The defects observed in vestibular organs likely contribute to the circling behavior of Sfswap mutant mice. We further analyzed the source of this reduction by examining early known markers of the vestibular cristae and maculae, Bmp4 and Lunatic Fringe (Lfng), respectively [30]. We found that the anterior stripe of Bmp4 expression, which corresponds to the presumptive horizontal and anterior canal cristae, is reduced in mutants at E10.5. The posterior spot, which will develop into the posterior crista, is small or indiscernible at this age (Figure 5D). Similarly, Lfng expression, which marks the future utricle, saccule, and neurogenic domain, is severely reduced in mutants at E10.5 (Figure 5D). These data suggest that the size reduction in the maculae and cristae in SfswapTg/Tg mice may represent an early developmental defect.
Finally, we tested for defects in mechanotransduction in the hair cells of SfswapTg/Tg mutant pups by intraperitoneal injection of the dye AM1-43, which can be taken up through mechanotransduction channels [31]. We found no differences in AM1-43 uptake in maculae, cristae, or cochleas between mutant and control mice (Figure S3). This data combined with the ABR and DPOAE data suggest that the remaining hair cells in SfswapTg/Tg mutant mice are likely to be functional.
SfswapTg/Tg mutants have a cochlear phenotype that is strikingly similar to that of Jag1 heterozygotes and point mutants. Jag1 heterozygous mutants display loss of the third row of outer hair cells, ectopic or extra inner hair cells, reduced expression of Bmp4 and Lfng at 10.5, disruptions in canal cristae and utricular macula, and small body size [14], [15], [32]. These defects can vary according to the genetic background [33]. We found that Jag1/+ mutants also have fewer supporting cells in a pattern similar to SfswapTg/Tg mutants (data not shown). These similarities lead us to hypothesize that Sfswap and Jag1 function in the same genetic pathway. To test this, we crossed Jag1 knockout mice (B6;129S-Jag1tm1Grid/J referred to hereafter as Jag1+/−) maintained on a C57BL/6 background, with our SfswapTg/+ mutants raised in an FVB/N background to obtain SfswapTg/+; Jag1+/− F1 progeny. These were then crossed again to SfswapTg/+ mice on an FVB/N background to generate SfswapTg/Tg; Jag1+/− mutants. Most of these compound mutants (10/14) exhibited semicircular canal truncations, but neither SfswapTg/Tg nor Jag1+/− mutants have canal truncations on this genetic background (Figure 6A). Jag1+/− mutants are known to have variations in semicircular canal defects depending on the genetic background [33]. To test if this is the case for the FVB/N background on which our SfswapTg mice were maintained, Jag1+/− mice were crossed one and two generations to wild-type FVB/N mice. Surprisingly, one generation was enough to completely suppress the canal truncations (Figure 6B). This indicates that there is a strong suppressor of canal truncations in the FVB/N background, and the canal truncations found in SfswapTg/Tg; Jag1+/− mice are therefore strongly indicative of an interaction between these two genes. To further test this hypothesis, we stained cochlear flat mounts from single and compound mutants with fluorescently labeled phalloidin to visualize stereocilia. We found that SfswapTg/Tg; Jag1+/− mutants have a more pronounced cochlear phenotype than either single mutant alone, showing a loss of outer hair cells extending into the second and first rows, increased inner hair cells throughout the length of the cochlea and the addition of a fourth row of outer hair cells in the apex (Figure 6C). In addition, compound mutant cochleas are significantly shorter than either Sfswap or Jag1 mutant cochleas (Figure 6D, Table 2). Quantitatively, the number of outer hair cells in the mid-turn and base of the cochlea are significantly fewer in the compound mutant than either mutant alone (Figure 6E, Table 2). For audiological measurements, we out-crossed the Jag1+/− mice to the FVB background for an additional three generations. These Jag1+/− mice were then crossed to SfswapTg/+ mice to generate compound heterozygotes that were then intercrossed to produce SfswapTg/Tg; Jag1+/−, Jag1+/− and SfswapTg/Tg mice. ABR and DPOAE measurements of compound mutants also showed increased threshold compared to those of Tg/Tg or Jag1 mutants (Figure 6F, G; Tables 3–6). Interestingly, we find that Jag1+/− mice have a significant increase compared to WT ABR and DPOAE thresholds; this difference is evident at high frequencies. However, compound mutants also have increased thresholds compared to Tg/Tg at frequencies below 20 kHz, a range that is only mildly affected in Jag1+/− mice. The defects found in SfswapTg/Tg; Jag1+/− cochlear hair cells, semicircular canals, and auditory responses are greater than expected for a simple additive effect, implying a synergistic relationship between Jag1 and Sfswap in inner ear development.
To further analyze the mechanism for Sfswap's interaction with Jag1, we analyzed expression of Jag1 in cochlear whole mounts and at the otocyst stage, but found no differences in expression (Figure 7A, B). We next examined expression of the downstream target Hey1 at E10.5 and found a significant reduction in expression (Figure 7C), despite normal Jag1 expression. To determine if transcripts for genes in the Notch signaling pathway are incorrectly spliced in Tg/Tg mutants, we performed RT-PCR of splice junctions for Notch pathway genes using inner ear mRNA. We analyzed Jag1, Notch1, Hes1, Hes5, Hey1, Hey2, HeyL, Mfng, Lfng, Rbpj, Delta1, Numb, NumbL, Maml1, Maml2, Maml3, MamlD1, Neuralized1A (Neurl1A), and Sfswap and found no differences in splicing in any gene except for Sfswap (Figure 7D and data not shown). We found slight reductions in levels of Neurl1A and Numb mRNA and significant reductions in MamlD1 mRNA visible by RT-PCR, all of which could potentially contribute to the Jagged1-like phenotype of SfswapTg/Tg mice. To test for splice differences in these genes that would not be detectible by RT-PCR, we performed Northern blots using brain RNA. However, we found no detectable differences in splicing in Sfswap mutants in Jag1, Neurl1A, MamlD1, or Numb (Figure S4).
Using a lentivirus-based insertional mutagenesis strategy, we have identified the putative splicing factor Sfswap as an essential gene for inner ear development. SfswapTg/Tg mutants exhibit circling and balance dysfunction associated with vestibular defects and also have a moderate (20–25 dB) hearing loss. The mutants exhibit a partial loss of the third row of outer hair cells and show ectopic inner hair cells, together with some missing pillar cells. All organs in the vestibular system are smaller than their wild-type counterparts, particularly the utricle and saccule, which are less than 50% of the size of wild-type organs. Our data are consistent with Sfswap regulating aspects of Jagged1 signaling. First, the cochlear phenotype of SfswapTg/Tg mice is strikingly similar to that seen in Jag1 heterozygous point mutants [14], [15]. Second, SfswapTg/Tg mice have otocyst patterning defects similar to those seen in Jag1 mutants [32]. Finally, we show that our Sfswap mutant allele exacerbates the phenotype of Jag1 heterozygous mice. We discuss these phenotypes in more detail below.
SfswapTg/Tg mutants have reduced levels of the early sensory markers Bmp4, Lfng, and Hey1 and significantly smaller sensory organs. Jag1 conditional mutants similarly have severely reduced vestibular structures [13]. Jag1 mutants also show changes in expression of sensory markers such as Bmp4, Sox2, Lfng and Hey1 at E10.25 [32]. Vestibular structures are also reduced in Bmp4 conditional or hypomorphic mutants [34], particularly in the horizontal lateral canal. We suggest the reduction of Bmp4 expression in our mutants may be due to defects in Jag1 signaling and leads to the smaller vestibular structures and reduced eminentia cruciata of the anterior crista that we observe in SfswapTg/Tg mutants.
The unique combination of loss of outer hair cells and ectopic inner hair cells that we observe in the cochleas of SfswapTg/Tg mutants has only been observed previously in heterozygous knockout or point mutants of Jag1 [14], [15]. To date, there is no definitive molecular explanation for this unique phenotype. Kiernan and colleagues [14] suggested that Jag1 may have two roles, an early one in which Jag1 helps specify sensory organs, and a later one in which Jag1 helps define hair cell and supporting cell identity through lateral inhibition. We did not see significant changes in the size of the cochlear prosensory domain in SfswapTg/Tg mutants, although if there are small changes in the prosensory domain, they would be difficult to discern using the markers and techniques currently available. Ectopic rows of inner hair cells can be seen in Jag1 heterozygotes and in apical regions of the Jag1 conditional homozygote cochlea [6], [13]. In SfswapTg/Tg mutants, we also observe an extra row of inner hair cells in the apical region of the cochlea. Since Jag1 is expressed at the boundary of the greater epithelial ridge and the prosensory domain where the first inner hair cells form, it is possible that a partial loss of Jag1-Notch signaling leads to the inappropriate formation of extra inner hair cells in Jag1 and SfswapTg/Tg mutants.
Although our data support a genetic interaction between Sfswap and Jag1, we currently have no evidence for a direct interaction. It is possible that Jag1 mRNA splicing or stability is regulated by Sfswap. However, we did not detect significant differences in the levels or splicing pattern of Jag1 transcripts, and we have not detected significant differences in Jag1 protein expression in the cochlea by immunofluorescence. Since the cochlear phenotype of Jag1 mutants can vary according to both the type of mutation (point mutation versus null) and the genetic background on which mutants are maintained [13], [14], [15], [33], it is clear that hair cell patterning in the cochlea is exquisitely sensitive to levels of Jag1. It is possible that SfswapTg/Tg mice are causing small changes in Jag1 expression below the limits of detection that are sufficient to affect cochlear patterning. Alternatively, it is possible that Sfswap is regulating the splicing or stability of modifiers of the Jag1 locus. We screened a variety of Notch pathway genes for possible expression changes in Sfswap mutant cochleas. Several genes showed significant changes by RT-PCR including the mouse neuralized homologue Neurl1A, the Notch transcriptional co-activator MamlD1, and Numb (Figure 7D). Neurl1A is an E3 ubiquitin ligase that has been shown to regulate turnover and endocytosis of Jag1 in vitro [35]. It is therefore possible that small changes in Neurl1A activity may be sufficient to disrupt hair cell patterning. Similarly, Numb has been shown to regulate Notch signaling through endocytosis and ubiquitinization of the intracellular domain of Notch [36], [37]. Interestingly, Numb has recently been shown to be broadly expressed in the rat cochlea during development and over-expression can modulate levels of Atoh1, suggesting a possible role in hair cell development [38]. Mastermind proteins are transcriptional co-activators that are essential for Notch signaling [39]. It is possible that a reduction in MamlD1 can affect Jagged1-Notch signaling through a reduction in canonical or non-canonical Notch signaling [40], [41]. It is also possible that the simultaneous reduction in two or more of these proteins results in the phenotypes seen in the Sfswap mutants.
Sfswap was discovered in Drosophila as a suppressor of the transposon-induced white-apricot mutation [16]. White-apricot mutants have a transposon insertion that disrupts the white transcript by splice inclusion and this disruption is partially suppressed by mutation of Sfswap through selective exclusion of the transposon [42], [43], [44]. Beyond this system, no evidence has been identified for the in vivo function or targets of Sfswap in Drosophila. In mice, in vitro studies have identified Fibronectin and CD45 as putative splice targets of Sfswap [21], although in vivo targets have yet to be confirmed. In humans, the nonsyndromic autosomal dominant deafness locus DFNA41 contains at least 100 genes, including SFSWAP [45]. The pathological mutation in this region has yet to be identified, making SFSWAP a potential candidate for future gene sequencing efforts.
Many genes undergo complex splicing patterns in the inner ear. These include several genes that are involved in ion transport including Atp2b [46], Kcnq4 [47], [48], Cav1.3 [49], BK channels [50], [51], [52], and P2X2 [53]. Spliced isoforms of these channels have been proposed to be involved in electrical differences in tonotopy, electromotility of outer hair cells, and differences between sensory and neuronal cells. Alternative splicing has been found to affect protein targeting [49], [54], [55], electrical properties channels [56], [57], [58], [59], cell viability [60], [61], and stereociliary organization [62] in the inner ear. There are also several examples of mutations in distinct alternative spliced isoforms that cause different pathologies in the ear. For example, loss of one isoform of PCDH15 results in stereociliary defects, while mutations in the other two isoforms of this gene have no significant effect on hair cells [62]. Similarly, mutations in harmonin are typically associated with Usher Syndrome 1C, whereas a mutation in an alternative exon results in non-syndromic deafness at the DFNB18 locus with normal vision [63]. Despite the significance of alternative splicing in the inner ear, only one splicing factor has been identified that functions in the inner ear [61]. Srrm4 is an SR-like protein that has recently been identified to be mutated in the Bronx waltzer (bv) mouse. Mutation at this locus results in degeneration of inner hair cells after E17.5 [64]. Gene ontology analysis of exons regulated by Srrm4 suggests this factor regulates splicing of genes involved in synaptic transmission and the secretory pathway. We have now identified a second SR-like protein, Sfswap, which has distinct and highly specific effects on patterning of the inner ear. Our work provides the first evidence that a putative splicing factor is necessary for establishing the size of sensory organs in the ear and for proper patterning of mechanosensory hair cells in the organ of Corti.
A detailed account of the generation of Sfswap transgenic mice by lentiviral insertional mutagenesis with a tyrosinase minigene is given in Text S1.
For RT-PCR, RNA was isolated from E15.5 inner ears using the Ambion PureLink RNA mini kit (12183018A) according to the manufacturer's instructions. cDNA was generated using the Superscript III First Strand system (Invitrogen 18080-051) according to manufacturer's instructions. RT-PCRs for exon splicing events were performed using exon tiling primers (Tables S1 and S2). See Text S1 for Northern blot protocol.
Plasmid probes for in situ hybridizations and Northern blots were obtained from Brigid Hogan (Bmp4),Gerry Weinmaster (Jag1), Elias Pavlopoulos (Neurl1A), Yutaka Hata (Numb, Addgene Plasmid 37012 [65]) or cloned using TOPO 2.1 vector (Life Technologies). PCR primers used to clone probes are as follows: Sfswap exons 1–4: F-GCTGTGTTGAAGTTGCGAAG and, R-CATCAGACGGGACGCTTAAT; Sfswap exons 15–18: F-AAAGGACCCGTTCCAGAAGT and R-CCACTGACTGACCCAGGAGT; beta-Actin: F- TGTTACCAACTGGGACGACA and R- AAGGAAGGCTGGAAAAGAGC; MamlD1 F-TCCATTTCCCATCTCCTCAG and R- AGCCTTCCAAAAGCTCTTCC.
Some in situ probe templates were generated through direct PCR with the addition of a T7 promoter (T7 sequence is underlined). These include Lfng: F-GTTCCGCTCTGTCCATTGC R- GGATCCTAATACGACTCACTATAGGGAGCCCACTATGGGCGACTTTC and Hey1: F-AGACCTTGGGGGACAGAGAT and R-GGATCCTAATACGACTCACTATAGGGAGAACGGTGAAATCCGTGAGAC.
For cryosections, heads were fixed in 4% paraformaldehyde overnight at 4 degrees. Heads were then washed briefly in PBS, immersed in 30% sucrose overnight, equilibrated in OCT, then embedded in OCT. Sections were taken between 8 and 14 µm. For P0 and older whole mounts, heads were fixed overnight in 4% paraformaldehyde. Cochleas, cristae, and maculae were then dissected out and processed. For immunohistochemistry, some tissue was first boiled for 10 minutes in 10 mM citric acid for antigen retrieval. All samples were washed 3 times in PBS, then 30 minutes in blocking buffer (PBS with 10% goat serum and 0.1% Triton-X or Tween-20). Tissue was then stained overnight in primary antibody diluted in blocking buffer. Tissue was then washed 3 times in PBS then incubated for 2 hours with the appropriate secondary antibody at 1∶400 dilution in blocking buffer. In tissues where Topro3 (1∶10,000, Invitrogen) or phalloidin (1∶200, Invitrogen) were used, they were added to the secondary antibody cocktail. Tissue was then washed 3 times in PBS and mounted in either Vectashield (Vector Labs) or Prolong Gold+Antifade (Invitrogen).
Antibodies used: Myo6 (Proteus), Prox1 (Millipore Bioscience Research Reagents), Parv (Sigma), Jag1 (Santa Cruz Biotechnology), p27Kip1 (Neomarker), p75 (Advanced Targeting Systems), Alexa 488 anti mouse and rabbit (Invitrogen), Cy5 anti mouse and rabbit (Jackson Immunolabs). In situ hybridization was performed as previously described [66], [67]. Digoxigenin labeled riboprobes were synthesized according to standard protocols [68].
Images were taken using a Zeiss LSM 510 confocal microscope, a Zeiss Axiophot microscope, or a Zeiss dissecting scope. Images were processed using Axiovision software or ImageJ, then further processed in Adobe Photoshop. Inner and outer hair cells were counted in P0 cochleas per 200 µm using Axiovision software. Cochlea lengths were measured at P0 using Axiovision software. Vestibular areas were measured on P0 vestibular flat mounts using ImageJ.
Paint fills were performed as previously described [69], [70]. In brief, E13.5–E16.5 heads were fixed in Bodian's fix. Heads were dehydrated in an ethanol series, and then cleared with methyl salicylate. Inner ears were filled with white gloss paint in methyl salicylate using a Picospritzer III pressure injector (General Valve Corporation) and stored and photographed in methyl salicylate.
Open field activity was measured using the Versamax System for automated activity recording. Mice were acclimated for at least 30 minutes to the testing room illuminated to 400 lux light and with 60 dB white noise before testing. The open field is a 40×40×30 cm Plexiglas arena. Tests were performed for 30 minutes during which time horizontal and vertical activity were measured via beam breaks. Chambers were cleaned with alcohol before and after each run to minimize interfering odors. Pre-pulse inhibition and startle responses were measured using a San Diego Instruments system. Mice were acclimated for at least 30 minutes in a separate nearby room before testing. Mice were placed in cylindrical restraint tube in a sound-attenuating chamber to minimize movement of the mouse and interfering noise. Mice were acclimated in the testing chamber to 70 dB white noise for 5 minutes before test pulses were delivered. Mice were then presented with 6 rounds of 8 test types in a pseudorandom order with 10–20 seconds between trials. These consist of no stimulus, startle at 120 dB for 40 ms, pre-pulse trials (74, 78, and 82 dB for 20 ms) and pre-pulse inhibition trials of each sound 100 ms prior to a 120 dB startle. Responses were recorded every 1 ms for 65 ms following the stimulus. Tg/Tg mice fail to present a threshold level startle response, so percent pre-pulse inhibition could not be calculated and only startle responses are reported. ABR and DPOAE were performed as previously described [71].
Significance was measured using students T-Test or ANOVA with repeated measures where appropriate using SPSS software. Graphs were generated using Prism.
All animal experiments in this study were carried out in accordance with the Institutional Animal Care and Use Committee at Baylor College of Medicine.
|
10.1371/journal.pntd.0004954 | Controlling Neglected Tropical Diseases (NTDs) in Haiti: Implementation Strategies and Evidence of Their Success | Lymphatic filariasis (LF) and soil-transmitted helminths (STH) have been targeted since 2000 in Haiti, with a strong mass drug administration (MDA) program led by the Ministry of Public Health and Population and its collaborating international partners. By 2012, Haiti’s neglected tropical disease (NTD) program had reached full national scale, and with such consistently good epidemiological coverage that it is now able to stop treatment for LF throughout almost all of the country. Essential to this success have been in the detail of how MDAs were implemented. These key programmatic elements included ensuring strong community awareness through an evidence-based, multi-channel communication and education campaign facilitated by voluntary drug distributors; strengthening community trust of the drug distributors by ensuring that respected community members were recruited and received appropriate training, supervision, identification, and motivation; enforcing a “directly observed treatment” strategy; providing easy access to treatment though numerous distribution posts and a strong drug supply chain; and ensuring quality data collection that was used to guide and inform MDA strategies. The evidence that these strategies were effective lies in both the high treatment coverage obtained– 100% geographical coverage reached in 2012, with almost all districts consistently achieving well above the epidemiological coverage targets of 65% for LF and 75% for STH—and the significant reduction in burden of infection– 45 communes having reached the target threshold for stopping treatment for LF. By taking advantage of sustained international financial and technical support, especially during the past eight years, Haiti’s very successful MDA campaign resulted in steady progress toward LF elimination and development of a strong foundation for ongoing STH control. These efforts, as described, have not only helped establish the global portfolio of “best practices” for NTD control but also are poised to help solve two of the most important future NTD challenges—how to maintain control of STH infections after the community-based LF “treatment platform” ceases and how to ensure appropriate morbidity management for patients currently suffering from lymphatic filarial disease.
| We present evidence of Haiti’s successful neglected tropical disease (NTD) program targeting lymphatic filariasis and soil-transmitted helminths and the methods used to achieve this success. By 2012, Haiti’s NTD program had reached full national scale, with such consistently good treatment coverage that the program is now able to stop treatment for lymphatic filariasis in much of the country. These findings are in line with the predictions and expectations of the global community for countries where high coverage is achieved for program implementation. In addition to the evidence of successful program outcomes, we present a detailed description of how the program was implemented—from facilitating the effectiveness of the drug distributors to improving drug logistics and supporting a well-informed population. These methods described can be used to inform the design of other mass drug administration programs and enhance the development of global “best practices” guidance.
| Neglected tropical diseases (NTDs) are a group of 17 parasitic, bacterial, and viral infections affecting more than 1 billion people globally [1]. Seven of these are known as the preventative chemotherapy NTDs because they can be eliminated or controlled by administering medicines to entire eligible populations or large segments of these populations in an effort to reduce transmission of infection and prevent disease.
Four preventative chemotherapy NTDs are endemic in Haiti—lymphatic filariasis (LF) and the three soil-transmitted helminth (STH) infections caused by Ascaris, Trichuris, and hookworm. LF is one of the world’s most debilitating parasitic diseases, causing lymphedema, elephantiasis, hydrocele (enlarged scrotum by fluid accumulation), and hidden internal damage to the lymphatic and renal systems of affected individuals [2]. Furthermore, as a disease associated with stigma, despair, hopelessness, embarrassment, ridicule, frustration, and economic burden [3], LF can also cause significant mental health complications that reach far beyond even its physical morbidity [4]. STH causes a wide range of intestinal symptoms and has also been associated with poor cognitive development and learning capacities in children [2]. Furthermore, STH can cause anemia in women of child-bearing age, which is both detrimental to the mother’s health during pregnancy and can lead to low birth weight [5].
Haiti is one of only four countries in the Americas where LF transmission still occurs, and it is home to the largest at-risk population in the region [6,7]. In Haiti, LF is caused by Wuchereria bancrofti filarial parasites, primarily transmitted by Culex quinquefasciatus mosquitoes [7,8]. In 2000, a nationwide mapping exercise reported an infection prevalence (assessed by filarial antigen) as high as 45% in children 6 to 11 years old, with infected children identified in 117 of the country’s present 140 communes (133 communes at that time) [9]. Nationwide “mapping” for STH infections was carried out in 2002, with all of Haiti’s ten departments (previously reported as nine due to redistricting) reporting >20% prevalence and two of these with >50% prevalence [10].
Addressing such NTD challenges is complex. Although health experts and international organizations have presented general recommended approaches for integrated programs targeting NTD control or elimination [11], the global portfolio of “best practices” can only grow through the addition of detailed records of individual national experiences. Indeed, the success of national NTD programs (or at a minimum, their avoidance of failure) depends on understanding the methods used to implement and assess programs in those countries where NTD programs are already mature or complete. In this regard, Haiti could make important contributions to the accumulating experiences of implementing integrated NTD programs [12,13]. This current report, while recognizing the importance of the contributions of all of Haiti’s principal NTD stakeholders[14], focuses on the efforts implemented by the Haitian government and a large project (ENVISION) supported by the U.S. Agency for International Development (USAID). ENVISION is led by RTI International and is being implemented in 19 countries globally [http://www.ntdenvision.org]; in Haiti, IMA World Health leads the ENVISION activities.
Treatment coverage is a measure of whether programs are on track to achieve program goals and control/elimination targets. Good coverage was reported (Table 1)– 100% geographical coverage was achieved in 2012 and national coverage targets were reached for both STH (≥75%) and LF (≥65%) across most communes. In communes supported by ENVISION, the average epidemiological coverage for LF MDA rounds ranged from 86% to 94% (Table 1). The number of communes with coverage below the 65% target ranged from one to five in any given year, with no systematically low-performing communes identified.
Post-treatment coverage surveys were used to validate the coverage routinely reported by CDDs. Results of the coverage surveys generally confirmed that coverage rates were at least 65%, and often substantially higher (Table 2).
KAP questions were added to the coverage surveys in both 2012 and 2013 (Table 3). Results showed that it was easy for people to reach treatment posts (96% of respondents in 2012 and 82% in 2013 reported that the treatment post was “easily accessible” or “not too far”) and that awareness levels on the program were high (95% of respondents in 2012 and 89% in 2013 reported knowing the MDA was going to occur). While knowledge of signs and symptoms of LF was good (85% of respondents in 2012 and 59% in 2013 correctly reported at least one sign or symptom of LF), knowledge on prevention methods was low (27% of respondents in 2012 and 47% in 2013 correctly reported at least one method to prevent infection). A similar study in Léogâne following a single MDA, conducted by MSPP with support from CDC, had found no significant association between knowledge of disease and treatment [25].
The factor most strongly associated with taking the treatment was access, with adjusted odds ratios of 30 and 56.2 for 2012 and 2013 respectively, p <0.001 (t-test). Caution should be exercised when interpreting the odds ratios not to overemphasize the significance of such high odds ratios. If an outcome is rare, odds ratios and relative risk would be roughly equivalent. Here, since the majority of the surveyed took the drug, the odds ratio and relative risk are different. The unadjusted relative risk is 2. Awareness of the program was also significantly associated with taking the drugs (odds ratios of 5.8 and 12.1 for 2012 and 2013 respectively, p <0.001 [t-test]). The association between knowledge of disease and taking treatment was inconclusive—knowledge was not associated with taking the treatment in 2012 but was significantly associated in 2013 –odds ratio of 1.8 (p = 0.002) for persons who could reports signs and symptoms and odds ratio of 2.7 (p <0.001) for ability to report prevention strategies.
Program impact is first measured at sentinel and spot-check sites—these are conducted at midterm after two to three MDA rounds and again before proceeding to conduct the more extensive TAS, following a minimum of five MDA rounds. Once a commune has completed a minimum of five rounds of MDA, with epidemiological coverage >65% and where sentinel site W. bancrofti antigen levels are below 2%, TAS should be implemented to determine whether MDA can be stopped [24]. For these TAS, communes may be grouped into larger evaluation units (EUs) based on similar epidemiologic characteristics. Results are reported for areas supported by ENVISION only, where the program was implemented as described in the methods section.
In 2014–2015, sentinel and spot-check surveys were conducted in 23 communes as part of either midterm or pre-TAS surveys (Table 4). In all communes, there was a marked decrease in infection prevalence as compared with initial mapping results. Results were below the threshold of 2% antigenemia in 19 of the 23 sites (Fig 4); 7 of those sites (conducting midterm assessments) had completed only three or four rounds of MDA, and all 7 had already reached the 2% threshold. MDA was continued in these areas based on current global guidelines to conduct a minimum of five rounds.
On the other hand, four sites had large reductions but failed to drop below the critical threshold, even after six or seven rounds (Table 4). The principal difference between these two groups is that the 19 sites where infection rates were reported as below the 2% threshold typically had lower prevalence rates at baseline; despite not meeting the criteria to stop MDA, the substantial decrease in prevalence (i.e., ranging from 28% to 39% at baseline to between 2.3% and 6.5% pre-TAS) among those four sites is viewed as a programmatic success (Table 4).
By mid-2015, TAS had been implemented in a total of 13 EUs, 10 of which achieved the prevalence threshold required to stop MDA. These 10 EUs are made up of a total of 44 communes in which, according to WHO guidelines, MDA can now be stopped, while post-MDA surveillance will continue for an additional four to six years. In the other three EUs (made up of three communes), MDA will continue for a minimum of two more years.
For STH, a national survey was conducted in 2013 to determine the STH infection prevalence in school children (aged from 6 to 16 years old), at least five years after the program of annual MDA with ALB+DEC had begun. This survey showed that STH prevalence had decreased in 9 of 10 departments, with STH prevalence at or below 10% in 3 departments and at or below 20% in another 6 departments. In the tenth department (Grand’Anse), which had an initial prevalence of 72% when surveyed in 2001, the decline was not as pronounced as in the other nine, and the prevalence of infection remained at 55%[26]. Although these results are encouraging and show significant progress in addressing the problems caused by STH in Haiti, they also highlight the question of how progress will continue after support for annual MDA targeting LF ceases.
Program costs. Detailed program cost assessments have been carried out for those communes where MSPP activities have been supported by ENVISION. Although relative costs for some of the program activities did vary slightly from year to year, drug delivery costs always predominated, especially when including social mobilization costs (Fig 5). Because ALB is provided by GlaxoSmithKline at no cost to the program, and since the DEC is quite inexpensive, the cost of medicines (i.e., the sine qua non of MDA strategies) to the national program remains relatively low.
Between 2010 and 2013, the per person treated cost of the program—characterized predominantly by MDA—has remained steady (US$0.35–0.37). However, in 2014 an increase in total budget was observed as the requirement for more comprehensive M&E efforts, including TAS surveys, became necessary.
By 2012, Haiti’s NTD program had reached national scale, with consistently good epidemiological coverage reported for MDA. Following TAS results that confirmed LF infection was below the targeted transmission thresholds, MDA could be stopped, resulting in a scaling down of geographical coverage. By 2015, 45 communes– 44 of which were in ENVISION-supported areas—had stopped MDA. Such results add to the growing body of global evidence supporting the feasibility and effectiveness of implementing WHO’s strategy to eliminate LF via once-yearly MDA for at least five years with a minimum of 65% epidemiological coverage [27].
To achieve these results, it is essential that MDA rounds are able to reach and sustain good coverage of the targeted population. In practical terms, the success of MDA programs hinges on two operationally critical elements: a population willing to accept the treatments being offered (community acceptance) and a health system able to deliver the treatments effectively to the community (MDA delivery). These in turn depend on a number of other essential factors, summarized in Fig 6 and described below.
The likelihood of taking the offered treatment has been positively associated in a number of studies with the community’s knowledge and understanding of the disease, its transmission, MDA strategy, and possible side effects; the community’s perceived risk of getting the disease; and history of side effects experienced [28–33]. In Haiti, however, knowledge of disease signs and symptoms was generally low and not strongly associated with taking part in MDA. The most important factors in determining whether people took the treatment offered them were (1) knowing that the MDA was to take place and (2) the ease of access to treatment posts. It is also of note that treatment in Haiti’s NTD program is directly observed, a factor associated in other settings with increased likelihood of community participation [29,34–38].
Reluctance to take treatment due to fear of minor side effects was low, reflecting the importance placed on informing the community about expected side effects—achieved through emphasis during CDD training and through dissemination of IEC messages. The program also benefitted from the fact that there were no reported serious adverse events.
This program has been able to extensively leverage pre-existing relationships among ENVISION, the field MSPP personnel, and the community leaders—all of whom emphasize the involvement of the community, thereby building on the foundation of existing strong social capital. Other studies also support the idea that drug distributors who are known and respected by the community will have better success in treating the population [34].
The factor most strongly associated with taking the drugs in this program was ready access to treatment (odds ratios of 30 and 56.2 for 2012 and 2013 respectively, p <0.001). The program’s attention to logistics—in terms of number and location of posts, as well as assurance that the posts are attended by well-trained and supported CDDs—accompanied by a strong supply chain that ensured the availability of treatments, were critical to the successful achievement of persistent high treatment coverage.
In light of the strong evidence presented on good coverage in MDA rounds obtained by the Haitian program, the excellent results of the assessment of impact on infection at sentinel, spot-check, and evaluation sites should not be surprising. LYMFASIM, a mathematical model predicting elimination of LF using an MDA approach, was developed based on a similar scenario in India, where C. quinquefasciatus is the vector and W. bancrofti the parasite, but where ivermectin was used instead of DEC. This model predicted that the number of rounds necessary to achieve elimination depends to a large extent on both coverage and the pre-program endemicity level [39]. This prediction was well corroborated in Haiti’s experience– 19 of the 23 sentinel sites tested in 2014–2015 reached below the critical 2% antigen prevalence threshold. The four sites that did not reach this threshold, despite at least six rounds of MDA with high epidemiological drug coverage (≥83%), had very high baseline antigen prevalence (28%–39%; Fig 4). Earlier studies in Léogâne, Haiti also reported slower declines—from 50% to only 14% following seven rounds of MDA [40].
As noted earlier, seven sentinel sites with initially much lower baseline prevalence had already reached below the <2% antigenemia threshold after only three to four rounds of MDA. These communes continued to implement another one to two rounds of MDA following current WHO guidelines. Researchers in Haiti’s La Tortue Island earlier reported a successful TAS survey and the stopping of MDA after just two rounds of treatment [16]. In the neighboring Dominican Republic, similar results—reduction of antigenemia to below threshold after three rounds of MDA—were found in the urban center of the country, again with low baseline prevalence and high program coverage [41]. There is potential for programmatic efficiencies if programs targeting LF elimination could be successful with fewer rounds of MDA than are now recommended; whether this would be sufficient to interrupt transmission will need to be determined by operational research.
In addition to completing efforts to eliminate LF, Haiti’s NTD program needs to sustain achievements of reduced STH burden. During recent years, the STH program has relied on the fact that ALB is administered to the whole population as part of the LF MDA. As the LF program progressively meets the criteria for stopping MDA, another platform will be needed for delivery of STH treatments. Haiti’s plan is to focus on school-based, STH-only MDA and to add water, sanitation, and hygiene activities.
In 2015, the InterAmerican Development Bank provided the MSPP with funds for a national deworming campaign. The InterAmerican Development Bank drew upon IMA’s expertise for support in training of drug distributors and working with department-level STH leads to formulate best-practice MDA strategies—illustrating the continued integration and collaboration between the two disease programs and the potential resources that have been developed in country and can be leveraged to ensure the sustainability of STH control activities.
As MDA activities are now reaching the stopping point, Haiti has started to turn its attention to morbidity control efforts for LF, building on hydrocele surgery efforts and lymphedema patient support groups that have been implemented in Léogâne [42]. The MSPP updated its national morbidity management and disability prevention (MMDP) strategic plan in early 2016, which includes rolling out the use of WHO’s new MMDP situation analysis tools to estimate the number of patients with hydrocele or lymphedema, identify platforms to support MMDP activities, identify strategies to mobilize patients, and identify the human resources and funding needs.
Finally, plans are also in place to work Hispaniola-wide on a malaria elimination plan, harnessing the momentum already gained for LF elimination and further increasing the probability of LF elimination though vector control [43].
In summary, in this mostly rural Haitian setting, a very successful MDA campaign was launched in 2001 with sustained high coverage rates, resulting in steady progress toward LF elimination (with more MDA rounds being needed in areas with high baseline prevalence and fewer where prevalence was initially low). A strong foundation for ongoing STH control has also been established. The strategies utilized with proven success by the program are summarized in Box 1. This list is an important contribution to the development of the global evidence base on MDA best practices.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.