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.1000921 | Is Protein Folding Sub-Diffusive? | Protein folding dynamics is often described as diffusion on a free energy surface considered as a function of one or few reaction coordinates. However, a growing number of experiments and models show that, when projected onto a reaction coordinate, protein dynamics is sub-diffusive. This raises the question as to whether the conventionally used diffusive description of the dynamics is adequate. Here, we numerically construct the optimum reaction coordinate for a long equilibrium folding trajectory of a Go model of a -repressor protein. The trajectory projected onto this coordinate exhibits diffusive dynamics, while the dynamics of the same trajectory projected onto a sub-optimal reaction coordinate is sub-diffusive. We show that the higher the (cut-based) free energy profile for the putative reaction coordinate, the more diffusive the dynamics become when projected on this coordinate. The results suggest that whether the projected dynamics is diffusive or sub-diffusive depends on the chosen reaction coordinate. Protein folding can be described as diffusion on the free energy surface as function of the optimum reaction coordinate. And conversely, the conventional reaction coordinates, even though they might be based on physical intuition, are often sub-optimal and, hence, show sub-diffusive dynamics.
| To understand dynamics of complex systems with many degrees of freedom, one often projects it onto one or several collective variables. Protein folding, the complex, concerted motion of a protein chain towards a unique three-dimensional structure, is one example of where such reduction of complexity is useful. It is usually assumed that the projected dynamics is diffusive. However, many experiments and simulations have shown that the projected dynamics is sub-diffusive, i.e., the mean square displacement grows slower than linear with time. It means that the dynamics has a memory; that the free energy surface together with diffusion coefficient do not properly define the dynamics; and that such projections cannot be used to accurately describe dynamics. Here, we show that if one carefully constructs the reaction coordinate by optimizing (maximizing) its free energy profile, one can use a simple (memory-less) diffusive description. Loosely speaking, when the complex dynamics is projected onto a simple coordinate, all the complexity of the original dynamics goes into the memory of the projected dynamics. If the dynamics is projected onto the (complex) optimum reaction coordinate, all the complexity of the original dynamics is in the reaction coordinate, and the projected dynamics is simple.
| A free energy surface (FES) projected onto one or a small number of coordinates is often used to describe the equilibrium and kinetic properties of complex systems with a very large number (100 to 1,000 or more) of degrees of freedom. Studies of protein folding are an important case where this type of projected surface has been introduced and coordinates such as the number of native contacts and radius of gyration have been used [1]–[3]. Protein folding then is described as diffusion on the projected free energy surface. Diffusive dynamics is characterized by means square displacement linearly growing with time, , where D is the diffusion coefficient. For a single reaction coordinate diffusive dynamics is completely specified by the free energy profile (FEP), i.e. the free energy as a function of the coordinate and coordinate-dependent diffusion coefficient, which conveniently can be computed from conventional and cut based free energy profiles [4]. Construction of a “good” reaction coordinate (i.e. the one that preserves systems dynamics) is challenging. In many cases, the standard progress variables (e.g. number of native contacts, radius of gyration, root mean square distance from the native structure) are not good reaction coordinates, because they do not preserve the barriers on the FES and thus may mask the inherent complexity of the latter [5]. A number of methods to construct good reaction coordinates have been suggested [4], [6]–[9].
Employing the Mori-Zwanzig formalism [10], [11] one can derive generalized Langevin equations, which describe system dynamics projected on the reaction coordinates. The generalized Langevin equation contains a memory kernel, which leads to non-Markovian dynamics and subdiffusion. Subdiffusion is characterized by the mean square displacement growing slower than that for diffusion, with exponent . To completely specify dynamics in this case one has to compute the memory kernel, which is not trivial, since it requires the solution of a multidimensional partial differential equation [12]. Long-term memory in correlation functions and anomalous diffusion in proteins was observed experimentally and theoretically [13]–[23]. This raises the question whether the folding dynamics of proteins can be described as simple diffusion on the projected free energy surface, as is often done, or if one has to use more sophisticated descriptions, e.g. generalized Langevin equations [24], [25], fractional Fokker-Plank equations [26] or multiscale state space networks [19]. Here we show that if the reaction coordinate is properly optimized, then the dynamics projected onto this coordinate is diffusive, while the same dynamics projected onto a sub-optimal coordinate is sub-diffusive.
The equilibrium folding dynamics of the Go model [27] of the N-terminal domain of phage -repressor protein is analyzed [28]. Structure-based Go models containing attractive native interactions and repulsive nonnative interactions correspond to perfectly funneled energy landscapes with energetic frustration completely absent [3], [29]. A trajectory of frames (saved with = 7.5 ps) was obtained by simulating with Langevin molecular dynamics at T = 323 K and contains about 100 folding-unfolding events. The saving interval of 7.5 ps is used below as the unit of time. Note that the timescales in the simulation do not correspond directly to the timescales of the folding dynamics of the real protein because the coarse-grained model of the protein without explicit representation of the solvent is employed. Relation between the folding timescales of coarse-grained models of proteins and that of real proteins is discussed in [30]. The protein has complex FES with five basins: denatured, native, native, intermediate and intermediate and two symmetrical folding pathways [28].
Optimum one-dimensional reaction coordinates are constructed by numerically optimizing the mean first passage time to the native basin for a sufficiently broadly chosen functional form of a reaction coordinate (see Methods). Two different functional forms of reaction coordinates are considered. For each coordinate we show the cut based free energy profile (FEP) together with the exponent ; the latter is used to distinguish between diffusive and sub-diffusive dynamics (see Methods). The coordinate dependent exponent describes how the mean absolute displacement grows with time, and can be determined from the distance between computed at two different sampling intervals (see Methods); the smaller is the distance, the higher is the exponent . is equal to 1/2 for diffusive and is less than 1/2 for sub-diffusive dynamics. Each coordinate is transformed to the natural coordinate (see Methods), so that the diffusion coefficient is constant and is equal to one and diffusive dynamics is completely specified by the FEP .
The first coordinate () generalizes the number of native contacts coordinate NNC as: , where is either 1 or −1, is the distance between atoms and and is the distance threshold, when contact between the atoms is considered to be formed; is the Heaviside step function, whose value is zero for a negative argument and one for a positive argument. Figure 1a shows and for the (sub-optimal) reaction coordinate, just initialized to the NNC, i.e., with sum over pairs of atoms (ij) in the set of native contacts. The value of gives the highest barrier for the transition state for the simple variants of NNC, where the distance threshold is the same for all the native contacts. The relatively large value (inter-atom distances between atoms in the native contacts are within to Å) may be explained by the fact that the optimal reaction coordinate should better distinguish between the denatured and native basins (rather than indicate a formed native contact), which happens around the transition state and sufficiently far from the native structure. On the FEP one can notice three basins: denatured , native ; the third basin consists of a number of overlapping free energy basins. The exponent shows that the dynamics is sub-diffusive. To confirm this Figure 2 shows the mean square displacement (MSD) as a function of time () averaged over pieces of the trajectory that start from the transition state (TS) (). The MSD grows approximately as (the mean absolute displacement as ), indicating sub-diffusive dynamics. The number of folding events computed with Kramer's equation (Eq. 4) is 1200, i.e. an order of magnitude more than the actual number of 100 events. It means that the reaction coordinate is “bad” and the computed folding free energy barrier is lower than the correct one. Limited structural information can be exploited by making the distance threshold proportional to the native distance for each native contact (ij), so that . However, it does not improve the reaction coordinate since the highest barrier for the transition state, obtained at (see Figure S1 in Text S1), is similar to that obtained with the constant threshold (Figure 1a).
Figure 1b shows and for the optimized reaction coordinate . The FEP is more informative now: one can distinguish the three basins, that were overlapping on Figure 1a. The free energy of the transition state () of the optimized reaction coordinate is higher than that for the sub-optimal one (Figure 1a). The relative position of the transition state for the optimum coordinate is shifted to the left compared to the NNC coordinate which may give a misleading impression that the transition states occupy different regions of the configuration space. The optimum and NNC reaction coordinates have different coordinate dependent diffusion coefficients. When the coordinates are transformed to the natural coordinate with diffusion coefficient equal to unity the same regions of the configuration space may occupy different positions. Figure 3 shows FEPs along the reaction coordinate, which is invariant to coordinate transformation, and can be used to compare different coordinates. measures the relative partition function of the coordinate segment between 0 and x. The transition states on Figure 1 correspond to those on Figure 3, since the cut free energy profiles are invariant under coordinate transformation [4]. The transition states on Figure 3 are located at the same position, i.e., they occupy the same region of the configuration space. for the optimum coordinates are uniformly higher than that for the corresponding sub-optimum ones. coordinate, however, is of limited use to correctly represent the dynamics since the diffusion coefficient is not constant, which leads to such artifacts as sharply peaked transition states.
The scaling exponent for the optimized reaction coordinate (Figure 1b) is no longer a constant. It is a bit higher than 0.5 at the TS region () and a bit lower than 0.5 in the denatured state and at the second barrier (), indicating diffusive dynamics. After the TS is around 0.25 indicating sub-diffusive dynamics. Values of higher than 0.5 (superdiffsion) are an artifact due to over-fitting of the trajectory by the reaction coordinate. The estimated number of folding events for the optimized reaction coordinate is 168, which is quite close to the actual number. Figure 2 shows MSD for the pieces of the trajectory starting from the TS (). The MSD grows linearly with time, confirming diffusive dynamics. The reaction coordinate can be optimized in another region, e.g. by maximizing the mfpt to go from the TS () to the native structure (). In that case dynamics in the region around the second barrier () becomes diffusive, while that at the TS is back to sub-diffusive. Optimization of the reaction coordinate inside the native basin has increased the exponent in the basin from 0 to 0.3, indicating that the dynamics in the basin is still sub-diffusive. This can be due to a relatively large value of the sampling interval () of 7.5 ps, at which MSD between two subsequent snapshots is close to an equilibrium value inside the native basin. Moreover, sub-diffusive dynamics inside the basins have relatively small influence on folding dynamics, which is determined mainly by diffusive dynamics at the transitions state regions. The Text S1 shows an all-atom structure based model of the lambda repressor protein where the optimum reaction coordinate is constructed so that the dynamics is diffusive for the whole coordinate, not just around the transition state.
The second coordinate is a linear combination of all interatom distances , where is the distance between atoms and . It was initialized to be the distance between atoms and (Figure 4a). The end to end distance (the distance between and atoms), often employed in single molecule experiments, does not separate the denatured and native basins; the free energy profile along the distance is barrier-less. Figures 4 and 2 show , and MSD for and lead to the similar conclusions. For the sub-optimal dynamics is sub-diffusive at the TS (), with the exponent steeply decreasing to zero just after the TS. The exponent means that the MSD has reached the equilibrium value (at this time scale and in this region of the reaction coordinate). The estimated number of folding events is about 8700. The optimized reaction coordinate (panel b) has a higher folding barrier and shows that the dynamics is diffusive at the TS () and estimates the number of folding events as 154.
Figure 5 shows for different values of the sampling interval for the optimal and sub-optimal coordinates . The constant distance between the profiles at fixed and different (small) means that is independent of and that (see also Figure 6).
The partition function of the cut based free energy profiles () at point is defined as the number of transitions through the point [4] (see Methods). For the sufficiently large sampling intervals , when the system “flies” ballistically over the TS barrier, i.e. no recrossing events are detected, the at the TS is equal to the total number of folding events (100 here). This value denoted as () is, evidently, the same for the optimal and sub-optimal coordinates. The optimum reaction coordinate has higher at the TS compared to the sub-optimal coordinate. Hence, (at the TS) estimated as (see Methods)(1)is higher for the optimum reaction coordinate than it is for the sub-optimal one. In other words, an inadequacy of a sub-optimal reaction coordinate (low ) which leads to faster kinetics is corrected by making the dynamics sub-diffusive (slower).
We assume here that is roughly a constant for the sampling intervals between and (i.e., ), which is validated by Figure 5. However the assumption evidently breaks down for the very small time scales, when the system follows Newtons equations of motion with meaning . Thus, the sampling interval should be chosen sufficiently large so that the dynamics is in the (sub)diffusive regime.
Figure 6 shows computed at the TS as a function of for different reaction coordinates. Initially, curves have a constant slope, which is close to diffusive for the optimized reaction coordinates and to sub-diffusive for the sub-optimal reaction coordinates. The slope changes when approaches the limiting value of . The latter is not strictly constant, though its dependence on is rather weak. As increases further (, the mean life time in the basins), the probability of the system to visit another basin undetected (between successive sampling events) increases as well and (the number of detected transitions) decreases. for different reaction coordinates fall on the same curve at sufficiently large , i.e. are the same when local differences between the coordinates become negligible.
However, Figure 6 shows that the ballistic time () for different reaction coordinates is slightly different, while in deriving Eq. 1 it was assumed to be constant. To take this into account we proceed as follows. The curve (Figure 6) is approximated by two straight lines as for less than the limiting value of () and constant (), where is the time when dynamics becomes ballistic. Define , and ; where and denote, respectively, and . The diffusion coefficient is set to unity by transforming the reaction coordinate to the natural coordinate, which means that . The time can be estimated as time when mean absolute displacement is about the barrier width (), i.e. . At this time . Eliminating from the two equations, one finds(2)Taking and , one obtains equal to 0.32, 0.39 and 0.49 for equal to −9.72, −8.34 and −7.13, respectively, in reasonable agreement with Figure 6. The ballistic times are 1924, 487 and 144, respectively. From Eq. 2 it follows that the higher is the free energy barrier the higher is the exponent and the closer is the dynamics to the diffusive one.
The two optimized reaction coordinates, while having very different functional forms, show very similar behavior (at the TS regions), e.g. the width and the height of the TS barrier is the same ( on Figure 1b and Figure 4b), the MSDs are identical (Figure 2) as well as dependencies (Figure 6). This, likely, indicates that the two coordinates have converged to and closely approximate the true reaction coordinate (at the TS region). The residual difference between the estimated and the actual numbers of folding events which is due to limited statistics and insufficient flexibility of the chosen functional forms, is relatively small so it does not affect the results. The fact that the diffusive character of dynamics is determined by the height of free energy barrier , rather than the chosen functional form of the coordinate indicates the robustness of the approach. It also means that the method of constructing the optimum reaction coordinate by optimizing its FEP () [4], [6] has an advantage over the other approaches [7]–[9], in that it guarantees that the optimum reaction coordinate has dynamics closest to diffusive. Distribution of folding times is single exponential and identical for all four coordinates because folding events can be detected with high likelihood by any sufficiently good order parameter.
The analysis suggests that the higher is the free energy profile the closer is dynamics to diffusive. Evidently, the most optimal reaction coordinate is the one which has its free energy highest for every value of reaction coordinate. Consider invariant parametrization of reaction coordinate, namely the partition function of the configuration space from the initial value to the position x . The optimum reaction coordinate is the one that attains or for any , assuming that for different values of can be varied independently. This defines the optimum reaction coordinate introduced in [4], which has the largest mean first passage time. Conversely, diffusive dynamics on the constructed reaction coordinate can serve as an indication of optimality of the reaction coordinate.
To illustrate that the results presented are robust with respect to particular choice of the protein or the interaction potential, a protein with different secondary structure content (-sheet) and an all-atom structure based model of the lambda repressor protein are analyzed in Text S1. The analysis confirms that the dynamics is sub-diffusive when projected onto a sub-optimum reaction coordinate and diffusive, when projected onto the optimum reaction coordinate.
Low free energy barrier per se does not mean that the dynamics is sub-diffusive, for example, a freely diffusing particle has flat free energy profile. Dynamics should be sub-diffusive, when the reaction coordinate is sub-optimal, i.e., the free energy barrier along the coordinate is much lower than the correct one. The latter is defined either as the highest barrier attained by the optimum reaction coordinates, or as a solution of the multidimensional minimum cut problem (), which locates the transition state [4].
The analysis above just considers the dynamics around the transition state, i.e., at the top of the free energy barrier. The conclusion that the higher the free energy profile the closer the dynamics to diffusive is likely to be valid in general, e.g., for the barrier-less folding proteins. The quantitative analysis exploits the fact that at the very large sampling intervals, when the system flies ballistically over the barrier, the two free energy profiles for optimal and sub-optimal reaction coordinates are very similar, because the two coordinates distinguish equally well between the basins. It can be extended to the following general qualitative argument. The two sufficiently good reaction coordinates likely differ significantly only at relatively small spatial scales with the large scale description of the dynamics being very similar. As the sampling interval increases, the characteristic change of the reaction coordinates during the sampling interval ()) increases as well. When ()) is comparable to the large scale, so that the relative difference between the coordinate is negligible, the description of the dynamics by the two coordinates is similar and results in similar free energy profiles. Since the distance between the higher profile and the joint profile at large sampling intervals is smaller than that for the lower profile, the dynamics in former case is closer to diffusive compare to the later. It is assumed that is valid for the whole range of from the small sampling intervals, when the dynamics start to manifests itself as (sub)diffusive to the large sampling intervals, where the profiles for the different reaction coordinates become very similar. This equation connects the dynamics and the free energy profiles at these different time scales.
The model of the protein employed in the analysis is relatively simple, thus allows for extensive simulation with large number of folding-unfolding events. More realistic simulation of protein folding would include explicit representation of solvent configuration degrees of freedom. The dynamics projected on the optimum reaction coordinate constructed by considering only protein degrees of freedom might be sub-diffusive because neglected solvent degrees of freedom could be important.
The analysis suggests that without specifying the reaction coordinate, the question why the dynamics is sub-diffusive is rather ill-posed. It is more appropriate to ask: is it possible, for a given trajectory, to construct the optimum reaction coordinate, so that the projected dynamics is diffusive?
In conclusion, we have shown that dynamics projected onto a reaction coordinate can be diffusive or sub-diffusive depending on the coordinate employed for the projection. If one has a flexibility in choosing the reaction coordinate, e.g. when describing protein folding, dynamics can be made diffusive (or close to it) by optimizing the reaction coordinate (making higher). When the coordinate describing the process is specified and can not be varied, for example, the donor-acceptor distance in the single molecule FRET or ET experiments [24], [31] or the mean square displacement in the neutron scattering experiments [13], the dynamics is likely to be sub-diffusive [13], [14], [31]. However, this does not necessarily mean that the dynamics per se is sub-diffusive. A properly chosen reaction coordinate (too complex to realize in experiment) may show that dynamics of transition between free energy basins is diffusive. A relatively small deficiency of the putative reaction coordinate (difference in 1 kT in free energy () of the folding barrier) is sufficient to make the dynamics sub-diffusive. Hence, one should model protein dynamics as diffusion on a putative reaction coordinate [32], [33] with care, because, it is very likely that the coordinate is sub-optimal, unless it has been specifically constructed (optimized) [5], [28].
The conventional way to construct the FEP, given the projection of a trajectory onto a reaction coordinate (the time-series of the value of the reaction coordinate) , is to compute a histogram and estimate the partition function (probability density) as , where is the number of time-series points in bin and is the size of the bin. The free energy can then be found as . The partition function of the cut based free energy profile [4] at point is defined as the number of transitions through that point, i.e. , where is the sampling interval and is the Heaviside step function; . Assuming that the is approximately constant on the distance of the mean absolute displacement , one can derive the following expression(3)where is the mean absolute displacement during sampling interval; for diffusive dynamics it gives .
A reaction coordinate (x) with a variable diffusion coefficient can be transformed to coordinate (y), called the natural coordinate [4], so that the diffusion coefficient is constant and equal to unity, by numerically integrating ; i.e. that .
Other approaches have been suggested to characterize diffusive dynamics by computing the free energy profile together with the coordinate dependent diffusion coefficient [32], [34], [35]. It is not clear, however, if they can be used to characterize the sub-diffusive regime.
It is reasonable to assume that any “bad” choice of reaction coordinate, when different parts of the configuration space overlaps at projection onto this coordinate, will result in faster kinetics, i.e. in a smaller mean first passage time (mfpt). Clearly, the longest mfpt is obtained on the original FES or from a projection where no such overlapping occurs. Hence, we define the optimum reaction coordinate as the one that has the longest mfpt, which can be computed by Kramer's equation [4](4)
The optimum reaction coordinates are constructed by numerically optimizing the mfpt functional for a sufficiently broadly chosen functional form of reaction coordinate. Starting with the initial set of parameters, which are sufficient to distinguish between the two free energy basins, the coordinate is iteratively improved by changing parameters and accepting the change if mfpt is increased. For the first reaction coordinate we pick a random pair of atoms , scan the whole parameter space for the pair ( and ) and select the one that gives the highest mfpt. For the second reaction coordinate we pick a random pair , scan the whole parameter space for the pair ( for and is a random number uniformly distributed between 0 and 1) and select the one that gives the highest mfpt. For the given values of parameters the mfpt is computed by first computing and and then numerically integrating Eq. 4. Alternatively one may minimize the number of transitions, the quantity related to mfpt as , where is the partition function of basin A and is the position of the transition state between basins A and B.
For subdiffusion, the mean absolute displacement no longer scales as , but rather as . The exponent (possibly coordinate dependent), can be determined by comparing at two different sampling intervals (see Eq. 3). For a trajectory with fixed length and varying sampling interval (when ) it is equal to(5)Since is invariant with respect to nonlinear coordinate transformation, the scaling exponent computed by Eq. 5 is also invariant, while computed from or are not invariant and are computed here after the coordinate has been transformed to the natural reaction coordinate.
|
10.1371/journal.pgen.1003168 | Approaching the Functional Annotation of Fungal Virulence Factors Using Cross-Species Genetic Interaction Profiling | In many human fungal pathogens, genes required for disease remain largely unannotated, limiting the impact of virulence gene discovery efforts. We tested the utility of a cross-species genetic interaction profiling approach to obtain clues to the molecular function of unannotated pathogenicity factors in the human pathogen Cryptococcus neoformans. This approach involves expression of C. neoformans genes of interest in each member of the Saccharomyces cerevisiae gene deletion library, quantification of their impact on growth, and calculation of the cross-species genetic interaction profiles. To develop functional predictions, we computed and analyzed the correlations of these profiles with existing genetic interaction profiles of S. cerevisiae deletion mutants. For C. neoformans LIV7, which has no S. cerevisiae ortholog, this profiling approach predicted an unanticipated role in the Golgi apparatus. Validation studies in C. neoformans demonstrated that Liv7 is a functional Golgi factor where it promotes the suppression of the exposure of a specific immunostimulatory molecule, mannose, on the cell surface, thereby inhibiting phagocytosis. The genetic interaction profile of another pathogenicity gene that lacks an S. cerevisiae ortholog, LIV6, strongly predicted a role in endosome function. This prediction was also supported by studies of the corresponding C. neoformans null mutant. Our results demonstrate the utility of quantitative cross-species genetic interaction profiling for the functional annotation of fungal pathogenicity proteins of unknown function including, surprisingly, those that are not conserved in sequence across fungi.
| HIV/AIDS patients, cancer chemotherapy patients, and organ transplant recipients are highly susceptible to infection by opportunistic fungal pathogens, organisms common in the environment that are harmless to normal individuals. Understanding how these pathogens cause disease requires the identification of genes required for virulence and the determination of their molecular function. Our work addresses the latter problem using the yeast Cryptococcus neoformans, which is estimated to cause 600,000 deaths annually worldwide in the HIV/AIDS population. We describe a method for determining gene function in which C. neoformans genes are expressed in deletion mutants of all nonessential genes of the well-studied model yeast S. cerevisiae. By examining the impact on growth (enhancement or suppression) we generated “cross-species” genetic interaction profiles. We compared these profiles to the published genetic interaction profiles of S. cerevisiae deletion mutants to identify those with correlated patterns of genetic interactions. We hypothesized that the known functions of S. cerevisiae genes with correlated profiles could predict the function of the pathogen gene. Indeed, experimental tests in C. neoformans for two pathogenicity genes of previously unknown function found the functional predictions obtained from genetic interaction profiles to be accurate, demonstrating the utility of the cross-species approach.
| Diseases produced by fungal infections are challenging to diagnose and treat, making these infections a major cause of morbidity and mortality worldwide [1], [2]. Genetics and genomics have led to the identification of numerous pathogen genes required for replication in the mammalian host [3]–[7]. Unfortunately, many, if not most, identified virulence genes lack in vitro phenotypes that could explain their effects in the host [3]–[8], and the predicted protein sequences often offer few clues to function. Thus, our power to identify pathogen genes required for disease far outstrips our ability to understand their molecular function in the host.
Historically, the expression of human genes in the model yeasts Saccharomyces cerevisiae and Schizosaccharomyces pombe has been used as a tool to identify specific genes and to determine their cellular function [9]–[14]. In a classic example, complementation of a fission yeast cdc2 mutant was used to identify human Cdk1 [11]. More recently, a number of groups have combined the expression of foreign genes with high-throughout tools available in S. cerevisiae to identify suppressor genes to obtain insights into the function of human proteins, ranging from those involved in neurodegeneration to cancer [9], [11], [14].
Likewise, expression of viral and bacterial proteins in yeast, coupled with subsequent genetic analysis, has proven to be informative. For example, the genes responsible for biosynthesis of the eEF2 modification diphthamide were identified in selections for resistance to the F2 fragment of diphtheria toxin [15]. Identification of S. cerevisiae gene deletion mutants hypersensitive to the expression of the Shigella virulence factor OspF, a type III secretion substrate, coupled with transcriptional profiling experiments, led to the identification of the cell wall integrity MAP kinase pathway as a likely target of OspF in yeast [10], [13]. Importantly, the latter study took advantage of phenotypic information for yeast deletion mutants available at that time to obtain clues to gene function [10].
The construction of a library of all nonessential gene deletions for S. cerevisiae [8] together with the development of genetic selections led to the development of the synthetic genetic array (SGA) method for quantitatively measuring genetic interactions on a genome scale [16], [17]. This approach has facilitated the systematic annotation of gene function in S. cerevisiae [18], [19]. Genetic interaction, or epistasis, measures the degree to which two genes affect each other [16], and is measured by comparing the phenotype of a double mutant to that of the two corresponding single mutants. Genes that act in the same pathway display similar patterns of genetic interactions with other genes [16]–[19]. Recently, the large-scale application of these methods led to production of a remarkable genome-scale genetic interaction map based on the analysis of ∼5.4 million gene pairs. Such a comprehensive genetic interaction dataset has only been described to date for the model yeast S. cerevisiae [19].
Below we test the utility a cross-species genetic interaction approach for fungal pathogen gene annotation that combines expression of pathogen genes in S. cerevisiae with genetic interaction profiling. We used genes from the human pathogen Cryptococcus neoformans, an opportunistic basidiomycete fungal pathogen that is very distantly related to the model yeasts S. cerevisiae and S. pombe. C. neoformans is the most common cause of fungal meningitis in humans, and among the most important causes of morbidity and mortality in AIDS patients, leading to ∼1 million infections and ∼600,000 deaths annually in sub-Saharan African alone [1]. Our laboratory previously generated a library of 1201 gene deletion strains and used a signature-tagged mutagenesis approach to identify genes required for pathogen fitness during experimental infection of mice [5]. In addition to identifying new genes required for the synthesis of known virulence factors, these studies identified several dozen genes required for virulence whose mutation failed to yield in vitro phenotypes that could explain its role in the host.
As a proof-of-principle, we expressed six C. neoformans genes of interest in each member of the S. cerevisiae deletion library and quantified their impact on fitness, thereby producing cross-species genetic interaction profiles. We exploited their similarities to existing S. cerevisiae knockout genetic profiles to predict possible functions for each C. neoformans protein. For two of these C. neoformans proteins, Liv6 and Liv7, we describe validation experiments that support the functional assignment. For Liv7, additional experiments connect its newly identified function to the evasion of phagocytosis, an important virulence trait. The cross-species genetic interaction profiling approach described here offers a generalizable avenue toward the functional annotation of pathogenicity factors of fungal agents of infectious disease.
We sought to develop a generic approach for developing testable hypotheses for the function of novel C. neoformans virulence genes by taking advantage of the methods and datasets that exist in S. cerevisiae. We created S. cerevisiae strains that each expressed a C. neoformans gene of interest (described further below). We crossed these to the S. cerevisiae gene deletion library using automated SGA methods and measured fitness of the progeny strains using high-throughput colony imaging methods [16], [17] (Figure 1). Measurements (n = 8 per double mutant) were converted into significance scores (S-scores) [20] (See Methods). We refer to these data as a “cross-species genetic interaction profile” which is the set of quantitative genetic interactions between strains expressing a particular C. neoformans gene and each S. cerevisiae deletion mutant. We calculated correlations between these cross-species profiles and the available genetic interaction profiles of deletion mutants in S. cerevisiae [21]. We reasoned that the expression of a C. neoformans gene could, in some cases, produce dominant-negative effect and produce genetic interaction profiles that correlate positively with those of S. cerevisiae gene deletions that function in the homologous pathway. Alternatively, the expression C. neoformans gene might have a dominant-positive effect, producing a profile that anti-correlates with those of S. cerevisiae deletions mutants in the same pathway. Scenarios on which both behaviors occurred could also be imagined. We further expected that the expression of some, but not all, of C. neoformans genes would produce profiles that would allow us to develop experimentally testable hypothesis for gene function.
We focused on six C. neoformans genes (Figure 2 and Table S1), four of which (LIV5, LIV6, LIV7, and LIV13) our previously work identified as necessary for growth in a murine infection model [5]. Two others, BLP1 and MEP1, are targets of Gat201 [22], a master transcriptional regulator of virulence [5], [22]. Blp1 is important for C. neoformans to evade phagocytosis by macrophages. Four of these genes (LIV6, LIV7, MEP1, and BLP1) lack S. cerevisiae orthologs. Several contain conserved domains identified by BLAST [23], but the function of these domains are poorly understood (Figure 2 and Table S1). The application of PHYRE, a threading-based structure prediction algorithm, provided information for only Liv6, which it predicts to be structurally related to a lectin [24].
We generated cross-species genetic interaction profiles using S. cerevisiae strains carrying two control constructs and six different bait constructs: pGPD (promoter-only control), pGPD-GFP (nonspecific protein control), pGPD-LIV5, pGPD-LIV6, pGPD-LIV7, pGPD-LIV13, pGPD-MEP1, and pGPD-BLP1. To ensure reasonable expression levels (see Materials and Methods) used the strong GPD1 promoter because the C. neoformans genome is GC-rich [25] compared to S. cerevisiae [26], which is anticipated to inhibit protein translational efficiency due to differences in codon usage and an increased propensity to form inhibitor RNA structures [27]. We calculated Pearson correlations (correlation score) to compare cross-species genetic interaction profiles with the previously described genetic interaction profiles of produced by crosses of 1712×3885 S. cerevisiae gene deletions [19]. To avoid potentially spurious correlations, we filtered out correlations with S. cerevisiae deletions whose profiles yielded significant correlations with either of the two control baits. Significance testing revealed that correlations with a value of greater than 0.08 are highly significant (P<0.001, two-tailed test, Bonferroni-corrected for multiple hypothesis testing). Quantile-quantile plots of the correlations with S. cerevisiae deletions versus standard normal quantiles revealed outliers on one or both tails for all baits (Figure S1). We focused on correlations that departed from the mean by at least three standard deviations (|Z|>3). This conservative strategy yielded from 2–15 hits, depending on the bait (Table 1). The profile of LIV7 displays the largest number hits, and their identities strongly points to a role in Golgi transport, a prediction whose validation via experiments in C. neoformans is described below. The LIV6 profile correlates positively and negatively with two S. cerevisiae genetic profiles, those of deletions in SYN8 and ECM21, respectively. Strikingly, both genes play a role in endosome transport and/or fusion [28], [29], predicting a role for Liv6 in these processes in C. neoformans. Support for this prediction via experiments in C. neoformans is also described in this paper. Several other profile hits were obtained, but have yet to validated. The Liv5 profile correlates with those of deletions affecting the cell cycle and autophagy [30]–[32] and the Liv13 profile negatively correlates with the genetic profiles of knockouts involved in alleviating protein folding stress [33]–[35]. The profile of the MEP1 metalloprotease correlates with that of a knockout in a S. cerevisiae metalloprotease of a different family, YBR075W [23], as well as proteins involved in nucleotide and RNA metabolism. Finally, the genetic interaction profile of the S. cerevisiae strain expressing Blp1 positively correlates with that of the deletion of an S. cerevisiae gene, ETR1, that has a role in fatty acid synthesis. This observation may be related to the Rare Lipoprotein A (RlpA) domain prediction for Blp1 (Table S1).
Liv7 (Figure 2) is a 330-residue protein that contains a DUF3752 domain, which is annotated as a conserved domain of unknown function [36]. The profile of the S. cerevisiae strain expressing LIV7 displays the strongest three positive correlations with the published genetic interaction profiles of S. cerevisiae gene deletions trs33Δ, tlg2Δ, and vps51Δ (Figure 3A). Strikingly, all three of these genes function in transport events that involve the Golgi apparatus (Figure 3B). Trs33 is one of two nonessential subunits of the TRAPP complex, an essential vesicle tethering complex involved in ER-to-Golgi transport [37]. Vps51 is a member of the GARP complex, another vesicle tethering complex that promotes endosome-to-Golgi transport and retrograde transport within the Golgi [38]. Tlg2 is a t-SNARE that is important vesicle fusion within the Golgi [39]. These data make a strong prediction that the function of the unannotated Liv7 protein is in transport events involving the Golgi apparatus. Below we describe experiments in C. neoformans that support this prediction and additional follow-up experiments that led us to find that the Liv7 protein is required for the suppression mannose exposure on the cell surface and the suppression of mannose-dependent phagocytosis by mammalian macrophages.
Given that Trs33 is a nonessential subunit of TRAPP, we anticipated that if Liv7 functions to promote TRAPP function in C. neoformans, that its gene deletion mutation should display a synthetic lethal or synthetic sick phenotype with a deletion of TRS33 in C. neoformans. We tested this prediction by creating single and double targeted knockouts of LIV7 and an ortholog of TRS33 we identified in the C. neoformans genome. We found that wild-type, liv7Δ, and trs33Δ strains all grow at approximately the same rate, with a doubling time of two hours (Figure 3C. In contrast, the liv7Δ trs33Δ double mutant cells display a severe growth defect, having a doubling time of four hours (Figure 3C). These data demonstrate that LIV7 and TRS33 interact genetically in C. neoformans, as inferred from analysis of the cross-species genetic interaction profiles described above. We also constructed a deletion in the gene coding for a member of the C. neoformans GARP complex, Vps52 (we were unable to delete the C. neoformans VPS51 gene), and found that it displayed a growth phenotype similar to that of the liv7Δ trs33Δ double mutant.
We next tested the hypothesis that LIV7 functions in the ER-Golgi system by using a chemical biology approach that takes advantage of the small molecule Brefeldin A (BFA). BFA is a fungal secondary metabolite that inhibits eukaryotic Sec7-family guanine nucleotide exchange factors that are involved in vesicle transport and themselves localize to the membranes of the ER and Golgi apparatus [40]–[42]. BFA blocks anterograde transport from the ER to the Golgi, fusion of ER and Golgi compartments, and loss of Golgi apparatus itself [40], [41]. We grew strains with and without a growth-inhibitory, sublethal concentration (40 µg/ml) of BFA (Figure 3D). Wild-type, liv7Δ, and trs33Δ show identical responses to BFA: a sharp increase in doubling time from two hours to over 12 hours (p≤0.01) (Figure 3D). liv7Δ trs33Δ mutants, which already exhibit slow growth (p≤0.01), do not show any further increase in their four-hour doubling time. The resistance to BFA exhibited by liv7Δ trs33Δ double mutants demonstrates that either Liv7 or Trs33 function is required for BFA to inhibit cell growth (Figure 3D). These data could be explained if Liv7 and Trs33 have a severe defect in the assembly and/or function of the Golgi apparatus (which we show to be the case below). In this scenario, the growth rate of such cells would thus not be affected by BFA since the have greatly reduced the target organelle most strongly affected by the drug. A more formal statement of such a model would be that in the absence of BFA, Liv7 and Trs33 redundantly promote growth (via a role in Golgi biogenesis), but in the presence of the drug, cells convert to a state in which either Liv7 or Trs33 inhibits growth (Figure 3E). This genetic behavior is analogous to that of the S. cerevisiae MAP kinase Kss1, which is converted from an inhibitor of filamentous growth to an activator via phosphorylation by the upstream MAP kinase Ste7 [43]. The vps52Δ mutant also displays resistance to BFA (Figure 3D).
To further test the hypothesis that Liv7 functions in the Golgi, we examined the colocalization of an mCherry-tagged version of Liv7 with compartment markers. The levels of Liv7 protein are low and we could not detect it by Western or microscopy under yeast culture conditions (data not shown). However, under the same tissue culture conditions we use to study pathogen phagocytosis (DMEM, 5% CO2, without shaking), we observed a punctate Liv7-mCherry signal that was well above background signal observed in an untagged control strain (Figure 4A–4C). To label the ER and Golgi, we briefly incubated cells with a fluorescent derivative of Brefeldin A (fBFA) [44] at sub-inhibitory concentrations (0.5 µg/ml for 40 min, 80-fold less than the minimal inhibitory concentration). To confirm that the compound was labeling the anticipated compartments, we stained a C. neoformans strain with fBFA carrying a mCherry-tagged version of the conserved Erd2 protein, which is found in both ER and Golgi compartments [45] and found that the fBFA signal colocalizes with the Erd2 signal (Figure S2). Importantly, the Liv7-mCherry colocalizes with the fBFA signal. The respective puncta co-localize in almost 100% cells that display signals for both fluorophores (Figure 4A–4C). As a control, we stained mitochondria with MitoTracker did not observe co-localization with Liv7-mCherry signal (Figure S3).
To test whether mutations LIV7 and TRS33 impact the formation of the ER and Golgi we stained single and double mutants with fBFA. Wild-type, liv7Δ, and trs33Δ strains showed similar cytoplasmic punctate staining (Figure 4D–4G). However, liv7Δ trs33Δ mutants did not exhibit detectable fBFA staining (Figure 4G), consistent with a severe defect in organelle formation. These data show that Liv7 is important in promoting organelle formation in cells lacking Trs33. Together with the impact of the mutants on BFA sensitivity and the colocalization of Liv7 with fBFA, these observations provide strong evidence for a role for Liv7 in Golgi function.
Key functions of the Golgi include the sorting and modification of proteins and the biosynthesis of polysaccharides. The cell surface of microbes often contain pathogen-associated molecular patterns (PAMPs), molecular signatures that are recognized by the mammalian immune system [46]. Previous studies of the human fungal pathogen Candida albicans has shown that there are mechanisms by which this pathogen masks PAMPs to order to avoid recognition by neutrophils [47]. To test whether LIV7 or TRS33 are involved in PAMP exposure, we examined the cell surface exposure of two well-established fungal PAMPs, mannose and β-glucan. These experiments were performed in tissue culture conditions, which modestly induces production of the C. neoformans polysaccharide capsule. In addition, we stained cells for the glucuronoxylomannan (GXM) component of the capsule and as well as the cell wall polysaccharide chitin. We used an antibody to detect glucuronoxylomannan (GXM) component of C. neoformans polysaccharide capsule (Figure 5A), the lectin CBP to detect chitin (Figure 4A), an antibody to detect β-glucan (Figure 5B), and the lectin concanavalin A (conA) to detect exposure of mannose (Figure 5C). Wild-type, liv7Δ, and trs33Δ all showed similar PAMP exposure, with modest staining of β-glucan and mannose under tissue culture growth conditions (Figure 5A–5C). We also observed modest staining using reagents that detect chitin and GXM (Figure 5A–5C).
In contrast, we observed strikingly different results in liv7Δ trs33Δ double mutant cells and in the vps52Δ mutant. Most remarkably, we observed a dramatic increase in mannose exposure in these mutants as measured by conA staining (Figure 5C). In contrast, GXM or β-glucan staining is virtually eliminated (Figure 5A, 5B). The chitin signal is reduced in intensity and localizes to a focus at the cell pole. The increase in conA signal cannot be explained by the lack of capsular GXM in the double mutant, as GXM- and capsule-deficient mutant strains cap10Δ [48] and cap60Δ [48], [49] do not exhibit this phenotype (Figure 5C). These data suggest that LIV7 and TRS33 act redundantly in the transport of molecules required to suppress the exposure of mannose on the cell surface and that the integrity of the GARP complex is also required for this process.
Mannose and mannoproteins (mannan) are highly immunogenic [50], and, consequently, masking their exposure would be expected to be critical for pathogen evasion of the host immune system. It is well-established that C. neoformans evades phagocytosis by macrophages (anti-phagocytosis), the first line of host immune defense, and that this attribute is important for mammalian infection [5], [51]. In prior work, we demonstrated that C. neoformans evades phagocytosis by at least two pathways, one requiring capsule production and a second that is independent of capsule production and programmed by the transcriptional regulators Gat201 and Gat204 [5], [22]. Strikingly, mutations that abrogate capsule formation and mutations in the capsule-independent pathway do not result in detectable exposure of mannose or β-glucan on the cell surface, suggesting that these pathways do not act by masking these known PAMPs, even though their exposure would be anticipated to activate phagocytic receptors on macrophages.
Since we observed a dramatic increase in mannose exposure in the liv7Δ trs33Δ double mutant, we anticipated that it would display high levels of phagocytosis. To test this, we cultured wild-type, liv7Δ, trs33Δ, and liv7Δ trs33Δ C. neoformans cells with RAW 264.7 cells, a murine macrophage cell line. To test the potential impact of opsonization, C. neoformans strains were treated or not with fetal bovine serum prior to incubation (Figure 6A). Wild-type C. neoformans displays a low level of phagocytosis (4% macrophages with associated C. neoformans cells) that increased (∼17%) upon opsonization (p≤5×10−4). As anticipated from their mannose exposure, liv7Δ trs33Δ mutants and vps52Δ mutants show high levels (∼80%) of phagocytosis regardless of opsonization (p≤2×10−3). The lack of anti-phagocytosis activity by liv7Δ trs33Δ cells and vps52Δ cells is not solely due to lack of GXM, as GXM mutants cap10Δ and cap60Δ show increased association with macrophages (p≤3×10−3) but not to the same extent as liv7Δ trs33Δ cells, and are still sensitive to opsonization (p≤5×10−3).
Surprisingly, even though there was no gross increase in mannose exposure in the liv7Δ single mutant, it displays a small (11%) but reproducible increase in phagocytosis without opsonization (p≤5×10−5) and no further increase with opsonization. In contrast, the trs33Δ mutant does not show this phenotype.
The single liv7Δ and trs33Δ mutants show distinct phagocytosis phenotypes yet the mannose exposure (as determined by conA staining) of both mutants is not distinguishable from wild-type. We hypothesized that liv7Δ cells might exhibit an increase in mannose or mannan on their surface not present in trs33Δ cells that is too subtle to detect by microscopy-based lectin staining assays. A functional prediction of this hypothesis is that the increase in phagocytosis of the liv7Δ mutant should be specifically blocked by an excess of free mannose. We performed phagocytosis assays using unopsonized C. neoformans cells and added either soluble mannose (to block recognition of mannose and mannans by macrophage mannose-recognition receptors) or laminarin (a control oligosaccharide that blocks recognition of beta-glucan) [52]. Strikingly mannose, but not laminarin, blocks the increased phagocytosis of liv7Δ mutants (p≤10−3) (Figure 6B). Mannose addition also partially rescues the anti-phagocytosis defect of liv7Δ trs33Δ cells (p≤10−3). Importantly, this treatment did not impact phagocytic index of gat204Δ cells (Figure 6B), a mutant we described previously that produces similar increase in phagocytosis, supporting the view that Liv7 and Gat204 function via distinct mechanisms [22].
The genetic interaction profile produced by the expression of LIV6 in S. cerevisiae shows positive and negative correlations with the corresponding profiles of the S. cerevisiae syn8Δ and ecm21Δ deletion mutants, respectively (Table 1 and Figure 7A). These genes act in endosome transport and/or fusion [28], [29], a process that mediates transport from either the plasma membrane or the late Golgi to the vacuole [53]. These correlations predict that Liv6 participates in endosomal functions in C. neoformans.
We first tested this prediction by assessing the impact of Liv6 on vacuole number. S. cerevisiae genes involved sorting to the vacuole include those that function in endosome biology and often impact vacuole number and morphology [54]–[56]. Vacuoles can be detected by staining with LysoTracker Green (Invitrogen), a dye that is taken up by the cell during endocytosis and fluoresces in acidified compartments, including endosomal vesicles, and typically strains the outer rims of vacuoles. Wild-type C. neoformans cells grown in yeast culture conditions and strained with LysoTracker show efficient uptake, many internal vesicles, and rim-stained vacuoles (Figure 7B). This pattern is remarkably similar to those reported for S. cerevisiae stained with LysoTracker or FM4-64 [56], [57], an older vital stain used to study protein sorting to the vacuole [57]. Strikingly, liv6Δ cells consistently exhibit a greater number of vacuoles than wild-type cells (p<0.005) (Figure 7C, 7D). Notably, the S. cerevisiae gene SYN8, whose deletion mutant's genetic interaction profile displays a positive correlation with the profile produced by LIV6 expression (Table 1), has been reported to function with another SNARE to promote normal vacuolar morphology [28]. The increase in vacuolar number seen in liv6Δ cells is highly specific, as knockout mutant in any of the bait genes did not exhibit a change in vacuole number (Figure 7D).
We next exploited the aminoglycoside antibiotic neomycin, which interferes with eukaryotic endosomal activity by binding phosphytidylinositol phosphates [58], [59] necessary for endosome function [60]. As a consequence, loss-of-function mutations in S. cerevisiae genes involved in endosome function [61], [62] are sensitive to neomycin [61]. Supporting a role for Liv6 in endosome function, we found that C. neoformans liv6Δ knockout mutants are sensitive to this drug (Figure 7E). Cells lacking LIV7 display a subtle reproducible neomycin resistance which could be due to altered cell permeability, a characteristic of neomycin-resistant S. cerevisiae strains [63], [64]. Knockout mutants in the other bait genes do not display a change in sensitivity to this compound. liv6Δ cells do not exhibit a growth defect on fluconazole, suggesting that their growth defect is specific to neomycin. Together, the changes in vacuole number and sensitivity to neomycin in produced by the liv6Δ mutation support the prediction from cross-species genetic profiles of a role for Liv6 in the endovacuolar system of C. neoformans.
Genetic approaches to understanding mechanisms of virulence in human fungal pathogens can efficiently identify genes necessary for pathogens to cause disease. However, a key roadblock to progress is the lack of tools that can help define the function of a gene product when its predicted sequence offers few clues to its biochemical function, a common occurrence. We described here a case study of a cross-species genetic interaction profiling approach to develop testable hypotheses for the function of fungal virulence factors of unknown function. Notably, this proof-of-principle study shows that the approach can provide information on fungal pathogenicity factors that lack S. cerevisiae orthologs. Although many studies have used S. cerevisiae to investigate the function of foreign genes [9]–[11], [13], [14], the cross-species genetic interaction profile used here represents an application of quantitative genetic profiling of foreign proteins in S. cerevisiae coupled with comparison to recently described genetic map of S. cerevisiae [19] to the problem of annotation of fungal virulence factors. Because S. cerevisiae is a fungus, we anticipate that this approach may be particularly useful for fungal genes but that the method may also find utility in the study of bacterial and viral proteins that impact conserved intracellular processes in eukaryotic host cells.
Our approach involves expression in S. cerevisiae of cDNAs encoding Cryptococcus neoformans virulence factors identified in systematic genetic screens; the generation of genetic profiles by assessing the effect of C. neoformans gene expression in the context of each nonessential S. cerevisiae deletion mutants; and, correlation analysis with the existing database of genetic interactions to develop testable functional hypotheses. As mentioned above, one mechanism whereby expression of a C. neoformans gene could produce impact S. cerevisiae would be “dominant-negative” effect thereby inhibiting the activity of an S. cerevisiae pathway. Our results with LIV7 in both S. cerevisiae and C. neoformans are consistent with this scenario. The expression of LIV7 in S. cerevisiae produces a profile that correlates with that of the S. cerevisiae trs33Δ deletion mutant, but in C. neoformans, the liv7Δ mutation produces a synthetic phenotype with the trs33Δ mutation. Alternatively, expression of a C. neoformans gene product could act in a “dominant-active” fashion to increase the activity of a pathway which might result in a negative correlation with the profile of a gene deletion in the corresponding pathway. With Liv6, we observed both positive and negative correlations that led us to test a role in endosome function. Although we have focused on the extensive deletion mutant genetic interaction dataset [19], comparisons of the cross-species profiles generated here with genetic interaction profiles produced using chemicals [65], [66] and/or overexpressed genes [67], [68] will likely be equally useful as these approaches are applied on a larger scale. Thus, the analysis of correlations between cross-species genetic interaction profiles and existing “within-species” genetic interaction profiles offers a tool for generating testable predictions for pathways in which foreign genes operate.
The genetic profiling studies and validation experiments described in this paper provide new information on two C. neoformans pathogenicity factors identified previously, Liv7 and Liv6. These proteins lack orthologs in S. cerevisiae and lack orthologs of known function in other species. Our studies of Liv7 suggest it functions in Golgi transport in a process that suppresses the exposure of the PAMP mannose on the cell surface (Figure 6C). The increased phagocytosis phenotype of the liv7Δ single mutant and its specific suppression by soluble mannose appears specific to liv7Δ cells and is specific to mannose versus other carbohydrates (Figure 6B). The anti-phagocytic properties of C. neoformans are critical for mammalian infection [5], [22], [69] and the capsule is important for the anti-phagocytosis activity of opsonized C. neoformans cells [51]. Our previous work identified a capsule-independent pathway necessary for anti-phagocytosis under unopsonized conditions [22]. The suppression cell surface exposure of PAMP mannose appears to represent a third anti-phagocytosis pathway (Figure 6C) since mannose does not rescue the anti-phagocytic defect of gat204Δ cells (Figure 6B), which are defective in the capsule-independent anti-phagocytosis pathway [22]. This argument is supported by the observation that cap10Δ and cap60Δ cells, which lack GXM [48], [49], do not exhibit increased conA staining (Figure 5C). We suggest that Liv7 is important for mammalian infection [5] because it inhibits macrophage recognition of mannose-containing patterns on the C. neoformans cell surface (Figure 6C). Although our studies of Liv6 point to a role in endosome biology that impacts neomycin resistance and vacuole number, understanding how this function relates to its role in pathogen fitness in the host will require further investigation. One possibility is that Liv6 is involved in the endocytic uptake of limiting factors required for proliferation from the host milieu.
One anticipates that functional annotation of fungal virulence factors identified genetically will continue to be a major challenge for the future. The approach described here represents one generic tool that could be applied to this problem on a larger scale. We expect that a substantial number of virulence genes of unknown function in fungal pathogens will impinge on conserved cellular processes and that their genetic profiling in S. cerevisiae could therefore yield testable functional predictions in a significant number of cases. The cross-species interaction profiling could also be useful for studying genes from highly virulent pathogens that are difficult to work with due to the requirement for extensive containment.
We inserted the GPD1 promoter region, our C. neoformans cDNA of interest, and a NAT resistance marker into the multicloning site of pRS316. For recombination into S. cerevisiae, we cut with a restriction enzyme that cleaved within the URA3 locus, the transformed the linearized vector into S. cerevisiae using standard lithium acetate-based transformation techniques. We verified expression of C. neoformans genes by extracting total RNA from log-phase S. cerevisiae cultures grown at 30°C in YNB, selecting for mRNA, and making cDNA as previously described [70]. Expression of C. neoformans genes was verified by qPCR performed as previously described [70].
We expressed each C. neoformans gene under the GPD1 promoter and we measured RNA by qRT-PCR (Figure S4). We then measured the levels of BUD1 mRNA in the same RNA preparation. BUD1 is a small GTPase expressed at low levels [71]–[73] along with its two co-regulators BUD2 and BUD5 [74], [75]. We used published data on the molecules of BUD1 RNA per cell averaged with co-regulators BUD2 and BUD5 [71]–[73] to estimate the number of RNA molecules per cell for C. neoformans genes from the ratios in Figure S4, then calculated its rank position compared to other S. cerevisiae genes. BLP1 and LIV6 were in the lowest 10% of genes with detectable RNA (∼5090 of ∼6580 genes had detectable RNA [71]–[73]). LIV5 and LIV7 were in the 10–20th percentile, as were the BUD genes. MEP1 was the best expressed of the C. neoformans genes, (∼35th percentile). LIV13 was expressed based on the increase in LIV13 primer products with and without RT (data not shown) but not compared to BUD1.
We performed SGA screens as described in Tong et al [16], [17] using a RoToR colony pinning robot (Singer Instruments). All screen plates were scanned on a flatbed scanner with autofocus. We extracted colony size data using the publicly available ScreenMill software [76]. We then adjusted the raw colony size data to control for plate position, edge effects, and slow growth of knockout mutants using the S-score method developed by Collins et al [20]. The final S-scores, one for each double mutant strain, indicate the strength of the genetic interaction (absolute value) and whether the interaction is synthetically sick (negative numbers) or buffered (positive numbers) [20]. We then adjusted S-scores so that they were on a scale between −1.0 and 1.0 and calculated the Pearson correlation between S-scores and ε-scores from Constanzo et al [19]. We calculated p-values of the Pearson correlations by calculating the Z-score of the Pearson correlation for each interaction, then using the Z-score to determine the p-value of each interaction.
C. neoformans was routinely grown in yeast culture conditions in either YPAD (1% yeast extract, 2% peptone, 2% glucose, 0.015% L-tryptophan, 0.004% adenine) or yeast nitrogen base (YNB) (Difco). Strain construction and genetic manipulation was previously described [5]. Whenever more than one knockout mutant for a single gene is shown, mutants were made by independent transformations. Growth curves (Figure 3A) were performed in YNB at 30°C by taking measurements of OD600 every two hours for 10 hrs. The growth curve was repeated three times and representative data are shown. When C. neoformans cells were grown in tissue culture conditions, they were first grown overnight to saturation in YNB, then washed once in 1× PBS and resuspended at a density of 1 OD600/ml (∼1.7×107 cells/ml) in DMEM, then incubated for the specified amount of time in 5% CO2 at 37°C.
Samples were grown in overnight in YNB at 30°C, then subcultured to OD600 = 0.2. BFA or DMSO (-BFA control) was added to each culture and the OD600 taken every hour for 10 hr. Doubling time was calculated over the interval from 4–8 hr. The treatment curve was repeated three times and the data shown are averages of the three experiments.
Samples were grown overnight in YNB at 30°C, then washed 3× in 1× PBS and resuspended at 1 OD/ml in DMEM, then incubated 16 hr under tissue culture conditions (5% CO2, 37°C). Samples were then either imaged (unstained samples) or fBFA (Life Technologies) was added to the medium to a final concentration of 0.5 µg/ml. fBFA samples were incubated 40 min, washed 1× in PBS, then imaged immediately.
C. neoformans cells were grown overnight under yeast culture conditions (yeast nitrogen base (YNB), 30°C with rotation), then subcultured to OD600 of ∼0.2 and grown to midlog phase. LysoTracker Green was added to a final concentration of 500 nM and incubated for five minutes with shaking at 30°C. Cells were then harvested and immediately imaged.
Strains were grown under tissue culture conditions for 12 hr. MitoTracker Green (Invitrogen) was added to a final concentration of 1 µM (from 1 mM stock in DMSO), incubated 30 min at 37°C, then imaged.
Samples were grown overnight in YNB at 30°C, then washed three times in 1× PBS and resuspended at 1 OD/ml in DMEM, then incubated 16 hr under tissue culture conditions (5% CO2, 37°C). Samples were then fixed for 15 min in 4% paraformaldehyde, washed three times in 1× PBS, and then used for staining. To stain with concanavalin A (conA) staining for mannose residues, cells were incubated 5 min in 50 µg/ml Alexa Fluor 594 (Invitrogen), washed once in 1× PBS, then imaged. Samples for staining for chitin and GXM were incubated with αGXM antibody mAb 339 (1∶1000) as previously described [5] and fluorescein-conjugated chitin binding domain (New England Biolabs) (1∶500) for 4 hr, then washed twice in 1× PBS and incubated with TRITC-conjugated donkey anti-mouse secondary antibody (Jackson ImmunoResearch) and fluorescein-conjugated chitin binding domain (1∶500) for 1 hr. Samples were then washed once and imaged using an Axiovert 200 M (Zeiss) microscope running Axiovision software. β-glucan staining was performed using the same procedure as GXM staining but with anti-β-glucan antibody (1∶1000) (Biosupplies Australia).
Phagocytosis assays were performed as previously described [5], [22]. RAW264.7 macrophages (2×104 cells/well) were seeded into 96-well tissue-culture treated plates in DMEM medium and allowed to adhere overnight. C. neoformans cells grown in YPAD medium were washed three times with PBS, then resuspended to a density of 107 cells/ml in PBS. 200 µl fresh DMEM was added to RAW264.7 cells. 5 µl C. neoformans culture (5×104 cells) were then added to each well for a multiplicity of infection of two yeast to one macrophage. Following 24 hr co-incubation, the macrophages were washed three times with PBS to remove unphagocytosed yeast and then fixed with 1% formaldehyde/PBS prior to visualization on an inverted light microscope. Percentage of yeast cell-associated macrophages was determined by counting the number of macrophages with yeast internalized or associated with their cell surface, divided by the number of macrophages counted. At least 200 macrophages were assayed per well, and each strain was tested in triplicate. If performing phagocytosis experiments under opsonizing conditions, C. neoformans cells were grown overnight in YNB, washed three times in 1× PBS, resuspended to a density of 107 cells/ml in either fetal bovine serum (opsonized samples; FBS) or 1× PBS (unopsonized samples), incubated for 30 min at 30°C on a shaking platform, washed once in 1× PBS, then resuspended at 107 cells/ml in 1× PBS and used to infect macrophages as above.
|
10.1371/journal.pbio.2004644 | Effective polyploidy causes phenotypic delay and influences bacterial evolvability | Whether mutations in bacteria exhibit a noticeable delay before expressing their corresponding mutant phenotype was discussed intensively in the 1940s to 1950s, but the discussion eventually waned for lack of supportive evidence and perceived incompatibility with observed mutant distributions in fluctuation tests. Phenotypic delay in bacteria is widely assumed to be negligible, despite the lack of direct evidence. Here, we revisited the question using recombineering to introduce antibiotic resistance mutations into E. coli at defined time points and then tracking expression of the corresponding mutant phenotype over time. Contrary to previous assumptions, we found a substantial median phenotypic delay of three to four generations. We provided evidence that the primary source of this delay is multifork replication causing cells to be effectively polyploid, whereby wild-type gene copies transiently mask the phenotype of recessive mutant gene copies in the same cell. Using modeling and simulation methods, we explored the consequences of effective polyploidy for mutation rate estimation by fluctuation tests and sequencing-based methods. For recessive mutations, despite the substantial phenotypic delay, the per-copy or per-genome mutation rate is accurately estimated. However, the per-cell rate cannot be estimated by existing methods. Finally, with a mathematical model, we showed that effective polyploidy increases the frequency of costly recessive mutations in the standing genetic variation (SGV), and thus their potential contribution to evolutionary adaptation, while drastically reducing the chance that de novo recessive mutations can rescue populations facing a harsh environmental change such as antibiotic treatment. Overall, we have identified phenotypic delay and effective polyploidy as previously overlooked but essential components in bacterial evolvability, including antibiotic resistance evolution.
| What is the time delay between the occurrence of a genetic mutation in a bacterial cell and manifestation of its phenotypic effect? We show that antibiotic resistance mutations in Escherichia coli show a remarkably long phenotypic delay of three to four bacterial generations. The primary underlying mechanism of this delay is effective polyploidy. If a mutation arises on one of the multiple chromosomes in a polyploid cell, the presence of nonmutated, wild-type gene copies on other chromosomes may mask the phenotype of the mutation. We show here that mutation rate estimation needs to consider polyploidy, which influences the potential for bacterial adaptation. The fact that a new mutation may become useful only in the “great-great-grandchildren” suggests that preexisting mutations are more important for surviving sudden environmental catastrophes.
| As genetic mutations appear on the DNA, their effects must first transcend the RNA and protein levels before resulting in an altered phenotype. This so-called “phenotypic delay” in the expression of new mutations could have major implications for evolutionary adaptation, particularly if selection pressures change on a timescale that is short relative to this delay, as may be the case for selection by antibiotics. The duration of phenotypic delay is an old but nearly forgotten question in microbiology[1–3]. Luria and Delbrück were interested in the delay because they expected it to affect the mutant distribution in the fluctuation test in their seminal work on the random nature of mutations [1]. They argued that if a mutant clone expressed its phenotype after G generations, then phenotypic mutants should be observed in populations in groups of 2G. Frequent observations of single-mutant populations, however, suggested G ≈ 0. They therefore concluded that the phenotypic delay is negligible [1,3,4]. This has remained the modus operandi [4], despite the fact that molecular cloning protocols imply a significant delay because they require a waiting time typically longer than a bacterial generation to express new genetic constructs [5].
To quantify the phenotypic delay more directly, the time point of occurrence of a mutation in a cell needs to be known, which has only become possible with modern methods of genetic engineering. Here, we use a recombineering approach to introduce mutations in E. coli within a narrow time window and find a remarkable phenotypic delay of three to four generations for three antibiotic resistance mutations. We identify the underlying mechanism as effective polyploidy, which reconciles the long phenotypic delay with Luria and Delbrück’s observations. Investigating the consequences of effective polyploidy and phenotypic delay, we find that mutation rate estimates need to be adjusted for ploidy. Moreover, resistance mutations that occur after exposure to antibiotics are much less likely to survive due to the multigenerational phenotypic delay, while preexisting mutations become a much more important contributor to survival.
To quantify phenotypic delay, we introduced each of four mutations at a specified time point in E. coli with an optimized recombineering protocol (Materials and methods), in which a single-stranded DNA (ssDNA) oligonucleotide carrying the point-specific mutation is transformed into bacteria by electroporation [6]. The ssDNA then binds reverse complementarily to its target on the bacterial genome as part of a lagging strand in an open replication fork [6], thereby introducing the mutation. Three mutations confer antibiotic resistance to rifampicin, nalidixic acid and streptomycin, respectively (RifR, NalR, and StrepR) [7]; the fourth mutation enables lactose prototrophy (lac+) [8] (S1 Table). After introduction of the mutations, the cells grew continuously without selection and were sampled over time. Sampled cells were subjected to “immediate” versus “postponed” selection to quantify, respectively, the frequencies of current phenotypic mutants and of genotypic mutants (that contain at least one mutant gene copy and eventually have some phenotypic mutant descendants), with their ratio called phenotypic penetrance (Fig 1). Phenotypic delay is quantified as the time in bacterial generations to reach 50% phenotypic penetrance.
Surprisingly, all three resistance mutations showed significant phenotypic delay at two selective concentrations of their respective antibiotic (Fig 2A). Reaching 50% phenotypic penetrance required five to six generations of postrecombineering growth. The frequency of genotypic mutants increased over the first one to two generations but eventually declined (Fig 2C). The transient increase may reflect the time window of introduction of the mutations. Discounting the first two generations, a phenotypic delay of three to four generations remains to be explained.
Phenotypic delay could result from multiple factors. Firstly, it could arise from the gradual replacement of wild-type proteins by mutant proteins following mutagenesis. Time may be required for sufficient protein turnover before the mutant phenotype can manifest. Another possibility is that cells are effectively polyploid due to multifork replication [9,10]. Recombineering incorporates the mutation into only one or some of the chromosomes starting from a single-strand mutant DNA [6], comparable to the occurrence of natural mutations. This yields effectively heterozygous cells that could produce both wild-type and mutant proteins from different chromosomes, which may prevent the onset of the mutant phenotype. Three generations could be the minimal time needed for a cell with one mutant copy out of eight chromosomes (comparable to previous estimates [9]) to produce the first homozygous mutant carrying only mutant alleles. Effective polyploidy is also compatible with the observed decline in genotypic mutant frequency (Fig 2C) because heterozygous mutants produce both mutant and wild-type descendants, such that the frequency of cells carrying at least one mutant gene copy will decline until all cells are homozygous.
To quantify the contribution of effective polyploidy, we used a lacZ reporter assay to visualize heterozygous mutants. We constructed three reporter strains with a disrupted lacZ gene inserted close to each resistance target gene and restored it through recombineering (S2 Table). Genotypic mutants were visualized by plating on indicator media where lac+ and lac− cells become blue and white, respectively. Heterozygous mutants generate sectored colonies, while homozygous mutants generate blue colonies, thus indicating the frequency of homozygous mutants amongst all genotypic mutants (Fig 2E). Comparing the estimated proportion of homozygous reporter mutants with the corresponding phenotypic penetrance of the resistance mutation reveals that phenotypic delay can be fully explained by effective polyploidy for NalR at 2x minimum inhibitory concentration (MIC) and for RifR (Fig 2B). Homozygosity precedes phenotypic penetrance by about 0.5 generations for NalR at 8x MIC and one generation for StrepR, suggesting that here, additional protein turnover may be involved.
These results also imply that these resistance mutations are genetically recessive to antibiotic sensitivity, which has also been described in previous studies based on co-expression assays [11,12]. The recessive nature of these antibiotic resistance alleles stems from their molecular mechanism: when the antibiotic molecule binds to its target protein, the resulting complex is capable of damaging the cell even in small quantities, essentially acting as toxins. As a result, the gene dosage of wild-type targets is critical. For instance, nalidixic acid-bound gyrase proteins can introduce DNA double-stranded breaks [13]. In particular, bacteria that overexpress gyrase become even more sensitive to quinolone antibiotics [14]. Although the exact killing mechanism of streptomycin remains a subject of debate, it is generally accepted that streptomycin-bound ribosomes damage the cell via mistranslation [15]. Finally, rifampicin-bound RNA polymerase appears to blockade the DNA, thereby preventing transcription even by drug-resistant RNA polymerases [12]. Although dominant RifR mutations have also been described [16], the mutation we tested here appears to be recessive.
In the case of lacZ mutations, we scored the frequency of lac+ phenotypic mutants on lactose-limited media. The ability to metabolize lactose is dominant to its inability [8]. Because any cell containing a lac+ allele can metabolize lactose and eventually form a colony, phenotypic penetrance, as expected, was always at 100% (Fig 2D), indicating that the observed phenotypic delay of resistance mutations is not an artifact of our protocol.
A further testable prediction of effective polyploidy is that inheritance of mutant alleles is asymmetrical: heterozygous mutants are expected to produce both wild-type and mutant offspring. For mutations with intermediate dominance, offspring progressing towards mutant homozygosity should show an increasingly mutant phenotype, while others show a transient phenotype because they inherit no mutant genes and their mutant proteins are diluted over subsequent divisions. Phenotypic delay would manifest as the time such mutations need to reach full phenotypic expression. To test this prediction, we repaired a disrupted YFP gene with recombineering, creating fluorescent mutants where the fluorescence intensity depends on the number of functional copies of this gene. We then tracked fluorescence as an intermediate-dominant phenotypic trait using single-cell imaging. As expected, we observed fully, transiently, and non-fluorescent offspring lineages from recombineering-treated cells (Fig 3A–3C, S1 Movie), consistent with effective polyploidy. Furthermore, fluorescence in mutant lineages increased monotonically and reached maximal intensity almost two generations after forming homozygous mutants (Fig 3C). This additional delay could be due to protein folding [17]. In total, we found 34 homozygous mutant lineages in 25 microcolonies. Six microcolonies spawned multiple, separate homozygous mutant lineages, thus corroborating previous findings that recombineering may modify multiple genome copies in one cell [18]. Overall, a median of five generations was required to form homozygous mutants, consistent with our lacZ reporter assay results (Fig 3D). These results provide direct visual support that effective polyploidy underlies phenotypic delay. A similar pattern has been observed previously in E. coli with fimbrial switching, a genetic modification that involves inversion of a promoter sequence on the bacterial genome [19].
Electroporation, as used in the recombineering protocol, may cause bacteria to form filaments due to stress [20]. Filamentation might exacerbate phenotypic delay by increasing the intracellular genome copy number. Our single-cell experiment revealed that filamentation was indeed frequent (Fig 3E). By directly observing the cell shape and the time point for onset of fluorescence, we estimated that 18 out of 34 homozygous mutant lineages incorporated the ssDNA mutation into a filamentous ancestor cell. Strikingly, lineages initiated by filamentous cells showed a distinctly different distribution of time to form the first homozygous mutant than nonfilamentous lineages (Fig 3F). Nonfilamentous lineages showed a median of 4 generations until homozygosity, as would be required for cells that incorporated the mutation in 1 out of 16 DNA single strands (i.e., 8 chromosomes) in less than 1 generation post recombineering. In contrast, filamentous lineages showed a median of 6.5 generations, and all lineages requiring more than 6 generations were filamentous. In conclusion, both filamentous and nonfilamentous cells exhibit a multigenerational phenotypic delay. Filamentation, however, can exacerbate phenotypic delay presumably by increasing the number of chromosome copies within a cell, explaining particularly long delays of extensive generations sometimes observed in our experiments.
Because phenotypic delay arises from effective polyploidy, one would expect genes further from the replication origin with lower ploidy than origin-proximal genes to show shorter phenotypic delay. However, we observed a similar phenotypic delay and distribution of time until homozygosity for all three tested genes (Fig 2) despite different distances from the origin. Further analysis revealed a strong negative correlation between distance to origin and the initial frequency of genotypic mutants induced by recombineering (Fig 4A and 4B). We hypothesized that these observations reflect the mechanism of mutagenesis by recombineering. Because open replication forks are required for recombineering, mutations could only be introduced during DNA replication [21]. The probability of successful mutagenesis on at least one open chromosomal target increases with ploidy, which itself increases during DNA replication. Therefore, we hypothesized that instead of mutating at low ploidy and thereby exhibiting shorter phenotypic delay, most of our observed origin-distal mutations were generated when their targets transiently reached higher ploidy either during normal DNA replication or due to cell filamentation. Therefore, consistent with our observations so far, origin-distal mutations would show phenotypic delay similar to origin-proximal mutations but reduced recombineering efficiency because origin-distal genes are less accessible for mutagenesis.
To further test this hypothesis, we inferred a minimal ploidy of the aforementioned target sites at the precise time point of ssDNA integration with an adapted lacZ reporter assay that measured the fraction of mutant cells in mutant colonies directly after recombineering (Materials and methods). Overall, the distributions of ploidy were similar for the three tested chromosomal locations at 5.5%, 9.7%, and 34.2% genome distance from the origin (Fig 4C). This result potentially explains why we observed no effect of chromosomal location on phenotypic delay. Furthermore, this principle should apply not only to recombineering but also to natural mutations that arise during DNA replication.
Fluctuation tests are widely used to estimate bacterial mutation rates by counting mutants exhibiting a selectable phenotype. Selection is typically applied to stationary phase cells [23], which are expected to be polyploid [10,22]. That polyploidy should affect the appearance of mutants in the fluctuation test was already pointed out quite early in the literature [24,25], but subsequently this consideration largely fell by the wayside.
When a mutation arises in an effectively polyploid cell, the first homozygous mutant descendant must appear as a single cell in the population (Fig 5). Therefore, in contrast with earlier interpretations [1], the frequent observation of singletons in the mutant distribution does not invalidate the existence of a substantial phenotypic delay. Furthermore, because heterozygous cells carrying recessive mutations do not exhibit the mutant phenotype, i.e., cannot form colonies on selective plates in the fluctuation test, these unobserved mutants should result in an underestimation of the mutation rate. Incidentally, the phage or antibiotic resistance mutations typically used in fluctuation tests are recessive [11,12,26].
The impact of polyploidy on mutation rate estimation from fluctuation tests using the modern “gold standard” maximum likelihood method [23] has not been examined. Although one recent study considered the “segregation lag” for recessive mutations resulting from polyploidy [27], corrections to mutation rate estimators were only derived for two simpler methods with limited range of accuracy and low statistical efficiency relative to the maximum likelihood method [23]. Furthermore, these derivations neglected the key point that not all descendants of heterozygous mutants will be mutants themselves.
We investigated the effect of polyploidy on observed mutant distributions, and thus estimated mutation rates, for both dominant and recessive mutations by simulating fluctuation tests. Our simulation model assumed fixed effective ploidy at the target by doubling and symmetrically dividing chromosome copies upon division according to a model of segregation that is justified for E. coli (Materials and methods). Importantly, this model leads to the shortest possible time to homozygosity and thus a conservative estimate of lag; however, other models of segregation in bacteria and archaea are possible [27,28]. From simulated cultures, we counted phenotypic mutants given either a completely dominant or a completely recessive mutation, assuming instant protein equilibration, and estimated mutation rates using standard maximum likelihood methods [23,29,30]. Other than polyploidy and dominance considerations, all modelling assumptions are the same as the standard approach (Materials and methods), allowing us to isolate the effect of ploidy. Under our model, ploidy c has two effects relative to monoploidy: (i) it increases the number of mutation targets and thus the per-cell mutation rate by a factor c, and (ii) it generates initially heterozygous mutants that, after a delay of log2c generations, produce one out of c homozygous mutant descendants (S1 Fig).
The mutation rate estimate at the mutational target can be compared to the actual per-copy rate μc and per-cell rate c·μc used in the simulations (Fig 6A and 6B). When c = 1 (monoploidy), as the standard method assumes, the estimate indeed reflects the per-copy or (equivalently) per-cell rate. For c > 1, the estimate is higher for dominant than for recessive traits. Surprisingly, for recessive traits, the estimate tends to coincide with the per-copy rate μc regardless of ploidy. For dominant traits, the estimate lies between the per-copy and per-cell rates, with confidence interval size increasing with ploidy. These patterns are robust across a range of parameter values (S2 Fig). In fact, these effects have a precise mathematical explanation (S1 Text section 2.2): the distribution of mutant counts in a polyploid population turns out to match the standard (monoploid) model with rescaled mutational influx in the case of a recessive trait (Fig 6C and S3A Fig) but fundamentally differs for a dominant trait (Fig 6D and S3B Fig). Therefore, for more commonly used recessive traits, estimates reflect mutation rates per target copy, which can be scaled up to per genome copy. This could explain why per-nucleotide mutation rates estimated from different targets do not differ significantly, despite differences in target location that potentially influence their copy number [31].
We then asked whether effective polyploidy impacts mutation rate estimates based on whole-genome sequencing (WGS) as well as fluctuation tests. WGS is typically conducted on evolved populations from mutation accumulation (MA) assays, which use single-cell bottlenecking to minimize selection [31]. Under the simplifying assumptions of fixed generation time and no cell death, we modeled an MA assay by tracking the single lineage that passes through each bottleneck and is ultimately sampled for sequencing (S1 Text section 3 and S4 Fig). Accounting for polyploidy, the per-cell mutation rate is c·μg, where μg is the per-genome-copy mutation rate. However, due to asymmetric inheritance, only a fraction 1/c of descendants from a mutant progenitor will eventually become homozygous mutants. Therefore, only a fraction 1/c of mutations arising in the focal lineage will ultimately be sampled, leaving the per–genome copy rate μg as the inferred mutation rate (S1 Text section 3). We therefore conclude that neither the fluctuation test nor WGS methods can accurately capture the per-cell mutation rate. Therefore, as the typical assumption is one genome per cell, neglecting polyploidy underestimates the total influx of de novo mutations in bacterial populations, which is relevant for adaptation.
Effective polyploidy has important consequences for evolutionary adaptation, both through the aforementioned increased influx of mutations and the masking of recessive mutations’ phenotype. Masking of deleterious recessive mutations is expected to increase their frequency in the standing genetic variation (SGV) and yield transiently lower, but eventually higher, mutational load in a fixed environment [28,32]. This higher standing frequency could promote adaptation to new environments should these mutations become beneficial. However, in an environment where mutations are beneficial, masking their effects should hinder adaptation. Previous theoretical studies addressing these conflicting effects of ploidy on adaptation [32,33] have not been linked to bacteria, nor have they specifically considered the chance of evolutionary rescue, i.e., rapid adaptation preventing extinction under sudden harsh environmental change (e.g., antibiotic treatment).
Rescue mutations may preexist in the SGV and/or arise de novo after the environmental shift during residual divisions of wild-type cells. The source of rescue mutations has implications for the optimal approach to drug treatment [34] and the preservation of genetic diversity following rescue [35]. To address the impact of effective polyploidy on rescue from SGV and de novo mutations, we developed a mathematical model of replication, mutations, and chromosome segregation in polyploid bacterial cells (Fig 7 and S1 Text section 4).
We first derived the frequency of mutants in the SGV at mutation–selection balance in the “old” environment, where mutations continually arise at rate μ∼c per chromosome copy per replication and, if expressed, carry a fitness cost s (S1 Text section 4.2). This yielded analytical expressions confirming that polyploidy increases the total frequency of a recessive mutant allele by masking its cost in heterozygotes. In contrast, the total mutant allele frequency is independent of ploidy if the mutation is dominant (Table 1 and S5 Fig).
Next, we considered the fate of the population upon shifting to a new, harsh environment (e.g., antibiotic treatment), where phenotypically wild-type (“sensitive”) cells have a low probability of successfully dividing while phenotypically mutant (“resistant”) cells have a higher probability. The population may already contain heterozygous and homozygous mutants in the SGV and additionally give rise to de novo mutants stochastically according to a Poisson process. We developed a branching process model to evaluate the probability that such mutations escape stochastic extinction, accounting in particular for the multiple cell divisions required until mutations segregate to homozygosity, with probabilities of successful division depending on whether these mutations are recessive or dominant. Finally, combining these model components yielded expressions for the probability of population rescue from SGV, PSGV, and from de novo mutations, PDN (S1 Text section 4.3).
These rescue probabilities depend strongly on ploidy, dominance, and other model parameters (Fig 8, S6, S7, S8 and S9 Figs). In the recessive case, if phenotypically wild-type cells cannot divide in the new environment (e.g., a perfectly effective antibiotic), then PSGV is independent of ploidy, reflecting the constant frequency of preexisting phenotypically mutant homozygotes (Table 1). This result is consistent with our above findings for fluctuation tests. If division of wild-type cells is possible (e.g., imperfect antibiotic efficacy), then PSGV increases with ploidy because heterozygotes may produce additional homozygous mutant descendants. On the other hand, PDN decreases with ploidy because de novo mutations require more cell divisions until segregation is complete and the mutant phenotype is expressed, which turns out to outweigh the increase in mutational influx (S1 Text). Therefore, although the overall probability of rescue remains similar as ploidy increases, rescue is increasingly from SGV rather than de novo mutations (Fig 8A and 8B). These qualitative patterns are robust to variations in the model parameters (S6 and S8 Figs). For dominant mutations, on the other hand, both PSGV and PDN increase with ploidy, and their relative contributions can show more complex patterns (Fig 8C and 8D, S7 and S9 Figs). In general, SGV makes a relatively larger contribution when the mutation has a low cost in antibiotic-free conditions (small s) and when the antibiotic is highly effective (low pS), in agreement with previous findings in the evolutionary rescue literature [36].
The phenotypic effect of a bacterial mutation cannot manifest instantaneously. Here, we therefore asked two questions: how large is this phenotypic delay, and what is its primary cause? We found a delay of three to four generations in the expression of three recessive antibiotic resistance mutations in E. coli and provided evidence that effective polyploidy is its primary cause.
Polyploidy is often regarded as a transient property limited to fast-growing bacteria, but this view has been challenged in recent years. Though ploidy tends to be higher during exponential growth (up to eight or 16 partial chromosome copies) [9], even during stationary phase, E. coli cells contain typically four and up to eight complete chromosome copies [10]. Environmental stresses can also induce multinucleated, polyploid cell filaments [37], in which adaptive mutations must overcome phenotypic delay before allowing population survival in deteriorating environments. A recent study exposing bacteria to low doses of the antibiotic ciprofloxacin showed that resistant bacteria can only emerge from mononucleated offspring cells that bud off from a long multinucleated cellular filament [37]. This observation can be explained by masking of the mutant phenotype in polyploid, heterozygous cells. Furthermore, obligate polyploid bacterial species ranging from free-living bacteria to clinically relevant pathogens have been discovered across six phyla [22,38,39]. This has, for instance, been recognized as a confounding factor in metagenomic studies of bacterial community structure by marker gene–based analysis [38]. Even within the same bacterial species, ploidy may vary in response to selection, as shown in a previous study that E. coli with resistance to camphor vapor also showed increased ploidy [40]. Therefore, we argue that polyploidy is broadly relevant for bacteria and will generally result in phenotypic delay of recessive mutations.
Dominance and polyploidy (whether effective or obligate) together affect the number of mutants observed in fluctuation tests and thus require reinterpretation of mutation rate estimates. Fluctuation tests typically use recessive antibiotic resistance mutations. Encouragingly, we found that the resulting estimates accurately reflect the per–target copy mutation rate, regardless of ploidy. Therefore, studies using fluctuation tests to compare per–target copy mutation rates across different conditions, e.g., [41], remain valid. Similarly, we showed that sequencing-based methods of mutation rate estimation from MA assays reflect per–genome copy rates. Therefore, effective polyploidy does not appear to explain the up to 10-fold difference in mutation rate estimates [31,42] obtained using these two different methods.
Importantly, however, neither method reflects the per-cell mutation rate and thus the total mutational influx in the population, which is proportional to ploidy. Indeed, our models suggest that fluctuation tests with recessive mutations or sequencing-based methods leave no detectable signal of ploidy in the data: that is, a polyploid population is indistinguishable from a monoploid population in these assays, even though their total mutational influx differs by a factor equal to the ploidy. Meanwhile, dominant mutations lead to fundamentally different mutant distributions in the fluctuation test, and neither the per-copy nor the per-cell mutation rate is accurately estimated. In conclusion, mutation rate estimates must be interpreted with caution, regardless of the method used.
The effects of polyploidy on number of mutational targets and phenotypic delay influence the evolutionary potential of populations to escape extinction under sudden environmental change such as antibiotic treatment. In particular, we showed that recessive rescue mutations are increasingly likely to come from the SGV as ploidy increases. This is due to the dual effects of masking the fitness cost of these mutations in the old (antibiotic-free) environment while decreasing the chance that de novo mutations survive in the new (antibiotic) environment until their beneficial phenotype is expressed. Our novel results for rescue are broadly in line with previous theoretical findings on the role of ploidy for adaptation [32] and highlight the point that these considerations are relevant to bacteria as well as eukaryotes.
Our theoretical results rest on several simplifying assumptions. Firstly, we examined only the cases of complete dominance or complete recessivity. More generally, gene dosage effects could result in intermediate dominance; in this case, we expect the effects on mutation rate estimation and rescue probability to be intermediate between the two extremes considered here. Secondly, while we examined rescue via single mutations, if multiple mutations with different dominance were available, populations at different ploidy levels may tend to evolve via different pathways [43]. Furthermore, while we exclusively considered chromosomal mutations, mutations on plasmids, particularly those with high copy number [44], should show similar effects, although segregation patterns and thus time to achieve homozygosity are likely to differ. Finally, models developed thus far have assumed constant ploidy, whereas future modeling efforts could incorporate the dynamically changing and environment-dependent nature of bacterial ploidy.
Given the manifold implications of a multigenerational phenotypic delay, we argue that effective polyploidy and the resulting phenotypic delay are essential factors to consider in future studies of bacterial mutation and adaptation.
All experiments were performed with strains derived from the wild-type E. coli MG1655 strain. A complete list of strains can be found in S2 Table. Cells were grown at 30°C in LB or in M9 media with 0.4% lactose. Antibiotics were purchased from Sigma-Aldrich. To prepare stocks, rifampicin was dissolved in DMSO to 100 mg/ml; nalidixic acid was dissolved in 0.3 M NaOH solution to 30 mg/ml; streptomycin and ampicillin were dissolved in MilliQ water to 100 mg/ml and filter sterilized. Rifampicin, streptomycin, and ampicillin stocks were kept at −20°C, while nalidixic acid was kept at 4°C. Ampicillin 100 mg/L was used for maintaining the pSIM6 recombineering plasmid. All antibiotic agar plates were prepared fresh before every experiment.
The MICs of rifampicin, streptomycin, and nalidixic acid were determined by broth dilution method in LB and found to be 12 mg/L, 12 mg/L, and 6 mg/L, respectively.
Our recombineering protocol was adapted from previous studies [6,21]. To ensure reproducibility, a detailed step-by-step protocol is provided in S1 Text section 1. In brief, E. coli harboring pSIM6 plasmids were grown into early exponential phase before heat activation at 43°C for 10 minutes to express the recombineering proteins. Activated cells were then repeatedly washed in ice-cold MilliQ water to remove residual salts. Concentrated salt-free cell suspension 50 μl was then mixed with approximately 200 ng of ssDNA before electroporation at 1.8 kV/mm. Immediately after electroporation, cells were resuspended in LB and recovered for 30 min at 30°C. After this initial recovery, cells were pelleted, then resuspended in fresh LB to continuously grow at 30°C for subsequent phenotyping.
From the resuspended population, approximately 2% of cells were sampled hourly for the first 10 hours and then at 24 hours. A time point at 48 hours was also included to control for factors that potentially prevent phenotypic penetrance from ever reaching 100%, such as low establishment probability of mutant cells. The sampled populations were appropriately diluted for optimal plating onto selective and nonselective plates. Total population size and thus generations elapsed in the sampled cultures was estimated from colony-forming units (CFU) on nonselective plates. To score the frequency of genotypic mutants, we replica-plated all colonies from the nonselective plates to selective plates for each tested time point. The frequency of genotypic mutants, Fg, was determined by the fraction of colonies from nonselective plates that could grow after postponed replica plating onto selective plates. The frequency of phenotypic mutants, Fp, was determined by the ratio of CFU from immediate plating on selective plates versus CFU on nonselective plates. Phenotypic penetrance was defined as P = Fp ÷ Fg. Phenotypic delay was then quantified as the time point at which phenotypic penetrance reaches 50%.
To quantify mutant homozygosity, i.e., the fraction of homozygous mutants among all genotypic mutants, we developed a lacZ-based visual assay. We constructed bacterial strains with a lacZ gene disrupted by a nonsense point mutation (E461X) [8] and inserted the broken lacZ within 5 kb of each antibiotic resistance target gene. These strains were subjected to recombineering with an ssDNA carrying the reverse point mutation (X461E) that restored the lac+ phenotype. The resulting phenotypic mutants were selected on M9-lactose media. Phenotypic mutants become blue on permissive media containing 1 mM IPTG and 40 μg/ml X-gal [5]. Heterozygous mutants with mixed lac+/lac− alleles form blue-/white-sectored colonies, whereas homozygous mutants form entirely blue colonies (Fig 2E). Plates with colonies were left at 4°C for 1 week to allow sufficient development of the blue color but before the blue pigment spreads too far to obscure sectored colonies. Counting sectored (s) and nonsectored (n) blue colonies, we determined mutant homozygosity as fhom = n/(s+n). Comparing fhom to the phenotypic penetrance P thus indicates to what extent phenotypic delay is attributable to effective polyploidy. Colony counting was performed using CellProfiler [45].
We constructed a strain with a constitutively expressed YFP gene disrupted by 3 consecutive stop codons. Recombineering corrected the stop codons. After electroporation and 30 min recovery at 30°C, 1 μl of appropriately diluted cell suspension was pipetted onto a small 1.5% UltraPure Low Melting Point agarose pad. After drying the pad for 1 minute, it was deposited upside down in a sealed glass-bottom dish (WillCo Wells, GWST-5040). Time-lapse microscopy was performed with a fully automated Olympus IX81 inverted microscope, with 100X NA1.3 oil objective and Hamamatsu ORCA-flash 4.0 sCMOS camera. For fluorescent imaging, we used a Lumen Dynamics X-Cite120 lamp and Chroma YFP fluorescent filter (NA1028). The sample was maintained at 30°C by a microscope incubator. Phase-contrast and yellow fluorescence images were captured at 5-minute intervals for 16 hours. The subsequent image analysis was performed with a custom-made MATLAB program (Vanellus, accessible at: http://kiviet.com/research/vanellus.php).
We performed the lacZ reporter assay, as described above, for three strains with the lacZ gene juxtaposed to each of the antibiotic resistance target genes. After the 30-min recovery following recombineering (before extensive growth), cells were plated directly onto LB agar with IPTG and X-gal. After 24 hours of incubation at 30°C, entire mutant colonies that contained blue color were picked. We started from colonies closest to the center of each agar plate and expanded outwards to eliminate picking bias. The picked colonies were diluted 104- to 105-fold in PBS before plating on average about 500 CFUs on fresh LB agar with IPTG and X-gal to infer the fraction of mutant cells in the given colony. This fraction was then used to deduce the minimal ploidy at the time of ssDNA integration based on a previous study [46]: a colony with one-quarter mutant cells, for instance, has minimal ploidy of 2 because it could have resulted from mutagenesis on 1 out of 4 DNA single strands. Actual ploidy may be higher if, for instance, 2 out of 8 single strands mutated in a cell of ploidy 4.
For simplicity, we assumed every cell has the same effective ploidy, i.e., copies of the gene of interest, over the relevant timescale. At each generation, chromosomes must therefore undergo one round of replication and be evenly divided between the two daughter cells. In E. coli, chromosomes appear to progressively separate as they are replicated and detach last at the terminus [9]. We therefore assumed segregation into daughter cells occurs at the most ancestral split in the chromosome genealogy. This assumption is conservative because it implies that mutant chromosomes always remain together, resulting in the fastest possible approach to homozygosity and thus the shortest phenotypic delay. Under this model, ploidy must take the form c = 2n (for n = 0, 1, 2, …), among which the number of mutant copies is j = 0 or 2i (0 ≤ i ≤ n), while the remaining c–j copies are wild-type. Note that other models of segregation are possible, e.g., random segregation in highly polyploid Archaea [28], which would lead to slower approach to homozygosity and corresponding effects on the evolutionary model results.
All simulations and inference were implemented in R. We wrote our own code to account for polyploidy, but in the future, our methods could potentially be integrated into recently published R packages for fluctuation analysis [47,48]. We simulated culture growth in nonselective media with stochastic appearance of spontaneous de novo mutations (for details see S1 Text section 3.1). We assumed a fixed per-copy mutation rate of μc per wild-type cell division, such that the per-cell mutation rate is μ = c μc for effective ploidy c. We neglected the chance of more than one copy mutating simultaneously, i.e., mutants always arose with the mutation in a single chromosome copy. Although natural mutations may initially arise in either single- or double-stranded form (for instance, mismatches versus indels following double-strand breaks), to be consistent with the standard model, we assumed mutations arose in double-stranded form (see discussion in S1 Text sections 2.1 and 2.3). The descendants of each de novo mutant were tracked individually, with mutant chromosomes segregating as described above and interdivision times either drawn independently from an exponential distribution or constant. We assumed no fitness differences between wild-type and mutant in nonselective media. In the case of c = 1 and exponential interdivision times, our model corresponds to the standard “Lea-Coulson” model [23,30], which is also the basis of the widely used software FALCOR [49].
Each simulated culture was initiated with 1,000 wild-type cells, and after 20 wild-type population doublings, the culture growth phase ended and phenotypic mutants were counted under the assumption of either complete recessivity (requiring all c chromosomes to be mutant) or complete dominance (requiring at least one mutant chromosome). Assuming (as standard) 100% plating efficiency and no growth of phenotypically wild-type cells under selective conditions, the number of colonies formed on selective plates equals the number of phenotypic mutants in the final culture. The mutant colony counts from 50 simulated parallel cultures were then used to obtain a maximum likelihood estimate (MLE) μ^ and 95% profile likelihood confidence intervals of mutation rate under the standard model, which in particular assumes that a de novo mutant and all its descendants are immediately phenotypically mutant. The best-fitting distribution of mutant counts was calculated from the standard model with mutation rate equal to μ^. While we implemented these calculations in R (code available on Dryad: http://dx.doi.org/10.5061/dryad.8723t), calculation of the likelihood under this model has been previously described [29,50] and has also been implemented in FALCOR [49].
We considered a population with effective ploidy c, in which mutations arise (again, in double-stranded form) in a proportion μ∼=cμ∼c of offspring in each generation. The definition of mutation rate used in the population genetics literature is subtly different from that used in fluctuation analysis and thus given different notation here (see S1 Text section 4.1). The mutation has relative fitness cost s in homozygotes, with the cost either completely masked (if recessive) or equal (if dominant) in heterozygotes. We extended deterministic genotype frequency recursions to incorporate chromosome segregation as described above and solved for the equilibrium frequencies of all heterozygous and homozygous mutant types (S1 Text section 4.2).
We modeled the fate of a population shifted to a harsh new environment, i.e., either extinction or rescue by mutants, stochastically using a multi-type branching process. Unlike in the fluctuation test simulations, where we neglected the chance that wild-type cells produce surviving lineages in the new environment, here we allowed a probability pS ≤ ½ that a phenotypically wild-type cell successfully divides before death to produce 2 offspring, while phenotypically mutant cells have corresponding probability pR > ½. Therefore, phenotypically wild-type cells cannot sustain themselves but have a nonzero chance of producing phenotypically mutant descendants either by segregation of mutant alleles in the SGV (modeled by mutation–selection balance as above) or de novo mutations during residual divisions in the new environment. We derived analytical approximations (S1 Text section 4.3) for the probability of rescue from SGV (PSGV) or from de novo mutations (PDN), which are not mutually exclusive.
All data were deposited in the Dryad repository (http://dx.doi.org/10.5061/dryad.8723t) [51].
|
10.1371/journal.pbio.0060191 | TSCOT + Thymic Epithelial Cell-Mediated Sensitive CD4 Tolerance by Direct Presentation | Although much effort has been directed at dissecting the mechanisms of central tolerance, the role of thymic stromal cells remains elusive. In order to further characterize this event, we developed a mouse model restricting LacZ to thymic stromal cotransporter (TSCOT)-expressing thymic stromal cells (TDLacZ). The thymus of this mouse contains approximately 4,300 TSCOT+ cells, each expressing several thousand molecules of the LacZ antigen. TSCOT+ cells express the cortical marker CDR1, CD40, CD80, CD54, and major histocompatibility complex class II (MHCII). When examining endogenous responses directed against LacZ, we observed significant tolerance. This was evidenced in a diverse T cell repertoire as measured by both a CD4 T cell proliferation assay and an antigen-specific antibody isotype analysis. This tolerance process was at least partially independent of Autoimmune Regulatory Element gene expression. When TDLacZ mice were crossed to a novel CD4 T cell receptor (TCR) transgenic reactive against LacZ (BgII), there was a complete deletion of double-positive thymocytes. Fetal thymic reaggregate culture of CD45- and UEA-depleted thymic stromal cells from TDLacZ and sorted TCR-bearing thymocytes excluded the possibility of cross presentation by thymic dendritic cells and medullary epithelial cells for the deletion. Overall, these results demonstrate that the introduction of a neoantigen into TSCOT-expressing cells can efficiently establish complete tolerance and suggest a possible application for the deletion of antigen-specific T cells by antigen introduction into TSCOT+ cells.
| T cells play critical roles in the immune response. While developing in the thymus (from whence T cells and their precursors, thymocytes, derive their name), thymocytes are selected for the ability to recognize harmful antigen (positive selection), while those that respond to antigens present in their own body are eliminated (negative selection). Dogma holds that the thymus is divided into different functional compartments to ensure that these contrasting selection processes occur efficiently: the cortex is thought to be responsible for positive selection and the medulla for negative selection. In this study, we made use of a novel transgenic mouse (carrying a LacZ marker in a small fraction of cells in the cortex) to test whether the cortex is really excluded from negative selection. We were able to show that the introduced LacZ “antigen” present only in the cortical cells leads them to eliminate any LacZ-reactive T cells from the immune repertoire and leads to tolerance of the LacZ “antigen” by the body's immune system. This process is highly efficient, such that a relatively tiny number of antigen molecules present in a small fraction of the cells in the thymic cortex can singularly perform proofreading of all developing thymocytes.
| T cell tolerance is established mainly in the thymus where the T cell population develops and learns by a process called negative selection to avoid harmful reactivity against self-antigens expressed in that thymus (reviewed in [1,2]). In the periphery, organ-specific tolerance can be established by various other mechanisms, including anergy [3], ignorance [4], and regulatory T cells [5]. Furthermore, antigen-presenting cells (APC) lacking costimulatory molecules in peripheral tissues initiate abortive immune responses [6].
The thymic microenvironment is organized and equipped to achieve efficient self-tolerance by providing stimulatory signals to developing self-reactive thymocytes. For a diverse T cell repertoire, this negative selection process occurs primarily in the thymic medullary compartment (reviewed in [7,8]). The major player among the hematopoietic cells is the dendritic cell (DC), which possesses a highly efficient antigen presentation capability. In addition, it is widely accepted that thymic medullary epithelial cells (mTEC) that express low levels of tissue-specific peripheral antigens in a promiscuous/ectopic fashion [9,10] can also initiate clonal deletion. Discovery of the AIRE gene and its expression in mTEC has led to an understanding of its critical regulatory role in the removal of autoreactive T cells, particularly against tissue-specific antigens expressed in the endocrine system (reviewed in [11,12]). However, AIRE is also expressed in non-mTEC, including thymic DC [13,14] and in cortical thymic epithelial cells (cTEC) from Rag-2–deficient thymus [15]. Furthermore, the cross-presentation pathway can participate in the CD4 and CD8 tolerance for the membrane-bound antigens [16]. Therefore, the natures of cell types responsible for the tolerance induction still remain unsettled.
The role of cortical epithelium in tolerance induction has been controversial (reviewed in [17–20]). Several experiments using thymus transplantation have clearly indicated that thymic epithelium exhibits toleragenic function [21–24]. In contrast, experiments using transgenic mice with targeting of major histocompatibility complex class II (MHCII) [25] or MHCI [26] molecules to the thymic cortical compartment (and the skin) using a fortuitous keratin 14 promoter led to the conclusion that cTEC are not capable of inducing tolerance. Such results have given rise to the idea that the thymic microenvironment is compartmentalized, with positive selection taking place in the cortex and negative selection in the medulla. If this is a true dogma, there will be autoimmune responses to the antigens specifically expressed in the cortical epithelial cells. However, when other antigens were targeted into cTEC using the same promoter, incomplete but significant tolerance to the specific antigens was observed [27,28]. In the case of a circulating antigen (C5), all types of thymic APC, including the cTEC, could effect efficient negative selection in vitro [29]. Finally, the question of the role of circulating peripheral DC in the induction of thymic tolerance has also been raised [30] and tested true [31].
Experiments regarding the ability of cTEC to efficiently present antigens have also been controversial. In early studies, the death of cortical thymocytes upon activation by antibody or peptides was interpreted as resulting from antigen presentation by the cortical stromal cells [32,33]. In addition, a study with purified thymic APC suggested that cTEC were able to present antigens to a self-reactive hybridoma, with an efficiency comparable to that of thymic DC [34]. However, later studies indicated that a cell line with cTEC properties was inefficient in processing antigens both in vitro and in vivo [35,36]. In contrast, Volkmann and his colleagues, using enriched stromal cell preparations from adult thymus, demonstrated that cTEC are able to present soluble antigens as efficiently as DC or mTEC in reaggregate cultures. In many, if not all, of the above studies, however, difficulties in interpretation still persist, in particular, because of a lack of sufficient understanding about the nature of the defined cTEC subpopulation under study, as well as the purity of the cells expressing the specific antigens that were used in the assays. More recently, Gray and colleagues reported that well-defined, purified cTEC, as well as mTEC, express costimulatory molecules and can stimulate naive T cells as much as thymic dendritic cells do in vitro [37]. Therefore, we felt it was necessary to re-evaluate the role of the cTEC subpopulation in central tolerance induction using a different model system, one perhaps better suited to more directly answering the question of whether subpopulation of cTEC can present endogenous antigens and whether this can lead to deletion of thymocytes.
Previously, in an effort to separate thymic epithelial cell (TEC) components, we introduced a new marker (Ly110), designated thymic stromal cotransporter (TSCOT), which is expressed in a specific TEC subpopulation. TSCOT is a putative 12-transmembrane protein, located mainly in the thymic cortex [38]. TSCOT is not expressed in any other tissues, as detected by quantitative reverse-transcription PCR (RT-PCR) [39]. It is also not expressed in thymocytes [38]. TSCOT+ thymic stromal cells are all MHCII+ and CDR1+/6C3+, well-defined cortical epithelial markers [40], with observable variations in levels during different developmental stages [41]. In this study, we introduce a new mouse model system called TSCOT delta LacZ (TDLacZ) that expresses a β-galactosidase (β-gal) in the TEC subpopulation. This model system constitutes a new tool for the study of TEC development and function. First, we were able to follow TSCOT-expressing TEC by β-gal activity assays or antibody staining and flow cytometry using an anti-TSCOT monoclonal antibody (mAb) [41]. LacZ enzymatic activity could also be assayed for the location of cells with a high degree of sensitivity, in both sections and the whole organism, and expression could be assessed in a quantitative manner. Second, because the protein is generated by an endogenous promoter, this system is designed to express normal doses of neoself-antigen relative to other competing cellular proteins. This is in contrast to some previous systems for the targeting of cortical expression, in which MHC molecules were displayed at unusually low levels [19,26]. Third, the absence of the TSCOT promoter activity in peripheral tissues precludes the involvement of recirculating DCs, which might deliver peripheral antigens to the thymus, and present them ectopically.
By targeting LacZ protein as a neoantigen within the TSCOT-expressing thymic epithelium, we were able to demonstrate that TSCOT+CDR1+ TEC alone, without any help from the mTEC or DC, is able to establish deletional tolerance in an AIRE-independent manner with a surprisingly high degree of efficiency.
We established a new system by knocking-in the LacZ gene into the TSCOT locus between two BamHI sites (Figure 1A). LacZ was transcribed in the same message with the 5′ portion of the TSCOT message, and translation of LacZ was facilitated by incorporating an internal ribosome entry site (IRES) sequence [42]. The targeting was confirmed by Southern blotting (Figure 1B). Northern blotting confirmed that the LacZ message was in a fusion transcript with the 5′ portion of the TSCOT message (Figure 1C).
The TDLacZ mice evidenced no distinguishable abnormalities with regard to thymic structure as the result of the deletion in TM5-TM12 portion of the TSCOT protein. In Figure 1E, we show that the similar thymic stromal patterns of the 2-wk-old homozygote and the wild type. The small difference in the fraction of stromal cell populations was within the experimental variations. There was also no difference detected in the profiles between hetero- and homozygote littermates of the various ages (unpublished data). The N-terminal portion including transmembrane spans 1–4 of the protein still remained expressed on the cell surface, as detected by flow cytometry (unpublished data). The only apparent difference was for the total thymocyte yield at 6 wk of age, which was slightly lower in about one-third of the TDLacZ homozygotes (Figure 1D). However, we failed to detect any reproducible differences in the profiles of thymocyte population except the individual variation. In addition, an analysis of 6-mo-old mice also showed no significant differences detected in the recovery of thymocytes and major profiles of CD25, CD44, CD4, and CD8 (unpublished data). When 5CC7 T cell receptor (TCR) Tg mouse was bred with TDL, no significant differences for the thymocyte populations were found in selecting or nonselecting background (F. Flomerfelt, unpublished data).
When β-gal activity was assessed in TDLacZ mice at embryonic day 11 (E11), the time at which thymus organogenesis is initiated, LacZ was already expressed in the two separated thymic rudiments, but it was not expressed in the wild-type littermates (Figure 1F). This expression was not detected in any other organs. At E16, when the thymus harbors mostly developing double-negative (DN) and double-positive (DP) cells, thymic expression of LacZ also was very clear (Figure 1G). In addition, endogenous β-gal activity appeared in the TDLacZ intestine at E16, as in the wild-type control (unpublished data). β-gal activity in thymus samples from newborn TDLacZ pups showed a gene dose dependency (Figure 1H).
We next located the LacZ-expressing cells in thymic sections. At the newborn stage, anti-LacZ antibody staining revealed the expression mostly in the thymic cortex as expected (Figure 2A). When the thymus had fully matured (8 wk of age), LacZ activity was also detected in the cortex (Figure 2B). This is consistent with our previous result that TSCOT protein and mRNA expression was located in the cortex [38]. After careful examination, we occasionally found LacZ staining extends to corticomedullary junction (unpublished data and see later). In an attempt to characterize the TSCOT-expressing cells in the mature thymus in greater detail, flow cytometric analysis was conducted using a TSCOT-specific mAb. Previously, we group the thymic stromal cell populations into at least five different subpopulations [43]. Three main population are cTEC as CDR1+UEA-1−MHCIIhiG8.8+, mTEC as CDR1−UEA-1+MHCIIhi or MHCIImedG8.8+, as well as nonepithelial population, nonTEC, CDR1−UEA-1−MHCII−G8.8−. As shown in Figure 2C, 35.6% of cTEC (CDR1+MHCIIhi) population expresses TSCOT, whereas none of the mTEC or nonTEC population expresses detectible levels of TSCOT. Although all of the TSCOT-expressing cells were positive for cortical marker CDR1 [41], a fraction of TSCOT+CDR1+ cells were found to express UEA-1 (unpublished data and see Discussion). Finally, we examined whether TSCOT mRNA was expressed along with FoxN1 and AIRE mRNAs (Figure 2D). TSCOT and FoxN1 were detectable only in the MHCII+CD45− epithelial compartment. In contrast, the AIRE message was detectable in both the epithelial and CD45+MHCII+ compartments as expected, supporting the previous result on the expression in hematopoietic stromal cells of the thymus [13–15].
Next, in order to measure sensitivity of tolerance induction, we estimated the average quantity of antigen expressed in one adult thymus by measuring the β-gal activity of the LacZ protein in purified thymic stromal cells. We isolated the cells from TDLacZΔ/Δ mice, and stained them with a mAb against TSCOT (Figure 3A). In this preparation using 28 animals, TSCOT+ cells (12.3%) corresponded to 1.2 × 105 cells. This calculates out to a total of about 4,300 TSCOT+ cells per thymus. When this cell preparation was lysed and the β-gal activity was evaluated (Figure 3B, and unpublished data), we were able to determine the LacZ concentration from a standard curve (2 × 10−11 M of 50-μl reactions). These numbers corresponded to 5,017 molecules of LacZ protein per TSCOT+ cell by the simple mathematical calculation of concentration × volume × Avogadro number/cell number; 2× 10−11 M × 50/106 × 6.02 × 1023 molecules in 1.2 × 105 cells in the experiment shown. In the second experiment, the final number was 6,825 molecules per cell.
TDLacZ and wild-type animals were immunized with recombinant LacZ protein and examined for a LacZ-specific polyclonal CD4+ T cell response. As shown in Figure 4A, a concentration-dependent LacZ-induced proliferative response was detected in purified CD4+ lymph node T cells from wild-type B6 animals, whereas T cells from TDLacZ mice clearly showed no response to LacZ. Both the heterozygous and homozygous animals showed a large reduction in proliferation (Figure 4B). LacZ-specific antibody responses were then evaluated by ELISA (Figure 4C). When whole LacZ protein was administered in CFA, wild-type mice produced both IgG1 and IgG2b isotypes specific for LacZ. In contrast, heterozygous and homozygous TDLacZ mice did not produce such antibodies (Figure 4C, top). In order to assess the possibility that this represented tolerance at the B cell level, we administered a GST-tagged loop portion of TSCOT (GST-Loop) in CFA and screened for specific antibody responses with a His-tagged loop protein (His-Loop) in an ELISA. In this case, with help provided by T cells specific for GST, both the heterozygous and homozygous TDLacZ mice made as much anti-loop IgG1 and IgG2b antibodies as the wild-type mice (Figure 4C, bottom). These results clearly show that the presence of LacZ expression in the subpopulation of TSCOT+ TEC was sufficient for the tolerization of LacZ-specific CD4+ T cells, and this tolerance is not due to the absence of whole TSCOT molecules in the animal.
Because AIRE is known to play a key role in the establishment of tolerance to antigens promiscuously/ectopically expressed in small amounts by mTEC [44–46], we investigated the possible role of AIRE in TSCOT+ TEC with regard to the induction of tolerance. We crossed the TDLacZ mouse with an AIRE-deficient mouse, and conducted the same proliferation assay for an anti-LacZ CD4+ T cell response to the LacZ protein. As shown in Figure 5, the AIRE-deficient mice displayed slightly enhanced anti-LacZ responses compared to the wild type, possibly due to the introduced cross-reactivity or autoreactivity. When one copy of LacZ was expressed by breeding the AIRE knock-out (KO) to the TDLacZ mouse, the anti-LacZ-specific proliferative response was clearly reduced. In the specific responses to 1 μg of antigen, the degrees of responses contributed by AIRE were similar (difference between wild type and AIRE KO vs. that between TDLacZ to TDLacZ AIRE KO). These results strongly suggest the presence of another pathway that AIRE does not play a major role in the induction of tolerance to an antigen expressed in TSCOT+ TEC.
We further assessed the presence or absence of selected costimulatory and adhesion molecules in the TSCOT-expressing cells. Although there has been reports that cortical epithelium does not express costimulatory molecule by histological analysis, we had reasons to believe that this conclusion may be false based on our observation of disparity between histology and flow cytometry [47]. As shown in Figure 6A–6C, flow cytometry revealed that TSCOT+ cells are all positive for MHCII, CD40, and CD54 expression. In more detailed analysis, the relative CD40 level of TSCOT+ cells was similar to that of some CD45+ cells (presumably dendritic cells), and higher than TSCOT− cells that contain TSCOT−cTEC and mTEC populations (Figure 6B, histogram). An important costimulatory molecule, CD80 was expressed in some TSCOT+ cells as shown in Figure 6D (CD45− gate and CD45−MHCIIhi gate where most of the TSCOT+ cells reside). In order to compare the relative levels of CD80 between mTEC and TCSOT+ cells, the multiparameter analyses in flow cytometry and in confocal microscopy were applied including LacZ staining with the stromal cells prepared from TDLacZ thymus. As seen in (Figures 6E and 6F, and S1). In both analyses, mean fluorescence intensity (MFI) of the relative levels of CD80 was higher in TSCOT+cTEC (CDR1+LacZ+) than in other cells mTEC (UEA-1+CDR1−) and TSCOT−cTEC (CDR1+LacZ−). CD86 was not detected under our conditions, possibly due to the trypsin-sensitive nature of this marker (unpublished data). These results clearly suggest that TSCOT+ cells can function as efficient APC.
To determine the specific mechanism for the observed tolerance, we utilized a monospecific TCR Tg mouse, BgII (D. Palmer, Marc R. Theoret, and N. Restifo, unpublished data; see Materials and Methods) that carries an anti-LacZ TCR transgene from a CD4+ LacZ-specific T cell clone [48]. This line was crossed with Rag1−/− to establish a monospecific TCR-bearing T cell population. When the BgIITg/Tg Rag1−/− mouse was crossed with the TDLacZΔ/Δ Rag1−/− mouse heterozygote for TCR Tg and TDLacZ, only CD4− CD8− DN cells were detected in the smaller thymus (Figure 7). The total number of thymocytes recovered was approximately 17.5% of what was recovered from a TCR Tg mouse. Most of the cells were arrested at the CD25hi CD44− (DN3) stage, similar to what was observed in a Rag1−/− mouse (Figure 7B). However, massive cell death in the DN as well as CD4 and/or CD8 cells was found only in the BgIITg+ TDLacZΔ/+ mouse (Figure 7C). By contrast, in the BgII Tg alone, the fraction of CD44− CD25− (DN4) cells had the highest DN subpopulation (Figure 7B). Thus, in the presence of LacZ, the TCR Tg thymocytes appeared to be substantially deleted at the post-DN3 stage as soon as they expressed their TCR at the cell surface (see Figures S2 and S3 for the expression of TCR gene and protein on the surface).
The pattern of thymic stromal cells (gated on CD45− cells) observed in the BgIITg/+ TDLacZΔ/+ Rag1−/− mice was also similar to that of a Rag1−/− mouse (Figure 7D). UEA1+ mTECs, which are prominent in the adult wild-type thymus, were barely detected, and thus the proportion of CDR1+ cTECs was greatly elevated. Therefore, BgIITg/+ TDLacZΔ/+ Rag1−/− mice do not harbor fully developed mTECs, yet they remain able to efficiently delete developing thymocytes.
Cross-Presentation by DC or mTEC Is Not Involved in CDR1+cTEC-Mediated Deletion
In order to exclude the possibility of cross-presentation by mTEC and DC in the induction of tolerance, we employed a clean reaggregate thymic organ culture system (RTOC) [49,50] using UEA-1– and CD45-depleted thymic stroma reconstituted with purified anti-LacZ TCR transgenic thymocytes. The stromal cells were prepared from a E14.5 fetal thymic organ culture, in the presence of 2-dGuo, which depletes DC, and the population was further depleted of CD45+ and UEA-1+ cells by magnetic bead separation. Using such cells isolated from wild-type or TDLacZ+/Δ thymus samples, anti-LacZ TCR-bearing DN and DP cells (1:10 ratio, similar to that of a normal thymus) from adult BgIITg/Tg animals were reaggregated with them and cultured for 5–6 d. As shown in Figure 7, the recovery of the thymocytes from the RTOC with TDLacZ cTECs was between 5%–20% of that achieved with the wild-type stroma. In addition, these cultures did not contain a significant number of CD4 single-positive (SP) cells (Figure 8A). When the DN and DP transgenic thymocytes were separately reaggregated with TDLacZ cTEC (Figure 8B), both subsets showed reduced cell numbers after culture, indicating that LacZ expression could also deplete the LacZ-responding DP cells in the RTOC. Although these results argue against the idea of impaired differentiation from the DN to the DP stage, they suggest that developing thymocytes are deleted once they react with antigens presented in thymic cortical epithelium.
In this report, we have examined the induction of tolerance to a TSCOT+ TEC subset-specific antigen. Our results demonstrate that a small amount of antigen (a few thousand/cell), present in a small number of TEC subset (4,000 cells/thymus), functions as a highly efficient tolerogen in midst of diverse repertoire. In addition, cortical marker CDR1+ TEC delete TCR+ Tg thymocytes without cross-presentation by either DC or mTEC. This type of tolerance was established, at least, partially in an AIRE-independent manner.
From the histological analyses, the thymic microenvironment is already known to be complex in nature, and to change during development [41,51–53]. In addition, histological analyses alone do not constitute a suitable method for delineating expression profiles for different compartments, because of the poor cortical staining [43]. Using a combination of flow cytometry with compartment-specific markers [43] and LacZ reporter staining, cortical expression of the TSCOT locus was confirmed at the newborn stage (Figure 2A and [37,41,51–53]). In the mature thymus, β-gal activity was also principally found in the cortex (Figure 2B), and TSCOT surface expression was exclusively detected in CDR1+ TEC populations, not in conventional mTEC or nonepithelial cells (Figure 2C). However, there are cases in which LacZ activity staining extends to the corticomedullary junction of mature TDLacZ thymus and TSCOT marker also stains uncharacterized minor population of cells (unpublished data). A part of these minor populations could be developing transitional TEC, but this possibility requires further detailed study with an improved technology that can handle an extremely small number of cells. Nonetheless, it is clear that TSCOT+ cells are not part of conventional medullary cells. TSCOT/LacZ was never detected in the conventional CDR1−UEA-1+ mTEC population (Figure 2C). Thus, we are able to dismiss the possibility that LacZ is ectopically expressed in the mTEC of the medulla. Furthermore, TSCOT expression was widely located in the Rag1−/− thymus [38], which lacks mature mTEC (Figure 7C and [54]). In case of BgII mouse on a Rag1−/− background, tolerance at the DN stage was very clear when there was antigen only in the cortical epithelium (Figure 7).
The notion of the exclusion of cortical epithelium in the induction of tolerance was derived from the transgenic expression of MHC molecules exclusively in the cortex of the thymus, using the K14 promoter [25,26]. However, the idea of an exclusive tolerance niche has been challenged: incomplete, but significant, tolerance was observed when other antigens were targeted to cTEC using the same promoter [27,28]. In addition, it has been clearly demonstrated that the K14-MHCII thymus is in fact autotoleragenic when self-antigens are presented by its own cortical epithelial cells [22]. Our current findings using a specific TSCOT promoter corroborate the notion that cTEC participate in the establishment of CD4 central tolerance in a highly efficient manner. There is no evidence for autoreactivity to specific antigens presented by the TSCOT+ cells. Instead, we found clear tolerance to LacZ antigen. Therefore, the TSCOT+ TEC niche of the cTEC subpopulation is not excluded from the tolerance induction so that it may avoid autoreactivity against its own cell.
Taking advantage of sensitive enzymatic activity, our estimate for LacZ under the control of the TSCOT promoter is about 6,000 (the average of two measurements) molecules per TSCOT+ TEC in homozygote thymus (Figure 3). This number is surprisingly similar to that of the estimation of mTEC derived from the indirect estimation [55]. However, one half of this amount in heterozygotes was sufficient to induce complete CD4 tolerance in the absence of mTEC (Figures 7 and 8) or DC cross-presentation (Figure 8). Previous accurate estimates [56] have suggested that recognition of only three to four peptide/MHC complexes by an immature thymocyte was sufficient to generate a negative selection event in transgenic mouse. Therefore, it remains a challenging question as to how such a high efficiency is achieved. The number of cTEC in the adult thymus is far less than that of mTEC [43]. The total number of TSCOT+ TEC, estimated from a large pool of adult thymuses, was only on the order of several thousand per thymus. In order to screen all of the developing thymocytes for potential autoreactivity, the frequency of cell encounters between cTEC presenting the specific antigen and thymocytes would have to be optimized, even considering the average 3-d period in which DP thymocytes remain in the cortex [57,58]. This could be accomplished in thymic nurse cells in which multiple thymocytes are found in association with one epithelial cell.
Several earlier papers had come to the conclusion that the thymic epithelium induced tolerance by the induction of anergy, rather than deletion [59,60]. In contrast, in our anti-LacZ TCR Tg × TDLacZ model, it is evident that deletion is the dominant mechanism (Figure 7). Deletion has also been observed in a number of other TCR transgenic systems [28,61]. Whether this is a normal physiological process or death subsequent to developmental arrest following the premature expression of a transgenic receptor at the DN stage was questioned [62]. However, DP thymocytes included in the RTOC were also deleted, arguing that cTEC-induced tolerance involves the specific deletion rather than the arrest at premature developmental stage (Figure 8B).
In the TDLacZ thymus, cortical epithelium expressing a specific antigen was able to tolerize quite thoroughly (Figures 7 and 8). This was somewhat surprising if thymic epithelial cells are poor presenters of antigen as concluded earlier on the poor expression of costimulatory molecules in cTEC [63]. However, we clearly show that TSCOT+ cTEC express high levels of MHCII, CD54, CD40, and CD80. Surface CD40 and CD80 proteins in particular are expressed at surprising levels that are even higher than those of mTEC (Figure 6). According to the RT-PCR result from the purified cells, CD40, CD80, and CD86 message levels are slightly higher in mTEC than cTEC ([37] and unpublished data). Among cTEC, TSCOT+ cells are the ones expressing more costimulatory molecules (Figures 6 and S1). Therefore, molecules on the TSCOT+cTEC can provide the environment for the highly efficient deletional tolerance of TCR bearing early thymocytes Figure S2) through a unique TSCOT+ cTEC antigen presentation process. As seen in figure 7C, massive apoptosis events in DN TCR transgenic cells in the presence of the LacZ antigen-bearing cTEC are also consistent with the deletional tolerance.
In order to determine the molecular mechanism underlying the induction of tolerance, we determined whether or not AIRE was involved. It has been fairly well established that AIRE is involved in mTEC-mediated tolerance induction by facilitating the expression of peripheral antigens in normal and genetically modified animals [12,64,65]. As a result of the introduction of one copy of the LacZ gene into AIRE-deficient animals, the LacZ-specific CD4 proliferative response was significantly reduced (Figure 5). Since AIRE expression was, mostly, assumed to be absent in the normal CDR1+ cTEC (except Rag1−/−), it is not surprising that AIRE was not directly involved in TSCOT+ TEC-induced tolerance. Instead, it may suggest that the AIRE-independent tolerance pathway exists in the TSCOT+ TEC. However, there was still the question of whether tolerance was induced by AIRE-expressing DC or mTEC, via a cross-presentation. The results obtained with RTOC using thymocytes from LacZ-specific TCR transgenic mice and purified CD45−UEA-1−CDR1+ cells from 2-dGuo–treated FTOC (Figure 7) show that neither DC nor mTEC are necessary for tolerance induction in vitro. The direct involvement of TSCOT+ TEC in deletional tolerance constitutes strong evidence for the capacity of direct antigen presentation [29,32–36]. More detailed studies will be required to identify the specific molecules that are involved in this type of antigen presentation.
It is generally accepted that negative selection requires specific conditions of either high-avidity interaction or prolonged signaling [20,66,67]. The quantitative aspects discussed above seem insufficient to explain negative selection by a simple affinity/avidity model for cTEC. The surface and cytoplasmic levels of MHCII in cTEC are not appreciably lower than in mTEC and MHCII molecules exist on cTEC as aggregates on the surface [43]. Thus, if a self-peptide was presented at sufficiently high concentrations to display multiple complexes in the same aggregate at any one time, these MHCII aggregates could potentially generate high-avidity signaling leading to thymocyte death. If so, then cortical epithelium could function directly in both negative and positive selection. In line with this notion, it has been shown that a single cTEC line can mediate both positive and negative selection [68]. If the amount of any antigen produced by a cTEC is low, then under normal conditions with a random loading mechanism for a diverse set of endogenous peptides [69], a single MHCII aggregate on the cTEC surface would likely contain only one peptide/MHC complex. This would hinder an avidity-based mechanism from operating as there would be no multimeric presentation. Although such monomeric presentation might be adequate for positive selection, it seems that it would be inadequate for negative selection. This raises the possibility that other mechanisms might exist for increasing the antigen density on cTEC. Such a mechanism might involve intercellular antigen transfer [70], in addition to sampling of other self-antigen pools [8,71]. However, the expression of costimulatory molecules on TSCOT+ cTECs is consistent with the idea that the presence of costimulation/second signals may distinguish negative from positive selection.
All mice were handled according to American Association for Accreditation of Laboratory Animal Care Regulations.
In order to generate mice carrying an inserted LacZ allele at the TSCOT locus, a 4.2-kb targeting vector was constructed by cloning IRES-LacZ with a neo-selectable marker from p1049 [42] between two BamHI sites in the first exon of the TSCOT gene. A Herpes Simplex Virus thymidine kinase (HSV-TK) expression cassette was positioned at the 5′ end of the construct in order to facilitate negative selection for homologous recombination. The targeted allele harbors an IRES-LacZ and PGK-neomycin expression cassette within the first exon, resulting in a small deletion (284 bp) between the two BamHI sites within exon 1. The mouse 129 embryonic stem cell line (R1) was electroporated with the construct, and the neomycin-resistant clones were screened in the laboratory of Dr. Hua Gu (National Institute of Allergy and Infectious Diseases [NIAID]/ National Institutes of Health [NIH]). Chimeras were generated by blastocyst injection, and one founder mouse was backcrossed to C57BL/6Tac.
To study antigen-specific CD4+ T cell responses to β-gal, a transgenic mouse strain on a C57BL/6 background was developed and named BgII. RNA was isolated from an I-Ab–restricted, β-gal–specific CD4+ T cell clone. Total mRNA was isolated using a Qiagen RNeasy kit, and the α and β TCR were amplified by 5′-Rapid Amplification of cDNA Ends (5′-RACE, Life Technologies) using constant region anti-sense primers a1 (5′-GGCTACTT TCAGCAGGAGGA-3′) and b1 (5′-AGGCCTCTGCACTGATGTTC-3′), respectively. The 5′-RACE products were amplified with nested TCR α and β constant region primers a2 (5′-GGGAGTCAAAGTCGGTGAAC-3′) and b2 (5′-CCACGTGGTCAGGGAAGAAG-3′), and cloned into pCR4TOPO TA sequencing vectors (Invitrogen). Genomic cloning PCR primers were designed based upon the method previously described [72]. The genomic variable domains were TA cloned into pCR4TOPO (Invitrogen), validated by sequencing, subcloned into TCR cassette vectors kindly provided by Dr. Diane Mathis (Harvard), and coinjected into fertilized C57BL/6 embryos (SAIC) yielding TCR transgenic founder which were then bred.
PCR genotyping.
Tail or ear samples were employed for genotyping, using the red Extract-N-Amp Tissue PCR kit (Sigma) and primers: for the TSCOT locus, Neo primer: ACCGCTATCAGGACATAGCGTTGG, 1C12 F1: TTACTCAAAGTGATGCTGGACTGG, 1C12 B2: CCGAGGGTTCCTTGGTACATTC; for the RAG1 locus, Neo primer: ACCGCTATCAGGACATAGCGTTGG, Rag-1 F: TCGTTTCAAGAGTGACGGGCAC, Rag-1 B: AATCCTGGCAATGAGGTCTGG; and for the AIRE locus, forward primer: GTCATGTTGACGGATCCAGGGTAGAAAGT, reverse primer: AGACTAGGTGTTCCCTCCCAACCTCAG. For the anti-LacZ TCR transgenic allele, BG2 Alpha F1: ACAACCCGGGATTGGACAG, BG2 Alpha R1: GTATAGCGGCCGCCTCCTAGTGCAATGGT, BG2 Beta F1: TATCTCGAGTCCTGCCGTGACCCTACTATG; BG2 Beta R1: CAGCCGCGGAACCCAACACAAAAACTATAC.
Antibodies used for flow cytometric analysis were as follows: for stromal cells, FITC-conjugated anti-mouse I-Ab (Ab) AF6–120.1 (BD Pharmingen), CDR1-PE (prepared by L. Lantz, NIAID flow cytometric facility), CD45 PE-Texas Red conjugate (Caltag), biotinylated Ulex europaeus agglutinin-1 (Vector Laboratories), streptavidin-APC (BD Pharmingen), CLVE1 anti-TSCOT mAb (prepared by Dr. L. Lanz, NIAID), and FITC-conjugated goat anti-Rat IgM (Jackson Laboratories). For thymocytes, FITC-conjugated anti-mouse CD4 (L3T4) (GK1.5) (BD Pharmingen), PE-conjugated anti-mouse CD44 (BD Pharmingen), mouse CD8α PE-Texas Red conjugate (Caltag), biotin-conjugated anti-mouse CD25 (BD Pharmingen), streptavidin-APC (BD Pharmingen), annexinV-FITC (Clontech), PE-conjugated anti-mouse CD69 (BD Pharmingen), CD4 PE-Texas Red conjugate (Caltag), biotin anti-mouse αβTCR (H57–597) (BD Pharmingen), FITC-conjugated anti-mouse CD44 (BD Pharmingen), PE-conjugated anti-mouse CD25 (BD Pharmingen), APC-conjugated anti-mouse CD4 (L3T4) (GK1.5) (BD Pharmingen), and APC-conjugated anti-mouse CD8α (BD Pharmingen). For LacZ staining, after treating with 30 mM chloroquine diphosphate to block endogenous β-gal activity for 30 min at 37 °C, 33 μM ImaGene Red C12RG substrate (Molecular Probes, ImaGene Red C12RG lacZ Gene Expression Kit I-2906) was used in FACS buffer for 20 min at 4 °C.
Either whole embryos or isolated thymuses were washed in PBS and fixed in 1% paraformaldehyde, 0.2% glutaraldehyde, 0.02% NP-40, 1 mM MgCl2 in PBS for 1 or 2 h on ice. Staining was conducted using X-gal solution with 100 mM d-galactose in 2 mM MgCl2, 5 mM potassium ferricyanide, 5 mM potassium ferrocyanide overnight at 37 °C [42]. For the sections, the thymuses were embedded in Tissue Freezing Medium (Triangle Biomedical Sciences). The 4-μm sections were fixed for 2 min in 1% formaldehyde, 0.2% glutaraldehyde, 0.02% NP-40 1 mM NaCl, then incubated with X-gal solution (1 part X-gal 40 mg/ml in dimethyl formamide, in 40 parts 2 mM MgCl2, 5 mM potassium ferricyanide, 5 mM potassium ferrocyanide in PBS) at 37 °C for 48 h. For antibody staining, the paraffin sections were stained with DAKO CSA reagent. For the LacZ Lysis assays, thymic cells from 30 TDLacZ mice or control C57BL/6 mice were partially purified via MACS CD45-bead sorting (Miltenyi Biotech). The cell pellets were lysed using Reporter Lysis Buffer from the β-Galactosidase Enzyme Assay System with Reporter Lysis Buffer (Promega). After lysis and centrifugation, the Assay Buffer was added to each supernatant as well as to enzyme aliquots for a standard curve, and then incubated for 30 min at 37 °C. Absorbance was then measured at a wavelength of 405 nm.
Isolated TEC (about 105) were washed in cold FACS buffer (PBS + 1% BSA), subsequently stained on ice with 2.4G2, APC-conjugated anti-CDR1, biotinylated anti-CD80 (B7–1, Armenian hamster IgG2κ) followed by streptabidin-Alexa568 (Molecular Probes). For the detection of LacZ-expressing cells, ImaGene Green C12FDG lacZ Gene Expression kit (Molecular Probes) was used. Stained samples were placed on the slides by cytospin at 1,200 rpm for 2 min. Images were collected on LSM 510 META (Zeiss), and analyzed with LSM Image Examiner (Zeiss) and Photoshop.
Three mice per group were immunized with the affinity-purified, LPS-removed recombinant proteins in Complete Freund's Adjuvant in one footpad and the base of the tail. Ten days after immunization, the mice were sacrificed and the inguinal, mesenteric, and para-aortic lymph nodes were collected and crushed in Iscov's Modified Dulbecco's Medium. The CD4+ cells were collected from MACS columns using anti-CD4 antibody (GK1.5) (purity of the cells were usually over 95%) and incubated at 37 °C with irradiated whole spleen cells and 0, 1, 10, or 100 μg of LacZ protein (in triplicate). The cells were pulsed with 3H-thymidine for the final 24 h of a 72-h incubation. The cells were harvested with a Brandel 96-well harvester, and thymidine incorporation into DNA was measured with a Wallac Trilux 1450 β-scintillation counter.
The mice were immunized intraperitoneally with either LacZ or purified recombinant GST-TSCOT-Loop protein in CFA, three times every other week. The mice were bled 3 d after the last injection, and the sera were incubated on His-LacZ protein or His-Loop–coated (5 μg/well) ELISA plates. The bound anti-LacZ Ab was detected with anti-mouse immunoglobulin isotype-specific antibodies conjugated with HRP and assayed with ABTS solution (Southern Biotechnology Associates) by following the manufacturer's description. Optical density (OD) was measured at 405 nm.
The thymic stromal cells were prepared by treating E14.5 fetal thymus samples with 2-dGuo for a week. UEA-1− and CD45− cells were then purified using biotinylated reagents and streptavidin-MACS beads. Thymocytes from anti-LacZ TCR transgenic mice were sorted for DP and DN cells and reaggregated with the stromal cells using protocols developed by Anderson and Jenkinson at the University of Birmingham, United Kingdom [49]. Recovered cells were counted and then analyzed by flow cytometry after 4–6 d of culture.
|
10.1371/journal.pbio.1000072 | Mammalian Kinesin-3 Motors Are Dimeric In Vivo and Move by Processive Motility upon Release of Autoinhibition | Kinesin-3 motors drive the transport of synaptic vesicles and other membrane-bound organelles in neuronal cells. In the absence of cargo, kinesin motors are kept inactive to prevent motility and ATP hydrolysis. Current models state that the Kinesin-3 motor KIF1A is monomeric in the inactive state and that activation results from concentration-driven dimerization on the cargo membrane. To test this model, we have examined the activity and dimerization state of KIF1A. Unexpectedly, we found that both native and expressed proteins are dimeric in the inactive state. Thus, KIF1A motors are not activated by cargo-induced dimerization. Rather, we show that KIF1A motors are autoinhibited by two distinct inhibitory mechanisms, suggesting a simple model for activation of dimeric KIF1A motors by cargo binding. Successive truncations result in monomeric and dimeric motors that can undergo one-dimensional diffusion along the microtubule lattice. However, only dimeric motors undergo ATP-dependent processive motility. Thus, KIF1A may be uniquely suited to use both diffuse and processive motility to drive long-distance transport in neuronal cells.
| Molecular motors transport a wide variety of cellular cargoes that are important for diverse cellular phenomena such as mitosis, polarity, motility, and secretion. Motor activity must be tightly regulated to ensure that ATP hydrolysis and processive motility occur only upon coupling to the correct cargo. In neuronal cells, Kinesin-3 motors drive the transport of presynaptic vesicles and other membrane-bound organelles along microtubule tracks. Yet the mechanisms of Kinesin-3 motor activation and motility remain controversial. In this study, we examine the regulation and motile properties of the Kinesin-3 motor KIF1A. We show that in the absence of cargo, KIF1A motors exist in a dimeric inactive state that is maintained by two distinct autoinhibitory mechanisms. This suggests a simple model for activation of dimeric motors upon cargo binding. We also show that dimeric motors can undergo two mechanisms of motility along microtubule tracks: one-dimensional diffusion and ATP-driven processive motility. This unique property may facilitate the ability of KIF1A to drive long-distance vesicular transport in neuronal cells.
| Kinesin motors drive the long-distance transport of membrane-bound cargoes along microtubules. Long-distance transport is particularly important in neuronal cells whose length and polarity require robust sorting and transport of cargoes to pre- and postsynaptic destinations. Transport of synaptic vesicle precursors to axon terminals is driven by members of the Kinesin-3 family, the mammalian KIF1A, and Caenorhabditis elegans Unc104 motors [1]. Loss of KIF1A or Unc104 function results in decreased synaptic vesicles in axonal growth cones, and early death [1]. Thus, understanding how kinesin motors are regulated to enable transport of the correct cargo to the proper cellular destination at the relevant time is an important biological problem.
In the absence of cargo, kinesin motors are kept inactive to prevent futile ATP (adenosine triphosphate) hydrolysis and motility. Two models have been proposed for how activity is suppressed in the absence of cargo. The first model posits that dimeric motors are regulated by an autoinhibitory mechanism. Autoinhibition typically involves a folded state that enables the motor's own tail domain to interact with and inhibit its motor domain. This model is based on a large body of work on the Kinesin-1 motor (formerly conventional kinesin or KIF5) [2–5]. In recent years, this model has received increasing experimental support from studies on kinesin motors involved in diverse functions such as epithelial polarity, intraflagellar transport, and mitosis [6–8]. Interestingly, autoinhibition may be a general model for motor regulation, as two well-studied members of the myosin family, nonmuscle myosin II and myosin V, exist in a folded inactive state [9–11]. Autoinhibition enables precise spatial and temporal regulation of motors and may be relieved by cargo binding [6,12], phosphorylation [8], or other mechanisms.
The second model states that motor activity is regulated by transition from a monomeric to dimeric state. Evidence for this model comes from studies on KIF1A/Unc104 motors where the full-length motors exist in a monomeric, inactive state [13–15]. Unc104 activity can be increased by forced dimerization or by an increase in the local concentration of the motor on liposomes [16–18]. Thus, cargo-induced dimerization would enable KIF1A/Unc104 motors to coordinate their two motor domains and step processively in a “hand-over-hand” fashion [19,20]. The cargo-induced dimerization model has gained support from recent studies on the myosin family member myosin VI [21–24]. In this case, the cargo-binding tail domain exists in two mutually exclusive situations: either directly inhibiting the catalytic head in the monomeric state or mediating dimerization on the cargo membrane [25].
The cargo-induced dimerization model for Kinesin-3 motors is primarily based on work with recombinant CeUnc104 proteins. Whether mammalian KIF1A motors are regulated by cargo-induced dimerization has never been tested. In addition, although a region of the KIF1A stalk domain has been shown to inhibit microtubule binding of the motor domain [26], it is unclear how this potential autoinhibitory segment fits in the context of the two models for motor regulation. Furthermore, the sequences that facilitate dimerization remain to be identified.
Here, we test the models for regulation of mammalian KIF1A motors and relate activity to the monomer/dimer state. We show that expressed and endogenous mammalian KIF1A motors exist in a dimeric, inactive state. Thus, cargo-induced dimerization is not a valid model for mammalian KIF1A motors. Rather, we provide support for an autoinhibition mechanism by identifying the sequences and mechanisms required for dimerization and inhibition. Finally, we show that only dimeric motors undergo processive, ATP-driven motility.
To study the regulation and motile properties of mammalian KIF1A under native conditions, rat KIF1A (Figure S1) was tagged with monomeric citrine (mCit), a variant of yellow fluorescent protein (FP), and expressed in COS cells. This approach has been used successfully to study Kinesin-1 motors and avoids potential problems associated with in vitro purification and/or labeling [2]. In live cells, motor activity can be determined using the nonhydrolyzable ATP analog AMPPNP (5′-adenylyl-beta,gamma-imidodiphosphate) to block the release of active kinesin motors from microtubules [2]. COS cells expressing mCit-KIF1A were transiently permeabilized with the bacterial toxin streptolysin O (SLO). Upon addition of AMPPNP, mCit-KIF1A did not become trapped on microtubules but remained diffuse and cytosolic (Figure 1A and 1B), indicating that KIF1A is not engaged with microtubules when expressed in mammalian cells. Identical results were obtained when the mCit tag was placed at the C-terminus of KIF1A (KIF1A-mCit, Figure S2). The inability to bind microtubules is inherent to KIF1A, as untagged and Myc-tagged versions of KIF1A also did not become trapped on microtubules (Figure S2). In contrast, active versions of Kinesin-1 rapidly became locked on microtubules upon addition of AMPPNP ([2] and Figure S2). We conclude that mammalian KIF1A is inactive in vivo.
KIF1A could be inactive due to either of the two proposed models for kinesin motor regulation. To distinguish between these possibilities, we first investigated whether expressed KIF1A motors exist as monomers or dimers using coimmunoprecipitation. COS cell lysates coexpressing mCit- and Myc-tagged KIF1A proteins were immunoprecipitated with control antibodies (immunoglobulin G [IgG]) or antibodies to the Myc tag. Coprecipitation of mCit-KIF1A with Myc-KIF1A was observed only with the Myc antibody (Figure 1C), suggesting that KIF1A exists in a dimeric state. As an alternative method, we used chemical crosslinking. Myc- or mCit-tagged versions of KIF1A were expressed in COS cells, and lysates were untreated or treated with dimethylpimilimidate (DMP). In the presence of DMP, both Myc-KIF1A and mCit-KIF1A motors migrated at approximately twice the molecular weight (MW) of the corresponding noncrosslinked proteins (Figure 1D, lanes 1–6). A known dimeric motor, the kinesin heavy chain (KHC) subunit of Kinesin-1, undergoes a similar DMP-induced mobility shift (Figure 1D, lanes 7 and 8). In contrast, mCit showed no shift in mobility upon DMP treatment (Figure 1D, lanes 9 and 10). These results indicate that the expressed KIF1A protein exists in a dimeric state.
We next probed the monomer/dimer state of KIF1A using fluorescence resonance energy transfer (FRET) stoichiometry, a method that calculates an average FRET efficiency (EAVE) and minimizes effects of variable donor and acceptor FP expression [2,27]. Control experiments show an EAVE = 0% for unlinked donor (monomeric cyan FP [mCFP]) and acceptor (mCit) FPs and an EAVE = 37% for mCFP and mCit linked by 16 amino acids (unpublished data). For donor and acceptor FPs placed on the N-terminus of KIF1A (“motor-to-motor” FRET), a low but measurable FRET efficiency was obtained (EAVE = 3.0 ± 1.2%, n = 45 cells, Figure 1E and 1F). This value is not compatible with a monomeric state of the motor. Rather, this value indicates that KIF1A motors exist in a dimeric state in live cells. The motor-to-motor FRET efficiency of KIF1A is comparable to that of the Kinesin-1 holoenzyme (EAVE = 2.1 ± 0.4%, n = 30 cells, Figure 1E and 1F) whose motor domains are “pushed apart” in the inactive state [2]. By analogy, it is thus possible that the two motor domains of inactive, dimeric KIF1A motors may be pushed apart as part of the regulatory mechanism. FRET stoichiometry was also used to probe the overall conformation of KIF1A. For donor and acceptor FPs localized at the N- and C-termini of a single KIF1A polypeptide (mCit-KIF1A-mCFP), the low but measurable “motor-to-tail” FRET (EAVE = 4.8 ± 1.1%, n = 39 cells, Figure S3) confirms that KIF1A is not a fully extended molecule. We conclude that KIF1A exists in a compact dimeric state but is not active for microtubule binding.
That KIF1A motors expressed under native conditions exist in a dimeric state was surprising as recombinant and endogenous KIF1A/Unc104 motors have been defined as monomeric kinesins based on hydrodynamic analysis. Thus, we tested whether endogenous KIF1A motors also exist in a dimeric state. We first used crosslinking of cytosolic extracts from rat brain and detergent extracts of murine cortical neurons. In the presence of DMP, the endogenous KIF1A proteins showed a reduced mobility (Figure 1G), suggestive of a dimeric state. Crosslinking analysis is dependent on the ability of the antibody to recognize the crosslinked species. Unfortunately, the KIF1A antibody was less able to recognize its antigenic sequence after crosslinking than were antibodies to the epitope tags (Figure 1D). Thus, we sought an alternative method to analyze the monomer/dimer state of endogenous KIF1A motors. Previous reports used hydrodynamic analysis of KIF1A as compared to MW standards [13]. As the shape of kinesin motors may influence their hydrodynamic properties, we sought to compare the sedimentation of endogenous KIF1A motors in sucrose gradients to motors with known oligomeric states. Cytosolic extracts of rat brain or detergent extracts of COS cells expressing dimeric Myc-KIF1A or mCit-KIF1A motors were separated by sucrose gradient sedimentation. The majority of the expressed Myc-KIF1A was found in fractions 5–7 (Figure 1H), whereas the majority of the expressed mCit-KIF1A was found in fractions 6 and 7 (Figure 1H) with the shift likely due to the mCit tag. The endogenous Kinesin-1 protein in rat brain cytosol was also found in fractions 6 and 7 (Figure 1H). Strikingly, the endogenous KIF1A protein in rat brain cytosol was found in fractions 5–7 (Figure 1H), a mobility identical to that of the expressed dimeric Myc-KIF1A and mCit-KIF1A proteins. It is interesting to note that Kinesin-1 (∼360 kDa) and KIF1A motors (∼380–440 kDa) sediment slower than the marker protein catalase (240 kDa) despite their larger size. Thus, sedimentation as compared to MW standards is not a reliable indication of motor mass. Taken together, these results indicate that endogenous KIF1A exists as a dimeric protein.
That KIF1A exists in a dimeric and inactive state suggests that dimerization is not sufficient for activation. Thus, we tested an autoinhibitory mechanism for KIF1A regulation as related to the domain structure and dimer state of the protein. In KIF1A, a coiled-coil (CC) segment adjacent to the motor domain, referred to as the neck coil (NC), is predicted by the COILS program only if a 14–amino acid window is used for analysis (Figure 2A). The NC is followed by a region of strong coiled-coil prediction (CC1), a forkhead-associated (FHA) domain, and coiled-coil segments CC2 and CC3 (Figure 2A).
Based on this domain structure, we created truncated versions of KIF1A (Figure 2B) by placing a mCit tag after CC2 [KIF1A(1–726)], after CC1 [KIF1A(1–491)], or after the NC [KIF1A(1–393)]. We first tested the ability of these constructs to bind to microtubules. KIF1A(1–726)-mCit remained cytosolic and did not localize to microtubules upon addition of AMPPNP (Figure 2C–2E), indicating that this protein retains the autoinhibited state of the full-length (FL) molecule. In contrast, deletion of the FHA and CC2 segments resulted in a motor, KIF1A(1–491)-mCit, that became trapped in a microtubule-bound state upon addition of AMPPNP (Figure 2C–2E). Thus, the FHA and CC2 domains contribute to autoinhibition by blocking productive interactions with microtubules. This is consistent with previous work on truncated KIF1A/Unc104 motors in vitro [26,28]. Interestingly, deletion of the CC1 domain resulted in a construct, KIF1A(1–393)-mCit, that accumulated on peripheral microtubules at steady state, suggesting that this may be a processive motor. Addition of AMPPNP resulted in an increase in microtubule localization, most notably in the central regions of the cell (Figure 2C). These results suggest that CC1 negatively regulates motility of KIF1A.
To directly analyze the processive motility of FL and truncated KIF1A motors, we used two assays. First, motility in vivo was indirectly assessed by the ability of KIF1A motors to accumulate at the tips of neurites in neuron-like CAD cells. Second, processive motility in vitro was directly measured using single-molecule motility assays and a total internal reflection fluorescence (TIRF) microscope. For the latter, KIF1A constructs were tagged with three tandem copies of mCit for improved signal and decreased photobleaching and photoblinking [29].
When expressed in differentiated CAD cells, KIF1A-mCit and (1–726)-mCit were diffusely localized in the cell body and did not concentrate at the ends of neuronal processes (Figure 2F). In addition, very few motility events were observed in vitro for KIF1A-3xmCit or (1–726)-3xmCit motors (Figure 2G and 2H, Table 1). These results indicate that truncation of the C-terminal half of KIF1A is not sufficient to relieve autoinhibition of microtubule binding (Figure 2C–2E) or processive motility (Figure 2F–2H). Truncation of the FHA+CC2 region also resulted in little to no processive motility for KIF1A(1–491) in vivo (Figure 2F) or in vitro (Figure 2G and 2H, Table 1). This was surprising as KIF1A(1–491) motors bound to microtubules in vivo (Figure 2C–2E). Thus, the FHA+CC2 region contributes to autoinhibition by blocking microtubule binding, but an additional mechanism(s) controls the ability of the motor to undergo processive motility. This second control mechanism may reside, at least in part, in the CC1 domain as KIF1A(1–393) showed a significantincrease in motility as these motors concentrated at the tips of neurites (Figure 2F) and displayed a large number of processive motility events in vitro (Figure 2G and 2H, Table 1). Analysis of the motile properties of KIF1A(1–393) motors gave an average speed of 1.36 ± 0.04 μm/s and an average run length of 1.24 ± 0.06 μm per event (Table 1), comparable to previous studies [13,14,16]. The few motility events observed for the FL, (1–726), and (1–491) motors occurred with decreased velocity and run lengths as compared to (1–393) and were likely due to a dynamic equilibrium between active and inactive states.
Taken together, these results indicate that two regions of KIF1A contribute to autoinhibition. First, the FHA and CC2 domains (amino acids 492–726) prevent microtubule binding (Figure 2), and second, the CC1 domain (amino acids 394–491) inhibits processive motility (Figure 3).
To further investigate the relationship between KIF1A domain structure, autoinhibition, and dimerization, we used three assays to test whether truncated motors exist in a dimeric state. We first used chemical crosslinking of KIF1A(1–726)-mCit, KIF1A(1–491)-mCit, and KIF1A(1–393)-mCit motors. In the presence of DMP, most of the (1–726)-mCit and (1–393)-mCit proteins shifted to a higher MW species (Figure 3A), consistent with a dimeric state. In contrast, very little (1–491)-mCit protein displayed a shift in mobility (Figure 3A). Rather, in the presence of DMP, (1–491)-mCit motors showed either the same mobility as the uncrosslinked species or a slightly increased mobility (Figure 3A), perhaps due to an intramolecular crosslink between the CC1 and NC domains [28].
We next used coimmunoprecipitation of Myc- and mCit-tagged truncated KIF1A motors. Coexpression of (1–726)-Myc and (1–726)-mCit resulted in precipitation of (1–726)-mCit by the Myc antibody but not a control antibody (Figure 3B), suggesting that KIF1A(1–726) exists in a dimeric state. Similar results were obtained for the (1–491) and (1–393) truncated motors (Figure 3B). Thus, (1–726) and (1–393) behaved as dimeric proteins by both chemical crosslinking and coimmunoprecipitation, whereas (1–491) showed a more varied behavior.
We then took the advantage of stepwise photobleaching of FPs to test whether truncated versions of KIF1A exist as dimers. This assay has the advantage of analyzing individual motors rather than ensemble averages. Lysates of COS cells expressing 3xmCit-tagged FL or truncated KIF1A motors were analyzed by TIRF microscopy. For each construct, the fluorescence intensity of 160–200 individual fluorescent spots was recorded over time. The number of bleaching steps was determined for each spot (Figure 3C) and then plotted in a histogram to show the population distribution (Figure 3D). In control experiments, the majority of KHC(1–891)-3xmCit motors bleached in four to six steps (Figure 3D and [29]), consistent with the presence of six FPs in the dimeric KHC molecule. A similar distribution of bleaching events was obtained for FL, (1–726), and (1–393) motors (Figure 3D), indicating a dimeric state for these KIF1A constructs. In contrast, the majority of KIF1A(1–491)-3xmCit motors bleached in two to three steps, consistent with a monomeric state (Figure 4D). Taken together, these data suggest that truncated KIF1A constructs that contain only the NC and CC1 domains exist as monomers that are not capable of processive motility.
To directly demonstrate that removal of the CC1 domain restores processive motility as well as the dimeric state, we used two-color TIRF imaging to simultaneously track mCit and mCherry fluorescence in lysates of COS cells coexpressing KIF1A(1–393)-3xmCit and KIF1A(1–393)-3xmCherry. That (1–393) motors move as dimeric molecules is indicated by observations of fluorescent spots labeled with both mCit and mCherry that move together in a linear fashion (representative track, Figure 3E). We also observed mCit and mCherry fluorescent spots with nonoverlapping motility events, as observed when KIF1A(1–393)-3xmCit and KIF1A (1–393)-3xmCherry motors were expressed separately (unpublished data). We believe that these mCit-only and mCherry-only fluorescent spots are also dimeric motors based on their fluorescence intensity. The average maximum mCit fluorescence of mCit-only spots was 513.4 ± 54.0 arbitrary units (n = 65 spots). This value is significantly greater than the average maximum mCit fluorescence intensity (356.8 ± 51.4 arbitrary units, n = 44 spots) of spots that colabeled and comigrated with mCherry. These results provide the first direct demonstration that KIF1A moves in a directed manner as a dimeric motor.
To identify the sequences required for dimerization, we generated KIF1A constructs containing various amounts of the NC region (1–369, 1–377, or 1–381, Figure 4A) or the full NC and several residues of the subsequent hinge region (1–393, Figure 4A). These constructs were designed to directly correspond to previously studied KIF1A motors (Figure 4A). The monomer/dimer state of the NC truncations was tested by crosslinking and photobleaching analysis. In the presence of DMP, the majority of (1–393)-mCit shifted to a higher MW species (Figure 4B). In contrast, (1–381)-mCit, (1–377)-mCit, and (1–369)-mCit showed no mobility change in the presence of DMP (Figure 4B). In photobleaching experiments, a large proportion of KIF1A(1–393)-3xmCit molecules bleached in four to six steps (Figure 4C and 4D), consistent with a dimeric state containing six FPs. However, KIF1A(1–381)-3xmCit and KIF1A(1–369)-3xmCit molecules bleached primarily in two or three steps (Figure 4C and 4D), indicating a monomeric state. These results suggest that the entire NC, as well as residues in the subsequent hinge region, are required for dimerization. This is consistent with a study of synthesized NC peptides where several residues beyond G387 were required to prevent dissociation [30].
We next set out to compare the microtubule-based properties of monomeric and dimeric KIF1A motors. We first confirmed that monomeric motors retain the ability to bind to microtubules. Indeed, upon AMPPNP treatment, monomeric (1–381)-mCit and (1–369)-mCit motors became locked on microtubules, similar to dimeric (1–393)-mCit motors (Figure 5A and 5B). We next tested the ability of the NC constructs to undergo processive motility in vivo. The dimeric motor (1–393)-mCit accumulated in neurite tips of differentiated CAD cells, whereas the monomeric motors (1–381)-mCit and (1–369)-mCit remained primarily in the cell bodies (Figure 5C). These results confirm that the full NC is required for dimerization as well as processive motility in mammalian cells.
Finally, we investigated the motile characteristics of (1–393)-3xmCit, (1–381)-3xmCit, and (1–369)-3xmCit motors by in vitro single-molecule motility assays. Motility events that persisted for at least five frames (500 ms) were analyzed to determine velocities and run lengths. As expected, a large number of motility events were observed for dimeric (1–393) motors (Figure 5D and 5E, Table 1). Surprisingly, motility events were also observed for the monomeric (1–381) and (1–369) motors (Figure 5D and 5E, Table 1) even though these motors were not processive in vivo (Figure 5C). However, the motility of the monomeric motors differed from that of the dimeric motor in three ways. First, monomeric motors showed a significant decrease in the number of observed motility events (Table 1). Second, qualitative analysis of the velocity and run-length histograms shows that only dimeric motors gave distributions (Gaussian and single exponential, respectively) characteristic of processive motors (Figure 5D and 5E). Third, dimeric motors underwent significantly longer run lengths (in some cases >6 μm) as dimeric (1–393) motors averaged 1.24 ± 0.06 μm/run, whereas monomeric (1–380) and (1–369) motors averaged only 0.42 ± 0.02 μm/run and 1.02 + 0.09 μm/run, respectively (Figure 5E, Table 1). Thus, monomeric motors are significantly impaired in their motile properties.
We then used two approaches to determine whether the motility of dimeric KIF1A(1–393) motors occurs by one-dimensional (1D) diffusion, as demonstrated for monomeric KIF1A motors, or by ATP-driven processive motility, as demonstrated for other dimeric kinesin motors [19,20]. Single-molecule motility assays were analyzed to include fluorescent spots visible and motile for at least three frames (300 ms) to ensure inclusion of diffusive events (representative times series, Figure 6A). This resulted in an expected decrease in the average run lengths of both dimeric (1–393)-3xmCit and monomeric (1–369)-3xmCit motors (0.90 ± 0.05 μm/run and 0.65 ± 0.04 μm/run, respectively, Figure 6C) as well as increases in the average velocities (1.88 ± 0.07 μm/s and 2.16 ± 0.10 μm/s, respectively, Figure 6B).
Our first approach was to compare the biophysical properties of a large number of motility events by plotting the data as a comparison of the velocities or run lengths against the duration of each motility event (time spent in one-directional motion). By this analysis, drastic differences between monomeric and dimeric mechanisms of motility became apparent. The motility events of dimeric (1–393)-3xmCit motors could be segregated into two classes: first, motile events lasting for short periods of time (<1 s) at a wide variety of speeds and distances (Figure 6D and 6E, top left panels, red circles) and second, motile events lasting for longer time periods (>1 s) at constant speeds of ∼1.2 μm/s (Figure 6D, top left panel, blue circle) and with run lengths directly dependent on the amount of time spent in motion (Figure 6E, top left panel, blue circle). We hypothesized that the first class of motility (Figure 6D and 6E, red circles) is 1D diffusion, whereas the second class of motility (Figure 6D and 6E, blue circles) is processive motility. Indeed, monomeric KIF1A(1–369)-3xmCit motors only moved for short periods of time (<1 s) at a wide variety of speeds and distances (Figures 6D and 6E, middle left panels, red circles), indicative of 1D diffusion. In contrast, the processive motility of dimeric KHC(1–891)-3xmCit motors was evident as these motors spent longer periods of time in motion (>1 s) at a constant speed (Figure 6D, bottom left panel, blue circle) and for distances directly dependent on the time spent in motion (Figure 6E, bottom left panel, blue circle). Thus, dimeric KIF1A(1–393) motors display motility properties of both 1D diffusion and processive motility.
Our second approach was to distinguish these motility classes by their ATP dependence. Diffusion of monomeric KIF1A motors occurs in the weakly bound or ADP (adenosine diphosphate) state [31], whereas processive hand-over-hand stepping of kinesin motors requires the energy of ATP hydrolysis. When single-molecule motility assays of monomeric (1–369)-3xmCit motors were carried out in the presence of ADP, no change in motility was observed. Monomeric motors continued to move only for short periods of time (<1 s) at various speeds and run lengths (Figure 6D and 6E, middle right panels, red circles), confirming that monomeric KIF1A(1–369)-3xmCit motors moved by 1D diffusion. In contrast, the processive motility of dimeric KHC(1–891)-3xmCit motors was abolished in the presence of ADP (Figures 6D and 6E, bottom right panels). The motile properties of dimeric (1–393)-3xmCit motors were also highly dependent on nucleotide. Dimeric (1–393)-3xmCit motors continued to display short motility events (<1 s) of various velocities and run lengths (Figure 6D and 6E, top right panels, red circles) in the presence of ADP, similar to the monomeric motors. However, dimeric motors were unable to undergo processive motility in the presence of ADP (Figure 6D and 6E, top right panels). This analysis demonstrates that for motility events that last only short time periods, it is not possible to distinguish 1D diffusion from processive, hand-over-hand motor stepping (overlap of red and blue circles). However, for events that last longer than 0.5–1 s, comparisons of velocity and run length to time spent in directional motility enable the separation of motility mechanisms. We conclude that KIF1A motors exist as dimeric molecules that can move by 1D diffusion but show processive motility only in the presence of ATP.
The motile properties of kinesin motors must be tightly coupled to the binding and transport of cargo. Two models have been proposed to account for the inactive state of kinesin motors in the absence of cargo. We show that the current model for Kinesin-3 motors, that inactive monomeric motors are activated by cargo-induced dimerization, is not valid for KIF1A as expressed and endogenous motors exist in a dimeric and inactive state. We show instead that regulation is due to autoinhibition by nonmotor regions. Thus, autoinhibition of dimeric motors appears to be a general mechanism for regulating molecular motors.
We provide four lines of experimental evidence (chemical crosslinking, coimmunoprecipitation, FRET, and photobleaching) to demonstrate that FL KIF1A expressed in mammalian cells exists as a dimeric protein. A concentration-dependent push to the dimeric state seems unlikely since crosslinking, coimmunoprecipitation, and photobleaching assays are carried out under dilute conditions. We also show that endogenous KIF1A behaves as a dimeric motor by crosslinking and sucrose gradient sedimentation. The sedimentation behavior of endogenous KIF1A was previously interpreted, based on comparisons to size standards, as indicating a protein of approximately 200 kDa and thus a monomeric molecule [13]. However, we show that the sedimentation behavior of endogenous KIF1A is consistent with a dimeric state when compared to Kinesin-1 and KIF1A motors whose oligomeric state has been confirmed in alternative assays.
The sedimentation behavior of KIF1A is not surprising considering the behavior of other kinesin motors in these assays. Values of 6.7 S have been reported for Kinesin-1 (∼340 kDa) [32], 8.6–9.8 S for the heterotrimeric Kinesin-2 motor KIF3A/KIF3B/KAP (∼300 kDa) [33,34], and 6.8–7.9 S for the homodimeric Kinesin-2 motor CeOSM-3 (∼160 kDa) [7,34]. Thus, the oligomeric state of kinesin motors cannot be conclusively determined in sucrose gradients, and perhaps not in other biochemical analyses, by using MW standards and without considering other variables such as protein conformation. In conclusion, our data provide strong support for the idea [6,35] that, like Kinesin-1 and Kinesin-2 motors, Kinesin-3 family members are dimeric motors.
Our results further show that the NC plays an important role in mediating dimerization. This is consistent with studies showing coil formation by peptides corresponding to the predicted NC regions of MmKIF1A and CeUnc104 [15,30]. Importantly, we show that several residues in the hinge segment C-terminal to the NC are required for dimerization and processive motility, providing an explanation for why previous constructs that lack a full NC resulted in monomeric motors [16,28,31,36,37]. A role for the hinge region in dimerization and/or processive motility has been noted for other kinesins, most notably fungal Kinesin-1 [38].
Our results demonstrate that KIF1A motors can exist in a dimeric but autoinhibited state. These results do not support a model for cargo-mediated dimerization of Kinesin-3 motors but rather suggest a simple model in which cargo binding relieves autoinhibition of dimeric motors. Using truncation mutants, we identified two regions that contribute to different mechanisms of autoinhibition. First, the FHA and CC2 domains inhibit the interaction of KIF1A with microtubules as KIF1A(1–491) motors can bind to microtubules but cannot undergo processive motility, consistent with previous work [39]. Second, the CC1 domain regulates processive motility as KIF1A(1–393) motors can both bind to microtubules and undergo processive motility. We show that CC1 blocks processive motility by interference with the formation of dimeric motors, perhaps due to an intramolecular interaction with the NC as seen in truncated CeUnc104 motors [28]. This interaction may contribute to the altered mobility of (1–491) motors upon crosslinking. Further studies are required to test whether CC1-mediated inhibition occurs in the FL motor. In this respect, the interaction of CC1 and NC could serve to keep the two motor domains separated and/or uncoordinated, similar to the role of kinesin light chain (KLC) in the autoinhibition of Kinesin-1 [2]. As two separate mechanisms are also involved in the autoinhibition of Kinesin-1 [2], the dual-inhibition mode may be a common way to inhibit the activity of kinesin motors in the absence of cargoes.
Autoinhibition is a common regulatory strategy used in diverse biological systems [40]. In terms of motor regulation, autoinhibition prevents futile ATP hydrolysis and allows rapid and specific control of motor activity both temporally and spatially. Autoinhibition has now been shown to regulate members of the Kinesin-1, Kinesin-2, and Kinesin-3 families (this study and [5–7]). Thus, the evolutionary development of new kinesin families by the attachment of novel cargo-binding domains onto the catalytic kinesin core may have simultaneously ensured that each family adopt a unique mechanism for autoinhibition.
Our finding that endogenous KIF1A motors exist in a dimeric state indicates that the physiological form that drives processive motion of cargoes in cells is likely to be the dimeric motor. We show that the motility of dimeric KIF1A motors is characterized by smooth, unidirectional movement along microtubules and occurs with an average velocity comparable to vesicles moving in vivo [39,41]. Whether dimeric KIF1A motors utilize a hand-over-hand mechanism like Kinesin-1 is a question for future experiments. That both monomeric and dimeric forms of KIF1A undergo 1D diffusion on microtubules likely indicates a nonspecific binding to the microtubule that enhances processivity.
The ability of dimeric KIF1A motors to undergo both diffusive motion and processive motility is unique among kinesins involved in vesicle transport. Yet, it is interesting that 1D diffusion along microtubules has been reported for all three motor classes. The homotetrameric Kinesin-5 motor Eg5 undergoes ATP-independent 1D diffusion in the absence of cargo and switches to ATP-dependent processive motility in the presence of cargo, resulting in the sliding antiparallel microtubules during formation of the mitotic spindle [42,43]. Similarly, mitotic centromere-associated kinesin (MCAK), a Kinesin-13 motor, uses ATP-independent 1D diffusion to reach microtubule plus ends where ATP-dependent catalytic activity results in depolymerization of microtubules [44]. Surprisingly, Myosin Va can undergo rapid diffusion along the microtubule lattice in vitro [45]. Finally, diffusive motion of the dynactin complex along microtubules may facilitate cytoplasmic dynein processivity [46]. KIF1A may thus be uniquely positioned among vesicular motors to drive long-distance motility using 1D diffusive motion to tether motors to microtubules, as well as ATP hydrolysis for force production.
FL or truncated rat KIF1A constructs tagged with mCit, 3xmCit, or Myc were generated using convenient restriction sites or PCR and cloned into vectors mCit-N1/C1 (i.e., Clontech's EYFP-N1/C1 vectors), 3xmCit-N1 [29], or pRK5-Myc, respectively. Rat KIF1A(1–393)-mCit was a gift from Gary Banker. All plasmids were verified by DNA sequencing. KHC(1–891)-3xmCit has been described previously [29]. The following antibodies were used: KIF1A (BD Transduction Labs), Myc (Sigma), and GFP (Invitrogen). HA (Covance) and Flag (Sigma) were used as control IgG.
COS and CAD cells were cultured, transfected, and processed for immunofluorescence or immunoprecipitation as previously described [47]. For crosslinking, cells were lysed in Borate buffer (50 mM NaBorate, 100 mM potassium acetate, 2 mM MgCl2, 1 mM EGTA, 1% Triton X-100, and protease inhibitors [pH 8.57]), cleared of insoluble material by centrifugation, then incubated for 30 min with 20 mM DMP (Sigma). An equal volume of 50 mM NH4Cl2 in PBS was added for 10 min to quench the reaction. Lysates were analyzed by SDS-PAGE and western blot. Sucrose density gradient centrifugation was performed as described [48]. Catalase (11.3 S, 256 kDa; Calbiochem) and bovine serum albumin (4.3 S, 66 kDa; Sigma) were used as standards.
Microtubule binding and FRET stoichiometry assays in live cells were carried out as described [2]. For microtubule binding in fixed cells, cells were treated with SLO and AMPPNP in the presence of taxol for 10 min and then fixed with 3.7% paraformaldehyde and processed for immunofluorescence. The Relocation Index was calculated on a frame-by-frame basis using ImageJ as described [2].
Motility and photobleaching assays were performed using a custom objective-type TIRF microscope as described [29]. Briefly, flow chambers were sequentially incubated with Cy5-microtubules, casein, and cell lysates in oxygen-scavenging buffer. Lysates were incubated with ATP or ADP for motility assays or with AMPPNP for photobleaching analysis. The 488-nm line of a tunable, single-mode, fiber-coupled argon ion laser (Melles Griot) was used for excitation. Images were captured every 100 ms. Two-color TIRF was achieved by combining a yellow diode, pumped, solid-state laser (593 nm; CrystaLaser) with the 488-nm laser using a dichroic mirror (Z488RDC). mCit and mCherry fluorescence emissions were first passed though a FF495/605 dual-band dichroic mirror (Semrock) and then projected separately onto each half of the charge-coupled device (CCD) camera by a Dualview beam-splitter (Optic Insights) equipped with a T585LP dichroic beam splitter and ET525/50M and HQ610LP emission filters (Chroma).
All of the fluorescence image analyses were analyzed as described [29] using ImageJ (NIH) and Origin. For single-molecule tracking, measurements for each construct come from at least two independent protein preparations and include motile events lasting at least five frames (500 ms) unless indicated otherwise. All data are presented as mean ± standard error (SE).
Additional methods are available in the supporting information (Text S1).
|
10.1371/journal.pntd.0007724 | Effect of insecticide-treated bed nets on visceral leishmaniasis incidence in Bangladesh. A retrospective cohort analysis | Visceral leishmaniasis (VL) is a parasitic disease, transmitted by the sand fly species Phlebotomus argentipes in the Indian sub-continent. Effective vector control is highly desirable to reduce vector density and human and vector contact in the endemic communities with the aim to curtail disease transmission. We evaluated the effect of long lasting insecticide treated bed nets (LLIN) and bed nets impregnated with slow-release insecticide tablet K-O TAB 1-2-3 (jointly insecticide-treated nets or ITN) on VL incidence in a highly endemic sub-district (upazila) in Bangladesh.
Several distributions of LLIN or K-O TAB 1-2-3 for self-impregnation of bed nets at home took place in Fulbaria upazila, Mymensigh district from 2004 to 2008 under three research projects, respectively funded by CDC, Atlanta, USA (2004) and WHO-TDR, Geneva, Switzerland (2006 & 2008). We included all households (n = 8142) in the 20 villages that had benefited in the past from one of these interventions (1295 donated LLIN and 11,918 local bed nets impregnated with K-O TAB 1-2-3) in the “exposed cohort”. We recruited a “non-exposed cohort” in villages with contemporaneously similar incidence rates who had not received such vector control interventions (7729 HHs from nine villages). In both cohorts, we visited all families house to house and ascertained any VL cases for the 3 year period before and after the intervention. We evaluated the incidence rate (IR) of VL in both cohorts as primary endpoint, applying the difference-in-differences method.
The study identified 1011 VL cases (IR 140.47/10,000 per year [py]) before the intervention, of which 534 and 477 cases in the intervention and control areas respectively. The IR was 144.13/10,000 py (534/37050) and 136.59/10,000 py (477/34923) in the intervention and control areas respectively, with no significant difference (p = 0.3901) before the intervention. After the intervention, a total of 555 cases (IR 77.11/10,000 py) were identified of which 178 (IR 48.04/10,000 py) in the intervention and 377 (107.95/10,000 py) in the control area. The intervention area had a significant lower IR than the control area during follow up, rate difference = –59.91, p<0.0001. The IR during follow up was significantly reduced by 96.09/10,000 py in the intervention area (p<0.0001) and 28.63/10,000 py in control area (p<0.0001) compared to baseline. There was a strong and significant overall effect of the ITN intervention, δ = –67.45, p <0.0001. Sex (OR = 1.36, p<0.0001) and age (OR = 0.99, p<0.0001) also had a significant effect on VL incidence. Male had a higher risk of VL than female and one year increase in age decreased the likelihood of VL by about 0.92%. Two third of the VL incidence occurred in the age range 2 to 30 years (median age of VL patients was 17 years).
VL incidence rate was significantly lower in the ITN intervention cohort compared to control in Bangladesh. Some bias due to more intense screen-and-treat activities or other interventions in the intervention area cannot be ruled out. Nonetheless, given their feasibility and sustainability, ITNs should be considered for integrated vector control during the maintenance phase of the VL elimination programme.
| Visceral leishmaniasis (VL) is a deadly parasitic disease, transmitted by the sand fly species Phlebotomus argentipes in the Indian sub-continent. Humans are the only proven reservoir of the parasite, Leishmania donovani. Effective vector control is highly desirable to reduce vector density and human and vector contact in the endemic communities to stop the disease transmission. We evaluated the effect of long lasting insecticide treated bed nets (LLIN) and bed nets impregnated with slow-release insecticide tablet K-O TAB 1-2-3 (jointly insecticide-treated nets or ITN) on VL incidence in a highly endemic sub-district (upazila) in Bangladesh. The nets were either donated or impregnated between 2004 to 2008 under three studies and defined as “exposed cohort” comparing their effect on VL incidence with “non-exposed cohort” (no donation of impregnated nets) for a 3 year period before and after the intervention. The study identified 1011 VL cases (IR 140.47/10,000 per year [py]) before the intervention, of which 534 and 477 cases in the intervention and control areas respectively. There was a strong and significant overall effect of the ITN intervention, δ = –67.45, p <0.0001. The VL incidence rate was significantly lower in the ITN intervention cohort compared to control in Bangladesh, though some bias cannot be totally ruled out.
| Visceral leishmaniasis (VL)—also known as kala-azar (KA) in the Indian sub-continent—is a deadly parasitic disease transmitted by the female Phlebotomus argentipes sand fly. In the South-East Asia Region, humans are the only proven reservoir of the parasite, Leishmania donovani. Kala-azar has been present in the Bengal territory (presently West Bengal, India, and Bangladesh) since the early 1800s [1] and gradually spread along the course of the Ganges and the Brahmaputra rivers, the major transport routes of Bengal. In what is today Bangladesh, KA was first described in 1824 in Jessore district [2], where an epidemic killed an estimated 75,000 people between 1824 and 1827 [1]. The historical records describe the classical picture of KA, with prolonged irregular fever, progressive emaciation despite good appetite, enlargement of liver and spleen and black coloration of skin [3].
In the late 1950s and early 1960s, WHO launched a malaria eradication programme throughout the South Asian sub-continent based primarily on Indoor Residual Spraying (IRS) of Dichlorodiphenyltrichloroethane (DDT). During this programme, KA almost disappeared as a collateral benefit [4]. However, within a few years after the end of the malaria eradication efforts, KA returned to Bihar and Bengal on both sides of the India-Bangladesh borders [5].
In Bangladesh, sporadic KA cases were reported again from the late 1960s onwards [6]. Between 1968 and 1980, 59 KA patients were reported, mostly from 5 districts (Sirajgang, Pabna, Mymensingh, Rajshahi, and Tangail) [7]. The numbers of KA cases soared from 1980 onwards, and a major outbreak occurred in Pabna district [1]. Between 1994 and 2013, the National Programme of Disease Control, Directorate General Health Services (DGHS), Government of Bangladesh reported 109,266 KA cases and 329 deaths [8]. Fifty percent of those cases were reported from just five sub-districts (Upazila) of Mymensingh district [8].
In 2005, three countries (Bangladesh, India, and Nepal), supported by WHO, launched a regional initiative to eliminate KA as a public health problem from the region and signed a Memorandum of Understanding (MoU) to this effect. The initiative aimed to reduce the KA incidence to one case per 10,000 population in each endemic sub-district by 2015 [9]. This deadline was later extended to 2017 [10], and 2020 in the London Declaration on Neglected Tropical Diseases [11]. Despite an impressive decrease in the number of cases in each country, WHO could not yet validate the KA elimination status in any of them and advocates for more intense and sustained control efforts and disease surveillance.
The intervention strategies in the elimination programme are based on case detection and treatment and integrated vector management (IVM) [9]. In Bangladesh, however, no specific sand fly control operations were carried out by the programme between 1999 and early 2012 [8,12]. It took a long time to register the required insecticides for the indication of sand fly control. The first indoor residual spraying (IRS) activity was conducted using deltamethrin 5WP in April/May 2012 in eight highly endemic Upazilas (sub-districts) [13]. Till today none of the countries was able to fully implement the IVM strategy in the region, as they tend to implement IRS for sand fly control only, and this in an independent way of any other vector control operation. Although well-performed IRS can reduce vector density dramatically, it remains operationally challenging and expensive, and its acceptance by the community is not always optimal. Several authors have highlighted its limitations, in terms of insecticide resistance, quality of implementation, occupational hazard, cost, sand fly adaptation, etc. [14–17]. Therefore, there is a need for alternative tools which are operationally easy to implement and cost-effective in terms of per household protected. The question of whether there are alternatives to IRS will only become more relevant in the post-elimination era. We briefly summarize here the evidence on P. argentipes control methods from Bangladesh and the region so far. Between 2002 and 2009 several epidemiological and entomological studies were conducted in the highly endemic area of Fulbaria, one of the Upazilas of Mymensingh district, either to assess the KA disease burden and its risk factors [18], or to evaluate the effectiveness of insecticide-treated nets as an alternative for IRS for controlling the P. argentipes sand fly [19–21]. Consistent use of non-treated local bed nets in summer was associated with reduced risk for VL in an observational study [18]. This study also showed that use of bed nets is acceptable in the rural community of Bangladesh and found a high percentage of households owning at least one bed net [18], similar to evaluations in India and Nepal [22].
Inspired by the effectiveness of insecticide treated bed nets for malaria control, several intervention studies evaluated either donated long-lasting insecticide-impregnated bed nets (LLIN) or local bed nets impregnated with slow release insecticide tablet K-O TAB 1-2-3, on entomological endpoints [19–21]. For ease of understanding, we regroup both interventions as “Insecticide-Treated Nets” (ITN) in the remainder of the text. The two multi-country intervention studies found significant reductions in sand fly density, ranging from 60% to 80%. A less pronounced 25% reduction of sand fly density was found in a cluster randomized trial (CRT) conducted in India and Nepal comparing households covered by LLIN with households where no LLINs were used, which were allowed to continue to use their own commercially available non-treated nets [23]. However, the CRT study in India and Nepal did not find any effect of the LLIN distribution on Leishmania donovani infection nor KA incidence, notwithstanding a high coverage of all household members and effective use of the LLINs [24]. Authors suggested this negative finding might be related to exposure outside the peridomestic environment due to changing sand fly behaviour, which was partly confirmed later [25]. Long-standing insecticide pressure because of the repeated IRS campaigns in India and Nepal might have forced sand flies to adapt again to the outdoor environment.
It is worthwhile to study the same question in Bangladesh though, as the sociocultural and environmental parameters are somewhat different. In Bangladesh, in contrast to India, no IRS was in place for a very long time in the KA endemic areas, so there was no insecticide pressure on the peridomestic sand fly habitat. Therefore, we set out to investigate the impact of ITNs on KA incidence in Bangladesh through a retrospective cohort analysis, as staging another CRT would raise ethical questions and would not be feasible in the present context of very low incidence rates near elimination.
The present study protocol was approved by the Ethical Review Committees of the Bangladesh Medical Research Council (BMRC) and the Special Program for Research and Training in Tropical Diseases/Regional Office for South-East Asia, World Health Organization (WHO/SEARO), India. Informed consent in the household survey was signed by the head of household before their voluntary participation in the study.
This study is a retrospective cohort analysis set in Fulbaria sub-district, Bangladesh. Fulbaria is located 111 km from the capital city Dhaka, and 23 kilometers away from the district headquarters respectively. Fulbaria has 13 unions (lowest administrative unit) and 116 villages. The Upazila occupies an area of 398.70 km2 including 14.76 km2 forest area.
In the Fulbaria population, we retrospectively defined two distinct cohorts; the exposed cohort was the one who benefited from an ITN intervention in the recent past, and the unexposed were those who did not. The first, “exposed cohort” was therefore composed of all the communities who had benefited previously from a LLIN or K-O TAB 1-2-3 distribution in 1 of three distinct studies (18–21). The non-exposed cohort (control) (i.e., families that did not receive any donated long-lasting nets or whose local nets were not impregnated), was composed of villages of similar population size with a comparable KA incidence rate in the corresponding study period of each of the three published studies (18–21).
Based on data from the epidemiological records of the Ministry of Health (passive surveillance data), we then ranked all the KA endemic villages of Fulbaria (excluding those already included in the exposed cohort), according to the number of reported cases, for the corresponding time period when the respective ITN interventions were carried out (2004, 2006, 2008). We then randomly selected nine endemic villages from the 15 top-ranked villages (Baddiyan bari, Balashawr, Palashtali, Deoli, Dhamar, Shibpur, Kathgara, Harirumbari, and Palashihata) (Fig 1), and included these nine communities (n = 7729) in the control cohort.
We describe here the different ITN interventions that took place in the exposed cohort. A first epidemiological study was conducted in a total of 506 households from three paras (sub-villages) namely Nadirpar, Lakxmipur, and Bamonbaid of Chouder village (Fig 1) of Fulbaria union, Fulbaria Upazila, Mymensingh district between 2002 and 2004 [18]. The study was funded by CDC, Atlanta, USA and implemented by icddr,b. After completion of the study, each HH was donated one unit of LLIN (manufactured by Vestergaard Frandsen Private Limited).
Between 2006 and 2007, a cluster randomized trial was conducted with four arms (3-intervention [arm-1: IRS using deltamethrin 5 WP, arm-2: LLIN and arm-3: environmental management] and 1-control arm where LLIN were donated after completion of the study period) in Fulbaria Upazila (Fig 1), Mymensingh district [19, 20]. This study involved a total of 596 households. The study was funded by the Special Programme for Research and Training in Tropical Diseases (TDR), WHO, Geneva, Switzerland and conducted by the National Institute of Preventive and Social Medicine (NIPSOM), Bangladesh. LLIN (manufactured by Vestergaard Frandsen Private Limited) were donated in two arms.
Last, a community-based study was conducted with 6967 households in Putijana union (Fig 1) of Fulbaria Upazila, Mymensingh district between 2007 and 2008 [21]. In this study, all existing bed nets at HH level were impregnated with slow release insecticide tablet K-O TAB 1-2-3 (0.4g deltamethrin in a 1.6 g tablet and a chemical binder) manufactured by Bayer Crop Science, Isando, South Africa. The study was supported by the Special Programme for Research and Training in Tropical Diseases (TDR), WHO, Geneva, Switzerland and conducted by National Institute of Preventive and Social Medicine (NIPSOM), Bangladesh.
We assessed the outcome “KA” independently from the original research projects, and in the same way for the intervention and control area, as follows. We exploited the full database for Fulbaria sub-district for the period of 2001 to 2011 to identify reported KA cases in the intervention and control areas and visited all affected communities in an exhaustive house to house survey. All households (HH) of both cohorts were visited between 2011 and 2012, and the head of the HH/responsible adult was interviewed in order to ascertain the number of reported KA cases in the period of three years before and after the intervention for the three distinct study sites described above, and in a matching time frame for the control cohort. The period of observation was seven years for each intervention, and the respective time windows were as follows: for the 2004 CDC funded study: 2001–2007; for the 2006 TDR study: 2003–2009; and for the 2008 TDR study: 2005–2011. Additionally, information about current bed net use and washing practices was also collected in the intervention area. Trained Research Assistants conducted the interview using a structured questionnaire.
A standard data entry interface was designed using Microsoft Office Access for entering study data. Data were checked and cleaned before analysis. Percentages were used to summarise the demographic and study variables. VL incidence rate was calculated for control and intervention areas (per 10,000 persons per year) for baseline and follow-up period separately. Z-test was used to compare the VL incidence rates between the intervention and control area, and p-values at the 0.05 significance level were used. Difference In Difference (DID) estimates (δ) were calculated to estimate the effect of the intervention at the community level. Binary logistic regression was used to calculate odds ratios for the effect of gender and age on VL incidence rate. STATA/MP 13.0 for Windows (StataCorp LP, College station, TX77845, USA) was used for data analysis.
Of a total of 15,871 HHs (71,973 population), 8142 HHs (37,050 population) and 7729 HHs (34,923 population) were included in the study as exposed and control cohort. Table 1 shows that their baseline characteristics are very comparable, including for the frequency of KA at baseline in household level.
In the household survey, we investigated the persistent use of bed nets in the intervention area. Of 8142 HHs that benefited at one point in the past from an ITN distribution in the intervention area, more than 92.2% HHs had at least one bed net in their house at the time of our household survey. Among those, 80.1% were ITNs, either self-impregnated with K-O TAB 1-2-3 or LLIN, the others were non-impregnated commercial nets (Table 2). About 33.9% HHs even had two nets, of which 88.9% were ITNs. However, 7.8% of all HHs did not have any bed net at the time of survey. More than 84.3% HHs (6864/8142) informed that they were always sleeping under a bed net. Sixty-five percent of all HHs reported that they felt impregnated nets were effective against mosquitoes along with other insects while about 32% HHs informed nets only effective against mosquitoes. About 82% HHs stated that there were less kala-azar (VL) cases in their community after the introduction of impregnated nets while about 17% respondents had no opinion. Only 12% of our respondents knew that kala-azar is transmitted by a sand fly bite, while the majority (74.4%) said it is transmitted by mosquitoes. Almost all HHs (98%) expressed a demand for ITNs, and the majority (72.7% respondents) asked for a free-of-cost distribution as a government donation (Table 2).
ITNs had been washed upto 5 times in 57.6%, 23.0% and 67.5% of HHs and 6–10 times in 29.6%, 54.1% and 32.4% in Putijana union; Chouder village; and Bhalukjan, Panch Kushmail, Neogi Kushmail, Baruka villages respectively (Table 3). Regarding washing practice of nets, 88.5%, 88.2% and 94.9% HHs in the Putijana union; Chouder village; and Bhalukjan, Panch Kushmail, Neogi Kushmail, Baruka villages reported that they washed their nets in the pond (Table 3) which is not recommended. In Putijana union, the majority of the respondents (98.8%) said that they dried their nets in direct sunlight (also not recommended), while this was 76.5% and 53.9% in Chouder village and Bhalukjan, Panch Kushmail, Neogi Kushmail, Baruka villages respectively.
In the house-to-house survey, we recorded a total of 1011 VL cases (140.47/10,000/year) in the three years preceding the respective research projects of which 534 (144.13/10,000/year) and 477 cases (136.59 per 10,000/year) in the intervention and control areas respectively (Table 4; Fig 2). The difference in incidence rate (IR) was not statistically significant (p = 0.3901). In the three years after the research projects, we identified a total of 555 KA cases (incidence rate 77.11/10,000/year) of which 178 (48.04/10,000 per year) in the intervention area and 377 (107.95/10,000 per year) in control area (Table 4; Fig 2). The area that benefited from ITN had a significantly lower incidence rate than the control area in the 3-years follow up period, the rate difference was –59.91, p<0.0001. The VL incidence rate during follow-up was significantly reduced both in the intervention and control areas, by 96.09/10,000/year in intervention area (p<0.0001) and 28.63/10,000/year population in control area (p<0.0001) compared to baseline. The effect of the intervention was strongly significant, δ = –67.45, p <0.0001. The estimated reduction of VL incidence rate by the intervention was 46.80% (p<0.0001).
Moreover, sex (OR = 1.36, p<0.0001) and age (OR = 0.99, p<0.0001) also had a significant effect on VL incidence. Male were more affected by VL than females. A one year increase in age decreased the likelihood of VL by about 0.92%. Seventy five percent of the VL incidence occurred in the age range of 2 to 30 years (median age of VL patients was 17 years).
The main finding of our analysis is that the introduction of ITN in rural highly endemic communities in Bangladesh was associated with a significantly greater reduction of the VL incidence, compared to unexposed communities that also experienced a reduction over time but of lesser size.
The present study confirmed that the use of bed nets is a common practice in the rural community of Bangladesh as observed by others [18]. We found that many households in the intervention cohort were still using the nets which had been distributed during the previous studies. A certain proportion of HHs (about 8%) were not having bed nets, and those were most likely the poorest families, are mostly related to poverty as it is well established that VL affects the poorest communities in the Indian sub-continent (ISC) [26–28]. It has been observed that high coverage of bed net use has community effect on vector sand fly in India and Nepal [23], similar impact found for malaria vector in Tanzania [29], so unavailability of bed net in the small number of HHs might not have negative impact. However, the study findings suggest that the washing practices of the ITNs require some change to preserve their effectiveness. Impregnated nets should not dry in the direct sun light as no UV protection is in place in the net. Large number of people dried their nets in the direct sun in the intervention areas which may have reduced their efficacy. It is hard to explain why HHs dried their nets in the direct sun though they were informed to dry nets in the shady place. The possible reason could be HHs want to make sure bed net get dried before sunset in the same day of wash as they may not have extra net. Moreover, it is also recommended that impregnated net should not be washed in ponds or rivers as deltamethrin (synthetic pyrethroid) is poisonous for aquatic animals especially for fishes [30]. Unfortunately very few people washed their nets using tube well water. It is well established that VL endemic communities are poorest of the poor, due to this reason many of the study families may not have own tube well which forces them to wash their bed nets in the pond as it is convenient.
The strength of our design is that we were able to control for a declining temporal trend by comparing the effect in the intervention area with that of a contemporaneous control area. We also acknowledge two important limitations of our study design. As it is non-experimental in nature, there could be other factors that explain the trend in IRs in the cohorts, such as e.g. a more intense screen-and-treat as the baseline IR were of the highest in the region, and communities might have been targeted preferentially by the programme. We believe the influence of such factor to be minor, as prior to 2009 the elimination programme in Bangladesh was not yet in full swing [12]. At the time, except for some training, little governmental control activities took place. VL patients were in theory entitled to receive all medication free of cost in the government health facilities, but in practice there were severe drug shortages of Sodium Stibogluconate [18] until the introduction of Miltefosine as first-line treatment option in 2009 [8]. It later appeared that one of the batches of Miltefosine supplied by the national programme was a fake drug with no active substance [31], so we may consider that the effect of case management was minimal during that period. Similarly, no sand fly control activities were conducted by the government up to early 2012 [8, 12] since banned of DDT in 1997 [32], as the registration process of deltamethrin for sand fly control took a long time [13]. However, in 2013 the national programme distributed two pieces of LLIN to each patient who had completed VL treatment between 2009 and 2011 [8]. Secondly, our comparison is a one-to-one comparison of one cohort compared to another, and given the erratic behaviour of VL in small areas, the lack of replicates limits the robustness of our findings. Randomization of a sufficient number of study units to either intervention or control cohorts would undoubtedly lead to less biased results, but in the given context of very low case incidence, the organization of such trial is deemed not feasible.
Unfortunately, very few studies evaluated the impact of ITN on VL incidence in the ISC. The only study evaluating the impact of local nets impregnated with slow release insecticide on VL in Bangladesh found a 66.5% incidence reduction after one year of use in a comparison of one intervention to one control area [33]. Our study showed a significant reduction of VL incidence after three year of use. In Sudan, another observational study found a 59% reduction of VL after using impregnated bed net [34] which is in line with our findings as well. However, our findings contrast with those of pair-matched cluster randomized trial of LLIN distribution in India and Nepal where no VL incidence reduction was found [24]. However, the same study showed a significant reduction on malaria incidence, and the LLIN reduced about 25% P. argentipes sand fly density at household level [23]. The difference between Bangladesh and India/Nepal could be that long-term DDT spraying in India and synthetic pyrethroid spraying in Nepal induced some adaptation of sand fly behaviour towards more outdoor resting or feeding behaviour which is partially supported by a study from India [25].
To eliminate or control a vector-borne disease it is highly important to reduce human-vector contact and vector density. Till today except for IRS no other interventions are included in the vector control strategy of the VL elimination initiative. In the MoU, it was noted that IVM should be adopted as regional strategy for vector control, and this requires more than one tool [9]. Operationally IRS is a more challenging and also more expensive method than ITN distribution. Studies in Bangladesh, India and Nepal identified that the impact of IRS is sub-optimal when it was carried out by the national programme [13, 16]. Furthermore, VL cases are sharply reducing in the countries so that it will not be sustainable to continue blanket IRS operations in all endemic sub-districts in the country. Health authorities in the region may no longer allocate enough funding for IRS because they have many other health priorities to respond to. It is worth to mention that Bangladesh and Nepal so far did not receive any external funding to control the VL vector in contrast to India (personal observation, RC).
In this regard, the present study provides observational evidence of the effect of ITNs in the absence of other governmental control interventions. Given the affordability of ITNs [15], their ease of implementation and their acceptability, they should be given consideration for inclusion in integrated vector management, definitely in the era of post-VL-elimination [35,36].
|
10.1371/journal.ppat.1002413 | Norovirus Regulation of the Innate Immune Response and Apoptosis Occurs via the Product of the Alternative Open Reading Frame 4 | Small RNA viruses have evolved many mechanisms to increase the capacity of their short genomes. Here we describe the identification and characterization of a novel open reading frame (ORF4) encoded by the murine norovirus (MNV) subgenomic RNA, in an alternative reading frame overlapping the VP1 coding region. ORF4 is translated during virus infection and the resultant protein localizes predominantly to the mitochondria. Using reverse genetics we demonstrated that expression of ORF4 is not required for virus replication in tissue culture but its loss results in a fitness cost since viruses lacking the ability to express ORF4 restore expression upon repeated passage in tissue culture. Functional analysis indicated that the protein produced from ORF4 antagonizes the innate immune response to infection by delaying the upregulation of a number of cellular genes activated by the innate pathway, including IFN-Beta. Apoptosis in the RAW264.7 macrophage cell line was also increased during virus infection in the absence of ORF4 expression. In vivo analysis of the WT and mutant virus lacking the ability to express ORF4 demonstrated an important role for ORF4 expression in infection and virulence. STAT1-/- mice infected with a virus lacking the ability to express ORF4 showed a delay in the onset of clinical signs when compared to mice infected with WT virus. Quantitative PCR and histopathological analysis of samples from these infected mice demonstrated that infection with a virus not expressing ORF4 results in a delayed infection in this system. In light of these findings we propose the name virulence factor 1, VF1 for this protein. The identification of VF1 represents the first characterization of an alternative open reading frame protein for the calicivirus family. The immune regulatory function of the MNV VF1 protein provide important perspectives for future research into norovirus biology and pathogenesis.
| This report describes the identification and characterization of a novel protein of unknown function encoded by a mouse virus genetically similar to human noroviruses. This gene is unique to the mouse virus and occupies the same part of the genome that codes for the major capsid protein. The protein that we have described as virulence factor 1 (VF1) is found in all murine norovirus isolates, absent in all human strains but is indeed expressed during infection. Its expression enables MNV-1 to establish efficient infection of its natural host through interference with interferon-mediated response pathways and apoptosis. Our data would indicate that the VF1 protein is multi-functional with an ability to modulate the host's response to infection. Murine noroviruses are frequently used firstly as a model to study human norovirus replication and pathogenesis, studies hampered by their inability to replicate in cell culture. Secondly, persistent infection of laboratory animals with murine norovirus may affect other models of disease using experimental mice. The role of VF1 in infection and pathology in the differential outcome of infection is the source of continued research in our laboratory.
| Collectively, the innate and adaptive immune systems result in a strong evolutionary pressure on pathogens to develop countermeasures to allow their continued existence. Therefore pathogens, including viruses, have evolved a multitude of mechanisms for evading the host response to infection, often by the expression of proteins that interfere with cellular antimicrobial response mechanisms [1]. The size of RNA virus genomes is thought to be limited by the error prone nature of the viral polymerase. As a likely direct consequence, RNA viruses have evolved a variety of mechanisms to increase the coding capacity of their genomes [2]. These include the use of ribosomal frameshifting where a proportion of translating ribosomes change the reading frame to produce proteins with common N-terminal but a different C-terminal from the read-through sequence [3]. Many viruses have also evolved to use a mechanism that creates overlapping reading frames through the use of two or more transcription initiation sites or translation start codons within the same RNA sequence [4], [5].
Murine norovirus (MNV) was identified in 2003 as a virus that caused a lethal infection in immunocompromised mice [6]. However, MNV is now known to be a widespread infectious agent of laboratory mice with a reported seroprevalence of 20-64% [7], [8]. MNV is currently the only norovirus which replicates efficiently in tissue culture, where it has a tropism for dendritic and macrophage cells [9]. The availability of immortalized macrophage cell lines such as the murine macrophage RAW264.7, has allowed significant advances to be made in understanding the life cycle of this virus. For the first time critical processes in the norovirus life cycle have been dissected e.g. the mechanism of tissue culture mediated attenuation of MNV-1 [10] , the requirement for dynamin II and cholesterol during virus entry [11], [12], the identification and functional requirement for RNA secondary structures in virus replication [13] and pathogenesis [14] as well as the induction of apoptosis during infection [15], [16]. In addition, MNV has allowed an unprecedented analysis of the immune response to norovirus infection [17],[18]–[21]. This broadening in understanding of norovirus replication has been facilitated greatly by the development of murine norovirus reverse genetics [22], [23] and its recent optimisation [24]. The role of murine norovirus in potential exacerbation or complication of other diseases, especially murine models of infection, has also been investigated. This is certainly warranted given the seroprevalence of MNV in animal houses. Studies with models of Crohn's disease [25] or bacterial induced inflammatory bowel disease [26] showed a significant impact of MNV and MNV infection prolongs the shedding of mouse parvovirus [27]. In contrast, MNV co-infection had little or no impact on murine CMV [28], Friend retrovirus infection [29] or models of diet induced obesity and insulin resistance [30]. These contrasts warrant further studies into the nature and mechanisms of interference observed in mouse models of disease. MNV has also provided a useful experimental system in determining the immune responses required for efficient norovirus vaccination [18],[31]. Collectively, this highlights both the relevance of MNV as a significant infectious agent in its own right and also the utility of MNV as a model for human norovirus. Continued research into what differentiates murine and human noroviruses and how norovirus infection affects the host cell is therefore of upmost importance to both fields of research.
Unlike other members of the Caliciviridae, which typically encode three open reading frames [6], our analysis and that presented during large scale sequencing of many MNV genomes [32] indicates the presence of a fourth potential ORF in the MNV genome (Figure 1A) In this study we demonstrate that the protein encoded by ORF4 is expressed during virus infection, is not essential for virus replication in tissue culture but plays a role in viral virulence and therefore represents a novel viral virulence factor. Based on the findings that it possesses anti-innate immune activity, contribute towards the regulation of virus induced apoptosis during infection and modulates the outcome of experimental infection of mice, we have described the ORF4 gene product as virulence factor 1 (VF1). The study provides important insights into the mechanisms of norovirus avoidance of the innate immune response and norovirus pathobiology.
The region of the genome encoding VF1 contains an intact reading frame in all available MNV sequences derived from different isolates or strains (Figure 1A, 1B and data not shown). In contrast to the typical 8–15% sequence divergence seen between MNV variants in the amino acid sequences of ORF1, ORF3 and the predicted single coding region of ORF2, variability is markedly suppressed in the predicted ORF4 and double coding region of ORF2 (3%; Table 1; Figure 1C). Almost all sequence variability between MNV variants in the single coding regions (ORF1 and ORF3) occurs at synonymous sites. dN/dS ratios, namely the substitution rates at non-synonymous and synonymous sites, ranging between 0.03–0.10 are indicative of strong negative selection. In the double coding region of ORF2 (i.e. the region which codes for both VP1 and VF1), the restricted variability that is observed occurs at synonymous sites in the ORF2 reading frame (dN/dS: 0.044) consistent with stronger sequence constraints in the conventional reading frame encoding the MNV structural protein than in the ORF4 gene (dN/dS ≈ 2). However, the elevated ratio relative to that of ORF2 arose through greater suppression of synonymous variability in this reading frame, rather than increased amino acid sequence variability. dN values were 0.04 and 0.03 in ORF2 and ORF4 respectively.
As well as suppressing variability, the existence of a second reading frame in ORF2 leads to altered codon usage by the ORF4/VF1 coding sequence. For example, there was a significant overrepresentation of the UUG triplet coding for Leu in ORF4 (15 from 33, compared to 17 from 128 in ORF1; p<0.001 in a 6×2 contingency table for the 6 synonymous Leu codons), whereas there were no differences in Leu codon usage between ORFs 1, 2 (single coding region) and ORF3. The program MLOGD identifies overlapping coding sequences by specific codon usage signatures arising from mutational constraints consequent to the requirement to maintain protein function in two putative genes [33]. The relative likelihood that a given sequence region is single-coding or double-coding was calculated using a codon usage table and nucleotide mutation and amino acid substitution matrices (Figure 1C). This analysis provides independent support for the existence of ORF4/VF1, independent of its effect on sequence variability and evolutionary conservation. The first methionine codon in ORF4 at 5069 lies two residues downstream from a stop codon in that reading frame, and is 2 and 4 residues away from Met codons in ORF1 (including the -1 frameshift). The ORF4 start codon is in a strong Kozak context (G at +4 and -3) and likely represents the translation start site of VF1.
The ability of the MNV-1 subgenomic RNA (sgRNA) to produce a protein from the open reading frame predicted to encode VF1 was examined by in vitro translation of a plasmid containing the entire MNV-1 sgRNA under control of a T7 RNA polymerase promoter (Figure 2A). A coupled transcription and translation reaction of the MNV-1 sgRNA produced three proteins and the identity of the major (VP1) and minor (VP2) capsid proteins were confirmed using immunoprecipitation (Figure 2A). Polyclonal antisera to a peptide from MNV-1 VF1 was generated in rabbits and used to confirm the identity of the VF1 protein product by immunoprecipitation (Figure 2A). Full length his-tagged VF1, purified from E.coli was poorly immunogenic, hence a modified immunization protocol that used a variety of forms of VF1 (described in Materials and Methods), followed by affinity purification was required in order to obtain reactive antisera. Immune sera from MNV-1 infected mice did not contain antibodies to VF1 as determined by western blot using recombinant his tagged VF1 (data not shown).
To examine the expression of VF1 during MNV-1 replication in tissue culture, the well established RAW264.7 cell culture system for MNV [9] was used and the production of VF1 analyzed by western blot (Figure 2B). Using a high multiplicity of infection (MOI of 5 TCID50/cell) infection, VF1 was readily detected as early as 9 hours post infection, appearing at the same time as the minor capsid protein VP2 (Figure 2B). In contrast, the viral RNA polymerase NS7 was detected as early as 6 hours post infection (Figure 2B). Whilst we were unable to detect VF1 and VP2 prior to 9 hours, this may simply be a reflection of the sensitivity of the antisera used in the assay, but may also reflect the kinetics of viral sgRNA synthesis, as this is likely to occur after the initial rounds of viral genomic RNA synthesis. VF1 and VP2 expression levels observed over the course of the infection were also significantly different, with VF1 being expressed to a higher degree than VP2. Whilst this may be a reflection of the differences in the ability of the antisera to detect both proteins, it is known that VP2 synthesis requires translation re-initiation at the end of VP1 [34] which is likely to produce reduced levels of VP2 relative to the other proteins expressed from the viral sgRNA.
To determine if VF1 was required for MNV-1 replication in tissue culture we used a recently developed reverse genetics system [22] to truncate the VF1 coding region at various positions. Three mutants were created containing single nucleotide changes that lead to the introduction of a stop codon in the VF1 coding region but which did not alter the VP1 coding sequence (Figure 3A): M1 containing the mutation T5118A truncating the VF1 protein at amino acid 16; M10 containing the mutation T5364A, truncating VF1 at amino acid 98; M20 containing the mutation G5655A, truncating VF1 at amino acid 195. All mutations were introduced at positions where it was possible to change the VF1 coding sequence without affecting the major capsid protein VP1. Single nucleotide substitutions were used due to the nature of the overlapping coding regions. The interruption of VF1/ORF4 was confirmed by in vitro coupled transcription and translation of a PCR product encompassing the sgRNA of each mutant compared to wild-type MNV-1 (Figure 3B). VF1 was readily detected after in vitro translation of the wild type sgRNA product as well as the sgRNA from the M20 mutant that encodes a C-terminally truncated form of VF1. VF1 was not detected after in vitro translation of the sgRNA from either the M1 or M10 VF1 truncations as expected (Figure 3B). Recovery of wild-type and VF1 mutant viruses was performed using fowlpox mediated expression of T7 RNA polymerase to drive the synthesis of MNV-1 RNA in cells transfected with full length cDNA constructs of MNV-1 as described [22]. As we have previously reported, the BHK cell line used during virus recovery, although permissive to virus replication, cannot be infected with MNV due to the lack of a suitable receptor [22], therefore the yield of virus from this system represents a single round of virus replication only. The initial yields of VF1 knockout or truncation viruses were comparable to that derived from wild-type cDNA (∼1–5×104 TCID50 per 35mm dish, data not shown), indicative that VF1 was not required for virus replication in tissue culture. Western blot analysis of cells infected with the sequence verified M1, M10, M20 viruses confirmed that VF1 was not expressed in cells infected with either M1 or M10, but low levels of VF1 were observed in M20 infected cells (Figure 3C). The levels of VP2 produced by the VF1 knockout viruses were comparable to the wild-type MNV-1 derived from cDNA, confirming comparative levels of infection (Figure 3C). It is possible that the truncation of VF1 in the mutant M20 results in some protein misfolding, decreasing the half-life of the resulting truncated protein. The growth kinetics of low passage, sequence verified M1, M10 and M20 viruses was examined by both single-step (data not shown) and multi-step growth curve analysis and were indistinguishable from that of the wild-type parental MNV-1 derived from cDNA (Figure 3D), indicating that VF1 is not required for MNV-1 replication in tissue culture.
The observation that all MNV isolates identified to date retain ORF4/VF1 and that repeated passage of wild-type virus in tissue culture does not result in the loss of VF1 (data not shown), indicates that although VF1 is not essential for virus replication in tissue culture, it confers some benefit to virus replication. To address this, we examined the stability of the mutations in the M1, M10 and M20 viruses following repeated low multiplicity of infection (0.01 TCID50 per cell), multi-cycle replication in tissue culture. We observed that the mutations M1 and M10 were under negative selection in tissue culture whereas M20 was stable (Figure 3E). Sequence analysis of the virus population after 5 low multiplicity, multicycle passages in tissue culture, subsequent to the initial amplification after reverse genetics recovery, demonstrated that the M1 virus, which at passage 1 contained the mutation T5118A introducing a stop codon at position 17 in VF1, had introduced the mutation A5118G by the 5th additional passage, restoring full-length VF1 production by the insertion of a tryptophan residue. Analysis of the M10 virus population, which had the mutation T5364A at the first passage, also indicated that the population was heterogeneous and that in a proportion the VF1 open reading frame was restored by the introduction of the mutation A5364G. As with the M1 virus, this mutation is predicted to result in the introduction of a tryptophan at position 99. In contrast however, sequence analysis of the M20 virus after repeated multicycle passage in tissue culture demonstrated that the introduced mutation (G5655A) was in fact stable (Figure 3E), which may indicate that the major functional domain lay within the 195 amino acids. Western blot analysis of cells infected with ‘passage 5’ stocks of M1 and M10 viruses indicated that, as expected from sequence analysis, VF1 expression was detectable (Figure 3F), although the levels were notably lower than observed in WT infected cells. This reduced level may be in part due to the effect of the amino acid change on VF1 protein stability, but clearly for the M10 virus population the heterogeneous nature of the M10 virus stock (Figure 3E) is likely a contributing factor. M20 virus stocks maintained the ability to express low levels VF1 as previously seen using the initial virus stocks (Figure 3C and 3F). In all cases, the level of virus replication was similar as determined by the expression of the minor capsid protein VP2 (Figure 3F) and virus titre (data not shown).
To gain further insights in the potential function of VF1, the localization of VF1 was examined by confocal microscopy. Due to the high degree of cross-reactivity of the VF1 antisera with endogenous host cell proteins (Figure 2B), fusions of MNV-1 VF1 to EGFP were used to examine VF1 localization in cells. Transfection of COS7 cells with cDNA constructs expressing either N or C-terminal fusions of MNV-1 VF1 with EGFP demonstrated a pattern of EGFP expression characteristic of mitochondrial localization (Figure 4A). This was confirmed via co-staining of cells with the mitochondrial vital stain Mitotracker (Invitrogen) (Figure 4A). Similar co-localization of VF1-GFP and mitochondria was observed in BHK and 293 cells (data not shown). The expression levels observed in cells transfected with the VF1-GFP fusion proteins were substantially lower than those observed in infected cells as expression was not detectable by western blot analysis with either α-VF1 or α-GFP antisera (data not shown).
To confirm the mitochondrial localization of VF1 during virus infection, mitochondria were purified from infected RAW264.7 cells at 15 hours post infection and analyzed for the presence of VF1 by western blot (Figure 4B). Whereas the well characterized host cell nucleic acid binding proteins PCBP1/2 were shown to be predominantly cytoplasmic as expected [35], VF1 was only detected in the mitochondrial fraction (Figure 4B). Apoptosis inducing factor 1, a predominantly mitochondrial protein was enriched in the mitochondrial fraction, confirming the validity of the purification procedure (Figure 4B).
RNA viruses frequently encode proteins that antagonize the innate immune response to infection. Mitochondria play a significant role in signaling innate immune responses through the well characterized mitochondrial antiviral signaling protein (MAVS), an integral membrane protein found in the outer mitochondrial membrane [36]–[39]. MAVS is a key adapter protein in the sensing of viral RNA by RIG-I and MDA5 that, in part, leads to IRF3 and NFΚB activation and the upregulation of antiviral genes such as IFN-Beta, CXCL10 and ISG54 [37]. Given the mitochondrial localization of VF1 in infected cells we assessed the activation of this sub-section of the innate immune response in both M1 and WT infected cells. RAW264.7 cells infected at a low MOI (0.1 TCID50 per cell) with the M1 VF1 knockout virus exhibited a greater induction of antiviral genes such as ISG54, CXCL10 and IFN-Beta in response to viral infection than those infected with the WT virus (Figure 5A and Figure 5B). Alterations to the levels of mRNA were calculated relative to uninfected cells using the standard ΔΔCT method with hypoxanthine phosphoribosyltransferase 1 (HPRT) (Figure 5A and Figure 5B) or actin (data not shown) mRNA levels used as endogenous controls. CXCL10, ISG54 and IFN-Beta mRNA levels were then normalized to the amount of viral RNA present in each sample in order to calculate the rate of induction of the innate immune response over time. This method of data normalization was also used to overcome variations often observed in the rate of virus replication seen in a variety of RAW264.7 cell clones (not shown). Normalizing the mRNA fold change to a constant amount of MNV RNA established that in all cases examined (CXCL10, ISG54 and IFN-Beta) the M1 infection causes a much more rapid induction of the innate immune response. For instance CXCL10 in M1 infected cells is induced 15.5 fold more quickly in response to the same amount of viral RNA than in WT infected cells. This value is calculated by comparing the slope/gradient for M1 and WT (Figure 5A) which represents the rate of induction of each gene. Significantly, the IFN-Beta and ISG54 mRNAs are also activated more quickly, 4.6 and 8.3 fold respectively, in M1 infected cells compared to WT equivalents. Of note, the total fold increase in CXCL10, ISG54 and IFN-Beta mRNA induced in M1 infected cells was significantly higher than that observed in WT cells at both 20 and 24 hours post infection (Figure 5A). These time points, and the eight hour window from 16 to 24 hpi, reflect the period of amplification for innate immune related gene activation following low MOI MNV-1 infection of RAW264.7 cells, since quantification of mRNA fold change at 16 hpi showed little or no increase (Figure 5A and 5B). The activation of the IFN-Beta mRNA in infected cells was also shown to correlate to increased protein production and secretion using ELISA (Figure 5B). The amount of IFN-Beta protein in the supernatants of infected RAW264.7 cells was significantly higher in M1 than WT infected cells at 24 hours post infection (Figure 5B). Protein production was again normalized to a constant level of viral RNA to demonstrate the relative response to WT and M1 replication.
Treating cells with poly (I:C), and analysis of gene expression, confirmed the sensitivity of RAW264.7 cells to dsRNA over an equivalent time course. Induction of ISG54, CXCL10 and IFN-Beta was demonstrated in poly (I:C) treated cells confirming their suitability for the investigation of innate immune responses to RNA stimuli (Figure S1). In addition, UV inactivated M1 and WT virus showed no significant induction of ISG54, CXCL10 and IFN-Beta when used in equivalent experiments and compared to mock infected cells (Figure S1). The IFN-Beta protein secretion in response to poly (I:C) and UV inactivated viruses was equivalent to that seen for the mRNA (Figure S1).
This ability of VF1 to antagonize the innate immune response was confirmed independently of infection using an IFN-Beta promoter driven luciferase assay. Murine embryonic fibroblast (MEF) cells were co-transfected with plasmid DNA expressing firefly luciferase under the control of an IFN-Beta promoter as well as expression constructs for RIGI, MDA5, MAVS or TBK1 whose ectopic over-expression has been shown to drive IFN-Beta production [40]. In addition these cells were co-transfected with either the empty vector or a plasmid expressing the MNV-1 VF1 protein. Over-expression of RIG1, MDA5, MAVS and TBK1 in cells transfected with the IFN-Beta promoter driven reporter resulted in an expected increase in luciferase production in all cases. However, this induction was significantly reduced in all instances where VF1 was co-transfected in comparison to the empty vector (Figure 5C). This indicates that VF1 in some way antagonizes the induction of IFN-Beta, correlating with the results observed in infected RAW264.7 cells.
MNV-1 infection is known to result in the induction of apoptosis via the down regulation of survivin, activation of caspases [41], as well as induction of cathepsin B activity [16]. Given our observation that VF1 localized with mitochondria and the key role mitochondria play in regulating apoptosis, we also examined how the absence of VF1 during infection affected the activity of the executioner caspases 3 and 7. Proteolytic activation of caspase 3 and 7, both of which play critical roles in the induction of apoptosis via the intrinsic cellular pathway, was also investigated (Figure 6A and 6C). As previously described [16], [41], a rapid increase in caspase 3/7 activity was observed in cells infected with wild type MNV-1 from 12 hours onwards (Figure 6A). Cells infected with MNV-1 lacking VF1 (M1), displayed significantly higher caspase 3/7 activity at 15 and 18 hours than cells infected with wild type MNV-1 (Figure 6A). Cells infected with the M1 virus lacking VF1 also displayed increased levels of the cleaved caspase 3 at 16 hours post infection (∼50% more than WT when quantified by densitometry) as determined by western blot analysis (Figure 6C). There was a notable alteration to the kinetics of viral protein production in the later stages of virus infection; whereas VP1 and NS7 levels continue to increase from 15 hours onwards in cells infected with WT MNV-1, the levels observed in cells infected with the M1 virus remained largely constant (Figure 6B). Importantly, the levels of infectious virus produced during infection were identical (Figure 3D). Induction of caspase 3/7 activities was shown to be due to virus replication as prior virus inactivation by UV treatment prevented virus induced caspase activity and the appearance of cleaved caspase 3 (Figure 6C).
The observed restoration of VF1 expression in the M1 and M10 viruses after repeated low multiplicity, multi-cycle replication in cell culture indicates that VF1 expression, although not essential for virus replication, confers some benefit to MNV-1 replication in cell culture. To examine if VF1 contributed to virus replication in vivo we examined replication in immunocompetent C57BL/6 mice. Whilst the isolate of MNV-1 used in this study and the only strain for which a reverse genetics system has been developed (CW1), is attenuated in the presence of a competent innate immune response [6], [42], low level virus replication in some tissues can be observed. Mice were inoculated with a high dose of low passage, sequenced verified WT and VF1 knockout viruses (M1) and the effect on body weight examined. As previously reported, no significant effect of MNV infection on weight loss was observed in this genetic background (Figure 7A). We also failed to detect robust levels of viral RNA in the small intestine, feces and spleen over the course of the experiment (data not shown). In contrast however, modest but significant levels of viral RNA were readily detected in the mesenteric lymph nodes of animals infected with WT MNV-1 at days 5 and 7 post infection (Figure 7B). In contrast, viral RNA was not detected in animals inoculated with the VF1 knockout virus M1. These data suggest that VF1 contributes to virus replication in vivo although virulence per se was not evident in this model even for the WT virus.
To examine if VF1 contributes to MNV virulence, the robust STAT1 -/- mouse model for MNV was utilized. However, in order to undertake these studies it was first necessary to generate VF1 mutant viruses in a cDNA backbone virulent in STAT1-/- mice. We have previously demonstrated that the single nucleotide mutation A5941G, changing glutamate 296 to lysine in the major capsid protein VP1, was sufficient to restore virulence to the tissue culture adapted strain of MNV-1, which is attenuated in STAT1-/- mice [10]. The VF1 mutation M1 was generated in an MNV-1 cDNA clone bearing two mutations (G2151A and A5941G) as this sequence more faithfully represents the consensus sequence in viruses isolated from infected STAT1-/- mice [6,referred to as CW1.P1 in 10]. Initial analysis of the levels of virus obtained after reverse genetics recovery of the VF1 mutant virus M1 in the virulent backbone (referred to herein as M1-v), demonstrated identical levels to the wild-type virulent virus (WT-v) of approximately 1-5×103 TCID50/ml (data not shown). This suggests that, as observed in the attenuated background, VF1 expression is not essential for MNV-1 replication in the STAT1-/- virulent backbone. To further verify this, multi-cycle growth kinetics analysis of low passage, sequence verified, WT-v and M1-v viruses in RAW264.7 cells was performed confirming equivalent growth kinetics (data not shown).
The ability of WT-v and M1-v viruses to infect and cause disease in the STAT1-/- mouse model was then examined by oral infection of age and sex matched mice. Oral inoculation of STAT1-/- mice with 1000 TCID50 units of low passage, sequence verified, wild-type virulent MNV-1 derived from cDNA (WT-v) resulted in the appearance of clinical signs (sunken eyes, reduced appetite, hunched inactivity and piloerection) as early as three days post inoculation. This was followed by a rapid and statistically significant (P<0.001) weight loss, when compared to animals inoculated with mock RAW264.7 cell lysate (day 4 onwards), and the development of more severe clinical signs culminating in significant weight loss (Figure 8). All WT-v infected mice succumbed to infection or were euthanized (because of disease severity limits being surpassed) by day 7 post infection. In stark contrast mice inoculated with the VF1 mutant virus (M1-v) showed a delayed onset of clinical signs. A statistically significant weight loss, compared to the mock-inoculated control group, was not observed until 6 days post infection (P<0.05, Figure 8). Of note, although the onset of M1-v associated disease was significantly delayed, all M1-v infected animals eventually succumbed to the infection or surpassed the severity limits of our trial. Experiments performed using a 10 fold higher dose (10,000 TCID50) also demonstrated that M1-v inoculated mice displayed a delayed onset of clinical signs including a statistically significant variation in body weight loss (two-way ANOVA). However this variation was markedly less than that observed at the lower infectious dose of 1000 TCID50 (Figure S2).
To examine if the distribution of virus replication differed between animals inoculated with WT-v or M1-v viruses, viral genome copies were quantified in various tissues at 3 and 5 days post infection by quantitative real-time reverse transcription PCR (qRT-PCR, Figure 9). Whilst >106 genome equivalents (gEq) per µg of total RNA could be readily detected in samples from mice infected with WT-v 3 days post infection, viral genome levels in M1-v infected animals were typically 104–105 fold lower (Figure 9A). For example, average levels in the spleen for WT-v infected animals were 1.7×109 gEq/µg of total RNA whereas M1-v inoculated animals showed an average of 2.6×104 qEq/µg of total RNA. Increased viral RNA replication was detected in all WT-v infected mice at day 5 post infection but only in a subset of the M1-v infected mice (Figure 9B). This subset correlated with those animals that had developed more significant clinical signs and had lost body weight at day 5 post infection.
Tissue samples from the spleen, small intestine and liver of mock, WT-v and M1-v infected mice were harvested at day 5 post infection for histopathological analysis on hematoxylin and eosin stained sections (Figure 10). Tissues from mock-infected mice were relatively normal for STAT1-/- mice (Figure 10). In contrast the WT-v infected tissues demonstrated reduced cellularity in spleen and liver, with foci of marked necrosis and apoptosis. Necrosis was evidenced by eosinophilia (dead cells staining bright pink) with pyknosis (nuclear condensation) and karryorrhexis (pyknotic nuclei become fragmented into several particles). Apoptosis was evidenced by cell rounding, a shrunken nucleus and in some cases cell fragmentation with some of the fragments containing apoptotic bodies. Blunting of the intestinal villi, as determined by measuring the villous height on the digital images taken at the same magnification and a comparison carried out on the mean of 8 villi, was apparent only in sections of the small intestine from animals infected with WT-v (Figure 10). Although the spleen from animals infected with M1-v appeared activated with partial paracortical hyperplasia, it was otherwise normal with little evidence of the necrosis or apoptosis evident in WT-v tissues (Figure 10). The liver revealed a partial loss of cellularity; however, evidence of apoptosis was again absent (Figure 10). In conclusion the lack of substantial pathology in M1-v infected mice at 5 days post infection correlated with our previous observations for differential viral RNA replication and weight loss in WT-v and M1-v infected mice.
The delayed virulence of VF1 knockout viruses in STAT1-/- mice was suggestive of reversion during replication in vivo. To examine this possibility further, sequence analysis of the viral population from all tissues in two animals displaying the highest degree of clinical signs (and viral genomes determined by qRT-PCR) was undertaken. The animals analyzed are highlighted in Figure 8B with a hash and asterisk indicating animals culled on days 5 and 7 respective. The animal culled on day 5 displayed similar disease onset to that of WT-v inoculated animals and had lost ∼12% of its initial body weight. The animal culled on day 7 for sequence analysis was removed from the study due to human end points being exceeded and had lost ∼7.6% of its initial body weight. Consensus sequence analysis, which under the conditions used could reproducibly detect reversion when ∼25% of the population had restored VF1, failed to detect any reversion in any of the tissues (Figure S3). In addition, tissues/samples containing the highest viral loads (4×106 to 4×107 copies per µg of RNA), namely the spleen and feces from the day 5 animal and the spleen from the day 7 animal, were PCR amplified and 10 individual clones sequenced. Of the 30 clones sequenced, none contained a mutation in VF1 that would lead to restoration of VF1 expression, further confirming the lack of detectable reversion upon replication a single pass in Stat1-/- mice in vivo (data not shown).
Studies on numerous RNA viruses have identified the use of overlapping open reading frames to maximize the coding capacity of their small RNA genomes [2]. These ORFs and their proteins typically play accessory roles in the viral life cycle such as modulating the host immune response to infection [43]-[45]. Frequently they dispensable for viral replication in immortalized cell lines; however, it is the in vivo setting that the true requirement for these proteins in the viral life cycle is apparent. This study indicates that MNV should now be added to the list of viruses that have adopted this strategy to maximize the coding potential of their genome.
Initial bioinformatic investigation of MNV complete genome sequences identified a conserved ORF overlapping with ORF2 (Figure 1), potentially translated from the sgRNA produced during infection. Traditionally the sgRNA is thought to encode only the major and minor capsid proteins, VP1 and VP2. However suppression of variability in this region and conservation of the alternate ORF was shown to be absolute in all available MNV sequences (Table 1). Although the resultant full length protein, VF1, was recalcitrant to high level expression and poorly immunogenic, polyclonal antibody specific to this protein was generated and used to confirm expression during infection (Figure 2). The efficient translation of this protein was confirmed by immmunoprecipitation following translation of the sgRNA in vitro (Figure 2).
This is the first confirmation of the expression of an internal open reading frame protein for any member of the Caliciviridae. The internal open reading frame encoding VF1 can be found in all currently published MNV genome sequences, highlighting the requirement for this feature in the MNV genome. The evolutionary conservation of ORF4 coding sequences and the marked suppression of sequence variability localising specifically to the area of overlap (Figure 1C) provides evidence independent of the in vitro data for a functional requirement to maintain an intact ORF4 reading frame. As indicated by the analysis of leucine and other synonymous codon usage, this selection pressure was sufficiently strong to drive unfavoured codons into the ORF4 coding sequence (Figure 1C), a feature exploited by MLOGD [33] to detect regions of multiple coding.
There were considerable similarities in the arrangement and translation contexts of the ORF2 and ORF4 genes of MNV with documented regions of multiple coding in other viruses. The MNV ORF4 has an initiating AUG triplet at position 5069 in a strong Kozak context (G at -3 and +4 [46]. It is positioned 13 bases downstream from the first AUG triplet of ORF2 (weak context; U at -3, A at +4) and 7 from the second (adequate context; A at -3 and +4). This arrangement of initiating codons in the MNV sgRNA transcripts is similar to viral [47] and eukaryotic [48] dicistronic mRNAs in which alternative weak context initiating codons around an initiating codon in a strong context (ORF4 in MNV) can be accessed by random forwards and backwards movements of the ribosome from its initial binding site, termed "leaky scanning". In this case, this would require a backwards movement to the second AUG triplet of ORF2, remarkably similar to the documented dicistronic expression of p206 (strong context) and p69 (weak context 7 bases upstream) from genomic RNA of turnip yellow mosaic virus [47]. This hypothesis is supported by the observation that noroviruses that lack ORF4 (genogroups 1–4) show a strong Kozak context around the second AUG triplet in ORF2 (A at -3, G at +4). The evolutionarily conserved nucleotide difference at position 5065 (+4) between MNV (A) and other noroviruses (G) may thus play a key enabling role in the hypothesised dicistronic expression of ORF4 and ORF2 by MNV.
The juxtapositioning of ORF4 at the start of the sgRNA gives an indication of the additional evolutionary constraints that this ORF, and the respective protein, must be under. This region of the genome contains multiple conserved cis-acting RNA elements that play an important role in the viral life cycle (unpublished observations). It is important to note however that the single nucleotide mutation introduced in this region to generate the M1 virus, did not affect the structure of these RNA elements as we have determined biochemically that nucleotide 5118 is positioned within a single-stranded region (data not shown). ORF4 also overlaps with the region of ORF2 that encodes the shell (S) domain of the major capsid protein, VP1. Dimerization of the S domain is thought to be integral for the development of the icosahedral core of the virus particle and is consequently the most conserved domain in VP1. As the S domain is buried inside the virus particle, it is unlikely to be under strong antibody selection pressure, unlike the more variable C terminus of VP1 which contains the protruding (P) domain. The contribution of all these factors is likely responsible for the low divergence observed between MNV VF1 sequences (Table 1).
Within the norovirus genus, ORF4/VF1 appears to be a feature unique to MNV as other noroviruses appear not to encode an equivalent open reading frame. The human noroviruses, which represent a significant cause of viral gastroenteritis in man, do not share the extensive suppression of synonymous site variability at the start of ORF2 that first indicated the presence of ORF4 (Figure 1C) [13]. Direct analysis of the available human norovirus sequences confirms that no such ORF exists (data not shown). A broader analysis of the Caliciviridae family indicates the presence of an equivalent open reading frame in some strains of human sapoviruses (data not shown) [49]. Although there is low sequence homology between the respective proteins (25% similarity, 18% identity) the presence of this alternative ORF indicates a potential conserved mechanism for maximizing coding potential. It is also possible that a common ancestor of all caliciviruses possessed an equivalent ORF, which has subsequently been lost in the case of the majority of caliciviruses. Although human noroviruses, as well as other members of the Caliciviridae, lack an equivalent ORF4 within the VP1 coding region of the sgRNA, we cannot at this point rule out functional duplication i.e. that the functions of MNV VF1 have been duplicated in human noroviruses by one of the other viral proteins. Further studies are therefore warranted to determine if human noroviruses and other members of the Caliciviridae also possess the ability to modulate the innate immune response.
The role of the VF1 protein in MNV-1 replication was examined using the permissive macrophage RAW264.7 cell line and the reverse genetics system developed previously in our laboratory [22]. A series of VF1 truncations, generated by inserting stop codons into ORF4, which importantly left the VP1 coding sequence unaltered, confirmed this protein was a classical viral accessory protein not required for replication. However, repetitive multicycle, low multiplicity of infection passage in the permissive RAW264.7 cell line resulted in a phenotypic reversion for the more severe truncations (M1 and M10) (Figure 3), demonstrating that VF1 expression and function benefits virus replication in cell culture. Whilst this observation was reproducible, it was clear that rapid phenotypic reversion did not occur as virus stocks generated by reverse genetics and subsequently amplified in cell culture by a single passage maintained the introduced mutations. The phenotypic reversion observed was the likely result of the multicycle nature of the infections as very low multiplicity of infections were used, resulting in multiple rounds of virus replication to occur in each pass in cell culture.
As the ‘powerhouses’ of the eukaryotic cell, viruses often modulate the function of mitochondria to maintain an intracellular environment beneficial for viral replication. The mitochondrial localization of VF1 (Figure 4) together with its apparent modulation of innate immune signaling and apoptosis (Figure 5 and 6) indicates that this protein may function, like so many other viral proteins, to facilitate viral replication and antagonize anti-viral mechanisms adopted by the cell. The beneficial functions of VF1 expression, e.g. delayed apoptosis and innate immune responses in infected cells, are apparent since analysis of VP1 protein levels produced during infection are clearly reduced at the later stages of infection in the absence of VF1 (Figure 6B). Surprisingly this does not appear to affect the final yield of virus (Figure 3D) or the levels of viral RNA produced during replication in RAW264.7 cells (data not shown). The RAW264.7 murine macrophage cell line is extremely permissive to infection and it is likely that the total pool of available VP1 protein in the M1 infected cells later in infection does not limit virion production. Of relevance is our observation that the IRF3 modulated gene ISG54, also known as IFIT2 or p54, is significantly upregulated in cells infected with a virus lacking VF1 (Figure 5A). The ISG54 protein (p54 or IFIT2) functions to repress cellular translation by binding and inhibiting the cellular eIF3c protein [50]. Previous work has indicated that norovirus translation initiation may require the eIF3 complex via a direct interaction of VPg [51]. The expression of the MNV VF1 protein may therefore delay or block the ISG54 mediated inhibition of cellular and viral VPg-dependent translation by preventing the induction of ISG54 mRNA at the point of mitochondrial mediated activation of the innate immune response [51]–[53]. ISG54 expression has also been linked to apoptosis induced via the mitochondrial pathway [54], again in agreement with our observed increase in apoptosis in the absence of VF1 expression (Figure 6). It is possible therefore that the increased apoptosis observed during virus replication in the absence of VF1 expression is as a result of the increased induction of the innate immune response.
It is well established that MNV is sensitive to an effective innate immune response: type I and II interferon are known to inhibit viral translation [20], a fact supported by the observed sensitivity of STAT1-/- mice to infection [6]. The role of STAT1 mediated, interferon-based, innate immune signaling in combating MNV-1 infection has been well characterized to prevent the progression of MNV-1 infection and dissemination to peripheral tissues [42]. We also observed this effect in our studies with immunocompetent C56BL/6 mice as only low levels of viral RNA were detected in the MLN (Figure 7). Previous work has also highlighted that at least part of the innate sensing of MNV infection by the host cell can be attributed to the MDA5 protein [17]. When activated, MDA5 signals the detection of viral RNA through the mitochondrial antiviral signaling (MAVS) adapter protein (also known as IPS-1, VISA or Cardif) embedded in the outer membrane of this organelle [36]–[38]. One downstream consequence of these signaling events at the mitochondria is the dimerisation and subsequent nuclear shuttling of IRF3. This activation of IRF3 results in the trans-activation of genes responsible for combating viral infection including interferon beta. In our studies, the regulation of genes specifically stimulated by virus infection was monitored by qPCR and ELISA and shown to be elevated in cells infected with a virus lacking the VF1 accessory protein (Figure 5A and Figure 5B). This potential role for VF1 as an antagonist of the innate immune response was investigated using co-expression studies with various auto-stimulatory components of this pathway which, after transient over-expression, are known to trigger interferon production (such as RIG-I, MAVS) (Figure 5C). In this instance, VF1 was shown to reduce the expression of a reporter protein under the direct transcriptional control of the interferon beta promoter. This occurred at the level of, or subsequent to TBK1 activation, since its stimulation of the IFN promoter was also inhibited by VF1 expression. The mechanism of action of VF1 therefore potentially involves modulation of the interaction of TBK1 with IRF3, or directly acts on IRF3 itself. Inhibition or degradation of IRF-3 is a frequent target of viral evasion strategies among both RNA and DNA viruses, including the Npro protein of pestviruses [55], [56], the P protein of rabies virus [57] and the G1 protein of hantaviruses [58]. Mitochondria also serve as a platform for the activation of IRF3 via TBK1 as the mitochondrial import protein Tom70, interacts with MAVS upon RNA virus infection and subsequently recruits the TBK1-IRF3 complex via Hsp90 [59]. The interaction of Tom70 with cytosolic chaperone Hsp90, which is itself constitutively associated with TBK1 and IRF3, plays a critical role in the activation of IRF3. Therefore another possible mechanism of VF1 function may be via the modification of the mitochondrial activation pathway or the formation of the MAVS-Tom70-Hsp90 complex.
The role of VF1 during in vivo infection was initially examined in a wild type immunocompetent mouse background. However, the strain of MNV-1 used during these studies and the only strain for which a reverse genetics system is currently available, namely CW1, is attenuated in this model, failing to produce any obvious clinical signs (Figure 7). It is worth noting however that this variant of MNV-1, although attenuated in STAT1-/- mice due to a glutamate at position 296 in the major capsid protein VP1 [10], is actually more representative MNV isolates from immunocompetent mice as they typically also contain glutamate at position 296 [32]. Using this variant of MNV-1 CW1 in this genetic background we observed viral RNA in the mesenteric lymph node at 5 and 7 days post infection with WT MNV-1 but not when animals were infected with a virus lacking VF1. These data add further strength to our hypothesis that VF1 expression is beneficial to virus replication as it delays the innate immune response and as a consequence, virus induced apoptosis. To examine the role of VF1 in MNV-1 virulence, the M1 VF1 truncation was first engineered into the virulent MNV-1 backbone, previously described by our laboratory [10]. This virus, which represents the closest available progenitor of the original isolate of MNV identified in 2003, causes a lethal infection in STAT1-/- mice [6]. Typically, one might expect to observe a restoration of virulence after infection of STAT1-/- mice with a virus that lacks an interferon antagonist, in this case VF1. One of the best-studied examples of this is in influenza virus where the lack of NS1 has no effect on virulence in STAT1-/- mice but in immunocompetent mice a deletion of NS1 results in attenuation [60]. However, In the case of the MNV-1 studies undertaken here, the in vivo analyses are complicated by the already attenuated nature of the strain used (CW1) in an immunocompetent host. Infection of STAT1-/- mice with either 104 or 103 TCID50 of WT-v or M1v demonstrated that MNV-1 lacking VF1 was partially attenuated in this system exhibiting delayed replication kinetics in the murine host (Figure 7, 8, 9 and 10). This manifested as a delay in both the onset of typical MNV-1 disease and the associated presentation of symptoms (weight loss, piloerection, anorexia, eye discharge that subsequently develop to ataxia, moribundity and death). Quantification of the viral RNA genomes in infected tissues at days 3 and 5 post infection, as well as the gross differences in MNV-1 related pathology at day 5 are testament to the debilitated replicative ability of this virus in vivo even in the absence of STAT1. The exact mechanism of this attenuation is unclear since all the inoculated animals (M1-v or WT-v) eventually developed disease and either succumbed to infection or had to be euthanized due to the established humane end points being surpassed. Detailed analysis of the function of VF1 in the avoidance of the innate immune response in vivo is likely to require the development of a reverse genetics system for a MNV variant capable of infecting immunocompetent mice. In the absence of this however, we are able to offer at least one possible explanation as to why we observed clinical disease in the STAT1-/- model even in the absence of VF1. In STAT1-/- mice, the absence of an intact STAT1-dependent interferon response pathway prevents the generation of robust autocrine and paracrine interferon responses. There are however many examples of virus infection leading to the induction of host genes classically defined as interferon stimulated genes (ISGs) in the absence of interferon and/or STAT1 mediated signalling; examples include LCMV [61] where the induction of ISG-49, ISG-54, and ISG-56 was observed in the absence of STAT1, and also HSV-1 which elicits an IRF3-dependent, but IFN-independent cellular antiviral response [62]–[64]. Direct IRF3 mediated responses are also known to protect against West Nile virus infection in both interferon dependent and independent mechanisms [65]. In addition, recent studies have highlighted that STAT2-mediated signalling may stimulate the expression of a subset of ISGs in the absence of STAT1 [66]. Therefore we would propose that during our studies in the STAT1-/- mouse model, it is likely that infection with the virus lacking VF1 leads to the induction of a subset of ISGs during virus replication at the primary site of infection, either directly via an unknown mechanism, or via STAT2. This limited response may slow virus replication, resulting in the delayed virus replication at the initial site of infection, reduced virus production and delayed onset of disease, all consistent with our observations. We would predict however that this limited response is not sufficient to clear virus after multiple rounds of infection. Our preliminary analysis would confirm that infection of STAT1-/- mice with virulent WT MNV-1, can result in the induction of ISGs, even in the absence of STAT1, as we observed increased levels of CXCL10 and ISG54 at 3 days post infection (Figure S4). The mechanism of ISG induction in the absence of STAT1-mediated signalling and how VF1 contributes to virulence in the absence of STAT1 will require further studies.
Expression of accessory proteins from alternate open reading frames can be found in many RNA viruses, many of which parallel the ability of VF1 to antagonize the innate immune system (discussed in more detail below). Many negative strand RNA viruses from the Paramyxovirus genus encode alternative proteins from internal ORFs in the phosphoprotein mRNA. These proteins, denoted C, play multiple roles in the viral life cycle that include the facilitation of RNA replication and control of the innate-immune response [67]. Sendai and measles virus mutants lacking C are viable in tissue culture but partially attenuated in vivo [68], [69]. This inability to replicate as efficiently as the wild-type virus in vivo is comparable to the observed results in this study for MNV-1 VF1 (Figure 7, 8, 9 and 10).
The influenza protein PB1-F2, another viral protein produced from an alternate open reading frame, provides additional evidence for the role of these accessory proteins in disease [43], [44], [70]. PB1-F2 is a recently discovered virulence factor, encoded by the PB1 gene segment, which interacts with mitochondria and stimulates apoptosis by facilitating cytochrome c release via interactions with ANT3 and VDAC1 [71], [72]. In addition PB1-F2 has been shown to affect influenza polymerase activity in the nucleus, modulate interferon responses during infection and, interestingly, to exacerbate secondary bacterial infections in vivo [71]. Despite the mitochondrial interactions of PB1-F2, it is unlikely that VF1 functions in an analogous manner since apoptosis was exacerbated in cells infected with a virus lacking VF1 (Figure 6). A recent report demonstrates a link between the innate immune response and apoptosis suggesting that both MAVS and IRF3 may play direct roles in stimulating apoptosis [73], [74]. ISG54 expression is also known to induce apoptosis [54]. It is possible that this exacerbation of apoptosis, in the absence of VF1, is a by-product of enhanced activation of the innate immune response in cells infected with the M1 virus (Figure 5 and 6) or an as yet uncharacterized direct or indirect modification of MAVS function by VF1. Interestingly, the HCV NS3/4A [75] and SARS-CoV NSP15 protein have been shown to be inhibitors of MAVS mediated apoptosis, identifying these proteins as potential orthologs of VF1 [73].
Other potential ARFP proteins have also recently been shown to associate with the mitochondria [76] as has the L* protein of Theiler's murine encephalomyelitis virus (TMEV) [77]. L* is only encoded by the TO subgroup of TMEV viruses where it is required for growth in macrophages and has been implicated in the establishment of persistence and the demyelination associated with TMEV [78]. This macrophage specific requirement for L* is paralleled, albeit to a lesser degree, for MNV VF1. In our study the M1 and M10 knockout viruses phenotypically reverted upon repeated, low multiplicity, multy cycle replication in the RAW264.7 murine macrophage cell line, highlighting that VF1 expression and function confers some benefit to MNV growth in this cell line (Figure 3E and 3F). This benefit is the likely combination of the observed increase in apoptosis and innate immune signalling observed in cells infected with a virus lacking the VF1 protein.
Numerous other examples of viral proteins that interact with the mitochondria during infection include the HBV X gene protein whose interactions are thought to stimulate apoptosis and play a role in the development of cancer in affected individuals [79], the HCMV UL37 protein which is thought to modulate Ca2+ signaling and apoptosis at the mitochondrial ER synapse [80], as well as three hepatitis virus proteases that have been shown to cleave MAVS (Hepatitis A, B and C), the adapter for RIG1 and MDA5 signaling, thereby antagonizing the innate immune response to infection [81]. Given the critical role that mitochondria play in cellular responses to infection and stress e.g. innate immune signaling, calcium homeostasis and regulation of apoptosis, it is maybe not surprising that so many viruses target this organelle.
Our studies have demonstrated that MNV-1 encodes a novel virulence factor from an alternate open reading frame in the sgRNA. This protein localizes to the mitochondria during infection and apparently inhibits the signaling events that take place in and around this organelle during infection of the host cell. This appears to affect the downstream activation of genes regulated by mitochondrial arm of the innate immune response and also the development of apoptosis in response to infection. A mutant virus lacking the ability to express VF1 does not replicate as efficiently in immuncompetent or STAT1-/- mice, which manifests as a delayed onset in the development of disease. However this does not protect the mice from developing serious disease highlighting the sensitivity of this specific model. The role of VF1 in the establishment of persistent MNV infection as well as the exact nature of VF1 interaction with the mitochondria and the mechanism by which this interaction modulates the function of this organelle is the subject of continued research in our laboratory. The identification and preliminary characterization of the MNV-1 VF1 protein provides a unique perspective on this widely used model pathogen and may provide additional insights into the mechanism of norovirus evasion of the immune system.
All of the STAT1-/- animals used in this study were maintained at an American Association of Laboratory Animal Care-accredited animal facility at UTSW Medical Center and the protocol was approved by the IACUC at UT Southwestern Medical Center (Permit number: 1151). Animal use adhered to applicable requirements such as the Animal Welfare Act (AWA), the Guide for the Care and Use of Laboratory Animals (Guide), the Public Health Service Policy, and the U.S. Government Principles Regarding the Care and Use of Animals. Studies involving C57BL/6 mice were performed at Imperial College London St Mary's Campus (PCD 70/2727) after ethical review by the Imperial College Ethical Review Panel and subsequent approval of the British Home Office (PPL 70/6838). All animal procedures and care in the UK conformed strictly to the United Kingdom Home Office Guidelines under The Animals (Scientific Procedures) Act 1986.
MNV-1 was propagated in the murine leukaemia macrophage cell line RAW264.7 using Dulbecco modified Eagle medium (DMEM) with 10% fetal calf serum (FCS), penicillin (100 U/ml), streptomycin (100 µg/ml) and 10 mM HEPES (pH7.6). COS7 and MEF cells were cultured in DMEM (with FCS and pen/step as above). Baby-hamster kidney cells expressing T7 DNA polymerase (BSRT7 cells) used during reverse genetics recovery of MNV-1 from cDNA clones, were obtained from Klaus Conzelmann (Ludwig-Maximilians-University Munich) [82] and cultured in DMEM (+FCS and pen/step as above) containing G418 at a concentration of 1 mg/ml. All cells were maintained at 37°C with 10% CO2.
Repeated attempts to immunize rabbit with full length his-tagged MNV-1 VF1, purified from E.coli, failed to illicit a robust immune response. Therefore a modified immunization regime that used a combination of peptide immunization followed by booster injections with various recombinant proteins was required. Full details are available upon request, but briefly animals were immunized with the peptide (PGKLTKLTPGSSKIL), representing amino acids 42–56 of VF1 conjugated to KLH, then boosted with the same peptide. This was followed by two subsequent booster injections using full length recombinant his-tagged VF1 expressed in and purified from E.coli. One booster injection with amino acids 42–70 (PGKLTKLTPGSSKILSSAPLVSFPSRLET) fused to a Cherry/his tag (Cherry-VF1-his), expressed and purified from E.coli using the Cherry express system (Eurogentec) was also performed. Animals were subsequently boosted again with the primary peptide immunogen conjugated to KLH, followed by a final boost with Cherry-VF1-his. VF1 specific antiserum was then affinity purified from sera on a column generated using the initial peptide immunogen. Antisera to the MNV-1 VP2 protein, the product of ORF3, was generated by immunization of rabbits with full-length recombinant his-tagged protein expressed and purified from E.coli. Note that some batch-to-batch variation of the anti-VP2 antisera was observed resulting in minor differences in the staining intensity of background non-specific bands. In some cases antisera was pre-adsorbed by prior incubation with membranes on which uninfected samples has been run to remove the non-specific reactivity. Antisera to the MNV-1 VP1 protein was kindly supplied by Skip Virgin (Washington University in St Louis) and was used as previously described [6].
The following 28 full length genomic sequences were downloaded from GenBank and used for bioinformatic prediction of open reading frames: DQ223042, EU854589, EU004665, FJ446720, FJ446719, AB435514, EU004660, EU004672, EU004679, EU004681, EU004682, EU004674, EF531291, EU004673, EF531290, DQ911368, EU004683, EU004670, EU004668, EU004663, EU004671, EU004664, EU004677, EU004676, DQ223041, DQ223043, EU004678 and EU004680. Sequences were selected based on showing >1% sequence divergence from all other sequences and thus representing different MNV isolates.
Synonymous and amino acid variability for each ORF coding sequence were calculated using the program Sequence distance in the Simmonic sequence editor as previously described [13]. Variability at each position was averaged over 11 adjacent windows of 50 codons incrementing by 3 bases/window. MLOGD [33], [83] was used through the web interface (http://guinevere.otago.ac.nz/aef/MLOGD/). The MNV phylogeny used to generate the Pairs files was created by PHYLIP version 3.62 [84] using DNADIST (Jukes-Cantor corrected distances) and NEIGHBOR programs. Likelihood scores were calculated for the existence of ORF4 in addition to ORFs 1, 2 and 3 (gene positions 6-5066, 5056-6678 and 6681-7307 added as annotation).
2×107 RAW264.7 cells were infected with MNV-1 at a MOI of 5 TCID50 per cell. After 12 h at 37°C, total cell lysates were prepared by washing in PBS and lysing directly into reducing SDS sample buffer. The mitochondria and cytosol of infected cells were separated using a mitochondria isolation kit for mammalian cells (Thermo scientific). The isolated mitochondria were directly suspended into reducing SDS sample buffer whilst due to the high volume the cytosolic fraction was concentrated using the UPPA-protein concentration reagent (G-Biosciences). Fractions were separated on a 15% SDS PAGE gel and analyzed by western blot using antibodies against VF1. A rabbit antibody against poly rC-binding protein1 and 2 (PCBP) acted as a control for the cytosolic fraction and the commercial goat anti-apoptosis inducing factor (AIF) antibody (D-20 from Santa Cruz Biotechnology) acted as a control for the mitochondrial fraction. VF1 and PCBP were detected using secondary HRPO conjugated anti-rabbit antibodies, whilst AIF was detected using a secondary HRPO conjugated donkey anti-goat antibody (Santa Cruz Biotechnology).
VF1 mutant viruses were generated by the insertion of stop codons at various positions within ORF4, which disrupted VF1 production without affecting the amino acid coding sequence of the major capsid protein VP1. The mutations were generated in the previously described MNV-1 cDNA clone pT7:MNV 3’Rz [22] by PCR mutagenesis (primer details available upon request) using KOD hot start DNA polymerase (Novagen). The VF1 mutant virus M1 contains a stop codon through mutation of T to A at genome position 5118 and hence translation of VF1 terminates after 16 amino acids. The VF1 truncated virus M10 contains a T to A mutation at genome position 5364 terminating VF1 translation after 98 amino acids, and the truncated M20 contains a G to A mutation at genome position 5655 terminating VF1 translation after 195 amino acids.
The VF1 expression plasmid pcDNA3.1+MNV-1 VF1 was generated by cloning the VF1 encoding sequence into the expression plasmid pcDNA3.1+, which contains a CMV and T7 promoter (primer details available upon request). GFP fusions of VF1 were generated by cloning the VF1 encoding sequence into the GFP plasmids pEGFP-N1 and pEGFP-C1 (Clontech).
ORF4 mutant viruses were recovered using reverse genetics as previously described [22]. Briefly, BHK cells expressing T7 polymerase (BSRT-7) were infected with FPV expressing T7 polymerase and were subsequently transfected with the MNV-1 full length clones (pT7 :MNV 3’Rz) containing the VF1 mutations M1, M10 or M20 (described above). 24 h post transfection cells were frozen and the clarified lysates were used to generate passage 1 and 2 stocks by infecting RAW264.7 cells at low MOI and freezing 48 hours post infection. The virulent WT and VF1 knockout viruses used in vivo were generated by the same means, although the virus underwent only a single pass in tissue culture to prevent the appearance of the tissue culture adapted mutations at genome positions 2151 and 5941 as described previously [10]. Viral titres were determined by TCID50 titration in RAW264.7 cells. Prior to use, all viruses were sequenced to ensure they contained the relevant mutations.
RAW264.7 cells were seeded at 3.2×105 cells per well in a 24-well plate and subsequently infected with the VF1 knockout (or truncated) viruses M1, M10 or M20 at an MOI of 0.01 TCID50 per cell. The assay was performed in triplicate for each virus. At given time points (0, 6, 12, 24, 48, 72 hours post-infection) the infections were frozen and upon thawing the viral titers were determined by TCID50. Protein samples over the given time course were also taken to analyze the kinetics of viral protein expression (data not shown).
RAW264.7 cells were seeded at 3.75×106 cells per well of a 6-well dish and were subsequently infected with ORF4 mutant viruses M1, M10 or M20 at a MOI of 0.01 TCID50 per cell. Note that the initial virus stock used for the infections were generated by reverse genetics recovery followed by a single passage in cell culture, so were effectively passage 1. The ORF4 region of the input viruses was sequenced prior to use. After 48 h the resulting ‘pass 1’ cultures were freeze-thawed and used to set up the subsequent low MOI (<0.01 TCID50 per cell) infections. Virus passage was continued for 5 cycles after which cells were infected at high MOI and RNA was isolated 12 h post infection. RT-PCR reactions were subsequently used to sequence the region of the genome encompassing ORF4 (primer details available upon request).
MEFs were seeded into 12 well dishes and transfected the following day using Lipofectamine 2000 (Invitrogen) with a reporter plasmid expressing firefly luciferase under the control of the complete IFN beta promoter. Where appropriate, cells were co-transfected with plasmids expressing VF1 or empty vector (pcDNA3.1) along with plasmids expressing cellular proteins (e.g. RIG-I etc. [85]). Empty plasmid was added to ensure each transfection received the same amount of total DNA. To normalize for transfection efficiency and to ensure lack of protein toxicity, the pRLTK Renilla luciferase reporter plasmid was added to each transfection. Samples were lysed 24 hours post transfection in passive lysis buffer (Promega) and activity measured using a dual luciferase reporter assay system (Promega) as described [85].
In order to analyze the virulence of wild type and VF1 mutant M1 viruses in mice, age (six to eight weeks, up to16 animals per group) and sex matched animals were orally inoculated (oral gavage) with the WT or the VF1 knockout viruses (diluted in DMEM to a total volume of 100 µl). Control mice (a group of 6) were inoculated with non-infected cell lysates prepared in an identical manner to the virus stocks. Mice were sourced from Taconic (STAT1 -/-) or Harlan (C57BL/6) and verified as MNV free at the beginning of the study. Control mice were verified as MNV free at the end of the study confirming barrier controls were effective. At various times post infection post infection tissue samples were taken post mortem. In addition a fecal sample was taken post mortem. All samples were snap-frozen and either stored in Trizol solution (Invitrogen) or RNA later (Ambion) before the RNA was extracted, as per the manufacturer's instructions. Remaining mice were left for the duration of the experiment or euthanized based on the following humane end points: a 20% loss of bodyweight, the development of severe symptoms (ataxia, moribundity), or the presence of moderate MNV-1 specific symptoms (continued weight loss, discharge from the eye) for 3 consecutive days. These limits were set in order to prevent undue suffering to the animal. In accordance with funding regulations and to minimize the number of animals used in these studies, the data for the control and WT-v inoculated groups of animals in Figure 7 has been reported in a previous study [14]. Note that both the previous experiment and that illustrated in Figure 8 were performed side by side.
Tissue samples were stored in Trizol reagent (Invitrogen) or RNa Later (Ambion) and after homogenization RNA was extracted according to the manufacturer instructions. Upon quantification, an aliquot of total RNA from each tissue sample was used for reverse transcription using AMV RT enzyme (Promega) and a primer specific for the genomic RNA of MNV-1 (IC464, CAAACATCTTTCCCTTGTTC). qPCR reactions were prepared using the MESA Blue qPCR MasterMix Plus for SYBR Assay (Eurogentech). Briefly, cDNA was mixed with 2X buffer and primers IC464 and IC465 (TGGACAACGTGGTGAAGGAT) prior to activation by incubation at 95°C for 10 min. Reactions were then subjected to 40 cycles of 94°C, 15 sec; 58°C, 30 sec; 72°C, 30 sec. Viral genome copy number was calculated by interpolation from a standard curve generated using serial dilutions of standard RNA representing nucleotides 1085 to 1986 generated by in vitro transcription and extrapolated back to per µg of input RNA. The limit of detection was determined either by the lowest dilution of control standard RNA reproducibly detected in the assay (Figure 7) or by incubating serial dilutions of standard RNA in RNA extracted from the tissues of mock infected animals (Figure 9). An equivalent protocol was used to determine the genome copy number in RNA samples extracted from infected RAW264.7 cells, using 500 ng of input RNA was used.
Analysis of mRNA levels in RAW264.7 cells was performed at low MOI (0.1 TCID50 units/cell). Cells were seeded at 2×105 cells per well in a 24 well dish, grown overnight at 37°C before being infected. RNA was harvested from infected cells at 16, 20 and 24 hours post infection using the GenElute RNA extraction kit (Sigma) as detailed by the manufacturer. RNA was quantified and diluted to a standard concentration before being reverse transcribed using MuMLV RT enzyme (Promega) using an oligo dT primer. SYBR green based qPCR was performed using an ABI 7900 HT real time PCR machine. The MESA Blue (Eurogentec) SYBR master mix was combined with sample cDNAs and optimised high efficiency mouse specific primers for the following genes: HPRT, CXCL10, ISG54 and Interferon Beta (primer details available upon request). Relative mRNA fold change was calculated using the ΔΔCt method e.g. normalising target mRNA levels based on an endogenous control (HPRT) before comparison with mock infected cells at equivalent time points. Where appropriate this fold change was then normalised to the level of MNV RNA detected in each sample in order to provide statistical information on mRNA or protein fold induction relative to MNV RNA and allow for the variance in the susceptibility of RAW264.7 cell clones to MNV infection (data not shown). Data handling and fold change was calculated using the SDS 2.3 and RQ programs from ABI. For the UV inactivation experiments high titre stocks of MNV were cross-linked for 15 minutes under high intensity UV light using a Spectrolinker XL-1500 (Spectronics Corporation). The artificial viral dsRNA analogue poly IC was used to confirm the suitability of RAW264.7 as a model for innate immunity, specifically their capacity to sense dsRNA. polyIC was added directly to the media at a final concentration of 25 µg/ml. RNA was harvested at 24 hours post addition and analysed by qPCR as detailed above. Analysis of IFN-Beta protein levels was performed using a murine IFN-B specific ELISA (PBL Interferon Source) as per manufacturer's instructions. Supernatants from infected or treated cells were centrifuged for 5 minutes at 2,000 rcf before analysis to remove any cellular debris.
For apoptosis assays, RAW264.7 cells were seeded at 5×105 cells per well of a 24-well plate and grown overnight at 37°C. Cells were infected with an MOI of 5 TCID50 per cell or treated with 5 µM staurosporine (Sigma) and at given time points (9, 12, 15, 18 and 21 h post infection) cells were PBS washed and lysed in 1 ml of 1 x cell culture lysis reagent (Promega). 100 µl of lysate was then incubated with 100 µl of Glo3/7 assay reagent (Promega) and after incubating at room temperature for 40 minutes, luminescence was read using a TD20/20 luminometer (Turner Designs). Samples were subsequently analyzed for protein content and the luminescence signal normalised to account for variations in the efficiency of cell lysis. Where UV-inactivated virus was used, virus stocks were UV-inactivated on ice for 20 minutes using a UV-crosslinker at 254 nm (Stratagene). Mock treated stocks were used as controls and were simply kept on ice for the same period of time prior to dilution and use in virus infections.
Levels of cleaved caspase 3, were compared by western blot analysis using antibodies from Cell Signalling Technology. Rabbit polyclonal antibodies generated against the viral polymerase (NS7) and major capsid protein (VP1) were used to control for equal amounts of virus, whereas a mouse monoclonal antibody to GAPDH (Ambion) was used to ensure equal protein loading.
|
10.1371/journal.pgen.1006808 | Impact of mutations in Toll-like receptor pathway genes on esophageal carcinogenesis | Esophageal adenocarcinoma (EAC) develops in an inflammatory microenvironment with reduced microbial diversity, but mechanisms for these influences remain poorly characterized. We hypothesized that mutations targeting the Toll-like receptor (TLR) pathway could disrupt innate immune signaling and promote a microenvironment that favors tumorigenesis. Through interrogating whole genome sequencing data from 171 EAC patients, we showed that non-synonymous mutations collectively affect the TLR pathway in 25/171 (14.6%, PathScan p = 8.7x10-5) tumors. TLR mutant cases were associated with more proximal tumors and metastatic disease, indicating possible clinical significance of these mutations. Only rare mutations were identified in adjacent Barrett’s esophagus samples. We validated our findings in an external EAC dataset with non-synonymous TLR pathway mutations in 33/149 (22.1%, PathScan p = 0.05) tumors, and in other solid tumor types exposed to microbiomes in the COSMIC database (10,318 samples), including uterine endometrioid carcinoma (188/320, 58.8%), cutaneous melanoma (377/988, 38.2%), colorectal adenocarcinoma (402/1519, 26.5%), and stomach adenocarcinoma (151/579, 26.1%). TLR4 was the most frequently mutated gene with eleven mutations in 10/171 (5.8%) of EAC tumors. The TLR4 mutants E439G, S570I, F703C and R787H were confirmed to have impaired reactivity to bacterial lipopolysaccharide with marked reductions in signaling by luciferase reporter assays. Overall, our findings show that TLR pathway genes are recurrently mutated in EAC, and TLR4 mutations have decreased responsiveness to bacterial lipopolysaccharide and may play a role in disease pathogenesis in a subset of patients.
| Esophageal adenocarcinoma (EAC) is a deadly human cancer that develops in the lower esophagus, which is exposed to reflux of acidic stomach contents and has reduced microbial diversity. Next generation sequencing studies have shown that accumulation of somatic mutations occurs along the metaplasia–dysplasia–carcinoma sequence in EAC, and it has been proposed that groups of genes involved in similar processes or pathways could have a cumulative effect. The Toll-like receptor (TLR) signalling pathway is critical for host-microbe interaction, and therefore mutations in these pathways could be important in the pathogenesis of EAC. Here we show that TLR pathway genes are recurrently mutated in 15% of EAC tumours and other solid tumour types exposed to microbial communities, including uterine endometrioid carcinoma (59%), cutaneous melanoma (38%), colorectal adenocarcinoma (27%), and stomach adenocarcinoma (26%). TLR mutant cases were associated with more proximal tumours and metastatic disease, indicating possible clinical significance of these mutations. A better understanding of how altered TLR signalling contributes to the inflammatory tumour microenvironment in EAC could help inform cancer prevention strategies.
| Esophageal adenocarcinoma (EAC) is increasing in incidence and has poor survival outcomes. The main risk factor for EAC is Barrett’s esophagus, a pre-malignant glandular epithelium that develops in the setting of gastro-esophageal reflux disease. Over time exposure to refluxed acid and bile in the lower esophagus leads to chronic inflammation, increased cell turnover, production of reactive oxygen species and DNA damage [1,2]. The combination of exposure to noxious substances and defects in DNA damage repair enable cancer cells to accumulate mutations, evidenced by characteristic mutational signatures [3,4].
Next generation sequencing studies have shown that accumulation of somatic mutations occurs along the metaplasia–dysplasia–carcinoma sequence in EAC [5,6] and potentially impair important cell functions. However, few genes are mutated in greater than 10% of cases, underlining the heterogeneous nature of point mutations and small indels in this type of cancer. As a result it is difficult to pinpoint the genes that are relevant to tumorigenesis using traditional approaches based on the frequency of mutations in a single gene. It has been proposed that groups of genes involved in similar processes or pathways could have a cumulative effect. An example is the SWI/SNF nucleosome remodeling complex (ARID1A, SMARCA4 and ARID2), for which gene members are mutated collectively in 20% of EAC tumors [6]. Computational tools such as PathScan [7] have been developed to identify cellular pathways that are targeted by somatic mutations above the background mutation rate.
Epidemiologic evidence has linked the rising incidence of EAC with the eradication of Helicobacter pylori [8,9], and studies have suggested that the esophageal microbiota are altered in Barrett’s esophagus [10–12] and EAC with decreased microbial diversity [13]. The Toll-like receptor (TLR) signaling pathway is a key component of the innate immune system and one of the main ways in which tumor cells interact with the microbiota. TLRs are pattern recognition receptors that bind unique molecular components of the microbiota and generate a pro-inflammatory innate immune reaction through nuclear factor kappa B (NF-κB) [14]. TLR signaling exerts an effect on epithelial cell function in the gastrointestinal tract, including stimulating repair of damaged enterocytes (lipopolysaccharide), enhancing cell proliferation, and triggering secretion of antimicrobial peptides (lipopolysaccharide and flagellin) [15]. Alongside TLRs, inflammasomes also contribute to inflammation through recognition of pathogen-associated molecular patterns. NLRP6 is a member of the NOD-like receptor family that plays a role in regulating inflammation and epithelial cell repair in the intestine and has been implicated in colorectal carcinogenesis [16,17].
Genome-wide association studies [18,19] and next-generation sequencing studies [5,6] have identified TLR4 gene mutations in solid tumors including EAC. We hypothesized that somatic mutations may collectively target the TLR signaling pathway in EAC and alter inflammatory signaling. We aimed to interrogate TLR pathway mutations and expression in a cohort of EAC and Barrett’s esophagus patients with clinical outcome data. We then investigated whether TLR4 mutations in EAC affect downstream inflammatory signaling using an in vitro model system. Finally, we aimed to determine the broader relevance of TLR pathway mutations in other cancers using TCGA data and the COSMIC database.
To determine whether the TLR signaling pathway is dysregulated through somatic mutations in EAC, we interrogated the mutational profiles of 171 EAC tumors and matched germline controls that were sequenced as part of the International Cancer Genome Consortium (ICGC) esophageal study. Non-synonymous somatic mutations affected the Toll-like receptor signaling pathway in 25/171 (14.6%) of EAC samples. Missense mutations (and two splice variants) were identified in TLR4 (5.8%), TRAF6 (1.8%), TLR7 (1.8%), TLR9 (1.2%), MYD88 (1.2%), IRAK4 (1.2%), LBP (0.6%), TRAF3 (0.6%), TLR5 (0.6%), and TLR2 (0.6%, Fig 1A, S1 Table). Applying the PathScan tool showed that genes in the TLR signaling pathway were significantly enriched for mutations (p = 8.7x10-5). Fig 1A presents the mutation and copy number status for known recurrent alterations in EAC in the context of TLR pathway mutated samples, and these events are compared with the remainder of the ICGC cohort in S1 Fig. As expected in EAC, most samples showed mutations in TP53, independent of whether or not they contain TLR pathway mutations. There was no significant enrichment of other known driver mutations in the TLR pathway mutated samples (S1 Fig). TLR pathway mutated tumors did not show significant differences in the numbers of total SNVs (Welch’s t-test, p = 0.134) or non-synonymous SNVs (p = 0.147) compared to wild-type tumors and did not show significant enrichment of any of the molecular subtypes recently defined on the basis of their dominant molecular signature (Fisher’s exact test, p = 0.57, S1 Fig) [20].
To investigate other potential sources of altered function of the TLR pathway genes, we investigated copy number and structural variants potentially affecting the TLR pathway in the ICGC cohort. Analysis of the copy number profiles of the TLR pathway genes indicated copy number gains for LBP and CTSK, while TLR3, RIPK1 and CD14 frequently showed fewer copies compared to the average ploidy (S2 Fig). Three samples showed homozygous deletions in TLR pathway genes: MYD88 in ICGC-30, IRAK1 in ICGC-24, and TLR7, TLR8 and IRAK1 in ICGC-10. Interestingly, several TLR pathway genes were affected by loss of heterozygosity. Loss of heterozygosity events overlapped with SNVs in TLR4 in four samples (ICGC-24, ICGC-156, ICGC-18 and ICGC-141) and in MYD88 in one sample (ICGC-78). Eleven out of 171 samples showed structural variants whose breakpoints overlap with TLR pathway genes or lead to a potential fusion with a TLR pathway gene (S2 Table). The following genes were affected: TLR1, TLR3, TLR9, TLR10, RIPK1, TOLLIP, TRAF3, TRAF6, TRAM1 and LBP. Most structural variants were classified as duplications or deletions and there were three translocation and two inversion events. No recurrent events were detected.
To ensure that the findings were generalizable we interrogated exome sequencing data from the Dulak et al. TCGA cohort [6] using our computational pipeline, and found a similar mutational spectrum with non-synonymous mutations in TLR pathway genes in 33/149 (22.1%, PathScan p = 0.05) of EAC samples (Fig 1B). Based on the combined data from both cohorts (Fig 1C and 1D), the most frequently mutated genes were TLR4 (5.3%) and TLR9 (3.1%), and these events appear to be mutually exclusive. We verified 11/11 TLR4 mutations and 2/2 TLR9 mutations from the ICGC cohort using PCR and Sanger sequencing (S3 Fig). We also examined the COSMIC database for 20 different cancer types (as defined in S3 Table), comprising a total of 10,318 samples, and found a high proportion of mutations in TLR pathway genes in endometrial carcinoma (188/320, 58.8%), cutaneous melanoma (377/988, 38.2%), colorectal adenocarcinoma (402/1519, 26.5%), stomach adenocarcinoma (151/579, 26.1%), lung adenocarcinoma (81/477, 17%), lung squamous cell carcinoma (81/497, 16.3%), head and neck squamous cell carcinoma (41/330, 12.4%), and esophageal carcinoma (combined adenocarcinoma and squamous cell carcinoma: 99/869, 11.4%, Fig 2A). The COSMIC database further showed a high proportion of TLR4 non-synonymous mutations in cutaneous melanoma (60/988, 6.1%), lung adenocarcinoma (22/477, 4.6%), stomach adenocarcinoma (26/579, 4.5%) and lung squamous cell carcinoma (17/497, 3.4%, Fig 2A and S4 Table). Next we categorized the different cancer types into two groups based on tumors that arise in body sites associated with significant microbiomes, including the oral and gastrointestinal tract, skin, urogenital tract and respiratory tract [21], and those with rare exposure to microbes. There was an increase in the frequency of TLR pathway mutations in cancer types that are highly exposed to microbes (p = 0.019, Wilcoxon rank-sum test, S4 Fig). This trend was also present when looking only at the fraction of TLR4 mutant tumors (p = 0.028).
Next we investigated the clinical relevance of TLR pathway mutations through correlation with clinical outcome data. In the ICGC EAC cohort, patients with TLR pathway mutations (n = 25) tended to have more advanced disease with metastases (Fisher’s exact test, p = 0.036, Table 1). TLR mutant tumors originated more proximally at the level of the gastro-esophageal junction or above (Siewert Type 1–2 or esophageal, Fisher’s exact test, p = 0.012). There was a trend towards decreased survival in patients with TLR mutations in comparison to wild-type although this did not reach statistical significance, possibly due to the limited number of TLR mutant cases and length of follow-up (S5 Fig). This trend was also seen when comparing patients with TLR pathway mutations to wild-type patients in the EAC cohort from the TCGA database, although again the clinical data is limited.
EAC frequently arises from the premalignant lesion Barrett’s esophagus through different degrees of dysplasia. To investigate the timing of TLR pathway mutations in disease pathogenesis, we examined samples from patients with Barrett’s esophagus adjacent to tumor (n = 24). The adjacent Barrett’s was non-dysplastic in 18/24 cases, contained low grade dysplasia in 2/24 cases and was indefinite for dysplasia in 2/24 cases. Only 2/24 Barrett’s samples showed mutations in the TLR pathway. A TRAF6 mutation was present in tumor and adjacent Barrett’s, indicating that the mutation had occurred early in carcinogenesis. In another patient, a TLR9 mutation was detected in Barrett’s but not the adjacent tumor. No other TLR pathway mutation was found in any of the Barrett’s adjacent to tumors, while three additional tumors had TLR pathway mutations.
Since TLR4 was frequently mutated in both EAC cohorts and other solid tumor types in the COSMIC database, we decided to characterize these mutations in greater detail. Both EAC cohorts identified missense mutations at amino acid position E439 (substitution to glycine) and F487 (substitution to leucine or valine, Fig 2B). The COSMIC database showed additional missense mutations at position E439 (two in stomach adenocarcinoma and one in cutaneous melanoma) and position F487 (two in stomach adenocarcinoma and one in esophageal carcinoma, Fig 2C). Seven tumors with TLR4 mutations had paraffin-embedded tissue available to evaluate mutant TLR4 protein expression using immunohistochemistry for TLR4 monoclonal antibody. Similar to wild-type tumors, the TLR4 mutant tumors showed combined membranous and cytoplasmic staining of TLR4 protein, with staining intensity ranging from weak to strongly positive (S6 Fig). None of the mutant tumors showed complete loss of TLR4 expression, which was anticipated since the missense mutations did not cause truncation of the protein. We hypothesized that the mutations could have a functional effect on TLR4 signaling, and this was supported by computational modeling using a published crystal structure for dimerized human TLR4 ectodomain with associated MD2 co-receptor (LY96 gene product) with LPS bound (PDB ID 4G8A [22]) and a hypothetical structure for the dimerized TIR domain based on TLR10 (PDB ID 2J67 [23], S7 Fig). Two structurally significant mutations affected the TIR domain (F703C and R787H), which is involved in downstream signaling and interaction with the adaptor molecule MyD88 [14]. The E439G mutation is also critically located at the TLR4 dimerization interface and may disrupt hydrogen bonds in the binding site of LPS and MD2 (Fig 2D). Further, amino acid sequence alignment against seven non-human species showed that the positions of the TIR domain mutations (F703 and R787) are evolutionarily conserved, along with amino acids L80, L498 and S570, and E439 is semi-conserved (S8 Fig). Overall the combination of crystal structure modeling, sequence alignment, and SNP prediction algorithms (SIFT and Polyphen) suggested that six of the verified TLR4 mutations could have a functional consequence: L80M, E439G, L498V, S570I, F703C and R787H (S5 Table).
To test our functional predictions for the TLR4 mutants, we performed site-directed mutagenesis and NF-κB luciferase reporter assays in HEK293 cells, a common model for measuring TLR signaling. The TLR4 mutants were stimulated first using the weak agonist synthetic monophosphoryl Lipid A (MPLA). There was a significant decrease in ligand-dependent signaling for 7/9 of the TLR4 mutations stimulated with synthetic MPLA (S9 Fig). A double mutation of E439G with F703C, representing a tumor with two TLR4 mutations, showed a further decrease in TLR4 signaling compared to F703C (p = 0.0052) but not E439G (p = 0.099). Stimulation with stronger TLR4 agonists, synthetic Lipid A and lipopolysaccharide (LPS), showed a significant decrease in signaling for four single mutants (E439G, S570I, F703C and R787H) and the double mutation E439G + F703C (Fig 3A). Western blotting for recombinant FLAG-TLR4 confirmed adequate expression of the different TLR4 mutants (Fig 3B). We also visualized recombinant TLR4-FLAG protein expression in HEK293 cells using confocal microscopy for mutants E439G and R787H, and there was no difference in expression or localization of TLR4-FLAG for either of the mutants in comparison to wild-type, suggesting that the decreased signaling was due to altered protein function rather than mis-folding and failure to reach the cell surface (Fig 3C).
Next, TLR4 mutants E439G, R787H and E439G+F703C were transfected into EAC cell lines stimulated with LPS for 24 hours, and secretion of the NF-κB dependent cytokine IL-8 was measured. OE33 and JH-EsoAd1 cells were selected because of their low endogenous TLR4 mRNA expression and ability to secrete measurable amounts of IL-8 (S10 Fig). The fold change of IL-8 secretion was significantly lower for mutants R787H and E439G+F703C in comparison to wild-type TLR4 (Fig 3D). In contrast to HEK293 cells, no significant decrease in TLR4 signaling was observed for mutant E439G stimulated with LPS in the EAC cell lines, suggesting that the strong agonist LPS was still able to trigger TLR signaling despite mutation of the ligand binding site.
Our experiments in cell lines suggest that TLR4 mutations impair TLR signaling and NF-κB activation in HEK293 cells and IL-8 secretion by EAC cell lines. We next tested whether there was any effect on gene expression in patient data using RNA-Seq data available for the TCGA cohort. Out of 89 samples with RNA-Seq data available, 17 samples had TLR pathway mutations and four had TLR4 mutations. Our analysis using DESeq2 [24] shows that expression of IL-8, NFKB2 and RELB was significantly elevated in the tumors compared to normal samples (n = 10), but no significant difference was observed when comparing tumors with mutations in the TLR pathway and wild-type tumors, possibly related to the small sample size (Fig 4A). Similarly, there was no significant difference in gene expression between TLR4 mutant and wild-type tumors. We next searched for alternative genes whose expression could be affected by TLR mutations in vivo. Tumors with TLR pathway mutations showed significant upregulation of NLRP6, GAST, TTC29 and C19orf69, and down-regulation of SFRP5, MYO18B, NAT8L, SHISA9 and IGFALS (Fig 4B). We validated our findings using RNA-Seq data from 23 independent tumors of which 2 were mutated in the TLR pathway, and significant upregulation was found only for NLRP6 (p = 0.004). NLRP6 is involved in pathogen recognition through the inflammasome pathway and has overlap in function with the TLR pathway. Additionally, quantitative RT-PCR was performed in 22 tumor samples with available RNA (11 of which contained TLR pathway mutations, including 5 TLR4 mutations). The results showed a similar trend, with upregulation of NLRP6 in TLR pathway mutated samples in comparison to wild-type tumors; however, this did not reach significance likely due to the small sample size (p = 0.172, S11 Fig). Of the samples with TLR4 mutations, R787H (relative expression 6.8) and L80M (relative expression 3.7) showed upregulation of NLRP6 compared to mean expression in wild-type tumors (relative expression 0.87, S6 Table).
Our results suggest that somatic mutations may collectively target the TLR signaling pathway in EAC (25% of cases) and other solid tumor types. TLR mutant cases were associated with more proximal tumors and metastatic disease, indicating possible clinical significance of these mutations. TLR4 was the most frequently mutated TLR gene in 5% of cases in the combined EAC cohorts. Only two TLR pathway mutations (TLR9 and TRAF6) were detected in the adjacent Barrett’s samples in the ICGC cohort. TLR4 mutations have been previously reported in high grade dysplasia [5], which implies that this mutation may be acquired later during disease progression. Although TLR4 is mutated in less than ten percent of EAC cases, it is one of the top twenty most frequently mutated genes in both the ICGC and TCGA cohorts. The most frequently mutated gene in this disease is TP53, which is mutated in approximately 70% of tumors [6]. A limited number of candidate driver genes have been identified that are mutated in greater than ten percent of EAC tumors, including CDKN2A, SMAD4, ARID1A, MYO18B and DOCK2 [5,6].
In addition to EAC, TLR pathway mutations were frequently observed in solid tumors arising in body sites exposed to microbiota [21] including the oral and gastrointestinal tract (colorectal adenocarcinoma, stomach adenocarcinoma and head and neck squamous cell carcinoma), skin (melanoma), urogenital tract (endometrial carcinoma) and respiratory tract (lung adenocarcinoma and squamous cell carcinoma). Regrettably, it was not feasible to analyze our whole genome sequencing data using pathogen discovery software such as PathSeq [25], because the concentration of microbial DNA is so low relative to human DNA in samples derived from esophageal tumor tissue and the results were unreliable with high levels of contaminants. This is a major limitation of analyzing low microbial biomass tissue samples with a WGS approach. Our recent work examining the microbiome in esophageal tumors and Barrett’s tissue samples used 16S rRNA gene amplicon sequencing and a microbial extraction protocol with rigorous reagent controls to maximize the microbial DNA yield and account for contaminants [13]. However, only one of the tumors from that study had a TLR4 mutation so it was not possible to further correlate the findings with our current study, and additional research is needed to investigate the link between TLR mutations and the microbiome in EAC.
Seven TLR4 mutants in our study showed decreased signaling in response to weak agonists, and four were hypo-responsive to strong agonists. This suggests that the mutations differentially affected signaling with weak versus strong agonists, which may be relevant to the different microbial antigens present in the tumor environment. The model system was also a contributing factor; for instance E439G was hyporesponsive to LPS in HEK293 cells but not EAC cell lines. In EAC cell lines, addition of the second mutation F703C was required to significantly reduce TLR4 signaling activity. Further, loss of heterozygosity events overlapped with four TLR4 mutations, R787H, E603D, T193K, and L498V. Of these, R787H showed decreased signaling in response to strong and weak agonists, L498V showed decreased signaling in response to weak agonist only, and E603D and T193K showed no significant change. It is conceivable that loss of heterozygosity could potentially compound the effect of TLR4 missense mutations and further decrease signaling. A possible mechanism is that defective TLR4 signaling may negatively impact epithelial cell repair, which is in part dependent on TLR4 stimulation, and potentially enable microbes to breach the epithelial barrier in the tumor microenvironment. Hold et al. found that the single nucleotide polymorphism c.896A>G (p.D299G) was associated with an increased risk of gastric cancer in two patient populations [19], and cells with this mutation have been shown to be hypo-responsive to lipopolysaccharide [26,27]. However, Hold et al. found no significant increased risk for EAC or esophageal squamous cell carcinoma with this germline polymorphism, and the D299G polymorphism was not identified in either of the ICGC or TCGA cohorts.
The relationship between TLR4 signaling and tumorigenesis is complex and involves both innate and adaptive immunity, with evidence showing that TLR4 signaling can enhance or suppress cancer development, depending on the model system. For example, in the setting of chemically induced colitis TLR4 deficient mice were protected from colon carcinogenesis [28], while villin-TLR4 mice overexpressing TLR4 in the intestinal epithelium were highly susceptible to cancer when treated with dextran sodium sulphate and azoxymethane [29]. Conversely, overexpressing TLR4 using the CD4-TLR4 transgene in the intestinal epithelium of APCMin/+ mice that are genetically susceptible to colon carcinoma reduced tumorigenesis by increasing apoptosis [30]. Additionally, analysis of TCGA expression data suggested that NLRP6 is upregulated in TLR mutant tumors, which may reflect cross-talk between the TLR pathway and NOD-like receptor signaling pathway. NLRP6 regulates inflammasome-dependent innate immune signaling and has been shown to inhibit TLR2 and TLR4-dependent activation of NF-κB and MAP-kinase signaling pathways [31]. Further understanding of how altered TLR signaling may contribute to the inflammatory tumor microenvironment in Barrett’s carcinogenesis could be helpful in cancer prevention strategies.
Ethical approval was obtained from the National Research Ethics Services Cambridgeshire Research Ethics Committee on behalf of all hospital centers in the OCCAMS/ICGC trial (REC 07/H0305/52 and 10/H0305/51). Written informed consent was obtained from all subjects prior to the collection of samples and recording clinical information.
EAC tissue samples with matched germline controls (n = 171 patients) were collected from 11 UK hospitals participating in Oesophageal Cancer Clinical and Molecular Stratification (OCCAMS) for the International Cancer Genome Consortium (ICGC). Ethical approval was obtained from the National Research Ethics Services Cambridgeshire Research Ethics Committee on behalf of all hospital centers in the OCCAMS/ICGC trial (REC 07/H0305/52 and 10/H0305/51). Written informed consent was obtained from all subjects prior to the collection of samples and recording clinical information. All tissue samples were flash frozen in liquid nitrogen and stored at -80°C. The tissue histology was reviewed by an expert gastrointestinal pathologist to ensure that tumor cellularity was greater than seventy percent in samples selected for sequencing.
DNA was extracted using the QIAGEN AllPrep kit and quantified using the Qubit Fluorometer. Whole genome sequencing was performed on the Illumina HiSeq platform in San Francisco, CA, USA. 100 bp paired-end sequencing was performed to an average depth of 50-fold for tumors and 30-fold for controls, achieving a Phred quality score of at least 30 for 80% of mapping nucleotide bases. WGS sequencing data has been deposited in the European Genome-phenome Archive under accession number EGAD00001001960: https://www.ebi.ac.uk/ega/search/site/EGAD00001001960.
Single-nucleotide variant calling was performed as described previously [20]. Briefly, reads were mapped to the human reference genome (GRCh37) using BWA 0.5.9 [32]. Mutations and Indels were called using Strelka 1.0.13 [33]. The resulting SNVs were filtered using a custom set of filters [20] and then annotated using the Variant Effect Predictor (VEP release 75) tool [34]. As described in Secrier et al. [20], for structural variant calling the reads were mapped to the GRCh37 reference genome, and Manta [35] was used to identify putative breakpoint junctions using discordant read pairs and split reads. Breakpoints and potential run-through events were annotated using the Ensembl GRCh37, version 75 gene annotation.
Copy number calling was performed as described by Secrier et al. [20] In short, absolute and minor copy number of genomic loci was estimated using ASCAT-NGS v2.1 [36]. Read counts at germline heterozygous positions were estimated by GATK 3.2–2 [37]. ASCAT estimates for tumor purity and average ploidy for each sample were used. Loss of heterozygosity and homozygous deletions were defined for loci with an estimated minor copy number of 0 and an estimated total copy number of 0, respectively. For each gene the highest copy number change is shown. The copy number alterations for genes were classified into gains, losses, amplifications and deletions based on the relative copy number rCN = log2(CN/ploidy), where CN is the copy number of the gene and ploidy is the average ploidy as estimated by ASCAT. The cut-offs for classification are: deletion–rCN< = -2, loss—rCN< = -1, no change—-1>rCN<1, gain—rCN> = 1, amplification—rCN> = 2.
We defined a list of genes that are specific to the TLR pathway based on the KEGG Toll-like receptor signaling pathway. PathScan [7] was used to assess whether mutations affecting the TLR pathway were significantly enriched in the ICGC cohort or the Dulak et al. TCGA cohort [6]. We used a Docker image (https://hub.docker.com/r/beifang/music/) of the Genome MuSiC 0.4 suite [38] to run PathScan. The overall background mutation rate was calculated on non-synonymous mutations using the human genome GRCh37 and all human Ensembl genes as a reference. The mutational consequences of each SNV predicted by VEP were converted to MAF format using the conversion scheme of vcf2maf (https://github.com/mskcc/vcf2maf/blob/master/vcf2maf.pl). The COSMIC database v78 (http://cancer.sanger.ac.uk/cosmic) was downloaded and interrogated for TLR4 and TLR pathway mutations. We only used mutations derived from high-throughput studies. Further, we selected only missense and nonsense substitutions, as well as frameshift insertions or deletions. Mutations were counted by cancer type as defined by the filters for tissue type and histology described in S3 Table.
Samples were grouped into molecular subtypes as defined by Secrier et al. [20] The mutational context of each mutation was defined using the UCSC hg19 reference genome and the R package SomaticSignatures [39]. The estimated mutational signatures from Secrier et al. were used to fit the exposures by non-negative least squares. The three molecular subgroups were defined by consensus clustering on the estimated exposures.
The ICGC TLR4 mutations were modeled using mutation functions in COOT (Crystallographic Object-Oriented Toolkit) [40] applied to the crystal structure for dimerized TLR4 ectodomain (PDB ID: 4G8A [22]) bound to MD2 and lipopolysaccharide (LPS) and a hypothetical structure for the Toll/interleukin-1 receptor (TIR) domain based on the TLR10 dimer (PDB ID: 2J67 [23]). The starting ectodomain PDB structure 4G8A was the common human variant (D299G and T399I), which was initially mutated back to D299 T399 before modeling the observed ICGC mutations. The TLR10 TIR domain template had a 35% identity (61% similarity with BLOSUM50 matrix) over the 143 residues of the TLR4 TIR domain. TLR4 ectodomain model energy minimization, was with FoldX 3.0 [41], and TIR dimer contacts were optimized using PyRosetta [42]. Contacting TLR4 residues less than 0.5 Å apart at surfaces were determined using UCSF Chimera and sorted by residue range in Mathematica(TM) to show dimer contacts (to other TLR4 chain), interactions with the ligand lipopolysaccharide (LPS) and adaptor MD2. Amino acid sequence alignment for human TLR4 was performed using Uniprot (http://www.uniprot.org/align/) and Espript (http://espript.ibcp.fr/ESPript/cgi-bin/ESPript.cgi) against seven non-human species.
Primers were designed to amplify an area approximately 200 bp in size surrounding TLR4 mutations using Primer3 (S7 Table). PCR amplification was performed with Q5 Hot Start High Fidelity Taq and the following conditions: denaturation at 98°C for 30 s, 35 cycles of 98°C for 10 s, 55–60°C for 10 s and 72°C for 10 s, followed by a final extension at 72°C for two minutes. The PCR product was sent for Sanger sequencing (Source BioScience, Cambridge, UK).
The plasmids pNF-κB-luc (firefly luciferase), pEFIRES-MD2, phRG-TK (renilla luciferase), pcDNA3-CD14, and a pCMV vector containing TLR4 cDNA with an N-terminal FLAG tag (pCMV-TLR4-FLAG) were provided by N. Gay [43]. HEK293 cells were cultured in Dulbecco’s Modified Eagle’s Medium (DMEM: Gibco, Life Technologies Ltd.) supplemented with 4.5 g/l glucose, 10% foetal bovine serum (FBS: HyClone Laboratories Inc.), penicillin (50 U/ml) and streptomycin (50 μg/ml) at 37°C, 5% CO2. The EAC lines OE33 (European Collection of Cell Cultures (ECACC)), SKGT4 (ECACC), OE19 (Sigma-Aldrich), OACP4C and OACM5.1 (both provided by W. Dinjens [44]) were cultured in RPMI-1640 medium (Sigma-Aldrich) supplemented with 10% FBS. JH-EsoAd1 cells were obtained from H. Alvarez [45] and grown in Minimal Essential Medium (MEM: Gibco) supplemented with 10% FBS. HEK293-hTLR4-MD2-CD14 cells (InvivoGen) were cultured in DMEM with 4.5 g/l glucose, 10% FBS, 50 U/ml penicillin, 50 μg/ml streptomycin and 100 μg/ml Normocin (InvivoGen). Selective antibiotic pressure was maintained with 10 μg/ml Blasticidin (InvivoGen) for hTLR4 and 50 μg/ml Hygrogold (InvivoGen) for MD2 and CD14 plasmids. FLO-1 cells (from D. Beer [46]) were cultured DMEM with 4.5 g/l glucose and 10% FBS.
Site-directed mutagenesis was performed using the Stratagene QuikChange II Kit and PfuUltra HF polymerase (Agilent Technologies, Inc.). Primers were designed using PrimerX with the mutation site at the center region (S7 Table). Mutagenesis was confirmed by Sanger sequencing the entire TLR4 cDNA. NOT1 and BamH1 (New England Biolabs) were used for cloning, and ligation was performed with T4 DNA Ligase (New England BioLabs) using an insert:vector ratio of 3:1.
HEK293 cells were seeded at 20,000 cells per well in 96-well plates and transfected with 0.2 μl jetPEI (Polyplus Transfection) per well and the expression plasmids (S8 Table), in triplicate. Forty-eight hours later, the cells were stimulated for six hours with 10 ng/ml LPS from E. coli O127:B8 (Sigma-Aldrich, L3129), 100 ng/ml synthetic Lipid A (PeptaNova, 24005-s) or 5 ng/μl synthetic monophosphoryl Lipid A (Enzo Life Sciences, ALX-581-205-C100) diluted in filtered DMEM with 10% FBS. Following lysis with passive lysis buffer (Promega) the luminescence of firefly and renilla luciferase was quantified with luciferin and coelenterazine reagents, respectively, using a FLUOstar OPTIMA microplate reader (BMG LABTECH).
Total RNA was extracted using the AllPrep mini kit (QIAGEN), and 1 μg of RNA was transcribed into cDNA using the Quantitect reverse transcription kit (QIAGEN) and diluted 1:5 in nuclease free water. Two μl of cDNA were amplified in a 10 μl reaction volume containing 500 nM of each primer and 5 μl SYBR Green I Master (Roche). The primers for TLR4 spanned across two exons and the sequences were 5’CCGACAACCTCCCCTTCTCA (forward primer) and 5’GGCTCTGATATGCCCCATCTTC (reverse primer). The sequences for NLRP6 primers were 5’cctggtgggtatgcttcgg (forward primer) and 5’ctcctgtagtgactgctcgc (reverse primer). The LC480 LightCycler 480 II (Roche) template for SYBR Green 384 well plates was used with 45 amplification cycles of 10 s each for 95°C, 60°C, 72°C. All reactions were performed in triplicate. The cycle threshold (CT) results were normalized by subtracting the average of three housekeeping genes: ribosomal protein S18 (RPS18: forward 5'ATCCCTGAAAAGTTCCAGCA, reverse 5'CCCTCTTGGTGAGGTCAATG), beta-actin (ACTB: forward 5’GGCACCCAGCACAATGAAGA, reverse 5’ AAGCATTTGCGGTGGACGAT) and glyceraldehyde 3-phosphate dehydrogenase (GAPDH: forward 5’GTCTCCTCTGACTTCAACAGCG, reverse 5’ACCACCCTGTTGCTGTAGCCAA). The delta-delta CT method [47] was used to compare relative mRNA expression between cell lines.
Protein lysate was separated using SDS-PAGE and blotted onto an Immobilon-P membrane (Fisher Scientific UK Ltd.). The membrane was probed with monoclonal rabbit antibodies for FLAG (1:1000, Sigma-Aldrich, F2555) and β-actin (1:10,000, New England Biolabs, 4970L), and incubated with peroxidase conjugated Pierce goat anti-rabbit IgG, H+L (1:2500, Fisher Scientific Ltd., 31462) secondary antibody. Bands were visualized using Amersham ECL Prime detection reagents (GE Healthcare UK Ltd.) developed on Kodak Biomax XAR film. Size of glycosylated TLR4 (120 kDa) and deglycosylated TLR4 (110 kDa) is based on da Silva Correia et al. [48].
Cells were seeded at 20,000 cells per well in 96-well plates and transfected with 0.25 μl Lipofectamine 2000 (Polyplus Transfection) per well and the expression plasmids (S9 Table). Twenty-four hours after transfection, cells were stimulated with 10 ng/μl monophosphoryl Lipid A or 40 ng/ml LPS diluted in culture medium. Secretion of IL-8 was measured using human Quantikine ELISA kits (R&D Systems Europe Ltd.) in duplicate.
Cells were seeded at 40,000 per well on coverslips in 24-well plates and transfected with 1 μl jetPEI per well and the appropriate transfection vectors (S10 Table), in triplicate. Forty-eight hours after transfection the cells were washed with PBS and fixed with 4% paraformaldehyde in PBS for 30 minutes at room temperature. The cells were washed twice and permeabilized with 0.2% Triton X-100 in PBS for 20 minutes at room temperature. The cells were washed three times with PBS and blocked with 2% bovine serum albumin and 5% chick serum in PBS for one hour at room temperature. The cells were stained with Alexa Fluor 488 conjugated DYKDDDDK (FLAG) Tag antibody (Cell Signaling Technology, 5407) diluted 1:200 in blocking buffer overnight at 4°C protected from light. The cells were washed three times with 0.1% TWEEN 20 in PBS (PBS-T) and incubated with Atto 594 Phalloidin (Sigma-Aldrich, 51927) diluted 1:100 in PBS for 20 minutes at room temperature protected from light. The cells were washed with PBS-T and stained with DAPI in PBS (1:5000) for 15 minutes at room temperature protected from light. The coverslips were inversely mounted on slides with hard-dry DPX mountant (Leica Biosystems). A Leica TCS SP5 II confocal microscope (Leica Microsystems) was used to image the cells.
RNA-Seq read counts per gene and variant calls for esophageal cancer were downloaded from the TCGA data portal (https://gdc.cancer.gov/). Only esophageal adenocarcinoma samples that had matched genotype data available and normal control samples were selected for the analysis. ICGC RNA-Seq bam files in Secrier et al. [20] for samples matching our cohort were also used. Reads overlapping RefSeq genes were counted in R using the GenomeAlignment package [49]. Entries with duplicated gene names were summarized to the median read count per sample. Only genes having at least one read count in any of the samples were retained. Differential expression analysis between TLR pathway mutated vs. wild-type, TLR pathway vs. control and wild-type vs. control samples was performed using DESeq2 [24].
The analysis for luciferase reporter assays was performed using the ANOVA test and the Tukey post-test in Graphpad Prism 6. Within Graphpad, the nonparametric Kruskal-Wallis test and Dunns post-test was used to compare means for ELISA assays. Kaplan-Meier curves for survival analysis were generated using the log-rank test. The Wilcoxon rank-sum test was used to compare the frequency of TLR mutant samples for microbiome exposed versus rarely exposed tumors in the COSMIC database. A significant p-value was defined as less than 0.05.
|
10.1371/journal.pcbi.1000901 | Similar Impact of CD8+ T Cell Responses on Early Virus Dynamics during SIV Infections of Rhesus Macaques and Sooty Mangabeys | Despite comparable levels of virus replication, simian immunodeficiency viruses (SIV) infection is non-pathogenic in natural hosts, such as sooty mangabeys (SM), whereas it is pathogenic in non-natural hosts, such as rhesus macaques (RM). Comparative studies of pathogenic and non-pathogenic SIV infection can thus shed light on the role of specific factors in SIV pathogenesis. Here, we determine the impact of target-cell limitation, CD8+ T cells, and Natural Killer (NK) cells on virus replication in the early SIV infection. To this end, we fit previously published data of experimental SIV infections in SMs and RMs with mathematical models incorporating these factors and assess to what extent the inclusion of individual factors determines the quality of the fits. We find that for both rhesus macaques and sooty mangabeys, target-cell limitation alone cannot explain the control of early virus replication, whereas including CD8+ T cells into the models significantly improves the fits. By contrast, including NK cells does only significantly improve the fits in SMs. These findings have important implications for our understanding of SIV pathogenesis as they suggest that the level of early CD8+ T cell responses is not the key difference between pathogenic and non-pathogenic SIV infection.
| Simian immunodeficiency viruses (SIV) are typically non-pathogenic in their natural hosts. However, if the same virus infects a non-natural host it often leads to AIDS-like symptoms. Therefore, comparing SIV infections in these two types of host might help explain the pathogenesis of SIV in non-natural hosts and thereby also that of HIV. We combined mathematical modeling with data on the levels of virus and immune cells early in infection, and compared both non-pathogenic SIV infections of sooty mangabeys and pathogenic SIV infection of rhesus macaques with respect to how the virus grows in them and to what extent it is controlled by the immune system. We found that the impact of the immune system on early virus replication is remarkably similar in both species. In particular, for both species virus replication can only be explained by the effect of CD8+ T cells. These findings have important implications for our understanding of SIV pathogenesis as they suggest that the impact of the early immune responses is not the key difference between pathogenic and non-pathogenic SIV infection.
| The simian immunodeficiency virus (SIV) occurs as a natural infection in several Old-world monkey species, such as sooty mangabeys (SM) or African green monkeys [1], [2]. In striking contrast to HIV infection of humans, SIV infection does not cause disease in natural hosts. The levels of virus replication, however, are similarly high in natural hosts and non-natural hosts such as rhesus macaques (RM), in which SIV causes AIDS-like symptoms. Comparative studies of SIV infection in natural and non-natural hosts provide the opportunity to investigate the interaction between the virus and the host immune system in pathogenic and non-pathogenic infection. Such a comparison might shed light on the mechanisms of disease progression in pathogenic SIV and by extrapolation on HIV.
Although natural and non-natural hosts allow similar levels of virus replication, there are interesting immunological differences: SMs do not exhibit the increased CD4+ T cell turnover and the generalized immune activation that is characteristic for the SIV infection of RMs or HIV-infection in humans [3], [4]. Thus, virus load alone cannot be the key to understanding pathogenesis. Silvestri and Feinberg [5] interpreted these findings in favor of the hypothesis that HIV disease progression is a result of generalized immune activation rather than of the destruction of CD4+ T cells by the virus alone. This view of HIV pathogenesis is a derivative of the immuno-pathological hypothesis [6]. Because primary HIV infection is a period critical for the future immune responses' capability of controlling the infection [7], [8], the potential differences between pathogenic and non-pathogenic SIV infection are likely to manifest themselves early in infection.
In both RMs and SMs, the early SIV infection is divided into three phases. The first phase is characterized by a sharp increase of virus load soon after infection. The second phase describes the decline of virus load that follows the initial peak viremia. The third phase finally describes the largely stable equilibrium virus load that eventually establishes after the decline. This stable virus load is also referred to as the viral set point. The characteristic pattern of virus load in primary SIV infection can be explained either through the delayed action of cellular immunity [9], [10] or through target cell limitation [11] or both. Note that in this context the term target-cell limitation refers to the hypothesis that the level of target cells on its own can explain the early virus-load dynamics [11]. Regoes et al. [9] investigated these hypotheses by fitting mathematical models to viral loads of SIVmac239-infected RMs that exhibited either normal or experimentally impaired cellular immunity as a result of co-stimulatory blockade. This analysis showed that target-cell limitation can explain the virus-load dynamics in the animals with impaired cellular immunity but not in those with a normal immune response. In the latter case, the models could explain the virus-loads only if cellular immunity is also taken into account. These results imply that target-cell limitation alone cannot explain the level of virus replication during primary SIVmac239 infection of RMs and thus suggest a role for cellular immunity in determining the post-peak decline of viremia.
In this article, we use the method of Regoes et al. [9] to analyze the early virus dynamics in non-pathogenic SIV infection of sooty mangabeys (SM). In particular, we sought to determine the roles that target-cell limitation, CD8+ T cell responses and NK cells play in primary infection of SMs, and to compare the impact of these factors with that in SIV-infected RMs. To this end we fit the measurements of virus load with population-dynamic models that differ as to whether they take factors such as cellular immunity or NK cells into account. Comparing the goodness of fit of these models, we can then evaluate the role of these factors in the primary infection of pathogenic and non-pathogenic SIV.
We used previously published data of experimental SIV infections (see Figure 1) to assess the relative importance of target-cell limitation, CD8+ T cells, and NK cells for controlling virus replication in primary SIV infection. To this end, we determined to what extent the ability of mathematical models to fit the early virus dynamics depends on the inclusion of these factors (see Figure 2). We start by showing that CD8+ T cells in combination with target-cells, but not target cells on their own, can explain the early SIV dynamics in RMs. Then we show that cellular immunity has a similar effect in early SIV replication of both RMs and SMs. Finally, we argue that NK cells only have an impact on the early replication in SMs.
The target-cell model aims to explain the virus-load dynamics through target-cell limitation only (equation 1), whereas the CD8+ T cell model takes both target-cell limitation and cellular immunity into account (equation 2). We use the density of proliferating CD4+ T-cells and of proliferating CD8+ T-cells as proxies for the size of the target cell population and for the strength of the specific cellular immunity. Comparing these two models assesses the relative role of target-cell limitation and cellular immunity in controlling the virus load: A good fit of the target-cell model and an only insignificant improvement in the CD8+ T cell model, would suggest that the virus load is mainly controlled by target-cell limitation. On the contrary, a bad fit of the target-cell model and a significant improvement in the CD8+ T cell model, would support the view that cellular immunity plays an important role.
The analysis of the RM data reconfirms the results of Regoes et al. [9] in an extended dataset. In particular, we find that target-cell limitation alone cannot explain the virus dynamics. For all animals except one (animal RPB8), the best fit of the target-cell model predicts a steadily increasing virus load (black lines in Figure 3), i.e. the fit fails to explain the characteristic peak and the subsequent post-peak decline exhibited by the data. Moreover, the quality of the fit is poor even for the animal for which the target-cell model can predict a viral load decrease. Adding specific cellular immunity to the target-cell model does significantly improve the fit for RMs (F-test, p = 2.8×10−18). Importantly, the CD8+ T cell model can explain the characteristic post-peak decline of the viral load (green lines in Figure 3).
The results of our analysis of the data from SIV infection of SMs are strikingly similar to those obtained for the rhesus macaques: The target-cell model fails to explain the virus dynamics for all eight animals (Figure 3), whereas the CD8+ T cell model provides a significantly better fit (F-test, p = 1.3×10−11), which can reproduce the qualitative patterns of the virus dynamics. The only exception is the animal FSS, for which both the target-cell and the CD8+ T cell model produce poor fits. The poor quality of these fits might be due to the fact that this animal exhibits a comparatively early increase of target-cell number and a comparatively late increase of CD8+ T-cell number (see Figure 1). The similarity of the results in SMs and RMs suggests that the relative importance of specific cellular immunity and target-cell limitation during early infection is comparable in pathogenic and non-pathogenic SIV hosts. In both cases, the temporal dependence of the viral load can only be explained if CD8+ T cells are taken into account.
Table 1 shows the best-fit estimates and the confidence intervals for the parameters of the CD8+ T cell model. The parameters r and k quantify the per-cell impact of target-cells and CD8+ T-cells on the viral replication rate (see equation 2). Both parameters are on average higher for sooty mangabeys: r roughly by a factor 6 and k by a factor 3. Furthermore, the intrinsic death rates of infected cells, a, were estimated to be 0 for most animals. This suggests that, for both SMs and RMs, most deaths of infected cells are caused by cellular immunity (see [9]).
The NK cell model and the CD8+ T cell & NK model are obtained from the target-cell and the CD8+ T cell model by adding NK cell number as an explanatory variable. We consider the fits of these extended models for two reasons: First, to test whether the above results are robust against adding NK cells to the model and, second, to investigate the role of an important effector mechanism of the innate immune system during primary SIV infection. In total, four types of statistical comparisons were performed (see Figure 2): Comparison i) between the target-cell model and the CD8+ T cell model is the one discussed above. Comparison ii) between the target-cell model and the NK model evaluates whether adding NK cells to target-cell limitation improves significantly the quality of fit. Comparison iii) between the NK model and the CD8+ T cell-NK model evaluates whether taking cellular immunity into account improves the fit of the NK model. Finally, comparison iv) assesses whether NK-cell number does significantly improve the fit of the CD8+ T cell model.
NK cell counts were available for 8 SMs (FWo, FYl, FWn, FFS, FRS, FSS, FUV, FWV) and 4 RMs (RPB8, RSO8, RYE8, RZS8). If the number of all NK cells is used as a proxy of NK cell activity, extending the target cell based model by NK cells (comparison ii) does improve the model fits significantly only for SM but not for RM (F-test, p = 0.016 and p = 0.24 for SM and RM, respectively). Extending the CD8+ T cell model by NK cells failed for both species to improve the model fits significantly (F-test, p = 0.98 and p = 0.33 for SM and RM, respectively). In contrast, extending the NK model by CD8+ T cells improves the fit significantly (F-test, p = 2.4×10−5 and p = 5.7×10−4 for RM and SM, respectively). If the number of proliferating NK cells is used as a proxy of NK cell activity, including NK cells again significantly improves the target-cell based model only for SM (F-test, p = 1.3×10−5 and p = 0.33 for SM and RM, respectively). In addition, including NK cell activity via this proxy also improves the CD8+ T cell model for SM (F-test, p = 0.00013 and p = 0.97 for SM and RM, respectively). These results suggest that NK cells play a role in the early infection of SM but not of RM.
The role of cellular immunity in early SIV/HIV infection has been a debated topic since the suggestion of Phillips [11] that early virus replication might be controlled by target-cell limitation. Several lines of evidence suggest however that cellular immunity is an important force for the control of early SIV replication. First, the post-peak decline of virus load coincides temporally with the rise of CTLs [12](although this is also consistent with the alternative explanation of [11]). Second, [10] have shown that the post-peak decline of virus-load is significantly weakened if CD8+ T-cells are depleted. Third, the ubiquitous selection for mutants that escape CTL response [13] also suggests an important role of cellular immunity. Fourth, it has been shown that the patients' ability to control HIV depends strongly on the alleles at the HLA and KIR loci [14], which control the action of CD8 T cells and NK cells, respectively. More recently, some of the authors of this paper [9] have shown that mathematical models can explain the early virus dynamics if they take both target-cells and CD8+ T-cells into account, but not if they take only target cells into account. Our study extends this previous work by considering the impact of NK cells, important effectors of innate immunity. In addition to the extended analysis of the early viral dynamics in pathogenic SIV infection, we here compare our results to non-pathogenic SIV infection in sooty mangabeys (SMs). This comparison has important implications for our understanding of pathogenesis.
Our analysis confirms the earlier finding of [9] that target-cell limitation alone cannot explain the virus dynamics in RMs. We find that, in SIV-infected sooty mangabeys, target-cell limitation is equally unable to explain the viral load dynamics during early infection. In both species, our model can only explain the virus dynamics if it takes cellular immunity into account. This suggests that specific cellular immunity plays an important role in determining the dynamics of virus replication during early infection in both species. We, however, also found that a model, which assumes a constant viral replication rate, independent of target cells, was unable to fit the virus-load data of all animals consistently (results not shown). This implies that, although target cells alone cannot explain the virus-load dynamics, in particular the peak and the post-peak decline, temporal variation of target cells is nevertheless important. Overall, our results indicate that the relative impact of target-cell limitation and specific cellular immunity is similar in RMs and SMs.
These results give rise to testable predictions. If, for example, one would selectively deplete NK cells during primary infection, the pattern of virus load should be affected in SM, but not in RMs. In contrast, selective depletion of CD8+ T cells is predicted to lead to a loss of control of virus replication in both species. Of note, all depletion experiments performed using an anti-CD8 antibody depleted CD8+ T cells as well as NK cells because both cell types express CD8 [10]. In RMs, treatment with a costimulatory inhibitor, which prevented the development of SIV-specific cellular and humoral immunity and reduced target cell levels, gave rise to target cell limited virus replication [9]. The similarity between the factors governing virus replication predicts that an analogous treatment of SMs would also lead to target cell limitation.
Our conclusions about the role of cellular immunity and target-cell limitations are based on several assumptions. First, the virus loads and the immune-cell densities were measured in the blood, which is not the main compartment of SIV replication and lymphocytes. Our analysis, therefore, relies on the assumption that the measurements in the blood reflect the situation in the whole body. In this context, it has been suggested that target-cell depletion in the gut might play an important role in the early SIV infection [15], [16], [17], [18], [19], [20], [21], [22]. However, a recent study has shown that in SIV infections of both SM and RM, the target-cell depletion in the gut occurs too early to explain the peak in virus-load [23]. Second, our models consider only the primary phase of SIV infection. Therefore, our conclusion that cellular immunity does not differ in pathogenic and non-pathogenic SIV, does only apply to this phase. It might thus be that cellular immunity at later phases plays a very different role in RMs and SMs, as suggested by numerous comparative studies [2], [3], [24], [25]. As discussed in Regoes et al. [9], it is difficult to extend the approach used here to later phases of infection, because immune-escape and antibody responses would require considerably more complicated models. Last, we cannot exclude that different cell compartments or cell types play the role of target cells in the SIV infections of sooty mangabeys and rhesus macaques. Indeed, our model fits result in larger replication rate constants, r, for SM than for RM, which either suggests a better target cell utilization in SM, or is an indication that Ki67+ CD4+ T cells do not play the same roles in SM and RM. Such an effect could systematically bias our analysis if our proxy (i.e. proliferating CD4+ cells) would be representative for target cells in one species but not in the other. Finally, the p values of the model comparisons rely on the assumptions of normality and independence, which might be violated in our data. Especially, autocorrelation in the virus-load and cell-numbers, might potentially lead to an overestimate of the degrees of freedom and thereby to an underestimate of those p-values. However, it should be noted that independently of the statistical evaluation, the least-squares approach is a simple and intuitive method to fit dynamical models to data, and these fits clearly (Figure 3) show that for all animals except one RM (RPB8), the best fit of the target-cell-limitation model fails to predict a post-peak decrease in virus-load. This suggests that our results regarding the CTLs are robust against these (in principle valid) statistical concerns. By contrast, adding NK cells to the model leads to smaller improvements of the fits and therefore these findings may be more vulnerable to potential autocorrelations.
One important caveat mentioned in the previous section is the uncertainty as to whether the measured cell populations (e.g. Ki67+ CD4+ T-Cells, Ki67+ CD8+ T-Cells, NK cells) can be identified with populations performing a specific function (target cells, cytotoxic T cells, cytotoxic NK cells). This potential problem is substantially alleviated by the way these measurements are integrated into our model. Specifically, the quality of fit as measured by the residual sum of squares, is invariant with respect to a linear transformation of the variables. I.e. if we measure the cell population x but the active population is x′ = a x-b we will obtain the same quality of fit regardless of whether we incorporate x or x′ into our model. Therefore it does not matter whether only a fraction of the measured cells is active or whether a constant number of the measured cells is inactive. For practical reasons, however, it is important that the fraction of the active cells is not too small relative to the inactive cells, because then the noise in the latter is likely to overwhelm the signal in the former. This reasoning implies that the comparison of the quality of fit of the different models (Figure 2) is much more robust than the parameter estimates (Table 1): In principle, the first type of analysis (model comparison) still works, even if the linear transformation (relating measured cell populations to the cell-populations performing a specific function) is different for each animal. By contrast, the second type of analysis (parameter estimation) requires that this transformation is similar in the animals compared. For these reasons, we conclude that not too much weight should be given to the parameter estimates, as they rely much stronger on a good match between measured cell populations and the populations actually performing a certain function, while we can assert that the model comparison is robust. The fundamental robustness of the method also explains why [9] found qualitatively similar results with Ki67+ CD8+ T-cells and tetramer positive T-cells as markers for SIV-specific cellular immunity.
As SIV infection is pathogenic in rhesus macaques but non-pathogenic in sooty mangabeys, our results can be interpreted in the context of current theories of SIV pathogenesis, in particular with respect to reasons underlying the absence of disease progression in SIV-infected SMs. While an initial study suggested that acute SIV infection of SMs is characterized by limited to absent T cell activation [25], a number of more recent studies that included a more comprehensive sample collection have shown very clearly that SMs exhibit substantial T cell activation during acute SIV infection [24], [26], [27], [28]. However, in marked contrast with SIV-infected RMs, sooty mangabeys are able to rapidly and dramatically reduce the level of T cell activation during the early chronic infection (i.e., starting at day 30 post inoculation) [24], [26], [27], [28]. Although our model comparison did not directly test differences in the antigenicity of SIV between SM and RM, our results are more consistent with the latter observations and suggest that the divergent outcome of SIV infection in RMs and SMs is not caused by differences in CD8+ T-cell response during the early stages of infection.
All the experiments on non-human primates from which these data are sampled have been approved by the Institutional Animal Care and Use Committee (IACUC). All these experiments have been described in previous publications.
The data analyzed in this article were generated in experimental infections of SMs infected with the viral strain SIVsmm and of rhesus macaques infected with the strains SIVmac (animals rbm, rvy, roz, ryt) or SIVsmm (animals RPB8, RSO8, RYE8, RZS8, RFT8). A detailed description of the experiments can be found in Garber et al. [29], Gordon et al. [30], and Mandl et al. [4]. For the sake of comparability, we consider the same time-window as Regoes et al., i.e. a window ranging from day 0 (start of infection) to day 30. In one of the rhesus macaques (animal RFT8) no SIV infection could be established. This animal was therefore excluded from further analysis. In total, we consider 8 SMs (all infected with SIVsmm) and 8 RMs (4 infected with SIVmac239 and 4 infected with SIVsmm).
Figure 1 shows the measurements relevant for this study: the virus-load, the density of proliferating CD4+ T-cells, the density of proliferating CD8+ T-cells, and the density of NK cells. The fraction of proliferating CD4+ and CD8+ T cells was assessed by staining for the nuclear antigen Ki67, which is expressed by cycling cells. We consider the density of proliferating CD4+ T-cells as representative for the size of the target cell population and the density of proliferating CD8+ T-cells as a surrogate measure for the SIV-specific cellular immunity. We will therefore refer to the density of proliferating CD4+ T-cells and of proliferating CD8+ T-cells also as “target cells” and “cellular immunity”, according to the functional role we assume these populations to play. Data on the density of NK cells were only available for all sooty mangabeys and for 4 out of 8 rhesus macaques (RPB8, RSO8, RYE8, RZS8).
The data were analyzed by using population dynamic models, which describe the virus dynamics as a function of target cells, CD8+ T-cells, and NK cells. The models are fitted to the virus load. Hereby, the measurements of target cells, CD8+ T-cells, and NK cells were used as explanatory variables. Importantly, the model does not aim to explain the measurements of these cell populations, but considers them only as factors that might explain viral replication. A detailed account of this approach can be found in [9].
In order to assess the role of target-cell limitation and cellular immunity in early SIV infection, we compared the fits of two nested models, which describe the virus dynamics by taking into account either target cells only or target cells and specific cellular immunity. These models are referred to as the target-cell model and the CD8+ T cell model, respectively. Mathematically, these models read(1)(2)where v is the virus load and T(t) and E(t) denote the number of proliferating CD4+ T-cells and of proliferating CD8+ T-cells, respectively. The parameters r, a, and k are chosen for each animal such that T(t) and E(t) give the best possible prediction of v (see below).
In order to test the impact of the non-adaptive immune system on our results, we extended the above models by adding NK cell number as an explaining factor. We incorporate the impact of NK cells by using two different proxies: either the total density of NK cells (characterized as CD3− CD20− CD16+ cells) or only the density of activated NK cells (i.e. Ki67+ NK cells). The second approach is identical to the one used of CD8+ and CD4+ T-cells. The first approach can be justified by the fact that, in contrast to CD8+ T-cells, NK cells do not recognize specific antigens. Thus, every NK cell can potentially inhibit virus replication by either killing infected cells or by IFN-gamma production [31], and their effect is most likely proportional to their level. We would like to emphasize that we do not assume that every NK cell is cytotoxic, or that every NK cell has anti-viral activity. We only assume that the impact of NK cells is proportional to their abundance (see discussion). The extensions of the target-cell model and the CD8+ T cell model are referred to as the NK-model and the CD8+ T cell & NK model. Mathematically these models read(3)(4)where N(t) denotes the number of NK cells and the parameter n is chosen according to the best fit criterion.
We illustrate the fitting-procedure for the CD8+ T cell & NK model: First the differential equation of the model (4) can be integrated to(5)If t0…tk , denote the time points for which measurements of v are available then the parameters r, k, a and n are chosen such that the residual sum of squares(6)is minimized. The integrals in the sum are computed from the measurements of the cell numbers T ,E, and N by first interpolating these measurements by a piecewise linear function, resulting in the functions T(t), E(t), and N(t), and then integrating these interpolating functions. As expression (5) is linear in the parameters r, k, a and n, the best fit can be found using a standard linear-model solver such as the lm() routine of the R language [32]. Biologically, the parameters r, k, a and n must be larger than or equal to 0. If the best fit of (5) does not fulfill these conditions, one or several of the parameters r, k, a and n is set to 0 and the fitting procedure is repeated with these reduced functions. From all the “reduced fits”, that one is chosen, which yields the minimal sum of squares while fulfilling the biological conditions.
The fits for the target-cell, the CD8+ T cell, and the NK model are obtained in a similar way as for the CTL-NK model. In formula (5) the parameters that do not occur in the differential equation of the model (i.e. equation 1, 2, or 3 for the target-cell, CD8+ T cell, and NK model respectively) are set to 0 and the remaining parameters are chosen such that the corresponding sum of squares (SSQtarget-cell, SSQCD8+ T cell, and SSQNK) is minimized.
We can statistically compare two of the above models, for instance model 1 and model 2, if they are nested, i.e. if model 1 results from model 2 by setting one of the parameters to 0. In such cases, model 2 will always provide a better fit than model 1, because model 1 is included as a special case in model 2. Whether this improvement in the quality of fit is significant can then be assessed by performing an F-test. The corresponding test statistic is
Here SSQi denotes the residual sum of squares of the model i, and dfi refers to the corresponding degrees of freedom. The p value that corresponds to the value of F is then calculated from the Fisher Distribution with degrees of freedom df1-df2 and df2, i.e F(df1-df2, df2). This comparison between models can be made either for each animal individually, or, as we mostly do in this article, for all animals of a species taken together. In the latter case, the residual sum of squares obtained by fitting the models to each animal and their corresponding degrees of freedom have to be summed to perform the F-test.
Figure 2 illustrates the statistical comparisons that are made in this article. The most important of these comparisons is the one between the target-cell model and the CD8+ T cell model (comparison i in Figure 2), which assesses the relative importance of target cells and specific cellular immunity for explaining the virus-load dynamics. If NK-cell counts are available, one can ask in addition whether taking NK cells into account improves the fit of the target-cell model (comparison ii), whether taking specific cellular immunity into account improves the fit of the NK model (comparison iii), and whether taking NK cells into account improves the fit of the CD8+ T cell model (comparison iv).
|
10.1371/journal.pntd.0007216 | Determinants for progression from asymptomatic infection to symptomatic visceral leishmaniasis: A cohort study | Asymptomatic Leishmania donovani infections outnumber clinical presentations, however the predictors for development of active disease are not well known. We aimed to identify serological, immunological and genetic markers for progression from L. donovani infection to clinical Visceral Leishmaniasis (VL).
We enrolled all residents >2 years of age in 27 VL endemic villages in Bihar (India). Blood samples collected on filter paper on two occasions 6–12 months apart, were tested for antibodies against L. donovani with rK39-ELISA and DAT. Sero converters, (negative for both tests in the first round but positive on either of the two during the second round) and controls (negative on both tests on both occasions) were followed for three years. At the start of follow-up venous blood was collected for the following tests: DAT, rK39- ELISA, Quantiferon assay, SNP/HLA genotyping and L.donovani specific quantitative PCR.
Among 1,606 subjects enrolled,17 (8/476 seroconverters and 9/1,130 controls) developed VL (OR 3.1; 95% CI 1.1–8.3). High DAT and rK39 ELISA antibody titers as well as positive qPCR were strongly and significantly associated with progression from seroconversion to VL with odds ratios of 19.1, 30.3 and 20.9 respectively. Most VL cases arose early (median 5 months) during follow-up.
We confirmed the strong association between high DAT and/or rK39 titers and progression to disease among asymptomatic subjects and identified qPCR as an additional predictor. Low predictive values do not warrant prophylactic treatment but as most progressed to VL early during follow-up, careful oberservation of these subjects for at least 6 months is indicated.
| Visceral Leishmaniasis (VL) or Kala-azar is a vector born disease, deadly if not treated. On the Indian subcontinent VL is caused by the protozoan parasite Leismania donovani, transmitted by an insect vector, sand fly of the Phlebotomus argentipes species, and considered an anthroponotic disease. Not every L.donovani infection progresses to clinical VL disease, and only a small minority of those infected will progress to disease and therefore not all those infected need to be treated. Importantly, diagnostic and treatment options have considerably improved over the past 10 years. There are several markers of infection in VL: antibody-tests as DAT, rK39-ELISA, markers of cellular immunity as the Quantiferon assay, and molecular markers as quantitative PCR. Also SNP/HLA genotyping has been shown to be associated with VL. However the factors that determine who will and who will not progress from infection to disease remain largely unknown. To try and elucidate the factors associated with progression to disease we identified a cohort of healthy recently infected persons in a highly VL endemic area of Bihar, India, and followed them up for three years. We also included in the follow-up an age and village matched group of initially seronegative controls. SNP/HLA genotyping was performed on all subjects to identify genetic predisposition. The only factors strongly and significantly associated with progression to disease turned out to be high DAT and/or rK39 titers and positive qPCR. The proportion progressing to disease was too low to merit preventive treatment. As most disease tends to occur early during follow-up, it is recommendable to closely follow up those with high antibody titers or testing qPCR positive over at least a 6-months period.
| Visceral leishmaniasis (VL) or kala-azar is the severest form of leishmaniasis and fatal if left untreated. More than 90% of global VL cases occur in just six countries: India, Bangladesh, Sudan, South Sudan, Brazil and Ethiopia [1]. India accounts for approximately 50% of the global burden of VL and is a signatory to a Tripartite Memorandum of Understanding (MoU) to achieve VL elimination from the South-East Asia Region (SEAR). The goal is to reduce the annual incidence of VL to less than 1 case per 10,000 population at the sub-district (block) level [2, 3]. This elimination target is expressed as a number of new clinical cases of VL per person-year. However, it is established that many L. donovani infections do not lead to a clinical episode of VL and that asymptomatic infections far outnumber the clinical cases [4]. A prospective study in India and Nepal showed a ratio of incident asymptomatic infection, measured by recent conversion in antibody tests, to clinical disease of 9 to 1 while in neighboring Bangladesh it was 4 to 1 [5, 6]. Mathematical modeling has suggested that transmission of L.donovani could be maintained by asymptomatically infected hosts [7, 8]. Therefore the study of asymptomatic infection is considered a key research priority to support the VL elimination initiative [9]. Currently, xenodiagnosis studies are ongoing in India and Bangladesh to establish whether asymptomatic carriers of L.donovani infection are infectious to sand flies, but there are other issues to address as well.
Mathematical modeling has suggested that detecting and treating clinical cases early enough is key to reducing their transmission potential [10, 11] and therefore, it would be useful if one could identify the infected persons who are most likely to progress to clinical disease. To date it is not established which are the best predictors for the development of active VL disease in somebody with a positive leishmanial infection marker but no signs and symptoms. A strong association has been observed between high baseline antibody titers and progression to VL in the subsequent 36 months in large cohort studies in India, Nepal and Bangladesh [12, 13]. In the above mentioned studies in India and Nepal, even stronger associations for progression to VL were observed with recent seroconversion to high antibody titers.
Whether an infection remains asymptomatic or progresses towards VL probably results from the complex interaction between genetic susceptibility and immune response of the host, combined with parasite, socioeconomic and demographic factors. A genome-wide association study (GWAS) carried out in India showed that HLA class II alleles, in particular, HLA-DRB1, are major genetic risk factors for VL. Sequence-based classical HLA typing and haplotype analysis suggest that risk allele(s) in India belong to HLA-DRB1*13/*14 allele groups and protective alleles to HLA-DRB1*15 allele group [14]. However, the relative contribution of these different factors to the development of VL is still not well understood. We assessed immunological and genetic markers for progression to active clinical disease in a large prospective cohort study in Bihar, India. Our aim was to identify markers that are predictors of progression from L.donovani infection to clinically symptomatic VL.
We conducted this prospective study in two high VL incidence areas of Muzaffarpur district, Bihar State, India from 2008–2015.
The review committee of the U.S. National Institutes of Health (NIH), as well as the Institutional Review Boards of the Institute of Medical Sciences, Banaras Hindu University, Varanasi, India, Institute of Tropical Medicine, Belgium and the University of Iowa reviewed the study protocol and gave ethical clearance for this study. Data was anonymized. All subjects provided written informed consent; in case of illiterate subjects, a thumb print plus a signature of an independent witness was obtained. For minors under the age of 18 years, informed consent was obtained from a parent or guardian.
The study was conducted in two areas. The first area (Area-1) had a total population of 19,634 divided over 11 villages with high VL incidence rates before 2009. Two house to house surveys were conducted in Area-1at a one-year interval between December 2009 and February 2011. All residents above two years of age who were present and gave their informed consent (or whose parents gave consent, for minors) were enrolled in the study. A capillary blood sample was obtained on pre-printed Whatmann filter paper in consenting participants and rK39-ELISA, and DAT tests were performed to detect antibodies against VL. Those individuals testing negative for both the tests in the first sero-surveywere re-tested in the second serosurvey in the following year. We defined seroconverters as subjects negative on rK39-ELISA and DAT in the first sero-survey but positive on either of the two assays during the second survey. For each seroconverter a control who was rK39-ELISA and DAT-negative on both survey rounds was recruited into the study, controls were group matched according to age (<10, 10–18 or >18 years) and village of residence.
An interim analysis in 2010 showed that, due to a declining incidence trend of VL, the target sample size of 600 seroconverters could not be achieved [12]. Therefore we selected an additional study area (Area 2) of 10,729 population living in 1,836 households in 16 geographically scattered villages. Villages were selected based on reported recent high VL incidence levels. In this area two similar serosurveys were conducted six months apart to recruit more seroconverters. This time we recruited four controls for each seroconverter, based on the same matching criteria [12, 15].
All seroconverters and their controls were interviewed to gather baseline demographic and medical history data, and were clinically examined. At the time of recruitment 5 ml of blood was obtained from seroconverters and controls for the following tests: DAT, rK39-ELISA, Quantiferon assay (IFN-γ release assay), SNP/HLA genotyping and quantitative PCR (qPCR). Both the seroconverters and controls were followed up monthly for the development of clinical symptoms of VL. In case of clinical suspicion (i.e. more than 2 week fever history and rK39 RDT positivity), VL was confirmed parasitologically by splenic smear and treated with Amphotericin B as per national guideline recommendation [16]. All the participants were followed up for a minimum of 3 years.
As a quality control measure for each seroconverter identified both the original sample and the sample of the follow-up survey were rerun on the same plate for rK39 ELISA as well as for DAT. Subjects who were intially classified as seroconverters, but for whom quality control serologic testing disagreed with the initial baseline negative or follow-up survey positive results, were kept in the cohort but were reclassified as controls. The assessment of the association between seroconversion and disease was done comparing the final validated set of seroconverters to all controls, as well as by comparing the final set of seroconverters to only the original controls.
To determine the probability of progression to disease as a function of baseline status for various markers, we calculated odds ratios and confidence intervals using logistic regression. The factors on which converters and controls had been matched, i.e. village of residence and age group, were included in the models. Persons who developed VL before their inclusion in the cohort were excluded from the main analysis. For baseline DAT and rK39 results, we constructed Kaplan-Meier survival plots after subdividing both markers into three categories. DAT titers were regrouped on a 0 to 8 scale, each step representing an increase of one titer step from undiluted (0), via 1:400 (1) up to 1:25,600 or above (8). As cut off for being labeled DAT positive we chose a titer of 1:1600 or above based on our prior studies of subjects in this area [12]. For further analysis we defined three categories, DAT negatives (titer < 1:1,600), moderately DAT positives (titer ≥1:1,600 but < 1:25,600) and strongly DAT positives (titer ≥ 1:25,600). For rK39 ELISA we used percentage points optical density. As we previously defined for subjects in the region, above 14 percentage points was considered positive. For further analysis we divided subjects into three categories based on percentage points (pp) optical density. Titers of ≤ 14 pp were considered negative, >14 pp up to ≤40pp was considered moderately positive, above 40 pp as strongly positive.
Altogether 1,606 subjects were enrolled, including 476 seroconverters and 1,130 controls. Among the 1,130 controls, 79 were originally classified as seroconverters but re-classified after quality control. Altogether 978 subjects (61%) were female; the proportion was the same among converters and controls. The youngest subjects were two years of age; the oldest was 88 years. The median age of the group of seroconverters was 25 years as compared to 24 years for controls.
Over an average 52 months of follow-up, 17 persons developed VL, eight in the group of seroconverters and nine among controls, resulting in an odds ratio of 3.1 (95% CI 1.1–8.3). Most cases arose early during follow-up, with a median follow-up duration of 5 months and a maximum of 50 months. Fifteen out of 17 cases occurred in area 2, the area with the highest reported incidence at the time of the baseline survey. The association between seroconversion and progression to disease was only observed in this region (OR 3.5, 95% CI 1.2–9.8), whereas in Area 1 there was no association, the odds ratio was 1.3 (95% CI 0.08–21.1).
Fig 1 shows the probability of progressing to VL independent from the case/control group status. When analyzing all subjects together–seroconverters as well as non-seroconverters-there was a strong association between DAT at the start of follow-up and progression to disease. In the persons with high DAT antibody titers, 8 out of 77 subjects (10.4%) developed VL, all within nine months of follow-up, resulting in an odds ratio of 19.1 (95% CI 4.4–57.1) when compared to DAT negatives. In the latter category only 8 out of 1,175 (0.7%) developed VL.
For rK39 ELISA, findings were similar as for DAT. Out of 44 subjects belonging to the high titer category, seven (15.9%) developed VL, all within nine months, resulting in an odds ratio of 30.3 (95% CI 9.6–95.2) when compared to rK39 ELISA negatives (Fig 2). In the latter category only 9 out of 1,416 (0.6%) developed VL.
Among 1,579 subjects tested with Quantiferon (IGRA assay) at the start of follow-up, 280 (17.7%) tested positive. Five out of sixteen VL cases occurred in this group, resulting in an odds ratio of 1.8 (95% CI 0.6–5.3).
Out of 1,604 subjects tested with qPCR, 68 were positive, i.e. exceeded the—1 parasite genomes/ml of blood. Out of those, six(8.8%) developed VL, compared to 11 out of 1,536 (0.7%) among qPCR negatives, resulting in an odds ratio of 20.9 (95% CI 6.5–66.8). All six progressed to disease within four months, four out of six even progressed within two weeks.
Of 957 subjects subjected to SNP genotyping, 380 (39.7%) were homozygous without the protective allele, 442 (46.2%) were heterozygous, and 135 (14.1%) were homozygous for the protective allele. With the first category (homozygous without the protective allele) as reference category, we found odds ratios of 0.61 (95% CI 0.2–1.8) and 0.62 (OR 0.13–3.1) respectively for the second and third category.
When looking at combinations of the three markers that were strongly associated with progression to disease, i.e. high DAT titers, high rK39 titers and qPCR positivity, we observed that relatively little gain in sensitivity was achieved by combining tests. Results are shown in Table 1. A high titer DAT identified 8 out of the 9 cases that were identified by combining the three markers.
The strong overlap between baseline high DAT titers, high ELISA titers and qPCR positivity among incident VL cases is also apparent from Table 2 below. This table also shows that for each of these markers VL cases among positives arose early during follow-up, an observation that was already visible in the Kaplan-Meier graphs. All cases among subjects that were not highly DAT and/or ELISA positive arose only after a minimum delay of six months. Among qPCR negatives the picture was similar though there was one exception of a case arising after just two months of follow-up.
Our data show very strong associations between being qPCR positive at baseline and subsequent progression to VL (OR 20.8, 95% CI 6.5–66.8), the same applies to having a high DAT titer (OR 19.1, 95% CI 4.4–57.1) or a high rK39-ELISA titer (OR 30.3, 95% CI 9.6–85.2). There was only a moderately strong association between seroconversion and progression to disease, (OR 3.1, 95% CI 1.1–8.3). There was no significant association between progression to disease and positivity in the IGRA test (OR1.8, 95% CI 0.6–5.3). SNP/HLA genotyping showed a trend towards a protective effect of the genes tested, but the association was weak and non-significant. Both heterozygous and homozygous individuals for the protective variants had lower odds of disease when compared to individuals without the protective variants with odds ratios of 0.61 (95% CI 0.2–1.8) and 0.62 (OR 0.13–3.1) respectively.
This study corroborates our previous findings of a strong association between high DAT and rK39 titers and subsequent progression to disease, as well as the findings by Chapman and others in Bangladesh [13, 15].
The main strength of our study is that we followed up a cohort of seroconverters on DAT and rK39-ELISA over a relatively long period, three years, and included a number of other potential markers of infection. We also performed SNP/HLA genotyping. One of the major difficulties in studies of “asymptomatically infected” is that there is no clear consensus about the case definition. L.donovani infection status can be measured by antibody, antigen or nucleic acid detection or else by markers of cellular immunity in combination with a clinical assessment of signs and symptoms. Asymptomatically infected persons have been defined in various studies as those who show no clinical signs or symptoms of VL but are positive in at least one marker of infection such as the Leishmanin Skin Test (LST), a marker of cell-mediated immunity [24–26]; an antibody detection test as the DAT, rK39 ELISA, or IFAT [5, 6, 12, 27], or a molecular marker as qualitative or quantitative PCR to detect Leishmania spp. DNA [5, 28–30]. Given the lack of agreement between these infection markers when measured cross-sectionally, the case definition of “asymptomatically infected” is a recurrent matter of discussion [31].
Medley et al. [11] pointed out how important early diagnosis of VL is, as it has an impact on individual prognosis as well as on curtailing transmission. Individuals with a high probability of developing clinical disease might be treated sooner if given an intense follow-up scheme. Based on our data it is, therefore, tempting to promote the high-titre DAT, high–titer rK39 ELISA, and qPCR as markers to identify persons at high risk for clinical VL. However, the relatively low positive predictive value of the markers (ranging from 8.8% for qPCR over 10.4% for high DAT titres and 15.9% for rK39) warrants a word of caution. As there is no easy and absolutely safe treatment available, an attitude of watchful waiting is probably best at this time, observing those persons closely to detect the first clinical signs early enough. If one decided to treat all the qPCR positives or all the high titre DAT positives, about nine out of ten persons treated would be treated without reason, while for the rK39 this amounts to 8 out of ten. Combining tests does not add much to sensitivity. It should also be noted that incident VL cases that were missed at baseline, i.e. that did not show high-titre rK39 and DAT, did not arise until 6 months after the start of follow-up; half of them arose only after 18 months of follow-up or later. These subjects may well have been infected during the follow-up period. Importantly, recent reports from Bangladesh indicate that incorporating rKR95 and rTR18 with rK39 in serological tests conferred a sensitivity of 84% and may enable simple and accurate detection of asymptomatic infection in surveillance [32].
Finally, the potential epidemiological importance of the group of people with at least one positive marker of infection but no symptoms remains elusive. Are they all truly “infected”–i.e. latent carriers of L.donovani—with potential for transmission of the parasite or is this, more plausibly, a mixed group of i) very recently infected persons, ii) established latent carriers and iii). Immune persons who cleared their infection? It is hoped that the xenodiagnosis studies will shed some light on this question in the near future.
In conclusion, healthy persons living in VL endemic areas who have high antibody titers or test positive to qPCR have an increased probability to progress to VL disease. Such probability is not high enough to merit prophylactic treatment but carefull follow-up is warranted as most of those who do progress to disease eventually do so within the first 6 months.
|
10.1371/journal.pbio.0050207 | Proboscidean Mitogenomics: Chronology and Mode of Elephant Evolution Using Mastodon as Outgroup | We have sequenced the complete mitochondrial genome of the extinct American mastodon (Mammut americanum) from an Alaskan fossil that is between 50,000 and 130,000 y old, extending the age range of genomic analyses by almost a complete glacial cycle. The sequence we obtained is substantially different from previously reported partial mastodon mitochondrial DNA sequences. By comparing those partial sequences to other proboscidean sequences, we conclude that we have obtained the first sequence of mastodon DNA ever reported. Using the sequence of the mastodon, which diverged 24–28 million years ago (mya) from the Elephantidae lineage, as an outgroup, we infer that the ancestors of African elephants diverged from the lineage leading to mammoths and Asian elephants approximately 7.6 mya and that mammoths and Asian elephants diverged approximately 6.7 mya. We also conclude that the nuclear genomes of the African savannah and forest elephants diverged approximately 4.0 mya, supporting the view that these two groups represent different species. Finally, we found the mitochondrial mutation rate of proboscideans to be roughly half of the rate in primates during at least the last 24 million years.
| We determined the complete mitochondrial genome of the mastodon (Mammut americanum), a recently extinct relative of the living elephants that diverged about 26 million years ago. We obtained the sequence from a tooth dated to 50,000–130,000 years ago, increasing the specimen age for which such palaeogenomic analyses have been done by almost a complete glacial cycle. Using this sequence, together with mitochondrial genome sequences from two African elephants, two Asian elephants, and two woolly mammoths (all of which have been previously sequenced), we show that mammoths are more closely related to Asian than to African elephants. Moreover, we used a calibration point lying outside the Elephantidae radiation (elephants and mammoths), which enabled us to estimate accurately the time of divergence of African elephants from Asian elephants and mammoths (about 7.6 million years ago) and the time of divergence between mammoths and Asian elephants (about 6.7 million years ago). These dates are strikingly similar to the divergence time for humans, chimpanzees, and gorillas, and raise the possibility that the speciation of mammoth and elephants and of humans and African great apes had a common cause. Despite the similarity in divergence times, the substitution rate within primates is more than twice as high as in proboscideans.
| An accurate and well-supported phylogeny is the basis for understanding the evolution of species. With the appropriate and adequate amount of data, it is possible not only to determine relationships among species, but also to date divergence events between lineages. In turn, divergence events can be correlated to environmental changes recorded in the fossil record to help understand mechanisms driving evolution. The power of these correlations, and arguments for particular environmental mechanisms driving evolutionary change, increases when the pattern is repeated across multiple taxa.
Sequencing complete mitochondrial genomes has become a powerful tool in the investigation of phylogenetic relationships among animal groups, principally mammals, birds, and fishes. Despite this potential, and the proliferation of ancient mitochondrial DNA (mtDNA) research, geneticists so far have succeeded in sequencing complete mitochondrial genomes from ancient DNA for only a few extinct species of moas [1,2] and the woolly mammoth [3,4]. For both groups, the complete sequences resolved long-standing evolutionary questions, which argues for an extension of such analyses to other species.
The living elephants comprise the last survivors of the Elephantidae, a once-flourishing sub-group of the Order Proboscidea that lived throughout much of Africa, Eurasia, and the Americas [5]. The evolution of Elephantidae, which includes the recently extinct mammoths, has been extensively studied in recent years using both modern and ancient sequences of mtDNA and nuclear DNA (nuDNA). For example, nuDNA sequences have been used to argue that the African forest elephant is a valid species (Loxodonta cyclotis), distinct from the African savannah elephant (L. africana) [6], with the two species having diverged by 2.6 million years ago (mya), though this view is disputed [7].
Two recent studies reported complete mtDNA genomes from the woolly mammoth (Mammuthus primigenius) [3,4] that provided strong evidence that mammoths were more closely related to Asian elephants than to African elephants. However, another study analyzing several nuDNA segments cautioned that the lack of a closely related outgroup is a problem in phylogenetic analyses of Elephantidae and argued that the relationship between mammoth and the living elephants is still unresolved [8]. To date, all genetically based analyses of elephant phylogenies have used dugong (Dugong dugon) and hyrax (Procavia capensis) as outgroups. These are the nearest living relatives of elephants, but they diverged from proboscideans some 65 mya [5,9], which severely limits their power as effective outgroups for assessing elephant genetic data.
In contrast, ancestors of the American mastodon, Mammut americanum, diverged from the Elephantidae lineage no earlier than 28.3 mya [10], which would make mastodon a much more appropriate outgroup for the Elephantidae. Moreover, mastodon fossils preserved in permafrost and dating to the late Pleistocene have been recovered in eastern Beringia (Alaska and Yukon). Their young age and preservation in permafrost means Beringian mastodon are excellent candidates for ancient DNA analyses [11,12]. The good biochemical preservation of mastodon samples was appreciated early in the study of ancient DNA. In 1985, Shoshani and colleagues determined the immunological distances of albumins among mammoth, mastodon, and African and Asian elephant [13].
We report here what is to our knowledge the first complete mitochondrial genome for mastodon. The sequence was derived from a tooth collected in northern Alaska, where Pleistocene bones are well-preserved in permafrost. Collagen from the root of the tooth was radiocarbon-dated to more than 50,000 y before present (BP), but the geological provenance suggests it is not older than 130,000 y (see Materials and Methods). This extremely old and complete sequence is of interest in its own right, but it also can be used to help resolve existing debates concerning the phylogeny of the Elephantidae and allows us to evaluate rates of molecular evolution within the Proboscidea, because it finally provides an appropriate outgroup for such analyses. Here, we use the mastodon mtDNA sequence to resolve relationships among African elephants, Asian elephants, and mammoths and to more accurately date their divergence times.
To determine the mastodon mtDNA genome sequence, we used 78 primer pairs, separated into two sets of 39 pairs each, on DNA extract from a mastodon molar originating from northern Alaska (Figure 1) and performed quadruplicate amplifications for each primer pair. In the first round of amplification, 48 primer pairs yielded at least one positive result, and for 36 primer pairs all four attempts were positive. To obtain the remaining fragments, we used four consecutive rounds of redesigning primers (see Protocol S1 for details). In each round, whenever possible, we used the mastodon sequences from flanking fragments as the basis for primer design. With three primer pairs, PCR controls or extraction controls yielded products of the correct size. All three cases of contamination occurred with primers for which it was not possible to design the primers in a way that they selected against the amplification of human DNA. The resulting products were cloned and sequenced and turned out to be exclusively of human origin. Two out of six clones obtained from a single amplification from mastodon extract using one of these primer pairs were also of human origin. We did not replicate part of the mastodon sequence in an independent laboratory, following the argument we have made previously [14], and in contrast to Cooper and Poinar [15], who argued that this is necessary in all ancient DNA studies. None of the determined sequence fragments was identical to any known sequence either from GenBank or determined previously in our laboratory. Moreover, to ensure authenticity of each individual fragment, apart from extensive internal replication, we compared each PCR fragment individually to all published sequences from GenBank and noted the similarity to both Elephantidae and hyrax and dugong mtDNA sequences. All amplified fragments showed between 78.1% and 98.6% identity to Elephantidae sequences and between 45.9% and 92.6% identity to hyrax and dugong sequences. Moreover, for individual fragments, the average identity to Elephantidae was 6.2% to 40% higher than that to hyrax and dugong. These results are consistent with the phylogenetic position of the mastodon and difficult—if not impossible—to explain by contamination. Therefore, we conclude that our results indeed represent authentic mastodon sequences.
The length of the mastodon mitochondrial genome is 16,469 bp. Thus, it is about 300–400 bp smaller than the mitochondrial sequences of the other proboscideans published to date. This difference in length is most likely an artifact due to the failure to obtain an amplification product that covers the complete tandem repeat in the control region. The mastodon mitochondrial genome contains 13 protein coding genes, 22 tRNA genes, two rRNA genes, and the control region, as expected for a placental mammal. Stop and start codons are shared with at least one of the other proboscideans for each of the protein coding genes, except ND3, ND4, and ND6. For these three genes, the mastodon sequences start with ATT, ATG, and GTG, whereas the corresponding genes of all other proboscideans start with ATC or ATA, GTG, and ATG, respectively. Differences in the annotation were observed in the other sequences notably because of stop codons being created by polyadenylation, a phenomenon widely present in mitochondrial genomes of other vertebrates [16]. The early stop of ND4 of the woolly mammoth sequences was not observed, and none of the proteins of the mastodon seems to be truncated. With a divergence time of at least 24 million years from any living relative, the mastodon sequence shows that the multiplex PCR approach can also be applied to taxa without sequence information from closely related species.
Partial DNA sequences that are claimed to be derived from mastodon are available from four studies. The sequences of Yang et al. [17] (228 bp of CYTB) and Joger et al. [18] (294 bp of CYTB) show substantially less divergence from mammoth and elephants than our sequence. The numbers of substitutions between their mastodon sequences and the Elephantidae are comparable to differences within the Elephantidae (results not shown). In contrast, our sequence clearly differs from the elephantid sequences. The results of Yang et al. have been recently questioned [19,20]. Debruyne et al. [19] concluded that Yang et al.'s mammoth sequence is a probable chimera of African and Asian elephant. Our study also rejects the results of Yang et al. for mastodon because the fragment of their CYTB mastodon sequence clusters closely within Elephantidae (see Figure S1A), as expected by the number of substitutions. For the same reason, the partial sequence of Joger et al. also seems to be derived from an Asian/African elephant contaminant (see Figure S1B). Finally, the unpublished 16S rDNA sequences obtained by Park et al. and Goldstein et al. (GenBank accession numbers AF279699 and AY028924, respectively) show 97% and 99% identity to primates using Blastn [21], which shows that these two sequences also were very likely derived from contamination.
In order to further investigate proboscidean evolution, we used the six existing whole mitochondrial genomes of Elephantidae (two African elephants, L. africana; two Asian elephants, Elephas maximus; and two woolly mammoths, Mammuthus primigenius) together with our newly obtained Mammut americanum sequence. As the mastodon lineage is evolutionarily much closer to Elephantidae than to extant outgroup species such as dugong or hyrax, multiple substitutions should not represent such a pronounced problem as is the case with the latter species. A comparison of the seven proboscidean mitochondrial genomes to those from hyrax and dugong shows that the number of substitutions separating mastodon from the Elephantidae is less than half that separating either hyrax or dugong from the Elephantidae (Table S1). Consistent with this result, the rates of synonymous (ds) and nonsynonymous (dn) substitutions are much lower when using mastodon as the outgroup (ds = 0.4; dn = 0.05). Using dugong or hyrax as the outgroup, ds is 1.2, showing substitution saturation [4], and dn is still 0.15. The transition/transversion ratio can also be used as a measure of substitution saturation [3,22]. Including the whole mitochondrial genome of the cow (Bos taurus) and the ostrich (Struthio camelus) reveals that dugong and hyrax have reached a plateau for both number of substitutions and transition/transversion ratio when compared to Elephantidae, whereas the mastodon has not (Figure 2). Thus, as previously noted [23], the mastodon represents a much better outgroup for inferring Elephantidae evolution than either dugong or hyrax.
The greater suitability of mastodon as an outgroup to the Elephantidae is further demonstrated by the results of the phylogenetic analyses. Using the mastodon sequence as an outgroup, we obtained higher support values for a sister group relationship of mammoth and Asian elephant than previous studies [3,4] and obtained the same tree topology from different methods of phylogenetic inference (Table 1). The bootstrap values for this relationship were between 94% (neighbor joining [NJ]) and 99% (maximum likelihood [ML]) with a Bayesian posterior probability of 1.00 (Table 1). These bootstrap values do not vary significantly among substitution models and are not dependent on whether the data are partitioned or unpartitioned (results not shown). Thus, at least with regard to mtDNA sequences, the relationship among mammoth and the living elephant species can no longer be seen as equivocal, as argued by some authors [8]. This result indicates that mastodon would also provide an excellent outgroup for phylogenetic analyses of Elephantidae using nuclear sequences [8] if it became possible to recover nuclear mastodon DNA sequences.
Using our complete mtDNA mastodon sequence, we were able to employ gene-by-gene phylogenetic analyses to explain why several earlier studies found a sister group relationship between African elephants and mammoths. The reconstructed phylogeny of the Elephantidae varied widely when we used each of the 13 protein coding genes and the two rRNAs individually. We recovered the mammoth–Asian elephant topology for the majority of the genes, but with lower support values (44%–90% for bootstraps and 0.42–1.00 for posterior probabilities). Other genes supported different tree topologies, sometimes with high bootstrap values or Bayesian posterior probabilities (up to 90% or 1.00; see Figure 3 and Table S2). In fact, when considering NJ trees alone, the majority (eight of 15) of the single-gene analyses in fact supported an incorrect topology. Some single-gene analyses resulted in different, yet well supported topologies when hyrax and dugong were used as the outgroup instead of mastodon [4]. These results indicate that studies based on a single gene can be misleading, and long sequences may often be necessary to obtain correct phylogenies (Figures 3 and S2; Protocol S3; see also, e.g., [24–27]).
Unlike the sequences of the nearest living outgroups of the Elephantidae, the dugong and the hyrax, the mastodon sequence has not yet reached substitution saturation (Figure S3). Consequently, the mastodon mtDNA sequence provides better estimates of the dates of divergence within the Elephantidae. Moreover, fossil evidence constrains the divergence of hyrax and dugong from proboscideans to earlier than ∼60 mya, the date for the oldest proboscidean fossil genus, Phosphatherium [28]. Previous attempts to date the Elephantidae divergence either used the two very distant outgroups, with limited success because the sequences violated the molecular clock assumption [4], or used the divergence of African elephant as calibration point [3], which yields more reliable relative divergence dates within Elephantidae, but does not provide independent estimates of divergence times within this group. Thus, the mastodon sequence allows independent dating of Elephantidae divergence times based on the sequence of a well-calibrated outgroup.
Using the best-fitting topology, we found the TN93 [29] model of substitution to be the simplest that fitted the data (results not shown). Using this substitution model, we were not able to reject the assumption of a molecular clock for the whole mtDNA genome sequences of the mastodon and the three Elephantidae species (2Δℓ = 6.8; p = 0.24). Using the Bayesian approach of Yang and Rannala [30], we evaluated the saturation level of our data by plotting the posterior means versus the width of the 95% credibility intervals (CIs) of the divergence times. The relationship is almost linear, and the coefficient of correlation of the linear regression has a value of 0.85 (Figure 4). Although a correlation of 0.85 suggests that the sequence data are highly informative it is still possible that the accuracy of divergence time estimates could be further improved by additional sequence data. When we used the paleontologically determined divergence date of 24–28 mya for mastodon [10,31], the divergence time of the African elephant turned out to be older than the 6 mya previously assumed [32]. In fact, we calculated it to have a posterior mean of 7.6 mya (95% CI 6.6 to 8.8 mya) when using the whole mitochondrial genome. The divergence between mammoth and Asian elephant also moves back in time to 6.7 mya (CI 5.8 to 7.7 mya; Figure 5; Table 2). Both dates are concordant with the presence of African elephant and Asian elephant fossils by 5.4–7.3 and 5.2–6.7 mya [33], respectively. Only about a million years separates the two divergence events, which is less than in humans, chimpanzees, and gorillas [34], for which extensive lineage sorting has been found [35,36]. Hence, we expect lineage sorting for the nuclear genome to also be problematic for Elephantidae, as claimed by Capelli et al. [8].
Our new estimates make the divergence of mammoth and African and Asian elephants even closer in time to the divergence of humans, chimpanzees, and gorillas, and other mammalian taxa. A number of environmental changes were occurring globally at that time, including the spread of grasslands and an increase in C4 plant biomass [37,38]. Further efforts should be given to studying the relationship between environmental changes and these phylogenetic events during the late Miocene.
Finally, our revised divergence dates also have bearing on the status of the African forest elephant. Based on nuDNA sequences, Roca et al. [6] argued that the African forest elephant represents its own species, L. cyclotis, distinct from the savannah elephant, L. africana. Using nuDNA sequence data and a divergence time between African and Asian elephants of 5 mya, they estimated the divergence between the nuDNA sequences from African savannah and forest elephants to be 2.63 ± 0.94 mya, and argued that both the deep divergence and the reciprocal monophyly between forest elephants and savannah elephants with regard to nuDNA support the distinction of the two forms as different species, a view also supported by microsatellite analyses [39]. However, based on extensive mtDNA analyses, both Eggert et al. [40] and Debruyne [7] disputed this view, as they found forest and savannah elephants being polyphyletic with respect to mtDNA sequences. In an extension of their earlier analysis, using both X and Y chromosomal and mtDNA sequences, Roca et al. [41] confirmed their view of a distinction on the species level and argued for unidirectional gene flow of mtDNA from forest elephants to savannah elephants. Although, complete mtDNA sequences for the two African elephant are not available, we can use our results to recalculate the divergence time between forest and savannah elephant inferred from the nuclear sequences. Using our estimate of 7.6 mya for the initial Loxodonta divergence increases the estimated divergence time to 4.0 mya. This date is older than the divergence times of many species pairs and hence supports the classification of African savannah and forest elephants as different species as proposed by Roca et al. [6,41].
The initial sequencing of the mammoth mitochondrial genome indicated that the substitution rate within Elephantidae is much lower than in humans and the African great apes [3]. We can use the mastodon sequence to determine whether this arose recently or in the more distant past. We compared the substitution rates of the mammoth, elephants, and mastodon, which had a most recent common ancestor 24–28 mya, to those of human, chimpanzee, gorilla, and baboon, which had a most recent common ancestor about 33 mya [42]. The substitution rate for the whole mtDNA genome was found to be more than twice as high in the four species of primates than in the four species of proboscideans. However, the distribution of rates among sites within the genome was almost identical (Table S3). It is not clear what causes the difference in rates. A possible explanation could be differences in body size, which in turn influence metabolic rates [43,44], but further studies on more species would be necessary to evaluate whether this is the major, or only, cause.
The sequence of a complete mtDNA genome obtained from a mastodon tooth extends the time frame for large-scale sequencing of ancient DNA substantially. Inclusion of this sequence in phylogenetic analyses confirms mammoth and Asian elephants as sister taxa and provides evidence for earlier divergences between Elephantidae species. The similarity of the divergence dates between Elephantidae species and between humans and African great apes suggests that a change in environmental conditions triggered speciation in African mammals beginning some 7.5–8 mya. Finally, we found further evidence that the mitochondrial substitution rate in proboscideans is considerably lower than in primates, and this difference manifested by at least 24 mya.
Although permafrost environment is especially suitable for long-term DNA survival, DNA sequences about 130,000 y old have also been reported from non-permafrost remains. Moreover, the age limit for preservation of plant DNA in permafrost environment is even older, currently at 300,000–500,000 y [45]. Even if the DNA from specimens is more fragmented than in the sample used in this study, the two-step multiplex procedure would allow reconstruction of long, continuous sequences such as complete mitochondrial genomes, substantially enlarging the possibilities of ancient DNA analyses.
For DNA extraction, we cut a ∼200-g sample along the tooth root of a mastodon molar (United States Bureau of Land Management–Alaska collection IK-99–237) collected in 1999 on the Ikpikpuk River (69° 22′ 10′′ N, 154° 40′ 46′′ W; Figure 1), which drains the central arctic coastal plain of northern Alaska. The tooth was a detrital find collected on a point bar along with numerous other large mammal bones of late Pleistocene affinity. These bones are eroded from cutbanks and deposited in large quantities on point bars along a number of rivers that drain the arctic coastal plain. The fluvial and alluvial sediments in these systems were mostly deposited during the late Pleistocene and early Holocene, but contain reworked bones of Pleistocene age [46–48]. There is a rare occurrence of sediments dating to the last interglacial period (∼130,000–100,000 y BP). To date, an assemblage of ∼3,000 bones has been collected, 312 of which have been radiocarbon-dated [48,49]. Most are within the radiocarbon range (<50,000–40,000 y), but about 20% have returned nonfinite dates (i.e., “greater than” ages ranging from 50,000 to 40,000 y BP). Radiocarbon analyses of a collagen extract from mastodon tooth IK-99–237 yielded a nonfinite 14C age of >50,000 y BP (Lawrence Livermore National Laboratory number CAMS91805), which places a lower limit on its age. In terms of its maximum age, there are only two bones in the entire assemblage from a taxon (Praeovibos) that dates to the middle Pleistocene, and even those two bones were re-worked into late Pleistocene sediments. Consequently, it is conservatively estimated that nonfinite-age bones in the assemblage date to between approximately 50,000 y BP and the end of the penultimate glaciation (about 150,000 y ago). Since the mastodon is principally an inhabitant of interglacial periods in Alaska, it is most likely IK-99–237 dates to the last interglacial period (i.e., its maximum age is probably ∼130,000 y BP), though we have no way of absolutely ruling out an age between 100,000 and 50,000 y BP. Bones in this assemblage are extremely well-preserved because they were entombed in permafrost sediments before being exposed by river erosion. This accounts for the exceptional preservation of DNA in IK-99–237.
Part of the tooth root was cut into small pieces and ground to a fine powder in a freezer mill (freezer mill 6750, Spex SamplePrep, http://www.spexcsp.com/sampleprep/) using liquid nitrogen. About 25 g of the powder was incubated overnight under constant agitation at room temperature in 700 ml of extraction buffer consisting of 0.45 M EDTA (pH 8.0) and 0.25 mg/ml proteinase K . After centrifugation the supernatant was concentrated to ∼50 ml using the Vivaflow 200 system (Vivascience, http://www.vivascience.com) with a polyethersulfone membrane with a molecular weight cut-off of 30,000 [50]. DNA was bound to silica using 40 ml of guanidinium thiocyanate buffer [51] and 100 μl of silica suspension for each of the five 10-ml aliquots of concentrated extraction buffer, with the pH adjusted to 4.0 using hydrochloric acid. After incubation for 3 h under constant agitation the silica was pelleted by short centrifugation and washed once with 1 ml of guanidinium thiocyanate buffer and twice with 1 ml of wash solution (51.3% ethanol, 125 mM NaCl, 10 mM Tris, and 1 mM EDTA [pH 8.0]). DNA was eluted using 50 μl of TE buffer for each of the five aliquots, resulting in ∼250 μl of extract. An extraction blank was carried alongside throughout all steps of extraction to monitor for possible contamination.
We designed 78 primer pairs using published sequences of African (Loxodonta1) and Asian (Elephas1) elephants, mammoth (Mammuthus1), and dugong (Dugong). The length of the targeted fragments varied between 139 and 334 bp (including primer), covering the entire mitochondrial genome of the mastodon except a repeat sequence in the control region.
Concordant with the initial paper describing the multiplex approach for DNA amplification from ancient samples [3], we divided the primer pairs for the first step into two sets to avoid amplification of the short overlapping fragments between adjacent amplification products [52]. In the second step, each primer pair was used individually, and to increase the specificity of amplification, in all except four cases, one primer per pair was “nested” compared to the first step, resulting in fragments between 139 and 324 bp long (including primers). Wherever possible, primers were designed to exclude amplification of human DNA.
Four primary amplifications (the multiplex step) were done for each of the two primer sets using two times 5 μl of 1:5 and 1:10 dilutions, respectively, of the extract, resulting in altogether eight primary amplifications. PCRs were conducted in a final volume of 20 μl consisting of 1× GeneAmp PCR Buffer II (Applied Biosystems, http://www.appliedbiosystems.com), 4 mM MgCl2, 1 mg/ml BSA, 250 nM of each dNTP, 2 U AmpliTaq Gold (Applied Biosystems), and 150 nM each of 78 primers (39 primer pairs per set). Initial denaturation and activation of the polymerase was done for 9 min at 94 °C followed by 30 cycles at 94 °C for 20 s, at 48, 50, or 52 °C for 30 s, and at 72 °C for 30 s. Aliquots of 5 μl of a 1:20 dilution of this primary amplification were used for the two sets of 39 individual secondary amplifications. These secondary PCRs were done in a final volume of 20 μl consisting of the same reagent concentrations as described above except that only 0.25 U of AmpliTaq Gold was used and the primer concentration was raised to 1.5 μM. Cycling conditions were the same as above except that between 30 and 40 cycles were performed. Extraction and water controls were carried along during all steps.
As not all amplifications were successful in the first attempt, several primers were redesigned after sequencing the flanking regions obtained from successful amplifications, and more primary amplifications were performed (see Results/Discussion and Protocol S1).
Amplification products were visualized on 2.0% agarose gels, and products of correct lengths were cloned using the Topo TA Cloning Kit (Invitrogen, http://www.invitrogen.com). In cases where amplifications showed visible primer dimers in addition to products of the correct length, the products were isolated from the gel and purified using the QIAquick Gel Extraction Kit (Qiagen, http://www.qiagen.com). After colony PCR [53] and purification of the products using the QIAquick PCR Purification Kit (Qiagen) and a Biorobot, a minimum of three clones per amplification (see Protocol S1) were sequenced on an ABI 3730 capillary sequencer (Applied Biosystems) using M13 universal primers.
All alignments were made using ClustalW [54] with default parameters. The D-loop sequences were removed from all genome sequences prior to analyses. We aligned the mastodon sequence to previously published sequences of two mammoths (Mammuthus1 and Mammuthus2), two African elephants (Loxodonta1 and Loxodonta2), and two Asian elephants (Elephas1 and Elephas2). For this alignment, we verified that the start and stop codons were aligned and that the open reading frames were preserved. For the concatenation of the protein coding genes, all genes were aligned individually and subsequently concatenated, ignoring the fact that some of them were overlapping. Hence, some nucleotides are duplicated in our concatenation. We examined publicly available mastodon sequences by aligning them with our sequence and constructing NJ trees using MEGA3.1 [55] with default parameters (see Protocol S2 for details on the program settings).
To test the suitability of our outgroup we estimated the degree of substitution saturation for the concatenated protein coding genes by computing the number of synonymous and nonsynonymous substitutions per site with the Pamilo-Bianchi-Li method implemented in MEGA3.1. We also used MEGA3.1 to compute the number of substitutions per site between proboscideans, dugong, and hyrax and to compute the transition/transversion ratio by aligning the seven proboscidean mtDNA genomes with those of dugong, hyrax, cow, and ostrich.
To solve the Elephantidae phylogeny we used MEGA3.1 to reconstruct trees using NJ and maximum parsimony. Initially, parameters similar to those used in [4] were chosen. We used the default parameters from MEGA3.1 with 10,000 bootstrap replicates. In addition, we built a Bayesian tree using MrBayes 3.1 [56]. The tree with maximum posterior probability was computed using one million iterations (see Protocol S1 for information on the options used). A GTR model [57,58] and a HKY85 model [59] of nucleotide substitution, both with gamma distributed rates of substitutions among sites, were used. The ML tree was constructed using Paup* [60] with an exhaustive search. For the ML tree, we performed 1,000 bootstrap replicates and chose a GTR model of substitution with gamma distributed rates among sites. Finally, we reconstructed the phylogeny for the seven sequences individually for each protein coding gene using MrBayes 3.1 and a HKY85 model of substitutions as well as by NJ as described above. The same analysis was performed for the concatenated sequence of the protein coding genes. However, the result of this analysis did not differ significantly from that of the analysis performed using the complete sequence (results not shown).
Because the different reconstructions revealed a single topology, we tested the different models of nucleotide substitutions against each other following the Felsenstein hierarchy in order to choose the best-fitting model. To this end, we compared the different models by likelihood ratio tests with the program baseml in PAML [61] using the alignment of the seven sequences without the D-loop, using gamma distributed rates among sites with eight categories, and considering a non-clock situation. The topology of the tree used for these analyses is shown in Figure 5 and identical to that in previous publications [3,4]. We reconstructed the tree once again using the above methods with TN93, the simplest model of substitution for which none of the more complex models fitted the data significantly better, but no significant difference from the previous results was observed (results not shown). To test whether the seven sequences evolved following a molecular clock we estimated the ML values for the phylogenetic tree of the seven sequences under both a non-clock assumption and a molecular clock assumption and compared the likelihoods of the trees obtained under the two assumptions using a likelihood ratio test (five degrees of freedom). As before, we assumed gamma distributed rates among sites and TN93 as a model of substitution.
Given the difficulty of precisely estimating calibration points and hence divergence times [62,63], we used a Bayesian approach and the program mcmctree to estimate the posterior means and the CIs of the divergence times. We used as a calibration point the divergence of the mastodon, estimated to be 24–28 mya [31] applying the lower and upper bound method. Given the unavailability of TN93, we chose HKY85, the most complex model implemented in mcmctree. The burn-in was set to 10,000, the number of samples to 100,000, and the sample frequency to five, as in [30], with four independent chains for each analysis. We chose wide priors for κ, the transition/transversion ratio, r, the substitution rate, and α, the parameter of the gamma distribution of rate variation among sites, to avoid too strong an influence of the priors on the posteriors. The 95% prior intervals were set to (0.00, 111.24), (0.02, 3.69), and (0.00, 1.23) for κ, r, and α, respectively. To compute the rate of substitution for primates we used the same approach for a four-taxon tree of baboon, gorilla, chimpanzee, and human. The calibration point was set to a lower and upper bound of 6–8 mya for the chimpanzee–human split [64] and an upper bound of 33 mya for the baboon divergence ([42] and references therein). We used the same disperse priors as before for κ, r, and α and the same options for the Markov chain Monte Carlo.
Since the use of a single model of evolution for the whole mtDNA sequence may result in errors, we partitioned the data into five subsets. We separated the first, second, and third positions of the codons, the rRNAs, and the tRNAs. The few noncoding sites were excluded to avoid overparametrization. The protein–protein overlapping fragments were classified under the second codon position, except for the seven sites between ND5 and ND6 that were excluded because the two genes are on different strands. Indeed we inverted the strands according to the transcription process. The phylogenetic tree was constructed using MrBayes 3.1 with a GTR model of substitution and one million iterations with unlinked partitions. We also allowed the partitions to have different rates. The divergence times using the partitions were computed using mcmctree with the options described above.
The GenBank (http://www.ncbi.nlm.nih.gov) accession numbers for the complete mitochondrial genome sequences of E. maximus, L. africana, and Mammuthus primigenius are NC_005129.1 for Elephas1 and NC_005129.2 for Elephas2, NC_000934.1 for Loxodonta1 and DQ316069.1 for Loxodonta2, and NC_007596.2 for Mammuthus1 and DQ316067.1 for Mammuthus2. For the partial sequences of Mammut americanum the GenBank accession numbers are U23737.1 (denoted Yang Mastodon), AY028924.1, and AF279699.1; for the partial sequences of Mammuthus primigenius the GenBank accession numbers are U23738.1 (denoted Yang Mammuthus1) and U23739.1 (denoted Yang Mammuthus2). The accession number for the whole mitochondrial genome of Mammut americanum determined in this study is EF632344. For the non-proboscideans, the GenBank accession numbers for the whole genomes used in this study are NC_003314.1 for D. dugon (Dugong), NC_004919.1 for P. capensis (Procavia), NC_006853.1 for B. taurus (Bos), Y12025.1 for S. camelus (Struthio), NC_001992.1 for Papio hamadryas (Baboon), NC_001645.1 for Gorilla gorilla (Gorilla), NC_001643.1 for Pan troglodytes (Chimpanzee), and NC_001807.4 for Homo sapiens (Human).
|
10.1371/journal.pcbi.1006626 | Coevolving residues inform protein dynamics profiles and disease susceptibility of nSNVs | The conformational dynamics of proteins is rarely used in methodologies used to predict the impact of genetic mutations due to the paucity of three-dimensional protein structures as compared to the vast number of available sequences. Until now a three-dimensional (3D) structure has been required to predict the conformational dynamics of a protein. We introduce an approach that estimates the conformational dynamics of a protein, without relying on structural information. This de novo approach utilizes coevolving residues identified from a multiple sequence alignment (MSA) using Potts models. These coevolving residues are used as contacts in a Gaussian network model (GNM) to obtain protein dynamics. B-factors calculated using sequence-based GNM (Seq-GNM) are in agreement with crystallographic B-factors as well as theoretical B-factors from the original GNM that utilizes the 3D structure. Moreover, we demonstrate the ability of the calculated B-factors from the Seq-GNM approach to discriminate genomic variants according to their phenotypes for a wide range of proteins. These results suggest that protein dynamics can be approximated based on sequence information alone, making it possible to assess the phenotypes of nSNVs in cases where a 3D structure is unknown. We hope this work will promote the use of dynamics information in genetic disease prediction at scale by circumventing the need for 3D structures.
| Proteins are dynamic machines that undergo atomic fluctuations, side chain rotations, and collective domain movements that are required for biological function. There is, therefore, a need for quantitative metrics that capture the dynamic fluctuations per position to understand the critical role of protein dynamics in shaping biological functions. A limiting factor in incorporating structural dynamics information in the classification of non-synonymous single nucleotide variants (nSNVs) is the limited number of known 3D structures compared to the vast number of available sequences. We have developed a new sequence-based GNM method, termed Seq-GNM, which uses co-evolving amino acid positions based on the multiple sequence alignment of a given query sequence to estimate the thermal motions of C-alpha atoms. In this paper, we have demonstrated that the predicted thermal motions using Seq-GNM are in reasonable agreement with experimental B-factors as well as B-factors computed using 3D crystal structures. We also provide evidence that B-factors predicted by Seq-GNM are capable of distinguishing between disease-associated and neutral nSNVs.
| A 3D structure is still required to computationally obtain protein dynamics, drastically limiting the extent to which conformational dynamics can be incorporated into genomic analysis. The reason for this is that there are exponentially more sequences than experimental structures. Currently, UniProtKB contains more than 100 million sequence entries, whereas the PDB reports the number of known 3D structures to be around 140,000 [1]. Furthermore, the number of known sequences is increasing at an exponential rate, compared to the much slower addition of new experimental PDB structures. This is due to the advent of high-throughput genomic sequencing, which is providing an unprecedented amount of data for genomic analysis. The vast amount of sequence data has driven the rapid classification of novel genetic variations through genome-wide association studies [2,3]. A large catalogue of non-synonymous single nucleotide variants (nSNVs) occurs in coding regions that can severely impact protein function, potentially leading to disease [4]. There are many in silico methods developed using evolutionary methodologies such as positional conservation and phylogeny and those that combine evolutionary approaches with biochemical and structural properties to diagnose neutral and disease associated nSNVs [5–11]. However, the accuracy of the majority of these in silico prediction methods is significantly lower for predicting the impact of nSNVs at highly evolving sites [12–16]. Protein dynamics can also be used to elucidate the functional impact of nSNVs and mechanisms of disease [5,17]. Our previous studies have evinced that a site-specific conformational dynamics analysis is capable of diagnosing nSNVs irrespective of evolutionary conservation [5,18,19] and recently has been incorporated as an additional feature for in silico prediction tools [20]. However, only a small fraction of the catalogued nSNVs in the coding regions (i.e. missense variants) have 3D experimental structures, [20], impeding broad application of protein dynamics in in silico tool predictions.
Coevolution, on the other hand, has become a valuable tool for its ability to predict structural contacts of 3D structures, particularly using global information through Potts models [21–27]. Coevolving residues are inferred from a multiple sequence alignment (MSA) of a given protein family, whereby if two given amino acids exhibit concordant patterns of evolution throughout the MSA then they are assumed to be in close spatial proximity in the folded 3D structure. This evolutionary principle can be leveraged so that sequence information can be used to describe protein topology, making de novo structure prediction possible [24,27]. It has been reported that only one correct contact for every 12 residues in a protein is necessary for accurate topology-level modeling [28]. In addition to structure prediction, coevolution analysis has also been used to identify critical interactions between protein complexes [22] important functional sites [24] and allosteric response [29]. The use of coevolution for structure prediction is largely possible for two reasons. First, the amount of sequence data for different protein families is sufficient to be leveraged by this technique to make predictions. Second, the methods for inferring coevolving residues from an MSA are becoming increasingly robust [30–34].
Inferring evolutionary couplings from an MSA are based on two primary approaches categorized as local [35–37] and global approaches [37–39]. The global approaches detangle direct evolutionary couplings from indirect couplings which enables them to capture spatial contacts [40]. Regardless of the method, the accuracy of detecting coevolving residues that correspond to structural contacts is fundamentally limited by the number of sequence homologs in the MSA. While most of the current methods use only the sequence homologs of the protein family belonging to target sequence, integrating multiple orthology protein families (i.e. families that share similar phylogeny and retain similar functions) was used to increase the number of homologs to produce a more accurate statistical inference [41]. RaptorX, leverages this joint family methodology; it uses an ultra-deep neural network combining coevolution information with sequence conservation information to infer 3D contacts and has produced higher accuracy than other methods [42–44].
In this paper, we will demonstrate the efficacy of our novel sequence-based GNM approach, called Seq-GNM, to estimate the dynamics profile of a protein with no a priori knowledge of its 3D structure. This de novo approach based on a Gaussian network model (GNM) enables the prediction of the magnitude of mean-square fluctuations of residues, which are proportional to the B-factors determined by X-ray crystallography experiments. However, instead of using a cutoff distance to determine 3D contacts as does the original structure-based GNM, we use coevolving residues (evolutionary couplings) in our model. We show that the theoretical predictions from our Seq-GNM are in reasonable agreement with experimental crystallographic B-factors as well as the values obtained from the structure GNM models that use spatial contacts. We also extend this analysis to determine the capacity of our model to assess the functional impact of nSNVs. We will demonstrate that the dynamics predicted by Seq-GNM can adequately classify disease and benign nSNVs across the proteome.
We considered a high-resolution protein (2.25 Å) that is involved in amino acid catabolism, acyl-CoA dehydrogenase (1JQI), as an example case to examine the B-factor profiles and predicted contact maps using Seq-GNM. Coevolution analysis using direct coupling analysis (DCA) has been shown to recapitulate accurate structural contact maps for a wide range of proteins [21,23,24,27,31,45]. As expected, the contact maps of Seq-GNM and structural GNM are similar (Fig 1). In a comparison of their B-factor profiles, both Seq-GNM and structural GNM exhibit good agreement with observed B-factors, capturing flexible and rigid positions. Using evolutionary coupling (EC) values obtained from RaptorX, the correlation between the Seq-GNM and observed B-factors is 0.77, whereas the correlation between the structural GNM and observed B-factors is 0.57 (Fig 1A). Similarly, using EC values obtained by EVcouplings produced a correlation of 0.60 between the Seq-GNM and observed B-factors (Fig 1B). The scores obtained from EVcouplings are still reasonable, yet relatively lower correlations compared to those obtained by the RaptorX. This is likely due to the relatively noisy contact map predictions by EVcouplings compared to the more reliable contact maps produced by RaptorX (we think this is due to their inclusion of multiple orthology protein families) [42].
The Seq-GNM produces a correlation with crystallographic B-factors of 0.60, which is within the same range as those produced by the GNM from structure of 0.57. Moreover, theoretical B-factor profiles obtained from both methods were able to identify the catalytic sites on all of the proteins.
As a further test of the efficacy of the Seq-GNM, we superimposed the predicted B-factors onto the structures of three diverse proteins– 5'(3')-deoxyribonucleotidase (2JAO), acyl protein thioesterase (1FJ2), and NADH-cytochrome b(5) reductase (1UMK)–to visually contrast the predicted B-factors with that of experiment. Fig 2 shows each protein color-coded according to their B-factor profile on a spectrum of blue–white–red, where blue represents the lowest B-factors (less mobility) and red represents the highest B-factors (more mobility). The left panel shows the experimental B-factors for each protein, while the right panel shows the theoretical values predicted by the Seq-GNM. We investigated whether secondary structure was a factor in how the B-factors were distributed across the protein, and if certain secondary structure domains would exhibit less agreement with experiment. In this context, the proteins were selected so that they had a variety of secondary structure components–2JAO contains primarily alpha helices, 1UMK is mainly composed of beta-sheets, and 1F2J is a combination of alpha helices and beta-sheets. For 2JAO, the exterior helices that are flexible (red) in the observed structure are all reproduced in the predicted structure. The one highly rigid (blue) helix in the observed structure was more flexible in the predicted structure but was still in overall agreement. There is a surprising amount of similarity between the observed and predicted structure of 1F2J, considering that it contains both alpha-helix and beta-sheet elements. Similarly, 1UMK showed good agreement, except for some miniscule differences. This gives further evidence that the magnitudes of residue fluctuations predicted by the Seq-GNM model is representative of the crystallographic B-factor profiles for many proteins.
In order to compare predicted B-factors with crystallographic B-factors, we extracted a subset of 39 structures that had a resolution better than 2.0Å to obtain more realistic crystallographic B-factors (unreliable B-factors are common for many PDB structures) [18,46]. The same cutoff of 2.0Å was used in an earlier study to compare GNM predicted B-factors with those determined by crystallography [47]. For all 39 structures, the Seq-GNM (using EC values from RaptorX) and structure GNM were used to estimate their B-factors, which were then compared with the observed B-factors by calculating the correlation for each protein. The mean correlation coefficient for the Seq-GNM was 0.53 while the mean correlation coefficient for the structure GNM was 0.58. The correlation of 0.58 for structural GNM of our smaller data set is consistent with the findings of Kundu et al. where 113 high-resolution structures (resolution <2.0 Å) were used and, the mean correlation coefficient with observed B-factors was 0.59 [47].
As shown in Fig 3A, boxplot distributions reveal that correlations are not significantly different between the sequence and structure GNM (p = 0.055 in a student t-test). The structure GNM appears to perform only slightly better than the Seq-GNM. Fig 3B shows the same distribution separated into 10 individual bins of size 0.1. The overall shapes of the two distributions are similar, except for the exaggerated relative lower second peak of the Seq-GNM at 0.4. It should also be noted that for these cases where Seq-GNM had low correlations, the EC threshold could be tuned to yield much higher correlations. If this were done on a case-by-case basis, the overall correlation distributions would be even more similar. Thus, the EC threshold may be used as a tuning parameter to enhance the correlation coefficient for purposes of model optimization.
Interestingly, for the cases where predicted B-factors by Seq-GNM yielded significantly better correlations with the experimental B-factors than those obtained by GNM from structures, we observed that biological units of these proteins are assigned as oligomeric forms. While predicted B-factors obtained using Seq-GNM does not retain this information, it successfully predicts the experimentally low B-factor values of interface positions as shown for protein 5'(3')-deoxyribonucleotidase (2JAO) and protein aldehyde Dehydrogenase 7A1 (2J6L) in Fig 4. It is indeed shown in earlier work of direct contact analysis that co-evolution can identify positions of protein interfaces and protein-protein interaction partners and successfully reconstruct protein complexes and interaction network [23,30,48]. Thus, it is not surprising to see that it yields good correlations with the experimental B-factors. Conversely, predicted B-factors from structure can only improve when the oligomeric structure is used for the GNM analysis.
Even when using high-resolution X-ray structures, there is still some uncertainty about the realistic nature of crystallographic B-factors. For this reason, we thought a more plausible way to determine the efficacy of the Seq-GNM was to compare it directly with the structure GNM. The structure GNM is a robust method to describe thermal fluctuations in a protein, and in many cases, it performs as good or better than the ANM or MD [47,49]. We systematically evaluated the performance of the Seq-GNM and structure GNM for the entire set of 139 structures and obtained the correlation coefficients for each protein (Fig 5).
The average correlation of B-factors between the Seq-GNM and structure GNM model is 0.63 when using EC contacts from RaptorX and 0.43 when using contacts from EVcouplings. As seen in Fig 5A, the distribution of correlation coefficients increases until 0.8, and then subsequently decreases. Interestingly, there are still an appreciable number of sequences yielding high correlations from 0.8 to 1.0. A distinguishing feature of the distribution is the pronounced peak in the bin from 0.7 to 0.8, indicating that significant fraction of our data set yields high correlations between 0.7 and 0.8. This is evidence that the Seq-GNM is efficiently capturing protein dynamics and supports the theory that ECs can be used as a substitute to 3D structure contacts in the GNM and still produce reliable dynamics profiles. The results of Seq-GNM based on contacts predicted by RaptorX usually yields B-factors that are closer to experimental B-factors as it uses structural information in its neural networks leading to better EC values and correlations with structure [44].
Crystallographic B-factors have previously been used to assess the impact of nSNVs on protein function [18,50–54]. A study [51] found that mutations on lysozyme that impaired function exhibited lower than average temperature factors, suggesting that rigid sites on the protein are more susceptible to destabilizing nSNVs than flexible sites [55]. Another study revealed a relationship between crystallographic B-factors and the impact of nSNVs on protein function [56]. A commonly used tool to diagnose neutral and disease associated nSNVs, PolyPhen-2, uses evolutionary information, structural information, and crystallographic B-factors in its prediction model [49]. These studies indicate that crystallographic B-factors can be used to predict the tolerance of a given residue to an nSNV (i.e., whether or not the occurrence of an nSNV would impact function).
We investigated whether B-factors predicted by the Seq-GNM were indicative of biological phenotype for nSNVs in the human population. A total of 738 nSNVs were mapped to the 139 enzymes, where 436 are disease-associated and 302 are neutral. S1 Table shows the number of disease and neutral nSNVs that occur on each protein. The Seq-GNM (using EC contacts from RaptorX and EVcouplings) was computed systematically for all 139 enzymes to obtain their dynamics profiles. The theoretical B-factors scores were converted into a percentile rank so that the values could be compared across different proteins.
We initially looked at two human enzymes, human lysozyme (PDB: 1C7P) and human cytochrome reductase (PDB: 1UMK). They were chosen because they were short proteins that each contain a disease and neutral nSNV. Human lysozyme is a glycoside hydrolase that functions in the immune system by causing damage to cell walls of bacteria. Human cytochrome b5 reductase is involved in many oxidation/reduction reactions including converting methemoglobin to hemoglobin [55].
Each structure is color-coded according to its theoretical B-factor profile on a spectrum of blue–white–red. Sites that exhibit high mobility (flexible) are red, and sites that have low mobility (rigid) are blue. Regions that are characterized by low mobility are usually important for maintaining stability and function, thus a mutation could act to destabilize the protein and impair its function. Fig 6A show the disease mutation I56T occurring on a rigid site with a B-factor of 0.0075. The neutral mutation T70N has a B-factor of 0.96 indicating that it is a highly mobile site. Both I56T and T70N occur on loop regions. Although loops are generally more flexible, three alpha-helical domains encompass the loop containing I56T, which implies that it may be involved in interactions that contribute to stabilizing the functional conformation. Thus, the I56T mutation may disrupt these critical interactions and impair the enzymatic function. In the case of cytochrome reductase (Fig 6B), the disease mutation R57Q is also on a rigid site with a B-factor of 0.14. Instead of being located near the core, R57Q is highly exposed protruding outwardly from a beta-barrel. However, since beta-barrels often harbor functional residues, the R57Q mutation may disrupt certain interactions critical for modulating function. The neutral mutation T116S is located on a loop and has a B-factor of 0.96, indicating that is it has a high mobility. In our earlier proteome wide study of over 100 human protein structures, we have shown that sites that are highly flexible (e.g., loop regions, or superficial sites) are typically more robust to mutations. Conversely, rigid sites are more susceptible to mutations that may disrupt function [18,19]. For these two cases, the B-factors produced by Seq-GNM successfully distinguished between the disease and neutral nSNVs, without using the 3D structures.
These findings prompted us to analyze the proteome-wide set of 139 enzymes to determine if the B-factors were indicative of phenotype for all 436 disease and 302 neutral nSNVs. The raw B-factor values were converted into a percentile rank (%B-factor) and then binned into 5 bins of size 0.2. We computed the observed-to-expected ratio of B-factors, where the expected values were based on the B-factor distribution of all 51,618 sites across all 139 proteins, and the observed values were based on the B-factors of the 436 disease sites. The same process was done for the 302 neutral nSNVs. Under the null hypothesis that predicted B-factor of the disease associated nSNVs yields similar distribution of all the positions gathered from 139 enzyme sequences, the ratio of expected and observed sites harboring disease mutations for each %B-factor bin should be close to 1, which would imply that B-factor does not distinguish sites that are prone to disease. This is the null hypothesis that disease sites are distributed uniformly between sites with low and high mobility. However, the null hypothesis was rejected for the 436 disease nSNVs (p <0.001). Fig 7 shows the observed-to-expected ratio plot of disease and neutral nSNVs, which indicates that disease nSNVs are overabundant at low %B-factor sites (<0.4) and under abundant at high %B-factor sites. Conversely, neutral nSNVs are overabundant at high %B-factor sites (>0.6) and under abundant at low %B-factor sites. This evidence suggests that the occurrence of an nSNV on a site with a low B-factor is likely damaging based on the position irrelative of the substitution. This is in agreement with our previous proteome-wide study showing that substitutions at rigid sites are more often associated with diseases [18]. Conversely, an nSNV on a high B-factor site is usually benign. Low B-factors usually signify a residue that is crucial for modulating functional motions (e.g., a hinge). Thus, mutations on these sites can severely impact function. High B-factor sites are more flexible (e.g., loops) and more robust to mutations. Fig 7 suggest that it is possible to use the predicted B-factors to discriminate between disease and neutral nSNVs using co-evolution obtained from only multiple sequence alignment. Moreover, it can be used as an additional feature for in silico predictions [12].
Predictive models were created using logistic regression as the classification algorithm, 80% of the data was used for training and 20% for testing for 10 randomized sets. Models were evaluated based on ROC curves and their respective area under curve (AUC), the best performance is labeled as AUC_max and average performance as AUC. Theoretical B-factors obtained by Seq-GNM, experimental B-factors, and evolutionary parameters were used as predictive variables for training and testing (Fig 8). Seq-GNM and experimental B-factors have similar performance (maximum AUC of best 0.76 and 0.75, respectively), with Seq-GNM overshadowing experimental B-factors on average (AUC of 0.69 and 0.60, respectively). The ~0.70 AUC of B-factors obtained from Seq-GNM is impressive, as it has been shown that majority of state-of-art methods also yields similar AUC in independent tests [5,13]. Moreover, incorporation of Seq-GNM as an additional feature with evolutionary parameters resulted in higher prediction performance. While the AUC scores obtained using the evolutionary features for classification gives 0.76, this is increased to 0.81 after including the B-factors of Seq-GNM (Fig 8C and 8D). This result also demonstrates the efficacy of Seq-GNM in disease prediction as a complementary metric to other metrics used as features in classifiers.
We also compared the performance of Seq-GNM with common in silico prediction tools like Polymorphism Phenotyping v2 (PolyPhen-2), and Sorting Intolerant from Tolerant (SIFT) [6,58]. The accuracy, sensitivity, and selectivity of disease predictions for nSNVs with experimental B-factors, B-factors from SIFT, PolyPhen-2, evolutionary parameters, and Seq-GNM are tabulated in Table 1. The accuracy of Seq-GNM using both EC values from EVcouplings and RaptorX is ~0.70. This accuracy is similar to using experimental B-factors for prediction (0.69) and also very close to prediction with evolutionary parameters (0.75), suggesting that Seq-GNM allows us to incorporate protein dynamics in nSNV predictions when the 3D experimental structures are not available. Moreover, accuracy of Seq-GNM approach is greater than SIFT (0.65) and PolyPhen-2 (0.64). Interestingly, Seq-GNM obtained by EVcouplings and RaptorX yields similar accuracies indicating that evolutionary couplings without the inclusion of structure could be utilized to predict B-factors to include as a feature to in silico prediction tools. Seq-GNM sensitivity (~0.90) surpasses other methods (0.80 for SIFT, 0.63 for PolyPhen-2, and 0.85 for evolutionary parameters), but it has a shortcoming in selectivity (~0.36) as other methods reach higher (~0.59). Conversely, training Seq-GNM combined with evolutionary parameters enhances the selectivity (0.66) to its highest value compared to others. Seq-GNM with evolutionary parameters predicted disease related nSNVs with accuracy 0.78 and sensitivity of 0.84, reaching beyond predictions of other metrics solely. These results suggest the incorporation of Seq-GNM with other prediction metrics can augment accuracy, sensitivity, and selectivity of prediction.
Prediction accuracy of Seq-GNM is further tested using 323 nSNVs (187 disease-associated, 136 neutral) of 22 proteins where their 3D experimental structures are not available (S2 Table). We used the trained classifier model of Seq-GNM B-factors for this test. While the B-factors obtained solely from Seq-GNM are used, it reached an accuracy, sensitivity, and selectivity of 0.82, 0.82, 0.83, respectively. This result further suggests that Seq-GNM allows us to incorporate protein dynamics as additional feature in in silico prediction tools without a known 3D structure.
While we and others [5,19,59–63] have shown that the integration of conformational dynamics into genomic analysis will help next generation of approaches to predict the impact of novel missense mutations on the human proteome, the inherent limitations in the availability of 3D structures compared to the vast number of sequences must be addressed. This begs the question: how can protein dynamics be used in genome-wide analysis to predict functional impacts of nSNVs? There is, therefore, a need to be able to obtain protein dynamics by leveraging only sequence information, without a priori knowledge of a 3D structure. For this reason, we have developed this novel method to estimate the dynamics profile of a protein by using only a sequence as input. The method uses the coevolution of amino acids through multiple sequence (which tend to be spatially close in the 3D tertiary structure) and a simple Gaussian network model (GNM) to obtain dynamics. The original GNM based on the 3D structure is well-known for its ability to describe residue dynamics profiles due to thermal motions in proteins (i.e., B-factors). We showed that our sequence-based GNM model is able to adequately reproduce the mean-square fluctuations (B-factors) calculated by the original GNM, particularly outperforms for the cases where biological functional state is oligomeric. Our estimates of B-factors for a proteome-wide set of proteins exhibited good correlation with the structure GNM. Moreover, our estimated B-factors were in reasonable agreement with crystallographic B-factors for many cases. To address the issue of how protein dynamics can determine the impact of nSNVs across the genome where there are no known 3D structures, we tested the ability of our predicted dynamics from the Seq-GNM to assess nSNV phenotypes. A plot of the observed-to-expected ratio of the predicted B-factors revealed distributions of disease and neutral nSNVs that are similar to those in a previous protein dynamics analysis work [18]. The predicted B-factors using the Seq-GNM was able to discriminate between disease and neutral nSNVs with an accuracy of 0.70 and incorporating the Seq-GNM predicted B-factors with evolutionary parameters increased overall accuracy to 0.78. This analysis demonstrates that the Seq-GNM makes it possible to obtain estimates of dynamics without using a 3D structure, which will allow for the integration of conformational dynamics into large-scale analysis of genomic variants.
A curated set of 139 structures was selected for several reasons. First, they have high query coverage (>80%) and sequence identity (>80%) as found from a BLAST search, and the structures had already been modeled using the Modeller software package [64] to account for any missing residues. Second, genetic variants were previously mapped onto these structures, such that the positions containing known nSNVs were already determined, enabling us to easily compare our results using sequence coevolution with the genetic variation data. A total of 738 genetic variants were obtained from the HumVar database [58], which was comprised of 436 disease and 302 neutral nSNVs. Finally, the structures were either monomers or the single-chain unit of a multimer with <600 residues, allowing for tractable calculations of residue coevolution using the RaptorX web server [42,44], and EVfold (EVcouplings) [21]. A table summarizing the dataset is presented in S1 Table.
The amino acid sequence from each of the 139 structures was used as input for the evolutionary coupling (EC) analysis. The choice of taking the amino acid sequence from the structure was done so that the predicted EC contacts could be compared directly to the experimentally observed structure contacts as verification that the model was producing realistic contact maps. Moreover, the theoretical B-factors predicted by our sequence-based model could be directly compared to the experimental B-factors for each protein. If the structure was unknown, however, sequence databases (e.g. UniProt, PFAM, etc.) could be used. The PDB sequences were given to the RaptorX web server [42,43], which computed the relative probability of each residue pair i, j of being in 3D contact based on their coevolution strength. The sequences were also used to generate MSAs using phmmer [65]. Using MSAs, DI values are calculated by EVcouplings. In order to ensure consistency between different proteins of varying lengths, we converted the raw scores into percentile ranks. We then used a threshold value, taking only the top scoring evolutionary couplings (i.e., the strongest couplings are more likely to be in spatial contact). An optimized threshold value was systematically evaluated and is discussed in the Methods.
The Gaussian network model (GNM) is an isotropic approach based on the contact topology of a crystal protein structure to obtain the equilibrium fluctuations of residues due to thermal motion. It uses a specified cutoff distance to define interacting pairs that are connected by springs with a single-parameter harmonic potential. In this structure-based GNM, the interacting residue pairs within the cutoff range are represented as contacts in the Kirchhoff (connectivity matrix).
In the proposed sequence-based GNM (Seq-GNM) approach we will instead use coevolving residue pairs (evolutionary couplings) as contacts in the Kirchhoff. In this way, the 3D structure is no longer a prerequisite to form a GNM. To construct the Kirchhoff, a threshold is defined where any evolutionary coupling scores above that threshold are sufficiently coupled such that they are spatially close in 3D structure. If a given evolutionary coupling pair meets the threshold criteria, it is assigned a value in the Kirchhoff for non-bonded contacts of –1 multiplied by its evolutionary coupling score (i.e., –1×ECscore). This will permit that the strength of each connection will attenuate proportionally to the evolutionary coupling strength. The Kirchhoff can be decomposed into the individual contributions from the bonded contacts representing the chain connectivity (Rouse chain) and that from the non-bonded contacts [56]. In the Seq-GNM the contribution of non-bonded contacts to the Kirchhoff is constructed according to
Γijnb={−1×ECscore,i≠jevolutionarycoupling0,i≠jnocoupling−∑i,i≠jΓij,i=j
(1)
For the local chain connectivity (Rouse chain), we don’t take into account evolutionary couplings, and matrix was constructed such that every residue pair i, i ± 1 to i, i ± 3 is in contact as
Γijcc={−1,i≠jand∑i|k=1,2,3Li,i±k0,i≠jelse–∑i,i≠jΓij,i=j
(2)
Then the overall Kirchhoff is the combination of the two contributions Γij=Γijcc+Γijnb. The vibrational dynamics due to thermal fluctuations can then be evaluated in the same way as the original GNM by inverting the Kirchhoff matrix. The magnitude of mean-square fluctuations is then written in terms of the inverse Kirchhoff as
⟨(ΔRi)2⟩≅[Γ−1]ii
(3)
This is proportional to the Debye-Waller temperature factors or B-factors, which describe the attenuation of X-ray scattering due to the thermal motions of atoms (Bi = 8π2⟨(ΔRi)2⟩/3). Here there is no single-parameter force constant as in the GNM obtained from structure [52], and the pair-wise interactions are simply the strength of the evolutionary couplings as given by their ranked scores. The theoretical predictions of our Seq-GNM can be compared to the predictions of the original GNM obtained from structure as well as observed crystallographic B-factors. A general workflow of our method is presented as a flow diagram in Fig 9.
To ensure consistency when analyzing different proteins with varying lengths, we converted the raw scores of evolutionary couplings (EC) into a percentile rank. We computed the Seq-GNM for all 139 structures using a constant threshold percentile rank EC value to assign contacts and measured the correlation between the B-factors predicted by our Seq-GNM to the GNM obtained from structure. We used only the top percentile EC scores predicted by RaptorX and EVcouplings as predicted contacts, because only certain fraction of high EC scores are true native contacts in 3D structure, largely due to noisy artifacts in the MSA such as the transitivity of correlations and phylogeny. To determine the optimal threshold value, we tested a range of threshold values from 0.92 to 0.99. A threshold value ≤0.92 yields superfluous contacts leading to a noisy contact map, and thus, a lower overall correlation (Fig 10). Conversely, a threshold value ≥0.99 gives a deficient number of contacts, which yields an excessively sparse contact map and a lower overall correlation. As Fig 10 shows, a threshold value of 0.98 produced the best overall correlation with the GNM from structure and, thus, was taken to be the optimal threshold value used in the analysis.
|
10.1371/journal.pntd.0001158 | Protection against Diarrhea Associated with Giardia intestinalis Is Lost with Multi-Nutrient Supplementation: A Study in Tanzanian Children | Asymptomatic carriage of Giardia intestinalis is highly prevalent among children in developing countries, and evidence regarding its role as a diarrhea-causing agent in these settings is controversial. Impaired linear growth and cognition have been associated with giardiasis, presumably mediated by malabsorption of nutrients. In a prospective cohort study, we aim to compare diarrhea rates in pre-school children with and without Giardia infection. Because the study was conducted in the context of an intervention trial assessing the effects of multi-nutrients on morbidity, we also assessed how supplementation influenced the relationship between Giardia and diarrhoea rates, and to what extent Giardia modifies the intervention effect on nutritional status.
Data were collected in the context of a randomized placebo-controlled efficacy trial with 2×2 factorial design assessing the effects of zinc and/or multi-micronutrients on morbidity (n = 612; height-for-age z-score <−1.5 SD). Outcomes measures were episodes of diarrhea (any reported, or with ≥3 stools in the last 24 h) and fever without localizing signs, as detected with health-facility based surveillance. Giardia was detected in stool by enzyme-linked immunosorbent assay. Among children who did not receive multi-nutrients, asymptomatic Giardia infection at baseline was associated with a substantial reduction in the rate of diarrhea (HR 0.32; 0.15–0.66) and fever without localizing signs (HR 0.56; 0.36–0.87), whereas no such effect was observed among children who received multi-nutrients (p-values for interaction 0.03 for both outcomes). This interaction was independent of age, HAZ-scores and distance to the research dispensary. There was no evidence that Giardia modified the intervention effect on nutritional status.
Although causality of the Giardia-associated reduction in morbidity cannot be established, multi-nutrient supplementation results in a loss of this protection and thus seems to influence the proliferation or virulence of Giardia or associated intestinal pathogens.
| Giardia intestinalis is a well-known cause of diarrhea in industrialized countries. In children in developing countries, asymptomatic infections are common and their role as cause of diarrhea has been questioned. In a cohort of rural Tanzanian pre-school children, we assessed the association between the presence of Giardia at baseline and subsequent diarrhea risk. The study was conducted in the context of a randomised trial assessing the effect of supplementation with zinc and other micro-nutrients on malaria, and half of the children daily received a multi-nutrient supplement. Surprisingly, we found that the presence of Giardia at baseline was associated with a substantial reduction in diarrhea risk. Multivariate statistical analysis showed that this protection could not be explained by differences in age or walking distance to the dispensary between children with and without Giardia. Because we cannot exclude that children differed in other (unmeasured) characteristics, we cannot draw firm conclusions about the causality of the observed association, but our findings support the view that the parasite is not an important cause of diarrhea in highly endemic settings. Striking was that the Giardia-associated protection was lost when children received multi-nutrients. Our data do not provide information about the mechanisms involved, but suggest that multi-nutrients may influence the compositionor pathogenicity of intestinal biota.
| In developed countries, Giardia intestinalis (syn. G. duodenalis, G. lamblia) causes diarrhea while the prevalence of infections in the general population usually does not exceed 5% [1]. In developing countries, however, asymptomatic infections are much more common, with prevalence values in pediatric populations typically being around 30% [2]–[4], and reports on their association with diarrhea are inconsistent. Some reported an association with acute [5] and persistent [6] diarrhea, whereas several studies found no association [6]–[10], or even that Giardia infection was associated with protection against acute diarrhea [9], [11]–[14].
Because the role of Giardia as diarrhea-causing agent is controversial and re-infection can occur rapidly in developing areas where it is highly endemic, it has been recommended that children with asymptomatic infection should not be treated in such settings [15], [16]. This notion is challenged, however, by findings from surveys [17]–[21] and a prospective cohort study [22] suggesting that such infections may impair linear growth, presumably by reducing intake and causing malabsorption of nutrients. In addition, in a prospective cohort study, it was found that episodes of Giardia with diarrhea but not diarrhea itself were associated with impaired cognition, perhaps because infection can lead to deficiencies of zinc and other micronutrients that have been associated with deficits in cognitive development [23].
In the current study, we aim to compare rates of diarrhea in children with and without Giardia infection. Because the study was conducted in the context of an intervention trial that assessed the effect of multi-nutrient supplementation on malaria, we also assessed to what extent the relationship between Giardia and diarrhoea rates was influenced by supplementation. In addition, we explore whether the presence of Giardia infection at baseline modifies the response of nutritional indicators to multi-nutrient supplementation.
This study was part of a randomized placebo-controlled trial in children aged 6–60 months, with the primary aim to assess the effect of supplementation with zinc and other micronutrients on malaria rates (ClinicalTrials.gov, NCT00623857). It was conducted in a rural area in Handeni District, Northern Tanzania that is highly endemic for malaria. In a pilot survey among children aged 6–72 months in 2006 (n = 304), we found a high prevalence of Giardia intestinalis (30%; assessed by microscopic examination of a single stool sample per child), and only few cases of Ascaris lumbricoides, Trichuris trichiura or Schistosoma intestinalis (3%, 5% and 0%, respectively) (unpublished results). Residents in the area virtually all comprise poor farmer families engaged in subsistence farming, with oranges being produced seasonally as cash crops. Families are living in self-constructed clay houses, with very few having pit-latrines. Water for drinking and household use is collected from central shallow wells. Few people boil water before drinking. Access to health-care was limited until we constructed a research clinic at a central location in the study area, which provided free primary care to study participants.
Details about study design will be published elsewhere. In brief, between February and August 2008, we recruited all resident children aged 6–60 months, and excluded those with height-for-age z scores >−1.5SD, weight-for-age z-score <−3SD, haemoglobin concentration <70 g/L and with signs of severe or chronic disease, until attaining the target number (n = 600) (Figure S1).
The trial had a 2×2 factorial design with children receiving either multi-nutrients with zinc (Table S1), multi-nutrients without zinc, zinc alone (10 mg), or placebo. The levels of magnesium and vitamin C in the multi-nutrient supplement were below the upper limits that were based on osmotic diarrhea and related gastrointestinal disturbances as critical endpoints [24]. Supplements were color-coded and administered daily by community volunteers. Intervention groups were similar in baseline characteristics.
At baseline (on the day of enrollment), we collected venous blood in tubes suitable for trace element analyses (Becton-Dickinson, Franklin Lakes, NJ) and a fresh stool sample for each child in a vial that was pre-filled with sodium acetate-acetic acid-formalin (SAF) and stored in a refrigerator immediately after collection. A second vial with unfixed feces was stored in liquid nitrogen (−196°C) for subsequent genotyping. Whole blood hemoglobin concentrations were measured immediately using a portable photometer (Hemocue, Ängelholm, Sweden). An aliquot of plasma was stored in liquid nitrogen. A clinical officer recorded reported symptoms and performed a physical examination using standardised forms. We computed anthropometric indices as the average of two recordings, taken on consecutive days.
We asked parents or guardians to bring their children to the research clinic if they noticed signs of illness. A clinical officer was on 24 h-duty and collected medical information on standardised forms that included a section on diarrhea.
A second survey, at 251 days (median; 95% reference range: 191–296 days) after enrolment, followed similar procedures as the baseline survey. Follow-up continued for all children until March 2009, when the study ended for all children simultaneously.
Stool samples were analyzed for the presence of Giardia-specific antigen by enzyme immunoassay (ProSpecT Giardia Microplate Assay, Oxoid, Basingstoke, UK). This test has a sensitivity and specificity of 93% and 99%, respectively, as compared to detection by microscopy in two sequential stool samples from individual subjects [25]. Plasma concentrations of ferritin, soluble transferrin receptor, folate and vitamin B12 were measured on a Beckman Coulter Unicel DxC880i system according to the manufacturer's instructions. Plasma concentrations of zinc and magnesium were determined by inductively-coupled plasma-mass spectrometry.
The study was approved by ethical review committees in The Netherlands and Tanzania (National Health Research Ethics Review sub-Committee). We sought written individual informed consent; parents or primary caretakers were invited to sign (or thumbprint if illiterate) the informed consent form in the presence of a member of the community as impartial witness (who countersigned the form).
Cases of diarrhea were defined as: a) all dispensary visits for parent- or guardian-reported loose or watery stools, with episodes being separated by at least 48 h of being without symptoms; or b) similar episodes with ≥3 loose or watery stools per 24-h period. Fever without localizing signs was defined as cases with reported fever that did not classify as malaria and were not accompanied by cough, diarrhea or other localizing signs. Thus cases of diarrhea and fever without localizing signs were mutually exclusive.
Data were analyzed using SPSS (v15·0 for Windows, SPSS, Chicago, IL, USA) and STATA (v11; College Station, Tx, USA). We report incidence rates and assessed group differences by Kaplan-Meier analysis with Tarone-ware test. Differences in the association between Giardia at baseline and morbidity outcomes between intervention groups were assessed by analysis within intervention strata, and directly by Cox regression analysis that included dummies for intervention groups and interaction terms. Cross-over between groups, whereby children who were initially infected became infection-free and vice versa in the course of the intervention period, may dilute potential effects of Giardia over time. For this reason, we restricted our primary analysis to first episodes, because an analysis of all events is probably more susceptible to such dilution of effect. However, because a substantial number of children experienced recurrent events and analysis of all events may better reflect total disease burden, we repeated these analyses based on all events, with robust estimates of the standard error to account for correlation between episodes within children. We explored potential confounding by adjusting for factors that were previously found to be prognostic for diarrhea and other morbidity outcomes (age, distance and height-for-age z scores). Children for whom Giardia infection status at baseline could not be determined were excluded from the analysis of the association between Giardia and disease rates.
Continuous outcome variables that were not normally distributed were log-transformed as appropriate. We used multivariate linear regression analysis with interaction terms to assess to what extent the effect of zinc and multi-nutrient supplementation (either alone or combined) on indicators of nutritional status depended on Giardia infection.
The study profile is shown in Figure 1. G. intestinalis was detected in 192 children (31%). We failed to obtain fresh stool samples for 54 children at baseline and for 50 children during the second survey, when 20 children (3%) were lost to follow-up (3 died; 2 were withdrawn by parents; 15 emigrated from the study area). Baseline characteristics are presented in Table 1. Children with Giardia infection were on average 3.3 months older than their uninfected peers, resided somewhat closer to the dispensary, had a lower prevalence of inflammation, marginally higher hemoglobin concentrations, as well as marginally lower plasma concentrations of soluble transferrin receptor and folate. All other biochemical indicators of nutritional status were similar, and we found no evidence that Giardia was associated with symptoms as reported by the mother (Table S2). The percentage of children who received antibiotic or anti-malarial treatment at baseline was similar in both groups.
At time of the second survey, 43% of children with Giardia infection at baseline no longer carried the parasite, while 23% of children who tested negative at baseline had become infected (Figure 1).
There were 3,268 clinic visits in a total follow-up time of 526 child-years. For 390 of these visits (12%), the parent or guardian reported diarrhoea, of which 223 episodes were accompanied by ≥3 loose stools in the past 24 h. Overall, the incidence of first episodes of diarrhoea was almost 50% lower among children with Giardia at baseline than among those without (Table 2). Similar effect estimates were obtained when including all episodes in the analysis (Table 3). When stratified by intervention group, the association between Giardia infection at baseline and diarrhea was similar in children receiving placebo as in those receiving zinc (in both groups the infection was associated with a protection). Likewise, this association was similar in both groups receiving multi-nutrients (no association in either group; Figure 2); we therefore combined children who received multi-nutrients with or without zinc (henceforth referred to as ‘with multi-nutrients’), as well as their peers who received zinc or placebo (‘without multi-nutrients’) as two separate groups.
Thus analyzed, Giardia infection at baseline was associated with a substantial increase in time to first diarrhea episode with ≥3 watery stools/24 h (p<0.001), but only so among children without multi-nutrients, whereas no association was found between Giardia infection and diarrhea in those receiving multi-nutrients (Figure 3; top panels, and Tables 2 and 3).
Adjustment for age and distance to the dispensary led to smaller but still substantial associations between baseline Giardia infection and diarrhea (all events), whilst interaction effects between infection and the multi-nutrient intervention remained virtually unchanged (Figure 4). Further adjustment for baseline factors previously found to be prognostic for diarrhea (height-for-age z-scores, sex, inflammation and use of mosquito nets) led to similar effect estimates (not shown). We also explored the association between Giardia infection and diarrhea within age classes in children without multi-nutrients; although the numbers of cases within these strata was low, all estimates pointed towards a protective association (HR: 0.36 [0.13 to 1.01], 0.81 [0.43 to 1.55], 0.19 [0.04 to 0.84] in children aged 6–17 months, 18–35 months and 36–60 months, respectively). Also, when restricting the analysis to children who were infected at both surveys (n = 98) and those who were never infected (n = 294), we observed very similar patterns and came to the same conclusions (not shown).
Similar patterns were seen for cases of fever without localizing signs: Giardia infection was associated with an increase in the time to first episodes of such fevers among those receiving zinc or placebo, but not among those receiving multi-nutrients. Adjusted estimates of hazard ratios (including all events) are shown in figure 3.
The effect of the multi-nutrients on height-for-age z-scores, hemoglobin concentrations and plasma transferrin receptor concentrations measured at the second survey tended to be greater in children without Giardia infection at baseline, whereas supplements seemed to have little effect in those who tested positive at baseline (Figure 5). The overall effects were rather small, and statistical evidence for differences in effect between children with and without Giardia was weak (p-values for interaction between Giardia and multi-nutrients: 0.13 [height-for-age z-scores], 0.24 [hemoglobin concentrations] and 0.32 [plasma soluble transferrin receptor concentrations]). Adjustment for age led to similar conclusions (not shown). For other indicators of nutritional status (plasma concentrations of zinc, magnesium, cobalamin, folate and ferritin), there was no evidence that Giardia infection influenced the effect of supplementation.
Giardia intestinalis infection at baseline was associated with a marked reduction in the rates of subsequent diarrhea among children receiving zinc or placebo, but not in those receiving multi-nutrients. Multi-nutrient supplementation among children with Giardia infection resulted in disease rates similar to those found in uninfected children. Similar patterns were observed for cases of fever without localizing signs.
Substantial cross-over occurred between groups in the course of the study, and this may lead to underestimates of differences between children with and without Giardia infection at baseline. Our Kaplan-Meier analysis indicates, however, that the protective association occurred almost from the start of the follow-up period, when presumably few cross-over cases had occurred.
Our study is limited by the observational nature of our data, which does not allow a conclusion that the protection observed was caused by Giardia infection. Although this association was still present after adjustment for age and other potentially confounding factors, we cannot exclude the possibility that children with Giardia infection differed from their uninfected peers in other unmeasured characteristics that are prognostic for diarrhea (e.g. sanitation, or previous or current exposure to other gastro-enteric pathogens [26]–[28] or health-care seeking behavior). We did not measure breastfeeding behavior, but it is unlikely that this could have explained the protective association found against diarrhea in children not receiving multi-nutrients: even in older children (aged 36–60 months), who are unlikely to be breastfed, Giardia infection was associated with a reduction in hazard rates by 81% (16% to 96%). Treatment with artemether at baseline (Table S2) may have had some effect on the prevalence or intensity of Giardia infections, which would argue against a causal role of the parasite in the observed protection [29].
Our findings support the view that the parasite is not an important cause of diarrhea in our study population. G. intestinalis comprises various genotypes, and its prevalence and its association with diarrheal symptoms seems to vary with geographic areas [30]. A recent study showed that Giardia infection was associated with protection against diarrhea, whereas G. intestinalis assemblage A was associated with acute diarrhea [13]. Thus, due to variation in genotypes and environmental factors, our findings may not apply to other populations, and further research is also needed to determine G. intestinalis genotype in this population.
It is not inconceivable that Giardia infection protects against diarrhea, for example by competing with or suppressing other enteric pathogens, or by inducing changes in mucosal immunity [e.g. 11], [31]. Chronic or repeated exposure to non-pathogenic Giardia genotypes may have induced immunity against more pathogenic genotypes. This cannot fully explain the protective effect observed, however, because the magnitude of the protective association found probably exceeds the Giardia-attributable fraction of diarrhea. Giardia infection may also be a marker of an unknown factor (e.g. previous exposure to other pathogens) that leads to protection against both diarrhea and fever without localizing signs.
Whatever the cause, Giardia-associated protection was no longer present when giving multi-nutrients. This interaction is supported by the magnitude of the differences between the subgroup effects, whilst the probability that the interaction is due to chance seems low. We believe it is highly unlikely that the estimates of the interaction effect are biased: because the intervention was randomly allocated it is improbable that an external factor (e.g. health care seeking behaviour) would coincidentally bias disease rates strongly towards a Giardia-associated protection in the zinc and placebo group, but not in both multi-nutrient groups.
Further studies are needed to evaluate how supplemental micronutrients influence the composition, proliferation and pathogenicity of intestinal biota, and the interaction of these biota with their host. Iron deserves special attention in view of findings that it can modify the profile of gut microbiota towards potentially more pathogenic [32], or enhance the virulence and invasion of Salmonella enteritidis [33], whilst a recent study suggests that supplementation with bovine lactoferrin, an iron binding-protein, reduced the prevalence of Giardia among in Peruvian preschool children [34]. A meta-analysis of intervention trials with iron showed a slightly increased risk of diarrhoea due to iron supplementation [35].
Our study findings do not support treatment of Giardia infections in symptom-free children, and question the benefit of providing multi-nutrient supplements in populations frequently exposed to diarrheal diseases.
In conclusion, Giardia infection at baseline was associated with a marked reduction in the rates of subsequent diarrhea. Our data suggest that it is a marker for the response in diarrhea to multi-nutrient supplements, that should be taken into the account when analysing trials assessing the effect of multi-nutrient supplementation on diarrhea.
|
10.1371/journal.pgen.1005023 | Co-chaperone p23 Regulates C. elegans Lifespan in Response to Temperature | Temperature potently modulates various physiologic processes including organismal motility, growth rate, reproduction, and ageing. In ectotherms, longevity varies inversely with temperature, with animals living shorter at higher temperatures. Thermal effects on lifespan and other processes are ascribed to passive changes in metabolic rate, but recent evidence also suggests a regulated process. Here, we demonstrate that in response to temperature, daf-41/ZC395.10, the C. elegans homolog of p23 co-chaperone/prostaglandin E synthase-3, governs entry into the long-lived dauer diapause and regulates adult lifespan. daf-41 deletion triggers constitutive entry into the dauer diapause at elevated temperature dependent on neurosensory machinery (daf-10/IFT122), insulin/IGF-1 signaling (daf-16/FOXO), and steroidal signaling (daf-12/FXR). Surprisingly, daf-41 mutation alters the longevity response to temperature, living longer than wild-type at 25°C but shorter than wild-type at 15°C. Longevity phenotypes at 25°C work through daf-16/FOXO and heat shock factor hsf-1, while short lived phenotypes converge on daf-16/FOXO and depend on the daf-12/FXR steroid receptor. Correlatively daf-41 affected expression of DAF-16 and HSF-1 target genes at high temperature, and nuclear extracts from daf-41 animals showed increased occupancy of the heat shock response element. Our studies suggest that daf-41/p23 modulates key transcriptional changes in longevity pathways in response to temperature.
| Temperature is a critical environmental factor that affects ageing in both cold-blooded and warm-blooded species. In invertebrate animals, lifespan varies inversely with temperature, with higher temperature resulting in faster development but shorter lifespan. This phenomenon has been usually attributed to passive changes in metabolic rate, but recent work suggests that this process is regulated. In this study, we identify the co-chaperone protein p23 in the nematode C. elegans as an important modulator of longevity in response to temperature. Co-chaperones bind to client proteins to assist in their folding or stabilize their shape, thereby regulating their activity. Remarkably, deletion of p23 results in animals that are long lived at high temperatures and short lived at low temperatures relative to normal wild type animals. Our experiments indicate that p23 regulates lifespan through the neurosensory apparatus. These in turn impinge on key longevity regulators that mediate the transcriptional outputs of insulin/IGF, heat shock response and steroidal signaling. These studies suggest that complexes formed by p23 play a central role in regulating longevity in response to temperature.
| Temperature dramatically impacts the lifespan of ectotherms, with lower temperatures typically extending and higher temperatures shortening life [1–3]. The conventional view is that temperature passively affects the rate of chemical reactions and metabolism, thereby influencing species longevity. An emerging body of evidence, however, indicates that changes in longevity in response to temperature also reflect a regulated process entailing important organismal adaptations [4–6].
Like other ectotherms, the nematode Caenorhabditis elegans shows clear temperature dependent influences on development and lifespan [2]. At the normal cultivation temperature (20°C), animals typically live three weeks. At low temperature (15°C) they live approximately ten days longer, and at high temperature (25°C) ten days shorter. A handful of identified loci have been recently shown to mediate changes in lifespan in response to temperature. At warm temperatures, signaling from thermotaxis neurons promotes normal lifespan, and activates steroid hormone signaling to maintain longevity [4]. Animals compromised in thermotaxis genes or steroid hormone production display shortened life. In addition, animals will mount a heat shock response when exposed to acute heat stress, which helps preserve organismal viability. Thermotaxis neurons regulate the organismal heat shock response through the heat shock transcription factor HSF-1 [7]. Overexpression of hsf-1 and downstream chaperones can extend lifespan [8,9]. In contrast, longevity at cool temperatures requires the cold sensitive TRPA-1 channel [5], which works through Ca2+ signaling, protein kinase-C PKC-2, and serum and glucocorticoid kinase SGK-1 to activate the forkhead transcription factor DAF-16/FOXO, a crucial regulator of longevity. Evidently the mechanisms governing longevity at low or high temperature appear somewhat different and few components affect both [4].
Temperature effects on organismal longevity are not limited to ectotherms. Notably, temperature sensing neurons involved in homeostatic control of core body temperature affect murine lifespan [6]. Higher temperatures in these neurons trigger a lowering of the core body temperature and correlate with extended lifespan. Several long-lived mouse models including the Ames Dwarf, Growth Hormone Receptor knockout, and FGF21 transgenic mice have associated a lower core body temperature reminiscent of nutrient induced torpor [10,11]. Furthermore cold exposure extends lifespan and suppresses tumorigenesis in rats [12]. These studies suggest an intimate but relatively unexplored relationship between nutrient and thermal sensing, metabolism and longevity.
At the cellular level, chaperones and co-chaperones facilitate protein folding and assembly, often in response to thermal stress [13,14]. One such co-chaperone is p23 [15]. p23 complexes with the HSP90 chaperone and inhibits its ATPase activity [16–18], thereby stabilizing association with client proteins such as steroid receptor transcription factors, heat shock factor, and others [19–25]. These interactions play an important role in regulating transcriptional events. p23 also displays HSP90 independent chaperone-like activity [26], and is implicated in various cellular functions [27]. Additionally the protein reportedly harbors prostaglandin E2 synthase (PGS) activity in vitro [28], although this is debated [29]. Interestingly, p23 upregulation has also been implicated in tumorigenesis presumably through its interactions with HSF1, steroid receptors or growth regulated kinases [30]. Indeed, several anti-cancer drugs being developed target chaperones and co-chaperones, highlighting the clinical importance of these pathways. p23 knockout mice reveal an early peri-natal lethal phenotype, with defects in lung and skin development, but its organismal roles remain largely unknown [29,31].
Here, we used C. elegans to explore the role of p23 function in metazoan biology. We found that daf-41, the C. elegans homolog of co-chaperone p23, has a novel role in regulation of lifespan at both high and low temperatures, as well as in the formation of the long lived dauer stage. Remarkably, daf-41 mutants provoked longevity at warm temperatures, but short lived phenotypes at cold temperatures, thus equalizing the temperature response. daf-41 interacted with insulin signaling, heat shock factor, and steroidal signaling to regulate lifespan by distinct mechanisms at different temperatures. Our findings implicate daf-41 as a central player in the thermal regulation of longevity.
DAF-41/ZC395.10 is the sole C. elegans homolog of co-chaperone p23/cytosolic prostaglandin E synthase-3. DAF-41/p23 is broadly conserved in evolution, with approximately 45% peptide similarity to the human homolog (Fig. 1A). The protein contains an N-terminal HSP20-like co-chaperone domain, implicated in HSP90 binding, and a C-terminal region, implicated in intrinsic chaperone activity (Fig. 1B). daf-41(ok3052) is a deletion allele and presumptive null that removes co-chaperone and chaperone domains; daf-41(ok3015) harbors an in-frame deletion of the co-chaperone domain, leaving the chaperone domain intact (Fig. 1B).
Under conditions of food scarcity, overcrowding and elevated temperature, C. elegans larvae enter the dauer diapause, a stage specialized for survival and dispersal. We found that both daf-41(ok3052) and daf-41(ok3015) alleles were constitutively prone to form dauer larvae (Daf-c phenotype) at elevated temperature, yielding approximately 25% dauer larvae at 25°C, and nearly 100% dauer larvae at 27°C (Fig. 1C-D). By contrast, mutants containing deletions of other putative prostaglandin synthase homologs, pges-2/mPGES2, gst-4/PGDS, had little observable dauer phenotype at these temperatures. By inference, the co-chaperone function of daf-41 may be more important for Daf-c phenotypes.
Many Daf-c loci in the dauer signaling pathways, such as the daf-2/Insulin receptor (InsR) mutant, provoke resistance to various forms of stress [32,33]. Similarly daf-41 mutants displayed resistance to oxidative stress induced by H2O2 challenge comparable in strength to daf-2/InsR mutants (Fig. 1E). daf-41 mutants exposed to heat stress at 35°C were also significantly resistant (Fig. 1E). Other PGS mutants had little effect on oxidative and heat resistance. Altogether these results demonstrate that daf-41 is a novel Daf-c gene involved in stress tolerance.
To clarify daf-41 function, we examined its expression pattern. A promoter fusion to gfp, daf-41p::gfp, revealed expression most prominently in anterior and posterior neurons (Fig. 1F) including amphids (e.g. ASE, AWC, ASI, ADL) and phasmid sensory neurons, as well as peripheral neurons and ventral cord motorneurons. We also observed strong expression in body wall muscle and pharynx, as well as occasional expression in vulva, seam and intestine (S1A-B Fig).
In the dauer signaling pathways, environmental cues are detected by the neurosensory apparatus, and integrated by cGMP, TGF-β and insulin/IGF signaling. Ultimately these pathways converge on steroidal signaling to mediate the choice between arrest at the dauer diapause or continuous development to reproductive adult [34]. To understand where daf-41 acts in the dauer signaling pathways, we performed genetic epistasis experiments for dauer formation at 27°C. We first combined daf-41 Daf-c mutations with dauer formation defective (Daf-d) mutations in TGF-β signaling (daf-5/Ski), insulin/IGF signaling (daf-16/FOXO), and steroidal signaling (daf-12/FXR) [35–39]. As expected, null mutation of the steroid receptor, daf-12—a master regulator of dauer formation—completely suppressed daf-41 Daf-c phenotypes. Mutation of daf-16 partially suppressed daf-41 Daf-c phenotypes, while daf-5 mutation had little or no effect (Fig. 2A). At 25°C, daf-41 Daf-c phenotypes were also suppressed by daf-12. These experiments reveal that daf-41 acts upstream of daf-16, and daf-12, and in parallel to daf-5, to prevent dauer formation, resembling loci acting early in the dauer signaling pathways (Fig. 2H) [40].
We next analyzed interactions between daf-41 and neurosensory machineries of thermotaxis and chemotaxis, which variously affect dauer formation. We made double mutants with neurosensory transduction mutants (daf-10/IFT122, osm-1 and osm-3), which are Daf-d at 25°C and often Daf-c at 27°C [41,42]. We also examined thermotaxis mutants (pkc-1, ttx-3) that modulate dauer formation dependent on temperature and signaling pathway (e.g. ttx-3 suppresses daf-7/TGF-β Daf-c phenotype at 25°C, but enhances it at 15°C) [43,44].
We found that mutations in the chemotaxis genes, daf-10, osm-1 and osm-3, significantly suppressed Daf-c phenotypes of daf-41(ok3052) at 25°C and partially at 27°C (Fig. 2B). (Precedence for weaker Daf-c mutants suppressing stronger Daf-c mutants has been seen previously [45]). Consistent with a role proximal to the chemotaxis machinery, daf-41(ok3052) worms were chemotaxis defective for isoamylalcohol, benzaldehyde and 2,4,5-trimethylthiazoline (S2A Fig). By contrast, the thermotaxis loci did not appreciably affect daf-41(ok3052) Daf-c phenotypes at the examined temperatures (S3A-B Fig), although pkc-1 had a minor effect at 27°C. Thus daf-41 might work closely to chemotaxis loci and downstream or parallel to the thermotaxis loci.
Neurosensory cilia normally contact the environment through sensilla in head and tail, and typically fill with the lipophilic dye DiI. Mutants with defective neurosensory cilia structure, including daf-10 and osm-3 fail to fill with DiI because dendritic endings are not exposed [46]. We found that daf-41 null mutants had normal DiI filling similar to wild type worms (S2B Fig). Altogether these results show that daf-41 affects chemotaxis function, but not sensory cilia structure.
In its capacity as co-chaperone, p23 is known to complex with HSP90 [16–18,47]. The C. elegans homolog of HSP90, daf-21, functions early in the dauer signaling pathways at the level of chemosensory processing, similar to daf-41. daf-21(p673) is a weak gain-of-function (gf) amino acid substitution that renders animals Daf-c [48]. Genetic epistasis studies show that Daf-c phenotypes are suppressed by daf-10 and other chemosensory mutants [40]. cGMP signaling represents another branch of the dauer signaling pathways that works at a similar level as daf-21. Mutations in the daf-11/transmembrane guanylyl cyclase provoke Daf-c phenotypes, which are similarly suppressed by chemosensory mutants [40,48]. Finally both daf-21 and daf-11 exhibit chemosensory deficits [49].
Because daf-41/p23 may regulate dauer entry through the chemosensory axis, we sought to dissect genetic interactions between daf-41, daf-21 and daf-11. Although both daf-41(ok3052) and daf-21(p673) yielded Daf-c phenotypes, daf-41;daf-21 worms did not show synthetic enhancement of dauer formation at 25°C (Fig. 2C). Furthermore, daf-41 phenotypes were weakly suppressed by daf-21 mutation at 27°C (conversely daf-21 phenotypes were weakly enhanced by daf-41) (Fig. 2D). Likewise the daf-41(ok3052) null allele had little effect on dauer formation of daf-11(m47) at 22.5°C, where daf-11’s Daf-c phenotypes were partially penetrant (Fig. 2E). Whereas mutants in independent pathways typically give strong synergy, the modest interactions observed above suggest that daf-41 could work in a proximal or overlapping pathway with daf-21 and daf-11.
In mammals, HSP90 complexes are known to negatively regulate the activity of the heat shock transcription factor HSF1 [23,24]. Recently, it was reported that the hsf-1(sy441) missense mutation suppresses the Daf-c phenotypes of daf-21 and daf-11 at 23°C [50], suggesting that dauer formation depends upon active hsf-1(+). We therefore analyzed genetic interactions between hsf-1 and daf-41 around this temperature. Surprisingly, instead of suppression, hsf-1(sy441) enhanced daf-41 Daf-c phenotypes at 22.5°C (Fig. 2F). Similarly the egg laying defect of hsf-1(sy441) was strikingly enhanced in the daf-41 mutant background (S4 Fig). These synthetic interactions suggest that daf-41 and hsf-1 could work closely together, or identify parallel pathways converging on the same process (Fig. 2H).
Because daf-41 mutants showed clear temperature dependent dauer and stress resistance phenotypes, we wondered whether daf-41 would influence ageing at various temperatures. Increasing temperature is well known to reduce longevity in ectotherms, including wild type C. elegans (Fig. 3A-C, Table 1). daf-41 mutants exhibited an altered temperature dependent longevity, revealing a leveling out of the lifespan curves to those typically seen at 20°C: animals were long lived at 25°C, normal lived at 20°C, and short lived at 15°C relative to wild type (Fig. 3A-E). Other PGS mutants showed normal temperature dependent lifespan phenotypes (Fig. 3A-C, Table 1).
To determine if the observed longevity phenotypes were due to lesions in daf-41, we performed rescue experiments with the daf-41(+) transgene. As expected, daf-41(ok3052) mutant animals harboring daf-41(+) were readily rescued for Daf-c phenotypes, while the pges-2(+) control transgene expressed under daf-41 regulatory elements did not rescue (Fig. 3F). We next tested the influence of the transgenes on ageing. First, we found that overexpression of daf-41 and pges-2 had no effect on ageing in both N2 and daf-41(ok3052) at 20°C (Fig. 3G and S5A Fig). However, the daf-41 transgene clearly reversed the daf-41(ok3052) longevity phenotype at 25°C, and partially rescued the short-lived phenotype at 15°C (Fig. 3G and S5B-C Fig). In sum, these results reveal that daf-41(+) regulates longevity in response to temperature. Consistent with a role in thermal regulated processes, daf-41 mRNA increased modestly with temperature (Fig. 3H).
Because metabolic rates increase with temperature and vary inversely with longevity, we wondered whether daf-41 mutants altered mitochondrial metabolism. When we measured O2 consumption of daf-41 mutants at different temperatures, however, we saw no significant differences from wild type (S5D Fig). We also examined reproductive potential of daf-41(ok3052) worms and found that mutants produced nearly the same number of progeny as wild type at 20°C (S5E Fig).
To better understand the nature of daf-41 longevity, we next performed genetic epistasis experiments with known longevity regulators, including daf-16/FOXO, hsf-1/heat shock factor, daf-12/FXR steroid receptor, and daf-10/IFT122 [8,51–54]. Intriguingly, daf-16 was epistatic at all temperatures. At 25°C daf-16(mgDf50) completely abolished the longevity of daf-41(ok3052) (Fig. 4B, Table 2). At 20°C daf-16 mutation reduced lifespan of daf-41(ok3052) similar to wild type N2 (Fig. 4A). At 15°C, daf-16 mutation did not further reduce the short lifespan of daf-41(ok3052) mutants (Fig. 4C). Moreover, daf-16 mutation had the effect of restoring temperature dependent regulation of lifespan to daf-41 mutants (Fig. 4D, Table 2). Although daf-16 mRNA levels were little affected by daf-41, several major target genes of daf-16, including sod-3, dod-3, and lipl-4 [55] were elevated in daf-41(ok3052) in a daf-16 dependent manner (Fig. 4E). No differences in DAF-16 nuclear localization were seen between WT and daf-41(ok3052), since both showed moderately elevated translocation at 25°C, (S6A-B Fig). These data support the notion that daf-41(+), either directly or indirectly inhibits the activity but not localization of DAF-16/FOXO at elevated temperatures. We therefore performed qRT-PCR analysis for insulin like peptides and found that expression of several ins genes were changed by temperature shift and in daf-41(ok3052) worms. In particular, ins-1, ins-5, ins-7, ins-10, ins-11, ins-12, ins-17, ins-18, ins-27 and ins-37 were up-regulated in daf-41(ok3052) worms at 25°C, and only ins-13 was down-regulated at 25°C (S7 Fig), consistent with modulation of IIS.
We next examined genetic interactions with the heat shock transcription factor hsf-1, which is required for normal lifespan at various temperatures [8,9]. Longevity of daf-41(ok3052) was completely abolished in the hsf-1(sy441) background at 25°C, suggesting the two genes work in a unified pathway (Fig. 5B, Table 2). By contrast, the daf-41;hsf-1 double mutant showed additive short-lived phenotypes at 15°C and 20°C (Fig. 5A, C and D, Table 2) presumably because hsf-1(sy441) is non-null. Although hsf-1 itself was not transcriptionally regulated, major target genes of hsf-1, including hsp-16.2, hsp-70, and hsp-4 [56], showed augmented expression in the daf-41 background compared to wild type at 25°C (Fig. 5E). These results indicate that daf-41(+) directly or indirectly inhibits the activity of HSF-1 at 25°C, to influence transcription and longevity. Recently, it has been reported that HSF-1 forms nuclear foci in response to heat shock but not by reduced IIS [57]. Similarly, we failed to detect foci formation of HSF-1 at 20°C and 25°C daf-41(ok3052) mutants (S8A-B Fig).
Given the genetic interactions with hsf-1 for dauer formation, longevity, and gene transcription described above, we wondered whether daf-41 affects assembly of HSF-1 nuclear complexes. To measure this, we prepared nuclear extracts from wild type and daf-41 mutants and performed electrophoretic mobility shift assays on the heat shock factor response element as established previously [58]. First we found that daf-41 mutation had no effect on mRNA or protein levels of HSF-1 (S8C Fig). Second, we observed that daf-41 nuclear extracts showed a clear and reproducible 1.5–2 fold higher occupancy of the heat shock factor response element, compared to WT controls (Fig. 5F). These results suggest that DAF-41(+) normally affects the assembly or disassembly of HSF-1 transcriptional complexes.
Because p23 and HSP90 are known to work together, we asked whether they would have similar effects on lifespan. At 25°C, both daf-41(ok3052) and daf-21(p673)gf worms were longer lived than wild type; however, the daf-41;daf-21 double mutant had longevity phenotypes more similar to the daf-21 single mutant (Fig. 5G, S9A-B Fig, Table 2). The convergent behavior at 25°C suggests they could work together at this temperature, similar to dauer. By contrast at 15°C, daf-21(p673) worms were extremely long-lived; this longevity was additive to the short-lived phenotype of daf-41(ok3052) mutant animals. The distinct behavior of daf-41/p23 and daf-21/hsp90 at 15°C, suggests they may regulate lifespan at this temperature through different pathways (S9C Fig).
The steroid receptor DAF-12 also influenced the longevity phenotypes of daf-41(ok3052) (Fig. 6A-D, Table 2). At 25°C, daf-12 mutation partially reduced the longevity of daf-41 mutants in an additive manner, but was not epistatic, i.e. daf-12 daf-41 did not give the same life span as daf-12 itself. The expression levels of the DA biosynthetic gene, daf-36/Rieske, and two DAF-12 target genes, cdr-6 and fard-1, were also increased at higher temperatures. Therefore, daf-12(+) could work in parallel, or function as a minor branch of daf-41 signaling to promote longevity. More interestingly at 15°C, daf-12 mutation suppressed the short lived phenotypes of daf-41, restoring near normal life span. At this temperature, cdr-6 and fard-1 genes decreased in expression in daf-41 mutants relative to WT (Fig. 6E). These interactions suggest that daf-41(+) could prevent life shortening properties of DAF-12 steroidal signal transduction at lower temperatures.
We also examined interactions with daf-10/IFT122, which affects neuronal cilia, and is long lived [46,54]. At 15°C and 20°C, daf-10 mutation did not affect daf-41 and showed additive phenotypes, i.e. daf-10 increased lifespan of both N2 and daf-41(ok3052) independent of temperature (Fig. 7A, S10A-C Fig, Table 2). At 25°C, however, daf-41 and daf-41;daf-10 had similar degrees of lifespan extension, suggesting that the two activities might converge on a common process.
Finally we examined the effect of genes involved in thermotaxis. These genes are implicated in the systemic heat shock response [7]. Moreover, recent work has shown that longevity arising from RNAi knockdown of the pat-4/integrin-linked kinase depends completely on genes involved in thermotaxis (ttx-3 and gcy-8), hsf-1, but not daf-16 [59]. For our studies, we used the soluble guanylyl cyclase (gcy) triple mutant, gcy-23(nj37) gcy-8(oy44) gcy-18(nj38), which is defective in thermotaxis due to signaling defects in the thermotaxis neurons [60]. As expected, we observed the gcy triple mutant to be short-lived at 25°C (Fig. 7B, S10F Fig, Table 2). Unexpectedly, the gcy triple mutant equally shortened the lifespan of daf-41 and wild type, suggesting parallel pathways. At 15°C, the gcy triple mutant did not further shorten daf-41 lifespan, possibly suggesting a convergent mechanism (S10D-E Fig, Table 2).
How temperature influences animal lifespan is not well understood. In this work, we demonstrate that daf-41/p23 co-chaperone PTGES3 homolog modulates longevity in response to temperature, and regulates entry and exit from the long-lived dauer stage. On the one hand, daf-41(+) promotes normal adult longevity at cold temperature (15°C). On the other hand, daf-41(+) limits lifespan at warm temperature (25°C), but has little influence on longevity at 20°C. Thus the overall effect seen in daf-41 mutants is an equalization of the lifespan curve so that animals appear as if they are living at 20°C. Surprisingly, no difference in respiratory rates (O2 consumption) between wild type and daf-41 animals was detected at these temperatures, suggesting that daf-41’s effect on longevity is likely through mechanisms independent of mitochondrial metabolism. These intriguing observations reveal that daf-41 plays a unifying role in regulatory mechanisms that modulate lifespan in reaction to temperature, and imply that co-chaperone/chaperone complexes may mediate this response.
Although the detailed molecular mechanism by which daf-41 mediates these temperature dependent effects is not entirely understood, our genetic epistasis experiments suggest that daf-41 may directly or indirectly impinge on several transcriptional longevity regulatory mechanisms, including DAF-16/FOXO, HSF-1, and DAF-12 steroidal signaling, to regulate lifespan in response to temperature. Several lines of evidence argue that daf-16/FOXO is a critical mediator in these circuits. At warm temperature, lifespan extension of daf-41 mutants was abolished in the daf-16 mutant background, at low temperatures the daf-41;daf-16 double mutants did not live shorter than daf-41 alone, suggesting convergent mechanisms. Overall daf-16 mutation restored temperature dependent lifespan regulation to daf-41 mutants. As further support for a link to FOXO function, daf-16 partly suppressed daf-41 Daf-c phenotypes, placing daf-16 downstream for both dauer formation and longevity. Consistent with working in a signaling pathway, we found that daf-41(+) negatively regulated transcription of DAF-16 targets (sod-3, dod-3, lipl-4) in response to warm temperature although no clear effect was seen at low temperatures with these genes. Several insulin-like peptides were up or down regulated in daf-41 mutants in response to warm temperature, suggesting that daf-41 has the potential to affect IIS. However, we could not detect obvious differences in DAF-16 nuclear localization, as is often seen with mutations that only weakly affect signaling (Henderson et al. 2001). Conceivably daf-41 could regulate components of IIS or DAF-16 complexes themselves.
Heat shock factor, hsf-1, often works in tandem with daf-16/FOXO [8,9] and consistently, was also required for daf-41 induced longevity at elevated temperatures. Accordingly, several hsf-1 target genes (hsp-16.2, hsp-70, hsp-4) were expressed in daf-41 mutants under these conditions. Although HSF-1 mRNA or protein levels were unchanged in the daf-41 background, nuclear extracts from daf-41 mutants showed increased occupancy of the heat shock response element. Importantly, this result suggests that daf-41(+) influences either directly or indirectly the assembly/disassembly of the HSF-1 complex. This finding is consistent with observation the p23-HSP90 complex has been shown to inhibit the activity of the HSF-1 complex in other systems [23,25]. The observed short-lived phenotypes of daf-41 with hsf-1 at low temperatures were additive, possibly because the hsf-1 allele is temperature sensitive and non-null [56].
daf-41’s interactions with nuclear receptor daf-12 suggest an intriguing role for steroid signaling. At 15°C, daf-12 mutation suppressed daf-41 short-lived phenotypes, suggesting that daf-41 modulates life extending effects of daf-12. Steroid signaling has previously been implicated in lifespan regulation albeit at high temperatures: daf-9 hypomorphs as well as thermotaxis mutants are particularly short-lived at 25°C, in a manner dependent upon daf-12 [4]. At these temperatures, thermotaxis loci as well as hsf-1(+) promote expression of the hormone biosynthetic gene daf-9/CYP27A1 [4,50]. These observations suggest that normal lifespan at 25°C depends on adequate stimulation of steroidal signaling, thereby preventing life shortening activities of the unliganded DAF-12. At 25°C, daf-12 mutation shortened the life span of daf-41 to a similar extent as wild type, suggesting parallel pathways. Alternately daf-12 could contribute partially toward daf-41 longevity, since daf-12 target genes as well as daf-36 increase expression at this temperature. At this point, it is unclear whether the activating or repressing functions of the receptor are responsible for longevity.
In mammals, p23 together with HSP90, immunophilins and other chaperones, binds to unliganded steroid receptors in the cytosol, maintaining the receptor in a primed state competent for rapid ligand binding; upon binding hormone the steroid receptor enters the nucleus where interacts with transcriptional coregulators. Thereafter, p23 helps disassemble transient transcriptional complexes to facilitate hormone sampling and repeated rounds of transcription [15,20–22,26]. The p23-HSP90 complex also regulates type II nuclear receptors, such as the thyroid receptor [20], which constitutively reside in the nucleus, a situation more resembling that of DAF-12. Further studies on the role of steroidal signaling at low and high temperature may help clarify the whether similar mechanisms are at work in C. elegans.
Interactions with chemosensory mutants suggest a role within the sensory apparatus. Indeed, daf-41 longevity at warm temperatures converged with daf-10/IFT122, and both mutants stimulate daf-16, suggesting they could work through a similar chemosensory mechanism [54]. Consistent with a proximal role in chemosensory signaling, daf-41 mutants exhibited chemosensory deficits and chemosensory mutants suppressed the daf-41 Daf-c phenotypes. Accordingly, daf-41 was expressed in several chemosensory neurons including ASE, AWC, ADL, and ASI but was not obviously expressed in the main thermosensory neurons AFD and AIY, although AWC and ASI also reportedly contribute to thermosensation [61]. Ultimately, tissue specific dissection of daf-41 activities may help illuminate what cells mediate these interactions. It is also noteworthy that daf-41 did not further shorten the lifespan of the gcy triple mutant at 15°C, suggesting it could work through thermotaxis circuits at low temperature.
p23 is known to interact directly with HSP90 to modulate various client proteins [15,16], and recently, C. elegans p23 and HSP90 have been shown to physically interact in vitro [47]. Consistent with the possibility of working together, both daf-41/p23 and daf-21/HSP90 act at the level of chemosensory processing with respect to dauer formation (this work; [49]), and modulated one anothers’ dauer and longevity phenotypes at 25°C. That daf-41 null and daf-21 gain of function have similar phenotypes could indicate that the wild type activities work in opposition for these processes. On the other hand, daf-41 and daf-21 had several divergent phenotypes: whereas daf-41 Daf-c phenotypes were enhanced by hsf-1 mutation (Fig. 2H, this work), daf-21 Daf-c phenotypes were suppressed [50]. Furthermore, whereas daf-41 mutants were short-lived at 15°C, daf-21 mutants were long-lived, and regulated life span differently. Thus some p23 phenotypes might arise from HSP90 dependent as well as independent processes, as has been noted previously [26,27]. We interpret these experiments with caution since daf-21 mutations are gain-of-function and non-null. Intriguingly, HSP90 in Candida albicans has been shown to regulate the switch to pseudohyphal growth in response to temperature [62], perhaps analogous to the thermal regulation of dauer formation or longevity.
Based on the observed genetic interactions we suggest the following model for longevity regulation. At elevated temperatures, p23 directly or indirectly inhibits the transcriptional activities of HSF-1 and perhaps DAF-16, thus limiting lifespan at these temperatures, while DAF-12/FXR and thermotaxis signaling work in parallel (Fig. 7C). Given the convergence with daf-10 longevity, we suggest that daf-41(+) might work through the chemosensory apparatus to impinge upon DAF-16. More speculatively, at lower temperatures, p23 stimulates DAF-16/FOXO and inhibits DAF-12/FXR, possibly through the thermotaxis apparatus, while HSF-1 works in parallel to promote lifespan extension.
Given the intriguing genetic interactions described here, it would be interesting to investigate whether DAF-41 and/or HSP90 binds and regulates the activities of HSF-1, DAF-16, or DAF-12 by protein-protein interaction. Alternately DAF-41 could interact with upstream components of these signaling pathways, including kinases, temperature sensitive channels, guanylyl cyclases, or cilia proteins. Conceivably, thermal regulation in these circuits could result from thermal influences on protein-protein interactions.
The results here and elsewhere reveal that regulation of longevity at different temperatures works by distinct mechanisms. This is perhaps not surprising, given that different stresses challenge the organism at low and high temperatures. Moreover, ectotherms must have evolved optima for growth and reproduction within a temperature range. With this in mind, we suggest that daf-41 could play distinct roles at low and high temperatures. Alternately daf-41 may be part of an adaptive response to temperature in which optima shifted towards higher temperatures have a consequent tradeoff at lower temperatures, and vice versa. Further elucidation of daf-41/p23 complexes and physiology in C. elegans should help illuminate the mechanism of thermal regulation of metazoan longevity.
Strains were obtained from the Caenorhabditis Genetics Center (CGC, USA) and National Bio Resource Project (NBRP, Japan). All strains were outcrossed at least 4 times to wild type N2 before further analysis. All strains in this work are itemized in S1 Table. Genotypes of mutants were confirmed by PCR, sequencing and phenotyping. Primer sets for genotyping are listed in S2 Table.
All lifespans were measured as previously described [52]. Strains for ageing experiments were maintained at the respective cultivation temperature of 15°C and 20°C for more than 3 generations before analysis, unless indicated otherwise. Progeny were collected by egg laying and bleaching. Ageing experiments were performed with and without 2′fluoro-5′deoxyuridine (FUdR, Sigma) to prevent contamination of next generation progeny and to reduce bagging phenotypes of Egl mutants. Strains were treated with FUdR as described previously [63]. For ageing experiments at warm temperature, worms were cultivated at 20°C until the young adult stage, then moved to 25°C to start the ageing analysis in order to bypass dauer formation and minimize internal hatching phenotypes. Each experiment started with more than 150 worms. Sterile, escaped, internally hatched, and exploded worms were censored on the day of loss. Experiments were performed at least 3 times and the mean lifespan calculated. Worms were transferred onto new OP plates every 2–3 days from the end of reproductive period, and scored for survival every 2–4 days. Statistical analyses were performed by Kaplan–Meier method with GraphPad Prism software (GraphPad Software, Inc.)
The regions of the daf-41/ZC395.10 promoter, coding region and 3' UTR, as well as pges-2 coding region and its 3' UTR were amplified by PCR. The promoter region was inserted in front of gfp in the L3781 plasmid and coding regions inserted after gfp. daf-41p::gfp::daf-41::3'UTR and daf-41p::gfp::pges-2::3'UTR plasmids were confirmed by sequencing and co-injected into daf-41(ok3052) worms with coel::RFP plasmid as an injection marker. Both transgenic strains were outcrossed with N2 to generate the following strains: N2; daf-41p::gfp::daf-41::3'UTR, N2;daf-41p::gfp::pges-2::3'UTR, daf-41(ok3052); daf-41p::gfp::daf-41::3'UTR, and daf-41(ok3052);daf-41p:: gfp::pges-2::3'UTR.
For dauer formation assays, all strains were maintained at 20°C and eggs collected by egg laying or bleaching. Greater than 50 eggs were transferred onto 3cm OP plates and cultured at 20°C, 22.5°C, 25°C and 27°C, respectively. Dauer formation fraction was typically scored at 60 hrs, and dauer exit ratio was scored at 84 hrs.
All strains were maintained at 20°C and Day 1 young adults were used for analysis. For heat stress analysis, 50–100 adult worms were transferred onto fresh 6 cm OP plates and shifted to 35°C. Fraction survival was scored 8 hrs after heat shock. For oxidative stress analysis, 50–100 adult worms were collected by washing off plates and transferred into 24 well plastic plates filled with 20mM of hydrogen peroxide (SIGMA) in M9 buffer. Experiments were performed with 5 biological replicates with 3 technical replicates for each mutant.
Synchronized worms were prepared at different temperatures. Worms were collected in TRIzol (Invitrogen) at L4 stage and frozen in liquid nitrogen. Total RNA was extracted by RNeasy Mini kit (QIAGEN) and Superscript III First Strand Synthesis System (Invitrogen) was used for cDNA generation. qRT-PCR was performed with Power SYBR Green master mix (Applied Biosystems) on a 7900HT Fast Real-Time PCR System (Applied Biosystems). ama-1 was used as internal control for mRNA quontification. For each analysis, qRT-PCR was performed with at least three biological replicates. Primer sequences are listed in S2 Table.
WT and daf-41(ok3052) worms were harvested and frozen down at day 1 of adulthood. Frozen worm pellets were first homogenized in an equal volume of 2X NPB buffer (20 mM HEPES, pH 7.6, 20 mM KCl, 3 mM MgCl2, 2mM EDTA, 0.5 M sucrose, 1 mM dithiothreitol, protease inhibitors, and phosphatase inhibitors) using a Kontes Pellet Pestle tissue grinder. The suspension was then centrifuged (4000 g, 5 min, 4°C) and the pellets were further homogenized by 20 strokes with pestle A of a Dounce homogenizer. The pellet was resuspended in NPB buffer with 0.25% NP-40 and 0.1% Triton-X100, centrifuged again, and washed three more times with the same buffer. The nuclear pellet was extracted with four volumes of HEG buffer (20 mM HEPES, pH 7.9, 0.5 mM EDTA, 10% glycerol, 0.42 M NaCl, 1.5 mM MgCl2, and protease inhibitors) at 4°C for 45 min. Finally, the nuclear fraction was collected by centrifugation at 14,000 g for 15 min at 4°C. Protein concentrations were determined with a Bradford assay kit (Bio-Rad, Hercules, CA).
EMSAs were carried out as previously described [58]. In brief, 1 μg of nuclear extract (described above) was mixed with 1 mg/mL of poly (dI-dC) and 1 nM of a biotin-labeled oligonucleotide containing the heat-shock element (HSE) sequence of hsp-16.1 [58], and incubated for 15 min at room temperature in binding buffer [20 mM HEPES, pH 7.6, 5 mM EDTA, 1 mM dithiothreitol, 150 mM KCl, 50 mM (NH4)2SO4, and 1% Tween 20 (v/v)]. After incubation, the samples were separated by native 3.5% PAGE and the HSF-1::HSE DNA complexes were visualized using a LightShift Chemiluminescent EMSA kit (Pierce, Rockford, IL).
Additional details for DiI staining, chemotaxis analysis, oxygen consumption measurement and microscopy analysis are described in S1 Text.
|
10.1371/journal.pbio.1001326 | Novel Role of NOX in Supporting Aerobic Glycolysis in Cancer Cells with Mitochondrial Dysfunction and as a Potential Target for Cancer Therapy | Elevated aerobic glycolysis in cancer cells (the Warburg effect) may be attributed to respiration injury or mitochondrial dysfunction, but the underlying mechanisms and therapeutic significance remain elusive. Here we report that induction of mitochondrial respiratory defect by tetracycline-controlled expression of a dominant negative form of DNA polymerase γ causes a metabolic shift from oxidative phosphorylation to glycolysis and increases ROS generation. We show that upregulation of NOX is critical to support the elevated glycolysis by providing additional NAD+. The upregulation of NOX is also consistently observed in cancer cells with compromised mitochondria due to the activation of oncogenic Ras or loss of p53, and in primary pancreatic cancer tissues. Suppression of NOX by chemical inhibition or genetic knockdown of gene expression selectively impacts cancer cells with mitochondrial dysfunction, leading to a decrease in cellular glycolysis, a loss of cell viability, and inhibition of cancer growth in vivo. Our study reveals a previously unrecognized function of NOX in cancer metabolism and suggests that NOX is a potential novel target for cancer treatment.
| Glycolysis is a cytoplasmic metabolic process that produces energy from glucose. In normal cells, the rate of glycolysis is low, and glycolysis products are further processed in the mitochondria via oxidative phosphorylation, a very efficient energy-producing process. Cancer cells, however, display higher levels of glycolysis followed by cytoplasmic fermentation, and reduced levels of oxidative phosphorylation. It was thought that increased glycolysis is associated with mitochondrial dysfunction, but how these phenomena are functionally linked was not known. Understanding how these processes are regulated will be essential for developing more effective anti-cancer therapies. Here, we show that induction of mitochondrial dysfunction by either genetic or chemical approaches results in a switch from oxidative phosphorylation to glycolysis. We further show that NADPH oxidase (NOX), an enzyme known to catalyze the oxidation of NAD(P)H, also plays a critical role in supporting increased glycolysis in cancer cells by generating NAD+, a substrate for one of the key glycolytic reactions. Inhibition of NOX leads to inhibition of cancer cell proliferation in vitro and suppression of tumor growth in vivo. This study reveals a novel function for NOX in cancer metabolism, explains the increased glycolysis observed in cancer cells, and identifies NOX as a potential anti-cancer therapeutic target.
| Development of selective anticancer agents based on the biological differences between normal and cancer cells is essential to improve therapeutic selectivity. Increased aerobic glycolysis and elevated oxidative stress are two prominent biochemical features frequently observed in cancer cells. A metabolic shift from oxidative phosphorylation in the mitochondria to glycolysis in the cytosol in cancer was first described some 80 years ago by Otto Warburg, who later considered such metabolic changes as “the origin of cancer” resulting from mitochondrial respiration injury [1]. It is now recognized that elevated glycolysis is a characteristic metabolism in many cancer cells. In fact, active glucose uptake by cancer cells constitutes the basis for fluorodeoxyglucose-positron emission tomography (FDG-PET), an imaging technology commonly used in cancer diagnosis. In addition, cancer cells exhibit elevated generation of reactive oxygen species (ROS), which disturb redox balance leading to oxidative stress [2]. However, despite these long-standing observations and clinical relevance, the biochemical/molecular mechanisms responsible for such metabolic alterations and their relationship with mitochondrial respiratory dysfunction remain elusive. Identification of the major molecular players involved in the metabolic switch in the context of mitochondrial dysfunction in cancer cells is important for understanding the underlying mechanisms and developing more effective treatment strategies.
For many years, studies of mitochondrial respiratory defect usually involve the use of ρ° cells, in which mitochondrial DNA (mtDNA) deletion is generated by chronic exposure of cells to the DNA-intercalating agent ethidium bromide [3]. While successful, the use of ρ° cells generated by this method as a model for metabolic study has potential complications due to possible nuclear DNA damage by ethidium bromide and thus may compromise data interpretation [4]. To investigate the relationship between mitochondrial dysfunction and alterations of cellular metabolism, it is important to establish a model system in which the mitochondrial function can be regulated without significant impact on the nuclear genome. Mitochondrion DNA polymerase gamma (POLG) is a key enzyme responsible for the replication of mtDNA [5],[6], which encodes for 13 critical components of the respiratory chain. Thus, it is possible to specifically impact the mitochondrial respiratory function by selectively suppressing POLG, which is not involved in nuclear DNA replication. Indeed, a dominant negative form of POLG (POLGdn), which contains a point mutation (D1135A) in the coding sequence, has been previously identified and demonstrated to have a negative impact on mtDNA replication, causing respiratory defect after transfection [7],[8]. Thus, it is possible to use a gene transfection strategy to selectively impact mitochondrial function without affecting nuclear DNA.
NOX are a group of membrane-associated enzymes capable of oxidizing NADPH or NADH to NADP+ or NAD+, leading to generation of superoxide by one-electron reduction of oxygen [9]. There are seven members of the NOX family in the human genome with unique patterns of cellular expression and conserved structure. They include NOX1 (also called Mox1), NOX2 (gp91phox), NOX3, NOX4, NOX5, and the dual oxidases DUOX1 and DUOX2. Activation of most NOX complexes requires proper assembling of multiple protein components, including p22phox (CYBA), Rac-GTPase (Rac1 and Rac2), p47phox, p67phox, p40phox, NOX organizer 1 (NOXO1), and NOX activator 1 (NOXA1) [10],[11]. NOX enzymes have been implicated in host defense, regulation of gene expression, ROS generation, and redox signaling [12]. In many cancers, NOX activities are increased and mRNAs are overexpressed [13]–[16], but the precise functions of NOX in cancer cell metabolism remain unclear.
In the current study, we adapted the molecular strategy using POLGdn to generate a tetracycline-inducible cell system as previously described [8] and used this model system and cancer cells with compromised mitochondrial respiration to investigate the relationship between mitochondrial respiratory dysfunction and metabolic alterations and to identify the key molecular players in the metabolic switch from oxidative phosphorylation to aerobic glycolysis. We discovered an unexpected metabolic function of NOX, which is important for maintaining high glycolytic activity in cells with mitochondrial respiratory dysfunction. We also found that cancer cells with mitochondrial dysfunction due to the expression of oncogenic Ras or a loss of p53 consistently exhibited elevation in NOX activity and were highly sensitive to NOX suppression. Importantly, the significant increase in the expression of p22phox, a main component of NOX complex, was found in human pancreatic carcinoma, where K-Ras aberrant activation is prevalent. Furthermore, suppression of NOX exhibited significant antitumor activity in vivo, suggesting that NOX can be a potential target for cancer therapy.
To investigate the relationship between mitochondrial dysfunction and metabolic changes, we first established an experimental cell model where the mitochondrial respiratory status can be regulated via a tetracycline-inducible system. This model is based on the fact that mtDNA replication is catalyzed by DNA polymerase γ (POLG) and that POLGdn can abolish mtDNA replication leading to respiration defects [7],[17]. A 4-kb DNA fragment containing the full-length POLGdn gene with a D1135A mutation was constructed into a mammalian expression vector pcDNA4/TO (Figure 1A), which was then transfected into human T-Rex 293 cells to generate a Tet/on inducible system as described in the Materials and Methods. As shown in Figure 1B, in the absence of doxycycline (Tet/off), the endogenous POLG was readily detected by POLG antibody, but the expression of exogenous POLGdn was not detectable (absence of the FLAG signal), suggesting that this experimental system was tightly controlled without detectable leakage. Addition of doxycycline (Tet/on) to the culture medium induced the expression of POLGdn (detected by anti-FLAG), which remained expressed for over 2 wk in the presence of doxycycline. The induced expression of POLGdn was also evident by the increase of the band intensity detected by anti-POLG antibody (Figure 1B).
The expression of POLGdn led to a severe decrease of mtDNA synthesis, as evidenced by a dramatic decrease of mtDNA 2 d after doxycycline induction (Figure 1C). To determine the mtDNA-encoded gene expression, Northern blot analysis was used to measure the level of mtDNA-encoded cytochrome c oxidase subunit II (COII) RNA. Figure 1D showed that the expression of COII RNA was decreased on day 2, almost depleted on day 6, and disappeared by day 10 after POLGdn induction. Western blot analysis showed a corresponding depletion of COII protein in a time-dependent manner (Figure 1E).
To evaluate the metabolic alterations subsequent to POLGdn induction, we first measured cellular oxygen consumption as an indicator of mitochondrial respiratory capacity [18]. Expression of POLGdn caused a time-dependent decrease in oxygen consumption, detectable on day 2 and dramatically decreased on day 6 (Figure 2A). Associated with this decrease in mitochondrial respiration, the Tet/on cells became highly glycolytic, as evidenced by a significant increase in glucose uptake (Figure 2B) and elevated production of lactate, a metabolic product of glycolysis (Figure 2C). Consistent with the active glycolysis, the Tet/on cells showed an upregulation of hexokinase II (HKII), a rate-limiting enzyme of the glycolytic pathway (Figure 2D). This increase in glycolysis seemed effective in compensating the decrease of ATP production in the mitochondria, since the overall cellular ATP levels only decreased moderately (Figure 2E), even on day 12 when respiration was severely suppressed.
Since the mitochondrial respiratory chain is a major site of cellular ROS generation, we examined if induction of mitochondrial dysfunction by POLGdn could lead to a change in cellular ROS. Flow cytometry analysis of cells stained with hydroethidine (HET), a relatively specific chemical probe for superoxide (O2−) [19] showed that the cellular O2− level was significantly decreased in the Tet/on cells at day 12 (Figure 3A). Consistently, MitoSox Red, a mitochondrial O2− indicator, revealed a significant decrease in mitochondrial O2− in the Tet/on cells (Figure 3B). In contrast, a general redox-sensitive probe DCF-DA, which detects cellular H2O2 and other ROS, revealed a significant increase in cellular ROS in the Tet/on cells (Figure 3C).
Because O2− can be converted to H2O2 by the mitochondrial superoxide dismutases (SOD2) or the cytosolic SOD1, we tested the possibility that the increase in cellular H2O2 and the decrease in O2− observed in the Tet/on cells might be a consequence of altered SOD activity. Western blot analysis showed that the mitochondrial SOD2 remained unchanged (Figure 3D), whereas the basal expression of cytosolic SOD1 was abundant in the Tet/off cells and was further increased by day 4 after POLGdn induction (Figure 3E). Concurrently, SOD1 activity was also increased (Figure 3F). Taken together, these data suggest that the increase in cellular ROS detected by DCF-DA was most likely due to elevated generation of O2− outside the mitochondria, and such cytosolic O2− was then converted to H2O2 by the elevated SOD1.
Since NOX is a membrane-associated enzyme capable of generating ROS outside the mitochondria [20], we then measured the membrane-associated NOX activity of the Tet/on cells in comparison with Tet/off cells, using standard NOX assay described previously [21],[22]. The mitochondrial respiratory defective cells consistently exhibited a significant increase in NOX activity, which was detected 2 d after POLGdn induction and remained high as long as the cells were maintained in Tet/on stage (Figure 4A). This increase in NOX activity was inhibited by 10 µM diphenyleneiodonium (DPI), a known inhibitor of NOX [23], to less than 10% of the original NOX activity (Figure 4B). In contrast, pharmacologic inhibitors of other ROS-generating molecules including NOS inhibitor Nω-nitro-L-arginine mrthyl ester hydrochloride (L-NAME, 100 µM), the mitochondrial respiratory chain complex I inhibitor rotenone (20 µM), and the xanthine oxidase inhibitor oxypurinol (100 µM) showed no effect on the NOX activity assay (Figure 4B). To determine if the increased NOX activity was due to an increase in gene expression, we used semi-quantitative RT-PCR to evaluate possible changes in RNA expression of various NOX components and showed that the mRNA levels of NOX1, NOXA1, and p47phox were significantly increased (Figure S1A–B, Text S1). Time-course analysis of NOXA1 and p47phox expression showed that the mRNA levels increased 2 d after Tet/on (Figure S1B), concurrent with the timing of NOX activity increase. This elevated gene expression was further quantitatively confirmed by qRT-PCR assay (Figure 4C).
The elevated NOX in cells with mitochondrial dysfunction induced by POLGdn suggests that NOX upregulation might be functionally important for these cells. We then used DPI to test if the POLGdn-expressing cells with mitochondrial respiratory defects might be more vulnerable to NOX inhibition. As shown in Figure 4D, the Tet/on cells were significantly more sensitive to DPI treatment than the Tet/off cells, evident by a substantial loss of mitochondrial integrity (ability to retain rhodamin-123), a decrease in cell viability (annexin-V/PI double staining, Figure 4E), and an inhibition of cell growth (Figure S1C). To test whether the elevated NOX gene expression and NOX activity were only limited to the POLGdn Tet/on system, we compared the respiration defective (ρ°) cell line HL60-C6F (C6F cells) [24] with its mitochondrial competent parental HL60 cell line and observed that the mitochondrial respiration-defective C6F cells showed a significant increase in NOX activity (Figure 4F) and elevated expression of NOX1, NOXA1, and p47phox (Figure 4G). Consistently, C6F cells were also more sensitive to DPI than the parental HL60 cells (Figure 4H).
The above observations suggest that the up-regulation of NOX might be important for the viability of cells with mitochondrial dysfunction. Since respiration-defective cells are highly dependent on glycolysis for cell survival, we tested the potential role of NOX on glucose metabolism by evaluating the effect of NOX knockdown on glucose uptake, ATP generation, and cellular NAD+ level. siRNA was used to specifically knockdown the expression of the critical NOX components NOX1 and p22phox in both Tet/on and Tet/off cells (Figure S2A and S2B). Interestingly, suppression of NOX by siRNA selectively impact cells with mitochondrial dysfunction (Tet/on), evidenced by a significant decrease in glucose uptake and reduced ATP generation, while cells with competent mitochondrial function (Tet/off cells) were not affected (Figure 5A–B), suggesting a potential role of NOX in glucose metabolism in cells that are highly glycolytic such as in the cells with mitochondrial defect. We further measured cellular NADH and NAD+ contents in the Tet/off and Tet/on cells in the presence and absence of NOX inhibition. As shown in Figure 5C, NADH and NAD+ were separated by HPLC and eluted at 9 min and 15 min, respectively. The chemical identities of NADH and NAD+ in these HPLC peaks were collected and confirmed by mass spectrometry analysis (Figure S2C and S2D). Quantitative analysis showed that the expression of POLGdn (Tet/on) caused a decrease in NAD+ by approximately 30% (75.5→53.6 ng/4×106 cells; Figure 5D), indicating an increase in NAD+ consumption by the high glycolytic activity when mitochondrial respiration was suppressed by expression of POLGdn. Inhibition of NOX by p22phox siRNA knockdown or by DPI caused an additional 40%–50% decrease in NAD+, with a corresponding increase in NADH (Figure 5D). Besides, cellular NADP+/NADPH ratio was significantly increased in cells with POLGdn expression (Figure 5E). The above data suggest that NOX might by important to maintain high glycolytic activity in the cells with mitochondrial defects by supplying additional NAD+ from oxidation of NADH. Indeed, the membrane-associated NOX preparation was able to oxidize either NADPH or NADH as the substrates, and the oxidase activity was decreased by p22phox gene knockdown (Figure 5F). These data might explain why p22phox knockdown could cause NAD+ level decrease in the Tet/on cells.
Because NOX generates ROS, which might then stabilize HIF-1α to stimulate glycolysis [25], we tested if altering NOX activity would cause changes in ROS and HIF-1α. We knocked down p22phox using siRNA and then induced POLGdn expression for 4 d. Cellular O2− and mitochondrial O2− levels did not show any change on day 4 (Figure S2E–F), while knockdown p22phox causes a detectable decrease in cellular H2O2 (Figure S2G), consistent with the role of NOX in generating ROS outside the mitochondria. HIF-1α was not induced in the POLGdn cells even after 2 wk of Tet/on, while hypoxia induced a significant increase of HIF-1α in the same cells (Figure 5G). These data together suggest that induction of mitochondrial dysfunction did not significantly change HIF-1α expression. Thus, HIF-1α seemed not to play a major role in the POLGdn Tet/on cells to promote glycolysis.
It is known that the tumor suppressor p53 plays an important role in maintaining mitochondrial function through transcriptional activation of SCO2 and the loss of p53 leads to a decrease in mitochondrial respiration and increase in lactate generation [26],[27]. We reasoned that if NOX were important for the survival of cells with mitochondrial respiratory defect, the p53-null cells would be expected to have elevated NOX activity and be sensitive to NOX inhibition. To test this possibility, we compared NOX activity in human colon cancer cells (HCT116) with wild-type p53 or p53−/−. As shown in Figure 6A, the p53−/− cells had a higher membrane-associated NOX activity. Further analysis by semi-quantitative RT-PCR revealed that the expression of the NOX components NOX1 and p67phox were increased in HCT116 p53−/− cells (Figure S3A). Quantitative analysis using qRT-PCR showed that NOX1 and p67phox in HCT116 p53−/− cells have about 2- and 4-folds increase in gene expression, respectively, compared to HCT116 p53+/+ cells (Figure 6B).
To evaluate the importance of increased NOX for the survival of HCT116 p53−/− cells, we tested their sensitivity to NOX inhibitor DPI in comparison with the HCT116 p53+/+ cells. As shown in Figure 6C, after DPI treatment for 24 h, HCT116 p53−/− cells exhibited a round-up morphology and detached, leading to a decrease in the number of viable cells when compared with HCT116 p53+/+ cells. To further confirm this different sensitivity, the effect of DPI on mitochondrial transmembrane potential was compared in both cell lines by flow cytometry using rhodamin-123 staining. Substantially more HCT116 p53−/− cells (32%) lost their mitochondrial transmembrane potential after DPI treatment for 24 h, compared to 17% observed in the HCT116 p53+/+ cells (Figure 6D). These data suggest that NOX was up-regulated in cancer cells with loss of p53 and that inhibition of NOX could be a therapeutic strategy to preferentially kill these cancer cells. Indeed, cell death measured by annexin-V/PI analysis showed that the HCT116 p53−/− cells were more vulnerable to DPI than the HCT116 p53+/+ cells (Figure S3B). To further confirm that p53−/− cells have higher NOX activity, H1299 cells were transfected with p53wt expression plasmid to generate p53wt stably expressed H1299-p53wt cell line (Figure S3C). We observed that forced expression of wild-type p53 in H1299 cells have significantly decreased NOX activity (p<0.001) (Figure S3D).
Malignant transformation by oncogenic Ras is known to attenuate mitochondrial function and promote glycolysis [28],[29]. We introduced an inducible K-rasG12V expression vector into the T-Rex 293 cells and observed that induction of K-rasG12V expression caused a 50% decrease in mitochondrial respiration [30]. Importantly, this also caused a significant increase in NOX enzyme activity (Figure 7A) and elevated gene expression of the NOX components (Figure S4A). In a separate experiment, stable transformation of human pancreatic ductal epithelial (HPDE) cells with K-rasG12V also induced NOX upregulation (Figure 7B and Figure S4B). Western blot analysis of the K-rasG12V stably transformed cells and two naturally occurring pancreatic cancer cell lines (AsPC1 and Panc-1) showed that the protein level of p22phox, an essential component of the NOX enzyme complex, was substantially higher in pancreatic cancer cells than in non-malignant HPDE cells (Figure 7C). Using the H-RasV12-transformed human ovarian epithelial cell pair [31],[32], we also found that NOX activity in H-RasV12-transformed cells (T72Ras) was significantly higher than that in their non-tumorigenic parental T72 cells (Figure 7D). The increase in expression of NOX1, NOX2, p22phox, and p47phox in the H-RasV12-transformed cells was demonstrated by semi-quantitative RT-PCR analysis (Figure S4C) and quantitatively confirmed by qRT-PCR (Figure 7E). To test if the elevated NOX activity is important for the survival of the H-RasV12-transformed cells, we compared the sensitivity of T72Ras cells and the parental T72 cells to DPI. The transformed T72Ras cells were more vulnerable to NOX inhibition by DPI, leading to a substantial decrease of mitochondrial transmembrane potential (Figure 7F).
To test the clinical relevance of the above observations, we analyzed the expression of p22phox in clinical specimens using primary pancreatic tissue microarrays containing 105 cases of stage II pancreatic ductal carcinoma samples and 94 benign pancreatic tissues (normal and pancreatitis tissues). As shown in Figure 7G, about 48% (50/105) of the pancreatic cancer tissues exhibited a high protein level of p22phox, whereas only about 7% (7/94) of the benign pancreatic tissues were positive for p22phox. These data suggest that p22phox was significantly higher in pancreatic carcinoma than in non-malignant tissues (p<0.0001, Fisher's exact test).
To test the role of NOX in cancer cell survival, we stably knocked down p22phox expression in human pancreatic cancer cells (Panc-1), using p22phox shRNA lentiviral particles (Santa Cruz). The p22phox protein was substantially decreased and NOX activity was significantly decreased (Figure 8A) in p22phox-shRNA stably expressed cells compared with Panc-1 cells transduced with the control shRNA lentiviral particals. The knockdown of p22phox led to a significant decrease in cell growth (Figure 8B) and colony formation capacity (Figure 8C). Suppression of NOX expression also significantly decreased glucose uptake and lactate generation in pancreatic cancer cells (Figure 8D and 8E). To test the effect of NOX suppression on tumor growth in vivo, we subcutaneously inoculated the Panc-1 cells bearing p22phox-shRNA into the left flank of athymic nude mice (n = 7) and Panc-1 cells bearing control shRNA into the right flank of the same 7 mice (5×106 cells/injection site). Suppression of p22phox expression significantly impaired tumor growth in vivo (Figure 8F and 8G). The average tumor volume in the p22phox-shRNA group was 53.4±20.9 mm3, compared to 277.6±77.8 mm3 in the control group. After 60 d of cell inoculation, the average tumor weight of the p22phox-shRNA xenografts was 17±5 mg, compared to 170±62 mg in the control group (Figure 8H and 8I). These data together suggest that NOX are essential for tumor growth in vivo.
We then used DPI, a chemical inhibitor of NOX [23], to test its potential therapeutic activity against pancreatic cancer in vivo. Athymic nude mice were inoculated with Panc-1 cells subcutaneously (Text S1). When the tumor grew to 100 mm3, the mice were divided into two groups for treatment with vehicle (PBS) as control or with DPI (1.5 mg/kg mouse, i.v., 5 times/week). Such treatment did not cause significant toxicity in the mice and there was no loss of body weight in the DPI-treated group (Figure S5A). The tumors grew progressively in the control group, whereas the DPI-treated group exhibited significant retardation in tumor growth (tumor volume 1,068±309.7 mm3 versus 139±80.4 mm3, Figure S5B). Some of the tumors in the treatment group showed complete regression (Figure S5C). The average tumor weights of the control group were significantly increased after about 10 wk compared with the DPI-treated group (478±98 mg versus 84±52 mg, p<0.01, Figure S5D). These data suggest that DPI could effectively inhibit pancreatic tumor growth without apparent toxic side effects.
In this study, we established a dominant-negative mitochondrial DNA polymerase γ inducible cell system, which enabled us to investigate the relationship between mitochondrial respiratory defect and metabolic alterations. Unlike ρ° cells derived by chronic exposure of cells to ethidium bromide that could also damage nuclear DNA, the POLGdn inducible cells provide a clean isogenic model system that allowed the induction of mitochondrial respiration suppression under well-defined conditions without causing direct nuclear DNA damage. This model made it possible to monitor metabolic alterations during the shift from oxidative respiration to high glycolysis and to examine the biochemical mechanisms in detail. The use of this cell system in the current study led to a novel finding that NOX was consistently up-regulated when mitochondrial respiration was suppressed by the expression of POLGdn, and this was further confirmed in cancer cells with loss of p53 or expression of oncogenic Ras. Further study showed that such NOX elevation is critical for the maintenance of high glycolytic activity in cells with mitochondrial respiratory defects. This conclusion is supported by multiple lines of evidence: (1) NOX expression and enzyme activity were elevated in cells with different degrees of mitochondrial respiration suppression induced by POLGdn, and in cancer cells with mitochondrial defect due to a loss of p53, or under the stress of Ras oncogenic signal. (2) In cells with mitochondrial respiratory defect, a suppression of NOX led to a decrease in NAD+ level, lower glucose uptake, and reduced ATP content, leading to loss of cell viability. (3) Knockdown of NOX enzyme component p22phox in Panc-1 cancer cells decreased glucose uptake and lactate generation, decreased cancer cell proliferation and colony formation capacity, and suppressed tumor growth in vivo.
Cells with mitochondrial respiration defect require a higher rate of glycolysis, and activation of NOX seems necessary to provide additional NAD+ to support the highly active glycolysis. Our study showed that suppression of NOX in these cells caused a decrease in NADH oxidation and lower cellular NAD+ level, a decrease in glycolysis and cellular ATP, and a loss of cell viability. Thus, NOX activation is important for providing additional NAD+ to support active glycolysis. This is a previously unrecognized function of NOX in energy metabolism. The production of NAD+ by lactate dehydrogenase (LDH) is traditionally thought to be the main pathway that maintains the supply of NAD+ for glycolysis. In normal cells with competent mitochondria and a moderate level of basal glycolytic activity, the NAD+ generated by LDH may be sufficient to support the glycolytic reaction catalyzed by GAPDH. However, in cells with mitochondrial dysfunction that require higher glycolytic activity, upregulation of NOX seems important to provide additional NAD+.
It is important to note that the POLGdn cells not only represent a model system to study mitochondrial respiration defect as shown in other ρ° cells, but it also allows us to investigate the metabolic changes with various degrees of mitochondrial dysfunction. The upregulation of NOX was observed not only at the stage when mitochondrial respiration was severely inhibited by POLGdn expressing, but we have also observed NOX activation in the early stage (day 2 of Tet/on) of mitochondrial dysfunction. NOX activation is also consistently observed in cancer cells with mitochondrial dysfunction, such as HCT116 colon cancer cells that lack p53 and thus have a disruption of cytochrome c oxidase complex (complex IV) in the mitochondria [26]. Synthesis of cytochrome c oxidase 2 (SCO2) is a transcriptional target of the tumor suppressor p53. The alterations of metabolic parameters including lactate generation and cellular ROS, NADH, and ATP levels observed in the HCT116 SCO2−/− cell model [33] were similar to those seen in the POLGdn-Tet/on cells in this study. Interestingly, the Ras oncogene has recently been shown to cause mitochondrial dysfunction by disrupting mitochondrial complex activity [30],[34], which is consistent with our findings in this study. Stable transfection of human pancreatic ductal epithelial (HPDE) cells by K-rasG12V, induced expression of K-rasG12V in T-Rex 293 Tet/on cells, or stable H-rasV12 transformation in human ovarian epithelial cells led to a significant increase in NOX enzyme activity. Moreover, the protein level of p22phox was found to be significantly elevated in human primary pancreatic cancer tissues. These observations together suggest that NOX up-regulation may be an important cellular adaptive response to mitochondrial dysfunction to enable the increase in glycolytic activity. As such, NOX upregulation may be significant in many pathological processes related to mitochondrial respiratory inhibition, especially in cancer cells where mitochondrial respiratory dysfunction and metabolic abnormalities are prominent.
The molecular mechanisms by which divergent triggers including POLGdn, oncogenic Ras, and loss of p53 cause upregulation of NOX remain unclear. In the study, we observed that the expression of various NOX family members was upregulated in cells with mitochondrial dysfunction. A recent study suggests that SIRT1, a NAD+-dependent histone deacetylase, is involved in the negative regulation of NOX1 expression [35]. It is possible that mitochondrial dysfunction induced by various factors would lead to a decrease in cellular NAD+ content due to active glycolysis that consumes NAD+, and therefore release the SIRT1 suppression on NOX1 expression. Further study is needed to investigate this possibility.
Conventionally, the membrane-bound NOX enzyme complex is considered as ROS-generating machinery in phagocytes involved in the defense against microorganisms and mediating certain inflammatory processes. Subsequently, nonphagocytic NOX family of proteins homologous to gp91phox and other subunits have been shown to generate ROS in nonphagocytic cells and have been thought to contribute to various cancer cell proliferation and progression [15],[16],[36],[37]. Our study revealed an important function of NOX in cellular energy metabolism, especially in cancer cells with mitochondrial dysfunction. The ability of NOX inhibition to suppress pancreatic cancer cell proliferation, to abrogate colony formation ability, and to significantly suppress tumor growth in vivo suggests the feasibility to target NOX for cancer treatment. Since pancreatic cancer is highly resistant to many anticancer agents currently used in clinic, targeting NOX may have significant clinical implications and merit further investigation.
T-Rex 293 cells containing the pcDNA6/TR vector were obtained from Invitrogen. Full-length POLGdn cDNA harboring D1135A mutation and a FLAG-tag was excised by restriction enzyme EcoRI digestion from the POLGdn plasmid described previously [8]. The POLGdn fragment was subcloned into a mammalian expression vector pcDNA4/TO (Invitrogen) at EcoRI site and verified by nucleotide sequencing. The plasmid with correct orientation was verified by HindIII or XhoI enzyme digestion. The resulting POLGdn plasmid was then transfected into T-Rex 293 cells using lipofectamine 2000 (Invitrogen) to generate Tet/on inducible cell line. POLGdn positive colonies were selected by Zeocin (250 µg/ml) for 15 d, and expression of POLGdn was verified by 1 µg/ml doxycycline induction and Western blot analysis using anti-FLAG antibody (Sigma).
Oxygen consumption, glucose uptake, and lactate generation were described previously [18]. For measurement of oxygen consumption, cells were trypsinized and resuspended in 1 ml fresh culture medium pre-equilibrated with 21% oxygen at 37°C followed by applying the cells to the sealed respiration chamber of a Clark-type oxygen measuring system (Oxytherm, Hansatech Instrument, Cambridge, United Kingdom) with constant stirring. To measure cellular glucose uptake, cells in exponential growth phase were washed with glucose-free medium and incubated in fresh glucose-free RPMI 1640 medium for 3 h followed by an incubation with 0.2 Ci/mL 3H-2-deoxyglucose for 1 h. The glucose uptake represented by 3H radioactivity was determined by liquid scintillation counting and normalized by cell number. To measure lactate production, cells in 80% confluent were replenished with fresh medium. Aliquots of the medium were removed at the indicated time for measurement of lactate, using an Accutrend lactate analyzer (Roche, Mannheim, Germany). At each time point, cell number was also counted for normalization of lactate generation. Cellular ATP contents were measured using CellTiter-Glo Luminescent Cell Viability Assay kit (Promega) according to manufacturer's recommendations.
Flow cytometry determination of ROS, mitochondrial membrane potential, and cell death was performed using a BD Biosciences FACSCalibur flow cytometer (Mountain View, CA) and analyzed using the CellQuest software (Becton Dickinson). Cellular superoxide level was measured by incubating cells with 200 ng/ml HET for 1 h and mitochondrial superoxide level was measured by incubating the cells with 5 µm MitoSox Red for 60 min. Intracellular H2O2 contents were measured by incubating cells with 4 µM DCF-DA at 37°C for 1 h and mitochondrial transmembrane potential was assessed by incubating cells with 1 µM Rhodamine-123 (Rho-123) for 30 min before flow cytometer analysis. Apoptosis was determined by using the annexin-V and PI double staining method.
NAD(P)H oxidase activity was measured by a lucigenin-derived chemiluminescence assay as described [21],[22]. Briefly, 5–7 µg homogenized protein was incubated with its substrate 100 µM NADH or NADPH in a phosphate buffer (50 mM, pH 7.0) containing 150 mM NaCl and 1 mM EGTA for 15 min, followed by an addition of 5 µM lucigenin for 15 min in dark. The chemiluminescent signal (photon emission) was measured using a Turner 20/20 luminometer (Turner Designs, Sunnyvale, CA). No activity could be measured in the absence of NADH or NADPH. Experiments were also performed with the following pharmacologic inhibitors: a flavoprotein inhibitor DPI, a NOS inhibitor L-NAME, a mitochondrial respiratory chain complex I inhibitor rotenone, or a xanthine oxidase inhibitor oxypurinol.
Superoxide dismutase 1 (SOD1) activity was assayed using a SOD assay kit (Cayman Chemicals) per the manufacturer's instructions. Briefly, cells were collected by centrifugation at 1,000× g for 10 min at 4°C, lysed, and centrifuged at 1,500× g for 5 min at 4°C. Cytosolic fraction containing SOD1 was obtained from further centrifugation at 10,000× g for 15 min.
Intracellular NAD+ and NADH were quantified by HPLC as described previously [38] with some modifications. Briefly, after trypsinization, 6×106 cells were suspended in DMEM medium without FBS and collected by centrifugation at 1,400 rpm for 5 min. Cells were then shock-frozen with liquid-nitrogen. The cell pellets were resuspended in 150 µl extraction buffer containing 7 volumes of ethanol and 3 volumes of 10 mM potassium phosphate buffer, pH 8.5. The cells were disrupted by sonication and then kept at room temperature for 30 min to release cellular contents. Lysates were cleared by centrifugation at 13,000 rpm at 4°C for 15 min. 100 µl of the supernatant were subjected to HPLC analysis using an anion-exchange column (Partisil-10 SAX, Whatman) at a flow rate of 1 ml/min from 100% buffer A (5 mM NaH2PO4, pH 4.0) to 100% buffer B (250 mM NAH2PO4+0.5 M NaCl, pH 4.75) over 20 min, followed by another 5 min isocratic 100% buffer A. Pure NAD+ and NADH were used as reference standards and for quantitative calibration. NAD+ and NADH were detected by UV absorbance at 260 nm and 340 nm with retention times of 8.9 min and 15.4 min, respectively. The peaks corresponding NAD+ and NADH were collected according to the standards' retention time, freeze-dried in a lyophilizor, and further confirmed by MALDI-MS analysis (Bruker Daltonics). In HPLC analysis, the amount of cellular NAD+ and NADH of each sample was quantified based on the peak area compared to the standard curve generated by NAD+ and NADH standards.
Immunohistochemical staining for p22phox was performed on 4-µm unstained sections from the tissue microarray blocks consisting of 105 stage II pancreatic ductal carcinoma and their paired non-neoplastic pancreatic tissue samples from patients who underwent pancreaticoduodenetomy at our institution. The use of clinical specimens for tissue array study was approved by the Institutional Review Board of MD Anderson Cancer Center. To retrieve the antigenicity, the tissue sections were treated at 100°C in a steamer containing 10 mmol citrate buffer (pH, 6.0) for 60 min. The sections were then immersed in methanol containing 0.3% hydrogen peroxidase for 20 min to block the endogenous peroxidase activity and were incubated in 2.5% blocking serum to reduce nonspecific binding. The sections were then incubated with a rabbit polyclonal antibody against p22phox (Santa Cruz, 1∶100 dilution) at 4°C overnight, washed, and then incubated with secondary antibody at room temperature for 60 min. Standard avidin-biotin immunohistochemical analysis of the sections was done according to the manufacturer's recommendations (Vector Laboratories, Burlingame, CA) and photographed using a digital camera attached to the microscope. The staining results were evaluated by a gastrointestinal pathologist. Since all the pancreatic cancer samples showed either negative or diffuse staining for p22phox, the levels of p22phox expression were graded based on the staining intensity as negative (0), weak (1), moderate (2), and strong (3). P22phox expression was categorized as p22phox-low (intensity score of 0 or 1) or p22phox-high (intensity score of 2 or 3).
The experiments with mouse xenografts were carried out according to the protocols approved by the Institutional Animal Care and Use Committee (IACUC) of the University of Texas MD Anderson Cancer Center. Each side of the seven 10-wk-old athymic nude mice received a subcutaneous injection of 5×106 Panc-1 cells bearing p22phox-shRNA (p22phox-shRNA, left flank) or control shRNA (c-shRNA, right flank). Tumor size and body weight were measured throughout the experiment. Moribund animals were sacrificed as mandated by the IACUC protocol, and the tumor weight was recorded. Tumor volume was calculated using the equation: tumor volume (mm3) = L * W *(L+W)/2 * 0.526.
The Kolmogorov-Smirnov test (Cell Quest Pro software, Becton-Dickinson, San Jose, CA, USA) was used to evaluate the significant difference between control and test samples in flow cytometry analysis. For comparison of the statistical differences of more than two groups, one-way ANOVA and Newman-Keul's multiple comparison test was used. All other statistical significant difference analyses were evaluated using Student's t test (Prism GraphPad, San Diego, CA). Statistical differences between p22phox expression in benign and malignant pancreatic tissue groups on microarray were evaluated by Fisher's exact test. A p value of less than 0.05 was considered statistically significant.
|
10.1371/journal.pgen.1004590 | Differential Responses to Wnt and PCP Disruption Predict Expression and Developmental Function of Conserved and Novel Genes in a Cnidarian | We have used Digital Gene Expression analysis to identify, without bilaterian bias, regulators of cnidarian embryonic patterning. Transcriptome comparison between un-manipulated Clytia early gastrula embryos and ones in which the key polarity regulator Wnt3 was inhibited using morpholino antisense oligonucleotides (Wnt3-MO) identified a set of significantly over and under-expressed transcripts. These code for candidate Wnt signaling modulators, orthologs of other transcription factors, secreted and transmembrane proteins known as developmental regulators in bilaterian models or previously uncharacterized, and also many cnidarian-restricted proteins. Comparisons between embryos injected with morpholinos targeting Wnt3 and its receptor Fz1 defined four transcript classes showing remarkable correlation with spatiotemporal expression profiles. Class 1 and 3 transcripts tended to show sustained expression at “oral” and “aboral” poles respectively of the developing planula larva, class 2 transcripts in cells ingressing into the endodermal region during gastrulation, while class 4 gene expression was repressed at the early gastrula stage. The preferential effect of Fz1-MO on expression of class 2 and 4 transcripts can be attributed to Planar Cell Polarity (PCP) disruption, since it was closely matched by morpholino knockdown of the specific PCP protein Strabismus. We conclude that endoderm and post gastrula-specific gene expression is particularly sensitive to PCP disruption while Wnt-/β-catenin signaling dominates gene regulation along the oral-aboral axis. Phenotype analysis using morpholinos targeting a subset of transcripts indicated developmental roles consistent with expression profiles for both conserved and cnidarian-restricted genes. Overall our unbiased screen allowed systematic identification of regionally expressed genes and provided functional support for a shared eumetazoan developmental regulatory gene set with both predicted and previously unexplored members, but also demonstrated that fundamental developmental processes including axial patterning and endoderm formation in cnidarians can involve newly evolved (or highly diverged) genes.
| The recent wave of genome sequencing from many species has revealed that most of the gene families known to regulate animal development are shared not only between humans and laboratory favorites such as mice, flies and worms, but also by evolutionarily more distant animals such as jellyfish and sponges. It is often assumed that genes inherited from a common ancestor remain largely responsible for regulating embryogenesis across these animal species, rather than more recently evolved genes. To address this issue we made an unbiased, systematic search for developmental genes in embryos of the jellyfish Clytia, selecting genes whose expression altered upon manipulation of the key regulator Wnt3, and comparing their expression in embryos specifically disrupted for Planar Cell Polarity. Identification of evolutionarily conserved and novel genes as developmental regulators was confirmed by demonstrating characteristic expression profiles for a sub-set of genes, and by gene knockdown studies. Conserved genes coded for members of many known signaling pathway and transcription factor families, as well as previously unstudied proteins. Nearly 30% of the identified genes were restricted to cnidarians (the jellyfish-sea anemone-coral group), supporting the idea that the appearance of new genes during evolution contributed significantly to generating animal diversity.
| A major challenge in biology is to understand how the current extraordinary diversity of animal forms has been generated during evolution. Specific goals are to determine which genes were employed to regulate developmental processes in the earliest multi-cellular animals, and how this set of regulators was expanded during the evolution of different animal branches by diversification of existing gene families or by the acquisition of new genes. To address these questions requires identification and functional analysis of developmental regulatory genes in species from right across the animal kingdom, covering not only the “bilaterian” (protostome plus deuterostome) branch including the classic experimental models such as mouse, zebrafish, Drosophila and Caenorhabditis, but also non-bilaterian phyla such as cnidarians, ctenophores and sponges, which have evolved many distinct forms and body plans.
Following the recent explosion of genome and transcriptome sequencing it has been widely noted that the majority of families of transcription factors and signaling pathway components uncovered as developmental regulators in bilaterian model species are represented in genomes of cnidarians, as well as ctenophores and to a lesser extent sponges [1]–[9]. This has fuelled the idea that a shared set (or “common toolkit”) of genes inherited from a common metazoan ancestor is used to regulate development in widely divergent species through differential deployment [1], [3], [10]–[17]. The common toolkit idea relies heavily on the assumption that conserved genes have retained largely equivalent developmental functions during the evolution of each animal lineage, for which evidence remains quite patchy. Comparing the expression territories, and in some cases functions, of gene orthologs in families of transcription factors such as Hox, Sox, Fox and T-box genes, and components in signaling pathways such as Wnt, TGFβ, FGF, Hedgehog, Notch etc, has provided some support for this assumption, but also for lineage specific modifications in gene repertoires for example through gene duplications and losses within the transcription factor families [10]–[21]. Another possibility is that novel regulatory genes have emerged within specific evolutionary lineages to contribute to generating animal diversity [14], [18]–[23]. The significant proportions of cnidarian-specific gene sequences in the fully sequenced genomes of Hydra (around 15%) and Nematostella (around 13%) is compatible with such a scenario in Cnidaria [14], [18], [22]–[26]. Detailed studies involving transcriptome comparisons in Hydra have shown that many cnidarian-specific genes are associated with specialized cell types, notably nematocytes (stinging cells) but also nerve and gland cells [22], [24]–[30], while others have been specifically implicated in intercellular signaling and regulating morphological processes [22], [27]–[31]. Furthermore, in a subtractive hybridization search for cnidarian-specific genes involved in embryogenesis, 30 of 88 distinct partial cDNA clones recovered did not match known bilaterian sequences, including one corresponding to a Hydra specific gene (HyEMB-1) expressed in the ovary and early embryo [31].
To gain a fresh perspective on the gene repertoires that regulate metazoan development, we employed a systematic unbiased comparative transcriptomics approach to identify potential regulators of embryonic patterning at gastrula stage in the cnidarian experimental model Clytia hemisphaerica [32]. Clytia is a typical hydrozoan species that includes a jellyfish form as well as a polyp form in its life cycle, unlike anthozoan cnidarians such as the popular sea anemone model Nematostella vectensis. After gastrulation, a torpedo-shaped “planula” larva is formed, whose organization shows the characteristic cnidarian body plan: a single “oral-aboral” axis and two germ layers. The outer ectoderm of the Clytia planula features ciliated epitheliomuscular cells for motility, and an internal endodermal (or “entodermal”) region including a population of interstitial stem cells (i-cells) specific to hydrozoans, which generate a variety of cell types for each germ layer [33]–[36]. Gastrulation proceeds by unipolar cell ingression to fill the blastocoel prior to endoderm cell epithelialization [37]. The gastrulation site derives from the egg animal pole and corresponds to the pointed oral pole of the larva, giving rise after metamorphosis to the mouth region of the polyp form [38].
Establishment of the oral pole in Clytia critically depends on Wnt/Fz signaling activity through the Wnt/β-catenin pathway. Maternally-provided transcripts for the ligand Wnt3 and the receptors Fz1 (activatory) and Fz3 (inhibitory) are pre-localized along the egg animal-vegetal axis to drive activation of this pathway on the future gastrulation site/oral side during cleavage and blastula stages [39], [40]. This activation establishes distinct regional identities characterized by specific sets of transcribed genes at the oral and aboral poles of the developing embryo, including those required for cell ingression at gastrulation. Fz-PCP signaling, dependent on the conserved transmembrane protein Strabismus (Stbm), is activated in parallel along the same axis to coordinate cell polarity in the ectoderm and to guide embryo elongation [41]. Since multi-member Wnt families with early polarized embryonic expression have also been uncovered in other cnidarians [42], [43], ctenophores and sponges [44]–[47] as well as in a range of bilaterian models [48], [49], it seems highly probable that Wnt/Fz signaling regulated embryonic patterning in ancestral metazoans, specifying the primary body axes and/or presumptive germ layer regions.
To identify genes potentially involved in Clytia embryogenesis without favoring gene families identified as developmental regulators from bilaterians, we compared transcriptomes at the onset of gastrulation between normal embryos and ones strongly “aboralized” by Wnt3 morpholino (Wnt3-MO) injection prior to fertilization [40]. In many animals gastrulation coincides with, or closely follows, a significant stepping up of transcription from the zygotic genome, taking over from an initial phase of development predominantly dependent on maternally supplied mRNAs and proteins. By comparing transcriptomes from undisturbed and Wnt3-MO early gastrulae by Digital Gene Expression (DGE) we compiled lists of significantly over- and under-expressed genes. These included orthologs of known conserved developmental regulators but also members of unexplored metazoan conserved gene families, and in addition many sequences restricted to cnidarians. Expression profiling for an unbiased subset of these transcripts systematically revealed spatially or temporally restricted expression profiles of four types. Further transcriptome and in situ hybridization comparisons with Fz1-MO and Stbm-MO embryos revealed expression-pattern-related differences in the responses of genes to disruption of Wnt/β-catenin versus PCP. Finally, roles in developmental processes for the identified genes, both conserved and cnidarian–restricted, were supported both by their characteristic expression patterns and by correlated phenotypes obtained following morpholino injection for a subset of 8 genes.
Overall our unbiased screen allowed systematic identification of developmental genes regulated by the Wnt/ß-catenin pathway and by Fz-PCP. It provided functional support for a shared eumetazoan developmental regulatory gene set with both predicted and previously unexplored members, while also showing that axial patterning and endoderm formation in cnidarians can involve taxon restricted genes.
To identify genes regulated transcriptionally in relation to Wnt dependent embryo patterning we compared transcriptomes from unmanipulated early gastrula stage embryos and from embryos injected prior to fertilization with a morpholino antisense oligonucleotide targeting Wnt3 [40]. Digital Gene Expression analysis (DGE) was performed using an Illumina HiSeq sequencing platform. The number of mapped reads onto a reference transcriptome data set was taken as a measure of transcript level, and the statistical significance of differences in these levels between samples assessed using the DEGseq package (Figure 1; see Materials and Methods for technical details). Plotting for each transcript the expression ratio between two samples against the global average expression (Figure 1A,C) allowed visualization of sets of transcripts that showed significant differential expression, defined as ones that cannot be accounted for by sampling variation according to Random Sampling Model. We used the MATR method [50], justified by the Normal distribution of the data (Figure 1B), to adjust the cutoff to take into account experimental noise, based on comparison of replicate samples (blue line in Figure 1A: compare with the red line delimiting the theoretical random distribution). For subsequent analyses we routinely used a corresponding “z-score” value as an index of significant differences between samples (see Methods).
Comparisons between Wnt3-MO and uninjected embryo samples (Figure 1C) identified 375 assembled transcript sequences as differentially expressed according to the z-score +/−3.3 cutoff, which corresponds to a probability threshold (p-value) of 0.01 (colored dots in Figure 1C). Detailed analyses were performed for a more restricted set of 179 sequences with z-scores of less than -5 or greater than +5 (see insert in Figure 1C; list of transcripts and their characteristics in File S1). We could eliminate transcripts whose expression levels were affected non-specifically by the morpholino injection procedure by comparing the Wnt3-MO embryo differentially regulated transcripts with those identified in embryo populations generated using morpholinos targeting two other genes, Fz1 and Fz3, which respectively activate and repress Wnt/β-catenin signaling leading to aboralized and oralized phenotypes, respectively [39]. Genes non-specifically affected by the morpholino injection procedure are expected to respond in the same way in all three experimental groups, whereas genes regulated specifically downstream of Wnt3 are expected to respond distinctly following Fz1-MO compared to Fz3-MO injection. Comparison between these groups allowed us to identify 4 sequences with high z-scores (>5) in Fz1-MO and in Fz3-MO (opposite phenotypes) as well as Wnt3-MO samples (purple dots in Figure 1D,E; DGE class 5 in File S1). Two of these code for Ubiquitin ligases, implicated in protein degradation, and one for a secreted cyclase, suggesting a possible association with lysis of damaged cells in injected embryos. In addition the Fz3 transcript was itself detected at high levels in Fz3-MO embryos, probably due to the stabilizing effect of the morpholino. An additional set of 10 transcripts were eliminated as coming from likely bacterial contaminants, because they clearly stood apart as strongly under-represented (z-scores <−5) in both Fz3-MO and Wnt3-MO samples (and also for Fz1-MO in 9 cases) compared with uninjected controls (blue dots in Fig 1D, E.). The sequences of these transcripts had no similarity with any known eukaryotic genes but rather included genes from bacteria. Contamination from bacteria may be higher in uninjected embryos due to reduced manipulation of the egg and thus more frequent retention of the jelly coat and associated contaminants.
After elimination of the 13 non-specifically affected sequences, our final validated transcriptome comprised 166 differentially expressed transcripts. 153 of these 166 had clear predicted full or partial ORFs, comprising 40 over-expressed in Wnt3-MO embryos and 114 under-expressed. Detailed analysis of these sequences (File S1) revealed conserved and novel genes.
We undertook detailed characterization of spatial expression and sequence analysis (Table 1) non-selectively for the top 20 under-expressed transcripts (Figure 2) and top 18 over-expressed transcripts (Figure 3) in Wnt3-MO early gastrulae. Expression territories for all 38 transcripts were determined by in situ hybridization at three stages: early gastrula, 24 hpf planula (just completed gastrulation, endoderm still undifferentiated), and 48 hpf old planula (cell differentiation ongoing in both endodermal and ectodermal regions). We found that almost all the in situ hybridization profiles could be assigned to one of four types, which we termed Oral (O), Aboral (A), Ingressing/Endodermal (IE) and Delayed expression (D) types, as described in more detail below and summarized in Figure 4. Briefly, O and A type profiles are characterized by polarized expression with respect to the developing oral-aboral axis at all stages, suggesting ongoing patterning roles during embryonic and larval development. The IE type profile corresponds to cells destined to contribute to the complex endodermal region including the i-cell stem cells and their derivatives. The D type profile transcripts were barely detectable in early gastrulae but showed at larval stages expression in diverse patterns in the ectoderm and/or later in the endoderm. Overall, our approach to identify new candidates for roles in cnidarian embryonic development was completely validated by these analyses. Without any selection based on sequence identity, all the transcripts we tested showed expression restricted in space and/or time during gastrulation and planula development.
Names were assigned to the analyzed transcripts on the basis of orthology and/or membership of known gene families (all phylogenetic analyses in File S2). Multiple members of known gene families were distinguished by suffixes designating the 4 main expression profile types: O, A, IE or D. Cnidarian-specific transcripts lacking any recognizable orthologs from non-cnidarian species in NCBI databases, and those with non-cnidarian orthologs that had not previously been characterized, were assigned novel names using the same suffixes, prefixed by “Weg” to denote differential expression in Wnt3-MO early gastrulae, or given names based on recognizable repeats when present.
The overall outcome of our in situ hybridization analyses was that transcripts identified as Wnt3-MO-underexpressed consistently showed Oral and Ingressing/Endodermal type expression profiles while the overexpressed ones all showed Aboral and Delayed type profiles. The significance level of the response did not, however, correlate with expression patterns (O versus IE or A versus D, respectively; see z-scores in Table 1). Remarkably, we were able in both cases to uncover a strong correlation when we included in the analysis the z-scores obtained for the Fz1-MO sample (Figure 7A). This could be demonstrated by plotting the z-scores calculated for the two experimental conditions (against non-injected) against each other and determining the position of all the transcripts analyzed in Figures 2 and 3, of genes with expression patterns characterized previously (Bra, Fz3) and of five additional examples selected from our primary list (FoxQ2c, Tbx; NotumO, sFRP-A, Gsc, WegO3; File S4; All patterns summarized in Table 1 and Figure 4). Amongst the Wnt3-MO embryo under-expressed transcripts (orange dots in Figure 7A), those with Fz1-MO z-values higher than -5.0, ie not significantly affected or only relatively weakly underexpressed in Fz1-MO embryos, tended to show the O type expression pattern (eleven of the thirteen examined transcripts in the dark orange “Class 1” zone). The others (pale orange “Class 2” zone) showed IE type expression profiles in eleven of the twelve cases. A similar strong correlation was found for the Wnt3-MO embryo-over-expressed transcripts (green dots in Figure 7). In this case, applying a Fz1-MO z-score value threshold of +5.0 we found that transcripts with higher z-values (grey “Class 4” zone) tended to show D or mixed D/A-type patterns (seven and three respectively of the eleven analyzed transcripts), while nine transcripts with z-scores less than 5.0 (green “Class 3 “zone) showed A-type patterns and the tenth (Notch) a mixed A/D pattern. In this Class3 zone, responses to Fz1-MO were quite variable, including moderate over-expression, unchanged expression and, in a few cases, under-expression (notably FoxQ2a and WegA1).
From these analyses we defined four “DGE classes” on the basis of z-score values in Wnt3-MO and Fz1 MO embryos, as indicated in Figure 7A. Although these classes strongly correlate with the four types of expression profiles (Figure 7; Table 1) there are exceptions, for instance ZnfO is categorized as Class 2 on the basis of z-scores but shows an oral type expression profile, while Sulf is categorized as Class 1 but shows endodermal expression.
Fz1 acts as a receptor for Wnt3 to activate Wnt/β-catenin signaling [39], [40], but is also thought to interact with the Clytia Strabismus protein to mediate planar cell polarity (PCP), necessary for cell alignment in the ectoderm but also axial elongation during larval development and endoderm formation [41]. We thus hypothesized that the differences in expression responses in Fz1-MO versus Wnt3-MO could be due to the specific involvement of Fz1 in PCP. To test this hypothesis we made additional comparisons using a transcriptome derived from early gastrula embryos in which PCP was specifically disrupted by a morpholino targeting Strabismus (Stbm-MO). Plotting the z-scores (in relation to uninjected embryos) of the Fz1-MO and Stbm-MO transcriptomes against each other revealed a striking similarity (Figure 7B). The linear positive correlation was especially clear between Fz1-MO and Stbm-MO z-scores for the Wnt3-MO over-expressed transcripts (i.e. DGE Classes 3 and 4; green and grey dots respectively in Figure 7B; Pearson correlation coefficient value 0.93). The separation between Class 1 and Class 2 transcripts on the basis of Stbm-MO responses was less strict, with Class 1 transcripts showing moderately increased or decreased levels in these embryos, compared to unaffected or reduced levels in Fz1-MO embryos (compare distribution of orange dots in Figure 7A and 7C). This can be explained by the requirement of Fz1 but not Stbm in Wnt/β-catenin signaling in the presumptive oral territory,
We validated the transcriptome comparison analyses by in situ hybridization on Fz1-MO and Stbm-MO early gastrula embryos (Figure 8) using a subset of the probes used to examine Wnt3-MO embryos (Figure 5). For each gene the expression patterns in the two morpholino conditions were strikingly similar: The Class 1/O-type pattern transcript Myb, and the Class 3/A-type pattern transcripts ZnfA and sFRP-A showed little change compared with non-injected controls (Figure 8A, E, F). ZnfO, assigned to DGE Class2 despite its O-type expression profile, showed undetectable expression at the early gastrula stage in both Fz1-MO and Stbm-MO embryos (Figure 8B) and thus indeed represents an axially-expressed gene atypically sensitive to PCP perturbation.
The weak change in levels of most axially-expressed genes along with the significant under-expression of FoxQ2a and WegA1 in both Fz1-MO and Stbm-MO early gastrula embryos (Figure 7A, C; File S1) revealed in this study may at first seem difficult to reconcile with the previous description of an “aboralized” phenotype including a slight expansion of the FoxQ2a expression domain in Fz1-MO embryos [39], but this can be explained by a difference in the timing of the two studies since the PCP effect is only transient. Thus, analysis of Stbm-MO embryos revealed that while aboral FoxQ2a expression is undetectable by in situ hybridization at the early gastrula stage it subsequently becomes restored, while conversely oral expression of Bra1 is transiently expanded but then becomes re-restricted to the oral pole of the planula [41].
The in situ analyses performed for Class 2/IE-type and Class4/D-type pattern transcripts also validated the DGE analyses. FoxA and Znf845 were barely detectable by in situ hybridization at the early gastrula stage (Figure 8C, D), while Botch1 and bZip were detected strongly across the embryo (Figure 8G, H). As in Wnt3-MO embryos (Figure 5) the signal in these latter cases was mainly detected in cells positioned on the basal side of the ectodermal epithelial layer.
We conclude that the relatively strong under-expression (Class 2) or over-expression (Class 4) of certain genes in Fz1-MO embryos is due in whole or part to disruption of PCP. This effect could reflect regulation of gene transcription by specific signaling pathways activated by PCP or be indirect, resulting from disturbed morphogenesis following failure of the ectodermal cells to align, to develop cell polarity and to undergo ciliogenesis [41].
To test whether the newly identified genes in Clytia were indeed involved with developmental processes as predicted by their expression patterns, we injected antisense morpholino oligonucleotides targeting a selection of identified genes. We included in this analysis transcripts representing each of the four expression profile types including cnidarian-restricted genes (WegO1, WegIE2, WegD1), candidate conserved developmental regulators (Bra1, Bra2, FoxQ2c, FoxQ2a, HD02) and the partly conserved transcript WegA1. For each morpholino tested, developmental defects observed at morphological (Figure 9) and cellular (File S8) levels were coherent with the corresponding expression patterns (Figures 2 and 3), confirming the usefulness of our approach to identify developmental regulators. Wherever possible (6/8 cases, see File S7 for details) morpholinos targeting two different sites in the transcript were used, and in each case similar phenotypes were observed.
Morpholinos to the three O-type expression pattern transcripts all showed defects in endoderm formation, consistent with endoderm fate specification in the oral territory [61], [62]. Morpholinos targeting the two Clytia paralogs Bra1 and Bra2 both significantly inhibited endoderm formation. Initial signs of cell ingression at the oral pole occurred with only a slight delay with respect to non-injected controls, but subsequent filling of the blastocoel was strongly retarded, such that by 24hpf Bra1-MO and Bra2-MO embryos (Figure 9C, D) resembled uninjected embryos at the onset of gastrulation (about 11hpf). Bra1-MO and Bra2-MO embryos then elongated somewhat and disorganized cells accumulated in the blastocoel to a variable degree, although often with a significant reduction in the amount of endoderm observed. Confocal microscopy confirmed that the residual ectodermal cells of both Bra1-MO1 and Bra2-MOe/i embryos accumulated in aboral regions and showed signs of epithelialization (File S8 C, D). A similar but much less severe delay in gastrulation was obtained following injection of morpholinos targeting the cnidarian-restricted gene WegO1, whose expression profile is very similar to that of Bra1 and Bra2 (Figure 2). Planulae showed a characteristic tapering of the oral half (Figure 9H), and confocal microscopy revealed that endoderm was reduced in this region (File S8 B).
Strikingly, morpholinos targeting the A-type profile transcript WegA1 generated an opposite phenotype from the O-type pattern morpholinos. At the onset of gastrulation, massive cell ingression initiated widely across the embryo (Figure 9F). This is reminiscent of the phenotype previously described for Fz3 MO [39]. During subsequent development, cells from the internal regions were expulsed in most embryos, so that by the planula stage, embryos were commonly smaller and consisted of accumulations of endodermal-type cells surrounded in some cases by a very thin ectoderm layer, in which the cells were stretched over the inner cell mass (Figure 9F; File S8 G, H, I).
Morpholinos targeting the two IE type pattern genes WegIE2 and FoxQ2c both caused only minor disruption of development prior to the end of gastrulation, but subsequent formation of the endodermal cell layer was affected, with in both cases a thin and uneven layer of endodermal cells observed at 48hpf surrounding a distended cavity containing cell debris (Figure 9R, S). WegIE2-MO embryos showed additional disorganization of the oral ectoderm. Confocal microscopy confirmed that the endodermal cell layers were severely disorganized (File S8 E,F).
Finally, morpholinos targeting the two D-type profile genes, which are strongly up-regulated at the early gastrula stage upon Wnt3, Fz1 or Stbm disruption, did not markedly disrupt gastrulation but resulted in highly aberrant morphology of the planulae (Figure 9T,U). WegD1-MO embryos showed a distended aboral end with the ectoderm then becoming highly folded, this effect extending along the length of the embryo in the most extreme cases. Injection of morpholinos targeting the ANTP family gene HD02 also resulted in elongated and very irregular shaped planulae. In both cases the interface between the ectoderm and endoderm layers was very irregular with confocal microscopy revealing mixing of cells from the two layers and an absent or highly disrupted basal lamina between them (File S8 P, T). In HD02-MO embryos, anti-tubulin staining revealed an abundance of neurite–like projections traversing irregularly this interface, contrasting with the well defined epithelial basal lamina and regular distribution of orthogonally extending neural projections in undisturbed planulae (File S8; compare K and O).
We used the strong correlation between DGE classes and expression patterns to assess the relationship between transcript identity and localization, using the 128 transcripts for which complete ORFs were present (Figure 10). The proportions of transcription factors and probable signaling pathway regulators were similar between DGE classes (12–21%; values not significantly different by Fisher's Exact Test). In contrast there was a significantly higher proportion of cnidarian-restricted sequences in DGE classes 1, 2 and 3 than in DGE class 4 which tend to show D-type expression profiles (around 30% vs 6%; Fisher's Exact Test p-value for this comparison = 0.04). This analysis suggests that while cnidarian-restricted developmental regulators contribute significantly to patterning at the early gastrula stage, expression of evolutionary ancient genes predominates during development of the larva following gastrulation.
This study successfully identified many potential developmental regulators from the cnidarian experimental model Clytia hemisphaerica by analyzing the transcriptome of early gastrula stage embryos aboralized by Wnt3 knockdown, providing a number of new insights into the evolution of developmental patterning mechanisms. Firstly, the key role of Wnt signaling in embryo patterning was confirmed since the identified genes all displayed one of four basic expression profiles, three associated with embryo patterning (through localized expression in the oral, aboral and presumptive endoderm regions) and one with planula formation. Expression profile types could be related to differential expression sensitivity to Wnt3-MO vs Fz1-MO or Stbm-MO, allowing us to separate genes expressed along the oral–aboral axis predominantly under Wnt/β-catenin signaling regulation from genes whose expression at the early gastrula stage is affected by Fz-PCP. Secondly, the identified genes included not only members of known conserved metazoan developmental gene families, but also previously uncharacterized or understudied conserved metazoan genes, providing novel candidates for evolutionary ancient roles in directing developmental processes. Finally, a number of cnidarian-restricted genes emerged as potential developmental regulators. Roles in larval patterning and morphogenesis were confirmed by morpholino analysis for 3 such genes as well as for one that shares a domain of unknown function with bilaterians. Overall our study illustrates the power of systematic transcriptomics-based screens, coupled with functional studies, to identify developmental genes in non-bilaterians and thus to help understand metazoan evolution and diversification.
Our findings confirmed the central importance of Wnt signaling in embryo patterning. The transcripts under–represented in the spherical, aboralized Wnt3-MO embryos were during normal development systematically found expressed either in the oral ectoderm or in cells that contribute to the endodermal region (defining O and IE type profiles respectively), while those from the over–represented set were detected either in the aboral ectoderm (A type profile) or generally repressed throughout the embryo at the early gastrula stage to be expressed in different patterns during planula larva formation (D type profile). The O and A type profile genes displayed sustained localized expression at the poles through gastrulation and larval development and are thus good candidates for roles in patterning along the oral-aboral axis, but may also include precociously expressed gene markers of larval cell types enriched at one pole.
We were intrigued to find that the four types of expression profile for Wnt3-MO-differentially expressed transcripts strongly correlated with four “DGE classes”, distinguished by the strength of the effect of Fz1-MO on the expression of the same genes. More specifically the axially expressed transcripts tended to show less extreme changes in expression in Fz1-MO early gastrulae than did IE and D-type profile transcripts (Figure 7A). We have shown previously that Wnt/β-catenin signaling activated by Wnt3 and Fz1 is a key regulator of gene expression along the oral-aboral axis [39], [40]. The relatively weak difference in expression of the axial genes in Fz1-MO relative to Wnt3-MO early gastrulae documented here could be explained, at least in part, by incomplete inhibition of this pathway by Fz1-MO compared with total extinction by Wnt3-MO, as revealed by β-catenin nuclear localization (compare Figures 3 in [39] and [40]). It is also conceivable that Wnt receptors other than Frizzleds such as RYK or ROR2 [63] could be partly responsible for mediating the Wnt3 responses in oral regions. Our Stbm-MO analyses demonstrate, however, that the main explanation for less marked changes in expression of ‘axial’ versus ‘non-axial’ genes in Fz1-MO embryos relates to the involvement of Fz1 in PCP. One aspect of this is that transient up-regulation of some oral genes and down-regulation for some aboral genes due to PCP disruption, as shown in Stbm-MO embryos[41], could in Fz-MO embryos counterbalance and dampen the effects of Wnt/β-catenin signaling. Concerning the non-axial genes the strong effects of PCP disruption could reflect direct signaling through ‘non-canonical’ intracellular pathways acting downstream of Fz/Dsh [64]. Given the transient nature of the effect, however, we favor the possibility that the effect is indirect, resulting from the developmental programs of the corresponding cell lineages being delayed or accelerated by a changed morphological environment. For cells of the presumptive endodermal region (IE type pattern), lack of detection at the early gastrula stage in Fz1-MO and Stbm-MO embryos could result from disruption of ingression behavior due to loss of polarity of oral ectoderm cells. Conversely the strong over-expression of the D-type profile genes at the early gastrula stage in Fz1-MO and Stbm-MO embryos suggests that epithelial PCP may have a significant effect in delaying the development of certain planula cell types. One attractive possibility is that Fz-PCP disruption affects apical-basal polarity of epithelial cells and thus the generation of new cell types through oriented asymmetric divisions, as has been recently demonstrated in Xenopus embryos [65]. Consistent with this hypothesis, cells expressing Botch1, bZip and Amt became prominent in basal regions of the epithelial ectoderm of early gastrulae when PCP was disrupted directly using Stbm-MO or Fz1-MO (Figure 8), or disturbed indirectly in the Wnt3-MO context [41](Figure 6). Furthermore several other D type profile transcripts (HD02, UNC, WegD2 and possibly also Notch and Botch2) also tended to be expressed in basal regions of the ectodermal and/or endodermal epithelia during planula development (Figure 3).
Our study provides further support for the well-known idea that a common set of transcription factors diversified from a common cnidarian-bilaterian ancestor has retained roles in regulating development in individual evolutionary lineages, with some families diversifying functions following lineage-specific gene duplications[4]–[6], [9]. Clytia orthologs of many of known developmental regulator genes were identified from our unbiased screen based on sensitivity to Wnt/Fz signaling. All those tested showed characteristic spatiotemporally restricted expression profiles, and for four examples from well-known transcription factor families, roles in developmental regulation were supported by functional studies based on morpholino injection. Analysis of the morphant phenotypes suggested that the two Clytia Brachury paralogs Bra1 and Bra2, expressed at the oral pole throughout larval development, both play important roles in controlling the progression of gastrulation. Expression around the blastopore has been proposed to be an ancestral metazoan characteristic of Brachury, which during bilaterian evolution became involved in the specification of various mesoderm and endoderm fates from these tissues [66] but with the ancestral role likely to have been in regulating morphogenetic movements [67]. In Clytia, although there is no blastopore, the relationship with the gastrulation initiation site is conserved, and our morpholino results suggest that morphogenetic movements upstream of endoderm specification are affected. The Hydra Bra1 and Bra2 orthologs have been shown to have subtly distinct roles in endoderm and ectoderm layers of the budding polyp [11], suggesting that while embryogenesis roles for these genes overlap, their functions at other life cycle stages have diverged. A morpholino targeting FoxQ2c, expressed in the developing endodermal region during planula formation, caused severe defects in the organization of the endodermal layer. As with Brachyury, gene duplications have expanded the FoxQ2 gene family in Cnidaria, and in this case the paralogs have adopted clearly distinct expression profiles, FoxQ2a having conserved the likely ancestral aboral (anti-blastoporal) expression [68] while FoxQ2b is only expressed in oocytes [15]. The final member of a known developmental transcriptional regulator gene family we tested functionally was HD02, a non-Hox member of the Antp homeodomain family [16], expressed particularly strongly in cells at the base of the ectoderm and endoderm layers during larval development (Figure 2P). The phenotypes following morpholino injections suggest that HD02 is involved directly or indirectly in regulating development of the neural network that develops at this site [69], perhaps dependent on the correct organization of the basal lamina. Further in depth studies will be required to explore this possibility, as well as to confirm and understand fully all the other morpholino phenotypes documented here.
On the basis of expression patterns it is likely that several other transcription factor genes identified in this study have developmental functions conserved through metazoan evolution. For example FoxA and FoxC are associated with distinct cell populations contributing to the endoderm region during gastrulation, as has also been reported for their Nematostella orthologs expressed in distinct regions of the developing pharynx [12], [70]. In bilaterian species orthologs of these Fox genes are associated with development of endoderm/axial mesoderm and mesoderm respectively [71]–[74].
As well as transcription factors from families such as T-box, Fox and Antp, our transcriptome comparison identified likely regulators of a variety of intercellular signaling pathways including Notch, FGF, TGFβ and Ras-MAPkinase. These included core components (ligands, receptors and secreted antagonists), but also less well known regulators acting in ligand or receptor processing and/or extracellular interactions, such as the Botch, Sulf and Notum proteins. Most strikingly we identified Clytia orthologs of known Wnt pathway regulators acting at all levels: Wnt ligands (WntX1A), receptors (Fz3, Fz2), members of three of the five families of secreted antagonists known from bilaterian models (Dkk1/2/4; Dan1; two sFRPs) [52], [75], MESD which specifically interferes with ligand co-receptor LRP5/6 [58], [59], an ortholog of the intracellular negative regulator Naked Cuticle [56], and also the two Notum family lipases and Sulf. Sulf enzymes act on cell surface Heparan Sulphate Proteoglycans and have been reported to modulate Wnt as well as Hedgehog, TGFβ and FGF signaling while Notum releases the GPI anchor of glycipans such as Dally [76]–[79]. The oral expression profile of all the positive Wnt pathway regulators from this and our previous study (five Wnt ligands, Axin and TCF) reinforces the notion that an active Wnt signaling source is maintained at the cnidarian embryo and larval oral pole [40], [42], [80] as it is at the equivalent ‘head organizer” site in the Hydra polyp [81]–[83]. Co-expression of orally expressed putative pathway inhibitors such as Clytia NotumO is consistent with a role in limiting the extent of Wnt activity, equivalent to its action in Drosophila imaginal discs [76] or during planarian head regeneration [84]. Most of the putative Wnt antagonists we identified, however, were expressed aborally in the gastrula and in aboral pole subdomains in the planula (demonstrated by in situ hybridization for Dkk1/2/4, Dan1, sFRP-A and NotumA, implied by DGE responses for sFRP-B and MESD), suggesting that Wnt signaling is inhibited actively at the aboral pole region in the larva. Future functional studies will be required to examine the functions of each Wnt regulator during Clytia development, and to unravel the interactions between them.
Our study uncovered many potential developmental regulators amongst gene families with orthologs and/or shared domains identifiable from the mass of available genomic and transcriptomic data across bilaterian species, but for which nothing is known about function or expression. These include zinc finger and helix-loop-helix domain transcription factors as well as putative novel signaling pathway components. The prominence of cell surface protein modifiers with known impact on one or several signaling pathways in our screen raises the possibility that some of the other uncharacterized conserved or cell surface proteins may function similarly. In this context it would be interesting, for example, to test the function of the ZpdA and Aat genes, which code for a likely cell surface glycoprotein and a membrane transport protein respectively. Uncovering developmental roles for such proteins in Clytia would open the way to explore the involvement of potential novel regulators of key embryonic and cellular processes in bilaterians, and the associated evolutionary and medical implications. WegA1 offers an interesting illustration of this possibility. WegA1-MO injection results in a spectacular developmental defect involving premature cell ingression (a process of epithelial-mesenchymal transition) at gastrulation, and a massive shift in the balance of ectoderm to endoderm formation. This finding implies that this previously unknown protein functions during normal development under the control of Wnt/β-catenin and PCP signaling to inhibit cell ingression in aboral territories. As well as the 135-amino acid, C-terminal DUF3504 domain the WegA1 sequence contains a putative nuclear localization signal. Whether it has true orthologs in bilaterians remains to be established.
Amongst the potential developmental regulators identified in our study, 29% were defined as cnidarian-restricted on the basis that they had no identifiable orthologs in any other metazoans. Previous surveys of available cnidarian genomic and transcriptomic data revealed about 25% in Clytia and 15% in the ‘polyp only’ cnidarian models Nematostella and Hydra [4], [14], [18], [23]. A few of these match genes previously known only outside Metazoa, and so represent ancient genes lost in bilaterian branches or gained by lateral gene transfer, while the others probably represent cnidarian innovations. Although more in depth studies of each gene are required, the characteristic phenotypes observed in our morpholino experiments support the stereotypical expression pattern data in suggesting roles in regulating developmental processes for these cnidarian-restricted genes: larval oral pole organisation for WegO1, endoderm formation for WegIE2 and epithelial organization for WegD1 respectively.
More than half of the cnidarian-restricted transcripts identified in our study contained secretion signal sequences. These are prime candidates for roles in cell-cell signaling, either as ligands or as modulators of ligand/cell surface/receptor interactions during axis establishment and gastrulation. Candidate receptors for such signaling molecules include the many unclassified 7tm receptors identified particularly amongst IE profile/DGE class 2 transcripts. With notable exceptions such as Frizzled and Patched, members of the 7tm superfamily, including the G–protein coupled receptors (GPCRs), have not been strongly implicated in developmental regulation in bilaterians. This family has expanded independently in cnidarians [14], so its exploitation for developmental signaling might represent a cnidarian specialty, a fascinating possibility to explore in future studies.
Intriguingly, almost all (35/37) of the cnidarian-restricted genes we identified belonged to the three DGE classes associated with regional expression and thus embryo patterning at the gastrula stage (Figure 10). Conversely, the DGE Class 4 transcript set contained a higher proportion of broadly conserved “ancient” genes. Recent studies have demonstrated that the extensive variation in modes of early embryogenesis between species correlates with expression of evolutionarily “newer” genes, while subsequent ‘phylotypic stages’ (corresponding to neurula and somatogenic stages in vertebrates and the germ-band segmentation stage in insects) are strongly conserved at the phylum level and tend to express more ancient genes [85], [86]. With the caveat that our analysis concerns only a small fraction of the transcriptome and provides only limited coverage of developmental stages, the observation that most (28/35) of the DGE class 1–3 (putative patterning) genes lacked counterparts in Nematostella or Hydra may reflect the widely divergent modes of early embryo patterning and gastrulation amongst cnidarian species [87]. In contrast several of the DGE class 4 genes, mostly “ancient”, appeared to be associated with epithelia development and in particular with formation of the basal lamina, a structure considered to be a major innovation in the animal lineage [88], [89] and highly conserved in all Eumetazoa. A temporal shift in expression from “new” to “old” genes between gastrula and larva in cnidarian species is consistent with the idea that the epitheliarized, planula stage-ciliated torpedo larva represents the phylotypic stage [90].
To conclude, from a methodological standpoint, our study demonstrates the power of rigorous unbiased transcriptomic approaches to obtain a fresh view of gene conservation and innovation in the evolution of animal diversity. It also illustrates how transcriptome comparisons can allow prediction of expression characteristics without doing large-scale in situ hybridization screens; The differential transcriptional responses in Fz1-MO and Stbm-MO embryos will be very useful for picking candidate genes for future studies targeted to particular developmental processes. From a theoretical standpoint, our findings provide strong support for the notion that many evolutionary-conserved genes are deployed across eumetazoans to regulate development, but also good evidence that developmental regulation in cnidarians may involve a significant number of taxon-restricted genes. Functional studies of the genes identified here in Clytia should provide a fruitful entry for exploring both these possibilities.
Eggs obtained by light-induced spawning of laboratory-raised medusae were microinjected with morpholino oligonucleotides prior to fertilization as described [39]. Previously unpublished morpholino sequences are provided in File S7. Use of genetically identical female medusae derived from a single individual laboratory polyp colony Z4B and males from a closely related colony [32] restricted the problems of sequence polymorphism. After culture at 18°C to the four cell stage, any unfertilized or abnormally-dividing embryos were removed. Early gastrula stage embryos, used for RNA extraction or fixed for in situ hybridization or confocal microscopy, were obtained after culture at 16°C overnight (17 hours). Planulae were fixed for in situ hybridization after 24 or 48 hours of culture at 18°C. Particular care was taken to use identical timing and temperature regimes for all experiments.
For each experimental condition, total RNA was extracted from batches of 900–1400 early gastrula stage embryos using RNAqueous kit (Life Technologies/Ambion, CA). RNA integrity was confirmed by formaldehyde gel electrophoresis. Two independent biological replicates were performed for the uninjected and Wnt3-MO conditions, and single samples for the other morpholino conditions. Estimated final embryo numbers in each sample, after removal of any showing arrested cleavage or irregular development, were as follows: Uninjected: each 1900; Wnt3-MO: each 2300, Fz1-MO: 900, Fz3-MO: 1600 and Stbm: 1400. Library construction and Illumina short-read (51 bp) sequencing was performed by GATC (Konstanz, Germany).
To quantify gene expression, the number of mapped reads onto a reference transcriptome data set was taken as a measure of transcript level. The reference transcriptome, comprising 24893 distinct (non-overlapping) assembled sequences, was built by combining, using CAP3 software, previous EST data [15], [32] and Illumina sequences from one of the untreated early gastrula samples generated in this study. Redundant sequence entries were eliminated by USEARCH (ver. 5.2.32_i86linux32). The longest predicted ORF from each sequence was used as the reference for read mapping. To reduce polymorphism, adaptator sequences and probable 5′ UTR sequences upstream of the first ATG in each cDNA contig were removed,
For each experimental condition approximately 80 million of 51pb Illumina reads were mapped on the reference transcriptome using the Bowtie command, with tolerance of two mismatches. Reads that matched to more than one reference sequence were not taken into account. Around 35% of the reads obtained for each condition could be mapped using this method. Statistical analysis was performed using the DEGseq R package [50] to determine for each transcript whether the observed ratio of transcript levels (M) between two samples is significant given the global average expression (A). The Random Sampling Model employed assumes a normal distribution for log2(C), where C is the number of counts, as confirmed for our data by a Q-Q plot (Figure 1B). M = log2(C sample1)-log2(C sample2) estimates the difference of expression between the conditions; A = (log2(C sample1)+log2(C sample2))/2 measures the average expression in the two conditions. A p-value was generated for each gene to determine whether the expression difference between samples was significant. A z-score was generated for each transcript, as a measure of the deviation from the random model (z-score = (Mobserved – Mexpected according to random sampling model)/Var(M expected according to random model). The MATR method used an estimation of the variation between duplicate embryo samples (calculated using the CTR method) to generate a second MA plot and to adjusts the z-score accordingly.
A R-script was devised to analyze automatically the six possible reading frames of each unique assembled transcript sequence and to predict the best ORF (“find. ORF” script downloadable at http://octopus.obs-vlfr.fr/R_scripts). Sequence comparisons were performed with both BLASTx with the whole sequence and BLASTp with predicted translated ORF against the “non-redundant’” (nr) NCBI database. Domain analyses (Files S1 and S3) were performed using Interproscan, SignalP for the detection of secreted peptide signals and TMHMM for the prediction of transmembrane domains. Gene identities (column 3 of File S1) were based on BLAST and domain analyses. Gene accession numbers are provided in File S1 and File S3.
To determine orthology of the transcript sequences studied in detail (Table 1) we searched for homologs by reciprocal BLASTp. When reciprocal blast and domain analysis (see above) gave unambiguous identities (non-multigene families), gene names were attributed directly (Sulf, Aat, Asparaginase, Amt, UCP). For certain multigenic developmental regulator families, we added our candidate sequence and the retrieved cnidarian sequences to alignments from previously published studies kindly provided by authors (see File S2 and acknowledgements). Where no existing appropriate alignments were available, sequences from a range of eumetazoan genomes (Drosophila melanogaster, Lottia giganta, Strongylocentrotus purpuratus, Xenopus laevis or Homo sapiens, H. magnipapillata and N. vectensis) were aligned using MUSCLE, the best fitting model of evolution was determined using ProtTest2.4, and phylogenetic analysis performed using PhyML3.0. The trees are available in File S2.
In cases where clear Hydra magnipapillata orthologs were identified, further analysis was performed using the Hydra vulgaris transcript dataset (HAEP) available at http://compagen.zoologie.uni-kiel.de/blast.html) [91]. The number of matching reads recorded in each separated cell population (endoderm, ectoderm and nanos-positive cells) was normalized with respect to total read number (File S5).
For cnidarian-specific sequences, WegO1, WegO2, WegIE2, WegA2, WegD2, Zpd, had no recognisable homologs in Hydra or in Nematostella genomes. For WegIE1, WegD1 we identified single orthologs in Hydra: (listed in File S5).
In situ hybridization probes were synthesized from cDNA clones corresponding to our EST collection when available. For the remaining sequences, cDNAs were cloned by PCR using the TOPO-TA cloning kit (Invitrogen). All sequences were verified before probe synthesis. DIG-labeled antisense RNA probes for in situ hybridization were synthesized using Promega T3/T7/Sp6 RNA polymerases, purified using ProbeQuant G-50 Micro Columns (GE Healthcare) and taken up in 100 µl of 50% formamide.
Gastrulae, 24hpf and 48hpf planula larvae were fixed in 3.7% formaldehyde/0.2% glutaraldehyde in PBS for 2 hours on ice, washed five times in PBST (PBS containing 0.1% Tween 20) for 10 minutes, dehydrated in PBST/50% methanol and stored in methanol at −20°C. In situ hybridization was performed using the InsituPro robot (Intavis). After rehydratation in PBST/50% methanol and three 5 minute washes in PBST, samples were transferred to the plate. The robot program was as follows: two 20 min washes in PBST; 20 min in PBST/50% hybridization buffer (5X SSC, 50% deionized formamide, 1% dextran sulfate, 1% SDS, 50 µg/ml tRNA, 50 µg/ml heparin); 20 min in hybridization buffer; 2 hours pre-hybridization in hybridization buffer at 62°C; 40 to 63 hours hybridization at 62°C with the denatured DIG-labelled RNA probe; four 30 min washes in 5X SSC, 0.1% Tween 20 and 50% formamide at 62°C; four 30 min washes in 2X SSC, 0.1% Tween 20 and 50% formamide at 62°C; two 20 min washes in 2X SSC, 0.1% Tween 20 at 62°C; two 20 min equilibration steps in MABT (100 mM maleic acid pH 7.5, 150 mM NaCl, 0.1% Triton X-100); 1 hour blocking in MABT/1% blocking reagent (Roche); 3 hours incubation with an alkaline phosphatase labeled anti-DIG antibody diluted 1/2000 in the blocking solution; seven 20 min washes in MABT; three 20 min washes in TMNT (100 mM Tris-HCl pH 9.4, 50 mM MgCl2, 100 mM NaCl and 0.1% Tween 20). The color reaction was performed manually in TMNT containing 0.08 mg/ml NBT and 0.1 mg/ml BCIP (Promega). Color development time varied from 1 hour to 1 day. Samples were then washed twice in water, three times in PBS, post-fixed in PBS/3.7% formaldehyde and washed three times with PBST before mounting in 40% glycerol.
For the selected candidate genes we addressed phenotype specificity by designing and testing several morpholinos targeting different parts of the sequence, discarding any that proved toxic to cell division during pre-gastrula development. We could only identify 1 non-toxic morpholino targeting FoxQ2c and WegA1, and none for FoxQ2a. For Bra2 one morpholino targeted the predicted AUG translation initiation codon and the other an exon-intron junction (all details in File S7). For each morpholino we first injected a range of concentrations into eggs prior to fertilization, and then assessed planula morphology for the lowest non-toxic dose at 24 h and 48 h. The cellular basis of the observed phenotypes was then further assessed by confocal microscopy. Images of in situ hybridization profiles and DIC images of live embryos were acquired on an Olympus BX51 microscope. For confocal imaging of cell boundaries using fluorescent phalloidins and nuclei using Hoechst 33358 or TOPRO-3 dyes, embryos were fixed, processed and imaged on a Leica SP5 microscope as described previously [39]. Microtubules were stained by immunofluorescence using anti-α tubulin rat monoclonal antibody YL1/2 (Sigma) followed by rhodamine-conjugated anti-rat Ig antibodies (Jackson Immunoresearch).
Total RNA from 60 Wnt3-MO injected and 60 non-injected early gastrulae was extracted using RNAqueous-Micro kit according to the manufacturer's instructions (Ambion, Warrington, UK). Genomic DNA was removed by a DNAse I treatment (Ambion) and this step was controlled for each RNA extract. First-strand cDNA was synthesized using 500 ng of total RNA, Random Hexamer Primers and Transcriptor Reverse Transcriptase (Roche Applied Science, Indianapolis, USA). Quantitative PCRs were run in quadruplicate and EF-1alpha used as the reference control gene. Each PCR contained 5 µl cDNA 1/400, 10 µl SYBR Green I Master Mix (Roche Applied Science), and 200 nM of each gene-specific primer, in a 20 µl final volume. PCR reactions were run in 96-well plates, in a LightCycler 480 (Roche Applied Science). Sequences of forward and reverse primers designed for each gene: EF-1alpha-F 5′ TGCTGTTGTCCCAATCTCTG 3′; EF-1alpha-R 5′ AAGACGGAGTGGTTTGGATG 3′; Bra-F 5′ GCAACACCCACAACAACAAC 3′; Bra-R 5′ TACGGGAAACATACGCCTTC 3′; NotumO-F 5′ GGGACATCTAAAACCCATGC 3′; NotumO-R 5′ CATGGATCTCGCATTGTGAC 3′; ZnfO-F 5′ TGCTGCTAACAACGACCAAC 3′; ZnfO-R 5′ TGGTGGAAGTGGAGATTGTG 3′; Mos3-F 5′ ATCTTACGTCCCGAACAACG 3′; mos3-R 5′ ATCCACCAATGGCAGCTTAC 3′; Znf845-F 5′ AGACCGACAGCATTTCATCC 3′; Znf845-R 5′ TGGCATTCCTTGCATACCTC 3′; Dkk1/2/4-F 5′ GGGCTTGTTCGTACTTTTCC 3′; Dkk1/2/4-R 5′ ATTCCATCCCACGACAACAC 3′; ZnfA-F 5′ CAACAACTTTCACCGAGCTG 3′; ZnfA-R 5′ TGTCTCTTGTGTTGCCAAGC 3′; Dan1-F 5′ CATGCCCGTTCATGAGAAAG 3′; Dan1-R 5′TTTTGGCTGTTCCCACTGTC 3′; NotumA-F 5′ TGCTGAAGGGTCGTACATTG 3′; NotumA-R 5′ CGTGTGTCCATTTTCAGTGC 3′; HD02-F 5′ TT AACAGCCCACCGAAACTC 3′; HD02-R 5′ CGTCGTGTTTTTCAGTGACG 3′.
For each gene studied an expression level N was calculated as 2−Ct, where Ct (Cycle threshold) represents the number of cycles required for the fluorescent signal to cross the threshold.
|
10.1371/journal.pgen.1007912 | Temporal and spatial regulation of protein cross-linking by the pre-assembled substrates of a Bacillus subtilis spore coat transglutaminase | In many cases protein assemblies are stabilized by covalent bonds, one example of which is the formation of intra- or intermolecular ε-(γ-glutamyl)lysil cross-links catalyzed by transglutaminases (TGases). Because of the potential for unwanted cross-linking reactions, the activities of many TGases have been shown to be tightly controlled. Bacterial endospores are highly resilient cells in part because they are surrounded by a complex protein coat. Proteins in the coat that surrounds Bacillus subtilis endospores are crosslinked by a TGase (Tgl). Unlike other TGases, however, Tgl is produced in an active form, and efficiently catalyzes amine incorporation and protein cross-linking in vitro with no known additional requirements. The absence of regulatory factors raises questions as to how the activity of Tgl is controlled during spore coat assembly. Here, we show that substrates assembled onto the spore coat prior to Tgl production govern the localization of Tgl to the surface of the developing spore. We also show that Tgl residues important for substrate recognition are crucial for its localization. We identified the glutamyl (Q) and lysil (K) substrate docking sites and we show that residues on the Q side of Tgl are more important for the assembly of Tgl than those on the K side. Thus, the first step in the reaction cycle, the interaction with Q-substrates and formation of an acyl-enzyme intermediate, is also the determinant step in the localization of Tgl. Consistent with the idea that Tg exerts a “spotwelding” activity, cross-linking pre-formed assemblies, we show that C30 is an oblong hexamer in solution that is cross-linked in vitro into high molecular weight forms. Moreover, during the reaction, Tgl becomes part of the cross-linked products. We suggest that the dependency of Tgl on its substrates is used to accurately control the time, location and extent of the enzyme´s activity, directed at the covalent fortification of pre-assembled complexes at the surface of the developing spore.
| The orderly recruitment of proteins during the assembly of complex macromolecular structures poses challenges throughout cell biology. During endospore development in the bacterium Bacillus subtilis at least 80 proteins synthesized in the mother cell are assembled around the developing spore to form a protective coat. Regulation of coat gene expression has been described in detail but it is unknown how the information encoded by the structures of the proteins guide their assembly. We have examined the assembly of a transglutaminase, Tgl, which introduces ε-(γ-glutamyl)lysil cross-links in coat protein substrates. We describe with molecular detail a substrate-driven assembly model that directs the enzyme to the locations of its substrates where, as we suggest, it exerts a “spotwelding” activity to fortify pre-assembled complexes. The catalytic cysteine, located in a tunnel that spans the Tgl structure, first forms an acyl enzyme intermediate with a glutamine (Q) donor substrate. Then, it engages a lysine (K) donor substrate to form the cross-linked product. We have identified the Q and K acceptor ends of the Tgl tunnel, and we show that substitutions in substrate recognition residues at the Q side impair assembly more strongly than at the K side. Thus, assembly of Tgl parallels its catalytic cycle, directing the enzyme to the pre-formed complexes that are to be cross-linked.
| Protein function is often restricted to specific cellular locations, in both eukaryotes or prokaryotes, and knowledge of the pathways governing protein localization is essential to understand protein function (reviewed by [1]). The formation of supramolecular protein assemblies, for instance, requires targeting pathways that direct the various components to the appropriate cellular location or locations at the right time [1]. In many cases, protein assemblies, whether static or dynamic, are formed through non-covalent interactions of their components. In other cases, protein assemblies are stabilized by covalent bonds, one example of which is the formation of intra- or intermolecular ε-(γ-glutamyl)lysil cross-links catalyzed by transglutaminases (TGases) [2]. TGases are involved in blood clotting processes, the organization of the extracellular matrix, tissue and bone mineralization, cell adhesion, stabilization of dermo-epidermal junctions, or the cross-linking of eye lens proteins [2–4]. Because of the potential for unwanted cross-linking reactions, the activity of the human and human-like enzymes is tightly controlled as illustrated by human factor XIIIa (FXIIIa), a TGase responsible for the cross-linking of fibrin in the last stages of the blood coagulation cascade [5]. Factor XIII (FXIII), the zymogenic inactive form of FXIIIa, exists in the plasma as an A2B2 heterotetramer, in which the A2 subunits correspond to the TGase zymogen, while B2 are non-enzymatic; most of the FXIII in the plasma is bound to fibrinogen, the fibrin precursor, by its B2 subunits [5]. Thus, the inactive TGase is associated with its substrate precursor during the coagulation cascade ensuring proper localization of the enzyme for the formation of the blood clot. FXIII activation will then be elicited by thrombin processing and dissociation of the B2 subunits. Fibrinogen is cleaved by thrombin and fibrin then self-assembles into protofibrils and fibers held together by non-covalent bonds, eventually forming a clot network. In a classical example of “spotwelding” activity, FXIIIa cross-links and thereby stabilizes the fibrin polymer, a process essential for haemostasis [2, 5]. Furthermore, the association of fibrin chains stimulates the activity of FXIIIa at specific cross-linking sites, a process that has been termed “assembly-driven regulation of cross-linking” [2, 5].
Species of the Bacillus and Clostridium genera and related organisms, have the ability to differentiate into dormant endospores (spores for simplicity) arguably one of the most resilient cell types found in nature [6–8] (Fig 1A). Spores are composed of three main concentric compartments: the core, which contains a copy of the genome; the cortex, composed of a modified form of peptidoglycan, and the coat, which surrounds the cortex [9–11] (Fig 1A). In the best-studied spore-forming organism, B. subtilis, the coat is assembled from over 70 polypeptides and has a role in spore protection and germination [9–13]. Perhaps not surprisingly, one of the proteins recruited to the coat layers, Tgl, is a TGase. Tgl shows the NlpC/P60 catalytic core characteristic of TGases and a dual, partially redundant catalytic dyad located in a tunnel that transverses the molecule from side to side [14]. In the reaction cycle, shared by all TGases, a glutamyl (Q or acceptor) substrate binds to the enzyme at one side of the tunnel and forms an acyl-enzyme intermediate with the catalytic Cys116 (Fig 1B). Only then, a lysil (K or donor) substrate approaches the enzyme from the opposite side of the tunnel and attacks the thiolester bond, leading to the formation of a ε-(γ-glutamyl)lysil isopeptide bond [2, 5] (Fig 1B). Importantly, unlike other TGases, Tgl is produced in an active form, and efficiently catalyzes amine incorporation and protein cross-linking in vitro with no known additional requirements [14, 15]. Tgl is a small and structurally simple TGase, and we have argued that it embodies the minimal structural requirements for protein cross-linking [14, 15]. The absence of regulatory factors raises questions as to how the activity of Tgl is controlled during spore coat assembly.
Sporulation takes place in a sporangium formed by a larger mother cell and a smaller forespore that will become the future spore (Fig 1A). Assembly of the coat is mainly a function of the mother cell and begins soon after asymmetric division with the expression of genes coding for early spore coat proteins under the control of RNA polymerase sigma subunit σE [9, 10, 16]. After asymmetric division the forespore is engulfed by the mother cell and becomes isolated from the external medium (Fig 1A). Engulfment completion coincides with the onset of transcription of the genes coding for late spore coat proteins, under the control of σK [9–11]. σK also triggers cortex polymerization; soon after, the spore development process is finalized through lysis of the mother cell and release of the spore (Fig 1A). The coat is formed by four main layers: a basement layer which sits on the cortex and is surrounded by an inner and outer coat layers and finally by the crust [9–11] (Fig 1A). Four morphogenetic proteins that localize to the forespore surface as engulfment begins, create an organizational scaffold for the construction of the different coat layers [10, 11, 16, 17]. Formation of the basement layer requires SpoIVA, tethered to the forespore membrane by a small peptide, SpoVM, which recognizes the positive curvature of the forespore outer membrane [18–21]. SafA and CotE, are necessary for inner and outer coat formation, respectively, and CotZ governs formation of the crust [22–27] (Fig 1C). These proteins are recruited by SpoIVA and bring to the spore surface the proteins that form the various coat layers. In a second step termed encasement, that requires SpoVM and another morphogenetic protein, SpoVID, the coat proteins migrate around the forespore, in successive waves determined in part by the activities of σE and σK and their auxiliary transcription factors [17]. SafA, CotE and CotZ appear to function as hubs, i.e., they may interact directly with several of the coat or crust proteins whose assembly they direct (Fig 1C). The C30 protein, for instance, interacts directly with SafA [27]. C30 is formed by internal translation initiation at Met codons 161 and 164 of the safA mRNA [28] (Fig 1D). Since it lacks the localization signals found at the N-terminus of the full-length protein, which include a LysM domain and a short region (region A) that both mediate an interaction with SpoVID, C30 relies on the interaction with SafAFL for assembly [28–30]. SafAFL and C30, as well as two other inner coat SafA-dependent proteins, YeeK and GerQ, are substrates for Tgl [31–35] (Fig 1C). SafAFL and C30, as well as Tgl itself, are found mainly in the cortex and inner coat [27, 35, 36]. SafA is a key factor in recruitment of Tgl to the inner coat, but is indispensable for its association with the cortex; this pathway is auto-regulatory, in that following recruitment by SafA, Tgl then cross-links SafAFL and C30 in both the coat and the cortex [35].
Here, we show how the activity of Tgl is spatially restricted to the surface of the developing spore. We show that assembly of Tgl depends on the prior assembly of its substrates. Conversely, we have uncovered residues in the vicinity of the active site of Tgl involved in enzyme-substrate interactions that are critical for the assembly of Tgl. We identify the Q and K sides of Tgl, and we show that Q-side residues make a more important contribution to assembly of the enzyme than K-side residues. Hence, the steps in Tgl assembly parallel the steps in the enzyme´s reaction cycle [35]. Supporting a model in which Tgl exerts a “spotwelding” activity to covalently fortify pre-assembled complexes, we show C30 is a hexamer that is cross-linked by Tgl into high molecular weight forms. Moreover, we show that the enzyme itself becomes part of the cross-linked products. Together, our results suggest that the dependency on the prior localization of its substrates spatially and temporally restricts the activity of Tgl, ensuring that pre-assembled complexes are cross-linked at the spore surface.
Four Tgl substrates are known, YeeK, GerQ, SafAFL, and C30 [31–35]. With the exception of YeeK, which is under σK control, all other known Tgl substrates are produced early in development, under the control of σE. In previous work we have shown that safA is important, but not the exclusive determinant, for the assembly of Tgl [35]. Therefore, we wondered whether other substrates played redundant roles in the assembly of Tgl. To test this possibility, we analyzed the localization pattern of a functional Tgl-CFP fusion in single or multiple mutants unable to produce the known Tgl substrates. The functional Tgl-CFP fusion was previously described; in it, Tgl is separated from the CFP moiety by a 20Å-long, rigid, α-helical linker [35]. In wild-type (WT) sporangia, and in line with previous results [35], Tgl-CFP initially localizes to phase dark spores, as two caps, simultaneously to both the mother cell proximal (MCP) and the mother cell distal (MCD) poles of the engulfed forespore (localization class A); Tgl-CFP then encases the spore forming a continuous ring of fluorescence (class B) [35] (Fig 2A and 2B). Thus, Tgl belongs to class V of coat proteins as defined by McKenney and Eichenberger, which are produced mainly under σK control and localize to both poles of phase dark spores before encasing the forespore [17]. The accumulation of Tgl at the spore poles, which may correspond to our two-cap (class A) pattern, has been described [37]. Here, and as a reference for the localization of Tgl-CFP in different mutants, the localization of the fusion protein was assessed by scoring sporangia according to classes A and B (as defined above; with one exception, described below, other localization patterns were not considered in this study). Since the main period of tgl expression is under σK control [31, 37–39], we analyzed the localization of Tgl-CFP 6 and 8 hours after sporulation initiation, when σK is active and the majority of the sporangia have visible phase dark or phase bright spores (Fig 2A). For the WT, classes A and B correspond to ≥ 98% of the sporangia with phase dark or phase bright spores scored, in agreement with previous work [35] (Fig 2B and 2C). In both the yeeK and gerQ mutants, the sum of classes A and B corresponded to >97% of the sporangia with either phase dark or phase bright spores, similar to the WT (Fig 2B and 2C). Thus, Tgl-CFP is recruited to the forespore surface in nearly all yeeK or gerQ sporangia. The main difference found for the yeeK mutant when compared to the WT was that for sporangia of phase dark spores the representation of class A (49%) and class B (51%) was nearly the same, whereas class A represented 61% of WT sporangia with phase dark spores (Fig 2B). Deletion of gerQ also increased the representation of class B in sporangia of phase dark spores (to about 60%). The increase in class B in sporangia of phase dark spores in either mutant suggests that the absence of YeeK or GerQ facilitates encasement by Tgl-CFP. This effect seems mainly mediated by gerQ, because in a yeeK gerQ double mutant, class B in sporangia of phase dark spores represents 69% (as compared to 62% for the gerQ single mutant) (Fig 2A and 2B). These results suggest that YeeK and GerQ, per se, have no major impact on the recruitment of Tgl to the forespore but somehow control the encasement step, which seems to be more efficient in their absence. In agreement with the inference that recruitment is largely unaffected in the gerQ or yeeK mutants, or in the gerQ yeeK double mutant, quantification of the forespore/mother cell fluorescence ratio in sporangia of phase bright spores revealed a median value of 1.00 for the WT and yeeK strains, a median value of 1.02 for the gerQ mutant and of 0.95 for the yeeK gerQ double mutant (Fig 2D). Thus, in what concerns the recruitment step, essentially no difference was found between the WT and yeeK or gerQ mutants, or the double yeeK gerQ double mutant.
We then examined the assembly of Tgl-CFP in a strain where SafAFL is produced, but because the internal Met codons 161 and 164 were changed to Ala codons, the independent production of C30 through internal translation of the safA mRNA is eliminated [28, 35]. In this strain, herein termed ΔC30 (carrying the safAM161A/M164A allele) the sum of classes A and B still represent the majority (>95%) of the sporangia of phase dark spores, but this number dropped to 89% for sporangia of phase bright spores. C30 thus plays a role in recruitment of Tgl. In agreement with this conclusion, the analysis of the forespore/mother cell fluorescence ratio for the ΔC30 mutant strain shows that Tgl-CFP accumulates to slightly higher levels in the mother cell (median value of 0.95, compared to 1.00 for the WT strain) (Fig 2D).
In sharp contrast, in the ΔsafA in frame-deletion mutant, unable to produce both SafAFL and C30, the sum of classes A and B represented only 5% of the sporangia of phase dark spores and 28% of those with phase bright spores (Fig 2A, panels k-q and Fig 2B and 2C). Thus, in the absence of SafAFL and C30, recruitment of Tgl-CFP to the spore surface is severely impaired. Note that in spite of the strong localization defect, the accumulation of Tgl-CFP does not differ significantly from the WT, as shown by immunoblot analysis of whole cell extracts (S1A Fig). Also, that localization improves in sporangia of phase bright spores, further suggests a delay in the assembly of Tgl. In keeping with these inferences, the median value of the forespore/mother cell fluorescence ratio for the ΔsafA strain (0.69) is considerably lower than for the WT (1.00), implying that the elimination of safA causes accumulation of Tgl-CFP in the mother cell (Fig 2D). The ΔsafA mutation, however, does not completely eliminate the localization of Tgl-CFP, suggesting that YeeK and GerQ may partially compensate for the absence of SafAFL and C30, and that their assembly is partly independent of safA. Alternatively, additional, as yet unknown factors, may control the localization of Tgl.
We then analyzed the localization of Tgl-CFP in a ΔsafA yeeK gerQ triple mutant, that is, in the absence of all the known substrates of Tgl (Fig 2A, panels r-b´). In line with the conclusion that SafAFL and C30 are the main factors involved in the recruitment of Tgl, the sum of classes A and B in sporangia of phase dark spores did not differ much from the ΔsafA single mutant (Fig 2B). Surprisingly, however, in sporangia of phase bright spores, not only the sum of classes A and B represented 45% of the sporangia scored (Fig 2C), but the forespore/cell fluorescence ratio for the triple mutant (0.86) is significantly higher than that of the safA mutant (0.69) (Fig 2D). One possible explanation is that the simultaneous absence of SafA, C30, YeeK and GerQ unmasks an alternative SafA-independent pathway that can recruit Tgl-CFP to the spore surface, albeit inefficiently. If so, one possible implication is that additional, as yet unknown Tgl substrates may exist. In previous work we presented evidence indicating that synthesis of the spore cortex peptidoglycan facilitates assembly of Tgl [35].
Finally, we examined the assembly of Tgl-CFP in a strain bearing a safA allele with a stop codon replacing the Phe155 codon; in this strain (termed ΔsafAFL, carrying the safAPhe155STOP allele) C30 is still formed through internal translation but production of SafAFL is eliminated (Fig 2A, panels c’-l’). Tgl-CFP was not detected in sporangia of phase dark spores, and classes A and B represented 16% of the sporangia of phase bright spores, a value lower than that found for the ΔsafA mutant (Fig 2C). Thus, even though the forespore/mother cell fluorescence ratio of ΔsafAFL and ΔsafA sporangia with phase bright spores is very similar (Fig 2D), assembly of Tgl-CFP is more compromised when SafAFL (but not C30) is absent (ΔsafAFL) than when both SafA and C30 are absent (ΔsafA). Because the localization of C30 is dependent on SafAFL [28, 29] one possible explanation is that C30 retains Tgl-CFP in a mislocalized pattern. Consistent with this supposition, in the majority of the ΔsafAFL sporangia carrying phase dark spores, Tgl-CFP accumulated as a dot of fluorescence in the MCP forespore pole (Fig 2A, panel c’) (NB: this pattern will be further discussed below). Together, the results suggest that the known Tgl substrates, as well as additional factors, play partially redundant roles in the assembly of the enzyme, with GerQ and YeeK mainly involved in enforcing the normal pattern of encasement, and SafA and C30 primarily involved in recruitment. This partial redundancy in the recruitment of Tgl-CFP to the surface of the developing spore explains why assembly of the enzyme cannot be completely abolished in any of the single mutants studied.
Since the assembly of Tgl is largely controlled by proteins that are also Tgl substrates, then, it should be possible to identify residues in Tgl, involved in enzyme-substrate interactions, important for the assembly of the enzyme. In addition, mutations in those residues of Tgl could possibly have more pronounced effects on the assembly of the enzyme than the absence of the known substrates individually. Tgl has a distinctive catalytic dyad composed by Cys116 and Glu187, the latter of which can be non-reciprocally substituted by Glu115 [14] (Fig 3A). The catalytic residues of Tgl are located within a tunnel that traverses the molecule from side to side and thus are not surface exposed (Fig 3B–3F). As has been seen for the animal-type TGases, Gln (acceptor) and Lys (donor) proteins that participate in cross-linking reactions (Q and K substrates) approach the active site from opposite sides of a tunnel [40, 41] (Fig 3B). Thus, both entrances of the Tgl tunnel act as substrate docking ports. Accordingly, previous work has shown that His200, located at one of the entrances of the tunnel (in the so-called back side of the enzyme) (Fig 3D and 3F), while not playing an essential role in catalysis, is important for the interaction with substrates and in that way for the overall activity of Tgl [14]. Thus, in searching for additional residues with a role in enzyme-substrate interactions, we focused our attention on four residues at the front entrance of the tunnel (Trp149, Tyr171, Arg185 and Asn188; Fig 3C and 3E), and two residues at its back entrance (Phe69 and Trp184; Fig 3D and 3F). With the exception of Arg185, all of the selected residues are well conserved among Tgl homologues [14]. Arg185 appeared interesting, however, because with other residues in its vicinity, it shows high thermal displacement parameters (B factors) suggesting flexibility [14]. Flexibility in this region of the protein could be important for substrate interactions and for displacement of the ceiling of the tunnel following catalysis, allowing release of a cross-linked product [14] (Fig 3E).
To examine the contribution of the selected residues on the overall activity of Tgl in vitro, each was independently substituted by Ala. TglWT and its variants were overproduced with a C-terminal His6 tag in E. coli by auto-induction [15]. All the proteins accumulated in E. coli and we have shown before that the His6 tagged-Tgl is functional, both in vitro and in vivo [15]. The activity of TglWT and its variants was then evaluated using BSA as the Q substrate, and the fluorescent primary amine dansylcadaverine as the K substrate. Enzymatic activity can be assessed in this manner by monitoring the formation of fluorescent BSA over time [14]. These assays were conducted at 50ºC, which we showed before to be the optimal temperature for labeling of BSA; presumably, increased conformational flexibility of the substrate makes surface exposed Q residues more prone to labeling [14].
For two of the mutant enzymes tested, TglW149A and TglY171A, no labeled BSA could be detected (Fig 3G and 3H; not plotted). The activity of TglR185A and TglW184A was significantly reduced compared to the WT, while TglF69A and TglN188A showed only residual activity (Fig 3G and 3H). In previous work we showed that no activity could be detected for TglH200A,TglE115A and TglC116A, while TglE187A retained considerable activity (about 60% when compared to TglWT), as Glu115 acts as a substitute in the absence of the E187 side chain [14]. In all, these results are in line with the view that residues at both entrances of the tunnel are involved in enzyme-substrate interactions and in that way important for the overall activity of Tgl.
While we could not detect accumulation of fluorescent BSA during the time of the amine incorporation assay with the TglW149A variant, we detected the accumulation of fluorescent TglW149A (Fig 3G). The auto-catalytic activity of Tgl [15] and other TGases has been described [42–46], and during our labeling assay, we could also detect formation of fluorescent Tglwt concurrently with the labeling of BSA (Fig 3G). Both TglW149A and Tglwt show similar auto-labeling activity during the first 60 minutes of reaction (Fig 3I). Thus, TglW149A is essentially functional although it may be somewhat unstable over long incubation times. In any case, and importantly, TglW149A shows an altered substrate specificity when compared to TglWT, as it no longer uses BSA but labels itself. This is in line with the view that residues at the two entrances of the Tgl tunnel, as seen for other TGases, are important for enzyme-substrate interactions (above). More importantly, however, because BSA or Tgl are the Q substrate in the BSA/Tgl-dansylcadaverine labeling reactions, Trp149 must be involved in the interaction with the Q substrate. It follows that the front side of the tunnel, where Trp149 is located, corresponds to the Q substrate docking site, while the back side of Tgl is the site of interaction with K substrates (Fig 3C and 3E).
Since the substrates of Tgl are important for the localization of the enzyme to the surface of the forespore, Tgl mutants with Ala substitutions of residues located at both entrances of the tunnel should exhibit impaired localization. We tested this prediction by introducing point mutations in tgl-cfp and transferring the resulting alleles to the non-essential amyE locus of the tgl::sp insertional mutant. We then analyzed the localization of the different forms of Tgl during sporulation using fluorescence microscopy (Fig 4A). The residues analyzed included not only those located at both entrances of the tunnel that were also tested for a role in enzyme activity (above), but also the residues that form the catalytic center of Tgl, hidden inside the tunnel, that we defined previously [14] (Fig 3).
We determined the forespore/mother cell fluorescence ratio for all strains, producing the WT and mutant forms of Tgl-CFP (Fig 4B). Of the four front side residues analyzed in this way, only the R185A substitution did not seem to impair assembly of Tgl-CFP (median values of the forespore/cell fluorescence ratio of 0.99 and 1.00, for TglR185A-CFP and TglWT-CFP, respectively) (Fig 4B). Assembly of TglW149A, TglY171A and TglN188A, however, was severely impaired (median values of 0.44, 0.47 and 0.51, for the W149A, Y171A, and N188A substitutions, respectively). Moreover, for these three proteins, fluorescence dispersed throughout the mother cell cytoplasm was seen (Fig 4A, panels e, f, h). Note that all these three residues, W149, Y171 and N188, are located at the Q-acceptor side of the tunnel, and that the surface of N188, in particular, forms the bottom of the front entrance of the tunnel. Forms of Tgl with single Ala substitutions of K-acceptor side (back) residues, TglF69A, TglW184A, and TglH200A, also show impaired assembly, but judging from the quantification of the forespore/mother cell fluorescence ratios, to a lesser degree than front side substitutions (median values between 0.73–0.94) (Fig 4B). Lastly, the assembly of TglC116A, TglE115A and TglE187A, with Ala substitutions of the catalytic residues buried within the Tgl tunnel, was also reduced as compared to the WT (median values between 0.80–0.91), but to a lesser degree than Q- (median values between 0.44–0.51, R185A excluded) or K-side mutants (median values between 0.73–0.94) (Fig 4C). The catalytic residues are not likely to be directly implicated in substrate docking. Q-donor substrates, however, form an acyl-enzyme intermediate with the catalytic Cys116 residue, in the first step of the reaction catalyzed by TGases. Therefore, while Q-side residues (located at the front entrance of the tunnel) make the most important contribution to the assembly of Tgl, this step may be rapidly followed by the formation of an acyl-enzyme intermediate. In any event, the results suggest that enzyme-substrate interactions are important for the assembly of Tgl. The analysis of the levels of the various forms of Tgl in spore coat extracts is in agreement with the microscopy analysis (S1B Fig; see also the S1 Text).
To test whether the recruitment defects observed for the different forms of Tgl-CFP resulted from protein instability, extracts from sporulating cells were fractionated into a mother cell and a forespore fraction and the presence of the various forms of Tgl assessed by immunoblotting (S1C Fig). We found that all forms of the enzyme showed reduced levels in the forespore fraction, but correspondingly higher levels in the mother cell fraction (S1C Fig; see also the S1 Text). Thus, the mutations did not seem to grossly alter the folding of Tgl. In addition, the total cell fluorescence was measured for the strains expressing the various Tgl-CFP fusions (Fig 4C; see also the material and methods section). Note that we have reported before that the functional Tgl-CFP fusion shows minimal processing and release of the CFP moiety [35]. Some of the point mutations introduced in tgl-cfp, for instance E115A and F69A, lead to whole cell fluorescence lower than that found for the WT enzyme. While this analysis may indicate that some of the mutations may slightly alter the folding of Tgl-CFP, we note that there was no direct correlation between the levels at which the different mutant forms of Tgl-CFP accumulate in the cell and their recruitment to the forespore: the total cell fluorescence values were the lowest for TglE115A-CFP (0.58) and TglF69A-CFP (0.67), but the recruitment of these forms was not among the most affected, with a median value of forespore/cell fluorescence ratio of ~0.80, while the three lowest values are between 0.44 and 0.51 (Fig 4B and 4C). Thus, the defects observed in the recruitment of the different mutant forms of Tgl-CFP do not seem a direct result of protein instability and likely reflect impaired enzyme-substrate interactions, with Q-side residues (located at the front entrance) making the most important contribution to the assembly of the enzyme.
One prediction arising from the results described in the preceding section is that at least some of the Ala-substituted forms of Tgl should show reduced cross-linking activity towards its natural substrates. We used C30 to test this prediction as this is the substrate whose absence, per se, more severely impairs assembly of Tgl (above). Thus, the cross-linking activity of Tgl mutant forms whose recruitment appears more compromised, with either Ala substitutions of residues at the Q and K entrances of the tunnel, were analyzed in parallel with TglWT and two variants with Ala substitutions of catalytic residues. C30 and Tgl (WT and mutant forms) were independently over-produced in E. coli with a C-terminal His6-tag, and the various proteins were purified and incubated together at 37°C. This is the temperature at which the microscopy experiments were conducted; note, however, that the optimal temperature for Tgl activity is 50°C [14, 15], and it was at this temperature that the BSA labeling assays were conducted (see also above). A cross-linking assay with Tgl and C30 or BSA as the substrates, conducted at 50º is shown in S3 Fig, for reference. TglWT efficiently cross-linked C30 into high molecular weight species, C30m (Fig 5A and 5B, parenthesis). For TglE115A, TglC116A, TglY171A and TglN188A, no activity could be detected [Fig 5A; note that the C30 species presumed to be a dimer, (C30)2, resistant to the SDS-PAGE conditions, is detected in all assays at time 0]. In contrast, TglH200A displayed significant activity when compared to the other forms of the enzyme, with the formation of high molecular weight products during the time of the assay (Fig 5A). TglW149A displays significant enzymatic activity but an altered specificity: it shows auto-labeling activity with dansylcadaverine but is unable to label BSA (Fig 3G and 3I; see above). Yet, our results also show that the assembly of TglW149A is compromised (Fig 4 and S1B Fig). The C30 cross-linking assays, however, show that TglW149A produces cross-linked species that are not detected with Tglwt or any of the other mutant forms of the enzyme (Fig 5A and 5B, asterisks). While these forms may be the result of intramolecular cross-linking, the observation shows that TglW149A interacts with C30 differently from TglWT. Presumably, then, altered interaction of TglW149A with C30 does not permit efficient recruitment of the enzyme in vivo.
In the three mutants, ΔsafA, ΔsafAFL and the triple mutant ΔsafA, yeeK gerQA, in which the class A/class B pattern of Tgl-CFP localization is more significantly reduced (quantification in Fig 2B–2D), Tgl-CFP is detected in the majority of sporangia with either phase dark or phase bright spores as a strong dot of fluorescence at the MCP pole of the forespore (Fig 2A). Quantification of the dot pattern for the three strains shows a dot in 80% of the sporangia with phase dark spores and 12% of those with phase bright spores for the ΔsafAFL mutant (Fig 5C). The dot pattern was less represented in the ΔsafA single mutant or the yeeK gerQ ΔsafA triple mutant (Fig 5C). We used transmission electron microscopy (TEM) to characterize further the phenotype of the ΔsafAFL mutant (Fig 5D). The TEM images of sporangia at an intermediate stage of sporulation shows the accumulation of coat material at the MCP forespore pole in ΔsafA sporangia and projecting into the mother cell cytoplasm, as previously noted [22] (Fig 5D, middle panel). This phenotype is corrected by safAwt at amyE (Fig 5D, left panel). Strikingly, accumulation of coat material was seen at the MCP forespore pole in ΔsafAFL sporangia (Fig 5D, right panel). In the ΔsafAFL strain, the C30 form, but not SafAFL, is produced [28]. Because C30 is not assembled in the absence of SafAFL [27, 28], one possibility is that C30 retains Tgl-CFP, along with other coat proteins, in the mother cell.
If so, TglWT and C30 should co-localize in ΔsafAFL sporangia. Conversely, Tgl variants whose assembly and the ability to cross-link C30 in vitro is impaired should not co-localize with C30. To test these predictions, we co-expressed C30-yfp along with tglwt-cfp in the ΔsafAFL mutant. Because the single fluorescence dot pattern of Tgl-CFP is more easily seen in sporangia with phase dark spores, the co-localization of Tgl-CFP and C30-YFP was only scored in these cells. We found that C30-YFP localizes as a bright dot of fluorescence in the majority of the sporangia (84%) and that Tglwt-CFP co-localized with C30 (in 76% of the sporangia scored) (Fig 6A and 6B). Forms of Tgl with Ala substitutions of K-side residues (F69A, H200A), or the catalytically inactive TglC116A-CFP form, also co-localized with C30-YFP (values varying between 69 and 70%, as indicated in Fig 6B), although a weak CFP fluorescence signal was also seen dispersed throughout the mother cell cytoplasm (Fig 6A). Strikingly, however, TglY171A-CFP and TglN188A-CFP (with Ala substitutions in Q-side residues) largely accumulate throughout the mother cell cytoplasm, and only co-localize with C30-YFP in 2 and 10% of the sporangia scored (Fig 6A and 6B). Note, however, that the C30-YFP dot at the MCP forespore pole is maintained in tglY171A-cfp (77% of the sporangia) or tglN188A-cfp (80%) sporangia. Thus, forms of Tgl with Ala substitutions of residues in the Q acceptor side of Tgl, which show impaired assembly, but much less so K-side residues, fail to co-localize with C30 in the mother cell cytoplasm. This observation lends strong support to the idea that assembly of Tgl is mainly determined by its interaction with the SafAFL/C30 substrates and that Q-side residues make a more important contribution to the assembly of Tgl than do K-side residues.
As a control for these experiments, we also wanted to test whether a SafA-dependent protein that is not a direct substrate of Tgl would still co-localize with C30 when the cells produce a form of Tgl, such as TglY171A whose localization is impaired. Since in the absence of SafAFL accumulation of some coat material is seen by TEM at the MCP forespore pole (Fig 5D; see above), we expected that such a protein would form dots at the MCP forespore pole regardless of the form of Tgl produced by the cells. A YaaH-GFP fusion was shown before to be dependent on SafA for localization to the inner coat [17, 25, 36] and YaaH is not a Tgl substrate [30–35]. We constructed a YaaH-YFP fusion and examined its localization as well as the localization of TglWT-CFP or TglY171A-CFP, in the strain producing C30 but not SafAFL (S4A Fig). In this experiment, Tgl-CFP formed a dot in 88% of the sporangia examined, while dots were only seen in 5% of the sporangia for TglY171A, the most common pattern being the fluorescence signal dispersed throughout the mother cell cytoplasm (S4B Fig, class b; see also above). For YaaH-YFP, however, no dots of fluorescence were seen at the MCP forespore pole; instead, YaaH-YFP localized as a ring around the forespore (S4A and S4B Fig). This result shows that mislocalization of C30 is not sufficient to retain all of the known SafA-dependent proteins in the mother cell. On the other hand, it strongly reinforces the specificity of the TglWT-C30 interaction and its requirement for the proper localization of Tgl.
The results in the preceding sections suggest a model in which Tgl is recruited to the surface of the developing spore by its pre-assembled substrates, with SafA and C30 making the most important contribution. This model raises the possibility that the activity of Tgl is itself controlled by the local concentration of the pre-assembled substrates. To investigate how the activity of Tgl varied with the concentration of its substrates, we conducted cross-linking assays using purified C30. In a first series of assays, C30, at a fixed concentration (25 μM) was incubated with various concentrations of Tgl, and the formation of cross-linked species monitored by SDS-PAGE. Total cross-linking was estimated by measuring the decrease in the representation of C30 over time, which is converted into high molecular weight forms (S5A Fig; see the material and methods for details on the quantification). Incubation of C30 alone results in the formation of the (C30)2, dimer (at about 60 kDa) that resists the denaturing conditions of SDS-PAGE (S5A Fig). Formation of this species was increased for long incubation times in the presence of the lowest concentration of Tgl tested (0.8 μM); at an intermediate concentration of Tgl (4 μM), the cross-linking activity increased linearly with time (Fig 7A). For the next concentration of enzyme tested (8 μM), activity increased linearly during the first 70 min of incubation but then reached a plateau (Fig 7A). For a higher concentration of Tgl (12.5 μM), however, activity increased linearly up to 120 minutes of incubation (Fig 7A). Together, these results suggest that the enzyme is somehow inactivated over time, in a manner that depends on the enzyme/substrate ratio.
If so, then the plateau effect seen at 8 μM of Tgl should be eliminated by decreasing the concentration of substrate. To test this, Tgl, at a final concentration of 8 μM, was incubated with progressively lower concentration of C30. Surprisingly, we found that the activity of Tgl decreased with an increase in substrate concentration (Fig 7B; representative gels are shown in S5B Fig). When C30 was present at 25 μM, the highest substrate concentration tested, enzyme activity also reached a plateau, after 60 min of incubation (Fig 7B). Inspection of the SDS-PAGE gels used to quantify the activity of Tgl show that the enzyme is depleted over time, in a manner that is dependent on the concentration of substrate (S5A and S5B Fig). This suggests that Tgl itself becomes cross-linked. To more precisely test this idea, we incubated Tgl alone (at 4 μM) and after 20 min, C30 was added to a final concentration of 25 μM. During the first 20 min of the reaction, the intensity of the band corresponding to Tgl, at 26 kDa, is slightly reduced (S5C Fig). When C30 is added (to 25 μM), however, the intensity of the Tgl band is rapidly reduced, concomitantly with the formation of the (C30)2 dimer (S5C Fig). The quantification of this effect is shown in Fig 7C. The decrease in Tgl when incubated alone is consistent with the previously described auto-cross-linking activity [14]) and with the auto-labelling activity detected in the presence of dansylcadaverine (above). Together, these results suggest that Tgl itself is cross-linked during the reaction, and that the ratio of Tgl/substrate will limit the duration and extent of C30 cross-linking.
C30 self-interacts in a yeast two-hybrid system [27], and it has been proposed that it forms high molecular weight complexes that are cross-linked by Tgl at the surface of the developing spore [35]. The oligomeric state of C30 in solution, however, is not know. Also, the cross-linking assays described in the preceding section indicate that Tgl is itself cross-linked to C30, which could possibly be facilitated if C30 forms an oligomeric structure. To gain insight into the oligomeric state of C30, the purified protein (S6A Fig) was analyzed by size exclusion chromatography (SEC) and small-angle X-ray scattering (SAXS). In SEC, purified C30 eluted as a main single sharp peak (ca 1 ml wide) corresponding to a molecular mass of 151.4 ± 7.5 kDa, i.e., the mass of a possible hexamer, herein termed (C30)6 (S6B Fig). The peak fractions were then analyzed by SEC-SAXS [47]. SAXS data indicate that (C30)6 is a globular particle in solution with a radius of gyration, Rg of 63.1 ± 0.1 Å and a maximum intramolecular distance, Dmax of 220.0 ± 10.0 Å (Fig 7C). The molecular weight estimation from the SAXS data is of 160.1 ± 15 kDa, in good agreement with the SEC data, and suggesting that in the absence of Tgl the particle is a stable hexamer. The C30 oligomer in solution can accommodate six molecules of Chymotrypsinogen A, a monomeric globular protein with a molecular mass similar to a C30 monomer (25.9 kDa) ([48]; S6C Fig). The smooth asymmetrical pair-wise distance distribution, P(r), indicates that (C30)6 is slightly aspheric (Fig 7E). The corresponding low-resolution shapes of (C30)6 reconstructed based on the P(r) [47] (inset in Fig 7D) fit the experimental data well (χ2 = 2.2), and provide an oblong envelope model (average NSD = 0.8; see also the material and methods section).
This analysis suggests that C30 forms a large oligomer, whose repeating unit may be the (C30)6 hexamer detected both in our SEC and SAXS analysis. C30 is detected as a dimer by SDS-PAGE even in the absence of Tgl (Figs 5A, 5B, 5A and S5B) suggesting that this species is highly stable. In the presence of Tgl, (C30)2 increases rapidly before higher molecular weight forms of the protein accumulate (Fig 5A and S5 Fig; see also above). This suggests that (C30)6 is a trimer of cross-linked dimers, although other architectures can explain the accumulation of cross-linked (C30)2. A cross-linked (C30)6 hexamer in turn, may be an intermediate in the formation of higher molecular weight species. The size of the complexes obtained in the presence of Tgl precluded, for the moment, a study by SAXS. In any event, we posit that (C30)6 and higher order oligomers accumulate at the surface of the developing spore, and that these pre-assembled complexes recruit Tgl. Cross-linking of these species by Tgl may be self-limiting, in that the enzyme becomes cross-linked into the forming structure, and thereby unable to conduct further catalysis.
We show that the recruitment of Tgl to the forming coat is largely controlled by its substrates, in what we refer to as substrate-driven localization. Of the four known Tgl substrates, SafAFL and C30 are the main factors controlling recruitment of Tgl, whereas YeeK and GerQ appear to act mainly to delay encasement. Encasement is controlled by the SpoVID protein, with which SafA directly interacts [49, 50]. Thus, spore encasement by Tgl may rely primarily on its interaction with SafAFL and C30, and it seems plausible that GerQ and YeeK compete with Tgl for a binding interface in SafA, or that in their absence SafA is more accessible to Tgl.
While the independent formation of C30 through internal translation is not essential for inner coat formation, the corresponding region in SafAFL, i.e., the C-terminal moiety of the protein (Fig 1B), is likely to have the main role in inner coat assembly. Consistent with the idea that the C30 region carries the morphogenetic information for inner coat assembly, not only does the overproduction of C30 in an otherwise WT background blocks sporulation [28] but, as we now show, production of C30 in the absence of SafAFL results in the retention of Tgl and the accumulation of coat material at the MCP forespore pole (Fig 5D and Fig 6). Deletion of safA causes a much less severe coat mislocalization phenotype [27, 36]. Possibly, only inner coat material is misassembled in a safA mutant, whereas C30 additionally retains outer coat material in the mother cell (Fig 5D). Nevertheless, since YaaH-YFP still forms a ring around the forespore not all inner coat proteins are retained in the mother cell when C30 is produced in the absence of SafAFL (S4 Fig). In any event, the retention of TglWT by C30 at the MCP forespore pole, but not of Tgl variants with Ala substitutions of Q-side residues (Fig 6), provides strong support for the substrate-driven localization model.
C30 self-interacts, C30 interacts with SafAFL and the latter interacts with itself via the C30 region [27–29]. It has been suggested that these interactions create a multivalent platform for binding of the inner coat proteins and that Tgl exerts mainly a “spotwelding” activity, fortifying pre-formed assemblies rather than catalyzing de novo protein polymerization [35]. Both types of activities, however, have been described for human and human-like TGases ([2]; see also below). Our finding that C30 is a hexamer in solution (Fig 7 and S6 Fig) provides strong support for both ideas (Fig 8). In addition, the Tgl-mediated cross-linking of (C30)6 into larger molecular weight species in vitro may represent a case of de novo protein polymerization. In any event, it seems likely that Tgl cross-links and fortifies in vivo of a scaffold formed through the interactions among the various forms of SafA (Fig 8).
The requirement for its substrates for the localization of Tgl then appears as a mechanism for directing the enzyme to the exact locations, within the coat, where its “spotwelding” activity is needed. SafAFL recruits C30, YeeK and GerQ to the inner regions of the coat [51]. Tgl itself is recruited to the cortex region and to the inner coat by SafA and we argued before that it most likely acts to cement the cortex/inner coat interface [27, 29, 35]. As we recently showed, SafA itself has an important role in proper assembly of the cortex/inner coat and inner/outer coat interfaces [30]. Since at least SafAFL, C30 and GerQ are produced earlier than Tgl, recruitment of the enzyme is governed by its time of synthesis, which is primarily under σK control. Thus, substrate-driven localization imparts both spatial and temporal control over the activity of Tgl, positioning the enzyme in the vicinity of its substrates, so that the correct proteins or assemblies are timely cross-linked, reminiscent of the “assembly-driven regulation of cross-linking” described for FXIII [2, 5, 52]. Incubation of Tgl with C30, while leading to cross-linking of C30, also brings about cross-linking of Tgl, in a manner that is influenced by the concentration of substrate (Fig 7A–7C and S5 Fig). Since Tgl is found as high molecular weight cross-linked species in both the cortex and inner coat of mature spores [35], it is tempting to speculate that Tgl is eventually immobilized through cross-linking (Fig 8), limiting its action during maturation of the coat.
The substrate-dependent localization of Tgl may also minimize unwanted protein cross-linking reactions. The crystal structure of Tgl shows the catalytic center of the enzyme located within a 15Å-long and ~6Å wide (on average) tunnel that crosses the molecule [14]. While the tunnel appears narrow, and Tgl may be a narrow-specificity enzyme, it is able to catalyze amine incorporation into and cross-link of non-physiological substrates such as BSA (Fig 3G and S3 Fig), especially at high enzyme/substrate concentrations, conditions that conceivably could also be found in the coat [14, 15].
Residues at the two entries of the tunnel are the docking sites for the Q and K substrates [14] (Fig 3). We present evidence suggesting that the Q-substrate docking site of the enzyme corresponds to its front side, where Trp149 is located, and we show that front (Q) side residues make more important contributions to the localization of Tgl than back (K) side residues (Fig 5A and 5B). Following binding to the enzyme, the Q- substrate forms a γ-glutamythiolester with the active site Cys116, a step that is also dependent on Glu187 or Glu115 [14]. The K-substrate then binds the acyl-enzyme intermediate and attacks the thiolester bond, re-generating the active site Cys 116 residue and resulting in protein cross-linking (Fig 1B). Formation of the acyl-enzyme intermediate is thought to be the rate-limiting step in the reactions catalyzed by TGases [2]. During coat formation, however, docking of the Q-substrate, appears to be the limiting step in the localization of Tgl (Fig 4; see also the final model in Fig 8). After this interaction, a second Tgl-substrate interaction is allowed at the back-side entrance of the tunnel, with the K substrate. That back-side residues do not compensate for Ala substitutions of front side residues, is in line with the K substrate interaction only occurring after docking of the Q-substrate and formation of the acyl-enzyme intermediate as proposed for the reaction cycle of TGases (Fig 1B). It seems possible that the K-substrate interaction does not immediately follows Q-substrate docking, as the cross-linking activity of Tgl is delayed until spore release [33, 53]. Although Tgl cross-links itself in the cortex and inner coat [35], as mentioned above, it is unlikely that the activity of Tgl is directly required for its localization. Not only the buried catalytic residues make a smaller contribution than Q- or K-side residues but TglC116A, inactive in vitro, is less impaired for localization than TglH200A, which is active in vitro (Fig 3 and Fig 5A and 5B). This is consistent with the idea that enzyme-substrate docking is the main mechanism controlling the localization of Tgl, with formation of the acyl-enzyme intermediate stabilizing the Q-substrate-enzyme interaction. With respect to C30, (C30)6 acts as both the Q and K donor substrate (Fig 8).
The substrate-dependent localization mechanism that we propose for Tgl is also akin to that of some PBPs, the localization of which is prevented by active site mutations, substrate reduction or masking by certain antibiotics, in that it efficiently couples sub-cellular localization and proper timing of enzyme activity [54–57]. It differs in that no substrates other than the terminal D-Alanyl-D-Ala of the peptidoglycan stem peptide are known for the PBPs. TGases, in contrast, have the potential to cross-link proteins other than their cognate substrates, and the time and site where they exert their activity is tightly regulated, as exemplified for FXIIIa. All known human and human-type TGases undergo complex activation mechanisms, and even another bacterial TGase, the MTG protein from Streptomyces mobaraensis, which is structurally characterized, is produced as an inactive zymogen [2–4, 58, 59]. In contrast, Tgl is synthesized in active form. The substrate-dependent localization of Tgl may thus the primary mechanism to ensure proper temporal and spatial control of spore coat protein cross-linking, as well as specificity and the extent of its action (Fig 8). Substrate-dependent localization may similarly control the activity of Tgl orthologues in other spore-forming bacteria.
The construction of all plasmids and strains is described in the supporting material; strains, plasmids and primers used in this study are listed in S1, S2 and S3 Tables.
Sporulation was induced by the resuspension method as described [60]. Briefly, cultures were grown in growth medium, and the cells collected at an OD600 of ~0.4, and transferred to a minimal medium in which sporulation is induced. Samples were taken 6 and 8 hours after resuspension (defined as the onset of sporulation), the cells collected by centrifugation (1 min at 2.400 x g, room temperature), and washed with 1 ml of phosphate-buffered saline (PBS). Finally, the cells were resuspended in 20 μl of PBS and applied to microscopy slides coated with a film of 1.7% agarose. Images were taken with standard phase contrast, CFP and YFP filters, using a Leica DM 6000B microscope equipped with an aniXon+EM camera (Andor Technologies), and driven by Metamorph software (Meta Imaging series 7.7, Molecular Devices).
Image analysis was conducted with Meta Imaging series 7.7 software and the determination of the values of fluorescence in the forespore or in the mother cell cytoplasm was conducted as described [35]. The individual values of the forespore/cell fluorescence ratio, (FFS/Fcell), or cell fluorescence, Fcell, were examined with GraphPad Prism 5 (GraphPad Software, Inc) which indicated that the values showed a non-normal distribution. Thus, for normalization purposes each individual value was divided by the median value of the wild type strain that had been grown on the same day and under the same conditions; the same normalization was conducted for the individual values of the wild type strain. The calculation of the normalized values of the forespore/cell fluorescence ratio, (FFS/Fcell)norm, or cell fluorescence, (Fcell)norm, was according to Eqs 1 and 2, respectively:
(FFS/Fcell)norm=(FFS/Fcell)median(FFS/Fcell)wt
(Eq 1)
(Fcell)norm=Fcellmedian(Fcell)wt
(Eq 2)
Samples were collected from DSM cultures at hour 4 of sporulation and processed from thin sectioning transmission electron microscopy essentially as described [61]. Samples were viewed on a Hitachi H-7650 microscope equipped with an AMT digital camera and operated at 100 keV.
Tgl (WT and mutant forms) and C30 were introduced in a pET derivative (Novagen) fused to a C-terminal His6-tag and protein over-production was achieved by auto-induction [15]. The different proteins were purified in 1 ml Ni2+-NTA Agarose columns (for details see the Supporting Material). Enzyme activity assays were conducted at 37ºC or 50ºC in Tris-HCl 0.1 M, pH 8.0, as described before [14]. The following concentrations were used: BSA (New England Biolabs), 60 μM; dansylcadaverine (Fluka), 0.5 mM; and Tglwt/mut (16 μM). BSA/Tgl labeling by dansylcadaverine was detected after SDS-PAGE resolution using a UV transilluminator (Chemidoc XRS, Biorad) and quantified using ImageJ 1.37v [62]. The values obtained for each time point were normalized by the fluorescence detected for the wild type enzyme after 120 min of incubation. All Tgl forms were purified and assayed at least three times independently.
Tgl (wild type or mutant forms, at the indicated concentrations) was incubated with C30 (at the indicated concentrations), at 37° C, in Tris-HCl 0.1 M, pH 8.0. Samples were taken at different times and resolved in 10% SDS-PAGE gels, which were subsequently stained with Coomassie brilliant blue. For graphical representation of the data, the stained SDS-PAGE gels were scanned and the amount of C30 or Tgl at each time point was quantified using ImageJ 1.37v [62]. The values obtained for each time point were normalized according to the following equation (where t represents the different time points, in minutes, in a cross-linking reaction):
(Proteincross‑linking)norm=(1‑proteinattimettproteinattimet0)*100
Synchrotron SEC-SAXS data were collected on the BM29 ESRF beamline (Grenoble, France) using an in-line HPLC system and used to generate an average low-resolution shape representation of C30 oligomer in solution. Please see S1 Text for details.
|
10.1371/journal.ppat.1003303 | Identification of Targets of CD8+ T Cell Responses to Malaria Liver Stages by Genome-wide Epitope Profiling | CD8+ T cells mediate immunity against Plasmodium liver stages. However, the paucity of parasite-specific epitopes of CD8+ T cells has limited our current understanding of the mechanisms influencing the generation, maintenance and efficiency of these responses. To identify antigenic epitopes in a stringent murine malaria immunisation model, we performed a systematic profiling of H2b-restricted peptides predicted from genome-wide analysis. We describe the identification of Plasmodium berghei (Pb) sporozoite-specific gene 20 (S20)- and thrombospondin-related adhesive protein (TRAP)-derived peptides, termed PbS20318 and PbTRAP130 respectively, as targets of CD8+ T cells from C57BL/6 mice vaccinated by whole parasite strategies known to protect against sporozoite challenge. While both PbS20318 and PbTRAP130 elicit effector and effector memory phenotypes in both the spleens and livers of immunised mice, only PbTRAP130-specific CD8+ T cells exhibit in vivo cytotoxicity. Moreover, PbTRAP130-specific, but not PbS20318-specific, CD8+ T cells significantly contribute to inhibition of parasite development. Prime/boost vaccination with PbTRAP demonstrates CD8+ T cell-dependent efficacy against sporozoite challenge. We conclude that PbTRAP is an immunodominant antigen during liver-stage infection. Together, our results underscore the presence of CD8+ T cells with divergent potencies against distinct Plasmodium liver-stage epitopes. Our identification of antigen-specific CD8+ T cells will allow interrogation of the development of immune responses against malaria liver stages.
| Vaccination against malaria is feasible, as demonstrated with radiation-attenuated sporozoite vaccine, which protects experimental animals and humans by targeting the clinically silent liver stages. Potent protection largely depends on CD8+ T cells, a type of white blood cell that is tailor-made to kill obligate intracellular pathogens. Malaria-infected cells display fragments of parasite proteins, which are then recognised and targeted by CD8+ T cells. How CD8+ T cells are activated following immunisation and how they execute protective functions are key considerations for vaccination. However, characterisation of CD8+ T cells is hampered by the lack of identified malaria protein targets. Of concern, the circumsporozoite protein, which is the basis of the most advanced malaria vaccine candidate (RTS,S), is not an essential target of CD8+ T cells induced by attenuated sporozoites in several mouse strains. In this study, we have made considerable advances by identifying for the first time, fragments of malaria proteins that are targeted by CD8+ T cells generated by vaccination in a relevant mouse strain, C57BL/6. Notably, CD8+ T cells against one of the target proteins elicit partial protection against infection. Our study exemplifies how immunisation by complex pathogens can be dissected to identify distinct antigens for subunit vaccine development.
| Malaria is responsible for an estimated 250 million episodes of clinical disease and 600,00 to 1.2 million deaths each year [1], [2]. Notwithstanding recent reductions in the burden of malaria in some endemic areas, sustained control, elimination or eradication of the disease will require a highly efficacious vaccine that prevents malaria transmission as well as reducing the burden of disease. As a benchmark in malaria vaccination, multiple immunisations of γ-radiation-attenuated Plasmodium sporozoites (γ-Spz) can protect both mice and humans against sporozoite challenge [3], [4]. The elicited protection targets the development of liver stages and completely prevents blood stage infection, resulting in sterile immunity. This experimental vaccine approach has now been replicated using other whole sporozoite immunisation strategies that include infection under drug cover and genetically arrested parasites [5]–[8]. Naturally acquired pre-erythrocytic immunity is likely multifactorial [9], involving both antibodies and T cells. However, CD8+ T cells are the prime mediators of protection after γ-Spz vaccination in mice [10], [11], and interferon (IFN)-γ is a signature of effector function [12].
How CD8+ T cells are primed, modulated, and maintained following immunisation, and how these cells execute protective functions, are key considerations for vaccine design and can only be addressed with antigen-specific tools. The circumsporozoite protein (CSP), the major surface protein of the sporozoite, has been at the forefront of vaccination studies for more 20 years – being the basis of RTS,S, the most advanced malaria vaccine to date [13]. Furthermore, CSP-specific responses have been the standard in measuring cellular responses to malaria liver stages in fundamental immunological studies in mice [14], [15].
Murine models of sporozoite immunisation have largely focused on two strains, BALB/c and C57BL/6 (B6). Immunisation with Plasmodium berghei (Pb) or P. yoelii (Py) γ-Spz induces highly protective, H2d-restricted CD8+ T cell responses to defined CSP epitopes in BALB/c mice [16], [17]. However, protection can also be obtained in the absence of PyCSP-specific T cells: (a) PyCSP-transgenic BALB/c mice - that are tolerant to CSP - can be completely protected by Py γ-Spz immunisation [18] and (b) there is cross-species immunity to sporozoites despite lack of cross-reactivity of the CSP-derived CD8+ T cell epitopes [19]. These data highlight the importance of non-CSP antigens in generation of protective immunity to liver stages. However, the paucity of liver-stage specific antigens for CD8+ T cells, and the limited availability of gene-targeted mice on the BALB/c background, has limited both the evaluation of subunit vaccine candidates in murine malaria models and the characterisation of the mechanisms underlying CD8+ T cell mediated protection.
In contrast to the ease of inducing protective immunity in BALB/c mice, B6 mice expressing H-2b can only be protected against Pb (or Py) infection by multiple rounds of γ-Spz immunisation. Most importantly, protection is entirely independent of CSP-specific CD8+ T cells [18]. Indeed, the CSP seems to contain no naturally processed and presented H-2b-restricted epitopes. We propose, therefore, that the Pb-B6 model is a more relevant model of liver stage immunity than the BALB/c model. It more closely resembles the situation in humans, where CD8+ T cell responses to the CSP are infrequent [9]. These immune-epidemiological findings in malaria-endemic areas are reflected by the fact that multiple immunisations are needed to elicit sterilising immunity [20] Moreover, B6 mice are particularly attractive for immunological studies due to the availability of a large collection of sub-strains with targeted gene deletions.
In order to develop a Pb-B6 model of antigen-specific CD8+ T cell-mediated anti-liver stage immunity, we employed an unbiased genome-wide approach for screening H-2b (Kb and Db) restricted Pb-derived peptides that are recognised by CD8+ T cells from B6 mice immunised with whole sporozoite immunisation strategies known to induce protection. Our results identify two novel liver stage immunogenic targets of effector CD8+ T cells in immunised B6 mice. Of these two, CD8+ T cell responses to PbTRAP confers partial efficacy against sporozoite challenge in vivo. Considering that P. falciparum TRAP (PfTRAP) is a major target of human malaria vaccine development, our results emphasize the translational relevance of the Pb-B6 model.
The data and tools available at the Immune Epitope Database and Analysis Resource (www.iedb.org) were used for the identification of putative CD8+ T cell epitopes [21]. Systematic epitope profiling has previously identified previously unrecognized CD8+ T cell responses to a number of viral infections including vaccinia, dengue and herpes viruses [22]–[24]. To assemble a genome-wide peptide library, Pb open-reading frames, based on published sporozoite and liver stage transcriptomic and proteomic data [25]–[29], were scanned in silico using artificial neural network methods [30] for major histocompatibility complex (MHC) Class I H2-Kb and Db restricted peptides. In addition, predictions were performed using stabilised matrix methods [31] on the entire Pb draft genome [32]. Finally, Pb orthologs of Pf proteins that were reported to be antigenic for either human antibodies or T cells from individuals immunised by irradiated Pf sporozoites (Pf γ-Spz) [33], [34] were also analysed.
From this in silico analysis, 600 unique peptide sequences (288 8-mers, 311 9-mers and one 10-mer), which correspond to >350 Pb antigens, were identified and subsequently produced by solid phase synthesis (Table S1: Summary of datasets and Table S2: Complete list of peptides). Individual peptides were tested and CD8+ T cell-derived IFN-γ was quantified by two complementary read-outs: (1) an enzyme-linked immunospot (ELiSpot) assay [35] (Figure 1A,C), and (2) direct peptide-stimulation followed by intracellular cytokine staining (ICS) (Figure 1B,C). Animals received two immunisation doses of one of four whole sporozoite vaccination strategies: (i) Pb γ-Spz [3] and live sporozoites (PbSpz) given concomitantly with anti-malaria drugs (ii) azithromycin (AZ) [8], (iii) primaquine (PQ) [7] or (iv) chloroquine (CQ) [6], [36]. As negative controls, CD8+ T cells were isolated from mice immunised with heat-killed sporozoites (PbHKSpz), known to elicit sub-optimal T cell responses [37], [38], and naïve mice.
Two peptides consistently elicited robust IFN-γ responses in a proportion of CD8+ T cells isolated across all four whole sporozoite vaccine strategies (Figure 1A,B) but not from PbHKSpz (Figure 1B) and naïve mice (data not shown). Several other peptides were weakly reactive during initial screens but were not confirmed upon re-screening.
The first peptide, VNYSFLYLF, contains motifs for Kb and is derived from amino acids 318–325 of the PbS20 protein [PbS20318 (or PbS20318–325)] (PBANKA_142920; gi: 40950503), an uncharacterised protein that is conserved in Pf (Figure S1). S20 was first identified as a sporozoite-specific gene in Py [26]. PbS20318 is located within a galactose oxidase (central domain) superfamily motif of the protein (Figure S1).
The second peptide, SALLNVDNL, is restricted for Db and is derived from amino acids 130–138 of the PbTRAP [PbTRAP130 (or PbTRAP130–138)] (PBANKA_134980; gi: 1813523) [39], also known as sporozoite surface protein 2 [40] or sporozoite gene 8 (S8) [26] (Figure S1). Conserved in Pf (Figure S1), TRAP is a secreted transmembrane protein of sporozoites that plays a vital role in parasite motility and invasion of hepatocytes [41]. PbTRAP130 is located within the von Willebrand factor type A domain (Figure S1A), the key motif for parasite locomotion and target cell entry [42]. Reactive CD8+ T cells to PbTRAP130 are considerably more abundant than those reactive to PbS20318 (Figure 1A–C). The identification of a PbTRAP-derived peptide as a target of CD8+ T cells in B6 mice is of interest since PfTRAP has been a major target for malaria vaccine development in humans [43]–[46]. Thus far, no CD8+ T cell epitope in TRAP has been identified in a murine model, meaning that fundamental studies on the induction, differentiation and long-term persistence of protective TRAP-specific cells following parasite immunisation could not be carried out yet.
PbS20318 and PbTRAP130 represent the first reported endogenously processed CD8+ T cell epitopes of malaria liver stages in the B6 model. To determine expansion and contraction of PbS20318- and PbTRAP130-specific CD8+ T cells over time, we quantified the responses in the spleen and the liver after one or two immunisations with Pb γ-Spz (Figure 2A). After a single immunisation, CD8+ T cell responses in both the spleen and the liver reach the highest magnitude on day 7 (Figure 2A,B). The responses were slightly decreased on day 14 as contraction of the response occurs but they remained quantifiable for up to 180 days after immunisation. The percentages of antigen-specific CD8+ T cells were generally higher in the liver that in the spleen. More robust responses were observed after two immunisations with Pb γ-Spz (Figure 2A,B) Polyfunctional analysis of PbS20318- and PbTRAP130-specific CD8+ T cells revealed the induction of IFN-γ positive cells and IFN-γ/tumour necrosis factor (TNF) double positive CD8+ T cells (Figure S2). Consistent with the generation of effector and effector memory responses, PbS20318 and PbTRAP130-specific IFN-γ-producing CD8+ T cells were immunophenotyped as CD62Llo, CD44hi, CD11ahi, and CD49dhi (Figure 3, S3). Cells stimulated with no peptide or cells from naïve mice stimulated with either peptide did not respond to either PbS20318 or PbTRAP130 (data not shown). Together, these results indicate that immunisation with Pb γ-Spz recruits antigen-specific CD8+ T cells to undergo differentiation, proliferation, and long-term persistence.
To determine the in vivo cytotoxic potential of PbS20318- and PbTRAP130-specific CD8+ T cells, we utilised an assay that allows the quantification of rapid killing of adoptively transferred target cells by activated CD8+ T cells in vivo [47]. CFSE-labelled and peptide-pulsed syngeneic targets were transferred to Pb γ-Spz-immunised mice 14 days after the last immunisation (Figure 4A). We observed considerable (∼90%) disappearance of PbTRAP130,-pulsed (Figure 4B,C), but not PbS20318-pulsed, target cells when transferred to mice that were immunised twice with Pb γ-Spz. To corroborate our finding that PbTRAP130-specific CD8+ T cells exhibit significant cytotoxic activity, we repeated the cell transfer to mice that were immunised only once and to naïve controls (Figure S4). Cytotoxicity against cells presenting PbTRAP130, but not PbS20318, was already apparent after a single immunisation.
To test whether the newly identified targets of CD8+ T cells contribute to protection, we first performed peptide-tolerisation experiments. This method of depleting antigen-specific CD8+T cells, being performed in a malaria model for the first time, was adapted from previous studies aimed at inducing and maintaining antigen-specific tolerance [48]–[51]. Mice were subjected to repeated high dose administrations of adjuvant-free PbS20318 and PbTRAP130 peptides prior to and during the immunisation protocol (two immunisations with Pb γ-Spz) (Figure 5A). These mice, along with sham-tolerised mice, were challenged with sporozoites and liver parasite loads were measured 42 hours later. As shown in Figure 5B, the levels of protection, i.e. very low parasite load, in PbS20318-tolerised mice were similar to sham-tolerised controls. In contrast, a significantly increased liver parasite load was observed in PbTRAP130-tolerised mice. These data indicate that a substantial degree of protection in whole sporozoite-immunised animals, measured by a reduction of parasite liver load over four orders of magnitude, can be attributed to PbTRAP130-specific responses.
These results are in agreement with the in vivo cytotoxicity experiments (Figures 4, S4), where we observed the potent in vivo cytotoxic function of PbTRAP130-specific, but not PbS20318-specific, CD8+ T cells. However, it remains to be determined whether parasite killing per se requires the lytic capacity of antigen-specific CD8+ T cells. To corroborate our finding that PbTRAP130-specific CD8+ T cells contribute to protection against malaria liver stages, the tolerisation experiments were also performed in mice that were immunised only once (Figure S5A). As shown in Figure S5B, the contribution of PbTRAP130,-specific CD8+ T cells to protection was already apparent after a single immunisation. Together, our results show a critical and immunodominant contribution of PbTRAP130-specific CD8+ T cells in parasite killing.
To determine the efficiency of tolerisation, we measured PbS20318- and PbTRAP130-specific CD8+ T cell responses in the tolerised mice (Figure 5C and S5C). Indeed, in spleens of mice injected with the respective peptides, IFN-γ secretion was greatly reduced and was comparable to naïve mice. Importantly, PbS20318-tolerised mice mounted PbTRAP130-specific CD8+ T cell responses comparable to non-tolerised immune mice. Similarly, PbTRAP130-tolerised mice mounted PbS20318-specific CD8+ T cell responses indistinguishable from controls. These results demonstrate that induction of tolerance was peptide-specific and did not interfere with induction of CD8+ T cell responses against other antigens. Of note, the absence of a CD8+ T cell response to one antigen did not increase the response to another, suggesting the lack of compensation in the immunodominance hierarchy in our infection model, in contrast to other infections [51]–[54].
At least three immunisations with Pb γ-Spz are needed to induce sterile protection in the B6 model [8], [55]. To ascertain whether the development of sterile immunity is dependent on responses to the identified antigens, peptide tolerisation experiments were repeated to mice that were immunised three times with Pb γ-Spz (Table S3); 14 days after the last immunisation, mice were challenged with sporozoites. Similar to control sham-tolerised mice, multiply immunised PbS20318- and PbTRAP130-tolerised mice were completely protected and did not develop patent parasitaemia. The results are reminiscent of observations in transgenic Balb/c mice tolerant to T cell responses to PyCSP where complete protection was achieved following three immunisations with Py γ-Spz despite an immunodominant and protective role for PyCSP after one and two immunisations [18]. Taken together, it is likely that additional antigens contribute to protection in both the B6 and Balbc models.
Since PbTRAP130-specific, but not PbS20318-specific, CD8+ T cell responses are cytotoxic in vivo and contribute to protection after one or two immunisations with Pb γ-Spz, we evaluated the immunogenicity and protective efficacy of PbTRAP in a heterologous prime-boost vaccine regimen with viral vectors. Priming with adenovirus (Ad) carrying a foreign antigen and boosting with orthopoxvirus modified vaccinia Ankara (M) expressing the same antigen has consistently been shown to induce strong T cell responses capable of inducing high levels of efficacy against intracellular pathogens [56]–[58].
Adenovirus chimpanzee serotype 63 (Ad) and Modified Vaccinia Ankara (M) vaccines expressing a mammalian codon-optimised fragment of PbTRAP were generated. Referred to as Ad-M PbTRAP combination vaccine, they were used to vaccinate B6 mice with an 8-week resting period between priming and boosting (Figure 6A). Since the recombinant PbTRAP vaccines contained sequences in addition to PbTRAP130, we used both PbTRAP130 and a pool of overlapping peptides to PbTRAP (PbTRAPpool) in stimulation assays to verify if other PbTRAP-derived sequences were able to induce T cell responses. As shown in Figure 6B–D, the frequencies of IFN-γ secreting CD8+ T cells were indistinguishable between the two stimulations; approximately ∼17% (range: 14%–50%) of the total CD8+ T cells produce IFN-γ specific for PbTRAP130 or PbTRAPpool. Consistent with the induction of effector responses, these activated IFN-γ-producing cells coincided with the modulation of the corresponding expression markers CD62Llo and CD11ahi (Figure 6B). Polyfunctional analysis revealed that the responses were predominantly IFN-γ positive cells and IFN-γ/TNF double positive cells (Figures 6C,D). This intracellular cytokine pattern of PbTRAP130-specific CD8+ T cells were similar to that measured by immunizations with irradiated sporozoites (Figure S2). Cells stimulated with no peptide did not respond to either PbTRAP130 or PbTRAPpool. No cytokine-producing CD4+ T cells were detected following PbTRAP130 or PbTRAPpool stimulation. These results demonstrated the induction PbTRAP130-specific CD8+ T cells by vaccination and confirmed that PbTRAP130 is the only T cell epitope in PbTRAP in this infection model.
To determine protective efficacy, mice vaccinated with Ad-M PbTRAP were challenged with PbGFP-Luccon. This permitted us to perform in vivo imaging in order to quantify hepatic parasite development after challenge and to subsequently follow the development of patent parasitaemia in the same animal. Mice vaccinated with Ad-M PbTRAP show significant efficacy against sporozoite challenge as shown by a considerable decrease (∼95% reduction in liver parasite load; range: 91%–99%) in liver parasite load as compared to mice given Ad-M vector controls (Figures 7). However, Ad-M PbTRAP-vaccinated mice did not develop sterile immunity; rather, they developed patent parasitaemia on day 3 after challenge. To further assess vaccine efficacy, we performed survival analysis by measuring time to reach 1% parasitaemia. This measurement has been reported to reflect the number of merozoites that egress from the liver under the assumption that the vaccine has no efficacy against malaria blood stage parasites [44]. Consistent with lower liver parasite load, Ad-M PbTRAP-vaccinated mice showed significant delay in parasite growth as compared to controls (Figure S6).
Finally, to establish whether the observed decrease in liver parasite load in mice vaccinated with Ad-M PbTRAP is solely mediated by CD8+ T cells, groups of vaccinated mice were administered CD8+ depleting antibodies or control rat IgG prior to sporozoite challenge (Figure 8). In addition, a group of mice vaccinated with Ad-M PbTRAP were given anti-CD4+ T cell antibodies. Efficacy was abrogated in the group that received anti-CD8+ T cell, but not anti-CD4 T cell, antibodies after vaccination. Taken together, these results provide a striking correlation between the high levels of PbTRAP130-specific CD8+ T cells and the CD8+ T cell-mediated efficacy elicited by Ad-M PbTRAP vaccination.
Over 40 years have passed since the observation that immunisation of mice with a whole organism vaccine, i.e. Pb γ-Spz, induces complete protection against normal sporozoite challenge [3]. Although a crucial role for CD8+ T cells in this protection was demonstrated more than 25 years ago [10], [11], information regarding the precise targets of these cells remains remarkably incomplete. This paucity of naturally processed parasite-specific epitopes of CD8+ T cell responses has thwarted efforts to characterise the induction of protective and possibly non-protective responses that can be rigorously studied in murine vaccination models for malaria liver stages [59]. CSP, the major protein that coats the sporozoite's surface, has been at the forefront of malaria vaccination studies for more than two decades, and CSP-specific responses have been the benchmark in measuring cellular responses to malaria liver stages [20]. However, PbCSP and PyCSP are targets of immunodominant and protective CD8+ T cell responses only in the BALB/c model [16], [60]. The need to identify non-CSP targets of CD8+ T cell responses comes from accumulating evidence suggesting that protection against malaria infection can be obtained in the absence of T cell responses to CSP [18], [19].
In this study, we employed an unbiased approach for screening Pb-derived peptides that are recognised by CD8+ T cells from B6 mice immunised with whole sporozoite strategies, which are known to induce protection. Out of 600 peptides predicted to contain H-2b motifs, we report for the first time the identification of two peptides as signature targets of CD8+ T cells from sporozoite-immunised B6 mice. Notably, we provide evidence that responses to PbTRAP confers partial efficacy in vivo. Thus, we anticipate that pre-clinical development of an anti-malarial vaccine based on whole sporozoites or sub-unit vaccines that incorporate TRAP will be facilitated by the identification of quantifiable signatures of effective immunisation.
Plasmodium parasites are complex pathogens with a ∼23 Mb genome [32], [61]. The identification of targets of CD8+ T cells has important implications for understanding the hierarchy and the scope of responses to complex pathogens. When we initiated this study, we expected the CD8+ T cell responses to be relatively widely dispersed over many sporozoite and liver stage antigens that together can account for the protection offered by sporozoite immunisation. However, we only identified two targets - PbS20318 and PbTRAP130. Our inability to identify additional targets is indicative of immune evasion once sporozoites invade target cells, a hallmark that might generally apply to parasitic infections. Coincidentally, vaccines against these complex pathogens remain elusive. In marked contrast, a similar epitope prediction approach revealed broad and abundant H2b-restricted CD8+ T cell epitopes of vaccinia virus [22].
One potential limitation in our in silico analysis is the inclusion of only the top hits of potential liver stage peptides. In addition, by utilising IFN-γ as a read-out of our assays, we cannot formally exclude other cytokines or cellular markers that may serve as additional signatures of protection. The quantification of IFN-γ expression by CD8+ T cells is a reliable read-out in murine and human vaccine studies. Consistent with the observation that few antigens are recognised by CD8+ T cells from immunised mice, an assessment of 34 candidate non-CSP sporozoite antigens in the BALB/c model failed to reveal any additional epitopes [62]. In experimental studies in humans, antigenic analysis of genomic and proteomic data revealed 16 new antigens recognised by volunteers immunised with Pf γ-Spz [33]. However, there was no significant overall difference between protected and non-protected volunteers, indicating that T cell proliferation is associated with pathogen exposure rather than protection.
By measuring the responses to PbS20318 and PbTRAP130, we report the first characterisation of the development of CD8+ T cell responses following sporozoite immunisation in B6 mice. PbS20318- and PbTRAP130-specific CD8+ T cells persist to long-term memory while displaying effector and effector memory phenotypes in both the spleens and livers of immunised mice. The ability of PbS20318 and PbTRAP130-specific CD8+ T cells to produce IFN-γ after peptide stimulation, coinciding with activation phenotypes, suggested their ability to exert cytotoxic functions in vivo. Remarkably, only PbTRAP130-specific, but not PbS20318-specific, CD8+ T cells, are able to lyse target cells pulsed with the respective peptides. These results provide the first evidence that liver stage infection evokes antigen-specific CD8+ T cells that are both cytolytic and non-cytolytic. In support of this notion, tolerisation induction via high dose intravenous injection of peptides revealed that PbTRAP130-specific, but not PbS20318-specific, CD8+ T cells contribute to parasite killing.
Owing to the complexity of immune responses to sporozoites, it was not surprising that PbTRAP130-tolerised mice that received multiple immunisations with Pb γ-Spz were completely protected against sporozoite challenge. The residual efficacy obtained in the challenged PbTRAP130-tolerised mice is likely due to responses to unidentified targets of both CD8+ and CD4+ T cells. It is noteworthy that in the Py-B6 and Py-C57Bl/10 (both expressing H2b) models, both CD8+ and CD4+ T cells equally participate in the inhibition of parasite development [63]. The results are in agreement with earlier findings on the immunodominant but imperfect CSP epitope in the Balb/c infection model [18].
In our prime-boost vaccination experiments with Ad-M PbTRAP in B6 mice, we were able to generate very high levels (14–50%) of circulating PbTRAP130-specific CD8+ T cells, yet complete protection was not achieved. This level of antigen-specific CD8+ T cells is not achieved following multiple immunisations with Pb γ-Spz in B6 mice, yet complete protection is attained (Figure 2). It is likely that additional antigens help consolidate the protection afforded by whole sporozoite immunisation. Very high levels of PyCSP- or PbCSP-specific CD8+ T cells generated by various vaccination strategies in Balb/c mice have been shown to induce complete protection [64]–[69]. We anticipate that generating high levels of PbS20318-specific CD8+ T cells through a similar Ad-M vaccine protocol is unlikely to confer any quantifiable protection in B6 mice. Further studies are needed to characterise the complex mechanisms of protection. In a recent study, it was suggested that strain-specific background genes in nonhematopoietic cells can control the threshold of antigen-specific CD8 T cells necessary for protection [70].
Taken together, our results raise the intriguing and important question as to what factors govern the protective efficacy of responding antigen-specific CD8+ T cells. Numerous possibilities exist to explain these findings, including quantitative and functional differences in CD8+ T cells, distinct expression of cognate proteins and/or MHC class I presentation. Future work is warranted to identify the underlying mechanisms that distinguish cellular correlates of sporozoite exposure (PbS20318-specific CD8+ T cells), from signatures of protection (PbTRAP130-specific CD8+ T cells). Our work suggests that the mechanisms involved can now be studied in a tractable animal model.
The outcome of our work has obvious relevance for vaccine development. We provide a clearly defined model system, in which to investigate fundamental aspects of the CD8+ T cell response and to manipulate this response to enhance protective immunity. Our identification of TRAP as major target of protective CD8+ T cells in the Pb-B6 model lends strong support for PfTRAP as leading vaccine candidate to elicit strong cellular immune responses. Initial and partial results from ongoing phase III clinical trials of the RTS,S/AS01 malaria vaccine candidate, which is based on PfCSP, demonstrated only modest efficacy [71], [72]. RTS,S/AS01 immunisation elicits high concentrations of anti-CSP antibodies [73] and induces CD4+, but not CD8+, T cell responses [74]. Towards a second-generation, >80% effective malaria vaccine, rational vaccine design to elicit superior and lasting immune responses is critical.
TRAP has been identified as a viable target for Pf vaccines. While PyTRAP was suggested as a target of CD8+ T cells in BALB/c mice [66], [75], the epitopes have been elusive so far. PfTRAP induces large numbers of polyfunctional CD8+ T cells in experimental vaccination studies [76], and these T cells are associated with partial but significant efficacy in human vaccine trials [43], [45]. However, when tested in larger Phase IIb trials in endemic areas, efficacy was lost despite moderate immunogenicity of the vaccine, suggesting that the T cell response that is induced is insufficient in some way [46]. Further work is needed to improve the immunogenicity of TRAP-based vaccines for malaria-exposed individuals. More recently, vaccination regimes with viral vectors have induced much stronger CD8+ T cell responses in phase I trials [76] and greater efficacy in controlled challenge studies (Ewer et al., submitted for publication). Based on our identification of a major protective CD8+ T cell epitopes in PbTRAP, we propose that the parasite-host combination Pb-B6 is an attractive and relevant model to further systematically explore subunit vaccination strategies, focusing on the TRAP antigen with the aim of increasing potent and durable sterilising immunity.
Animal procedures were performed in accordance with the German ‘Tierschutzgesetz in der Fassung vom 18. Mai 2006 (BGBl. I S. 1207)’, which implements the directive 86/609/EEC from the European Union and the European Convention for the protection of vertebrate animals used for experimental and other scientific purposes. Animal protocols were approved by the ethics committee and the Berlin state authorities (LAGeSo Reg# G0469/09). Experiments performed at the University of Oxford were performed under license from the United Kingdom Home Office under the Animals (Scientific Procedures) Act 1986 and approved by the Animal Care and Ethical Review Committee. Female B6 mice 6–8 weeks of age were purchased from either Charles River (Sulzfeld, Germany) or Harlan (Derbyshire, UK).
The complete life cycles of two parasite strains, Pb (strain ANKA, clone cl15cy1) and PbGFP-Luccon (strain ANKA, clone 676m1cl1), were previously cloned and maintained by continuous cycling between rodent hosts and Anopheles stephensi mosquitoes [77]. Sporozoites were isolated from salivary glands. For Pb γ-Spz, the radiation dose was 1.2×104 cGy, and mice were immunised intravenously with 1.5×104 parasites. For Pb infection and treatment strategies, 1.5×104 sporozoites were injected intravenously while anti-malaria drugs were given intraperitoneally: 160 mg AZ/kg (days 0, 1 and 2) [8], 60 mg PQ/kg (days 0, 1 and 2) [7] and 40 mg CQ/kg (days 0–7) [6]. For PbHKSpz, parasites were subjected to 95°C for 15 minutes [38]. Immunised mice were challenged with 104 Pb sporozoites. In some experiments involving imaging, mice were challenged with 5×103 PbGFP-Luccon normal sporozoites. For mice receiving >1 immunisations, sporozoites were administered >7 days apart.
Peptide libraries were generated by solid phase synthesis (Peptides and Elephants, Potsdam). PbS20318 (VNYSFLYLF) and PbTRAP130 (SALLNVDNL) peptides were additionally synthesised at large scale. Peptides were synthesised as peptide amides and lyophilised peptides were resuspended in DMSO at a concentration of at least 1 mg/ml and stored at −80°C. Purity was confirmed by mass spectrometry. Peptides were diluted in PBS and tested individually at 10 µg/ml for the ELISpot assay and T cell stimulations with the exception of Figure 6 in which peptides were used at 5 µg/ml. A 20-mer peptide library (pooled) spanning the entire sequence of PbTRAP was generated, in which each peptide was overlapped 10 amino acids to another peptide (Mimotopes, Victoria, Australia). The final concentration of the peptide pool used for stimulations was 5 µg/ml for each peptide. For the tolerisation experiments, mice received 300 µg peptide (either PbS20318 or PbTRAP130) on day −7 and 100 mg on days −4 and −1 as described [51]. Mice were immunised on day 0 and received 100 µg peptide weekly until the end of the experiment.
AdCh63 and MVA vaccines expressing a mammalian codon-optimised fragment of PbTRAP [39] (GeneArt, Regensburg, Germany) were constructed and propagated based on previously published viral vectors [76], [78]. The TRAP signal peptide sequence was replaced with the human tissue plasminogen activator signal peptide sequence and the 3′ transmembrane region deleted. The viral vectors, referred to as Ad-M PbTRAP as a combination vaccine, were administered intramuscularly in endotoxin-free PBS at a concentration of 109 viral particles for Ad PbTRAP and 106 plaque-forming units for M PbTRAP.
ELiSpot assay was performed as described [35] with minor modifications. CD8+ T cells were purified from spleen cells by positive selection using mouse CD8 microbeads (MACS Miltenyi Biotec, Bergish Gladbach, Germany). Syngeneic spleen cells from naïve B6 mice were coated with peptides and were used as antigen-presenting cells. Anti-IFN-γ [AN18] and biotin-anti-IFN-γ [R4-6A2] were obtained from Mabtech (Nacka Strand, Sweden).
For T cell stimulations followed by ICS, splenic and liver-infiltrating lymphocytes were incubated with peptides for 5–6 hours in the presence of Brefeldin A (eBioscience, California, USA), followed by standard surface and intracellular staining procedures. Data was acquired using either a LSRII or a LSRFortessa (BD Bioscience, Heidelberg/Oxford). Antibodies for stainings were obtained from eBioscience: anti-mouse CD8 [53-6.7], CD62L [MEL14], CD44 [IM7], CD11a [M17/4], CD49d [R1-2], IFN-γ [R4-6A2], TNF [MP6-XT22] and interleukin (IL)-2 [JES6-5H4].
Data analysis was performed using FlowJo 7.6.3 (Tree Star Inc., Oregon, USA). SPICE 5.21, a gift from Dr. Mario Roederer (NIAID, NIH, Maryland, USA) was used to analyse polyfunctional data.
The cytotoxic potential of antigen-specific CD8+ T cells was assayed as described [47].
Livers were excised 42 hours after challenge and total RNA was isolated. cDNAs, generated by reverse transcription, were used as templates for quantitative real-time PCR [8] of Pb 18S rRNA (gi: 160641) and GAPDH sequences (gi: 281199965) using the Applied Biosystem Step One Plus Real Time PCR System (Darmstadt, Germany). Relative parasite loads were calculated using the ΔΔCt method.
For in vivo imaging, an IVIS200 imaging system (Caliper Life Sciences, Chesire, United Kingdom) was utilised to monitor parasite development in the liver 42 hours after challenge [78]. Mice were anaesthetised, injected subcutaneously with D-luciferin (Synchem Laborgemeinschaft OHG, Felsberg/Altenberg) (100 mg/kg in PBS), and 8 minutes later, imaged for 120 s at binning value of 8 and fields-of-view (FOV) of 12.8 cm. Bioluminescence in the liver was quantified using Living Imaging 4.2 software (Caliper Life Science) and expressed as total flux of photons per second of imaging time.
The development of patent parasitaemia was determined based on Giemsa-stained blood smears. Relationships between log percentage parasitemia and time after challenge were plotted. Kaplan Meier analysis was performed to compare the parasite growth rate, and protection was measured as a delay in reaching 1% parasitaemia [44].
Statistical analysis (see Figure Legends), unless otherwise specified, was performed using Prism 5.0c (GraphPad Software Inc., CA, USA). Mixture model calculations were performed using Stata 12 (StataCorp LP, TX, USA).
|
10.1371/journal.pntd.0001439 | Dengue Virus Type 4 Phylogenetics in Brazil 2011: Looking beyond the Veil | Dengue Fever and Dengue Hemorrhagic Fever are diseases affecting approximately 100 million people/year and are a major concern in developing countries. In the present study, the phylogenetic relationship of six strains of the first autochthonous cases of DENV-4 infection occurred in Sao Paulo State, Parana State and Rio Grande do Sul State, Brazil, 2011 were studied. Nucleotide sequences of the envelope gene were determined and compared with sequences representative of the genotypes I, II, III and Sylvatic for DEN4 retrieved from GenBank. We employed a Bayesian phylogenetic approach to reconstruct the phylogenetic relationships of Brazilian DENV-4 and we estimated evolutionary rates and dates of divergence for DENV-4 found in Brazil in 2011. All samples sequenced in this study were located in Genotype II. The studied strains are monophyletic and our data suggest that they have been evolving separately for at least 4 to 6 years. Our data suggest that the virus might have been present in the region for some time, without being noticed by Health Surveillance Services due to a low level of circulation and a higher prevalence of DENV-1 and DENV- 2.
| Dengue virus infections are a major concern in developing countries, affecting approximately 100 million people/year. The virus has four immunologically related serotypes (DENV-1, DENV-2, DENV-3 and DENV-4) associated with human disease. The virus is widespread in tropical and Sub-Tropical areas of Asia, Africa and Americas. The virus is transmitted by mosquito bites, and is primarily associated with Aedes aegypti as its main vector. To understand the reemergence of DENV-4 in Brazil in 2010–2011 we carried out a Bayesian phylogenetic analysis of the envelope gene sequences sampled in Brazil in 2011. Our results indicate that the studied samples are close related to strains circulating since 1981, when DENV-4 was first introduced in South America, but have gone trough recent evolution for at least 4 to 6 years. Our results also suggests that the virus may have penetrated Brazilian population earlier than 2010, indicating that the virus could have been present but not detected due a higher prevalence of DENV-1 and DENV- 2 and the failure of the surveillance system to locate the milder disease commonly associated with DENV-4.
| Dengue virus (DENV) is a single stranded RNA virus, with four immunologically related serotypes (DENV-1, DENV-2, DENV-3 and DENV-4) associated with Dengue Fever (DF) and Dengue Hemorrhagic Fever (DHF) [1].
The virus is widespread in tropical and Sub-Tropical areas of Asia, Africa and Americas. The virus is transmitted by mosquito bites, and is primarily associated with Aedes aegypti as its main vector [2].
The disease affects, approximately, 100 million people/year, causing 250,000 cases of DHF with a case fatality rate up to 15%, and is a major concern for Public Health authorities around the globe, primarily in developing countries [2].
Historically, the State of Sao Paulo, Brazil, has been suffering dengue outbreaks since 1990 when DENV-1 was introduced in the area. Subsequent epidemics were detected in 1997 and 2002, caused by DENV-2 and DENV-3, respectively, with increasing casuistic and detection of severe cases of DHF or Shock Syndrome [3]–[5].
DENV-4 had a brief circulation in Brazil in 1982 in the Northwestern region of Brazilian Amazon in a focal epidemic. No further cases of infection had been registered in the country until 2008, when the virus was detected in three patients, who had no international traveling history, in Manaus [6].
After this episode, the Brazilian Ministry of Health implemented the use of the NS1 ELISA test in 16 states in order to increase the percentage of viral isolates and the determination of the serotypes circulating in the country. Before the screening with the NS1 ELISA test, virus isolation was obtained in only 10% of samples submitted to isolation. With the screening of samples the percentage of detection of serotype rose to 82% [7]. The introduction of the NS1 ELISA assay as a tool for screening positive samples led to an important increase in the success of virus isolation. In São Paulo State, only 33.3% of the total of the samples inoculated in 2008 resulted in successful virus isolation, while in 2009 and 2010, 85.7% succeeded. The number of São Paulo state counties that sent samples for isolation also increased from 0.9% in 2008 to 10.2% in 2009 (Bisordi I, 2011, unpublished data).
DENV-4 reemerged in the country in 2010 in the municipalities of Boa Vista and Cantá in Roraima State [8]. The virus spread to different geographic regions of Brazil with cases of infection registered in the North (Roraima, Amazonas, Pará), Northeast (Bahia, Pernambuco, Piauí) and Southeast (Rio de Janeiro, Sao Paulo) [9].
Despite the importance of the virus distribution, little is known about its rate, pattern of spreading and evolution. Each serotype represents a cluster of different genetic lineages constantly evolving and changing within the population [10].
In the present work, six strains of the first autochthonous cases of DENV 4 infection occurred in Sao Paulo State and Rio Grande do Sul State, Brazil, in 2011 were studied using a Bayesian Phylogenetic approach. Nucleotide sequences of the envelope gene were determined and compared with the corresponding sequences of representative strains of the known DENV-4 genotypes. The main objectives of the present study are the identification of the genotypes of the newly introduced strains, the examination of the phylogenetic relationships between strains and the estimation of emergence time of DENV-4 strains.
The specimens analysed in this study were retrieved from a collection formed from materials received for diagnostic purposes in the Instituto Adolfo Lutz. The samples were sent by reference hospitals and the patients names are confidentially anonymized, and only reference numbers were used during the diagnostic procedures and in the analysis that originated this study.
All new DENV-4 strains characterized in this study were isolated directly from patient serum and detected by RT-PCR between February and March of 2011. The origin of the strains are detailed in Table 1.
Twenty microliters of the patients blood or serum were inoculated in tubes seeded with cultured cells of Aedes albopictus, clone C6/36. Indirect immunofluorescence assay (IFA) with polyclonal anti-flavivirus antibodies and anti-mouse immunoglobulin conjugated (fluorescein isothiocyanate – Sigma) were performed [11]. The positive samples were typed by IFA with monoclonal antibodies to DENV (Biomanguinhos).
Total RNA was extracted from the supernatant fluid of C6/36 infected cells using the commercial kit QIAamp® Viral RNA (Qiagen Inc., Ontario, CA), according to the manufacturer's instructions.
One step RT-PCR was performed employing the protocol described by Lanciotti et al, [12] in the presence of a set of primers targeting the complete envelope gene sequence, described by Lanciotti et al [13]. RT-PCR products were purified and directly sequenced using the Big Dye v.3.1 terminator chemistry. Sequences were determined using the Applied Biosystems 3130XL DNA sequencer.
All nucleotide sequences of the envelope gene for DENV-4 serotype generated for this study are deposited in GenBank under accession numbers JN092553 and JN848496–JN848500 (Table 1).
Sequences representative of the known genotypes I, II, III and Sylvatic for DEN4 were retrieved from GenBank and included in the phylogenetic analysis for comparison with the sequences generated in this study (Table 1). Sequence alignment was performed using the BioEdit software [14].
The Bayesian inference method available in the software BEAST v. 1.6.2 was used in order to analyze the phylogenetic relationship of the strains of this study [15]. The analysis of phylogenetic relationships and evolution, encompassed the entire Envelope gene, including six DENV-4 strains generated in this study and 107 sequences retrieved from GenBank (Table 1).
Each sequence of the corresponding data set was dated and maximum clade credibility (MCC) tree was generated. The internal nodes were inferred using a Markov Chain Monte Carlo (MCMC) Bayesian approach under a GTR model with Gamma-distributed rate variation (γ) and a proportion of invariable sites (I), using a relaxed (uncorrelated lognormal) molecular clock. Previously published data [10], [16] suggest that dengue evolution generally approximates a molecular clock with occurrence of minor differences in rate. Four independent MCMC runs of four chains each were run for 10 millions generations. Convergence of parameters during MCMC runs were assessed by their Effective Sample Size (ESS) reaching values above 150 as calculated with Tracer V 1.5 [15]. We used a Bayesian skyline coalescent prior to estimating population dynamics through time and access an estimative of evolutionary rate and the time of the most recent common ancestor (TMRCA) in the Envelope gene analysis.
A fragment of 1487 nucleotides representing the entire sequences encoding the envelope gene was determined from 6 strains of DENV-4 and further aligned with other 107 envelope gene sequences retrieved from GenBank.
The phylogenetic relationships among those strains were reconstructed by Bayesian analysis with a relaxed (uncorrelated lognormal) molecular clock model. The analysis generated a MCMC phylogenetic trees (Fig. 1). All samples sequenced in this study were located in Genotype II, and coupled with samples from the Caribbean region and northern South America (Fig. 1). In general, the group is strongly supported (posterior probability of 0.98) with Internal relations within the clade showing a lower support, most likely due the higher homology of the samples, which hinders the separation, but the isolated strains are monophyletic in origin, supported by a high posterior probability (0,99).
The isolated strains in this study are monophyletic and our data suggest that they have been evolving separately for at least 4 to 6 years. Nonetheless, they are quite similar and relatively unchanged in relation to the DENV-4 introduced originally in the Caribbean region and northern South America.
The relaxed molecular clock estimated after the analysis of the envelope gene encopassed a time of evolution for DENV-4 of 50–60 years and an average replacement rate of 2.0037×10−3 Subs/Site/Year, considering an Effective Sample Size of 334.79 calculated in Tracer 1.5. The replacement rate of the branch of the isolated strains is of 1.238×10−3 Subs/Site/Year, and the branch originated within 4 to 6 years probably diverging from virus circulating in Venezuela as the closest sister branch reunited Venezuelan strains supported by a posterior probability of 0,99.
All sequenced strains were encompassed in genotype II, with a high medium posterior probability (0.98), slightly lower in the terminal clades due to the genetic similarity of samples which hinders the separation. The isolated strains formed a strongly supported monophyletic branch (posterior probability of 0,99).
Not all Brazilian samples included in this study belonged to genotype II. The sequence AM 1619, from Manaus, 2008, retrieved from GenBank, grouped with genotype I. Our data also support the recent circulation of DENV-4, genotype I, reported in Manaus County in 2008 [6].
The studied period of evolution of DENV-4 after the analysis of the Envelope gene was estimated between 50–60 years, with an average replacement rate of 2.0037×10−3 Subs/Site/Year, considering an Effective Sample Size of 334.79. This estimative is supported by previously published data [10], [17]. Our result strongly suggests that the introduction of genotype II in South America occurred between 30–35 years ago, most probably through the Caribbean region or the northern South America. These results corroborate previously published data, since the first cases associated with DENV-4 from the American Continent are dated around 1982, in the Caribbean islands [10], [13], [18].
These data indicate that Dengue evolution approximates a molecular clock with minor rating variances. It is interesting to observe that raising ratings are mostly associated with increasing case occurrences or the emergence of the virus in a new region, meaning that the virus, when confronted with a susceptible population, undergoes an explosion of diversity.
These phenomena were previously reported concerning Dengue and other Flaviviruses [1], [19]–[21]. The clade directly associated with the studied strains showed a replacement rate of 1.238×10−3 Subs/Site/Year, slightly under the average rate. However, the rates observed within the clade formed by the isolated strains show higher replacement rates when compared with the sister branches (Figure 2). Such findings may indicate that the virus started to evolve more quickly, suggesting that it may have recently found a susceptible population and is spreading.
The DENV-4 samples, sequenced in this study, represent a recent emergence of a viral strain circulating in South America around 20 to 25 years ago. Results suggest that a local evolution has been taking place for about 4 to 6 years. These data could indicate that the virus might have been present in the region for some time, without being noticed by Health Surveillance Services due to a low level of circulation and a higher prevalence of DENV-1 and DENV- 2. It is possible that, since DENV-4 is associated with a milder disease [22], [23], the human cases may have been below the line of screening, going unnoticed. It is probable that the recent efforts to increase the success of virus isolation and serotyping allowed the study of a greater number of cases that otherwise would not have been serotyped, enabling the notification of less prevalent serotypes.
However, the hypothesis of a recent introduction cannot be ruled out, but it would imply in multiple recent introductions of the virus, in a very short period of time, in relatively distinct areas, or a single introduction event in a significantly important area that facilitated the virus introduction in new areas. The simultaneously occurrence of DENV-4 in different Brazilian States, forming a strongly supported clade, in the beginning of 2011, favors a recent emergence of the virus followed by a quickly introduction. However, such occurrence did not provide any clue to substantiate whether the virus was widespread but circulating in a low level, or circulating in a restricted area and subsequently taken to new localities with susceptible hosts.
The isolated strains are monophyletic in origin and the molecular clock supports a local evolution, but by no means it indicates where that evolution occurred. It may have occurred in northern Brazil, and the virus quickly were introduced in Southern region due the constant human traffic. As the closest branch in our phylogenetic analysis is formed by Venezuelan strains of DENV-4, a Venezuelan origin of Brazilian DENV-4 may be a plausible hypothesis.
Either way, the virus may have evolved in an imperceptible manner in an undisclosed place, it was not reported and later emerged subtly and spread fast among a susceptible population. The recent DENV-4 cases reported elsewhere may represent a cryptic circulation that was only recently detected. The analysis of more sequences from a broader geographical perspective, encompassing other Brazilian regions, is crucial in order to understand how the virus evolved and how it got widespread.
The reemergence of DENV-4 should be a concern for Health authorities since there are evidences that the replacement of a dominant circulating genotype is associated with the rising of a previously rare lineage. These phenomena were observed in Puerto Rico [24] and could be a plausible scenario in Brazil.
The authors indicate the necessity to study the phylodynamics of Dengue virus and the dynamics of genotypes and serotypes circulation and substitution in the population.
It is equally necessary to extend the efforts of virus isolation and sequencing towards the mosquito population. The mosquitoes are a reliable source of information on circulating virus, as mosquitoes do not depend on medical screening or the spontaneous search for medical services by the symptomatic patients.
Our results indicate the recent circulation of DENV-4 in São Paulo.
|
10.1371/journal.pcbi.1002700 | Multistationary and Oscillatory Modes of Free Radicals Generation by the Mitochondrial Respiratory Chain Revealed by a Bifurcation Analysis | The mitochondrial electron transport chain transforms energy satisfying cellular demand and generates reactive oxygen species (ROS) that act as metabolic signals or destructive factors. Therefore, knowledge of the possible modes and bifurcations of electron transport that affect ROS signaling provides insight into the interrelationship of mitochondrial respiration with cellular metabolism. Here, a bifurcation analysis of a sequence of the electron transport chain models of increasing complexity was used to analyze the contribution of individual components to the modes of respiratory chain behavior. Our algorithm constructed models as large systems of ordinary differential equations describing the time evolution of the distribution of redox states of the respiratory complexes. The most complete model of the respiratory chain and linked metabolic reactions predicted that condensed mitochondria produce more ROS at low succinate concentration and less ROS at high succinate levels than swelled mitochondria. This prediction was validated by measuring ROS production under various swelling conditions. A numerical bifurcation analysis revealed qualitatively different types of multistationary behavior and sustained oscillations in the parameter space near a region that was previously found to describe the behavior of isolated mitochondria. The oscillations in transmembrane potential and ROS generation, observed in living cells were reproduced in the model that includes interaction of respiratory complexes with the reactions of TCA cycle. Whereas multistationarity is an internal characteristic of the respiratory chain, the functional link of respiration with central metabolism creates oscillations, which can be understood as a means of auto-regulation of cell metabolism.
| The mitochondrial respiratory chain shows a variety of modes of behavior. In living cells, flashes of ROS production and oscillations accompanied by a decrease of transmembrane potential can be registered. The mechanisms of such complex behavior are difficult to rationalize without a mathematical formalization of mitochondrial respiration. Our most complete model of mitochondrial respiration accounts for the details of electron transport, reproducing the observed types of behavior, which includes the existence of multiple steady states and periodic oscillations. This most detailed model contains hundreds of differential equations, and such complexity makes it difficult to grasp the main determinants of its behavior. Therefore the full model was reduced to a simplified description of complex III only, and numerical bifurcation analysis was used to study its behavior. Then the evolution of its behavior was traced in a sequence of models with increasing complexity leading back to the full model. This analysis revealed the mechanism of switching between the modes of behavior and the conditions for persistence in a given state, which defines ATP production, ROS signaling and destructive effects. This is important for understanding the biochemical basics of many systemic diseases.
| The electron transport chain links the central carbohydrate energy metabolism with ATP synthesis (see Fig. 1). It transforms the free energy released by the oxidation of NADH and succinate into a form of transmembrane electrochemical potential (ΔΨ), which is used for ATP synthesis [1]. Reactive oxygen species (ROS) are byproducts of electron transport [2]. They play the roles of both metabolic signals and destructive agents [3]–[8]. These key roles of electron transport in cellular metabolism motivate the great interest in understanding the details of the dynamics of this process. Electron flow through the chain of carriers is controlled by levels of substrates (NADH, succinate), levels of tricarboxylic acid (TCA) cycle intermediates, the rate of ATP consumption, oxygen availability, etc [9]. However, many interesting dynamical properties of the electron transport and linked ROS production are determined by the intrinsic properties of the electron transport chain, such as the structural and functional links between carriers (topology of the system) and values of parameters, e.g. reaction rate constants. Such intrinsic properties can determine physiologically important modes of respiratory chain operation and how transitions between these modes occur. This relationship between intrinsic properties and dynamics can be understood by analyzing a detailed model of electron transport. We have reported elsewhere that the Q-cycle mechanism of electron transport in respiratory complex III exhibits bistability [10], i.e. two stable steady states may exist under the same microenvironmental conditions (corresponding to two stable steady state solutions of a system of ordinary differential equations (ODEs) at the same parameter values). The importance of bistability resides, in particular, in the fact that it can be a main determinant of the destructive effects of hypoxia/reoxygenation [3], [11]. Bistability also occurs in a model of the whole respiratory chain that agrees quantitatively with the measured forward and backwards fluxes through the respiratory chain [12].
Current experimental techniques make it possible to monitor the behavior of a single mitochondrion in living cells [13]. This has provided evidence that mitochondria can operate in numerous qualitatively distinct modes. They can persist in a steady state characterized by a high value of ΔΨ and a low rate of ROS production, can switch to another steady state characterized by a low value of ΔΨ and a high rate of ROS production, and can also enter into a mode of sustained oscillations [13]–[15]. The two qualitatively different steady states can be associated with the normal physiological state and a pathological one that can be approached after hypoxia/reoxygenation. The oscillatory behavior is probably very important for intracellular signaling, as it was found for Ca2+ signaling [16]. An application of a systematic method that reveals qualitatively different modes of model behavior and corresponding regions of relevant problem parameters, i.e. a bifurcation analysis of a mathematical model describing mitochondrial electron transport, would give insight into these physiologically important modes of mitochondrial functioning and the mechanism of switching between modes. The objective of the present study is to advance in this direction.
This paper presents a bifurcation analysis of four increasingly complicated models. The first two of these describe only complex III (Fig. 2), in forms simplified compared with those previously presented [10]. The last two include elements of the respiratory chain, shown in Fig. 1, that were previously modeled in [12]. For respiratory complex III, the models assume that the core of the complex contains four redox sites: cytochrome b with its two hemes, bH and bL, cytochrome c1, and the Rieske protein containing iron-sulfur center (bH-bL-c1-FeS). In addition, the core can bind a two-electron carrier ubiquinone either in the matrix (Qi-Qi-) or cytosolic (-Qo-Qo) side of the inner mitochondrial membrane. Symbols are repeated to represent the two electrons. This gives four different configurations of complex III: (bH-bL-c1-FeS, bH-bL-c1-FeS-Qo-Qo, Qi-Qi-bH-bL-c1-FeS, Qi-Qi-bH-bL-c1-FeS-Qo-Qo). The models take into account that respiratory complexes constituting the respiratory chain consist of a number of fixed in space electron carriers that can be either reduced (red) or oxidized (ox). A possible state of a complex is defined by a combination of redox states of individual carriers constituting it. The model variables correspond to these states. The repeated symbols each correspond to three possible states: either with two valence electrons and two corresponding protons (ubiquinol), one valence electron (semiquinone), or no valence electrons (ubiquinol). Thus the configuration Qi-Qi-bH-bL-c1-FeS-Qo-Qo has 144 possible states, each corresponding to a variable in the model.
These large numbers of variables make it difficult even to write down explicitly the corresponding systems of ordinary differential equations (ODEs). Therefore an algorithm is designed and implemented to automatically compute the values of the right hand sides of the ODE system at each step of a numerical solution of a corresponding initial value problem (IVP), based on the rules formulated in accordance with the reaction mechanism [10]. Two complimentary techniques are combined for numerical bifurcation analysis to systematically search for various types of model behavior and for intervals of parameter values corresponding to multiple steady state solutions or oscillatory solutions: 1) using IVP solvers to solve IVPs for our ODE systems, and 2) using the numerical bifurcation analysis software CL-MATCONTL [17], [18] (Text S1) to study the corresponding large equilibrium systems.
A numerical bifurcation analysis of the respiratory chain was first conducted for a model of complex III simplified to 145 equations (referred to further as model 145) as described in Methods, with basic set of parameters listed in Table 1. This model accounts for only one configuration of complex III, namely the core (consisting of cytochrome b with its two hemes bH and bL, cytochrome c1, and the Rieske iron-sulfur (FeS) center) together with ubiquinone molecules bound at both inner (i) and outer (o) sites. This configuration is denoted as Qi-Qi-bH-bL-c1-FeS-Qo-Qo. Binding/dissociation of quinones in this model is accounted for by the replacement of reduced bound molecules at Qi by oxidized bound molecules and oxidized bound molecules at Qo by reduced bound molecules (see Methods). Numerical continuation of steady state solutions, as a function of succinate concentration, (proportional to VmSDH, eq.(1)), performed with CL_MATCONTL (as described in Methods), revealed an interval of parameter values with multiple steady state solutions (Fig. 3, orange curves) enclosed between two limit points (LP) indicating the points of fold bifurcations. This interval of the parameter values for which multiple steady state solutions exist, corresponds to two curve segments of stable steady states and one curve segment of unstable steady states in between.
The method of numerical bifurcation analysis that we used allowed us to accurately and rigorously determine the bifurcation behavior of the whole system, without any simplifications. At the same time, the main process underlying the bifurcation behavior can, in some cases, be heuristically identified by reducing a system by taking into account different time scales. Such a reduction of Model 145 (see Text S2) points to binding/dissociation of ubiquinone coupled with its reduction/oxidation as the main process responsible for the multistationarity of complex III (Fig. S1 in Text S2).
A gradual increase of succinate from low concentrations towards the interval of multistationarity leads to the lowest branch of steady states for quinol (QH2) as shown in Fig. 3A. Under such conditions complex III functions to give high electron flow and high ÄØ (upper branches in Figs. 3B and 3C). It should be noted that, although the concentrations of QH2 that constitute the lower branch of steady states in Fig. 3A are low, they nevertheless are sufficient to maintain high levels of ΔΨ. The inset in Fig. 3C shows that ΔΨ drops to 0 as the concentration of succinate decreases to 0. This drop is a consequence of QH2 deficiency. A decrease of succinate concentration from 15 to 0 corresponds to the almost linear decrease of QH2 levels from 0.02 to 0 nmol/mg prot. The above described “active” state, that provides highest electron flow, is characterized by low levels of semiquinone (SQ) at the quinol oxidase site (Qo) (Fig. 3D).
If succinate concentration surpasses the right limit point on Fig. 3A, the rate of ubiquinone reduction by succinate dehydrogenase outstrips the maximal capacity of its oxidation by complex III, therefore Q is almost completely converted into QH2, as Fig. 3A shows. This is the biochemical mechanism of the bifurcation. The lack of an electron acceptor at Qi site results in a decrease of electron flow through the Q-cycle, a decrease of ΔΨ, and an increase in levels of SQ at Qo. If then succinate concentration decreases, the blocked electron transport cannot produce a sufficient amount of acceptors Q to activate electron flow. When the system is in such a blocked state, even decreased succinate dehydrogenase activity is sufficient to maintain low levels of Q and keep the system blocked. If initially the electron carriers are oxidized, the solution of an initial value problem for the ODE system approaches an “active” steady state, and if initially the carriers are reduced, the solution approaches a “blocked” steady state.
Model 145 has 13 parameters. The number of parameters is much less than the number of equations because parameters characterize the types of electron transport reaction, which is a much smaller number than the number of combinations of redox states of carriers (the number of equations). Different states participate in reactions of the same type with the same parameters, but they cannot be combined (the system cannot be reduced without additional simplifications) because the whole set of reactions for each state (variable) is different.
Fig. 3 shows curve segments of multiple steady state solutions corresponding to an interval of values of one parameter. The shape and size of these curve segments depends on the values of other parameters. The width of this interval may be smaller, or the interval may even disappear. Fig. 3 shows an interval of relative succinate concentrations corresponding to multiple steady states, obtained using an algorithm (described in Methods) that scans all the parameters with the objective of finding as large interval as possible.
Including in our model all of the four configurations of complex III and an explicit description of quinone binding/dissociation increases the number of equations to 257 (see Methods). This more detailed model also has a region of multiple steady state solutions. Application of the same algorithm maximizing the region of relative succinate concentration characterized by multiple steady states resulted in the interval that is larger than the one in the case of model 145, as is shown for ΔΨ in Fig. 4. However, the qualitative behavior of the two models remains similar. Evidently, model 145 faithfully accounts for the main properties of complex III determined by the Q-cycle mechanism.
Model 267 is obtained by adding to model 257 equations that account for reactions taking place in complex I (described in Methods). This extended model contains almost all of the essential components of the respiratory chain model that we used for the analysis of experimental data [12]. Using it enables us to start the numerical bifurcation analysis for a “real” set of parameter values (Table 2) that reproduces the measured dynamics of NADH reduction in the presence and absence of rotenone, and the maximal and state 4 respiration rates, when mitochondria are fueled by succinate or pyruvate/malate [12]. Numerical continuation of steady state solutions for ΔΨ as a function of succinate concentration uncovers the existence of an interval of parameter values with multiple solutions, in the form of an S-shaped curve, enclosed between two limit points (Fig. 5). There is also a Hopf bifurcation point in the vicinity of one limit point. The sustained oscillation, which can be simulated in the parameter interval between the left limit point and Hopf bifurcation, has very small amplitude (inset in Fig. 5). The mechanism of the fold bifurcations in this case is similar to that discussed for the simplest model. Similar to the case presented in Fig. 3, the stable steady states with the lowest values of ÄØ have the highest levels of SQ radicals at the Qo site; this may be a physical basis for high ROS production rates.
The measurements that were used to find the given set of parameters were performed in a suspension of isolated mitochondria. The rate of electron transport from cytochrome c1 to cytochrome c is the parameter most affected by the procedure of isolation, since it depends on the structure of intermembrane space, which is significantly changed after the isolation. In intact mitochondria the value of this parameter is expected to be higher than in isolated mitochondria. Increasing its value by less than an order of magnitude increases the interval of multiple steady state solutions to infinity (Fig. 6). Starting from an initially oxidized state, the system approaches a steady state, which, with numerical continuation with the substrate concentration as a parameter, results in the upper branch in Fig. 6. This branch does not contain any bifurcation points. However, at a high substrate concentration, and starting from a reduced state, the system approaches another steady state located on a different branch marked blue in Fig. 6. The lower segment of this branch is stable. A decrease of the substrate supply parameter ultimately leads to a limit point, and the upper segment of unstable steady states starts at this point. Note, this branch of unstable steady states is not connected to the upper branch of stable steady states. Similar to the cases analyzed above, the steady states corresponding to low ΔΨ values (lowest blue branch) are characterized by practically complete reduction of the free ubiquinone pool and maximal levels of free radicals SQ at the Qo site. Similarly to the cases analyzed above, the steady states corresponding to high ΔΨ (yellow branch) are accompanied by an oxidized free ubiquinone pool and low levels of SQ at the Qo site.
The change of a single parameter that characterizes interaction between cytochromes c1 and c (kc1c) gives a qualitatively different behavior of model 267, as seen in Fig. 6 from the comparison between the blue curve and the orange curve, which is redrawn from Fig. 5. Such a difference in behavior can be induced by swelling/shrinking of mitochondria, as was mentioned above, but also hypoxia/reoxygenation can induce a similar change of parameters and, thus, similar effects. Hypoxia in our model can be simulated as a kc1c decrease.
Assume that before hypoxia the system effectively functions at some point on the yellow curve (Fig. 6). Suppose the change of kc1c induced by hypoxia transforms the properties of the system so that the orange curve becomes the continuum of its steady states. If before hypoxia the functional steady state was to the right of the rightmost limit point in orange curve, then after hypoxia the system evolves until it reaches a steady state in the lower segment of orange curve (coinciding with the lower segment of blue curve). If then the system is re-oxygenated, and the blue and yellow curves again become the continuum of steady states, it remains in the same state now located in the low segment of blue curve. Thus, hypoxia and re-oxygenation change the state of the system. Before hypoxia it generated high ΔΨ (yellow curve in Fig. 6) and slowly produces ROS, whereas after re-oxygenation it stays in a state characterized by low ΔΨ (blue curve with points in Fig. 6) and rapidly produces ROS.
Further extension of the ODE model to 272 equations, as described in Methods, by including the reactions of the TCA cycle with the parameters listed in Table 3, allowed us to study the interaction of the respiratory chain with central carbohydrate metabolism. In the extension, pyruvate is accounted for as a substrate for the TCA cycle, which provides succinate for complex II and NADH for complex I. Using parameter values verified by fitting the measured dynamics of NADH and respiration rates [12], this model predicts the existence of wide range of pyruvate concentrations with two stable steady state segments (Fig. 7), as well as those described above with respect to succinate.
Similarly to the cases considered above, the redox state of free ubiquinone pool determines the dynamics of the system. If the free ubiquinone pool (Fig. 7A) is not completely reduced, the electron flow through the respiratory chain (Fig. 7B) and ÄØ (Fig. 7C) is high, and the level of SQ at the Qo site (Fig. 7D), determining the ROS production rate, is low. On the other hand, if ubiquinone is practically completely reduced, the electron flow through the respiratory chain and ÄØ is low, and the level of SQ at the Qo site is high. However, the bifurcations which determine a switch between the two steady state branches are different from those considered above. Specifically, when the system is in an oxidized state, the increase of the pyruvate concentration leads to a Hopf bifurcation. Oscillations in a neighborhood of this bifurcation point have insignificantly small amplitude (see Fig. 8A). An increase of the control parameter makes the amplitude greater (inset in Fig. 8B), but the trajectory comes to the zone of attraction of the lower segment of stable steady states (Fig. 7C) and approaches one of these states (Fig. 8B). As the pyruvate concentration decreases, the system stays in this “reduced” stable curve segment until it reaches the limit point at ∼0.003 mM, where a curve segment of unstable steady states starts.
An increase of the cytochrome c1 to cytochrome c electron transition rate (kc1c) to 782 s−1 and an increase in VmSDH (equation (5)) from 171 (Table 3) to 1714 nmol/s/mg changes the bifurcation behavior as is shown in Fig. 9A. Although the bifurcation diagram here is also a basic S-shaped curve producing multi-stationarity, the entire segment of steady states between the two Hopf bifurcation points is unstable. Stable oscillations of high amplitude appeared between these Hopf bifurcation points (Fig. 9B and 9C). ÄØ oscillates between 160 and 20 mV; such changes can be measured and, probably, this mechanism underlies the observed behavior [19]. As Figs. 9B and 9C show, ÄØ and the level of SQ at the Qo site (defining the rate of ROS production) oscillate in counter-phase. This also corresponds to the behavior monitored in intact mitochondria [14]. Variation of parameters can significantly change the durations of phases of low or high potential (high or low ROS production rate, respectively). This could be a basis of the ROS signaling [20].
The mechanism of oscillations arises from an interaction of the ubiquinone reduction/oxidation with the TCA cycle. The switch from the “oxidized” curve segment of steady states to the “reduced” one is accompanied by a decrease of the electron flow, and, as a consequence, an increase of the NADH levels (decrease of NAD+). Since the conversion of pyruvate in the TCA cycle requires NAD+, the production of succinate slows down. The high levels of NADH are maintained for some time due to reverse electron transport through complex I reducing NAD+. A decrease of the substrate production in the TCA cycle and reverse electron transport result in an accumulation of a sufficient amount of the electron acceptor ubiquinone that activates electron transport, which results in switching back to the curve segment of the “oxidized” steady states, and then the cycle repeats again.
The change from multi-stationarity shown in Fig. 7 to an oscillatory behavior shown in Fig. 9 is, in part, the result of a change of the rate constant kc1c, which accounts for interaction between cytochromes c1 and c. This rate constant depends on the volume of the intermembrane space, where the interaction takes place. The intermembrane volume can be controlled experimentally in a suspension of isolated mitochondria, and the correspondence of the model predictions and the measured ROS production rates under variations of the intermembrane space can be experimentally verified. The change of the ROS production rate (characterized by the level of SQ at the Qo site), predicted for an “oxidized” state of model 272 with an increase of succinate concentration, is shown in Fig. 10A. The shape of the lower curve segment of steady state concentrations of SQ bound at the Qo site depends on parameter kc1c. Fig. 10A shows a superposition of curve segments of stable steady states obtained at two different values of kc1c. At a low value of this parameter (∼260 s−1 as shown in Table 1), increasing the succinate concentration above 1 mM takes the system past the Hopf bifurcation point (similar to that shown in Fig. 7D), and it switches to the upper curve segment of SQ stable steady state concentrations (blue curve). Increasing kc1c shifts this Hopf bifurcation point to infinity, so that the upper branch of stable steady states, although it still exists, becomes inaccessible from the lower branch in the space of succinate concentrations (orange curve).
In Fig. 10A, the lower branch of steady states obtained at a higher value of kc1c crosses the lower branch obtained at a lower value of kc1c. At low succinate concentrations, the levels of SQ at the Qo site are higher when kc1c is high. At high succinate concentrations, this relationship is reversed. We have confirmed this experimentally, registering the rate of the ROS production as a measure of the SQ concentration at the Qo site. It is expected that kc1c decreases if the intermembrane space increases, thus diluting the concentration of cytochrome c that is included implicitly in this parameter. The light scattering technique allows measuring changes in the volume of the mitochondrial matrix and, implicitly, the intermembrane space. Light scattering is higher in KCl than in sucrose of the same osmolarity (Fig. 10B). This indicates that the mitochondrial matrix is more compact in KCl than in a sucrose solution. The outer membrane is permeable for both solutes, hence the total mitochondrial volume, which it restricts, must be the same. Therefore the intermembrane space, estimated as the difference between total and matrix volumes, is greater in KCl media. Thus, mitochondria incubated in KCl are characterized by lower values of kc1c than those incubated in sucrose. The experimental results shown in Fig. 10C are consistent with the model. Indeed, in the media with sucrose, the rates of ROS production driven by low succinate concentrations (100–500 uM) are higher than those in KCl supplemented media. In the range of succinate concentrations ≥500 uM, the situation was reversed: ROS production in KCl-based media exceeded that observed in sucrose-based media.
Bifurcation behavior, as revealed by the numerical bifurcation analysis of complex III models, is inherent in the Q-cycle mechanism of electron transport. The main process underlying fold bifurcations in the considered models of complex III was found to be reduction/oxidation and coupled binding/dissociation of ubiquinone in accordance with the Q-cycle mechanism (Text S2). We show here that a decrease of ÄØ accompanied by an increase of ROS production rate can take place as a consequence of perturbations in the respiratory chain operation. The most critical element in such bifurcation behavior is that the same metabolite is reduced at the Qi site and oxidized at the Qo site.
The interaction of complex III with complex I increases the width of the maximal interval of multiple solutions, see Fig. 5, or may even make it infinite, as shown in Fig. 6. The width of this interval is sensitive to the parameter (kc1c) that characterizes the combined processes of electron transport from cytochrome c1 to c and further, ultimately reducing molecular oxygen. Thus, it can represent the availability of oxygen and, in this case, the comparison of Figs. 5 and 6, given in Results, demonstrates how hypoxia and reoxygenation may perturb the system to a state of a very high ROS production.
Further extending the model by including into it the reactions of the TCA cycle preserves the interval of parameters where multiple steady state solutions exist. In particular, there are two stable steady states at the parameter values defined in [12] by fitting measured experimental data, as shown in Fig. 7. In experiments performed previously [10], we confirmed that isolated mitochondria incubated with high succinate concentration can persist in one of two different steady states. Mitochondria can be switched from a high ROS production state to a low one by transient incubation with ADP, and then back to a high ROS production state by transient hypoxia. Another experimental confirmation of the predicted behavior of the electron transport chain is the consistency between the predicted curves of steady state levels of SQ at Qo attained at various concentrations of succinate for two different swelling conditions and the measured curves of ROS production rate (Fig. 10).
Stable oscillations that can be obtained at the parameter values in a neighborhood of the Hopf bifurcation point have insignificantly small amplitude, and the region of their stability appears to be so small that it is practically undetectable. However, an increase of the values of the two parameters, which characterize succinate dehydrogenase activity and the rate of combined reactions upstream from cytochrome c1, results in the appearance of an interval of succinate concentrations where high-amplitude oscillations exist and are stable (Fig. 9). Feedback interaction of the multistationary respiratory chain operation with the TCA cycle creates oscillations. NADH, as a common metabolite, provides such feedback (see “Mechanism of oscillations” in Result section).
The parameters shown in Tables 1–3 that were used for model 272 were determined from the best fit to the data from experiments performed in vitro in isolated mitochondria. One can expect that the volume of the intermembrane space increased after the procedure of isolation. Such a change of the intermembrane space dilutes cytochrome c, and thus decreases the rate of interaction of cytochromes c1 and c. Natural spatial variability of succinate dehydrogenase activity contributes to an uncertainty in its estimated value. Our results show that the change in these parameters, which can be expected in mitochondria of living cells, compared to the isolated ones, results in an oscillatory mode of operation.
In fact, in mitochondria of living cells, flashes and oscillations of ROS production accompanied by the counter-phase changes of ΔΨ can be measured either as a response to laser excitation [13], [14], or as a spontaneous mitochondrial activity [15], [19]–[25]. Usually, the measured in vivo decrease of ÄØ that accompanies the ROS flashes was ascribed to either a ROS-induced mitochondrial permeability transition (MPT) [13], [14] or a ROS-activated inner membrane anion channel [26]. Our study opens a new direction in the investigation of the MPT that is of great importance for understanding intracellular signaling and regulation. In particular, it can help to solve the question: why is respiration inhibited during the MPT, when ÄØ is low and cytochrome c still remains in the intermembrane space? Our hypothesis is that the MPT is secondary with respect to the change in the steady state of respiration; it happens when the electron transport chain switches to the “reduced” steady state, where respiration is inhibited by the mechanism considered above. There is more evidence supporting this hypothesis. Matrix pH is well known to be important for the MPT and models considering it as a main factor governing opening/closure of the MPT pore describe the observed events of the MPT [27]. However, the link between the change of matrix pH and the MPT was described phenomenologically; the concrete mechanism of the pH effect on the MPT remains elusive. The models presented here points to the mechanisms by which pH can affect electron transport: protons are explicitly involved in reduction/oxidation of ubiquinone, which is the main process defining the bifurcation. Alkalinization of the matrix must slow down SQ reduction at Qi site and, thus, block electron transport and facilitate the switch to the “reduced” state. If the change to a reduced steady state of the electron transport chain induces the MPT, this provides a concrete mechanism of pH-induced MPT, though it requires further investigation.
Moreover, in some cases, ROS sparks and a decrease of ΔΨ may be a direct consequence of the functional organization of the electron transport, and may not require the involvement of other mechanisms. Thus, many experimentally observed effects, such as excessive ROS production after hypoxia/reoxygenation, or oscillations of ROS production and of ΔΨ, can be explained as a consequence of intrinsic properties of the respiratory chain and its interaction with the central metabolic pathways.
These qualitatively different modes of behavior are manifestations of the same mechanism of electron transport, determined by its quantitative characteristics. Understanding the qualitatively different types of behavior requires a quantitative analysis of electron transport and the linked reactions of the central metabolism. The method presented here can be used for such an objective. However, the simplifications of reality used in our models should be taken into account. In particular, they represent complex III as a monomer, whereas it is known that the functional form is dimeric in living cells [28]. The functional link of monomers at the level of cyt bL was analyzed based on the edge-to-edge distance between cyt bL hemes of the two monomers [29]. It was found that the intermonomer interaction can affect the rate of electron transport, especially in the energized states and when the bH heme is reduced because of a lack of electron acceptors at the Qi site. Using our method to model the dimer would require solving ODE systems containing roughly the square of the number of equations that we analyze here. Such systems can be constructed, but solving them would create computational problems. Performing a preliminary analysis of a simplified model of the bc1 dimer containing cyt bL and bH, and Qo binding sites, we found that, despite intermonomer interactions, which quantitatively affect the kinetic behavior of complex III, qualitatively, the dimer demonstrates the same types of bifurcation behavior as the monomer in the situations analyzed in [29].
Another limitation of our models concerns the values for the parameters. The basic set of parameters shown in Tables 1–3 originally was taken from [30] and was then modified by fitting experimental data [12]. In principle, the rate constants can be determined based on the distances of electron tunneling [31], [32]. However, the resolution of 2.1 Å in the determination of distances [33] and uncertainties in other parameters necessary for such determination (as indicated e.g. in [29]) can result in great variation in the values obtained for the rate constants. These uncertainties can be greater than an order of magnitude. The rate constants that we used are inside the range of possible variations admitted by the estimation based on the known distances between the electron carriers.
It should be noted also that the TCA cycle is introduced in the model in a very schematic manner, however, keeping the stoichiometry of respiratory substrates, i.e. succinate and NADH production from pyruvate. Most of the reactions are lumped together and specific reaction mechanisms are not considered. Many of the enzymes performing consecutive reactions form multienzyme complexes [34], [35] (not considered here), where local concentrations of the metabolites can be different from their average concentrations in the matrix. Therefore we did not require average metabolite concentrations at equilibrium to be consistent with respective equilibrium constants. Simulation of a more realistic mechanism of the TCA cycle reactions might affect the bifurcation behavior of the whole system, since, as is indicated above, the appearance and location of Hopf bifurcations and related oscillations is a probable consequence of the interaction of the TCA cycle with reduction/oxidation of ubiquinone. This warrants further investigation.
As is shown by our analysis, perturbations in metabolite concentrations or oxygen availability may induce critical changes in the modes of mitochondrial behavior that result in huge changes in ATP synthesis and ROS production. Changes in the energy supply or signaling or damaging events can have crucial consequences on cell operation in general. We have presented a general overview of the possible modes. At the same time, this approach opens a way to study effects of specific disease conditions on mitochondrial functioning, and to predict mitochondria related disease progressions. In particular, the primary disorder in the chronic obstructive pulmonary disease (COPD) is a decreased capacity of an organism to take up oxygen. The limits of oxygen uptake are measured clinically, and such limitations can be simulated by changing the value of the respective mitochondrial parameter, provided that the other model parameters are determined for the given tissue. In this way the role of the mitochondrial component in a disease progression can be elucidated. A similar approach can be developed, for instance, for diabetes, which results in an essential change in a substrate supply and composition. The effects of such a substrate change on the mitochondrial state and consequences for the whole cell functioning can be predicted. In this way the approach developed here opens a way to better understand progressions of many systemic diseases.
All procedures involving animals were approved by Children's Hospital of Pittsburgh and were in compliance with “Principles of Laboratory Animal Care” and the current laws of the United States.
A detailed model of the respiratory chain is described elsewhere [12]. The general algorithm for constructing the ordinary differential equations (ODEs) for the model accounts for all the possible redox states of the respiratory complex III [10] interacting in accordance with the well accepted Q-cycle theory. It assumes that the core of the complex contains four redox sites: cytochrome b with its two hemes, bH and bL, cytochrome c1, and the iron-sulfur containing Rieske protein (bH-bL-c1-FeS). Each of these redox sites can carry one valence electron. The core can bind the two-electron carrier ubiquinone either on the matrix (Qi) or cytosolic (Qo) side of the inner mitochondrial membrane (bH-bL-c1-FeS-Qo-Qo, Qi-Qi-bH-bL-c1-FeS, Qi-Qi-bH-bL-c1-FeS-Qo-Qo), giving four different configurations of the complex. The model describes binding/dissociation of ubiquinone/ubiquinol that results in interconversion of these four configurations. Bound electron carriers, as well as core redox centers occupy fixed binding sites and have fixed interactions within the respiratory complex. Therefore the probability that a complex is found with a given combination of reduced/oxidized states, including the states of the four redox sites and the states of the substrates, is considered as a variable of the model.
The oxidized state of each redox site is coded as a binary “0” and the reduced state as a binary “1”. In this way various combinations of reduced and oxidized states of carriers can be represented as a four-digit binary numbers with values from 0 to 15 representing redox states of the core (bH-bL-c1-FeS), six-digit binary numbers with values from 0 to 63 representing the redox states of each of two configurations containing one ubiquinone, (Qi-Qi-bH-bL-c1-FeS and bH-bL-c1-FeS-Qo-Qo) and eight-digit binary numbers with values ranging from 0 to 255 representing the redox states of the configurations containing two ubiquinones (Qi-Qi-bH-bL-c1-FeS-Qo-Qo). The algorithm constructs an ODE system for all the configurations and their redox states. This system accounts for the transitions of electrons between carriers resulting in oxidation of the donor (1→0) and reduction of the acceptor (0→1), and binding/dissociation of ubiquinone/ubiquinol. These reactions are simulated in accordance with the well accepted Q-cycle theory and are described in detail in [10].
All the models analyzed here consider the following electron transitions performed by complex III:
With one exception (described below) all the models analyzed here consider the following reactions of binding/dissociation of ubiquinone/ubiquinol to/from complex III, (eq(12–16) in [10]),:
The rate constants of the reactions listed above, which were used as a base set of values for the analysis presented in the figures, are shown in Table 1.
The numerical continuation algorithm requires that the Jacobian matrix for the model differential equations is nonsingular and its rows are linearly independent. Earlier versions of our model [10], [12] used two binary digits to model the state of ubiquinone, giving 4 combinations (00, 01, 10, and 11), although there are only 3 physically distinct states. In fact, 01 and 10 represent the same state (semiquinone with one valence electron). The algorithm was originally designed so that only the state “01” can be produced, and the amount of configurations containing “10” as a state of semiquinone always was zero. The presence of equations describing such zero-concentration states did not change the result of numerical integration of the initial value problem. However, it did make the Jacobian matrix for the system singular as it contained linearly dependent rows. To perform a bifurcation analysis using CL_MATCONTL, such subsidiary equations had to be eliminated. We modified the algorithm for constructing the ODEs so that it does not include the equations for the states containing “10” semiquinone.
The model simulates electron flow from succinate that reduces ubiquinone. Succinate concentration implicitly defines the maximal rate of ubiquinone reduction:(1)here VmSDH = k·S, where S is succinate concentration, k is a constant. Q is ubiquinone concentration. This reaction is accounted for as a term in the differential equation for ubiquinone, which also participates in other electron transport reactions of the respiratory chain. The structurally fixed reduction and oxidation of ubiquinone in complex III respectively on the matrix and cytosolic sides results in the translocation of protons and generation of a transmembrane electric potential.
The conservation of the total amount of complex III and the total amount of ubiquinone is taken into account. The model constructed in this way contains 257 equations (255 equations for the redox states of complex III, and one each for ubiquinone (oxidized), and transmembrane potential).
This and all other models considered here account for a proton leak through the membrane that is exponentially dependent on ΔΨ:(2)where F = 96500 c/mol is the Faraday constant, R = 8.3 J/(mol×K) is the gas constant, T = 298 K is temperature. Outside (Ho = 0.0001 mM) and matrix (Hi = 0.00005 mM) proton concentrations are considered to be fixed due to high buffer capacity, but ΔΨ is a variable whose dynamics are described by a differential equation that takes into account proton translocations coupled with electron transport [10] and a proton leak described by (2).
In accordance with the Q-cycle mechanism, ubiquinol bound in the Qo site is oxidized giving its electrons to the FeS center of the Rieske protein and cytochrome bH and releasing its protons into the intermembrane space. Then the ubiquinone formed is released. The next pair of electrons can be transported only after the next ubiquinol is bound to the same Qo site. For simplicity, release of ubiquinone and binding of new ubiquinol can be combined and described as an exchange of ubiquinone with ubiquinol at the Qo site. Similarly, the release of ubiquinol and binding of ubiquinone at Qi site can be combined. In this way, all the reactions of the Q-cycle can be described considering only one configuration of complex III that contains two bound ubiquinones (Qi-Qi-bH-bL-c1-FeS-Qo-Qo). After removing “zero-states” containing “10” semiquinone (as described above) and taking into account the conservation of the total contents of complex III and ubiquinone, the above model is reduced to 145 equations.
The model of 257 equations simulates the whole set of reactions of Q-cycle. The reduction of the number of equations from 400, as explained above, does not change the biological model, but only simplifies its mathematical representation. Therefore we use this simplified set of equations as part of an extended mathematical description of electron transport in the respiratory chain. In addition to the reactions performed in complex III as described above, this extended model with 267 equations accounts for electron transport reactions performed by complex I, as described elsewhere [12]:
The rate constants of the reactions listed above, which were used as a base set of values for the analysis presented in the figures, are shown in Table 2.
Substrate supply in model 267 is treated in the same way as in model 257, with the exception that succinate oxidation depends on NAD+. This accounts for NAD+-dependent reactions of succinate production in TCA cycle, lumped in the model with SDH and results in NADH production, which is oxidized by complex I:(3)NAD+ is a variable of the model and its dynamics are described by a following differential equation that is included in the ODE system describing the dynamics of various redox states of complexes I and III. It accounts for the rate of NAD+ production as a result of NADH oxidation by complex I, and the rate of NAD+ consumption in the TCA cycle reactions reducing it into NADH:(4)Mass conservation NAD++NADH = const is taken into account (with const = 16 nmol/mg prot).
In this model, consisting of 272 equations, the production of the substrates of respiration, succinate and NADH, is considered in more detail, although still in extremely simplified form. The expressions for some of the reaction rates lump several reactions together and account phenomenologically for the substrates rather than real reaction mechanisms.
Succinate dehydrogenase now accounts for succinate (suc), since it is a real variable of the given model:(5)The fumarate oxidation and malate dehydrogenase (MDH) reactions forming oxaloacetate (oa) assume that fumarate and malate are represented as a single pool (mal):(6)The citrate synthase reaction assumes that pyruvate and acetyl CoA are combined in a single pool (pyr):(7)Transport of pyruvate assumes a constant cytosolic concentration (Cpyr) and a variable mitochondrial concentrations (pyr):(8)A number of TCA cycle reactions from citrate (cit) to succinate are combined. The whole set depends on citrate as input substrate and NAD+:(9)Succinate exchange to fumarate/malate assumes constant external concentrations (Csuc, Cmal):(10)Succinate entry when it is added externally is modeled by:(11)Malic enzyme transforms malate into pyruvate:(12)The parameter values for the reactions listed above are shown in Table 3.
Dynamics of the new variables (suc, mal, oa, pyr, cit) are described by following ODEs incorporated in the model:(13)(14)(15)(16)(17)The dynamics of NAD+ are now described in a more complex way compared to Eq. (4)(18)The TCA cycle reactions from citrate to succinate reduce two molecules of NAD+ for each succinate produced, but the model also accounts for a molecule of NADH produced by transformation of pyruvate into acetyl CoA, which is not included explicitly.
A numerical approach used to systematically search for various types of model behavior was implemented by combining two complimentary techniques: 1) using initial value problem (IVP) solvers to solve IVPs for our ODE systems, and 2) using the numerical bifurcation analysis software CL-MATCONTL to study the corresponding large equilibrium systems.
The first method consists in numerical integration of two IVPs using the same parameter values but different initial states. One initial state is oxidized and one is reduced. Integration is continued until an approximate steady state is reached. A trajectory is considered to have reached an approximate steady state if the time derivatives of all variables were less than 1.0e-12 nmol·(mg prot)−1·s−1. The numerical solution is obtained using the DASSL method [36], as implemented in the NAG Fortran library (http://www.nag.co.uk/numeric/fl/fldescription.asp). This Fortran code is incorporated within our C++ program. If the steady states reached from the two different initial states are different, this indicates that the given parameters correspond to multiple stable steady state solutions. The algorithm implemented in our software automatically scans parameter regions to find such points of multiple steady states. It detects multiple steady states even if they are located in disconnected branches (as in Fig. 6), but it can detect only stable equilibrium solutions. Parameter values corresponding to multiple steady state solutions are found using this method, and the interval of the chosen parameter is maximized (optionally) as described below. Within the interval thus obtained, we apply the second method to find steady state solution(stable and unstable) and exactly locate and characterize bifurcation points.
The second method consists in numerical continuation and bifurcation analysis of equilibrium solutions to our ODE systems by CL-MATCONTL (Text S1) [17], [18], a MATLAB package for bifurcation analysis of large equilibrium systems. These equilibrium systems are viewed as systems of nonlinear algebraic parameter dependent equations. To compute their solution branches one uses pseudo arch length (numerical) continuation, which is a technique to compute a sequence of consecutive points approximating the desired solution branch using Newton type methods. This allows an accurate computation of both stable and unstable equilibrium solutions on a branch. CL-MATCONTL requires an initial point on the solution branch to start continuation and can compute only a connected solution branch. It can miss a whole branch of steady states, if the complete solution consists of more than one unconnected branches (as in Fig. 6). At each point on the solution branch the Jacobian of the system is computed, and a bifurcation is detected and located when an eigenvalue of the Jacobian crosses the imaginary axis, See Text S1 for more details. Combining both the IVP and continuation methods allows one to find unconnected branches of steady states and various types of steady states and bifurcation points.
|
10.1371/journal.pntd.0000881 | High Affinity Human Antibody Fragments to Dengue Virus Non-Structural Protein 3 | The enzyme activities catalysed by flavivirus non-structural protein 3 (NS3) are essential for virus replication. They are distributed between the N-terminal protease domain in the first one-third and the C-terminal ATPase/helicase and nucleoside 5′ triphosphatase domain which forms the remainder of the 618-aa long protein.
In this study, dengue full-length NS3 protein with residues 49 to 66 of NS2B covalently attached via a flexible linker, was used as bait in biopanning with a naïve human Fab phage-display library. Using a range of truncated constructs spanning the NS2B cofactor region and the full-length NS3, 10 unique Fab were identified and characterized. Of these, monoclonal Fab 3F8 was shown to bind residues 526 through 531 within subdomain III of the helicase domain. The antibody inhibits the ATPase and helicase activites of NS3 in biochemical assays and reduces DENV replication in HEK293 cells that were previously transfected with Fab 3F8 compared with mock transfected cells.
Antibodies such as 3F8 are valuable tools for studying the molecular mechanisms of flaviviral replication and for the monospecific detection of replicating dengue virus in vivo.
| Dengue virus is the most prevalent mosquito transmitted infectious disease in humans and is responsible for febrile disease such as dengue fever, dengue hemorrhagic fever and dengue shock syndrome. Dengue non-structural protein 3 (NS3) is an essential, multifunctional, viral enzyme with two distinct domains; a protease domain required for processing of the viral polyprotein, and a helicase domain required for replication of the viral genome. In this study ten unique human antibody fragments (Fab) that specifically bind dengue NS3 were isolated from a diverse library of Fab clones using phage display technology. The binding site of one of these antibodies, Fab 3F8, has been precisely mapped to the third α-helix within subdomain III of the helicase domain (amino acids 526–531). The antibody inhibits the helicase activity of NS3 in biochemical assays and reduces DENV replication in human embryonic kidney cells. The antibody is a valuable tool for studying dengue replication mechanisms.
| Dengue virus belongs to the Flaviviridae family and is the etiological agent of dengue fever, dengue hemorrhagic fever and dengue shock syndrome. It is the most prevalent arthropod transmitted infectious disease in humans and has four antigenically distinct viral serotypes (DENV 1–4) [1]. The genome of dengue viruses comprises a positive single stranded RNA of 11kb. Post-translational processing of the polyprotein gives rise to three strucural proteins (C, prM and E) and seven non-structural proteins (NS1, NS2A, NS2B, NS3, NS4A, NS4B and NS5). The processing of the amino terminal region of the polyprotein is carried out by host signal peptidases, while processing of the 2A-2B, 2B-3, 3-4A and 4B-5 sites is catalysed by the two-component viral protease NS2B/NS3 [2], [3].
DENV NS3 is a multifunctional enzyme with three known catalytic activities segregated into two distinct domains (Figure 1). The serine protease lies within the N-terminal 180 amino acid residues of the 618 amino acid protein. The central hydrophillic portion of the intergral membrane protein NS2B (residues 49–96) is required for protease activity [4]–[6]. The ATPase/helicase and nucleoside 5′-triphosphate activities are localised in the remaining C-terminal domain. There appears to be cross-talk between the two domains; the helicase activity is approximately 30-fold higher in the full-length NS3 protein than in the domain and the affinity of the full-length protein for ATP is 10 fold lower than that of the helicase domain alone [7], [8]. Recent crystal structures of full-length NS3 from DENV and the related flavivirus Murray Valley encephalitis virus, reveal that the protease and helicase domains are linked by an interdomain linker (residues 169–179 in DENV) as illustrated in Figure 1 [8], [9].
Infection with one DENV serotype results in immunity to that serotype only; this protection is thought to be due to neutralizing antibodies, DENV-specific memory T cells, or a combination of the two. While the T-cell response is directed against several DENV proteins, NS3 appears to be the dominant target for CD4+ and CD8+ T cells, and multiple human T cell epitopes have been mapped onto NS3 (reviewed in [10]). Interestingly DENV NS3 also elicits a specific antibody response in humans. A study of acute sera from patients infected with DENV-2 or DENV-4 showed that although anti-E (envelope) antibodies were the most abundant, anti-NS3 antibodies were widely detected, particularly in those with secondary infections [11].
Given the vital role NS3 plays in viral replication, and the specific T- and B-cell responses observed towards NS3 in DENV infected patients, well characterised anti-NS3 antibodies would be vaulable tools for studying viral replication in detail and detecting DENV infected cells. There are very few reports describing the production of monoclonal antibodies specific for NS3, and those that have used hybridoma technology in mice [12]–[15]. In three of these studies anti-NS3 antibodies were isolated following immunization with recombinant NS3 [12]–[14] but the fourth study innoculated mice with DENV-1 virus (purified from suckling mouse brain) and selected hybridomas that produced anti-NS3 antibodies [15]. These antibodies were then used to immunize mice that were subsequently challenged with DENV-1. Intriguingly an increase in survival, albeit equivocal, was noted with four of the monoclonal antibodies tested.
Recombinant antibody technology (phage display) is a powerful alternative to conventional antibody techniques that permits the selection of high-affinity antibodies specific for the target protein [16], [17]. Antibody fragments are expressed on the surface of filamentous phage linking the antibody protein with its encoding DNA sequence within the phage. Phage displaying antibodies that bind the target protein are enriched by several rounds of selection and amplification (bio-panning), and the resulting antibody fragments can be produced recombinantly in E. coli without the need for immunization. This study describes the identification and characterisation of human Fab antibody fragments that bind DENV NS3 using phage display. The panel of antibodies have different specificity patterns towards the NS3 protease and helicase domains, and NS3 proteins from DENV 1–4. We have evaluated the ability of the antibodies to inhibit the protease, helicase and ATPase activities catalysed by NS3 in vitro and DENV replication using cell based assays, and have identified one Fab, designated 3F8, that recognises a conserved epitope on subdomain III of the NS3 helicase domain. This antibody is cross-reactive with all four serotypes, and binds NS3 with high affinity. It can be used as a tool to study the DENV replication complex or could potentially be developed as a therapeutic.
A schematic of the constructs used in this study is shown in Figure 1A. The pET32b expression contructs for DENV4 NS2B18NS3 full-length (NS3 protein residues 1-618 linked with residues 49–66 of NS2B via a Gly4-Ser-Gly4 linker) and DENV4 NS3 helicase (residues 172–618) have been described previously [8], [18]. Expression constructs for DENV1-3 pET32b NS2B18NS3 full-length proteins were kind gifts from the Novartis Intitute for Tropical Diseases, Singapore, and comprise the same corresponding residues as the DENV4 full length construct.
The DENV4 protease domain constructs were amplified from the DENV4 NS2B18NS3 full-length construct. For DENV4 NS2B18NS3pro169 (NS3 protein residues 1–169 linked with residues 49–66 of NS2B via a Gly4-Ser-Gly4 linker) the forward primer 5′-CCACGCGGTTCTCATATGGCAGACTTGTCACTA-3′ and the reverse primer 5′-TTCATAATCTGGATCCCCAATTCATTCAGCTTGCGT-3′ were used. The underlined sequence corresponds with the NdeI and BamHI sites, respectively. DENV4 NS2B18NS3pro185 (NS3 protein residues 1–185 linked with residues 49–66 of NS2B via a Gly4-Ser-Gly4 linker) was amplified using the same forward primer as above and the reverse primer 5′-GTAAGTCCATTATGGATCCTCTTTACTTTCGAAAAATG-3′. The PCR fragments were digested with NdeI and BamHI and cloned into pET14b (Novagen). Escherichia coli BL21 Codon Plus (DE3)-RIL cells (Stratagene) transformed with pET32b or pET14b expression constructs were grown in ZYM5052 autoinduction medium [19] for 4 hours at 37°C followed by 20 hours at 18°C.
The PCR fragment for the DENV4 GST-NS2B47 expression construct (encoding residues 49–96 of NS2B fused at the N-terminus with Glutathione S-transferase) was generated with the forward primer 5′-GTGGTGGATCCGCAGATCTGTCACTAGAG-3′ and reverse primer 5′-CAGTGAATTCAAAAGTCATATCATATTGGTTTCCTC-3′ (BamHI and EcoRI sites are underlined) and the previously constructed DENV4 pET15b CF47-NS3 protease vector [6] as template. The PCR product was digested with BamHI and EcoRI and cloned into pGEX-4T-1 (GE Healthcare). E. coli BL21 Codon Plus (DE3)-RIL cells transformed with the DENV4 GST-NS2B47 construct were grown in in ZYM5052 autoinduction medium at 37°C for 4 hours followed by 16°C for 20 hours.
The DENV1-4 NS2B18NS3 full-length and DENV4 NS3hel proteins were purified by immobilised metal-ion affinity chromatography (IMAC) and size exclusion chromatography (SEC) as described previously [8], [18]. SEC was performed in 20 mM Tris pH 7.5, 150 mM NaCl, 3 mM β-mercaptoethanol and 5% glycerol. The pET14b DENV4 NS3 protease proteins were purified using the same protocol except the SEC buffer was 20 mM Hepes pH 7.5, 250 mM NaCl, 3 mM β-mercaptoethanol and 5% glycerol.
DENV4 GST-NS2BCF47 cell pellets were lysed in 20 mM Tris pH 7.5, 200 mM NaCl and the clarified lysate was incubated with Glutathione Sepharose 4B (GE Healthcare) for 2 hours at 4°C. Beads were washed extensively in lysis buffer and the protein was eluted in a buffer containing 20–50 mM reduced glutathione followed by dialysis into 20 mM Tris pH 7.5, 200 mM NaCl and 1 mM β-mercaptoethanol for storage.
Purified DENV4 NS2B18NS3 full length was dialysed into phosphate buffered saline (PBS) pH 7.5 prior to biotinylation. The protein was incubated with a 20-fold molar excess of the biotin reagent on ice for 2 hours according to the manufacturers instructions (Thermo Fisher Scientific). The reaction was stopped with 100 mM glycine and excess biotin was removed by SEC in PBS pH 7.5.
Library screening was performed with a naïve human fab phage display library HX02 (Humanyx Pte Ltd, Singapore) displayed in a modified pCES1 vector [20]. The amber stop codon prior to bacteriophage gene III in pCES1 has been removed and replaced with a SalI site. An additional SalI site has been placed at the C-terminus of gene III such that following SalI digestion and religation (and the concurrent formation of a TAA stop codon) soluble Fab can be expressed in both suppressor and non-suppresor strains of E. coli.
Library panning was essentially performed as decribed previously [21] but streptavidin megnetic beads (Invitrogen) were used to immobilise the antigen (biotinylated DENV4 NS2B18NS3). The concentration of DENV4 NS2B18NS3 was 200 nM in the first round and reduced to 40 nM and 10 nM in rounds two and three, respectively. The number of input phage in each round was constant at 1×1012 pfu while washing was increased from six times with PBS-T (0.1% Tween-20) in round one to 14 times with PBS-T in rounds two and three. Bound phage were eluted with 100 mM triethylamine and used to infect E. coli TG1 cells. Phage were resuced with M13K07 helper phage and amplified on 2xTY (tryptone-yeast) agar plates supplemented with 100 µg/mL ampicillin and 25 µg/mL kanamycin. Plates were scraped with Tris-buffered saline (TBS) and phage was concentrated from the supernatant by polyethylene glycol-NaCl precipitation.
Following the third round of selection individual TG1 clones were rescued with M13K07 and screened by enzyme-linked immunosorbent assay (ELISA) for reactivity against DENV4 NS2B18NS3 full-length (coated at 5 µg/mL in PBS pH 7.5). An anti-M13-horse radish peroxidase (HRP) conjugate (GE Healthcare) was used for detection and clones with an absorbance value two times higher than background levels were considered positive. To assess clone uniqueness a BstN1 restriction digest was performed following PCR amplification of the Fab coding region of the phagemid. Clones with unique DNA fingerprints were subject to automated sequencing.
Phagemids from unique Fab-phage clones were digested with SalI to remove the gene III coding sequence and re-ligated with T4 DNA ligase. The resulting plasmids were transformed into E. coli Top10 F' cells (Invitrogen) for expression and periplasmic extraction. Cell pellets were resuspended in chilled lysis buffer (120 mM Tris pH 8.0, 0.3 mM EDTA and 300 mM surose) and incubated on ice for 30 minutes for periplasmic extraction. Magnesium chloride (2.5 mM) was added to the clarified extract prior to IMAC purification. Fab were further purified by SEC (S200 10/300 column) if required.
For ELISA Maxisorb Immunoplates (Nunc) were coated with the relevant NS3 antigen (0.25 µM) in PBS pH 7.5 and blocked with 5% skim milk in PBS-T. Blocked wells were incubated with purified Fab (100 nM unless otherwise stated) at room temperature for 1 hour. Plates were washed with PBS-T and incubated with an anti-c-myc HRP conjugate (Roche) for detection.
Binding affinities of the Fab for DENV4 NS2B18NS3 were determined by surface plasmon resonance (SPR) using a Biacore 3000 instrument (GE Healthcare). All experiments were conducted at 25°C in HBS-EP (10 mM Hepes, 150 mM NaCl, 3.4 mM EDTA, 0.0005% P-20, pH 7.4). Biotinylated full-length NS3 protein was captured on a streptavidin (SA) sensorchip at a flow rate of 10 µl/min. For screening, the 10 Fab (100 nM) were injected across the flowcells, in replicates, at 10 µl/min for 1 min and allowed to dissociate for 2.5 min. Regeneration of the surface was achieved by a 30 second pulse with 15 mM HCl. The Fab that showed the best apparent KD in screening were selected for kinetic analysis. Kinetic parameters were measured by varying the molar concentration of each Fab (3.9–500 nM) and injecting these across the flowcells, in duplicates, with the same conditions used in the screening. Raw sensorgram data were aligned, solvent-corrected and double-referenced using the Scrubber II software (BioLogic Software, Campbell, Australia). A simple 1∶1 model, with or without the mass transport coefficient, was used for global kinetic analysis as appropriate.
NTPase assays were conducted as previously decribed [7]. DENV4 NS2B18NS3 full-length (4.8 nM) was preincubated with 1 µM Fab for 30 minutes at room temperature in 90 µl of reaction buffer (50 mM Tris pH 7.4, 2 mM MgCl2, 1.5 mM DTT, 0.05% Tween 20, 0.25 ng/µl bovine serum albumin). Poly U (1 µg, average length 200–250 bases) was added and a further 5 minute incubation at 37°C was performed before initiating the reaction with 10 µl of ATP. The reaction was carried out at 37°C for 30 minutes after which the malachite green reagent (Sigma) was added and absorbance (630 nm) was measured. The amount of phosphate released was determined with a standard curve and all assays were carried out in triplicate.
Protease activity was determined for NS2B47NS3pro185 (NS3 protein residues 1–169 linked with residues 49–96 of NS2B via a Gly4-Ser-Gly4 linker) based on protocols published by Li et al. [6] as detailed in the supporting information (Figure S1 in Supporting Information S1).
Helicase activity assays were performed as published [7], [22]. The substrate was prepared by annealing an 18-mer DNA oligo (5′-GCCTCGCTGCCGTCGCCA-3′) with a 38-mer RNA oligo (5′-UGGCGACGGCAGCGAGGCUUUUUUUUUUUUUUUUUUUU-3′). The 5′ end of the DNA was labelled using T4 polynucleotide kinase and [γ−32P]ATP. Reactions (10 µl) contained 50 mM Tris-HCl pH 7.4 supplemented with 5 nM of the DNA:RNA duplex, 500 nM DENV4 NS2B18NS3 full-length, 1.75 µM Fab (NS3:Fab ratio 1∶3.5), 4 units of RNAsin, 2 mM MgCl2, 1 mM DTT, 0.5% Tween and 0.25 µg/mL BSA. Fab and NS3 were preincubated in assay buffer for 10 minutes prior to initiating the reaction with 5 mM ATP. Assays were performed at 37°C for 30 minutes and were resolved on a 10% native polyacrylamide gel and autoradiographed using a Pharos FX system. Signal intensity was quantified with Quantity One software (Biorad).
Statistical analysis of all assay data was performed using paired t-tests. The results were considered statistically significant if p<0.05.
HEK293 cells were maintained at 37°C in a CO2 incubator in Dulbeco's modified eagle's medium (DMEM) containing 10% fetal calf serum and 1% penicillin-streptomycin. DENV2 (Eden 3295) was propagated in C6/36 cells prior to infection. For Fab transfection, 5×104 cells per well were transfected with 1 µg Fab (3F8, or the non-NS3 binding control Fab 3F6) using the TurboFect protein transfection reagent (Fermentas) and control cells were mock-transfected with TurboFect according to the manufactures instructions. Cells were infected four hours post transfection with DENV2 (Eden 3295) at an MOI of 1.0 in fresh media. For immunofluorescence cells were fixed 48 hours post infection using methanol, and incubated with 4G2 mouse monoclonal antibody for two hours at room temperature followed by a goat-anti-mouse secondary antibody conjugated with Alexa-488. Coverslips were mounted using ProLong Gold antifade reagent with DAPI (Invitrogen). Cells were visualized by fluorescence microscope using the 20X objective.
For plaque assay, media supernatants were collected 48 hours post-infection and virus titers (plaque forming unit per ml, PFU/ml) were determined by a plaque assay on BHK-21 cells as previously described [23]. Western bots were performed using the 4G2 mouse monoclonal antibody and an anti-His-tag antibody. Anti-GAPDH antibody was used as a loading control.
The Ph.D-12 random dodecapeptide library was purchased from New England Biolabs. Panning was performed as described in the New England Biolab Instruction Manual. Purified 3F8 (240 pmol) was mixed with 1×1011 pfu phage for 20 minutes at room temperature. Phage that bound 3F8 were isolated using anti-c-myc resin (Thermoscientific). Resin was washed 10 times and bound phage were eluted with 200 mM Glycine-HCl pH 2.2. The amplified eluate was enriched by two further rounds of selection. To minimize target unrelated peptides, phage that bound 3F8 were isolated using magnetic Ni-NTA agarose beads (Qiagen) in the second round. In the third round phage were pre-incubated with anti-c-myc resin and Ni-NTA magnetic beads in a ‘negative’ selection step prior to incubation with 3F8. Individual phage clones were purified following the third round of biopanning and tested for reactivity in an ELISA with an anti-M13-HRP conjugate. Single stranded DNA was isolated from positive clones using an iodine buffer (10 mM Tris pH 8.0, 1 mM EDTA, 4 M NaI) and sequenced using the M13 (-96gIII) primer provided in the library kit.
The epitope identified by peptide phage display was verified by competition ELISA using an array of overlapping 15-mer peptides purchased from Mimotopes that span subdomain III of the NS3 helicase domain (DENV-2 strain 16681). 3F8 at a concentration of 0.6 nM was preincubated with 3 nM of peptide for 30 minutes at room temprature before being transferred to an immunoplate previously coated with DENV2 NS2B18NS3 (5 µg/mL in PBS pH 7.5). For control, full-length DENV2 NS2B18NS3 was used as a competing reagent. Bound Fab was detected with an anti-c-myc-HRP conjugate and all measurements were repeated in duplicate and the mean value taken.
The NS3 full-length proteins (DENV 1–4) and the DENV4 NS3 domain constructs were purified at yields of between 5–10 mg/litre of culture (Figure 1A). DENV4 NS2B18NS3 was biotinylated using a 20 fold excess of the biotin reagent, bound onto streptavidin magnetic beads, and used to screen for binding against the naïve human Fab phage library. After three rounds of selection 480 TG1 clones were screened for their ability to bind DENV4 NS2B18NS3 by ELISA. A BstN1 digest of the 150 positive clones enabled grouping of clones with similar DNA fingerprints. Sequencing confirmed the identification of 10 unique clones. Sequence analysis with IMGT/V-QUEST showed that all the variable heavy chain (VH) sequences belong to the VH3 or VH4 gene families while the variable light (VL) sequences selected are derived from a larger number of gene families (Vκ2, Vκ3, Vλ1, Vλ2, Vλ6) (Table S1 in Supporting Information S1). The heavy chain complementary determining region 3 (CDR3) has been shown to have the most influence over antibody binding specificity [24]. CDR3 heavy chain sequences of the anti-NS3 clones selected are diverse in length and composition.
The phagemids of the 10 unique clones were digested with SalI to remove the gene III sequence and enable expression of the soluble Fab in E. coli Top 10F' cells. The expressed Fab have a hexa-histidine and c-myc tag at the C-terminus of the CH domain and were purified from the periplasm of Top10 F' E. coli cells by IMAC. Binding of the Fab to the antigen used in panning (DENV4 NS2B18NS3 full-length) and the DENV4 NS3 domain constructs was confirmed in an ELISA incorporating 100 nM Fab and an anti-c-myc-HRP conjugate (Figure 2A). Three Fab clearly bind the helicase domain of NS3 (3F4, 7 and 8), while a further two Fab (3F3 and 16) also appeared to bind the helicase domain, although the signal at 100 nM was low. An ELISA using 1000 nM Fab confirmed the helicase specificity of 3F3 and 16 (data not shown). Two Fab (3F10 and 11) gave positive signals with DENV4 NS2B18NS3pro169, DENV4 NS2B18NS3pro185, and DENV4 GST-NS2B47 indicating they bind to the 18 residues of NS2B (49 through 66) required for maintaining stability of the protease domain as these are the only residues present in all three domain constructs. 3F9 binds both DENV4 NS2B18NS3pro169 and DENV4 NS2B18NS3pro185 but has a higher signal with the construct ending at residue 185. This suggests the epitope for 3F9 spans the protease domain (up to residue 169) and residues 170 through 185. X-ray crystallography data shows that residues 169–179 form the 10-residue linker located between the protease and helicase domain in the DENV4 NS3 full-length structure and 3F9 may bind this linker region (Figure 1B). The two remaining Fab (3F12 and 14) bind only the the full-length NS3 protein. The epitopes for these Fab maybe structural epitopes located at the interface of the protease and helicase domains, that are only present in the full-length protein.
An ELISA performed with NS2B18NS3 full-length from all four DENV serotypes and 100 nM Fab (Figure 2B) showed that three of the helicase specific Fab (3F4, 7 and 8) are cross-reactive with all the serotypes tested. 3F10 binds DENV-2, DENV-3 and DENV-4, while 3F3 and 3F9 bind DENV-2 and DENV-4. Three Fab (3F11, 12 and 14) were DENV-4 specific. The signals for 3F16 are low but indicate it cross-reacts with DENV-3 and DENV-4.
An ELISA using serial dilutions of the purified Fab showed that all Fab gave concentration-dependent binding curves with 3F4, 7, 8, 10 and 11 recognising NS3 at concentrations of less than 100 nM (Figure 3A). 3F8 in particular has a high ELISA signal at 15 nM which suggests a high binding affinity. 3F3, 9, 12, 14 and 16 appear to bind to NS3 with lower affinity.
To further probe the affinity of the Fab for NS3, kinetic rates and affinity constants were measured in real-time using SPR. Biotinylated DENV4 NS2B18NS3 was immobilised on a SA sensor chip and, for initial screening, all 10 Fab were injected at a concentration of 100 nM. Binding responses were observed for five of the Fab at 100 nM (3F4, 7, 8, 10 and 11), while the remaining Fab showed no binding at 100 nM (3F3, 9, 12, 14, 16). The five Fab that showed a binding response also gave the highest signals in the affinity ELISA against NS3 (Figure 3A). Kinetic studies were performed on the binding Fab over a concentration range of 3.9–500 nM (Figure 3B). The Fab ranged 70-fold in their affinity for DENV4 NS3, with the highest affinity observed for 3F8 (KD 10.5 nM). This was followed by 3F7 (94.9 nM), 3F11 (95.6 nM) and 3F4 (98 nM). 3F10 had the lowest affinity of those measured with a KD of 670 nM (Table 1). This contrasts with the ELISA binding curve where 3F10 reaches maximum signal at a lower concentration than 3F11. An ELISA measures endpoint binding while SPR is in real time. The differences in ka (on rate) and kd (off rate) of the Fab may explain this discrepancy.
DENV NS3 utilizes ATP to drive the unwinding of the RNA duplex during replication. To determine the inhibitory characteristics of the Fab, DENV4 NS2B18NS3 full-length was pre-incubated with 1 µM of each Fab for 30 minutes at room temperature and ATP hydrolysis was monitored in a colorimetric assay [7]. Of the 10 Fab tested, 3F8 was the only antibody that significantly (p<0.05) reduced the amount of phosphate released (60.6±1.7 µM) compared with NS3 alone (105.6±7.6 µM) as seen in Figure 4A.
Protease activity was examined with a DENV2 NS2B47NS3pro185 construct which contains the 47 amino acids from NS2B sufficient for activating the protease domain. 3F10 inhibits DENV2 protease activity in a dose-dependent manner with activity significantly reduced to 49.5% of control at 1.2 µM (Figure S1 in Supporting Information S1). The epitope for 3F10 is contained within the 18 residues of NS2B that form an N-terminal β-strand distal from the NS3 protease active site (Figure 2A) suggesting allosteric inhibition of the NS2B-NS3 protease.
Helicase assays were performed with a 32P-labelled DNA:RNA duplex prepared by annealing an 18-mer DNA oligonucleotide with a 32-mer RNA oligonucleotide. As seen in Figure 4B, with an ATP concentration of 5 mM and a 100-fold molar excess of NS3 relative to nucleotide duplex a significant level of unwinding was observed, in keeping with results published previously [7], [22]. The effect of 3F4, 7 and 8 (the three cross-reactive, helicase specific Fab) on unwinding was assessed at a NS3:Fab molar ratio of 1∶3.5. Both 3F4 and 3F8 significantly (p<0.05) reduced NS3 helicase activity (60 and 53% of control, respectively), while 3F7 had no effect on unwinding.
The helicase specific Fab 3F8 was chosen for further characterisation. This Fab has superior affinity to the other helicase specific Fab, and inhibits both the ATPase and unwinding activities of NS3. The cross-reactive ELISA in Figure 2b demonstrates 3F8 reacts with NS3 from DENV1–4. To determine whether the affinity of the cross-reactions is comparable between serotypes an ELISA using serial dilutions of 3F8 was performed (Figure S2 in Supporting Information S1). The binding curves for DENV1, 2 and 4 overlay while the curve for DENV3 is shifted to the right suggesting lower reactivity with this serotype. A similar trend was seen by SPR where 3F8 was immobilised using amine coupling and kinetic rates and affinity constants were measured for NS3 from the four serotypes (Table S2 in Supporting Information S1). The affinity of DENV3 NS3 was lower (KD 38.0 nM) than for NS3 from the other serotypes (DENV1 KD 6.6 nM, DENV2 11.0 nM, DENV4 16.7 nM). Nevertheless, 3F8 binding is in the nanamolar range across serotypes. Western blots were performed to demonstrate 3F8 specifity. The Fab recognises a band of the expected molecular weight of NS3 (70 kDa) in DENV2-infected C6/36 cells but not in uninfected cells demonstrating the detection is robust and specific (Figure S2 in Supporting Information S1).
As NS3 is an intracellular viral enzyme, to observe the effects 3F8 has on DENV replication it must first be delivered across the cell membrane. A protein transfection reagent was used to ensure sufficient 3F8 cell penetration is achieved in HEK293 cells. As shown in Figure 5A, immunofluorescence with the dengue antibody 4G2 shows punctate staining in HEK293 cells infected with DENV2, indicating that 48 hour post-infection most cells were infected with dengue virus where as control cells show no staining. Cells that were transfected with 3F8 prior to infection with DENV2 (Infection +3F8) show less staining for dengue suggesting reduced viral replication in these cells compared with mock transfected cells, and cells transfected with a non-binding Fab control (Infection +3F6). To confirm this plaque assays were performed with supernatants from infected cultures. Cells transfected with 3F8 prior to DENV2 infection showed a two-log decrease in released virus compared with DENV2 infected cells and cells infected with control Fab. Uninfected control cells show no viral replication (Figure 5B). Western blot shows bands in cells transfected with 3F8 using an anti-His antibody indicating successful transfection of His-tagged Fab. A western blot with the anti-dengue 4G2 antibody shows cells transfected with 3F8 have decreased band density verifying the immunofluorescence and plaque assay results (Figure 5C).
Peptide inserts obtained after three rounds of panning with the Ph.D-12 random dodecapeptide library and 3F8 were sequenced. Sequences were obtained for twenty phage clones and compared to the sequence for DENV2 NS3 helicase. The DENV2 serotype sequence was chosen for comparison as the peptide array used in the competetion ELISA (described below) was DENV2. Despite using solution panning methods to minimise plastic binding phage, and alternating the affinity resin (between anti-c-myc and Ni-NTA resin) to minimise resin binding phage, a large proportion of target-unrelated peptides were obtained [25]. Eighteen of the twenty sequences were highly hydrophobic and the phage clones had no affinity for 3F8 in an ELISA. However, one clone showed sequence identity to DENV2 NS3 helicase. The 12-mer peptide sequence was DETPMRGETRKV, the residues with identity to NS3 helicase are underlined. A second clone also displayed sequence identity with DENV2 NS3 helicase although there were fewer matching residues coompared with the first clone (LSPVQRNNVAII). Both clones had affinity for 3F8 in a phage ELISA and represent possible 3F8 epitopes (data not shown).
To determine which peptide-phage sequence (phagotype) contained the true epitope a competition ELISA was perfomed using an array of overlapping 15-mer peptides from DENV2 NS3 helicase subdomain III. The first phagotype was contained within peptide 105 and the second phagotype within peptide 113. As shown in Figure 6A, peptides 105 and 106 strongly compete with 3F8 for binding to DENV2 NS2B18NS3, whereas peptide 113, nor any other peptide from subdomain III, do not compete. The sequences for peptides 104, 105 and 106 suggest the lysine (K531) in the RGExRK motif identified by peptide phage display is essential for 3F8 binding as peptide 104 terminates at R530 and does not compete in the ELISA.
The 3F8 epitope identified maps to residues 526–531 in the third α-helix (α3′′) in subdomain III of NS3 helicase. An alignment of subdomain III from the four Dengue serotypes (Figure 6B) shows the epitope residues are strictly conserved with the exception of R526 which is replaced with a similar, positively charged residue (K526) in DENV3. Interestingly, this conservative substitution appears to have some effect on 3F8 reactivity. Signals in western blot and ELISA are lower for NS3 from DENV3 (Figure S3 in Supporting Information S1), and the affinity of 3F8 for DENV3 NS3 was reduced compared with the other serotypes (Table S2 in Supporting Information S1).
Surface accessibity derived from an apo DENV2 NS3 helicase crystal structure [7] is shown underneath the alignment in Figure 6B. Of the residues in the RGExRK motif R526, G527, E528 and K531 are highly accessible (blue) while R530 is buried (white) suggesting it may not interact with 3F8 in the structural epitope. The position of the epitope (green) on the surface of DENV2 NS3 helicase is shown in Figure 6C. It is distal from the ATP binding site (blue) but is in close proximity to residues involved in RNA binding (red). The ATP and RNA binding sites were previously proposed and/or observed in the DENV2 and DENV4 crystal structures [7]. The distance between the α-carbons of K531 in the 3F8 epitope and R538 and T264 in the RNA binding tunnel is 10.5 and 14.9 Å, respectively.
A naïve human Fab-phage library has been successfully employed to isolate antibody fragments with specificity for DENV4 NS2B-NS3. They can be broadly grouped based on their domain specifity; those that bind the NS3 helicase domain (3F3, 4, 7, 8 and 16), those that bind the 18 residues of NS2B required for correct folding and stability of the protease domain (3F10 and 11), and a final Fab that binds the 10 residue linker between the protease and helicase domains of NS3 (3F9). The helicase domain appears to have dominated the selection process with the majority of Fab binding this domain. This is most likely due to the size of the helicase domain (50 kDa) relative to that of the protease domain (20 kDa), rather than a lack of epitopes on the protease domain since a recent study in which the NS3 protease domain from West Nile Virus was subject to phage display identified several protease specific Fab [26], albiet with a synthetically expanded library.
From the 10 antibodies originally isolated, three of the helicase specific Fab (3F4, 7 and 8) cross-react with NS3 from the four DENV serotypes (Figure 2B). Of these, 3F8 was the most promising for detailed characterisation as it binds NS3 with 10-fold higher affinity (KD 10.5 nM) compared with 3F4 (KD 98.0 nM) and 3F7 (KD 94.9 nM). The epitope of 3F8 has been mapped to residues 526–531 in subdomain III of the helicase domain (Figure 6C). Subdomains I and II are observed across the SF2 superfamily of helicases. However subdomain III is unique to the flaviviruses [7] and it has been shown to influence DENV NS3 activity, with mutation of a single arginine (Arg513) to alanine in α2″ producing a defective helicase [27]. An alignment of flavivirus NS3 sequences shows the 3F8 epitope (RGExRK) is essentially conserved across several members of the flaviviridae including Japanese encephalitis virus, West Nile virus, Tick Borne encephalitis virus and Yellow fever virus suggesting the antibody will cross react with NS3 from these species (Figure S3 in Supporting Information S1). However, the epitope is not observed in hepatitis C virus, and the largest differences in both sequence and structure between DENV and hepatitis C virus NS3 have been observed in subdomain III [7].
The 3F8 epitope is distal from the ATP binding pocket, and 10–15 Å from the subdomain I end of the RNA tunnel. The impact 3F8 has on NS3 enzymatic activities was assessed using biochemical assays. 3F8 reduces NS3 catalysed unwinding of a double stranded DNA:RNA substrate (Figure 4C). When 3F8 is bound, accessibility to the RNA binding site may be reduced, especially when the size of a Fab molecule (50 kDa) is considered. Allosteric effects induced by 3F8 binding may also contribute to the reduced activity observed. Large quaternary changes have been observed in NS3 upon RNA binding [18] and these may be hindered in the presence of a strongly binding subdomain III antibodies such as 3F8. The helicase and ATPase activities of NS3 are linked with ATP hydrolysis being induced by RNA binding. This drives translocation of NS3 in a 3′ to 5′ direction along an RNA substrate. Interestingly, 3F8 was the only antibody to reduce ATP turnover by NS3 (Figure 4A). It is possible 3F8 binding induces a conformational change in the distal ATP binding pocket, or alternatively it may restrict RNA binding and, in turn, reduce ATP turnover. The ATPase assay includes poly(U) to stimulate NS3 ATPase activity [28].
The reduced enzyme activity observed in biochemical assays translated to a cell-based system with 3F8 reducing DENV replication in HEK293 cells (Figure 5B). While the direct effect 3F8 has on NS3 activity will contribute to this reduced replication, protein-protein interactions must also be considered. As part of the viral replicative complex, NS3 interacts with the RNA-dependent RNA polymerase NS5. The region for the NS5 interaction has been mapped to subdomains II and III of NS3 [29], and 3F8, by binding subdomain III may interfere with this interaction. NS3 co-localises with the integral membrane protein NS2B in infected cells [30] and Fab binding may also have steric effects in this context, inhibiting protein rearrangments that occur in close proximity with the endoplasmic reticulum membrane during the replication cycle. More studies are required to understand how the various DENV proteins assemble to form the molecular machinery necessary for RNA replication, and a panel of epitope mapped, NS3 and NS5 specific antibodies such as 3F8 may prove powerful tools in such studies.
The high specificity and affinity of 3F8, together with its ability to inhibit DENV replication may be further exploited in therapeutics. However as NS3 activity is intracellular, transport through the cell membrane is a major hurdle in developing anti-NS3 antibodies for clinical use. Experimental approaches such as coupling the antibody with a transport protein [31], or expressing the antibody fragment as an intracellular protein (intrabody) using gene therapy vectors maybe considered to overcome the challenges associated with delivery of biomolecules into cells.
|
10.1371/journal.pcbi.1000610 | Alu and B1 Repeats Have Been Selectively Retained in the Upstream and Intronic Regions of Genes of Specific Functional Classes | Alu and B1 repeats are mobile elements that originated in an initial duplication of the 7SL RNA gene prior to the primate-rodent split about 80 million years ago and currently account for a substantial fraction of the human and mouse genome, respectively. Following the primate-rodent split, Alu and B1 elements spread independently in each of the two genomes in a seemingly random manner, and, according to the prevailing hypothesis, negative selection shaped their final distribution in each genome by forcing the selective loss of certain Alu and B1 copies. In this paper, contrary to the prevailing hypothesis, we present evidence that Alu and B1 elements have been selectively retained in the upstream and intronic regions of genes belonging to specific functional classes. At the same time, we found no evidence for selective loss of these elements in any functional class. A subset of the functional links we discovered corresponds to functions where Alu involvement has actually been experimentally validated, whereas the majority of the functional links we report are novel. Finally, the unexpected finding that Alu and B1 elements show similar biases in their distribution across functional classes, despite having spread independently in their respective genomes, further supports our claim that the extant instances of Alu and B1 elements are the result of positive selection.
| Despite their fundamental role in cell regulation, genes account for less than 1% of the human genome. Recent studies have shown that non-genic regions of our DNA may also play an important functional role in human cells. In this paper, we study Alu and B elements, a specific class of such non-genic elements that account for ∼10% of the human genome and ∼7% of the mouse genome respectively. We show that, contrary to the prevailing hypothesis, Alu and B elements have been preferentially retained in the proximity of genes that perform specific functions in the cell. In contrast, we found no evidence for selective loss of these elements in any functional class. Several of the functional classes that we have linked to Alu and B elements are central to the proper working of the cell, and their disruption has previously been shown to lead to the onset of disease. Interestingly, the DNA sequences of Alu and B elements differ substantially between human and mouse, thus hinting at the existence of a potentially large number of non-conserved regulatory elements.
| Identifiable repeat elements cover a very large fraction of the human and mouse genomes, and even though they are quite diverse at the sequence level, they can be assigned to a fairly small number of families [1]. Alu and B elements belong to the Short Interspersed Nuclear Element (SINE) family, members of which exist in several mammalian genomes, where they have spread in great copy numbers [2]–[4]. Alu elements, the most abundant class or repeat elements in the human genome, originated in the duplication and subsequent fusion of the 7SL RNA gene at the beginning of the radiation of primates [5],[6]. B1 elements belong to the same repeat family and have also descended from the 7SL RNA. Following the primate-rodent split, copies of Alu and B1 elements have amplified and duplicated independently in the two genomes while accumulating mutations [4],[7]. The extent of the acquired mutations is such that extant instances of archetypal Alu and B1 elements bear little resemblance to one another or to the original 7SL RNA gene.
In earlier work, the Alu distribution in the human genome was studied in terms of several genomic features in order to understand how they spread in the genome: it was shown that Alu elements are predominant in R bands and inversely distributed with respect to L1 elements [8], correlated with GC-rich parts of the genome [9],[10] as well as gene and intron density [10]–[12], and enriched in isochores [11], segmental duplications [13] and transcription factor binding sites [14]. Another study of Alu, B1 and related SINE elements across mammalian genomes demonstrated their presence in primates, rodents, and tree-shrews and their absence in other mammals [15]. There have also been attempts to associate Alu elements with functional classes of genes. In [16], Alu elements located on chromosomes 21 and 22, were found to be over-represented in a limited set of functional classes. Housekeeping genes vs. tissue-specific genes were also found to have preferences for Alu elements [17]. In [14], the authors considered for their analysis only 5 kb upstream of known genes, and a limited set of functional classes for over-representation or under-representation of Alu elements.
In what follows, we extend previous work by studying and comparing the distributions of extant instances of both Alu and B1 elements, as well as related B2 and B4 elements (from this point on, we will be referring to B1, B2 and B4 elements collectively as “B elements”) in both upstream and intronic regions of known protein-coding genes, in order to contribute to the understanding of the evolutionary history of these elements. More specifically, we test whether their current distributions in the human and mouse genomes are a result of positive or negative selection across functional classes of genes.
Following the primate-rodent split, Alu and B elements spread throughout the human and mouse genomes: Alu elements currently number ∼1.1 million copies and cover about 5.4% of the human genome (in the sense orientation), while B elements number ∼1.2 million copies and cover about 3.6% of the mouse genome (in the sense orientation).
We studied Alu and B element densities separately for all combinations of: (a) distance from gene transcript start positions, (b) direction (upstream vs. downstream), and (c) orientation (sense vs. antisense). In the case of downstream direction, we computed Alu and B element densities separately for intronic and exonic regions. For a detailed description of the computation method and all relevant definitions, see Methods section. Our results demonstrate that Alu and B elements are significantly over-represented in the upstream regions of genes, and that the highest densities are observed within the window ending at 16 kb upstream of gene transcript start positions. For a detailed explanation of how we determine significance and how we compute p-values for all cases of over-representation and under-representation, see Methods. Similarly, Alu and B elements are significantly over-represented in the intronic downstream regions of genes, and, just as in the upstream case, the highest densities are observed in the window ending at 16 kb downstream of the gene transcript start positions. However, in introns, the over-representation is significantly more pronounced in the antisense orientation. Finally, there is a significant under-representation of Alu and B elements in exons and the effect of distance is not as pronounced as in the upstream and intronic downstream cases.
These results are shown in detail in Figure 1 for Alu elements in human and in Figure 2 for B elements in mouse: we plot Alu and B element densities upstream and downstream of known genes as a function of distance from the gene transcript start positions. Green and red curves correspond to Alu and B densities in the sense and antisense orientation respectively. In the downstream case, we distinguish between exonic and intronic regions.
We first associated Alu elements to functional classes by performing a genome-wide analysis on the latest release of the human genome annotations and applying a distribution-free statistical test with multiple hypothesis testing correction. Unlike the analysis in [14], where only 5 kb upstream of known genes were considered, we examined the 0 kb–16 kb window for the upstream analysis, i.e. the window where we find that the Alu density is maximized (see above). In addition, we: (a) examined the possibility that intronic instances might also be linked to specific functional classes, and (b) treated sense and antisense orientations separately. As a result, we were able to associate with Alu elements at least four times more functional classes than we would have been able to, had we only considered 5 kb upstream regions. Finally, after determining the functional associations, we conducted additional computational experiments to pinpoint the most likely explanation for the observed functional biases.
We applied the following statistical test in order to determine potential biases in the positioning of Alu elements within upstream and intronic regions of genes belonging to specific functional classes. After labeling each gene's upstream or intronic region with the GO terms attributed to the corresponding spliced transcripts, we tested whether Alu densities are significantly higher in the upstream or intronic regions of genes associated with certain GO terms. Density is defined as the fraction of the upstream or intronic region of a given gene that is covered by Alu instances. For a more formal definition of density and a detailed description of the statistical method used here we refer the reader to the Methods section.
Using this approach we found that upstream and intronic Alu instances are not randomly distributed, but instead are located, significantly more frequently than expected, inside upstream and intronic regions (in either the sense or antisense direction) of genes belonging to specific functional classes, i.e. GO terms. In Table 1, we report these functional classes at GO hierarchy level six or greater. In Supplemental Table S1, we report the entire list of GO terms and the associated p-values.
In order to validate our computational findings, we searched the existing literature for experimental evidence linking Alu elements to specific functions and compared them to the GO terms listed in Table 1 (or in the full list of significant GO terms found in Supplemental Table S1). Alu elements have been shown to be involved in DNA repair [18], to play a role in alternative splicing, RNA editing and translation regulation [19],[20], to repress transcription following heat shock [21], and to affect genomic organization and evolution, through insertion mutation and recombination [4],[22]. For most of these functions, we were able to find related significant GO terms: DNA repair, RNA splicing, translation, chromatin remodeling, and DNA recombination. In Figure 3, we verify that for these GO terms, the Alu density of associated genes in upstream and intronic regions is significantly higher than we would expect in a randomly chosen set of genes. Interestingly, most of the functional classes reported in Table 1 have not previously been linked to Alu elements, suggesting potential novel regulatory roles for these elements.
In search for the most likely interpretation of the functional biases of Alu instances in upstream and intronic regions reported in Table 1, we explored three alternative scenarios, and conducted further computational experiments in order to prove or disprove them. One possible explanation for our findings could be that Alu elements were selectively retained through natural selection in the genes of these functional classes, because they play a positive role in the function of these genes and offer a selective advantage. Had these insertions been neutral, no functional biases would have been observed in our analysis. If, on the other hand, these insertions had had a negative impact, they would have been selected against during evolution, considering that insertions in upstream regions of genes, where regulatory signals are located, could easily disrupt normal function. Not surprisingly, an obvious case of negative selection is found in the exonic regions where not only Alu elements are under-represented (see Figure 1), but also no functional biases are observed, in other words, the negative selection of Alu elements in exonic regions is active across all functional classes.
A second possible explanation could be that mobile elements in general possess either an insertion or a tolerance bias towards these functional classes of genes. In other words, either mobile elements may be preferentially inserted in genes belonging to these functional classes, or genes in these functional classes may tend to tolerate mobile element insertions better than the rest of the genes. To corroborate or refute these hypotheses, we tested whether other types of mobile repeat elements are enriched in the same functional classes as Alu elements and, in general, we found no significant overlap: 22% with LINEs and 1% with ERVs, 1% with LTRs and zero for all other mobile element families. Even in the case of LINEs, where we observed the highest overlap, none of these common classes is related to DNA repair, recombination, chromatin remodeling, splicing or translation. In addition, we analyzed the three main Alu subfamilies and discovered significantly fewer functional biases for the recently inserted Alu elements (see following section), thus demonstrating that these functional biases are crystallized as Alu elements survive longer inside the genome, and after some of these elements have been retained. In summary, we conclude that Alu elements share little in common in terms of functional biases with either older or younger mobile elements, and we can therefore rule out the tolerance and preferential insertion hypotheses, a conclusion that is in fact consistent with previous findings [23],[24].
A third alternative explanation could be that certain Alu instances were selectively lost after the initial random spreading, and, in fact, this scenario corresponds to the prevailing hypothesis. However, when we tested whether Alu elements are under-represented in the upstream or intronic regions of genes of specific functional classes, we found no such bias. This suggests that Alu instances have been lost randomly across functional classes.
Based on the above analysis, we conclude that, as described in the first scenario, there has been a positive selection of Alu elements in the upstream and intronic regions of the genes that belong to the functional classes reported in Table 1. This finding suggests that Alu elements likely play an active role in the entire set of functions listed in Table 1, and not only in the small subset which has already been reported in the literature.
B1 and Alu repeat families both descended from an initial duplication of the 7SL RNA gene [4] before the primate-rodent split, i.e. more that 80 million years ago. However, after the primate-rodent split, Alu and B elements spread independently, accumulated mutations and, over time, substantially diverged from the 7SL RNA sequence from which they originated [4],[7]. Consequently, extant B1 elements should be very different from Alu elements at the sequence level. We confirmed the lack of sequence similarity between Alu and B1 elements in two ways. First, in Figure 4, we show that the average pair-wise similarity among Alu elements is 71.5±11.1%, whereas the expected similarity is 45.3±4.4% as determined using shuffled versions of the Alu sequences. The average pair-wise similarity for B1 elements is 70.1±10.8%, whereas the expected similarity is 45.1±2.5%. In contrast, the average pair-wise similarity between extant Alu monomers and B1 elements is only 51.1±4.7% and very close to the expected similarity value of 44.2±2.7%. Second, using human/mouse whole-genome alignments we found that Alu and B1 elements are located overwhelmingly in non-conserved regions of the human and mouse genomes: the percentages are ∼99.9% in the case of Alu elements and ∼96.4% in the case of B elements (∼95.8% for B1, ∼96.9% for B2 and ∼96.5% for B4 elements).
Next, we applied the same statistical analysis used in the previous section, in order to look for enrichment of B elements in specific functional classes of genes. Given that, as shown above, the sequences of B elements are so different from those of Alu elements, and that the current distribution of Alu and B elements has been shaped independently in the each of the two genomes through initial random spreading and subsequent loss of certain copies, one would expect that the functional associations of B elements in upstream and intronic regions of genes would be different from the ones described in the previous section. However, we found that the set of functions associated with B elements contains 83.2% of the functions associated with Alu elements (expected = 12.2±2.0%). The fact that this result is observed independently in the mouse genome further strengthens our claim that these two types of SINE elements have been selectively retained in genes of certain functional classes, rather than selectively lost from certain genes.
Nevertheless, we examined an alternative scenario: since Alu and B elements are found in non-conserved regions of human and mouse, we tested whether certain functional classes of genes tend to have non-conserved upstream and intronic regions (effectively defining the differences between these two organisms), and whether these functional classes overlap with those associated with Alu and B elements. We found that the set of GO terms associated with non-conserved regions and the set of GO terms associated with Alu elements share only five entries in the combined sense/antisense intronic regions, and zero in the combined sense/antisense upstream regions. The common GO terms in the intronic case are generic high-level terms (e.g. metabolism, binding, etc.), and do not include DNA repair, recombination, chromatin remodeling, splicing or translation. Therefore, we conclude that lack of conservation of Alu and B elements does not explain the observed functional biases.
Human Alu elements belong to one of three main sub-families AluS, AluJ and AluY, with approximately 660,000, 283,000 and 148,000 copies respectively in the human genome. We repeated the above GO term analysis separately for each Alu sub-family and found that all three Alu sub-families are significantly over-represented in the upstream and intronic regions of genes of certain functional classes. Using the same cutoff on the adjusted p-values, we obtained 244 significant GO terms for the oldest AluS sub-family, 200 for the AluJ sub-family and 116 for the youngest AluY sub-family. The relationships of these three sets to one another are depicted in the form of a Venn diagram in Figure 5. A qualitative interpretation of the Venn diagram is that the AluS GO term set is an approximate superset of the AluJ set (86.0% of the AluJ set members are also members of the AluS set; expected overlap is 7.7±1.6%), which in turn is an approximate superset of the AluY set (93.1% of the AluY set members are also members of the AluJ set; expected overlap is 6.6±2.3%). The AluY set is 100% covered by the AluS set. The computed p-values for all sub-families, for both upstream and intronic regions, and for both sense and antisense orientations can be found in the Supplemental Tables S2 and S3.
Similarly, in the mouse genome there are B1, B2 and B4 elements with approximately 417,000, 363,000 and 390,000 copies respectively. Using the same method and cutoff, we found 293, 260 and 232 significant GO terms for B1, B2 and B4 elements respectively. Unlike Alu sub-families, where the number of significant GO terms increased with the age of the sub-family, here all three types of elements have comparable numbers of significant GO terms associated with them. Also, pair-wise intersection of these lists of GO terms show high similarities, measured using the Jaccard coefficient between each pair of sets: 65.6% between B1 and B2 (expected similarity = 4.6±0.8%), 54.0% between B1 and B4 (expected similarity = 4.4±0.9%), and 56.2% between B2 and B4% (expected similarity = 4.0±0.9%). The computed p-values for B1, B2 and B4 elements, for both upstream and intronic regions, and for both sense and antisense orientations can be found in the Supplemental Tables S4 and S5.
Almost all instances of Alu elements in human (95%) are conserved in the chimpanzee genome, i.e. they are included in human-chimpanzee whole-genome alignments. After repeating the GO analysis in the chimpanzee genome, we concluded that 81% of the identified significant GO terms are identical to the significant GO terms identified in human. Similarly, B elements are conserved between mouse and rat genomes: 50% of B element instances in mouse have a conserved counterpart in rat. Even though the level of conservation between mouse and rat B elements is not as high as between human and chimpanzee Alu elements, 90% of the significant GO terms identified in rat genome are identical to the significant GO terms identified in mouse. The results of the chimpanzee and rat analyses can be found in the Supplemental Tables S6 and S7 for chimpanzee and Tables S8 and S9 for rat. In conclusion, our findings show that there exists a human-chimpanzee conservation of Alu elements and a mouse-rat conservation of B elements on the sequence level. More importantly, there exists a conserved functional connection between all four organisms, independent of the level of cross-species sequence conservation of these elements.
Our analyses reveal that both upstream and intronic regions in human and mouse are significantly enriched in Alu and B elements respectively. Surprisingly, we find that Alu and B elements are significantly enriched across similar functional classes in human and mouse, even though these two types of elements have spread independently in the two genomes, following the primate-rodent split. In contrast, we find no depletion across functional classes, a finding which suggests that the final distribution of Alu and B elements across the two genomes is unlikely to be the result of a selective loss of some of their randomly retrotransposed copies. A simpler explanation suggests that they have been selectively retained in the upstream and intronic regions of genes belonging to the functional classes presented in Table 1, presumably because they offered some selective advantage (for example more binding sites to help increase the complexity of regulatory networks, or more transcript splice variants) thus increasing each organism's chances of survival. Indeed, a subset of the functional associations we uncovered in this paper has been reported in the literature, thus supporting the merit of our computational approach, while the majority of the functions are novel and suggest possible avenues to specific experimental tests.
Most importantly, our analysis suggests that SINEs are implicated in gene regulation effected through the upstream and intronic regions of specific genes, and contributes to an increasing body of literature attributing functional relevance to repeat elements which were initially ‘dismissed’ and labeled “junk DNA” [25]. Indeed, soon after the advent of genomic sequencing, reports of mobile elements that were exapted into novel genes and regulatory elements through retrotransposition [26]–[28] or exonization [29] started appearing in the literature. Individual instances of various types of repeat elements were shown to cause disease but to also drive genomic evolution in a positive manner [4],[22]. Recent reports also discuss findings suggesting that the role of mobile elements in genomic evolution, organization and cell process regulation may be significantly more important than previously thought [30]–[34].
Interestingly, the sequences of Alu and B elements are not conserved between human and mouse. For nearly three decades, most searches for regulatory elements made explicit or implicit use of the assumption of equivalence between sequence conservation and function. However, recent work has shown that the human genome regions can be classified into three broad categories with respect to the extent of their evolutionary conservation and their coding potential: (a) sequences that are under strong evolutionary constraints (∼5% of the human genome [35],[36]); (b) conserved non-exonic sequences that are more frequent than expected [37] but do not necessarily comprise functional elements [38]; and (c) non-conserved, non-exonic sequences, a category with an unexpected high number of functional elements [39]. Such findings increasingly question whether sequence conservation is a necessary and sufficient condition for function. Indeed, recent publications have revealed the existence of regulatory elements that are not conserved between human and mouse [33], [40]–[45].
Recent studies suggest that RNA silencing pathways including endogenous siRNA and piRNA pathways provide an adaptive defense in the transposon arms race [46], raising the possibility of a connection between RNAi pathway genes and Alu/B element insertions. Key proteins in these pathways, such as Argonaute and PIWI, are categorized as “gene silencing” proteins in the GO hierarchy, a term that is, in fact, identified by our statistical method as significant in the case of antisense upstream B element instances in mouse (see Supplemental Table S4), thus revealing a possible connection among genes that participate in the RNAi pathways and Alu/B elements.
In closing, it is worth emphasizing that, in our analysis, antisense intronic regions are significantly more enriched in Alu and B elements than sense intronic regions, unlike upstream regions, where no significant difference is observed between sense and antisense. In view of this finding, and taking into account previous work showing evidence of widespread occurrence of antisense transcription in introns [47],[48] as well as correlation of non-coding antisense intronic RNA levels with tumor differentiation [49], it is reasonable to conjecture that antisense intronic sequences may play an important role in regulation. Conceivably, this conjectured activity may be coordinated with instances of Alu and B elements located upstream of protein-coding genes. Taken together, these findings hint at the existence of a potentially very complex web of interactions among upstream regions, introns, and repeat elements in the context of cell process regulation.
Data sources. We obtained genome chromosome sequences and genomic region coordinates for transcripts, exons, introns as well as Gene Ontology (GO) annotations (biological processes and molecular functions) from ENSEMBL release 52. Human/mouse pair-wise alignments and repeat regions corresponding to the same genome assembly versions (NCBI36 for human and NCBIM37 for mouse) were obtained from UCSC Genome Browser.
Computing densities and associated p-values. We define density of a given type of elements (for example Alu or B elements) in a given genomic region as the fraction of the region that is covered by the instances of these elements. We calculated the densities of Alu and B elements in genomic regions obtained from all combinations of: (a) distance from gene transcript start positions (1, 2, 4, 8, 16, 32, 64, 128, 256 and 512 kb), (b) direction (upstream and downstream), and (c) orientation (sense and antisense). Each genomic region was identified as follows:
The expected Alu and B element densities were calculated on the entire human and mouse genome respectively. All density calculations were performed using resampling and the results are shown as mean and standard deviation on Figure 1 for human and Figure 2 for mouse. P-values were computed in each case using Student's T test between the observed and expected, or between sense and antisense in the intronic downstream case. For a wide range of distances (i.e. 4 kb–256 kb), both upstream and downstream, the p-values are practically zero.
Identifying significant GO terms and computing adjusted p-values. The following definitions are necessary for the rest of the section. A genomic locus x is a quadruplet (xc,xs,xa,xb) containing information about its chromosome, strand, and start and stop coordinates. A genomic region is a set of genomic loci. The overlap θ(x,y) between two genomic loci x and y is θ(x,y) = min(xb,yb)−max(xa,ya), if xc = yc and xs = ys, and 0 otherwise. The overlap θ(Q,R) between two genomic regions Q and R is the sum of overlaps θ(x,y) of all possible pairs (x,y) of genomic loci where x is in Q and y in R. The density δ(Q,R) of region Q in reference region R is defined as the overlap θ(Q,R) divided by the total length of reference region R, i.e. the sum of the length of the region's loci.
In order to determine which GO terms are significantly enriched in Alu/B elements, the following information is used as input to our algorithm:
For each gene g, we compute the density δ(g) = δ(Q,r(g)) of test region Q in the reference region r(g) of gene g. For each GO term t, we also compute the average density δ(t) of test region Q across the set of reference regions R(x) = { r(g) | g in G(t) }, i.e. the set of reference regions of genes associated with GO term t. Then, we calculate the p-value of δ(t) as the probability p(t) that value δ(t) is drawn from the null distribution. The null distribution of GO term density values is estimated using N = 1,000,000 randomized experiments designed to redistribute the test region loci Q within the reference regions r(g), while satisfying the following criteria:
All these criteria can be satisfied by simply permuting the density values δ(g) across genes of the same chromosome and strand. Then, the p-value p(t) for each GO term t is calculated as the number of randomized experiments where the randomized density δ'(t), as computed based on the permuted δ'(g) densities, exceeds or is equal to the observed density value δ(t), divided by the total number of experiments.
Since we carry out only 1,000,000 randomized experiments, p-values smaller than 1e-06 needed to be approximated for presentation purposes in the Supplemental tables, and this was achieved by approximating the tail of the null distribution with an exponential distribution. We point out that all the results presented in the manuscript regarding significance are based on the exact p-values and not on the approximated ones. Finally, in order to estimate the false discovery rate (FDR), we computed the adjusted p-values (q-values) according to the method presented in [50]. Two approaches were evaluated: (a) all hypothesis tests were considered as one family, and (b) each level of GO hierarchy was considered as a separate family. The difference of the outcomes of the two approaches was negligible. We also note that for a given repeat element family we analyzed the upstream sense/antisense and intronic sense/antisense regions simultaneously under the same random permutation experiment, i.e. we collected all the gene densities in all four types of regions together, in order to estimate the number of significant GO terms at 1% FDR.
|
10.1371/journal.ppat.1003182 | Genome-wide Prediction and Functional Validation of Promoter Motifs Regulating Gene Expression in Spore and Infection Stages of Phytophthora infestans | Most eukaryotic pathogens have complex life cycles in which gene expression networks orchestrate the formation of cells specialized for dissemination or host colonization. In the oomycete Phytophthora infestans, the potato late blight pathogen, major shifts in mRNA profiles during developmental transitions were identified using microarrays. We used those data with search algorithms to discover about 100 motifs that are over-represented in promoters of genes up-regulated in hyphae, sporangia, sporangia undergoing zoosporogenesis, swimming zoospores, or germinated cysts forming appressoria (infection structures). Most of the putative stage-specific transcription factor binding sites (TFBSs) thus identified had features typical of TFBSs such as position or orientation bias, palindromy, and conservation in related species. Each of six motifs tested in P. infestans transformants using the GUS reporter gene conferred the expected stage-specific expression pattern, and several were shown to bind nuclear proteins in gel-shift assays. Motifs linked to the appressoria-forming stage, including a functionally validated TFBS, were over-represented in promoters of genes encoding effectors and other pathogenesis-related proteins. To understand how promoter and genome architecture influence expression, we also mapped transcription patterns to the P. infestans genome assembly. Adjacent genes were not typically induced in the same stage, including genes transcribed in opposite directions from small intergenic regions, but co-regulated gene pairs occurred more than expected by random chance. These data help illuminate the processes regulating development and pathogenesis, and will enable future attempts to purify the cognate transcription factors.
| The genus Phytophthora includes over one hundred species of plant pathogens that have devastating effects worldwide in agriculture and natural environments. Its most notorious member is P. infestans, which causes the late blight diseases of potato and tomato. Their success as pathogens is dependent on the formation of specialized cells for plant-to-plant transmission and host infection, but little is known about how this is regulated. Recognizing that changes in gene expression drive the formation of these cell types, we used a computational approach to predict the sequences of about one hundred transcription factor binding sites associated with expression in either of five life stages, including several types of spores and infection structures. We then used a functional testing strategy to prove their biological activity by showing that the DNA motifs enabled the stage-specific expression of a transgene. Our work lays the groundwork for dissecting the molecular mechanisms that regulate life-stage transitions and pathogenesis in Phytophthora. A similar approach should be useful for other plant and animal pathogens.
| Eukaryotic pathogens typically employ specialized structures for dissemination and infection. Most filamentous fungi and oomycetes, for example, proliferate in their hosts as vegetative hyphae, which generate spores that are used to reach new infection sites [1]. The spores of many plant pathogens, especially those with a biotrophic disease stage, germinate to form structures known as appressoria that are used to breach the host epidermis. Transitions between these stages requires the precise control of transcription, which is accomplished through interactions between transcription factors and their binding sites (TFBSs) in DNA [2]. Some transcription factors and their cognate TFBSs have been identified in filamentous fungal and oomycete pathogens [3]–[6], but relatively little is known about the structure or regulation of their promoters compared to those of model saprophytes and animals. Studies in Saccharomyces cerevisiae have shown that its promoters typically contain only a small number of regulatory sequences located a few hundred bases upstream of the transcription start site [7]. This contrasts with metazoans, where genes are also controlled by more distant motifs, which often bind many transcription factors and exert long-range effects across chromatin domains [8].
Identifying TFBSs in the promoters of pathogen genes is an important step towards characterizing the networks that regulate growth, differentiation, and pathogenesis. The classic strategy for identifying regulatory motifs by promoter mutagenesis is laborious and not suited to genome-wide application, especially in non-model systems which include most plant and animal pathogens. In recent years, bioinformatic analyses enabled by genome sequencing and expression profiling have helped accelerate the discovery of promoter motifs in model organisms. Typically, co-regulated promoters are searched for over-represented motifs using methods that include enumerative search, expectation maximization, or Gibbs Sampling algorithms [9]–[12]. Comparative genomics also offers methods for predicting motifs by searching cross-species promoter alignments for phylogenetic footprints, i.e. regions of conservation [13]. The over-representation and evolutionary approaches have both been used with some success, since many of the resulting motifs resemble those identified by traditional methods [10], [13]–[17].
Relatively little is known about the organization and function of promoters in oomycetes, a group of eukaryotes that includes important pathogens of plants and animals. Studies of the potato late blight pathogen Phytophthora infestans and relatives revealed a novel genome structure comprised of gene-dense and gene-sparse regions [18]. P. infestans grows by extending tubular hyphae which then form sporangia, each of which can release multiple biflagellated zoospores [19]. In response to external cues, the motile zoospores transform into walled cysts which extend germ tubes that form infection structures called appressoria. In prior studies, we used the traditional strategy of promoter mutagenesis to identify three motifs directing transcription during sporulation and zoosporogenesis [20]–[22].
Resources enabling high-throughput promoter analysis have recently been developed for P. infestans, including a genome sequence and microarray data [18], [23]. In this report, we combine bioinformatic and functional approaches to identify TFBSs involved in stage-specific expression. More than 100 motifs associated with five life-stages were identified based on over-representation analysis. Most are high-confidence candidates since they also showed conservation in related species or positional bias within promoters. Functional testing of six motifs using reporter genes in P. infestans transformants confirmed their predicted activities.
To obtain data for planning the strategy for motif discovery, the genome-wide distribution of intergenic distances, GC-content in intergenic regions, gene orientations, and stage-specific expression patterns were analyzed. Previous researchers reported that the P. infestans genome is partitioned into gene-dense and gene-sparse regions [18]. We repeated that analysis, incorporating data on gene orientation (Figure 1A). We focused on the 67% of genes that had 5′ intergenic distances of <2 kb, since their transcriptional regulatory sequences would be more likely to interact with those of adjacent genes. Of gene pairs separated by <2 kb, 41% are transcribed from a common intergenic region, with transcripts in the 5′ to 5′ orientation; in such cases the median intergenic distance was 430 nt. Since 5′ UTRs in P. infestans average about 41 nt [24], this implies that two functional promoters can reside within as little as 300 nt. By comparison, median intergenic regions in S. cerevisiae, Arabidopsis thaliana, and Homo sapiens are 0.45, 1.5 and 35 kb, respectively [25]–[27]. The remaining 59% of P. infestans genes are transcribed in the same orientation, with the 3′ end of one gene being adjacent to the 5′ end of its neighbor; their median intergenic distance was 441 nt.
We also studied if intergenic regions varied in size depending on how genes were expressed, as this would help indicate the best search space for motifs. This took advantage of a prior microarray study of five developmental stages [23]. Five sets of promoters from 100 genes induced strongly (>7.5-fold) in each of the stages were assembled. These were from genes up-regulated in sporangia compared to hyphae (“sporangia promoter set”), sporangia chilled for 1 hr to stimulate zoosporogenesis versus untreated sporangia (“cleavage set”), zoospores versus chilled sporangia (“zoospore set”), and germinating cysts forming appressoria versus zoospores (“germinating cyst/appressoria set”). A hyphal set was also developed from genes with higher mRNA levels in hyphae than the other stages. In addition, 150 constitutive genes were identified for which mRNAs varied by less than 25% between stages. Each gene model was curated manually, guided by EST data and sequences from Phytophthora ramorum and Phytophthora sojae. Corrections were applied to the 5′ ends of 13% of the P. infestans gene models.
The resulting data suggested that stage-induced promoters are larger than those from constitutive genes. This involved sorting the genes into groups with the different expression patterns, and then calculating median 5′ intergenic distances for the subset that were closely spaced, i.e. 2 kb or less from another gene (Figure 1B). Although each dataset spanned a broad range, the median 5′ intergenic region of constitutive genes was the smallest at 317 nt. Values from the inducible genes ranged from 373 for the sporangia set to 616 for the hyphal set. This resembles trends in other species, presumably since variably-expressed genes bind more transcription factors [25].
To develop background models for evaluating the statistical significance of motif frequencies, we also measured AT content 1-kb upstream of ATG codons. This averaged 49.6%, but rose to 54% near the start of genes (Figure 1C). The profile of the curve in the figure may reflect the small size of the typical P. infestans promoter, if its functional regions have a uniform AT content. Alternatively, the core promoter (the site that nucleates the assembly of a functional preinitiation complex; [28]) may be more AT-rich than other upstream regions, where most stage-specific TFBSs are expected. Genome-wide, intergenic regions are 49.3% AT compared to 46.1% for coding sequences [18].
In light of the close proximity of most P. infestans genes, we examined whether genes influenced the transcription of their neighbors. This would be relevant to motif discovery since a TFBS between two genes might influence both. For the analysis, we mapped expression patterns along P. infestans supercontigs and also calculated correlations between adjacent gene pairs.
For mapping expression patterns, we linked features in the microarrays to gene models in the P. infestans assembly, upon which the transcription patterns were plotted. This is illustrated in Figure 2A for a representative portion of Supercontig 1 (not drawn to scale), in Figure 2B for four selected regions (drawn to scale), and in Figure S1 for all genes. In these figures, genes showing >2-fold higher mRNA levels than average in one of the five stages are color-coded based on the stage with the maximum level; for example, green means highest in sporangia. Genes with the same stage-induced pattern were not typically adjacent, for example there were no physical clusters of genes having peak expression in sporangia. We also calculated the probability of adjacent genes having the same stage-induced pattern, focusing on 3744 pairs of neighboring genes as well as a subset of 2937 genes residing within 2 kb of each other; while P. infestans encodes about 17,797 genes, not all were represented or yielded signals on the arrays. With a few exceptions, adjacent genes showed unrelated patterns of stage-specific induction. Most exceptions involved tandemly repeated gene families, which would be expected to be co-expressed since both promoter and coding regions were likely to have undergone duplication. This occurred most for genes induced in the germinating cysts with appressoria stage; only for this stage were co-induced genes clustered more than expected by random chance at a 95% confidence interval. This was attributable to tandemly duplicated sets of β-glucanases, protein kinases, glucose transporters, and bZIP transcription factors, among others. One example is presented in the lower right portion of Figure 2B, which illustrates three co-expressed β-glucanases (PITG_03511, PITG_03512, and PITG_03513). A second example is an array of genes annotated as glucose transporters (PITG_13001 to PITG_13007). Such observations are consistent with prior studies that showed that genes induced in this pre-infection stage are rapidly evolving and prone to duplication [29].
In addition to the above which focused on the distribution of stage-induced patterns, we also measured global expression between gene pairs since this might detect subtle interactions. There was a weak tendency for pairs to be co-expressed, with an average Pearson correlation coefficient (r) of 0.11. Moreover, the distribution of r values between gene pairs and pairs from a scrambled dataset were significantly different based on a Kolmogorov-Smirnov test (p<0.001). The expression of 375 pairs were highly correlated (r>0.8) and 87 were anticorrelated (r<−0.8). Of the 375 co-regulated pairs, 53% were transcribed in the same direction, and 55% of these represented duplicated genes. In contrast, only about 10% of the co-regulated 5′-to-5′ genes were duplicated. There was little correlation between 5′ intergenic distance and co-regulation (r = −0.09).
The scheme illustrated in Figure 3 was used to identify candidate stage-specific TFBSs. In brief, the five sets of stage-induced promoters were searched for motifs that were over-represented compared to total P. infestans promoters. A search space of 1-kb of 5′ sequences was selected since this should include most TFBSs, based on the data in Figure 1B. The motifs were then tested for positional bias, orientation bias, and evolutionary conservation.
Each of the five stage-induced datasets were searched separately for motifs using BioProspector, MEME, and YMF. These programs were selected since they employ independent methods and scored well in prior comparisons [11], [30]. We focused on promoters from genes induced >10-fold between developmental transitions (443 genes in the five promoter sets); this fold cut-off was raised compared to our earlier analyses to reduce noise in motif discovery. We also focused on motifs detected by at least two of the programs, allowing degeneracy at two sites. About 145 motifs fit this requirement, which were consolidated to 107 by joining those that were similar in sequence and had similar patterns of over-representation. Based on a p-value threshold of 10−2, 103 showed significant over-representation in at least one stage, which is shown in heat-map format in Figure 4; the motif sequences and number of hits in each dataset are in Table S1. The overall AT content of the 103 motifs was 49.9%, five were palindromes, and lengths ranged from 6 to 9 nt.
Approximately 80% of the 103 motifs were linked strongly to one developmental stage or two consecutive stages, and are therefore good candidates for binding sites for transcription factors that determine stage-specific expression. Based on the number of stages for each motif that passed the p-value threshold of 10−2, about 52 motifs were specific for a single stage. Examples include motif 82 (M82), which was significantly over-represented in sporangia-induced promoters (p = 10−8) but not the other sets, M95 which associated only with germinating cysts with appressoria (p = 10−12), and M99 which associated only with hyphae (p = 10−12).
About 21 motifs were significantly over-represented in promoters from two sequential stages, and 10 from three sequential stages. About half of the 21 were over-represented in the germinated cyst/appressoria and hyphal promoters, such as M87 (p = 10−14 and 10−13, respectively). This was not unexpected, since cyst germ tubes are very similar to hyphae and transition into hyphae. Several motifs were over-represented in sporangia and cleaving sporangia-induced promoters, such as M43 (p = 10−11 and 10−7, respectively). This was also not surprising since these stages are separated only by a 1-hr cold treatment, and many mRNAs induced in sporangia continue to rise during zoosporogenesis and/or during the zoospore stage. Accordingly, some motifs such as M64 were also over-represented in the sporangia, cleaving, and swimming zoospore promoters (p = 10−4 and 10−3, and 10−3 respectively). Likewise, several motifs were over-represented in hyphal and sporangia-induced promoters, such as M86 (p = 10−6 and 10−8, respectively). This may be explained by the fact that oomycete sporangia develop directly from hyphae, or that some tissue samples used for microarray analysis were not very synchronous. Regardless of the explanation, the approximately 80 motifs that associated with promoters from one or two sequential life stages are all good candidates for sites that bind transcription factors with stage-specific activities.
Two motifs matched the three promoter sites that were shown previously by mutagenesis to be needed for stage-specific transcription. M97, which was over-represented in the cleavage promoter set, is a close match (in the reverse orientation) to a site required for inducing the NifC gene during that stage [20]. Sporangia-associated motif M43 is an exact match to the site required for inducing the Pks1 gene during sporulation [22], and a close match (in the reverse orientation) to the region needed to induce Cdc14 during sporulation [21].
Not all motifs were associated only with consecutive stages. About 15 were over-represented in promoters from nonconsecutive stages, or both developmental or constitutive promoters (Figure 4, Table S1). One example is M60, which was over-represented in sporangia and swimming zoospore-induced promoters (p = 10−16 and 10−13, respectively) but not the intervening stage of cleaving sporangia (p = 0.8). A total of six motifs (M1, M8, M13, M60, M64, M67) occurred more in total promoters than expected by random chance; these may act as general enhancers.
Many transcription factors need to act at a certain distance from the transcription start site or other regulatory locations, and therefore their TFBSs concentrate at a certain site within promoter space [31], [32]. Whether any of the motifs exhibited this bias was determined by mapping them within 200-nt bins from the relevant promoter set; 65 motifs were found to have positionally biased distributions (column “Pos. Bias” in Figure 4, Table S1). This may be an underestimate, since convincing evidence of bias could not be drawn for low-frequency motifs.
Data for 18 representative positionally biased motifs are shown in Figure 5. About one-third, such as M8 and M38, had distributions matching overall promoter size as shown in Figure 1A indicating that these TFBSs lack a strong positional bias. In contrast, motifs such as M68 and M83 tended to reside 200–600 nt upstream of the start codon. Others such as M30, M39, M53 (not shown in Figure 5), M57, and M82 were found closer to the transcription start site. M30 and M39 do not match known oomycete core promoter motifs, but M53 resembled the Inr or Initiator [33]. Interestingly, M53 was over-represented in promoters induced in the germinating cyst with appressoria stage.
As a control, we observed that similar biases were not observed in total promoters, where most matches may be false hits. This is illustrated at the base of Figure 5 for three representative motifs, M87, M95, and M97. These had biased distributions in induced promoters (Figure 5, second row from bottom), but very different patterns in total promoters (Figure 5, bottom row). Due to variation in AT-content across promoters (Figure 1C), the controls are not expected to have similar values in each bin. As AT-rich motifs, hits to M87 and M95 due to random chance are more common in the 3′ portion of total promoters, which are AT-rich. The opposite was observed for GC-rich motifs such as M97.
Some transcription factors must orient in a certain direction to fulfill their regulatory function. In S. cerevisiae, for example, 47% of TFBSs were found to have an orientation bias [32]. As shown in the column labeled “F/R bias” in Figure 4 and Table S1, this was the case for 50 of the 98 non-palindromic motifs (52%) from P. infestans, using a p-value threshold of 10−2. Cleaving sporangia-associated motif M33, for example, was detected 87 times in the forward orientation in the 95 cleavage-induced promoters but only 50 times in the reverse orientation. After correcting for the false discovery rate, the bias is even greater at 71 versus 27.
About 78% of P. infestans motif candidates were judged to be conserved in orthologous promoters from P. ramorum or P. sojae. A conclusion about whether a motif was conserved was developed by aligning promoters from about five genes; a match in at least some was considered to be indicative of conservation. Assessments for each motif are shown individually in the “Evol. Con.” column in Figure 4 and in Table S1, and results for all motifs are summarized in Figure 6A. Evidence for conservation between P. infestans and both of the other two species was obtained for 52% of motifs. About 10% of motifs were conserved between P. infestans and P. ramorum only, and 16% between P. infestans and P. sojae only. In 11% of cases, motifs were absent from both P. ramorum and P. sojae. In 11% of cases, the motifs were detected at new locations in one or both species; this was taken as an ambiguous result, since while promoter rearrangements are common [34] they are hard to distinguish from false hits.
Figure 6 shows representative alignments where conservation was detected. For cleavage-induced P. infestans gene PITG_16321 and its P. ramorum and P. sojae orthologs, for example, perfect matches to M51 were detected in the same location in all three species. The PITG_16321 alignment also reflects the common relationship seen between orthologous promoters: two to four sequence blocks are typically conserved. One usually spans the transcription start site, which in this case contains an Initiator-like sequence at −71 in P. infestans. Other conserved regions are typically found 40–200 nt upstream. For PITG_16321 these are the M51-containing block at −177, and another at −114. As will be shown later, the −114 block and conserved nucleotides a few bases to the left and right of M51 do not determine stage-specific expression.
Results for three ortholog sets containing sporulation-associated motif M58 are also shown in Figure 6. For PITG_03886, M58 is conserved perfectly in the three species. In PITG_09960, a three-way match also exists allowing for one base change in P. sojae; this was scored as a positive hit, since TFBSs often vary between species [35]. The PITG_14222 alignments show M58 at the same location in P. infestans and P. sojae, but 80-nt upstream in P. ramorum. Since the latter could be a false hit, our scoring scheme classifies M58 in PITG_14222 as conserved only between P. infestans and P. sojae. Since M58 was at the expected location in P. ramorum and P. sojae orthologs of PITG_03886 and PITG_09960, however, its overall classification is “conserved”.
Nearly all of the motifs demonstrated one or more characteristics typical of authentic TFBSs besides over-representation, such as interspecific conservation, positional bias, orientation bias, or palindromy (Figure 7). Of the 103 motifs, 101 had at least one of these features in addition to over-representation, 78 had at least two, and 25 had three. These classifications help indicate which motifs have the highest likelihood of having a function, in addition to suggesting how they interact with the transcriptional apparatus.
The two motifs lacking these additional characteristics were M1 and M66. These may still be real TFBSs, since not all experimentally confirmed sites exhibit positional bias or directionality, or reside in the same location in orthologs. The two motifs were over-represented in at least one developmental stage with p-values ranging from 10−4 to 10−5 and thus are unlikely to be false hits.
A few of the “high-confidence” motifs were close in sequence. These were M93 (TACATGTA) and M94 (TACCGGTA), which are palindromes differing only at the two central bases, M32 (AGC[AG]CAAG) and M34 (AGCTGAAG) which also differ at the two central bases, and M16 (AAATAAA) and M91 (TAAATAA) which overlap. As mentioned earlier, most motifs from BioProspector, MEME, and YMF had been merged if they differed at two or fewer sites and were over-represented in the same stages. The six motifs remained unmerged since their biases and/or probability distributions and varied. For example, M16 but not M91 was over-represented in hyphal-induced promoters (p = 10−6 and p = 10−1, respectively), and M93 was more over-represented than M94 in hyphal promoters (p = 10−26, p = 10−2).
Six of the motifs were subjected to experimental analysis to see if they could drive β-glucuronidase (GUS) expression with the expected stage-induced pattern in transformants of P. infestans. As described below, each yielded the expected pattern. First analyzed was M51, which was predicted to confer expression during zoosporogenesis (i.e. sporangial cleavage). Interestingly, M51 is flanked by two sequence blocks that are conserved in P. ramorum and P. sojae, which are labeled LB and RB in Figure 6. These flanking sequences were not over-represented in cleaving sporangia promoters, but we considered the possibility that our definition of M51 was smaller than the authentic TFBS.
Initial experiments showed that at least part of the LB-M51-RB region was required for zoosporogenesis-specific expression. Plasmid pDEL312, which contains a 312 nt promoter fused to GUS, yielded expression in sporangia treated at 10°C for 1-hr to induce the cleavage of sporangia into zoospores, but not sporangia maintained at 22°C; for this plasmid and others described below, similar results were observed in multiple transformants. The zoosporogenesis-specific activity of the promoter fragment was shown first by histochemical staining (as in Figure 8), and later by RNA blot analysis in which bands of the expected size were detected only in the chilled samples (Figure 9). No activity was seen in hyphae. Indistinguishable results were obtained using a 500 nt promoter (not shown). pDEL187, which lacked bases upstream of the LB-M51-RB region, showed the same staining pattern and gene induction was confirmed by RNA blot analysis. pDEL104, which lacks the LB-M51-RB block, showed no expression.
Subsequent experiments specifically tested the functions of LB, M51, and RB by mutating those regions within pDEL187, and led to the conclusion that only M51 conferred stage-specific expression. Similar results were obtained from histochemical staining (not shown) and RNA blot analysis (Figure 9). Specifically mutating LB had no effect on stage-specific expression (pMUT1), while altering M51 prevented expression (pMUT2). As a control, we showed that the native gene in the transformants was induced in sporangia by the cold-treatment. Mutating RB did not block cold-induction of the transgene, although its basal expression seemed to be slightly elevated (pMUT3). Next, oligonucleotides containing the LB-M51-RB block or M51 alone were fused to the NifS minimal promoter, which contains a transcriptional start site but is not expressed on its own [21]. As shown in Figure 9, a fusion of LB-M51-RB to the minimal promoter, separated by a 37-nt spacer of random DNA, drove the normal chilling-specific expression of GUS (pOLIGO1). As a final and definitive test, an oligonucleotide containing M51 alone was shown to also confer this wild-type pattern to transformants (pOLIGO2).
Since the above experiments indicated that at least some motifs could act autonomously, we next tested five other predicted stage-specific motifs by fusing them one at a time to the NifS minimal promoter. Motifs M39, M58, M64, and M75 were most over-represented in the sporangia stage, and each resulted in the sporulation-specific accumulation of GUS (Figure 8). No staining was seen in nonsporulating hyphae. The effects of the motifs were subtly different, however. Transformants containing M58 showed GUS staining at the earliest stage; these showed expression within hyphae soon after cultures were stimulated to sporulate, and then later in sporangiophore and mature sporangia. This illustrated in Figure 8 where the three panels show (left to right) staining within a small segment of a hypha in a sporulating culture, immature sporangia (lacking basal septa and papilla), and mature sporangia. Transformants containing M39 and M64 first exhibited GUS staining in hyphal-like structures that are presumed to be sporangiophores, and then in mature sporangia. In contrast, expression driven by M75 seemed to be activated at a later stage, since staining was first observed in sporangia near maturity. It should be noted that while M58 was most over-represented in the sporangia stage (p = 10−8), it was also over-represented in hyphae (p = 10−6) and constitutive promoters (p = 10−3); perhaps it binds a transcription factor which does not become activated until sporulation is induced.
Also tested as a fusion with the minimal promoter was M95, which was associated with transcription in germinated cysts and appressoria. This resulted in the accumulation of GUS in germinated cysts, including their germ tubes and appressoria (Figure 8). Staining was first observed 2 hours after encystment. No expression was observed in hyphae, sporangia, chilled sporangia, or zoospores. As described later, this motif is associated with the expression of many pathogenesis-related proteins.
Further support for the motifs was provided by electrophoretic mobility shift assays (EMSA) involving M51, M58, and M75. As shown in Figure 10, each motif bound a protein from nuclear extracts of cleaving sporangia (M51) or sporangia (M58, M75). Binding appeared to be specific based on comparing different unlabeled competitors. These included a specific competitor (same sequence as the labeled probe), a nonspecific competitor (a random sequence), and mutated competitor (same as the labeled probe, but with the motif mutated). In each case the nonspecific and mutated competitors had little effect in reducing the binding of the labeled probe, compared to the specific competitor. For M51, several bands were detected, which was suggestive of a multi-protein complex or the binding of proteins of different sizes.
Several classes of proteins have been identified that play roles in pathogenesis, of which many are secreted and sometimes induced during infection [29], [36], [37]. To assess the usefulness of our data for understanding how such genes are regulated, we checked their promoters for the motifs, focusing on motifs associated with the germinating cyst/appressoria stage. This involved analyzing the main classes of genes annotated by Raffaele et al. [29] as potentially encoding secreted pathogenicity factors, of which many are induced during plant infection. As shown in Table 1, four motifs associated with the germinating cyst/appressoria stage and ten linked to both the hyphal and germinating cyst/appressoria stages were over-represented (p<0.05) in such genes. These included those encoding plant cell-wall degrading enzymes, glucanase inhibitors, Nep1-like (NLP) toxins, PcF toxins, elicitins and elicitin-like proteins (potential sterol carriers), proteases, protease inhibitors, and RXLR effectors. As expected, motifs linked to stages such as sporangia and zoospores were typically under-represented (Table S2).
Not all genes in each group contained a germinated cyst/appressoria motif in their promoters, however. For example, such motifs were in only 225 of the 493 RXLR promoters, with 66 containing M95, 163 having M101, and 49 having M103 (Table S2). As only some RXLR genes are induced during infection [18], [29], [38], we checked for a correlation between motif and expression pattern. RXLR genes with a germinated cyst/appressoria motif were more likely to be infection-induced than those without; many had more than one motif, with a correlation between the degree of induction and motif number (r = 0.27, p = 0.04). Crinkler genes, which are not typically infection-induced but are considered to encode pathogenicity factors due to the ability of some to produce necrosis in plants [18], had M93 as the sole over-represented motif. M93, a palindrome, is over-represented in both germinated cyst/appressoria and hyphal-induced genes and occurs 4150 times within P. infestans promoters. Its abundance suggests that it is associated with general growth and not specifically with pathogenesis.
We assessed the extent to which the presence of a motif predicts a gene's expression pattern. This involved searching promoters of all 7,862 genes on the microarrays for motifs associated with sporangia-induced genes and germinated cyst/appressoria genes, using 500-nt of DNA upstream of the start codon as the search space. We then compared motif frequencies in promoters induced by >5-fold at each stage versus non-induced promoters. The 99 sporangia-induced and 103 germinated cyst/appressoria-induced promoters used originally for motif discovery were excluded from these analyses, to test if our earlier results extended to all P. infestans genes.
Each of the 11 sporangia-associated motifs occurred more often in the induced promoters than non-induced controls (Table 2). On average, each motif was 66% more likely to occur in an induced promoter, with individual motifs showing a 21 to 100% enrichment. For example, M8 was found in 14.3% of induced promoters compared to 11.6% of non-induced promoters, representing a 23% enrichment. It is important to note that hits due to random chance are always expected to greatly exceed the number of functional TFBSs for reasons elaborated upon in Discussion [39].
Most of the 15 motifs linked to the germinated cyst/appressoria stage were also over-represented in that stage when the total microarray data were analyzed (Table 2). Each was on average 19% more likely to be in an induced promoter compared to controls. It is notable that four of the motifs were not enriched in the genome-wide set of induced promoters, including M95. Since our functional tests showed that M95 conferred expression in the germinated cyst/appressoria stage, it is possible that M95 binds a bifunctional transcription factor or has its activity mediated by other transcription factors.
We also checked for the association of a sporulation-associated motif with expression pattern in ten genes from P. infestans that were not on the microarrays. Motif M8 was chosen for this exercise simply since it was first on the list in Table 2. We identified promoters containing M8, used RT-qPCR to measure mRNA in sporangia and nonsporulating hyphae, and assessed if M8 was within the orthologous promoter from P. sojae (Table 3). Of six P. infestans genes in which the P. sojae ortholog also contained the motif, five were induced by >2-fold in sporangia. This was significant (p = 0.004), compared to the likelihood of this fraction of genes being induced by random chance. In contrast, none of the four P. infestans genes that lacked M8 in their P. sojae ortholog was induced based on the 2-fold cutoff.
Genome-wide searches for promoter motifs shared by co-expressed genes have been performed in model animals, plants, and fungi [14]–[17], but only on a limited scale in pathogens [40], [41]. The strategy seemed attractive for Phytophthora since its modest transformation efficiencies make motif discovery through traditional means challenging [42]. The success of our approach was shown not only by the identification of 100 putative TFBSs, but the fact that all six motifs tested performed as predicted in functional assays. Nearly all motifs also exhibit at least one feature typical of TFBSs besides over-representation such as positional bias, orientation bias, or evolutionary conservation. Discovering the motifs, which include several associated with pathogenicity factors, is a key step towards understanding the networks that control development and host infection in Phytophthora and similar approaches should be useful in other pathogens.
Several features contributed to our approach by increasing the sensitivity of our searches and reducing false positives. First, our requirement that motifs be identified by two of three algorithms served as a stringent filter. Second, we focused on promoters that show large changes, which was possible since major shifts in mRNA levels occur during the P. infestans life cycle as about 12% of genes change by >100-fold in the stages addressed by this study [23]. Third, gene models were manually curated to accurately define the search space. Finally, since intergenic distances are typically small in P. infestans, most regulatory regions were probably within the 1-kb search space.
Our analysis of the overall transcriptional landscape of P. infestans has also helped illuminate the structure and function of its promoters; few promoter studies have previously been performed in the entire Kingdom Stramenopila, which includes diatoms and brown algae in addition to oomycetes [43]. Remarkably, the median intergenic distance within gene-dense regions of P. infestans is even less than that of most yeasts [25], [44]. The ratio of adjacent P. infestans genes that are transcribed in the same direction versus from a shared or adjacent promoter region is 1.43, which is higher than that of S. cerevisiae and A. thaliana [45]. This presumably reflects functional constraints associated with having small adjacent promoters, which is reflected in the co-expression or anti-correlated profiles of about 10% of adjacent P. infestans genes. Excluding cases where co-expressed pairs are duplicated genes, most adjacent genes in P. infestans are nevertheless transcribed independently. Our prediction of directionality for more than half of the motifs helps to explain how this independence is mediated. The predominant mechanism for regulating transcription in Phytophthora may also not involve chromatin-level effects, which in yeasts and metazoans are inferred to extend up to 4 kb and tens of kilobases, respectively [46]. Nevertheless, of the approximately 300 transcription factors annotated within each Phytophthora genome, several belong to families associated with chromatin remodeling [47].
Relative simplicity in transcriptional regulation in P. infestans is also implied by our finding that each of six stage-induced motifs tested conferred tissue-specific expression with a minimal promoter. Combinatorial control, not counting transcription factor heterodimerization, thus may not be a principal feature of stage-specific regulation in oomycetes, unlike other eukaryotes with complex genomes [48], [49]. Since position effects in P. infestans make it challenging to compare transgene expression between transformants [50], our data are silent on roles of other TFBSs in quantitative expression. The potential involvement of only a few TFBSs per gene is consistent with our observation of limited blocks of similarity between P. infestans, P. ramorum and P. sojae promoters, as shown in Figure 6. As the three organisms are relatively distant in molecular phylogenies [51] and have significant morphological differences, it would be useful to know if the orthologs had similar patterns of expression.
Our analyses of motifs associated with sporangia and germinated cyst/appressoria stages (Table 2) suggests that the occurrence of a motif has utility in predicting expression pattern. However, it is important to recognize the limitations of this approach. Since TFBSs are short and often degenerate, they occur by random chance in great abundance. Moreover, TFBS function depends on chromatin structure and often the co-occurrence of other TFBSs. Some transcription factors are also bifunctional, leading to different outcomes depending on post-translational modification or co-regulators [52]. Because of such complications, Wasserman and Sandelin [39] posited the “futility theorem” which states that the great majority of predicted TFBSs lack function: it is thus futile to predict expression based on the occurrence of a promoter motif. Our experience in P. infestans was more encouraging, however. For example, while the presence of a sporulation-associated motif was a weak forecaster of expression pattern, the predictive value was fairly strong if the motif was conserved in another Phytophthora. The limits of predictions based solely on motif presence can be illustrated for the 11 sporulation-associated motifs in Table 2. Based on their average size and base composition, and using 500-nt of promoter sequences as a search space, about 8250 promoters should contain one or more of the 11 motifs by random chance. Extrapolating from microarray data, however, only about 2200 genes are sporulation-induced, so random hits exceed functional TFBSs by a 4 to 1 ratio. Nevertheless, a future can be envisioned where better predictions of expression based on motif occurrence alone may be possible. In S. cerevisiae, a network model that integrated expression patterns of 2,587 genes under 255 conditions of growth and development with 666 TFBS definitions using AND, OR, and NOT logic resulted in fairly good predictions of expression of about 3/4 of the genes [53], [54].
Inferences about the complexity of the networks that control development and pathogenesis in P. infestans may be drawn from our observation that roughly 10 to 20 motifs were linked to each stage of the life cycle. This is consistent with observations that show that sporangia and zoospore formation involves several steps and signaling pathways [20], [55], [56]. Characterizing transcription factors that bind the motifs will help reveal details of these pathways, and enable chromatin immunoprecipitation studies to confirm the target genes [57]. Studying the transcription factors may also lead to strategies for blocking diseases, by interfering with the expression of proteins used for overcoming host barriers and defenses.
Expression data were from a prior study that used Affymetrix microarrays to measure mRNA during the stages addressed by this paper [23]. Reliable expression calls were detected for 12,463 of the 15,650 sequences targeted by the arrays. Since the microarrays predated the current draft genome which is based on strain T30-4 (available from the Broad Institute of Harvard and MIT), we linked the microarray sequences to annotated T30-4 genes using BLASTN, but excluded genes on small contigs to reduce errors in analyses of intergenic distances and co-expression. By selecting the best hit with >97% identity, 7,862 genes with reliable expression data in the five life-stages were matched, including 3944 adjacent gene pairs of which 2937 were within 2 kb of each other. These were mapped to the assembly, omitting unexpressed or missing genes.
Datasets of P. infestans promoters included 1-kb of sequences 5′ of predicted open reading frames. Total promoters were downloaded from the Broad Institute database, and then subsets were extracted using custom scripts. Sets included promoters from the differentially-expressed gene sets described above, in which mRNA levels were induced by at least 7.5-fold compared to the prior development stage (p<0.05 based on replicates). Sets of at least 100 promoters were used for calculating 5′ intergenic distances. For identifying over-represented motifs, analyses were limited to genes induced >10-fold, which corresponded to 99, 95, 46, 103, and 100 in the sporangia, cleavage, zoospore, germinated cyst, and hyphal sets, respectively. Prior to extracting promoters, gene models were examined and corrected as needed (changing 8, 14, 17, 3, and 9 promoters, respectively). This mostly involved eliminating introns that contradicted EST evidence, or spanned regions that when converted to exons maintained the reading frame and had high similarity to P. ramorum and P. sojae orthologs. In addition, a constitutive promoter dataset was established from 150 genes that showed <25% variation between the stages. Promoters from P. ramorum and P. sojae were extracted from genome assemblies downloaded from the Virginia Bioinformatics Institute.
Stand-alone versions of MEME (version 4.3.0; [12]), YMF (version 3.0; [9]), and BioProspector (release 2; [10]) were employed. MEME ran with minimum and maximum widths of 5 and 8, respectively, using 5 iterations. Gap opening and extension costs were 11 and 1, respectively, any number of repetitions were allowed, and the E-value cut-off was 10−5. YMF used a value of 8 for lenOligo (the number of non-spacer characters) and output was sorted by z-score. BioProspector used a value of 8 for motif width, with the 100 top motifs reported per run. The program was run 10 times on each set of promoters and a PERL script was used to eliminate redundant motifs.
Initial outputs (382 motifs from BioProspector from the five stage-induced promoter datasets, 450 from MEME, and 1261 from YMF) were submitted to a PERL script to detect motifs detected by at least two programs. These were then merged to eliminate redundancy, allowing degeneracy at two sites. P values for over-representation of the final motifs were calculated based on a hypergeometric distribution, using Fisher's Exact test. The locations, numbers, and orientations of each motif were extracted from the datasets using custom Perl scripts, with tests for orientation bias employing Chi square. Motifs that were positionally biased were identified by counting the number of hits per 200-nt bin, extending 1-kb upstream of the start codon, and checking for deviations from a random allocation model using a 2 by 5 Fisher's Exact test.
Candidate orthologs were identified in P. capsici, P. ramorum, and P. sojae genome databases (from the Joint Genome Institute of the U. S. Department of Energy) using BLASTP. Their promoters were then extracted, and aligned with CLUSTALW using gap opening and extension penalties of 10 and 0.1, respectively, and DIALIGN using default parameters [58], [59]. After preliminary tests, P. capsici was omitted since the version of its genome assembly available at the time contained too many gaps and erroneous gene models. Putative three-way orthologs (P. infestans, P. ramorum, P. sojae) were identified for 66% of genes. Up to five genes (mean = 4.6) were typically examined for each motif. If a motif appeared in the alignment at the same position in at least one comparison it was scored as being evolutionarily conserved. A score of “ambiguous” was given for motifs found in a different location (including by searching in both orientations); on average, 8% of promoters would have a false positive. Conservation at the same site in ortholog sets for all genes was never expected: gene models in the different species often started at different locations, errors may have occurred in selecting orthologs, not all orthologs might have similar expression patterns, and promoter rearrangements are common during evolution.
Stable transformants were generated from isolate 1306 (from tomato in California, USA) using a liposome-assisted protoplast method as described [42], except that Extralyse (Laffort, Bordeaux, France) was used as the β-glucanase. Non-sporulating mycelia were obtained by inoculating clarified rye-sucrose broth with a sporangial suspension (104/ml), followed by 48 hr incubation at 18°C. Sporangia were obtained from rye-sucrose agar cultures by adding water, rubbing with a glass rod, and passing the fluid through a 50 µm mesh to remove hyphal fragments. To induce cleavage, sporangia were placed in 100 mm glass culture dishes resting on ice (internal temperature 8–10°C) for 60 min. Germinated cysts were obtained by allowing the chilled sporangia to release zoospores, to which CaCl2 was added to 0.5 mM followed by vortexing for 1 minute and incubation at 18°C for up to 9 hr. Gene expression analyses involved RNA blotting and β-glucuronidase (GUS) staining as described [21].
Constructs for testing promoters were based on pNPGUS, which is an improved version of pOGUS [60], and pNIFS-NPGUS. Each contains a promoterless GUS gene and an nptII selectable marker driven by the ham34 promoter. The improvements in pNPGUS included the addition of additional cloning sites upstream of GUS (the polylinker from pBS-KS2+) and translational stop codons upstream of the polylinker to reduce the number of cryptic transcripts with GUS activity. pNIFS-NPGUS contains a 74-nt minimal promoter from the NifS gene of P. infestans [21], [61] cloned into XmaI and EcoRI sites of the polylinker. Promoter fragments were inserted into pNPGUS or pNIFS-NPGUS as fragments amplified by polymerase chain reaction, or by ligating double-stranded oligonucleotides into the XbaI and XmaI sites of the vectors. Oligonucleotides used for cloning are listed in Table S3.
Nuclear protein isolation and EMSA were as described [22], except that heparin agarose was not used for the extractions. EMSA involved mixing 5 µg of nuclear protein with 1 µg poly dI-dC and 1.6 ng of 32P-labeled probe in buffer containing 1 mM dithiothreitol for 15 min at room temperature followed by 30 min on ice, followed by electrophoresis at room temperature on a 4.5% acrylamide gel. For competition assays, protein was incubated with unlabeled DNA for 15 min and then the labeled probe for 30 min on ice. Double-stranded oligonucleotides generated using the sequences in Table S3 were used as probe and cold (unlabeled) competitors. Mutated competitors were altered for the predicted motifs (A↔C, G↔T), and the nonspecific probe was a random sequence.
qRT-PCR employed DNAse-treated RNA, pooled from two biological replicates, which was reverse-transcribed using oligo-dT with a first-strand synthesis kit from Invitrogen (Carlsbad, CA, USA). Amplifications employed hot-start Taq polymerase with primers targeted to the 3′ regions of genes, typically yielding amplicons of 100 to 125 nt, using SYBR Green as a reporter. Reactions were performed in duplicate using the following conditions: one cycle of 95°C for 8 min, and 35 cycles of 95°C for 20 s, 55°C for 20 s, and 72°C for 30 s. Controls lacking reverse transcriptase and melting curves were used to test the data. Results were normalized based on primers for a constitutively expressed gene encoding ribosomal protein S3a, and expression was determined by the ΔΔCT method.
|
10.1371/journal.pgen.1005089 | Mutation in MRPS34 Compromises Protein Synthesis and Causes Mitochondrial Dysfunction | The evolutionary divergence of mitochondrial ribosomes from their bacterial and cytoplasmic ancestors has resulted in reduced RNA content and the acquisition of mitochondria-specific proteins. The mitochondrial ribosomal protein of the small subunit 34 (MRPS34) is a mitochondria-specific ribosomal protein found only in chordates, whose function we investigated in mice carrying a homozygous mutation in the nuclear gene encoding this protein. The Mrps34 mutation causes a significant decrease of this protein, which we show is required for the stability of the 12S rRNA, the small ribosomal subunit and actively translating ribosomes. The synthesis of all 13 mitochondrially-encoded polypeptides is compromised in the mutant mice, resulting in reduced levels of mitochondrial proteins and complexes, which leads to decreased oxygen consumption and respiratory complex activity. The Mrps34 mutation causes tissue-specific molecular changes that result in heterogeneous pathology involving alterations in fractional shortening of the heart and pronounced liver dysfunction that is exacerbated with age. The defects in mitochondrial protein synthesis in the mutant mice are caused by destabilization of the small ribosomal subunit that affects the stability of the mitochondrial ribosome with age.
| Mitochondria make most of the energy required by eukaryotic cells and therefore they are essential for their normal function and survival. Mitochondrial function is regulated by both the mitochondrial and nuclear genome. Mutations in nuclear genes encoding mitochondrial proteins lead to mitochondrial dysfunction and consequently diminished energy production, a major symptom of metabolic and mitochondrial diseases. The molecular mechanisms that regulate mitochondrial gene expression and how dysfunction of these processes causes the pathologies observed in these diseases are not well understood. Messenger RNAs encoded by mitochondrial genomes are translated on mitochondrial ribosomes that have unique structure and protein composition. Mitochondrial ribosomes are a patchwork of core proteins that share homology with prokaryotic ribosomal proteins and mitochondria-specific proteins, which can be unique to different organisms. Mitochondria-specific ribosomal proteins have key roles in disease however their functions within mitochondria are not known. Here we show that a point mutation in a mammalian-specific ribosomal protein causes mitochondrial dysfunction, heart abnormalities and progressive liver disease. This mouse provides a valuable model to elucidate the pathogenic mechanisms and progression of metabolic diseases with age, while enabling a more thorough understanding of mitochondrial ribosomes and protein synthesis.
| Mitochondria are composed of proteins encoded by the nuclear and mitochondrial genomes. Most of the mitochondrial proteins including the ribosomal proteins and translation factors that are responsible for the expression of the mitochondrial genome are synthesized on cytoplasmic ribosomes and imported into mitochondria post-translationally. In chordates, the mitochondrial genome encodes 22 tRNAs, 2 rRNAs and 11 mRNAs that are translated on mitochondrial ribosomes (mitoribosomes) into 13 polypeptides, all members of the oxidative phosphorylation complexes [1]. Mutations in mitochondrial genes or nuclear genes coding for mitochondrial proteins result in mitochondrial dysfunction and impaired energy production that cause mitochondrial diseases (reviewed in [2]). The most common cause of mitochondrial diseases are defects in the translational machinery (reviewed in [3]), however the mechanisms of mitochondrial protein synthesis are not well understood.
Mammalian mitoribosomes are 55S particles consisting of a 28S small subunit that includes the 12S rRNA and ~ 29 proteins and a 39S large subunit, which contains the 16S rRNA and ~ 48 proteins [4–7]. Mitoribosomes are distinct from their bacterial and cytoplasmic counterparts; they have reduced RNA content and an increased number of proteins [8]. The mitoribosome consists of proteins that share homology with bacterial ribosomes, some of which can have mitochondria-specific extensions, and additional, mitochondria-specific proteins that decorate the ribosomal surface, the mRNA entrance site and form some of the bridges linking the small and large subunits [6,7,9,10]. The increased ribosomal protein content does not entirely compensate for the loss of rRNA as many of the mitochondria-specific ribosomal proteins do not replace the missing RNA helices but instead have unique positions on the exterior of the mitochondrial ribosome [11,12]. Recent cryo-electron microscopy (cryo-EM) reconstructions indicate that the additional protein elements may fulfil roles necessitated by the unique features of the mitochondrial leaderless mRNAs, likely in their recognition, as well as facilitating the translation of the particularly hydrophobic proteins they encode and making the contacts between the small and large ribosomal subunits [6,7,9,10,12]. The mitoribosomes are located inside the matrix, the site of the transcriptome, and associate closely with the mitochondrial inner membrane through MRPL45 [7]. This positioning allows for co-translational insertion of the hydrophobic proteins, which are translated by the mitoribosome, into the inner membrane and their assembly into oxidative phosphorylation (OXPHOS) complexes [13].
There is very little known about the functions and roles of the mitochondria-specific ribosomal proteins, therefore characterizing their function within the ribosome and mitochondria will provide valuable insights into mitochondrial translation. The mitochondrial ribosomal protein of the small subunit 34 (MRPS34) has been identified as one of 15 mitochondria-specific proteins that are part of the small ribosomal subunit [14,15]. Although MRPS34 has been found localized to mitochondria and associated with the human homolog of the Drosophila discs large tumor suppressor protein (hDLG) [16], its role in mitochondria and protein synthesis has not been identified or characterized. Here we investigated the role of MRPS34 in mice carrying a homozygous mutation in the nuclear gene encoding this protein that causes a significant decrease of this protein. MRPS34 is required for protein synthesis of all 13 mitochondrially-encoded polypeptides and stability of the 12S rRNA and specific mRNAs. Dysfunction in the efficiency of mitochondrial protein synthesis leads to reduced mitochondrial oxygen consumption and respiratory complex activity in the mutant mice indicating that MRPS34 is essential for the stability of actively translating ribosomes. The Mrps34 mutation causes tissue-specific molecular and pathological changes that result in alterations in fractional shortening of the heart and pronounced liver steatosis that leads to fibrosis with age. Mitochondrial dysfunction caused by the Mrps34 mutation is likely caused by decreased levels of mitochondrial ribosomal subunits and translationally competent mitoribosomes.
A mouse line carrying an ENU-induced T203C point mutation in the Mrps34 gene that converts a leucine residue at position 68 to proline (Fig. 1A) was identified by whole exome sequencing [17]. Sanger sequencing shows that the Mrps34mut/mut mice are homozygous for the mutation that is absent in age and littermate matched control Mrps34wt/wt mice (Fig. 1B). The leucine residue at position 68 is conserved in vertebrates (Fig. 1A), and mutation to proline is predicted to disrupt the formation of an alpha helix. To determine whether the mutation caused changes in the abundance of the MRPS34 protein we carried out immunoblotting of mitochondrial lysates isolated from liver and heart of Mrps34 mut/mut and Mrps34wt/wt mice. The MRPS34 protein was reduced in the heart and liver of the mutant mice (Fig. 1C), indicating that the mutation causes instability of the protein. The MRPS34 protein is expressed in all examined tissues including brain, colon, heart, kidney, liver, thymus, pancreas, skin and testis and the Mrps34 mutation results in decreased levels of the protein in these tissues (S1 Fig). Mutations in nuclear genes encoding proteins that are part of the translational machinery have been shown to cause mitochondrial diseases with varying age of onset and diverse clinical pathologies that affect a range of different tissues [2,3]. Next we sought to determine the effects of the Mrps34 mutation, in young (6–8 week) and aged (28–30 week) mice, on mitochondrial function and the downstream effects on energy metabolism and disease pathology.
The mitochondrially-encoded 12S and 16S rRNAs form the scaffolds for the mitochondrial ribosomes that use 22 mitochondrial tRNAs to translate the 11 mRNAs [12]. Therefore we analyzed if the Mrps34 mutation affected the steady-state levels and stability of mitochondrial RNAs in heart and liver by northern blotting. The 12S rRNA levels were reduced in hearts and livers of the young Mrps34mut/mut mice, and the 16S rRNA level was unaffected (Fig. 2A), suggesting that the MRPS34 is required for the stability of the 12S rRNA. In addition, we observed that the levels of the mt-Nd5 mRNA were decreased in the livers, but not in the hearts of the young mutant mice (Fig. 2A), indicating that the Mrps34 mutation causes tissue specific effects on mitochondrial RNA metabolism. The 12S rRNA was also reduced significantly in the heart and livers of the mutant aged mice (Fig. 2B). Furthermore, the levels of specific mitochondrial mRNAs, mt-Co1, mt-Nd1 and mt-Nd5 were decreased in the livers of aged mice. However, in the hearts of aged Mrps34wt/wt and Mrps34mut/mut mice only mt-Nd5 was decreased but not mt-Co1 or mt-Nd1 suggesting that in the liver the stability of specific mRNAs is more severely affected by the Mrps34 mutation with age (Fig. 2A and 2B). The levels of mitochondrial tRNAs were unaffected in heart and liver mitochondria, suggesting that MRPS34 is necessary for the stability of 12S rRNA and for the stability of specific mRNAs.
Since MRPS34 is a ribosomal protein we analyzed how decreased levels of this protein, as a result of the Mrps34 mutation, affect translation by measuring de novo protein synthesis of the 13 mitochondrially-encoded polypeptides in heart and liver mitochondria from Mrps34mut/mut and Mrps34wt/wt mice. The overall decrease of mitochondrial protein synthesis in the young and aged mutant mice compared to controls (S2 Fig) indicates that MRPS34 is required for mitochondrial protein synthesis. To investigate the effects of the mutation on the rate of protein synthesis between heart and liver mitochondria we measured mitochondrial translation over time. Interestingly we observed that the initial rate of translation in control liver mitochondria is faster compared to that of heart mitochondria (Figs. 3A, S2). In addition, we observed that the Mrps34 mutation causes a decrease in mitochondrial protein synthesis (Fig. 3A), however the initial rate in translation may account for the severity of the molecular changes found in the liver compared to heart mitochondria.
Next we analyzed the effects of the Mrps34 mutation on the steady-state abundance of mitochondrial proteins by immunoblotting. The levels of the mitochondrially-encoded COXI and COXII were reduced in the livers and hearts of young Mrps34mut/mut mice compared to controls (Fig. 3C). In the aged mutant mice we observed a more pronounced decrease in COXI and COXII in both the heart and liver, suggesting that the effects of decreased MRPS34 levels on the steady state levels of these proteins are cumulative and the molecular defects have a late onset in the heart (Fig. 3D). In addition, we observed reduction in nuclear-encoded mitochondrial proteins, such as NDUFA9 and COXIV, possibly in a retrograde response to the decreased levels of mitochondrially-encoded proteins in the livers of aged Mrps34mut/mut mice (Fig. 3D).
We investigated the effects of Mrps34 mutation on the abundance of the mitochondrial respiratory complexes by blue native polyacrylamide gel electrophoresis (BN-PAGE). The reduction in the abundance of the respiratory complexes in hearts of Mrps34mut/mut mice is more apparent in the aged mice compared to the young (Fig. 4A and 4B). The reduction of the respiratory complexes was more significant in the livers of Mrps34mut/mut compared to Mrps34wt/wt mice (Fig. 4A and 4B) and this was confirmed by immunoblotting of each complex following BN-PAGE (Fig. 4C and 4D). Complexes I and IV were reduced in the hearts of both young and aged mice, whereas in the livers, Complexes I, III, IV and V were reduced (Fig. 4C and 4D). Taken together these findings provide evidence that MRPS34 is required for protein synthesis and that decreased mitochondrial translation as a result of the Mrps34 mutation can have varied effects on the abundance of mitochondrial proteins and respiratory complexes in different tissues.
Because we observed a decrease in the abundance of mitochondrial respiratory complexes we measured their enzyme activities in heart and liver mitochondria from young and aged Mrps34mut/mut and Mrps34wt/wt mice. The activities of the respiratory complexes were not significantly decreased in hearts of young Mrps34mut/mut mice (Fig. 5A), although there was a trend towards a decrease in in the activity of Complex IV, likely as a result of its decreased abundance (Fig. 4C). The activities of Complexes III and IV were significantly decreased in livers of the young mutant mice (Fig. 5B), consistent with the observed decrease in the levels of these complexes relative to those in control mice. In the hearts of aged Mrps34mut/mut mice only the activity of Complex IV was decreased (Fig. 5C), whereas the activities of both Complex III and IV were most significantly affected in the livers of these mice (Fig. 5D) as a result of reduction in the abundance of these complexes identified by immunoblotting (Fig. 4C and 4D). Measurements of oxygen consumption confirmed that mitochondrial respiratory function was affected more in the livers than hearts of young Mrps34mut/mut mice and this reduction was more dramatic with age in both tissues (Figs. 5E, 5F, and S3). Consistent with the observed molecular changes in response to the Mrps34 mutation, mitochondrial dysfunction was more pronounced in the liver compared to the heart of Mrps34mut/mut mice (Fig. 5E), suggesting that these two organs have different capacities to cope with changes in translational efficiency at different ages as observed when we measured mitochondrial translation (Fig. 3A).
Mutations in genes encoding mitochondrial ribosomal proteins have been shown to cause cardiomyopathy in patients that result in impaired mitochondrial protein synthesis and consequently reduction in the levels and activities of respiratory complexes [18,19]. The Mrps34mut/mut mice appeared similar to their Mrps34wt/wt mice littermates at birth and no developmental or fertility differences were observed compared to the control mice. However, with age the Mrps34mut/mut mice develop physiological changes that affect multiple tissues to varying extents. We observed slight reduction in vision in the Mrps34mut/mut compared to the Mrps34wt/wt mice, measured using optokinetic drum experiments, although the optic nerve was not affected (S4A–S4C, S4H Fig). The motor coordination, strength, and balance between the Mrps34wt/wt and Mrps34mut/mut mice was not significantly different (S4D–S4G Fig), although a slight difference at day 4 may suggest a potential motor learning deficit. We observed centralization of nuclei in the muscle of aged Mrps34mut/mut mice and decreased COX activity (S4I Fig). To investigate if reduced Complex IV activity, as a result of the Mrps34 mutation, affects heart function we carried out echocardiography on the Mrps34mut/mut and Mrps34wt/wt mice. We found that the hearts of mutant mice have increased fractional shortening, a thickening (hypertrophy) of the posterior wall and associated decreased oxygen consumption (Fig. 6A) that can be a common consequence of mitochondrial dysfunction. However, because we found greater mitochondrial dysfunction in the livers of Mrps34mut/mut mice we also investigated the effects of the mutation on the morphology and function of the livers in these mice compared to Mrps34wt/wt mice. Morphological examination revealed increased lipid accumulation in the livers of the young mutant mice and this was more pronounced in the aged mutant mice, suggesting that they have developed liver steatosis (Fig. 6B). Oil red O staining revealed extensive accumulation of lipid droplets in the livers of Mrps34mut/mut mice that was significantly exacerbated with age (Fig. 6C and 6D) and correlated with increased levels of alanine aminotransferase (ALT) in the serum of these mice (Fig. 6E and 6F), which is a marker of liver dysfunction, commonly associated with liver steatosis [20]. Extensive liver dysfunction can cause fibrosis, therefore we used Gomori’s trichome to stain the livers of Mrps34mut/mut and Mrps34wt/wt mice (Fig. 6B and 6C). In the young Mrps34mut/mut mice infiltration of collagen was found around portal tracts (Fig. 6B, arrows) compared to control mice, that was more pronounced in the aged Mrps34mut/mut mice, disrupting the morphology of the liver, encapsulating the tissue (arrow) and forming nodules that are markers of liver fibrosis (Fig. 6C). Our findings indicate that the Mrps34 mutation causes mitochondrial dysfunction that can affect multiple tissues, making these mice a model system to investigate how nuclear mutations can cause pathology with varying severity in different tissues.
To understand the role of MRPS34 within the mitochondrial ribosome we investigated the effects of the Mrps34 mutation on the levels of mitochondrial ribosomal proteins from the small and large subunit in Mrps34mut/mut compared to Mrps34wt/wt mice by immunoblotting. It has previously been shown that the loss of certain mitochondrial small ribosomal subunit proteins can disrupt the assembly of the small subunit and cause the loss or reduction of other small ribosomal subunit proteins [21], while loss of other small subunit proteins does not [22], potentially revealing their roles in ribosome assembly. We investigated the levels of MRPS16, MRPS25, and MRPS35 mitochondrial small ribosomal subunit proteins, and found that their levels closely paralleled those of MRPS34 (Fig. 7A), suggesting that the Mrps34 mutation causes destabilization of the small ribosomal subunit. This reduction in the small ribosomal subunit proteins is consistent with the reduced levels of the 12S rRNA (Fig. 2A and 2B). Although in the young mutant mice the levels of the large ribosomal subunit protein MRPL44 were slightly increased by the Mrps34 mutation (Fig. 1C), likely as a compensatory response to decreased small ribosomal subunit proteins (as previously observed in [23]), we observed that in the aged mutant mice the MRPL44 and MRPL23 proteins were also decreased compared to controls (Fig. 7A) both in heart and liver mitochondria. As ribosomal proteins are required in tightly regulated ratios, decrease in MRPS34 over time leads to reduction in other mitoribosomal proteins. Therefore we conclude that MRPS34 is required for the steady-state levels of small subunit ribosomal proteins, and thereby the stability of the small ribosomal subunit, which is necessary for decoding of mitochondrial mRNAs when coupled with the large subunit and consequently translation of the mitochondrially-encoded proteins.
Next we investigated if the residual levels of the mutant MRPS34 protein can interact with the mitochondrial ribosome using sucrose gradients followed by immunoblotting for this protein in heart and liver mitochondrial from the mutant and control mice (Fig. 7B). We observed that there are significantly reduced levels of the MRPS34 protein in both heart and liver mitochondria, however the mutant protein can associate with the small ribosomal subunit and constitute part of the actively translating mitoribosome.
To determine if MRPS34 is required for the stability of the mitoribosome we analyzed the mitochondrial ribosome profile of the large and small subunit, the monosome and polysome on sucrose gradients followed by immunoblotting of ribosomal marker proteins (Fig. 7C and 7D). In heart and liver mitochondria from young Mrps34mut/mut mice we observed a decrease in the presence of the small ribosomal proteins MRPS16 and MRPS35 and consequently decrease in the presence of actively translating mitochondrial ribosomes. Similarly, we observed a decrease in the actively translating mitochondrial ribosome in young heart and liver mitochondria when we immunoblotted for the large ribosomal subunit markers MRPL44 and MRPL23 and a small increase in the presence of the large subunit consistent with the slight increase in the steady state levels of these proteins (Fig. 1C), that are more apparent in heart mitochondria (Fig. 7C). The decreased levels of the small ribosomal subunit result in decreased formation of mitochondrial ribosomes.
In liver and heart mitochondria from aged mutant mice we observe decreased levels in both the small and the large ribosomal subunits and consequently reduced levels of the actively translating mitoribosomes (Fig. 7D). In the aged Mrps34mut/mut mice we found that both the ribosomal proteins from small and large mitochondrial subunits are redistributed compared to control mice and we observe reduced polysome formation in the mutant mice, indicating destabilization of mitoribosomes and polysomes with age (Fig. 7D), likely as a result of decreased MRPS34 levels (Fig. 7A and 7B). These results indicate that reduced abundance of the MRPS34 protein affects the rate of translation through destabilization of the small ribosomal subunit and the 12S rRNA that are required for mitoribosome formation.
The mitochondrial ribosome is a unique molecular machine that has been honed through evolution to cope with the compaction of the mitochondrial genome [24]. The lack of substantial untranslated regions has given rise to leaderless mitochondrial mRNAs that are somehow recognized by the mitoribosome for translation. These ribosomes have acquired additional mitochondria-specific proteins that predominantly decorate the surface of the ribosome, giving clues that these proteins may play a role in specific recognition and recruitment of mRNAs to the mitoribosome, modulation of translation according to mitochondrial environmental changes, and the exclusive translation of hydrophobic membrane proteins [7,10]. The mitochondria-specific proteins are largely uncharacterized and understanding their roles in the mitoribosome would provide insight into the unique features of mitochondrial protein synthesis that are different from cytoplasmic and prokaryotic systems. The MRPS34 protein was identified as a constituent of the small ribosomal subunit [6,25], it shares little homology to other proteins and is conserved from humans to zebrafish. Here we show that a point mutation in the mouse Mrps34 gene, causing a leucine to proline change in a conserved alpha helix, results in a significant reduction of the MRPS34 protein in heart and liver mitochondria. Consequently, mitochondrial protein synthesis is impaired in the mice homozygous for the Mrps34 mutation leading to pathologies with varying severities in different tissues.
Decreased levels of MRPS34 affect the stability of mitochondrial RNAs and this is particularly pronounced in the liver. The 12S rRNA is specifically decreased in the hearts and livers of the mutant mice, although this reduction is more profound in the liver, particularly with age, indicating that MRPS34 is required for the stability of the 12S rRNA. Although the mutation causes significant reduction of MRPS34, the remaining protein is assembled within the small subunit of the mitoribosome. However the significant decrease in 12S rRNA levels as a result of the MRPS34 reduction suggests that this protein may be required early in the assembly of the small ribosomal subunit and is likely a 12S rRNA-binding protein. We found that reduction in MRPS34 levels affects the abundance of small ribosomal subunit proteins. In the young mice the proteins of the large ribosomal subunit were present in substantial amounts, suggesting that the large subunit was still fully assembled despite the significant loss of the small subunit. This is further corroborated by the reduction in abundance of the 12S rRNA, which is readily degraded unless incorporated into a ribosomal subunit [26], and the uninterrupted presence of the 16S rRNA. This finding suggests that there is no regulatory cross-talk monitoring the levels of the mitochondrial large and small subunit proteins, instead there seems to be a compensatory mechanism that increases the level of large ribosomal proteins in response to loss of small ribosomal proteins. However this may not be long lasting as with age persistent loss of the small ribosomal subunit also leads to reduction of the large ribosomal subunit proteins and destabilization of the actively translating ribosomes in the Mrps34mut/mut mice. In the mutant mice we observed re-distribution of mitochondrial ribosomal proteins suggesting that the assembly of mitochondrial ribosomes is compromised when there are insufficient levels of MRPS34. The differential distribution of the ribosomal subunits in the mutant compared to control mice suggests that there is accumulation of ribosome assembly intermediates due to the reduced MRPS34 levels. In the young mice this re-distribution is observed solely in the small ribosomal subunit proteins indicating that the small ribosomal subunit is dissociating while the large remains stable. This may result in the formation of inaccurately assembled ribosomes, which are easily dissociated into subunits or are unable to bind to docking sites on the inner mitochondrial membrane (IMM) affecting protein synthesis and OXPHOS biogenesis. As a result of this we find that over time, in the aged mice, the stability of the mitochondrial ribosome is affected causing a reduction in polysome formation. These changes may account for the more obvious mitochondrial dysfunction in the aged mutant mice. Furthermore, as the ribosomal proteins are required in a particular stoichiometry, loss of a ribosomal protein would destabilize this balance and would result in decreased assembly of de novo mitochondrial ribosomes. This is confirmed by our data where persistent decrease in MRPS34 in the aged mutant mice leads to reduction of actively translating ribosomes as a result of a decrease in small and large ribosomal subunit proteins.
We found decreased levels of specific mRNAs, including mt-Nd1, mt-Nd5 and mt-Co1 in the livers of mutant mice, while only the mt-Nd5 mRNA was reduced in the hearts of aged mutant mice. The levels of other mRNAs such as mt-Co2 and the tRNAs were unaffected in both heart and liver, and with age. Taken together these findings reveal that MRPS34 is necessary for the stability of specific mitochondrial mRNAs and may indicate that the active translation of certain mRNAs is linked to their stabilities. The different effects on the mitochondrial mRNAs between heart and liver reflect tissue-specific changes that have been found in mitochondrial disease patients previously [18,27,28], as well as recently in a mouse model of mitochondrial disease [29]. In the heart, mitochondrial RNAs account for at least 30% of the total RNA [30] reflecting the dependence of the heart on OXPHOS. It is likely that in the heart there is excess production of mitochondrial mRNAs relative to the required threshold for normal function, such that decreases in mRNA levels do not compromise mitochondrial function in the short term. This is not the case in the liver where the decrease in mitochondrial RNAs has more profound effects on energy metabolism. Furthermore we found that MRPS34 is more abundant in the heart compared to the liver of control mice, which may protect the heart from a more pronounced decline in function in the mutant mice. We find that the levels of the other mitoribosomal proteins in control mice is higher in the heart compared to the liver, likely due to the higher mitochondrial content in the heart. Finally we observe that the liver has a faster initial rate of mitochondrial translation that is compromised in the mutant mice and may contribute to the more pronounced defect in the liver compared to the heart. It could be that after the initial burst of translation in the liver the levels of mitoribosomes or their recycling become limiting so that subsequently translation plateaus, whereas in the heart the higher abundance of mitoribosomes contributes to the steady rate of translation. Unlike the heart, the liver is highly proliferative and it may require rapid bursts of mitochondrial translation for its normal function and during regeneration.
Mitochondrial diseases caused by mutations in nuclear genes encoding mitochondrial proteins can affect the function of many different tissues with varying severity (reviewed in [2,3]). Furthermore, nuclear mutations in mitochondrial proteins can result in remarkably heterogeneous defects with varying severity in different tissues that are poorly understood currently [2]. We observed similar varying defects in different tissues in the Mrps34mut/mut mice, which have a fractional shortening of their hearts, centralized nuclei in their muscles and increased lipid accumulation in their livers that causes liver steatosis. Hepatopathies have been identified along with a range of different symptoms in many mitochondrial disorders, although hepatopathy was the main consequence of a mutation in the tRNA 5-methylaminomethyl-2-thiouridylate methyltransferase (TRMU) gene that produces the enzyme responsible for 2-thiouridylation of the tRNAGlu, tRNAGln and tRNALys, causing a severe but reversible infantile hepatopathy [31–33]; the mutation in the Mrps34 gene is the first to show pronounced liver defects in mice providing the means to investigate the contribution of mitochondrial function to hepathopathy in the future. Many mutations in nuclear genes that encode protein components of the translational machinery result in compromised biogenesis of specific or all mitochondrial respiratory chain complexes and lead to decreased OXPHOS and some of these have been shown to cause accumulation of lipids in hepatocytes [3,34]. Although we observe overall decrease in mitochondrial protein synthesis as a result of the Mrps34 mutation, the greatest reduction is in the activity of Complex IV in both heart and liver, and this reduction is greater in the liver. Decrease in the mt-Co1 mRNA and consequently the COXI protein, that is necessary for the biogenesis of Complex IV [35] likely contributes to this pronounced decrease in its abundance and activity.
Our work has shown that MRPS34 plays a role in maintaining the stability of the small ribosomal subunit and the 12S rRNA, which are necessary formation of actively translating mitoribosomes. In addition, MRPS34 is required for the stability of specific mRNAs, indicating that the mitochondria-specific ribosomal proteins might have unique roles in mitochondrial RNA metabolism. Because the Mrps34 mutation in mice is not embryonic lethal it has provided a means to investigate how mitochondrial dysfunction can lead to disease in the whole body and the progression of the disease with age. Establishment of mouse models where mitochondrial dysfunction causes disease are particularly important for understanding the causes of human disease pathology, identification of drug targets and for the development of future treatments for these diseases.
Male age- and litter-mate matched (6–8 weeks ‘young’ and 30 weeks ‘aged’) wild-type (Mrps34wt/wt) and homozygous (Mrps34mut/mut) ENU mutant mice on a C57BL/6J background were obtained from the Australian Phenomics Facility. The Mrps34 mice were bred onto a C57BL/6J background for 8–10 generations. Animals were singly housed in standard cages (45 cm × 29 cm × 12 cm) under a 12-h light/dark schedule (lights on 7 a.m. to 7 p.m.) in controlled environmental conditions of 22 + 2°C and 50 + 10% relative humidity. Normal chow diet (Rat & Mouse Chow, Speciality Foods, Glen Forrest, Western Australia) and water were provided ad libitum. The study was approved by the Animal Ethics Committee of the UWA (AEC 03/100/526) and performed in accordance with Principles of Laboratory Care (NHMRC Australian code for the care and use of animals for scientific purposes, 8th Edition 2013). Behavioral and motor tests are described in Supplemental Methods (S1 Text).
Tissues were homogenised in 100 μl of 100 mM Tris, 2 mM Na3VO4, 100 mM NaCl, 1% Triton X-100, 1 mM EDTA, 10% Glycerol, 1 mM EGTA, 0.1% SDS, 1 mM NaF, 0.5% deoxycholate, 20 mM Na4P2O7, pH 7.4 containing PhosSTOP Phosphatase Inhibitor Cocktail and EDTA-free Complete protease inhibitor cocktail and the supernatant was collected after centrifugation at 10,000 g.
Mitochondria were isolated from homogenized hearts and livers and isolated by differential centrifugation as described previously [36] with some modifications. Livers were homogenized in buffer containing 250 mM sucrose, 5 mM Tris, 1mM EGTA, pH 7.4 with EDTA-free Complete protease inhibitor cocktail (Roche) and hearts were homogenized in 210 mM mannitol, 70 mM sucrose, 10 mM Tris, 0.1mM EDTA pH 7.4 containing EDTA-free Complete protease inhibitor cocktail.
Liver (1.2 mg of protein) and heart (0.8 mg of protein) mitochondria were lysed with 1% n-Dodecyl β-D-maltoside in 10mM Tris-HCl, pH 7.4, 260 mM sucrose, 100 mM KCl, 20 mM MgCl2 in the presence of RNasin and protease inhibitors for 30 min, the lysate centrifuged at 10,000 g for 45 min at 4°C, the clarified lysate was loaded on a continuous 10–30% sucrose gradient (in 10 mM Tris-HCl, pH 7.4, 100 mM KCl, 20 mM MgCl2 in the presence of RNasin and protease inhibitors) and centrifuged at 71,000 g in an Optima Beckman Coulter preparative ultracentrifuge as described before [6]. Fractions were collected and precipitated with 30% trichloroacetic acid, washed in acetone, and the entire fraction was resolved by SDS-PAGE. Protein markers of the mitochondrial ribosomal subunits were detected by immunoblotting as described below.
RNA was isolated from heart and liver mitochondria using the miRNeasy Mini kit (Qiagen) incorporating an on-column RNase-free DNase digestion to remove all DNA. RNA (5 μg) was resolved on 1.2% agarose formaldehyde gels, then transferred to 0.45 μm Hybond-N+ nitrocellulose membrane (GE Lifesciences) and hybridized with biotinylated oligonucleotide probes specific to mouse mitochondrial tRNAs, mRNAs and rRNAs. The hybridizations were carried out overnight at 50°C in 5x SSC, 20 mM Na2HPO4, 7% SDS and 100 μg.ml-1 heparin, followed by washing. The signal was detected using streptavidin-linked infrared antibody (diluted 1: 2,000 in 3x SSC, 5% SDS, 25 mM Na2HPO4, pH 7.5) using an Odyssey Infrared Imaging System (Li-Cor).
Specific proteins were detected using rabbit polyclonal antibodies against: MRPL44, MRPL23, MRPS16, MRPS35, MRPS25 (Proteintech, diluted 1:1000), MRPS34 (Sigma, diluted 1:1000) and mouse monoclonal antibodies against: porin, NDUFA9, Complex II, Complex III, COXI, COXII, COXIV and Complex V subunit (Abcam, diluted 1:1000), in Odyssey Blocking Buffer (Li-Cor). IR Dye 800CW Goat Anti-Rabbit IgG or IRDye 680LT Goat Anti-Mouse IgG (Li-Cor) secondary antibodies were used and the immunoblots were visualized using an Odyssey Infrared Imaging System (Li-Cor). Tissue specific immunoblotting analysis was performed on a Proteintech mouse tissue blot (Cat. No. M10005).
Mitochondrial de novo protein synthesis was analyzed using Expres35S Protein Labelling Mix [35S] (14 mCi, Perkin–Elmer) as described before [23].
Liver and heart mitochondrial lysates were resolved by BN-PAGE to detect the respiratory complexes by Coomassie staining as described previously [8,37] or by immunoblotting as described above.
The enzyme activities of all five respiratory complexes and citrate synthase were measured in a 1 ml cuvette at 30°C using a Perkin Elmer lambda 35 dual beam spectrophotometer as described in [38].
Mitochondrial respiration was evaluated as O2 consumption in isolated heart and liver mitochondria according to Kuznetsov et al. Mitochondria were supplemented with substrates for either complex I (10 mM glutamate/malate, Sigma), II (10 mM succinate, Sigma) or III (1 mM TMPD/1 mM ascorbate, Sigma). After addition of 1 mM adenosine diphosphate (ADP, Sigma) to the recording chamber, State 3 respiration activity was measured. ADP independent respiration activity (State 4) was monitored after addition of oligomycin (2 μg/ml, Sigma).
Echocardiographic studies to measure left ventricular function were performed on mice under light methoxyflurane anesthesia with the use of an i13L probe on a Vivid 7 Dimension (GE Healthcare). Echocardiographic measurements were taken on M-mode in triplicate from each mouse and the quantitative measurements represent the average. M-mode recordings were made at a sweep speed of 200 mm/s. Measurements of left ventricular end diastolic diameter (LVEDD), left ventricular end systolic diameter (LVESD), fractional shortening (FS), left ventricular posterior wall in diastole (LVDPW), left ventricular posterior wall in systole (LVSPW), intraventricular septum in diastole (IVDS), and intraventricular septum in systole (IVSS) were made. Fractional shortening (FS) was calculated by the formula [(LVEDD-LVESD)/EDD] x 100.
Fresh sections of the liver and muscle were frozen in Optimal Cutting Temperature (OCT) medium, sectioned and stained with Haematoxylin and Eosin, Gomori’s Trichrome and Haematoxylin, Oil Red O and Haematoxylin, COX or NADH stains. Images were acquired using a Nikon Ti Eclipse inverted microscope using a Nikon 40x objective and Oil Red O staining was quantified as described previously [39].
|
10.1371/journal.pcbi.1005837 | A cyber-linked undergraduate research experience in computational biomolecular structure prediction and design | Computational biology is an interdisciplinary field, and many computational biology research projects involve distributed teams of scientists. To accomplish their work, these teams must overcome both disciplinary and geographic barriers. Introducing new training paradigms is one way to facilitate research progress in computational biology. Here, we describe a new undergraduate program in biomolecular structure prediction and design in which students conduct research at labs located at geographically-distributed institutions while remaining connected through an online community. This 10-week summer program begins with one week of training on computational biology methods development, transitions to eight weeks of research, and culminates in one week at the Rosetta annual conference. To date, two cohorts of students have participated, tackling research topics including vaccine design, enzyme design, protein-based materials, glycoprotein modeling, crowd-sourced science, RNA processing, hydrogen bond networks, and amyloid formation. Students in the program report outcomes comparable to students who participate in similar in-person programs. These outcomes include the development of a sense of community and increases in their scientific self-efficacy, scientific identity, and science values, all predictors of continuing in a science research career. Furthermore, the program attracted students from diverse backgrounds, which demonstrates the potential of this approach to broaden the participation of young scientists from backgrounds traditionally underrepresented in computational biology.
| Computational biology research is frequently conducted by virtual teams: groups of scientists in different locations that use shared resources and online communication tools to collaborate on a problem. It is imperative that the next generation of computational biologists can easily work in these interdisciplinary, distributed settings. However, most undergraduate research training programs are hosted by a single institution. In this report, we describe a new summer undergraduate research program in which students conduct biomolecular modeling research with the Rosetta software in research groups around the world. The students each conducted their own research project in a university-based group while collaborating with other students and members of the Rosetta Commons at a distance using everyday tools such as Slack, Skype, GitHub, and Google Hangouts. When compared with in-person summer research training programs, students report similar or even improved outcomes, including the development of a sense of community and increases in their scientific self-efficacy, scientific identity, and science values. Furthermore, our program attracts a diverse group of students and thus has the potential to help broaden participation in computational biology.
| Computational biology is an interdisciplinary field, and many computational biology research projects are performed by distributed international teams of scientists. In the coming decade, it will be imperative for computational biologists to collaborate within these virtual communities [1,2]. However, few undergraduate programs expose students to a distributed research environment. Introducing new training paradigms is one way to facilitate research progress in computational biology. In this work, we describe the Rosetta Research Experience for Undergraduates (REU), a program in biomolecular structure prediction and design in which students conduct research in a distributed environment. We detail the structure of the program designed to expose students to a virtual community and describe student research experiences from the first two cohorts.
Undergraduate research experiences are important avenues for recruiting and preparing the next generation of scientists [3]. Hands-on lab experiences encourage creativity and expose students to problem-solving frameworks [4]. Students who spend significant time in the lab learn to perform new techniques, collect data, interpret findings, and formulate new research questions [5,6]. Lab experiences can shape students’ perceptions about careers in research [7]. Through undergraduate research experiences, students gain access to professional mentors who provide career support needed to retain a diverse group of students in science and engineering. Undergraduate research can also serve as an introduction to fields such as computational biology, which are not well represented in undergraduate degree programs or courses, especially at institutions that serve large proportions of students from underrepresented backgrounds.
In the United States, REU sites are funded by the US National Science Foundation (NSF) and serve as a major mechanism for involving undergraduates in science research. Most REU sites offer 10-week summer programs designed to engage 8–10 undergraduates in meaningful research [8] and to recruit students, especially those from underrepresented backgrounds, into graduate education and research-related careers [9]. Students participate in hands-on lab or field research experiences, complemented by journal clubs, sessions for writing and presentation peer review, and information sessions about graduate education and research-related career options. In general, REU sites are hosted by a single department, program, center, or institution.
This REU structure is inherently limiting for computational biology because computational biology research is performed by geographically-distributed teams of scientists with varied academic backgrounds ranging from mathematics and computer science to cellular and molecular biology. In addition, scientific projects depend on shared computing resources, data sets, and codebases. To be successful in computational biology, students need to develop interdisciplinary research skills such as the ability to formulate integrative research questions and communicate with researchers in other fields [10]. These distinctions require rethinking how to structure REUs to meet the unique needs and challenges of computational biology.
We created a new REU program within the Rosetta Commons, a group formed to enable close collaboration between 52 (and growing) labs developing the Rosetta software suite for biomolecular structure prediction and design. The Rosetta Commons labs are united by a set of core challenges, including (1) sampling macromolecular conformational space, (2) improving energy functions, (3) utilizing advanced computing resources, (4) improving code organization and algorithm efficiency, and (5) disseminating the tools to academic and industry labs. To tackle these challenges, community developers from a broad range of fields have contributed tens of thousands of revisions to the master version of Rosetta from their development branches. Collaborating scientists have tackled a wide range of science and engineering challenges, from RNA folding [11] to the refinement of structures using NMR data [12] to designed proteins [13,14], interfaces [15–17], protein nanomaterials [18,19], mineral binders [20], and antibodies [21,22]. The public has also engaged in Rosetta-mediated science through the Berkeley Open Infrastructure for Network Computing (BOINC)-distributed computing platform [23] and game-playing applications such as Foldit [24].
The Rosetta collaboration is an appropriate environment for a geographically-distributed computational biology REU for two key reasons. First, the problem-solving approaches are highly interdisciplinary. For instance, X-ray crystallography and NMR were originally developed in physics and chemistry, and sequencing and protein expression originated in biology. Second, labs at different institutions are already connected by online communication tools. In particular, the GitHub code-sharing platform [25], Slack team messaging [26], and an in-house benchmarking server allow developers to work on a common source in their own branch, request code review, tag collaborators, comment on developments, and easily share their work.
In this report, we describe the implementation and evaluation of the Rosetta biomolecular modeling REU, the first REU situated within a globally distributed scientific community. We describe our strategies for recruiting a diverse cohort of students and explain the implementation of the three program phases: (1) one week of intensive, hands-on learning about computational methods development, (2) eight weeks of research at different Rosetta labs, and (3) one week at the Rosetta annual conference. We discuss strategies we used to keep students connected while they conducted their research. We describe early evaluations of the program and student outcomes. Finally, we discuss the program goals as they align with grand challenges in undergraduate science education, and we postulate next developments therein.
A primary goal of the Rosetta REU was to attract and retain underrepresented groups in computational science, chemistry, engineering, and the biosciences. We took a two-pronged approach to recruit a diverse cohort. First, we promoted the program via email to several organizations, including the Society of Women Engineers (SWE), Hispanic Association of Colleges and Universities (HACU), the Society of Hispanic Professional Engineers (SHPE), the National Society of Black Engineers (NSBE), and the American Indian Science and Engineering Society (AISES). We reached out via email to local universities with diverse populations. We also partnered with diversity programs, including Minority Access to Research Careers (MARC) and the Leadership Alliance, by asking them to distribute the program information and recommend potential participants.
Second, we reached out to attendees at two affinity group conferences. For the last three years, we have sent a delegation of two faculty plus 6 to 10 female scientists from multiple Rosetta labs to the Grace Hopper Celebration of Women in Computing. The two faculty led a Student Opportunity Lab round-table to present “Computational Molecular Biophysics: Design Your Future.” In addition, the delegation hosted a booth at the career exposition with demonstrations and information. At this event, we collected over 40 resumes annually and eventually recruited three students through this outreach. We recently replicated this effort with an initiative to minority students by attending the Annual Biomedical Research Conference for Minority Students (ABRCMS). At the conference, we collected between 40 and 60 resumes and followed up with these students, encouraging them to apply for the program via email, eventually enrolling one program participant.
The program was open to all undergraduate science, mathematics, and engineering students who had not graduated before the summer session. To apply, students submitted an online application that included a personal statement, summary of research and computing experience, resume, transcript, lab assignment preferences, and contacts for three reference letters. In the personal statement, students were asked to explain why they are interested in the REU program and how the projects fit with their interests and talents. The experience statement required students to summarize their academic achievements, special skills, academic honors, and other creative work.
We sought both computer science majors with no previous biology experience and life science majors with wet lab experience but limited computational background. Previous experience was not required but preferred, to increase the likelihood of student success in the program. The applications were evaluated by a panel of two professors and two graduate students. The criteria for evaluating applications are detailed in S1 File. After selection, we contacted students to confirm their interest, and then we asked the student and the assigned faculty to meet via Skype to discuss project ideas and again confirm their interest in working together.
To provide students with a foundation in computational methods development, we initiated the program with one week of hands-on practice at Rosetta Boot Camp [27]. Rosetta Boot Camp is an in-person workshop designed to teach software development skills and Rosetta3 library [28] concepts to new graduate students and postdoctoral fellows. We adapted this workshop for undergraduates by emphasizing skills not taught in traditional courses yet necessary to begin research. We also structured the boot camp to achieve a 4:1 student-to-teacher ratio and to promote collaboration between students. A set of detailed learning objectives is listed in S1 File.
To achieve the learning objectives, students participated in a combination of lecture and lab activities. First, interactive lectures were used to introduce concepts (Table 1). Then, students collaboratively worked on two types of activities (Table 2). The first set focused on skills needed to write, test, debug, and version-control code. The second set (marked by an asterisk in Table 2) walked students through the creation of a complex conformational sampling protocol. In the first lab, they wrote an application to perturb and minimize a structure using core Rosetta modules. In subsequent labs, they refined this protocol to more carefully control how perturbation propagated through the structure, dividing structures by secondary structure elements, and eventually incorporating the cyclic-coordinate-descent (CCD) loop-closure algorithm [29] to improve the likelihood that perturbations would result in low-energy conformations. They connected their protocol to the job-distributor machinery in Rosetta and to RosettaScripts: two parts of Rosetta that many students would work with during their internships (Fig 1).
The workshop was led by a primary instructor and two student teaching assistants, including alumni of the program and a student volunteer from the Rosetta community. Students prepared by completing readings and short C++ homework assignments. During the week, students worked in groups on the lab activities to encourage sharing of complementary knowledge. This was crucial because both cohorts comprised students with diverse academic backgrounds. Finally, we assessed the students’ progress through code review, short-answer concept tests, and assignment completion.
Over the next eight weeks, each student conducted a research project in one of the 52 Rosetta Commons labs, typically under the supervision of a senior graduate or postdoctoral researcher in the lab. The students remained connected with each other and other participating research groups through several channels discussed below.
Each summer, the Rosetta Commons members convene to discuss the newest science to emerge from the collaboration. This meeting is held in Washington state and involves about 250 people from the 52 Rosetta labs, plus invited speakers. The first two days are held on the University of Washington campus and are meant to facilitate discussion on software and ongoing technical challenges. The following three days occur at the Sleeping Lady Conference Center in Leavenworth, Washington, and consist of scientific presentations, small group discussion, posters, and leadership and team meetings.
Students attended the full conference, which allowed them to reconnect with one another in person, network with other researchers at the conference, and learn about the wider field of computational biology. Each student presented a poster of their research accomplishments and received feedback on their work. Finally, we held a debriefing session for the cohort in which we solicited feedback about the program.
We hosted eight interns during the summer of 2015 and eight interns during the summer of 2016 in 14 different Rosetta Commons labs. We also educated a diverse cohort of students: across both cohorts, 63% of students were female, 13% were African American, and 13% were Hispanic. The students conducted a diverse set of scientific projects described in Table 3.
Rosetta REU students have already shared their work with the scientific community in the format of formal presentations and publications. All students shared the outcomes of their scientific projects at the Rosetta conference. Two students have presented their work at other scientific meetings, and one student is an author on a conference paper [31,32]. In addition, two students contributed code to the main Rosetta repository; their contributions are already being distributed to end users. These scientific deliverables demonstrate that students can conduct high-level research projects in the eight-week time span.
Informally, we observed that the interns helped to advance the research of the host lab. For example, one intern used a newly developed framework for modeling protein glycosylation [33] to create models of antibody constant regions with different mutations and glycosylations that affect binding to antibody receptors and immune stimulation [31]; this work continues in the host lab and has enabled new collaborations with experimental labs. Another intern examined the computer–human interface for the protein-folding game FoldIt [24] to measure how three-star rating systems affect game player persistence [32]. One student designed co-assembling multicomponent protein crystals, and the host lab invited him back for a second summer to continue the research.
Most of the students who participated in the REU program are now pursuing careers in science. Of the twelve alumni who have completed their BS degree, six students are now PhD candidates in fields ranging from chemical engineering to computer science and molecular biology. Two are working in the pharmaceutical industry, one is working in an academic research lab, and one is working as a high school mathematics teacher. One is currently applying to medical school, and three from the 2016 cohort are currently applying to graduate school (as of fall 2017).
To evaluate our virtual REU model, we surveyed both cohorts of students at the end of each summer about their sense of community, scientific self-efficacy, scientific identity, and the extent to which their personal values aligned with scientific values [34–36]. These outcomes are indicators of the students’ integration into their scientific community and predictors of their likelihood to continue in science research–related career paths, especially for students from backgrounds traditionally underrepresented in the sciences [35]. We compared the responses of our students with responses from students in two in-person, computational life science REU programs.
Post-program survey data (Fig 2) show that both cohorts matched the “sense of community” of other programs. Interview comments reinforce the strong community even across distributed virtually-linked labs (see S1 File). Similarly, the data revealed that our program matched outcomes for scientific self-efficacy, scientific identity, scientific values alignment, and their intentions to pursue a science research–related career.
In this report, we presented a summer research experience that involves undergraduates in distributed computational biology research. We also attracted a diverse cohort, demonstrating the potential of this approach to broaden participation by students from traditionally underrepresented backgrounds. After the first two cohorts, we pooled our experiences to identify strengths and weaknesses in the program. Here, we elaborate on these takeaways and recommend directions for improvement.
A primary challenge of our program was teaching students with varied academic backgrounds. Most undergraduate science programs do not include quantitative courses beyond prerequisite calculus [37]. Furthermore, computational biology degree programs are still new [38] and seldom available at institutions that primarily serve students from underrepresented backgrounds. Therefore, we anticipated that students would vary in their preparation to do computational work.
At boot camp, we prepared to support students with a high instructor-to-student ratio (1:4). We also arranged the students around a conference table intended to facilitate collaboration while working on lab activities. One hurdle was teaching the Unix command line because half of the students had no prior experience. This knowledge is critical because most molecular modeling programs are controlled from the command line. Initially, we tried to pair students with and without experience. However, we found that the more experienced student felt held back. In the future, we plan to include more Unix preparation in the homework preceding boot camp. We also hope to integrate strategies that encourage patience when working in teams with mixed backgrounds.
For future work, we also plan to further develop the boot camp learning objectives (see S1 File). Undergraduate boot camp was derived from a workshop intended for new graduate students and postdoctoral fellows. Thus, the week is packed with technical details about C++ language features and the mathematics underlying Rosetta algorithms. However, we postulate that skills required for an eight-week internship may differ. For instance, students are more likely to apply the tools and analyze results rather than develop new protocols from scratch. Furthermore, undergraduates may benefit from developing more transferable skills. In the future, we plan to revisit the objectives and potentially rebalance toward more general computational biology skills rather than those specific to Rosetta.
The Rosetta REU program is a “proof of principle” example that undergraduates can perform research in a distributed setting. We found that students made strong connections within the cohort that matured into an internal collaboration network during the eight-week research period. A few students even contributed code and commented on ongoing projects via the GitHub [25] code-sharing platform. All these findings are reinforced by survey reports that students experienced a strong sense of community.
Forming strong bonds between students is a top priority of the program. As the program continues, we are aiming to help mentors better guide and connect with their students during the eight-week research period by drawing more from evidence-based mentoring practices [39–41], and we want students to leverage weak ties [42] in the Rosetta community. Students were given access to several collaboration tools, including the Slack [26] channel and developer mailing lists. However, we observed that the students used these tools sparingly. In scientific communities, weak ties are critical because reaching out of one’s inner network increases the probability that knowledge transfers are more novel. One possibility of encouraging students would be to scaffold using community resources during boot camp rather than introducing them at the end. This way, students can begin using the tools under instructor guidance, gain confidence, and then apply them.
Another goal of the Rosetta REU program was to foster an inclusive culture. Diversity is critical to the creativity and productivity of teams [43]; however, recruiting a diverse cohort remains a challenge, especially in computer science and mathematics [44]. To address this goal, we attended affinity conferences and reached out to affinity groups, thereby adding more applications to our pool. Sending student and faculty representatives to these conferences also allowed our students and faculty to learn strategies to confront the confidence gap [45] and unconscious bias [46]. Overall, this also increases awareness of these issues not only within our small group but also amongst the larger Rosetta community.
We postulate that the diversity of the REU cohort also contributed to the strong sense of community. In addition, our recruiting efforts at Grace Hopper and ABRCMS strengthened our community of women in the Rosetta Commons, and by rotating the attending faculty, more received education and awareness of gender issues in the field. Upon returning to the labs, these conference delegates have sparked other diversity efforts, including broader conference activities, Lean In Circles [47], and monitoring of conference speaker diversity. In the future, we will continue to engage in affinity conferences and take home new practices for fostering and encouraging diversity and inclusiveness in virtual cohorts and the Rosetta community overall.
|
10.1371/journal.pntd.0000574 | Prevalence and Spatial Distribution of Entamoeba histolytica/dispar and Giardia lamblia among Schoolchildren in Agboville Area (Côte d'Ivoire) | New efforts are being made to improve understanding of the epidemiology of the helminths and intensifying the control efforts against these parasites. In contrast, relatively few studies are being carried out in this direction for the intestinal protozoa. To contribute to a better comprehension of the epidemiology of the intestinal protozoa, prevalence, and spatial distribution of Entamoeba histolytica/dispar and Giardia lamblia, and their association with drinking water supplies, were determined in the Agboville department in southeast Côte d'Ivoire.
Stool samples were taken from more than 1,300 schoolchildren in the third year of primary education (CE1) from 30 primary schools and preserved in SAF (sodium acetate-acetic acid-formalin). The samples were analyzed by formalin-ether concentration. Then, a survey questionnaire addressed to schoolchildren and school directors was used to collect data on water supplies. Prevalence of E. histolytica/dispar and G. lamblia were, respectively, 18.8% and 13.9%. No particular focus zone was observed in the spatial distribution of the two species. Significant negative association was observed between use of tap water and high prevalence of E. histolytica/dispar infection (OR = 0.83, p = 0.01). High prevalence of G. lamblia infection was positively associated with use of ponds as the source of drinking water (OR = 1.28, p = 0.009).
These two species of pathogenic protozoa are present with substantial prevalence in this area of Côte d'Ivoire. Although their spatial distribution is not focused in any one place, determination of the population segments with the highest levels of infection will help to target the chemotherapeutic fight. To reinforce treatment with chemotherapeutic agents, tap water should be made available in all the localities of this area.
| According to WHO, intestinal amoebiasis caused by Entamoeba histolytica is the third principal parasitic disease responsible for mortality in the world. This protozoal parasite infects approximately 180 million individuals throughout the world, among whom 40 to 110 thousand die from it each year. Giardiasis, caused by another protozoan parasite, Giardia lamblia, infects approximately 200 million individuals throughout the world, is a frequent cause of diarrhea in children, and can have negative impact on growth and development. Unfortunately, these intestinal protozoa are taken into account in few epidemiologic studies. The investigation we carried out to determine prevalence and spatial distribution of these infections shows the importance of these parasites in the Agboville department in southeast Cote d'Ivoire. Determination of spatial distribution of these parasites will help to focus delivery of chemotherapy in this area. In addition, our description of the relation of sources of drinking water with these parasitic infections will contribute to the development of an integrated treatment program for these parasites in this area of Côte d'Ivoire. This work will help make the population and political powers aware of the importance of these parasites and the need for safe drinking water in all localities of this area.
| Although intestinal parasites seem to raise much less interest than do AIDS and tuberculosis, they are a major public health problem in tropical regions [1]. In 2002, WHO estimated the number of people infected by digestive tract parasites at 3.5 billion and the number of people made ill by them at 450 million [2]. Whereas much effort is being made toward a better comprehension of helminth epidemiology [3],[4], relatively few equivalent studies are done on intestinal protozoa. This is surprising, because intestinal amebiasis caused by the protozoan Entamoeba histolytica is the third-greatest parasitic disease responsible for death in the world after malaria and schistosomiasis [5],[6]. It affects approximately 180 million people, of whom 40,000 to 110,000 die each year [7]. Giardiasis, caused by Giardia lamblia, is a frequent cause of diarrhea [8],[9] that can have a negative impact on growth and development of children [10] and affects approximately 200 million people worldwide [11]. These parasitic diseases are found in all the major regions of Africa [12]–[14] and were reported in Côte d'Ivoire by surveys carried out in the west of the country [15]–[18]. Giardia cysts were reported in an investigation on an epidemic of diarrhea that occurred in the village of Offoumpo in Agboville area [19]. Other studies in this area also reported a high prevalence of certain protozoal species such as E. histolytica [20]. In the same area, N'Guessan et al. found that the very high rate of blood in feces is associated with intestinal schistosomiasis [21]; these authors thought that blood in feces could also be due to other diseases such as amoebiasis. A parasitological survey should help to establish the existence of amoebiasis and assess the probable contribution of E. histolytica to the occurrence of fecal blood in the Agboville area.
Treatment of giardiasis and intestinal amoebiasis relies on derivatives 5-nitro-imidazoles such as the metronidazole, marketed since 1959 [22]. To date, some resistant cases of G. lamblia to these products have been reported. Unfortunately, no new drug is under development for specific treatment of intestinal protozoa [22]. In order to reduce or delay development of resistance, certain authors recommend avoidance of mass treatments in favor of targeted treatments and greater effort put into prevention [22],[23]. Collection of epidemiological data is necessary to develop fight effective strategies against these parasites.
The main objective of this study was to estimate the prevalence of intestinal protozoa in the feces of schoolchildren in the Agboville area. The secondary objectives were to establish spatial distribution of E. histolytica and G. lamblia in this area and to determine the relationship between these parasites and household water sources. The results should facilitate evaluation of the endemic level of these parasites and to know if infection risk is focused in an area or is widely spread, and consequently whether massive or focal measures of parasite control are required.
The study was carried out in the Agboville area, southeast Côte d'Ivoire (3°55′ and 4°40′ West and 5°35′ and 6°15′ North). The area is rugged and consists of numerous valleys with swamps. It is a forested region and the climate is of equatorial type with two rainy seasons and two dry seasons. Its average annual rainfall is between 1,298 and 1,739 mm with temperature ranging between 25 and 26.6°C [24]. This zone covered by a dense hydrographic network made up of two rivers (Agnéby and Mé). The tributaries and streams are numerous and conduct water to some villages; there are also many isolated rivers. The Agboville area has 103 villages and the population is estimated at 244,865 people, most of whom are farmers. The main crops are cocoa, coffee, and food products as in west of the county. Populations are supplied with water by traditional dug wells, boreholes, taps (which are supplied by wells or public water delivery systems), rivers, and ponds.
The study population consisted of schoolchildren. The study presented here used two surveys: first, a comprehensive parasitological survey in all the primary schools of one education inspection in the Agboville area that fulfilled our inclusion criteria (i.e., that they were registered in one of the schools of the Agboville inspection); second, a questionnaire survey to collect data on water sources.
Institutional approval of the study protocol was granted by Abidjan-Cocody University (IRB 09-2003). The study received ethical clearance by the Ministry of Public Health in Côte d'Ivoire. Then we obtained the oral consent of teachers and parents of pupils according to the principles of the Declaration of Helsinki, before beginning the data collection. The consent was oral because the majority of the parents cannot read nor write. Documentation of this oral consent was initialed and dated by the examiner according to data collection forms approved by the IRB. It was also approved by the organization of parents of pupils. Participation of pupils was voluntary. Those who refused to give fecal samples or to answer the questionnaire were simply excluded from the study. At the end of the parasitological survey all schoolchildren were treated without cost with albendazole for soil-transmitted helminth infections and intestinal protozoa.
The list of all 30 primary schools in the area was provided by the inspector. Then, in the last week of October 2004, he explained the aim and the procedures of the study to the school directors and requested class lists with name, age, and sex of each pupil. Sampling was done from all voluntary schoolchildren of the third grade class of the inspection during the last two weeks of November 2004. After the children were given an explanation of the stool sample process, they received plastic, covered 125 ml transparent tubes into which they placed their samples. Tubes were given labels to identify the sample, then placed in racks and transported to the laboratory of the major diseases of Agboville (the state-run laboratoire des grandes endémies d'Agboville). To preserve the samples, 1 to 1.5 g of stool was placed (by wooden stick) in another tube containing 10 ml of sodium acetate–acetic acid–formalin (SAF) solution carrying the same label as the corresponding tube. Thereafter, tubes were shaken vigorously to mix feces and SAF [25],[26]. SAF tubes were transported to the laboratory in Abidjan, where stool samples were analyzed by formol-ether concentration [27] and examined by microscopy. All the species of intestinal protozoa and helminths observed were recorded. Slides were read semiquantitatively for intestinal protozoa (1+, 2+, 3+ according to parasitic load of microscopic fields) and quantitatively for helminths (eggs were counted systematically).
Two weeks before the parasitological survey, the questionnaire was distributed in all 30 schools. This questionnaire was used in previous studies in Côte d'Ivoire [28],[29]. It takes into account other aspects but only data on water supply sources were considered for this study. It consists of two parts, one sent to school directors and the other to teachers. The part sent to school directors was filled in by them. The part sent to teachers was used to collect data from the students. The teachers followed the instructions that accompanied the questionnaire and interviewed pupils separately, one after another, in an empty classroom to avoid the influence of others on their responses. Answers to the questions were “O” for “yes”, “N” for “no” and “—“ for “I do not know.” Completed questionnaires were collected during the parasitological investigations.
During the parasitological investigation, geographical coordinates and altitude of each school were recorded with GPS (global positioning system; Magellan 315, Thales Navigation, Santa Clara, California, United States); then information on roads were identified on-site and those on rivers were observed on maps. Data were used to develop geo-referenced files of the Agboville area from existing maps, ArcView (Redlands, California, United States), and MapInfo. The prevalence of parasites (by species) was then incorporated into the digital map.
Data were double entered and validated with EpiInfo 2002 (US Centers for Disease Control and Prevention, Atlanta, Georgia, United States). Two age groups, 6–10 and 11–12 years, were performed. Chi-square (χ2) tests were conducted with STATISTICA 6.0 (StatSoft, Data Analysis Software System, Tulsa, Oklahoma, United States), to determine the relationship between parasites and the children's age and sex with a confidence interval (CI) of 95%. The relationship between the prevalence of different species of parasites was evaluated by the Pearson correlation coefficient (r) and its significance (p-value) by linear regression carried out with STATISTICA 6.0. Associations between parasite prevalence and water supply sources were examined by logistic regression conducted with STATA 9.0 (Stata, College Station, Texas, United States).
All 30 schools of the Agboville inspection participated in the study. Out of 1,500 schoolchildren who were registered on the class lists, 89 did not provide stool samples (27 were absent during the study and 62 refused to participate) and 37 did not complete the questionnaire (Figure 1). Consequently, 1,398 schoolchildren (93.2%) provided stool samples and answered the questionnaire. Only these were included in the analysis. Eight species of intestinal protozoa, including two pathogenic species, were found in the stool samples. E. histolytica/dispar was found in 263 pupils (18.8%) (Table 1) and G. lamblia was found in 195 (13.9%). 2.9% of the pupils were infected by both species and 29.7% were infected by at least one of them. In addition to these pathogenic species, six nonpathogenic species were found among the samples: Entamoeba hartmanni, Entamoeba coli, Endolimax nana, Iodamoeba butschlii, Chilomastix mesnili, and Blastocystis hominis. The most common species were E. nana and E. coli, with respective prevalence of 65.5% and 62.3%. Concerning the prevalence of protozoal infection by age and sex, we found a significant association between the prevalence of G. lamblia and sex (χ2 = 7.32, df = 1, p = 0.006) and between the prevalence of C. mesnili and age (χ2 = 4.25, df = 1, p = 0.037). No other significant association with age and sex were found.
We found that only 118 (8.4%) students carried no protozoal species. However, among the infected schoolchildren, 326 (23.3%) were infected by one protozoal species and the remaining children (68.2%) had multiple infections. In the multiple-infection group, 2.5% were infected by E. histolytica/dispar, G. lamblia and E. coli and 12.3% by E. histolytica/dispar, E. nana, and E. coli.
The five sites studied in the town of Agboville were so close that they merged into a single point on the map. Students in all the study sites in this area were infected by these two parasites (Figure 2). E. histolytica/dispar prevalence varied from 4.2% in Agboville town to 40.8% in Oress-Krobou; prevalence exceeded 20% in almost half of the study locations (13 of 30). The villages with the greatest prevalence were distributed throughout the Agboville area. The prevalence of G. lamblia ranged from 2.0% in the village of Loviguié to 26.8% at Kouadjakro. Eight localities had prevalence that exceeded 20%: Séguié, Boka Oho, Kouadjakro, Babiahan, Ery-Makouguié, Grand Moutcho, Gbéssé, and Anno. These villages were also distributed throughout the area without focus zone. In addition, Seguié, Boka Oho, Kouadjakro, Ery-Makouguié, Grand Moutcho, and Gbéssé were the most infected localities by both pathogenic species.
The spatial distribution of E. histolytica/dispar and the rate of blood in stools cannot be superimposed, even if in certain localities prevalence coincides.
Localities where blood in stools is found in more than 20% of the schoolchildren were identified during a study carried out by Guéssan et al. [21]. In our study, prevalence of the three intestinal parasites species suspected to cause blood in stools was evaluated in these localities (Table 2). In Gbéssé, Séguié, and Offompo, only E. histolytica/dispar had a prevalence higher than 20%. In these locations, this species is likely the most responsible for the observed blood in stools. In Ery-Makouguié and Oress-Krobou, E. histolytica/dispar together with Ancylostoma spp. has a prevalence higher than 20%; both species could account for the observed rate of blood in stools. In Odoguié and Yadio, E. histolytica/dispar could have slightly contributed to blood in stools, given the higher prevalence of Ancylostoma spp. and S. mansoni. In all other localities the prevalence of E. histolytica/dispar is lower than 20% so this species seems not have contributed to blood in stools there.
Water sources in the Agboville area consist of taps, dug wells, boreholes, rivers, backwaters (shallows area adjacent to or part of a river, where clothing is often laundered), ponds, and reservoirs. Most people supply themselves with water from several sources at the same time. Dug wells are the main source in both towns and villages.
A significant negative association was observed between use of tap water and a high prevalence of E. histolytica/dispar (odds ratio OR = 0.84, 95% CI = 0.73–0.96). High prevalence of G. lamblia infection had a significant positive association with use of pond water (OR = 1.28, 95% CI = 1.06–1.53). The results show also a significant positive association between high prevalence of non-pathogenic E. coli and E. nana with, respectively, backwaters and rivers. These two nonpathogenic intestinal protozoa had a significant negative association with use of tap water (Table 3).
The study was carried out among schoolchildren because they were one of age groups the most exposed to intestinal parasites and were generally accessible. Those of third grade primary school (CE1) were chosen because they were the youngest pupils able to answer to questions without difficulty and could be followed over several subsequent years.
The method used for stool analysis, formol-ether concentration [25]–[27], did not allow a distinction between E. histolytica and E. dispar, so these parasites were indicated by E. histolytica/dispar. More specialized methods now exist to distinguish them [30],[31] but remain inaccessible in the majority of developing countries [32]. The prevalence of this parasite complex in our study (18.8%) is identical to that obtained by Heckendorn et al. in 2002 in the town of Agboville [20]. In addition, our extended areas of sampling show that beyond Agboville town, the parasite complex infects people in the wider area (including villages) beyond the town Agboville, and maintains its level of infection in the population. In the Man area, in west Côte d'Ivoire, prevalence of the complex E. histolytica/dispar is even lower, with a rate of 11.3% [33]. Distinction between the two species E. histolytica and E. dispar could led to a weaker prevalence of the pathogenic species [30],[31]. In Agboville town, in an analysis of only microscopically positive samples by PCR, the ratio E. histolytica to E. dispar was 1∶46 [20]. On the basis of this ratio, prevalence of the pathogenic species (E. histolytica) in our study could be about 0.4%. However, studies have shown a significant association between this complex and diarrhea in Nigeria [32], so the high prevalence of the E. histolytica/dispar complex as a contributor to illness must nevertheless be considered, even if it is controversial.
Prevalence of G. lamblia in Agboville area was 13.9%. This is above other estimates for Côte d'Ivoire: 10.8% in the Man area [18] and 1.4% in Toumodi in central Côte d'Ivoire [34]. The higher prevalence of this parasite in the Agboville area could be due to higher rainfall [24].
Protozoal infection was associated with age and sex for two species. The 6- to 10-year age group was the most infected by C. mesnili. This has been observed in the west [17] and in other African countries [1],[35] and is due to the risky behavior and relatively poorer hygiene measures in this age group. G. lamblia infection was associated with sex, with girls more highly infected. Where surface water is used for household activities, girls are more vulnerable as indicated by Brelet [36].
Concerning polyparasitism, our results are comparable to those of Keiser et al. obtained in western Côte d'Ivoire [17]. The observed multiple infections could be explained by the facts that many species of protozoa have the same mode of transmission and that hygiene is poor in these areas. E. histolytica/dispar and G. lamblia were found in samples from all the localities studied. This cosmopolitan distribution of these parasites has been reported by some authors [7],[31]. Localities of high prevalence are distributed throughout the Agboville area. The even spatial distribution of E. histolytica/dispar is identical to that observed in the Man area in western Côte d'Ivoire. In contrast, three focal zones were observed in the spatial distribution of G. lamblia in the Man area, contrary to the Agboville area [33]. The even distribution of these parasites in the Agboville area shows that transmission is not related to the physical environment of the area but to the fact of specific parameters of each locality. Eradication efforts should thus take into account the entire area without stratification and emphasize improvements in hygiene conditions. Among the 13 localities with E. histolytica/dispar prevalence over 20%, Offompo, Grand Moutcho, Oress Krobou, Yadio, Ery-Makouguié, Odoguié, and Gbéssé also have a fecal blood rate of over 20% [21]. This blood may be due to S. mansoni or Ancylostoma spp., but E. histolytica/dispar is likely to be a cause only in Séguié, Offompo, and Gbéssé. In Azaguié and Agboville, the prevalence of pathogenic protozoan infection is low, as is the blood rate in stool, certainly because these localities benefit from a distribution network of safe drinking water and hygiene conditions are better.
The socioeconomic status of the populations has not been taken into account in this study. However, recent work has shown that income levels of people influence the distribution of intestinal helminths [37]. Other factors, such as drinking water sources, could play decisive role in the occurrence of these parasites. Therefore, water sources were explored in this study. Negative ratio values between high prevalence of E. histolytica/dispar, E. coli, and E. nana and use of tap water (OR less than 1) show that contamination with these parasites decreases when use of tap water increases. Tap water usually undergoes chemical treatment to remove a number of infectious agents before being distributed to people. These precautions provide relatively good water quality.. Its consumption contributes to the reduction of infection by protozoa. Positive odds ratio (OR greater than 1), obtained between high rates of infection by G. lamblia, E. coli, and E. nana species and use of ponds, backwaters, and rivers as sources of household or drinking water, show that the prevalence of these parasites increase when the use of these sources increases. In Agboville, as in majority of developing countries, hygiene conditions are poor and could support propagation of G. lamblia through pond water contamination by human feces. In addition, animals such as rats bathe or drink in ponds and then leave many Giardia cysts [38]. These water sources are usually highly polluted, especially in rainy seasons [39], contaminated by rain runoff charged with parasite cysts from animals and human droppings. Consumption of these exposed waters, in an area with high rainfall like Agboville, would be the basis for population-wide parasite infection. As in Offompo village, which has few or no toilets at all [19], other localities studied lack toilets. In villages, when toilets exist, they are not often used and people defecate in the open. This observation was made during a study conducted in a village in Senegal, where 24% of the subjects defecated in the open, despite the existence of toilets [40]. This behavior in the population favors the spread of protozoal cysts.
In order to limit the development of resistant strains of pathogenic intestinal protozoa, some authors recommend focusing preventive efforts and to target chemotherapy [22],[23]. For parasitosis control, spatial distribution is important [41]. In the Agboville area, chemotherapy treatment should target the most infested populations. Sanitation education of the population, especially on the risks of surface water use and precautions to be taken, must accompany this treatment. The importance of providing communities with safe drinking water should also be impressed upon communities and authorities. Implemented together, targeted chemotherapy and provision of safe drinking water will allow better control of these parasites in the study area.
Intestinal protozoa are common in the Agboville area of Côte d'Ivoire with a high prevalence of the pathogenic species E. histolytica/dispar and G. lamblia. Polyparasitism was highly prevalent in this area. No major focus zone was observed in the spatial distribution of both species. This result shows that control or eradication efforts against these intestinal protozoa must take into account the whole area, with urgent chemotherapy treatments delivered to the most-infected population segments. A significant negative association was observed between infection with E. histolytica/dispar and household use of tap water. G. lamblia was significantly associated with household use of pond water. Parasites prevalence decreases when tap water is used and increases when surface water is used. This work will help to make populations and political powers aware of the importance of these parasites and the need for safe drinking water in all the localities of this area. It can also contribute to develop an integrated control program against these parasites in this area of Côte d'Ivoire, including prophylactic and chemotherapy measures.
|
10.1371/journal.pgen.1007144 | Brassinosteroids regulate root growth by controlling reactive oxygen species homeostasis and dual effect on ethylene synthesis in Arabidopsis | The brassinosteroids (BRs) represent a class of phytohormones, which regulate numerous aspects of growth and development. Here, a det2-9 mutant defective in BR synthesis was identified from an EMS mutant screening for defects in root length, and was used to investigate the role of BR in root development in Arabidopsis. The det2-9 mutant displays a short-root phenotype, which is result from the reduced cell number in root meristem and decreased cell size in root maturation zone. Ethylene synthesis is highly increased in the det2-9 mutant compared with the wild type, resulting in the hyper-accumulation of ethylene and the consequent inhibition of root growth. The short-root phenotype of det2-9 was partially recovered in the det2-9/acs9 double mutant and det2-9/ein3/eil1-1 triple mutant which have defects either in ethylene synthesis or ethylene signaling, respectively. Exogenous application of BR showed that BRs either positively or negatively regulate ethylene biosynthesis in a concentration-dependent manner. Different from the BR induced ethylene biosynthesis through stabilizing ACSs stability, we found that the BR signaling transcription factors BES1 and BZR1 directly interacted with the promoters of ACS7, ACS9 and ACS11 to repress their expression, indicating a native regulation mechanism under physiological levels of BR. In addition, the det2-9 mutant displayed over accumulated superoxide anions (O2-) compared with the wild-type control, and the increased O2- level was shown to contribute to the inhibition of root growth. The BR-modulated control over the accumulation of O2- acted via the peroxidase pathway rather than via the NADPH oxidase pathway. This study reveals an important mechanism by which the hormone cross-regulation between BRs and ethylene or/and ROS is involved in controlling root growth and development in Arabidopsis.
| Both brassinosteroids (BRs) and ethylene have been known to control root growth and development. ROS have been also reported to play an important role in root development. However, the relationship between BRs and ethylene or ROS in root growth and development was not addressed before. In this study, a det2-9 mutant defective in BR synthesis was identified from an EMS mutant screening, displaying a short-root phenotype which is result from the hyper-accumulation of ethylene and superoxide anions (O2-). Exogenous BR apply showed that BRs either positively or negatively regulate ethylene biosynthesis in a concentration-dependent manner. Different from the BR induced ethylene biosynthesis through stabilizing ACSs stability, we found that the BR signaling transcription factors BES1 and BZR1 interacted with promoters of ACS7, ACS9 and ACS11 to repress their expression, indicating a native regulation mechanism under physiological levels of BR. The BR-modulated control over the accumulation of O2- acted via the peroxidase pathway rather than via the NADPH oxidase pathway. This study provides new insights into how brassinosteroids control root growth through the cross-regulation with ethylene synthesis and ROS.
| Roots are important plant ground organs, which absorb water and nutrients to control plant growth and development. In higher plants, root growth is maintained by coordinating cell proliferation and differentiation [1–3]. Plant hormones have been known to play a crucial role in the regulation of root growth [4]. Recent studies in the Arabidopsis root have shown that different hormones control organ growth by regulating specific growth processes such as cell proliferation, differentiation or expansion in distinct tissues. Plant hormones such as auxin, cytokinin, abscisic acid, brassinosteroids, ethylene and gibberellins have been shown to be involved in root growth through a range of complex interactions. The activities of these hormones during root growth progression depend on cellular context and exhibit either synergistic or antagonistic interactions. For example, ethylene enhances inhibition of root cell elongation through upregulating the expression of ASA1 and ASB1 to enhance auxin biosynthesis in Arabidopsis seedlings [5]. Furthermore, ethylene regulated root growth was also mediated through modulating the auxin transport machinery [6]. In addition, cytokinin was also found to control root growth through transcriptional regulation of the PIN genes and thus influencing auxin distribution [7]. The balance between auxin and cytokinin signaling is crucial during root growth. In Arabidopsis, cell division and cell differentiation largely determines root meristem size, which is under the control of cytokinin and auxin through an ARR1/SHY2/PIN circuit [1]. All these studies suggest that hormonal cross-talk plays a pivotal role in the regulation of root growth.
The brassinosteroids (BRs) represent a class of phytohormones involved in a wide variety and developmental processes including root development [8–12]. BR, detected by the BRI1 receptor, activates the transcription factors BES1 and BZR1, which in turn govern the transcription of a large number of genes [13–16]. BRs are known to participate in root growth and development, because mutants impaired with respect to either the synthesis or signaling of BR develop foreshortened roots [17, 18]. However, excessive application of bioactive BR hampered normal development of plants [19]. Therefore, a finely tuned cellular regulation of BR levels is important for the development of plant. It has been found that BR deficient conditions elicit the expression of BR biosynthesis genes, while increase in endogenous BR concentration lead to feedback regulation of the expression of BR metabolic genes to maintain the homeostasis of BR [20]. Recent studies demonstrate that BR interacts with plant hormones such as abscisic acid, gibberellins, auxin and cytokinin to regulate plant growth and development [21–23]. BR interacts with ethylene to regulate the gravitropic response of the shoot, and is involved in ethylene-controlled processes in the hypocotyl of both light- and dark-grown seedlings [24, 25]. Exogenously supplied BR enhances the stability of type 2 of the enzymes 1-aminocyclopropane-1-carboxylate synthase (ACS5 and ACS9) and thus increasing ethylene production, thereby modulating the hypocotyl growth of etiolated seedlings [26]. Though both BRs and ethylene have been reported to regulate root growth and development, it is still unknown if there is a cross-regulation between BRs and ethylene in this process.
In addition to plant hormones, the regulation of root growth has also been tightly linked to reactive oxygen species (ROS). Root growth is profoundly affected by endogenously generated ROS. While ROS were initially believed to merely represent a damaging by-product of the plant’s stress response [27], they have been now recognized as signaling molecules [28]. For example, ROS have been shown to be important for balancing cell proliferation and differentiation during root growth, and have been proposed to adopt a signaling role during lateral root formation [29, 30]. It has been reported that ROS produced in mitochondria of root tip cells in response to the hormone abscisic acid (ABA) are responsible for regulating the root’s meristematic activity [31]. A BR receptor-mediated increase of the cytosolic concentration of calcium ions (Ca2+) regulates ROS production, thereby reducing the length of the hypocotyl in dark-grown seedlings [32, 33]. Though BRs have been reported to regulate many plant biotic and abiotic stresses through the regulation of ROS homeostasis [27, 34], the role of the cross-regulation between BRs and ROS in root growth is largely unknown.
Here, the participation of BR in root growth and the extent of its cross-regulation with both ethylene and ROS signaling were investigated by characterizing a novel A. thaliana det2 mutant allele (det2-9) selected on the basis of its short-root phenotype, which proved to be defective with respect to BR synthesis. A key observation was that the det2-9 mutant accumulated more ethylene and ROS than the wild type. The increased accumulations of both ethylene and ROS caused the short root phenotype in det2-9. This study reveals a mechanism about how BRs regulate root growth through a cross-regulation with ethylene and ROS signaling.
To identify novel determinants involved in the control of root growth, an ethyl methane sulfonate (EMS)-mutagenized Arabidopsis population was screened by monitoring root length and elongation. One mutant was subsequently named as short root 5 (sr5) (Fig 1A and 1B). The length of the mutant root was only 23% of the one of a wild type (WT) seedling at 7–8 days post germination. A longitudinal zonation pattern analysis showed that the size of its root apical meristem (RAM) was significantly smaller than the WT control (Fig 1C). Both meristem zone (MZ) and transition zone (TZ) in the RAM were substantially reduced in size. Cortical cells in the mutant mature zone were significantly shorter than those in WT, and cell number in the MZ was strongly reduced (Fig 1D and 1E and Table 1). The number of cells formed by the RAM in sr5 was 1.8 fold fewer than that in the WT control. The length of the mutant’s RAM was 67% of WT’s, and both its MZ and TZ were reduced in size (Table 1). The compromised RAM in the mutant was accompanied by an increased cell cycle time which displayed 1.4 times longer than that in the WT control (Table 1). The signal obtained from the mitotic cyclin B1;1 G2/M transition marker pCYCB1;1::GUS was much weaker in the mutant than the WT control (Fig 1F), an indication that cell proliferation was inhibited in sr5. The conclusion was that the mutant’s short root derived from both a reduced MZ cell number and a smaller cell size in the mature zone.
When positional cloning was employed to identify the site of the sr5 mutation, a position on chromosome 2 flanked by the markers W20 and W22 was identified (S1A Fig). Sequencing of the genes present in the critical genomic region revealed that the mutant had a point mutation causing a G-to-A transition at nucleotide position 107 after ATG in At2g38050 (DET2). The root growth and seedling morphology of the reported det2-1 mutant were indistinguishable from those of sr5 (S1B and S1E Fig). Since the F1 hybrid sr5 x det2-1 retained the short-root phenotype (S1C Fig), it was concluded that the sr5 mutation likely involved a lesion in DET2. Moreover the DET2 promoter driving DET2 cDNA fused to GFP-GUS (pDET2::DET2-GFP-GUS) complemented the short-root phenotype in sr5 (S1D Fig), suggesting that the G107A mutation in DET2 led to the short-root phenotype in the sr5 mutant. In seedlings carrying the transgene, GUS activity was detected both in the shoot and the RAM (S2 Fig). DET2 encodes a steroid 5α-reductase acting in the BR synthesis pathway. The phenotype of sr5, which was similar to that observed in det2-1 grown in darkness, was rescuable when the plants were treated with exogenous BR (eBL) (S1B Fig). Since there are eight alleles of det2 mutant have been reported, we renamed the sr5 mutant as the det2-9 which was used for most of the analysis in this study. We compared the expression levels of some BR induced genes between det2-9 and det2-1 through Q-PCR analysis. The results showed that, though both mutants displayed reduced expression levels of TCH4, BAS1, IAA17 and IAA19, the det2-9 mutant has a higher expression level of these BR-induced genes than the det2-1 mutant (S3 Fig). Consistently, the det2-9 mutant had a weaker phenotype compared with the det2-1 mutant (S1 Fig), indicating that the point mutation at position 107 might not be a null allele.
A RNA-seq approach was applied to compare the det2-9 root transcriptome with that of the WT, a total of 1,480 and 1,116 genes were found to be, respectively, up- and down-regulated (Fig 2A). Among the differentially expressed genes, based on the GO analysis we found that there is a statistically significant enrichment in genes annotated as being linked to secondary metabolic process and response to stimulus (P<0.01). It is not surprising for this enrichment considering the dwarf phenotype of the mutant. Though the previous research has shown that exogenously supplied BR can enhance the production of ethylene [26], the ethylene biosynthesis and ethylene response factors were up-regulated in det2-9 according to our RNA-seq analysis (Fig 2B). These genes included 1-aminocyclopropane-1-carboxylate synthase encoding genes, 1-aminocyclopropane-1-carboxylate oxidase encoding genes and ethylene response factor encoding genes (S1 Dataset). According to our GO analysis, we also found that many of genes belong to GO:0000302 (response to reactive oxygen species) were up-regulated significantly in det2-9 (Fig 2C), which was in contrast with the previous reports showing that BR could induce the generation of H2O2 [27, 34]. To confirm the RNA-seq results, we performed a quantitative real time-PCR (qRT-PCR) assay on a selection of 20 ethylene related genes which were differentially transcribed in WT and det2-9 seedlings grown in light and dark growth conditions (Fig 2D and S4 Fig). Though the expression changes of most of ethylene related genes were confirmed, we also found that some genes, for example ACS6, ERF6 and ERF17, had little agreement between the transcript abundance by RNA-seq and qRT-PCR analysis. Considering three independent repeats were done for the confirmations, the results of qRT-PCR analysis are more reliable. These results suggest that both ROS and ethylene signaling were enhanced in the det2-9 mutant.
Ethylene signaling in the det2-9 mutant was monitored by the expression of the pEBS::GUS ethylene signaling reporter. The strength of the GUS signal was considerably higher in the mutant than that in the WT (Fig 3A), suggesting that an enhanced level of ethylene signaling occurred in det2-9. The increased ethylene response in det2-9 was abolished by the presence of 10 nM eBL during seedling growth (Fig 3A). The ethylene content was considerably higher in the det2-9 mutant than that in WT seedlings (Fig 3B and 3C). The transgene line pDET2::DET2-GFP-GUS/det2-9 complemented the higher level of ethylene observed in det2-9 (Fig 3C). Treatment with the BR synthesis inhibitor propiconazole (PPZ) also resulted in higher ethylene content in light-grown WT seedlings, while eBL (10 nM, a concentration which partially rescued the short-root phenotype in det2-9) treated WT or bes1-D (a mutant which displays an enhanced BR signaling response) light-grown seedlings both showed a reduction in ethylene content (Fig 3B). A similar profile of ethylene content was also observed when seedlings were grown in darkness (S5 Fig). All these above chemical treatment experiments and mutant analysis suggest that both exogenous applied low levels of BR and native BR signaling negatively regulated ethylene biosynthesis.
In addition, we also observed that root growth was inhibited gradually by eBL at concentrations ranging from 10 to 5000 nM (Fig 4A and 4B). While the hypocotyl length was unchanged when treated with low concentration of eBL (<500 nM) but reduced sharply when the concentration of eBL greater than 500 nM (Fig 4A). Furthermore, dark-grown seedling hypocotyls treated with higher concentration of eBL (≥500 nM) displayed a typical “triple response”, indicating the enhanced ethylene response (Fig 4A). Therefore, we further examined the effects of BR on ethylene production using different concentrations of eBL. The results showed that ethylene content was greatly reduced in seedlings treated with low concentration of eBL (10 or 100 nM) while it was strongly increased when the concentration of eBL greater than 500 nM (Fig 4D). Consistently, both GUS staining analysis with the pEBS::GUS transgene line and an examination of ethylene response factors (ERFs) expression using qRT-PCR analysis show that low concentrations (10–100 nM) of BR inhibits ERF expression while high concentrations (≥500 nM) of BR enhanced expression (Fig 4C and 4E), consistent with the change ethylene levels. In summary, BR either positively or negatively regulate ethylene biosynthesis depends on the levels of BRs.
When det2-9 mutant seedlings were grown on a medium containing either silver nitrate (AgNO3, an antagonist of ethylene signaling) or 2-aminoethoxyvinyl glycine (AVG) (an inhibitor of ethylene synthesis), the root growth of det2-9 mutant was partially rescued, producing root lengths almost double than those developed by the non-treated mutant seedlings. However, both treatments inhibited the root growth of WT seedlings (Fig 5A and 5B). In addition inhibition of ethylene signaling by AgNO3 rescued the cortical cell length in det2-9 (S6 Fig). On the other hand, the root cell elongation and root growth of the mutant seedlings was found to be more sensitive to ACC (a precursor of ethylene synthesis) (Fig 5A and 5B and S6A Fig). In addition, both WT and det2-9 mutant seedlings displayed similar cell numbers in root meristem under the same treatment with either AgNO3 or ACC (S6B Fig). This result suggests that the BR deficiency caused short-root phenotypes in det2-9 was mediated by the effect of ethylene signaling on root cell elongation. Consistently, the octuple acs mutant CS16651 (acs2-1/acs4-1/acs5-2/acs6-1/acs7-1/acs9-1/amiRacs8acs11), ein2-5 and the ein3/eil1-1 double mutant, which have defects in either ethylene biosynthesis or ethylene signaling, were less affected by the PPZ treatment than WT (Fig 5C). The short-root phenotype of det2-9 was partially recovered in the det2-9/acs9 double mutant and det2-9/ein3/eil1-1 triple mutant (Fig 5D–5G). These results indicate that the short-root phenotype in det2-9 partly result from enhanced ethylene biosynthesis and ethylene signaling.
A promoter analysis showed that promoters of ACS6, 7, 9, 11, along with ACO1 and 3 (all these genes were strongly up-regulated in det2-9, Fig 2) contained a BRRE and/or an E-box, the binding sites for BES1 and BZR1 (Fig 6A). The direct interaction of the ACSs by BES1 or BZR1 was confirmed by a chromatin immunoprecipitation (ChIP)/qPCR analysis in FLAG-tagged BES1 or YFP-tagged BZR1 transgenic lines (Fig 6A). A series of yeast one-hybrid assays were conducted to further verify whether any of these promoters was regulated directly by either BES1 or BZR1. The outcome was that in yeast BES1 interacted with the promoters of ACS7 and 9, while BZR1 did so with the promoters of ACS9 and 11 (Fig 6B). Neither of the two transcription factors interacted definitively with the ACS6 (Fig 6B), ACO1 or the ACO3 promoter (S7 Fig). The trans-activity of BES1 or BZR1 with the ACS promoters was further demonstrated in a transient dual LUC expression assay in A. thaliana mesophyll protoplasts. The over-expression of both BES1 and BZR1 strongly repressed the activity of ACS promoters (Fig 6C), confirming that either BES1 or BZR1 can repress ACS7, 9 and 11 gene expression in vivo.
A previous study reported that short-term treatment with eBL resulted in dephosphorylation of BES1 (its active form) in WT and in det2-1, but not in bri1-5 or bin2-1 signaling BR mutants [35]. Therefore we further investigate the expression of ACSs in BR signaling mutants including bri1-116 and bin2-1. qRT-PCR results showed that the expression of ACS6, 7, 9 and 11 increased in the det2-9 (see also Fig 2D), bri1-116 and bin2-1 mutant compared with WT (S8 Fig). The expressions of these four ACS genes decreased when treated with eBL to a lower extent in bri1-116 or bin2-1 mutants compared with the det2-9 mutant (S8 Fig). These results indicate that BR signaling pathway is required for the BR-mediated repression of ACS gene expression, via direct regulation by the BES1 and BZR1 transcription factors.
The transcriptomic analysis in det2-9 mutant roots identified genes responding to ROS as BR-targets (Fig 2C). Therefore, we further analyzed det2-9 mutant for defects in ROS using the nitroblue tetrazolium (NBT) staining method to detect the presence of O2- in vivo [36]. The NBT signal was higher in the det2-9 mutant than that in the WT control (Fig 7A), while there was no clear difference when 3,3’-diaminobenzidine (DAB) staining was used to visualize the level of H2O2 present [37] (S9 Fig). This suggests that det2-9 accumulated O2- but not H2O2. Treatment with eBL substantially reduced the extent of the O2- hyper-accumulation in det2-9 (Fig 7A). Meanwhile, the BR-signaling defective mutant bri1-116 hyper-accumulated O2-, while BR-signaling enhanced plants (p35S::BRI1-GFP or bes1-D) accumulated less superoxide anion in their roots compared with WT plants (Fig 7B), indicating that BR signaling suppresses the accumulation of O2-. Therefore, root growth analysis was done using det2-9 mutant seedlings were exposed to two different O2- scavengers, namely superoxide dismutase (SOD) [38] and 1,3-dimethyl-2-thiourea (DMTU) [39]. The root length in det2-9 was significantly increased in the presence of 0.65U/ml SOD, while the same treatment inhibited root growth in WT seedlings (Fig 7C). Similarly, a concentration of 0.1 to 2 mM DMTU treatment, which has no effect on root growth in WT seedlings, could significantly increase root lengths in det2-9 (Fig 7D).
The accumulation of O2- in det2-9 can be the result of the activation of two signaling pathways: peroxidase or NADPH oxidase. When the NADPH oxidase pathway was blocked by the presence of either diphenylene iodonium (DPI) [40] or ZnCl2 [41], det2-9 mutant roots were insensitive to any treatment (Fig 8A and 8B). Consistent with this result, the abundance of transcripts of the four NADPH oxidase genes (RBOHC, D, F and G) was identical in det2-9 and WT (S10A Fig). The root growth response to PPZ treatment of three mutants rbohD, rbohF and rbohD/F was also similar to the one of WT (S10B Fig). NBT staining showed that the BR deficiency-induced O2- hyper-accumulation by PPZ treatment was unaffected in both plants harboring p35S::NADPHD-GFP and the rbohD/F double mutant (S10C Fig). And O2- hyper-accumulates in det2-9/rbohD and det2-9/rbohD/F mutants similarly to det2-9, compared with WT (S10D Fig). These experiments allow us to conclude that the hyper-accumulation of O2- in det2-9 did not involve the NADPH oxidase pathway. So attention was focused on the peroxidase pathway [29], by treating seedlings with either salicylhydroxamic acid (SHAM) [42] or 1,10-phenanthroline (1,10-Phe) [43], inhibitors of peroxidase activity. The root length of det2-9 was significantly increased by both treatments, whereas root growth of WT was slightly inhibited (S11A and S11B Fig). NBT staining showed that the levels of O2- in det2-9 reduced sharply when treated with SHAM or 1,10-Phe but no obvious changes were observed when treated with DPI or ZnCl2 (Fig 8C), which was consistent with the NBT staining observed in det2-9/rbohD and det2-9/rbohD/F mutants compared with WT and det2-9 (S10D Fig). When the transcription of genes encoding peroxidase was investigated, no clear-cut differences were visible between the mutant and WT (S12 Fig), but peroxidase activity was much stronger in the det2-9 mutant and was reduced when seedlings were treated with exogenous BR (Fig 8D). Thus the hyper-accumulation of O2- in det2-9 was likely the effects of an increased peroxidase activity.
Given that the level of both ethylene and O2- was enhanced in det2-9, the question arose as to whether ethylene and ROS interacted with one another. O2- accumulation was initially assayed in WT and det2-9 plants treated with either AVG or ACC (Fig 9A). NBT staining showed that the ACC treatment had a positive and AVG had a negative effect on superoxide anion accumulation in WT roots (Fig 9A). This indicates that ethylene induces an accumulation of O2- in Arabidopsis. NBT staining also showed that p35S::EIN3-GFP accumulated more O2- than WT, but when treated with PPZ, the extent of O2- accumulation was similar among p35S::EIN3-GFP, ein3/eil1-1 and WT (S13 Fig), indicating that there was another pathway independent from ethylene participating in O2- accumulation when BR synthesis was blocked with PPZ treatment. In the det2-9 mutant, there was no clear increasement for ACC-induced superoxide anion accumulation, but the AVG treatment reduced it, which is also an indication that the increase in superoxide anion accumulation was at least partially dependent on ethylene production in det2-9. Since peroxidase activity in det2-9 was higher than that in WT (Fig 8D), an experiment was conducted to compare peroxidase activity in plants carrying p35S::EIN3-GFP, the ein3/eil1-1 double mutant and WT. The result showed that the peroxidase activity was not clearly affected in p35S:EIN3:GFP, ein3/eil1-1 compared with the wild-type control (Fig 9B), indicating that ethylene signaling pathway is unlikely to activate the POD pathway for BR-regulated accumulation of O2- in det2-9 mutant.
To investigate whether the O2- accumulation can alter normal ethylene signaling, we compared the expression level of pEBS::GUS when treated or not with methyl viologen (MV, a superoxide anion propagator) treatment. As shown in Fig 9C, the expression level of pEBS::GUS reporter was considerably induced when treated with MV, suggesting that an O2- accumulation can increase ethylene content. We then measured the primary root growth of ein2-5, ein3/eil1-1 and wild type when treated or not with MV. Mutants in ethylene signaling were more resistant than WT to the negative effects of MV on root growth (Fig 9D). The expression of genes encoding ACS and ACO, analyzed by qRT-PCR, increased when treated with MV (S14 Fig). These results further indicated that O2- accumulation can alter normal ethylene production.
The BRs are well recognized as promoters of cell elongation, also in addition to their involvement in the de-etiolation response, where the opening of the apical hook is thought to require a decrease in the level of ethylene synthesis [17, 44, 45]. It was found that BR enhances ethylene production through the synergistic interaction with eto1 and eto3 [46]. Another study from the same lab showed that supplying BR exogenously promotes ethylene synthesis in the A. thaliana seedlings via stabilizing ACS5 and ACS9 protein [26]. This BR-induced ethylene production was also observed in mung bean and maize [47, 48]. However, in jujube fruit, 5 μM BR-treated fruits caused a significantly lower level of ethylene during storage and the inhibition fruit ripening [49]. These contradictory results indicate the complicated effects of BR on ethylene synthesis. However, all these observations are based on the chemical treatment with BR.
In this study, a new mutant allele of DET2, det2-9, was identified based on the short-root phenotype (Fig 1 and S1 Fig). DET2 encodes a steroid 5α-reductase involved in BR biosynthesis, catalyzing the formation of campestanol with campesterol as substrates. Since another allele det2-1 and other mutants such as cpd and dwf4 which have defects in different steps of BR biosynthesis also displayed short-root phenotype [50–52], it is unlikely that campesterol accumulation caused the short-root phenotype. The short-root phenotype is most likely a result of the reduced BRs synthesis in det2-9 since externally applied BRs could largely rescued the mutant phenotypes (S1 Fig). Through genetic analysis and chemical treatment, we found that the short-root phenotype in det2-9 was partly the resulted of over accumulation of ethylene leading to enhanced ethylene signaling (Figs 3 and 5). The ethylene content was considerably higher in the det2-9 mutant than that in WT seedlings (Fig 3B). Treatment with the BR synthesis inhibitor propiconazole (PPZ) also resulted in higher ethylene content in light-grown WT seedlings, while eBL (10 nM, a concentration which enhances the growth of root in det2-9) treated WT or bes1-D (a mutant which displays an enhanced BR signaling response) light-grown seedlings both showed a reduction in ethylene content (Fig 3B). A similar profile of ethylene content was also observed when seedlings were grown in darkness (S5 Fig). Transcriptional profiling showed that a number of ACS genes were up-regulated in the det2-9 mutant both in light and dark growth conditions, consistent with its increased level of ethylene (Figs 2B, 2D and 3 and S4 and S5 Figs). Since the BR signaling transcription factors BZR1 or BES1 bind to ACS promoters to repress their expression (Fig 6), we analyzed BR signaling pathway by using bri1-116 or bin2-1 mutants and found that the BR-mediated down-regulation of ACS genes was greatly reduced in these two mutants compared with det2-9 (S8 Fig), further indicating BZR1 or BES1 mediated BR signaling negatively regulates the expression ACS transcription factors. This result indicates that native physiological levels of BRs negatively regulate ethylene production through BZR1 or BES1 mediated transcriptional regulation of ACSs.
Since our results and Zhu et al.’s observations [49] are in contrast to other reports which showed that BRs enhanced ethylene biosynthesis [26, 46–48], we further did dosage-dependent assay to test the effects of BR on ethylene productions and root growth. Not surprisingly, root growth was inhibited gradually by eBL at concentrations ranging from 10 to 5000 nM (Fig 4A and 4B). However, ethylene content was greatly reduced in seedlings treated with low concentration of eBL (10 or 100 nM) while it was strongly increased when the concentration of eBL greater than 500 nM (Fig 4D). Root growth analysis under both treatment suggests that both high and low levels of ethylene cause a short-root phenotype (Fig 4D), which is consistent with the previous reports. In the octuple acs mutant (CS16651), which has only 10% ethylene level compared with WT, a reduced root growth phenotype was observed [53]. The acs9 mutant also displays a short root phenotype (Fig 5E). The high levels of BR (500 nM or 1000 nM BR) induced ethylene production is also consistent with the previous reports in Arabidopsis [26, 46]. This study together with previous reports clearly showed that BRs either positively or negatively regulate ethylene biosynthesis in a concentration-dependent manner to control root growth. Certainly, since BR can also interact with other plant hormones such as auxin, ABA, cytokinin and jasmonic acid to regulate myriad aspects of plant growth and developmental processes in plants [54, 55], externally applied BR treatment caused root-growth phenotype might be also result from the interaction between BR and other plant hormones.
ROS represent not only a by-product of stress response, but also influence growth and development in response to both internal developmental signals and external environmental cues [28]. The contrasting ROS status in the cell proliferation and the cell differentiation zones has recently been shown to be an important driver of root growth [29]. A mitochondria localized P-loop NTPase was also reported to regulate quiescent center cell division and distal stem cell identity through the regulation of ROS homeostasis in Arabidopsis root [56]. It has been pointed out that ABA-promoted ROS regulates root meristem activity [31]. In cucumber plants exposed to exogenous BR, H2O2 accumulates as a result of an increased activity of NADPH oxidase [27], while in tomato, the same result is achieved by the up-regulation of RBOH1 [57]. BR has been documented as inducing a receptor-dependent increase in cytosolic Ca2+, which stimulates NADPH oxidase-dependent ROS production [32, 58]. Thus, although the participation of BR in root growth and development is accepted, its interaction with ROS signaling has not been systematically explored to date. Here, a key finding was that the det2-9 mutant hyper-accumulated O2-, which in itself likely contributed to the short root phenotype (Fig 7). BR inhibited the synthesis of O2- via the peroxidase (Fig 8C and 8D and S11 Fig) rather than via the NADPH oxidase (Fig 8A and 8B and S10 Fig) pathway. These results suggest that H2O2 and superoxide anion respond dissimilarly to BR in A. thaliana seedlings. While the level of H2O2 rises rapidly upon exposure to exogenous BR, the one of the superoxide anion is repressed. In addition, the hyper-accumulation of ethylene displayed by det2-9 contributed to a rise in the superoxide anion content in a peroxidase-independent manner (Fig 9A and 9B).
In summary, according to this study together with the previous reports, a proposed model was given in Fig 10. We suggest that BR inhibits ethylene synthesis by activating the transcription factors BZR1 and BES1 under low levels. These transcription factors bind directly to the ACS promoters, thereby suppressing ACS expression and damping the level of ethylene synthesis under normal growth conditions. While high levels of BR induce ethylene biosynthesis either through increasing the stability of ACSs or influencing auxin signaling regulated ethylene production [47, 59, 60]. The possible regulation mechanism of BES1/BZR1’s activity under different levels of BR maybe refer to the regulation mechanism of ARF3 under different levels of auxin. Recent study has found that ARF3 acts as a repressor or activator depends on auxin concentration [61]. At the same time, BR inhibits the synthesis of O2- via the peroxidase pathway, but not NADPH oxidase pathway, which serves to regulate the growth of the A. thaliana seedling root. The accumulation of the O2- is also partially controlled by ethylene signaling in a peroxidase-independent manner and the O2- accumulation can enhance ethylene signaling by increasing the expression of ACSs and ACOs. Understanding how ethylene mediates BR signaling to control the accumulation of the O2- represents a logical follow-up research target.
All of the A. thaliana mutants and/or transgenic lines utilized are in a Col-0 background; the following have been described elsewhere: det2-1 [62], bes1-D [13], bri1-116 [63], bin2-1 [64], p35S::BRI1-GFP [65], ein2-5 [66], ein3/eil1-1 [67], acs9 [68], p35S::BZR1-YFP [69], pNP::BES1-FLAG [70], p35S::EIN3-GFP [71], and p35S::NADPH-GFP [72]. And rbohD, rbohF, rbohD/F all described in Torres’ paper [73]. The octuple acs mutant (CS16651, acs2-1/acs4-1/acs5-2/acs6-1/acs7-1/acs9-1/amiRacs8acs11) [53] was obtained from the Arabidopsis Biological Resource Center (ABRC, Columbus, OH, USA), and the marker lines pCYCB1;1::GUS [74] and pEBS::GUS [75] from early research. The 1501-bp upstream region from the DET2 start cordon and the cDNA of DET2 were amplified and linked to the GFP-GUS reported in gateway vector PKGWFS7.1 [76] to obtain pDET2::DET2-GFP-GUS reporter construct. Prior to germination, the seed was surface-sterilized by fumigation in chlorine gas, held for two days at 4°C on solidified half strength Murashige and Skoog (MS) medium, then transferred to a growth room providing a 16 h photoperiod and a constant temperature of 20°C.
Root tips were imaged by laser-scanning confocal microscopy. The number (obtained from a count of cells in the cortex file extending from the quiescent center to the TZ) and length of cortical and mature epidermal cells were obtained from microscope images using ImageJ software. The criteria for defining the MZ and TZ were those described by Napsucialy-Mendivil et al. [77]. The cell production rate was based on the rate of root growth and the length of fully elongated cells, and the cell cycle time on cell production and the number of cells present in the MZ, as described by Napsucialy-Mendivil et al. [77]. The number of cells displaced from the cell proliferation domain (Ntransit) during a 24 h period was estimated from the equation Ntransit = (24 ln2 NMZ)/T, where NMZ represents the number of cells in the RAM MZ and T means cell cycle time in hours. Histochemical GUS staining was performed according to the method described by Gonzalez-Garcia et al. [78].
The mapping population was the F2 generation of the cross sr5 x Landsberg erecta. Genomic DNA was extracted from each F2 seedlings showing the sr5 phenotype. Simple sequence length polymorphism markers were used for the initial genome-wide linkage analysis, following Lukowitz et al. [79]. To enable fine mapping, 22 PCR-based markers were designed to target the relevant region of the A. thaliana genome sequence.
RNA was isolated from the roots of six day-old det2-9 and WT seedlings using the TRIzol reagent (Invitrogen, Carlsbad, CA, USA) and treated with DNase I to remove contaminating genomic DNA. The preparation was enriched for mRNA by introducing magnetic beads coated with oligo (dT). The resulting mRNA was fragmented into fragments of about 200 nt, and the cDNA first strand was then synthesized via random hexamer priming. After synthesizing the second strand with DNA polymerase I, the ds cDNA was purified using magnetic beads coated with oligo (dT) and End reparation is then performed. Adaptors were then ligated to each end of the fragments, and the products were size-selected by gel electrophoresis. Finally, the fragments were amplified based on the adaptor sequences, purified using magnetic beads coated with oligo (dT) and dissolved in the appropriate amount of Epstein-Barr solution. The concentration and integrity of the ds cDNA was monitored using a 2100 Bioanaylzer device (Agilent Technologies Japan Ltd.). The cDNA was then sequenced using an Ion Proton platform (www.thermofisher.com). Low quality and adaptor sequences were removed and the remaining sequences were then aligned to the A. thaliana genome sequence using SOAP2 software. Individual transcript abundances were expressed in the form of the number of reads per kilobase per million reads (RPKM), and differentially transcribed genes were identified using the thresholds FDR≤0.001 and |log2|≥1 [80].
The RNA template required for qRT-PCR was isolated using an RNeasy PlantMini kit (Qiagen, Hilden, Germany) following the manufacturer’s protocol. After treating with DNase I to remove contaminating genomic DNA, a 2 μg aliquot was reverse-transcribed using a Transcriptor First Strand cDNA Synthesis kit (Roche, Basel, Switzerland), following the manufacturer’s protocol. The subsequent qRT-PCRs were run on a MyiQTM Real-time PCR Detection System (Bio-Rad, Hercules, CA, USA) using FastStart Universal SYBR Green Master mix (Roche, Basel, Switzerland). Each sample was represented by three biological replicates, and each biological replicate by three technical replicates. The reference sequence was AtACTIN2 (At3g18780). Primer sequences are given in S2 Dataset.
Ten seedlings were placed in a 100 mL vial containing 50 mL solidified half strength MS either with or without eBL or PPZ, and immediately capped. The vials were held under a 16 h photoperiod and a constant temperature of 20°C. After seven days, a 10 μL sample of the headspace was subjected to gas chromatography using a GC-6850 device equipped with a flame ionization detector (Agilent Technologies Japan Ltd.).
The coding sequences of BES1 and BZR1 were inserted separately into the EcoRI-XhoI cloning site of pGADT7 (Takara, USA), while the promoter sequences of ACS6, 7, 9, 11, ACO1 and 3 were inserted into the cloning site of pAbAi. The primer sequences used in the construction of the various constructs are given in S2 Dataset. Each of the constructs (including an empty vector for control purposes) was transferred separately into yeast Y1HGold using the PEG/LiAc method. The yeast cells were plated onto SD/-Ura/-Leu medium containing various concentrations of Aureobosidin A to allow for a highly stringent screening of interactions. The procedure followed the manufacturer’s protocol given for the Matchmaker Gold Yeast One-Hybrid Library Screening System (www.clontech.com).
Ten day old transgenic plants were used for the ChIP assay following Gendrel et al. [81]. The quantity of precipitated DNA and input DNA was detected by qPCR. For each ACS promoter, primers were designed to amplify a fragment of length ~70–150 bp lying within the 2 kbp of sequence upstream of the transcription start site. The relevant primers are given in S2 Dataset. Enrichment was calculated from the ratio of bound sequence to input.
The BES1 or BZR1 coding sequences were amplified and the resulting sequences introduced into pBI221 to place them under the control of the CaMV 35S promoter. The ACS promoter sequences were amplified and introduced into the pGreenII0800-LUC reporter vector. Both recombinant plasmids were then transferred into A. thaliana protoplasts. Firefly luciferase (LUC) and renillia luciferase (REN) activities were measured using the Dual-Luciferase Reporter Assay System (www.promega.com). LUC activity was normalized against REN activity [82]. Details of all primers used are given in S2 Dataset.
The roots of five day-old seedlings were immersed for 15 min in 2 mM NBT in 20 mM phosphate buffer (pH 6.1). The reaction was stopped by transferring the seedlings into distilled water. The material was then imaged under a light stereomicroscope.
Tissue peroxidase activity was measured by a spectrophotometric analysis (420 nm) of the formation of purpurogallin from pyrogallol in the presence of H2O2. The roots of nine day-old seedlings were harvest and weighted. Tissue homogenate was prepared using 9 times phosphate buffer and then centrifuged for 10 min in 3500 rpm. The supernatant was used for peroxidase activity measurement. A single unit of enzyme was defined as the amount catalyzed and generated 1 μg pyrogallol by 1.0 mg fresh tissues in the reaction system at 37°C. Peroxidase activity was calculated from the formula provided with the peroxidase assay kit (Jiancheng Bioengineering Institute, Nanjing, China).
Sequence data for genes used in this study can be found in the Arabidopsis Genome Initiative or GenBank/EMBL databases under the following accession numbers:
DET2 (At2g38050), BES1 (At1g19350), BZR1 (At1g75080), ACS1 (At3g61510), ACS2 (At1g01480), ACS4 (At2g22810), ACS5 (At5g65800), ACS6 (At4g11280), ACS7 (At4g26200), ACS9 (At3g49700), ACS11 (At4g08040), WOX5 (At3g11260), ARF10 (At2g28350), ARF16 (At4g30080), ARR1 (At3g16857), SHY2 (At1g04240), BRI1 (At4g39400), CYCB1;1 (At4g37490), ACO1 (At2g19590), ACO2 (At1g62380), ACO3 (At2g05710), ACO4 (At1g05010), ACO5 (At1g77330), ERF6 (At4g17490), ERF13 (At2g44840), ERF17 (At1g19210), ERF104 (At5g61600), ERF105 (At5g51190), EBS (At4g22140), EIN3 (At3g20770), EIL1 (At2g27050), RBOHC (At5g51060), RBOHD (At5g47910), RBOHF (At1g64060), RBOHG (At4g25090), TCH4 (At4g57560), BAS1 (At2g26710), IAA17 (At1g04250), IAA19 (At3g15540), ACTIN2 (At3g18780).
|
10.1371/journal.pcbi.1005206 | Extending Integrate-and-Fire Model Neurons to Account for the Effects of Weak Electric Fields and Input Filtering Mediated by the Dendrite | Transcranial brain stimulation and evidence of ephaptic coupling have recently sparked strong interests in understanding the effects of weak electric fields on the dynamics of brain networks and of coupled populations of neurons. The collective dynamics of large neuronal populations can be efficiently studied using single-compartment (point) model neurons of the integrate-and-fire (IF) type as their elements. These models, however, lack the dendritic morphology required to biophysically describe the effect of an extracellular electric field on the neuronal membrane voltage. Here, we extend the IF point neuron models to accurately reflect morphology dependent electric field effects extracted from a canonical spatial “ball-and-stick” (BS) neuron model. Even in the absence of an extracellular field, neuronal morphology by itself strongly affects the cellular response properties. We, therefore, derive additional components for leaky and nonlinear IF neuron models to reproduce the subthreshold voltage and spiking dynamics of the BS model exposed to both fluctuating somatic and dendritic inputs and an extracellular electric field. We show that an oscillatory electric field causes spike rate resonance, or equivalently, pronounced spike to field coherence. Its resonance frequency depends on the location of the synaptic background inputs. For somatic inputs the resonance appears in the beta and gamma frequency range, whereas for distal dendritic inputs it is shifted to even higher frequencies. Irrespective of an external electric field, the presence of a dendritic cable attenuates the subthreshold response at the soma to slowly-varying somatic inputs while implementing a low-pass filter for distal dendritic inputs. Our point neuron model extension is straightforward to implement and is computationally much more efficient compared to the original BS model. It is well suited for studying the dynamics of large populations of neurons with heterogeneous dendritic morphology with (and without) the influence of weak external electric fields.
| How extracellular electric fields—as generated endogenously or through transcranial brain stimulation—affect the dynamics of neuronal populations is of great interest but not well understood. To study neuronal activity at the network level single-compartment neuron models have been proven very successful, because of their computational efficiency and analytical tractability. Unfortunately, these models lack the dendritic morphology to biophysically account for the effects of electric fields, and for changes in synaptic integration due to morphology alone. Here, we consider a canonical, spatially extended model neuron and characterize its responses to fluctuating synaptic input as well as an oscillatory, weak electric field. In order to accurately reproduce these responses we analytically derive an extension for the popular integrate-and-fire point neuron models. We show that the dendritic cable acts as a filter for the synaptic input current, which depends on the input location, and that an electric field modulates the neuronal spike rate strongest at a certain (preferred) field frequency. These phenomena can be successfully reproduced using integrate-and-fire models, extended by a small number of components that are straightforward to implement. The extended point models are thus well suited for studying populations of coupled neurons with different morphology, exposed to extracellular electric fields.
| Extracellular electric fields in the brain and their impact on neural activity have gained a considerable amount of attention in neuroscience over the past decade. These electric fields can be generated endogenously [1–3] or through transcranial (alternating) current stimulation [4–6], and can modify the activity of neuronal populations in various ways [1, 7–9]. Although the fields generated by this type of noninvasive brain stimulation are rather weak (≤1 V/m [4, 5]) and do not directly elicit spikes, they can modulate spiking activity and lead to changes in cognitive processing, offering a range of possible clinical interventions [10–12]. How external fields lead to changes of the membrane voltage in single cells has been studied in detail [13–15]. However, their effects on population spike rate and the underlying mechanisms are largely unexplored.
Computational models of neurons exposed to electric fields offer a useful tool to gain a better understanding of these mechanisms. Multi-compartment models of neurons are well suited for corresponding investigations at the level of single cells and small circuits [16] but are too complex for a purposeful application in large populations. Single-compartment (point) neuron models of the integrate-and-fire (IF) type are well applicable to study the dynamics of large neuronal populations, due to their computational efficiency and analytical tractability [17]. However, typical IF model neurons lack the dendritic morphology required for a biophysical description of electric field effects. Furthermore, even in the absence of an extracellular field, the dendritic morphology strongly shapes neuronal response properties [18].
In this contribution, we extend the popular class of IF point neuron models to quantitatively account for morphology dependent modulations of neural activity due to: (i) dendritic influences on the integration of synaptic inputs and (ii) the effects of extracellular electric fields. Furthermore, we describe how oscillatory electric fields affect neuronal subthreshold and spiking activity and identify field-induced spike rate resonance. Specifically, we considered a canonical spatial pyramidal neuron model which consists of a somatic compartment and one (apical main) passive dendritic cable, and which is exposed to in-vivo like fluctuating synaptic input and an electric field. Based on that model we analytically derived an extension for the classical leaky and the refined exponential, [19], IF point neuron models in order to exactly reproduce the subthreshold dynamics of the spatial model for arbitrary parametrizations. We then evaluated the extended IF models by quantitatively comparing their spiking activity with the spiking activity of the corresponding spatial model. Finally, we used these models to study the effects of an oscillating electric field (due to the presence of the dendritic cable) on the spike rate dynamics.
Our derivation of the extended point neuron model consists of two steps. We first calculate the somatic membrane voltage of a ball-and-stick (BS) model in response to subthreshold synaptic inputs at the soma and the distal dendrite and to a time-varying, spatially homogeneous, extracellular electric field. This involves solving a generalized cable equation [20]. Second, we seek to exactly reproduce this voltage response in the point neuron model by deriving additional model components (see Fig 1): two linear temporal filters, one for each input location, to be applied to the “raw” synaptic input and one additional input current to describe the field effect. The model components are given in analytical form and depend on the parameters of the BS model and the electric field. We refer to the model equipped with the new components as the extended point (eP) neuron model. We first derive this extension for the well-known leaky IF (LIF) neuron model, and present the extension adapted for the exponential IF (EIF) neuron model in a separate section.
The BS neuron model consists of a lumped soma attached to a passive dendritic cable of length L. The dynamics of its membrane voltage, when receiving synaptic inputs at the soma, Is(t), and the distal dendrite, Id(t), and when exposed to a spatially homogeneous external electric field, E(t), are governed by the cable equation:
c m ∂ V BS ∂ t - g i ∂ 2 V BS ∂ x 2 + g m V BS = 0 , 0 < x < L , (1)
subject to the boundary conditions:
C s ∂ V BS ∂ t - g i ∂ V BS ∂ x + G s V BS - G s Δ T e V BS - V T Δ T = I s ( t ) - g i E ( t ) , x = 0 , (2) ∂ V BS ∂ x = I d ( t ) g i + E ( t ) , x = L , (3)
at the soma (x = 0) and the end of the dendrite (x = L). VBS denotes the deviation of the membrane voltage from rest, Vrest, VBS(x, t) ≔ VBS,i(x, t) − VBS,e(x, t) − Vrest, where VBS,i and VBS,e are the intra- and extracellular potentials. The effects of a spike are described by the IF-type reset condition for the soma:
if V BS ( 0 , t ) ≥ V s then V BS ( 0 , t ) : = V r (4)
and by a short refractory period of length Tref during which VBS(0, t) is clamped at the reset value Vr. Spike times are defined by the times at which the somatic membrane voltage VBS(0, t) crosses the spike voltage value Vs from below. cm denotes the membrane capacitance, gm the membrane conductance, and gi the internal (axial) conductance of a dendritic cable segment of unit length. Cs and Gs are the somatic membrane capacitance and leak conductance. The exponential term with threshold slope factor ΔT and effective threshold voltage VT approximates the somatic sodium current at spike initiation [19]. For details see Methods.
In the proposed IF point neuron extension, that is, the eP model, the deviation of the membrane voltage, VeP, from rest is governed by
C eP d V eP d t + G eP V eP - α G eP Δ T e V eP - V T Δ T = [ L s * I s ] ( t ) + [ L d * I d ] ( t ) + I E ( t ) , (5)
and by the reset condition:
if V eP ≥ V s then V eP : = V r ′ , (6)
where VeP is clamped to V r ′ for the duration of the refractory period Tref after every spike. CeP and GeP are the membrane capacitance and leak conductance. The scaling factor α ensures an equal membrane voltage response to the depolarizing current described by the exponential terms in both models (BS and eP). We consider two versions of these models separately. First, we treat the LIF versions in detail, where we omit the exponential terms in Eqs 2 and 5; specifically, by taking the limit ΔT → 0 (and setting Vs = VT). In the subsequent part we then consider the (full) EIF versions of the BS and eP models. Below we explain in detail how the components of the point model extension are derived: the linear input filters Ls(t), Ld(t), the additional input current equivalent to the field effect, IE(t), and, in case of the (full) EIF type models, the scaling factor α. The analytical expressions of these model components are given in Eqs 10, 13, 20 and 21 (for the LIF case), and in Eqs 22–26 (for the EIF case). To mimic the remaining depolarization along the dendritic cable after each spike, we choose an elevated reset voltage for all point neuron models: V r ′ = ( V r + V T ) / 2.
For comparison we also use a point neuron model (of LIF and EIF type, respectively) without the extension, that is, Ls(t) = Ld(t) = δ(t) and α = 1, and we fit the parameters of that model to best reproduce the activity of the BS model for equal synaptic inputs (details see below). We refer to this model as the P model.
We first consider the BS and eP model neurons of the LIF type (i.e, ΔT → 0, Vs = VT) receiving subthreshold synaptic input at the soma in the absence of an electric field (E(t) = 0, IE(t) = 0, Id(t) = 0). To avoid ambiguity we use the superscript Is for the membrane voltage variables in this case. The somatic membrane voltage response of the BS model (Eqs 1–3) can be calculated as (see Methods)
V ^ BS I s ( 0 , ω ) = I ^ s ( ω ) C s i ω + G s + z ( ω ) g i tanh ( z ( ω ) L ) , (7) z ( ω ) = g m + g m 2 + ω 2 c m 2 2 g i + sgn ( ω ) i - g m + g m 2 + ω 2 c m 2 2 g i , (8)
where . ^ indicates the temporal Fourier transform and ω = 2πf denotes angular frequency. The somatic membrane voltage response of the eP model (Eq 5) is given by
V ^ eP I s ( ω ) = L ^ s ( ω ) I ^ s ( ω ) C eP i ω + G eP . (9)
The dendritic filter Ls required to exactly reproduce the somatic membrane voltage response of the BS model, i.e., V ^ e P I s ( ω ) = V ^ B S I s ( 0 , ω ), must then be equal to ratio of the impedances of both models:
L ^ s ( ω ) = C eP i ω + G eP C s i ω + G s + z ( ω ) g i tanh ( z ( ω ) L ) , (10)
where z(ω) is given by Eq 8. In the following, we choose the membrane capacitance and conductance of the eP model to be equal to the corresponding somatic quantities of the BS model, CeP = Cs, GeP = Gs. To see the necessity of the filter, let us consider the P model (no dendritic filter, L ^ s ( ω ) = 1) whose subthreshold response is given by
V ^ P I s ( ω ) = I ^ s ( ω ) C P i ω + G P . (11)
Because of the additional frequency-dependent term in the denominator of Eq 7 compared to Eq 11, it is not possible to adjust the parameters CP and GP of the P model such that V ^ P I s ( ω ) = V ^ B S I s ( 0 , ω ) for all frequencies ω. The somatic response of the BS model can only be approximated in this case.
Fig 2A shows the impedances, Z m I s ( ω ) ≔ V ^ m I s ( ω ) / I ^ s ( ω ), m ∈ {BS|x = 0, eP, P}, of the three neuron models for an example set of parameter values for the BS model. The two parameters of the P model (CP and GP) were determined by matching the steady-state somatic voltage, Z P I s ( 0 ) = Z B S I s ( 0 ), and minimizing the mean squared distance between Z P I s and Z B S I s over the visualized range of input frequencies. The impedance of the eP model matches the impedance of the BS model exactly while the impedance of the P model deviates substantially, in particular for larger frequencies.
Fig 2B–2D show the amplitudes and phases of the input filter L ^ s ( ω ) for various sets of parameters for the BS morphology. L ^ s ( ω ) is always a high-pass filter, which attenuates the somatic inputs at lower and amplifies them at higher frequencies. This effect is more pronounced for a larger dendritic and a smaller somatic compartment. It becomes stronger with increasing ratio of dendritic over somatic size. Nevertheless, the filter does not differ qualitatively for changes in neuron morphology.
We next compare how well the point neuron models eP and P reproduce the spiking activity of the BS model neuron. For this purpose we consider an in vivo-like fluctuating synaptic input current Is(t) described by an Ornstein-Uhlenbeck process (see Methods). The model outputs are compared over a range of values for the input mean I s 0 and standard deviation σs. The parameter values of the P model were adjusted to best reproduce the spike train of the BS model (see Methods for details). Fig 3A displays the time series of the somatic membrane voltage of the three models in response to the same input currents—a weak (small I s 0, σs) and a strong current (large I s 0, σs). For both input currents, the eP model well reproduces the somatic voltage dynamics of the BS model. Consequently, the spike times are also well reproduced. There is, however, a mismatch between the voltage traces during short periods (of less than approximately 10 ms duration) after spikes have occurred. This discrepancy is a result of the remaining dendritic depolarization after a spike has occurred in the BS model, which is only approximated by the elevated reset voltage V r ′ (see section Models above) in the point neuron models. In comparison, the P model performs worse in reproducing the BS membrane voltage dynamics, particularly the fast fluctuations are poorly recovered. This is expected from the mismatch in the impedance for high frequencies (cf. Fig 2A).
In Fig 3B–3E we compare spiking activity in terms of spike coincidences and spike rates for a wide range of input parameters. We used the spike coincidence measure Γ which quantifies the similarity between two spike trains for a given precision of 3 ms (see Methods). The maximum value of 1 indicates an optimal match, i.e., spike times always coincide, a value of 0 corresponds to pure chance, i.e., the degree of coincidences for two Poisson spike trains with equal rates. The P model was fitted to the BS model for each input (in terms of I s 0, σs) separately. The parameters of the eP model, on the other hand, are constant and do not depend on the input at all. The eP model very accurately reproduces the BS spike times for small spike rates (Γ ≥ 0.9 for small I s 0 and σs), see Fig 3B and 3E. This performance decreases only slightly for increasing σs (noise dominated input) and somewhat stronger for increasing I s 0 (mean dominated input). Generally, Γ decreases with increasing spike rates. This can be attributed to the transient periods after spikes during which the dendritic cable is still loaded and the membrane voltages of both neuron models deviate. Those periods do not depend on the spike rate and therefore have a stronger deteriorating effect when the interspike intervals are smaller. In addition, when σs is small the model neurons spike repetitively in a rather clock-like manner, with comparable rate but most likely out of phase due to mismatches caused by the membrane voltage resets. This helps understand the rather low values of Γ for mean dominated inputs. The spike rate of the BS model is also reproduced quite well by the eP model, which underestimates it only slightly (Fig 3D). Spike coincidence and spiking rate reproduction of the eP model can be improved even further by additionally tuning the reset voltage V r ′ using Γ or the spike rate distance as a cost function. The P model, in comparison, is substantially worse in reproducing the spike times at small spike rates and only slightly better than the eP model for large spike rates (Fig 3B and 3C). The spike rate of the BS model is slightly overestimated by the P model (Fig 3D). Even though the parameters of the P model were optimized in an input-dependent manner the eP model leads to an improved reproduction of the BS spiking activity overall.
In summary, the dendritic cable implements a high pass filter for inputs at the soma. Due to the derived filter for somatic inputs, the eP model—without having fitted any of its parameters—well reproduces the BS model dynamics for subthreshold and suprathreshold inputs. Notably, the computation time required for the BS model was at least 25 times that of the eP model, using measurements on a single core of a desktop computer.
We next consider subthreshold synaptic input at the distal dendrite instead of somatic input, but otherwise the same setup as in the previous section. Here we use superscipt Id for the membrane voltage variables to better distinguish from the previous scenario. The somatic membrane voltage response of the BS model can be expressed as (see Methods)
V ^ BS I d ( 0 , ω ) = I ^ d ( ω ) sech ( z ( ω ) L ) C s i ω + G s + z ( ω ) g i tanh ( z ( ω ) L ) , (12)
where z(ω) is given by Eq 8. In order to reproduce that voltage response using the eP model, for which V ^ e P I d ( ω ) = L ^ d ( ω ) I ^ d ( ω ) / ( C e P i ω + G e P ) (cf. Eq 9), we obtain
L ^ d ( ω ) = ( C eP i ω + G eP ) sech ( z ( ω ) L ) C s i ω + G s + z ( ω ) g i tanh ( z ( ω ) L ) . (13)
As in the previous section, we choose CeP = Cs, GeP = Gs. In contrast to the somatic input filter Ls the filter Ld for distal inputs exhibits low pass properties for various BS morphologies, see Fig 4A. The shape of this filter is largely independent of the soma size. Compared to the attenuation of low frequency in case of somatic input, the filter gain for high frequency dendritic input is much lower. This results in a stronger filtering effect for dendritic inputs than for somatic inputs.
An evaluation of the distal input filter in terms of reproduction of BS spiking activity (Γ and rates) is shown in Fig 4B–4E for a range of input mean I d 0 and standard deviation σd values. For comparison we used the P model (without filter) whose parameters were tuned to best reproduce the spike train of the BS model for each input (i.e., (I d 0, σd)-pair) separately. The eP model very accurately reproduces the BS spike times for small spike rates (Γ ≥ 0.9 for small I d 0 and σd). The accuracy drops somewhat as I d 0 increases, which can be explained as in the previous section. Interestingly, the performance does not deteriorate with increasing spike rate in general; it remains high if the noise intensity σd is sufficiently strong (Γ ≥ 0.8 for σd ≥ 80 pA, independent of I d 0 in the considered range). The spike rate of the BS model is somewhat underestimated by the eP model (Fig 4D). It should be noted that the spike rate reproduction could be substantially improved by an increased reset voltage value V r ′, as the remaining dendritic depolarization after spikes is more pronounced in case of distal input compared to somatic input. The computational speed-up of the eP model here is the same as in the previous section. The P model, in comparison, is less accurate across all inputs (Fig 4B–4D), even though its parameters depend on the input.
In summary, the dendritic cable implements a low pass filter for inputs at the distal dendrite, and due to the corresponding derived filter the eP model reproduces the BS model dynamics for subthreshold and suprathreshold inputs much better than the P model.
We now consider an extracellular electric field—in addition to the synaptic inputs—to which the neuron is exposed to. We characterize the effects of that field on the subthreshold somatic membrane voltage and spiking dynamics of the BS neuron and we determine an explicit expression for the additional input current of the extended point neuron model to reproduce these effects. The electric fields we are interested in are oscillatory, spatially uniform on the neuronal scale and weak such as induced by transcranial brain stimulation [6]. In the following, we consider a field with amplitude E1 and angular frequency φ,
E ( t ) = - ∂ V BS , e ∂ x ( t ) = E 1 sin ( φ t ) . (14)
Recall that VBS,e(x, t) is the extracellular potential. The BS subthreshold somatic membrane voltage response to this field, V B S E ( 0 , t ), is determined by Eqs 1–3. Using the temporal Fourier transform the solution can be expressed analytically as
V ^ BS E ( 0 , ω ) = E ^ ( ω ) g i [ sech ( z ( ω ) L ) - 1 ] C s i ω + G s + z ( ω ) g i tanh ( z ( ω ) L ) , (15)
where z(ω) is given by Eq 8 (see Methods). Note, that we again neglect the exponential current in this section (LIF case, but see next section for the EIF case). In the time domain this yields
V BS E ( 0 , t ) = | A ( φ ) | sin φ t + arg ( A ( φ ) ) , (16) A ( φ ) = E 1 g i [ sech ( z ( φ ) L ) - 1 ] C s i φ + G s + z ( φ ) g i tanh ( z ( φ ) L ) , (17)
where arg(x) denotes the argument of the complex number x. The overall subthreshold response in presence of the electric field and synaptic input can be decomposed as
V ^ BS ( 0 , ω ) = V ^ BS I s ( 0 , ω ) + V ^ BS I d ( 0 , ω ) + V ^ BS E ( 0 , ω ) , (18)
with V ^ B S I s ( 0 , ω ), V ^ B S I d ( 0 , ω ) and V ^ B S E ( 0 , ω ) given by Eqs 7, 12 and 15. For the eP model, on the other hand, we have
V ^ eP ( ω ) = L ^ s ( ω ) I ^ s ( ω ) + L ^ d ( ω ) I ^ d ( ω ) + I ^ E ( ω ) C eP i ω + G eP . (19)
To guarantee an equal subthreshold response in both models, i.e., V ^ e P ( ω ) = V ^ B S ( 0 , ω ), we obtain the following expression for the additional input current,
I E ( t ) = | B ( φ ) | sin φ t + arg ( B ( φ ) ) , (20) B ( φ ) = E 1 g i ( C eP i φ + G eP ) [ sech ( z ( φ ) L ) - 1 ] C s i φ + G s + z ( φ ) g i tanh ( z ( φ ) L ) , (21)
where we set CeP = Cs and GeP = Gs (as in the previous sections). It should be noted that these results are not restricted to sinusoidal field variations, as considered here, and can be easily adjusted for any time-varying or constant description of the electric field using its Fourier transform.
The equivalent input current IE(t) as well as the somatic subthreshold sensitivity to the field, |A(φ)|/E1 and the phase shift between oscillating membrane voltage and field, arg(A(φ)), with A(φ) from Eq 17, are shown in Fig 5. Interestingly, the amplitude of IE(t) increases with increasing field frequency (Fig 5A), while the sensitivity decreases (Fig 5B). The sensitivity curve changes quantitatively, but not qualitatively, with varying neuronal morphology (Fig 5B). Specifically, its dependence on the field frequency becomes more pronounced with increasing ratio of dendritic size over somatic one. The cable length has the strongest impact in this respect. Notably, the morphology parameters can be adjusted such that the sensitivity curve well matches with empirical results obtained from rat hippocampal pyramidal cells in vitro. The phase shift between the somatic membrane voltage and field oscillations also depends on the field frequency. It exhibits an anti-phase relation for slow oscillations, and decreases with increasing frequency (Fig 5B).
We next assess how the electric field affects spiking activity for a range of field frequencies using the BS and eP models. For that purpose, we simulated both model neurons subject to the field and noisy synaptic input at the soma or at the distal dendrite. The synaptic drive alone is strong enough to cause stochastic spiking with rate r0. The oscillatory field leads to an oscillatory spike rate modulation quantified as r1(φ)sin(φt + ψ(φ)) around the constant baseline spike rate r0 (see Methods for details). Note that this spike rate modulation measure is related to the frequently used spike field coherence measure.
The amplitude r1 and phase shift ψ of the spike rate modulation for various somatic inputs (in terms of I s 0 and σs), a range of field oscillation frequencies and two field strengths are shown in Fig 6. The eP model well reproduces the spike rate dynamics of the BS model exposed to the field for all considered field and input parameter values. The amplitude r1 increases linearly with increasing field magnitude E1. In contrast to the subthreshold sensitivity to the field (cf. Fig 5B), the spike rate modulation exhibits a clear resonance in the beta and gamma frequency bands across the different inputs. In other words, the spike rate oscillations are strongest for field oscillations of that frequency range. The amplitude peak is more pronounced for stronger inputs and most prominent when the input is dominated by its mean (large I s 0, small σs). This resonance amplitude rapidly increases with increasing baseline spike rate—by increasing both, mean and standard deviation of the background input from small values—and saturates at about r0 = 30 Hz (Fig 6, center). The resonance frequency shifts rather gradually from the beta to the gamma range as the baseline spike rate increases from a few spikes per second to about 60 Hz. The phase shift ψ varies around π, depending on the input and field frequency. Note that ψ = π implies that the probability of spiking is largest at the trough of the field oscillation. This results from the orientation of the field, which, in case of E(t) = E0 > 0, induces a (hyper-)polarized somatic membrane voltage.
To examine the importance of the specific shape of IE(t), we also considered an alternative sinusoidal input current IE(t) = I1 sin(φt + ϕ) for the eP model. Note that the amplitude and phase shift of that current are constant across different field frequencies. Using that current, the typical resonance of the spike rate modulation due to the field cannot even roughly be reproduced (Fig 6).
Let us now inspect spike rate modulation due to the field in presence of distal dendritic inputs instead of somatic ones. In Fig 7 the results are shown for various distal inputs (in terms of I d 0 and σd). Interestingly, for all considered distal dendritic inputs, spike rate modulation amplitudes increase monotonically with the field frequency over the whole considered range (up to 1 kHz, see Discussion for an explanation). Similarly as for somatic inputs, modulation is strongest for mean dominated (large I d 0, small σd) distal inputs, and the phase shift ψ varies around π. Overall, the eP model well reproduces the modulation observed in the BS model.
In the previous sections, we considered only capacitive and leak currents through the neuronal membrane; the model extension presented there applies to the LIF type model neurons. Here, we consider the BS and eP models described by Eqs 1–3, 5 without neglecting the exponential term, that approximates the voltage dependent sodium current at spike initiation. That is, we derive and evaluate the model extension for model neurons of the EIF type.
To derive the required model components Ls(t), Ld(t), α and IE(t) we linearize the exponential terms in Eqs 2 and 5 around a baseline voltage value V0 and then proceed similarly as above. Specifically, we calculate the subthreshold somatic membrane voltage response of the BS model, using the (temporal) Fourier transform, and obtain four response components: V ^ B S ( 0 , ω ) = V ^ B S I s ( 0 , ω ) + V ^ B S I d ( 0 , ω ) + V ^ B S Δ T ( 0 , ω ) + V ^ B S E ( 0 , ω ), where V B S I s, V B S I d and V B S E denote the voltage response components to Is, Id and E, respectively, and the additional term V B S Δ T is due to the (linearized) exponential term. These four voltage response components are given by the explicit expressions Eqs 41–43 in the Methods section. For the eP model, on the other hand, we can also calculate the subthreshold membrane voltage response in the Fourier domain, V ^ e P ( ω ), given by Eq 48. By requiring equal subthreshold responses, V ^ e P ( ω ) = V ^ B S ( 0 , ω ), we obtain the following explicit expressions for the components Ls, Ld, α and IE, considering the electric field defined in Eq 14:
L ^ s ( ω ) = C eP i ω + G eP 1 - α e V 0 - V T Δ T C s i ω + G s 1 - e V 0 - V T Δ T + z ( ω ) g i tanh ( z ( ω ) L ) , (22) L ^ d ( ω ) = C eP i ω + G eP 1 - α e V 0 - V T Δ T sech ( z ( ω ) L ) C s i ω + G s 1 - e V 0 - V T Δ T + z ( ω ) g i tanh ( z ( ω ) L ) , (23) α = G s G s + tanh ( L / λ ) g i / λ , (24) I E ( t ) = | B ( φ ) | sin φ t + arg ( B ( φ ) ) , (25) B ( φ ) = E 1 g i C eP i φ + G eP 1 - α e V 0 - V T Δ T [ sech ( z ( φ ) L ) - 1 ] C s i φ + G s 1 - e V 0 - V T Δ T + z ( φ ) g i tanh ( z ( φ ) L ) , (26)
where z(ω) is given by Eq 8. The scaling factor α guarantees that the voltage response component caused by the exponential term, V B S Δ T, is reproduced. In other words, α ensures that the spike initiation current, described by the exponential term, leads to an equal steady state in both models. Note that the two filters for EIF neurons and those for LIF neurons depend on input frequency in qualitatively the same way (by comparing Eqs 22 and 23 with Eqs 10 and 13).
We assessed the reproduction of BS spiking activity by the extended EIF model for somatic inputs using the spike coincidence factor Γ and estimated spike rates (Fig 8). Here again the parameter values of the P model were adjusted to maximize ΓBS,P for each input separately. The range of input parameter values was chosen to obtain similar spike rates as in Fig 3. Despite the linearization in the derivation, the eP model achieves a correct reproduction of the BS spike trains (Γ ≥ 0.7 for a wide range of input parameters). In particular, ΓBS,eP is large for small spike rates (small I s 0 and σs) and decreases for increasing I s 0 (towards mean dominated input), see Fig 8A and 8D. The eP model tends to underestimate the firing rate of the BS model (Fig 8C). This discrepancy in the rate could be reduced by optimizing the point model reset voltage, V r ′, to better account for the remaining dendritic cable depolarization in the BS model. Similarly, an improved performance of the eP model in terms of spike train reproduction could be achieved by tuning this reset voltage. The P model, on the other hand, rather poorly reproduces the BS spiking dynamics for small input noise intensity (Γ ≤ 0.6 for σs ≤ 30 pA, see Fig 8A). Overall, also in presence of the exponential term the eP model clearly outperforms the simpler P model for small spike rates (ΓBS,eP − ΓBS,P ≥ 0.3 for small I s 0 and σs) and achieves similar performance for high spiking activity (Fig 8B).
The reproduction of spiking activity of the BS model was also assessed for distal dendritic inputs. The range of input parameters (I d 0 and σd) was adjusted to obtain similar BS spike rates as for the LIF case. The eP model performs well, in particular for small spike rates or sufficiently strong noise intensity; its performance decreases in the mean driven regime (Fig 9A). On the contrary the P model fails to reproduce the BS spiking activitiy (see Fig A in S1 Text for more details).
In summary, the somatic and distal dendritic input filters obtained for EIF neurons are qualitatively similar to the ones obtained for LIF neurons. The eP model, in contrast to the P model, well reproduces the BS model dynamics for subthreshold and suprathreshold inputs—also for the EIF case.
Spike rate modulations due to an oscillatory electric field using EIF type model neurons for synaptic background input at the soma or distal dendrite are displayed in Fig 9 (see also Fig B and Fig C in S1 Text for additional parameter values of the background input). Similarly to the LIF case, spike rate modulation amplitudes do not decrease monotonically with the field frequency. For somatic background input, we find spike rate resonance in the beta and gamma frequency range, similarly as shown by LIF type models. However, in case of distal dendritic input, EIF neurons exhibit resonance peaks in the high gamma frequency band, in contrast to LIF neurons, whose resonance frequency is substantially larger (see Discussion for an explanation). For both input locations the spike rate modulations shown by the BS model are well reproduced by the eP model and resonance amplitudes are stronger for large spike rates (i.e., large I s 0, σs and large I d 0, σd, respectively).
In this contribution we presented an extension for IF point model neurons to accurately reflect the filtering of synaptic inputs caused by the presence of a dendrite and the effects of weak, oscillatory electric fields on neuronal activity. Based on a canonical BS neuron model, we analytically derived additional components for LIF and EIF point neuron models to exactly reproduce the subthreshold voltage dynamics of the spatially extended BS neuron.
These new components consist of (i) two linear filters applied to synaptic inputs depending on their location (soma or distal dendrite) and (ii) an additional input current quantifying the field effect on the membrane voltage. The EIF point model requires an additional scaling parameter to accurately match the BS voltage dynamics. Exhaustive evaluations for suprathreshold in-vivo like fluctuating inputs demonstrated that the BS spiking activity is well reproduced by the extended point neuron model in both cases (LIF and EIF). Optimizing the parameters of the standard LIF and EIF models without the derived extension components, however, does not suffice to adequately reproduce the BS model dynamics.
Due to their computational efficiency the extensions of the point neuron models are well suited for application in large networks to investigate, for example, the effects of neuronal morphology and electrical fields on neuronal spiking activity at the population level. Additionally, our methodological results serve as a building block to derive mean-field descriptions for the collective (spike rate) dynamics of large coupled populations [17, 21, 22, 23, chapter 4.2]. An implementation of the presented models using Python (for the eP model) and NEURON (for the BS model) is freely available at https://github.com/nigroup/IF_extension.
Below, we summarize our results on the obtained input filters and the field effects on neuronal dynamics.
We have demonstrated that synaptic input is integrated at the soma in distinct ways due to the presence of the dendrite, depending on the input site. Distal dendritic input is low-pass filtered (cf. Fig 4A), in accordance with previous results [24], whereas somatic input is high-pass filtered (cf. Fig 3B). The latter effect is consistent with recent measurements from Purkinje cells and with theoretical results [18] which show a similar change in somatic impedance due to the presence of a dendritic tree (Fig. 4 in [18], in comparison with Fig 2A here). Consequently, the presence of a dendrite can lead to an enhanced neuronal spiking response to high-frequency somatic inputs [18], which may be further amplified by the dendritic effect on the sharpness of spikes at the axon initial segment [25]. The derived IF model extension enables efficient analyses of the BS spike rate response to modulations of the input current—which are, however, not within the scope of this paper.
There are two different strategies for taking into account complex neuron morphologies in models while keeping numerical simulation computationally efficient. One option is to reduce the number of compartments while retaining important properties of the dendritic tree [26]. Alternatively, one can extend point neuron models with temporal kernels which are calibrated to reproduce the somatic membrane voltage response to synaptic inputs as observed in complex morphological cells [27, 28]. Our approach is of the latter type, with the advantage that the temporal kernels (filters) are analytically derived from the underlying morphological BS model.
A similar extension for point model neurons to reproduce dendritic input integration of model cells with complex morphology has been recently proposed in [29]. Using the Green’s function formalism a synapse model was developed, whose computational complexity practically allows for only a small number of synaptic input locations. Based on the BS model we were able to derive input filters for point model neurons using only the Fourier transform (without having to rely on the Green’s function) and these filters are simple to implement.
We have demonstrated that our extended model outperforms the simpler point neuron model in terms of spike train reproduction. Overall, it performs well for suprathreshold inputs, particularly in case of distal inputs and for somatic inputs that are not too strong. That performance could be further enhanced by optimizing the reset voltage to better reflect the remaining dendritic membrane depolarization in the BS model after each spike, as was mentioned previously.
In our study we have considered passive dendrites. Nonlinear (spike-generating) currents along the dendrite, which cause nonlinear synaptic input integration [30–32], could be incorporated using our approach in a “quasi-active” framework [24]. This would involve solving the cable equation with linearized nonlinear components, similarly as for the exponential terms used here (EIF case).
We investigated in detail the effects of a spatially homogeneous, oscillating, weak electric field, as induced by transcranial electrical stimulation, on the activity of the BS neuron. Such a one-dimensional spatial (cable plus soma) model provides a good approximation for neurons with elongated (apical) dendrites exposed to a uniform extracellular electrical field as long as the dendritic (apical main) cable is not substantially smaller than its electrotonic length [33, chapter 2.5]. Following the somatic doctrine [6], we focused on the effects of the field that are due to the polarization of the membrane voltage at the soma. We analytically calculated the subthreshold voltage response, whose properties are in accordance with electrophysiological observations: the response magnitude scales linearly with the field amplitude [13], as shown by the sensitivity in Fig 5. This sensitivity is of the same order of magnitude as that measured in pyramidal cells [15], i.e., around 0.30 mm for low frequency fields, and decreases with increasing field frequency in a morphology dependent manner [14]. For non-uniform electric fields, e.g., as generated by point source stimulation, however, the sensitivity can be roughly constant for frequencies up to at least 100 Hz [8]. Interestingly, such a behavior can also be observed for a uniform field in case of a rather short dendritic cable (cf. Fig 5B).
While polarization effects due to direct current fields have been extensively studied [34–36], the effects of time-varying fields are less well understood. The response of the subthreshold membrane voltage to time-varying fields has been calculated in [37] for a finite dendritic cable with leaky currents at one end, and in [38] for a spatially non-uniform field. Using a one-dimensional cable model [33] showed that the electrotonic length is a key quantity that determines the neuronal subthreshold response to an electric field. Specifically, elongated neurons are less sensitive to high frequency fields than compact ones. How the voltage response to an input current at a particular location along the cable depends on input frequency is largely determined by the membrane time constant. In case of an electrical field, however, which corresponds to symmetrical stimulation at both ends of the cable, the voltage response is also strongly influenced by currents flowing through the low-resistant intracellular medium. This results in an enhanced high frequency response to an extracellular field when compared to an input current [33, chapter 5].
Nevertheless, a somatic compartment was not considered in these studies. Using the BS model we have shown that the relative size of the soma compared to the dendritic cable substantially affects the neuronal sensitivity to the field.
Further, we found frequency-dependent spike rate modulation (and hence, spike field coherence) caused by the electric field. Unlike neuronal subthreshold sensitivity, spike rate modulation amplitude did not decrease with the field frequency and its precise relationship to field frequency depended on the synaptic input location. Spike rate modulation exhibited a clear resonance in the beta and gamma frequency bands in presence of only somatic inputs (cf. Fig 6 and B in S1 Text), whereas for only distal dendritic inputs, spike rate modulation amplitudes are strongest at much larger frequencies (cf. Fig 7 and C in S1 Text). This can be linked to a theoretical result showing that the response of single-compartment model neurons to high frequency inputs is stronger for larger autocorrelation times of a fluctuating synaptic input current [39]. Since fluctuating synaptic inputs arriving at the distal dendrite are low-pass filtered, the autocorrelation time of the corresponding input current felt by the soma is increased (or rather limited from below). Spike rate resonance frequencies were lower for EIF neurons as compared to LIF neurons, in particular for background inputs only at the distal dendrite. This may be explained by the fact that the presence of the exponential term, describing the spike initiating sodium current, decreases the rate response to high frequency inputs [19] (see also the analytical results in [18]). In all cases, the amplitude of the modulation also depended on the input strength (input mean and noise intensity), but its relationship to field frequency was not strongly affected by the input parameters. Recently it has been shown that Purkinje neurons, due to their large dendritic trees, exhibit spike rate resonance at rather high frequencies in response to somatic input modulations and in the presence of noisy dendritic input [18, Fig 5], which is qualitatively similar to the field-induced resonance effects described here (cf. Figs 7 and 9C). It should be noted, however, that an oscillatory (spatially uniform) external field corresponds to oscillatory input currents with opposite sign at the soma and the distal dendrite, respectively (cf. Eqs 2 and 3). The effects of the field can thus not be easily anticipated from those of an input current modulation at the soma alone. Furthermore, the dendritic membrane surface compared to the somatic one for Purkinje cells [18] is substantially larger than that of pyramidal neurons as considered here, which additionally impedes to directly relate the results.
Existing experimental studies on the modulation of neuronal activity by extracellular fields have considered a small number of field frequencies (see [40] for a review). Therefore, our results on spike rate resonance are currently not completely confirmed and may be regarded as predictions. In accordance with our findings weak alternating electric fields (of 30 Hz) have been shown to increase the spiking coherence of pyramidal cells in rat hippocampal slices [41], where this increase was proportional to the subthreshold membrane polarization. Moreover, spatially uniform extracellular fields with high-frequency components entrained spiking activity in ferret primary visual cortex more effectively than fields that only contain low-frequency components [1, Fig. S6]. Our predictions on spike rate modulation by an oscillating electric field are thus in agreement with current knowledge and are informative for future experimental studies. Those results may further be of potential interest for the design of transcranial electrical stimulation protocols.
Regarding the point model extension, we analytically derived an expression for an input current to reproduce the effect of the field as extracted from the biophysically grounded BS model. The amplitude and phase of this input current depend on the parameters of the BS neuron and the electric field. Previously, simple phenomenologically obtained input currents have been used for point neuron network simulations, with either constant amplitudes (across frequencies) [1, 9] or amplitudes fitted to electrophysiological data [7]. Interestingly, the latter study used an input current whose magnitude decreases with increasing frequency, in contrast to the equivalent current we obtained (whose magnitude increases with frequency up to 10 kHz). The neuronal subthreshold sensitivity in that study and the ones shown here, however, are similar. This apparent discrepancy in the currents describing the field effect may be explained by the impedance of the applied model neurons, which naturally influences the equivalent input current. In [7] the model parameters (and thus the impedance) were not fitted to real cells; hence it is unlikely that the model impedance matched with the impedance of the cells from which the current amplitudes were estimated [15]. The successful reproduction of the BS spike rate modulation due to the field by the eP model presented here supports the high-pass properties of the equivalent input current.
In the present study, we derived an extension for point neuron models of the LIF and EIF types. Additional model variables with slow dynamics [42] may also be included in this framework, in order to reflect, for example, effects of slowly deactivating potassium channels that mediate spike rate adaptation and associated characteristic neuronal response properties [43, 44]. In that case, a separation of timescales argument could be used to derive the model extension.
The results we extracted from a canonical spatial neuron model provide insight into the effects of cellular morphology on synaptic input integration and the impact of extracellular electric fields on neuronal activity. In particular, the presented point model extension, which is straightforward to implement and efficient to simulate, shall give rise to comprehensive computational investigations of neuronal population activity entrainment due to transcranial stimulation.
The subthreshold voltage dynamics of the eP model is specified by Eq 5 which is complemented by the reset condition 6 together with a refractory period (see Models in the section Results).
|
10.1371/journal.pbio.2000225 | Phylosymbiosis: Relationships and Functional Effects of Microbial Communities across Host Evolutionary History | Phylosymbiosis was recently proposed to describe the eco-evolutionary pattern, whereby the ecological relatedness of host-associated microbial communities parallels the phylogeny of related host species. Here, we test the prevalence of phylosymbiosis and its functional significance under highly controlled conditions by characterizing the microbiota of 24 animal species from four different groups (Peromyscus deer mice, Drosophila flies, mosquitoes, and Nasonia wasps), and we reevaluate the phylosymbiotic relationships of seven species of wild hominids. We demonstrate three key findings. First, intraspecific microbiota variation is consistently less than interspecific microbiota variation, and microbiota-based models predict host species origin with high accuracy across the dataset. Interestingly, the age of host clade divergence positively associates with the degree of microbial community distinguishability between species within the host clades, spanning recent host speciation events (~1 million y ago) to more distantly related host genera (~108 million y ago). Second, topological congruence analyses of each group's complete phylogeny and microbiota dendrogram reveal significant degrees of phylosymbiosis, irrespective of host clade age or taxonomy. Third, consistent with selection on host–microbiota interactions driving phylosymbiosis, there are survival and performance reductions when interspecific microbiota transplants are conducted between closely related and divergent host species pairs. Overall, these findings indicate that the composition and functional effects of an animal's microbial community can be closely allied with host evolution, even across wide-ranging timescales and diverse animal systems reared under controlled conditions.
| Studies on the assembly and function of host-microbiota symbioses are inherently complicated by the diverse effects of diet, age, sex, host genetics, and endosymbionts. Central to unraveling one effect from the other is an experimental framework that reduces confounders. Using common rearing conditions across four animal groups (deer mice, flies, mosquitoes, and wasps) that span recent host speciation events to more distantly related host genera, this study tests whether microbial community assembly is generally random with respect to host relatedness or "phylosymbiotic," in which the phylogeny of the host group is congruent with ecological relationships of their microbial communities. Across all four animal groups and one external dataset of great apes, we apply several statistics for analyzing congruencies and demonstrate phylosymbiosis to varying degrees in each group. Moreover, consistent with selection on host–microbiota interactions driving phylosymbiosis, transplanting interspecific microbial communities in mice significantly decreased their ability to digest food. Similarly, wasps that received transplants of microbial communities from different wasp species had lower survival than those given their own microbiota. Overall, this experimental and statistical framework shows how microbial community assembly and functionality across related species can be linked to animal evolution, health, and survival.
| A large body of literature has documented genetic and environmental influences on the composition of host-associated microbial communities [1–10]. Although environmental factors are considered to play a much larger role than host genetics and evolutionary history [11], host influences and their functional consequences are poorly elucidated and thus require systematic study across host–microbiota systems. Several outstanding questions remain regarding the nature of host effects on microbiota assembly. Are host–microbiota associations stochastically assembled, or might there be deterministic assembly mechanisms that predict these associations? How rapidly do microbiota differences form between closely related host species, and are interspecific microbiota differences prone to decay over evolutionary time? Can host-driven assembly of the microbiota be isolated from confounding variables such as diet, age, sex, and endosymbionts? If there are microbiota differences between species, are they functional in an evolutionarily informed manner, such that mismatches between host and interspecific microbiota lead to reductions in fitness or performance, particularly when interspecific microbiota transplants are conducted between older host species pairs?
If host-associated microbial communities assemble stochastically through environmental acquisition with no host-specific influence, then microbiota compositions across related host species will not differ from expectations based on random community assemblies and dispersal limitations. Therefore, in a common environment, microbiota will form independent of host species (Fig 1A), and any interspecific differences in microbiota composition would be arbitrary. In contrast, if hosts influence a sufficient amount of the composition of the microbiota, then under controlled rearing conditions, intraspecific microbial communities will structure more similarly to each other than to interspecific microbial communities (Fig 1B). Similarly, if microbial communities are randomly established or are not distinguishable with regard to host evolutionary relationships, then dendrograms illustrating beta diversity distance relationships between microbial communities will not parallel the phylogeny of the host species (Fig 1C). However, if microbial communities are distinguishable, then hosts with greater genetic divergence may exhibit more distinguishable microbiota. In this case, there will be congruence between the host phylogeny and microbiota dendrogram (Fig 1D). As this outcome is not likely due to coevolution, cospeciation, or cocladogenesis of the entire microbial community from a last common ancestor, "phylosymbiosis" was proposed as a new term that does not necessarily presume that members of the microbial community are constant, stable, or vertically transmitted from generation to generation [1,12]. Rather, phylosymbiosis refers to an eco-evolutionary pattern in which evolutionary changes in the host associate with ecological changes in the microbiota.
Phylosymbiosis leads to the explicit prediction that as host nuclear genetic differences increase over time, the differences in host-associated microbial communities will also increase. Indeed, phylosymbiosis has been observed in natural populations of sponges [13], ants [10], bats [14], and apes [15,16]. However, other studies on termites [17], flies [18–20], birds [21], and mice [22] have not observed strict patterns of phylosymbiosis or host-specific microbial signatures. In natural population studies, determining the forces driving phylosymbiosis is equivocal, as both environmental and host effects can covary and contribute to microbiota assembly. Importantly, major effects of the environment, age, or sex may overwhelm the ability to detect phylosymbiosis. Indeed, diet is a stronger determinant of whole microbial community structure than genotype in lab-bred mice [23]. Additionally, conjecture about the formation of host-specific communities should be resolved in a wider context, especially their functional significance, as microbiotas may be inconsequential to host biology or uniquely situated for certain host genotypes and fitness. Thus, the prevalence and functional significance of phylosymbiosis is uncertain and requires reductionist approaches to discriminate among the frequently confounded variables of host, environment, development, sex, and even endosymbiont status.
Here, we quantify phylosymbiosis under laboratory conditions to control for environmental and host rearing variation. Prior investigations of phylosymbiosis have not typically controlled for these confounding variables, with the exception of male Nasonia wasps [1,2] and Hydra [5,24]. Specifically, we reared 24 species in the laboratory while controlling for sex (virgin females), age, diet, and endosymbionts, thus removing major environmental variables and isolating the contribution of host species on microbiota assembly. The experimental systems, or “host clades,” span four species of Nasonia parasitic jewel wasps, six species of Drosophila fruit flies, eight species of Anopheles, Aedes, and Culex mosquitoes, and six species of Peromyscus deer mice. An externally derived dataset with seven members of the hominid lineage [16] provides another mammalian and multigenus clade for reference and facilitates examination of natural populations in which phylosymbiosis was previously documented. Together, the five host clades include 31 distinct taxa and span a range of estimated divergence times from 0.2–108 million y. Last, we test the hypothesis that phylosymbiosis represents a functional association through a series of microbial transplants with autochthonous (intraspecific) and allochthonous (interspecific) microbiota in Nasonia and Peromyscus. We expect that an experimentally mediated disruption of phylosymbiosis will have functional costs that may lower host fitness or performance in an evolutionarily informed manner. Our findings demonstrate that a consistent set of controlled experimental and bioinformatic approaches in comparative microbiota studies can isolate host-driven phylosymbiosis.
Phylosymbiosis predicts that host clades will harbor distinguishable microbial communities (e.g., jewel wasps versus fruit flies versus deer mice, etc.) and that more closely related host clades will exhibit more similar microbial communities (e.g., insects versus mammals). Indeed, at a broad scale, we found that host clades harbored relatively distinct microbial communities (Fig 2A, ANOSIM, R = 0.961, p < 1e–6). Furthermore, there was significant microbiota differentiation between the mammalian and invertebrate host clades in the principle coordinates analysis (PCoA) (Fig 2A, ANOSIM, R = 0.905, p < 1e–6). The PCoA shows insect groups separating along two dimensions of a plane, with the mammals distinguished orthogonally from that plane in a third dimension, suggesting that variance in insect microbial communities is fundamentally different than that in mammals. As is well established, the gut communities of mammals were dominated by the bacterial classes Clostridia (Firmicutes) (Fig 2B, hominid 42%, Peromyscus 37%) and Bacteroidia (Bacteroidetes) (Fig 2B, hominid 15%, Peromyscus 37%), while the insect clades were dominated by Proteobacteria (Fig 2B, Drosophila 78%, mosquito 69%, Nasonia 77%). This same bacterial divide is also seen in the network analysis, with significant clustering of the insect microbial communities around Proteobacteria, and the mammal microbial communities around subsets of shared and unique Firmicutes and Bacteroidetes (G-test, p < 1e–6, Fig 2C). Microbial diversity as measured by the Shannon index [25] was approximately 35% higher in mammalian hosts compared to insects, indicating more diverse symbiont communities among the mammalian clades (Fig 2D; Nested analysis of variance [ANOVA]: phylum effect [mammals versus insects]: F1,302 = 419.82, p < 0.001; clade effect nested within phylum: F3,298 = 18.46, p < 0.001; species effect nested within clade and phylum: F26,272 = 7.94, p < 0.001).
We implemented a random forest classifier (RFC) supervised learning algorithm to quantify the degree to which individual microbial communities can be classified into their respective host clade. RFC models show a strong ability to classify microbial communities to their correct host clades based on OTUs (98.5% classification accuracy) (S1 Table). Additionally, models distinguish mammals and insect samples with high accuracy (95.9% classification accuracy) (S1 Table). Cross-validation prevents overfitting by ensuring that classification accuracy is assessed using only samples excluded from model training. We also used RFC models to identify the most distinguishing bacterial taxonomic level for both interclade distinction and the divide between mammals and insects. Genera provided the strongest ability to predict host clade (99.0% classification accuracy) (S1 Table); however, the major groups of insects and mammals were better distinguished by family-level community classification (98.3% classification accuracy) (S1 Table). Taken together, these results illustrate that evolutionary relationships of the host clades broadly covary with differences in microbial communities. While differentiation of the five clades could in part be attributable to varied experimental conditions for each animal group (since they were reared separately), clustering of the vertebrate microbial communities from the insect microbial communities is independent of rearing conditions and suggests a host-assisted structuring of microbial communities.
Phylosymbiosis predicts that an individual’s microbial community will exhibit higher similarity to communities of the same host species than to those from different host species. The degree of similarity can be variable but should correlate with genetic relatedness of the host species. Pairwise comparisons of beta diversity distances between all individuals within each host clade reveal that the average distance between microbial communities within a species is always less than between species (S1 Fig). Summarized beta diversity also reveal lower intraspecific versus interspecific distances, with significant differences observed for all clades (Fig 3A, Each dataset: Mann–Whitney U, p < 1e–6).
We next evaluated intraspecific microbiota clustering through Bray–Curtis beta diversity interrelationships with PCoA and statistically assessed the strength of interspecific microbiota distinguishability with ANOSIM (Fig 3B). Visualization of the first three principle components revealed that individual samples clustered around their respective species’ centroid position. In all host clades, each host species harbored significantly distinguishable microbial communities (Fig 3B, ANOSIM p < 0.001 for all host clades). Notably, the ANOSIM R-values of interspecific microbiota distinguishability within a host clade positively correlated with the maximal age of divergence of the species in the host clades (Fig 3C, Regression Analysis Log Transformed Clade Age, R2 = 0.92, p = 0.006; Untransformed Clade Age, R2 = 0.70, p = 0.048). Thus, host clades with higher total divergence times between species had stronger degrees of microbiota distinguishability, while less diverged host clades exhibited less microbiota distinguishability. For example, with an estimated host divergence time of 108 million y [27], mosquitoes showed the greatest distinguishability of their microbiota. Conversely, in Nasonia jewel wasps, which only diverged between 200,000 and 1 million y ago [28], the relative strength of clustering was less distinct but still statistically significant. The three intermediate aged clades showed corresponding intermediate levels of clustering: Drosophila had an estimated divergence time of 62.9 million y [29], hominids diverged 9 million y ago [30], and Peromyscus diverged 11.7 million y ago [31]. Therefore, the phylosymbiotic prediction that host species will exhibit significant degrees of specific microbiota assembly was supported in these observations, even under highly controlled conditions in the laboratory models. Microbiota specificity was maintained among very closely related and very divergent species, and a connection was observed between the magnitude of host genetic divergence and microbiota similarity.
As microbiota clustering was supported within species across all five animal clades, it should be possible to model the strength of how well communities of bacteria predict their host species and how specific members of the microbiota affect these predictions. We therefore used RFC models trained on the microbiota of each host clade to evaluate classification accuracy (i.e., the percentage of assigning microbiota to their correct host species) and the expected predicted error (EPE, i.e., the ratio of model accuracy relative to random classification). RFC results indicated that the operational taxonomic units (OTUs) for Drosophila and Peromyscus and genus taxonomic levels for hominid, mosquito and Nasonia have the highest classification accuracies, with significant EPE observed for all clades (EPE > 2, S1 Table). At the genus level, the mosquito and Drosophila host clades exhibited the strongest results (mosquito, classification accuracy = 99.8%, EPE = 558.9; Drosophila, classification accuracy = 97.2%, EPE = 31.7). Other host clades demonstrated significant but comparatively lower strength models. The reduced predictive power of these models may be due to a number of factors, such as a lower number of host species (Nasonia, classification accuracy = 88.7%, EPE = 13.4), uneven sample representation from each species (hominid, classification accuracy = 53.4%, EPE = 2.1), and lower sequencing coverage (Peromyscus, classification accuracy = 61.4%, EPE = 2.5).
To determine the most distinguishing genera of the bacterial community, we examined the resulting loss of model classification accuracy when each genus was excluded from RFCs (S2 Table). Distinguishability within the Drosophila, Nasonia, and mosquito clades was driven primarily by genera in Proteobacteria, which represent five (14.0% model accuracy), seven (11.3% model accuracy), and eight (18.2% model accuracy) of the top ten genera, respectively. Three of the ten most distinguishing genera in Drosophila females are from the Acetobacteraceae family (9.5% model accuracy), previously recognized to be “core” microbiota members [19,32]. Three of the twenty most distinguishing genera in Nasonia females were closely related symbionts from the Enterobacteriaceae family (genera: Proteus, Providencia, Morganella; 3.1% model accuracy), consistently found in our previous studies of Nasonia males [1,2]. Eight genera from the phylum Proteobacteria dominate mosquito female distinguishability, primarily three Gammaproteobacteria of the order Pseudomonadales (8.2% model accuracy), and three Betaproteobacteria of the family Comamonadaceae (5.9% model accuracy). Hominid interspecific distinguishability was driven by the phylum Firmicutes, particularly of the order Clostridiales that contains three of the most distinguishing genera (1.5% model accuracy). The genus Allobaculum conferred nearly double the distinguishing power of any other bacteria in Peromyscus (3.8% model accuracy), and it is associated with low-fat diet, obesity, and insulin resistance in mice [33]. As may be expected, genera of the abundant phyla Firmicutes and Bacteroidetes dominated the majority of distinguishability in Peromyscus (10.6% model accuracy), but genera from Proteobacteria in the family Helicobacteraceae comprised four of the top eleven genera (4.4% model accuracy). Overall, microbiota composition can be used to predict host species with high accuracy, and genera commonly observed in other studies of these host clades underlie interspecific distinguishability.
The major prediction of phylosymbiosis is that phylogenetic relatedness will correlate with beta diversity relationships of microbial communities among related host species. Microbiota dendrograms were constructed by collapsing individual samples to generate an aggregate microbial community for each species and then by comparing relationships of their beta diversity metrics. The matching cluster and Robinson–Foulds tree metrics were utilized to calculate host phylogenetic and microbiota dendrogram topological similarity, with normalized distances ranging from 0.0 (complete congruence) to 1.0 (complete incongruence; [34]). Matching cluster weights topological congruency of trees, similar to the widely used Robinson–Foulds metric [34,35]. However, matching cluster takes into account sections of subtree congruence and therefore is a more refined evaluation of small topological changes that affect incongruence. Significance of the matching cluster and Robinson–Foulds analyses was determined by the probability of randomized bifurcating dendrogram topologies yielding equivalent or more congruent phylosymbiotic patterns than the microbiota dendrogram. Additionally, using the same methodology, matching cluster and Robinson–Foulds metrics were evaluated for Bray–Curtis, unweighted UniFrac [36], and weighted UniFrac [36] beta diversity dendrograms at both 99% and 97% clustered OTUs (S2 Fig). The cytochrome oxidase I (COI) gene was used to construct the phylogeny for each host clade, which compared well to established phylogenetic or phylogenomic trees for all species included in the study (Nasonia [27]; Drosophila [28]; hominids [29]; mosquitoes [26]). Peromyscus was further resolved with an additional marker (arginine vasopressin receptor 1A [AVPR1A]) to reflect the latest phylogenetic estimates [37,38].
Nasonia female wasps exhibited an equivalent phylogenetic tree and microbial community dendrogram, representing exact phylosymbiosis (Nasonia wasps, Fig 4A). These results parallel previous findings in Nasonia males [1,2]. Despite congruency, the Nasonia clade has limited topological complexity with only four species, therefore resulting in a relatively marginal significance. Mice also show nearly perfect congruence, with the exception of Peromyscus eremicus (Fig 4B). Drosophila fruit flies (Fig 4C) showed the lowest topological congruency but were still moderately significant. Four of the six species show correct topological relationships, while the microbial community relationships of Drosophila pseudoobscura and D. erecta are topologically swapped. These results are different from previous findings in Drosophila that utilized a different experimental design, set of taxa, and sequencing technology [19]. However, the evidence for phylosymbiosis is tentative in Drosophila as, unlike other clades, there is no significant congruence for either unweighted or weighted UniFrac metrics (S2 Fig). Previous studies detected no pattern of phylosymbiosis across Drosophila species [19], which could be attributed to Drosophila’s constant replenishment of microbes from the environment [18,20] or the dominance by the bacterial genus Acetobacter, which is important for proper immune and metabolic development [19]. The two additional clades, mosquitoes and hominids, showed significant phylosymbiosis (Fig 4D and 4E). Specifically, the mosquitoes showed accurate separation of Culex and Aedes genera from Anopheles, and the topological departures from phylosymbiosis appeared in two of the bifurcations between closely related species. The hominid microbial community dendrogram reflects the correct branching of Gorilla from Homo sapiens, followed by bonobos and chimpanzees, with the exception that one of the chimpanzee subspecies grouped more closely with the bonobo lineage. These results are similar to previous observations that the relationships of the microbial communities parallel those in the host phylogeny [16]. With the exception of Drosophila, which yielded variable evidence for host–microbiota congruence, significant degrees of phylosymbiosis were observed across clades with varying tree similarity metrics and microbiota beta diversity analyses.
Microbiota–host distinguishability and topological congruence does not strictly imply that the phylosymbiotic associations are fitness directed, though it naturally follows that a particular host species may be more ideally suited for an autochthonous versus allochthonous microbiota. We therefore performed a series of microbial transplants to test the prediction that inoculated microbiota from a different species would decrease aspects of host performance or fitness in contrast to inoculated microbiota from the same species. Moreover, if there is selection on host–microbiota interactions such that microbiotas are uniquely or better situated for resident host backgrounds, then transplanted microbiota from a divergent species could drive more pronounced reductions in host functions than transplanted microbiota from a closely related species.
In Peromyscus, we followed a previously established protocol [39] to transplant the microbial communities from six rodent donor species into a single recipient species, P. polionotus, as well as a control group in which the microbial communities from P. polionotus were introduced to intraspecific individuals of P. polionotus. Inventories of fecal microbiota from donor and recipient mice revealed that portions of the donor microbiota successfully transferred. The estimated amount of transplanted OTUs and their relative abundance ranged from 6.5%–26.2% and 11.4%–40.7%, respectively, when analyzed at the 99% OTU cutoff level. Variation in the transfer of foreign microbes was dependent on donor species and its divergence from the recipient species (S3 Fig). We then measured dry matter digestibility, or the proportion of food material that is digested by the animal. Consistent with selection on host–microbiota interactions, mice that were inoculated with microbial communities from more distantly related hosts exhibited decreased dry matter digestibility (Fig 5). These results were only significant when the group receiving feces from P. eremicus donors was removed (Fig 5). Notably, the microbiota of P. eremicus is not congruent with our predictions of phylosymbiosis (Fig 4). Thus, only the taxa showing phylosymbiosis exhibited the functional trend with digestibility. Distantly related donor species (Neotoma lepida and Mus musculus) did not drive significance, as the correlation remained statistically significant when investigating only Peromyscus donors (excluding P. eremicus; Fig 5).
In the most extreme cases in which mice were inoculated with the microbial communities from P. californicus or M. musculus, there was approximately a 3% decrease in dry matter digestibility, which is on par with the decrease in digestibility observed as a result of helminth infections in Peromyscus [40]. Animals must consume more food to meet energy demands when faced with decreases in digestibility. Indeed, mice inoculated with microbial communities from P. californicus or M. musculus exhibited significantly higher food intakes than the control group (S4 Fig; Tukey’s honest significant difference (HSD) test: p = 0.001 for P. californicus to P. polionotus; p = 0.044 for M. musculus to P. polionotus). The mice inoculated with the microbes from P. eremicus performed just as well, if not better, than the control groups in terms of dry matter digestibility (Fig 5) but still had slightly higher food intakes (S4 Fig).
In Nasonia, we used an in vitro rearing system to transplant heat-killed microbial communities from three Nasonia donor species into larvae of N. vitripennis or N. giraulti [41]. We then measured the survival of the recipients from first instar larva to adulthood. In both N. vitripennis and N. giraulti hosts, interspecific microbiota transplantations exhibited significant decreases in survival to adulthood when compared to intraspecific microbial transplantations (Fig 6). Specifically, N. giraulti with a N. vitripennis microbiota yielded a 24.5% average survival decrease in comparison to a N. giraulti microbiota (Fig 6A, Mann–Whitney U, p = 0.037). Interestingly, N. giraulti with a microbiota from the more closely related N. longicornis exhibited a similar but nonsignificant survival reduction (23.7%, Fig 6A, Mann–Whitney U, p = 0.086). N. vitripennis with a N. giraulti or N. longicornis microbiota exhibited a 42.6% (Fig 6B, Mann–Whitney U, p < 0.0001) and 23.3% (Fig 6B, Mann–Whitney U, p = 0.003) average survival decrease in comparison to a N. vitripennis microbiota, respectively (Fig 6A, Mann–Whitney U, p < 0.0001). Comparisons were also made between noninoculated hosts and those inoculated with interspecific backgrounds (N. giraulti background: N. vitripennis inoculum p = 0.07, N. longicornis inoculum p = 0.26; N. vitripennis background: N. giraulti inoculum p = 0.001, N. longicornis inoculum p = 0.15).
Under phylosymbiosis, host-associated microbial communities form, in part, as a result of interactions with the host rather than through purely stochastic processes associated with the environment. Specifically, we predicted that given closely related animals reared in controlled environments, the relationships of the microbiota would be congruent with the evolutionary relationships of the host species. Previous evidence for phylosymbiosis under controlled regimes existed in Nasonia [1,2] and Hydra [24], and wild populations of sponges [13], ants [10], and apes [15,16] also exhibited this pattern. Here, in a comprehensive analysis of phylosymbiosis in a diverse range of model systems, we report the widespread occurrence of this pattern under strictly controlled conditions as well as a functional basis in the context of host digestive performance in mice and survival in wasps. These results represent the first evidence for phylosymbiosis in Peromyscus deer mice, Drosophila flies, a variety of mosquito species spanning three genera, and Nasonia wasp females with the inclusion of N. oneida. Previous studies in Nasonia measured male phylosymbiosis and did not include N. oneida [1,2]. By rearing closely related species from the same host clade in a common environment, and by controlling age, developmental stage, endosymbiont status, and sex, the experiments rule out confounding variables that can influence microbiota relationships in comparative analyses. Eliminating these variables is important because they often substantially correlate with interspecific differences. Thus, our findings demonstrate that a uniform experimental and bioinformatic methodology can excavate host effects on phylosymbiosis from other potentially confounding variables in comparative microbiota studies.
We observed marked differences in microbial diversity and community structure between mammalian and invertebrate host clades. Mammalian communities were more diverse and dominated by Bacteroidetes and Firmicutes, while insect-associated communities were less diverse and primarily dominated by Proteobacteria. These results are consistent with previous microbial inventories conducted in mammals and insects [6,42]. Together, these findings suggest large-scale differences in the host–microbiota interactions between mammals and insects. These differences across host phyla could be due to a variety of possibilities, including host genetics, diet, age, and rearing environment.
To remove confounding variables that structure host–microbiota assemblages and to rigorously test phylosymbiosis, we utilized an experimental design within four host clades that isolated the effects of host evolutionary relationships from other effects (i.e., diet, age, rearing environment, sex, endosymbionts). We found that host species consistently harbored distinguishable microbiota within each host clade. Additionally, we found significant degrees of congruence between the evolutionary relationships of host species and ecological similarities in their microbial communities, which is consistent with the main hypothesis of phylosymbiosis. These results importantly expand previous evidence for this eco-evolutionary pattern and demonstrate that related hosts reared under identical conditions harbor distinguishable microbial assemblages that can be likened to microbial community markers of host evolutionary relationships. It is conceivable that recently diverged species (i.e., those younger than several hundred thousand years) would have less genetic variation and fewer differences in microbiota composition. Furthermore, divergent hosts may have vast differences in physiology that overwhelm the likelihood of observing phylosymbiosis. Surprisingly, we observed phylosymbiosis to varying degrees in all host clades, and the age of clade divergence positively correlates with the level of intraspecific microbiota distinguishability. Thus, as host species diverge over time, microbial communities become more distinct [1,12], and the limits of detecting phylosymbiosis may occur at extreme scales of incipient or ancient host divergence times.
The mechanisms by which phylosymbiosis is established requires systematic investigation. Perhaps the most apparent regulator of host–microbiota interactions is the host immune system. A previous study of phylosymbiosis in Hydra demonstrated that antimicrobial peptides of the innate immune system are strong dictators of community composition, and expression of antimicrobial peptides are necessary for the formation of host-specific microbiota [5]. Furthermore, genome-wide association studies in humans [43], mice [8], and Drosophila [44] have identified a large immune effect in which host immune genes can explain variation in microbial community structure. Interestingly, host immune genes often exhibit rapid evolution and positive selection compared to genes with other functions [45,46]. While this trend is often explained by the host–pathogen arms race [45], it is also likely due to host evolutionary responses for recruiting and tending a much larger collection of nonpathogenic microbes.
Other host pathways may also underlie the observed species-specific microbiota signatures. Hosts produce glycans and mucins on the gut lining that may serve as biomolecular regulators of microbial communities [47,48]. For example, knocking out the gene for α1–2 fucosyltransferase inhibits production of fucosylated host glycans on the gut surface and significantly alters microbial community structure [49]. Additional knockout studies have demonstrated the roles of circadian clock genes [50], microRNAs [51], and digestive enzymes [52] in determining microbial community structure. These various physiological systems might also interact with one another and may have even evolved in tandem to regulate microbial community structure.
Alternatively, rather than hosts “controlling” their microbiota, microbes may be active in selecting which host niches to colonize. For example, hosts have been compared to ecological islands, where environmental selection of the microbiota through niche availability may occur [53]. However, given the large number of studies that demonstrate the role of microbes in improving host performance [54], we find it unlikely that hosts would assume a solely passive role in these interactions. An elegant study allowed microbial communities from various environments (soil, termite gut, human gut, mouse gut, etc.) to compete within the mouse gut [55]. This study found that a foreign community of the human gut microbiota exhibited an early competitive advantage and colonized the mouse gut first. Later, the mouse gut microbiota dominated and outcompeted the human gut microbiota [55]. Thus, community assembly is not a monolithic process of host control but likely a pluralistic combination of host control, microbial control, and microbe–microbe competition. In this context, both population genetic heritability and community heritability measurements of the microbiota will be useful in prescribing the varied genetic influences of a foundational host species on microbiota assembly [56].
The acquisition route of microbes could also influence our understanding of phylosymbiosis. If phylosymbiosis is observed when the microbiota is acquired horizontally from other hosts, the environment, or some combination of the two, then phylosymbiosis is presumably influenced by host-encoded traits such as control of or susceptibility to microbes. However, maternal transmission of microbes is argued to be a common trend in animals [57]. For example, sponges exhibit vertical transmission of a diverse set of microbes in embryos [58]. Transmission of full microbial communities is unlikely in most systems, given that the communities of developing animals tend to exhibit markedly lower diversity and distinct community structure compared to adults [1,59,60]. Thus, it is improbable that phylosymbiotic relationships are explained simply by community drift over host evolutionary divergence. There could be a subset of microbial taxa that are more likely to be transmitted from mother to offspring that in turn affect what other microbes colonize. For instance, in humans, the family Christensenellaceae is situated as a hub in a co-occurrence network containing several other gut microbes and has a significant population genetic heritability [61]. When Christensenella minuta was introduced into the guts of humanized mice, the microbial community structure was significantly altered [61]. This microbe, as well as others, can therefore be likened to a keystone taxa or "microbial hub" that can impact community structure despite low abundance [61–63]. Thus, one could hypothesize that phylosymbiotic relationships in some systems may be driven by host transmission of microbial hubs that determine whole community structure through ensuing microbe–microbe interactions. However, further work is needed to test this hypothesis.
The congruent relationships between hosts and associated microbial communities are likely maintained through their positive effects on host performance and fitness but could be neutral or harmful as well. While the importance and specificity of hosts and microbes in bipartite associations has been demonstrated on host performance [64], it is unclear whether such effects commonly occur for hosts and their complex microbial communities. If they exist, disruption of phylosymbiosis via hybridization or microbiota transplants should lead to reduced fitness or performance. For instance, hybridization experiments demonstrate negative interactions or "hybrid breakdown" between host genetics and the gut microbiota that drives intestinal pathology in house mice [65] and severe larval lethality between N. vitripennis and N. giraulti wasps [2]. Furthermore, transplant experiments show that all microbes are not equal for the host. An early study demonstrated that germ-free rabbits inoculated with a mouse gut microbiota exhibited impaired gastrointestinal function compared to those given a normal rabbit microbiota [66]. Together, these functional studies and others suggest that interactions between hosts and their microbiota are not random and instead occur at various functional levels.
Here, we add an evolutionary component to these ideas by demonstrating that microbial communities from more evolutionarily distant hosts can be prone to more pronounced reductions in host performance or fitness. Specifically, Peromyscus deer mice inoculated with microbial communities from more distantly related species tended to exhibit lower food digestibility. The exception to this trend was the P. eremicus to P. polionotus group, which did not exhibit any decrease in digestibility. It should be noted that P. eremicus also did not follow phylosymbiosis (Fig 4B), which may explain the departure from our expected trend in digestibility. For example, deviations from phylosymbiosis could be due to a microbial community assembly that is inconsequential to host digestibility. Therefore, transferring a nonphylosymbiotic community between host species may not yield performance costs.
An alternative explanation for our results could be that hosts are acclimated to their established microbiota, and the introduction of foreign microbiota either elicits a host immune response or disrupts the established microbiota, thus decreasing digestibility. One technique to distinguish between adaptation and acclimation would be to conduct experiments in germ-free P. polionotus recipients. However, the derivation of germ-free mammals is a difficult and expensive process [67] and has not been conducted for Peromyscus. Earlier studies utilizing germ-free mammals demonstrate that microbial communities from evolutionarily distant hosts negatively impact gastrointestinal function [66] and immune development [68], thus supporting our hypothesis of functional matching between host and the gut microbiota.
Additionally, among very closely related species, Nasonia exposed to interspecific microbiota have lower fitness than those exposed to intraspecific microbiota. While this experiment utilized heat-killed bacteria to avoid shifts in the microbiota composition during media growth, the protocol is sufficient to test the predictions of phylosymbiosis. First, isolated microbial products can exert drastic effects on eukaryotic partners. For example, a sulfonolipid purified from bacteria can induce multicellularity in choanoflagellates [69]. Additionally, the insect immune system can respond with strain-level specificity to heat-killed bacteria [70]. Therefore, we hypothesize that each Nasonia host species evolved to the products of their own gut microbiota rather than those of gut microbiota from related host species. Together, results from the Peromyscus and Nasonia functional experiments reveal the importance of host evolutionary relationships when considering interactions between hosts and their gut microbial communities and ultimately the symbiotic processes that can drive adaptation and speciation [71,72]. The molecular mechanisms underlying the functional bases of phylosymbiosis in various systems demand further studies
Overall, we have established phylosymbiosis as a common, though not universal, phenomenon under controlled rearing with functional effects on host performance and survival. It is worth emphasizing again that this term is explicit and different from many other similar terms, such as coevolution, cospeciation, cocladogenesis, or codiversification [73]. While cospeciation of hosts and specific environmentally or socially acquired microbes—e.g., hominids and gut bacterial species [74] or the bobtail squid and Vibrio luminescent bacteria [75]—could contribute in part to phylosymbiosis, concordant community structuring with the host phylogeny is not dependent on parallel gene phylogenies but instead on total microbiota compositional divergence. Phylosymbiosis does not assume congruent splitting from an ancestral species because it does not presume that microbial communities are stable or even vertically transmitted from generation to generation [1,12]. Rather, phylosymbiosis predicts that the congruent relationships of host evolution and microbial community similarities could have varied assembly mechanisms in space and time and be newly assembled each generation (though see our discussion of transmission routes above). Moreover, the findings here imply that across wide-ranging evolutionary timescales and animal systems, there is a functional eco-evolutionary basis for phylosymbiosis, at least under controlled conditions.
It may be difficult to detect phylosymbiosis in natural populations because of extensive environmental variation that overwhelms the signal. We suggest that one way to potentially overcome this challenge is to start with laboratory-controlled studies that identify (i) phylosymbiotic communities and (ii) the discriminating microbial taxa between host species. Resultantly, investigations can test whether these microbial signatures exist in natural populations, albeit perhaps in a smaller fraction of the total microbiota that is mainly derived by environmental effects. Another advantage of controlled studies is that the functional effects, both positive and negative, of a phylosymbiotic community assembly can be carefully measured in the context of host evolutionary history.
Procedures involving functional microbiota transplants in Peromyscus mice were approved by the University of Utah Institutional Animal Care and Use Committee under protocol 12–12010. Mice obtained from the Peromyscus Genetic Stock Center were reared under IACUC approved protocols, and only fecal samples were directly utilized. While our paper contains data for several primate species, this data was conducted by another research group, has been previously published, and is now publicly available. Thus, there was no requirement of approved protocols for the primate species.
Nasonia were reared as previously described [2]. Four strains were used: Nasonia vitripennis (strain 13.2), N. longicornis (IV7U-1b), N. giraulti (RV2x(u)), N. oneida (NAS_NONY(u)). To collect individuals for microbiota analysis, virgin females were sorted as pupae into sterile glass vials and collected within the first 24 h of eclosing as adults. Subsequently, they were rinsed with 70% ETOH for 2 min, a 1:10 bleach solution for 2 min, followed by two rinses in sterile water. Individuals were then placed in 1.5 ml tubes and flash frozen in liquid nitrogen. They were then stored at –80°C until DNA extractions. Fifty individuals were collected per strain.
Nine strains of Drosophila were obtained from the University of California San Diego Drosophila Species Stock Center. Six strains were used in the microbiome analysis because they were Wolbachia-free: Drosophila melanogaster (Strain Dmel, stock number 14021–0248.25), D. simulans (Dsim, 14021–0251.195), D. yakuba (Dyak, 14021–0261.01), D. erecta (Dere, 14021–0224.01), D. pseudoobscura (Dpse, 14011–121.94), and D. mojavensis (Dmow, 15081–1352.22). The three strains that tested positive for Wolbachia (method described below) were: D. sechellia (14021–0248.25), D. ananassae (14021–0371.13), and D. willistoni (14030–0811.24). All strains were reared on a cornmeal media (Drosophila Species Stock Center: http://stockcenter.ucsd.edu/info/food_cornmeal.php) with a sterile Braided Dental Roll (No. 2, Crosstex, Atlanta, Georgia, US) inserted into the surface of the media. All stocks were incubated at 25°C with a 12-h light–dark cycle and monitored every 24 h. Every 14 d, stock vials were cleared of any emerged adults, and 6 h later, ten virgin females and three males were transferred to new food vials. This conditioning on the same food was done for five generations before setting up media vials for sample collection. For each of the six strains, five virgin females were mated with two males and allowed to oviposit for 24 h; afterwards, the parents were removed and the vials were incubated as per above.
After 12 d, vials were cleared and virgin females were collected every 4–6 h over a 36-h period. All females were rinsed with 70% ETOH for 2 min, a 1:10 bleach solution for 2 min, followed by two rinses in sterile water. Individual adult flies were then placed in 1.5 ml tubes and flash frozen in liquid nitrogen. They were then stored at –80°C until DNA extractions. Approximately 25–30 virgin adult females were collected per strain.
Mosquitoes were acquired from the Malaria Research and Reference Reagent Resource Center as eggs on damp filter paper within 24 h of being laid. Eight strains were used: Anopheles funestus (strain name FUMOZ), An. farauti s.s. (FAR1), An. quadrimaculatus (GORO), An. arabiensis (SENN), An. gambiae (MALI NIH), Aedes aegypti (COSTA RICA), Ae. albopictus (ALBO), and Culex tarsalis (YOLO F13). Eggs were floated in 350 ml of sterile water with 1.5 ml of 2% yeast slurry and autoclaved within a sterile and lidded clear plastic container. Containers were enclosed within a larger sterile clear container and placed inside an incubator set at 25°C with a 12-h light–dark cycle and monitored every 24 h. After 48 h, the hatched larvae were sorted out and 100–150 of each species were placed in new sterile water (150 ml) with 30 mg of powdered koi food (Laguna Goldfish & Koi all season pellets). Water level was maintained at 150 ml, and larvae were fed 30 mg of powdered koi food every day for a total of 13 d. All pupae were discarded (frozen and autoclaved) on day 10, and new pupae were collected every 12 h on day 11, 12, and 13. Water samples were also collected and frozen for microbial analysis on day 11.
To collect individuals for microbiota analysis, pupae were sorted according to sex, and all females were rinsed with 70% ETOH for two min, then 1:10 bleach solution for two min, followed by two rinses in sterile water. Individual pupae were then placed in 1.5 ml tubes and flash frozen in liquid nitrogen. They were then stored along with their corresponding water sample at –80°C until DNA extractions. Ten to 25 individuals were collected per strain.
Fecal samples were collected from the Peromyscus Genetic Stock Center at the University of South Carolina. Six stock species of Peromyscus were used: P. maniculatus (stock BW), P. polionotus subgriseus (PO), P. leucopus (LL), P. californicus insignis (IS), P. aztecus hylocetes (AM), and P. eremicus (EP). All mice were reared using their standard care practices at the stock center on the same mouse chow diet. Cages were cleaned at regular intervals for all species, and all species were caged within the same facility. Individuals from nonmating cages of females (five to six per cage) were used for collections.
Fecal pellets were collected on a single morning from individual mice directly into a sterile tube and placed on dry ice before being stored at –80°C for 24 h. Samples were then shipped overnight on dry ice and again stored at –80°C until DNA extractions. One to three pellets from 15 individuals were collected per strain.
In order to eliminate the introduction of confounding factors and exclude any subjects that had a pinworm infection at the time of sample collection, we conducted a screen to confirm the pinworm status of each mouse. Pinworm status was confirmed by PCR. Primers utilized to amplify the 28S rDNA D1 and D2 domains of multiple pinworm species were developed and confirmed with positive DNA samples of Syphacia obvelata and Aspiculuris tetraptera (received from the Feldman Center for Comparative Medicine at the University of Virginia). The C1 primer 5ʹ-ACCCGCTGAATTTAAGCAT-3ʹ and the D1 primer 5ʹ-TCCGTGTTTCAAGACGG-3ʹ were amplified under the following reaction conditions: 94°C for 1 min; 35 cycles of 94°C for 30 s, 55°C for 30 s, 72°C for 30 s; and a final elongation time at 72°C for 2 min. The resultant samples were then visualized on a 1% agarose gel. Of the 84 fecal specimens analyzed, 8 of the samples showed amplification at 750 bp corresponding to the expected amplification size of the pinworm DNA sequence. For confirmation, the 750 bp bands were extracted using a Wizard Gel Extraction Kit (Promega Corporation, Madison, Wisconsin, US) and sequenced (GENEWIZ, Inc, New Jersey, US). Sequence results confirmed the presence of Aspiculuris tetraptera infection, and these 8 samples and were excluded from further analysis.
The presence or absence of Wolbachia was checked using two replicates of three individuals per species. DNA extraction was performed with PureGene DNA Extraction Kit (Qiagen), and fragments of the 16S rDNA gene were PCR amplified using primer set WolbF and WolbR3 [76]. Only stock strains that were Wolbachia negative were used in the experiments.
Individual insects (and the mosquitoes’ corresponding water samples) were mechanically homogenized with sterile pestles while frozen within their collection tube. The samples were then thawed to room temperature for 30 s and flash frozen again in liquid nitrogen with additional mechanical homogenization. The samples were finally processed using the ZR-Duet DNA/RNA MiniPrep Kit (Zymo Research, Irvine, California, US). Samples were then quantified using the dsDNA BR Assay kit on the Qubit 2.0 Fluorometer (Life Technologies).
The PowerSoil DNA isolation kit (Mo Bio Laboratories, Carlsbad, California, US), was utilized to extract DNA from 20 mg of mouse fecal material per sample according to manufacturer’s protocol after being mechanically homogenized with sterile pestles while frozen within their collection tube. Samples where then quantified using the dsDNA BR Assay kit on the Qubit 2.0 Fluorometer.
Total genomic DNA was quantified using dsDNA HS Assay kit on the Qubit. Using two μl of DNA, a 20 μl PCR reaction of 28S general eukaryotic amplification was conducted on each sample, with only 25 cycles. Products were purified using Agencourt AMPure XP, quantified using the dsDNA HS Assay kit on the Qubit, and compared to the amount of 16S amplification from the same DNA volume and PCR reaction volume as previously described [2]. PCR amplification of the bacteria 16S rRNA was performed with the 27F 5ʹ-AGAGTTTGATCCTGGCTCAG-3ʹ and 338R 5ʹ-GCTGCCTCCCGTAGGAGT-3ʹ “universal” bacterial primers with the NEBNext High-Fidelity 2X PCR Master Mix; duplicate reactions were generated per sample, which were pooled together postamplification. For sequencing runs 1 (Peromyscus) and 2 (Nasonia, mosquito, and Drosophila), 16S PCR products that were made into libraries had their concentrations normalized relative to about 1,000 ng/ml and 2,000 ng/ml of the 28S quantity for library prep respectively.
Using the Encore 384 Multiplex System (NuGEN, San Carlos, California, US), each samples’ 16S product was ligated with Illumina NGS adaptors and a unique barcode index (after the reverse adaptor). The samples were then purified using Agencourt AMPure XP and quantified using the dsDNA HS Assay kit on the Qubit. Samples were subsequently pooled.
Each pooled library was run on the Illumina MiSeq using either the MiSeq Reagent Kit V2 or V3 for paired-end reads. Run 1 was conducted at the University of Georgia Genomics Facility and run 2 was conducted at Vanderbilt Technologies for Advanced Genomics (VANTAGE).
Sequence quality control and OTU analyses were carried out using QIIME version 1.8.0 [77]. Forward and reverse paired-end sequences were joined and filtered if they met the following criteria: they fell below an average Phred quality score of 25, contained homopolymer runs or ambiguous bases in excess of 6 nucleotides, or were shorter than 200 base pairs. Sequences were also removed if there were errors in the primer sequence or if barcodes contained errors and could not be assigned to a sample properly. A total of 5,065,121 reads passed quality control for the meta-analysis, with an average read length of 310 ± 48 nucleotides. Drosophila: 648,676 reads, average length 315 ± 23. hominid: 1,292,542 reads, average length 247 ± 38. mosquito: 664,350 reads, average length 328 ± 19. Nasonia: 864,969 reads, average length 322 ± 15. Peromyscus: 295,752 reads, average length 347 ± 12.
Chimeric sequences were evaluated and removed using the UCHIME algorithm [78] for the intersection of de novo and GreenGenes 13_5 non-chimeras [79]. The sequences were then clustered into OTUs at 94%, 97%, and 99% similarity using the USEARCH open-reference method [80]. OTUs were mapped at the respective percent against the GreenGenes 13_5 database and screened for a minimum group size of two counts, with dereplication based on full sequences [79]. Representative sequences were chosen as the most abundant representative in each OTU cluster and aligned using GramAlign [81]. A phylogenetic tree of the representative sequences was built in QIIME [77] with the FastTree method and midpoint rooting [82]. Taxonomy was then assigned to the OTU representatives with the UCLUST method against the GreenGenes 13_5 database [79]. OTU tables were constructed in QIIME [77] and sorted by sample IDs alphabetically.
OTU tables were screened to remove any OTUs classified as chloroplast, unassigned, and Wolbachia. Individual samples were assessed for low sequence coverage affecting community profiles and diversity as well as for processing errors based on minimum count thresholds assessed against group means. Following rarefaction, counts were subsequently chosen as the highest rarefaction number allowed by the smallest sample’s count representation in each respective clade and the meta-analysis. Alpha diversity was measured using Shannon and Chao1 metrics generated with the QIIME alpha_rarefaction script. Plots of alpha diversity at a range of rarefied levels were used to assess and remove samples with low diversity.
The PCoA (Fig 2A) components for the meta-analysis were constructed using the QIIME jackknifed_beta_diversity script. The OTU table first underwent rarefaction, followed by the computation of Bray–Curtis beta diversity distances for each rarefied table. PCoA plots of the first three coordinate dimensions were generated using a custom Python script. Individual samples are each depicted as a point and are colored by host clade of origin.
The community profile (Fig 2B) for the meta-analysis was generated using a custom Python script and BIOM tools [83]. OTU tables were first converted to relative abundance for each sample, and bacterial taxonomy was collapsed at the class level. Bacterial classes were sorted alphabetically, and a stacked bar chart representing the relative abundance for each sample was constructed.
The network analysis (Fig 2C) was visualized using Cytoscape [84]. OTU tables were first collapsed by bacterial taxonomy at the genus level, and QIIME’s make_otu_network script was used to construct connections between each bacterial genus to individual hosts based on relative abundance. Network files were then imported into Cytoscape, where the network was computed using an edge-weighted force directed layout. Nodes were colored by host clade, and connections were colored by key bacterial phylum observed in high abundance (i.e., Actinobacteria, Bacteroidetes, Firmicutes, Proteobacteria) and gray for additional phylum.
Alpha diversity plots (Fig 2D) were prepared using the Phyloseq package [85]. OTU tables collapsed by host species were imported into Phyloseq, and the plot_richness function was used to generate box-and-whisker plots of Shannon alpha-diversity. Plots were colored by host clade of origin.
Microbiota dendrograms were constructed using the QIIME jackknifed_beta_diversity script. OTU table counts were first collapsed by host species of origin to get representative species microbiota profiles. The pipeline script performed 1,000 rarefactions on each table and calculated Bray-Curtis beta diversity distances for each. Bray–Curtis distance matrices were UPGMA clustered to give dendrograms of interspecific relatedness. The role of 97% versus 99% OTU clustering cutoffs and weighted and unweighted UniFrac beta diversity measures (S2 Fig) were evaluated for Robinson–Foulds and matching cluster congruence with host phylogeny.
Host phylogenetic trees were constructed using sequences for each host species’ cytochrome oxidase gene downloaded from the NCBI. COI was chosen as a highly conserved molecular marker, and it is widely used for interspecific phylogenetic comparison [86]. Sequences were initially aligned using Muscle v3.8.31 [87]. Gap positions generated through inserts and deletions were removed, and overhanging sequence on 5ʹ and 3ʹ ends were trimmed. Models of molecular evolution were evaluated using jModelTest v2.1.7 [88], and the optimal model was used for final alignment and tree building in RaxML v8.0.0 [89]. The Nasonia and Peromyscus clades were carried out using the same methodology—except for final alignment and tree building in PhyML v3.0 [90]—and for Peromyscus the AVPR1A gene was concatenated with COI to further resolve the phylogeny. All trees are concordant with well-established phylogenies from literature references noted in the Results section.
Quantifying congruence between host phylogeny and microbiota dendrogram relationships (Fig 4) was carried out with a custom Python script and the TreeCmp program [91]. The topologies of both trees were constructed, and the normalized Robinson–Foulds score [35] and normalized matching cluster score [34] were calculated as the number of differences between the two topologies divided by the total possible congruency score for the two trees. Next, 100,000 random trees were constructed with the same number of leaf nodes, and each was compared to the host phylogeny. The number of trees which had an equivalent or better score than the actual microbiota dendrogram were used to calculate the significance of observing that topology under stochastic assembly. Normalized results of both statistics have been provided to facilitate comparison. Matching cluster and Robinson–Foulds p-values were determined by the probability of 100,000 randomized bifurcating dendrogram topologies yielding equivalent or more congruent phylosymbiotic patterns than the microbiota dendrogram.
Within each clade, the Bray–Curtis distances calculated by the jackknife_beta_diversity script (Fig 3A) were separated by those that compared microbiota within a host species and those that compared between host species. The box-and-whisker plots were constructed in Python. Coloring indicates host clade of origin, and all intraspecific and interspecific distances are represented for each clade. These distances were then compared between the groups using a nonparametric, two-tailed Mann–Whitney U test implemented in SciPy [92,93].
To evaluate intraspecific clustering (Fig 3B), the ANOSIM test was used to calculate the distinguishability of Bray–Curtis distances based on species of origin. Bray–Curtis distance matrices were generated using the QIIME jackknifed_beta_diversity script on tables of individuals rarefied 1,000 times. The QIIME script compare_categories was used to calculate ANOSIM scores using the Bray–Curtis distance matrix and host species as categories. 1,000 permutations were used to calculate the significance of clustering for each clade. Three-dimensional PCoA plots were generated in Python using components generated from Bray–Curtis distance matrices in QIIME, and the first three components are shown. Points are colored by host species within each clade, and colors correlate with the species labels in Fig 4 for reference.
A general linear regression was performed to test the correlation between age of clade origin and the intraspecific clustering measured through ANOSIM R-statistic scores. Cladogenesis Age was Log10 transformed to normalize the distance scale between samples (1, 10, 100 MYA). The regression was carried out in Stata v12.0 to determine the coefficient (R2) and significance (p-value).
OTU tables were first collapsed at each bacterial taxonomic level (i.e., phylum… genus) using the QIIME script summarize_taxa. Then, both the raw OTU table and each collapsed table underwent ten rarefactions to an even depth using the QIIME script multiple_rarefactions_even_depth. RFC models were constructed with the supervised_learning script for 1,000 rounds of ten-fold Monte Carlo cross validation on each table. At each level, the results were collated and averages were taken for the ten rarefied tables. Host species were used as the category for RFC model distinguishability, testing the ability to assign samples to their respective host species. The average class error for each clade was subtracted from 100 to get the percent accuracy of the models at each taxonomic level. The same methodology was used for constructing RFC models for the meta-analysis, with the only exception being that host species, host clade, and vertebrate or invertebrate categories were tested for distinguishability.
|
10.1371/journal.pgen.1000942 | Bulk Segregant Analysis by High-Throughput Sequencing Reveals a Novel Xylose Utilization Gene from Saccharomyces cerevisiae | Fermentation of xylose is a fundamental requirement for the efficient production of ethanol from lignocellulosic biomass sources. Although they aggressively ferment hexoses, it has long been thought that native Saccharomyces cerevisiae strains cannot grow fermentatively or non-fermentatively on xylose. Population surveys have uncovered a few naturally occurring strains that are weakly xylose-positive, and some S. cerevisiae have been genetically engineered to ferment xylose, but no strain, either natural or engineered, has yet been reported to ferment xylose as efficiently as glucose. Here, we used a medium-throughput screen to identify Saccharomyces strains that can increase in optical density when xylose is presented as the sole carbon source. We identified 38 strains that have this xylose utilization phenotype, including strains of S. cerevisiae, other sensu stricto members, and hybrids between them. All the S. cerevisiae xylose-utilizing strains we identified are wine yeasts, and for those that could produce meiotic progeny, the xylose phenotype segregates as a single gene trait. We mapped this gene by Bulk Segregant Analysis (BSA) using tiling microarrays and high-throughput sequencing. The gene is a putative xylitol dehydrogenase, which we name XDH1, and is located in the subtelomeric region of the right end of chromosome XV in a region not present in the S288c reference genome. We further characterized the xylose phenotype by performing gene expression microarrays and by genetically dissecting the endogenous Saccharomyces xylose pathway. We have demonstrated that natural S. cerevisiae yeasts are capable of utilizing xylose as the sole carbon source, characterized the genetic basis for this trait as well as the endogenous xylose utilization pathway, and demonstrated the feasibility of BSA using high-throughput sequencing.
| Ethanol made from fermentation of lignocellulosic biomass by baker's yeast can be considered “carbon neutral” and is one alternative to fossil fuels for powering vehicles. One of the recognized requirements for cost-effective and energy-efficient cellulosic ethanol production is the need to convert the sugar xylose—a major component of cellulosic biomass—into ethanol; however, it has traditionally been thought that baker's yeast cannot ferment xylose. We sought to investigate this assumption by looking at close relatives of baker's yeast from around the world to see if any had an intrinsic ability to grow on xylose. We identified a number of yeasts, many of them used in winemaking, that grow very slowly on this sugar, and studied one in detail. We determined that in this particular yeast the ability to grow on xylose is due to the presence of a single gene, which we named XDH1. This gene is not present in the typical laboratory strains of baker's yeast, but appears to be very common in natural wine yeasts. This gene could be useful in continuing efforts to make yeasts that can efficiently ferment xylose to ethanol.
| It is clear that society has a responsibility to address the anthropogenic causes of climate change. Current estimates indicate that about 95% of the world's energy comes from burning fossil fuels [1], which is the leading contributor of carbon dioxide emissions. Combustion of liquid fossil fuels for transportation is responsible for a large fraction of these carbon dioxide emissions in the United States, second only to electricity generation (U.S. Environmental Protection Agency). For these reasons, creating “carbon neutral” liquid transportation fuels should be an important part of global efforts to reduce carbon emissions.
One solution already in widespread use is bioethanol fermented from sugar cane (Brazil) or cornstarch (U.S.) by various strains of Saccharomyces cerevisiae [2], which is used as a major component or additive to liquid transportation fuels. For bioethanol to become a sustainable, economically viable commodity, and not to compete with food sources, it is necessary to move away from sugar cane or corn biomass toward lignocellulosic biomass sources such as corn stover or other agricultural wastes, wood byproducts, or dedicated fuel crops such as Miscanthus or switchgrass [3]–[5]. However, there are technical challenges that must be overcome before this is possible. For sugar cane and corn biomass, the predominant sugars are glucose and/or fructose, both of which are readily fermented to ethanol by various S. cerevisiae yeast strains, usually wild isolates that are particularly suited for large-scale fermentations [6], [7]. However, in lignocellulosic biomass sources, the second most abundant carbohydrate after glucose is xylose, the major pentose of hemicellulose. There is as yet no known strain of Saccharomyces that is able to convert xylose to ethanol as efficiently as glucose. Because the mass proportion of hemicellulose ranges from 20–50% in common agricultural lignocellulosic biomasses, finding both a cost-effective and energy-efficient conversion of xylose to ethanol is a critical hurdle [8].
The budding yeast Saccharomyces cerevisiae is the microorganism of choice for industrial fermentations for a variety of reasons, mainly due to its high ethanol productivity both aerobically and anaerobically, its high ethanol and low pH tolerance, and its resistance to many of the harmful compounds in a typical biomass hydrolysate. Despite recent evidence that some natural S. cerevisiae can grow, albeit poorly, on xylose [9], it has generally been reported that both natural and laboratory S. cerevisiae strains do not ferment xylose [10]–[12] leading to the assumption that they cannot, without recourse to genetic engineering, be utilized for efficient conversion of lignocellulose to ethanol. While S. cerevisiae strains were shown to be able to ferment the xylose isomer xylulose and to possess genes putatively encoding enzymes capable of xylose reduction (GRE3, GCY1, YPR1, YDL124W, YJR096W), xylitol oxidation (XYL2, SOR1, SOR2), and xylulose phosphorylation (XKS1) (Figure 1), there have been a number of experimental observations indicating that S. cerevisiae could not ferment xylose [13]–[15]. Such observations include low levels of gene expression of the endogenous enzymes, poor transport of xylose, redox cofactor imbalances, and insufficient flux through the pentose phosphate shunt [16], [17]. Despite these issues being well characterized in laboratory strains of S. cerevisiae, little is known about natural variation within Saccharomyces yeasts as it relates to xylose utilization which, as has already been shown [9], is likely to be relevant to this phenotype.
A significant amount of progress has been made over the last 30 years toward solving these problems, with much of the work focused on introducing foreign xylose pathway enzymes into S. cerevisiae: either the genes that code for xylose reductase [XR], xylitol dehydrogenase [XDH], or xylulokinase [XK] from the xylose-utilizing fungus Pichia stipitis [18]–[22], or genes coding for a xylose isomerase [XI] from other fungi and bacteria [23]–[28]. There have also been efforts to increase or adjust xylose pathway enzyme activities (XR, XK, XDH) [15], [29]–[33] and pentose phosphate flux [34], [35], reduce redox imbalances [36]–[41], and use directed evolution or random mutagenesis to increase xylose utilization [42]–[45]. Despite this large body of work, the fermentation of xylose to ethanol in these strains is still much slower than that of glucose, and there is still significant room for improvement in xylose fermentation, as well as co-fermentation of xylose and glucose, by S. cerevisiae for industrial scale applications.
As mentioned above, it has been determined that some natural strains of Saccharomyces cerevisiae are capable of growing on xylose, contrary to the notion that S. cerevisiae does not recognize this pentose as a usable carbon source [9]. It is also well characterized that there is abundant natural genetic and phenotypic variation within S. cerevisiae and closely related species [46]–[51]. In this work, we have screened a large number of wild, industrial and laboratory yeast strains to determine if other xylose-utilizing strains of Saccharomyces already exist in nature, and if so, to determine the genetic basis or bases for the phenotype. We screened 647 strains, and found a number of different Saccharomyces yeasts, predominantly wine yeasts, which are capable of utilizing xylose, albeit modestly. Through the application of high-throughput sequencing to Bulk Segregant Analysis [BSA] [52], we were able to identify the gene responsible for xylose utilization in a wine strain of S. cerevisiae, which encodes a novel putative xylitol dehydrogenase that we named XDH1. We observed that this gene is present in many different wine strains and is responsible for xylose utilization in these strains, however we have identified other strains in our screen that appear to have an independent genetic basis for their xylose utilization. We also carried out transcriptional profiling to characterize gene expression patterns during xylose utilization in wine strain derivatives and determined the contribution of native S. cerevisiae xylose pathway enzymes to the phenotype we observed. These data suggest that the putative enzyme encoded by XDH1 works in combination with the native xylose pathway to permit natural S. cerevisiae strains to recognize and utilize xylose.
To identify natural Saccharomyces species/strains that are able to utilize xylose, we screened each strain in our yeast collection for the ability, when placed in liquid medium with xylose as the sole carbon source, to increase in optical density [OD] after several days of incubation at 25°C. We measured the OD of 647 strains (Table S1) in a sealed 96-well plate format with constant, orbital shaking (see Materials and Methods). The collection largely comprises S. cerevisiae strains from various sources, including wine, brewing, baking, laboratory and clinical isolates, but it also contains other Saccharomyces sensu stricto yeasts and various hybrids between them. Of the 647 strains tested, we identified 38 strains that had some observable increase in OD (Table 1). These “xylose-positive” strains were predominantly (29/38) S. cerevisiae wine yeasts (although not all wine yeasts were xylose-positive), with the remainder being interspecific hybrids within the sensu stricto group. These xylose-positive hybrid strains generally reached higher OD in xylose media compared to the S. cerevisiae wine strains. Figure 2A shows a typical S. cerevisiae wine strain profile as well as the profile from one of the best hybrids, comparing growth in a xylose-containing medium to the same medium with no carbon source. While increase in OD does not provide evidence for fermentation of xylose to ethanol, or even of cell division, these data do show that there are natural Saccharomyces yeasts capable of utilizing xylose to accumulate biomass.
To understand the genetic basis of this xylose utilization we chose to focus on the wine strains because many could be sporulated and crossed to a laboratory strain of S. cerevisiae, and we could thus determine the segregation pattern of the phenotype. Twenty-five of these xylose-positive S. cerevisiae strains could be sporulated and tetrads dissected (Table 1). Note that because the strains have a wild-type HO gene, the spore products obtained after tetrad dissection are actually fully homozygous diploids, due to self-mating of the haploid spore during its growth on the dissection plate. In 8/25 of the xylose-positive S. cerevisiae strains, all of the spore products were xylose-positive (e.g. Simi White, Figure 2B), while in 2/25 the trait segregated 2 xylose-positive: 2 xylose-negative (e.g. Lalvin AC, Figure 2C). In the remaining 15 strains, including three strains from which no xylose-positive spores were recovered, spore viability was so poor that no complete tetrads were obtained, and thus the segregation pattern(s) could not be identified. We then took xylose-positive spore products from all of the strains from which such spores could be obtained, and crossed them (see Materials and Methods) to a laboratory S. cerevisiae strain, S288c. We observed that all of the resulting diploids were xylose-positive, indicating that the phenotype is dominant (data not shown). The resulting strains were then sporulated, and in those strains where a segregation pattern could be established, the xylose-positive trait segregated to produce two positive and two negative spores, suggesting that a single gene was responsible for the xylose-positive trait (e.g. Simi White, Figure 2D). To determine if the same locus is responsible for xylose utilization in these various wine strains, we crossed xylose-positive spores between the various wine strains and determined the segregation pattern of the xylose phenotype in the progeny of these crosses. In all of the crosses that were performed, the xylose-positive phenotype segregated 4∶0 in six tetrads (Figure 2E); this defines a cohort of at least 9 wine strains containing a single complementation group (locus) responsible for the phenotype (Figure 2F). These data indicate that a single, dominant locus is responsible for permitting xylose utilization in these S. cerevisiae strains and suggest that this mechanism of xylose utilization is common to all of the xylose-positive wine yeasts that we identified. These data also suggest that this locus may be identical by descent, consistent with evidence that wine strains are very closely related and have probably only diverged a few thousand years ago [46], [49], [50].
To determine the genomic location of the gene that permits xylose utilization we conducted BSA [53] using Affymetrix yeast tiling arrays. BSA works by taking advantage of DNA sequence polymorphisms between different strains and of the fact that it is relatively easy to pool large numbers of meiotic spore products (segregants) in yeast. Pooling segregants based on their phenotype allows the region of the genome responsible for the phenotype to be detected because DNA polymorphisms in regions unlinked to the responsible locus will segregate randomly and be “evened” out, while sequences or polymorphisms either directly responsible for the trait, or very closely linked to it, will be present in all positive segregants and absent in all negative segregants. In our case, the Simi White wine strain carrying the locus responsible for xylose utilization was crossed to a laboratory strain; the wine strain was previously estimated to carry DNA polymorphisms relative to the laboratory strain at a level of approximately .5% [54]. Spores from the Simi White/S288c diploid were screened for the xylose utilization phenotype and 39 positive spores were combined into one pool and 39 negative spores into another pool, and genomic DNA [gDNA] was isolated from each pool. We then hybridized the positive and negative gDNA pools to tiling microarrays (based on the S288c reference genome) with the expectation that regions of the genome derived from Simi White will hybridize less robustly to the array because of the DNA polymorphisms between Simi White and S288c. Log2 ratios of probe intensities were calculated (negative/positive), and a peak was evident by visual inspection in the chromosome XV right subtelomeric region that corresponds to less robust hybridization to the microarray of the positive pool gDNA (Figure 3). We confirmed the localization of the xylose-positive trait to this region by linkage analysis using strains from the yeast deletion collection, showing that the xylose-positive trait co-segregated meiotically with PHR1 (YOR386W), YOR378W, and YOR365C (2). We cloned a 10 kilobase [kb] region of the genome distal to PHR1 (containing YOR389W, YOR390W, HSP33, YOR392W, ERR1, and PAU21) from haploid, xylose-positive segregants of Simi White (GSY2469) and Lalvin AC (GSY1362) and independently transformed an S288c-based laboratory strain (FY2) with the constructs, but neither the Simi White nor the Lalvin AC derived constructs conferred a xylose-positive phenotype (data not shown), suggesting that the responsible gene was not within this 10kb region. Because yeast telomeric regions are susceptible to amplifications, insertions and translocations [55], we instead considered the possibility that the trait of interest may lie in an insertion distal to the subtelomeric sequences present in the S288c reference genome.
To identify whether there is an insertion on chromosome XV that contains the gene responsible for the xylose utilization phenotype, we repeated BSA using Illumina high-throughput sequencing on the same Simi White gDNA pools, as well as four additional pools, containing 19 positives and 16 negatives derived from a Lalvin AC/S288c cross and 16 positives and 16 negatives from a SIHA Activ-Hefe 4/S288c cross. We chose BSA over sequencing individual isolates to enrich for sequences responsible for (or tightly linked to) the xylose-positive phenotype, as there are likely to be many other novel sequences in the wine strains that are not present in the S228c genome but are unrelated to the xylose phenotype. Simi White positive and negative pools were sequenced to approximately 50× coverage of the S288c genome (∼17M mapped reads per pool), and the Lalvin AC and SIHA pools were sequenced to ∼25× coverage (∼8M mapped reads per pool) (Table S3). The 36 base pair sequence reads were aligned to the S288c reference genome using the software program MAQ [56].
To determine if any sequences were present in the positive pool that were not present in the negative pool, we performed de novo assembly of the reads that did not map to the S288c reference genome. Because de novo assembly with short sequence reads is challenging, it is important to have deep coverage and include only high quality sequence reads. To achieve this coverage and quality, we compiled all of the high-quality unmapped reads (where “high quality” reads were defined as those that did not contain any uncalled bases) from all three positive gDNA pools and used the software program Velvet [57] to perform the assembly. We then used MAQ to independently align the unmapped reads from all six gDNA pools (positive and negative) to the Velvet contigs created from the positive pools. We identified 9 individual contigs with a combined length of approximately 55kb that had no or very few reads map to them from the three negative pools. We designed primers that would amplify each of these 9 contigs and performed linkage analysis to confirm that these contigs are linked to the xylose-positive trait and yor365cΔ (Table S4). We then determined that there were approximately 28 open reading frames [ORFs] (>100 amino acids) within these 9 contigs and that a number of the ORFs are homologous to sugar metabolism genes, including a xylitol/sorbitol dehydrogenase homolog (Figure 4). The presence of a large insertion relative to the S288c reference genome containing these ORFs within the right sub-telomeric region of chromosome XV has independently been recently observed in the EC1118 wine strain genome sequence [54], [58]. The total size of the insertion is 65kb, indicating that de novo assembly identified most of the region. These data, combined with our observation that none of the previously annotated genes in the S288c reference genome distal to PHR1 were able to confer the xylose phenotype, strongly suggested that the xylose utilization trait resided in this telomeric insertion.
Of the ORFs within the chromosome XV insertion, the putative xylitol/sorbitol dehydrogenase was particularly interesting to us because it has homology to xylitol dehydrogenases from S. cerevisiae and other species (Figure S1), and we hypothesized that this gene was a likely candidate for the xylose utilization trait. We amplified this gene from both Simi White and Lalvin AC, along with approximately 400 bases of upstream and downstream sequences, and cloned it into the CEN/ARS vector pRS316 [59] to create pGS104 and pGS105. When either of these constructs were transformed into S288c, they were sufficient to permit xylose utilization in this previously non-xylose-utilizing laboratory strain (Figure 5A and data not shown). The phenotype is dependent on the presence of the plasmid containing the gene, as the xylose phenotype was lost when the transformants lost the plasmid (Figure 5A). These data show that this gene, which we have named XDH1, is sufficient to permit xylose utilization in an otherwise wild type, but xylose-negative strain.
To show necessity of XDH1 for the phenotype, we created a deletion strain (xdh1Δ) and measured xylose utilization as before. Two Simi White derivatives (GSY2468/9) were transformed with a KanMX deletion cassette containing sequences (∼400 bases) immediately up and downstream of XDH1. The deletion strains (GSY2472/1) were confirmed by PCR. Deletion of XDH1 completely abrogated the phenotype (Figure 5B). We crossed the deletion strain to another haploid derivative of Simi White and confirmed that the deletion always segregates in opposition to the xylose-positive phenotype in 9 tetrads tested (data not shown). These data prove that XDH1 is not only sufficient but also necessary for xylose utilization.
Having shown that XDH1 is responsible for xylose utilization in at least two S. cerevisiae wine strains (Simi White and Lalvin AC), and also considering our observation that all the other wine strains we were able to test appeared to be in the same complementation group, we sought to determine whether XDH1 is present in all of the xylose-positive S. cerevisiae strains and other Saccharomyces hybrids that we initially identified in our screen. To test for the presence of XDH1 in those strains, we performed colony PCR on all of the xylose-positive strains that were identified in the screen (Table 2). In 33/38 xylose-positive isolates, XDH1 was present. Interestingly, the 5 xylose-positive strains from which we could not amplify XDH1 were all recorded as being either S. bayanus or hybrids between S. bayanus and S. cerevisiae.
Some of the positive strains from our screen were heterozygous for xylose utilization, because when sporulated, the trait segregated to produce two positive and two negative spores (or some number of each type in cases where there were not enough viable spores to determine a distinct segregation pattern) (Table 1). We performed colony PCR on some of these spores to test for the presence of XDH1 (Table 3). Among the meiotic progeny of these heterozygotes, every xylose-positive segregant contained this gene. In four cases (Lalvin AC, PDM, SIHA Activ-hefe 4, and WE14) the presence of XDH1 segregated with the xylose-positive spores, while the negative spores did not contain XDH1. Surprisingly, we found instances where some negative spores did contain the XDH1 gene. In one instance, one of the two negative spores contained XDH1, while the other negative spore did not (ATCC66283, note that the four spores not from the same tetrad). In the four other cases (Montrachet, BDX, Fermichamp, French White), all the negative spores tested positive by PCR for XDH1. We sequenced XDH1 and approximately 200 bases up and downstream of the ORF from all spores of two of these heterozygous tetrads (Fermichamp tetrad 1A–D, BDX tetrad 1A–D) and did not observe any DNA sequence polymorphisms between the xylose-positive and negative spores (data not shown). This suggests that there may be another locus that is epistatic to XDH1 in these strains. Overall, the ubiquity of XDH1 in the xylose-positive strains is consistent with the hypothesis that this gene is necessary for xylose utilization in natural S. cerevisiae strains.
As described above, there are genes encoding putative xylose pathway enzymes in the S288c reference genome, and it has previously been suggested that the major XR contributors are GRE3, YPR1, and YJR096W [14]. It has also been observed that co-over-expression of GRE3 and XYL2, which encodes a putative XDH, can confer a xylose-positive phenotype [14], [15]. To assess the contribution of these and the other endogenous xylose genes to our xylose phenotype, we deleted either singly or in various combinations these genes from a haploid, xylose-positive Simi White derivative (GSY2469) and assessed the growth phenotypes of the various deletion mutants (Figure 6).
To test the contribution of each of the five putative xylose reductase genes, we introduced deletions of each of them individually in the XDH1 background. Only GRE3 significantly affected the phenotype, and none of the xylose reductase genes, when deleted individually, completely abrogated the phenotype (Figure 6, XR). We also tested sufficiency for each of the reductases by creating quadruple deletion mutants, leaving only one putative reductase gene intact (Figure 6, XR). The only two putative xylose reductases that alone contributed significantly to the ability to utilize xylose in our background were GRE3 and YPR1. The other three putative xylose reductases are insufficient by themselves to allow xylose utilization (YDL124W, GCY1, YJR096w). We also created a gre3Δ ypr1Δ double deletion in which the phenotype is almost completely removed (Figure 6, XR), though these data are not inconsistent with the other three putative xylose reductases contributing some residual XR activity. These data together suggest that both GRE3 and YPR1 are the major contributors to XR activity in a natural S. cerevisiae derivative.
Next, we tested the contribution of three putative xylitol dehydrogenases to the observed phenotype (Figure 6, XDH). Interestingly, when each potential XDH was deleted individually in the XDH1 background (sor1Δ, sor2Δ, xyl2Δ), the deletion mutants showed an improved xylose utilization phenotype relative to the positive control. Furthermore, when all three were deleted together (sor1Δ sor2Δ xyl2Δ), the phenotype was further enhanced (Figure 6, XDH). These data suggest that these putative xylitol dehydrogenases may actually be hampering the ability of this strain (and possibly all non-xylose utilizing S. cerevisiae strains) to utilize xylose, and thus are consistent with our newly identified Xdh1 protein being responsible for the presumptive xylitol dehydrogenase step of the canonical xylose utilization pathway.
Finally, we introduced an xks1Δ deletion into the XDH1 background, which encodes the putative xylulokinase, which is responsible for the phosphorylation of the fermentable metabolite xylulose to xylulose-5-phosphate [60], [61]. Deletion of XKS1 completely removed the ability of this strain to utilize xylose (Figure 6, XK), suggesting that the canonical pathway in this strain is responsible for metabolizing xylose and that XKS1 encodes the sole xylulokinase necessary for the xylose utilization phenotype we observe.
In addition to understanding how the endogenous xylose pathway genes contribute to the xylose phenotype, we sought to characterize how the presence or absence of xylose in the growth medium affected the S. cerevisiae transcriptional program over time, within the genomic context of presence or absence of the XDH1 gene. To do so, we measured mRNA levels in three pairs of sister spores from a Simi White strain that was backcrossed twice to S288c. Each pair of spores was from an independent tetrad, and contained one XDH1-containing spore (“positive”, GSY2465, 2466, 2469) and one spore that does not contain the XDH1 gene (“negative”, GSY2464, 2467, 2470). We pre-grew each of the six spores in YPD and used these cultures to inoculate minimal medium with or without 2% xylose as the sole carbon source (where the absence of xylose is the “no carbon” condition). Samples were taken from these cultures beginning immediately after inoculation (t = 0) and continuing every 8 hours for 72 hours. We then assayed relative RNA abundance versus a pooled reference, containing equimolar amounts of each sample, using Agilent yeast catalog arrays. The gene expression measurements (Log2(sample/reference)) were averaged among the three positive spores and the three negative spores at each time point.
To determine if the endogenous xylose pathway responds to the presence of xylose in the xylose-positive strain, we qualitatively compared the expression levels of all the putative xylose-pathway genes that are present in the S. cerevisiae S288c genome (Figure 7). In positive spores the putative xylose reductase genes are up-regulated compared to the reference only in the presence of xylose, while in the negative spores the xylose reductase genes are repressed under all conditions; the only exception is YDL124W, which appears to be up-regulated vs. the reference in all spore types and all growth conditions. The pattern of expression for the putative XDH XYL2 is similar to that of the xylose reductase genes; it is highly expressed across the time course in the positive strain in the presence of xylose, but is repressed over the time course in the positive strain in the no carbon medium and in both the xylose and no carbon media in the negative strain.
Interestingly, the sorbitol dehydrogenases SOR1 and SOR2, suggested to have the biochemical ability to oxidize xylitol, are highly expressed compared to the reference in the positive strain both in the presence and absence of xylose, and are strongly repressed vs. the pooled reference in the negative strains in both conditions across the time course. Because there is only one nucleotide difference between the coding sequences of SOR1 and SOR2, the probes on the array for these genes are only different by 1 base out of 60 and thus there is likely to be cross-hybridization of the mRNA's from the two SOR genes. It is also possible that there is hybridization of XDH1 mRNA to these probes, as there are only a few differences between XDH1 and the SOR1/2 probes on the microarray (6 for SOR1 and 7 for SOR2). Although we cannot determine which of the mRNA's (SOR1, SOR2 or XDH1) are hybridizing to the probes, it is nevertheless obvious that there is a distinct difference between the positive and negative spores in the expression levels of at least one of these putative dehydrogenase genes. No striking difference in the expression level of the xylulokinase, XKS1, was observed between any conditions or between any spores. The lack of change in the expression of XKS1 is somewhat unsurprising, as it has been previously reported that low levels of XKS1 are sufficient to allow xylose metabolism, while over-expression can enhance xylose fermentation in an engineered strain [29], [62]. Taken together, these data strongly suggest that the presence of XDH1 in the positive spores permits continued expression of some members of the endogenous xylose pathway when grown in xylose.
To further understand the transcriptome-wide response of these strains, we identified genes that changed significantly across the time course, compared these genes with other microarray datasets to identify any clear physiological responses, and looked for categories of functional enrichment within groups of up or down-regulated genes. Using Significance Analysis of Microarrays [SAM] [63] with a false discovery rate of 1%, we identified a list of 1266 genes whose expression levels were significantly changed over time. Specifically, we carried out a SAM analysis using the two-class (paired timecourse) option to identify genes whose expression changed over time within the positive spores, comparing the xylose to the no carbon condition. Next, we identified genes whose expression changed over time when comparing the positive to the negative spores in the presence of xylose, again using SAM with a two-class (paired timecourse) option. From the union of these two gene lists, we removed genes whose expression levels changed significantly over time within the negative strain, comparing the xylose to the no carbon condition (another two-class, paired timecourse analysis). Using this strategy, we generated an inclusive list of genes whose expression values change over time due to differences between the positive and negative strain, or due to differences between the presence and absence of xylose specifically in the positive strain. To identify the physiological responses that are associated with these gene expression differences, we retrieved data for these 1266 genes using HIDRA [64] from three other yeast microarray experiments [65]–[67] and organized the genes by K-means clustering with K = 10 [68] (Figure 8, Datasets S1, S2). For consistency with the other datasets, each of the four time-course experiments performed in this work were zero-transformed. To the right of the experiments from this paper are, respectively, a measure of how each gene's expression level correlates with increased growth rate [65], a gene expression time course over the diauxic shift [66], gene expression across a set of carbon sources (ethanol, sucrose, fructose, glucose, galactose, and raffinose) [67], and a series of time courses in various conditions including starvation, steady state growth, and other stresses [67]. We observed 5 groups (labeled on the right of the heat map) that appear to be strongly driven by similarity of the positive strain in 2% xylose to either growth rate or a stress response. For example, the genes in groups 1 and 4 (Figure 8) are more highly expressed over the time course in the positive strain in xylose when compared to the positive strain in no carbon source or the negative strain in either condition, and these genes also show a positive correlation with growth rate. As expected, when GO::TermFinder [69] is used on these groups to look for functional enrichment of biological processes, we observed processes known to be up-regulated in conjunction with a higher growth rate. Specifically, group 1 was significantly enriched for vesicle-mediated transport (GO:0016192, p = 2.11e-8) and cellular localization (GO:0051641, p = 3.13e-8) among others (Dataset S3) and group 4 is enriched for translation (GO:0006412, p = 2.26e-41) and ribosome biogenesis (GO:0042254, p = 5.87e-23) along with related processes (Dataset S4). Group 5 shows the same pattern, but largely with the opposite response, meaning that these are genes whose expression is negatively correlated with growth, and we observed that they are expressed at a lower relative level in the positive strain in xylose when compared to the no carbon condition or the negative strain in either condition; but we observed no functional enrichment in this group. Interestingly, within group 5 there is a small group of genes (labeled ‡) whose expression is induced over time relative to the reference in the positive strain in xylose, and repressed over time in the other conditions. This group includes SNO4, THI4, and HSP32, which are genes all at least putatively involved in thiamin biosynthesis. Thiamin biosynthesis is known to be important for sugar metabolism, and is a pathway in which higher expression of certain components has likely been selected for in a variety of industrial yeasts [7]. There is also a small group of genes within group 1 (labeled †) that behaves differently than the rest of the group, as it is strongly repressed relative to the reference in the positive strain in xylose. Within this group of seven genes, four of them could be involved in intracellular redox balancing as they all use NADP(H) as a cofactor (TRR1, OYE2, GDH1, ADH6). In general, these three groups suggest that XDH1 in the positive strain permits a “growth-like” transcriptional response in the presence of xylose, whereas in the absence of xylose or the absence of XDH1 the strains are exhibiting an expression pattern consistent with lack of growth and starvation (e.g. groups 4 and 5). We also observed two other groups that did not fit this pattern, but instead the positive strain in xylose exhibited a response more akin to various stresses. For example, in group 2 we observed lower relative expression in the positive strain in xylose compared to the other three conditions despite the fact that these genes are all strongly correlated with growth rate, and included functional enrichment for RNA metabolism (GO:0016070, p = 1.56e-6) and ribosome biogenesis (GO:0042254, p = 1.33e-5) (Dataset S5). Instead, they appear to be more similar to the expression patterns in strains experiencing nitrogen depletion, stationary phase, diamide, DTT, or hydrogen peroxide treatment, and 37°C heat shock. We observed a similar response in group 3, in which the expression level is opposite what we might expect if growth rate was the main cause of the expression differences but similar if the strains were exhibiting an environmental stress response. Interestingly, this group was enriched for pentose metabolic process (GO:0019321, p = 5.7e-3) and response to oxidative stress (GO:0006979, p = 7.88e-3) (Dataset S6). These data suggest that despite the fact that this set of genes is normally repressed in response to a higher growth rate, some of these genes may be responding to the presence of xylose.
There were also three groups of genes that did not have an obvious visual relationship with either growth rate or stress response. Group (a) appears to be more highly expressed in the positive strain in xylose compared to no carbon or the negative strain in either condition. While this group contains no functional enrichment using GO::TermFinder, it does contain a number of genes related to carbon metabolism, including PFK1, PFK2, PGI1, GCR1, and GND1. The final two groups (b and c) both appear to be expressed at a lower level in the positive strain in xylose compared to the other three conditions. Both groups have functional enrichment for various processes related to transcription and its regulation (Datasets S7, S8). In general genes in these three groups (a–c) show larger magnitude expression changes (induction or repression relative to the reference) in the non-growth conditions than in the positive strain in the presence of xylose. These clusters could support the conclusion that in the absence of xylose or the absence of XDH1, strains are exhibiting a response (perhaps starvation) that is simply not induced in the presence of xylose in the positive strain. In summary, these microarray data suggest that the positive strain in the presence of xylose is capable of “growth” when compared to the negative strain or lack of xylose, but it is still exhibiting a less pronounced stress-like response. These data are not inconsistent with the positive strain recognizing and using xylose as a carbon source.
In this work we have shown that naturally occurring strains of Saccharomyces cerevisiae are capable of utilizing xylose without engineering or directed evolution, and have determined the genetic basis for this phenotype. While it has been known for many years that the xylose isomer xylulose is fermentable by S. cerevisiae, it has generally been thought that this species is incapable of metabolizing xylose. However, recent work has shown natural genetic variation for xylose utilization does exist, and that natural selection and breeding can improve xylose utilization in natural strains of S. cerevisiae [9]. By screening through many industrial and clinical isolates, we discovered variation within this species that permits utilization of this sugar, fermentation of which is an important prerequisite for the efficient generation of ethanol from lignocellulosic biomass sources. We have also shown that this ability to utilize xylose by Saccharomyces is conferred by the presence of a single gene, a novel putative xylitol dehydrogenase that we have named XDH1. This gene is both necessary and sufficient to permit xylose utilization in the normally non-xylose-utilizing S288c laboratory strain, and is absent from the reference genome sequence of S288c.
We also characterized the transcriptional response of one of our xylose-utilizing strains of S. cerevisiae to xylose in the presence and absence of XDH1. While these data do not allow us to draw conclusions as to whether or not this gene permits actual fermentation (rather than simply utilization) of xylose, we can make a number of observations. First, it is clear that the endogenous xylose pathway is capable of responding at the transcriptional level to the presence of xylose when this novel XDH is present. Secondly, we can infer that this sugar and its downstream metabolites are likely being funneled into central carbon metabolism via the pentose phosphate pathway as is consistent with what has previously been observed. This suggests that industrial or laboratory strains of S. cerevisiae may be more poised to ferment this pentose than previously thought, implying that we can better harness the standing genetic potential that already exists in nature and use it in combination with directed evolution and metabolic engineering to make an industrially applicable xylose fermentation strain.
The idea that Saccharomyces might be more “ready” to ferment xylose than previously thought is further supported by our genetic dissection of the xylose metabolic pathway endogenous to S. cerevisiae. We corroborated previous data that shows the xylulokinase encoded by XKS1 is functional and supports metabolism of xylose. We also demonstrated that GRE3 and YPR1, encoding two aldo-keto reductases, are each sufficient to allow xylose utilization in our strain background. The observation that a novel xylitol dehydrogenase is responsible for the xylose utilization phenotype, and the observation that the genes in the reference strain encoding enzymes putatively thought to oxidize xylitol (SOR1, SOR2, XYL2) are in fact detrimental to the phenotype, further support that the idea of a redox imbalance in S. cerevisiae favoring xylitol production over further metabolism is true [70], [71]. Finally, our results also suggest that some property of the XDH1 is able to reduce the cofactor imbalance and may be capable of pushing xylitol through the xylose metabolic pathway.
We also discovered Saccharomyces sensu stricto interspecific hybrids in our screen that appear to robustly utilize xylose by a mechanism independent of XDH1. Some of these strains are even more effective at utilizing xylose than the S. cerevisiae wine strains we have characterized here, and we are currently attempting to identify the locus (or loci) responsible for these other xylose phenotypes. Based upon the results in Table 2 that show S. bayanus xylose-positive strains that do not possess XDH1, it is likely that there is at least one other trait that is as yet unidentified. There may also be additional components of the xylose utilization pathway for which hypomorphic alleles exist in natural strains, as XDH1 is present in xylose-negative segregants of some xylose-positive strains we identified. We also suggest that the only other previously described [9], [72] xylose phenotype native to S. cerevisiae is likely to be XDH1-dependent, given that wine strains were included in the initial breeding. Because we and others have assayed strains that only contain a small sample of the variation that likely exists in the Saccharomyces gene pool, it is likely that there is additional variation present in nature that may be able to contribute to a xylose-positive phenotype.
Finally, we have developed a novel application of high-throughput sequencing for quickly mapping an unknown trait by BSA. Because we were able to identify a clear segregation pattern for our phenotype of interest, in this case a single locus, we were able to easily pool segregants and use sequencing to narrow down the genomic location using the high frequency of polymorphisms that segregated with our locus. Applying sequencing technology in addition to tiling arrays was critical as our phenotype resided in a region of the genome that is not present in the reference genome. Given that the number of genes responsible is small, we suggest that this application of high-throughput sequencing could be used broadly for associating other unknown genotypes to well-characterized phenotypes. It will be particularly applicable to other species that have small genomes and for which the genome sequence or tiling arrays are not readily available, or for such species that may contain variation not captured in their respective reference genomes.
While effective conversion of xylose to ethanol in an industrial setting by Saccharomyces yeasts has not yet reached its full potential, much progress has been made recently. We suggest that uncovering and studying the genes responsible for xylose utilization in wild strains of Saccharomyces may contribute directly to further improvements in lignocellulosic biomass fermentation. Additionally, the functions of these genes might continue to shed light on problematic areas in the metabolism of xylose, helping to inform directed evolution and metabolic engineering approaches.
Strains used in this study are shown in Table S1 and Table S6. In order to cross diploid HO/HO wine strains to a haploid S288c strain, wine strains were transformed with either pGS35 (CEN/ARS, KanMX) or pGS36 (CEN/ARS, Hph) and the resulting transformants carrying the plasmid were sporulated (Hph is the gene that permits hygromycin B resistance). Spores were mixed with a haploid ho S288c strain carrying either pGS35 if the wine strain carried pGS36 or vice versa, and plated onto YPD plates supplemented with G418 (200µg/mL) and hygromycin B (150µg/mL).
To screen for xylose utilization, single colonies were pre-grown to saturation at 25°C in YP with 2% glucose and then diluted 1∶50 into YP with 2% xylose (Sigma) or no carbon source. 100µL cultures were grown for 5 days at 25°C in a sealed 96-well plate and absorbance was read at 595nm every 15 minutes in a TECAN Genios plate reader with orbital shaking. Xylose positives were identified by visual inspection of increasing OD in xylose compared to no carbon source, and were confirmed by retesting in both YP and Minimal [73] media.
Because growth on xylose is not exponential, we did not calculate a doubling time. Instead, to quantify xylose utilization we calculated a slope (change in OD over time) across the linear range of OD increase, from 20 to 80 hours in a typical TECAN growth experiment following the initial trehalose growth. Growth curves were done in at least triplicate (see Table S6 for all deletion strains), and a t-test was used to determine significant differences in rate of OD increase between deletion strains and “wild type” xylose positives.
To analyze gene expression, cultures were pre-grown to saturation in YP with 2% glucose and diluted 1∶50 into a 1.1L culture of minimal medium [73] with 2% xylose or no carbon source. 100mL samples were collected starting immediately after inoculation (t = 0) and at subsequent 8 hour intervals for 72 hours by filtering with 0.45µm analytical test filter funnels (Nalgene) and were snap frozen in liquid nitrogen. RNA was extracted using a modified version of the hot phenol protocol, as described [74], [75]. A pooled reference sample was created by combining 350ng of each of the 120 RNA samples (10 time points for 6 strains in 2 conditions). 325ng of each total RNA sample and reference were labeled with Cy dyes (Amersham) using the Agilent Low RNA Input Linear Amplification Kit, and hybridized to Agilent Yeast Gene Expression Arrays (v2, 8x15K) for 17 hours at 65°C at 10rpm in a hybridization oven (Shel Lab). Arrays were scanned at 5µm resolution on an Agilent Scanner, and Agilent Feature Extraction v9.5.3.1 was used for extraction of data from the scanned images, and data normalization and calculation of log2 ratios. Gene expression data have been deposited in the GEO database with accession number GSE19121.
Xylose-positive segregants of Simi White (Lallemand), Lalvin AC, and SIHA Activ-Hefe 4 were crossed once to S288c (GSY147), and the resulting diploids were then sporulated. F2 segregants were scored for xylose utilization in the TECAN plate reader as described above. 1.5mL of overnight YPD culture of each segregant grown was spun down, resuspended, and frozen in 300µL of sorbitol solution (0.9M sorbitol, 0.1M Tris pH 8, 0.1M EDTA). Samples were pooled by phenotype at this stage and genomic DNA was extracted as described [76]. The pools contained 39 positives and 39 negatives for Simi White, 19 positives and 16 negatives for Lalvin AC, and 16 positives and 16 negatives for SIHA.
Genomic DNA was labeled as described [77], and microarray-assisted BSA was done using Affymetrix GeneChip S. cerevisiae Tiling 1.0R Array basically as described [52], [78]. Briefly, a ratio of the log2 intensities for the perfect match probes was plotted across every chromosome for each nucleotide. The plots for each chromosome were scanned visually for local peaks in intensity. Tiling array data have been deposited in the GEO database with accession number GSE19121.
The same pools of genomic DNA were used for BSA by sequencing. 5µg of genomic DNA were prepared for sequencing using the Illumina Genomic DNA Sample Kit. Flow cells were prepared using the Illumina Standard Cluster Generation Kit v2, and samples were sequenced on the Illumina Genome Analyzer II. GAII data were analyzed with the Illumina 1.3.2 pipeline, and reads (with qualities) were aligned to the S288c genome with MAQ v0.7.1 [56] using default parameters. Reads from the positive pools that did not align to the reference genome were combined, and reads that contained any uncalled bases (“N”> = 1) were removed from further analysis. De novo assembly was performed on this filtered set of un-mapped reads using Velvet v0.7.55 [57] with default parameters and hash length = 13. All raw high throughput sequence data have been deposited in the SRA database with accession number SRP001391.
The novel XDH was cloned into the NotI site of pRS316 [59] from GSY2469 (Simi White derivative) and GSY1362 (Lalvin AC derivative) by PCR using primers that contained NotI restriction sites. Primers are listed in Table S5. FY2 (S288c) was then transformed with the resulting plasmids (pGS104 and pGS105) via a slightly modified lithium acetate method [79]. Plasmids are listed in Table S6. Growth was assayed as described above in the TECAN plate reader.
Plasmid loss experiments were done as follows. The original transformant that was used to generate a TECAN growth curve was also streaked for single colonies on a YPD plate. These were grown and replica plated onto YPD and SC-URA plates, and colonies were picked from the YPD plate that either retained the plasmid (grew on the SC-URA replica plate) or lost the plasmid during mitosis (did not grow on the SC-URA replica plate) and were tested again in the TECAN.
Homologous recombination was used to create a disruption of the novel XDH. Primers are listed in Table S5. Briefly, KanMX6 was amplified from pFA6-KanMX6 [80], and approximately 400 bases up and downstream of the XDH homolog were amplified separately using primers that overlapped with the 5′ and 3′ primers used to amplify KanMX6. The three fragments were joined using Phusion DNA polymerase (Finnzymes) and the resulting deletion cassette was integrated into GSY2469 and GSY2468 (Simi White derivatives) by lithium acetate transformation. Correct integration of the deletion was confirmed by PCR and by showing opposing segregation of G418 resistance and the xylose trait in a cross to another xylose-positive haploid derivative.
To genetically dissect the endogenous xylose pathway, deletions of the xylose pathway genes were crossed into a haploid Simi White derivative that was previously backcrossed twice to S288c (GSY2469). Diploid strains heterozygous for deletions of GCY1, GRE3, YPR1, YJR096W, XYL2, and XKS1 were purchased from Invitrogen. Deletions of SOR1 and SOR2 were not available from the deletion collection as they are in large genomic regions of essentially 100% identity. Deletions were constructed as described [81], except with approximately 80 bases of homology to the regions immediately up and downstream of the SOR1/2 open reading frames rather than 40. Transformants were crossed to pgu1Δ and lrg1Δ to differentiate between sor1Δ and sor2Δ. Segregation of deletions was tracked by colony PCR when creating strains with more than two deletions, as the deletions are all marked with G418R. Primers are listed in Table S5, and strains are listed in Table S6.
|
10.1371/journal.pcbi.1002310 | Translating Clinical Findings into Knowledge in Drug Safety Evaluation - Drug Induced Liver Injury Prediction System (DILIps) | Drug-induced liver injury (DILI) is a significant concern in drug development due to the poor concordance between preclinical and clinical findings of liver toxicity. We hypothesized that the DILI types (hepatotoxic side effects) seen in the clinic can be translated into the development of predictive in silico models for use in the drug discovery phase. We identified 13 hepatotoxic side effects with high accuracy for classifying marketed drugs for their DILI potential. We then developed in silico predictive models for each of these 13 side effects, which were further combined to construct a DILI prediction system (DILIps). The DILIps yielded 60–70% prediction accuracy for three independent validation sets. To enhance the confidence for identification of drugs that cause severe DILI in humans, the “Rule of Three” was developed in DILIps by using a consensus strategy based on 13 models. This gave high positive predictive value (91%) when applied to an external dataset containing 206 drugs from three independent literature datasets. Using the DILIps, we screened all the drugs in DrugBank and investigated their DILI potential in terms of protein targets and therapeutic categories through network modeling. We demonstrated that two therapeutic categories, anti-infectives for systemic use and musculoskeletal system drugs, were enriched for DILI, which is consistent with current knowledge. We also identified protein targets and pathways that are related to drugs that cause DILI by using pathway analysis and co-occurrence text mining. While marketed drugs were the focus of this study, the DILIps has a potential as an evaluation tool to screen and prioritize new drug candidates or chemicals, such as environmental chemicals, to avoid those that might cause liver toxicity. We expect that the methodology can be also applied to other drug safety endpoints, such as renal or cardiovascular toxicity.
| Translational research involves utilization of clinical data to address challenges in drug discovery and development. The rationale behind this study is that the side effects observed in clinical trial and post-marketing surveillance can be translated into a screening system for use in drug discovery. As a proof-of-concept study, we developed an in silico system based on 13 hepatotoxic side effects to predict drug-induced liver injury (DILI), which is one of the most frequent causes of drug failure in clinical trial and withdrawal from post-marketing application, and also one of the most difficult clinical endpoints to predict from preclinical studies. We first identified 13 types of liver injury which yielded high prediction accuracy to distinguish drugs known to cause DILI from these don't. To effectively apply these 13 hepatotoxic side effects to the drug discovery process for DILI, we developed in silico models for each of these side effects solely based on chemical structure data. Finally, we constructed a DILI prediction system (DILIps) by combining these 13 in silico models in a consensus fashion, which yielded >91% positive predictive value for DILI in humans. The DILIps methodology can be extended in applications for addressing other drug safety issues, such as renal and cardiovascular toxicity.
| Drug-induced liver injury (DILI) poses a significant challenge to medical and pharmaceutical professionals as well as regulatory agencies. It is the leading cause of acute liver failure, which has a high mortality rate (30%) as treatment is limited due to the availability of livers for transplantation [1]. Although many dangerous drugs are identified during animal testing thus protecting humans from this damage, a consortium determined that about half of the drugs that cause human hepatotoxicity were not identified as having this potential in nonclinical animal testing [2]. Many drugs have been withdrawn from the market or have received restrictions and warnings due to DILI [3]. DILI information and guidance for pharmaceutical industries has been released by regulatory agencies such as the U.S. Food and Drug Administration (FDA) (http://www.fda.gov/downloads/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/UCM174090.pdf), European Medicines Agency (EMA) (www.ema.europa.eu/pdfs/human/swp/15011506en.pdf) and Health Canada (http://www.hc-sc.gc.ca/dhp-mps/alt_formats/pdf/consultation/drug-medic/draft_ebauche_hepatotox_guide_ld-eng.pdf), highlighting both the significance and difficulties in DILI research. In the FDA, the Critical Path Initiative identified DILI as a key area of focus in a concerted effort to broaden the agency's knowledge for better evaluation tools and safety biomarkers (http://www.fda.gov/ScienceResearch/SpecialTopics/RegulatoryScience/ucm228131.htm).
Determining the potential for a drug candidate to cause DILI in humans is a challenge. First, the standard pre-clinical animal studies do not effectively predict DILI events in humans. In one notorious example, five subjects in a phase 2 clinical trial experienced fatal hepatotoxicity induced by fialuridine, an investigational nucleoside analogue that showed no liver damage in animal studies [4]. Out of 221 pharmaceuticals, the overall concordance of liver toxicity in humans and experimental animals is as low as 55%, which is in sharp contrast with the concordance of other target organs such as the hematological (91%), gastrointestinal (85%), and cardiovascular (80%) systems [2]. Secondly, even well-controlled clinical trials fail to accurately predict post-marketing DILI events. The main reason for this is the statistical power of the trials – the risk of severe DILI of an idiosyncratic nature is very low per exposed subject, while clinical trials are usually carried out with only several thousand patients [5], rendering them significantly underpowered to predict rare DILI events.
To enhance the predictability of DILI, novel approaches have been explored by many researchers. Notable examples include (a) development of new DILI biomarkers [6], (b) introduction of high-content screening [7], (c) adoption of more sensitive animal models [8], [9], [10], and (d) utilization of toxicogenomics [11]. Most of these investigations are focused on developing biomarkers using either animal or in vitro models for predicting DILI in humans. This still would involve synthesis of the drug and elaborate testing. An in silico approach could inform chemists at the earliest point in the drug discovery pipeline and enable them to select the best chemical structures.
We hypothesized that there exists a distinct set of liver side effects that can be used to characterize the DILI risk of drugs in humans. We identified 13 types of hepatotoxicity (hepatotoxic side effects or HepSEs) from the organ levels of hepatobiliary disorders in the Medical Dictionary for Regulatory Activities (MedDRA) ontology (http://www.meddramsso.com/). We found that these 13 HepSEs can discriminate DILI drugs from non-DILI drugs with high accuracy (∼83%). Since the side effects are clinical observations obtained either from clinical trials or from post-marketing surveillance with limited utility in drug discovery, we developed quantitative structure-activity relationship (QSAR) models for each of the HepSEs. We then constructed a DILI prediction system (DILIps) based on the 13 HepSE models with an improved prediction strategy using a “Rule of Three” (RO3) criterion (incriminated by 3 or more HepSE models). The systems were evaluated in several external test sets with performance surpassing most in silico models in the field. We screened the entire drug list using the DILIps and evaluated the RO3 drugs in terms of therapeutic use and drug targets.
Figure 1 is an overview of the approach taken. First, the identification and assessment of HepSEs were performed. We used the SIDER database [12] to identify drugs and associated side effects. Out of 1450 side effects in the database, we selected only those that were caused by more than 20 drugs (an arbitrary cut-off). This yielded 473 side effects. The distribution of 888 drugs over 473 side effects and vice verse were depicted in Supplementary Figure S1, indicating that over 90% drugs were associated with at least 10 side effects. These side effects were then directly mapped onto low level terms of MedDRA. The terms were linked to the system organ classes (SOC) level according to the hierarchical structure of MedDRA (Supplementary Table S1) in order to determine the terms' attributes at the organ level. Finally, we considered side effects defined by the MedDRA ontology as related to the hepatobilliary disorders SOC term as HepSEs, and identified 13 HepSEs: bilirubinemia, cholecystitis, cholelithiasis, cirrhosis, elevated liver function tests, hepatic failure, hepatic necrosis, hepatitis, hepatomegaly, jaundice, liver disease, liver fatty, and liver function tests abnormal.
We evaluated these 13 HepSEs for their ability to differentiate drugs that do and do not cause DILI using the Liver Toxicity Knowledge Base Benchmark Dataset (LTKB-BD) [13] and PfizerData [14]. For both datasets, we used only the drugs that they had in common with SIDER. There are several differences between two datasets to call a drug as DILI or non-DILI (see Materials and methods), including (1) LTKB-BD is based on the FDA-approved drug labeling while PfizerData is according to the case reports; (2) two datasets apply different criteria for DILI assessment; and (3) drugs are grouped differently between two datasets. To obtain an objective evaluation for 13 HepSEs, we took the following actions to select DILI positive and negative drugs from two datasets: (a) in LTKB-BD [13], Most-DILI-Concern drugs were classified as positive while No-DILI-Concern drugs were classified as negative; and (b) in PfizerData [14], drugs with evidence in human toxicity were considered DILI positive while drugs with no evidence in any species were considered DILI negative. Defining a drug as causing DILI if it was positive in any of the 13 HepSEs, this approach yielded 91% and 74% accuracy for LTKB-BD and PfizerData, respectively.
It is important to note that the 26 MedDRA SOCs are not all strictly related to human organs in a conventional sense. For example, “investigations” and “general disorders and administration site conditions” are not organs (the complete list of MedDRA SOC is available in Supplementary Table S1). Some side effects with DILI indication are resided in a SOC other than the hepatobilliary disorders. For example, the SOC of “investigations” include the “elevate liver enzyme” and “alkaline phosphatase increased”, both are conventional DILI indicators. Moreover, some side effects in the SOC of “general disorders and administration site conditions” could also be the manifestations of DILI. Thus, we conducted a permutation test with the purposes of confirming that the 13 HepSEs do in fact have significant performance over the chance to distinguish DILI drugs from non-DILI drugs. We randomly selected 3, 5,…, 21 side effects from the 473 side effect pool with each selection repeated 20,000 times. As shown in Figure 2, the classification accuracy of the 13 HepSEs, indicated by the red dot, was considerably higher than the average accuracy for each of the sets of randomly selected side effects, demonstrating that the observed classification accuracy for the 13 HepSEs was not due to chance.
As illustrated in step 2 (Figure 1), QSAR models were developed for each of the 13 HepSEs to enable their use in screening new drug candidates computationally. The QSAR models developed from the drugs related to each of these 13 HepSEs had high prediction accuracy (>93%) in a leave-one-out cross-validation (LOO-CV) process (Table 1).
Based on the 13 HepSE models, we further developed the DILIps (step 3 of Figure 1, left box). Using the same classification rule described above (i.e., drugs incriminated by any of the 13 HepSEs models are considered as DILI positives), we applied DILIps to three external validation sets. The validation sets of LTKB-BD and PfizerData contain drugs not used in developing the 13 HepSE models. For the O'Brien et al. dataset [15], the severe and moderate hepatotoxicity drugs were combined as DILI positive drugs while the non-toxic drugs were defined as DILI negative drugs (only the drugs not used by the 13 HepSE models were included). As summarized in Table 2, the DILIps exhibited a reasonable prediction performance for three external validation sets with the prediction accuracy between 60–70%.
Identifying the drugs of severe DILI potential with high confidence has an important application since these drugs are likely withdrawn from the market or restricted in use with black box warning (BBW) due to the serious public health concern. We assume that the number of models calling a drug causing DILI is positively correlates with the drug's severity for DILI and to the extension of the confidence to make such a call. We generated a union set based on the three validation datasets listed in Table 2. We removed three drugs having an inconsistent DILI assignment among three datasets (only three drugs were removed: moxisylyte, carbidopa and terfenadine), i.e., positive in one dataset and negative in another. This process resulted in 145 DILI positives and 63 DILI negatives (see Supplementary Table S2). We used this union set to assess how many HepSE models to be combined likely identify high risk DILI drugs (i.e., withdrawal or BBW drugs) with high positive predictive value (PPV).
Specifically, for each of the possible HepSE combination models requiring a drug to be incriminated by N HepSE models (“Rule of N”), we calculate PPV and the number of drugs retained by each of the HepSE combination models. As depicted in Figure 3, the PPV reaches a maximum of 91.3% when N = 3. Additionally, the percentage of high risk DILI drugs reached a local maximum at N = 3. Therefore, we established the RO3 criterion in the DILIps for identifying drugs that might cause severe DILI with high confidence (step 3 of Figure 1, right box). The number of drugs meeting the RO3 is 23, dramatically decreased from 100 (RO1) and 49 (RO2), which was expected when the optimization was tilted toward increasing PPV. In order to identify the drugs of severe DILI potential with high confidence, the trade-off was accepted in the context of an application. Therefore, the RO3 was selected to carry out further study.
We applied the RO3 criterion to the drugs (small molecules only) in DrugBank to investigate which therapeutic categories were most likely associated with DILI (represented by the graph at the right of step 4, Figure 1). Figure 4(a) shows the drug distribution across 14 therapeutic categories as defined by Anatomic Therapeutic Class (ATC) (http://www.whocc.no/atcddd/) with the RO3 positive drugs highlighted in red. The enrichment of the RO3 drugs in each therapeutic category was determined by Fisher's exact test. We found that two therapeutic categories (i.e., anti-infective for systemic use and musculoskeletal system drugs) were significantly associated with drugs that cause DILI (p-value = 5.00E-11 and 0.002, respectively). To confirm the findings, we carried out the same analysis for drugs in the SIDER database that met the RO3. As shown in Figure 4(b), the same two therapeutic categories were also found to be significantly associated with drugs that cause DILI (p-value = 8.94E-8 and 2.36E-7, respectively). Both results demonstrated that care must be taken when drugs are developed with existing targets in these two categories.
The findings are consistent with real-world observations; for example, non-steroidal anti-inflammatory drugs (NSAIDs, a subcategory of anti-infectives for systemic use) are often associated with DILI. A good example is didanosine (Videx®) which is an antiviral drug used to treat human immunodeficiency virus (HIV) infection. On Jan 29th, 2010, the FDA notified healthcare professionals and patients about a rare but serious complication in the liver known as non-cirrhotic portal hypertension in patients using the drug. Subsequently, a black box warning was added to the drug label to warn doctors and consumers of this risk. Didanosine can also cause lactic acidosis and severe hepatomegaly with steatosis, and has resulted in several fatal cases (http://dailymed.nlm.nih.gov/dailymed/drugInfo.cfm?id=23496).
It was important to determine if the drug target is related to the drug's likelihood of causing DILI. Accordingly, we investigated the drugs that were RO3 positive from DrugBank in the target space using network analysis as illustrated on the left side of Figure 1, step 4. These drugs were associated with 134 human protein targets. In the network analysis, we considered that two protein targets are directly related (connected with an edge in network analysis) if one or more drugs were associated with both targets.
As depicted in Figure 5, the network contains two large modules (Modules #1 and #2) with several small modules. There are 72 targets in Module #1 associated with 125 RO3 positive drugs, and 23 targets in Module #2 associated with 8 drugs. We conducted toxicity function and pathway analyses using Ingenuity Pathway Analysis (IPA, http://www.ingenuity.com/) for both modules. In each module, particularly Module #1, the biological functions related to disease and disorder were investigated to assess if the targets of the drugs meeting the RO3 have a relationship with hepatic system diseases or disorders. As shown in Table 3, liver injury and disease related functions enriched in Module #1 were hepatic system disorder, jaundice, liver cancer, hepatocellular carcinoma, and hepatitis C. All the liver injury and disease functions are under the hepatic system diseases branch of the top toxicity functions in IPA. The other significant toxicity functions of Module #1 can be found in Supplementary Table S3.
We also found that every drug in the two largest modules was associated with more than three targets on average. Note that drugs are prone to having multiple side effects if they interact with multiple targets since different targets may invoke different side effects [16], [17]. We conducted text mining to verify the association of 13 HepSEs and 134 targets identified by RO3 positive drugs. We identified 45 proteins associated with eight HepSEs in a co-occurrence analysis (Figure 6 and Supplementary Table S4). Most of these targets are associated with hepatitis, while targets such as PTGS2/COX-2 (prostaglandin-endoperoxide synthase 2) and ABCD1 (ATP-binding cassette, sub-family, and member 1) are related to multiple HepSEs.
One application of translational science involves utilization of clinical data to address challenges in drug discovery. The key concept of this study is that the side effects observed in clinical trials and post-marketing surveillance can be translated for use in drug discovery. As a proof-of-concept study, we developed the DILIps to address one of the most difficult clinical endpoints to predict from preclinical studies, that is DILI. The DILIps contains three distinct and sequential approaches. First, we identified 13 HepSEs based on the MedDRA ontology, which provided excellent discrimination of a drug's potential to cause DILI (91% and 74% accuracy for LTKB-BD and PfizerData, respectively). Secondly, HepSE-based QSAR models were developed by using all 888 drugs in SIDER, which were highly predictive as compared to published models [14], [18], [19] and offered a robust translation of clinical observation (i.e., side effects) using in silico techniques to the drug discovery/preclinical testing aspect of drug development. Next, we developed DILIps by combining these 13 HepSE QSAR models, which yielded 60–70% prediction accuracy for three independent validation sets. Lastly, a RO3 criterion was implemented in DILIps, which had >91% confidence for identification of drugs that might cause severe DILI.
The DILIps is a modular system; each of its components can be replaced by other methods or constructed using different variables. For example, besides selecting 13 HepSEs from the hepatobiliary disorders category in MedDRA, we also examined the effect of including additional two DILI related terms from the investigation SOC category, or selecting 14 DILI relevant terms as suggested by an expert (Supplementary Table S5). Both yielded similar performance compared to the 13 HepSE-based approach. Given the fact that each MedDRA category is a stand-alone ontology and other options did not yield exceptional performance, we choose the terms under hepatobiliary disorders as representative types of DILI in this study. For the second component of the DILIps, we developed HepSE-based QSAR models because chemical structure data were readily available for the entire set of 888 drugs in SIDER, providing a sufficiently large sample from which to build the HepSE-based models. Other technologies, such as gene expression microarrays, might be able to construct better HepSE models. However, the data from these technologies was not available for the complete set of SIDER drugs. With different choices in components 1 and 2, the criterion in component 3 of DILIps could be altered to optimize DILI classification using different consensus approaches instead of RO3. Therefore, the DILIps is subject to change and improvement when new data, technology, and knowledge are available.
Development of predictive models for drugs that might cause DILI in humans has been an active research field, with much of the work being done using QSARs. However, the DILI labels used in these studies are from different sources, some focused on case reports and others developed using text mining. Furthermore, the methods used to develop the models are also different. Thus, it is difficult to compare these methods. For example, Greene et al. [14] developed Derek for Windows (DfW), a knowledge-based expert system, to predict a drug's potential to cause DILI using the DILI classification scheme developed by Pfizer. The system has 56% overall accuracy with 73% specificity and 46% sensitivity. Fourches et al. [18] applied text mining for DILI reported in different species using MEDLINE abstracts, suggesting that the concordance of liver effects is low (i.e., 39–44%) between different species. They also developed QSAR models using a text mining approach to define DILI classification with external prediction accuracies ranging from 56 to 73%. Very recently, Ekins et al. [19] developed a Bayesian model based on DILI endpoint from cellular imaging predictions [7], which gave a concordance of 60%, sensitivity of 56%, and specificity of 67%.
Development of DILI models for humans is always confronted by two distinct but related challenges: (1) a comprehensive drug list with DILI annotation is usually not available, and (2) there is no authoritative assessment of whether a drug causes DILI or not. In this study, we compiled three large datasets from our LTKB project. We used only the drugs of the opposite extremes in DILI classification (positive or negative in relationship to DILI) by removing drugs with ambiguous call. The RO3 criterion of DILIps reached >91% positive predictive value for a combined drug list from these three literature datasets. We also applied DILIps for the drugs with ambiguous call and the results are available from Supplementary Table S6.
The translation of clinical observations to evaluation of drugs earlier in the drug development pipeline is a goal of translational medicine [20]. DILI is an endpoint influenced by several important factors, and it is difficult to adequately predict with a single model. The SIDER database has collected clinical observation data (side effects) from drug labels and the scientific literature, which allows the linkage of disease endpoints and related symptom profiles. This, in turn, provides an opportunity to combine drug information and patient information into a unified prediction method, a focus of this study. The HepSEs provide a new direction to predict DILI based on the consensus of multiple clinical endpoints (side effects) using an in silico method.
Elucidation of therapeutic uses, drug targets, and pathways related to DILI from a systematic perspective is of great use in drug discovery and pharmacovigilance. By applying the RO3 criterion to the entire drug space defined by DrugBank, we constructed a DILI landscape in terms of therapeutic and drug target space. We do acknowledge that the findings from this excise are dependent on the accuracy in annotation in DrugBank.
We identified two therapeutic categories (i.e., anti-infectives for systemic use and musculoskeletal system drugs) in which the drugs have a high risk for causing DILI. This is consistent with the general understanding that, for example, NSAIDs (a subcategory of anti-infectives for systemic use) are often associated with DILI and have been subject to a broad range of studies looking into drug-specific, therapeutic class-specific, and genetic-specific effects [21]. Another possibility is that these drugs may have higher exposure rates; they are widely used by many people over prolonged periods, which may inadvertently increase the risk of DILI. The RO3 positive criterion was able to identify most “bad actors” among NSAIDs including celecoxib, diclofenac, diflunisal, ibuprofen, leflunomide, and rofecoxib. Most of them are PTGS2 (COX-2) protein inhibitors. This gene is also involved in several hepatic system pathways such as hepatic system disorder, liver cancer, and hepatocellular carcinoma. COX (Cyclooxygenase) is an enzyme that is responsible for formation of important biological mediators called prostanoids. Pharmacological inhibition of COX can provide relief from symptoms of inflammation and pain. However, more and more reports indicated that the selective inhibition profile of COXs can cause certain serious adverse drug reactions. A classic example is rofecoxib (brand name Vioxx®), which was withdrawn in 2004 because of the risk of heart attack caused by selective inhibition of COX-2. Rofecoxib was also associated with DILI [22]. Another example is lumiracoxib, a selective COX-2 inhibitor developed for the symptomatic treatment of osteoarthritis and acute pain. Concern over hepatotoxicity has contributed to the withdrawal or non-approval of lumiracoxib in most major drug markets worldwide [23]. Therefore, the study of the relationship between drug target and DILI, such as COX selectivity and DILI, may provide new insights into DILI at a molecular level [24].
We also found that DILI drugs often involve multiple targets, which is often associated with drugs applied in multiple therapeutic categories [25]. Drugs interacting with multiple targets are considered “dirty” since they have a potential to initiate different adverse reactions. On the other hand, these drugs may also hold the potential to be repositioned for use outside of their original therapeutic indications. One such example is diclofenac, which is used to relieve pain, tenderness, swelling and stiffness caused by osteoarthritis, rheumatoid arthritis, and ankylosing spondylitis. Diclofenac is labeled with four different ATC codes (i.e., four different therapeutic uses) and associated with a number of targets categorized by DrugBank, including prostaglandin G/H synthase 1 and 2, the cytochrome P450 family (2C18/2E1/2C19/1A2/2C8/2D6/2C9/3A4/1A1/2B6), the UDP-glucuronosyltransferase family (1–1,2B7), prostaglandin G/H synthase 1, etc. Several case-control studies have been carried out to investigate the role of polymorphisms in the gene encoding regions of the aforementioned drug-metabolizing enzymes and transporters to determine susceptibility to diclofenac-induced hepatotoxicity [26], [27], [28], [29], [30]. Diclofenac has been withdrawn in several countries due to liver injury and other adverse drug reactions, including ulcers, bleeding, and ulcerations in the stomach and intestinal linings [31]. Diclofenac induced liver injury causes a number of side effect patterns, including cirrhosis, hepatic failure, hepatic necrosis, hepatitis, jaundice, all of which were included in our set of 13 HepSEs.
DILI is associated with two distinct but related parameters: drug properties and patient susceptibility. Some drugs are more likely to cause DILI, while some patients are more likely to show DILI. The DILIps is primarily capable of addressing the former challenge with an aim to enhance DILI identification in drug discovery. Identifying genetic variations and their associated protein products that contribute to DILI is another important research area, but one that requires the costly and time-consuming collection of samples from large numbers of affected individuals. Study of the genetic risk factors to DILI or other conditions usually involves the identification of genes associated with key disease mechanisms and immunological reactions using genotyping approaches. The network analysis conducted in this study connected DILI drugs with pathways and targets and might contribute to the identification of mechanisms that relate a patient's genetic predisposition and DILI. There are a small number of genetic risk factors identified for DILI, most are associated with a drug interaction with a specific HLA (human leukocyte antigen system) polymorphism within the major histocompatibility complex (MHC) such as lumiracoxib (HLA-DRB1*15∶01) [23], antituberculosis chemotherapy (HLA-DQB1*02∶01) [32], ticlopidine (HLA-A*33∶03) [33], ximelagatran (HLA-DRB1*07∶01) [34], flucloxacillin (HLA-B*57∶01) [21], and amoxicillin-clavulanate (HLA-DRB1*15∶01) [35]. Other genetic risk factors such as those involving drug metabolizing enzymes are exemplified by CYP2C8*4 (diclofenac), CYP2E1*1A (isoniazid), GSTT1-M1 (troglitazone), and UGT2B7*2 (diclofenac) are also reported [36], [37], [38].
Drug safety is a key area of focus in the FDA. Modernizing safety evaluation has been advocated by the FDA's recent initiative on advancing regulatory science with a proposal of incorporating both in vitro and in silico methodologies in drug development and safety assessment [39]. The DILIps follows the same philosophy that underlies this new initiative at the FDA. It could be a predictive system for FDA to utilize and reference when hepatotoxicity issues arise during the various stages of the regulatory review process. It could also serve as a proof-of-concept approach of using predictive systems for drug safety to support the FDA's regulatory science. While the DILIps was developed for DILI, its methodology can be applied equally well to address other drug safety issues, such as renal and cardiovascular toxicity.
|
10.1371/journal.pntd.0006608 | Vitamin D status contributes to the antimicrobial activity of macrophages against Mycobacterium leprae | The immune system depends on effector pathways to eliminate invading pathogens from the host in vivo. Macrophages (MΦ) of the innate immune system are armed with vitamin D-dependent antimicrobial responses to kill intracellular microbes. However, how the physiological levels of vitamin D during MΦ differentiation affect phenotype and function is unknown.
The human innate immune system consists of divergent MΦ subsets that serve distinct functions in vivo. Both IL-15 and IL-10 induce MΦ differentiation, but IL-15 induces primary human monocytes to differentiate into antimicrobial MΦ (IL-15 MΦ) that robustly express the vitamin D pathway. However, how vitamin D status alters IL-15 MΦ phenotype and function is unknown. In this study, we found that adding 25-hydroxyvitamin D3 (25D3) during the IL-15 induced differentiation of monocytes into MΦ increased the expression of the antimicrobial peptide cathelicidin, including both CAMP mRNA and the encoded protein cathelicidin in a dose-dependent manner. The presence of physiological levels of 25D during differentiation of IL-15 MΦ led to a significant vitamin D-dependent antimicrobial response against intracellular Mycobacterium leprae but did not change the phenotype or phagocytic function of these MΦ. These data suggest that activation of the vitamin D pathway during IL-15 MΦ differentiation augments the antimicrobial response against M. leprae infection.
Our data demonstrates that the presence of vitamin D during MΦ differentiation bestows the capacity to mount an antimicrobial response against M. leprae.
| A key function of MΦ is to recognize, phagocytose and mount an antimicrobial response against microbial pathogens to defend the host. In humans, monocytes are recruited to the site of infection and differentiate into MΦ upon the onset of microbial infection. The MΦ phenotype and function are determined by the cytokine profile of the microenvironment in which the monocyte enters. Additionally, vitamin D is known to trigger direct antimicrobial responses against invading pathogens in MΦ, but also disrupts the differentiation of immune subsets within the myeloid lineage. Therefore, we investigated whether vitamin D status during MΦ differentiation influenced either phenotype or function. Here, we found that the IL-15 MΦphenotype is sustained regardless of vitamin D status. In contrast, antimicrobial MΦ differentiated in the presence of vitamin D exhibited a robust expression of an antimicrobial peptide, relative to MΦ differentiated in the absence of vitamin D. The antimicrobial MΦ armed with cathelicidin prior to M. leprae challenge demonstrated a strong antimicrobial response against the invading pathogen. Our study reveals that the presence of sufficient levels of vitamin D prior to microbial infection contributes to effectively reduce the viability of the pathogen in MΦ.
| The MΦ is a sentinel of the innate immune system that serves as the first line of defense to recognize and destroy invading microbes. In human MΦ, activation by a toll-like receptor 2/1 (TLR2/1) ligand or interferon-γ (IFN-γ) triggers a direct antimicrobial response that depends upon the level of available vitamin D [1–3]. The vitamin D-dependent antimicrobial pathway involves the induction of IL-15 and IL-32, the conversion of 25D3 to bioactive 1,25-dihydroxyvitamin D (1,25D3) and subsequent activation of the vitamin D receptor (VDR) to induce the expression of the antimicrobial peptides including cathelicidin, autophagy and phagolysosomal fusion [2, 4–8]. This antimicrobial pathway is not induced in MΦ if the levels of 25D are not sufficient.
Macrophages demonstrate phenotypic heterogeneity which confer distinct functions in the innate immune system [9]. IL-15 MΦdemonstrate a vitamin D-dependent antimicrobial profile which includes the expression of CAMP mRNA [4, 10]. In contrast, primary human monocytes treated with IL-10 differentiate into phagocytic macrophages (IL-10 MΦ), which readily take up bacteria but weakly express the vitamin D-dependent antimicrobial pathway [10]. These MΦ subtypes can be identified by a specific cell surface phenotype, both IL-15 MΦ and IL-10 MΦ express CD209 but only IL-10 MΦ express CD163. As such, IL-15 MΦ and IL-10 MΦ are differentially identified in the polar forms of leprosy caused by the intracellular bacterium M. leprae, correlating with the different outcomes of infection.
In addition to its role in MΦ antimicrobial function, vitamin D has long been recognized to affect the differentiation of diverse cell types, including cells of the myeloid lineage [11]. Activation of the VDR converts malignant myeloid leukemia cells into non-proliferating monocytes or MΦ [12–15]. Dendritic cells differentiated in the presence of 25D3 or 1,25D3 demonstrate aberrant differentiation and decreased antigen presentation in vitro [16, 17]. MΦ differentiated in vitamin D have also demonstrated a change in phenotype and phagocytic function in vitro [15, 18]. Most of these studies were performed by adding non-physiological concentrations of the bioactive form of 1,25D3, such that the ability of the differentiating cell to utilize physiologic concentrations of 25D3 has not been substantially investigated. Although controversy still exists on the normal concentrations of 25D, we used the Endocrine Society Clinical Practice Guidelines which define vitamin D deficiency as below 20ng/mL (50nM), insufficiency as 21–29 ng/mL (52.5nM-72.5nM), sufficient levels as more that 30ng/mL (75nM) [19]. In humans, 1,25D levels are regulated to be constant, such that the available level of 25D determines the amount of bioactive 1,25D that is generated in an activated MΦ and is therefore key to innate immune function [3]. Therefore, the aim of our work is to study the effects of physiological levels of 25D3 during IL-15 MΦ differentiation, function and antimicrobial response against M. leprae.
Experiments with three or more measurements were analyzed using One Way ANOVA or with Student-Newman-Keuls Method (*P<0.05, **P<0.01, ***P<0.005, ****P<0.001) for pairwise analyses using GraphPad Prism 7 software. Error bars represent the standard error of the mean between individual donor values. A two-tailed student’s t-test was used to compare two different experimental conditions.
This study was conducted according to the principles expressed in the Declaration of Helsinki and was approved by the Institutional Review Board (IRB) of the University of California at Los Angeles (UCLA). Human peripheral blood from healthy donors was acquired with informed consent (UCLA Institutional Review Board #11–001927). All adult subjects provided written informed consent. Peripheral blood mononuclear cells (PBMCs) were isolated from the blood of healthy donors using Ficoll-Paque (GE healthcare) and monocytes were purified with plastic adherence as previously described [20].
Adherent monocytes were cultured in the presence of IL-15 (R&D Systems, 200ng/ml) or IL-10 (R&D Systems, 10ng/ml) for 48 hours using Serum Free MΦ media (SFM) (Gibco) at 37°C and 5% CO2. Cell phenotypes were consistent with previously published data [10].
The following antibody clones were used per manufacturers’ protocol for flow cytometry: CD209 (DCN46), CD163 (GHI/61), CD16 (3G8), CD14 (M5E2), and CAMP/LL37/FALL39/Cathelicidin Antibody (OSX12). Differentiated MΦwere harvested and stained as previously described [1, 4, 5, 20].
RNA was harvested using TRIzol reagent (Life Technologies) via phenol-chloroform extraction, followed by RNA cleanup and DNase digestion using the RNeasy Miniprep Kit (Qiagen) as previously described [21]. cDNA was synthesized using iScript cDNA synthesis kit (Bio-Rad) and stored at -80°C. Primer sequences were used as follows: CYP27B1 F: ACC CGA CAC GGA GAC CTT C, CYP27B1 R: ATG GTC AAC AGC GTG GAC AC; CAMP F: TGG GCC TGG TGA TGC CT, CAMP R: CGA AGG ACA GCT TCC TTG TAG C H36B4 F: CCA CGC TGC TGA ACA TGC T, H36B4 R: TCG AAC ACC TGC TGG ATG AC. Real-time PCR was performed using SYBR Green (Kapa Biosystems) according to the manufacturers’ protocol. cDNA levels were normalized with H36B4 as the housekeeping gene. Relative CAMP mRNA levels were normalized to IL-15 MΦ differentiated in the absence of vitamin D and shown as fold-change (FC). Relative CYP27B1 mRNA levels were normalized to IL-15 MΦ baseline levels and shown as fold-change (FC) as previously described [1, 6, 10, 21].
Primary human monocytes were seeded onto chamber slides (BD falcon) and treated with IL-15 with or without the presence of vitamin D. The cells were fixed and permeabilized using fixation/permeabilization solution kit (BD Bioscience) as indicated by the manufacturer. Cells were blocked with 10% human serum for 20 minutes and stained with CAMP/LL37/FALL39/Cathelicidin Antibody (OSX12) antibody at 10μg/mL overnight. The monolayers were washed three times with cold-PBS and stained with a biotinylated-horse anti-mouse antibody (Bio-Rad) at 10μg/mL at room temperature for one-hour. The monolayers were washed again three times with cold-PBS and stained with streptavidin-conjugated to Alexa Fluor 488 (Invitrogen) protected from light as previously described [1, 5]. The cells were washed with PBS and sealed with ProLong Gold antifade reagent with DAPI (Invitrogen). Microscopy images were analyzed with the SP8-SMD confocal microscope (Leica) at the Advanced Microscopy Laboratory Macro-Scale Imaging Laboratory, California Nanosystems Institute, UCLA [22].
IL-15 MΦwere differentiated in the presence or absence of vitamin D in a 24-well tissue culture plate (Corning). The cells were harvested and fixed with 4% paraformaldehyde for 15 minutes at room temperature in a V-bottom plate (Corning). After fixation, the MΦ were permeabilized with 0.5% saponin (Sigma) in PBS and quickly washed with a series of PBS washes. The cells were then stained using same staining protocol as described above. The cells were acquired with a BD LSRII in the Janis V. Gorgi Flow Cytometry Core Laboratory at UCLA. All analysis was done using FlowJo 10.4.2 software.
Cells were infected with PE-labeled M. leprae and harvested 24 hours post infection (PI). The cells were blocked with 10% human serum for 20 minutes at room temperature and stained with an anti-CD14 antibody as indicated by the manufacturer for 30 minutes on ice. The cells were then fixed with 4% paraformaldehyde for 15 minutes and analyzed by flow cytometry as previously described [4, 20]. The cells were acquired with a BD LSRII in the Janis V. Gorgi Flow Cytometry Core Laboratory at UCLA. All analysis was done using FlowJo 10.4.2 software. The same samples were also analyzed using IdEAS Software, ImageStream (Amnis) as explained below.
The ImageStreamX MarkII imaging flow cytometer from Amnis Corporation was used for acquisition at 60X magnification. Anti-human CD14 antibody conjugated to PacBlue channel 1 was detected on the 405 nm laser at 30.00 mW, and PE labeled M. leprae channel 3 was detected off the 488 nm laser set at 70.00 mW. Acquisition was set to collect 5000 objects from the single cell population (Aspect Ratio Bright Field channel 4 vs. Area Bright Field channel 4).
Data was analyzed using the Amnis IDEAS software. A compensation matrix was first created using single-stained CD14 labeled macrophages and PE calibrite beads (BD). Compensated data was then applied to a template with a gating hierarchy. Focused cells were first selected from the higher population of the Gradient RMS histogram; single cells were then chosen using the same gating strategy applied during acquisition and lastly, PE and PacBlue double positive cells were then applied to the Spot Count Wizard. Two populations containing 10 or more cells, each expressing high or low values of spots (single PE labeled M. leprae bacterium) were manually selected. The Spot Count Wizard has the ability to measure uptake and count spots, thus providing individual bacterial counts per cell.
IL-15 MΦwere differentiated with or without the indicated amount of 25D for 48 hours. The MΦ were then infected with M. leprae overnight at an MOI of 10. The cells were washed with SFM to remove extracellular bacteria and treated with IL-15 and the indicated amounts of 25D. The infection progressed for 24 hours, 48 hours and 120 hours and all the material in the well was harvested into a 15mL conical tube. To ensure that all the material in the well was harvested, each well was washed with a series of cold PBS-EDTA washes and accumulated into their respective 15mL conical tube. Each 15mL conical tube was placed into a centrifuge for 300xg for 10 mins at 4°C. The supernatants were carefully removed and the genomic DNA was isolated as previously described [21, 23]. We compared RLEP DNA levels of M. leprae with the H36B4 levels of the IL-15 MΦto measure bacterial burden using real-time PCR using SYBR Green as indicated by the manufacturer (Kapa Biosystems) [24, 25]. The following primers sequences were used: RLEP F: GCA GCA GTA TCG TGT TAG TGA A, RLEP R: CGC TAG AAG GTT GCC GTA T; H36B4 F: CCA CGC TGC TGA ACA TGC T, H36B4 R: TCG AAC ACC TGC TGG ATG AC. The bacteria burden at each time point was normalized to IL-15 MΦdifferentiated without vitamin D to quantify relative bacteria burden.
To investigate the effect of 25D3 on MΦ differentiation, we used SFM, allowing us to control the amount of 25D3 in the culture. SFM contains neither 25D3 nor any other vitamin D analogues, such that 25D3 can be added at defined concentrations. Thus, SFM has an advantage over fetal calf sera or human sera that have varying amounts of 25D3. Previously, IL-15 and IL-10 were shown to induce the differentiation of monocytes into distinct MΦ populations, however, these experiments were performed using FCS, which contains low levels of 25D3. Thus, it was unclear whether vitamin D status affects the differentiation of monocytes into IL-15 MΦ and IL-10 MΦ.
All experiments here involve MΦs derived from cytokine treated monocytes as previously reported [10]. Monocytes were cultured with either IL-15 or IL-10 for 48 hours in SFM with or without the addition of 25D3 (10-8M 25D3). This is equivalent to the physiologic concentration in vitamin D sufficient serum of 10-7M 25D3, which is then diluted to 10% serum in cell cultures [1, 26]. Both IL-15 and IL-10 induced CD209 expression as assessed by flow cytometry, but only IL-10 induced CD163 expression (Fig 1A), similar to differentiation in FCS [10]. Examining co-expression of CD209 and CD163, we found that IL-15 induced CD209+CD163- MΦ, whereas IL-10 induced CD209+CD163+ MΦ, accounting ~80% of cells. We also found that the IL-15 MΦand IL-10 MΦ derived in SFM express the MΦ specific marker CD16. The average surface expression of CD16 increased in IL-15 MΦ when differentiated in 25D3 to similar levels seen on IL-10 MΦ, but was not significant (p = 0.07). In addition, we observed that 25D3 status did not significantly alter the surface expression of CD209, CD163, CD16 or the coexpression of CD209+CD16+ whether derived using IL-15 or IL-10 (Fig 1A). The frequency of CD14 was expressed on IL-15 MΦ, with a small but significant enhancement by 25D3, to the level expressed on IL-10 MΦ (Fig 1A). This was also reflected in an increase in cellular abundance as measured by the change in mean fluorescence intensity (ΔMFI) for MΦ derived in IL-15 but not IL-10 (Fig 1B). Overall, SFM supported the differentiation of monocytes by IL-15 and IL-10 into divergent MΦ phenotypes, which were generally similar whether differentiated in the presence or absence of 25D3.
Although 25D3 did not dramatically affect the differentiation of MΦ by phenotype, we next investigated whether the presence of 25D3 during differentiation affected MΦ function. The induction of the antimicrobial protein cathelicidin is essential for the vitamin D-dependent antimicrobial response against intracellular mycobacteria in infected MΦ [5, 27]. To determine whether 25D3 status during MΦ differentiation results in activation of the vitamin D-dependent antimicrobial pathway, we treated monocytes with IL-15 and IL-10 in the presence of increasing concentrations of 25D3 during differentiation and measured CAMP mRNA levels after 48 hours by qPCR. Conditioning during differentiation of both IL-15 MΦ and IL-10 MΦ in SFM supplemented with increasing level of 25D3 resulted in a significant dose-dependent induction of CAMP mRNA (Fig 2A). At all concentrations of 25D3, the CAMP mRNA expression was more robust in IL-15 MΦ relative to IL-10 MΦ. We observed a ~315-fold induction of CAMP mRNA in SFM supplemented with 10-8M 25D3 as compared to media without 25D3 in IL-15 MΦ, and ~500-fold induction of CAMP mRNA in SFM containing 10-7M 25D3 (Fig 2A). In comparison, we observed a ~80-fold induction of CAMP mRNA in SFM containing 10-8M 25D3 in IL-10 MΦ and a ~100-fold induction of CAMP mRNA in SFM containing 10-7M 25D3 (Fig 2A). The baseline values of CYP27B1 mRNA expression was not significant between IL-15 MΦand IL-10 MΦ(S1 Fig).
We determined whether the induction of CAMP mRNA was associated with expression of cathelicidin protein using intracellular flow cytometry. The CAMP mRNA expression levels in IL-15 MΦ correlated with both the frequency of cathelicidin and the cathelicidin protein abundance as measured by ΔMFI (Fig 2B and 2C). The average frequency of cathelicidin was ~13% in IL-15 MΦ derived in 10-8M 25D3 and ~30% in 10-7M 25D3 supplemented SFM (Fig 2B). The ΔMFI was ~45 AU in IL-15 MΦ derived in 10-7M 25D3 and ~80 AU in 10-8M 25D3 supplemented SFM (Fig 2C).
Representative fluorescence microscopy images of IL-15 MΦ conditioned in 25D3 indicates that cathelicidin protein accumulates in the intracellular vesicles proximal to the host nucleus, but not in IL-15 MΦdifferentiated in no 25D3 (Fig 2D). These data collectively indicate that cathelicidin mRNA and protein expression directly correlate with 25D3 status during IL-15 induced MΦ differentiation.
An important function of antimicrobial MΦ is the phagocytosis of pathogens to contain microbes in the host in vivo. However, it is unclear how vitamin D status may alter the phagocytic function of the MΦ during M. leprae infection. To assess phagocytic function, IL-15 MΦwere conditioned with or without 25D3, infected with PE labeled-M. leprae for 24 hours, stained for the MΦ specific marker CD14 and phagocytosis was analyzed by flow cytometry and image stream flow cytometry. The efficiency of M. leprae infection in IL-15 MΦdifferentiated in the absence of vitamin D or in the presence of either 10-8M 25D3 or 10-7M 25D3, was not statistically different, although somewhat greater in culture in which no 25D3 was present (Fig 3A). Image stream flow cytometry analysis of the same samples demonstrated a frequency of infection of CD14+mLEP+ cells ranging from 30%, 35%, to 25%, when conditioned with no vitamin D, 10-8M 25D3 or 10-7M 25D3, respectively (Fig 3B). Using an unsupervised spot counting function of image stream flow cytometry, we determined the frequency of CD14+ cells containing varying numbers of intracellular M. leprae and no effect of 25D3 was observed (Fig 3C). Images from image flow cytometry analysis show the number of bacteria per MΦ (Fig 3D). Cells that contained either 7 or 8 bacteria all showed large clumps of bacteria in which were difficult to interpret as an accurate number of bacteria. Overall no significant difference was observed in the number of bacteria per cell. These data indicate that vitamin D status does not alter the phagocytic capacity of IL-15 MΦ.
After the phagocytosis of invading pathogens, a major function of MΦ is to effectively mount an antimicrobial response to defend the host. However, it is unclear whether sufficient levels of 25D3 will provide MΦ with the capacity to mount an antimicrobial response. To investigate whether 25D3 status during MΦ differentiation affects the antimicrobial response, we simultaneously measured the kinetics of CAMP mRNA induction and antimicrobial activity against M. leprae in IL-15 MΦ. IL-15 MΦ were conditioned with or without 25D3, infected with M. leprae for 24, 48, and 120 hours, at which time both RNA and DNA were harvested.
The levels of CAMP mRNA in M. leprae infected IL-15 MΦ at 24 hours were relatively low as compared to the previous experiments in which CAMP mRNA was measured in uninfected MΦ, to the extent that CAMP mRNA was not detectable in some donors at this time point. However, the cathelicidin protein colocalized with M. leprae in IL-15 MΦdifferentiated in 25D3, but not in IL-15 MΦdifferentiated in no 25D3 24 hours post M. leprae infection (Fig 4A). The low level of CAMP mRNA expression was not different whether the MΦ were differentiated in the presence or absence of 25D3 (Fig 4B). One possibility for the absence of CAMP mRNA but presence of cathelicidin protein at 24 hrs post infection is that upon infection with M. leprae the CAMP mRNA is downregulated yet the protein was already synthesized during differentiation.
At 48 hours after M. leprae infection, CAMP mRNA expression was approximately 1000 fold in the IL-15 MΦ differentiated in 25D3, at either 10-8M 25D3 or 10-7M 25D3 (Fig 4C). At 120-hours post infection, the relative CAMP mRNA remained significantly increased in the MΦ differentiated in 25D3, approximately 200–330 fold greater than in MΦ differentiated in the absence of 25D3 (Fig 4D).
The M. leprae burden was measured in the infected IL-15 MΦ according to the level of bacterial DNA. The M. leprae burden in infected IL-15 MΦ was not affected by the presence of 25D3 during differentiation as assessed at 24 or 48 hours (Fig 4E and 4F). In one of the five donors, we noted a reduction in bacterial burden at 48 hours. However at 120 hours post infection the relative bacteria burden significantly decreased to 0.67 and 0.44 in IL-15 MΦ differentiated in 10-8M and 10-7M 25D3 compared to no 25D3, respectively (Fig 4G). The decrease in viability of M. leprae at 120 hours was not due to differences in macrophage number, as H36B4 levels remained constant. Antimicrobial activity against M. leprae was detected in IL-15 MΦ differentiated in 25D3 in all five donors. These data indicate that the presence of 25D3 during the IL-15 MΦ differentiation program and throughout M. leprae infection contributes to the vitamin D-dependent antimicrobial response against by M. leprae.
The ability of human MΦ to mount an effective response against intracellular mycobacteria depends in part upon their ability to induce the vitamin D-dependent antimicrobial pathway [1, 2, 5, 28]. Although sufficient levels of vitamin D are required for optimal MΦ effector function, previous studies have indicated that myeloid cell differentiation and function can be altered by vitamin D bioavailability. Here, we investigated whether the level of 25D influences MΦ differentiation and programming of an antimicrobial response against M. leprae. The distinct phenotypes of IL-15 MΦ and IL-10 MΦ were largely sustained during differentiation from monocytes regardless of 25D3 status, yet onlyIL-15 MΦdifferentiated in the presence 25D3 robustly triggered the expression of CAMP mRNA and cathelicidin protein levels in a dose-dependent manner. Vitamin D status did not alter the phagocytic function of IL-15 MΦ, but a significant decrease in bacteria burden against M. leprae was observed at 120 hours post-infection. These data indicate that 25D3 status during IL-15 MΦdifferentiation permits the induction of an antimicrobial response against intracellular M. leprae.
It is important for the host to mount an antimicrobial response against intracellular mycobacteria before the bacteria employ evasion mechanisms that help establish infection and progress to clinical disease [22, 23]. A key finding of the present study was that the addition of 25D3 during the IL-15 induced differentiation of monocytes into MΦ led to a robust induction of the vitamin D-dependent antimicrobial pathway, including the induction of cathelicidin and an antimicrobial response against M. leprae. We detected a 315-fold induction of CAMP mRNA in IL-15 MΦ differentiated in the presence 10-8M 25D3 as compared to SFM without 25D3 and a 1/3 reduction in the M. leprae burden in infected cells. Previously we have shown that IL-15 MΦ supplemented with 10-8M 25D3 post-differentiation demonstrated a 5-fold increase in CAMP mRNA and a ~50% reduction in avirulent M. tuberculosis (H37ra) viability [4]. However, these IL-15 MΦ were differentiated in 10% 25D insufficient FCS (16nM). These data collectively suggest that 25D levels during IL-15 MΦ differentiation facilitate their antimicrobial function as part of the innate immune response. There are situations that allow the pathogen to escape the vitamin D antimicrobial response. For example, genetic polymorphisms in the VDR have been associated with increased susceptibility to mycobacterial infection [29]. M. leprae evades the vitamin D antimicrobial response via the induction of a microRNA that targets the pathway [23], and by induction of type 1 interferon leading to IL-10 and subsequent suppression of the vitamin D pathway [22]. These data imply that upon the onset of microbial challenge, monocytes that are recruited to the site of infection are dependent on the presence of sufficient levels of 25D to differentiate into powerful IL-15 MΦthat fend of M. leprae evasion mechanisms and effectively reduce bacterial viability [1, 5, 30].
The addition of 25D3 during the IL-15 induced differentiation of monocytes into MΦ affected antimicrobial function, but we observed little change in cell phenotype. Regardless if the MΦ were differentiated with or without 25D3, we found that the distinct phenotypes of IL-15 MΦ and IL-10 MΦ were largely not affected. In particular, the IL-15 MΦ were CD209+CD163- and the IL-10 MΦ were CD209+CD163+. Only IL-15 MΦ differentiated in 25D3 demonstrated both a significant increase in CD14 frequency and cellular abundance. Although CD14 is a marker that identifies VDR-activated MΦs [31], we have no evidence that the differences in CD14 expression directly affected function as phagocytic capacity was not affected. In contrast to our findings, the addition of 1,25D3 during differentiation of monocytes into MΦ by macrophage colony-stimulating factor (M-CSF) decreased phagocytic function and the release of pro-inflammatory cytokines [18]. Similarly, the addition of 25D3 or 1,25D3 during differentiation of monocytes into dendritic cells by granulocyte-macrophage colony-stimulating factor (GM-CSF) plus IL-4 decreased the expression of DC-specific surface markers CD1a, CD80, CD86, and MHC class-II, as well as antigen presentation capacity [16, 17]. In these experiments the levels of 1,25D3 were supraphysiologic, although 25D3 was added at physiologic levels. Taken together with our findings, these findings suggest that although physiologic levels of 25D may alter the differentiation of DC, it permits MΦ differentiation and enhances MΦ antimicrobial function.
In the present study we determined that clinically sufficient levels of 25D3 led to a functional difference in IL-15 MΦ, relative to MΦ differentiated in the absence of 25D3. In humans, there is a range of 25D levels that can be classified from deficient (45nM) to sufficient (98nM) [2]. Previously, we compared the ability of African American sera and Caucasian sera to induce the expression of the mRNAs encoding the antimicrobial peptides cathelicidin and beta-defensin 2 and found that African American sera was less capable to induce the antimicrobial peptides ex vivo due to the relatively lower 25D sera levels [1, 2]. Both exogenous 25D supplementation to African American sera ex vivo and 25D supplementation to vitamin D deficient individuals in vivo significantly enhanced CAMP mRNA expression in activated monocytes and MΦ in vitro [1, 2, 32].
Our data suggest that people with higher levels of vitamin D will derive MΦ with more antimicrobial function that could prevent the establishment of infection; however, testing the effects of vitamin D status on the prevention of infection by M. tuberculosis is challenging. The ability to acquire a large enough population with differential 25D levels randomly will be difficult as 25D status strongly correlates with season [33], as such there are few studies that investigate the interaction of 25D status with infection by the pathogen. Deficient levels of 25D in household contacts of TB patients demonstrated either increased latent TB incidence or positive tuberculoid skin tests [33, 34]; however the number of patients that acquire active TB is unclear [35–37]. These data support continued and more thorough investigation into whether vitamin D supplementation of deficient and insufficient individuals in vivo can enhance the MΦ antimicrobial response against mycobacterial infections and contain the spread and outcome of disease.
In conclusion, we found that vitamin D-dependent antimicrobial MΦ differentiated in the presence of sufficient levels of 25D3 sustain a MΦ phenotype and exhibit an antimicrobial response against M. leprae. Our model indicates that vitamin D-dependent antimicrobial MΦ differentiated in the presence of sufficient 25D are capable of intrinsic microbicidal activity against infection. In contrast, the same MΦ differentiated in the low levels of 25D require the addition of exogenous 25D to induce activity [1, 2, 5, 32]. These data suggest that sufficient levels of 25D at the site of microbial infection allow recruited monocytes to differentiate into vitamin D-dependent antimicrobial MΦ with the capacity to effectively reduce the viability of intracellular bacteria. Future clinical trials that study the relationship between vitamin D supplementation and susceptibility to microbial infection will determine if the prophylactic effects of vitamin D are therapeutically beneficial.
|
10.1371/journal.ppat.1004173 | The Cytoplasmic Domain of Varicella-Zoster Virus Glycoprotein H Regulates Syncytia Formation and Skin Pathogenesis | The conserved herpesvirus fusion complex consists of glycoproteins gB, gH, and gL which is critical for virion envelope fusion with the cell membrane during entry. For Varicella Zoster Virus (VZV), the complex is necessary for cell-cell fusion and presumed to mediate entry. VZV causes syncytia formation via cell-cell fusion in skin and in sensory ganglia during VZV reactivation, leading to neuronal damage, a potential contributory factor for the debilitating condition of postherpetic neuralgia. The gH cytoplasmic domain (gHcyt) is linked to the regulation of gB/gH-gL-mediated cell fusion as demonstrated by increased cell fusion in vitro by an eight amino acid (aa834-841) truncation of the gHcyt. The gHcyt regulation was identified to be dependent on the physical presence of the domain, and not of specific motifs or biochemical properties as substitution of aa834-841 with V5, cMyc, and hydrophobic or hydrophilic sequences did not affect fusion. The importance of the gHcyt length was corroborated by stepwise deletions of aa834-841 causing incremental increases in cell fusion, independent of gH surface expression and endocytosis. Consistent with the fusion assay, truncating the gHcyt in the viral genome caused exaggerated syncytia formation and significant reduction in viral titers. Importantly, infection of human skin xenografts in SCID mice was severely impaired by the truncation while maintaining the gHcyt length with the V5 substitution preserved typical replication in vitro and in skin. A role for the gHcyt in modulating the functions of the gB cytoplasmic domain (gBcyt) is proposed as the gHcyt truncation substantially enhanced cell fusion in the presence of the gB[Y881F] mutation. The significant reduction in skin infection caused by hyperfusogenic mutations in either the gHcyt or gBcyt demonstrates that both domains are critical for regulating syncytia formation and failure to control cell fusion, rather than enhancing viral spread, is severely detrimental to VZV pathogenesis.
| Varicella zoster virus (VZV) infects the human population globally, causing chickenpox in children and shingles in adults. While those afflicted with shingles experience severe pain that might last from weeks to months, the cause is not known. Biopsies of VZV infected skin and specimens of nerve ganglia collected at autopsy from patients with shingles at the time of death contain multi-nucleated cells, indicating that the virus is able to cause fusion between infected cells. Since the destruction of nerve cells that results from this process is likely to contribute to the pain associated with shingles, it is important to understand how the virus causes infected cells to fuse. We find that VZV cell-cell fusion is regulated by the intracellular facing domain of glycoprotein H (gH), a viral protein present on the surface of infected cells. This regulation was dependent upon the physical length of the domain, not a specific sequence. Loss of this regulation increased cell-cell fusion causing the formation of larger multi-nucleated cells that limited the ability of the virus to effectively spread in human skin. Our study provides new insight into how VZV manipulates host cells during infection and controls the spread of the virus in tissues.
| Varicella Zoster Virus (VZV) is a ubiquitous human pathogen that causes varicella (chickenpox) in children and zoster (shingles) in adults [1]. Primary infection with VZV initiates at the mucosal epithelium following contact with respiratory droplets or skin vesicle fluid from infected individuals [2]. Viral dissemination in the host occurs by T cell-associated viremia resulting in the infection of skin cells, formation of lesions (chickenpox), and the establishment of latency in neurons of sensory nerve ganglia [3]. Reactivation of VZV from latently infected neurons causes shingles, potentially leading to postherpetic neuralgia (PHN), a condition characterized by severe pain that can last from days to months and in rare cases, for years [4], [5].
Entry of enveloped viruses, including herpesviruses, into a host cell requires fusion of the virion envelope with the host cell membrane [6]. Some enveloped RNA viruses, such as respiratory syncytial virus and DNA viruses, also induce fusion of cell membranes between the infected cells resulting in the formation of a multi-nucleated cell called a syncytium [7], [8]. For VZV, syncytia formation is a hallmark of infection observed in skin lesions as well as trigeminal ganglia taken from cadavers when the individual had zoster at the time of death [9], [10]. Fusion between neurons and their satellite cells in ganglia has been postulated to contribute to the extensive damage caused by VZV reactivation in sensory nerve ganglia and to be a factor for PHN [11]. Mechanisms that regulate VZV syncytia formation can be assessed in cultured melanoma cells and examined for their role in pathogenesis using the human skin and dorsal root ganglia xenografts in the severe combined immunodeficiency (SCID) mouse model [11], [12].
The minimal herpesvirus proteins required for fusion have been determined using virus-free assays that utilize cell-cell fusion as a surrogate for virion envelope and cell membrane fusion. The requirements for fusion in vitro differ among the herpesviruses but consist of a core set of glycoproteins that includes glycoprotein B (gB) and the heterodimer of glycoproteins H (gH) and L (gL) [13]. Among alphaherpesviruses, transient expression of gB and gH-gL of VZV and pseudorabies virus (PrV) has been demonstrated to be necessary and sufficient for inducing cell fusion in transfected cells [14], [15], with the caveat that the last eight amino acids (834-841) of VZV gH must be removed for the detection of enhanced fusion under these conditions [15]. Other herpesviruses require additional accessory proteins for fusion, such as gD of herpes simplex virus-1 (HSV-1) and gp42 of Epstein Barr Virus for certain cell types [16], [17]. While transient expression of these glycoproteins induces cell fusion in vitro, herpesvirus replication does not trigger syncytia formation, with the exception of VZV and a few naturally occurring HSV variants [18], [19]. Given the differences in requirements for herpesvirus fusion and virally induced syncytia formation, it is necessary to define the functional role of the fusion machinery components of specific herpesviruses independently.
Similar to other herpesviruses, the fusion machinery core of VZV consists of gB, gH, and gL, which are expressed from open reading frame (ORF) 31, 37, and 60, respectively [13], [20]. While the crystal structure of the gB/gH-gL fusion complex has not been determined, X-ray crystallography of the individual glycoprotein components of other herpesviruses has provided insight into their function and roles in fusion. The crystal structure of HSV-1 gB determined the ectodomain to have characteristics of a type III fusion protein [21], whereas the crystal structure of HSV-2 heterodimer gH-gL found the ectodomain to have a “boot-like” structure that differed from other known fusogenic proteins [22]. The current model proposes that gB is expressed as a trimer on the virion surface with two internal fusion loops per monomer and acts as the primary fusogen during herpesvirus entry whereas the gH-gL heterodimer facilitates the fusion potential of gB [13]. However, it remains unclear whether or how gB and gH-gL interact with each other during fusion and while HSV gB/gH-gL complexes have been identified in infected and transfected cells [17], [23], it is not certain if they represent the functional fusion unit. Furthermore, the absence of structural data of a pre-fusion gB form to complement the post-fusion model of gB has limited the understanding of the changes in gB conformation required to trigger fusion. Nonetheless, VZV gB is presumed to be the primary fusogen and gH, together with gL, are essential components of the complex with all three proteins being critical for viral entry and syncytia formation based upon the proposed model.
VZV gH is essential for replication as shown by deleting ORF37 from the VZV genome and inhibiting VZV infection with antibodies against gH in cultured cells and human skin xenografts in vivo [24]–[26]. Initially expressed as a 100 kilodalton (kDa) immature polypeptide, gH forms a heterodimer with gL, which is necessary for its maturation to a 118 or 130 kDa form [20], [27]. gH is processed in the Golgi apparatus and traffics to the cell surface, where it then undergoes endocytosis and returns to the trans-Golgi network (TGN) for incorporation into the virion envelope [28]–[32]. gH is predicted to have a single membrane-spanning hydrophobic region with the N-terminus facing the extracellular space when the protein is expressed on the virion envelope or the infected cell surface and the C-terminus projecting toward the viral capsid or the cytoplasm [27].
Our previous work investigating the role of the VZV gH ectodomain in cell fusion led to the identification of motifs in domain I and III that were important for in vitro cell fusion and skin pathogenesis [27]. Other in vitro studies of the gHcyt identified the 835YNKI838 sequence as a functional YXXΦ endocytosis motif that contributes to cell fusion when HeLa cells infected with vaccinia virus (VV) expressing T7 polymerase are cotransfected with gH and gL vectors [33]. However, the role of the gHcyt in VZV syncytia formation during infection and pathogenesis has not been investigated.
In a recent study, we identified an immunoreceptor tyrosine-based inhibition motif (ITIM) in the VZV gB cytoplasmic domain (gBcyt) that has a regulatory role in cell fusion and syncytia formation [34]. Substitution of the tyrosine residue in the ITIM with phenylalanine (Y881F) caused an induction in cell fusion at levels significantly greater than the wildtype when the gB mutant was coexpressed with gH[TL], a gH construct lacking amino acids 834-841, and gL in vitro. Recombinant viruses with wildtype gH and the gB[Y881F] mutation caused aggressive syncytia formation and reduced viral spread and replication kinetics. Severe impairment of skin pathogenesis in vivo was also observed. How the regulation of fusion by the gBcyt relates to the gHcyt has not been fully explored.
In this study, we report that regulation of VZV syncytia formation by the cytoplasmic domain of gH is critical for skin pathogenesis and depends on the length of the domain, not its specific amino acid sequence. Furthermore, both the length of the gHcyt and the gBcyt ITIM must be preserved to prevent exaggerated syncytia formation and allow for effective propagation of VZV. The cytoplasmic domains of these fusion complex proteins contribute differentially to regulation of VZV induced syncytia formation, with both having essential contributions to skin pathogenesis.
Institutional Animal Care and Use Committee (IACUC) review of all research involving animals was performed and procedures were approved by the Stanford University Administrative Panel on Laboratory Animal Care (Protocol ID: 11130). Stanford University complies with all federal and state regulations governing the humane care and use of laboratory animals, including the USDA Animal Welfare Act, and the Stanford University Assurance of Compliance with Public Health Service Policy on Humane Care and Use of Laboratory Animals. Acquisition and use of fetal material has been reviewed by the Stanford University Administrative Panel on Human Subjects in Medical Research and the scope of use does not meet the criteria for research involving human subjects. Anonymized fetal material is provided by the non-profit tissue supply organization Advanced Bioscience Resources, Inc. (ABR) in accordance with applicable federal and state regulations.
Human melanoma cells, LDL-GFP melanoma cells, and human embryonic lung fibroblasts (HELFs) were propagated in minimal essential media (MEM) supplemented with 10% fetal bovine serum (FBS) (Invitrogen), nonessential amino acids (100 µM; Omega Scientific), and antibiotics (penicillin, 100 U/mL; streptomycin, 100 µg/mL; Invitrogen) [27]. CHO-K1 Cre cells, which stably express Cre recombinase, were propagated in F-12K Nutrient Mixture with Kaighn's modification (Invitrogen) supplemented with 10% FBS, penicillin, and puromycin (8 µg/mL; Invitrogen) [27].
gH constructs were generated from the pME18s-gH[WT] vector (wildtype gH) or the pME18s-gH[TL] vector, which was a gift from Tadahiro Suenaga and Hisashi Arase (Osaka University, Osaka, Japan) [15], [27]. Primers containing the desired mutation were used to amplify two PCR products using AccuPrime Taq (Invitrogen). The amplicons were digested with either KpnI or AccI restriction enzymes (New England BioLabs), and blunt end ligated with digested pME18s-gH[TL] or the pME18s-gH[WT] vector. Ligated products were electroporated into TOP10F' Electrocomp E. Coli (Invitrogen). Clones were sequenced using the pME18s-KPN1 primer to confirm the mutation. (Table S1: primer list)
The Y835A, Y835F, TL, Δ834-841, V5, and 834StopV5 mutations were generated in the self-excisable pOka-DX BAC as previously described [27]. Briefly, the mutations were generated in the shuttle vector, pCR669V1-gH-Kan [27]. The modified vector was digested with NaeI and PmeI enzymes (New England BioLabs) and the fragment of interest was inserted into pOka-DX-ΔORF37 BAC, which lacks ORF37 [27], by RED recombination. Fragment insertion was confirmed by PCR using primers {37}F65680-65700 and {37}R68697-68717. The pOka-TK-GFP-gB[Y881F] and pOka-TK-GFP-gH[Δ834-841] BAC constructs were generated from the pOka-TK-GFP-ΔORF37 BAC and -ΔORF31 BAC, respectively, using digested fragments from shuttle vectors following previously described methods [34]. The replacement of TK(ORF36) with TK-EGFP was performed by RED recombination using the previously generated pOka-DX-ΔORF37 and -ΔORF31 BAC and a PCR product amplified with primers 5′-TK-EGFP and 3′-TK-EGFP. The pOka-TK-GFP- gB[Y881F]/gH[Δ834-841] BAC was generated by deleting ORF31 in the pOka-gH[Δ834-841] BAC, replacing the gene with the gB[Y881F] cassette, followed by replacing the ORF36 with the TK-EGFP cassette. To confirm no spurious recombination, vectors were digested with HindIII enzyme (New England BioLabs), separated by gel electrophoresis, and compared to pOka-DX BAC. Recombinant viruses were generated by calcium chloride transfection of melanoma cells with mutant BAC DNA. The mutations in the viral genome were confirmed by PCR and sequencing. Recombinant viruses were passed in HELFs until the MiniF- sequences Cat and SopA were not detectable by PCR (Table S1), as previously described [35]. PCR of ORF31 further confirmed VZV genomic DNA.
Briefly, CHO-K1 Cre cells were transfected with gH expressing vectors and the pcDNA3.1-gL vector, using Lipofectamine 2000 (Invitrogen) according to the manufacturer's instructions and lysates were harvested at 24 hours post transfection (hpt) [27]. Lysates from melanoma cells infected with recombinant VZV were harvested at 48 hours post infection (hpi). Immunoprecipitation and western blotting for gH used anti-gH monoclonal antibody SG3 (Meridian Life Science). Western blotting of viral proteins and cellular proteins was performed on infected-melanoma lysates as a control for infection and for sample loading. Membranes were probed for gE (mouse mAb 8612 anti-gE; Millipore), IE63 (rabbit anti-IE63; a kind gift of William Ruyechan, State University of New York, University at Buffalo, Buffalo, NY), and alpha-tubulin (clone B-5-1-2 mouse anti-α-tubulin; Sigma). All primary antibodies were detected using secondary horseradish peroxidase-conjugated antibodies to either anti-mouse or anti-rabbit and ECL Plus Detection Kit (GE Healthcare Bio-Sciences).
Cell-cell fusion was measured using the quantitative Cre reporter assay described previously [27], [34]. Briefly, CHO-K1 Cre cells were transfected with equimolar amounts with pCAGGs-gB, a gift from Tadahiro Suenaga and Hisashi Arase (Osaka University, Osaka, Japan) [15], pcDNA3.1-gL, and pME18s-gH mutants and co-cultured with LDL-GFP melanoma cells. The frequency of GFP positive cells indicating fusion events was quantified on a modified Digital FACStar running Diva hardware and software (BD Bioscience). Analysis was performed using FlowJo and all results were normalized to gB/gH[TL]-gL. A negative control that contained only pME18s- and pcDNA3.1-empty vectors was used to establish background levels of GFP expression, which were then subtracted from the fusion frequency data obtained with the test constructs. Experiments were performed at minimum in duplicate.
Analysis of gH surface expression was performed as previously described [27]. Briefly, CHO-K1 Cre cells were transfected with equimolar amounts of pcDNA3.1-gL and pME18s-gH. Cells were fixed at 24 hours post transfection, stained with anti-gH SG3 antibody, and detected with anti-mouse Alexa Fluor 488 antibody. Analysis was performed on a FACSCalibur controlled by CellQuest Pro (BD Bioscience) and gH expression was quantified with FlowJo (Tree Star). Experiments were performed at minimum in triplicate.
Confocal microscopy of melanoma cells transfected with pcDNA3.1-gL and either pME18s-gH[WT] or gH mutant constructs was performed as previously described [27]. For infection, melanoma cells were inoculated with 500 plaque forming units (PFU) of recombinant virus. For TK-GFP-BAC transfection, melanoma cells were transfected with pOka-TK-GFP-gB[Y881F], -gH[Δ834-841], and -gB[Y881F]/gH[Δ834-841] BACs using Lipofectamine 2000. Cells were fixed with 4% paraformaldehyde at post 24 hours for non-BAC transfected and infected cells and 72 hours for BAC transfected cells. Cells were probed for gH, early endosomes and the trans-Golgi-network using anti-gH mAb SG3 (mouse), anti-EEA1 mAb (rabbit; Novus Biological), and anti-TGN46 pAb (sheep; AbD Serotec), respectively. TK-GFP-BAC-transfected cells were probed with anti-IE62 (mouse; Chemicon International) and anti-ORF23 (polyclonal rabbit; [36]). Primary antibodies were detected with secondary antibodies, anti-mouse Alexa Fluor 555 (Invitrogen), anti-rabbit Alexa Fluor 488 (Invitrogen), and anti-sheep Alexa Fluor 647 (Invitrogen). Nuclei were stained with Hoechst 33342 (Invitrogen). Images were captured with a Leica SP2 AOBS Confocal Laser Scanning Microscope. Channel merging and image processing was performed with ImageJ and Photoshop.
Melanoma cells were seeded in 6-well plates and infected with 500 PFU of pOka or recombinant gH mutant viruses. Images of syncytium were taken at 24, 36, and 48 hours post infection. Prior to each time point, infected cells were incubated with fresh media supplemented with Hoechst 33342 at 1∶1000 dilution for 10 minutes. Brightfield and fluorescence microscopy images were taken with an AX10 Microscope (Zeiss) equipped with an X-Cite Series 120 fluorescence excitation light source (Lumen Dynamics). Images were aligned and processed in Photoshop to enhance contrast for the cytoplasmic regions of the syncytium. To quantify the number of nuclei per syncytium at 36 hours post infection (hpi), 15 syncytia were randomly selected and the nuclei were visually counted in a single plane from images taken by a light microscope.
As previously described [37], melanoma cells were infected with 1000 PFU of recombinant VZV and harvested at 24 hour intervals to examine replication kinetics. Viral titration was performed by 10-fold dilution on melanoma cells in triplicate. To access plaque size, stained plaques (n = 30) from titration plates at four days post infection were captured with an AX10 Microscope (Zeiss) with the plaques outlined and area calculated using ImageJ, as described [34].
To examine infected cells, 12 mm glass coverslips were seeded with 2×105 melanoma cells and infected with 1000 PFU of pOka-gH[Δ834-841] or pOka virus. To examine cells transfected with pOka-TK-GFP-gB[Y881F], -gH[Δ834-841], and -gB[Y881F]/gH[Δ834-841] BACs, 100 mm dishes seeded with 1.2×106 melanoma cells were transfected with 30 µg of BAC DNA using Lipofectamine 2000. At 24 hours post transfection, cells were harvested and sorted for GFP positive cells using a BD InFlux Special Order sorter for enrichment. Sorted cells were then added to a cloning ring placed upon a 12 mm glass coverslip preseeded with 4×105 melanoma cells. At 48 hours post reseeding for transfected cells and 72 hpi for infected cells, samples were fixed with 2% glutaraldehyde and 4% p-formaldehyde in 0.1M sodium cacodylate buffer pH 7.2 for 20 minutes. Cells were then treated with 1% osmium (OsO4) for one hour and stained with 1% uranyl acetate overnight. Cells were dehydrated with multiple incubations of increasing concentrations of ethanol (50%, 70%, 95%, and 100%) followed by acetonitrile. Epon infiltration was performed by subsequent incubations with 1∶1 (Epon/Acetonitrile), 2∶1 (Epon/Acetonitrile), and 100% Epon. Samples were incubated at 65°C to allow for polymerization. Glass coverslips were dissolved by incubation in 49% hydrofluoric acid for 20 minutes. Areas of interest were trimmed. Ultrathin sections were prepared using an ultratome (Leica Microsystems) and placed upon formvar carbon film 100Mesh copper grids (Electron Microscopy Sciences). Grids were stained with 4% uranyl acetate and 0.2% lead citrate. Sections were visualized using a JEOL 1400 transmission electron microscope at 80 kV and digital photographs were captured with a Gatan Multiscan 701 digital camera.
Skin xenografts were prepared in homozygous CB-17 scid/scid mice using human fetal skin tissue obtained according to federal and state regulations as previously described [12]. Briefly, implanted skin xenografts were inoculated with HELFs infected with wildtype pOka or mutant viruses. The inoculum titer for each virus was determined to confirm similar PFU/mL concentrations. Xenografts were harvested at 10 and 21 days post infection (dpi) and viral titers were determined. Confocal microscopy of sectioned infected skin xenografts was performed as previously described [34]. Briefly, 5 µm sections were deparaffinized and treated with citrate-based-Antigen Unmasking Solution (Vector) at high temperature following the manufacturer's instructions. Sections were then probed with primary antibodies, anti-gE Mab8612 (mouse; EMD Millipore), anti-ORF23 (polyclonal rabbit; [36]), and anti-TGN46 (sheep; AbD Serotec). Primary antibodies were detected with anti-mouse Alexa Fluor 555, anti-rabbit Alexa Fluor 488, and anti-sheep Alexa Fluor 647. Nuclei were stained with Hoechst 33342. Images were aligned and processed with Photoshop.
Amino acid sequences of gH homologues were obtained from protein searches on PubMed. Sequences were submitted to TOPCONS (http://topcons.cbr.su.se/) to identify transmembrane and cytoplasmic domains [38]. Hydrophobic residues in the predicted cytoplasmic domains were manually identified.
All quantitative results were analyzed by either one-way or two-way ANOVA to determine statistical significance using Prism (Graphpad Software).
Expression of VZV gB, gL, and gH with amino acids 834-841 of its cytoplasmic domain absent (gH[TL]) because of a E834Stop mutation has been demonstrated to produce a high frequency of cell fusion events by the quantitative Cre reporter assay [27], [34]. This assay allows for the identification of domains that are required for fusion regulation based on enhanced fusogenicity (hyperfusogenicity) conferred by the mutations [34]. In contrast to gH[TL], the fusion frequencies for wildtype gH (gH[WT]) were only marginally greater than the vector alone indicating that the full length gH protein limits fusion under these conditions (Figure 1A and 2). Co-expression of gH[TL] with only gL did not induce fusion, further confirming that gB was a critical component of the cell fusion machinery. Expression levels and the molecular weights of gH[TL] were similar to gH[WT] in CHO cells, based on immunoprecipitation and western blots (Figure 1B), demonstrating that amino acids 834-841 in the gHcyt do not affect the synthesis, stability, or the post-translational processing of gH into its 118 or 130 kDa mature form in the transfected CHO cells used to evaluate cell fusion.
To determine whether the gHcyt was critical for cell fusion regulation in combination with the gBcyt, the fusogenicity of the single substitution, gB[Y881F], and double substitution, gB[Y881/920F], were examined with gH[TL] and gH[WT] (Figure 2). When co-expressed with gH[TL] and gL, the gB[Y881F] and gB[Y881/920F] mutants induced levels of fusion at 60% and 137% greater than wildtype gB (gB[WT]), respectively, which was consistent with our previous report [34]. In contrast, both gBcyt mutants had fusion levels that were limited or below the level of detection when they were co-expressed with gH[WT]. Thus, the hyperfusion phenotype of gB[Y881F] and gB[Y881/920F] is dependent upon the absence of amino acids 834-841 of gH, suggesting that the cytoplasmic domains of gB and gH might function together in regulating cell fusion.
To further characterize regulation of cell fusion by the gHcyt, mutants were constructed that replaced the eight amino acids of 834-841 with either the 14 amino acid V5 epitope (gH[V5]) or the 10 amino acid cMyc epitope (gH[cMyc]) (Figure 1A). To confirm that there was no read through of the E834Stop nonsense mutation or disruption of the expression plasmid by the V5 sequence, a mutant that contained both the E834Stop and V5 epitope (gH[834StopV5]) was also constructed. All three gH mutants had protein expression profiles similar to gH[WT] in transfected CHO cells (Figure 1B). However, despite the absence of amino acids 834-841, the V5 and cMyc mutations resembled gH[WT] in having very limited capacity to induce cell fusion (Figure 3A). Similar to the increase in fusion levels observed by gH[TL] relative to gH[WT], gH[834StopV5] also had increased fusion in comparison to its corresponding control, the gH[V5]. Although a slight increase in fusion levels was observed between gH[834StopV5] and gH[TL], possibly because of the V5 sequence, the enhanced fusion levels caused by the gH[834StopV5] mutation demonstrated that read through past the 834Stop codon, like the gH[TL] mutation, did not occur. This strongly suggests that the gHcyt length is important for regulating cell fusion.
To further determine the importance of the physical length of the gHcyt rather than its sequence for regulating cell fusion, a series of truncation mutants was constructed: gH[Δ834-841], gH[Δ836-841], gH[Δ838-841], and gH[Δ840-841], containing eight, six, four, and two amino acid deletions from the carboxyl terminus of the gHcyt, respectively (Figure 1A). None of the truncation mutations changed the protein expression profile of gH in CHO cells (Figure 1B). The incremental deletions of the gHcyt caused stepwise increases in gB/gH-gL mediated cell fusion rather than an all-or-none increase in fusion levels that would be indicative of deletion of a specific functional motif (Figure 3A). As expected, the gH[Δ834-841] mutant induced levels of cell fusion similar to gH[TL], confirming again the importance of amino acids 834-841 for fusion regulation. This also demonstrated that the induction of cell fusion by gH[TL] was not caused by the type of mutation, whether nonsense or deletion, that was employed to eliminate expression of amino acids 834-841. The remaining Δ836-841, Δ838-841 and Δ840-841 deletions exhibited 56%, 73%, and 99% less cell fusion when compared with gH[TL], respectively (Figure 3A). Thus, this data provided further evidence that regulation of gB/gH-gL mediated cell-cell fusion by the gHcyt depends on the length of the cytoplasmic domain and not the specific sequence or intrinsic properties of the amino acids in the domain.
The presence of hydrophobic residues in the cytoplasmic domain of viral fusion-related proteins has been suggested to have a role in regulating cell fusion [39]. The predicted gHcyt of alphaherpesvirus homologues, including VZV, contain highly hydrophobic residues (leucine, isoleucine, and, phenylalanine) (Figure S1), suggesting that the presence of hydrophobic residues in the gHcyt might be a regulatory factor for cell fusion. To test this hypothesis, two gH mutants were constructed, gH[834(IV)4] and gH[834(QN)4], in which amino acids 834-841 were substituted with a series of alternating isoleucine and valine residues or glutamine and arginine residues, respectively (Figure 1A). The 834(IV)4 substitution was highly hydrophobic, while the 834(QN)4 was highly hydrophilic (Figure S2A). Similar to gH[WT], both mutations induced little or no detectable fusion (Figure 3A), which demonstrated that fusion regulation by the gHcyt was not dependent upon the presence of hydrophobic residues and further confirmed that the regulation was motif independent.
Given that expression of gH on the surface of the cell would be expected to influence cell fusion, FACS analysis with anti-gH antibody was performed on CHO cells expressing the gH[WT] or gH mutants along with gL (Figure S3) (Figure 3B). gH[TL] and gH[WT] protein were not detected on the cell surface in the absence of gL, which was consistent with previous findings [27]. Surface expression levels of the gH mutants with gL could be separated into three distinct groups, with mutants having significantly greater, equal, or reduced surface expression compared to gH[WT]. gH[TL], gH[834StopV5], gH[Δ834-841], and gH[Δ836-841] exhibited surface expression levels that were 12%, 20%, 24%, and 19% greater than gH[WT], respectively. gH[V5] and gH[834(QN)4] mutants had surface expression levels equivalent to gH[WT], while gH[cMyc], gH[Δ840-841], and gH[834(IV)4] had surface expression levels that were 15%, 30%, and 10% less than gH[WT] (Figure 3B). To determine if there was a correlation between the surface expression and cell fusion effects of these mutants, the mean levels of their surface expression were plotted against their cell fusion levels on a scatter plot along with gH[WT] and gH[TL]. For both the greater and lesser surface expression groups, no significant correlation was found between their surface expression and cell fusion levels. A Pearson R value of 0.53 and R2 value of 0.28 was determined for the mutants with increased surface expression relative to gH[WT] with a P value of 0.29 (Figure 3C) and while a high R value of 0.713 and R2 value of 0.54 was calculated for those mutants with reduced surface expression, the P value was 0.15, making the correlation not statistically significant (Figure 3D). Thus, while gB-gH/gL mediated cell fusion requires the presence of gH on the surface of the cell, greater levels of surface expression did not correlate with the induction of cell fusion and the induction or extent of fusion was not a result of increased gH surface expression.
To determine if the changes in cell fusion associated with the gHcyt mutations might be due to altered intracellular trafficking, the subcellular localization of gH mutants in melanoma cells co-expressing gL was examined by confocal microscopy. As expected from previous reports, gH[WT] had a pattern of juxta-nuclear localization as well as a punctate localization in the cytoplasm [27] (Figure 4). Co-staining with organelle markers confirmed the co-localization of gH with the trans-Golgi network (anti-TGN-46) at the juxta-nuclear position and with some early endosomes, detected with anti-EEA1 antibody, appearing as punctae in the cytoplasm. gH[WT]-like subcellular localization was also observed for the gH[V5], gH[cMyc], gH[834StopV5], gH[Δ834-841], gH[Δ836-841, gH[Δ838-841], and gH[Δ840-841] mutants (Figure 4) and for gH[834(IV)4] and gH[834(QN)4] (Figure S2B). Thus, amino acids 834-841 of the gHcyt did not have a functional role in gH intracellular trafficking or its internalization from the cell surface.
The role of the 835YNKI838 motif was examined in the virus-free transfection system by constructing two mutants with the tyrosine (Y835) residue substituted with either alanine (Y835A) or phenylalanine (Y835F). The alanine substitution fully disrupted the residue, while the phenylalanine substitution served to determine if the hydroxyl group had a functional role in endocytosis (Figure 1A). Neither substitution affected levels of gH protein expression or processing compared to gH[WT] (Figure 1B). Similar to gH[WT], both gH[Y835A] and gH[Y835F] were able to colocalize with EEA-1 markers, indicating that the protein was present in endocytic vesicles (Figure 4, white arrows). Thus, gH endocytosis was not prevented by disrupting the YXXΦ motif.
Similar to gH[WT], both the Y835A and Y835F substitutions induced little to no detectable fusion in the Cre reporter fusion assay (Figure 3A). Cell surface expression levels were different between the mutants, with gH[Y835A] having levels of surface expression greater than gH[WT] and gH[Y835F] having levels similar to the gH[WT] (Figure 3B). This further confirmed the lack of correlation between the induction of cell fusion observed in the Cre reporter fusion assay and increased gH surface expression. These results demonstrated the 835YNKI838 endocytosis motif in the gHcyt did not contribute to regulation of gB/gH-gL mediated cell fusion in the absence of other VZV proteins and was consistent with evidence that the regulation of gB/gH-gL mediated cell fusion depended on the physical length of the gHcyt and not the presence of specific residues or motifs.
To determine the role of the gHcyt in VZV-induced syncytia formation and replication kinetics, six recombinant viruses were generated by inserting the mutations, Y835A, Y835F, V5, 834StopV5, Δ834-841, and TL into a parental Oka (pOka) BAC [28]. Infectious virus was recovered from all of the BACs with mutant gH. To determine if amino acids 834-841 of gH were critical for syncytia formation regulation, syncytium morphology was examined in VZV infected-melanoma cells at 24, 36, and 48 hours post infection (hpi). Mutant viruses expressing gH without amino acids 834-841, including pOka-gH[TL] and pOka-gH[Δ834-841], produced syncytia with a larger number of nuclei as early as 24 hpi, while few to no syncytia were observed at the same time point in cells infected with pOka (Figure 5). At 36 hpi, pOka-gH[TL], pOka-gH[Δ834-841], and pOka all exhibited syncytia formation, but the morphology of the pOka-gH[TL] and pOka-gH[Δ834-841] syncytia was significantly different compared to pOka. Melanoma cells infected with either of the gH mutants exhibited not only an observable increase in the number of nuclei per syncytium, but also had extended cytoplasm in the syncytia (Figure 5, white arrows). When quantifying the number of nuclei at 36 hpi, pOka-gH[Δ834-841] syncytium had 122±38.5 nuclei compared to 40±16.8 nuclei for pOka syncytium (Figure S4). At 48 hpi, there was a further increase in number of nuclei and enlargement of the extended cytoplasm within each syncytium. Syncytia formation and morphology of pOka-gH[834stopV5] was similar to the viruses with the TL and Δ834-841 mutations (data not shown). The syncytia morphology in melanoma cells infected with pOka-gH[V5], pOka-gH[Y835A], and pOka-gH[Y835F] viruses was similar to pOka at 36 and 48 hpi, indicating that these mutations had no effect on syncytia formation or morphology (Figure S5). The number of nuclei in pOka-gH[Y835A] induced syncytium at 36 hpi was 36±13.8, which was comparable to pOka (Figure S4). Thus, the truncation of the gHcyt caused an exaggerated syncytia formation phenotype in VZV infected melanoma cells, correlating with the increase in fusion levels observed in the Cre reporter fusion assay.
To determine if the gHcyt truncation inhibited propagation of the virus, replication kinetics of the gHcyt mutant viruses and pOka were compared in melanoma cells. Viruses with deletions that induced high levels of fusion in the Cre reporter fusion assay (834StopV5, Δ834-841, and TL) and exaggerated syncytia formation had a statistically significant 0.5–1 log10 decrease in viral titers between two and five days post infection (dpi) (Figure 6A), indicating that amino acids 834-841 of the gHcyt were important for viral propagation. In contrast, titers of viruses with mutations in the gHcyt that failed to induce fusion or altered syncytia formation in vitro (Y835A, Y835F, and V5) were similar to pOka (Figure 6B). The decrease in titer of pOka observed at six dpi was a result of a >95% infection of the monolayer. Thus, effective viral propagation is dependent upon canonical regulation of syncytia formation by the gHcyt.
To elucidate the cause of the reduction in infectious virus production, the plaque size, viral protein expression, and gH intracellular trafficking of the gHcyt mutant viruses were examined. The plaque sizes for pOka-gH[834StopV5], pOka-gH[Δ834-841], and pOka-gH[TL] viruses were similar to pOka and the other mutants (Y835A, Y835F, and V5) at four dpi (Figure 6C and D), indicating that deletion or substitution of amino acids 834-841 of the gHcyt did not inhibit viral spread to its neighboring cells in vitro but caused a premature fusion of infected cells. The maturation of gH during infection was not inhibited by either deletions or substitutions of the gHcyt as demonstrated by detection of both the 118 and 130 kDa molecular weight forms like pOka (Figure 7) indicating that the dysregulation of syncytia formation was not related to altered gH maturation. In contrast to transfected CHO cells, the levels of expression of the 130 kDa molecular weight form of gH[WT] and gH mutants during infection of melanoma cells was less than the 118 kDa form (Figure 1B). This difference in expression of the 130 kDa form did not noticeably affect fusion, since the exaggerated syncytia formation phenotype of pOka-gH[834StopV5], pOka-gH[Δ834-841], and pOka-gH[TL] was consistent with the data of the Cre reporter assay. Differences in glycosylation in mammalian cells lines [40] or posttranslational modification during infection that could affect gH processing or stability of the 130 kDa form did not alter the fusion phenotype. Intracellular trafficking of gH during infection of melanoma cells was not affected by the gHcyt substitutions when examined by confocal microscopy (Figure 8). Consistent with the localization of transiently expressed gH[WT] and gL, gH mutant proteins colocalized with markers for the TGN (anti-TGN46) and a subset of the early endosomes (anti-EEA1), indicating that gH was endocytosed during infection. Expression of VZV immediate early and early proteins were unaffected by the gHcyt mutations as demonstrated by pOka-like expression levels of immediate early 63 (IE63) and glycoprotein E, an early gene product, in western blots from infected cell lysates (Figure 7). Because changes were not detected in viral spread, gH protein expression and trafficking or synthesis of essential viral proteins, the reduction in infectious virus production by pOka-gH [TL], -gH[Δ834-841], and -gH[834StopV5] mutants was attributed to the exaggerated syncytia formation.
To determine if viral particle formation or egress was affected by the deletion of amino acids 834-841, electron micrographs were taken of pOka and pOka-gH[Δ834-841] infected melanoma cells at 48 hpi. Typical virus particles were observed on the surface of pOka-gH[Δ834-841]-infected cells (Figure 9G and J) and found to be similar to those observed on pOka-infected cells (Figure 9A and D) demonstrating that the deletion did not inhibit egress of virus particles. The pOka and pOka-gH[Δ834-841] viral particles had similar size and morphology (Figure 9B and H) indicating that the deletion did not affect particle formation. Nucleocapsids of pOka-gH[Δ834-841] formed crystalline arrays (Figure 9E) like pOka (Figure 9K) and no accumulation of nucleocapsids were observed in the perinuclear space, indicating that the gHcyt deletion did not inhibit nuclear egress (Figure 9C and I). pOka and pOka-gH[Δ834-841] particles were also found in cytoplasmic vesicles (Figure 9F and L).
To examine if removal of the fusion regulatory elements of both the gBcyt and gHcyt further exacerbated syncytia formation during infection, a BAC with amino acids 834-841 of gH deleted and the Y881 in the gBcyt ITIM substituted with phenylalanine was constructed using a pOka BAC that expressed ORF36, the thymidine kinase, tagged with EGFP at the C-terminus (pOka-TK-GFP). Infectious virus was not produced in melanoma cells transfected with the double mutant BAC, pOka-TK-GFP-gB[Y881F]/gH[Δ834-841]. In contrast, infectious virus was obtained from pOka-TK-GFP-gH[Δ834-841] and -gB[Y881F] BACs and melanoma cells infected with these viruses exhibited exaggerated syncytia phenotypes that were similar to their non TK-GFP counterparts. While infectious virus was not recovered from pOka-TK-GFP- gB[Y881F]/gH[Δ834-841], exaggerated syncytia formation occurred in BAC-transfected melanoma cells at 72 hours post transfection (Figure 10). To examine if the gB[Y881F]/gH[Δ834-841] mutations modified localization of immediate early proteins or capsid proteins, cells were stained for IE62 and capsid protein, ORF23, respectively. While accumulated nuclear localization of IE62 was observed for melanoma cells transfected with pOka-TK-GFP BAC and all three mutants, only pOka-TK-GFP, pOka-TK-GFP-gH[Δ834-841], and pOka-TK-GFP-gB[Y881F] had cytoplasmic localization of IE62, suggesting that the gB[Y881F]/gH[Δ834-841] mutations prevented the cell from entering the late stages of VZV infection. In addition, nuclei that had punctate IE62 expression were also observed in melanoma cells transfected with pOka-TK-GFP, pOka-TK-GFP-gH[Δ834-841], and pOka-TK-GFP- gB[Y881F] BACs consistent with early viral replication (Figure 10, white arrows). This pattern was not seen with melanoma cells transfected with pOka-TK-GFP-gB[Y881F]/gH[Δ834-841] BAC, indicating that the exaggerated syncytia formation either limited viral replication or spread. Capsid protein, ORF23, was detected in nuclei after transfection of all three mutants and pOka-TK-GFP, although levels were slightly reduced in cells transfected with pOka-TK-GFP- gB[Y881F]/gH[Δ834-841]. Syncytia formation was also more apparent at 72 hours post transfection, with pOka-TK-GFP-gB[Y881F] and -gB[Y881F]/gH[Δ834-841] having a rosette-like clustering of nuclei in the center of the syncytia compared to pOka-TK-GFP. In contrast, both pOka-TK-GFP-gH[Δ834-841] and pOka-TK-GFP exhibited smaller syncytia at 72 hours post transfection, suggesting that the exaggerated syncytia phenotype associated with pOka-TK-GFP-gH[Δ834-841] occurred later in infection while that induced by pOka-TK-GFP-gB[Y881F] either occurred earlier or was further enhanced (Figure 10).
To determine if the pOka-TK-GFP-gB[Y881F]/gH[Δ834-841] BAC was able to generate virus particles, melanoma cells transfected with the mutant BAC were examined at 72 hours post transfection by electron microscopy (Figure 11). Consistent with the confocal micrographs, syncytia were observed in melanoma cells transfected with pOka-TK-GFP, pOka-TK-GFP-gH[Δ834-841], pOka-TK-GFP-gB[Y881F], and pOka-TK-GFP-gB[Y881F]/gH[Δ834-841] BACs. In contrast to pOka, gH[Δ834-841], and gB[Y881F], nucleocapsids and virus particles were not observed in cells transfected with the pOka-TK-GFP-gB[Y881F]/gH[Δ834-841] BAC. This correlated with the reduced ORF23 expression and limited IE62 localization observed in the confocal micrographs. The failure of the pOka-TK-GFP-gB[Y881F]/gH[Δ834-841] BAC to generate viral particles while still being able to induce syncytia formation suggested that VZV-induced cell-cell fusion is the result of viral protein synthesis but does not require a full viral replication cycle. Furthermore, these data indicate that loss of both the gBcyt and gHcyt fusion regulatory elements generates an intracellular environment that limits viral production and spread.
To determine effects of the gHcyt mutations and dysregulation of syncytia formation on VZV pathogenesis in vivo, human skin xenografts were infected with pOka, pOka-gH[Δ834-841], pOka-gH[Y835A], pOka-gH[Y835F], pOka-gH[V5], and pOka-gH[TL] viruses at similar inoculum titers (Figure 12A). At 10 dpi, viral titers from skin xenografts infected with pOka-gH[Y835A] and pOka-gH[Y835F] mutants were not significantly different from pOka, indicating that the 835YNKI838 motif had no role in skin pathogenesis. The pOka-gH[V5] virus also replicated at similar levels to pOka. In contrast, infectious virus was not recovered from any of the six xenografts infected with pOka-gH[TL] or from five of the six pOka-gH[Δ834-841]-infected xenografts; the titer of the sixth implant was <1 log10 (Figure 12B). Similar results were observed in xenografts recovered at 21 dpi after inoculation of pOka-gH[Y835A], pOka-gH[Y835F], and pOka-gH[V5] viruses, which had titers that were again comparable to pOka. pOka-gH[TL] virus was not isolated from five of the six xenografts with the sixth having a titer of <1 log10. The pOka-gH[Δ834-841] infected implants showed some limited levels of viral replication with four of the six implants having a mean titer of 1.5 log10, approximately 1.3 log10 lower than the mean titers of pOka-infected xenografts. Thus, the gHcyt played a critical role in VZV skin pathogenesis and was dependent on amino acids 834-841 of gH.
To further examine VZV skin pathogenesis, confocal microscopy was performed on sections from xenografts at 21 dpi that yielded the highest viral titers. Infected cells were identified by ORF23 and gE expression, with the skin layers visualized by staining for TGN46 and nuclei (Figure 13). As previously reported [34], the lesion in the pOka-infected xenograft showed penetration and spread of the virus across the epidermal and basal layer into the dermal layer (Figure 13, top left). Infection was also visible in hair follicles and syncytia were identified scattered throughout the lesion (Figure 13, A). Similar to pOka, the single pOka-gH[Δ834-841]-infected xenograft that had a titer of 2 log10 also showed penetration and spread through all three layers of the skin (Figure 13, B). In addition, differences in syncytia morphology and size were not apparent between pOka and pOka-gH[Δ834-841]. Lesions from xenografts infected with pOka-gH[Y835A], pOka-gH[Y835F], and pOka[gH-V5] viruses as well as ORF23 and gE staining were all morphologically similar to pOka. Lesions were not identified in the single xenograft from which pOka-gH[TL] was recovered, which correlated with the extremely low viral titer (Figure 12B).
Syncytia formation is an inherent characteristic of VZV replication in vitro and in vivo. In this study, amino acids 834-841 of the VZV gH cytoplasmic domain were identified to be critical for the regulation of gB/gH-gL mediated cell-cell fusion with truncation of the domain causing exaggerated syncytia formation during infection. While this exaggerated syncytia formation only caused a moderate reduction of viral titers in cultured cells, skin pathogenesis was significantly impaired. This data and our prior studies show that this regulation of cell fusion is mediated by two separate mechanisms involving both the gHcyt and gBcyt, with loss of either resulting in limited viral replication and spread in vivo. Eliminating both blocks VZV replication completely. Thus, the gHcyt, along with the gBcyt, plays a critical role in skin pathogenesis through their shared capacity to maintain regulation of VZV-induced cell fusion.
This study elucidates the functional domains of gH and gB important for cell fusion and demonstrates that the regulation of syncytia formation is necessary to support effective skin pathogenesis. In the case of VZV, the current model of gB acting as the primary fusogen is supported by the requirement for gB as demonstrated by the inability of gH[TL] and gL in the absence of gB to induce the high levels of fusion observed with gB/gH[TL]-gL. Regulation of cell fusion by gH was determined to depend on the length of its cytoplasmic domain with truncations of the domain causing increased cell fusion, indicative of a loss in regulation. Deletions and substitutions of amino acids 834-841 of the gHcyt confirmed that the regulation was not dependent upon a unique motif or specific biochemical properties such as hydrophobicity. Given that direct protein interactions commonly depend upon the presence of a specific motif or domain, it is unlikely that the gHcyt functions by a direct interaction with either a host or viral protein [41]. However, while it might not interact directly, the gHcyt is proposed to operate as a gate keeper using its physical length to control access of functional domains within neighboring proteins.
The gB/gH-gL mediated cell-cell fusion is also regulated by the gBcyt ITIM which has the potential to be phosphorylated at Y881 [34]. VZV mutagenesis to substitute the tyrosine residue with a phenylalanine (gB[Y881F]), which prevents phosphorylation of the residue, resulted in exaggerated syncytia formation in melanoma cells and significantly reduced skin pathogenesis. In contrast to infected cells, the inability of the gB[Y881F] mutation to induce high levels of fusion when expressed with gH[WT] in the Cre reporter assay strongly supports the notion that the full length gHcyt has complex fusion regulatory properties that are apparent during VZV infection. While a direct interaction of gB and gH-gL has not been confirmed, cell fusion requires both to be present on the cell surface and it is also probable that these two transmembrane glycoproteins, and thus their cytoplasmic domains, are located within close proximity on cell membranes. As a gate keeper, the gHcyt could function as a regulator of phosphorylation of the ITIM of the gBcyt by controlling either kinase or phosphatase access to Y881 residue. Therefore, the absence of the gHcyt gate keeper function in the gH[TL] enables gB[Y881F] to readily enter a dysregulated state and enhance fusion. This dysregulated state was detrimental to the virus as melanoma cells transfected with pOka BAC containing the dual gB[Y881F]/gH[Δ834-841] mutation were unable to effectively establish a progressive infection, while retaining the capacity to induce exaggerated syncytia formation. These data confirm that losing syncytia regulation significantly limits viral replication and spread. The gBcyt might also contribute to regulation of the early steps of cell fusion, as demonstrated by the LL871/872AA substitution in the HSV-1 gBcyt initiating fusion significantly more rapidly than the wildtype, while having surface expression levels that were similar to the wildtype [42]. The VZV gBcyt also contains a similar dileucine motif (904LL905). Further investigation is needed to determine if the motif has a similar functional regulatory role in fusion initiation. Together, these data establish that VZV gB and gH participate in a complex regulatory system to limit syncytia formation.
Regulation of fusion between the virion envelope and cell membrane during entry has been postulated to be different from cell-cell fusion [43]. Our observations support this concept because while truncating the gHcyt increased cell fusion when transfected with gB and gL in vitro and caused exaggerated syncytia formation when introduced into the virus, the morphology of the viral particles and their presence on cell surfaces was similar to pOka. Since regulation of cell-cell fusion also depends upon phosphorylation of the gBcyt ITIM, identical regulation for both virion envelope and cell fusion would require the presence of kinases and phosphatases with the capacity to alter tyrosine phosphorylation within the virion as well as in infected cells. While mass spectrometry studies have not been performed on extracellular VZV virions, studies of HSV-1 virions have not shown incorporation of phosphatases or tyrosine kinases [44]. It has also been noted that the lipid and protein composition of the viral envelope membrane would be expected to be different from the cell membrane, which could also change the regulation of fusion [45]. While further studies are necessary, our findings support differences in how VZV gH and gB contribute to the regulation of cell fusion and virion entry.
In contrast to VZV, syncytia formation is not a canonical characteristic of HSV-1 infection which suggests that the fusion machinery for HSV-1 controls cell fusion more tightly than VZV. While syncytial (syn) strains of HSV-1 have been identified with mutations in either gB (UL27), gK (UL53), UL20, or UL24 [46], [47], HSV syn viruses with mutations in the HSV gHcyt have not been reported. Regulation of syncytia formation by the HSV gHcyt has been demonstrated by the reduction of the syncytial phenotype in cells either infected with the HSV-1 gB[V855A] syn mutant virus when the gHcyt was truncated by nine amino acids or a gH null/gB[V855A] syn mutant virus was trans-complemented with a plasmid expressing the truncated gH [43], [48], [49]. This was also observed with in vitro fusion studies as expression of the syncytial gB[A885V] mutant with the truncated gH reduced its hyperfusion levels from 200% to 150% of wildtype fusion levels [50]. These findings suggest the V855A substitution in the HSV-1 gBcyt inhibited the regulatory control of the gBcyt over cell fusion that was then rescued by the truncation of the HSV gHcyt. This supports the model that cell fusion is jointly regulated by the gBcyt and gHcyt. In vitro cell fusion studies involving transient expression of HSV-1 glycoproteins have provided further evidence of the role of the gHcyt in fusion and its regulation. Cell fusion was inhibited when the HSV-1 gHcyt was either completely deleted and replaced with analogous domains from other glycoproteins or when both the gH transmembrane and cytoplasmic domains were substituted with a glycosylphosphatidylinositol (gpi)-addition and expressed with gB, gD, and gL [51][52]. Similar to VZV gH, truncation of the final six amino acids of HSV-1 gHcyt increased levels of cell fusion by 50% relative to the wildtype gH when transfected with gB[WT], gD, and gL [50]. Whether inserting this gH mutation into the viral genome would produce a HSV syn mutant is not known. Also consistent with our observations about the VZV gHcyt, the length of the HSV-1 gHcyt might be important for its contribution to regulation of cell fusion [53]. An insertion of five amino acids at 824L, increasing the length of the gHcyt from 14 to 20 amino acids, inhibited cell fusion in the presence of gB[WT]. Nevertheless, in the context of infection, VZV and HSV regulation of cell fusion and syncytia formation differ since truncation of VZV gH caused exaggerated syncytia formation, while HSV gH truncation reduced syncytia formation by the gB syn mutant. Thus, the VZV gB/gH-gL core fusion machinery has characteristics that change its fusogenicity capacity compared to HSV.
While the VZV gHcyt has been predicted to be 18 amino acids in length, the HSV gHcyt is only 14 amino acids long and no motifs or sequences are conserved between the two homologues. Furthermore, induction of cell fusion by HSV in vitro requires not only gB/gH-gL, but also gD. These differences in sequence length and requirements for in vitro cell fusion make it challenging to compare the processes of HSV-1 and VZV cell fusion. However, it is apparent that the gHcyt domains of both viruses perform a significant function in cell fusion regulation which is linked to the regulation of their gB counterparts.
The gH 835YNKI838 motif does not mediate gB/gH-gL cell-cell fusion in vitro, virus-induced syncytia formation, and spread in cultured cells or skin pathogenesis. Disruption of the motif also did not alter gH expression or trafficking in melanoma cells that were transiently expressing gH-gL or infected with recombinant mutant VZV. These results are in contrast to an earlier report which indicated that the 835YNKI838 motif was a functional endocytosis motif based on effects of an identical tyrosine to alanine substitution examined by confocal microscopy and using a biotin endocytosis assay. While confocal microscopy was used to examine endocytosis in both studies, the first observations were based upon infecting HeLa cells with recombinant vaccinia virus (VV) expressing the T7 polymerase and co-transfecting with pTM1 plasmids, a T7-driven vector, expressing gH and gL [29]. Given that VV can induce significant changes in the cytoskeletal network of host cells [54] and the cell proteome [55], the modified intracellular environment and the additional VV proteins might have contributed to the inhibition of endocytosis of gH[Y835A]. In addition, we have shown that gH-gL co-expression to be insufficient for cell-cell fusion [27], [34] in contrast to a previous report [56]. HeLa cells infected with VV expressing gH with the Y835A substitution and gL showed increased syncytia size, which might also be explained by VV effects [33]. Vaccinia virus expresses its own acid induced membrane fusion protein on the surface of the infected cell [57] and regulates actin polymerization in the host cell, which might prime the cell to be more fusogenic. Our approach to examine the 835YNKI838 motif more accurately models the function and activity of gH since the virus-free transfection system does not have additional molecules that could potentially affect gH trafficking. Furthermore, gH was also examined under the context of infection, which allows for usual expression of other VZV proteins that could facilitate gH trafficking in infected cells by mechanisms other than the 835YNKI838 motif.
The truncation of the VZV gHcyt increased syncytia formation and significantly reduced skin pathogenesis. A similar phenotype was observed with the gBcyt ITIM mutation indicating that exaggerated syncytia formation limits the ability of VZV to spread effectively in the skin. This demonstrates a direct link between skin pathogenesis and cell-cell fusion. The epidermis consists of three main layers, with the granular layer facing the surface, followed by the spinous layer and the inner basal layer. The epidermis is constantly in flux due to continued cell shedding that is quickly replaced by differentiated keratinocytes [58]. Exaggerated syncytia formation is proposed to limit the ability of the virus to effectively penetrate the skin, causing the lesion to remain near the surface (Figure 14A). If the rate of penetration is less than the rate of differentiation and migration of keratinocytes from the basal layer, then the epidermis would drive the infected cells out through the granular and spinous layers, preventing spread to deeper layers of the spinous and basal layer. When syncytia formation is regulated appropriately as in pOka-infected tissue, the virus is able to effectively spread and penetrate the skin creating the typical cutaneous lesions (Figure 14B). While this model requires further study, the present study demonstrates that tight regulation of syncytia formation through complementary functions of the cytoplasmic domains of gB and gH is critical for VZV pathogenesis in skin.
|
10.1371/journal.pmed.1002033 | Inter-pregnancy Weight Change and Risks of Severe Birth-Asphyxia-Related Outcomes in Singleton Infants Born at Term: A Nationwide Swedish Cohort Study | Maternal overweight and obesity are associated with increased risks of birth-asphyxia-related outcomes, but the mechanisms are unclear. If a change of exposure (i.e., maternal body mass index [BMI]) over time influences risks, this would be consistent with a causal relationship between maternal BMI and offspring risks. Our objective was to investigate associations between changes in maternal BMI between consecutive pregnancies and risks of birth-asphyxia-related outcomes in the second offspring born at term.
This study was a prospective population-based cohort study that included 526,435 second-born term (≥37 wk) infants of mothers with two consecutive live singleton term births in Sweden between January 1992 and December 2012.
We estimated associations between the difference in maternal BMI between the first and second pregnancy and risks of low Apgar score (0–6) at 5 min, neonatal seizures, and meconium aspiration in the second-born offspring. Odds ratios (ORs) were adjusted for BMI at first pregnancy, maternal height, maternal age at second delivery, smoking, education, mother´s country of birth, inter-pregnancy interval, and year of second delivery. Analyses were also stratified by BMI (<25 versus ≥25 kg/m2) in the first pregnancy.
Risks of low Apgar score, neonatal seizures, and meconium aspiration increased with inter-pregnancy weight gain. Compared with offspring of mothers with stable weight (BMI change of −1 to <1 kg/m2), the adjusted OR for a low Apgar score in the offspring of mothers with a BMI change of 4 kg/m2 or more was 1.33 (95% CI 1.12–1.58). The corresponding risks for neonatal seizures and meconium aspiration were 1.42 (95% CI 1.00–2.02) and 1.78 (95% CI 1.19–2.68), respectively. The increased risk of neonatal seizures related to weight gain appeared to be restricted to mothers with BMI < 25 kg/m2 in the first pregnancy. A study limitation was the lack of data on the effects of obstetric interventions and neonatal resuscitation efforts.
Risks of birth-asphyxia-related outcomes increased with maternal weight gain between pregnancies. Preventing weight gain before and in between pregnancies may improve neonatal health.
| The increasing prevalence of overweight and obesity has epidemic proportions, also among pregnant women.
Maternal overweight and obesity increase the risks of pregnancy complications and adverse neonatal outcomes, including severe birth asphyxia disorders.
An independent association between maternal body mass index (BMI) and birth asphyxia would be further supported if maternal weight gain between pregnancies increases the risks of birth asphyxia in the following pregnancy.
Data from the nationwide Swedish Medical Birth Register were used, which included 532,858 second-born infants born at term (≥37 weeks gestation) to mothers having their first and second infants between 1992 and 2012.
We estimated the associations between inter-pregnancy weight change and the risks of a low Apgar score (0–6) at 5 min (rate 5.4/1,000), neonatal seizures (rate 1.2/1,000), and meconium aspiration (rate 0.7/1,000).
The risks of a low Apgar score consistently increased with maternal weight gain. Compared with mothers with stable weight (−1 to <1 kg/m2 change in BMI between pregnancies), infants born to mothers who gained 4 kg/m2 or more between pregnancies had a 33% to 78% increased risk of low Apgar score, neonatal seizures, and meconium aspiration.
A study limitation is that we could not investigate the potential impact of obstetric and neonatal interventions.
The risk of birth asphyxia increases with maternal weight gain between pregnancies. Consequently, preventing weight gain between pregnancies could reduce the risk of birth asphyxia and improve infant health.
| Maternal overweight and obesity during pregnancy increase the risks of severe maternal and infant complications [1–4]. In Sweden, the proportion of women with overweight and obesity (body mass index [BMI] ≥ 25 kg/m2) in early pregnancy increased from 26% in 1992 to 38% in 2010 [5]. In the US, 58% of women between 20 and 39 y of age were overweight or obese in 2011–2012 [6]. As recently stated by WHO, the prevalence of maternal obesity must be reduced in order to improve maternal, fetal, and neonatal health [7].
Maternal overweight and obesity increase the risks of severe neonatal complications, including major malformations, preterm birth, neonatal morbidities, and low Apgar score (0–6) [3,8–11]. In term non-malformed infants, low Apgar score is commonly caused by birth asphyxia [12].
We have previously demonstrated a linear relationship between maternal BMI in early pregnancy and the risks of low Apgar scores at 5 and 10 min and birth-asphyxia-related neonatal morbidity in infants born at term [8]. Furthermore, we have found that the risk of asphyxia-related infant mortality in term infants increases with the degree of maternal obesity [5]. Risks being influenced by a change of exposure (i.e., maternal BMI) over time would be consistent with a causal relationship between maternal BMI and adverse outcomes of the offspring.
In this nationwide Swedish cohort study, we examined whether changes in maternal BMI between first and second pregnancies influenced the risks of birth-asphyxia-related outcomes in the second-born offspring, including low Apgar score (0–6) at 5 min, neonatal seizures, and meconium aspiration.
Ethics approval for this study was obtained from the Research Ethics Committee at Karolinska Institutet in Stockholm, Sweden (number 2012/4:9). Informed consent was not required as all data were anonymous.
Between January 1992 and December 2012, the nationwide Swedish Medical Birth Register included 533,535 mothers with first and second live singleton term births (≥37 completed weeks). We excluded 165 mothers with no information on inter-pregnancy interval and 512 mothers with no information on mothers’ country of birth. In analyses of low Apgar score in the second infant, we excluded another 6,423 mothers (1.2%) where either the first- or the second-born infant had missing information on Apgar score at 1 or 5 min.
The Swedish Medical Birth Register started in 1973 and contains prospectively collected data on more than 98% of all births in Sweden. The quality of data is considered high [13]. Information on sociodemographic factors, maternal and infant anthropometry, and Apgar scores are collected on standardized forms used in antenatal, obstetric, and neonatal care in Sweden. Maternal and neonatal diagnoses, including pregnancy complications and neonatal morbidities, are classified by physicians according the Swedish version of the International Classification of Diseases (ICD). The ninth version (ICD-9) was used between 1992 and 1996, and the tenth version (ICD-10) thereafter. The standardized forms are forwarded to the Swedish Medical Birth Register when the mother and infant are discharged from hospital. Information on the mother’s country of birth and level of education were obtained from the Swedish Register of Total Population and the Swedish Register of Education, respectively. Individual cross-linkage of registries was possible using the personal identification number [14], a person-unique identifier assigned to all Swedish citizens at birth or at naturalization.
Maternal BMI (kg/m2) was calculated for each of the two consecutive pregnancies. At the first antenatal visit, which occurs in the first trimester in 90% of pregnancies [13], the woman´s weight is measured wearing light indoor clothes and barefoot, and data on self-reported height are recorded. Inter-pregnancy change in BMI was calculated as the difference in BMI between the second and first pregnancies. We categorized the inter-pregnancy BMI change as <−2 kg/m2 (i.e., BMI loss greater than 2 kg/m2), −2 to <−1 kg/m2, −1 to <1 kg/m2 (stable weight), 1 to <2 kg/m2, 2 to <4, and ≥4 kg/m2. A 1-kg/m2 change in BMI corresponds to change of 2.8 kg in a woman of average height (167 cm). Based on BMI in the first pregnancy, mothers were categorized as underweight (BMI < 18.5 kg/m2), normal weight (18.5–24.9 kg/m2), overweight (25–29.9 kg/m2), obese grade I (30–34.9 kg/m2), and obese grade II-III (≥35 kg/m2) [15].
The inter-pregnancy interval was calculated as the time difference between the date of birth of the first infant and estimated date of conception of the second infant (i.e., date of birth of the second infant minus gestational age + 14 d). Gestational age was estimated primarily based on early second trimester ultrasound, which is offered to all pregnant women and which 95% of women accept [16]. When dating from an ultrasonic scan was not available, gestational age was estimated using information on date of last menstrual period. Obesity-related disorders were identified in the Swedish Medical Birth Register based on ICD codes: preeclampsia—ICD-9 codes 642E–642G, ICD-10 codes O14–O15; chronic hypertension—ICD-9 codes 401–405, 642C, 642M, ICD-10 codes O10–O11; gestational diabetes—ICD-9 code 648W, ICD-10 code O244; pregestational diabetes—ICD-9 code 250, ICD-10 codes E10–E14, O241–O243. Covariates were categorized according to Table 1.
We estimated the risks of severe asphyxia-related outcomes, including low Apgar score, meconium aspiration, and neonatal seizures, at the second birth by change in BMI from the first to the second pregnancy. A low Apgar score was defined as a score between 0 and 6 points at 5 min after birth. A diagnosis of neonatal seizures was based on ICD-9 code 779.0 or ICD-10 code P90, and a diagnosis of meconium aspiration was based on ICD-9 code 770.1 or ICD-10 code P24.0.
Rates of low Apgar score, meconium aspiration, and neonatal seizures were calculated as the number of infants with these outcomes per 1,000 births. Logistic regression analyses were used to calculate odds ratios (ORs) with 95% confidence intervals for all outcomes. Mothers whose first infant had a low Apgar score, meconium aspiration, or neonatal seizures were excluded from the analysis of the respective outcome in the second pregnancy to avoid bias in case of a tendency to repeat adverse pregnancy outcomes. Inter-pregnancy weight change was categorized as presented above in the regression model, but also treated as a continuous variable. Multivariate models were restricted to second births with complete data on inter-pregnancy weight change and covariates. We adjusted for maternal BMI in the first pregnancy, maternal height, maternal age at second delivery, smoking habits in the second pregnancy, inter-pregnancy interval, mother´s education, mother’s country of birth, and year of second birth, categorized according to Table 1. Given the long study period, spanning over 21 y, we categorized year of second birth (not shown in Table 1) into intervals as 1992–1996, 1997–2001, 2002–2006, and 2007–2012.
In sensitivity analyses (S1 Table), mothers with obesity-related disorders (chronic hypertension, preeclampsia, or any type of diabetes) were excluded from the regression analyses. A sensitivity analysis was also performed to explore whether pregnancies of mothers with missing data on inter-pregnancy weight change differed from those with information on inter-pregnancy weight change (S2 Table). To explore any effect modification by first pregnancy BMI on the association between exposure and outcomes, analyses were also stratified by maternal BMI in the first pregnancy (BMI < 25 or BMI ≥ 25 kg/m2). Interaction terms were introduced in the multivariate models, and a p-value of less than 0.05 for the interaction term was considered statistically significant.
The total number of infants with a low Apgar score (0–6 points) at 5 min was 2,824 (rate 4.08/1,000). Corresponding numbers (rates) for neonatal seizures and meconium aspiration were 658 (1.23/1,000) and 372 (0.70/1,000), respectively (Table 1). Rates of all birth-asphyxia-related outcomes increased with maternal BMI in the first pregnancy, inter-pregnancy interval, smoking, chronic hypertension, preeclampsia, any type of diabetes, and generally also with maternal age at second delivery (Table 1). Maternal education was inversely correlated with all outcomes. There was a U-shaped relationship between birth weight and gestational age and rates of low Apgar score and neonatal seizures, while the rate of meconium aspiration increased with gestational age. Rates of low Apgar score, neonatal seizures, and meconium aspiration in second-born offspring were essentially similar in offspring of mothers with missing data on BMI in the first pregnancy and the total population.
Rates of low Apgar score, neonatal seizures, and meconium aspiration increased with inter-pregnancy weight gain (Table 2). Compared with mothers with stable weight, the risk of a low Apgar score was 26% increased in offspring of mothers who gained 2 to <4 kg/m2 and 33% increased in offspring of mothers who gained ≥4 kg/m2. Risk of neonatal seizures was increased by more than 40% in offspring of mothers who gained 2 to <4 and ≥4 kg/m2. The risk of meconium aspiration was 78% higher in offspring of mothers who gained ≥4 kg/m2 compared with offspring of mothers with stable weight.
We investigated whether the effect of weight change on asphyxia-related outcomes differed between offspring of mothers who were underweight/normal weight and overweight/obese in the first pregnancy (BMI <25 and ≥25 kg/m2, respectively) (Table 3). In underweight/normal weight mothers in the first pregnancy, the risk of low Apgar score increased with weight gain, while the corresponding association was less evident in infants of mothers who were overweight/obese in the first pregnancy. However, the test for an interaction between maternal BMI and low Apgar was not significant (p = 0.12).
The risk of neonatal seizures increased with weight gain in offspring of mothers who were underweight/normal weight in the first pregnancy, but not in offspring of mothers who were overweight/obese in the first pregnancy (test for interaction; p = 0.004). In underweight/normal weight mothers, offspring of mothers with a weight gain of ≥4 kg/m2 had a doubled risk of neonatal seizures compared with offspring of mothers with stable BMI.
Stratifying the analyses by maternal BMI in the first pregnancy (<25 and ≥25 kg/m2) demonstrated a more than doubled risk of meconium aspiration in offspring of underweight/normal weight mothers who gained ≥4 kg/m2. In contrast, in overweight/obese mothers, weight gain was not associated with increased risk of meconium aspiration in offspring. However, the test for interaction was not significant (p = 0.77).
In order to investigate whether associations between maternal weight gain and asphyxia-related outcomes were influenced by obesity-related disorders, we repeated the analyses after excluding offspring of mothers with chronic hypertension, preeclampsia, or any type of diabetes. Restricting the analyses to offspring of mothers without obesity-related disorders did not change the risks of low Apgar score, neonatal seizures, or meconium aspiration (S1 Table).
We also investigated weight gain and the risks of birth-asphyxia-related outcomes in the offspring of mothers who were underweight/normal weight (BMI below 25 kg/m2) in the second pregnancy (S2 Table). Compared with the offspring of mothers with stable weight, the risks of low Apgar score and neonatal seizures were more than doubled in offspring of mothers who were normal weight in the second pregnancy but who had gained ≥4 kg/m2 between pregnancies.
Finally, we investigated whether mothers with missing information on inter-pregnancy weight gain, and hence not included in the analyses, differed from mothers with information on weight gain with respect to other maternal characteristics or birth outcomes. The proportions of women who were overweight, obesity grade I, and obesity grade II-III in the first pregnancy were 20.5%, 5.2%, and 1.7%, respectively, for mothers with known weight change and 19.3%, 4.9%, and 1.9%, respectively, for mothers with unknown weight change (due to missing data on second pregnancy BMI) (S3 Table). Also, the distributions of maternal covariates (as listed in Table 1) were comparable between mothers with missing data on inter-pregnancy weight change and mothers with an inter-pregnancy weight change of −1 and <1 kg/m2 (the reference group) or an inter-pregnancy weight gain of 1 to <2 kg/m2. The proportions of mothers with obesity-related diseases and rates of birth-asphyxia-related outcomes were similar in mothers with and without data on inter-pregnancy weight change (S3 Table).The proportions of women who were overweight, obesity grade I, and obesity grade II-III in the second pregnancy were 25.3%, 7.6%, and 1.4%, respectively, for mothers with information on inter-pregnancy weight change and 24.2%, 7.1%, and 1.4%, respectively, for mothers with unknown weight change (due to missing data on first pregnancy BMI).
In this population-based cohort study we found that the risks of severe birth-asphyxia-related outcomes in the second offspring born at term increased with inter-pregnancy maternal weight gain. The risk increases were primarily found in offspring of mothers with BMI < 25 kg/m2 in the first pregnancy. The observed increments in risks of low Apgar score, neonatal seizures, and meconium aspiration remained essentially the same after exclusion of offspring of mothers with obesity-related diseases.
Data from epidemiological studies support that maternal BMI and changes in maternal BMI influence the risks of maternal complications, preterm delivery, and infant mortality [8–11,17–23], We have previously reported that the risks of severe birth-asphyxia-related complications increase with maternal overweight and obesity [8]. To our knowledge, this is the first study to assess whether these risks are influenced by changes of exposure (i.e., change in weight) over time.
The pathophysiology underlying the associations between maternal overweight/obesity and birth-asphyxia-related outcomes in offspring is likely to be complex and multifactorial. Obesity in pregnant women is accompanied by inflammation in maternal and placental tissues, impaired microvascular function, oxidative stress, and marked insulin resistance [24,25], changes that may contribute to the increased risk of birth asphyxia and other complications. It has also been proposed that an altered gut microbiota in obese women adversely influences maternal metabolism, which in turn may affect fetal health [26].
Fetal macrosomia is the most prevalent complication in pregnancies with maternal obesity [4]. There is a linear association between maternal BMI and fetal macrosomia and between maternal BMI and measures of hyperinsulinemia in cord blood [27], independent of maternal glucose values. Results from both experimental and clinical studies strongly suggest that fetal hyperinsulinemia is a risk factor for fetal hypoxia [28–31]. Thus, it is possible that fetal hyperinsulinemia is of pathophysiological importance for the increased risk of birth asphyxia in pregnancies with maternal obesity. Fetal macrosomia also increases the risks of traumatic delivery and shoulder dystocia, which increase the risk of birth asphyxia [32]. Obesity-related disorders, including chronic hypertension, preeclampsia, and diabetic diseases, are associated with increased risks of fetal hypoxia and low Apgar score [33,34]. However, excluding offspring of mothers with obesity-related diseases did not substantially change the weight-gain-related risks of asphyxia-related outcomes in our study.
Compared with offspring of normal weight mothers, offspring of overweight and obese mothers are at increased risks of birth-asphyxia-related neonatal outcomes [8]. However, increments in risks associated with inter-pregnancy weight gain were primarily restricted to offspring of mothers with BMI < 25 kg/m2 in the first pregnancy. The same pattern has also been demonstrated for complications during pregnancy [17]. Interestingly, in offspring of mothers with BMI < 25 kg/m2 in the second pregnancy, risks of low Apgar score and neonatal seizures were more than doubled in offspring of mothers who gained ≥4 kg/m2 between pregnancies (S2 Table). One could speculate that the less pronounced effect of weight gain in women with established overweight/obesity may reflect a metabolic adaption over time to an increased fat mass. Furthermore, women with overweight or obesity in early pregnancy accumulate less fat in pregnancy than underweight/normal weight women [25]. Thus, the same absolute increase in weight would reflect a larger relative increase in fat mass in women of underweight/normal weight compared with women who already were overweight. It is also possible that the distribution of added fat mass between pregnancies differs between lean women and overweight women, with a relatively larger increment in visceral fat in lean than in overweight women. However, rates of birth-asphyxia-related outcomes also increased with weight gain in offspring among women who were overweight or obese in the first pregnancy. Thus, the statistical power to detect risk increments could have been insufficient.
The primary strengths of our study are the population-based design, including a large number of births, and the prospectively recorded data on exposures and outcomes. The large cohort enabled us to investigate the impact of a range of inter-pregnancy weight change categories on the risks of neonatal outcomes related to birth asphyxia. The prospectively collected data limited the risks of selection and information bias. We were also able to analyze these risks stratified by maternal BMI in the first pregnancy and to adjust for several potential confounders. We used maternal BMI as a proxy for maternal fat mass. This assumption is justified as there is a strong correlation (r2 = 0.84) between BMI and fat mass in early pregnancy [35].
Some limitations of the present study should be noted. In the present study, inter-pregnancy weight change was calculated as the difference in early pregnancy BMI between the two first consecutive pregnancies. We did not have information on gestational weight gain and do not know when the weight gain occurred. The distribution of fat may differ if weight is gained during or after pregnancy. Longitudinal studies demonstrate that women with high gestational weight gain are likely to retain this weight into subsequent pregnancies [36] and that high weight gain in the first pregnancy is an important risk factor for future overweight and obesity [37].
In spite of having data on many key confounders, we cannot rule out the possibility of residual confounding by unmeasured maternal factors driving the relationship between inter-pregnancy weight change and risk of birth asphyxia. For example, women who gain weight between pregnancies may also have a less healthy life style in other aspects than women with stable weight. We also lacked specific information on obstetric interventions and neonatal resuscitation efforts. Therefore, the impact of these factors on the risk of birth asphyxia could not be investigated.
We studied neonatal conditions related to birth asphyxia. Severe birth asphyxia is commonly defined as an Apgar score of 0–3 at 5 min in combination with cord blood acidosis and neurological symptoms like neonatal seizures [38]. However, in epidemiological studies, a frequently used definition of birth asphyxia is an Apgar score of 0–6 at 5 min [39,40]. The validity of this definition is supported by the fact that an infant with an Apgar score of 4–6 at 5 min has a 45 times higher risk of neonatal death and a 31 times higher risk of cerebral palsy compared to infants with an Apgar score of 7–10 at 5 min [41,42]. We studied only infants born at term, as preterm birth itself is a common reason for low Apgar score [43]. Maternal BMI in early pregnancy also influences risk of preterm birth [10]. In addition, the possibility of selection bias should be considered given that the analyses were restricted to offspring of mothers with two children.
Information on inter-pregnancy weight change was missing in 19% of the study population. If pregnancies with information on inter-pregnancy weight change differ from pregnancies without this information, results could be biased. However, the distributions of maternal covariates, overweight, and obesity were similar in the first and second pregnancies of women with and without information on inter-pregnancy weight change, and rates of asphyxia-related outcomes were also similar.
Given the high prevalence of maternal overweight and the possible long-term consequences of birth asphyxia, our results have substantial public health relevance, as even modest weight increases in normal weight women may impact offspring outcomes on a population level. However, our finding that inter-pregnancy weight gain influences the risks of birth-asphyxia-related outcomes should be confirmed in other populations. Encouraging women to normalize BMI before pregnancy and to avoid weight gain between pregnancies is likely to be an important measure to improve infant health.
|
10.1371/journal.pgen.1006896 | MiR-1254 suppresses HO-1 expression through seed region-dependent silencing and non-seed interaction with TFAP2A transcript to attenuate NSCLC growth | MicroRNAs (miRNAs) are a class of small non-coding RNAs, which direct post-transcriptional gene silencing (PTGS) and function in a vast range of biological events including cancer development. Most miRNAs pair to the target sites through seed region near the 5’ end, leading to mRNA cleavage and/or translation repression. Here, we demonstrated a miRNA-induced dual regulation of heme oxygenase-1 (HO-1) via seed region and non-seed region, consequently inhibited tumor growth of NSCLC. We identified miR-1254 as a negative regulator inhibiting HO-1 translation by directly targeting HO-1 3’UTR via its seed region, and suppressing HO-1 transcription via non-seed region-dependent inhibition of transcriptional factor AP-2 alpha (TFAP2A), a transcriptional activator of HO-1. MiR-1254 induced cell apoptosis and cell cycle arrest in human non-small cell lung carcinoma (NSCLC) cells by inhibiting the expression of HO-1, consequently suppressed NSCLC cell growth. Consistently with the in vitro studies, mouse xenograft studies validated that miR-1254 suppressed NSCLC tumor growth in vivo. Moreover, we found that HO-1 expression was inversely correlated with miR-1254 level in human NSCLC tumor samples and cell lines. Overall, these findings identify the dual inhibition of HO-1 by miR-1254 as a novel functional mechanism of miRNA, which results in a more effective inhibition of oncogenic mRNA, and leads to a tumor suppressive effect.
| It is generally accepted that miRNAs bind to 3`UTR of target mRNAs and direct post-transcriptional gene silencing (PTGS) via its seed sequence. Here we report a new dual regulatory mechanism of miRNA. We described that miR-1254 repressed HO-1 at post-transcriptional level by directly targeting HO-1 3’UTR via its seed sequence and also inhibited HO-1 transcription by suppressing the transcriptional factor AP-2 alpha (TFAP2A) via its non-seed sequence. MiR-1254 induced cell apoptosis and cell cycle arrest in human non-small cell lung carcinoma (NSCLC) cells by inhibiting the expression of HO-1, consequently suppressed NSCLC cell growth. Moreover, in vivo mouse xenograft studies also supported the inhibitory effect of miR-1254 on NSCLC growth. These findings identify the dual regulation of miR-1254 on HO-1 as a novel functional mechanism of miRNA, which results in a more effective inhibition on the oncogenic mRNA, and leads to a suppressive effect on NSCLC growth, thus significantly advance our understanding of miRNA-directed gene regulation.
| MicroRNAs (miRNAs) are a class of small non-coding RNAs, which direct post-transcriptional gene silencing (PTGS) and play important regulatory roles in a vast range of cellular processes including cell differentiation, proliferation, apoptosis, and migration [1–3]. Aberrant expression of miRNAs may play an important role in tumorigenesis or cancer development through dysregulation of tumor-associated genes [4–9]. It is generally accepted that miRNA binds to 3’-untranslated regions (3’-UTRs) of target mRNAs via its seed sequence (position 2–8), resulting in degradation or translational repression of the target mRNA in mammalian cells.
For a majority of miRNAs, as few as 6nt of the seed sequence matching with the target mRNA is required for functional interaction. However, besides the canonical interaction between seed region of miRNA and the 3’-UTR of target mRNA, more and more evidence show that non-canonical miRNA-target sites can be functional as well [10–13]. For example, imperfect matches of miRNA seed region with the target can be compensated by supplemental components in near-perfect sites and function in target cleavage [12, 14–16]. Studies in mouse brain shows that only 73% of the Ago-mRNA interactions can be explained by seed matches for Ago-bound miRNAs, while the rest 27% have no predicted seed matches [17, 18]. Regions outside the seed sequence may also be necessary or sufficient to direct different non-canonical regulations [11]. For example, functional “centered sites” in miRNA which have 11–12 continuous Watson–Crick pairs complementary to the target mRNA were identified by analyzing microarray data [19]. Subsequent study demonstrated that 11-mer matches of miRNA “centered sites” to the target mRNA with single mismatches or GU wobbles also form hybrids but only a small proportion leads to a repression [20], which indicates additional mechanism besides sequence complementarity may also be necessary.
The most classical function of miRNA is to induce post-transcriptional gene silencing, through either mRNA cleavage and/or translational repression. The function of miRNA has been extended to transcriptional levels, either directly or indirectly. Several studies demonstrate sequence complementarity between miRNA and target gene promoter lead to gene silencing at transcriptional level [21–27]. MiR-552 is found with a dual inhibition on CYP2E1 expression, targeting both CYP2E1 promoter via its non-seed sequence and CYP2E1 mRNA 3`UTR region via its seed sequence, respectively. It consequently induces a dual inhibition of the target mRNA at both transcriptional and post-transcriptional levels, which represents a model of effective gene regulation by miRNA [28]. Here, during our study of miRNAs regulating heme oxygenase-1 (HO-1) expression, we found another type of “dual regulation” including indirect transcriptional silencing and direct post-transcriptional inhibition.
HO-1 is a rate-limiting enzyme that metabolizes heme to generate carbon monoxide (CO), ferrous iron, and biliverdin; biliverdin is subsequently reduced to bilirubin by biliverdin reductase [29, 30]. Although the physiological HO-1 expression is only found in normal liver and spleen, HO-1 is highly induced in different types of tumors, including melanoma [31], glioblastoma [32], pancreatic cancer [33], prostate cancer [34] and non-small-cell lung cancer [35]. A growing number of studies have demonstrated that HO-1 modulated tumor growth by regulating apoptosis and cell cycle, stimulating angiogenesis, and inhibiting or terminating inflammatory response [36, 37]. Here we report that miR-1254 is a miRNA which down-regulated HO-1 via two distinct mechanisms. On the one hand, miR-1254 directly targets HO-1 via its seed sequence and represses HO-1 expression at post-transcriptional level. On the other hand, miR-1254 targets transcription factor AP-2 alpha (TFAP2A) which is a transcriptional activator of HO-1, via its non-seed sequence, and consequently represses HO-1 expression at transcriptional level. MiR-1254 induces cell apoptosis and cell cycle arrest in NSCLC cells by inhibiting the expression of HO-1, consequently suppresses NSCLC cell growth. HO-1 expression is inversely correlated with miR-1254 level in human NSCLC tumor samples and cell lines. Collectively, these findings identify the dual inhibition of HO-1 through miR-1254 as a novel functional mechanism of miRNA, which results in a more effective inhibition of oncogenic mRNA, and leads to a tumor suppressive effect.
HO-1 over-expression has been reported to be involved in tumor growth and malignant progression [31–35], previous study from our laboratory has demonstrated that HO-1 is down-regulated by miR-1304 in lung cancer cell lines [37]. In order to find more effective miRNAs, we explored the miRNAs that potentially bind to the 3’UTR of HO-1 using bioinformatics tools. Twenty-six miRNAs were predicted to target HO-1 using all three databases (TargetScan [38], miRanda [39] and PITA [10]) (Fig 1A, left). To confirm whether these miRNAs target the 3`UTR of HO-1, we constructed a dual luciferase reporter (psi-HO1) by cloning human HO-1 3`UTR into the psiCHECK2 vector. Through co-transfection of 26 miRNAs individually with psi-HO1 reporter into HEK293 cells, we found that 11 miRNAs potently reduced the luciferase activity of psi-HO1 reporter (Fig 1A, right), indicating that these miRNAs potentially targeting HO-1 3`UTR and inhibiting HO-1 expression. To corroborate this finding, we transfected these miRNAs into human NSCLC A549 cells, western blot assays showed that miR-1254 had the strongest and most stable inhibitory effect on HO-1 protein expression (Fig 1B). We transfected different doses of miR-1254 into A549 cells, and found that miR-1254 suppressed HO-1 expression at both protein and mRNA levels in a dose-dependent manner (Fig 1C). However, the inhibition at mRNA and protein levels were not perfectly consistent, in low doses of miR-1254 (0.5~1nM), the inhibition at mRNA level is stronger than protein level, while in higher doses of miR-1254 (6~12.5nM), more dramatic inhibition was found at protein level, these results indicated that the suppression of HO-1 at mRNA and protein levels was probably achieved through different mechanisms. As a miRNA with the most dramatic inhibitory effects on HO-1 protein expression, and potentially functioning through multiple mechanisms, miR-1254 attracted our further interest.
We next sought to study the regulation of HO-1 expression by miR-1254 at both mRNA and protein levels in lung cancer cell lines. Human NSCLC A549 and NCI-H1975 cells were transfected with miR-1254 mimics for 48 hours, and then the expression levels of mature miR-1254, HO-1 protein and mRNA were examined. Taqman microRNA assay confirmed that miR-1254 mimics were successfully transfected into the cells and the level of mature miR-1254 was increased (Fig 2A, top). Consequently, the mRNA and protein levels of HO-1 were down-regulated in the cells transfected with miR-1254 mimics compared to that with the negative control oligonucleotides (Fig 2A bottom, 2B and 2C). We induced the expression of HO-1 in miR-1254-transfected and non-transfected cells via treatment with 20 μmol/L hemin, which is previously described as a HO-1 inducer [40]. In the presence of hemin, we found that HO-1 expression was greatly increased, and transfection of miR-1254 mimics inhibited the expression of induced HO-1 as well, to a lesser extent (Fig 2B and 2C). Taken together, our results demonstrated that miR-1254 mimics inhibited the expression of HO-1 in lung cancer cells. We subsequently explored whether the endogenous miR-1254 in NSCLC cells functions in the maintenance of HO-1 expression. MiR-1254 specific antisense oligonucleotides (Anti-1254) were used to down-regulate the level of endogenous miR-1254. Our results showed that the protein (Fig 2D, left and middle) and mRNA (Fig 2D, right) levels of HO-1 were up-regulated in A549 and NCI-H1975 lung cancer cells transfected with Anti-1254, compared to that with the negative control antisense oligonucleotides. In addition, CRISPR/Cas9 method was used to delete the genomic sequence of miR-1254 and further determine the endogenous relationship between miR-1254 and HO-1 (Fig 2E). The results showed that deletion of miR-1254 genomic sequence by CRISPR/Cas9 diminished endogenous miR-1254 level in A549 cells (Fig 2F). We established miR-1254 +/- and miR-1254 -/- cell lines derived from single clones, and examined HO-1 mRNA and protein levels in wild-type (WT), miR-1254 +/- and miR-1254 -/- cells. As expected, the expression of HO-1 at both mRNA and protein levels are negatively correlated with the level of miR-1254 in a dose-dependent manner (Fig 2G). Taken together, all these results suggest that both overexpressed and endogenous miR-1254 inhibit the expression of HO-1.
Since miRNAs usually direct post-transcriptional gene silencing through seed sequence binding to the 3’-UTR region of the target mRNA, we investigated the post-transcriptional effects of miR-1254 on HO-1. Predicted by TargetScan, miR-1254 seed region was complementary to the sequences from 1166–1172 in 3′UTR of HO-1 mRNA (Fig 3A). We generated dual luciferase reporter constructs containing wild type or mutant HO-1 3’UTR with mutations in miR-1254 potential binding site. Co-transfection of the wild type reporter with miR-1254 mimics resulted in a decrease of the luciferase activity in HEK293 cells (Fig 3B, lane1 and 2). As expected, the effect of miR-1254 mimics on luciferase activity was abolished in cells co-transfected with the reporter containing mutation in its binding site (Fig 3B). We further examined the mRNA and protein levels of HO-1 in A549 cells transfected with either miR-1254 or its seed sequence mutant (5`mt). Intriguingly, miR-1254-induced inhibition of HO-1 at mRNA level was not affected (Fig 3C) and the inhibition at protein level was only partly abolished (Fig 3D) upon transfection with 5`mt in A549 cells.
Given that miR-1254 reduced HO-1 mRNA level in A549 and NCI-H1975 cells (Fig 2A–2C and Fig 3C and 3D), we determined HO-1 mRNA half-life to preclude the contribution of miR-1254 on HO-1 mRNA stability. A549 and NCI-H1975 cells were transiently transfected with miR-1254 and control oligonucleotides (nc), and after 4 hours, cells were treated with 5 μg/ml actinomycin D, which is an established inhibitor of mRNA transcription [41, 42], for different times. As shown in Fig 3E, the half-life of HO-1 mRNA in A549 or NCI-H1975 cells transfected with miR-1254 was unaffected compared to the control cells, suggesting that miR-1254 did not affect HO-1 mRNA stability. This indicates that the inhibition of HO-1 at mRNA level is independent on the seed region of miR-1254, and provides further evidence to support our hypothesis that miR-1254 suppresses HO-1 expression through multiple mechanisms.
To test the hypothesis that whether miR-1254 suppresses HO-1 at the transcriptional level, first, we performed chromatin immunoprecipitation (ChIP) assays in A549 cells to examine the binding of RNA polymerase Ⅱ (pol-Ⅱ) on HO-1 promoter, and found that pol-Ⅱ but not an IgG control enrichment on HO-1 promoter fragment was reduced by miR-1254 (Fig 3F). Second, we constructed a HO-1 promoter (PGL-HO1) reporter by cloning a ∼1.5 kb human HO-1 promoter into the firefly luciferase vector PGL4.10, and we found that expressing miR-1254 greatly inhibited the luciferase activity of the reporter in HEK293 cells (Fig 3H, lane1 and 2). These findings suggest that miR-1254 represses HO-1 at transcriptional level. Moreover, mutation in miR-1254 seed region did not abolish the inhibition on the luciferase activity of the reporter, which indicated the transcriptional regulation functioned through the non-seed region (Fig 3H, lane 3). We designed different mutants of miR-1254 with mutations in non-seed region, and transfected them into HEK293 cells separately. As shown in the results, the mid region mutant mmt-6 (mmt) and the 3`region mutant 3mt-5 (3’mt) eliminated the inhibition of miR-1254 on HO-1 promoter activity (S1A Fig and Fig 3H, left, lane4 and 5). So we choose the mmt and 3’mt mutants in the following studies (Fig 3G). We examined the effect of miR-1254 non-seed region mutation (mmt and 3mt) on HO-1 mRNA in A549 (Fig 3I) and NCI-H1975 (S4A Fig, left) cells using qRT-PCR. Consistently with dual luciferase report assay, both mmt and 3mt abolished the inhibitory effects on HO-1 mRNA expression. These results suggest that miR-1254 inhibits HO-1 transcription via its non-seed sequence, while the seed region is additionally responsible for the inhibition at post-transcription level, which may consequently induces a dual inhibition of HO-1.
The novelty of miR-1254 non seed region induced transcriptional gene silencing (TGS) on HO-1 inspired us to explore the exact functional mechanism of its non-seed sequence. Our laboratory has reported that miR-552 binds to CYP2E1 promoter region via its non-seed sequence and induces TGS of CYP2E1 [28]. We analyzed the sequence alignment of miR-1254 with HO-1 promoter using miRBase database [39] and RNA hybrid [43], there were 6 potential binding sites in the fragment within 1.5kb upstream from the transcription start site (TSS) (S2A Fig, top). Non-denaturing PAGE experiment was performed to test the binding ability of miR-1254 with these 6 sites in vitro, and data showed that site 2 had the highest possibility to form hybrids with miR-1254 (S2A Fig, bottom). Next we used CRISPR/Cas9 to knockout site 2 in HO-1 promoter, however, the inhibition of miR-1254 on HO-1 mRNA level was not affected (S2B and S2C Fig). These results suggest that site 2 is not the functionally targeting motif in HO-1 promoter. Subsequently, we cloned 6 fragments of HO-1 promoter with varying length (S3A Fig) into the firefly luciferase vector PGL4.10 and individually co-expressed with miR-1254 in HEK293 cells. Data showed that expressing miR-1254 greatly inhibited the luciferase activity of the reporter containing the shortest fragment 1(only 150 bp upstream from TSS) (S3B Fig). However, when we mutated both of the two potential binding sites in the fragment based on sequence alignment, the inhibition of miR-1254 on HO-1 promoter was not abolished (S3C Fig). Altogether, the results suggested that miR-1254 might not suppress HO-1 transcription via directly targeting HO-1 promoter.
Previous study showed that miRNAs can also induced DNA methylation and consequently induced transcription inhibition [44, 45]. We tested this possibility by treating A549 cells with 1μM Decitabine, a DNA demethylation drug for 48 h, after transfection with miR-1254. As shown in S3D Fig, DNA methylation was abolished, however, the inhibitory effect of miR-1254 on HO-1 transcription still existed.
After excluding the possible mechanisms that miR-1254 directly targets HO-1 promoter or induces DNA methylation of HO-1 CpG islands we hypothesized that miR-1254 inhibits the transcription factors of HO-1 and consequently induces TGS. Plenty of regulatory elements have been identified in the promoter region of HO-1, targeted by transcriptional factors such as nuclear factor (erythroid-derived 2)-like 2 (Nrf2) [46, 47], activating protein-1 (AP-1) [48], up-stream stimulatory factor (USF) [49], nuclear factor-κB (NF-κB) and transcription factor AP-2(TFAP2) [50]. Since expressing miR-1254 inhibits the luciferase activity of the reporter containing the shortest fragment 1(only 150 bp upstream from TSS), and mutation of the binding sites on HO-1 promoter did not abolish the inhibitory effects, it’s possible that miR-1254 inhibit HO-1 transcription through targeting transcriptional factors. Predicted by TRANSFAC, TFAP2A and USF1 binding sites can be searched in this region. However, binding sites of NFκB, Nrf2 or AP-1 could not. Previous studies have demonstrated that the binding sites of NFκB are near to the binding sites of TFAP2A [46, 50], so we examined the expression of NFκB as well as TFAP2A and USF1. The results showed that TFAP2A but not USF1 or NF-κB was inhibited by miR-1254 and the inhibition effect were abolished by miR-1254 non-seed region mutants (mmt or 3mt), but not by seed region mutant (5mt), which indicated that TFAP2A may be involved in miR-1254-induced HO-1 TGS in A549 (Fig 4A) and NCI-H1975 cells (S4A Fig, right). RNA interference knockdown of TFAP2A (si-TFAP2A) potently reduced the HO-1 promoter luciferase reporter activity (Fig 4B, left). To further study the role of TFAP2A in miR-1254 regulation of HO-1, we cloned the coding sequence of TFAP2A into pTT5 vector (pTT5-TFAP2A) and then co-transfected it with PGL-HO1 into HEK293 cells. As shown in the data, the PGL-HO1 luciferase activity was substantially increased by the co-transfection with pTT5-TFAP2A in a dose-dependent manner (Fig 4B, right). We confirmed these results in A549 (Fig 4C) and NCI-H1975 cells (S4B Fig) with TFAP2A knocked down using chemically synthesized siRNA against TFAP2A (si-TFAP2A), the results consistently showed that HO-1 protein expression was decreased by si-TFAP2A. These results suggest that TFAP2A is a major transcription factor that functions in HO-1 activation in these cells.
Then, we examined the changes at mRNA and protein levels in A549 cells with CRIPSR/Cas9-modified miR-1254 knockdown. The results showed that the mRNA (Fig 4D) and protein (Fig 4E) expression levels of TFAF2A were dramatically increased in a dose-dependent manner in miR-1254-knockdown cells, but not the other two transcription factors of USF1 and NF-κB (Fig 4D), consistently with the changes of HO-1 expression at mRNA level (Fig 2G). These results suggested that the endogenous miR-1254 inhibited TFAP2A expression at both mRNA and protein levels. We then performed western blot assay to examine the effects of miR-1254 and its mutants on protein levels of TFAP2A and HO-1 in NSCLC cells. Consistently, the protein levels of both TFAP2A and HO-1 were inhibited by miR-1254 in both A549 (Fig 4F, lane1 and 2) and NCI-H1975 (S4C Fig, lane1 and 2) cells. MiR-1254 mmt basically lost the inhibition on TFAF2A and HO-1 protein expression, while miR-1254 5mt maintained the inhibitory function on TFAP2A, and partially attenuated the protein reduction of HO-1 (Figs 4F and S4C).
To confirm whether miR-1254 inhibits HO-1 mRNA expression through down-regulating TFAF2A, we co-transfected miR-1254 mimics with pTT5-TFAP2A into A549 (Fig 4G) and NCI-H1975 cells (S4D Fig), the results showed that over-expression of TFAP2A rescued miR-1254-induced inhibition on HO-1 expression. We further performed chromatin immunoprecipitation (ChIP) assays in A549 cells, and found that the enrichment of TFAP2A but not an IgG control on HO-1 promoter fragment was reduced by miR-1254, consistently with the enrichment results of polymerase Ⅱ (pol-Ⅱ) (Fig 4H). Collectively, these findings support that TFAP2A is a transcriptional activator for HO-1 in NSCLC cells and miR-1254 represses HO-1 transcription through targeting TFAP2A.
Our results suggested miR-1254 directs TGS of HO-1 expression via targeting the transcriptional activator TFAP2A, possibly through its non-seed region. We then elucidated the regulatory mechanism through which miR-1254 suppressed TFAP2A expression. We cloned 3`UTR of TFAP2A into the luciferase reporter psi-CHECK2 (psi-TFAP2A) and co-transfected with miR-1254 mimics or its mutants with mutation in different regions into HEK293cells. Consistent with the results on HO-1 mRNA expression, the luciferase activity of psi-TFAF2A was inhibited by miR-1254, and the inhibitory effect was abolished by miR-1254 non-seed region mutants (mmt or 3mt) but not seed region mutant (5mt) (Fig 5A). Then, we analyzed the sequence alignment of miR-1254 with TFAP2A 3`UTR. Predicted by RNA hybrid, miR-1254 non-seed region was complementary to the sequence in 3′UTR of TFAP2A mRNA (from 264–257) (Fig 5B). When mutations were introduced into the 8 nt sequence in TFAP2A mRNA 3’-UTR complementary to non-seed sequence of miR-1254, miR-1254 could not suppress the activity of the mutant reporter any longer (Fig 5C). Moreover, mutations were also introduced in seed and non-seed region of miR-1254 (5’mt, mut), qRT- PCR was performed to test their effects on HO-1 in A549 cells. As shown in the results, mRNA expression of TFAP2A was rescued when miR-1254 non-seed region were mutant (Fig 5D).
CRISPR/Cas9 was used to knockout the endogenous binding site of miR-1254 on TFAP2A 3`UTR genomic sequence (Fig 5E). The results showed that the inhibitory effect of miR-1254 on TFAP2A (Fig 5F, left) and HO-1 (Fig 5F, right) mRNA level were completely abolished in CRISPR/Cas9-modified A549 cells (CRISPR-sites). In addition, western blot showed that miR-1254 could not suppress TFAP2A protein expression any longer, however, miR-1254 still strongly inhibited HO-1 protein expression in CRISPR-sites cells (Fig 5G). These results suggest that miR-1254 binds to TFAP2A 3`UTR via its non-seed sequence through an 8 nt-contiguous Watson–Crick pairs and miR-1254 represses HO-1 expression at post-transcriptional level by directly targeting HO-1 3’UTR via its seed sequence. Altogether, we found that miR-1254 suppresses HO-1 expression at mRNA and protein levels through different mechanisms, and dependent on different regions.
As described previously, it has been found that HO-1 plays a vital role in promoting cell survival in several types of cancer [31–35]. It is highly possible that miR-1254 regulates human lung cancer cell growth through modulating the expression of HO-1. In order to investigate the effects of miR-1254 on lung cancer cell growth, miR-1254 mimics were transfected into A549 and NCI-H1975 cells. Trypan blue staining showed that miR-1254 over-expression for 3 days markedly decreased the number of A549 (Fig 6A) and NCI-H1975 cells (S5A Fig). MTT assay was used to examine the effects of miR-1254 on cell viability. Our results showed that the viability of A549 and NCI-H1975 cells transfected with miR-1254 mimics was clearly decreased compared to those transfected with negative control oligonucleotides (Figs 6B and S5B). In the colony formation assay, transfection with miR-1254 mimics inhibited the colony-forming activity of both A549 and NCI-H1975 cells, while transfection with negative control oligonucleotides has no such effects (Figs 6C and S5C). To determine the relationship between HO-1, miR-1254 and cell survival, a combination study was carried out whereby cells were first transfected with miR-1254 mimics, followed by treatment with 20μM hemin chloride [40]. Trypan blue staining, MTT assay and colony formation assay revealed that the decrease of cell viability due to miR-1254 transfection could be rescued by inducing HO-1 expression with hemin chloride in A549 (Fig 6A–6C) and NCI-H1975 (S5A–S5C Fig) cells. These data demonstrated that miR-1254 suppresses the growth of NSCLC cells by repressing the expression of HO-1.
To explore the precise path by which miR-1254 reduced the NSCLC cell number, we cloned the coding sequence of HO-1 into pTT5 vector (pTT5-HO1). The plasmids were transfected into NSCLC cells to overexpress HO-1 and test the rescue effects on cell proliferation, cell cycle and apoptosis. Western blot analysis of HO-1 and proliferating cell nuclear antigen (PCNA) indicated that miR-1254 over-expression in A549 cells reduced cell proliferation, as expected, the restoration of HO-1 expression strongly overrode the repression effects (Fig 6D). Flow cytometry combining with PI staining and Annexin V-FITC/PI staining assay were used to analyze the cell cycle and apoptosis, respectively. The results showed that enforced miR-1254 expression led to more than 10% S phase cell cycle arrest and a significantly higher percentage of apoptotic cells. Consistently, HO-1 re-expression attenuated miR-1254-induced S phase cell cycle arrest and cell apoptosis in A549 (Fig 6E and 6F) and NCI-H1975 (S5D Fig and S5E Fig). Altogether, these results implied that miR-1254 suppress the growth of NSCLC cells by inducing cell cycle arrest and cell apoptosis, and with a mechanism of inhibiting HO-1 expression.
MiR-1254 has been reported to be down-regulated in breast cancer cells. Over-expressing of miR-1254 could inhibit breast tumor growth and overrides tamoxifen resistance [51]. Our in vitro results suggested miR-1254 suppressed NSCLC cell growth, we also study the in vivo effects using mouse xenograft model. We established A549 cell line stably over-expressing miR-1254 (A549/miR-1254) by lentiviral transduction. Western blot showed that TFAP2A and HO-1 protein levels in A549/miR-1254 cells were markedly decreased compared with the cells over-expressing negative control oligonucleotides (A549/Cont cells) (S5F Fig). Then, A549/miR-1254 and A549/Cont cells were subcutaneously injected into nude mice respectively. We found that over-expression of miR-1254 in A549 cells significantly reduced tumor growth in nude mice compared with control cells (Fig 7A). These results additionally support our original finding that miR-1254 has inhibitory effects on NSCLC growth.
We examined the HO-1 mRNA level in 34 paired frozen NSCLC tumor samples and normal lung tissue specimens. Using quantitative reverse transcription–PCR (qRT–PCR), we found that the expression levels of HO-1 in tumor were significantly higher than those in normal lung tissues (Fig 7B). We analyzed the expression of TFAP2A in 57 paired tumor and normal samples from patients with NSCLC in The Cancer Genome Atlas (TCGA) database, and found that TFAP2A was significantly induced in tumor samples, consistently with HO-1 (Fig 7C). To characterize whether miR-1254 is involved in HO-1 regulation in human NSCLC, qRT–PCR was used to examine the levels of both miR-1254 and HO-1 mRNA in the same set of human NSCLC specimens. We found an inverse correlation between the level of miR-1254 and HO-1 mRNA expression in these tumors (Spearman’s R = − 0.4686, P = 0.0052<0.01 = (Fig 7D). The negative correlation between HO-1 and miR-1254 was also observed in multiple human lung cancer cell lines (Fig 7E). The results indicate that miR-1254 may be a negative regulator of HO-1 in human NSCLC patient samples and cell lines.
In general, miRNAs bind to 3`UTR of target mRNAs and direct PTGS via its seed sequence, however, we and other groups demonstrated that miRNAs can also function through its non-seed sequence [19, 28]. For example, 11–12 nt Watson–Crick paring between the center of the miRNA and the “centered sites” in target was proved to be functional in target suppression. [19, 20]. Our results suggest that miR-1254 binds to TFAP2A 3`UTR via its non-seed sequence through 8 contiguous Watson–Crick pairs effectively inhibits TFAF2A and its target gene HO- 1 expression. When we screened the miRNAs which potentially target HO-1 3’-UTR, which indicated the effect of post-transcriptional gene regulation, miR-1254 was not the one which has the most dramatic effect on HO-1 3’-UTR luciferase reporter, however, miR-1254 is the one which suppresses HO-1 protein expression most effectively, which also indicates a transcriptional inhibition in addition to the post-transcriptional gene silencing. Our data showed that HO-1 expression was negatively regulated by miR-1254 at transcriptional and post-transcriptional levels via its non-seed sequence and seed sequence, with indirect and direct mechanisms, respectively. The dual inhibition of miR-1254 we found here may represent a novel regulatory mechanism of miRNA, which results in a stronger and more stable suppression on target gene expression.
Previous studies have reported that miR-1254 expression is dysregulated in human breast cancer[51], retinoblastoma[52] and NSCLC[53, 54]. MiR-1254 was identified as a circulating miRNA downregulated in NSCLC patient serum, compared with healthy control [54]. However, in another study, miR-1254 was detected upregulated in early-stage NSCLC tumor samples, and is considered as a candidate for serum-based biomarker [53]. Those studies suggested that miR-1254 might be involved in tumor progression, but the exact function is largely unknown. As miR-1254 has a very high GC content (62.5%), it usually gives low reads in RNA-seq data, which makes it more difficult to be investigated [51].
HO-1 expression is widely up-regulated in various types of tumors and consequently impacts tumor development by promoting cancer cell growth, invasion and metastasis [55]. Previous studies showed that HO-1 can be regulated at both transcriptional and post-transcriptional levels [37, 56]. Several miRNAs have been reported as regulators of HO-1[37, 40, 57], and seven unreported miRNAs with inhibitory activity on HO-1 were screened out in our current work as shown in Fig 1B. These miRNAs are inactivated in different types of cancer cells under most circumstances. In the present study, we demonstrated that HO-1 is regulated by miR-1254 at both mRNA and protein levels in human lung cancer cells. MiR-1254 directly targets HO-1 3’-UTR via its seed sequence and represses HO-1 expression at post-transcriptional level. In parallel, miR-1254 suppresses TFAP2A, which is a transcriptional activator of HO-1, via its non-seed sequence, and consequently represses HO-1 expression at transcriptional level. This dual regulatory mechanism by miR-1254 at both transcriptional and post-transcriptional levels has the potential to lead a more effective inhibition effect on HO-1. Moreover, there are many other oncogenes which are similarly regulated with HO-1, they could be regulated both by miR-1254 and TFAP2A directly. ChIPBase was used to predict the transcriptional targets of TFAP2A and TargetScan was used to predict miR-1254 targets. The intersection of the two populations has 347 targets in total including HO-1. Importantly, our findings indicate that miR-1254 induces cell apoptosis and cell cycle arrest of NSCLC cells through inhibiting the expression of HO-1, consequently suppressing the NSCLC cell growth. The clinical data showing that HO-1 mRNA level is inversely correlated with miR-1254 in human NSCLC tumor samples, indicating that miR-1254 may be a negative regulator of HO-1 in physiological conditions. Collectively, our findings identify miR-1254 as an inhibitor of HO-1 with dual regulatory mechanisms via different sequence regions, and functioning in NSCLC cell growth inhibition (Fig 7F).
Despite the new functional mechanism of miRNA non-seed sequence, many details in the mechanism remain far more elusive. For example, is the non-seed region-dependent transcriptional or post-transcriptional gene regulation a universal effect of miRNA? How many miRNAs have dual regulation on their targets like miR-1254? In terms of the components and functional mechanism of RISC, what’s the difference between non-seed sequence-modified and seed region modified-gene regulation? Addressing these questions will give us a further insight into miRNA-modified gene regulation and also a better understanding of miRNA functions in the oncogenic signaling network.
For studies using human data, the study was approved by the ethics committees of Shanghai Pulmonary Hospital (approval number: K17-136) and an informed consent was obtained from all participants.
For studies using animal data, all experiments were performed according to the National Institutes of Health Guide for the Care and Use of Laboratory Animals, the guidelines approved by the Institutional Animal Care and Use Committee of the Shanghai Institute of Materia Medica (approval number: 2017-01-RJ-136).
All cancer samples were obtained from Shanghai pulmonary hospital (Shanghai, China) and were stored in liquid nitrogen until analysis. All experiments were conducted in accordance with the Declaration of Helsinki.
Human lung adenocarcinoma cell lines (A549 and NCI-H1975) were obtained from the American Type Culture Collection (ATCC, USA). Cells were cultured in RPMI-1640 (Gibco, USA) medium supplemented with 10% fetal bovine serum (FBS). HEK293 cells were purchased from ATCC and cultured in DMEM medium supplemented with 10% FBS. Cells were maintained in a humidified incubator at 37°C with 5% CO2.
Transient transfection was performed using Lipofectamine 2000 (Invitrogen, Carlsbad, CA, USA) according to the manufacturer’s instructions. 50nM of small interfering RNA and 25nM miR-1254 mimic or antisense oligonucleotide was used. In the TFAP2A rescue experiment, 500 ng of plasmid DNA was used in a 6-well plate, and in the HO-1 rescue experiment, 100 ng of plasmid DNA was used in a 6-well plate.
The plasmid pcDNA3.1-C5U (CCAR1 5`UTR) was a kind gift from Dr. Tao Zhu (School of Life Sciences, University of Science and Technology of China, Hefei, Anhui 230027, China). We cloned CCAR1 5`UTR into the lentiviral vector pCD513B-1 (pCD513B-1-1254) to over-express miR-1254. Human TFAP2A and HO-1 cDNA without 3′-UTR were cloned into pTT5 (obtained from Yves Durocher Lab) separately to construct the expression vectors. The 3`UTR of human HO-1 and TFAP2A was amplified via PCR using the genomic DNA of A549 and the PCR fragment was cloned into the psi-CHECK-2 vector separately (Promega, Madison, WI, USA). The promoter of human HO-1(~1.5kb) was also amplified via PCR using the genomic DNA of A549 and the PCR fragment was cloned into the PGL-4.10 vector. The primers used are listed in Supplementary S2 Table. All constructs were confirmed via DNA sequencing. miR-1254 mimic, anti-miR-1254 and small interfering RNAs targeting TFAP2A, or their respective negative control RNAs were purchased from GenePharma (Shanghai, China). The sequences of the RNA oligonucleotides are provided in Supplementary S1 Table.
The mutated plasmid was cloned using the KODPlus-Mutagenesis Kit (Toyobo, Osaka, Japan) as previously reported[58]. All the primers were shown in Supplementary S2 Table. DNA sequencing confirmed the nucleotide sequence of these plasmids.
Total RNA was extracted from cells using Trizol reagent (Invitrogen, USA) according to the manufacturer’s protocol. For HO-1 and TFAP2A expression, reverse transcription was performed with PrimeScript RT Master Mix (TaKaRa Biotechnology, China) following the manufacturer’s handbook. Quantitative real-time PCR (qPCR) was performed with QuantiNova SYBR Green PCR kit (Qiagen, USA) and analyzed on Rotor-Gene Q 2plex HRM System (Qiagen, USA).The relative HO-1 and TFAP2A mRNA levels were analyzed by normalizing the threshold cycle (Ct) value to that of internal loading control, β-actin. The primers are provided in the Supplementary S2 Table.
To quantify mature miR-1254, total RNA was reversely transcribed and amplified using TaqMan MicroRNA assay kit (Invitrogen, USA) according to the manufacturer's instructions. U6 snRNA were used as an internal loading control.
Total protein lysates were prepared from tumor cells and separated by 10% SDS-PAGE, transferred to PVDF membranes (Millipore, USA) and incubated with a primary antibody. HO-1 polyclonal antibody was purchased from Enzo Life Sciences, TFAP2A and PCNA antibodies were obtained from ABclonal Biotechnology, β-actin (Santa Cruze, USA) or α-tubulin (Cell Signaling Technology, Beverly, MA) was used as an internal control. The band densities were quantified by ImageQuant software (GE Healthcare, UK).
Cells seeded in 6-well plates were co-transfected with miR-1254 mimics (25 nM) or negative control and reporter constructs (200ng) using Lipofectamine 2000. Cell extracts were prepared 48h after transfection, and the luciferase activity was measured using the Dual-Luciferase Reporter Assay System (Promega).
Cell number was measured using Vi-cell XR cell viability analyzer (Beckman coulter, USA). Cell viability was determined using MTT assay. Briefly, cells were harvested following 24 h of transfection and plated at 2 × 103 cells per well in 96-well plates. After incubation, 20 μl MTT reagent (5.0 mg/mL) was added into each well and incubated in the dark at 37°C for 4 h. Then, 100 μl dissolution buffer was added into each well and incubated overnight. Absorbance was measured at 570 nm using a microtiter plate reader (Bio-Tek Instruments, USA).
Twenty-four hours after transfection with miR-1254 mimics or negative control oligonucleotides, the NSCLC cells were seeded in 6-well plates and grew for two weeks for the colony formation assay. The cells were then washed with PBS, fixed with methyl alcohol, and stained by Gimsa and then photographed using Typhoon FLA 9500(GE Healthcare, UK). Colonies were counted by ImageQuant TL (GE Healthcare, UK).
Cells were transfected with 500 ng indicated plasmid DNA and 25nM miRNA oligonucleotides in a 6-well plate. Apoptotic cells were examined using an Annexin V-FITC Apoptosis Detection Kit (BD Biosciences, USA). The cells were harvested and then stained with 5 μl of annexin V-FITC and 5 μl of PI for 15 min at room temperature in the dark. The cells were measured by the BD FACS flow cytometer (BD Biosciences, USA).
A549 cells (1 × 106) were transfected with miR-1254 mimics or control oligonucleotides at a 25 nM final concentration. Forty-eight hours later, cells were cross-linked with 1% formaldehyde for 10 min at 37°C and chromatin immunoprecipitation (ChIP) assay was performed using the ChIP Assay Kit from Upstate (Millipore). Five micrograms of anti-RNA polymerase II antibody (Millipore) and anti-TFAP2A antibody (ABCam) were used for each assay. No antibody (input) and normal rabbit IgG (Santa Cruz) were used as controls. Quantitative real-time PCR data were normalized to chromatin input and expressed as fold changes relative to the values in the cells transfected with negative control RNA oligonucleotides (nc). Primers are listed in Supplementary S2 Table.
The px330-mCherry and px330-GFP vectors were a kind gift from Dr. Hui Yang (Institute of neuroscience, Chinese academy of sciences, Shanghai, China) CRISPR/Cas9-modified nucleotide deletion was performed as previously described [28, 59]. Two sgRNAs were cloned into px330-mCherry and px330-GFP vectors, respectively. The sgRNA sequences are as follow: CRISPR-1254-left, 5`-caccgCCCAGCTACTTGGGAAGCTG-3`; CRISPR-1254-right, 5`-caccGTGTGTGTAAGGTTGCAGCT-3`; CRISPR-sites-left, 5`-caccgCACACCCCTGTGCCCTCATG-3`; CRISPR-sites-right, 5`-caccgACGGCCTGTTCTGTTCTCTT-3`. The plasmids were co-transfected into A549 cells (1 μg each) in a 6-well plate, and positively transfected cells were isolated using flow cytometry. The genome modification of each single cell used in the following studies was confirmed via DNA sequencing. Primers are listed in Supplementary S2 Table
To estimate the mRNA decay rates, transcription was inhibited by adding 5 μg/ml actinomycin D in medium [41, 42]. RNA was extracted at the indicated times and analyzed by qRT-PCR. The ratio of HO-1 mRNA to β-actin in each sample was calculated and used to determine the relative amount of specific mRNA remaining in each sample.
Animal studies were performed according to the National Institutes of Health Guide for the Care and Use of Laboratory Animals. Stable miR-1254–overexpressing A549 cells (A549/miR-1254) were harvested by trypsin, washed with PBS, and resuspended in Matrigel:RPMI medium (1:1); 1 million A549/miR-1254 cells and corresponding control cells were subcutaneously injected into the nude mice. Tumor volumes were calculated from the length (a) and the width (b) by using the following formula: volume (millimeters3) = ab2/2.
All statistical analyses were performed using GraphPad Prism software (version 5.01; GraphPad Software, Inc, CA, USA). The data are shown as the mean values with standard error of mean (SEM), and P<0.05 was considered significant. All experiments were performed independently at least three times. The significance of differences between two groups was measured by Student’s t test. One-way analysis of variance (ANOVA) was used to measure the significance of comparisons between more than two groups.
|
10.1371/journal.pcbi.1000143 | Falling towards Forgetfulness: Synaptic Decay Prevents Spontaneous Recovery of Memory | Long after a new language has been learned and forgotten, relearning a few words seems to trigger the recall of other words. This “free-lunch learning” (FLL) effect has been demonstrated both in humans and in neural network models. Specifically, previous work proved that linear networks that learn a set of associations, then partially forget them all, and finally relearn some of the associations, show improved performance on the remaining (i.e., nonrelearned) associations. Here, we prove that relearning forgotten associations decreases performance on nonrelearned associations; an effect we call negative free-lunch learning. The difference between free-lunch learning and the negative free-lunch learning presented here is due to the particular method used to induce forgetting. Specifically, if forgetting is induced by isotropic drifting of weight vectors (i.e., by adding isotropic noise), then free-lunch learning is observed. However, as proved here, if forgetting is induced by weight values that simply decay or fall towards zero, then negative free-lunch learning is observed. From a biological perspective, and assuming that nervous systems are analogous to the networks used here, this suggests that evolution may have selected physiological mechanisms that involve forgetting using a form of synaptic drift rather than synaptic decay, because synaptic drift, but not synaptic decay, yields free-lunch learning.
| If you learn a skill, then partially forget it, does relearning part of that skill induce recovery of other parts of the skill? More generally, if you learn a set of associations, then partially forget them, does relearning a subset induce recovery of the remaining associations? In previous work, in which participants learned the layout of a scrambled computer keyboard, the answer to this question appeared to be “yes.” More recently, we modeled this “free-lunch learning” effect using artificial neural networks, in which the synaptic strength between each pair of model neurons is a connection weight. We proved that if forgetting is induced by allowing each weight value to drift randomly, then free-lunch learning is almost inevitable. However, if, after learning a set of associations, forgetting is induced by allowing each connection weight to decay or fall toward zero, then relearning a subset of associations decreases performance on the remaining associations. This suggests that evolution may have selected physiological mechanisms that involve forgetting using a form of synaptic drift rather than synaptic decay, because synaptic drift yields free-lunch learning, whereas decay does not.
| The idea that structural changes underpin the formation of new memories can be traced to the 19th century [1]. More recently, Hebb proposed that “When an axon of cell A is near enough to excite B and repeatedly or persistently takes part in firing it, some growth process or metabolic change takes place in one or both cells such that A's efficiency, as one of the cells firing B, is increased” [2]. It is now widely accepted that learning involves some form of Hebbian adaptation, and a growing body of evidence suggests that Hebbian adaptation is associated with the long-term potentiation (LTP) observed in neuronal systems [3]. LTP is an increase in synaptic efficacy which occurs in the presence of pre-synaptic and post-synaptic activity, and can be specific to a single synapse. One consequence of Hebbian adaptation is that information regarding a specific association is distributed amongst many synaptic connections, and therefore gives rise to a distributed representation of each association.
In [4], participants learned the layout of letters on a “scrambled” keyboard. After a period of forgetting, participants relearned a subset of letter positions. Crucially, this improved performance on the remaining (i.e., nonrelearned) letter positions. However, whereas relearning some associations shows evidence of FLL in some studies [4]–[6], this is not found in not all studies [7]. This discrepancy may be because the many studies performed to investigate this general phenomenon use a wide variety of different materials and procedures, with some measuring recall and others measuring recognition performance, for example. However, within the realms of psychology, one relevant effect is known as part-set cueing inhibition.
Part-set cueing inhibition [8] occurs when a subject is exposed to part of a set of previously learned items, which is found to reduce recall of nonrelearned items. However, [9] showed that a learned row of words was better recalled if the cues consisted of a subset of words placed in their learned positions than if cue words were placed in other positions. In this case, part-set cueing seems to improve performance, but only if each “part” appears in the spatial position in which it was originally learned. This position-specificity is consistent with the FLL effect reported using the “scrambled keyboard” procedure in [4] but has no obvious concomitant in network models (e.g., [4],[10],[11]).
If the brain stores information as distributed representations, then each neuron contributes to the storage of many associations. Therefore, relearning some old and partially forgotten associations should affect the integrity of other associations learned at about the same time. As noted above, previous work has shown that relearning some forgotten associations does not disrupt other associations, but partially restores them. This FLL effect has also been demonstrated in neural network models ([10],[12]), where it can accelerate evolution of adaptive behaviors [13]. Crucially, in [12], the proof that relearning some associations partially restores other associations assumes that forgetting is caused by the addition of isotropic noise to connection weights, which could result from the cumulative effect of small random changes in connection weights. In contrast, here we prove that if forgetting is induced by shrinking weights towards zero, so that weights “fall” towards the origin, then relearning some associations disrupts other associations.
The protocol used to examine FLL here is the same as that used in [4] and [12] and is as follows (see Figure 1). First, learn a set of n1+n2 associations A = A1∪A2 consisting of two subsets A1 and A2 of n1 and n2 associations, respectively. After all learned associations A have been partially forgotten, measure performance error on subset A1. Finally, relearn only subset A2 and then remeasure performance on subset A1. FLL occurs if relearning subset A2 improves performance on A1.
In order to preclude a common misunderstanding, we emphasize that, for a network with n connection weights, it is assumed that n≥n1+n2 ; that is, the number of connection weights on each output unit is not less than the number n1+n2 of learned associations. Using the class of linear network models described below, up to n associations can be learned perfectly (see [12]).
The proofs below refer to a network with one output unit. However, these proofs apply to networks with multiple output units, because the n connections to each output unit can be considered as a distinct network, in which case our results can be applied to the network associated with each output unit.
Each association consists of an input vector x and a corresponding target value d. For a network with weight vector w, the response to an input vector x is y = w·x. We define the performance error for input vectors x1,…,xk and desired outputs d1,…,dk to be(1)where yi = w·xi is the output response to the input vector xi. By putting X = (x1,…,xk)T, d = (d1,…,dk)T andwe can write Equation 1 succinctly as(2)
The two subsets A1 and A2 consist of n1 and n2 associations, respectively. Let w0 be the network weight vector after A1 and A2 are learned. When A1 and A2 are forgotten, the network weight vector changes to w1, say, and the performance error on A1 becomes Epre = E(X;w1,d). Finally, relearning A2 yields a new weight vector, w2, say, and the performance error on A1 is Epost = E(X;w2,d). Free-lunch learning has occurred if performance error on A1 is less after relearning A2 than it was before relearning A2 (i.e., if Epost<Epre).
Given weight vectors w1 and w2, a matrix X of input vectors, and a vector d of desired outputs, define(3)which we shall also refer to simply as δ.
In previous work [12], we assumed that the “forgetting vector” v (defined as v = w1−w0) has an isotropic distribution. Here we shall assume instead that the post-forgetting weight vector w1 is given by(4)for some (possibly random) scalar r, so that(5)and therefore(6)The interpretation of Equation 6 is that forgetting consists of making the optimal weight vector w0 “fall” towards the origin by a falling factor 1−r.
We provide theoretical results, and compare these with results obtained using computer simulations. In essence, our theoretical and simulation results indicate that falling weights induce negative FLL, which decreases with the square of the falling factor 1−r.
Our two main theorems are summarised here, and proofs are provided in the Methods section. These theorems apply to a network with n weights which learns n1+n2 associations A = A1∪A2, and then after partial forgetting, relearns the n2 associations in A2.
We prove that if n1+n2≤n (so that, in general, the associations A1 and A2 are consistent) and the joint distribution of (X1,d1) is isotropic (where X1 and d1 are the matrix of inputs and the vector of desired outputs for subset A1 of associations) then the expected value of δ is negative (recall that δ is defined in Equation 3). We then prove that the probability P(δ<0) that δ is negative approaches unity as n1 approaches ∞.
For every non-zero value of r, the expected value of δ given r is negative. More precisely,(7)with equality only in trivial cases, and where the constant of proportionality is guaranteed to be positive. Thus, the expected amount of FLL is negative (or zero).
From a physiological perspective, the case r<1 is obviously of interest because it represents synaptic weight decay. However, from a mathematical perspective, Theorem 1 applies to every value of r, and so it also holds for r>1. In other words, any movement of the weight vector w along the the line connecting w0 to the origin yields an expectation of negative FLL, in accordance with Theorem 1.
Under mild conditions on the distributions of the input/output pairs (X1,d1) and (X2,d2),(8)where x and are any columns of and , respectively, and
Theorem 2 implies that, if (i) the number (n1) of associations in A1 is a fixed non-zero proportion ( n1/n ) of the number n of connection weights, (ii) E[∥d1∥2]E[∥d2∥−2] is bounded as n → ∞, and (iii) γ(n) → 0 as n → ∞ then P(δ>0) → 0 as n → ∞, i.e., the amount of FLL is negative, with a probability which tends to 1 as n → ∞.
For example, if we assume that (i) each input vector x = (x1,…,xn) is chosen from an isotropic Gaussian distribution and (ii) the variance of xi is then γ(n) = 2/n, , and E[∥d1∥2]E[∥d2∥−2] = n1/(n2−1). This ensures that P(δ>0) → 0 as n → ∞.
Simulation was carried out on a network with n input units and one output unit. The set A of associations consisted of k input vectors (x1,…,xk) and k corresponding desired scalar output values (d1,…,dk). Each input vector comprised n elements x = (x1,…,xn). The values of xi and di were chosen from a Gaussian distribution with unit variance (i.e., ). A network's output yi is a weighted sum of input values , where xij is the jth component of the ith input vector xi, and each weight wj is the connection between the jth input unit and the output unit.
Given that the network error for a given set of k associations is , the derivative of E with respect to w yields the delta learning rule , where η is the learning rate, which is adjusted according to the number of weights.
However, in order to save time, we used an equivalent learning method. Learning of the k = n associations in A = A1∪A2 was performed by solving a set of n simultaneous equations using a standard method, after which the weight vector w0 was obtained; this provided perfect performance on all n associations. Partial forgetting was induced by making weights “fall” towards the origin w1 = rw0, after which performance error was Epre. Relearning the n2 = n/2 associations in A2 was implemented with k = n2 as above, after which performance error was Epost.
In each simulation, each value in each input vector xi, and each target value di was chosen from the same isotropic gaussian distribution with unit variance. There were 100 input units, and one output unit. The subsets A1 and A2 each consisted of 50 associations. The value of δ = Epre−Epost was obtained in each of 100 simulations, using a different random seed for each simulation. In Figure 2, the mean of 100 values of δ is shown for various values of the falling factor 1−r.
We present a brief account of the geometry which underpins the results reported here, for a network with two input units and one output unit, as shown in Figure 3A. This network learns two associations A1 = (X1,d1) and A2 = (X2,d2).
Figure 3B provides a geometric example of how relearning A2 increases the error on A1. After learning A1 and A2, w = w0. The effects of forgetting and relearning can be seen by ignoring the ± superscripts and subscripts for now. After partial forgetting, w = w1, and performance error Epre = p2. Relearning A2 yields w2, the orthogonal projection of w1 on to L2, and performance error is Epost = q2. FLL occurs if δ = Epre−Epost>0, or equivalently if p2−q2>0 (see [12], Appendices A–C for proofs). Forgetting here consists of reducing w0 by a factor r<1, so that w1 = rw0.
The plus and minus signs in Figure 3B refer to two versions and of association A1, in which X1 is the same and the target d1 has the same magnitude, but opposite signs: and .
We now find the expected change in error induced by relearning a given association A2. After learning followed by forgetting, the change in error on after relearning A2 is . After learning followed by forgetting, the change in error on after relearning A2 is . Using similar triangles in Figure 3B,(9)(10)Therefore, the total change in error on and induced by relearning A2 (on different occasions) is(11)(12)(13)Irrespective of the precise value of the target output value d1 in A1, if the distribution of d1 is isotropic then +d1 is as probable as −d1. If the total change in error for two instances ( and ) of A1 is −2(1−r)2e2 then the expected change (conditional on e ) is E[δ|e] = −(1−r)2e2. Therefore, if forgetting is induced by falling weight values, then the expected change in error E[δ]<0.
We have proved and demonstrated that, in one of the simplest forms of neural network model, relearning part of a previously learned set of associations reduces performance on the remaining non-relearned associations. This result is in stark contrast to our previous results, which proved that relearning induced partial recovery of non-relearned items [12]. The only difference between these two studies is the way in which forgetting was induced.
An obvious physiological concomitant of Hebbian learning is long-term potentiation (LTP), which seems to underpin learned behaviors [14]. LTP can last for hours, days or even months, and usually follows an exponential decay [3]. However, some forms of LTP do not seem to decay [15], and have been shown to be stable for up to one year [16]. Such stability is remarkable, but from a statistical point of view, would almost certainly be accompanied by random fluctuations which would have a cumulative effect over time; and indeed, fluctuations are apparent in the stable LTP reported in [16]. Crucially, it is not known if the forgetting of learned behaviors is caused by decaying efficacy at many synapses, or by the cumulative effect of random fluctuations in stable LTP-induced synaptic efficacies. Here, decaying efficacy is analogous to weight values that fall toward zero in network models, whereas the cumulative effect of random fluctuations is analogous to the addition of random noise, or drifting, of weight values in network models.
Given a choice between forgetting via synaptic weights that fall towards zero and weights that drift isotropically, has evolution chosen drifting or falling? If all other things were equal then forgetting via synaptic drift would seem to be the obvious choice. This is because drifting ensures that relearning a subset of associations improves performance on other associations, whereas falling decreases performance. However, other things are rarely equal. The expected magnitude of weights increases with drifting but decreases with falling. (Consider a hypersphere centered on the origin, with radius ∥w0∥ . Simple geometry shows that more than half of all directions emanating from w0 yield a new weight vector w1 which lies outside the hypersphere, and therefore E[∥w1∥]>E[∥w0∥] (assuming, for example, that all vectors w1−w0 have the same length).) This decrease in weight magnitudes effectively reduces neuronal firing rates, which reduces metabolic costs relative to costs incurred by synaptic drift. Synaptic drift therefore confers mnemonic benefits, but these benefits come at a metabolic price. Thus the increased fitness gained from the mnemonic benefits of synaptic drift must be offset against their metabolic costs. In essence, even free-lunch learning comes at a price.
We proceed by deriving expressions for Epre, Epost, and for δ = Epre−Epost. We prove that if n1+n2≤n then the expected value of δ is negative. We then prove that the probability P(δ<0) that δ is negative approaches unity as n1 approaches ∞.
Given a c×n matrix X and a c -dimensional vector d, let LX,d be the affine subspaceof . If X and d are consistent (i.e., there is a w such that Xw = d) thenGiven weight vectors w1 and w2, a matrix X of input vectors, and a vector d of desired outputs, definewhere Epre = E(X;w1,d) and Epost = E(X;w2,d). Let be any element of LX,d. Then (14)
If Xi has rank ni then transposing the QR decomposition of (or, equivalently, using Gram–Schmidt orthonormalisation of the rows of Xi) givesfor unique ni×ni and ni×n matrices Ti and Zi with Ti lower triangular with positive diagonal elements, and . Simple calculation shows that, for any weight vector w, and are orthogonal. Since , it follows that the matrix represents the operator that projects orthogonally onto the image of . Because(15)the image of is contained in that of . As both these images have dimension ni, they must be equal, and so represents the operator which projects orthogonally onto the image of .
Now suppose that X and d are consistent, where
Then, after the network has learned A1 and A2, the weight vector w0 satisfies(16)(If, as below, n1+n2≤n, X2 and d2 are consistent, and (X1,d1) has a continuous distribution then Equation 16 holds with probability 1.)
We now assume that forgetting is induced by weight values “falling” towards the origin at zero, i.e., forgetting consists of shrinking the weight vector w0 by a (possibly random) factor r towards the “dead state” 0. Thus the post-forgetting weight vector w1 is given by(17)and so the “forgetting vector” v = w1−w0 is(18)
The form of forgetting given by Equation 17 is very different from that investigated in [12], where v has an isotropic distribution and is independent of (X1,d1) and (X2,d2).
Let w2 be the orthogonal projection of w1 onto L2. Then
Manipulation gives(19)and so(20)
Then Equations 14, 16, and 18–20 yield(21)
In this section we assume that the distribution of (X1,d1) is isotropic, i.e., that (UX1V,Ud1) has the same distribution as (X1,d1) for all orthogonal n1×n1 matrices U and all orthogonal n×n matrices V. Then taking the conditional expectation of Equation 21 for given X2, d2, and r gives the following theorem.
If
then(22)where x is any column of .
If 1.-3. of Theorem 1 hold then(23)with equality if and only if either r = 1 or d2 = 0.
Corollary 1 says that (apart from trivial exceptions) the expected amount of FLL is negative.
To obtain Theorem 2, it is useful to have some moments of isotropic distributions. Let x be isotropically distributed on . Then Equations 9.6.1 and 9.6.2 of Mardia and Jupp (2000), together with some algebraic manipulation, yield(24)(25)as in Equations A.14 and A.15 of [12].
The other tool used in proving Theorem 2 is the formula(26)for any random variables X,Y,Z for which these quantities exist. Equation 26 is an application to the conditional distribution of Y|Z of the standard conditional variance formula that is given in Equation 2b.3.6 on page 97 of [17].
Taking the expectation and variance of Equation 21 as only d1 varies and using Equation 24 gives(27)(28)
Taking the expectation of Equation 28 as only X1 varies and using Equation 24 gives(29)
We now suppose that(30)
Then taking the variance of Equation 27 as only X1 varies and using Equation 25 gives(31)
Adding Equations 29 and 30 and using Equation 26 yields(32)
To obtain an upper bound on the conditional probability of FLL (i.e., on P(δ≥0|X2,d2,r)), we use Chebyshev's inequality, which states that, for any random variable Y and any positive value of t
Applying Chebyshev's inequality to the conditional distribution of δ(w1,w2,X1,d1) given (X2,d2,r), taking t = E[δ(w1,w2;X1,d1)|X2,d2,r], and noting that (by Equation 23) t≤0, we obtain(33)
Substituting Equations 22 and 32 into Equation 33 gives(34)where
For any positive-definite symmetric matrix A and vector x, diagonalization of A, together with the fact that x+1/x≥2 for positive x, yields(35)
Combining Equations 34 and 35 with the fact that gives(36)
Taking the expectation of Equation 36 over X2 yields(37)where x and are any columns of and , respectively.
Taking the expectation of Equation 37 over d2 and r yields the following theorem.
If (a) conditions 1.-4. of Theorem 1 hold, (b) the columns of are distributed independently, (c) X2, d2, and r are independent, (d) the distribution of (X2,d2) is isotropic, and (e) E[∥d2∥−2] is finite then(38)where x and are any columns of and , respectively, and
If the conditions of Theorem 2 hold andwhere x and are any columns of and , respectively, then
Thusprovided that n1/n and n2/n are bounded away from zero.
|
10.1371/journal.pmed.1002597 | Chile’s 2014 sugar-sweetened beverage tax and changes in prices and purchases of sugar-sweetened beverages: An observational study in an urban environment | On October 1, 2014, the Chilean government modified its previous sugar-sweetened beverage (SSB) tax, increasing the tax rate from 13% to 18% on industrialized beverages with high levels of sugar (H-SSBs) (greater than 6.25 grams [g] sugar/100 milliliters [mL]) and decreasing the tax rate from 13% to 10% on industrialized beverages with low or no sugar (L-SSBs) (less than 6.25 g sugar/100 mL). This study examines changes in beverage prices and household beverage purchases following the implementation of the tax reform.
We used longitudinal data collected between January 1, 2013, and December 31, 2015, from 2,000 households. We defined the pretax period as January 1, 2013, to September 30, 2014, and the posttax period as October 1, 2014, to December 31, 2015. We conducted a pre–post analysis for changes in prices and purchases, with the latter examined by volume and calories. We compared posttax changes in prices and purchases to a counterfactual, defined as what would have been expected in the posttax period based on pretax trends. All results are stated as comparisons to this counterfactual. We linked beverages at the bar code level to nutrition facts panel data collected by a team of Chilean nutritionists who categorized them by taxation level and beverage subcategory, which included carbonated and noncarbonated H-SSBs and concentrated, ready-to-drink L-SSBs and untaxed beverages. We reconstituted concentrated beverages and analyzed all beverages using as-consumed volumes and calories. Posttax monthly prices of H-SSBs increased, but these changes were small. Prices of carbonated H-SSBs increased by 2.0% (95% confidence interval [CI] 1.0%–3.0%), while those of noncarbonated H-SSBs increased by 3.9% (95% CI 1.6%–6.2%). Prices of L-SSB concentrates decreased after the tax by 6.7% (95% CI −8.2%–−4.6%), and prices of ready-to-drink L-SSBs increased by 1.5% (95% CI 0.3%–2.7%). Households decreased monthly per capita purchases of H-SSBs by 3.4% by volume (95% CI −5.9%–−0.9%) and 4.0% by calories (95% CI −6.3%–−1.9%), and this change was greater among high socioeconomic status (SES) households. The volume of household purchases of L-SSBs increased 10.7% (95% CI 7.5%–13.9%), while that of untaxed beverage purchases decreased by 3.1% (95% CI −5.1%–−1.1%). The main limitation of this study was that there was no control group, so we were unable to assess the causal impact of the tax.
The modifications of Chile’s SSB tax were small, and observed changes in prices and purchases of beverages after the tax were also small. Our results are consistent with previous evidence indicating that small increases in SSB taxes are unlikely to promote large enough changes in SSB purchases to reduce obesity and noncommunicable diseases (NCDs).
| Sugar-sweetened beverage (SSB) taxes have emerged as a strategy to prevent a continued rise in obesity prevalence and noncommunicable diseases (NCDs).
Recent studies in Mexico and the United States have shown that SSB taxes are associated with increased prices of taxed beverages and reduced purchases of those beverages, particularly among households of low socioeconomic status (SES).
Little evidence exists on the effects increasing or decreasing an existing SSB tax has on beverage prices and purchases.
On October 1, 2014, Chile increased its tax on industrialized beverages with high levels of sugar (H-SSBs) from 13% to 18% and reduced its tax on industrialized beverages with low or no sugar (L-SSBs) from 13% to 10%. Beverages such as plain and flavored milk, 100% fruit juices, and unflavored water remained untaxed.
Understanding the effects of changes to existing SSB taxes is critical for informing future SSB tax policy.
We used a pre–post design to estimate changes in prices and purchases (volumes and calories) of beverages in Chile after the October 1, 2014, modifications of SSB taxes with data collected from January 1, 2013, to December 31, 2015. We compared observed posttax changes in beverage prices and purchases to the counterfactual, or what would have been expected in the posttax period based on pretax trends. All results are stated in terms of a comparison to this counterfactual.
Our model links household beverage purchase data to nutrition facts panel data using bar code information to classify each beverage purchased by tax level and subcategory.
We found that after the tax increase on H-SSBs, prices of H-SSBs increased, although the increases were small. In contrast, after the tax decrease on L-SSBs, prices decreased in some categories but increased in others. Prices of untaxed beverages also increased.
After the tax increase on H-SSBs, purchases of H-SSBs decreased, although the declines were small. After the tax decrease on L-SSBs, purchases of L-SSBs increased. Despite no change in the tax rate, purchases of untaxed beverages decreased.
High-SES households showed larger declines in purchases of H-SSBs after the tax than did low-SES households.
After a small increase in the tax rate on H-SSBs, changes in prices and purchases of these beverages were small. This is consistent with previous evidence indicating that small increases in SSB taxes are unlikely to produce large changes in SSB purchases. Further evidence is needed to understand the small change in prices due to the tax reform.
High-SES households showed larger declines in purchases of H-SSBs than did low-SES households. Further research is needed to understand the differential response to this tax by SES.
After a decrease in the tax rate on L-SSBs, purchases of these beverages increased a large amount. Further research is needed to understand the health consequences of the shift toward these beverages, which frequently contain artificial sweeteners.
| In response to the increasing global burden of obesity and related chronic diseases in the last decade, taxes on industrially produced sugar-sweetened beverages (SSBs) have emerged as a regulatory strategy to prevent the continued rise of obesity [1,2]. Research has shown that raising the prices of SSBs leads to significant decreases in SSB purchases [3–6], and recent studies in Mexico and cities in the US (e.g., Berkeley and Philadelphia) indicate that SSB taxes reduce purchases of SSBs, with a larger impact among lower socioeconomic status (SES) populations [7–11].
However, limited evidence exists on the impact of tax rate changes to existing SSB taxes. In high-income countries and low- and middle-income countries with rapid income growth, households might be unaware of small changes in tax rates due to higher median incomes, and commercial beverage companies might choose not to switch prices proportionally to the tax change [12]. In addition, access to safe tap water could affect choices available to consumers. For example, in Mexico, which leads in bottled water consumption worldwide [13] and has relatively limited access to safe tap water in many areas, the main substitution for SSBs in the first year after the tax implementation was water purchases [9]. In contrast, in high-income countries, free, clean water is readily available.
Chile, recently categorized as a high-income country, is an interesting case study to explore changes in prices and purchases after a change in the SSB tax rate. Chile has a high prevalence of obesity and type 2 diabetes [14–16] and recently became the country highest in SSB sales per capita [17]. Beverage taxes in Chile have a long history beginning in 1979, when the Chilean government introduced specific ad valorem taxes on alcoholic and nonalcoholic industrialized beverages. Beverage concentrates and all ready-to-drink industrialized beverages with any artificial flavoring, sweeteners, or dyes were subject to a common 15% tax rate. In 1985, the tax rate was cut to 13%. Roughly 30 years later, Chile modified its beverage tax again as part of a major tax reform that was announced and introduced in April 2014, approved in September 26, 2014, and implemented on October 1, 2014. This tax reform included an increase in the SSB tax rate that was intended to reduce purchases of SSBs and prevent continued increases in obesity and related noncommunicable diseases (NCDs). As a result, the tax rate on SSBs with greater than 6.25 grams (g) of sugar per 100 milliliters (mL) (e.g., sodas, industrialized juice drinks) increased from 13% to 18%. For SSBs with less than 6.25 g sugar per 100 mL (including powdered and concentrated beverages with added sugar and beverages containing artificial sweeteners, flavors, or dyes), the tax rate was reduced to 10%. Other beverages, such as plain milk and flavored sweetened milk-based drinks, 100% fruit juices, and unflavored water, remained untaxed.
The Chilean SSB tax structure is unique for two reasons. First, it creates a price differential between high-sugar and low-sugar SSBs, and second, it taxes beverages with artificial sweeteners (such as diet soft drinks) and flavored unsweetened beverages. In contrast, the SSB taxes in both Berkeley and Mexico apply a single rate to all nondairy and nonalcoholic beverages containing added sugar and do not apply to beverages with zero added sugar or those that contain only artificial sweeteners. The Philadelphia SSB tax is similar to Chile’s in that it also applies to artificially sweetened beverages but different in that it applies a single rate to all taxed beverages, regardless of sugar level. The United Kingdom SSB tax is also two tiered, with high-sugar SSBs taxed at a higher rate than low-sugar SSBs. However, this tax is based on the added sugar content of the beverage, and thus, artificially sweetened beverages containing no added sugar are not taxed. To our knowledge, the Chilean tax is the first that applies a two-tiered rate to SSBs and includes any artificially sweetened or flavored unsweetened beverages.
It is unclear how changing an existing SSB tax affects prices or purchases compared to the introduction of a new tax. Small sales taxes on SSBs (mean 5.2%) in jurisdictions across the US suggest that small relative increases in the tax rate may not lead to meaningful changes in sweetened beverage consumption [18,19]. However, these sales taxes are applied at the checkout and thus may not affect consumer purchasing decisions, whereas the Chilean SSB tax is included in the shelf prices of beverages. Understanding how relatively small increases in an SSB tax affect prices and purchases is critical to inform future SSB taxation policy.
The objectives of this study are to (1) analyze whether the average prices of beverage purchases changed after the tax implementation and (2) analyze whether the volumes and calories from beverage purchases changed after the tax implementation, controlling for household covariates and secular trends overall and by SES.
We used a longitudinal data set of household food purchases from January 2013 to December 2015 that we obtained from Kantar WorldPanel Chile (see the STROBE Checklist in S1 Table). The data are based on weekly purchases of fast-moving consumer goods by households from cities with more than 20,000 inhabitants, representing 74% of the urban population. The total sample is 2,000 households. Interviewers visited households weekly to collect data on food purchases using a handheld bar code scanner. First, information on purchases was collected either by scanning products’ bar codes on the packages or by using a codebook to assign bar codes tor bulk products or other products without bar codes. Second, households were instructed to keep all receipts so that interviewers could match purchases each week and determine the store where they were purchased (specifically for frequently consumed products). Finally, interviewers checked household pantries, and households stored empty product packages in a bin between interviews to ensure that products were not counted twice. Information collected for each beverage purchase included volume or weight, bar code expenditure, price per unit, retail channel, brand, package size, and date of purchase. A comparative analysis using the 2011–2012 Household Budget and Expenditure survey showed that households in the Kantar WorldPanel represent the average purchases of urban households [20].
For the price analysis, we sorted brands within each beverage subcategory based on their average market shares (i.e., the monthly average proportion of total sales) during the 2013–2015 period. We aggregated brands with market shares lower than 3% into a combined brand within each category. Brands with more than a 3% market share remained separate. We then classified products by market share as either low (less than or equal to 10% market share) or high (greater than 10% market share). Next, we defined package sizes for each individual brand. Ready-to-drink beverages we categorized as small (less than 2,500 mL) or large (greater than or equal to 2,500 mL). Beverages sold as liquid concentrates, powders, or dry leaves we categorized as small (less than 5 liters [L], diluted) or large (greater than or equal to 5 L, diluted). Hereinafter we refer to a product as the unique combination of brand and package size within each beverage category.
Our analysis used data on beverage prices as reported directly by households. Unlike data on prices collected from food stores, prices obtained from purchase data not only reflect changes in prices due to industry behavior but also differences in household preferences for certain beverage types or geographic differences in what products are available in the store (e.g., local brands may be available in some regions but not in others) [31]. In other words, the only prices captured are for beverages that are purchased, making it difficult to analyze whether the industry changed shelf prices in response to the tax modification. To deal with this complexity, we attempted to approximate prices consumers would see in a store by exploiting the geographic and socioeconomic variations across households. We defined a market as the pool of unique purchases that belong to the same region and SES group in the same month (using the product classifications described above). Thus, the average product-level prices in each market reflect the average price in a month for a given product that a household of a given SES and region would be likely to see in a store. We excluded from the analysis products that were not purchased both before and after the tax implementation (0.4% of all observations).
First, we examined sociodemographic characteristics of the households participating in the survey by SES status and separated households that exited the sample at any point between 2013 and 2015. Next, we calculated the average and median adjusted real prices at the market-month level, including the number of products within each beverage category. Finally, we calculated average purchases by household/month both by volume (mL) and calories (kilocalories [kcal]) purchased per capita per day for each beverage category and the percentage of nonzero household/month observations.
We conceived the initial study design in early July 2016 and purchased the Kantar data at the end of the same month. Between July 2016 and May 2017, master’s-level nutritionists cleaned and categorized the data, and we conducted descriptive analyses on volumes of purchases. Prior to any analysis, the initial analysis plan included:
In July 2017, we introduced two changes to the original analysis plan. First, we realized that, unlike in the Mexican tax evaluation, in Chile, no studies had examined changes in prices due to the tax implementation. Understanding changes in prices seemed critical to explaining any changes in beverage purchases (or lack thereof) after the SSB tax. Thus, we added a price analysis component, which included analysis of how estimated changes vary by product characteristics and market-level SES. Second, for both analyses, we tested several models to determine the best model specification for our data. As a result, we used a correlated random effects tobit model for volumes and a linear random effects model for prices. However, despite these modifications, the goal of the analysis remained the same: to compare observed posttax changes in prices and purchases to a counterfactual, or what would have been observed in the posttax period as predicted by pretax trends.
Finally, we introduced a few additional changes as requested during the review process, including the addition of 95% confidence intervals (CIs), sensitivity analyses on model specification, and reorganization of the tables in the main text. As requested, we also included the statistical significance of our results accounting for multiple hypotheses testing.
After a preliminary and descriptive analysis of the data set, we developed a pre–post analysis model to address the lack of a proper control group, since the tax was implemented at a national level. The goal of this analysis is to capitalize on changes in the average mean that occur before and after an intervention based on a break in time (the date of tax implementation). This approach is consistent with previous studies of Mexico’s SSB tax [9,10,32]. In this framework, for each analysis, we constructed a counterfactual, which represents the average predicted change in an outcome variable in the posttax period based on pretax trends. For each outcome, we calculated the difference between the observed trend and the counterfactual, on average, during the posttax period. Thus, each analysis is stated in terms of a comparison of what was observed after the tax compared to what might have been observed without a tax based on pretax trends (which is similar to an interrupted time series design). To ensure that our counterfactual captured the effect of the intervention rather than other contemporaneous changes, we controlled for time-varying confounders, seasonality, and national and regional trends, explained in more detail in the purchase analyses section. We calculated the statistical significance for each estimate adjusting for multiple hypotheses testing using the Sidak–Dunn correction [33]. Additionally, we reported nonconservative 95% CIs for each estimate. We estimated all models using Stata v.14.1.
We conducted all analyses at the product-market level using the natural logarithm of prices (assuming that prices follow a log-normal distribution). The unit of each observation is the price of each product per 1,000 mL in a market in a given month. We used a linear random effects model with standard errors clustered at the market level. This model uses a tax period indicator variable (pretax versus posttax) to create a break in the average mean purchases before and after the tax implementation. We controlled for linear and quadratic aggregated trends, seasonality, and regional economic covariates. Additionally, we examined whether price changes vary by low versus high market-level SES, package size, and the brand’s market share using interactions with the tax indicator variable.
After estimating the model, we calculated the adjusted predicted changes in price level and its corresponding 95% CI, comparing what was observed in the posttax period to the counterfactual (with the tax indicator variable set to zero). To obtain the predicted values, we back-transformed the conditional means, applying Duan smearing factors [34].
For our purchases analysis, the unit of observation was monthly per capita volumes or calories of beverages purchased by the household. We used a tobit model to explicitly recognize the large proportion of nonpurchases of each subcategory and to minimize biases introduced by the large proportion of nonpurchasers. We clustered standard errors at the household level.
Additionally, to adjust for unobserved heterogeneity by household (such as underlying differences in household preferences), we allowed for correlated random effects at the household level using the Chamberlain–Mundlak device [35,36]. We also included time-varying controls: household size, household composition by gender and age group, head of household education (no formal education, middle school, high school, college or more) and working status (unemployed, working, studying), indicator variables for region and SES (low and midlow versus midhigh and high), linear and quadratic aggregated trends, and quarterly seasonal indicator variables. We interacted SES with the tax indicator variable to analyze changes in effects by SES. We excluded from the calories analysis the beverage categories with no or very low sugar content (10% tax rate beverages), since the calorie content of these beverages is too low to analyze differences.
We conducted several sensitivity analyses to determine the best model specification using information criterion and goodness of fit for model selection. First, we tested different specifications for the trends, time breaks, and seasonality in each model and the significance of the time-varying covariates in each specification. Second, we tested the sensitivity of our estimates to autocorrelation (i.e., the similarity of the observations of each household over time) using the Arellano–Bond estimator [37]. We also estimated a static hurdle model to relax the assumptions imposed by the tobit model (i.e., that covariates have the same effect on both the probability of purchase and the amount purchased). Finally, as noted above, we examined interactions of the tax indicator with market-level SES, package size, and brand size for the price analyses and household-level SES for the purchase analyses. For each of these subgroup analyses, we report the p-value of a t test on whether the difference in the absolute predicted mean changes between the two groups in the posttax period is statistically different from zero.
Eighty-five percent of households were in the sample for all 36 months. Our analytic sample includes 1,795 unique households, 64,620 household monthly observations, and 114,003 market-month product observations.
Household characteristics in our sample are summarized in S3 Table. In this sample, low-SES households, compared to high-SES households, were more likely to live in Santiago and have an employed household head, a larger overall household size (including more children but fewer adults), and less education. Households were more likely to leave the survey if the household head was younger and slightly more educated, although we found no significant differences in household composition or SES. Given this, our estimates were slightly biased toward older households, conditional on being surveyed in January 2013.
Unadjusted average monthly market prices by beverage subcategory showed major differences between subcategories (S4 Table). We found the highest prices for noncarbonated H-SSBs. L-SSB concentrates were the least-expensive beverage type. We found the smallest (relative) differences in prices across SES among carbonated H-SSBs.
S5 Table shows average purchases (in volume and calories) at the household-month level. Carbonated H-SSBs (soft drinks) and untaxed beverages were purchased most frequently. Purchases of H-SSBs represent 18.0% of the total beverage purchases by volume and 55.0% of the total beverage purchases by calories in our data. Purchases of L-SSB concentrates were relatively large (10.0% of total beverage purchases) but contributed to a minimal proportion of calories purchased (0.2%). High-SES households purchased more of all beverage categories compared to low-SES households except for carbonated H-SSBs (2.3 versus 2.8 L/capita/month for high-SES and low-SES households, respectively).
The distribution of package size and brand size (i.e., market share) for all taxed products (weighted by total sales) are in S1 and S2 Figs, respectively. These results confirm that the use of 10% as a cutoff to define high versus low market share and the use of 2,500 mL to define small versus large package size was appropriate.
Monthly adjusted mean prices and purchases (volume) by beverage category are in S3 and S4 Figs, respectively. We found no obvious changes in prices or purchases after tax implementation, with one exception: the price of L-SSB concentrates decreased significantly following tax implementation. Finally, we tested normality in the distribution of prices (in logs) using the Shapiro–Wilk test on a random sample of products (S6 Table) and found that this assumption cannot be rejected.
Table 1 shows changes in average adjusted real prices at the product level (actual versus counterfactual estimates) by subcategory and tax regime by SES market and overall. Compared to the counterfactual, prices of carbonated H-SSBs rose by 2.0% (95% CI 1.0%–3.0%), and prices of noncarbonated H-SSBs rose by 3.9% (95% CI 1.6%–6.2%).
Compared to the counterfactual, L-SSB concentrates saw a sizable price reduction following the tax cut (−6.7%, 95% CI −8.2%–−4.6%), while prices of ready-to-drink L-SSBs increased slightly (1.5%, 95% CI 0.3%–2.7%). After the tax, the prices of untaxed beverages increased by 1.8% compared to the counterfactual (95% CI 0.7%–2.9%).
We found no significant differences in price changes for H-SSBs or L-SSBs by SES market. However, among untaxed beverages we found a significant increase in price for the low-SES market but no change in price for the high-SES market (p-value = 0.000).
S7 Table shows changes in average adjusted real prices at the product level (actual versus counterfactual estimates) by package size and market share. Overall, we found larger price variations in small packages relative to large packages (particularly for L-SSB concentrates). In terms of market share, we noted significant differences only in the price changes among L-SSB concentrates (p-value = 0.000). While L-SSBs with large market shares increased prices (despite the cut in the tax rate), brands with small market shares experienced a significant decrease in prices.
Absolute and relative changes in household beverage purchases by volume and calories are in Tables 2 and 3, respectively. Compared to the counterfactual, posttax household purchases of H-SSBs decreased in both volume (−3.4%, 95% CI −5.9%–−0.9%) and calories (−4.0%, 95% CI −6.3%–−1.9%). At the subcategory level, carbonated H-SSB purchases did not decrease by volume but did decrease 3.0% by calories (95% CI −5.2%–−0.8%) relative to the counterfactual. Noncarbonated H-SSBs decreased 8.2% by volume (95% CI −13.6%–−3.0%) and 8.9% by calories (95% CI −13.6%–−4.2%) relative to the counterfactual.
In contrast, the volume of household purchases of L-SSBs increased 10.7% (95% CI 7.5%–13.9%) relative to the counterfactual. By subcategory, we observed a 9.4% increase in purchase volume of L-SSB concentrates (95% CI 6.0%–12.8%) and a 12.3% increase in ready-to-drink L-SSBs (95% CI 8.4%–16.2%) relative to the counterfactual. Purchases of untaxed beverages decreased 3.1% by volume (95% CI −5.1%–−1.1%) and 5.3% by calories (95% CI −8.1%–−2.5%).
We found that posttax changes in H-SSB purchases varied by household SES (p-value = 0.004 for volume and 0.006 for calories). In general, high-SES households showed larger declines in H-SSB purchases after the tax than did low-SES households. Specifically, compared to their respective counterfactuals, high-SES households reduced purchases of H-SSBs by 6.4% by volume (95% CI −9.3%–−3.5%) and 6.5% by calories (95% CI −9.1%–−3.9%), whereas low-SES households showed no change in the volume or calories of H-SSB purchases.
We also found that posttax changes in the volume of L-SSB purchases varied by household SES (p-value = 0.006), though the size of the difference between SES groups was small. Specifically, relative to their respective counterfactuals, high-SES households increased the volume of L-SSB purchases by 10.8% (95% CI 7.8%–13.8%), while low-SES households increased the volume of L-SSB purchases by 9.5% (95% CI 5.8%–13.2%). Finally, there were also statistically significant differences in posttax changes of untaxed beverage purchases in high- versus low-SES households by volume (p-value = 0.035) but not by calories (p-value = 0.091). Specifically, relative to their respective counterfactuals, high-SES households had no change in the volume of untaxed beverages purchased, while low-SES households decreased the volume of untaxed beverage purchases by 4.3% (95% CI −6.6%–−2.0%).
The results presented in this paper reflect the best model fit based on Akaike information criteria (AIC) [38] and goodness of fit (R-squared), taking into account the particular characteristics of each model. A summary of results of different model specifications and AIC for selected outcomes is in S8 Table. We note that alternative model specifications provide less precise results compared to our chosen model specification, although results follow similar patterns. The model with a break in trends and intercepts provides more precise estimates for some outcomes. However, in this model, there are no significant differences in the pretax and posttax trend estimates; thus, results are to be interpreted with caution. For analyses of the volume of purchases, we found that a static hurdle model provided similar though less precise results due to an inability to capture household unobserved time-invariant heterogeneity (in household preferences, for example). Finally, we noted the lack of significance of the pre–post indicator when we switched it to April 2014, when the law was introduced in the National Congress of Chile, which is suggestive evidence that there was no significant anticipatory behavior prior to the tax implementation.
In October 2014, Chile increased its tax rate on H-SSBs from 13% to 18% and reduced its tax rate on L-SSBs from 13% to 10%, representing an 8% spread between L-SSBs and H-SSBs after tax implementation. This study found that after the 5% tax increase on H-SSBs, prices increased by 2.0% for carbonated H-SSBs and 3.9% for noncarbonated H-SSBs relative to their respective counterfactuals. After the 3% tax decrease on L-SSBs, price changes were heterogeneous, with prices decreasing for concentrated L-SSBs by 6.7% and for ready-to-drink L-SSBs by 1.5% relative to their respective counterfactuals.
Commensurate with the small increase in prices of H-SSBs after the tax relative to the counterfactual, we found a small decrease in household purchases of H-SSBs (−3.4% by volume and −4.0% by calories), with most of the declines coming from noncarbonated H-SSBs (−8.2%). Overall household purchases of L-SSBs increased relative to the counterfactual (10.7% for volume), with larger changes among ready-to-drink L-SSBs (14.3%).
The price increases of H-SSBs were small (2.0% for carbonated and 3.9% for noncarbonated H-SSBs relative to the respective counterfactuals) compared to the tax increase of 5% on these beverages. This is different from what was observed in Mexico, where shelf prices of carbonated SSBs increased proportionally to the size of the tax [32]. However, there are several important differences between this study of Chile’s tax and studies of Mexico’s tax. First, the current study examined only available prices of beverages that were purchased, whereas in Mexico and in most price analyses, the data come from the food store environment, allowing comparison of shelf prices for the same products over time. Second, changing store prices has administrative costs, which might have prevented some brands from increasing prices, especially considering the relatively small size of the tax hike in Chile (half the size of the tax implemented in Mexico). Further research is needed to determine whether the Chilean SSB tax hike led to a proportional change in shelf prices.
We also found that for H-SSBs and L-SSBs, price changes varied across beverage subcategories when we considered differences in brand size, package size, and market-level SES. While it is outside the scope of this article to explain commercial beverage company behavior, in Chile, companies commonly maintain a complex portfolio of carbonated H-SSBs and carbonated L-SSBs (e.g., regular and diet soda), which allows for a cross-subsidy between tax categories to maintain profits. This could explain the small absolute difference in the prices of ready-to-drink L-SSBs and carbonated H-SSBs after the tax. In contrast, L-SSB concentrates and noncarbonated H-SSBs tend to be manufactured by different companies, preventing them from cost-shifting from one category to another. This may explain why among these beverage types the observed price changes were consistent with the tax changes (i.e., the price of L-SSB concentrates decreased after the tax, consistent with the tax rate decrease on these beverages, while the price of noncarbonated H-SSBs increased after the tax, consistent with the tax rate increase on these beverages). Similar industry responses were observed in Mexico, where price changes varied geographically and by package size [32,39]. Future research should examine changes in beverage company behavior due to this complex tax structure.
With regard to household purchases of H-SSBs, after the tax, there was an overall decline of 3.4% by volume relative to the counterfactual. This reflected substantial heterogeneity by beverage subcategory, with an 8.2% decline in volume for noncarbonated H-SSBs and a 2.6% decline in volume for carbonated H-SSBs. Although these posttax declines in volume purchased are larger than the corresponding changes in price, these findings are consistent with previous studies of price elasticities (i.e., sensitivity to price) of SSBs in Chile, which range from −1.30 to −1.37 [4,40]. In other words, based on these elasticities, for a 2.0% increase in price, we would have expected consumers to decrease H-SSB purchases by approximately 2.6%, consistent with our results.
Despite these declines in purchases, however, this tax is likely to have a small impact on preventing excess caloric intake or adverse health effects associated with SSB intake. This is because the absolute reductions in H-SSB purchases were small. For example, compared to the counterfactual, the overall volume of purchases of H-SSBs decreased by 108 mL per person per month (equivalent to approximately a third of a can of soda). Similarly, calories purchased decreased by 3.0% and 8.9%, respectively, for carbonated H-SSBs and noncarbonated H-SSBs compared to their respective counterfactuals. This reflects absolute reductions of only 34 calories per person per month for carbonated H-SSBs and 21 calories per person per month for noncarbonated H-SSBs compared to their respective counterfactuals. To put this in context, Mexico’s SSB tax was about 10%, which translated to a 6% decline in volume of purchases relative to the counterfactual in the first year after tax implementation but represented a larger absolute decline of 360 mL per person per month compared to the counterfactual (with no results yet on corresponding calorie changes) [9]. In Chile, these small changes in volume and calories purchased seem unlikely to have large effects on obesity and NCD risk, but more research is needed to understand any potential implications for dietary intake and weight gain.
On the contrary, purchases of L-SSBs increased considerably in the posttax period (10.7% relative to the counterfactual). This translates into absolute increases of 281 mL per person per month (or roughly three-fifths of a can of soda) beyond what would be expected based on pretax trends. We were unable to estimate changes in calories purchased for L-SSBs, because many of them had very few or no calories. While the absolute changes in volume of L-SSB purchases were small, the relative percentage increase may be concerning, particularly for high consumers, given the lack of evidence regarding the long-term health effects of consumption of artificial sweeteners [41].
There are several possibilities that explain the smaller declines in H-SSB purchases after Chile’s tax compared to Mexico’s tax. First, as noted above, the estimated relative price increases on H-SSBs were smaller in Chile than in Mexico. Second, in Mexico, the SSB tax was accompanied by a prolonged, multiyear advocacy campaign that included coalition-building across key scientific and consumer advocacy organizations and paid and earned media campaigns, public demonstrations, press conferences, scientific forums, and civil society forums [42]. Such campaigns have the potential to influence consumers’ social norms, attitudes, and purchasing behaviors independently of the regulation itself, as was shown in a decline in SSB consumption in Mexico aligned with the initiation of the public discourse [9,43]. For example, a recent study found that a public awareness campaign to decrease SSB intake via TV advertising, digital marketing, outdoor advertising, social media, and earned media was associated with an accelerated decrease in SSB purchases [44]. In Chile, the SSB tax modification was a small component of a major fiscal reform and as such lacked significant advocacy campaigns. This might have limited potential changes in social norms and attitudes related to SSB consumption.
Even though the Chilean tax modification did not affect previously untaxed beverages, such as milk or 100% fruit and vegetable juices, this study found that prices of these beverages increased 1.8% in the posttax period relative to the counterfactual. Our estimates suggest that the changes in the relative prices of taxed beverages affected the markets of untaxed products, particularly in low-SES markets (which experienced a 3.2% increase in prices of untaxed beverages relative to the counterfactual).
Commensurate with this price increase, household purchases of untaxed beverages decreased in the posttax period (−3.1% by volume and −5.3% by calories relative to the counterfactual).
Interestingly, the largest reduction in calories purchased after the tax change came from untaxed products rather than taxed H-SSBs. One possibility is that consumers substituted L-SSBs for untaxed beverages, which is consistent with the finding that after the tax decrease on L-SSBs, untaxed beverages became relatively more expensive. Because different untaxed beverage subcategories have varying amounts of sugar (such as lactose in plain milk and added sugar in sweetened flavored milk) and beneficial nutrients, like calcium in milk or vitamin C in 100% juice, it will be useful for future research on SSB taxes to examine the changes in purchases that occur across various untaxed beverage subcategories and the potential implications of these changes for health.
This study found significant differences in changes in price after the tax by market-level SES for untaxed beverages but not for H-SSBs or L-SSBs. It was not clear why changes in the price of untaxed beverages would vary by market-level SES. We were unable to tell whether these differences in the change in price by SES market were due to differential price changes by retailers and manufacturers or differences in consumer behavior, as noted earlier. For example, if low-SES households change how often they buy products or favor bulk discounts on large products, it could lead to variations in the price changes across markets. More research is needed to explore differential price changes by SES market after the tax to understand the implications for both consumer behavior and industry behavior.
We found that high-SES households had a larger decline in the volume of H-SSB purchases (−6.4%) than did low-SES households (−1.6%) relative to their respective counterfactuals. This was mainly driven by differences in purchases of carbonated H-SSBs. While high-SES household volume purchases of H-SSBs decreased by 7.2%, low-SES household volume purchases did not change. Low-SES households showed larger declines in noncarbonated H-SSBs than high-SES households, but the difference was not statistically significant. In addition, it is important to note that carbonated H-SSBs represent a larger proportion of SSB purchases than do noncarbonated H-SSBs, so the SES differences for carbonated H-SSBs are especially important.
The result—that high-SES households had bigger changes in purchases of H-SSBs—contrasts with the Mexican first-year SSB tax evaluation, which found a 9.2% decline in low-SES households compared to a 5.6% decline for high-SES households [9]. This result is also contrary to expectations that lower-SES households will respond more to price increases due to a higher price sensitivity in this group. There are several potential explanations for these counterintuitive findings. Low-SES households could be more likely to intentionally avoid the tax by changing purchasing strategies, such as making larger but less frequent purchases [45]. Also, low-SES households in Chile are higher consumers of H-SSBs (relative to high-SES households) and thus may be less likely respond to price changes, because they have a stronger preference for these beverages [46]. While the absolute reductions in H-SSB purchases were small for all SES groups, the higher changes observed among high-SES households are concerning. Considering that SSB taxes are expected to produce larger changes in behavior among low-SES households and that this expectation is often used to rationalize the use of such taxes to reduce SES-related disparities in obesity and NCDs, future research should examine the drivers of this differential response by SES and potential implications for diet and health.
Finally, we found differences in changes in posttax purchases by household SES. Low-SES households showed larger reductions in untaxed beverage purchases by volume, but high-SES households showed larger reductions in untaxed purchases by calories. These differences in response by household SES could be the result of changes in the types of untaxed beverages purchased, since this category is very heterogenous and includes beverages with calories, such as milk and 100% juice, but also beverages with no calories, such as plain bottled water. We were unable to examine differences by beverage subcategory to better explain these findings due to the low amount of purchases in some of these subcategories.
The most important limitation of this study is our inability to assess a causal relationship between the tax modifications and changes in prices or purchases due to the potential presence of other simultaneous trends affecting underlying preferences and due to the inability of household food purchase data to capture all beverages consumed (particularly those consumed out of the home). While we did adjust for economic factors and secular trends, we cannot rule out shifts in preferences or norms that may have occurred concurrently with the tax implementation. To address our inability to capture all beverages using household purchase data, we will conduct future research using dietary recall data to capture the full range of beverages consumed. In the meantime, we note that the trends observed here are consistent with those in Euromonitor sales data, which reflect all beverages sold in the country [47]. In terms of external validity, we note that our sample is more likely to represent urban and older households and therefore does not completely reflect changes in purchases among younger households and the rural population. However, a comparative analysis of our data shows that the distribution of purchases by beverage group is consistent with those reported in the 2011–2012 Chilean Budget and Expenditure Survey. Furthermore, this study examines only the first posttax year in Chile. It is possible that subsequent years will see additional purchasing changes, as occurred in Mexico [10].
An additional limitation of this study is that, as previously noted, the prices reflected products purchased and not those available in stores (i.e., what was available in the market before and after the tax). However, we were able to address this issue using models that consistently captured variability in prices, allowing us to accurately estimate the average market changes while also exploring heterogeneity across product and brand characteristics. We also explicitly accounted for the censored nature of purchasing data by estimating a tobit model that predicts both the choice to purchase and the amount purchased. Finally, we were able to examine differential changes in both prices and purchases by SES (at the market level for price and at the household level for purchases). An understanding of differences in tax responses by SES is important for understanding how the tax may affect existing disparities in diet and health.
Additional research is needed to understand the observed changes in prices and purchases and the differential responses by SES. For example, one possibility is that companies changed their marketing strategies to mitigate the effects of the price increase. A second possibility is that companies could have changed the nutritional formulation of products to avoid the tax. We also want to note that a subsequent Chilean law in July 2016 implemented front-of-package warning labels, restrictions on marketing to children, and restrictions on sales in schools for all foods and beverages containing high levels of sugar, sodium, saturated fat, or calories, which could have had additional influence on both commercial beverage industry behavior and consumer behavior in the time period leading up to the implementation of those regulations. Given the substantial regulatory changes underway in Chile’s food environment since 2014, a critical question for future research is whether this small SSB tax modification, along with the newer marketing and media controls and front-of-package labels, will result in sustained changes in Chileans’ dietary intake with potential downstream effects for SES disparities in obesity and NCDs.
Finally, given the large and increased consumption of ready-to-drink L-SSBs and L-SSB concentrates, we recommend that the effects of these beverages on health be further investigated in the Chilean population, especially considering that they are relatively cheaper than their high-sugar alternatives and often bear significantly more front-of-package marketing, such as nutrition and health claims [48].
The small increase (5%) in the Chilean tax rate on H-SSBs was followed by small decreases in prices and purchases of these beverages, with high-SES households showing larger declines in purchases than low-SES households. The small decrease (3%) in the tax on L-SSBs was followed by heterogeneous changes in prices and increases in purchases of these beverages. Prices of untaxed beverages increased, and purchases of these beverage decreased, despite the lack of a tax or a change in the tax rate on these beverages. Our results are consistent with previous evidence indicating that small tax rates on SSBs are unlikely to promote the changes in SSB purchases needed to reduce obesity and NCDs.
|
10.1371/journal.ppat.1006401 | Suppression of IL-12p70 formation by IL-2 or following macrophage depletion causes T-cell autoreactivity leading to CNS demyelination in HSV-1-infected mice | We have established two mouse models of central nervous system (CNS) demyelination that differ from most other available models of multiple sclerosis (MS) in that they represent a mixture of viral and immune triggers. In the first model, ocular infection of different strains of mice with a recombinant HSV-1 that expresses murine IL-2 constitutively (HSV-IL-2) causes CNS demyelination. In the second model, depletion of macrophages causes CNS demyelination in mice that are ocularly infected with wild-type (WT) HSV-1. In the present study, we found that the demyelination in macrophage-intact mice infected with HSV-IL-2 was blocked by depletion of FoxP3-expressing cells, while concurrent depletion of macrophages restored demyelination. In contrast, demyelination was blocked in the macrophage-depleted mice infected with wild-type HSV-1 following depletion of FoxP3-expressing cells. In macrophage-depleted HSV-IL-2-infected mice, demyelination was associated with the activity of both CD4+ and CD8+ T cells, whereas in macrophage-depleted mice infected with WT HSV-1, demyelination was associated with CD4+ T cells. Macrophage depletion or infection with HSV-IL-2 caused an imbalance of T cells and TH1 responses as well as alterations in IL-12p35 and IL-12p40 but not other members of the IL-12 family or their receptors. Demyelination was blocked by adoptive transfer of macrophages that were infected with HSV-IL-12p70 or HSV-IL-12p40 but not by HSV-IL-12p35. These results indicate that suppression of IL-12p70 formation by IL-2 or following macrophage depletion causes T-cell autoreactivity leading to CNS demyelination in HSV-1-infected mice.
| Several mouse models of multiple sclerosis (MS) are now available. We have established two new mouse models. In the first model, ocular infection of different strains of mice with HSV-IL-2 recombinant virus causes CNS demyelination. In the second model, CNS demyelination was induced by different strains of wild type HSV-1 in the absence of macrophages. In the present study, we found differences in T-cell reactivity in the two models. However, both models exhibited an imbalance in IL-12p35 and IL-12p40. The requirement for formation of the IL-12p70 dimer in prevention of demyelination was supported by adoptive transfer experiments. These results suggest a pathological role for macrophages in these models of virus-induced MS in which suppression of IL-12p70 formation by IL-2 or following macrophage depletion causes T-cell autoreactivity leading to CNS demyelination.
| Multiple sclerosis (MS) is due to degradation of the myelin sheath [1] and visual disorders due to demyelination of the optic nerve is the early sign of individuals diagnosed with MS [2,3]. Thus, optic neuritis can be used as an early factor for detection of MS. Both genetic and environmental factors are implicated in development of optic neuritis and MS [4–8]. Considerable evidence supports the concept that dysregulation of IL-2 plays a critical role in the development of MS [9–18]. We therefore developed a model of MS in which we combined altered expression of IL-2 with an environmental signal, HSV-1 infection. In this model, ocular infection of mice with HSV-IL-2 recombinant virus caused demyelination in the brain, spinal cord, and optic nerve [19,20]. Ocular infection with parental, wild-type (WT) viruses, or with recombinant HSV-1 expressing either IFN-γ or IL-4, did not induce CNS demyelination. Similar results were obtained following delivery of rIL-2 protein, IL-2 DNA or IL-2 synthetic peptides prior to infection with different strains of HSV-1 [21]. Thus, the HSV-IL-2 offers a new and different small animal model for MS that integrates an environmental (viral) signal [19,20,22–24]. In this HSV-IL-2 model, the production of IL-2 by HSV-IL-2 is similar to the increases in IL-2 that have been observed in MS and there was increased T-cell autoreactivity leading to the CNS demyelination.
The second model arose from the finding that ocular infection of macrophage-depleted mice with WT HSV-1 leads to demyelination in the absence of an external source of IL-2. CNS demyelination did not occur in macrophage-intact mice that were ocularly infected with WT HSV-1 in the absence of an external source of IL-2 [19,20,22–24] and CNS demyelination did not occur on depletion of T cells, B cells, dendritic cells (DCs), or natural killer (NK) cells following ocular infection with WT HSV-1 [22].
The identification of these two closely related models provided the opportunity to use a comparative analysis approach to identify the mechanisms by which macrophages may contribute to, or modulate, demyelination in the context of ocular viral infection. Macrophages are mononuclear phagocytes that play critical roles in development, tissue homeostasis and the resolution of inflammation [25]. The wide variety of functions exhibited by macrophages include cytokine secretion and antigen presentation, and cytotoxicity as well as phagocytosis. Macrophage infiltrates are an integral component of the immune defense system. They are central to innate immunity and contribute to the intersection between innate and adaptive immunity. A number of factors are known to "activate" or engage macrophages in these activities, including viral infection. Following infection of naive mice with HSV-1, macrophages are the major infiltrates of the eye [26–28], and play a central role in both enhancement and blocking of inflammation in the eye [29–33]. In addition to their phagocytosis, antigen presentation and cytokine production [34,35], macrophages are the major source of IL-12 production [36,37], a cytokine reported to be involved in stimulation of both T cells and NK cells [38–40]. IL-12 also has been shown to enhance the TH1 response [41–43].
The results of our current study demonstrate that: 1) In the HSV-IL-2 model, demyelination was eliminated on depletion of FoxP3-expressing cells, when macrophages were present but not when macrophages were depleted. In contrast, the macrophage-depleted HSV-1 infected mice did not show demyelination when Foxp3-expressing cells were depleted. However, in the absence of macrophages FoxP3- T cells caused demyelination; 2) In both models, macrophages played a critical role in prevention of autoimmunity; 3) Either suppression of macrophages by IL-2 or their absence caused an imbalance of T cells and the development of autoaggressive T cells; and 4) Adoptive transfer of macrophages over-expressing IL-12p70 or IL-12p40, but not IL-12p35, blocked HSV-1 induced CNS demyelination in a dose-dependent manner. Collectively, our results suggest that macrophages play a major role in protection against HSV-1-induced CNS demyelination.
To investigate the potential contribution of macrophages in the HSV-IL-2-induced CNS demyelination model, we ocularly infected FoxP3DTR mice with HSV-IL-2 or parental virus in the presence and absence of macrophage depletion. As we had established that demyelination does not occur in the WT HSV-1-infected macrophage-depleted mice until after day 10 post-infection (PI) [20], we analyzed demyelination on day 14 PI. Demyelination was assessed in the optic nerve, spinal cord and brain by luxol fast blue (LFB) staining. Representative photomicrographs of optic nerve, spinal cord and brain are shown in Fig 1. We confirmed that the macrophage-depleted FoxP3DTR mice that were infected with parental virus developed demyelination in the optic nerve (Fig 1A, macrophage depleted), brain (Fig 1B, macrophage depleted), and spinal cord (Fig 1C, macrophage depleted), whereas the macrophage-intact FoxP3DTR mice infected with the parental virus did not develop detectable demyelination in the optic nerve (Fig 1G, no depletion), brain (Fig 1H, no depletion) or spinal cord (Fig 1I, no depletion).
The macrophage-intact FoxP3DTR mice infected with HSV-IL-2 developed demyelination in the optic nerve (Fig 1J, no depletion), brain (Fig 1K, no depletion) and spinal cord (Fig 1L, no depletion), further confirming our previous findings generated using BALB/c, C57BL/6, SJL/6, and 129SVE mice. However, we found that macrophage-depleted FoxP3DTR mice infected with HSV-IL-2 also developed demyelination in the optic nerve (Fig 1D, macrophage depleted), brain (Fig 1E, macrophage depleted), and spinal cord (Fig 1F, macrophage depleted). These results indicated that although macrophages can protect against WT HSV-1 infection-induced CNS demyelination, HSV-IL-2-induced demyelination occurs independently of the presence or absence of macrophages in FoxP3DTR mice. In addition, the demyelinated lesions were detected as both focal and diffuse areas in the white matters of HSV-IL-2-infected mice.
IL-2 is required for the induction of Foxp3 expression and the differentiation of Treg cells in the thymus [44]. We had found previously that the induction of CNS demyelination by WT HSV-1 in macrophage-depleted mice can be blocked by depletion of FoxP3-expressing cells [22]. To determine the contribution of FoxP3, macrophage-intact FoxP3DTR mice were depleted of Foxp3-expressing cells and ocularly infected with HSV-IL-2 or parental virus. In this model, the depletion of Foxp3-expressing cells blocked the HSV-IL-2-induced demyelination in the optic nerve (Fig 2A, FoxP3 depleted), brain (Fig 2B, FoxP3 depleted) and spinal cord (Fig 2C, FoxP3 depleted) of infected mice. The protocol used to deplete the FoxP3-expressing cells did not contribute to demyelination as no demyelination was observed in the optic nerve (Fig 2D, FoxP3 depleted), brain (Fig 2E, FoxP3 depleted) and spinal cord (Fig 2F, FoxP3 depleted) of the macrophage-intact, FoxP3-depleted mice that were infected with parental virus. We then investigated whether FoxP3 can protect against HSV-IL-2-induced CNS demyelination in the absence of macrophages. FoxP3DTR mice were depleted of both macrophages and FoxP3 prior to infection with HSV-IL-2 or parental virus. Surprisingly, HSV-IL-2 infection in the context of concomitant depletion of both FoxP3-expressing cells and macrophages resulted in demyelination in the optic nerve (Fig 2G, FoxP3 and macrophage depleted), brain (Fig 2H, FoxP3 and macrophage depleted), and spinal cord (Fig 2I, FoxP3 and macrophage depleted). Moreover, the level of demyelination was similar to that observed in FoxP3-expressing FoxP3DTR mice infected with HSV-IL-2 in the absence of macrophage depletion. Replication of HSV-IL-2 in the eye of depleted mice was similar to that of parental virus. This suggests that depletion of both macrophages and FoxP3 had no direct effect on virus replication in vivo. Thus, these results indicated a complex interaction, in which the FoxP3-expressing cells contribute to blockade of CNS demyelination in HSV-IL-2 mice in the presence of macrophages, but do not block the demyelination in the absence of macrophages.
Previously, we found that both CD4+ and CD8+ T cells contribute to HSV-IL-2-induced CNS demyelination in macrophage intact mice [20,23] whereas in HSV-1-infected macrophage-depleted mice the demyelination can be blocked by CD4+ T cells alone [22]. To directly address whether T cells contribute to HSV-1-induced demyelination in the absence of macrophages, we depleted wt mice of macrophages and CD4+/CD8+ T cells or mock depleted prior to infection with HSV-IL-2 or parental virus. The HSV-IL-2 infected mice that were depleted of both macrophages and CD4+/CD8+ T cells did not show any signs of demyelination in the optic nerve (Fig 3A, T cells and macrophage depleted), brain (Fig 3B, T cells and macrophage depleted) or spinal cord (Fig 3C, T cells and macrophage depleted). Similarly, no demyelination was detected in mice infected with the parental virus that were depleted of both macrophages and CD4+/CD8+ T cells (Fig 3D, 3E and 3F, parental virus). In contrast, as we reported previously [20], demyelination was detected in mock depleted mice infected with the HSV-IL-2 (Fig 3G, 3H and 3I, mock depleted) but not in mice infected with parental virus (Fig 3J, 3K and 3L, mock depleted). These results suggest that, in the absence of macrophages, depletion of T cells can block CNS demyelination after infection with HSV-IL-2 or WT HSV-1.
Collectively, our results supported the concept that macrophages play a major role in protection against HSV-1-induced CNS demyelination. To begin to identify the potential mechanisms involved, we compared the changes in the levels of mRNA in the brains of mice that were infected ocularly with HSV-IL-2, HSV-IL-4 or parental virus. We determined the mRNA levels of the IL-12 subunit genes (IL-23p19, IL-27p28, IL-35EBI3, IL-12p35, IL-12p40), IL-12 receptor genes (IL-12rβ1, IL-12rβ2, IL-23r, IL-27r, gp130), and markers of immune cells (CD4, CD8, FoxP3, IFN-γ, as well as CD11b and F4/80). In parallel, we determined the mRNA levels for the astrocyte marker gene, glial fibrillary acidic protein (GFAP), and demyelination marker genes (NSE, S-100, MAG, MBP, PLP, MOG). On day 5 PI, RT-PCR was performed on total RNA from individual brain. The levels of each mRNA relative to the baseline seen in the uninfected mouse brain is shown in Fig 4 and a summary of the differences between HSV-IL-2 infected and parental virus-infected mice is provided in Table 1.
The levels of CD11b and F4/80 mRNAs were similar in the macrophage-intact mice infected with HSV-IL-2, parental virus or HSV-IL-4 (Fig 4A). In terms of the IL-12-associated mRNAs, no differences were detected in the brains of mice infected with HSV-IL-2, HSV-IL-4 or parental virus groups in the mRNA levels of the IL-12 receptors (IL-12rβ1, IL-12rβ2); IL-23r, IL-27r, and gp130 (Fig 4B); or the IL-12 subunits IL-23p19, IL-27p28 and IL-35EBI3 mRNAs (Fig 4C). However, the levels of IL-12p35 mRNA, were significantly lower than baseline in the brains of mice infected with parental or HSV-IL-4 virus, but significantly higher than baseline in the IL-HSV-IL-2 infected mice (Fig 4C). Conversely, the levels of IL-12p40 mRNA were significantly higher than baseline in the parental and HSV-IL-4 infected mice but significantly lower than baseline in the HSV-IL-2 infected mice (Fig 4C). We therefore extended these experiments to analysis of the levels of IL-12p35 and IL-12p40 mRNAs in macrophage-depleted mice infected with WT HSV-1. In these mice, the levels of IL-12p35 mRNA were significantly lower than in the macrophage-intact HSV-IL-2-infected mice model but significantly higher than in the macrophage-intact mice infected with parental virus or HSV-IL-4 (Fig 4C).
In terms of the mRNA levels of T cell-associated molecules, the levels of FoxP3 mRNA were similar in the macrophage-intact mice infected with HSV-IL-2, parental virus or HSV-IL-4 (Fig 4A). However, we found that in the brains of macrophage-intact HSV-IL-2-infected mice the levels of CD4 mRNA were significantly lower than in mice infected with parental virus or HSV-IL-4 (Fig 4A) whereas the levels of CD4 mRNA in the brains of macrophage-depleted HSV-IL-2-infected mice were similar to those seen in macrophage-depleted mice infected with parental virus or HSV-IL-4 (Fig 4A). The levels of CD8 mRNA in the macrophage-intact HSV-IL-2-infected mice were significantly lower than those seen in macrophage-intact mice infected with parental virus or HSV-IL-4 (Fig 4D). In the macrophage-depleted mice infected with HSV-IL-2 virus, the levels of CD8 mRNA were similar to the levels in the macrophage-intact mice, and were significantly lower than those seen in macrophage-depleted mice infected with parental virus or HSV-IL-4 (Fig 4D). The levels of IFN-γ mRNA in the macrophage-intact HSV-IL-2 infected mice were similar to the levels in the macrophage-depleted HSV-IL-2 infected mice but significantly lower than parental virus or HSV-IL-4 infected groups, with HSV-IL-4-infected groups having the highest level of IFN-γ expression (Fig 4D).
These results suggest that the presence of IL-2 has a direct effect on the levels of IL-12p35, IL-12p40, CD4, CD8 and IFN-γ mRNAs, while depletion of macrophages affects the levels of IL-12p35, CD8, and IFN-γ mRNAs in the brain on day 5 PI. Therefore, we used the same protocol to determine the levels of these mRNAs in the spinal cord and the brain on day 10 PI in macrophage-intact mice infected with HSV-IL-2 or parental virus and macrophage-depleted mice infected with parental virus. We found that at day 10 PI, the levels of IL-12p40 mRNA in the brains of the mice infected with HSV-IL-2 or parental virus were similar and were higher than in the macrophage-depleted mice infected with parental virus (Fig 5A, brain). In contrast, the levels of IL-12p40 mRNA in the spinal cords of HSV-IL-2 infected mice were significantly lower than those in mice infected with WT parental virus and was similar to that of macrophage-depleted and infected mice (Fig 5A, spinal cord). IL-12p35 mRNA expression was suppressed in HSV-IL-2 infected mice brain compared with parental infected or macrophage-depleted mice (Fig 5B, brain), and similar patterns were observed in spinal cord of infected mice (Fig 5B, spinal cord). CD4 (Fig 5C), CD8 (Fig 5D), and IFN-γ (Fig 5E) mRNAs levels were suppressed in HSV-IL-2 infected mice brain and spinal cords compared with parental-infected or macrophage-depleted mice. In addition, the patterns of CD4 (Fig 5C), CD8 (Fig 5D), and IFN-γ (Fig 5E) mRNAs expression were similar in brain versus spinal cord of infected mice. These results for day 10 PI suggest that HSV-IL-2 has a suppressive effects on IL-12p35, IL-12p40, CD4, CD8, and IFN-γ mRNAs expression and is similar to that of their expression of day 5 PI, while macrophage depletion only affected IL-12p40 mRNA expression level. In summary, our results showed similar mRNA expression profiles for IL-12p40, CD4, CD8, IFN-γ, and GFAP but not IL-12p35 in brain and spinal cord of each group.
The levels of GFAP mRNA were significantly lower in the macrophage-intact HSV-IL-2-infected mice than macrophage-intact mice infected with HSV-IL-4 or parental virus (Fig 4E, GFAP). The levels of GFAP mRNA were similar in the macrophage-depleted mice HSV-IL-2 infected mice to that of the macrophage-depleted mice infected with HSV-IL-4 or parental virus (Fig 4E). The levels of NSE, S-100, MAG, MBP, PLP, and MOG mRNAs were similar in the brains of the HSV-IL-2, HSV-IL-4 and parental virus infected mice (Fig 4E). In the brain of infected mice, GFAP expression was the lowest for HSV-IL-2 infected mice followed by macrophage-depleted and infected mice compared with parental virus (Fig 5F, brain), while GFAP expression was similar between groups in spinal cord of infected mice (Fig 5F, spinal cord).
The qRT-PCR studies described above suggested that an imbalance of IL-12p35 and IL-12p40 may contribute to the HSV-IL-2-induced CNS demyelination. We found previously that both HSV-IL-2-induced demyelination and the demyelination induced by WT HSV-1 in the absence of macrophages can be blocked by either IL-12p70 DNA or HSV-IL-12p70 recombinant virus [22–24]. These data raised the possibility that the IL-12p70 arm of the macrophage response plays a key role in mitigating CNS demyelination. They suggested a hypothetical model in which suppression of macrophage IL-12p35 and IL-12p40 signaling by IL-2 in the macrophage-competent HSV-IL-2 infected mice, and the lack of IL-12p70 due to macrophage depletion play a key role in the CNS demyelination in these models of MS. To test this hypothesis, we used an adoptive transfer strategy in which bone marrow (BM)-derived macrophages infected with different recombinant HSV-IL-12 viruses were transferred into recipients that were subsequently ocularly infected with HSV-IL-2. The macrophages were infected with HSV-IL-12p35, HSV-IL-12p40, HSV-IL-12p70, or parental virus, or mock-infected then were injected intravenously (IV) into female C57BL/6 mice. Two weeks after adoptive transfer of 1 X 106 cells, the recipient mice were infected ocularly with HSV-IL-2. Fourteen days PI, the mice were sacrificed and the optic nerve, spinal cord and brain post-fixed and stained with LFB. The presence or absence of demyelination in each tissue was determined (Table 2). We found that all of the mice that received macrophages infected with HSV-IL-12p35 or WT HSV-1, or macrophages that were mock infected, developed demyelination in the optic nerve, brain or spinal cord. In marked contrast, most of the mice that received macrophages infected with HSV-IL-12p70 or HSV-1L-12p40 were protected from demyelination in the optic nerve, brain and spinal cord. The mice that received macrophages infected with HSV-IL-12p70 showed better protection than the mice that received macrophages infected with HSV-IL-12p40. In addition, the adoptive transfer of macrophages infected with HSV-IL-12p70 or HSV-1L-12p40 resulted in better protection in the optic nerve and brain of the recipient mice than in the spinal cord.
We then repeated the experiment using a higher dose (1 X 107) macrophages infected with HSV-IL-12p70 or HSV-IL-12p40 and, as control, transfer of 1 X 107 DCs infected with HSV-IL-12p70. Representative photomicrographs are shown in Fig 6 and a summary of the results is provided in Table 2. No demyelination was detected in optic nerve, brain and spinal cord sections of mice that received 1 X 107 macrophages infected with HSV-IL-12p70 (Fig 6, Table 2) or HSV-IL-12p40, whereas demyelination occurred in the mice that received HSV-IL-12p70-infected DCs or mock-infected macrophages prior to infection (Fig 6, Table 2).
Image analysis of the stained tissue sections suggested that the extent of demyelination differed amongst the experimental groups and the CNS tissue. In those mice that received 1 X 106 macrophages, the area of demyelination in the brains of mice was significantly larger in the mice that received HSV-IL-12p35-infected macrophages or mock-infected macrophages than the area of demyelination in the mice that received HSV-IL-12p40- or HSV-IL-12p70-infected macrophages (Fig 7, Brain). Moreover, the area of demyelination in the brains of mice that received HSV-IL-12p70-infected macrophages was lower than the area of demyelination in the brains of mice that received HSV-IL-12p40-infected macrophages (Fig 7, Brain). As described above, no demyelination was detected in brains of mice that received 1 X 107 HSV-IL-12p70-infected macrophages (Fig 7, Brain, Arrow: no demyelination). In those mice that received 1 X 106 macrophages, the area of demyelination was somewhat higher in the spinal cords of mice that received HSV-IL-12p35-infected macrophages than the area of demyelination in the spinal cord of mice that received mock-infected macrophages (Fig 7, Spinal cord). The level of demyelination in the spinal cord of mice that received HSV-IL-12p40-infected macrophages were similar to the level of demyelination in the spinal cord in mice that received HSV-IL-12p70-infected macrophages (Fig 7, Spinal cord). In contrast to the spinal cords, the area of demyelination in the optic nerves was somewhat lower in recipient mice that received HSV-IL-12p35-infected macrophages than the area of demyelination in the optic nerve of mice that received mock-infected macrophages (Fig 7, Optic nerve). However, the level of demyelination in the optic nerves of mice that received HSV-IL-12p40-infected virus were similar to the levels of demyelination in the optic nerves of mice that received HSV-IL-12p70-infected macrophages (Fig 7, Optic nerve). As described above, no demyelination was detected in spinal cord and optic nerve of mice that received 1 X 107 macrophages infected with HSV-IL-12p70 virus (Fig 7, Spinal cord, Optic nerve, Arrow: no demyelination).
We reported previously that CNS demyelination occurred following ocular infection of mice with HSV-IL-2 virus, while WT viruses, HSV-IFN-γ or HSV-IL-4 did not induce CNS demyelination [19,20,22–24]. In addition, severity of CNS demyelination in HSV-IL-2-infected mice was sex-dependent [20]. Similar results were reported for MS patients [45] and EAE model of MS [46]. Thus, in this study we used female mice for all our experiments. We also have shown that macrophages, but not B cells, DCs, NK cells or T cells, mediated self-tolerance and protection against autoimmunity following ocular infection with WT HSV-1 in a manner similar to that of HSV-IL-2 in HSV-1-infected mice [22–24]. Previously we have shown that macrophages play a significant role in blocking CND demyelination in mice infected ocularly with WT HSV-1 [22].
In the current study, we show an imbalance of IL-12p35 and IL-12p40 mRNA levels in the CNS of macrophage-intact HSV-IL-2-infected mice on days 5 and 10 PI that does not occur in parental virus-infected mice. A similar imbalance was observed in macrophage-depleted mice infected with parental virus. The results suggested that this effect was specific for these IL-12 subunits, as the transcription of other members of the IL-12 family that are involved in formation of IL-23, IL-27, and IL-35 were not affected. These results are is similar with our previous findings that IL-12p35–/–and IL-12p40–/–mice developed CNS pathology following ocular HSV-1 infection with WT viruses and that this demyelination did not occur when each knockout strain was reconstituted with its missing gene [22]. Moreover, this pathology was not detected in HSV-1-infected IL-23p19–/–mice or in EBI3–/–mice. Our previous data and the present study suggest that both p35 and p40 subunits of IL-12 are required for protection from CNS demyelination. Although the accumulation of macrophages around demyelination plaques suggests that they may play a pathologic role [47], it is possible that their accumulation simply reflects their phagocytic function. Our current results demonstrate that macrophages, through their production of IL-12p70, play a central role in protection from virus-induced demyelinating immunopathology.
Injection of mice with IL-12p70 DNA prevented development of CNS demyelination in macrophage-depleted mice. However, IL-12p35 or IL-12p40 DNA alone, or together had no protective effect on prevention of CNS demyelination in WT macrophage-depleted mice, indicating that control of CNS demyelination is dependent on the IL-12p70 heterodimer [22]. Similarly, we have shown previously that demyelination induced by HSV-IL-2 can be blocked by either injection of IL-12p70 DNA or a recombinant HSV-1 expressing IL-12p70 [23,24]. In the current study, we have shown that transfer of macrophages infected with HSV-1 recombinant virus expressing IL-12p70 or IL-12p40, but not IL-12p35 protected HSV-IL-2 infected mice from CNS demyelination in a dose-dependent manner. Higher demyelination in macrophages infected with IL-12p35 may be due to higher expression of IL-7 by microglia and macrophages. Previously it was reported that IL-12p35, but not IL-12p40, subunit of IL-12p70 is involved in the induction of IL-7 in microglia and macrophages [48]. Furthermore, increase of IL-7 expression were reported for individual with MS and in EAE model of MS [49,50]. In contrast, DCs infected with IL-12p70 or mock-infected macrophages did not block demyelination. Previously, we have shown that both macrophages and DCs can be infected with HSV-1 but the virus does not replicate and does not increase apoptosis or cell death in infected macrophages or DCs [51]. Thus, our results suggest that macrophages carrying IL-12p70, but not DCs or macrophages without IL-12p70, can compensate for the suppressive effects of IL-2 on the IL-12p70 components.
In the current study, we found that the effects of depletion of FoxP3-expressing cells on demyelination was highly dependent on the experimental model. In HSV-IL-2-infected mice, depletion of FoxP3-expressing cells blocked demyelination in mice, whereas depletion of macrophages as well as FoxP3-expressing cells did not block demyelination. In macrophage-depleted parental HSV-1-infected mice, demyelination was blocked following depletion of FoxP3-expressing cells. Previously, we had found that in WT HSV-1-infected mice, the absence of CD25 also blocked demyelination in macrophage-depleted mice [22] but not in the HSV-IL-2 model of demyelination [23]. These results are consistent with the reports by other investigators that depletion of Treg cells can result in enhanced immune responses against some infectious agents [52] and that Treg cells can enhance tissue damage and autoimmunity [53–58]. The reports that IL-2 can expand and induce Treg cells in vivo [59] and in vitro [60], are in line with our present study showing that depletion of FoxP3-expressing cells blocked CNS demyelination by HSV-IL-2 and required both IL-2 and viral infection. Thus, IL-2 can modulate effector and Treg cell function in the presence of HSV-1 infection. However, the absence of demyelination in the mice that were depleted of FoxP3-expressing cells infected with HSV-IL-2 was dependent on the presence of macrophages. In the mice that were depleted of FoxP3-expressing cells and macrophages and infected with HSV-IL-2, the depletion of both CD4 and CD8 was required for blocking demyelination independent of CD25. In contrast, as we reported previously demyelination in the absence of macrophages in mice infected with WT virus can be blocked by the absence of FoxP3, CD4, or CD25 [22]. In the HSV-IL-2 infected mice that were depleted of their macrophages but not in macrophage depleted mice infected with parental virus, the level of both CD4 and CD8 expression were reduced significantly. Thus, IL-2 signaling may be involved with contraction of T-cell responses in the HSV-IL-2 infected mice. Previously, it was shown that IL-2 signaling enhances susceptibility of T cells to apoptosis [61]. In addition, IL-2 impairs T follicular helper (Tfh) cells [62], while enhancing induced Treg (iTreg) [60].
We found previously that between days 3 and 7 PI, HSV-IL-2-infected mice exhibit a mixed TH1 + TH2 response, whereas mice infected with HSV-IFN-γ exhibit a TH1 response [19]. Similarly, we have shown that mice infected with HSV-IL-2 had an imbalance of TH1/Tc1 cytokines as compared with WT HSV-1 or recombinant viruses expressing IL-4 or IFN-γ [19]. In the present study, the levels of IFN-γ were significantly reduced in the brain of HSV-IL-2 infected mice as well as in macrophage-depleted mice. Thus, these data suggest that suppression of IL-12p70 formation by combination of IL-2 and HSV-1 infection shifts the immune response from a TH1 response, which could promote T cell autoreactivity and induction of demyelination. Surprisingly, the levels of both CD4+ and CD8+ T cells in HSV-IL-2 infected mice were reduced as compared with parental and HSV-IL-4 viruses. This reduction of T cells in the CNS of HSV-IL-2 infected mice may have been accompanied by a skewing of the population from protective T cells to pathogenic T cells. With regards to macrophage-depleted and HSV-1 infected mice, we found that depletion of macrophages affected the CD8+ but not CD4+ T cells. Thus, this imbalance of T cells also may be responsible for generation of pathogenic CD4+ T cells in macrophage-depleted mice that were infected with WT HSV-1.
Previously, we reported that depletion of macrophages enhanced infection of GFAP+ astrocytes in the spinal cords of HSV-1 infected mice as compared to mock-depleted mice [22]. Image analyses of HSV-1-infected mice revealed a significantly higher GFAP burden in the spinal cord white matter and grey matter of macrophage-depleted vs. mock-depleted mice. In contrast, in the present study we observed a significant suppression of GFAP mRNA expression in the brains of HSV-IL-2-infected mice but not in the brains of other groups. The suppression of GFAP mRNA in the HSV-IL-2-infected mice on day 5 PI suggests that the astrocytes are not activated. However, by day 10 post infection GFAP mRNA expression was significantly lower in the brains of HSV-IL-2-infected and macrophage-depleted mice compared with parental virus, while its expression in the spinal cords of all group was similar but lower than on day 5 PI. The discrepancy, between our two models of demyelination with regards to the level of GFAP mRNA expression could be due the presence of IL-2 in our HSV-IL-2 model of demyelination. Similar to the results of our current study, varicella zoster virus (VZV) infection has been shown to downregulate GFAP mRNA expression in vitro [63]. Additionally, it was reported that loss of astrocytes occurs before that of CNS demyelination [64]. In this report, we demonstrate that GFAP expression was significantly affected by HSV-IL-2 infection or after depletion of macrophages and infection with WT HSV-1. This suggests that a relationship exists between astrocytes and IL-2 or astrocytes and macrophages that control CNS pathology. Our results are supported by recent evidence that interruption of astrocyte function exacerbates pathogenesis of CNS diseases [65]. We propose that suppression of GFAP on astrocytes in the absence of macrophages or in the presence of IL-2 following infection with HSV-1 affect IL-12p70 expression thus leading to autoreactivity of T cells and thus CNS demyelination. However, this suppressive effect can be reversed by IL-12p70 or IL-12p40. In contrast to this study, IL-2 treatment has been reported to increase GFAP expression and induce inflammation and macrophage infiltration [66]. The discrepancy between our results and this study is probably due to the HSV-1 infection in our study.
Despite the presence of demyelinated plaques in the CNS of HSV-IL-2 infected mice, no significant change was observed in the mRNA levels of the demyelination marker genes (NSE, S100, MAG, MBP, PLP, MOG). These results suggest that the effects of HSV-IL-2 on demyelination are not executed at the transcriptional level. Any of these genes alone or in combination have been associated with degradation of myelin by activated T cells in the CNS of infected mice. Recently, we compared MOG35–55, MBP35–47, and PLP190–209 induced models of EAE with our HSV-IL-2-induced MS model [67]. CNS pathology in MOG treated mice was similar to that of HSV-IL-2 treated mice but both were different from MBP or PLP injected mice. The similarity of our HSV-IL-2 model of demyelination to the MOG-induced model of demyelination may suggest that HSV-IL-2 autoreactive T cells affect the MOG component of myelin. However, the contributions of other members of myelin, such as MBP and PLP alone or in combination, to CNS demyelination in our model cannot be ruled out.
In summary, our results suggest that suppression of IL-12p70 causing the FoxP3+ T cells to become autoreactive leading to demyelination of the CNS in the infected mice. However, in contrast to our study, previous study found that the absence of the Foxp3+ T cells causing autoimmunity in both humans and mice [68–70]. We feel that these contrasting results most likely stem from the infection of HSV-1 in the presence of IL-2 over-expression or in the absence of macrophages in our two models of MS.
All animal procedures were performed in strict accordance with the Association for Research in Vision and Ophthalmology Statement for the Use of Animals in Ophthalmic and Vision Research and the NIH Guide for the Care and Use of Laboratory Animals (ISBN 0-309-05377-3). Animal research protocol was approved by the Institutional Animal Care and Use Committee of Cedars-Sinai Medical Center (Protocols #2841 and 6134).
Plaque-purified HSV-IL-2, HSV-IL-4, dLAT2903, HSV-IL-12p35, HSV-IL-12p40, and HSV-IL-12p70 were grown in rabbit skin (RS) cell monolayers in minimal essential medium (MEM) containing 5% fetal calf serum (FCS) as we described previously [19,20,22–24,71,72]. dLAT2903 is the parental virus for HSV-IL-2, HSV-IL-4, HSV-IL-12p35, and HSV-IL-12p40 and is referred to as parental virus. Female C57BL/6 mice of 6 weeks of age were purchased from the Jackson Laboratory (Bar Harbor, ME). C57BL/6-FoxP3DTR mice were a gift from Alexander Y. Rudensky (Memorial Sloan Kettering Cancer Center, New York) and were bred in the Animal Facility at the Cedars-Sinai Medical Center and we only used female C57BL/6-FoxP3DTR mice for this study.
As we described previously [19,20,22–24,71,72], mice were infected ocularly in both eyes with 2 x 105 PFU per eye for each virus. Each virus was resuspended in 2 μl of tissue culture media and administered as an eye drop. No corneal scarification was performed prior to infection. No behavioral changes were observed between infected animals.
Liposome-encapsulation of dichloromethylene diphosphonate (Cl2MDP) was purchased (ClodronateLiposomes.org, Netherland) and depletions were carried out as we described previously [22,73]. Briefly, mice were injected twice with 100 μl of the mixture, once intraperitoneally (i.p.) and once subcutaneously (s.c.), on days -5, -2, +1, +4, and +7 relative to ocular infection with HSV-1.
Optic nerves, brains, and spinal cords of experimental and control mice were removed at necropsy on day 14 PI, embedded in OCT (Tissue-Tek, Sakura Finetek, Torrance, CA) for cryosectioning, and stored at -80°C as we described previously [20].
Transverse sections of ONs, brains, and SCs, 10 μm thick (spaced 50μm apart), were prepared using a Leica CM3050S cryostat, air-dried overnight, and fixed in acetone for 3 min at 25°C [74]. The presence or absence of demyelination in infected mice was evaluated using LFB (Sigma-Aldrich) staining of formalin-fixed sections of CNS as we described previously [19,20,22–24]. Demyelination in each section was confirmed by monitoring adjacent sections.
Female C57BL/6-FoxP3DTR mice were depleted of FoxP3 by treatment with diphtheria toxin (DT) (Sigma-Aldrich, Saint Louis, MO) as described previously [22,75]. Briefly, the mice were administered DT at 72 and 24 h before ocular infection, followed by four additional treatments on days +1, +3, +5, and +7 PI. Efficiency of FoxP3 depletion in spleens were monitored by flow cytometry analysis before ocular infection and 7 days after ocular infection. After three depletions, more than 97% of FoxP3+ T cells were depleted.
Each mouse received an i.p. injection of 100 μg of purified GK1.5 (anti-CD4) and 100 μg of 2.43 (anti-CD8) monoclonal antibodies (NCCC, Minneapolis, MN) in 100 μl of PBS, -5 and -2 days before ocular infection as we described previously [23]. The injections were then repeated on days +1, +4, +7, and +10 relative to ocular infection. Control mice were depleted using an irrelevant monoclonal antibody of the same isotype. The efficiency of CD4+ and CD8+ T-cell depletion was monitored by flow cytometry of splenocytes 24 h after the second depletion and before ocular infection. After the second depletion, more than 95% of CD4+ T cells and CD8+ T cells were depleted from spleen.
Brain and spinal cord from individual mice were collected on day 5 or 10 PI, immersed in RNAlater RNA stabilization reagent (Qiagen, Valencia, CA) and stored at -80°C until processing as we described previously [73,76]. The mRNA expression levels of IL-23p19, IL-27p28, IL-35EBI3, IL-12p35, IL-12p40, GFAP, NSE, S-100, MAG, MBP, PLP, MOG, IL-12rβ1, IL-12rβ2, IL-23r, IL-27r, gp130, CD4, FoxP3, CD11b, F4/80, CD8, and IFN-γ were determined using commercially available TaqMan Gene Expression assays (Applied Biosystems) with optimized primers as described below. In all experiments, GAPDH was used for normalization of transcripts. In this study we looked at the mRNA expression level and not protein expression.
Primer probe sets consisted of two unlabeled PCR primers and the FAM dye-labeled TaqMan MGB probe formulated into a single mixture. All cellular amplicons included an intron-exon junction to eliminate signal from genomic DNA contamination. The assays used in this study were as follows: 1) IL-23 p19, ABI assay I.D. Mm00518984_m1—amplicon length = 61 bp, 2) IL-27 p28, ABI assay I.D. Mm00461164_m1—amplicon length = 69 bp, 3) IL-35 Ebi3, ABI assay I.D. Mm00469294_m1—amplicon length = 123 bp, 4) IL-12 p35, ABI assay I.D. Mm00434165_m1—amplicon length = 68 bp, 5) IL-12 p40, ABI assay I.D. Mm01288990_m1—amplicon length = 105 bp, 6) GFAP, ABI assay I.D. Mm01253033_m1—amplicon length = 75 bp, 7) NSE, ABI assay I.D. Mm00469062_m1—amplicon length = 76 bp, 8) S-100, Mm00485897_m1—amplicon length = 69 bp, 8) MAG, ABI assay I.D. Mm00487538_m1—amplicon length = 94 bp, 9) MBP, ABI assay I.D Mm01262035_m1—amplicon length = 83 bp, 10) PLP, ABI assay I.D. Mm00456892_m1—amplicon length = 67 bp, 11) MOG, ABI assay I.D. Mm00447824_m1—amplicon length = 93 bp, 12) IL-12rβ1, ABI assay I.D. Mm00434189_m1—amplicon length = 60 bp, 13) IL-12rβ2, ABI assay I.D. Mm00434200_m1—amplicon length = 74 bp, 14) IL-23r, ABI assay I.D. Mm00519943_m1—amplicon length = 72 bp, 15) IL-27r, ABI assay I.D. Mm00497259_m1—amplicon length = 69 bp, 16) gp130, ABI assay I.D. Mm00439668_m1—amplicon length = 89 bp, 17) CD4, ABI assay I.D. Mm00442754_m1—amplicon length = 78 bp,18) FoxP3, ABI assay I.D. Mm00475164_m1—amplicon length = 80 bp, 19) CD11b, ABI assay I.D. Mm00434455_m1—amplicon length = 69 bp, 20) F4/80, ABI assay I.D. Mm00802529_m1—amplicon length = 92 bp, 21) CD8, ABI assay I.D. Mm01182108_m1—amplicon length = 67 bp, 22) IFN-γ, ABI assay I.D. Mm00801778_m1—amplicon length = 101 bp, and 23) GAPDH, ABI assay I.D. Mm999999.15_G1 –amplicon length = 107 bp. Additionally, a custom-made primer and probe set was used for LAT as follows: forward primer, 5'-GGGTGGGCTCGTGTTACAG-3'; reverse primer, 5'-GGACGGGTAAGTAACAGAGTCTCTA-3'; and probe, 5'- FAM-ACACCAGCCCGTTCTTT-3'–Amplicon Length = 81 bp. Quantitative real-time PCR (qRT-PCR) was performed using an ABI ViiA 7 Sequence Detection System (Applied Biosystems) in 384-well plates as we described previously [73,76].
Six-week-old female C57BL/6 mice were used as a source of bone marrow (BM) for the generation of mouse DCs and macrophages in culture as we described previously [51]. Briefly, BM cells were isolated by flushing femurs and tibiae with PBS. Pelleted cells were resuspended briefly in water to lyse red blood cells and stabilized by adding complete medium (RPMI 1640, 10% fetal bovine serum, 100 U/ml penicillin, 100 μg/ml streptomycin, 2 mM L-glutamine). The cells were centrifuged and resuspended in complete medium supplemented with murine CSF1 (100 ng/ml; Peprotech, Rocky Hill, NJ) to grow macrophages, while to grow DCs the media was supplemented with murine GM-CSF (100 ng/ml; Peprotech). The cells were plated in non-tissue culture plastic petri dishes (1 bone per 10 cm dish) for 5 d at 37°C with CO2. After 5 d, the media was removed, and the adherent cells recovered by incubating the cells for 5 min at 37°C with Versene (Invitrogen, San Diego, CA). Cells were washed, counted, and plated onto tissue culture dishes for use the following day.
Monolayers of macrophages were infected with 1 PFU/cell of dLAT2903, HSV-IL-12p35, HSV-IL-12p40, or HSV-IL-12p70, and monolayers of DCs were infected with 1 PFU/cell of HSV-IL-12p70. One hour after infection at 37°C, virus was removed and the infected cells were washed three times and fresh media was added to each well. The monolayers at 24 h PI were harvested, washed, and counted for subsequent adoptive transfer experiments.
Each recipient female C57BL/6 mouse was injected once intravenously (i.v.) with 1 X 106 or 1 X 107 infected macrophages in 100 μl of MEM. Similarly control mice received uninfected macrophages or infected DCs. Recipient mice were ocularly infected two weeks after transfer with HSV-IL-2 virus.
The numbers of plaques and size of plaques on multiple LFB stained fields were evaluated in a blind fashion for each treatment group by inspection of serial sections of CNS tissues. The amount of myelin loss in the stained sections of brains, SCs and ONs was measured using NIH Image J software as described previously [67]. The areas of demyelination (clear-white) to normal tissue (blue) were quantified using 75 random sections from the brain and SCs or 30 sections from ONs of each animal. Demyelination in each section was confirmed by monitoring adjacent sections. The percentage of myelin loss was calculated by dividing the lesion size by the total area for each section.
Level of demyelination in experimental and control groups were compared using Fisher’s exact tests. Student’s t test was performed for comparison of means of differences using Instat (GraphPad, San Diego). Results were considered statistically significant when the "P" value was <0.05.
|
10.1371/journal.pcbi.1000858 | Categorial Compositionality: A Category Theory Explanation for the Systematicity of Human Cognition | Classical and Connectionist theories of cognitive architecture seek to explain systematicity (i.e., the property of human cognition whereby cognitive capacity comes in groups of related behaviours) as a consequence of syntactically and functionally compositional representations, respectively. However, both theories depend on ad hoc assumptions to exclude specific instances of these forms of compositionality (e.g. grammars, networks) that do not account for systematicity. By analogy with the Ptolemaic (i.e. geocentric) theory of planetary motion, although either theory can be made to be consistent with the data, both nonetheless fail to fully explain it. Category theory, a branch of mathematics, provides an alternative explanation based on the formal concept of adjunction, which relates a pair of structure-preserving maps, called functors. A functor generalizes the notion of a map between representational states to include a map between state transformations (or processes). In a formal sense, systematicity is a necessary consequence of a higher-order theory of cognitive architecture, in contrast to the first-order theories derived from Classicism or Connectionism. Category theory offers a re-conceptualization for cognitive science, analogous to the one that Copernicus provided for astronomy, where representational states are no longer the center of the cognitive universe—replaced by the relationships between the maps that transform them.
| Our minds are not the sum of some arbitrary collection of mental abilities. Instead, our mental abilities come in groups of related behaviours. This property of human cognition has substantial biological advantage in that the benefits afforded by a cognitive behaviour transfer to a related situation without any of the cost that came with acquiring that behaviour in the first place. The problem of systematicity is to explain why our mental abilities are organized this way. Cognitive scientists, however, have been unable to agree on a satisfactory explanation. Existing theories cannot explain systematicity without some overly strong assumptions. We provide a new explanation based on a mathematical theory of structure called Category Theory. The key difference between our explanation and previous ones is that systematicity emerges as a natural consequence of structural relationships between cognitive processes, rather than relying on the specific details of the cognitive representations on which those processes operate, and without relying on overly strong assumptions.
| For more than two decades, since Fodor and Pylyshyn's seminal paper on the foundations of a theory of cognitive architecture (i.e., roughly, the component processes and their modes of composition that together comprise cognitive behaviour) [1], the problem of explaining systematicity has remained unresolved [2] despite numerous Classicist and Connectionist attempts [3]–[7]. In general terms, the problem of systematicity for a theory of cognition is to explain why various cognitive abilities are intrinsically connected in the sense that the capacity to exhibit some abilities is indivisibly linked to the capacity to exhibit some other related abilities. Why, for example, is it the case that if one has the ability to infer that John is the lover from John loves Mary, then one also has the ability to infer that Mary is the lover from Mary loves John, where both abilities involve a common relation, loves? That is to ask, in general: what is it about our cognitive system that necessitates a particular group-oriented distribution of cognitive capacities, whereby you don't find people with the capacity for some but not all the behaviours pertaining to the same group (excluding, of course, individuals who lack a particular capacity for reasons clearly unrelated to normal development, because of brain damage for example)? Although the debate over what systematicity implies for a theory of cognition has many aspects (see [2]), the generally accepted common ground is that: systematicity is a property of some (though not all) components of human cognition; a complete theory of human cognitive architecture must include an explanation for this property; and no theory of cognition has a satisfactory explanation for it. In the remainder of this section, we outline the systematicity property and the main problem it still poses for existing theories, what is required for a theory to explain it, and how our approach meets those requirements.
The systematicity problem consists of three component problems:
These problems are logically independent—one does not necessarily follow from another [8], and so a theory is required to explain all three, though for some theories an explanation for one property may entail explanations for others.
Classicists and Connectionists employ some form of combinatorial representations to explain systematicity. For Classicists, representations are combined in such a way that the tokening of complex representations entails the tokening of representations of their constituent entities, so that the syntactic relationships between the constituent representations mirror the semantics ones—systematicity is a result of a combinatorial syntax and semantics [1]. For Connectionists, representations of complex entities are constructed more generally so that their tokening does not necessarily imply tokening constituent entity representations [5], [6]. An example of a Classicist's representation of John loves Mary would be loves (John, Mary), and a Connectionist representation would be a tensor product so that the vectors representing John, loves, and Mary do not literally appear anywhere in the tensor representation. We refer to the former as classical compositionality, and the latter as connectionist (or, functional) compositionality.
In general, a Classical or Connectionist architecture can demonstrate systematicity by having the “right” collection of grammatical rules, or functions such that one capacity is indivisibly linked to another. Suppose, for example, a Classical system with the following three rules:
G1:
G1 provides the capacities to generate all four representations (i.e., John loves John, John loves Mary, etc.), and these capacities are indivisibly linked, because absence of any one of those rules means the system cannot generate any of those representations. In no case can the system generate one without being able to generate the others. So, this Classical architecture has the systematicity of representation property with respect to this group of four propositions. A tensor product [9], or Gödel numbering [5] scheme is a functionally compositional analogue of this explanation. Systematicity of inference follows from having additional processes that are sensitive to the structure of these representations. For Classical architectures, at least, compositionality of representation also follows, because the semantic content of a complex representation is built up from the semantic contents of the constituents and their syntactic relationships [8]. Aizawa [2], [8] disputes whether a Connectionist architecture can also demonstrate compositionality of representation. Regardless, though, neither Classicism, nor Connectionism can derive theories that provide a full account of systematicity [2].
A demonstration of systematicity is not an explanation for it. In particular, although grammar G1 has the systematicity of representation property, the following grammar:
G2:does not. This architecture cannot generate a representation of the proposition Mary loves John even though it can generate representations of both John and Mary as agents and patients, and the John loves Mary proposition. The essential problem for Classical theory—likewise Connectionist theory—is that syntactic compositionality by itself is not sufficient without some additional assumptions for admitting grammars such as G1 that have the systematicity property, while excluding grammars such as G2 that do not. An explanation for systematicity in these cases turns on the nature of those additional, possibly ad hoc assumptions.
To further clarify what is required of a theory to explain systematicity [1], [3], Aizawa [2] presents an explanatory standard for systematicity and the problem of ad hoc assumptions, which we follow, by analogy with the Ptolemean (geocentric) versus Copernican (heliocentric) explanations for the motions of the planets (see [10] for a review). The geocentric explanation for planetary motion places the Earth at the center of the other planets' circular orbits. Although this theory can roughly predict planetary position, it fails to predict periods of apparent retrograde motion for the superior planets (i.e. Mars, Jupiter, etc.) across the night sky without the assumption of epicycles (i.e., circular orbits with centers that orbit the Earth). This additional assumption is ad hoc in that it is unconnected with the rest of the theory and motivated only by the need to fit the data—the assumption could not be confirmed independently of confirming the theory. The heliocentric explanation, having all planets move around the Sun, eschews this ad hoc assumption. Retrograde motion falls out as a natural consequence of the positions of the Earth and other planets relative to the Sun. Tellingly, as more accurate data became available, the geocentric theory had to be further augmented with epicycles on epicycles to account for planetary motion; not so for the heliocentric theory.
The theory of planetary motion, of course, does not end there. The heliocentric theory, with its circular orbits, cannot explain the elliptical motion of the planets without further assumptions, and so was superseded by Newtonian mechanics. Newtonian mechanics cannot explain the precession of planetary orbits, and was in turn superseded by Einstein's theory of relativity. In each case, the superseding theory incorporates all that was explained by the preceding theory. Evaluating competing theories in this manner has an extensive history in science, and so one may expect it to be a reasonable standard for an explanation of systematicity in cognitive science.
Aizawa [2] notes that although philosophers of science may not have a precise definition for the concept of an ad hoc assumption, one can nonetheless usefully characterize the idea by analogy with generally accepted examples, such as the assumption of epicycles, which we just mentioned. Another example Aizawa uses is the Creationist versus Darwinian theory of speciation, where the appeal to a supernatural being to explain the existence of different species is an ad hoc assumption. The general sense in which a theory fails to provide a satisfactory explanation by its appeal to ad hoc assumptions is when those additional, so called auxiliary, assumptions are unconnected to the core assumptions and principles of the theory, motivated only by the need to fit the data, and cannot be confirmed independently of confirming the theory. In this sense, the core theory has no explanatory power for the particular phenomenon of interest. Note that an auxiliary assumption is not necessarily ad hoc, nor is it precluded from subsequent inclusion into the set of core assumptions of the modified theory. Orthogonal experiments may provide confirmatory data for an auxiliary assumption, independent of the theory in question. Observations of the Jovian moons would have been the sort of independent confirmatory evidence for epicycles, had such data been available at the time, to justifiably include it as one of the core assumptions. However, the assumption that all heavenly bodies are governed this way ultimately proved untenable. The kind of theory sought here is one where systematicity necessarily follows without requiring such ad hoc assumptions. This characterization guides our analysis of the problem posed by the systematicity property, and our explanation for it.
The problem for Classical and Connectionist theories is that they cannot explain systematicity without recourse to their own ad hoc assumptions [2]. For Classicism, having a combinatorial syntax and semantics does not differentiate between grammars such as G1 and G2. For Connectionism, a common recourse to learning also does not work, whereby systematicity is acquired by adjusting network parameters (e.g., connection weights) to realize some behaviours—training set—while generalizing to others—test set. Learning also requires ad hoc assumptions, because even widely used learning models, such as feedforward [11] and simple recurrent networks [12], fail to achieve systematicity [13]–[17] when construed as a degree of generalization [18], [19]. Hence, neither Classical nor Connectionist proposals satisfy the explanatory standard laid out by Fodor and Pylyshyn [1] and Fodor and McLaughlin [3] (see also [20], Appendix), and further articulated by Aizawa [2]. Ironically, failure to meet this criterion was one of the reasons Classicists rejected Connectionist explanations for systematicity. The import of Aizawa's analysis is that the same shortcoming also befalls Classicism, and so an explanation for systematicity is still needed. In this regard, it would appear that the 90s were also the “lost decade” for cognitive science.
In hindsight, the root of the difficulty that surrounds the systematicity problem has been that cognitive scientists never had a theory of structure to start with (i.e. one that was divorced, or at least separated from specific implementations of structure-sensitive processes). In fact, such a theory has been available for quite some time, but its relevance to one of the foundational problems of cognitive science has not previously been realized. Our category-theory based approach addresses the problem of ad hoc assumptions because the concept of an adjunction, which is central to our argument, ensures that the construct we seek not only exists, but is unique. That is to say, from this core assumption and category theory principles, the systematicity property necessarily follows for the particular cognitive domains of interest, because in each case the one and only collection of cognitive capacities derived from our theory is the systematic collection, without further restriction by additional (ad hoc) assumptions.
Category theory is a theory of structure par excellence [21]–[23]. It was developed out of a need to formalize commonalities between various mathematical structures [24], and has been used extensively in computer science for the analysis of computation [25]–[28]. Yet, despite computationalism being the catchcry of many psychologists since the cognitive revolution, applications of category theory to cognitive psychology have been almost non-existent (but, see [29], [30] for two examples). Our explanation of systematicity is based on the concept of an adjunction, which depends on the concepts of category, morphism, product, functor, and natural transformation. So, in this section, we provide formal definitions of these concepts. (For further explanation of some category theory concepts in the context of cognition, see [30].)
An adjunction is a formal means for capturing the intuition that a relationship between mathematical objects is “natural”—additional constructs are unnecessary to establish that relationship (see also [23], p2). The mathematical notion of being natural dates back at least to [24], and the technical aspect is given starting where we define natural transformation. In the current context of meeting the explanatory standard for systematicity, identifying a suitable adjunction means that no further (ad hoc, or arbitrary) assumptions are needed to define the relationship between a particular cognitive architecture and a desired group of cognitive capacities. Such constructs look natural (once understood), but it is the mathematical criterion that definitely establishes naturality.
A category consists of a class of objects ; a set of morphisms (also called arrows, or maps) from to where each morphism has as its domain and as its codomain, including the identity morphism for each object ; and a composition operation, denoted “”, of morphisms and , written that satisfy the laws of:
The most familiar example of a category is , which has sets for objects and functions for morphisms, where the identity morphism is the identity function and the composition operation is the usual function composition operator “”. Another example, where continuity is important, is the category of metric spaces and continuous functions.
Certain morphisms have important properties that warrant giving them names. Two such morphisms, which we will refer to later, are called isomorphisms and homomorphisms. A morphism is an isomorphism if there exists a morphism , such that and . If exists, then it is the inverse of , also denoted as .
Homomorphisms pertain to categories whose objects have additional internal structure, such as groups. For example, the category has groups for objects, and the morphisms are group homomorphisms. A group consists of a set of elements, and an associative binary operation , satisfying identity and inverse axioms. That is, has an identity element , and for each , an inverse element , such that and . A group homomorphism is a morphism , such that , for all . Homomorphisms in other categories (e.g., graph homomorphisms) are defined analogously.
A product of two objects and in a category is an object together with two morphisms and , such that for any pair of morphisms and , there is a unique morphism , such that the following diagram commutes:(1)where a broken arrow indicates that there exists exactly one morphism making the diagram commute. To say that a diagram commutes is to mean that the compositions along any two paths with the same start object and the same finish object are the same. So, in this diagram, and , where and are sometimes called projection morphisms. A product object is unique up to a unique isomorphism. That is, for any other product object with morphisms and there is one and only one isomorphism between and that makes a diagram like this one commute. Hence, is not unique, only unique with respect to another product object via isomorphism. This characteristic has an important consequence for our explanation of systematicity, which we present in the Results section. An essential characteristic of a product object is that the constituents and are retrievable via the projection morphisms. is also written , and since is uniquely determined by and , is often written as , and the diagram used in defining a product then becomes(2)
In , is (up to isomorphism) the Cartesian product (, , ), where , , and is the product function , sending to , so that and . The “maps to” arrow, , indicates the action of a function on a domain element, so is equivalent to . ( refers both to a general product in any category with products and the more specific Cartesian product in the category .)
The categorical concept of product is a very general notion of combinatoriality. Not surprisingly, then, Classical and Connectionist notions of combinatoriality can be seen as special cases of categorical products. A grammar like G1 (Introduction), for instance, can be used to realize the Cartesian product of the set of agents and the set of patients (i.e. by employing the first production without the loves symbol). A categorical product can also be realized by including suitable rules for inferring the agent and patient from this Cartesian product. (A grammar like G2 cannot realize a Cartesian product, or categorical product; in fact, it realizes a union of two partial products.) Similarly, a Connectionist method such as the outer product of two vector spaces with suitable projections from the outer product space to the original vector spaces also realizes a categorical product. However, an explanation for systematicity requires more than just realization, and as we shall see, additional category theory concepts are needed.
A functor is a structure-preserving map between categories and that associates each object in to an object in ; and each morphism in to a morphism in , such that for each object in ; and for all morphisms and for which compositions and are defined in categories and , respectively. The following diagram shows the details of a functor:(3)where dashed rectangles encapsulate the categories, and arrows between morphisms are omitted. The object and morphism components of a functor are sometimes explicitly distinguished as and , respectively. Otherwise, the functor component is implicitly identified by its argument.
Functor composition and isomorphism are defined analogously to morphisms (above). That is, the composition of functors and is the functor , sending all objects in to objects in ; and morphisms in to morphisms , such that identity and composition are respected. That is, ; and . A functor is an isomorphic functor, if and only if there exists a functor such that and , where and are the identity functors sending objects and morphisms to themselves in the respective categories.
Theories of cognition employ some form of representation. Functors provide a theoretical basis for constructing representations. For example, computational systems often employ lists of items, such as numbers. In category theory, lists can be modeled as monoids from the category whose objects are monoids, and morphisms are monoid homomorphisms [28]. A monoid is a set , with an associative binary operation , and an identity element , such that for all . A list monoid [28] is the set of all ordered lists constructed from set by concatenation operator , where the identity element is the empty list (so that, e.g., ). (It is worth noting that strings, e.g., lists of characters, of length 2 over the set are denoted , and strings of length denoted . In computer science, often means “match anything”, hence the notation can be read as strings of any length .) Lists can be constructed from sets by the functor , as indicated in the example diagram(4)where is the object part of (i.e., ) and is the morphism part (i.e., ), so that, e.g., (i.e., morphism is mapped to monoid homomorphism , which we will refer to as ). (For simplicity, we have omitted composition with a second morphism in each of the categories and functor mappings, as was shown in Diagram 3.) So, for example, . The examples pertaining to lists were adapted from [28] (Chapter 2), where in [28] corresponds to our . We choose to label the object component of the functor rather than to emphasize the fact that the functor constructs a set of lists of numbers from a set of numbers, not just a single list containing those numbers.
The two different sorts of arrows in Diagrams 3 and 4 highlight the constructive nature of functors. The objects are (co)domains with respect to the morphisms within categories, but are themselves elements of larger objects (in general, the class ) with respect to the morphisms between categories. In programmer parlance, was “lifted” from being a function over numbers to become a function over lists of numbers. In this way, functors provide a means for constructing new representations and processes from existing ones in a structurally consistent manner.
Notice that the definition of functor does not dictate a particular choice for monoid homomorphism as part of the definition of . A natural choice is to define so that functions applied to one-item lists result in one-item lists (i.e., ). Another choice that turns out to also respect the definition of a functor includes two copies of each transformed element (i.e., ). In this case,So, and in particular are monoid homomorphisms. In fact, there are many possible monoid homomorphisms that could be chosen to define this functor. Consequently, in the case of an architectural component of a cognitive system, there are many possible ways of constructing structurally consistent representations and processes from existing ones. We need to find a principled way to choose the “right” monoid homomorphism. In the context of explaining systematicity, a similarly principled choice is necessary. To narrow the choice down to a particular monoid homomorphisms, and hence a particular representational scheme, we need two additional category theory concepts: natural transformation and adjunction.
A natural transformation is a structure-preserving morphism from domain functor to codomain functor that consists of for each object in , such that , as indicated by the commutative diagram in the category (5)
Again for expository purposes, we include the source category and functor arrows, which are usually left implicit in such diagrams. When a transformation is natural in the technical sense it seems natural in the intuitive sense, for mathematicians. In fact, category theory was founded in an attempt to formalize such intuitions [24]. We will return to this point about naturality, in the Discussion, as it pertains to an explanation of systematicity without reliance on ad hoc assumptions.
A natural transformation is a natural isomorphism, or natural equivalence if and only if each is an isomorphism. That is, for each there exists a such that and . Natural transformations also compose, and the composition of two natural transformations is also a natural transformation. Just as there are identity morphisms mapping objects to themselves, and identity functors mapping categories to themselves, there are also identity natural transformations, , mapping functors to themselves. And, so, the composition of a natural isomorphism (isomorphic natural transformation), , with its inverse, , is an identity natural transformation, i.e., .
Functors preserve structure between categories; natural transformations identify the similarities between functors. For our purposes, functors construct new representations and processes from existing ones; natural transformations identify the similarities between constructions. A simple example that is closely related to the functor example, illustrating this perspective, involves list reversal as indicated by the commutative diagram(6)where the domain and codomain objects of each morphism are sets of lists, such as ; and is essentially with (co)domain the set instead of the monoid . As the diagram illustrates, squaring a reversed list is the same as reversing a squared list. So, there is a non-trivial (i.e. non-identity) relationship between the list monoid construction functor () and itself. The functor constructing the lists in Diagram 6 is closely related to in that the returned object is just the underlying set of the monoid , forgetting the binary operation and the identity element. The underlying set can also be extracted by a functor from the category , as we will see in the next section. This example shows how two ways of constructing individual lists, via the functor, are related by the list reversal natural transformation, .
Although their associated diagrams look similar, there is an important difference between functor and natural transformation pertaining to the equality constraint that defines the relationships between object elements. For a functor, the equality constraint is local to the codomain of the transformation, i.e. the relationships between object elements within the constructed category. And so, the elements of the objects in the new category are only indirectly related to the elements in the corresponding objects of the source category by the categories' common external structure (i.e. inter-object relationships). For a natural transformation, the equality constraint spans the transformation, involving object elements mapped by both domain and codomain functors. And so, the two functors are directly related to each other by the internal structure of their associated objects (i.e. the relationships between object elements within an object). As part of a theory of cognitive architecture, there is a tension between the freedom afforded by functorial construction on the one hand—allowing an architecture to transcend the specific details of the source elements to realize a variety of possible representational schemes for those elements—and the need to pin down such possibilities to specific referents on the other. This tension is resolved with adjunctions.
An adjunction consists of a pair of functors , and a natural transformation , such that for every and there exists a unique , such that , indicated by the following commutative diagram:(7)where the functors are implicitly identified by (co)domain categories (left subdiagram) and (right subdiagram). The two functors are called an adjoint pair, , where is the left adjoint of , and is the right adjoint of ; and natural transformation is called the unit of the adjunction.
The left and right functors of an adjoint pair are like “inverses” of each other, but unlike an isomorphic functor whose composition with its inverse sends all objects and morphisms to themselves, the returned objects and their elements of a composition of left and right adjoints are related to the argument (source) objects and their elements by a natural transformation. For categories and , the adjoint pair , consisting of functor that constructs the free monoid on the set , and then “forgetful” functor returns the underlying set of monoid , are related by an injection. The injection is called an insertion of generators, whose component at , , sends each element of to the corresponding element (one-item list) in . The elements together generate the set (i.e. is the alphabet from which the set of all “words” is constructed where each is mapped to ). In this context, is the unit of this adjoint pair.
The effect of on objects has just been given; the effect on morphisms is as follows: if is a function, then is defined as follows:(cf. [25], p.111–112). Note that is the functor defined in the Functors section.
Monoid is “free” in the informal sense that there are no missing or extra bits in the construction used to satisfy commutativity. The precise definition of free is as follows. Given the forgetful functor , and an object of , is free on if there is a morphism such that for any morphism , there exists a unique morphism such that , indicated in the following commutative diagram:(8)
However, not just any monoid generated from a set is a free monoid. For instance, the monoid (i.e. addition modulo 2) in the diagram(9)is not the free monoid on any set , because the only homomorphism, , maps 0 and 1 to , which does not make the diagram commute for . That is, . (It is easy to show that the free monoid on the empty set is . So is not the free monoid on the empty set, either.) Other free objects, such as the free group on a set are defined analogously (see [21]). A simple example of a free monoid as may be employed by a cognitive system is a primitive form of counting, where is the free monoid counter, having elements , on singleton set . This monoid is isomorphic to addition over the natural numbers, i.e. the monoid .
From free objects we get an alternative (equivalent) definition of adjunction: consider functor from the original definition. If for every object , is free on with morphism , then functor , with morphism mappings defined so that , is the left adjoint of , and is the right adjoint of [31].
Yet another (equivalent) definition of adjunction, favoured by category theorists for its conceptual elegance, highlights the symmetry between a pair of adjoint functors: a bijection (one-to-one correspondence) between the set of morphisms from object to in category and the set of morphisms from object to in category . So, identifying the unique morphism in one category means that it is associated with one and only one morphism in the other category.
In the list construction example, the unit of the adjunction is the injection sending each element in the set to the one-item list in the set of all lists constructed from , as shown in the following diagram:(10)where the left adjoint, , constructs the free monoid on the set ; and the right adjoint, , returns the underlying set, , of a list monoid, as mentioned earlier. In this way, given , the only homomorphism in the constructed category making the diagram commute is . The definition for arrow is essentially the same as , except that its (co)domain is a set, not a monoid. Other monoid homomorphisms that could have been chosen as part of the functor definition, such as , are excluded by and the commutativity property of the adjunction, because .
Since this arrangement works for any morphism in , it can also be used to define a particular list length function from a family of analogous “length” functions as indicated in the following commutative diagram:(11)where monoid is the set of non-negative integers with addition as the operator and 0 as the identity element; is a constant function sending every element to the number 1; and / are functions returning the number of items in a list. As in the previous example, the definition of functor affords other choices for “length”, such as , where is a list. This arrow is also a monoid homomorphism, since , where and are the lengths of lists and , respectively. Again, however, the morphism and the commutativity property force the usual choice for length function (i.e. ), and excludes others such as , because .
A general pattern emerges from this use of adjunction. Functor construction may afford multiple choices for particular morphisms (processes) in the constructed category, but a principled choice is obtained through the commutativity property of the adjunction. This arrangement means that we are not committed a priori to a particular representational scheme; i.e., we do not have to make an ad hoc assumption about what that representational format should be. Given that an architecture has the capacity for an instance of the group of computations under consideration, then necessarily it applies to all other computations in that group. In the case of list length, for example, may indeed be the “correct” choice when we require the length of a list of characters in number of bytes for characters that are 2-byte unicodes (i.e. the characters appearing in the extended set that includes other special symbols and language scripts requiring two bytes for unique identification). So, to paraphrase, a computational architecture with the capacity to count the length (in bytes) of some lists of 2-byte unicodes necessarily has the capacity to compute byte lengths for all other unicode lists. In this way, the explanation for the “systematicity of list length” has two parts: existence is afforded by the possible list length functions; and uniqueness is afforded by the commutativity property of the adjunction. Without the adjunction, the choice of construction is by ad hoc assumption. Our explanation for the systematicity of human cognition follows this pattern.
With these formal concepts in hand, we now proceed to our explanation of systematicity. We apply our explanation in two domains: systematicity with respect to relational propositions, and systematicity with respect to relational schemas. Then, we contrast our explanation with the Classical and Connectionist ones.
For expository purposes, we develop our adjoint functors explanation from its components. One may wonder whether a simpler category theory construct would suffice to explain systematicity. For this example domain, the components of this adjoint have some systematicity properties, but in and of themselves do not explain systematicity—just as for Classicism and Connectionism, having a property is not the same as explaining it. This bottom-up approach motivates the more complex category theory construct from which the systematicity properties necessarily follow. Our approach has three steps. First, we show a categorical product that has the systematicity of representation and systematicity of inference properties. However, a product of two objects may afford many isomorphic product objects that do not also have the compositionality of representation property. Second, we show that the product functor provides the principled means for constructing only those products that also have the compositionality of representation property. There may, however, be several products that have the compositionality property, but which differ in semantic content by having different orders between identical sets of constituents. So, a principled choice is needed to determine the product. So, third, we show that the diagonal functor, which is left adjoint to the product functor, provides that principled choice by the commutativity property of the (diagonal, product) adjoint functor pair. For concreteness, we refer to the category , but our explanation does not depend on this category.
(If we require an explanation of systematicity with respect to ternary relational propositions, then a ternary product is employed. The explanation for systematicity extends analogously, where the diagonal and product functors involve object triples. We may also need to explicitly represent a symbol for a relation, such as Loves. In this case, an object representing the relation symbol is paired with the product object representing the related entities. We address this situation in the next section. For present purposes, we omit relation symbols, since the relation is constant across the instances considered here and nothing essentially changes by its omission.
First, suppose objects (say, agents) and (patients) are sets containing representations of John and Mary, denoted as . Although and are the same set of members, we maintain distinct names to keep track of the distinction between member pairs. (The assignment of elements to objects is itself an assumption, but not an ad hoc one for our theory, as explained in the next section and in the Discussion.) A categorical product of these two sets is the Cartesian product of and , which is the set of all pairwise combinations of elements from and , together with projections and for retrieving the first and second constituents in each case. That is, , , and . By definition, the Cartesian product generates all pairwise combinations of elements from and , therefore this Cartesian product has the systematicity of representation property. Moreover, by definition, the categorical product affords the retrieval of each constituent from each representation (otherwise it is not a product), therefore the categorical product also has the systematicity of inference property. In this case, from the categorical product definition takes the role of input, so in terms of Diagram 2 inferring John as the lover from John loves Mary is just , where JM is the input and is the input-to-product object map, whose unique existence is guaranteed by definition.
The Cartesian product, however, is not the only product object that satisfies the definition of a categorical product of and . An alternative product has as the product object, and and as the projections. Indeed, for this example, any four-item set together with the appropriate projections for retrieving the constituents would suffice. However, these alternatives do not have the compositionality of representation property: the semantic contents of these representations, whatever they may be, are not systematically related to each other, or the semantic content of John, or Mary. Hence, categorical products, in themselves, do not necessarily provide an explanation of systematicity.
Second, for any category that has products (i.e. every pair of objects in has a product), one can define a product functor (or, , in the ternary case), that is from the Cartesian product of categories, , itself a category, to , where , , as indicated by the following diagram:(12)recalling that our functor diagrams explicitly identify the object component, , but not the morphism component, , of the functor. In this case, the semantic contents of these elements are systematically related to each other and their constituents John and Mary. This categorical construction is an instance of Classical compositionality, whereby the constituents , are tokened wherever the compositions are tokened. As such, it has the compositionality of representation property.
Although the product functor has the compositionality of representation property, it introduces a different problem: , where and is also a valid product, but the semantic content of is not the same as . That is because they have different order relationships between their constituents even though the corresponding constituents are identical. Thus, a principled choice is required to determine whether, for example, John loves Mary should map to (John, Mary), or (Mary, John). Otherwise, one can define an architecture that does not have the systematicity of inference property by employing both products to correctly infer Johnas the lover in John loves Mary via , yet incorrectly infer John as the lover in Mary loves John via , where position within the product triple identifies the relevant projection. The assumption that architectures employ only the first product is ad hoc just like the assumption that Classical architectures employ grammars such as G1, but not G2. So, a principled choice is needed to determine the product.
Third—final step, this problem brings us to the second aspect of our explanation foreshadowed in the Introduction (i.e. uniqueness). Again, as we saw with lists, a particular construction is specified through the left adjoint functor. The left adjoint to the product functor is the diagonal functor (or, , in the ternary case), where , as indicated by the following diagram:(13)
The (diagonal, product) adjoint pair is indicated by the following commutative diagram:(14)(see [28] Example 2.4.6). In this manner, the John loves Mary family of cognitive capacities is specified by the commutative diagram(15)where and are the agent and patient maps from the set of proposition inputs into the set containing all the possible constituent representations. Here, we explicitly consider the case of equality, so that . When , and have different codomains, since , so the conflict between these products does not come into play, therefore the adjunction is not required and the product functor is sufficient. With the understanding that sets and are equal, we maintain the notational distinction for clarity in the subsequent text. Given as the morphism used by the architecture to map proposition inputs to their corresponding internal representations, then the definition of an adjunction guarantees that is unique with respect to making Diagram 15 commute via . That is, , where is the input for proposition John loves Mary. The alternative construction is excluded because . Having excluded by the commutativity property of the adjunction, the only two remaining ways to map the other inputs (i.e. and ) are equal. So, given that the architecture can represent John loves Mary as via and infer John as the lover via from the product , then necessarily it can represent Mary loves John and infer Mary as the lover using the same morphisms. That is, , or .
This explanation works regardless of whether proposition John loves Mary is represented as (John, Mary) via , or (Mary, John) via . In the latter case, the adjunction picks out just the construction , and hence , because it is the one and only one that makes the following diagram commute:(16)
That is, ,but . Given that the architecture can represent John loves Mary as via and infer John as the lover via from the product , then necessarily it can do so for Mary loves John using the same morphisms. That is, , or .
If we need to explicitly represent a symbol for a relation, such as Loves, the product object is paired with an object, say , representing the context in which the entities are related. The object representing the relation in this case is . This situation may arise where we need an explanation for systematicity that involves multiple similar relations, e.g., loves, likes, dislikes, and hates, where the capacity for instances of each of these relationships is co-extensive. That is, if one can represent John loves Mary and John likes Mary, then one can also represent the other six combinations, such as Mary loves John and Mary likes John. If one can represent John loves Mary, but not John likes Mary, then one can represent Mary loves John, but not Mary likes John. In this case, there is a category of relation symbols whose objects, , are symbols referring to each relation (e.g., loves, likes, etc.), and whose morphisms, , are just the identity morphisms for each object. (Such a category is called a discrete category.) Each relation, in this case, is a pair . Hence, the capacity to represent instances of the loves and likes relations extends to the other instances for both relations.
For these situations, the diagonal and product functors have extensions. The extension to the diagonal functor is: , such that and . The product functor is: , such that and . The adjunction, which is an extension of the one shown in Diagram 15, is shown in the following commutative diagram:(17)
In this situation, provides the explicit context in which entities are related.
Under the assumption that these relation symbols belong to a different category, then cases such as loves loves loves cannot be generated. Note that supposing different objects for these entities is not an ad hoc assumption for our theory. does not contain members such as John or Mary, and likewise (or, ) does not contain relation symbols, because they refer to different types of entities with respect to the theory—Loves refers to a relation, which is at the level of objects in our theory, whereas John and Mary refer to entities in a relationship, which are members of objects.
In summary, products may have the systematicity of representation and inference properties (see also Discussion), but may not have the compositionality of representation property. Product functors construct products that have the compositionality property, but there may be more than one product with this property. The possible presence of multiple products requires a principled choice for fixing the product. That choice is provided by the (diagonal, product) adjoint functor pair. Importantly, the unit of the adjunction, , is not a free parameter of the explanation, it defines the specific adjunction in part; and there is no choice in representational format (i.e. left-right, or right-left constituent order)—the given capacity to represent a proposition fixes the same order for all the other propositions. The same situation also applies for the explicit (multiple) relational propositions domain. Hence, systematicity is a necessary consequence of this (extended) adjoint pair without recourse to ad hoc assumptions, and so meets the explanatory standard set by Aizawa [2], and Fodor and Pylyshyn [1], for this domain.
Another domain in which humans exhibit systematicity is relational schema induction. This domain is more complex than the previous one in that the intrinsic connection is between relations, rather than within one. In the relational schema induction paradigm [32], participants are required to do cue-response prediction over a set of stimuli, such as letters and shapes, whose relationships conform to a group-like structure. For example, participants are shown (trigram, shape) pairs generated from a set of four trigrams (e.g., NEJ, POB, KEF, BEJ) and two shapes (e.g., square, circle), and are required to predict the response trigram, also from the same trigram set. Suppose, for example, a participant is presented with NEJ and square. After making a prediction, the correct response trigram is presented. This procedure is repeated with a new cue-response trial. The first two responses are not predictable prior to the feedback provided by the correct trigram. Hence, the first two trials are regarded as “information” trials. Each block of eight trials (i.e. all possible trigram-shape combinations) is repeatedly presented until a certain criterion level of correct performance is reached (e.g., correct responses to all eight trials in a block). Each set of eight cue-response pairs (i.e., four trigram times two shapes) constitutes a task instance. Once participants reach criterion a new task instance of eight cue-response pairs was randomly generated from a larger pool of possible trigrams and shapes (task instance examples are shown in Tables 1 and 2). The crucial data for this paradigm are the performances on subsequent task instances. When subsequent task instances conformed to the same structure, albeit with different stimuli, mean response error over the 48 participants was at or near optimal level: 2.00 errors per eight trials for the sequence of task instances conforming to the Klein group, and 2.67 for task instances conforming to the cyclic-4 group—two information trials are needed to determine the assignment of novel stimuli to structural elements [32]. The results provide another example of systematicity of human cognition: given that a person can correctly do one task instance and the information trials from the new task instance, then necessarily they can predict trials of all others, with the usual provision for a distinction between competence and performance.
This task is modelled as the category of sets with actions, (cf. [25], 6.3.1, and [33] Definition 5.2), that has objects for task instances, where is a set of states indicated by trigrams, is a set of “actions” indicated by shapes, and specifies the action of a shape on a trigram resulting in a trigram. The morphisms in this category consist of pairs of maps and , such that the following diagram commutes:(18)where the identity morphism is the pair of identity maps , and compositions are defined component-wise. In our example, the set consists of four elements representing the four trigrams, and the set consists of two elements representing the two shapes.
For the purpose of finding a suitable adjoint, we need to see how is naturally embedded in a monoid. Recall that a monoid consists of a set and a binary associative operator that satisfies closure: i.e., for all , whenever is defined, and there is an identity element , such that . In terms of our ASets (i.e. objects in ), the monoid identity corresponds to a “shape” whose action is to do nothing at all to the trigrams on which it acts: it leaves them unchanged. (However, this shape was not included in the experiments [32].)
The adjoint functor pair used for this domain consists of the forgetful functor , which returns the underlying sets, i.e. and , and its left adjoint, the free functor , which constructs ASets. The (free, forgetful) adjoint is shown in the following commutative diagram:(19)where and, for the instance of interest to us, and are the (trigram, shape) pairs of sets for the first and second tasks (respectively), as defined for example in Tables 1 and 2 so that , , etc. Full details and a proof that is an adjoint functor pair are provided in Text S1.
Our explanation for systematicity in this domain follows the now familiar pattern, where monoids model the relationships between actions in each task instance. (Though our argument employs monoids, nothing essential changes if instead we use semigroups, or groups, where for example each task instance is extended with two additional shapes, one explicitly corresponding to the identity element, and the other to the remaining element in the Klein, or cyclic-4 group. For these cases, the proofs of adjointness can be extended to involve free semigroups and free groups, respectively.) Given an ASet modelling the first task instance and an ASet modelling the second task instance, there is more than one homomorphism from the first to the second, only some of which afford the correct responses to the stimuli in the second task instance. For example, one homomorphism has the following trigram and shape mappings: , , , , , and . Basically, the <1?show=[to]?>first table collapses to a table with one row and two columns. It is straight forward to check that it is indeed a homomorphism, for example, . However, this homomorphism does not yield the correct responses to some of the stimuli in the second task instance. For example, all predictions to trigrams REZ and JOQ are no longer possible. Thus, a principled choice is required to select only those homomorphisms that indeed result in models for the second task instance. That choice is determined by and the commutative property of the adjunction. That is, having obtained the first task instance, and given the two information trials of the second task instance that identify the correspondences between task stimuli, then there is one and only one homomorphism making the diagram commute, so that correct responses are obtained from the remaining trials of the second task instance. And so, systematicity is a necessary consequence of this adjunction.
Some readers may be interested in developing alternatives, or extensions to existing theories to address the systematicity problem in light of our explanation, so it is worth formally characterizing how our approach differs from previous ones. The difference between our category theory explanation and Classical/Connectionist approaches to systematicity may be characterized as higher-order versus first-order theories. Category theory also provides a formal basis for this distinction in terms of more general n-category theory (see, e.g., [34]). Though the concerns of n-category theorists go way beyond what we need here, some elementary aspects of the theory are used to formalize the difference between why our adjoint functors explanation addresses the systematicity problem and why the Classical or Connectionist approach does not.
Notice that the definitions of functor and natural transformation are very similar to the definition of a morphism. In fact, functors and natural transformations are morphisms at different levels of analysis: a natural transformation is a morphism one level above functors as we shall see. For n-category theory, a category such as is a 1-category, with 0-objects (i.e. sets) for objects and 1-morphisms (i.e. functions) for arrows. A functor is a morphism between categories. The category of categories, , has categories for objects and functors for arrows. Thus, a functor is a 2-morphism between 1-objects (i.e. 1-categories) in a 2-category. A natural transformation is a morphism between functors. The functor category, of functors from to , has functors for objects and natural transformations for arrows. Thus, a natural transformation is a 3-morphism between 2-objects (i.e. functors) in a 3-category. (A 0-category is just a discrete category, where the only arrows are identities, which are 0-morphisms.) In this way, the order of the category provides a formal notion of explanatory level.
Classical or Connectionist compositionality is essentially a lower-levels attempt to account for systematicity. For the examples we used, that level is perhaps best described in terms of a 1-category. Indeed, a context-free grammar defined by a graph is modelled as the free category on that graph containing sets of terminal and non-terminal symbols for objects and productions for morphisms [31]. By contrast, our category theory explanation involves higher levels of analysis, specifically functors and natural transformations, which live in 2-categories and 3-categories, respectively. Of course, one can also develop higher-order grammars that take as input or return as output other grammars. Similarly, one can develop higher-order networks that take as input or return as output other networks (e.g., networks whose connectivity is dynamic, such as cascade correlation [35]). However, the problem is that neither Classical nor Connectionist compositionality delineates those (higher-order) grammars or networks that have the systematicity property from those that do not. Likewise for our category theory explanation, not just any functor, nor just any natural transformation accounts for systematicity. If the explanation was left at either of these levels, then our approach would also succumb to the same problem that befalls Classicism and Connectionism—i.e. the problem of having to stipulate, ad hoc, just which functors or natural transformations account for the systematicity property. Rather, it is a natural transformation between an identity functor and a composition of two other functors () that defines the adjunction that accounts for systematicity relative to the particular domain of interest. In this formal sense, a crucial difference is that there is also a between-levels aspect to our explanation.
Our adjoints explanation of systematicity has essentially two parts: (1) existence, showing how a particular connection between cognitive capacities is possible from a functorial specification of the architecture; and (2) uniqueness, explaining why that particular connection is necessary because it is the one and only one that satisfies the commutativity property of the adjunction. In contrast, the Classical and Connectionist explanations only provide an account of existence, but not uniqueness. That is, some grammars/networks afford the required intrinsic links between capacities and some do not, just like some functorial constructions do and some do not; but, for Classicism or Connectionism, there is no further explanation determining only those grammars or networks yielding systematicity (other than by ad hoc assumption), whereas for the category theory explanation the adjunction specifies only the systematic functors. So, our explanation meets the explanatory standard laid out by Aizawa.
To be regarded as a theoretical explanation for systematicity, such an explanation should be potentially falsifiable. Our explanation could be challenged by an alternative theory that accounts for systematicity (without ad hoc assumptions) in a way that does not require, or implement an adjunction. This possibility would not falsify our explanation as such, but may provide an alternative theory that is preferred on other grounds. Alternatively, there may exist a domain in which humans exhibit systematicity but for which there does not exist a relevant adjunction. Hence, the category theory approach we have put forward is in principle falsifiable.
The unit of an adjunction is a natural transformation between functors. The sense in which a transformation is natural is that the transformation does not depend on a particular “basis”. A mathematician's example is to contrast the dual of a vector space with the, natural, double dual (dual of the dual) of a vector space—the former depends on a specific set of basis vectors chosen ad hoc, the latter does not. The analogue, here, is that our explanation of systematicity is natural in that it does not depend on a particular representational scheme (i.e., constituent order for relational propositions). Hence, the explanation does not depend on ad hoc assumptions about internal representations. Contrast this explanation with the Classical one, which must assume a particular grammatical form (e.g., G1 over G2) to fit the data.
In addition to explaining systematicity, our category theory approach has further implications. According to our explanation, systematicity with respect to binary relational propositions requires a category with products. A category theory account has also been provided for the strikingly similar profiles of development for a suite of reasoning abilities that included Transitive Inference and Class Inclusion, among others [30]—all abilities are acquired around the age of five years. The difference between the difficulties of younger children and the successes of older children (relative to age five) across all these reasoning tasks was explained as their capacity to compute (co)products. (A coproduct is related to a product by arrow reversal—see, e.g., [28] for a formal definition.) Therefore, our explanation implies that systematicity is not a property of younger children's cognition. Some support for this implication is found on memory tasks that require binding the background context of memorized items [36], though further work is needed to test this implication directly.
Our explanation for systematicity in regard to binary relational propositions does not depend on , it only requires a category with products. For example, the categories of topological spaces and continuous mappings, and of vector spaces and linear mappings [21] could also be used. These possibilities imply that an explanation of systematicity does not depend on a particular (discrete symbolic, or continuous subsymbolic) representational format. Thus, a further benefit is that our approach opens the way for integration of other (sub/symbolic) levels of analysis.
Though some effort is needed to provide a category theory explanation for systematicity, even for a relatively simple domain such as relational propositions, the potential payoff is that our explanation generalizes to other domains where an appropriate adjunction is identified. This sort of tradeoff has been noted elsewhere in the context of a category theory treatment of automata [25]. We sketch one possibility in the domain of context-free grammars. Languages conforming to context-free grammars can be modelled as the free category on the directed graph that defines the grammar, whose vertices are sets of terminal and non-terminal symbols, and edges are transitions [31]. The left adjoint is the functor from the category of directed graphs and graph homomorphisms to the category of categories and functors (category homomorphisms). The right adjoint is the forgetful functor , which returns the underlying graph (i.e. the arrows, forgetting their compositions). The explanation here is analogous to our explanation for relational schemas. The problem Aizawa raised with respect to Classicism is avoided here because systematicity is not derived from individual grammars, but homomorphic relationships between grammars.
Having provided an explanation of systematicity in terms of the rather abstract category theory concept of adjoint functors, one may wonder what this explanation means for a more typical conception of cognitive architecture in terms of internal representations and processes, and their realization in the brain. Human cognition is remarkable in that it affords the ability to think about things that have no sensory access (e.g., “a dog that is one lightyear long …”); yet reason about such entities as if they were grounded in our everyday experience (“… is smaller than a dog that is two lightyears long”). However, these two aspects must be reconciled: unbridled abstraction means that one can no longer determine what a particular internal representation is supposed to refer to; yet blinkering the system with over-narrowly defined representations curtails one's ability to think outside the box. These aspects appear in the form of functors and natural transformations in category theory. The adjunction is the category theory way of bringing them into precise “synchrony”, or co-ordination, so that we may think abstractly about very specific things.
The realization of computational processes in the brain is classically conceived as a physical instantiation mapping from computational states to brain states, where the syntactic relationships between computational states correspond to physical relationships between brain states via such maps (see [1], p13). Category theory affords a similar, but more general and formal treatment in terms of functors. Diagrams of categories are formally defined as functors that map graphs (i.e. the shape of the diagram) to categories (see, e.g., [37]). Analogously, a categorial cognitive system would involve a functor from a categorial computational model to a brain system.
Up to this point, we have not considered the relatively new Bayesian approach to cognitive modelling (see, e.g., [38], [39] for summaries) because, to our knowledge, a Bayesian explanation for systematicity has not yet been articulated. Nonetheless, the hierarchical Bayesian approach offers a significant advance with the ability to learn a diverse range of structures, such as lists, trees, and other (acyclic or cyclic) graphs, from data [40]. An important aspect of this approach is that structural form (or the type of structure) is encoded as prior beliefs by hyperparameters in the higher layers, and instances of those structures are encoded as parameters in the lower layers in so far as they conform to the constraints imposed by the data (environment). In this way, the architecture is not required to presume one particular structure to induce a group of behaviours from data. The hierarchical Bayesian approach affords the sort of higher-order theory that our analysis in the previous section implies. However, the question for the Bayesians is essentially the same as for the Classicists and Connectionists: that is, to articulate the Bayesian architectural principles from which systematicity necessarily follows. As the approach currently stands, systematicity depends on a number of factors including the available data, network connectivity, and optimization parameters. A Bayesian network with independently modifiable parameters for representing the distributions of constituents in each argument position of a relation may not have the systematicity property in the absence of data with, say, Mary in the patient position (so called strong systematicity [18]), simply because there may be no (prior) information available to determine the value of the associated parameters. Hyperparameters may enable a dependency between lower level parameters so that the acquisition of one entails the acquisition of another. Still, systematicity may not necessarily follow from hyperparameters alone: for example, one can envisage a network where partial hyperparametrization links some but not all behaviours within the group, analogous to the problem that was raised with respect to classical compositionality.
All theories make certain assumptions. The question is whether those assumptions are extrinsic to the theory and carry the essential explanatory burden (i.e. they are ad hoc). In our case, one may question whether supposing that an object contains representations of John and Mary is not itself an ad hoc assumption, for the Cartesian product does not necessarily represent all possible combinations of mental representations [41] (e.g., generates representations corresponding to John loves Mary and Mary loves Mary, but not John loves John). Our explanation for systematicity of binary relational propositions is a consequence of the (diagonal, product) adjoint (Diagram 15), not a specific categorical product. Though the categorical product is a component of the explanation, the particular product is derived from the adjunction, not chosen independently of it. Where the constituent entities are of the same sort, and so belong to the same object () in our theory, the diagonal functor generates the object pair , and the product functor takes and generates the product object , hence cases like cannot occur in this formulation. The assumption that relation symbols belong to a different category than the related arguments precludes the generation of intrinsically unconnected cases, such as loves loves loves. Typing, in this sense, shares some of the explanatory burden, but types are not extrinsic to our theory. An element cannot exist without belonging to an object (its type) in a category, by definition. Hence, types are intrinsic to the theory. Moreover, the explanatory burden is also born by the adjunction in our example domains. Even with typing, there must still be a principled choice for the order of those constituents, when they involve the same objects, which is provided by the adjunction. And, given that adjunctions are central to category theory, neither the assumption of types, nor our use of adjunction can be regarded as ad hoc for the purpose of explaining systematicity in these domains. Classicism also makes a distinction between atomic and molecular representations, as a core assumption [1]. However, even under core assumptions that are equivalent to ours—John and Mary belong to the same word classes, which differ from loves—systematicity does not necessarily follow, as exemplified by grammar G2. Hence, the critical difference between our explanation of systematicity and the Classical approach is the adjunction.
This assumption of typing, though, is acute for quasi-systematic domains, where cognitive capacity may extend to some but not all possible constituent combinations, which appear to be particularly prevalent in language (see [41]). For these cases, we would also need category theory-derived principled restrictions to products. Equalizers and pullbacks (see [30] for an application to cognitive development) are two ways to restrict (product) objects, in the same arrow-theoretic style. Products, pullbacks and equalizers are all instances of the general, formal concept of a limit in category theory. The existence of adjoint functors is closely linked to the existence of limits in the respective categories (cf. adjoint functor theorems [21], p210–214), which suggests that an appropriate adjunction can also be found for domains that require an explanation for quasi-systematicity.
Needless to say, our category theory explanation is not the final word on a theory of cognitive architecture. For our approach (and Classicism), where the assignment of elements to objects (and, words to word classes) is asserted, there is also the broader question of why they get assigned in a particular way. This question pertains to the acquisition of representations, whereas the systematicity problem pertains to their intrinsic connections. Incorporating category theory into the Bayesian approach may provide a more integrative theory in this regard. A connection between category theory and probability has been known for some time (see [42]), and category theory concepts have been incorporated into the development of probabilistic functional programming [43]. A potentially fruitful line of future research, then, may be to identify a suitable adjunction with respect to, say, a category of Bayesian models, if such a category exists.
From a category theory perspective, we now see why cognitive science lacked a satisfactory explanation for systematicity—cognitive scientists were working with lower-order theories in attempting to explain an essentially higher-order property. Category theory offers a re-conceptualization for cognitive science, analogous to the one that Copernicus provided for astronomy, where representational states are no longer the center of the cognitive universe—replaced by the relationships between the maps that transform them. |
10.1371/journal.pntd.0005014 | Plasmodium vivax Reticulocyte Binding Proteins Are Key Targets of Naturally Acquired Immunity in Young Papua New Guinean Children | Major gaps in our understanding of Plasmodium vivax biology and the acquisition of immunity to this parasite hinder vaccine development. P. vivax merozoites exclusively invade reticulocytes, making parasite proteins that mediate reticulocyte binding and/or invasion potential key vaccine or drug targets. While protein interactions that mediate invasion are still poorly understood, the P. vivax Reticulocyte-Binding Protein family (PvRBP) is thought to be involved in P. vivax restricted host-cell selectivity.
We assessed the binding specificity of five members of the PvRBP family (PvRBP1a, PvRBP1b, PvRBP2a, PvRBP2b, PvRBP2-P2 and a non-binding fragment of PvRBP2c) to normocytes or reticulocytes. PvRBP2b was identified as the only reticulocyte-specific binder (P<0.001), whereas the others preferentially bound to normocytes (PvRBP1a/b P≤0.034), or showed comparable binding to both (PvRBP2a/2-P2, P = 0.38). Furthermore, we measured levels of total and IgG subclasses 1, 2, 3 and 4 to the six PvRBPs in a cohort of young Papua New Guinean children, and assessed their relationship with prospective risk of P. vivax malaria. Children had substantial, highly correlated (rho = 0.49–0.82, P<0.001) antibody levels to all six PvRBPs, with dominant IgG1 and IgG3 subclasses. Both total IgG (Incidence Rate Ratio [IRR] 0.63–0.73, P = 0.008–0.041) and IgG1 (IRR 0.56–0.69, P = 0.001–0.035) to PvRBP2b and PvRBP1a were strongly associated with reduced risk of vivax-malaria, independently of age and exposure.
These results demonstrate a diversity of erythrocyte-binding phenotypes of PvRBPs, indicating binding to both reticulocyte-specific and normocyte-specific ligands. Our findings provide further insights into the naturally acquired immunity to P. vivax and highlight the importance of PvRBP proteins as targets of naturally acquired humoral immunity. In-depth studies of the role of PvRBPs in P. vivax invasion and functional validation of the role of anti-PvRBP antibodies in clinical immunity against P. vivax are now required to confirm the potential of the reticulocyte-binding PvRBP2b and PvRBP1a as vaccine candidate antigens.
| In parallel with the tremendous reduction in malaria burden, Plasmodium vivax (Pv) is now the predominant malaria species in the Asia-Pacific and Americas. Pv can only invade young erythrocytes (reticulocytes) and this restriction is thought to involve the Reticulocyte-Binding Protein family (PvRBP). Given their predicted role, PvRBPs are potentially interesting vaccine targets. However, the acquisition of immunity to Pv in general (PvRBPs in particular) is poorly understood, hindering vaccine development. Here, we show that out of five PvRBPs, only one (PvRBP2b) binds exclusively to reticulocytes. Furthermore, we measured antibody levels to all six PvRBPs in a cohort of young Papua New Guinean children, assessing the relationship between antibodies to PvRBPs and risk of malaria disease. Both total and specific antibody subclass levels (IgG1 and IgG3) to the reticulocyte-specific binder PvRBP2b, and the non-specific binder PvRBP1a were strongly associated with lower risk of clinical disease. Our findings indicate a diversity of roles of PvRBPs in erythrocyte invasion and highlight their importance as targets of the naturally acquired immunity to Pv. Functional studies of the role of PvRBPs in reticulocyte invasion will be required to fully understand the potential of PvRBP1a and PvRBP2b as vaccine candidates.
| The two major malaria parasites, Plasmodium falciparum and Plasmodium vivax, differ in their ability to invade human erythrocytes. While P. falciparum invades both mature (normocytes) and young erythrocytes (reticulocytes), P. vivax can only invade the latter [1]. This differential specificity is believed to be mediated by distinct ligand-receptor interactions, though the exact mechanisms remain to be elucidated [1]. For P. falciparum, the merozoite invasion of erythrocytes is a multistep process [2] mediated by the binding of the erythrocyte binding-like (EBL) and reticulocyte binding-like (PfRh) protein families to receptors on the surface of the host cell [3]. For P. vivax, the only ligand-receptor interaction identified to date is between the Duffy binding protein (PvDBP) and the Duffy antigen receptor for chemokines (DARC) [4]. The recently identified P. vivax erythrocyte-binding protein (PvEBP) also shares a Duffy binding-like domain [5]. However, the presence of DARC in both normocytes and reticulocytes does not explain the restricted host-cell selectivity of P. vivax. The recent observation that P. vivax can invade Duffy-negative cells also indicates the existence of alternative pathways of invasion [6, 7].
The P. vivax reticulocyte binding brotein family (PvRBP) is composed of 11 members [5, 8, 9], and although their precise roles remain largely unknown, their homology to the much better characterized PfRh protein family suggests that they may be important invasion ligands [3]. Members of the PvRBP family have been implicated in erythrocyte binding, and in some cases in reticulocyte recognition [8, 10]. Variation in expression of PvRBPs genes in different parasite isolates have been described, suggesting that these genes may be redundant in function [11]. The relatively high degree of polymorphism observed in the genes encoding PvRBPs also indicates that they are important for parasite survival and may be under immune selection [10, 12, 13]. Collectively, this suggests that the understudied PvRBP family may be of key importance for P. vivax invasion, and like their better studied P. falciparum homologues, potential targets for a vaccine targeting blood stage infections [14].
Antibodies to several P. vivax merozoite proteins have shown associations with reduced risk of vivax-malaria in naturally exposed individuals [15]. Among them, antibodies to PvDBPs have been target of extensive study [15]. While PvDBP is a promising vaccine candidate, several challenges to vaccine development remain, including the presence of highly polymorphic, immuno-dominant epitopes in the DARC-binding region II, and the need to elicit high titers to achieve strain-transcending blocking [14]. It is therefore likely that PvDBP would need to be combined with further antigens targeting alternative invasion ligands. Members of the PvRBP family are recognized by antibodies from vivax-positive patients [11, 16], and populations living in endemic areas [17, 18]. Yet, an in-depth characterization of the immune responses to these proteins, as well as the role of antibodies to PvRBPs in the acquisition of immunity to malaria is lacking.
In this study, we have assessed the erythrocyte-binding profiles of five members of the PvRBP family and their specificity to normocytes or reticulocytes, identifying a member of the PvRBP family that exclusively binds to reticulocytes. Furthermore, we measured levels of total and IgG subclasses to these five PvRBPs and a non- erythrocyte binding protein fragment of a sixth PvRBP in a cohort of young Papua New Guinean (PNG) children with well-characterized differences in exposure. We identified an association between reduced risk of vivax-malaria and antibodies to two of the PvRBPs, including the reticulocyte-specific binder. Our results provide important insights into the acquisition of immunity to PvRBPs in young children, highlighting this protein family as an interesting target to be further evaluated for their potential as P. vivax vaccine antigens.
Proteins included in this study were PvRBP1a (amino acids [aa] 160–1170), PvRBP1b (aa 140–1275), PvRBP2a (aa 160–1135), PvRBP2b (aa 161–1454), PvRBP2cNB (aa 501–1300) and PvRBP2-P2 (aa 161–641) Their expression and purification have been described in details elsewhere [10, 11]. Despite several attempts, we were unsuccessful in expressing recombinant PvRBP2c that includes the conserved erythrocyte-binding domain in E. coli using both native and refolding methods. Thus, the PvRBP2cNB fragment included in this study does not contain the erythrocyte-binding domain, which encompasses residues 128 to 429. An SDS-PAGE of PvRBP2-P2 recombinant protein is shown in S1 Fig; the purity and stability of the remaining PvRBPs have been verified and presented in a previous publication [11].
Antibody production was performed at the Walter and Eliza Hall Institute Monoclonal Antibody Facility as previously described [10].
96-well flat-bottomed plates (Maxisorp, Nunc) were coated with each of the 6 PvRBPs (65 nM/well) in individual wells and incubated for two hours. For 65 nM of protein in 100 μL, we added 0.8 μg, 0.9 μg, 0.7 μg, 1 μg, 0.6 μg and 0.4 μg for PvRBP1a, PvRBP1b, PvRBP2a, PvRBP2b, PvRBP2cNB and PvRBP2_P2 respectively. Plates were blocked with 5% skim milk/0.1% Tween-20 for one hour. After washing, specific anti-PvRBP polyclonal antibodies (1 mg/mL stock) were added at halving serial dilutions (from 1:2000 to 1:64000) for one hour. Plates were washed three times before the addition of HRP-goat anti-rabbit secondary antibodies (1:2000 dilution) for one hour. Azino-bis-3-ethylbenthiazoline-6-sulfonic acid (ABTS liquid substrate; Sigma-Aldrich) was used to detect HRP activity. 1% SDS was used to stop the reaction and absorbance was measured at 405 nm. All experiments were performed at room temperature. All washes were done in PBS/0.1% Tween-20, and dilutions of antibodies in 0.5% skim milk/0.1% Tween-20. Samples were tested in duplicates.
Reticulocytes were enriched from whole blood and the erythrocyte-binding assays performed as described previously [10]. Binding of PvRBPs was detected using 0.025 mg/mL of corresponding anti-PvRBP rabbit IgG.
Antibody reactivity to the 6 PvRBPs in naturally-exposed individuals was assessed in samples from a longitudinal cohort of 264 children (1–3 years old) undertaken in Ilaita, East Sepik Province, PNG [19]. Children were enrolled between March-September 2006, and followed for up to 16 months. Blood samples were collected every eight weeks and at episodes of febrile illness. All P. vivax infections were genotyped, allowing the determination of the incidence of genetically distinct blood-stage infections acquired during follow-up (i.e. the molecular force of blood-stage infections, molFOB) [20]. Samples collected at enrolment from 224 children that completed follow-up were included in the present study (median age 1.7, inter-quartile range [IQR] 1.3–2.5).
To measure antibody levels in the cohort of PNG children, purified proteins were conjugated to Luminex Microplex microspheres (Luminex Corp.) as described elsewhere [21, 22], using the following concentrations per 2.5 x 106 beads: PvRBP1a = 3 μg/mL; PvRBP1b = 11.4 μg/mL; PvRBP2a = 6.7 μg/mL; PvRBP2b = 0.2 μg/mL; PvRBP2cNB = 0.8 μg/mL; PvRBP2-P2 = 5.4 μg/mL.
Bead-array assays were performed as previously described [23]. Plasma samples were diluted 1:50 in PBS, and secondary antibody donkey F(ab’)2 anti-human IgG Fc R-PE (1 mg/ mL, Jackson Immunoresearch); mouse anti-human IgG1 hinge-PE (0.1 mg/ mL, clone 4E3, Southern Biotech); IgG2 Fc-PE (0.1 mg/ mL, clone HP6002, Southern Biotech); IgG3 hinge-PE (0.1 mg/ mL clone HP6050, Southern Biotech); or IgG4 Fc-PE (0.1 mg/ mL, clone HP6025, Southern Biotech) diluted 1:100 in PBS was added to detect total, IgG1, IgG2, IgG3 or IgG4 respectively. A dilution series of a pool made of serum collected as part of an earlier study with immune adults living in different villages of high malaria transmission in East Sepik Province, PNG, was included on each plate as positive controls.
To correct plate-to-plate variations, the dilutions of the PNG adult pool were fitted as plate-specific standard curves using a 5-parameter logistic regression model [22, 24]. For each plate, median fluorescence intensity (MFI) values were interpolated into relative antibody units based on the parameters estimated from the plate’s standard curve.
Associations with age, exposure and correlations between antibody levels of different subclasses and/or different antigens were determined using Spearman’s rank correlation, and differences by infection status using Mann-Whitney U tests. Generalized estimating equation (GEE) models with exchangeable correlation structure and semi-robust variance estimator were used to analyze the relationship between antibodies to PvRBPs and prospective risk of P. vivax episodes (defined as axillary temperature ≥ 37.5°C or history of fever in preceding 48 hours with a concurrent parasitaemia >500 P. vivax parasites/μl) over the 16 months of follow-up [22, 25]. For this, antibody levels were classified into tertiles (cut-off values are shown in Table 1), and analyses were done comparing the incidence rate ratio (IRR) of clinical malaria in those with medium and high versus low antibody levels. Children were considered at-risk from the first day after the blood sample for active follow-up was taken. For each child, the molFOB was calculated as the number of new blood-stage genetically distinct P. vivax clones acquired/year-at-risk, and square-root transformed for better fit [20]. Adjustments were made for seasonal trends, village of residency, age, and molFOB. In order to study the breadth of anti-PvRBP antibodies, for each antigen antibody levels stratified into tertiles were scored as 0 for low, 1 for medium and 2 for the high tertiles, respectively. Scores were then added up to reflect the breadth of anti-RBP antibodies, yielding a median score of 6 (IQR 2–9).
All analyses were performed using STATA version 12 (StataCorp) or R version 3.2.1 (http://cran.r-project.org).
Ethical clearance was obtained from the PNG Medical Research and Advisory Committee (MRAC 05.19), and the Walter and Eliza Hall Institute (HREC 07/07). Written informed consent was obtained from the parents or guardians all children participating in the cohort study.
Except for the PvRBP2cNB fragment, all PvRBPs proteins expressed encompass the conserved erythrocyte-binding domain [10]. As such, PvRBP2cNB binding serves as a control for background signal in this flow cytometry-based assay. Binding was significantly higher for all binding fragments except for PvRBP2b in normocytes (not stained with thiazole orange, TO-) and PvRBP1b in reticulocyes (stained with thiazole orange, TO +) (both P>0.05). Among the five binding PvRBPs tested we found three types of binding profiles: i) binding preferentially to normocytes: PvRBP1a (Fig 1, TO- vs. TO+: P = 0.034) and PvRBP1b (P = 0.017); ii) binding to both normocytes and reticulocytes: PvRBP2a (P = 0.38) and PvRBP2-P2 (P = 0.38); and iii) binding only to reticulocytes: PvRBP2b (P < 0.001).
We assumed that the pooled serum from hyper-immune PNG adults represented the equilibrium antibody levels to all proteins achievable under life-long natural exposure. Therefore, by comparison with IgG levels observed in PNG children, we determined how many children have already achieved IgG levels that were >50%, >25% or >10% of the adult levels (Table 1).
Although, semi-immune, young PNG children were reactive to all six PvRBPs tested, there were differences in the immunogenicity of different proteins (Table 1; S2 Fig). Whereas, 47% and 95% of children had reached >50% and >10% of the hyper-immune adult levels for PvRBP2-P2, respectively, only 7% and 20% reached the same levels for antibodies targeting the non-red cell binding PvRBP2cNB fragment (P < 0.001). The other proteins were intermediately immunogenic, with antibodies to PvRBP2a more rapidly acquired than those to PvRBP2b, PvRBP1a and PvRBP1b (Table 1; S2 Fig). To each PvRBP, total IgG levels correlated moderate to strongly with IgG levels to the other PvRBPs (rho = 0.49–0.82, P < 0.001), with the strongest correlation between PvRBP1b and PvRBP2b (rho = 0.82) (S1 Table).
Rabbit polyclonal antibodies against the different PvRBP constructs showed strong recognition of the specific constructs but only limited cross-reactivity (Fig 2). The exception was antibodies raised against the PvRBP2cNB fragment, which showed low cross-reactivity with PvRBP1b and PvRBP2a. This indicated that the high correlations in naturally acquired total IgG levels are therefore likely to reflect co-acquisition rather than cross-reactivity (Fig 2).
While antibodies to PvRBP1a, PvRBP1b, PvRBP2b and PvRBP2cNB increased moderately with age (rho = 0.15–0.29, P < 0.001–0.025), no such association was found for the two most immunogenic proteins, PvRBP2a and PvRBP2-P2 (Fig 3A; S2 Table). Total IgG levels to all proteins except PvRBP1b were significantly higher in the 124 children (55.4%) that had a current, PCR-detectable P. vivax infection (P < 0.001–0.006) (S2 Table). To better understand the effect of age on the acquisition of antibodies to PvRBPs, we stratified children by the presence of infection at sample collection. After stratification, increase in total IgG to PvRBP1b, PvRBP2b, PvRBP2cNB and PvRBP2-P2 with age were stronger in children with current infection (rho = 0.19–0.33, P < 0.001–0.039) suggesting that antibodies to PvRBPs are strongly reflective of recent exposure (S2 Table).
Given the young age and large heterogeneity in exposure among children [20], age is not a good proxy for life-time exposure. As a better measure of life-time exposure to malaria, we therefore calculated the molFOB (described in methods) [20]. This estimated P. vivax life-time exposure was a substantially better predictor of antibody levels than age, and total IgG levels to all 6 PvRBPs significantly increased with increasing life-time exposure (rho = 0.18–0.36, P < 0.001–0.006) (Fig 3B; S2 Table). For PvRBP1a and PvRBP2a, the effect of life-time exposure to P. vivax was observed only in children free of infection at study start (P = 0.008–0.022), again indicating that the effect of recent infections on antibodies to PvRBPs is strong (Fig 3C; S2 Table).
The breadth of anti-PvRBP antibodies (described in methods) was higher in children with concurrent P. vivax infection (median in children free of infection = 4.5, IQR = 2–8 versus median in infected children = 7, IQR = 4–10; P < 0.001), and increased with increasing estimated life-time exposure (rho = 0.29, P < 0.001).
In hyper-immune PNG adults, four different patterns of IgG subclass reactivity to PvRBPs were observed i) predominant IgG1 with sub-dominant IgG3: PvRBP2-P2; ii) predominant IgG3 with sub-dominant IgG1: PvRBP1a; iii) predominant IgG1 with sub-dominants IgG2+IgG3: PvRBP2a, PvRPB2b and PvRBP2cNB; and iv) IgG1+IgG2+IgG3 with no obvious dominance: PvRBP1b. There were no detectable levels of IgG4 to any of the proteins (Table 1; S3 Fig).
In comparison to adults, young children had already acquired substantial IgG1 levels to all 6 PvRBPs. Apart from the less immunogenic PvRBP2cNB fragment, >20% and >62% of the children had reached >50% and >10% of the IgG1 levels seen in hyper-immune adults to the different PvRBPs. PvRBP1a, PvRBP2-P2 and PvRBP2a also had detectable levels of IgG3, but only PvRBP2-P2 had a similarly high prevalence of IgG3 as for IgG1 (20.5% and 63.4%, respectively) (Table 1; S3 Fig), indicating that IgG1 antibodies were acquired faster than IgG3. The predominance of IgG1 subclass is further highlighted by the generally stronger correlations of total IgG with IgG1 (rho = 0.91–0.94, P < 0.001) than IgG3 (rho = 0.55–0.71, P < 0.001) (S1 Table). Despite the narrow age group in the cohort, there was a weak indication of polarization towards IgG3 with increasing age to PvRBP1a (rho = -0.19, P = 0.005) and PvRBP2-P2 (rho = -0.16, P = 0.015), evidenced as a decrease in the IgG1/IgG3 ratio. Children had no detectable IgG2 or IgG4 to any of the proteins (S3 Fig).
IgG1 to all 6 PvRBPs (P ≤ 0.003) were higher in infected children, although only moderately to PvRBP1b and PvRBP2cNB (P = 0.06). Similarly, those with a current infection had higher IgG3 to PvRBP1a, PvRBP2a and PvRBP2-P2 (P ≤ 0.002) (S2 Table). For the PvRBPs with dominant IgG1 subclass, the effect of age and exposure mostly mimic that observed for total IgG (S2 Table). For PvRBP2-P2 and PvRBP1a IgG3, but not IgG1, still increased with age (rho = 0.19–0.21, P = 0.002–0.005) (S2 Table).
Over the 16 months follow-up of the PNG cohort, each child had an incidence rate of 1.25 (95%CI 1.08–1.45) P. vivax episodes/year at risk. Following adjustment for confounders, total IgG to all 6 PvRBPs tested were associated with protection against P. vivax malaria (IRR 0.52–0.69, P < 0.001–0.016) (Fig 4; S3 Table).
To further understand the contribution of antibodies to specific PvRBPs in the protective effect observed, we fitted a multivariate model to account for the fact that antibodies to the 6 PvRBPs were co-acquired and therefore highly correlated. In multivariate analysis, only total IgG to PvRBP1a (IRRM 0.71, 95%CI 0.53–0.94, P = 0.019; IRRH 0.63, 95%CI 0.44–0.88, P = 0.008), and PvRBP2b (IRRM 0.73, 95%CI 0.54–0.99, P = 0.041; IRRH 0.63, 95%CI 0.44–0.90, P = 0.011) remained associated with reduced risk of vivax-malaria, suggesting that antibodies to these two PvRBPs are important correlates of naturally-acquired protective immunity.
After adjusting for confounders, both IgG1 and IgG3 to PvRBP1a (IgG1 IRR 0.48, P < 0.001; IgG3 IRR 0.51–0.67, P < 0.001–0.011) and PvRBP2a (IgG1 IRR 0.66, P = 0.010; IgG3 IRR 0.51, P < 0.001–0.011), and only IgG3 to PvRBP2-P2 (IRR 0.60, P = 0.002) were associated with reduced risk of vivax-malaria. IgG1 to PvRBP1b (IRR 0.52, P < 0.001) and PvRBP2b (IRR 057–0.59, P < 0.001) was also associated with protection. No association was found for the non-binding PvRBP2cNB fragment, although it was observed for total IgG levels in univariate analysis (Fig 4; S3 Table).
In a model combining both IgG1 and IgG3 levels to a given antigen, both IgG1 and IgG3 remained significantly associated with protection for PvRBP1a (IgG1 IRRH 0.55, 95%CI 0.38–0.80, P = 0.001; IgG3 IRRM 0.69, 95%CI 0.51–0.92, P = 0.011), and only IgG3 for PvRBP2a (IRRH 0.51 0.36–0.72, P < 0.001) and PvRBP2-P2 (IRRH 0.60, 95%CI 0.43–0.84, P = 0.002). When combining IgG1 and IgG3 to all PvRBPs in a multivariate model, however, only IgG1 to PvRBP1a (IRRH 0.56, 95%CI 0.39–0.80, P = 0.001) and PvRBP2b (IRRM 0.64, 0.46–0.90, P = 0.011; IRRH 0.69, 95%CI 0.48–0.97, P = 0.035) remained associated with reduced risk of vivax-malaria.
There was a very strong association between increasing total IgG to the repertoire of PvRBPs, and increase in protection against vivax-malaria. For each increase in one unit of the breadth score (see methods for detailed description), a reduction of approximately 8% in the incidence rate of P. vivax episodes was observed (IRR 0.92, 95%CI 0.89–0.96, P < 0.001). Considering the repertoire of IgG1 antibodies, the reduction in risk was of approximately 7% (IRR 0.93, 95%CI 0.90–0.97, P < 0.001). The effect of the IgG1 repertoire is no longer significant after IgG1 to PvRBP1a (IRRH 0.53, 95%CI 0.33–0.85, P = 0.008) and PvRBP2b (IRRM 0.63, 95%CI 0.44–0.89, P = 0.010, IRRH 0.65, 95%CI 0.42–1.00, P = 0.053) is accounted for, again highlighting importance of these proteins in naturally-acquired protective immunity.
Advances in understanding P. vivax biology and the acquisition of immunity to this parasite, as well as the development of vaccines against P. vivax lag much behind what has been achieved for P. falciparum [14, 26, 27]. This is largely due the lack of a stable in vitro culture system for P. vivax that makes functional studies very challenging.
In this study, we investigated whether six recombinantly expressed members of the PvRBP family are involved in P. vivax host-cell specificity by testing their erythrocyte-binding preferences. We identified that only PvRBP2b binds solely to reticulocytes. In previous reports, PvRBP1a and PvRBP2c have also been described as reticulocyte-specific binders [8]. Our recombinant PvRBP1a however binds preferentially to normocytes. One explanation for this observation is that C-terminal regions outside of the recombinant construct that is present on native protein governs reticulocyte specificity. In addition, PvRBP1a forms a complex with PvRBP2c [8] in parasites and this complex may be responsible for reticulocyte-binding. Unfortunately, as we were unsuccessful in expressing recombinant PvRBP2c with its binding domain, we were unable to confirm its erythrocyte-binding profile. The molecular mechanisms by which PvRBP2b mediates specific reticulocyte binding, and its reticulocyte-specific receptor are yet to be elucidated.
Since our ability to do functional assays with P. vivax is constrained due to the lack of in vitro culture, we sought to investigate whether antibodies to the 6 PvRBPs are targeted by natural-immunity in a population of young children from PNG. The 6 PvRBPs were recognized differently. The PvRPB2cNB fragment had the lowest immunogenicity of all and antibodies to this fragment did not have a strong association with risk of vivax-malaria. This may be a consequence of the absence of the erythrocyte-binding region. The strongest protective effect was observed for total IgG to PvRBP1a and PvRBP2b. As there is poor cross-reactivity between antibodies targeting the binding regions of different PvRBPs, it is highly likely that these antibodies may have additive or even synergistic effects.
IgG1 and IgG3 were the predominant IgG subclass to PvRBPs in PNG children. Interestingly, the dominant IgG subclass to PvRBP1a was different between children (IgG1) and adults (IgG3). For PvRBP1a and PvRBP2-P2, there was also some early evidence of switching to IgG3 with increasing age. Exposure to malaria parasites, among other antigenic and host characteristics, seems to play a major role in determining the predominant IgG response and for P. falciparum merozoite antigens, both age and transmission intensity have been previously associated with switching in predominance of the IgG subclass towards IgG3 [28, 29]. Both IgG1 and IgG3 subclasses are cytophilic, T-cell dependent, and bind strongly to Fcγ, mediating phagocyte activation and complement fixation [30, 31], and predominance of IgG1 and/or IgG3, in variable ratios, is common for several P. vivax [32–34] and P. falciparum merozoite proteins [35–37]. For the young PNG children included in this study, a reduction in risk of vivax-malaria was observed with IgG3 to PvRBP1a, PvRBP2a and PvRBP2-P2 but, ultimately, it was IgG1 to PvRBP1a and PvRBP2b that showed the strongest associations with protection. Adults also had detectable levels of non-cytophilic IgG2 to most PvRBPs tested however, the significance of this finding remains to be investigated. In P. falciparum IgG2 antibodies to EBA175 were shown to be short-lived [38], but nevertheless correlated with lower parasitemia. High levels of IgG2 to RESA and MSP2 have also been associated with a lower risk of P. falciparum infection [39], indicating that although uncommon, IgG2 antibodies might be important for immunity against malaria.
Antibodies to PvRBP2b have also been previously associated with lower parasitaemia in clinical cases [11]. Both proteins seem to be under selective pressure and, in comparison to PvRBP1a, PvRPB2b is less polymorphic and with highly conserved regions, which may be beneficial for vaccine development [12, 13]. Both PvRBP1a and PvRBP2b are less genetically diverse than PvRBP2c [12, 13]. The existence of antigenic diversity in the different PvRBP genes however, has never been investigated. A study with Brazilian samples identified two regions of PvRBP1a (aa 431–748 and 733–1407) as the most immunogenic with predominant IgG1 response, but the relationship between antibodies to the different regions and protection was not explored [18]. Further studies of different regions of both PvRBP1a and PvRBP2b molecules would be important to identify the main epitopes targeted by protective antibodies.
The major limitation of this study is the lack of inclusion of a recombinant protein containing the binding domain of PvRBP2c, precluding both the investigation of the red-cell binding characteristics of PvRBPc and determining relative importance of antibodies targeting red-cell binding fragments of all main PvRBP proteins. Nevertheless, this is the first study where antibody levels to PvRBPs were investigated in samples from a well-designed longitudinal cohort study, which made it possible to adjust for other factors that confound the relationship between antibody acquisition and risk of disease, most importantly the heterogeneity in individual exposure to P. vivax blood stage infections [20]. The findings of this study provide further insight into P. vivax host-specificity and naturally-acquired immunity to PvRBPs in children.
While the molecular functions of PvRBPs in P. vivax invasion are not well understood, the role of the PfRh family, which are homologs of PvRBPs in P. falciparum, have been well characterized in parasite invasion [3]. Several members of the PfRh family have been implicated in recognition of red blood cells, signaling events or creating a pore in the red blood cell membrane during invasion [40–44]. In particular PfRh5 has been a focus of intense research as a leading blood stage vaccine candidate due to its essential function in P. falciparum invasion [45, 46]. Apart from gene structure and sequence homology, PvRBP2a and PfRh5 also adopt a similar structural fold within their erythrocyte-binding domains [10, 47, 48]. Using this structural scaffold with varied surface properties PvRBPs and PfRhs are able to mediate alternate receptor engagement. Monoclonal antibodies against PfRh5 results in the strong inhibition of parasite growth across multiple strains and Aotus nancymaae monkeys immunized with anti-PfRh5 vaccine are protected against severe infection [42, 49, 50]. It is likely that the erythrocyte-binding domain of PvRBPs will be able to elicit antibodies that have the ability to block P. vivax invasion.
Our results underline the key role of PvRBPs in parasite–host interactions and highlight their potential as P. vivax vaccine candidate antigens. Further immuno-epidemiological studies in broader age groups from areas of different transmission intensities, as well as functional studies in vitro or in animal models, and a better understanding the molecular function of all PvRBPs, including the PvRBP1a/PvRBP2c complex in P. vivax invasion are now required to validate and prioritize one or several PvRBPs for development as vaccine candidates.
|
10.1371/journal.ppat.1004161 | Ly49C-Dependent Control of MCMV Infection by NK Cells Is Cis-Regulated by MHC Class I Molecules | Natural Killer (NK) cells are crucial in early resistance to murine cytomegalovirus (MCMV) infection. In B6 mice, the activating Ly49H receptor recognizes the viral m157 glycoprotein on infected cells. We previously identified a mutant strain (MCMVG1F) whose variant m157 also binds the inhibitory Ly49C receptor. Here we show that simultaneous binding of m157 to the two receptors hampers Ly49H-dependent NK cell activation as Ly49C-mediated inhibition destabilizes NK cell conjugation with their targets and prevents the cytoskeleton reorganization that precedes killing. In B6 mice, as most Ly49H+ NK cells do not co-express Ly49C, the overall NK cell response remains able to control MCMVm157G1F infection. However, in B6 Ly49C transgenic mice where all NK cells express the inhibitory receptor, MCMV infection results in altered NK cell activation associated with increased viral replication. Ly49C-mediated inhibition also regulates Ly49H-independent NK cell activation. Most interestingly, MHC class I regulates Ly49C function through cis-interactions that mask the receptor and restricts m157 binding. B6 Ly49C Tg, β2m ko mice, whose Ly49C receptors are unmasked due to MHC class I deficient expression, are highly susceptible to MCMVm157G1F and are unable to control a low-dose infection. Our study provides novel insights into the mechanisms that regulate NK cell activation during viral infection.
| We previously identified a viral murine cytomegalovirus (MCMV) strain whose variant m157 immunoevasin can bind the inhibitory Ly49C NK cell receptor in addition to activating Ly49H receptor in B6 mice. Here we show that simultaneous engagement of the two receptors by m157 hampers NK cell activation. Most Ly49H+ NK cells lack Ly49C in B6 mice, as a result, NK cell population efficiently controls MCMV infection; however, the anti-viral response is reduced in transgenic mice where all NK cells express Ly49C. MHC I masks Ly49C through cis-interactions, which restricts its inhibitory function. Indeed, B6 Ly49C Tg, β2m ko mice, that are deficient for MHC I, are highly susceptible to low dose of MCMV. Our findings indicate that both the NK cell repertoire and MHC I molecules control susceptibility vs resistance to viral infections.
| In humans, cytomegalovirus (CMV) is a pathogen responsible for causing significant mortality in immunocompromised patients [1] and in individuals lacking Natural Killer (NK) cells [2]. Mouse cytomegalovirus (MCMV) is a natural pathogen of mice. The similarities in structure and biology between human and mouse CMV make the latter a widely utilized model for human infection [3]. The study of MCMV has provided valuable insights into how the immune system responds to infection, and has helped to define the immune evasion mechanisms used by CMV to ensure that viral replication proceeds. NK cells play a crucial role in the early control of MCMV infection in resistant mouse strains; they limit viral replication and mortality during acute infection. The ability of NK cells to control viral infection is tightly regulated by their activating and inhibitory receptors [4]. Activating NK cell receptors include activating forms of killer cell immunoglobulin-like receptors (KIRs) in humans, and Ly49 receptors in mice. Both humans and mice express CD94/NKG2C which recognizes MHC class I molecules, and NKG2D which can be triggered by stress-induced ligands. NK cells also possess inhibitory receptors specific for MHC class I that permit discrimination of normal healthy cells from diseased ones, such as virus-infected cells, that display reduced MHC class I expression. These receptors include KIR in humans and members of the Ly49 family in mice, and LIR-1 and CD94/NKG2A in both species (reviewed in [5]). Inbred strains of mice express distinct NK cell receptor repertoires; NK cell receptors are encoded within a polygenic cluster in which each receptor gene is subject to polymorphism between the mouse strains; this variability results in resistance or susceptibility to specific viral infections.
Ly49H is the activating receptor responsible for resistance to MCMV infection in C57BL/6 (B6) mice [6]–[8]. Ly49H binds specifically to the m157 viral protein encoded by laboratory MCMV strains (Smith and K181) and triggers cytotoxicity and cytokine production [9], [10]. Arase et al showed that m157 binds to the inhibitory Ly49I receptor in 129/J mice, but not in B6 mice, while 129/J mice lack Ly49H [9]; this repertoire results in susceptibility to MCMV infection in the 129/J strain. In laboratory settings, immunological pressure through Ly49H was evidenced by the rapid selection of viral mutants producing m157 variants that escape recognition by this receptor [11]. Sequence analysis of m157 in a panel of MCMV isolates collected from a wild mouse population showed that only two isolates were identical to the laboratory MCMV strains (Smith and K181) [11], [12]. In addition, unlike the laboratory strains many of the viral isolates with m157 variants were able to replicate to high titers in resistant B6 mice [11]. We previously identified an MCMV strain (G1F) that was isolated from mice trapped in the wild; its m157 sequence shares over 93% homology with Smith and K181 strains but the protein displays a unusual binding profile to Ly49 receptors [13]. In addition to Ly49H, m157G1F can bind Ly49C in B6 and BALB/c mice [13], [14].
Inhibitory Ly49 receptors are thought to play a crucial role during NK cell education. Mechanisms of NK cell education are still unclear and different models co-exist. The current consensus states that NK cells expressing inhibitory receptors specific for self MHC class I molecules are fully educated (“licensed” [15] or “not disarmed” [16]) and have a greater response potential than NK cell subsets that lack such receptors. However, recent studies showed that Ly49C− NK cells are fully functional in B6 mice and indeed they dominate the Ly49H-dependent response to MCMV infection [17]. These results indicate that inhibition triggered by Ly49C binding to H-2 Kb overrides the responsive advantage gained by “licensing” and suggest that the inhibition mediated by Ly49C binding to H-2Kb regulates Ly49H-dependent NK cell activation. These data emphasize the need for a better understanding of the regulation of NK cells and how this impacts on NK cell function in the context of viral infection.
The high variability of the m157 sequence is not without similarities with human CMV, whose genome contains highly polymorphic loci that encode proteins (immunoevasins) with the potential to affect virulence through immune evasion [18]. The complex interactions of viral proteins with the host immune system are critical for determining a viral strain infectivity and pathogenicity. We undertook to study whether a viral immunoevasin able to bind multiple NK cell receptors can modulate the anti-viral immune response and to define the regulatory mechanisms. Our recent findings that m157G1F binds to two NK cell receptors with opposite functions in B6 mice, provide a unique opportunity to study, in a natural infection model, how the integration of competing signals determines tolerance versus killing and ultimately the outcome of an important viral infection in vivo.
We previously demonstrated the ability of m157 from the MCMVG1F strain (m157G1F) to bind both Ly49H and Ly49C in B6 mice [13]. Here, we analyzed the binding properties of m157G1F and aimed to determine whether it can bind the two receptors simultaneously. We transfected BWZ.36 cells to express either Ly49HB6 (BWZ-Ly49H) or Ly49CB6 (BWZ-Ly49C) and clones expressing the receptors at similar levels were selected to test m157G1F-Fc binding, by flow cytometry. Firstly, we measured the binding obtained at increasing concentrations of the m157G1F-Fc. Titration curves were similar for Ly49H and Ly49C expressing cells (Fig. 1A). No binding was detected to the parental BWZ.36 cells (data not shown). These results suggest that m157G1F binds to Ly49H and Ly49C with similar affinities. Next, we compared the kinetics of binding of m157G1F to these two receptors. Ly49H or Ly49C-expressing cells were incubated with saturating concentrations of m157G1F-Fc, as determined above, for 30 sec to 40 min. Fluorescence obtained after 40 min provided the maximal intensity value (100%), while the background to be subtracted was measured in the absence of fusion protein; the percentages of binding achieved over increasing incubation periods were calculated accordingly. Binding of m157G1F to Ly49C was found to occur slightly more quickly than to Ly49H, with 50% binding reached after 5 and 7 min, respectively (Fig. 1B). Thirdly, we measured the stability of the interactions. For this purpose, we used the optimal binding conditions previously determined, and analyzed dissociation rates. Excess amounts of anti-m157 antibody (6H121) prevented re-association of the fusion proteins to the receptors after they detached. We found that 50% of m157G1F that was initially bound to Ly49C had dissociated after approximately 20 min while over 50% was still attached to Ly49H after 90 min (Fig. 1C). These results suggest that m157G1F dissociates more quickly from Ly49C than Ly49H. A comparative binding analysis of m157K181 and m157G1F to Ly49H was also conducted as a control; our results showed similar binding characteristics of the two m157 variants to Ly49H (Fig. S1). This data confirmed our previous findings [13].
Overall, these results suggest that m157G1F binds to Ly49H and Ly49C with similar affinities but interactions with Ly49H are more sustained.
We previously showed that m157G1F binds to NK cells that are stained with the 5E6 mAb (i.e. NK cells expressing Ly49C and/or I) in B6 mice [13]. Here, we further analyzed m157G1F binding to NK cell subsets expressing Ly49H and/or Ly49C. Splenic NK cells were collected from the following mouse strains: B6 (H-2b), Cmv1r (H-2d background expressing B6 alleles in the NK cell receptor gene locus) [19] and B6 β2m ko (defective for MHC class I expression). Purified NK cells were stained with Ly49C-specific 4LO3311 and Ly49H-specific 3D10 antibodies in order to discriminate the respective NK cell subsets. We used the 4LO3311 antibody that is specific for Ly49C rather than 5E6, which also recognizes Ly49I [20]. 4LO3311 failed to stain B6 NK cells while it stained weakly Cmv1r NK cells; in contrast, NK cells from B6 β2m ko mice showed strong staining with 4LO3311 (Fig. 2A, left panels). In B6 mice, self-ligands for Ly49C are MHC class I molecules H-2 Kb; Ly49C can engage H-2 Kb on other cells (trans-interaction) as well as H-2 Kb expressed on the same NK cell membrane plan (cis-interaction) [21]. Cis-binding of the Ly49A inhibitory receptor with MHC class I has been shown to restrict the interactions with molecules presented in trans [22]. Likewise, the staining patterns detected on NK cells isolated from the three mouse strains analyzed were consistent with masking of Ly49C due to cis-interactions with H-2 Kb. In order to determine whether the differential staining illustrated in Fig. 2A were indeed due to cis-masking, we treated NK cells with an acid buffer; this treatment releases the β2m from MHC class I heavy chain, and disrupts cis-interactions [21], [22]. 4LO3311 bound to Ly49C on acid stripped B6 NK cells, while a more intense staining was achieved on Cmv1r NK cells; binding to B6 β2m ko NK cells remained unchanged (Fig. 2A, right panels). These results indicated that Ly49C is masked due to cis-interactions with MHC class I molecules in B6 mice. The existence of a partial masking in Cmv1r suggests that Ly49C also binds to H-2d MHC class I in cis but with a lower affinity than to H-2b molecules.
Next, we analyzed the binding of m157G1F to NK cell subsets expressing Ly49H and/or Ly49C (Fig. 2B). Using B6 β2m ko cells where Ly49C is free of cis-masking, m157G1F binding was detected to both Ly49H and Ly49C-expressing NK cell subsets. In B6, m157G1F binding was achieved only in the NK cell subsets expressing Ly49H, while in Cmv1r, m157G1F displayed weak binding to the Ly49C+, Ly49H− (Ly49C+,H−) NK cell subset. Acid stripping improved m157G1F binding to the Ly49C+,H− subset from B6 and Cmv1r and also resulted in enhanced binding to the subset co-expressing both receptors. The absence of m157G1F binding to untreated Ly49H− NK cells, despite a fraction of these cells expressing Ly49C, indicated that m157G1F cannot bind Ly49C when interactions in cis made it unavailable. As a control, we also incubated untreated and acid treated NK cells with the secondary antibody and with streptavidin in the absence of the first m157-Fc incubation step. We did not detect any fluorescence above the background, indicating the specificity of our multi-step staining in both untreated and acid treated NK cells (Fig. S2).
Consistent with the data we obtained using cell lines expressing NK cell receptors (Fig. 1), these results indicate that m157G1F can bind to NK cells expressing Ly49C and/or Ly49H. However, the ability to engage Ly49C is limited by MHC class I molecules (H-2b>H-2d) due to cis interactions.
We previously demonstrated that m157G1F induces intracellular signals upon binding to both Ly49H and Ly49C [14]. Here, we analyzed whether m157G1F differentially activates NK cells that express Ly49H only or coexpress Ly49H and Ly49C. B6 NK cells were exposed to target cells expressing m157G1F (RMA m157G1F) or to parental RMA cells and analyzed for degranulation and production of IFN-γ (Fig. 3A). Upon exposure to RMA m157G1F, Ly49H+,C− NK cells degranulated (∼40% LAMP1+) and produced IFN-γ (∼15% IFN-γ+). NK cell activation was Ly49H-dependent as it was abolished in the presence of blocking 3D10 antibodies. However, co-expression of Ly49C did not alter the response. We hypothesized that cis-masking of Ly49C by H-2 Kb was responsible for the absence of inhibition. We therefore examined NK cells isolated from B6 β2m ko mice, where Ly49C receptors are not masked in cis (Fig. 2A). We used RMAS cells (MHC class I deficient) expressing m157G1F (RMAS m157G1F) as targets instead of RMA m157G1F cells in order to prevent competition between m157 and Kb for binding to Ly49C, thus, ensuring that inhibition through Ly49C would only be due to m157 binding. RMAS m157G1F cells strongly activated Ly49H+,C− NK cells but not the Ly49H+,C+ subset. As above, activation was Ly49H-dependent. Blocking of Ly49C enabled Ly49H+,C+ NK cell activation, albeit not as strongly as the Ly49H+,C− subset (Fig. 3B).
Therefore, NK cell exposure to m157G1F-expressing cells triggers optimal Ly49H-dependent NK cell functions when Ly49C is not co-engaged. These results demonstrate that Ly49C can restrict Ly49H-dependent activation upon engagement of m157, but this inhibition is itself regulated in cis by MHC class I molecules.
Next we evaluated whether m157 binding to Ly49C affects the stability of cell conjugates formed between NK cells and their targets thereby preventing subsequent killing. B6 NK cells were expanded in IL2-containing medium and four subsets were sorted based on Ly49C and Ly49H expression (1. Ly49H−,C−, 2. Ly49H−,C+, 3. Ly49H+,C+ and 4. Ly49H+,C−). After further expansion, to eliminate residual fluorescence resulting from the sorting step, NK cells and target cells were labeled using intracellular dyes, CFSE and CMTMR respectively. NK cells were then exposed to various target cells and the formation of stable cell conjugates was assessed. Because B6 NK cells interactions with RMA cells do not result in killing, RMA cells were used as negative controls. RMA m157G1F were excluded as targets because they express Kb that can bind Ly49C and it would have been impossible to determine whether the effects of Ly49C on conjugate stability were due to engagement of m157 or Kb. Instead, we used MHC class I-deficient RMAS m157G1F cells. RMAS cells were also tested; they are killed by B6 NK cells due to “missing self recognition”, as described by [23], which is Ly49H-independent. Each of the four NK cell subsets indicated in Fig. 4A was exposed to target cells for 5 and 20 min and then the percentage of conjugated NK cells was analyzed by flow cytometry. In all four subsets, the fraction of NK cells conjugated with RMA cells remained below 20% (Fig. 4A, left panel). Using RMAS as targets, conjugation levels with NK cells expressing Ly49C increased to 40%, while they remained around 20% with the Ly49C− NK cells (Fig. 4A, middle panel). In B6 mice, Ly49C is involved in NK cell education [15], its expression resulting in a higher reactivity. This is consistent with the increased conjugate rates obtained when Ly49C+ NK cell subsets were exposed to RMAS cells. Interestingly, increased formation of conjugates by Ly49H−,C+ NK cells with RMAS (up to 40%, Fig. 4 A middle panel) were abolished when this subset was exposed to RMAS m157G1F (Fig. 4A right panel). This suggests that binding of m157G1F to Ly49C compensated for the absence of Kb and prevented “missing-self recognition”. Incubation of Ly49H+ NK cells with RMAS m157G1F resulted in high conjugation rates regardless of whether Ly49C was co-expressed (Fig. 4A, right panel). We hypothesized that the similar frequencies of cell conjugates detected in Ly49H+,C+ and Ly49H+,C− B6 NK cells with RMAS m157G1F (which suggested an absence of inhibitory effect mediated by Ly49C) was due to cis-masking by MHC class I molecules. Therefore, we isolated NK cells from B6 β2m ko mice to address this hypothesis. Although β2m ko mice are deficient for MHC class I expression, their NK cells undergo an education process that enables them to respond to MCMV infection as efficiently as wild type mouse cells (our unpublished results and [24]), while they are tolerant toward cells displaying altered MHC I expression. As described above, we sorted four NK cell subsets and exposed them to RMAS and RMAS m157G1F cells. The percentages of NK cells conjugated to RMAS cells remained around 20% for the four subsets. This result was consistent with NK cell tolerance toward MHC class I-devoid RMAS cells (Fig. 4B, left panel) in β2m ko mice. Incubation of RMAS m157G1F with the subsets expressing Ly49H resulted in increased conjugates, although to a lesser extent than seen with B6 NK cells (Fig. 4A and B). In addition, the Ly49H+,C+ NK cell subset demonstrated a reproducibly lower conjugation rate with RMAS m157G1F than the subset expressing Ly49H only (Fig. 4B, right panel).
These results suggest that “missing-self recognition” and Ly49H-dependent signals synergize, resulting in more conjugated NK cells (70% when NK cells were from B6 wt mice, Fig. 4A), while the interactions exclusively due to missing self-recognition (Fig. 4A middle panel) or to Ly49H engagement (Fig. 4B right panel) resulted in more modest conjugation rates (40%). Thus, m157G1F binding to Ly49C disrupts the cell conjugates formed in the context of “missing self recognition”, but a simultaneous engagement of Ly49H can override this inhibition.
Interactions between NK cells and target cells that trigger killing events require the formation of stable cell conjugates [25]. During this process, cytoskeleton polymerization orients the cytotoxic granules toward the immunological synapse. Polymerization of the cytoskeleton actin into F-actin can be detected by microscopy using fluorescent phalloidin. B6 β2m ko NK cells were sorted according to their Ly49H and Ly49C expression as above, then were incubated with CMTMR-labeled RMAS or RMAS m157G1F cells and tested for actin polymerization. NK cell subsets expressing Ly49H but not Ly49C displayed a strong polarization of their cytoskeleton toward RMAS m157G1F (fig. 4C, upper right panel) whereas NK cells co-expressing Ly49H and Ly49C did not (fig. 4C, upper left panel). This process was m157-dependent as incubation with parental RMAS cells did not trigger actin polymerization (fig. 4C, lower panels). Quantification of F-actin within the immunological synapse further supported this result, confirming cytoskeleton accumulation toward m157G1F expressing cells within the Ly49H+,C− NK cell subset (fig. 4D). Thus, m157 binding to Ly49C reduces the level of cytoskeleton actin polymerization and accumulation toward the target cell.
We previously showed that the MCMV K181 laboratory strain in which m157 was substituted by m157G1F (MCMV m157G1F), replicates in B6 mice at a similar rate to the MCMV K181 wt virus [13]. In B6 mice, Ly49C is mostly unavailable for binding m157 due to cis-interaction with H-2 Kb molecules, as illustrated in Fig. 2. We analyzed MCMV m157G1F replication in a H-2d background using Cmv1r mice (Ly49H+) where cis-interactions are weak (Fig. 2). Mice were infected with either MCMV K181 wt, MCMV m157G1F or with a recombinant virus with m157 deleted (MCMV Δm157) and measured viral titers in the spleen, liver, lungs and salivary glands at various times post infection. Viral loads we measured in Cmv1r mice were similar to those observed previously in B6 mice (Fig. 5 and [13]). Only MCMV Δm157, which escapes NK cell surveillance, replicated at high titers while MCMV K181 wt and MCMV m157G1F were similarly controlled. A possible explanation is that, although weak in Cmv1r, cis-masking of Ly49C restrained sufficiently its ability to engage m157G1F, and so impeded its inhibitory effect over Ly49H-dependent activation. In addition, as most Ly49H+ NK cells do not co-express Ly49C in Cmv1r mice (as in B6) and therefore cannot be inhibited by m157, we proposed that this subset ensured virus elimination and compensated for a likely impaired response of the Ly49H+,C+ subset. To test this hypothesis, we generated a Ly49C transgenic mouse strain (B6 Ly49C Tg) in which all NK cells express this inhibitory receptor.
We generated the B6 Ly49C Tg mouse strain as indicated in the Materials and Methods section. Analysis of blood-borne NK cells indicated than over 95% expressed Ly49C (Fig. 6A). Ly49C was also found in over 50% of T cells and in most NKT cells; it was also expressed in a small fraction of B cells (<20%, data not shown), but in none of the other leukocyte populations (data not shown). The transgenic mice and negative littermates had splenic and bone marrow compartments of similar size (Fig. S3). Wild-type and transgenic spleens contained comparable frequencies of B cells, NKT cells and DCs; the T cell fraction was slightly reduced in the transgenic mice, while monocytes/macrophages, neutrophils and eosinophils were slightly increased (Fig. S3). Most importantly, the size of the Ly49H+ NK cell subset was unchanged (Fig. S4). A fraction of Ly49C receptors was found to be masked by cis interactions (Fig. S4).
Ly49C Tg and negative littermates were infected with MCMV m157G1F and viral titers and NK cell activation tested after 4 days. The spleens in transgenic mice were not as enlarged as in wt mice upon infection, consistent with previous studies in Ly49H-devoid mouse strains [26], [27], although the percentage of splenic NK cells was similar (Fig. 6B). The frequency of Ly49H+ NK cells in wt and transgenic mice were also identical (Fig. 6C). CD27 dissects peripheral NK cells into two major subsets: naïve NK cells express CD27, while this marker is downregulated in activated and more mature NK cells [28]. We found that the frequency of the CD27− cells within Ly49H+ NK cells was higher in B6 wt than in Tg mice, while it was similar in Ly49H− NK cells (Fig. 6B). This suggests that Ly49H dependent activation was reduced in Ly49C Tg mice. We also analyzed the expression of the activation marker CD69 in Ly49H positive or negative NK cells. We found that CD69 was more increased in Tg than in wt mice, most particularly in Ly49H− NK cells (Fig. 6B). This sustained activation did not require Ly49H and was most likely due to higher cytokine levels associated with uncontrolled viral proliferation. Viral loads were measured in the spleen, liver and lungs after 4 days. A high dose of virus (5×104 PFU) was used so that viral titers found in wt mouse tissues would be just above the detection limit. Ly49C Tg mice had increased viral titers in the spleen and lungs; there was a trend for increased hepatic loads although the difference was not statistically significant (Fig. 6D). Thus expression of Ly49C on all NK cells results in an inhibitory effect of the immunoevasin m157 leading to higher viral replication in vivo.
Ly49C cis-masking was not as complete in B6 Ly49C Tg as in wt mice (Fig. 6C); we predicted that if more Ly49C receptors were rendered available in the absence of cis-interactions, the inhibitory effect on NK cells would be more intense and would lead to even more severe viral replication.
We crossed B6 Ly49C Tg mice with B6 β2m ko mice and obtained B6 Ly49C Tg, β2m ko double mutant mice. These were used to test in vivo the anti-viral NK cell response in the absence of Ly49C cis-masking. As in B6 Ly49C Tg mice, expression of the Ly49C transgene in the double mutant mice was limited to NK cells, NKT cells, most T cells (Fig. 7A) and a small fraction of B cells (data not shown). Ly49C was not detected elsewhere, and had little impact on the leukocyte population sizes (Fig. S5). B6 Ly49C Tg, β2m ko mice were infected along with B6 β2m ko and B6 Ly49C Tg using 5×104 PFU MCMVm157G1F, as previously. B6 Ly49C Tg, β2m ko mice were highly susceptible to viral infection with 3 out of 4 of these mice succumbing to infection by day 3. Infection with a lower inoculum (5×103 PFU) was performed and the viral titers were measured in the spleen, liver and lungs after 4 days (Fig. 7B). Splenic viral loads were higher in B6 Ly49C Tg, β2m ko mice than in B6 Ly49C Tg mice. Viral replication was completely controlled in B6 β2m ko mice, indicating that the absence of cytotoxic CD8 T cells due to the lack of MHC class I was not responsible for the impaired immune response at this early stage of acute infection.
We concluded that inhibitory Ly49C receptor is most efficient at negatively regulating NK cell functions when it are not engaged through cis-interactions and that targeting of this receptor by the m157 immunoevasin severely impairs the Ly49H-dependent response.
Our results illustrated in Fig. 4A,B indicated that m157G1F binding to Ly49C could modulate “missing self recognition”. We investigated this possibility further in TC1 congenic mice; this strain has a B6 background, but has BALB/c alleles in the Ly49 cluster of the NKC and thus lacks Ly49H, as described in [19]. Missing-self recognition dependent killing in the B6 background is mainly mediated by Ly49C+ NK cells [16]. We hypothesized that RMAS m157G1F target cell killing by TC1 NK cells would be impaired due to Ly49C-mediated inhibition, even though the targets were devoid of MHC class I molecules.
TC1 NK cells were expanded in IL-2 containing medium and tested against RMAS and RMAS m157G1F in an in vitro killing assay. As expected, NK cells killed RMAS targets. By contrast, they spared RMAS m157G1F cells (Fig. 8A, left panel). Addition of excess concentrations of Ly49C-blocking antibody enabled NK cells to kill RMAS m157G1F targets (Fig. 8A, right panel). These results suggest that engagement of Ly49C by m157G1F induces inhibitory signals that compensate for the absence of MHC class I, and thereby prevents NK cell cytotoxicity. We then tested in vivo whether m157G1F could affect the control of MCMV infection mediated by NK cells in a Ly49H-independent manner. Some MCMV genes down-regulate MHC class I expression in infected cells [29]. Previous studies showed that levels of expression of MHC class I can affect NK cell responses during MCMV infection [30], [31]. We hypothesized that NK cell killing of infected cells with reduced MHC class I expression would be prevented by m157G1F binding to Ly49C. We infected Ly49H-devoid mice with MCMV m157G1F or with MCMV Δm157 and compared the replication of the two viruses; infected mice displayed similarly high titers in the spleen, liver and lungs after 4 days (Fig. 8B). We hypothesized that strong inflammatory conditions would promote a better anti-viral NK cell response. Administration of alpha-galactosyl ceramide (αGC) at the time of infection improves NK cell mediated control of MCMV in Ly49H-deficient mice; this effect is NKT cell independent [32]. Next, we administered αGC in combination with MCMVm157G1F or Δm157 and measured viral titers 4 days later. Treatment using αGC resulted in reduced replication of Δm157 virus in all organs (Fig. 8B). TC1 mice treated with αGC had similar viral titers following infection irrespective of which of the two viruses was used (Fig. 8B). We predicted that the absence of an effect associated with m157G1F expression was due to cis-masking of Ly49C. Therefore, we repeated the experiment in CT6 mice (H-2d, Ly49H−) [19], where cis-interactions are weak, as shown in Fig. 2. In this H-2d background, although Ly49C is not responsible for self recognition, m157G1F can bind the inhibitory receptor [13] and induce inhibitory signals. Infection with the Δm157 virus combined with αGC treatment was almost completely controlled, while the protection conferred by αGC was only partial after infection with the m157G1F virus. We also tested B6 β2m ko mice whose NK cell are tolerant to cells with reduced MHC class I expression and where Ly49C availability is not hampered by cis-masking. Only the Δm157 virus was tested in this mouse strain as the m157G1F virus would have triggered a dominant Ly49H-dependent response. Contrary to those measured in TC1 mice, identical Δm157 virus titers were measured with and without αGC treatment (Fig. 8B). These results are consistent with a Ly49C-mediated inhibition of a Ly49H-independent NK cell response; it is achieved upon binding of the m157 immunoevasin and is regulated through cis-interactions.
Altogether, our results demonstrate that engagement of the inhibitory Ly49C receptor by the m157 immunoevasin inhibits NK cell activation leading to failure to control MCMV infection, and that Ly49C inhibitory function is modulated by MHC I molecules through cis-masking.
We previously showed, and confirmed in this study, that m157 from the K181 and the G1F MCMV strains bind Ly49H with similar affinities [13]. Here, we analyzed the kinetics of binding of m157G1F to Ly49H or Ly49C. We found that the kinetics of association of m157G1F to Ly49C was slightly quicker than to Ly49H. In addition, the binding profiles suggested similar affinities for the two receptors, consistent with the results obtained by surface plasmon resonance analysis [14]. Kinetics of dissociation of m157 indicated that interactions with Ly49H were more sustained. This could suggest that after some degree of initial inhibition, Ly49H activation would proceed unchallenged when NK cells coexpress the two receptors. However, since Ly49C is not internalized upon engagement, unlike Ly49H, it is likely that the Ly49C receptors that have dissociated from m157 will establish new interactions and induce inhibitory signals as long as the NK cell-target pairs remain conjugated. Nevertheless, these different dissociation rates could lead to competition in a mixed population where some NK cells express exclusively Ly49H or Ly49C that would ultimately favor Ly49H+ engagement and thus lead to an efficient anti-viral NK cell response. This could explain why infection with either MCMV K181 or MCMV m157G1F resulted in similar viral replication in B6 [13] and in Cmv1r mice as shown here.
Analysis of m157 binding to NK cells revealed another degree of complexity in the parameters at play due to NK cell-borne MHC class I molecules able to interact in cis with Ly49C [21]. These cis-interactions impeded detection of Ly49C+ NK cells using the anti-Ly49C antibody 4LO3311 and prevented m157G1F binding. This lack of binding was reversed by acid treatment. Identical binding patterns obtained with the 4LO3311 antibody and m157 are consistent with the findings that both bind the stalk domain of Ly49C [14], [20]. Ly49C binding to Kb in trans requires a back-folded configuration, while binding in cis requires an extended configuration [33]. Interestingly, our recent findings show that m157G1F binds the Ly49C stalk region when the receptor adopts an extended configuration [14]. As the residues involved in Ly49C interactions with Kb in cis and m157 are distinct [14], one would predict that Ly49C may simultaneously bind both m157 and Kb. However, our data indicate that these two interactions are mutually exclusive, as no m157 binding was detected when Ly49C was engaged by MHC class I in cis, and no inhibition of NK cell functions was achieved. Hence, Ly49C-dependent inhibition of NK cell response upon infection with MCMVm157G1F was not as severe when the receptors could be masked in cis (infection of Ly49C Tg mice vs Ly49C Tg, β2m ko mice). It is possible that structural constraints in the Ly49C stalk domain that are incompatible with binding of m157 result from cis-binding to Kb.
Inhibitory Ly49 receptors, such as Ly49C, contain intracellular immunoreceptor tyrosine-based inhibitory motifs (ITIM). When phosphorylated, ITIMs recruit phosphatases that subsequently suppress phosphorylation-based activation signaling [34]. Trans-binding to inhibitory Ly49 induces ITIM phosphorylation, however, it is likely that binding in cis does not generate such a tonic inhibitory signal. Rather, cis-interactions restrict the pool of Ly49 receptors that are available for functional interaction with ligands on target cells [33]. Our functional assays highlighted an important role for cis interaction in regulating anti-viral NK cell responses. Using B6 NK cells where cis-interactions largely prevent binding of m157 to Ly49C, we detected an un-inhibited Ly49H-dependent NK cell activation characterized by IFN-γ production and degranulation. Conversely, activation of the Ly49H+,C+ NK cell subset was almost completely abolished in B6 β2m ko NK cells where Ly49C/MHC cis-interactions are absent. These functional results correlated with the binding patterns of the m157-Fc and supported a role for MHC class I molecules in controlling Ly49C inhibitory functions in cis. To determine whether Ly49C could control the stability of cell conjugates formed through Ly49H engagement, we measured the formation of conjugates when Ly49H+ NK cells were exposed to m157-expressing targets and analyzed F-actin polymerization and polarization towards the targets; these events indicate NK cell activation and lead to formation of the cytolytic synapse that is required to mediate killing [35], [36]. Fewer conjugates were formed and cytoskeleton polymerization did not occur in B6 β2m ko NK cells expressing Ly49C. Thus, engagement of Ly49C in trans leads to inhibitory signals that destabilize cell conjugation and impairs cytoskeleton reorganization.
Our initial pathogenesis studies showed equivalent viral replication in B6 [13] and Cmv1r mice infected with MCMV expressing m157K181 or m157G1F, despite the fact that m157G1F binds both inhibitory and activating receptors on NK cells, while m157K181 binds only activating receptors. We produced a transgenic mouse in which all NK cells expressed Ly49C. Expression of the activating Ly49H receptor remained unchanged; however, these mice had an increased number of NK cells in the spleen. Despite the presence of more NK cells expressing Ly49H, infection of B6 Ly49C Tg mice resulted in higher viral loads compared to those seen in B6. Analysis of NK cells in infected B6 Ly49C Tg mice showed that only 30% displayed detectable Ly49C, consistent with the regulatory role played in cis by MHC class I molecules. To determine the effect of cis-masking on Ly49C-induced NK cell inhibition, we bred B6 Ly49C Tg with B6 β2m ko mice and generated B6 Ly49C Tg, β2m ko double mutant mice. These mice proved to be highly susceptible to MCMV infection and when infected with the usual dose of 5×104 PFU of virus succumbed to infection within 3 days. This is consistent with a LD50 of 6.4×104 PFU seen in BALB/c mice (that are devoid of Ly49H receptor) infected with the MCMV Smith strain [37] and suggests that Ly49C can completely obliterate Ly49H function in the absence of cis-masking. These results confirm the regulation exerted by cis interactions on NK cells functions in vivo.
Beside their inhibitory function in mature NK cells, inhibitory Ly49 receptors play a crucial role during NK cell education. NK cells acquire functional competence (are licensed) upon Ly49 interactions with MHC class I molecules [15]. Educated cells elicit stronger responses than NK cell subsets devoid of self-specific receptors [15], [16]. In B6 mice, licensing requires engagement of the inhibitory receptor Ly49C and signaling via the intracellular ITIM. Ly49C+ NK cells from B6 wt but not β2m ko mice produce IFN-γ in response to NK1.1 stimulation [15]. However, another study indicated that Ly49H+ NK cells robustly proliferate and produce IFN-γ in both B6 β2m ko and wt mice in response to MCMV infection [24]. We addressed NK cell responsiveness in situations where Ly49C plays a role (B6 origin) or not (B6 β2m ko origin) in NK cell education and confirmed that the NK cell response to MCMV does not require education through Ly49C. Furthermore, it was shown that unlicensed NK cells (Ly49C−) dominate the response to MCMV infection [17]. In B6 mice, Ly49H-dependent NK cell activation was shown to be stronger in the Ly49C/I− subset, while Ly49C/I+ NK cells responded better to MHC class I deficient targets. Our experiments were conducted is similar settings to those described by Orr et al. albeit our target cells expressed m157G1F and not m157Smith; we obtained similar IFN-γ production and degranulation in Ly49C+,H+ and Ly49C−,H+ NK cells from B6 mice. However, using B6 β2m ko NK cells, the Ly49C-expressing subset elicited a limited response, which was increased upon blocking of Ly49C using antibodies, even though it remained weaker than with the Ly49H+,C− NK cells. These results suggest that the Ly49C+ subset had a lower intrinsic activity. The mechanisms responsible for the hyporesponsiveness of the Ly49C+ subset remain to be clarified.
Our data illustrated in figure 8A show a decreased killing of the untransfected RMA-S target cells when excess concentration of Ly49C blocking antibody was added to TC1 NK cells in vitro; this result was surprising because the blocking treatment should not interfere with binding to Ly49C of any ligand present on RMA-S in these settings. Instead, we propose that the decreased killing of RMA-S cells was due to a differential clusterization of the Ly49C receptor in untreated and anti-Ly49C treated NK cells. Indeed, it has been shown that restraints due to cis interactions limit the redistribution of Ly49A to the synapse, whereas unbound Ly49A dampens NK cell activation [38], [39]. Similarly, Ly49C receptors engaged in cis by MHC I molecules are likely to be contained in micro-clusters. Addition of excess concentration of blocking Ab may disrupt cis interactions, resulting in a re-localisation of the freed Ly49C to the immune synapse next to activating receptors where they can interfere with activating signaling pathways. Indeed, ITIM-bearing MHC I specific receptors such as Ly49C require activating receptors to be in the vicinity to perform their inhibitory function [40].
Interestingly, Ly49C engagement by m157G1F had an inhibitory effect on a Ly49H-independent pathway, most likely due to “missing-self recognition” (Fig. 4 and 8). This activation pathway plays a role in NK cell mediated control of MCMV infection [30]. To counter this, MCMV encodes a number of proteins (m04, m06, m152) that interfere with MHC expression [41]–[43]. In mouse backgrounds where Ly49C is involved in NK cell education, such as B6 mice, Ly49C+ NK cells are expected to represent a major subset that responds to infected cells with reduced MHC class I expression. We found that MHC class I-deficient target cells expressing m157 (RMAS m157G1F) avoided killing by B6 NK cells in vitro. Similarly, in vivo, mice lacking Ly49H controlled MCMVΔm157 infection whereas the m157G1F-expressing virus escaped immune surveillance. We propose that infected cells displaying reduced MHC class I expression were ignored by NK cells as a result of m157G1F binding to Ly49C. The response pattern we described here is not without analogies to that reported by Jonjic and colleagues [30] who showed that in BALB/c but not in BALB H-2b mice, NK cells control an MCMV mutant lacking the m04 immunoevasin. NK cell activation was likely due to missing-self recognition, as NK cells in BALB/c β2m ko mice failed to keep viral replication in check. Babic proposed that m04 stabilizes MHC class I and improves binding to Ly49A to a lesser extent to Ly49C. In the absence of m04 inhibitory signals were reduced, which allowed a better elimination of the virus. This resembles our results obtained with MCMVm157G1F: inhibition was induced by m157 through Ly49C, whereas a Δm157 virus (like Δm04 virus) failed to induce inhibitory signals. We verified the m04 gene sequence in the MCMVm157G1F recombinant virus and did not find any alterations (data not shown), which confirmed that the effects we noted were due to m157.
Our study provides a novel understanding of the parameters that direct how efficiently NK cells deal with viral infections. The variety of receptors encoded within the NKC, combined to a high level of sequence polymorphism between mouse strains, are critical determinants of whether NK cells can recognize and eliminate virus-infected cells. In addition, the host H-2 haplotype and MHC sequence polymorphism determine the possibility of interactions in cis that regulate NK cell receptor functions. Our in vivo results highlight why the stochastic expression of distinct NK cell receptors, which result in various NK cell subsets, is so important to host survival. In the ongoing race between the host and CMV, cis-binding of inhibitory NK cell receptors by MHC class I may have provided the host with a mechanism of countering viral immune evasion. Held and colleagues showed the role of cis-interactions between Ly49A and H-2 Dd in education and in regulating NK cell functions (reviewed in [33]). This mechanism could provide a strong evolutionary gain to mice that possess a suitable NK cell repertoire and specific sets of MHC class I molecules. We should also consider the possibility that an evolution-driven selection of individuals having developed Ly49 able to establish cis-interactions with MHC might have resulted from viruses producing immunoevasins able to inhibit NK cells through their self-specific receptors. A recent study demonstrated that the human inhibitory NK cell receptor LIR-1 interacts in cis with HLA-G; this affects the receptor accessibility for HCMV immunoevasin UL18 [44]. These results highlight the relevance of our work and its possible translation into human settings. Our findings provide novel insights into the mechanisms that dictate susceptibility to viral infections, such as HCMV, and thus may also indicate new immunotherapeutic approaches for regulating NK cell responses in settings of infection as well as possibly cancer and transplantation.
All animal experiments were performed after approval by the Animal Experimentation and Ethics Committee of the University of Western Australia (AEC approval # RA/3/300/76 and RA/3/100/1079) and in accordance with the guidelines of the National Health and Medical Research Council of Australia (Australian code of the care and use of animals for scientific purposes 8th edition, 2013; ISBN: 1864965975).
Inbred C57BL/6 (B6) mice were obtained from the Animal Resources Centre (Perth, WA, Australia). BALB.B6-CT6 (H-2d, Ly49H−), BALB.B6-Cmv1r (H-2d, Ly49H+) and B6.BALB-TC1 (H-2b, Ly49H−) congenic mice and B6 Ly49C Tg, B6 Ly49C Tg, β2m ko and B6 β2m ko recombinant mice were bred at the University of Western Australia Animal Care Services (Perth, WA, Australia). Female mice aged 7–12 weeks were used for all experiments, except for virus stock preparation where 3 week-old mice were used. All animal experiments were performed in pathogen-free conditions.
Splenic NK cells isolated from a B6 mouse were expanded in IL-2 for 12 d. Ly49C+ NK cells were sorted and mRNA was isolated using pureLink RNA minikit. A 907 bp long cDNA was amplified by PCR using the following primers Forward primer: CTC CAC CAG CAT CAC TCC G and Reverse primer: CAA GAA ACG AAT AAG GAT CAA CTC. A 850 bp product harboring Sal1 and BamH1 ends was generated by amplifying the ly49cB6 cDNA using the following primers Forward primer: TAT ATG TCG ACC TCC ACC AGC ATC ACT CCG and Reverse primer: TAT ATG GAT CCT TAA TCA GGG AAT TTA TCC. The ly49cB6 insert was inserted between the Sal1 and BamH1 restriction sites into the modified transgene cassette described in [45] and generously provided by Prof. W. Held (Ludwig Center for Cancer research, Lausanne, CH). The cassette was microinjected in fertilized C57BL/6 eggs (WEHI transgenic mouse facilities, Melbourne-Bundoora, VIC). Five founders were obtained and backcrossed to C57BL/6 mice. Appropriate offspring were identified by FACs analysis for Ly49C expression and checked by PCR; selected mice from a same dam were inter-crossed, four transgenic lines were generated among which one that displayed a homogenous Ly49C expression on over 95% NK cells was maintained.
BWZ.36, RMA and RMAS cells were grown in RPMI 1640 containing 5% FCS (Invitrogen), glutamine, sodium pyruvate, penicillin, gentamicin and 2-ME. COS-7 cells and Platinum-E cells were grown in DMEM (Invitrogen) containing 10% FCS, glutamine, sodium pyruvate, penicillin and gentamicin. Primary mouse embryonic fibroblasts (MEFs) were produced by trypsin dispersion of 15–17 day-old embryos from ARC/S mice as previously described [46]. MEFs and M210B4 cells were maintained in MEM 10% NCS (Invitrogen). Immediately prior to infection the culture medium was replaced with media supplemented with 2% FCS. For in vitro NK cell expansion, splenocytes were passed through a nylon wool column and then cultured in DMEM containing 10% FCS, glutamine, non essential amino acids sodium pyruvate, penicillin and gentamicin (all from Invitrogen) and 250 ng/ml recombinant human IL-2 (Cell Sciences, Canton, MA, USA).
The CP197 construct containing the Smith m157-Fc behind the CD150L sequence cloned in the CDM8 vector [9] was kindly provided by Prof. Lewis Lanier (University of California, San Francisco, CA USA). Soluble m157-Fc fusion proteins for the MCMV G1F isolate was constructed as described in [13].
The MCMV laboratory strains used were Smith (originally obtained from E.S. Mocarski, Stanford University) and K181-Perth (K181). Origin of wild-derived isolate G1F has been previously reported in [13]. Tissue culture virus stocks were produced by propagation in M210B4 cells and titers determined by standard plaque assay as previously described [47]. Virulent salivary gland viral (SGV) stocks were prepared by infecting 3-week-old female BALB/c mice i.p. with 1×103 PFU of tissue-cultured passaged virus, and then preparing homogenates of salivary glands at 17 days post infection. Secondary SGV stocks were prepared by infecting 3-week-old female BALB/c mice with 1×103 PFU of the primary SGV stock. Viral DNA was produced from infected MEFs as previously described [48].
In order to produce m157 gene swap viruses we used a ‘BAC recombineering’ approach essentially as previously described [49]. The K181 BAC pARK25 DNA [50] kindly provided by A. Redwood (University of Western Australia). Generation of substitution mutants in which the K181 m157 gene has been replaced with the G1F m157 (MCMVm157G1F) sequences and deletion mutant in which the m157 gene has been deleted (MCMVΔm157) have been described in [13].
BWZ HD12 cells were generously provided by Prof. Wayne Yokoyama (Washington University, St. Louis, MO). BWZ HD12 reporter cells were transduced with the pMX-s-IRES-Ly49CB6 construct to generate BWZ HD12C cells as indicated in [13]. Clones expressing a range of levels of Ly49HB6 and Ly49CB6 were selected following limit dilution cloning, including clones which expressed essentially equal levels of these two receptors. For construction of Ly49H/C chimeric cells which express the Ly49HB6 transmembrane and cytoplasmic domains fused to the extracellular domains of Ly49CB6, Ly49HB6 was PCR amplified using forward primer (A): 5′-TATATCTCGAGATGAGTGAGCAGGAGGTCAC and reverse primer 5′-TCATTGATTTCTTGTTTGTGTTGA. Ly49C B6 was amplified using forward primer 5′-CAAACAAGAAATCAATGAAACTCT and reverse primer (B): 5′-TATATGCGGCCGCTTAATCAGGGAATTTATCC. PCR products Ly49HB6 plus Ly49CB6 were sewed together by PCR using primers A and B above. The resultant PCR product was digested with XhoI and NotI and cloned into pMX-s-IG plasmid. The Plat-E packaging cell line was transfected with the plasmids and resulting retroviral supernatant used to transduce the BWZ-DAP12 reporter cells. Expression of Ly49C was confirmed by FACS.
The antibodies used for analysis for flow cytometry were directed against the following proteins: CD3ε (145.2C.11), CD4 (RM4-5), CD8α (53.6.7), CD11b (M1/70), CD19 (1D3), CD44 (1M7), CD49b (DX5), CD69 (H1-2F3), CD107a (1D4B), H-2 IA/E (M5/114.15.2), IFN-γ (XMG1.2), mouse IgG1 (A851), Ly6C (AL21) and Fluorochrome-conjugated Streptavidin were obtained from BD BioScience (San Diego, CA). Anti-CD27 (LG7F9) and Siglec H (440c) were obtained from eBioScience (San Diego, CA). Anti-CD11c (N418) was obtained from BioLegend (San Diego, CA). Anti-human IgG (Fc part)-Biotin by Jackson ImmunoResearch Laboratories (West Grove, PA). Anti-Ly49H Ab 3D10 producing hybridoma was kindly provided by Prof. W.M. Yokoyama (Washington University, St. Louis, MO) and anti-Ly49C Ab 4LO3311 producing hybridoma was kindly provided by Prof. S. Lemieux (INRS, Laval, Canada). Purified anti-Ly49C and Ly49H antibodies were prepared in-house and conjugated to fluorochrome using antibody-labeling kit according to the manufacturer's recommendations (Molecular Probes).
Staining with m157G1F-Fc fusion proteins was done as follows: cell lines or NK cells were incubated with m157G1F-Fc for 40 min, or shorter time when indicated, followed by anti human IgG (Fc part)-Biotin for 30 min and by a third step with Streptavidin-APC-Cy7 for 15 min. Non specific binding was provided in the absence of the initial incubation with the fusion protein. Live cells were gated based on exclusion of propidium iodide (PI, Sigma). Detection of Ly49C in B6 and Cmv1r NK cells was performed after release of the cis-interactions achieved by a mild acid treatment. Cells were washed twice in PBS and resuspended for 4 min at ambient temperature in Citrate buffer (0.133 M Citric acid, 0.066 M Na2HPO4, pH 3.3) at a density of 10.106 cells/ml. The acid treatment was stopped by adding an excess of PBS 5% FCS; the cells were then stained for flow cytometry analysis. Acid treatment did not affect cell viability as determined by PI exclusion. For the detection of intracellular IFN-γ and of degranulation (LAMP1 expression), cells were cultured at 37°C in presence of monensin (Sigma-Aldrich, St Louis, MO). For LAMP-1 staining, anti-CD107a antibody (1D4B, BD Bioscience) was added into the culture medium as described in [51]. After the culture, cells were stained with fluorochrome conjugated antibodies directed to surface markers, the cells were then fixed with Cytofix/Cytoperm and permeabilized with Perm/Wash according to the manufacturer's recommendations (BD Bioscience). Permeabilized cells were incubated with anti-IFN-γ and anti-Ly49H antibodies for 40 min at room temperature. Intracellular detection of Ly49H was required to detect cells that had internalized this receptor upon exposure to m157-expressing target cells. Antibody-labelled cells were analyzed using a BD FACSCanto (BD Biosciences).
For in vitro assays, splenic cells were isolated and NK cells were expanded for 3–4 days in IL-2 containing medium; they were briefly treated in a mild acidic buffer, as above, and were stained with anti-NK1.1 (PK136), CD3ε (145.2C.11), Ly49H (3D10), Ly49C (4LO3311) fluorochrome-conjugated antibodies for cell sorting using a FACs Aria II cell sorter (BD Biosciences). Sorted NK cells were then placed back in IL-2 containing medium and expanded another 3 d to allow elimination of receptor-bound antibodies.
The recombinant viruses MCMVm157G1F, MCMVΔm157 and parental K181 MCMV virus were used to infect intra peritoneally BALB.B6-Cmv1r mice at the dose of 1×104 PFU and viral titers in the target organs (spleen, liver, lungs and salivary glands) were determined at designated times post inoculation (p.i.). The mice were sacrificed and organs were collected, homogenized in cold MEM 2% Neonatal Calf serum (NCS, Gibco) and centrifuged at 3000 rpm for 15 min at 4°C. The supernatants were stored at −80°C and viral titers were quantified by standard Plaque assay on M210B4 cells by standard plaque assay as described [47].
B6.BALB TC1, BALB.B6-CT6 mice were infected intraperitoneally with 1×104 PFU MCMVm157G1F or MCMVΔm157; B6 β2m ko were infected with 1×104 PFU MCMVΔm157. Infection was done in combination, or not, of an intraperitoneal injection of 2 µg α-galactosylceramide (αGC) (KRN7000, Kirin Brewer Co, Ltd, Japan). Mice were sacrificed 4 days p.i., the spleen, liver and lungs were collected and the viral titers within these organs was determined by standard plaque assay as indicated above.
Sorted NK cells were labeled with 5 µM 5-(and -6)-(((4-chloromethyl) benzoyl) amino) tetramethyl-rhodamine orange (CMTMR, Molecular Probes, Carlsbad, CA) and target cells were labeled with 0.5 µM 5- (and 6-) carboxyfluorescein diacetate succinimidyl ester (CFSE, Molecular Probes) according to the manufacturer's recommendations. 105 NK cells were mixed with 2×105 targets in a final volume of 100 µl of DMEM 5% FCS and were centrifuged for 2 min at 50 g. After incubation at 37°C for 5 or 20 min, the conjugates were gently resuspended in 100 µl ice-cold 4% paraformaldehyde. Percentages of conjugated NK cells were then determined on a BD FACs Canto flow cytometer.
72 to 96 hours after sorting, Ly49H+,C+ and Ly49H+,C− NK cells from B6 β2m ko mice were exposed to CMTMR labelled RMAS (top panels) or RMASm157G1F. 1×105 sorted NK cells were mixed with 2×105 CMTMR labeled targets in a final volume of 100 µl in DMEM 5% FCS and were centrifuged for 2 min at 50 g. After incubation at 37°C for 15 min, conjugates were gently resuspended and allowed to adhere on poly(L-lysine) coated microscopic slides (Lomb Scientific, Taron Point, NSW) for 3 min at 37°C and then were fixed with 4% paraformaldehyde for 10 min. After washing twice in PBS, slides were extracted with 0.1% Triton X-100 in PBS for 5 min at room temperature and were washed twice again in PBS. The slides were then stained with 1 unit of Phalloidin-FITC (Molecular Probes) diluted in PBS 1% BSA for 30 min at 37°C. After two washes in PBS, slides were mounted using Fluoromount (Sigma Aldrich) and analyzed on a fluorescence microscope. Polymerization of cytoskeleton actin as detected by phalloidin was quantified by using ImageJ software (National Institutes of Health; http://rsbweb.nih.gov/ij/); F-actin mean density was measured at the immune (IS) synapse and within the rest of a conjugated NK cell in order to determine the enrichment at the IS. A value close to 1 indicated that no polarization of the cytoskeleton toward the target cell was detected.
IL-2 activated NK cells were tested for cytotoxicity by standard 51Cr release assays [52]. Briefly, target cells were labeled with 100 µCi 51Cr (Perkin Elmer, Rowville, VIC Australia); radioactivity-labelled target cells were mixed with NK cells in triplicate at various E∶T ratios and incubated for 4 h at 37°C. Target cells incubated in Triton 1% provided the total radioactive content (maximal release) while targets cultured alone provided the spontaneous radioactive release values. Specific killing was determined by measuring the radioactivity released in cell-free supernatant according to the following formula: Specific killing = ((sample value – spontaneous release)/(maximal release – spontaneous release)×100). Radioactivity released in the culture medium upon target cell lysis was counted in cell-free culture supernatant using a Wallac Wizard γ-counter (Turku, Finland).
For statistical analysis, the nonparametric Mann-Whitney test was performed using the statistical software package GraphPad Prism (La Jolla, CA, USA).
|
10.1371/journal.pgen.1002791 | GWAS Identifies Novel Susceptibility Loci on 6p21.32 and 21q21.3 for Hepatocellular Carcinoma in Chronic Hepatitis B Virus Carriers | Genome-wide association studies (GWAS) have recently identified KIF1B as susceptibility locus for hepatitis B virus (HBV)–related hepatocellular carcinoma (HCC). To further identify novel susceptibility loci associated with HBV–related HCC and replicate the previously reported association, we performed a large three-stage GWAS in the Han Chinese population. 523,663 autosomal SNPs in 1,538 HBV–positive HCC patients and 1,465 chronic HBV carriers were genotyped for the discovery stage. Top candidate SNPs were genotyped in the initial validation samples of 2,112 HBV–positive HCC cases and 2,208 HBV carriers and then in the second validation samples of 1,021 cases and 1,491 HBV carriers. We discovered two novel associations at rs9272105 (HLA-DQA1/DRB1) on 6p21.32 (OR = 1.30, P = 1.13×10−19) and rs455804 (GRIK1) on 21q21.3 (OR = 0.84, P = 1.86×10−8), which were further replicated in the fourth independent sample of 1,298 cases and 1,026 controls (rs9272105: OR = 1.25, P = 1.71×10−4; rs455804: OR = 0.84, P = 6.92×10−3). We also revealed the associations of HLA-DRB1*0405 and 0901*0602, which could partially account for the association at rs9272105. The association at rs455804 implicates GRIK1 as a novel susceptibility gene for HBV–related HCC, suggesting the involvement of glutamate signaling in the development of HBV–related HCC.
| Previous studies strongly suggest the importance of genetic susceptibility for hepatocellular carcinoma (HCC). However, the studies about genetic etiology on HBV–related HCC were limited. Our genome-wide association study included 523,663 autosomal SNPs in 1,538 HBV–positive HCC patients and 1,465 chronic HBV carriers for the discovery analysis. 2,112 HBV–positive HCC cases and 2,208 HBV carriers (the initial validation), and 1,021 cases and 1,491 HBV carriers (the second validation), were then analyzed for validation. The fourth independent samples of 1,298 cases and 1,026 controls were analyzed as replication. We discovered two novel associations at rs9272105 (HLA-DQA1/DRB1) on 6p21.32 and rs455804 (GRIK1) on 21q21.3. HLA-DRB1 molecules play an important role in chronic HBV infection and progression to HCC. The association at rs455804 implicates GRIK1 as a novel susceptibility gene for HBV–related HCC, suggesting the involvement of glutamate signaling in the development of HBV–related HCC.
| Hepatocellular carcinoma (HCC) is the sixth common cancer and the third common cause of cancer mortality worldwide [1]. The incidence rate of HCC varies considerably in the world, with the highest in East, Southeast Asia and Sub-Saharan Africa, and China alone accounts for approximately half of HCC malignancies [1], [2]. Major risk factors for HCC are chronic infections with the hepatitis B or C viruses, and exposure to dietary aflatoxin B1. Hepatitis B virus (HBV) infection is particular important, because of its coherent distribution with the HCC prevalence [1], [2]. However, it is known that only a minority of chronic carriers of HBV develop HCC [3], and the chronic HBV carriers with a family history of HCC have a two-fold risk for HCC than those without the family history [4], strongly suggesting the importance of genetic susceptibility for HBV-related HCC.
A number of candidate genes were investigated by genetic association studies to evaluate their roles in the susceptibility to HCC [5]. However, the findings from these studies are inconclusive due to moderate evidence and lack of independent validation. Recently, a genome-wide association study (GWAS) of HBV-related HCC was performed [6], in which 355 HBV–positive HCC patients and 360 chronic HBV carriers were used for the genome-wide discovery analysis, and the top 45 SNPs from the discovery analysis were further evaluated in additional 1,962 HBV–positive HCC patients and 1,430 controls (both chronic HBV carriers and population controls) as well as 159 trios. The study identified KIF1B as a novel susceptibility locus (top SNP rs17401966) on 1p36.22. Further study with better design and bigger sample size was recommended for identifying additional susceptibility loci for HCC [7], [8]. These motivate us to carry out a GWAS with a large sample size in Chinese population to discover novel susceptibility loci for HCC.
We performed a genome-wide discovery analysis by analyzing 523,663 common autosomal SNPs in two independent cohorts of the Han Chinese: 480 cases and 484 controls from central China and 1058 cases and 981 controls from southern China (Table S1 and Figure S1). The principal component analysis (PCA) confirmed all the samples to be Chinese, but indicated moderate genetic mismatch between the cases and controls in the cohort of southern China (Figure S2). To minimize the effect of population stratification, we performed the genome-wide association analysis using PCA-based correction for population stratification. After the adjustment by the first principal component, the λgc of the genome-wide association results is 1.013 for the cohort of central China, 1.003 for the cohort of southern China and 1.012 for the combined samples. Furthermore, for all the three genome-wide analyses of central, southern and combined samples, the quantile-quantile (QQ) plot of the observed P values revealed a good overall fit with the null distribution (Figure S3). Taken together, these results clearly indicate that the final association results from our genome-wide discovery analysis are free of inflation effect due to population stratification.
The genome-wide discovery analysis revealed multiple suggestive associations (P<10−5) on 2q22.1, 6p21.32, 11p15.1 and 20q12 (Figure S4 and Table S2). To validate these findings, 39 SNPs were selected according to their overall association evidence in three GWAS analyses as well as their consistencies of association between the two independent GWAS samples (Central and Southern China) (see the Methods for the selection criteria). The 39 SNPs were genotyped in additional 2,112 HBV–positive HCC cases and 2,208 HBV carriers (Phase I validation) (Table S1). Of the 39 SNPs, only 3 (rs9272105 on 6p21.32, rs11148740 on 13q21.32 and rs455804 on 21q21.3) were validated, showing consistent association between the GWAS discovery and Phase I validation samples (Table S3). These 3 SNPs were then genotyped in additional 1,021 HBV–positive HCC cases and 1,491 HBV carriers (Phase II validation). The Phase II validation analysis (Table 1) confirmed the associations at rs9272105 on 6p21.32 (OR = 1.41, P = 7.63×10−9) and rs455804 on 21q21.3 (OR = 0.83, P = 3.63×10−3), but not the association at rs11148740 on 13q21.32 (Table S3).
For both rs9272105 and rs455804, no heterogeneity of associations were observed among the GWAS and validation samples (P>0.05), and the associations in the combined GWAS and validation samples achieved genome-wide significance (P<5.0×10−8) (rs9272105: OR = 1.30, P = 1.13×10−19 and rs455804: OR = 0.84, P = 1.86×10−8) (Table 1). As a replication, these two SNPs were genotyped in the fourth independent samples of 1,298 cases and 1,026 controls from central China, which further confirmed the associations at rs9272105 (OR = 1.25, P = 1.71×10−4) and rs11148740 (OR = 0.84, P = 6.92×10−3) (Table 1). When combining all the five groups of samples, the two SNPs resulted in a 28% increased, and a 16% decreased risk for HCC development (rs9272105: OR = 1.28, P = 5.24×10−22 and rs455804: OR = 0.84, P = 5.24×10−10) (Table 1), respectively. The associations at the two SNPs remained genome-wide significant after adjusting for age, gender, smoking and drinking (Table S4A). Furthermore, stratification analysis by age, gender, smoking and drinking status revealed similar ORs for rs9272105 and rs455804 among subgroups, except that the association at rs9272105 showed a stronger effect in the non-smoking group than the smoking one (OR = 1.38 vs. 1.19, P for heterogeneity = 0.004) (Table S4B). Pair-wise interaction analysis among these two SNPs, smoking and drinking status did not reveal any significant interaction (data not shown). The samples used in the GWAS, validation and replication analyses are summarized in Table S1, and the multi-stage design of the whole study is shown in Figure S5.
We further investigated the association of HLA alleles in our GWAS samples through imputation. After QC filtering (see the Methods), 37 HLA alleles were successfully imputed, and 5 alleles showed nominal association (P<0.05) (Table S5 and Table 2). Further stepwise conditional analysis revealed that only two DRB1 alleles showed independent associations (DRB1*0405: OR = 0.69, P = 6.18×10−4; DRB1*0901: OR = 0.82, P = 3.62×10−3) (Table 2). Conditioning on rs9272105 could abolish the associations of the DRB1 alleles, and conditioning on the DRB1 alleles could weaken, but not eliminate, the association at rs9272105 (Table 2). The haplotype analysis of rs9272105 and the two DRB1 alleles revealed consistent result, showing that both the DRB1 alleles sit on the haplotypes carrying the protective G allele of rs9272105 (Table S6). Taking together, there seems to be additional risk effect beyond the ones carried by the DRB1 alleles.
We further explored whether the SNPs rs9272105 and rs455804 play any role in HBV infection. First, we compared the frequencies of these 2 SNPs between 408 non-symptomatic HBV carriers and 521 symptomatic chronic HBV patients from southern China (GWA scanned). The analysis revealed a protective effect at rs9272105 (OR = 0.80, P = 1.67×10−2) on the development of symptomatic chronic hepatitis B, but no association at rs455804 (Table S7A). Furthermore, we genotyped these 2 SNPs in 1,344 individuals with HBV nature clearance and compared their frequencies with those in 4,183 asymptomatic HBV carriers (all from the Central China). The analysis also revealed a protective association at rs9272105 for HBV chronic infection (OR = 0.88, P = 3.78×10−3) (Table S7B).
SNP rs9272105 is located between HLA-DQA1 and HLA-DRB1 on 6p21.32 (Figure 1A). SNP imputation in the GWAS discovery samples revealed additional SNPs showing association, but rs9272105 remained to be the top SNP within the region (Figure 1A). The residual association at rs9272105 after conditioning the association effects of the HLA alleles DRB1*0405 and *0901 suggests that there may be additional risk effect beyond the DRB1 alleles in Chinese population. The associations of the DRB1 alleles revealed by this study are consistent with the previous reports that HLA-DQ/DR alleles associated with HCC risk [9], [10]. In addition, we investigated the previously reported HBV infection-associated SNPs rs3077, rs9277535, rs7453920, and rs2856718 within the HLA DP/DQ region [11], [12] with HCC development in our GWAS samples. By imputation, we found the evidence of the association at rs9277535 with HCC (rs9277535: OR = 0.85, P = 7.9×10−3). However, there is no linkage disequilibrium (LD) between rs9277535 and our SNP rs9272105 (r2 = 0.016 according the HapMap CHB+JPT samples), suggesting that the associations at rs9277535 and rs9272105 may be independent.
The HLA-DQ locus has also been shown to be associated with HCV-related HCC in a Japanese GWAS (rs9275572, OR = 1.30, P = 9.38×10−9) [13]. SNPs rs9275572 and rs9272105 are 79 kb away from each other and in weak LD (D′ = 0.43, r2 = 0.08 in the HapMap CHB samples). The SNP rs9275572 did not show any association with HBV-related HCC in our GWAS discovery samples (OR = 0.93, P = 0.24) (Table S8 and Figure S6B). In addition to HLA-DQ, MICA (rs2596542) on 6p21.33 and DEPDC5 (rs1012068) on 22q12.3 were also identified as independent susceptibility loci for HCV-related HCC in Japanese population [13], [14]. But, our GWAS discovery analysis did not reveal any supportive evidence for these two loci (rs2596542: OR = 1.06, P = 0.36; and rs1012068: OR = 1.06, P = 0.37) (Table S8 and Figure S6C and S6D). We also evaluated the power of our GWAS discovery samples and found that our samples should have sufficient power for detecting the previously reported associations at rs9275572 (power = 94%), rs2596542 (power = 92%) and rs1012068 (power = 94%). Taken together, the disparity of associations may suggest the different genetic background of the susceptibilities for HCV- and HBV-related HCC. Further studies will be required to confirm the genetic heterogeneity of HCV- and HBV-related HCC.
The association of rs9272105 (HLA-DQA1/DRB1) with HBV infection is consistent with the extensive reports on the association of HLA-DRB1 with HBV infection where both protective and risk DRB1 alleles for HBV infection and outcome were identified [11], [12], [15]–[19]. Intriguingly, our study has revealed that the variant allele of rs9272105 showed a protective effect for HBV infection (OR = 0.88) and the progression to chronic symptomatic hepatitis B, but a risk effect for the development of HCC (OR = 1.30). Further studies will be needed to demonstrate whether the opposite associations of HBV infection and HBV-related HCC progression at rs9272105 are due to different causal variants within the HLA class II region.
SNP rs455804 is located within the first intron of GRIK1 that is the only gene within the LD region of the association (Figure 1B), strongly implicating GRIK1 as a novel susceptibility gene for HBV-related HCC. SNP imputation of the region did not reveal any SNPs that showed stronger association than rs455804. GRIK1 encodes CLUR5, which is involved in the glutamate signaling, as one of the ionotropic glutamate receptor, kainite 1 protein (GLUR5), a subunit of ligand-activated channels and involved in glutamate signaling. Our discovery of the association of GRIK1 with HCC has enhanced the emerging evidences for the important role of glutamate signaling pathway in cancer development. Glutamate has been shown to play a central role in the malignant phenotype of gliomas through multiple molecular mechanisms [20]. Inhibition of glutamate release and/or glutamate receptor activity can inhibit the proliferation and/or invasion of tumor cells in breast cancer [21], laryngeal cancer [22], and pancreatic cancer [23], and ionotrpic glutamate receptor (GLUR6) was also suggested to play a tumor-suppressor role in gastric cancer [24]. Recently, the exome sequencing analysis revealed that GRIN2A (encoding the ionotrpic glutamate receptor (N-methyl D-aspartate) subunit 2A) was mutated in 33% of melanoma tumors, clearly indicating the involvement of glutamate signaling in melanoma development. Finally, SNPs within GRIK1 have also been found significantly associated with paclitaxel response in NCI60 cancer cell lines, and may play a role in the cellular response to paclitaxel treatment in cancer [25]. Consistent with the previous observations, our discovery of GRIK1 as a HBV-related HCC susceptibility gene has suggested the importance of glutamate signaling in HBV-related HCC development, and, although still speculative, has highlighted the glutamate signaling pathway as a potentially novel target for the treatment of HCC.
We also assessed the previously reported susceptibility locus KIF1B on 1p36.22 (rs17401966) for HBV-related HCC [6]. Our GWAS discovery analysis did reveal the consistent result for the association at rs17401966, but the strength of association in our GWAS discovery sample (OR = 0.90) is much weaker than the previously reported one (OR = 0.61) (Table S8). SNP imputation in our GWAS discovery samples did not reveal any stronger association than the association at rs17401966 within the LD region surrounding the 1p36.22 locus (Figure S6A).
Previous studies have clearly shown the existence of subpopulation structure of Chinese Han population along the north-south axis, and further demonstrated that geographic matching can be used as a good surrogate for genetic matching, and PCA-based correction is very effective in controlling the inflation effect of population stratification [26]. In the current study, all the cases and controls were matched by their geographic origin of residence. Moreover, the GWAS discovery samples were from central and southern China, while all the validation and replication samples were from central China. Our PCA analysis indicates that while there was mild population stratification in the sample of southern China, the cases and controls from central China were well matched without any indication of population stratification. In our study, the PCA-based correction was used in the GWAS analysis, and all the validation and replication analyses were from central China. Therefore, our findings should be free of adverse effect of population stratification in Chinese population.
In conclusion, the current GWAS identified two biologically plausible, novel loci on 6p21.32 and 21q21 for HBV-related HCC. These findings highlight the importance of HLA-DQ/DR molecules and glutamate signaling in the development of HBV-related HCC.
The genome-wide discovery analysis was performed by genotyping 731,442 SNPs in 1,575 HBV positive HCC patients and 1,490 HBV positive controls derived from two independent case-control cohorts of 500 cases and 500 controls from Central China (Shanghai) and 1,075 cases and 990 controls from Southern China (Guangdong). The first stage validation samples included 2,112 HBV–positive cases and 2,208 HBV–positive controls recruited from Jiangsu. The second stage validation samples consisted of 1,021 HBV–positive cases and 1,491HBV carriers recruited from Shanghai. The replication samples of 1,298 HBV–positive cases and 1,026 HBV carriers were recruited from Central China (Shanghai and Jiangsu). (Table S1 and Figure 1) All the samples are Han Chinese and partially participated in the previously published studies [27], [28]. The diagnosis of HCC was confirmed by a pathological examination and/or α-fetoprotein elevation (>400 ng/ml) combined with imaging examination (Magnetic resonance imaging, MRI and/or computerized tomography, CT). Because HCV infection is rare in Chinese, we excluded HCC with HCV infection. Cancer-free HBV+ control subjects from central China were recruited from those receiving routine physical examinations in local hospitals or those participating in the community-based screening for the HBV/HCV markers and frequency-matched for age, gender, and geographic regions to each set of the HCC patients. Almost all these community-based controls are asymptomatic HBV carriers. Similarly, cancer-free control subjects from southern China are all HBV+, and 408 of them were asymptomatic HBV carriers and 521 were symptomatic chronic hepatitis B patients. All the HBV+ controls were positive for both HBsAg and antibody to hepatitis B core antigen (anti-HBc), and negative for anti-HCV.
We also recruited a HBV natural clearance cohort form Jiangsu Province (Zhangjiagang and Changzhou cities) through a population based screening for the HBV/HCV markers in 2004 and 2009, respectively (58,142 persons). Subjects with HBV natural clearance were negative for HBsAg and anti-HCV, positive for both antibody to hepatitis B surface antigen (anti-HBs) and anti-HBc. About 9,610 subjects with HBV natural clearance were identified. No history of hepatitis B vaccination was reported for these people. Then, we randomly selected 1,344 HBV natural clearance people without self-reported history of cancer in the current study. The age for the 1,344 people were 52.6±10.2 years, and 217(16.2%) were females.
We collected smoking and drinking information through interviews. Those who had smoked an average of less than 1 cigarette per day and less than 1 year in their lifetime were defined as nonsmokers; otherwise, they were considered as smokers. Individuals were classified as alcohol drinkers if they drank at least twice a week and continuously for one year during their lifetime; otherwise, they were defined as nondrinkers. At recruitment, the informed consent was obtained from each subject, and this study was approved by the Institutional Review Boards of each participating institution.
We performed standard quality control on the raw genotyping data to filter both unqualified samples and SNPs. The samples with overall genotype completion rates <95% were excluded from further analysis (26 subjects). Eight subjects were excluded as they showed discrepancy between the recorded and genetically inferred genders. An additional 21 duplicates or probable familial relatives were excluded based on the IBD analysis implemented in PLINK (all PI_HAT>0.25). SNPs were excluded when they fit the following criteria: (i) not mapped on autosomal chromosomes; (ii) had a call rate <95% in all GWA samples or in either of Central cohort study or Southern study samples; (iii) had minor allele frequency (MAF) <0.05 in either of Central cohort study or Southern study samples; and (iv) genotype distributions deviated from those expected by Hardy-Weinberg equilibrium (P<1×10−5 in either of Central cohort study or Southern study samples). We detected population outliers and stratification using a principal component analysis (PCA) based method. After removing MHC SNPs on chromosome 6 from 25–37 Mb, PCA was performed by using common autosomal SNPs with low LD (r2<0.2) in the reference samples of the HapMap project (YRI (n = 90), CEU (n = 90), CHB (n = 45) and JPT (n = 44)) as the internal controls and our 3,010 participants of the GWAS discovery samples (after removal of samples with low call rates, ambiguous gender, and familial relationships). Projection onto the two multidimensional scaling axes is shown in Figure S2A. 7 outliers (more than 6 standard deviations) were identified and excluded. Finally, 523,663 autosomal SNPs in 1,538 cases and 1,465 controls, consisting of 480 cases and 484 controls from Central China and 1,058 cases and 981 controls from Southern China, were retained for association testing (Table S1).
SNPs for the first stage validation were selected based on the following criteria: (i) SNP had P joint≤1.0×10−4 in the analysis of the combined GWA samples or either the Central China sample or the Southern China sample, and had a consistent association in the two participant studies, meaning that the ORs from the two samples are both either above or below 1; (ii) only SNP with the lowest P value was selected when multiple SNPs showed a strong LD (r2≥0.8). As a result, a total of 39 SNPs were included in the first stage validation. 3 SNPs that were significantly associated with HCC risk in the first validation stage were further genotyped in the second stage validation samples. Genotyping in the two validation samples were done by using the iPLEX platform (Sequenom) or the TaqMan assays (Applied Biosystems). The primers and probes were available upon request (Table S9). Laboratory technicians who performed genotyping experiments were blinded to case/control status. For TaqMan assay, ten percent of random samples were repeated, and the reproducibility was 100%. The 2 validated SNPs were genotyped in another independent replication using the same method.
Population structure was evaluated by the PCA in the software package EIGENSTRAT 3.0 [26]. PCA revealed one significant (P<0.05) eigenvector which was included in the logistic regression with other covariates of age, gender, smoking and drinking status for both the genome-wide discovery analysis and the joint analysis of the combined discovery and replication samples. Ancestral origin checking by PCA confirmed all the samples to be Han Chinese and further demonstrated moderate genetic stratification between the cases and the controls of the Southern cohort (Figure S2). The genome-wide association analysis was therefore performed in logistic regression using PCA-based correction for population stratification and by treating the samples of two cohorts as independent studies. The genomic-control inflation factor (λgc) after adjustment by the first PC was calculated for the Central cohort samples (λgc = 1.013), the Southern cohort samples (λgc = 1.003) and the combined GWAS discovery samples (λgc = 1.012). Consistently, the QQ plot of the observed P values also showed a minimal inflation of genome-wide association results due to population stratification (Figure S3).
Statistical analyses were performed by using PLINK 1.07 [29] and R 2.11.1. The Manhattan plot of −log10P was generated using Haploview (v4.1) [30]. Untyped genotypes were imputed in the GWAS discovery samples by using IMPUTE2 [31] and the haplotype information from the 1000 Genomes Project (ASN samples as the reference set) and HapMap3 (CHB and JPT samples as the reference samples). The regional plot of association was created by using an online tool, LocusZoom 1.1. P value was two-sided, and OR presented in the manuscript was estimated by using additive model and logistic regression analyses if not specified.
To impute classical HLA alleles, we used 180 phased haplotypes from the HapMap CHB and JPT samples as our reference panel. This panel comprised dense SNP data and HLA allele types at 4-digit resolution for the HLA class I (HLA-A, B, C) and II (DQA1, DQB1 and DRB1) genes as previously described [32]. Genotypes, probability and allelic dosages were then imputed separately in the two discovery samples of Central and Southern Chinese using the BEAGLE program. Association testing was performed by using a logistic regression model on the best-guessed genotypes and allelic dosages. The results were checked for consistency between the two methods, and the results from best-guessed genotypes were presented.
|
10.1371/journal.pgen.1006854 | A genetic framework controlling the differentiation of intestinal stem cells during regeneration in Drosophila | The speed of stem cell differentiation has to be properly coupled with self-renewal, both under basal conditions for tissue maintenance and during regeneration for tissue repair. Using the Drosophila midgut model, we analyze at the cellular and molecular levels the differentiation program required for robust regeneration. We observe that the intestinal stem cell (ISC) and its differentiating daughter, the enteroblast (EB), form extended cell-cell contacts in regenerating intestines. The contact between progenitors is stabilized by cell adhesion molecules, and can be dynamically remodeled to elicit optimal juxtacrine Notch signaling to determine the speed of progenitor differentiation. Notably, increasing the adhesion property of progenitors by expressing Connectin is sufficient to induce rapid progenitor differentiation. We further demonstrate that JAK/STAT signaling, Sox21a and GATAe form a functional relay to orchestrate EB differentiation. Thus, our study provides new insights into the complex and sequential events that are required for rapid differentiation following stem cell division during tissue replenishment.
| Adult tissue/organ function is maintained by stem cells. Key question in stem cell biology is how the pool of stem cells can be robustly expanded yet timely contracted through differentiation according to the need of a tissue. Over the last years, the mechanisms underlying stem cell activation have been extensively studied, while the genetic control of progenitor differentiation, especially during regeneration, is still poorly understood. Using the fruit fly Drosophila midgut as model, we investigate the cellular changes and the genetic program required for efficient progenitor differentiation during intestinal regeneration. We first detect the presence of extended cell-cell contact between a stem cell and its differentiating daughter in regenerating intestine, compared to homeostatic conditions. The extended cell-cell contact is consolidated by cell adhesion molecules and enhances Notch signaling in the differentiating progenitors leading to their fast differentiation into enterocytes. We further uncover a genetic program, involving the JAK/STAT and Dpp signaling, the Sox21a and GATAe transcription factors, which acts in the differentiating progenitors to instruct their terminal differentiation. Thus, our study presents an integrated view of stem cell differentiation during tissue regeneration and the findings here are likely to apply to mammals.
| In metazoans, the digestive tract supports organismal growth and maintenance. Genetic disorders or microbial dysbiosis that prevent the digestion and absorption of nutrients are major causes of morbidity and mortality in humans. In mammals, mature intestinal cells are short-lived and constantly replaced by newborn differentiated cells. This is ensured by the existence of fast-cycling intestinal stem cells (ISCs) [1]. Although ISC division is important, failure in or improper differentiation into mature intestinal cells can equally cause a wide range of disorders that compromise organ function, such as intestinal cancer [2] and microvillus inclusion disease [3].
There is a great extent of similarity in intestinal functions and maintenance between flies and mammals [4]. Over the past decade, research has revealed the extreme plasticity of the Drosophila ISCs. For instance, stem cell activity and epithelial renewal can be adjusted in response to i) changes in nutrient availability [5–7], ii) physiological requirements for reproduction [8–10], iii) aging [11–13], iv) intestinal damage or infection [14–16], and v) body injury [17, 18]. Thus, both local and remote signals coordinate ISC activity to ensure intestinal homeostasis.
In the adult Drosophila midgut, ISCs differentiate into either polyploid absorptive enterocytes (ECs) or diploid secretory enteroendocrine cells (EEs) (Fig 1A). Recent studies indicated that EC and EE are generated through distinct mechanisms [19–21]. A post-mitotic and intermediate differentiating cell called enteroblast (EB) is differentiated into EC in a Notch-dependent manner [22, 23], while the production of EE through a so-far not molecularly characterized enteroendocrine mother cell (EMC) requires only low levels of Notch signaling [24]. ISCs and EBs (referred to as progenitor cells) reside basally next to the visceral muscles, while ECs cover the apical brush border (Fig 1B). In both flies and mammals, Notch signaling plays the same central roles in the choice of an absorptive or secretory fate in the intestinal lineages [25] [26]. Drosophila ISCs express the Notch ligand, Delta (Dl), which turns on Notch activity in its sibling cells for EB fate commitment [23, 25] (Fig 1C). Moreover, JAK/STAT signaling [14, 27], the transcription factors Escargot (Esg) [28–30], Sox21a [31, 32], GATAe [33], and Dpp signaling [34–36] have recently been shown to regulate progenitor differentiation. While stem cell proliferation has been the focus of most studies, the cellular mechanisms that mediate proper conversion of the expanded stem cell pool into mature intestinal cells especially during regeneration, are currently missing. Moreover, an integrated view of intestinal regeneration has not been established.
Here, we investigate the cellular and genetic basis underlying efficient differentiation of progenitor cells during intestinal regeneration. Our data uncover that enhanced cell-cell contact between an ISC and its differentiating daughter, consolidated by cell adhesion molecules, is required for efficient Notch signaling and rapid progenitor differentiation into EC during regeneration. We further identify a regulatory cascade involving, sequentially, JAK/STAT signaling, Sox21a and GATAe, that functions in EBs and is required for rapid differentiation. Our integrated study of intestinal regeneration provides new insights into stem cell differentiation that likely apply to other systems.
To understand the molecular and cellular mechanisms underlying intestinal regeneration, we analyzed the behaviors of progenitors in the gut of flies orally infected with the gram-negative bacterium Erwinia carotovora carotovora 15 (Ecc15). Oral infection with Ecc15 causes damage to the intestinal epithelium, which is quickly repaired through activating ISC proliferation and progenitor differentiation to maintain tissue integrity [15]. Unless otherwise noted, we focused our study on the anterior midgut, a region where the relatively low overall cell density allows better identification of individual cells. Interestingly, progenitor pairs with extended contact were observed in intestines following infection (Fig 1D and 1E), visualized by β-catenin staining (Armadillo (Arm) in Drosophila). Most differentiating EBs were in extended contact with at least one cell with strong esg>GFP signal (Fig 1F). Furthermore, ingestion of dextran sulfate sodium (DSS), which damages the intestinal epithelium and activates regeneration [16], also led to formation of progenitor pairs with extended cell-cell contact (S1A Fig). Since extended progenitor contact was not observed in basal homeostatic conditions, these results indicate that increased progenitor contact area is a general feature of the regenerating intestine.
Both Ecc15 infection and DSS treatment activate stem cell proliferation. To investigate if division of stem cells is required for the formation of extended progenitor contact, we used colcemid, a microtubule-depolymerizing drug, which blocks dividing cells in metaphase. The presence of colcemid suppressed the formation of extended contact that is normally induced by Ecc15 infection (S1B and S1C Fig). This suggests that during regeneration stem cells first proliferate before generating progenitors with increased cell-cell contact. However, the formation of extended cell-cell contact was not affected in Sox21a mutant gut where EB to EC differentiation is blocked (S1D and S1E Fig). Moreover, clusters of ISCs induced upon expression of a Notch RNAi using the progenitor specific driver esgTS also showed extended contact (S1F Fig). Thus, the formation of increased cell contact during regeneration appears to be a stem cell intrinsic behavior, which occurs independently of developmental signals regulating terminal differentiation.
Dl/Notch signaling plays a central role in determining the ISC and EB cell fate and is further involved in EB differentiation into EC. Since this signal transduction requires cell-cell contact between the signaling sending and receiving cells, the change in contact area likely affects Dl/Notch signaling dynamics [37]. We hypothesized that the extended progenitor contact observed in epithelial damage-induced regenerating intestines could promote efficient differentiation by enhancing Dl/Notch signaling. This led us to further analyze Notch signaling state in progenitor pairs of regenerating intestines by applying cell-type specific markers.
To unambiguously identify ISCs, we used an endogenous Dl-GFP fusion line with an ISC restricted expression [38] (Fig 1G). EBs were visualized by Su(H)-lacZ, a reporter gene of Notch activity [39, 40]. We first confirmed the increase in progenitor contact upon bacterial infection (Fig 1G and 1H). In line with previous results [5, 25, 41], ISCs in unchallenged conditions largely undergo asymmetric division, which generates another self-renewing ISC (Dl-GFP+) and a committed EB (Su(H)-lacZ+) (Fig 1G and 1I). When we quantified all progenitor combinations, including single Dl-GFP+ cells, Dl-GFP+—Dl-GFP+ pairs, Dl-GFP+—Notch+ pairs, Notch+—Notch+ pairs and single Notch+ cells, in both unchallenged and bacteria-infected (Ecc15, 12 hours post infection) intestines, we uncovered a significant increase of the Dl-GFP+—Dl-GFP+ pairs in infected guts (Fig 1I and 1J). Notably, the ratio of Dl-GFP+—Dl-GFP+ pairs increased from 5% in unchallenged intestines to around 40% in infected guts. This change was accompanied by a reduction in the proportion of single Dl-GFP+ cells in regenerating intestines, suggesting that most of them had recently divided. We also observed a drop in the ratio of Dl-GFP+—Notch+ pairs, from 62% to 42%. Collectively, these data are consistent with the notion that increased contact directly arises from newborn progenitor pairs that have just completed mitosis.
Interestingly, 57% of the Dl-GFP+—Dl-GFP+ pairs had one cell showing weak but specific Notch activity in Ecc15-infected intestines as revealed by the expression of the Su(H)-lacZ reporter (Fig 1K–1M; S2A and S2B Fig). Use of an antibody against Dl confirmed that both cells, including the one with weak Su(H)-lacZ expression, were indeed stem cells as defined by the expression of the Dl marker (S2C Fig). We further excluded the possibility of EE differentiation, since the EE marker Prospero (Pros) was never observed in such Dl-GFP+—Dl-GFP+ pairs with Notch activity (n>50) (S2D Fig). Importantly, ISCs undergoing mitosis were never found to express Su(H)-lacZ reporter (n>30) (S2E Fig), indicating that Notch activity was established in one cell of a newly formed Dl-GFP+—Dl-GFP+ pair after mitosis (S2F and S2G Fig). Although Dl-GFP+—Dl-GFP+ pairs formed during infection can arise either from single Dl-GFP+ cells or from Dl-GFP+—Notch+ progenitor pairs, these results support a symmetric expansion of the Dl+ cell pool followed by diverting a subset of them to be committed into EB and quickly differentiated, likely due to the presence of the extended cell-cell contact (Fig 2L).
We and others have previously shown that the transcription factor Sox21a is both necessary and sufficient for the differentiation of EB to EC [31, 32]. This was supported by the observations that ISC progenies are blocked at the EB stage in the absence of Sox21a, while overexpressing Sox21a in progenitors induced their precocious differentiation into mature EC. However, it is not yet established through which mechanisms Sox21a regulates progenitor differentiation, especially during regeneration following intestinal damage.
To further analyze the role of Sox21a in EB differentiation, we first monitored its expression levels in intestinal cells at different stages of their differentiation using a sGFP (superfolder green fluorescent protein)-tagged Sox21a transgene that is controlled by its own regulatory sequences [42]. Since it has been shown that Sox21a is expressed in both ISC and EB, cells that express Sox21a-GFP but not Su(H)-lacZ reporter are expected to be ISCs. As expected for a transcription factor, Sox21a-GFP signal was localized to the nuclei (Fig 2A and 2B). In unchallenged 5–7 day-old adults, Sox21a-GFP was expressed in both ISC and EB, with higher levels in ISC than in EB within an ISC-EB pair (Fig 2A, 2A’ and 2C). Oral infection with Ecc15 induced a marked increase of Sox21a-GFP expression in the progenitors (Fig 2B and 2B’). However, quantification of the relative intensity of Sox21a-GFP levels between cells within each ISC-EB pair indicated that regenerating intestines now expressed stronger Sox21a-GFP in the EB than in its sibling ISC (Fig 2C). Careful examination further revealed that Sox21a-GFP levels started declining in middle-sized EBs at a stage when Notch activity remained high. Moreover, Sox21a-GFP expression was totally shut down before the Notch activity reporter (Fig 2B).
Intestinal regeneration can be triggered by activating JNK or Ras/MAPK signaling in progenitors [12, 43, 44]. Interestingly, activating either pathway in progenitor cells with esg-Gal4 tub-Gal80ts for 36 hours also elevated Sox21a-GFP expression (S3A–S3E Fig), consistent with a previous observation [45]. However, akin to infection, stronger Sox21a-GFP expression was again seen in EBs, supporting the specific role of Sox21a in EB for EC differentiation. Collectively, this analysis shows that Sox21a displays a dynamic expression pattern, which coincides with the process of intestinal progenitor differentiation.
As shown previously [31], expressing Sox21a in progenitor cells with the esgTS driver for only 36 hours led to their differentiation into EC (Fig 2D–2F). Thus, over-expressing Sox21a provides a useful framework to unravel the sequence of events underlying stem cell differentiation into mature ECs. Taking advantage of this approach, we monitored the cellular changes resulting from overexpressing Sox21a in progenitors. Cells that were precociously differentiating towards ECs (termed “differentiating EBs”) showed an increase in cell size, became polyploid and expressed Pdm1, a marker for ECs [31]. Differentiating EBs accounted for around 40% of the esg>GFP+ progenitors expressing Sox21a, while they were barely seen in midgut progenitors of wild type flies (Fig 2G). They retained a weak esg>GFP signal and were located at a similar basal position as genuine ISC/EB (S3F Fig). In addition, each esg>GFP+ nest contained a slightly increased number of cells (Fig 2F and 2H), consistent with the notion that Sox21a also promotes a low level of ISC proliferation.
In homeostatic guts, most progenitor cells have limited membrane contact (Fig 2D). In Sox21a overexpressing guts, nearly all the differentiating EBs displayed an extended plasma membrane contact, with at least one cell with a strong esg>GFP signal (Fig 2E and 2E’; S3F Fig). Increased membrane contact was also reflected by the increased GFP signal at the membrane in progenitor cells expressing the membrane-tethered mCD8::GFP (Fig 2E’). To quantify changes in membrane contact, we measured contact area relative to the size of the smaller cell in more than 200 progenitor pairs. Maximal cell-cell contact between progenitors was reached as early as 36 hours after Sox21a expression was induced (Fig 2I and 2J). Similarly to the situation observed with infection, expressing Sox21a using the progenitor driver esg-Gal4 tub-Gal80ts also induced Dl-GFP+—Dl-GFP+ pairs with one cell exhibiting Notch activity with a high frequency (Fig 2K).
The observation that over-expressing Sox21a induces progenitors to differentiate towards ECs rather than EEs [31], suggests that Sox21a could enhance Notch activity. To test this idea, we analyzed how Sox21a impacts Notch signaling. Interestingly, Dl-GFP signal was enriched at the extended ISC-EB contact area upon Sox21a expression, in contrast to their even distribution in wild-type ISCs (Fig 3A and 3B). Furthermore, transcriptomic analysis using FACS-sorted progenitors revealed a four to seven fold increase of Dl mRNA in intestinal progenitors expressing Sox21a for only 12 or 24 hours (Fig 3C; S1 Table). The increase of Dl mRNA levels was further validated by quantitative PCR (qPCR) from dissected guts (S4 Fig). Thus, expressing Sox21a in progenitors increased Dl transcription and the amount of Dl protein at the ISC-EB interface, which are both expected to reinforce efficient Notch signal transduction. Finally, knocking down Dl with two independent RNAi lines both suppressed Sox21a-induced differentiation (Figs 3D and 3E and 4A; S5A and S5B Fig). Collectively, our data indicate that Sox21a mediates rapid differentiation at least in part by up-regulating Dl and by increasing progenitor contact, two features likely to enhance Notch signaling.
Extended progenitor contact in regenerating guts suggested that cell adhesion molecules might be involved in rapid progenitor differentiation. E-cadherin (E-cad) forms a complex with β-catenin at adherens junctions and mediates homophilic cell-cell adhesion [46]. Using an endogenous E-cad-GFP fusion, we first confirmed that E-cad was enriched at the ISC-EB interface in Sox21a expressing intestine. Similarly, the localization of Arm was almost identical to E-Cad (Fig 4B and 4C). Simultaneous depletion of either E-cad or arm for 36 hours completely suppressed the precocious differentiation phenotype seen in intestines expressing Sox21a (esgTS>Sox21a) (Fig 4A; S5C and S5D Fig). In these intestines, progenitor pairs displayed reduced contact confirming an essential role of E-cad in the formation of extended cell-cell contact during rapid differentiation. Since E-cad can impact Wg/Wnt signaling [47], the phenotype observed was possibly due to a requirement of Wg signaling for Sox21a-induced differentiation. However, neither activating (with constitutively active β-catenin) nor blocking Wg signaling (with dominant-negative Pangolin) affected Sox21a-induced differentiation (S5H and S5I Fig). We conclude that E-cad-mediated cell-cell adhesion is required for Sox21a-induced differentiation, independently of Wg signaling, in line with a previous report [48].
Furthermore, another cell adhesion molecule, Connectin (Con), which mediates homophilic cell-cell adhesion both in vitro and in vivo [49, 50], was up-regulated in progenitor cells expressing Sox21a in our progenitor-specific transcriptomic analysis (Fig 4D; S1 Table). Using an antibody against Connectin, we confirmed its enrichment at the extended contact between progenitors in esgTS>Sox21a gut (Fig 4E; S6A Fig). Like E-cad, Connectin was also crucial for Sox21a-induced differentiation, as its depletion abolished the occurrence of differentiating EBs (Fig 4A and 4F; S5E Fig).
To determine to which extent cell-cell adhesion can contribute to differentiation, we also overexpressed Connectin in progenitors. As expected, overexpressing Connectin using esgTS altered the morphology of progenitor cells with the formation of interconnected progenitors, sometimes forming big clusters (Fig 4G; S6A–S6C Fig). These changes were also associated with an increase in ISC proliferation but not a blockage of EB differentiation (Fig 4G’; S6H Fig). Consequently, progenitor tumors were not observed in esgTS>Connectin intestines despite the presence of large esg>GFP+ clusters. Surprisingly, many progenitors in esgTS>Connectin intestines displayed the same characteristics of differentiating EBs of esgTS>Sox21a intestines, including a weak esg>GFP signal, increased cell size, and extended contact with neighboring esg>GFPstrong progenitors (Fig 4G’). Moreover, regions with clusters of esg>GFP+ cells were devoid of EEs and progenitors in such regions were differentiating towards ECs as judged from their large cell size and polyploid nuclei (Fig 4G and 4G’; S6B and S6C Fig). Further experiments indicated that Connectin overexpression in the progenitors increased the expression levels of Notch activity reporter Su(H)-lacZ (Fig 4I) and promoted EB-EC differentiation rather than EE differentiation (Fig 4H–4L; S6I and S6J Fig). Therefore, an increase in cell adhesion by over-expressing Connectin in progenitors can promote their differentiation toward ECs, by enhancing Notch activity. Nevertheless, results obtained with the esgF/O system [14] did not support an essential role of E-cad or Connectin in basal intestinal turnover (S6K–S6M Fig). Unexpectedly, knocking-down Connectin in the progenitors induced both mild stem cell proliferation and progenitor differentiation in the absence of a challenge (S6D–S6H Fig), suggesting Connectin may have other functions in the maintenance of ISC under normal conditions. We conclude that the formation of extended cell contact between progenitors through adhesion molecules is required for Sox21a-induced rapid differentiation. Importantly, we show that this process is specifically required for the rapid differentiation of progenitors but not for basal low-level intestinal turnover during homeostatic conditions. Thus, our study reveals specific mechanisms that have evolved to accelerate the differentiation program.
Having analyzed the cellular changes that enhance Notch signaling activity in the ISC-EB transition during intestinal regeneration, we went on to investigate how Sox21a contributes to the processes of differentiation from EB to EC. In Drosophila, Notch deficient stem cells over-proliferate, leading to the formation of tumors composed mostly of ISCs and intermingled with EEs [22, 23, 25, 51], while the over-activation of Notch signaling drives progenitors to differentiate into ECs [22, 25]. The similarities between the function of Notch signaling and that of Sox21a in terminal differentiation led us to investigate their relationship.
While Sox21a is expressed in both ISC and EB, Notch activity is only found in EB [22, 23]. Thus, Sox21a expression in ISC is independent of Notch signaling. In addition, Notch activity was not blocked in a Sox21a mutant [31]. Several observations indicate that Sox21a and Notch signaling function interdependently for terminal differentiation. We first observed that Sox21a was required for the differentiation of progenitors into ECs upon Notch over-activation (Fig 5A and 5B), and for the formation of tumors composed of ISCs and EEs in Notch deficient clones (Fig 5C–5E). Conversely, the forced differentiation of progenitors into ECs upon expression of Sox21a was blocked in the absence of functional Notch signaling (S7A–S7D Fig). Thus, over-expression of Sox21a cannot overcome the requirement of Notch signaling for the differentiation of EB. Collectively, this led us to conclude that Sox21a and Notch signaling encompass two parallel systems that need to cooperate to ensure terminal differentiation.
Both Sox21a and JAK/STAT have been shown to be mandatory for the EB-EC differentiation, and knockdown of either of these two factors results in the formation of EB containing tumors. Mechanistically, the formation of these tumors is caused by a feed-back amplification loop whereby the differentiation-defective EBs secrete growth factors stimulating ISC proliferation [14, 27, 31, 32]. We have previously shown that over-expression of Sox21a can partially rescue the differentiation defect caused by the loss of JAK/STAT in MARCM clones, suggesting that Sox21a functions downstream of JAK/STAT signaling in EB differentiation [31]. We further analyzed the relationship between JAK/STAT signaling and Sox21a by using this time an EB specific driver, Su(H)GBETS (Fig 6A–6D). Knocking down Sox21a specifically in EB caused the accumulation of EBs and an increase in ISC mitosis resulting in strong tumor formation, recapitulating the Sox21a mutant phenotype (Fig 6D–6F). In contrast, EB-specific depletion of Stat92E, the gene encoding the JAK/STAT transcription factor, had only mild consequences with the formation of small-sized EB tumors (Fig 6C, 6E and 6F). Unexpectedly, expressing a dominant-negative form of Domeless (DomeDN), the receptor of the JAK/STAT signaling cascade, using the same EB driver did not affect the differentiation process and did not lead to any tumor formation (Fig 6B, 6E and 6F). However, silencing the JAK/STAT pathway in both ISC and EB with esgTS, using the same UAS-DomeDN or the UAS-Stat-RNAi constructs led to the formation of massive progenitor tumors (Fig 6G and 6H). The observations that EB tumor formation requires the inactivation of JAK/STAT in both ISC and EB, and that knock-down of Sox21a only in EBs is sufficient to cause tumors, are consistent with the notion that JAK/STAT signaling is required earlier than Sox21a in the course of EB-EC differentiation. Supporting this hypothesis, we observed that expressing Sox21a could suppress tumor formation caused by loss of JAK/STAT signaling in progenitors (Fig 6I–6L), indicating that Sox21a can promote progenitor differentiation in the absence of JAK/STAT signaling. Thus, the observations that JAK/STAT signaling functions earlier than Sox21a, and that Sox21a can override the differentiation blockage caused by the loss of JAK/STAT signaling confirm and extend our previous observation based on mosaic analysis [31] that Sox21a acts downstream of JAK/STAT pathway in EB-EC differentiation.
The role of the Dpp signaling pathway in EB differentiation has been controversial, with studies supporting that it is essential to this process while others suggest it is dispensable [34–36]. The observation that several genes encoding components of Dpp signaling, including the receptor thickveins (tkv), the transcription factors schnurri and Mothers against dpp (Mad), are down-regulated in Sox21a mutant EBs, supported a role of Dpp signaling in EB to EC transition (Fig 7A). We therefore explored its role in the differentiation process and its relationship with Sox21a. Interestingly, Dpp signaling was specifically induced in differentiating EBs in esgTS>Sox21a intestines as revealed by a reporter gene of Dpp signaling, Dad-GFPnls (Fig 7B and 7C). The expression of the Dad-GFPnls reporter was much stronger in differentiating EBs than in mature ECs that were already differentiated prior to the activation of Sox21a (Fig 7C), highlighting a role of Dpp signaling in the EB to EC transition. Importantly, Dpp signaling was mandatory for Sox21a-induced rapid differentiation, as depleting the key component Mad abolished progenitor differentiation of esgTS>Sox21a intestines (Fig 4A; S5F Fig). These results support a role of the Dpp pathway in EB differentiation downstream of Sox21a. Nevertheless, overexpressing the Dpp transcription factor Schnurri, which is known to promote progenitor differentiation into EC in the midgut [36], did not rescue the differentiation defect of Sox21a mutant clones (Fig 7D and 7E). We conclude that Dpp signaling is required downstream of Sox21a for rapid differentiation but is not sufficient to rescue Sox21a deficiency.
GATAe encodes a transcription factor, which is expressed in all the cell types of the fly midgut. It has recently been shown to be important for ISC proliferation and, to a lesser extent, for EB differentiation [33]. It also has a role in ECs to maintain the regionalization of the intestine [33, 52]. Our RNA-seq experiments with FACS-sorted EBs revealed that the expression of GATAe was decreased in the absence of Sox21a (Fig 7A), suggesting that this transcription factor could contribute to the differentiation program downstream of Sox21a. We therefore investigated in further detail the role of GATAe in the differentiation process and its relationship with Sox21a.
We first observed that over-expressing GATAe under the control of esgTS driver led to the precocious differentiation of progenitors into Pdm1-positive ECs (Fig 8A–8C), consistent with the notion that GATAe can promote progenitor differentiation [33]. The fast differentiation of progenitors was supported by the presence of many mature EC that still kept residual esg>GFP signal (Fig 8B), reminiscent of progenitor cells in esgTS>Sox21a intestines. Consistent with [33], loss of GATAe did not block terminal differentiation in basal conditions, since both Pdm1-positive ECs and Pros-positive EEs were found in stem cell clones deficient for GATAe (S8A and S8B Fig). Similarly, EB-specific depletion of GATAe using the EB-specific driver Su(H)GBETS did not lead to EB tumors, but rather to a slight accumulation of late-stage EBs as judged from their appearance. These results suggest that GATAe may contribute to the rate of EB differentiation. To test this idea, we analyzed the contribution of GATAe to rapid epithelial turnover induced by ingestion of bacteria. Both wild-type flies and flies with EB-specific depletion of GATAe were orally infected with Ecc15 for 2 days, a treatment that increases the pool of progenitors undergoing differentiation, and were further let to recover for 3 days. At this timepoint, the midgut of Su(H)GBETS>w control flies subjected to bacterial challenge had already returned to a homeostatic condition where only nascent EBs with small nuclei were found (Fig 8D). In sharp contrast, accumulation of EBs was observed along the midgut of Su(H)GBETS>GATAe-IR flies (Fig 8E). These EBs had a larger cell size than EBs found in wild-type flies suggesting that they were stuck in the process of differentiation. Consistent with a role of GATAe for accelerated EB differentiation, simultaneously depleting GATAe abolished progenitor differentiation in esgTS>Sox21a intestines (Fig 4A; S5G Fig). Moreover, expressing GATAe with the esgTS driver suppressed the differentiation defect and tumor formation induced by the loss of either Sox21a or Stat92E (Fig 8F–8L), revealing that GATAe acts downstream of Sox21a and JAK/STAT. Collectively, our study not only confirms that GATAe contributes to the differentiation process [33], but further reveals its critical role during regeneration as opposed to basal conditions. Our data also show that JAK/STAT-Sox21a-GATAe forms a sequential relay orchestrating the EB-EC differentiation process.
Key questions in stem cell biology are how the pool of stem cells can be robustly expanded yet also timely contracted through differentiation to generate mature cells according to the need of a tissue, and what are the underlying mechanisms that couple stem cell proliferation and differentiation. Over the last years, the mechanisms underlying intestinal stem cell activation have been extensively studied in both flies and mammals [1, 4], while the genetic control of progenitor differentiation, especially during regeneration, has only recently begun to be understood [26, 28, 31].
The transcription factor Sox21a has recently been the focus of studies in fly intestines [31, 32, 45]. Using a Sox21a-sGFP transgene, we uncovered its dynamic expression pattern in intestinal progenitors. Higher levels of Sox21a were found in ISC during homeostatic conditions but in EB during regeneration, supporting the roles of Sox21a in both ISC maintenance and EB differentiation at different conditions. The highly dynamic expression pattern of Sox21a revealed by this sGFP-tagged transgene per se argues against accumulation and perdurance of GFP fusion protein. Indeed, immunostaining using an antibody against Sox21a also indicated stronger Sox21a expression in ISC in homeostatic condition and global activation of Sox21a in progenitors under DSS-induced regeneration [45]. However, Chen et al., (2016) suggested that Sox21a levels are always higher in EB than in ISC by applying another antibody against Sox21a. The inconsistency between these studies may have arisen from the differences in EB stages examined or the sensitivity of respective detection approaches.
In this study, we have analyzed the cellular processes required for efficient progenitor differentiation during regeneration and uncovered three main findings revealing: i) the importance of extended contact between a stem cell and its differentiating daughter, ii) the existence of specific mechanisms allowing fast differentiation during regeneration, and iii) the characterization of a genetic program instructing the transition from EB to EC. These results together led us to propose a molecular framework underlying intestinal regeneration (Fig 9) that is discussed below step by step.
By studying the mechanisms of Sox21a-induced differentiation, we found that ISC establishes extended contact with its differentiating daughter within a progenitor pair. Increased interface contact was not only observed upon Sox21a expression but also during regeneration after bacterial infection and DSS-feeding. Since the presence of extended contact is rare in intestinal progenitors under homeostatic conditions, we hypothesize that extended contact between progenitors is related to increased epithelial renewal as a mechanism to elicit optimal juxtacrine Notch signaling to accelerate the speed of progenitor differentiation. The observations that down-regulation of the cell adhesion molecules E-Cadherin or Connectin suppresses rapid progenitor differentiation upon regeneration, and that overexpression of Connectin is sufficient to promote differentiation, underline the importance of increased cell-cell contact in rapid differentiation. Our study shows that one early role of Sox21a is to promote the formation of this contact zone, possibly through transcriptional regulation of Connectin. Further studies should identify the signals and pathways leading to the change of contact between progenitors to adjust the rate of differentiation.
Intestinal progenitors with extended contact in non-homeostatic midguts have been observed in some studies [14, 41], but their role and significance have not been analyzed. Previous studies have also shown that progenitor nests are outlined by E-Cadherin/β-Catenin complexes [23, 48], yet it was not known whether different degrees of progenitor contact are associated with their ISC versus EB fate. Consistent with our results, recent modeling analyses suggested a positive correlation between the contact area of progenitor pairs and the activation of Notch signaling [53, 54]. Thus, it seems that an increase in the contact area between intestinal progenitors is a hallmark of progenitors that are undergoing accelerated differentiation towards ECs. Another study by Choi et al. (2011) has suggested an inhibitory role of prolonged ISC-EB contact to restrict ISC proliferation. Collectively, these studies and our findings suggest that the strong contact between ISC and EB promotes on one hand the efficient differentiation of EBs into mature intestinal cells while on the other hand preventing stem cells from over-dividing. Thus, we hypothesize that alteration in the contact zone provides a mechanism for ensuring both the appropriate speed of differentiation and the timely resolution of stem cell proliferative capacity.
A second finding of our study consists in revealing the existence of specific mechanisms accelerating differentiation for tissue replenishment. In addition to the extended contact discussed above, we observe a difference in the pattern of ISC division between homeostatic and highly regenerative intestines. The modes of ISC division in Drosophila have been the topic of intense discussion, and the general consensus is that it is associated with an asymmetric cell fate outcome, in which one cell remains an ISC and the other engages in differentiation [5, 41, 55, 56]. In line with these previous studies, our results support the notion that asymmetric cell division is the most prevalent mode of ISC division under homeostatic conditions, where the rate of epithelial renewal is low. However, our use of ISC- and EB-specific markers shows that upon rapid regeneration an ISC divides into two cells both expressing the ISC marker Dl-GFP but with one cell showing weak Notch activity. Similarly to other Notch-mediated cell-fate decision systems [57], our study suggests that the two resulting Dl-GFP+ cells from a symmetric division stay in close contact and compete for the stem cell fate. While our study is not the first to postulate the existence of symmetric ISC division [5, 55], the use of reliable ISC- and EB-specific markers allows us to better visualize this process. Applying a dual-color lineage tracing system to unravel the final fate of respective cells in a Dl+—Dl+ pair could reinforce the existence of symmetric stem cell division. This is nevertheless technically challenging to apply here since all the current available lineage-tracing settings require a heat shock to initiate the labeling, which affects intestinal homeostasis.
Importantly, we show that the genetic program required for fast intestinal regeneration differs from the one involved in basal intestinal maintenance. Our study indicates that GATAe, Dpp signaling, and the cell adhesion molecules E-cadherin and Connectin are not critical for progenitor differentiation when the rate of epithelial renewal is low, whereas their roles become crucial upon active regeneration. We speculate that many discrepancies in the literature can be reconciled by taking into consideration that some factors are required only for rapid differentiation but not in basal conditions. For instance, the implication of Dpp signaling in differentiation has been disputed, since Zhou et al. (2015) focused on bacterial infection-induced regeneration while the other two studies dealt with basal conditions [34–36]. Our study here points to a clear role of Dpp signaling in the differentiation process upon regeneration. Therefore, better defining the genetic program that allows adjusting the speed of differentiation would be of great interest.
Cell fate determination and differentiation involve extensive changes in gene expression and possibly also gradual change of cell morphology. The EB to EC differentiation in the adult Drosophila intestine provides a model of choice to study this process. This transition includes changes in cell shape, an increase in cell size, DNA endoreplication leading to polyploidy and the activation of the set of genes required for EC function (S9A–S9C Fig). In this study, we have integrated a number of pathways (Notch, JAK/STAT and Dpp/BMP) and transcription factors (Sox21a and GATAe) into a sequential framework. We further show that Sox21a contributes to the EB-EC transition downstream of JAK/STAT but upstream of Dpp signaling and GATAe. The recurrent use of several factors, namely JAK/STAT, Sox21a and GATAe at different processes including ISC self-renewal and EB-EC differentiation is likely to be a general feature during cell fate determination, and somehow also complicates the study of differentiation. Future work should analyze how each of the factors interacts with the other in a direct or indirect manner. It would be interesting as well to further study how these factors shape intestinal regionalization as the gut exhibits conspicuous morphological changes along the length of the digestive tract [52, 58].
Several of the findings described are likely to apply to the differentiation program that takes place in mammals. Since Notch signaling plays major roles in stem cell proliferation and cell fate specification from flies to mammals [57, 59], it would be interesting to decipher whether in mammals changes in progenitor contact also impact differentiation speed and whether a specific machinery can accelerate progenitor differentiation when tissue replenishment is required.
Fly strains were kept on a standard medium (maize flour, dead yeast, agar and fruit juice). esg-Gal4, tub-Gal80TS, UAS-GFP (referred to as esgTS); Su(H)GBE-Gal4, tub-Gal80TS, UAS-GFP (referred to as Su(H)GBETS); esg-Gal4, tub-Gal80TS, UAS-GFP; UAS-Flp, Act>>Gal4 (referred to as esgF/O); MARCM tester FRT2A: y,w,hsFlp; tub-Gal4, UAS-CD8::GFP; FRT2A, tub-Gal80; MARCM tester FRT82B: y,w,hsFlp, tub-Gal4, UAS-nlsGFP;;FRT82B, tub-Gal80; Su(H)-lacZ, Sox21a6, UAS-Sox21a, UAS-Sox21a-RNAi (BL53991), UAS-Stat92E-RNAi (BL31318 and 35600) and UAS-N-RNAi (VDRC100002) have been described before [31]. Dl-GFP, UAS-hep, UAS-RafACT, UAS-lacZ, UAS-NICD, UAS-armS10, UAS-PanDN, UAS-E-cad-RNAi (BL27689), UAS-Con-RNAi (BL28967), UAS-Dl-RNAi (BL28032 and 34322), UAS-arm-RNAi (BL31304 and 31305), UAS-Mad-RNAi (BL31315), UAS-GATAe-RNAi (BL 34907), UAS-Notch-RNAi (VDRC, KK) and UAS-mCherry-RNAi were obtained from Bloomington Drosophila Stock Center (BDSC). Sox21a-GFP was from VDRC stock center. UAS-shn, UAS-Stat92E and UAS-domeDN (gift from Michael Boutros), FRT82B, NeurIF65 (gift from Allison Bardin), Dad-GFPnls (gift from Fisun Hamaratoglu), UAS-Con (gift from Rob White) and FRT82B, GATAe1 (gift from Takashi Adachi-Yamada) were also used. UAS-GATAe was generated in this study.
Dl-GFP (BL59819) encodes an endogenously GFP-tagged Dl protein resulting from recombination mediated cassette exchange of a Mi{MIC} insertion in the Dl coding intron [38]. This line is homozygous lethal. Sox21a-GFP transgenic line derives from a GFP-tagged fosmid clone containing a large genomic region including Sox21a [42]. In most cases, the driver lines (esgTS or Su(H)GBETS) were crossed to the w1118 strain, UAS-mCherry-RNAi, or UAS-lacZ, and the progenies were used as control for overexpression experiments. w1118 flies carrying one copy of esgTS (esgTS>w) were used as wild type to visualize the contact between progenitors in different conditions.
Erwinia carotovora carotovora15 (Ecc15) was grown in LB medium at 29°C with shaking overnight, and harvested by centrifugation at 3000g at 4°C for 30 minutes. The pellet was then suspended in the residual LB, and bacterial concentration was adjusted to OD600 = 200. Flies older than 3 days were first dry-starved in an empty tube for 2 hours, and then transferred into a classical fly food vial containing a filter paper that totally covers the food and was soaked with a solution consisting of 5% sucrose and Ecc15 at OD200 (1:1), or 6% DSS (average MW 40 kDa, sigma) treated flies were kept at 29°C until dissection. Colcemid treatment was done as reported previously [34]. 200ug/ml colcemid (Sigma) was added to 5% sucrose to pre-treat the flies for 12 hours, and then an Ecc15 infection was performed in the presence of 200ug/ml colcemid.
Flies were transferred overnight into a classical fly food vial containing a filter paper soaked with a solution consisting of 5% sucrose to clean the digestive tract. Then, 10–15 intestines of mated adult females were dissected in phosphate-buffered saline (PBS), and fixed for at least one hour at room temperature in 4% paraformaldehyde (PFA) in PBS. They were subsequently rinsed in PBS+0.1% Triton X-100 (PBT), permeabilized and blocked in 2% BSA PBT for one hour, and incubated with primary antibodies in 2% BSA PBT for overnight at 4°C. After one hour of washing, secondary antibodies and DAPI were applied at room temperature for two hours.
Primary antibodies used are: mouse anti-Pros (DSHB, 1:100), mouse anti-Arm (DSHB, 1:100), mouse anti-Dl (DSHB, 1:100), mouse anti-βPS (DSHB, 1:100), mouse anti-Con (DSHB, 1:4), rabbit anti-pH3 (Millipore, 1:1000), Chicken anti-GFP (Abcam, 1:1000), rabbit anti-βGal (Cappel, 1:1000), mouse anti-βGal (Sigma, 1:1000), and Rat anti-mCherry (Life Technologies, 1:500). Alexa488-, Alexa555- or Alexa647-conjugated secondary antibodies (Life Technologies) were used. Nuclei were counterstained by DAPI (Sigma, 1:10’000). All the images were taken on a Zeiss LSM 700 confocal microscope at BIOP in EPFL. Images were processed using Image J and Adobe Photoshop software. Shown in figures are maximal intensity projections of all the confocal z stacks.
To generate the UAS-GATAe construct, the following primers (caccATGGTCTGCAAAACTATCTC and TTAGTTATTCGATGATCGCTC) were used to amplify the 2.2kb GATAe-PA coding regions from cDNA clone LD08432 purchased from DGRC. The PCR product was first cloned into pENTR-D-TOPO (Life Technologies) vector, and then swapped into pTW destination vector to make UAS-GATAe. Transgenic flies were established by standard P element-mediated germ-line transformation (BestGene Inc.). At least three independent transgenic lines were tested for expression level.
The TARGET system was used in combination with the indicated Gal4 drivers to conditionally express UAS-linked transgenes [60]. Flies were grown at 18°C to limit Gal4 activity. After 3–5 days at 18°C, adult flies with the appropriate genotypes were shifted to 29°C, a temperature inactivating the temperature-sensitive Gal80’s ability to suppress Gal4, and dissected after indicated time of transgene activation.
Mosaic analysis with a repressible cell marker (MARCM) technique was used for clonal analysis [61]. For clone induction, 3-5-day-old flies with the appropriate genotypes were heat-shocked for 30 min at 37.5°C in a water bath. The flies were immediately transferred into a new tube and kept at 25°C or indicated temperature until dissection. Overexpression experiments were performed by combining the UAS-linked transgenes with the FRT2A, the FRT2A, Sox21a6, or the FRT82B, NeurIF65 chromosome. Note that UAS-linked transgenes were only expressed in the clones indicated by the presence of GFP.
esgTS virgin females were crossed to either w1118 as control or UAS-Sox21a for overexresspion at 18°C. Eclosed flies were maintained at 18°C for 5–7 days. Around 50 flies for each biological replicate were dissected in ice-cold 1xPBS made with DEPC-treated water under a dry-ice chilled dissecting microscope, within a one-hour time frame. Proventriculus, hindgut and midgut/hindgut junction were removed to collect only midgut esg>GFP positive cells. Two biological replicates were performed, and the activation of Sox21a expression was done by shifting esgTS>Sox21a flies to 29°C for 12 hours and 24 hours, respectively. Cell dissociation, FACS sorting, total RNA isolation and mRNA amplification were performed as described [31]. RNA-seq was performed on a Hi-Seq2000 (Illumina) with 100 nt single-end sequencing, and sequencing data was analyzed as described before [31]. Sequencing data will be deposited in public database prior to publication.
Total RNA was extracted from dissected whole guts (12–15 guts per sample) using Trizol and cDNA was synthesized using the PrimeScript RT reagent Kit (TaKaRa). 0.5μg total RNA was used for reverse transcription with oligo dT, and the 1st strand cDNA was diluted 10–20 times with water to be further used in real time PCR. Real time PCR was performed in triplicate for each sample using SYBR Green (Roche) on a LightCycler 480 System (Roche). Expression values were calculated using the ΔΔCt method and relative expression was normalized to Act5C. Results are shown as mean ± SEM of at least 3 independent biological samples. Statistical analysis was performed in Prism Software using the unpaired t test. Primers used for qPCR are as follows.
All analyses were done with GraphPad Prism. Unpaired t test were used unless otherwise noted. p values are indicated by *p < 0.05, **p < 0.01, ***p < 0.001, ns: p > 0.05. Shown are means and SEM. Data are representative of at least three experiments. Each dot represents one gut except Fig 5E. Sample size is also indicated in the figures.
|
10.1371/journal.pntd.0005099 | ‘If an Eye Is Washed Properly, It Means It Would See Clearly’: A Mixed Methods Study of Face Washing Knowledge, Attitudes, and Behaviors in Rural Ethiopia | Face cleanliness is a core component of the SAFE (Surgery, Antibiotics, Facial cleanliness, and Environmental improvements) strategy for trachoma control. Understanding knowledge, attitudes, and behaviors related to face washing may be helpful for designing effective interventions for improving facial cleanliness.
In April 2014, a mixed methods study including focus groups and a quantitative cross-sectional study was conducted in the East Gojjam zone of the Amhara region of Ethiopia. Participants were asked about face washing practices, motivations for face washing, use of soap (which may reduce bacterial load), and fly control strategies.
Overall, both knowledge and reported practice of face washing was high. Participants reported they knew that washing their own face and their children’s faces daily was important for hygiene and infection control. Although participants reported high knowledge of the importance of soap for face washing, quantitative data revealed strong variations by community in the use of soap for face washing, ranging from 4.4% to 82.2% of households reporting using soap for face washing. Cost and forgetfulness were cited as barriers to the use of soap for face washing. Keeping flies from landing on children was a commonly cited motivator for regular face washing, as was trachoma prevention.
Interventions aiming to improve facial cleanliness for trachoma prevention should focus on habit formation (to address forgetfulness) and address barriers to the use of soap, such as reducing cost. Interventions that focus solely on improving knowledge may not be effective for changing face-washing behaviors.
| Facial cleanliness is a core component of the SAFE (Surgery, Antibiotics, Facial cleanliness, and Environmental improvements) strategy for trachoma control. We conducted a mixed methods study in a trachoma hyperendemic region of rural Ethiopia to better understand knowledge, attitudes, and behaviors related to face washing. Overall, knowledge of the benefits of face washing was high, and participants reported regularly engaging in face washing practices. However, the use of soap for face washing varied more between communities. Participants cited cost and forgetting to use soap as the primary barriers to using soap for face washing. Trachoma prevention, including keeping flies from landing on children’s faces, was a commonly-cited motivator for face washing discussed in focus groups. Given the near-universal knowledge of the benefits of face washing, interventions focused on changing face washing behavior for trachoma control should focus on habit formation and removal of barriers to the use of soap rather than simply educational interventions.
| Trachoma is the leading cause of infectious blindness globally.[1–3] Caused by the bacterium Chlamydia trachomatis, trachoma is thought to be transmitted by direct contact from infected persons and clothing, as well as the moisture-seeking fly Musca sorbens.[4,5] Currently endemic in 53 countries[6], trachoma is estimated to result in blindness or severe vision loss in more than 2 million people[1], with the majority of cases found in sub-Saharan Africa.[1] Despite large reductions in the burden of trachoma in the past several decades[1], trachoma remains an important cause of blindness primarily among individuals living in poor, predominantly rural areas.[6–9]
The cornerstone of trachoma control is the SAFE (Surgery, Antibiotics, Facial cleanliness, and Environmental improvements) strategy.[6] Mass antibiotic distributions have been shown to be effective at reducing the prevalence of trachoma.[10,11] However, while antibiotics may lead to local control of trachoma, alone they may not be sufficient for trachoma elimination in places with hyperendemic infection.[11] Multiple observational studies have demonstrated an association between poor facial hygiene, including the presence of flies on a child’s face, and trachoma.[12–16] It is possible that improvements in hygiene, and especially facial hygiene, may alter the transmission dynamics of trachoma and create more favorable conditions for trachoma elimination.
The use of soap for face washing has been shown to be associated with decreased risk of trachoma in some[16–19] but not all[20] studies. Soap may decrease the bacterial load on children’s faces, which could decrease the probability of transmission of trachoma. A recent meta-analysis of observational studies demonstrated that use of soap was associated with a lower prevalence of trachoma.[16] However, soap specifically for face washing is rarely included or advocated for in trachoma elimination campaigns.
The association between poor facial hygiene and trachoma suggests that interventions to promote facial cleanliness may be helpful in reducing trachoma prevalence and ultimately achieving trachoma elimination. These interventions will benefit from understanding current knowledge, practices and beliefs related to face washing. Here, we analyze knowledge, beliefs, and practices related to face washing, and their relation to trachoma, in a mixed methods study in a trachoma-hyperendemic region of rural Ethiopia.
This study took place in a rural agrarian region in the Goncha Siso Enese woreda of East Gojjam, Amhara, Ethiopia. The communities in this study were participating in a series of cluster-randomized trials testing different mass drug administration strategies for trachoma elimination beginning in 2006. Each community has approximately 275 residents. These communities received mass azithromycin distributions annually or biannually between 2006 and 2013.[11] Methods for these trials are described in detail elsewhere.[11] At baseline, the prevalence of trachoma in children 1–10 years old was 48.5% and 15.5% in children 11 years and older.[11] For the present study, we selected five communities that were within a one-hour walk from the farthest place a four-wheel drive vehicle could reach. All households in each community included in this study were eligible to participate in the quantitative survey. Before and during the study, all communities continued to receive the prescribed government package of hygiene promotion activities. In this report, the five communities are labeled Community A, B, C, D, and E to protect anonymity of the communities.
The quantitative and qualitative surveys were designed to gain an understanding of existing knowledge and behaviors in relation to face washing, and to identify gaps between knowledge and behaviors.
Of 279 eligible households, 264 households had data available on face washing for children and 279 for using soap for face washing. 154 of the survey respondents were female and 123 were male. The majority of heads of household were male (211); 66 households had a female head-of-household. Median age of the respondent in the household survey was 36 years (IQR 30 to 50). The majority of heads of households (98%) were farmers, and the median number of children in each household was 3 (IQR 2 to 4). A total of 105 individuals participated in 15 focus groups. Focus group participants ranged in age from 18 to 60 years, with a mean age of 35 (Table 1). The majority of focus group participants had no formal education.
Universally, participants endorsed daily face washing of their children, typically in the morning, although many participants also indicated that more frequent washing would be beneficial for their children.
Whereas there was universal endorsement of face washing among participants in quantitative and qualitative data, the reported usage of soap for face washing varied widely by community, from 4% in Community C to 82% in Community E (Table 2).
To gain a deeper understanding of face washing behaviors, focus groups explicitly probed for reasons behind face washing. Participants in all communities mentioned fly control as a reason for face washing, especially for children. Participants noted both that face washing was necessary when flies were seen on children’s faces, but also that regular face washing prevented flies from landing on faces.
Many participants cited trachoma prevention as a reason for face washing for fly control. The benefits of fly control were seen as beneficial for the general health and wellbeing of children, as well.
In this mixed methods study, we document high prevalence of reported daily face washing among a rural population in a hyperendemic area for trachoma in Ethiopia. However, despite face washing being a common practice, soap was less commonly used as part of face washing routines. Face washing is a key component of the SAFE strategy for trachoma prevention, and use of soap may improve the ability of face washing to prevent trachoma transmission.[17,18] The use of soap for face washing in this study varied widely by community. Previous studies have demonstrated clustering of active trachoma and trachoma infection at the household and village level.[21,22] Geographic clustering of trachoma is likely due to both increased probability of transmission in areas with higher trachoma prevalence as well as shared characteristics such as environmental or climate factors.[23,24] The results of the present study suggest that there may be shared behavioral characteristics within villages that may also contribute to geographic clustering.
Although use of soap varied widely by community, focus group participants from all communities reported high levels of knowledge of the importance of soap. The focus groups suggested that economic barriers are important in limiting the regular use of soap for face washing, indicating economic interventions may be important for improving face washing with soap behaviors. In Community C, which had the lowest reported use of soap for face washing, focus group participants explained that while they knew the benefits of using soap for face washing, they were simply not in the habit of doing so. These results suggest that, in this community, individuals are further along the knowledge-attitude-behavior continuum in terms of behavior change.[25] As such, interventions promoting face washing with soap may be more effective if they focus on habit formation and practice rather than improving knowledge, as the community members already have knowledge of the benefits of using soap for face washing.
There is conflicting evidence of the relationship between distance to water source and trachoma.[16,26] Theoretically, increased distance to water source may reduce face washing behaviors because of water security in households. It is possible, however, that the inclusion of other factors in multivariable models (such as hygiene practices) obscures the relationship in some studies. In this study, we found an association between shorter distance to water source and face washing of children in the household. In addition, we noted a dose-response relationship, with households that reported longer times for water collecting less frequently reporting face-washing children. Similarly, households in which the survey respondent reported that the household had an adequate supply of water for their needs more often reported face washing of all children. The results indicate that face-washing behaviors may be facilitated by access to adequate water supply. Future work should consider the role of distance to water source on hygiene behaviors, as it is plausible that households with greater access to water have differential hygiene behaviors.
Overall in this sample, participants had high health literacy related to trachoma. Focus group participants generally believed that face washing would help prevent trachoma. The communities in which this study was conducted were in a region that is hyperendemic for trachoma, and as such received mass drug administrations for trachoma and participated in trachoma trials for the eight years prior to the present study.[27–29] High levels of trachoma knowledge may be related to participants’ involvement in these studies. This knowledge of trachoma and the fact that it could cause blindness in their children was likely a motivator for face washing behavior, explaining high coverage of daily face washing in this population. There is also the possibility that participants noted trachoma prevention as a motivation for face-washing because they felt it was the ‘correct’ answer and not because it was a true motivation. These results may not be generalizable to areas that are hyperendemic for trachoma that have not experienced this intensity of trachoma programming. Future work may be needed in trachoma study or program-naive populations to determine if face-washing predictors and behaviors differ.
The results of this study must be considered in the context of several limitations. Face washing behaviors in both the focus group discussions and the quantitative survey were collected via self-report. Although face washing is not necessarily a stigmatized behavior, it is possible that individuals’ responses may have been influenced by social desirability bias, as participants may have responded in ways which they perceived to be “correct”. There may have been alternative explanations, for example soap getting into children’s eyes, that were not discussed because participants perceived it was an incorrect answer. We anticipate that any outcome misclassification arising due to social desirability bias would be non-differential with respect to various predictors, and as such would, on average, bias towards the null in our regression models. In addition, while participants discussed their knowledge and current behaviors related to face washing and using soap, responses may not necessarily reflect motivations, and may rather reflect rationalization or normative reasons. Importantly, as qualitative data were collected as focus group discussions, individual responses may have been influenced by the responses of other members in the focus group. These results therefore should not be interpreted on the individual level, but instead represent community-level knowledge and behaviors. It is possible that some individual behaviors were masked in group discussions if some individuals did not want to discuss behaviors that were outlying from the rest of the group. Future work with individual interviews may yield additional insights into face washing and other hygiene behaviors in this region.
This study provides important insights into face washing knowledge, attitudes, and behaviors for intervention development in a trachoma hyperendemic region of rural Ethiopia. Overall knowledge of the benefits of face washing was high, and the use of soap for face washing varied widely by community. Water access was associated with reduced odds of all children in the household washing their faces, but was not discussed during focus group discussions. Barriers to face washing with soap included cost and forgetting to use soap. Interventions for face washing that include habit formation, which may help to address forgetfulness, and address structural barriers to accessing soap, like cost, may be important for increasing facial cleanliness and ultimately trachoma control in hyperendemic regions.
|
10.1371/journal.ppat.1000311 | An Antiviral Response Directed by PKR Phosphorylation of the RNA Helicase A | The double-stranded RNA-activated protein kinase R (PKR) is a key regulator of the innate immune response. Activation of PKR during viral infection culminates in phosphorylation of the α subunit of the eukaryotic translation initiation factor 2 (eIF2α) to inhibit protein translation. A broad range of regulatory functions has also been attributed to PKR. However, as few additional PKR substrates have been identified, the mechanisms remain unclear. Here, PKR is shown to interact with an essential RNA helicase, RHA. Moreover, RHA is identified as a substrate for PKR, with phosphorylation perturbing the association of the helicase with double-stranded RNA (dsRNA). Through this mechanism, PKR can modulate transcription, as revealed by its ability to prevent the capacity of RHA to catalyze transactivating response (TAR)–mediated type 1 human immunodeficiency virus (HIV-1) gene regulation. Consequently, HIV-1 virions packaged in cells also expressing the decoy RHA peptides subsequently had enhanced infectivity. The data demonstrate interplay between key components of dsRNA metabolism, both connecting RHA to an important component of innate immunity and delineating an unanticipated role for PKR in RNA metabolism.
| Our manuscript explores the immune response to viral infection by investigating events triggered by the protein kinase PKR. This sentinel kinase is constitutively expressed in all cells as an inactive protein that is subsequently activated by viral RNA produced during an infection. The active kinase perturbs viral replication by phosphorylating protein substrates in the cell. In this manuscript we identify a novel substrate for PKR, an essential helicase, RHA. Viruses usurp this helicase to replicate their own genome. We demonstrate that phosphorylation of RHA by PKR perturbs the ability of the helicase to bind viral RNA. Correspondingly, PKR prevents the capacity of RHA to enhance expression of genetic elements encoded by the human immunodeficiency virus (HIV). Juxtaposed to this, HIV virions packaged within cells that also express protein fragments of RHA, demonstrated to interact with PKR as decoy substrates, have enhanced infectivity. These fragments of RHA occur within a protein domain previously established to bind RNA but increasingly recognized to mediate protein–protein interactions. This supports an emerging role for these protein domains to coordinate the cell's response to pathogen-associated RNA. The findings identify a new cell-signaling pathway important in the response to viral infection.
| The primary detection of viral infection is by the host innate immune system, with the recognition of viral double-stranded RNA (dsRNA) a crucial early function. Responses to dsRNA are mediated by several protein receptors that recognize this pathogen-associated molecular pattern (PAMP). Most important of these receptors are the Toll-like receptor-3 (TLR3), two caspase recruitment domain (CARD)-containing helicases, retinoic acid inducible gene-I (RIG-I) and the related IFN inducible helicase-I (IFIH-I), and the protein kinase R (PKR). These dsRNA receptors are spatially separated within the cell to respond to either intra- or extra-cellular dsRNA. Moreover, the outcome of the ensuing antiviral response triggered by each receptor differs between cell compartments [1]. Consequently, a full contingent of pattern recognition receptors is required for immune competence. TLR3 is located on the cell surface or in the endosome compartment, and upon sensing dsRNA recruits the cytoplasmic adaptor Toll/IL-1R (TIR) domain-containing adaptor-inducing IFNβ (TRIF), via shared TIR homologous regions to mediate antiviral responses [2]–[4]. Adaptor signaling leads to IFN regulatory factor (IRF)3 and IRF7 activation and type-I IFN production [5],[6]. RIG-I and IFIH-I are cytoplasmic receptors which sense dsRNA and subsequently transmit a signal via helicase and CARD domains, respectfully. Activated RIG-I/IFIH-I associate with a mitochondrial anchored CARD adaptor, IPS-1 (also called MAVS, Cardif, or VISA), to activate NFκB and IRF3 and induce IFNβ [7]–[10]. Alternatively, dsRNA-binding at the amino terminus of PKR activates the kinase, resulting in the phosphorylation of the α subunit of the eukaryotic translation initiation factor 2 (eIF2α) and inhibition of protein translation within infected cells [11]. In addition, PKR evokes cellular responses by modulating cell-signaling pathways. The mechanisms by which PKR functions as a signaling molecule have not been fully delineated. However, PKR has been shown to mediate the responses to other PAMPs, including bacterial LPS, as well as stress stimuli such as IFNγ, TNFα, mitomycin C, and serum deprivation by inducing degradation of inhibitor κB (IκB), IRF1 expression, indirectly mediating STAT1 phosphorylation, and triggering apoptotic pathways [12],[13]. These broad responses are not reconciled with a narrow mechanism involving translational control through eIF2α. However, few other PKR substrates are known that account for these cellular responses.
PKR has two domains, a C-terminal catalytic domain and an N-terminal regulatory domain. The N-terminus encodes tandem RNA-binding motifs (RBMs). The RBMs not only recognize dsRNA to activate PKR, but also serve as an autoinhibitory domain, as well as mediating dimerization to form the fully active kinase molecule. These observations suggest an additional function for RBMs as protein–protein interaction domains. Support for this comes from other proteins identified to interact with PKR. The protein activator of PKR (PACT) encodes three RBMs, and the structurally similar transactivating response (TAR)-RNA binding protein (TARBP) interacts with PKR to, conversely, inhibit the kinase. That RBMs might mediate interactions between proteins, particularly other RBMs, highlights an emerging concept that has consequence for coordinating the dsRNA response in cells. In this way the RBM might be considered as a signaling domain, analogous to the TIR domain of TLRs or the CARD domain of RIG-I and IFIH-I to mediate homo- or heterotypic protein interactions.
Here, we identify an interaction between PKR and another protein encoding RBMs, the RNA helicase A (RHA). RHA is an essential DEAH-box protein that exhibits both RNA and DNA helicase activity [14]. The association is demonstrated to be via the helicase RBMs. Importantly, biochemical analysis shows RHA is a substrate for PKR, and demonstrates that phosphorylation modulates the helicase association with its nucleic acid substrate. The consequences of these observations are examined in relation to RHA's previously established role in retroviral infection. PKR is shown to mediate transcriptional activity and HIV-1 infectivity via phosphorylation of RHA. These findings identify a novel function for PKR, delineating a new cell signaling pathway to target in anti-HIV-1 therapy, and highlighting a process by which proteins that respond to dsRNA may be coordinated.
To detect proteins interacting with PKR, the kinase was immunoprecipitated from isogenic, pkr-null mouse embryonic fibroblasts (MEFs) either silenced or expressing human PKR from the native promoter elements at physiologic levels. The cells were stimulated with the synthetic, dsRNA mimic polyinosinic-polycytidylic acid (pIC) to activate the kinase. A 140 kDa protein band was coimmunoprecipitated with a monoclonal anti-PKR antibody (Figure 1A). Mass spectrometric analysis of this protein identified 11 different peptide sequences that matched the amino acid sequence of murine RHA (Table 1).
Immunoprecipitation and Western blot analysis were conducted to verify this protein association. The protein interaction was confirmed in MEFs isolated from wild-type and pkr null mice by performing the reciprocal immunoprecipitation, using a anti-RHA monoclonal antibody, then detecting coimmunoprecipited PKR (Figure 1B). This experiment was repeated in the rescued pkr null MEFs expressing the human kinase, with and without stimulation of the cells with pIC to activate PKR. Coimmunoprecipitation of PKR and RHA only occurred with pIC treatment, demonstrating that the protein interaction is dependent upon activation of the kinase (Figure 1C). Notably, pIC appears to modulate the conformation of RHA as the protein is not immunoprecipitated by its own antibody from untreated cell lysates. This supports a previous suggestion that the constitutively expressed helicase is maintained in an inactive conformation, likely via the protein's RBM, under basal conditions [15].
In vitro binding studies were conducted to map the association between PKR and RHA. Six different glutathione-S-transferase (GST) fusion constructs, together spanning the full helicase, were used in a GST-pull down experiment to determine which region of RHA associated with PKR. This experiment showed PKR associated exclusively with the first 263 amino acids of RHA that encodes its two RBMs (Figure 2A). Since the first RBM, encoded between amino acids 1 to 79, did not associate with PKR, it appears that the second RBM is the interacting region, or both RBMs are required. Efforts to map the region of PKR that interacts with RHA were inconclusive. Binding assays with the isolated RBM or kinase domains of PKR showed the N-terminal RBMs interacted specifically with RHA. However, the truncated C-terminus of PKR bound non-specifically to the control (beads only) in the assay conditions. Subsequent analysis suggests the RBM of RHA interacts with both C- and N-terminal domains of PKR (see below).
RHA encodes several domains that bind dsRNA, located in the helicase domain, the C-terminal RGG box, as well as the two RBMs at the N-terminus, that did not associate with PKR. This implies that the PKR-RHA interaction is direct and not through mutual association with dsRNA. Moreover, it has been established that PKR dissociates from dsRNA upon activation, autophosphorylation and dimerization [16]. Consequently, an indirect association between PKR and RHA, bridged by dsRNA, is unlikely. However, to unequivocally establish that the two proteins interact directly, several experiments were conducted.
First the ability of a 16 bp dsRNA molecule to block the association between in vitro synthesized PKR and the N-terminal 263 amino acid GST-fusion construct of RHA was measured. This short dsRNA, although able to bind to a single RBM, is not long enough to interact with two RBMs from separate proteins [17]. The results (Figure 2B) showed that the 16 bp dsRNA did not perturb the association between PKR and RHA, even when present at considerable molar excess. Next, alternative cell treatments that did not use pIC to activate PKR were evaluated and coimmunoprecipitations performed. Accordingly, treatment of the human monocytic cell line, THP-1, with LPS, demonstrated to activate PKR [18], induces the association between the kinase and RHA (Figure 2C). Finally, triggering of PKR with its protein activator PACT [19] effected association between PKR and RHA. HEK293T cells were cotransfected with expression constructs for RHA and wild-type PACT or a mutant (ΔPACT) that does not activate PKR and treated with actinomycin-D to stimulate PACT, then PKR was immunoprecipitated. Figure 2D shows the wild-type and not the mutant PACT increased the association between PKR and the helicase. Together these data demonstrate that PKR interacts directly with the N-terminal region of RHA, and that this interaction is dependent upon activation of the kinase.
To investigate the possible consequences of this interaction, we tested whether RHA is a substrate for PKR-mediated phosphorylation. An in vitro kinase assay was performed with proteins coimmunoprecipitated from MEFs using a monoclonal antibody against PKR. The results (Figure 3A) show RHA bound to PKR at physiological ratios was phosphorylated by the associated PKR in a subsequent kinase assay. To measure phosphorylation in vivo RHA was directly immunoprecipitated with antibodies to RHA, or coimmunoprecipitated with an anti-PKR antibody from HeLa cells treated with pIC. The resulting immune complexes were probed by western blot for phosphorylated residues using anti-phosphoserine and anti-phosphothreonine antibodies. Figure 3B shows only the RHA in complex with PKR had detectable phosphorylated serine and threonine residues. Detection of phosphorylated RHA specifically associated with the kinase, strongly suggesting direct phosphorylation of RHA by PKR. The antibody used to immunoprecipitate RHA in this experiment (ab2627 from Abcam) was raised against a synthetic peptide derived from within residues 100 to 200 of human RHA. This is within the region of RHA demonstrated to associate with PKR (between residues 80 to 263, Figure 2A). As this antibody and PKR interacts with the same region of RHA mutual association is excluded. Consequently, PKR is not coimmunoprecipitated with this anti-RHA antibody (Figure 3B). This confirms the GST-pull down experiments and further narrows the region mediating the interaction between the RHA and PKR.
Phosphorylation of RHA by PKR was confirmed in an in vitro kinase assay, using purified recombinant PKR and RHA (Figure 3C). To examine the possible functional consequences of phosphorylation by PKR we mapped the region of RHA that is modified. Accordingly, an in vitro kinase assay was conducted with truncated GST-fusion constructs of RHA and recombinant PKR. Figure 3D demonstrates that RHA is phosphorylated within the 263 amino acid region previously demonstrated to interact with PKR. Hence, the RBM of RHA must interact with the catalytic kinase domain of PKR. This is consistent with previous evidence showing other RBMs, for instance from PACT, are phosphorylated by PKR [20],[21].
As the N-terminal 263 amino acid region of RHA regulates the association with dsRNA, it seemed evident that addition of a negatively charged phosphate group to this region would perturb RHA's interaction with dsRNA. To test this hypothesis we measured the relative affinity of the phosphorylated or unphosphorylated RHA peptide for pIC. The 263 amino acid RHA peptide was either taken directly from an in vitro synthesis reaction, or subsequently phosphorylated by PKR in an in vitro kinase assay following synthesis. Approximately 19% of the total 35S-labeled RHA peptide was recovered in a pIC pull down. In contrast, none of the phosphorylated RHA peptide, evidenced as a 32P-labeled product, bound pIC (Figure 4). Consequently, phosphorylation of the RHA peptide by PKR inhibited pIC binding. Given that phosphorylation perturbs RHA's association with its nucleic acid substrate, we would expected PKR should have a profound effect upon the function of the helicase in vivo.
Our data supports a model in which PKR regulates RHA by phosphorylating its RBD thereby decreasing its affinity for RNA. To determine the in vivo consequence of such regulation, we investigated PKR's effect on the reported ability of RHA to regulate transactivation of the HIV-1 LTR [22]. Accordingly, transcription of an LTR-EGFP reporter construct was measured in HEK293T cells in which PKR was depleted by RNA interference (RNAi). Additional control small interfering RNAs (siRNAs) against RHA, EGFP, and as an alternative target Lamin A/C, were cotransfected with the reporter construct to gauge the RHA dependence of transactivation, the efficacy of RNAi, and to account for non-specific effects of RNAi, respectively. Since the HIV-1 LTR RNA has been reported to bind and activate PKR, no further activating stimulus was used [23]. As depletion of PKR can increase gene expression by reduced phosphorylation of the eIF2α translation factor, we delineated specific regulation of LTR-transactivation by normalizing reporter protein levels to an internal constitutive Renilla reporter. Western blot analysis confirmed the specific release of the reporter gene (GFP) relative to the constitutively expressed GAPDH, and verified appropriate targeting of each siRNA against PKR, RHA, and as a control GFP (Figure 5B). The control siRNA to Lamin A/C did not affect reporter protein levels and is not shown. Depletion of RHA by siRNA confirms the role of the helicase in LTR-regulated gene expression (Figure 5A). Significantly, depletion of PKR increased EGFP expression. The timing of transcriptional release, beginning at 48 hours, conforms to the anticipated kinetics of the removal of PKR from the cell, as the kinase has an approximate half-life of 48 hours (B. R. G. Williams, unpublished results).
This effect of PKR on the LTR reporter system was further tested using three PKR constructs with different catalytic activity. The requirement for kinase activity for PKR control of RHA-mediated LTR expression was assessed by comparing the relative affect of wild-type PKR, and two mutant PKR proteins either; catalytically active but modified to preclude eIF2α regulation by substitution of the threonine residue to an alanine at position 487 (T487A) [24], or a kinase dead construct modified by substitution of a lysine residue to a arginine at position 296 (K296R). Expression of wild-type PKR reduced RHA-dependent transcription of the LTR-EGFP reporter. Conversely, expression of the catalytically inactive PKR-K296R promotes RHA-dependent transcription of the reporter (Figure 5C). This construct (K296R) dimerizes with endogenous PKR, so acts as a dominant negative to directly inhibit PKR's regulation of RHA as well as general protein translation, via wild-type PKR phosphorylation of eIF2α. The relative contribution of these two mechanisms was explored by expressing the mutant PKR-T487A that mediates association with eIF2α. This construct is catalytically active, so will phosphorylate RHA, but is incapable of regulating translation. Accordingly, expression of the PKR-T487A showed an intermediate affect on the LTR-driven reporter, reflecting direct inhibition of RHA-mediated induction of the reporter without the wild-type PKR-mediated regulation of global protein translation (Figure 5C). The relative contribution of PKR's direct regulation of RHA juxtaposed to indirect effect upon translation, demonstrated with either PKR mutation (K296R or T487A), is made more clear when the RHA-dependent transcription of the HIV-1 LTR reporter (EGFP) is normalizing against a constitutive reporter (Renilla luciferase). This normalization shows the catalytically inactive PKR-K296R does not affect RHA-mediated transactivation of the HIV-1 LTR, while the catalytically active PKR-T487A construct reduces transactivation of the HIV-1 LTR. This data shows that PKR negatively regulates RHA transactivation of the retroviral reporter gene by direct phosphorylation control of RHA.
The preceding data predicts over expression of the RBD of RHA would perturb PKR function by acting as a decoy substrate. This prediction is supported by reporter assays in HEK293T cells that show RHA-regulated LTR expression increases with increasing amounts of the RHA RBD (Figure 6A). This rescue effect of RHA's RBD in the reporter assays should extend to full viral infection. To test this, we measured the capacity of constructs that encoded RHA's RBD, and two truncated constructs, of each separate RBM within this domain, to enhance HIV-1 infection in the peripheral blood mononuclear cell line MT-2. As a previous report had demonstrated that RHA becomes incorporated into the HIV-1 virion during replication [25], infectious virus was produced in cells co-expressing the three RHA peptides (RBD, RBM1, and RBM2, encoding residues 1-263, 1-76, and 169-263, respectively), and the virions produced were titrated onto the mononuclear cells. In keeping with the reporter assays, expression of the RBD significantly increased HIV-1 infectivity. Notably, a truncation construct from the first RBM (RBM1) that did not associate with, and was not phosphorylated by, PKR (Figures 2A and 3D), did not alter HIV-1 infectivity. In contrast, the construct encoding the second RBM (RBM2), predicted to be the substrate for PKR, did enhance viral infectivity. In fact, this peptide was more potent than the domain that encompassed both motifs (Figure 6B). As these truncated constructs have no helicase activity and lack other domains demonstrated to enhance retroviral replication, increased virus infectivity is presumed to be due to the demonstrated association and inhibition of PKR. However, an alternative mechanism is conceivable whereby RHA's RBD might recruit other cellular factors to enhance viral replication. This was assessed by measuring the activity of reverse transcriptase in infections with HIV-1 produced with the control plasmid or each of the RHA constructs. Importantly, the RHA peptides did not increase viral replication, as measured by the activity of the viral reverse transcriptase enzyme (Figure 6C). These experiments validate the preceding data in a cell infection system, substantiating a consequence of the interaction between PKR and RHA for the cell's innate immune response to HIV-1 infection.
The innate immune response is the primary shield against microbial infection and directs the subsequent adaptive response. The protein kinase PKR was identified some 30 years ago as a sentinel kinase that is constitutively expressed in all cells as a monomer that subsequently dimerizes to form the active enzyme. We show here that RHA is a novel substrate for PKR and explore points of significance that arise from the finding.
PKR's interaction with RHA identifies a novel mechanism by which the previously established translational regulator can also modulate transcription. This function of PKR identifies an antiviral pathway that represents a plausible target for treatment of established retroviral infections. Consistent with this antiviral mechanism, RHA is positively associated with viral replication. RHA transactivates the Bovine Viral Diarrhea virus by binding to the terminal nontranslated regions of the viral RNA genome [26]. RHA also positively regulates expression of the HIV-1 transactivation response region [27] (and in this study). In addition, RHA mediates release of retroviral transcripts from the splicesome and transports the RNA from the nucleus. Correspondingly, the helicase has been shown to associate with the constitutive transport element (CTE) of type D-retroviruses and Rev Response elements of HIV-1 and to associate with cellular mRNA export receptors TAP, SAM68, and HAP95 [28],[29]. Of particular relevance to this study, RHA also associates with the HIV-1 gag protein and becomes incorporated into the HIV-1 particle [25]. Our data demonstrates that coexpression of HIV-1 provirus with RHA peptides that are substrates for PKR subsequently enhances viral infectivity. Significantly, the truncated RHA peptides do not encode any helicase activity and are therefore incapable of transactivating the HIV-1 LTR sequences. Appropriately, no benefit to virus replication was observed by co-expressing RHA peptides. We contend the observed increased infectivity, without increased replication, is due to inhibition of the ensuing antiviral response mediated by PKR, through the interaction between RHAs second RBM and PKR. Therefore an additional function of RHA possibly exploited by HIV-1 is to dampen the primary host immune response. Such a role adds weight to the previous observed incorporation of RHA into the virion.
RHA was initially identified as a homolog of the Drosophila melanogaster maleless gene that regulates chromosomal dosage compensation, a function essential for survival of male larvae [30]. Deletion of rha-1 in Caenorhabditis elegans indicates that the helicase controls germ cell proliferation and development [31]. Embryonic lethality of rha null mice at day 11 of gestation shows that the helicase is also essential for development in mammals [32]. Several lines of evidence suggest that RHA may also have a role in the immune response. The helicase appears as an auto-antigen in the auto-immune disease systemic lupus erythematosus [33]. In addition, RHA associates with the transcription cofactor and histone acetyltransferase CBP, and the transcription factor NFκB, both potent factors in immune responses [34],[35]. Furthermore, the rha gene promoter contains regulatory elements that control induction of this constitutively expressed protein during cellular immune responses, including an Interferon Stimulatory Response Element. Interestingly, immunohistochemistry of IFNα-treated cells shows accumulation of the helicase within promyelocytic leukemia nuclear bodies that are involved in transcription of IFN-induced genes [35]. Accordingly, RHA may not only be induced by IFN, but could also regulate its downstream effects. Therefore appropriation of RHA by viruses during their replication would not only boost viral transcripts, but may also blunt the innate immune response.
Our observations of interplay between RHA and PKR strengthen the perspective that helicases are key signaling molecules. Helicases had been thought of as terminal proteins in signal cascades that elicit appropriate responses by remodeling RNA and DNA. The data here underpin previous findings with RIG-I and IFIH-I to support a primary role for helicases as immediate players in the innate immune response [36]. We demonstrate that just as the CARD domains of these helicases and their associated adaptor molecules mediate signal transduction, the RBM of RHA mediates the association with PKR. Importantly, by identifying an inhibitory effect of phosphorylation on the function of RHA, we present compelling evidence of this association, with resulting effect upon the enzyme's function. Correspondingly, peptides within RHA's RBD, that interacts with PKR, enhance the infectivity of HIV-1. The data support a paradigm by which the function of a class of RNA-responsive proteins are coordinated or exacerbated by interaction via their RBMs. The consequence of this could be considerable, as at least 17 human proteins have been described that encode RBMs. Moreover, gene deletion studies highlight the importance of these proteins. Disruption of the RBM-containing ribonuclease Dicer, TARBP, the adenosine deaminase ADAR-1, and, as discussed, RHA, is embryonically lethal in murine models [32], [37]–[39]. Similarly, mice null for the PKR-activator PACT, spermatid perinuclear RNA-binding protein (STRBP), and the testis-specific mRNA editor TENR, which all encode RBMs, have retarded growth, increased mortality and/or reduced fertility (G.C. Sen, unpublished results; L. Saunders and G.N. Barber, unpublished results; [40]–[42]. PKR has previously been reported to associate with four members of this family of proteins. As mentioned the kinase is activated by PACT, and conversely inhibited by TARBP, in addition to the nuclear factor of activated T-cells, NF90, as well as the antiviral protein ADAR1 [43],[44]. The association here between PKR and RHA via their RBMs strengthens an emerging paradigm whereby this motif acts as a signaling domain to coordinate the dsRNA-response as has been identified for the CARD domains of the cytoplasmic helicases RIG-I and IFIH-I, or the TIR domains of TLRs and their adaptor molecules.
Full-length RHA and truncated GST-RHA fusion plasmids were constructed as described by Nakajima et al. [34]. RHA was expressed for protein purification as a recombinant baculovirus as described by Lee et al. [30]. Wild-type PKR was expressed as described previously by Gabel et al. [16]. Truncated RHA constructs, encoding the N-terminal 262 amino acid (pRHARBD), the first 76 amino acids (pRHARBM1), or residues 169 to 262 (pRHARBM2), were generated in pCMVFlag (Sigma). Other plasmid constructs were gifted by others as listed in the acknowledgments.
Gene silencing was achieved though RNA interference using the chemically synthesized siRNAs, AAAUUUUCUGUAUGCCUGG, CAGCCAAAUUAGCUGUUGA, AATGTTCTTCTGGAAGTCCAG, and GCUGACCCUGAAGUUCAUCUU, targeting rha, pkr, lamin A/C, and egfp transcripts (Dharmacon). All other reagents were purchased from Sigma unless otherwise indicated.
Adherent cells were maintained in DMEM, while suspension cells were cultured in RPMI supplemented with 10% fetal bovine serum and cells were grown at 37°C with a humidified 95% air, 5% CO2 atmosphere. Murine (C57/BL6) PKR null MEFs were transformed with the pBeloBAC construct encoding an approximately 60 kbp genomic fragment that encompassed the gene and promoter elements of human pkr as described previously [45],[46].
Reporter assays were performed in HEK293T cells at 20–60% confluency in 6-well dishes (Falcon). Cells were transfected using the calcium phosphate method with 300 ng pLTR-EGFP, 2 ng pSV2tat72, 10 ng of a control reporter pβactin-RL and 4 nM siRNA per well. Cells were collected 24, 48, and 72 hours after transfection, Assays to measure the effect of the pRHARBD were performed in HEK293T cells cultured in 24 well dishes transfected with 50 ng of pLTR-EGFP, 10 ng pβactin-RL and 0, 10, 20, 40, 80, 160, or 640 ng of either pRHARBD or pCMVFlag DNA. The cells were cultured for 62 hours. HEK293T cells were washed with phosphate-buffered saline (PBS), and lysed in Promega's passive lysis buffer for fluorescence and luciferase measurements using a Wallac Victor3 plate reader (Perkin-Elmer). Fluorescence values were normalized to the total protein level quantified using the Bradford assay (BioRad) and also compared to an internal reporter quantified by Renilla luciferase assay (Promega). All experiments were performed in triplicate and independently replicated a minimum of three times.
PKR was activated in MEFs by adding 100 μg/ml pIC to the culture supernatant for 2 hours. Alternatively, THP-1 cells were treated with 10 μg/ml E. coli LPS for 2 hours as described by Gusella et al. [47]. Finally, HEK293T transfected with pcDNA-PACT/ΔPACT were temporally treated with Actinomycin-D as described by Peters et al. [20].
HIV-1 particles were produced by polyethylenimine transfection of HEK293T cells with 5 μg of pNL4-3-Luc-RE proviral DNA, 2.5 μg of pNLA1, and 2 μg of each RHA construct (pRHARBD, pRHARBM1, or pRHARBM2). Viral particles were harvest after 36 hours, purified from the supernatant and concentrated by ultracentrifugation through 20% sucrose, using ultracentrifuge at 87,000×g for 1 hour at 4°C in a Beckman centrifuge, and virus pellets were eluted in PBS, and quantitated with the HIV-1 Antigen p24-CA MicroELISA Vironostika system (Organon Teknika). Equivalent amounts of virus were used to infect 1×106 MT-2 cells maintained in RF10 (Gibco/BRL), supplemented with 2 mM L-glutamine and 24 μg/ml gentamicin for 2 hours at 37°C. Residual virus was removed by washing with PBS and cells were resuspended in RF10, then cultured at 37°C for 48 hours, before washing with PBS and harvesting in Cell Culture Lysis Reagent (Promega). The success of a single round of infection was determined by the level of luciferase activity, measured by luciferase assay (Promega) using a Fluorostar plate reader (BMG). HIV-1 infectivity was assessed in three independent experiments with two or four replicates at each occasion.
Ten μl of non-concentrated viral supernatant was mixed with 10 μL of 0.3% NP40, followed by addition of 40 μL reverse transcriptase (RT) reaction cocktail containing 5 μg/ml of the template primer poly(rA)-(dT)15 (Amersham Pharmacia Biotech), in 50 mM Tris-HCl (pH 7.8), 2 mM DTT, 5 mM MgCl2, 7.5 mM KCl, and 0.5 mCi α33P-dTTP. Following incubation for 2 hours at 37°C, 8 μL of the reaction mixture was spotted onto DEAE81 ion-exchange paper (Whatman) and washed six times in 300 mM NaCl and 30 mM sodium citrate to remove unincorporated α33P-dTTP. RT activity was determined by the level of α33P-dTTP using a Wallac 1450 Microbeta-Plus liquid scintillation counter (Perkin-Elmer).
Human PKR was immunoprecipitated using the mouse monoclonal antibody 71/10 [48]. PKR was detected in Western blot with multiple redundant antibodies, including a rabbit monoclonal antibody YE350 from Abcam (for human PKR), and rabbit polyclonal antibodies D20, and B10 from Santa Cruz Biotechnology. Activation of PKR in vivo was confirmed by detecting phosphorylation of eIF2α using a rabbit anti-phospho-eIF2α (Ser51) antibody from Stressgen. GAPDH and GFP were detected in Western blots using mouse monoclonal antibodies from Chemicon and Roche, respectively. Endogenous RHA was detected in Western blot analysis and immunoprecipitated from whole-cell lysates using a rabbit polyclonal antibody [30], a rabbit polyclonal antibody ab26271, and a mouse monoclonal ab54593 from Abcam. Recombinant HA-tagged RHA was immunoprecipitated and detected in Western blots using the monoclonal antibody HA.11 from Covance. Phosphorylated amino acids were detected using rabbit polyclonal anti-phosphoserine, and mouse monoclonal anti-phosphothreonine antibodies from Zymed Laboratories (Invitrogen). Cells were collected in lysis buffer (50 mM Tris-HCL [pH 7.4], 150 mM NaCl, 50 mM NaF, 10 mM β-glycerophosphate, 0.1 mM EDTA, 10% glycerol, 1% Triton X-100, and protease inhibitors). Immune complexes were resuspended in loading buffer (125 mM Tris-HCl [pH 6.8], 4% SDS, 20% glycerol, 10% β-mercaptoethanol, 1% Bromophenol Blue) for separation by SDS-polyacrylamide electrophoresis (SDS-PAGE). Separated proteins were visualized by staining with BioRads Coomassie, or Silver Stain Plus reagents. Stained protein bands were excised from the gel and analyzed by Maldi-ToF. Alternatively, separated proteins were electrophoretically transferred to either Immobilon-P, or Immobilon-FL membrane (Millipore) for immunoblotting using horseradish peroxidase-linked secondary antibody and ECL from Amersham, or fluorescently labeled (680 and 800 nm) secondary antibodies (Invitrogen, Molecular Probes), respectively. Fluorescently labeled antibodies were detected and quantitation using the Odyssey infrared imaging system (Li-Cor). Replicate experiments to quantitate RHA phosphorylated by PKR in vivo recorded mean values of phosphorylated residues on RHA coimmunoprecipitated with PKR of 1.6+/−0.9 for phosphoserine and 5.7+/−1.3 for phosphothreonine. No phosphorylated residues were detected with these phospho-specific antibodies in RHA directly immunoprecipitated. The values of total RHA coimmunoprecipitated with the anti-PKR antibody were 55.3+/−5.8, while that directly immunoprecipitated with the anti-RHA antibody was 111.5+/−13.2 in this experiment.
GST and His-tagged proteins were purified from E. coli and Sf-9 insect cells on either glutatione-Sepharose 4B beads (Amersham) or Ni-NTA agarose (Qiagen) according to the manufacturer's protocols. To map protein interactions, PKR was synthesized in an in vitro coupled transcription–translation system (Promega) with 35S-methionine (NEN-DuPont), then incubated in cleared E. coli lysate with protease inhibitors with GST-fused RHA constructs for 2 hours at 4°C. Supernatant, containing unbound proteins, was removed after 500×g centrifugation. Recovered beads were rinsed five times with bead-binding buffer (50 mM K3PO4 [pH 7.5], 150 mM KCl, 1 mM MgCl2, 10% glycerol, 1% Triton X-100 and protease inhibitors). The resin-bound proteins were eluted with loading buffer and separated by SDS-PAGE, then visualized by autoradiography.
The experiment on the effect of RNA on the interaction between PKR and the 263 amino acid GSTRHA-fusion peptide was conducted as above with an additional step. Approximately 20 μg of the GSTRHA peptide was incubated with a 16 bp dsRNA at 10, 100, or 1000-fold excess for one hour prior to addition of 35S-labeled PKR. The 16 bp dsRNA was synthesized in vitro using T7 RNA polymerase then gel purified from an SDS-PAGE gel.
To measure the relative affinity of unphosphorylated or phosphorylated RHA for RNA, the 263 amino acid N-terminus of RHA was synthesized in vitro with either 35S-methionine during the synthesis reaction or γ32P-ATP in a PKR kinase assay. Labeled proteins were incubated in binding buffer (20 mM Tris-HCL [pH 7.4], 200 mM NaCl, and 5 mM DTT) with pIC conjugated to agarose beads (Promega) for an hour, then washed with binding buffer five times, eluted with loading buffer and separated by SDS-PAGE. Recovered proteins were detected by exposure to a phosphor screen, imaged with a Storm-840 scanner, and quantified with ImageQuant software (Molecular Dynamics).
For kinase assays, full-length RHA, truncated GSTRHA fusion proteins, and the PKR substrate B56α [49] were incubated with recombinant PKR in 30 μl DBGA buffer (10 mM Tris-HCl [pH 7.6], 50 mM KCl, 2 mM [CH3COO]2Mg4H2O, 7 mM β-mercaptoethanol, 20% glycerol), 20 μl of DBGB buffer (2.5 mM MnCl2 in DBGA), 5 μl of ATP mixture (10 μM ATP and 1.5 μCi of γ32P-ATP/ml), and 5 μl of pIC (12 ng/μl) at 30°C for 10 minutes. Phosphorylated proteins were denatured in loading buffer and separated by SDS-PAGE, then visualized by autoradiography. The levels of total proteins in the SDS-PAGE gel were visualized by staining with Coomassie blue. In vivo phosphorylation of RHA was detected as described above (Immune analysis).
Protein were excised from SDS-PAGE gels and washed in 50% ethanol, 5% acetic acid, reduced and alkylated with DTT and iodoacetamide. The gel slices were dehydrated in acetonitrile and dried in a speed-vac, then digested in 20 ng/ml Trypsin in 50 nM ammonium bicarbonate overnight at room temperature. Released peptides were extracted from the polyacrylamide with 50% acetonitrile with 5% formic acid. The extract was evaporated for LC-MS analysis using a Finnigan LTQ linear ion trap mass spectrometer. Two μl volumes of the extract were injected and the peptides eluted from the column by acetonitrile in a 50 mM acetic acid gradient at a flow rate of 0.2 μl/minute. The microelectrospray ion source was operated at 2.5 kV. Samples were also analyzed by Maldi-ToF. Data collected in the experiment was used to search the NCBI non-redundant database with the search program TurboSequest.
|
10.1371/journal.pntd.0004508 | Evaluation of Parasiticide Treatment with Benznidazol in the Electrocardiographic, Clinical, and Serological Evolution of Chagas Disease | Chagas disease is one of the most important endemic parasitic diseases in Latin America. In its chronic phase, progression to cardiomyopathy has high morbidity and mortality. The persistence of a normal electrocardiogram (ECG) provides a similar prognosis to that of a non-diseased population. Benznidazole (BNZ) is the only drug with trypanocidal action available in Brazil.
A group of 310 patients with chronic Chagas disease who had normal ECGs at the first medical visit performed before 2002 were included. There were 263 patients treated with BNZ and 47 untreated. The follow-up period was 19.59 years. Univariate analyses showed that those treated were younger and predominantly male. As many as 79.08% of those treated and 46.81% of those untreated continued with normal electrocardiograms (p <0.0001). The occurrence of electrocardiographic abnormalities and relevant clinical events (heart failure, stroke, total mortality, and cardiovascular death) was less prevalent in treated patients (p <0.001, p: 0.022, p: 0.047 respectively). In multivariate analyses, the parasiticide treatment was an independent variable for persistence of a normal ECG pattern, which was an independent variable in the prevention of significant clinical events. The immunofluorescence titers decreased with the parasitological treatment. However, the small number of tests in untreated patients did not allow the correlation of the decrease of these titers with electrocardiographic alterations.
These data suggest that treatment with benznidazole prevents the occurrence of electrocardiographic alterations. On the other hand, patients who develop ECG abnormalities present with more significant clinical events.
| Twenty years of follow-up of patients with Chagas disease treated with benznidazole is presented in this paper. The persistence of a normal electrocardiogram (ECG) provides a similar prognosis to that of a non-diseased population. Benznidazole (BNZ) is the only drug with trypanocidal action available in Brazil. A group of 310 patients with chronic Chagas disease who had normal ECGs at the first medical visit performed before 2002 were included. There were 263 patients treated with BNZ and 47 untreated. The occurrence of electrocardiographic abnormalities and relevant clinical events (heart failure, stroke, total mortality, and cardiovascular death) was less prevalent in treated patients. In multivariate analyses, the parasiticide treatment was an independent variable for persistence of a normal ECG pattern, which was an independent variable in the prevention of significant clinical events. The immunofluorescence titers decreased with the parasitological treatment. However, the small number of tests in untreated patients did not allow the correlation of the decrease of these titers with electrocardiographic alterations. These data suggest that treatment with benznidazole prevents the occurrence of electrocardiographic alterations. On the other hand, patients who develop ECG abnormalities present with more significant clinical events.
| Chagas’ disease (CD), described by Carlos Chagas in 1909[1], and caused by a parasite–Trypanosoma cruzi, is one of the most important endemic diseases in Latin America, where there are 10 million people infected (about two million in Brazil). The vectorial transmission has historically been the most important. The disease may also be conveyed by blood transfusion, be congenital, or be transmitted orally (this is the most prevalent today in Brazil), among other types of transmission[2][3]. With globalization, many Latin Americans migrated to other continents, carrying this disease and transmitting it through blood transfusion to the inhabitants of non-endemic countries. Therefore, CD is now present in North America, Europe, Asia, and Oceania, and is becoming a worldwide public health problem[4].
After contamination, the acute phase occurs, characterized by severe inflammation and intense parasitemia, although with limited clinical impact and low mortality. This phase lasts for approximately 8 to 10 weeks, followed by the chronic phase with a decrease of parasitemia and inflammation, but not to extinction. Sixty to 70% of patients remain in the indeterminate form (positive serum reaction, no clinical signs, normal electrocardiogram, normal Chest X-ray, normal esophagogram, and normal barium enema). A total of 40 or 30% evolve to clinical forms, with isolated or concomitant heart, esophagus, and colon involvement[2].
The electrocardiogram (ECG) is a very important tool in monitoring patients with CD. Maguire et al[5], following a population of CD patients for seven years, showed that those who maintained a normal ECG, evolved in a similar way to individuals without the disease. This simple test has important prognostic value, and usually is sufficient for clinical follow-up[6][7].
The parasite's role in the chronic phase remains unclear, even one hundred years after the description by Carlos Chagas[8][9][10]
Parasiticide treatment is controversial as to its indication in the chronic phase and as to its real benefits. The criteria for assessing the possible medication benefits and certainty of a cure are not unanimous among authors. Benznidazole is the only drug in Brazil with proven parasiticide action. It is available in 100 mg tablets and the dose recommended for acute patients or children, is 10 mg/kg/day for 60 days of treatment, and in the chronic phase, 5 mg/kg/day, also for 60 days. Major side effects are dermatitis, which occurs in 30% of cases, and polyneuropathy, which is less prevalent. Patients usually tolerate well the side effects described. Significant leukopenia and liver damage are rarely observed, and the occurrence of agranulocytosis is exceptional[11][12][13][14].
The BENEFIT study that randomly evaluated the treatment with BNZ in 2854 (1431 BNZ and 1423 placebo) patients with chronic Chagas cardiomyopathy, NYHA class I, II, III, (97% class I and II), followed for the short period of 5.4 years, showed a significant decrease in parasitemia (PCR test) in the BNZ group. However, no difference in the occurrence of events during this period (death, resuscitated cardiac arrest, insertion of a pacemaker, or an implantable cardioverter–defibrillator (ICD), sustained ventricular tachycardia, cardiac transplantation, new heart failure, stroke or transient ischemic attack, or a systemic or pulmonary thromboembolic event)[15].
Immunofluorescence titers are stable in untreated patients. However, treated patients showed a decrease of the titers. Negativity of the immunofluorescence titers is infrequent, but may occur persistently after more than a decade of treatment[16][17].
This is a retrospective study that analyzes the electrocardiographic, clinical, and serological evolution of patients with chronic Chagas’ disease, with or without treatment with BNZ, and who had a previous normal ECG.
In our database, we evaluated patients with CD confirmed by two or more serum reaction techniques (immunofluorescence, hemagglutination, direct agglutination, and enzyme-linked immunosorbent assay—ELISA). All subjects needed to have a normal ECG at the first medical visit at the institution, done before 2002. The decision to include patients who had had their first visit before 2002 was due to the fact of having a minimum follow-up of 10 years. There was no age limit for inclusion because all patients who satisfied the inclusion criteria were evaluated.
Most of them did not undergo a digestive tract examination (barium swallow or barium enema). Therefore, we did not refer to them as having the indeterminate form, but as asymptomatic patients without heart disease. In the loss of follow-up cases, we attempted to find them by phone, telegram, through social networks, and contact with neighbors or family, which is part of the institution's usual approach to call patients for medical visits when necessary. We do not have the success percentages in each category because there are no specific records.
The ECG of the last visit needed to have been between 2011 and 2013 to have a minimum electrocardiographic evolution of 10 years. In 2013, data collection was ended.
When there was no electrocardiographic tracing during this period, the patient was invited to do so, and those who agreed signed the Informed Consent Form (ICF). The report of the first and last electrocardiographic tracings was done without patient identification or date of execution by the Dante Pazzanese Institute of Cardiology Section of Electrocardiography, according to the Guidelines for Electrocardiographic Analysis and Reports Issued by the Brazilian Society of Cardiology – 2009 [18].
While examining the medical records, patients who presented with significant co-morbidities that could influence electrocardiographic alterations during evolution were not included, such as:
Some patients remained untreated and others received BNZ at a dose of 5 mg/kg/day for 60 days. This treatment took place before 2002. The attending physician, without prior standardization, made treatment indications. If the patient had received BNZ and discontinued treatment due to a side effect, he would remain in the treated group, as per the principle of intention-to-treat.
The first and last titers of quantitative immunofluorescence of the patients were compared using trypomastigote forms of the parasite [19] (values above 1/40 are considered positive). The quantitative immunofluorescence method is used systematically in patients treated with BNZ. Only this reaction was available.This serum reaction is part of the clinical follow-up routine of these patients. A new collection of a blood sample was requested when there was only one immunofluorescence test, and in patients who agreed signed the ICF.
In all patients included in this study, we analyzed:
For the statistical analysis, the SPSS software version 19 was used. The quantitative variables were described by mean and standard deviation. To compare quantitative variables, the Mann-Whitney test was used for variables with non-Gaussian distribution. Qualitative variables were presented as absolute and relative frequency (%). The analysis of the relationship between the variables used Fisher's exact test or chi-squared. The "t" test was not used because the variables did not follow a normal distribution in the Kolmogorov-Smirnov test.
Univariate logistic models were made for each variable and those with a significance level less than 0.15 were included in the multivariate model.
Logistic regression was done considering:
The accepted significance level was 95%.
This study is registered at the Dante Pazzanese Institutional Ethics Committee.
All adult patients gave written informed consent to participate in the study.
From a database with approximately 1500 patients with CD, 527 had a normal ECG at their first medical visit. Of these, 379 met the inclusion criteria. Three hundred and ten patients were found (81.80%), and 69 (18.20%) could not be reached and were therefore excluded. Table 1 shows baseline characteristics of the group of patients found and the group not found.
The presence of DLP (described in the medical records) was more prevalent in the group of found patients than in the group not found, with 27.40% and 2.90%, respectively (p <0.001). It was not possible to detect a statistically significant difference between groups for other variables.
We followed the 310 patients included in the study for a period of 10 to 46 years (19.59 ± 6.46), with a median of 18 years; 50% of these patients were followed for 15 to 23 years. Age at the last visit varied from 30 to 84 years (57.80 ± 10.07). Only 107 (34.52%) were male, and 231 (74.52%) were white. Two hundred and sixty-three patients (84.84%) received BNZ and 47 (15.16%) did not. The characteristics of the two groups are shown on Table 2.
Treated patients were younger (56.07 years x 68.89 years, p <0.0001), predominantly male (36.90% vs. 21.30%, p: 0.045), had left the endemic area more recently (16.77 years vs 19.65 years, p: 0.012), and 208 (79.08%) maintained normal ECGs, compared to 22 (46.81%) of the non-treated individuals (p <0.0001).
Among the treated patients, 55 (20.92%) had ECG changes, as follows: Right Bundle Branch Block in 21 (38%), nonspecific changes in ventricular repolarization in 20 (37%), and Blockage of the anterior superior division of the left branch in 11 (20%). Among untreated patients, 25 (53.19%) had worsening of the ECG: Right Bundle Branch Block in two (8%), nonspecific changes in ventricular repolarization in four (16%), and Blockage of the anterior superior division of the left branch in six (24%). Other changes detected had low prevalence.
There were no statistically significant differences between groups in the other variables.
The side effects observed in treated patients were dermatitis in 92 patients (34.98%), polyneuropathy in 12 (4.56%), and others (dyspepsia, insomnia, leukopenia less than 4000/mm3) in eight (3.04%). Twenty-six (9.89%) patients abandoned treatment due to side effects. The analysis of relevant events (heart failure, stroke, and cardiac death or due to any cause) described in the medical records were HF in eight cases (2.58%), stroke in four (1.29%), and 12 deaths (3.87%), in which six were men and six were women. The date of occurrence of these events was not available; only if they had occurred or not. Therefore, a longitudinal analysis was not possible, so only the logistic regression was done. In six of them it was possible to assume that the cause was due to CD (1.93% of all patients studied). In only one case was it not possible to determine the cause of death.
Among the 80 patients who had worsening of the ECGs, eight (10%) died and among the 230 who maintained normal ECGs, four (1.7%) died (p: 0.002). The cause of death related to CD occurred in five (6.25%) patients with ECG alterations and in only one (0.43%) with a normal ECG (p: 0.001). The eight cases of HF occurred in patients with ECG alterations. Among the four cases with stroke, two (2.5% of 80) had ECG alterations, and two (0.9% of 230) did not (p = 0.274). Combined outcomes (HF, stroke, and death) occurred in 24 cases (7.74%), 16 of them (20%) with ECG alterations and eight (3.48%) with normal ECGs (p <0.0001).
Table 3 shows the occurrence of events in patients untreated and treated with BNZ. It shows that patients treated with BNZ had fewer cardiac deaths and fewer total deaths.
There were two or more results of the immunofluorescence test in 171 patients (Table 4), 11 (6.43%) untreated and 160 (93.57%) treated. These results remained stable in untreated patients (232.72 ± 104.02 and 254.54 ± 93.41), whereas in the treated individuals, the titers decreased (144.90 ± 109.80 and 70.25 ± 74.70: p < 0.0001). In the 112 patients who remained with normal ECGs and without any relevant clinical outcomes, the titers of the first and last reactions were 127.50 (± 104.60) and 63.21 (± 65.95), respectively (p <0.001). Titers decreased 39.93% in patients who had ECG alterations and 50.43% in those with normal ECGs (p: 0.863). The difference in years from first to last serology in those who had ECG alterations was 14.18 (± 4.09) years, and in those who maintained a normal ECG it was 14.04 (± 5.01) years (p: 0. 15).
The negativity of the immunofluorescence titer (<1/40) occurred in 60 patients treated with BNZ (37.50%), with an average of 14 years follow-up, and in none of the untreated individuals.
In the multivariate analysis (Table 5) with dependent variables, the occurrence of combined events (heart failure, stroke, and total mortality) and independent variables, treatment with BNZ, follow-up time, males, white ethnicity, and age, it was observed that with the withdrawal of the ECG from this model, the parasiticide treatment was the only protection against events.
Table 6 assesses another logistic regression model, analyzing the dependent variable, persistence of a normal ECG, with the independent variables, treatment with BNZ, follow-up time, male, white ethnicity, and age. In this model treatment with BNZ and white ethnicity favored the persistence of a normal ECG, while the evolution of time (less than average) favored the appearance of ECG alterations.
The etiological treatment of CD remains a controversial subject due to the lack of well-conducted studies to determine the importance of parasiticide treatment. In the database of our institution, we selected patients who had normal ECGs at the first visit, and they fulfilled the inclusion criteria for this study. Despite an exhaustive search using the available resources, it was not possible to contact 69 patients. Analyzing the two groups (found and not found) by two decades of follow-up, there was no significant difference between them except for the prevalence of dyslipidemia in the group included in this study. Therefore, it would be possible to consider that patients included or not included do not evolve differently, and it was assumed that if all patients were included, the results would not have been different.
The analysis of 310 patients shows that the mean follow-up was nearly two decades (in no case less than 10 years). The majority of patients were treated. This difference is due to the fact that the etiological treatment of patients with CD is a routine in our institution. The decision to indicate treatment was made by each medical doctor with the agreement of the patient. Side effects of BNZ were well tolerated despite the fact that dermatitis is very prevalent in about 1/3 of the cases. This was why treatment discontinuation due to intolerance occurred in a small percentage (9.89%). Literature data shows treatment discontinuation due to side effects between 4% and 30%, depending on the prescribed dose [20]. Viotti et al., in an analysis of 283 patients, found 13% of treatment discontinuation [12].
When we compared the groups treated and not treated with BNZ, we noticed that the treated patients were younger. This is due to the previous observation that the older patients who had normal ECGs, rarely presented with electrocardiographic alterations, and therefore, these patients might not have received treatment with BNZ. The prevalence of males among the treated patients was interpreted as a casual observation. Treated patients had definitively left the endemic area where they lived and had been contaminated more recently. It had already been noted that the shorter the period of time away from the region where the infection occurred, the more the treatment should be indicated, because of the unpredictability of clinical evolution in a period of less than 20 years.
There are few studies about the possible benefits of etiological treatment and most of them are non-randomized or not placebo-controlled. Viotti et al., analyzing 566 patients in a chronic phase of CD and without HF (283 treated with BNZ, 5 mg/kg/day for 30 days and 283 untreated) observed electrocardiographic worsening in 4.2% of treated patients and 14.1% of untreated individuals (p: 0.002), during an average period of 9.8 years of follow-up [12]. In 21 years of follow-up, Fabbro et al. evaluated 54 patients treated with BNZ or Nifurtimox and 57 untreated patients, and noted electrocardiographic worsening in 3.7% of treated patients and 15.8% of the untreated individuals (p <0.05) [21]. In a study of 58 patients, 29 treated with BNZ and 29 untreated, after 13 years of follow-up, Machado de Assis et al. found that patients in the indeterminate form of the disease who received BNZ had clinical deterioration in 17.4%, while 56.5% (p <0.05) of the untreated ones in the same condition worsened [22].
In our study, at the end of nearly two decades of follow-up, 79.08% of the patients treated with BNZ remained with normal ECGs, while only 46.81% of patients untreated remained with normal ECGs. These data are consistent with the literature, suggesting the possibility that the BNZ administered in chronic patients with normal ECGs can prevent the onset of electrocardiographic alterations.
Despite the follow-up of 19.59 years when patients with CD and normal ECGs were analyzed, few relevant clinical events were expected due to the slow progression of the disease [6]. De Lana et al. observed 28 chronic CD patients without treatment, 22 in the indeterminate form, and noticed clinical worsening in 2% per year, whereas global clinical worsening was 0.5% per year when initially in the indeterminate form [23]. Among the patients studied, combined outcomes occurred in 24 cases (7.74%), which is consistent with the literature data, considering that all our patients began the follow-up with normal ECGs.
Patients treated with BNZ annually go through a quantitative immunofluorescence test with trypomastigotes [19]. In this group, only 171 patients had at least two immunofluorescence tests, 160 treated with BNZ and 11 untreated. Zauza et al., analyzing 140 chronic patients without parasiticide treatment after 10 years of follow-up, found a statistically significant increase in immunofluorescence titers in patients with progressive clinical deterioration, especially in the age range of 20–59 years [24]. In our study, titers of immunofluorescence of the 11 untreated patients remained stable after 14 years of follow-up. Evaluating 13 patients treated with BNZ and followed during a period of four years, Andrade et al. observed a decline in antibody titers compared to pre-treatment [25]. In a 13-year follow-up of 58 chronic patients, 29 treated with BNZ and 29 untreated, Machado de Assis et al. did not observe negativity of the antibody titers in any case, but noted falling titers in the treated group, especially those treated in the indeterminate form [22]. Viotti et al. followed for 36 months 53 chronic patients treated with BNZ and 89 untreated, and observed a decrease in antibody titers in 64% of those treated and 21% of those untreated (p <0.001). The negativity of serology occurred in 40% of those treated and in 7% of the untreated patients (p <0.001) [17]. In our study, we observed that the 160 patients treated with BNZ along a period of 14 years between the first and last tests, had a decline from the titers of the first antibody test to the last. Since there were only five patients with ECG worsening and combined events, we could not make a better evaluation of the decrease in titers of these patients. In the group of 112 patients treated whose ECGs remained normal and without any clinical events, the decrease of immunofluorescence titers was significant. Despite this variation, it was not possible to identify through the antibody titers the patients who will eventually have ECG alterations or not. Contrary to the observations of Viotti et al [17], 37.50% of our treated patients and none of the untreated patients showed negative immunofluorescence reactions after 14 years of follow-up.
Most important in the follow-up of patients with CD is the possibility of progression to clinical forms, specifically heart disease. In a 1983 publication, Maguire et al. showed the importance of the ECG in CD. Analyzing 431 patients in the chronic phase during a period of seven years, the authors concluded that individuals younger than 60 years who had normal ECGs had mortality rates similar to those of the healthy population [5].
In our study, CD patients who received BNZ remained with normal ECGs in almost 80% of the cases during a mean follow-up of two decades. In the multivariate analysis that considered the occurrence of significant combined clinical events (HF, stroke, and mortality), treatment with BNZ was an independent variable with statistical significance. In the same logistic regression model, the fact that the white patients had been shown as an independent variable favoring the occurrence of events should be analyzed with caution because of the high prevalence of white ethnicity in our study, and this data was not analyzed with more specific criteria. Finally, in another logistic regression model, treatment with BNZ was the independent variable in maintaining a normal ECG. The white ethnicity in this analysis was an independent variable in maintaining a normal ECG, and a follow-up time shorter than the average was an independent variable favoring the occurrence of ECG abnormalities. The study considered white ethnicity as a confounding factor because in one model it favors the onset of clinical events and in another model it favors the maintenance of a normal ECG. Therefore, its value is questionable.
The data obtained in our study may lead us to suppose that CD parasiticide treatment is beneficial, since the patients prevalently maintained normal ECGs, and this fact is important in a better prognosis. Many papers have been published suggesting these same results, highlighting the importance of the parasite in the maintenance of myocardial inflammation, and the elimination or minimization of its presence should be the approach followed [26].
We would like to emphasize that the randomized, placebo-controlled BENEFIT study that showed equal outcomes between patients who received BNZ or placebo, had an average follow-up period of only 5.4 years, and unlike our study, the BENEFIT study enrolled patients with established heart disease. Our observation of patients with normal ECGs at the first visit and with a follow-up of two decades had a different approach and it cannot be compared with the BENEFIT study.
The adequate and appropriate assessment of the scientific hypothesis in question should be through a randomized controlled clinical trial, which did not happen in this study. However, the data obtained provides useful information because of the large number of patients evaluated during a two-decade follow-up.
From the data obtained, it could be suggested that treatment with BNZ prevents the appearance of ECG abnormalities, and patients with normal ECGs have fewer combined events. Treatment with BNZ decreases immunofluorescence titers. Therefore, only 37.5% of the treated patients showed negativity of the immunofluorescence titers.
|
10.1371/journal.pntd.0006934 | Antivirus effectiveness of ivermectin on dengue virus type 2 in Aedes albopictus | Dengue fever is the most rapidly spreading mosquito-borne viral disease over the past 50 years, with a 30-fold increase in global incidence. Dengue vector control is a key component for the dengue control strategy, since no absolutely effective vaccine or drug is available yet. However, the rapid rise and spread of mosquito insecticide resistance have become major threats to the efficiency of insecticide-based vector control activities. Thus, innovative vector control tools are badly needed. This study aims to confirm the antivirus effectiveness of ivermectin on dengue virus type 2 (DENV-2) in Aedes albopictus (Skuse, 1894), then to explore its potential use in the combating to the dengue epidemics.
Aedes albopictus were first infected with DENV-2 in human whole blood, and at the fourth day after infectious blood feeding, they were divided into eight groups. Seven of them were held for six days with access to 0, 2, 4, 8, 16, 32 and 64 ng/ml ivermectin, respectively, and the last one was set as a historical control group, which was stored at -80°C until being detected at the same time with the other groups. Each mosquito was detected using real-time fluorescent RT-PCR kit. DENV-2 RNA concentration (copies/ml) and infection rate in each group were compared.
Both of quantitatively and qualitatively inhibiting effects of ivermectin have been detected in this study. Generally, DENV-2 replicated well in Aedes albopictus without ivermectin intervention, whose virus loads exhibited significantly higher when the mosquitoes were holding from 4 days to 10 days after infectious blood feeding. In contrast, with the treatment of ivermectin, the infection rate was reduced by as much as 49.63%. The regression equation between infection rates (Y2) and ivermectin concentration log2 values (X2) was obtained as Y2 = 91.41–7.21*X2 with R2 = 0.89.
Ivermectin can directly or indirectly inhibit DENV-2 multiplication in Aedes albopictus. Moreover, the actual concentration for application in zooprophylaxis needs to be confirmed in the further field trials.
| Dengue fever is one of neglected vector-borne tropical diseases with a 30-fold increase in global incidence recently. In 2012, World Health Organization set a goal to reduce dengue mortality by at least 50% by 2020. Being faced with more challenges in the dengue control programs, such as the increase of dengue outbreaks, lacking absolutely effective vaccine, rise of vector insecticide resistance and so on; innovative vector control tools are urgently needed for current control programs on dengue fever. To find a new avenue in vector control, we for the first time assessed the inhibiting effectiveness of ivermectin on dengue virus type 2 (DENV-2) inside Aedes mosquitoes. We found that about 80% Aedes albopictus mosquitoes were effectively infected with DENV-2 without treatment of ivermectin. But in the groups of ivermectin treatment, the infection rate of DENV-2 and the median of virus loads were significantly reduced by up to 49.63% and 99.99%, respectively. Both quantitatively and qualitatively inhibiting effects of ivermectin were detected. We found out that ivermectin was able to effectively inhibit the DENV-2 multiplication in Aedes albopictus, which may gave us a hint that using ivermectin in some control programs as a zooprophylaxis to block dengue epidemic through inhibiting DENV-2 in field Aedes mosquitoes.
| Dengue fever is the most rapidly spreading mosquito-borne viral disease over the past 50 years, with a 30-fold increase in global incidence [1]. To reverse the growing trend, comprehensive technical strategies involving diagnosis and case management, integrated surveillance and outbreak preparedness, sustainable vector control and future vaccine implementation are necessary. Apart from the other technical elements, effective vector control is a critical component to achieve and sustain reduction of morbidity attribute to dengue. There are well-documented and various historical examples of dengue elimination or significant reduction through control of Aedes aegypti (Linnaeus and Hasselquist, 1762) [2]. While bioassay demonstrates that resistance to organophosphates and pyrethroids are widespread in Aedes aegypti and Aedes albopictus [3–9]. Therefore, innovative vector control tools are badly needed for current control programs on dengue fever [1, 10]. Many new tools in vector control have been developed, such as insecticide-treated materials [11–14], lethal ovitraps [15, 16], spatial repellents [12, 14], genetically modified mosquitoes [17–19], Wolbachia-infected Aedes spp. [20], and so on. But effective tools able to block the transmission of dengue inside vector are still lacking. Therefore, we are trying to find an innovative avenue to inhibit dengue virus development inside Aedes mosquito in order to block the cycle of dengue transmission.
Two significant progresses in the tools to block the transmission of dengue inside vector benefit from the advances in genetic engineering technology and molecular biology. One is the discovery of cytoplasmic incompatibility (CI) induced by the intracellular bacteria Wolbachia (Hertig and Burt, 1924), which has enhanced replacement in the control programs [21, 22]. CI is a reproductive phenotype induced by bacterial endosymbionts in arthropods. Measured as a reduction in egg hatchability resulting from the crossing of uninfected females with bacteria-infected males, CI increases the frequency of bacteria-infected hosts by restricting the fertilization opportunities of uninfected hosts in populations [23]. Markedly reduced severity of dengue virus infection has been found in Aedes albopictus infected with Wolbachia [21, 24]. The other one is the introduction of genetic-based strategies, which has the goal to eliminate or reduce mosquito densities below transmission threshold through population suppression or to establish mosquito populations that are refractory to the pathogen through population replacement and/or modification [25]. Genetically modified Aedes aegypti mosquitoes that activate the conserved antiviral JAK/STAT pathway in the fat body tissue have been developed, and the modified population inhibits infection with several dengue virus (DENV) serotypes [26], but its use encounters regulatory barriers and public opposition in some countries. Few drugs have been tested to inhibit the virus transmission inside mosquito, although some drugs against dengue virus effectively in vitro have been reported, such as quercetin [27], ivermectin [28–30], dasatinib [31], pyran naphthoquinones [32], mycophenolic acid [33, 34], castanospermine [34], deoxynojirimycin [35, 36]. From these drugs, we choose ivermectin as an available compound for the investigation by considering following three facts: (i) ivermectin has been used for about 30 years for treatment of parasitic infections in human since 1988 [37], and ivermectin mass drug administration (MDA) to humans has been suggested as a possible vector control method to reduce Plasmodium transmission [38–40]; (ii) ivermectin has the ability to target exophagic and exophilic vectors [40, 41] with a different mode of action [42, 43] from the currently used insecticides [44], and then avoid known mosquito behavioral and physiological resistance mechanisms [45]; (iii) ivermectin is an inhibitor for the development of dengue virus in cells [28–30]. The purpose of this investigation is to further determine ivermectin efficacy against dengue virus type-2 (DENV-2) in Aedes albopictus, and explore its potential application as an innovative vector control tool.
The study was approved by the ethical review committee of National Institute of Parasitic Diseases, Chinese Center for Disease Control and Prevention, and approval document number was 20160627. Moreover, no specific permits were required for the described field studies. The studies did not involve endangered or protected species.
C6/36 mosquito cell and BHK-21 cell lines, derived from Aedes albopictus and Baby Hamster Syrian Kidney respectively, were used in this study. The cell lines were maintained and propagated in Dulbecco’s Modified Eagle Medium (DMEM) (Gibco by Life technologies, Australia) containing 10% (v/v) fetal bovine serum (FBS) (Gibco by Life technologies, Australia) and 1% (v/v) Penicillin-Streptomycin (Gibco by Life technologies, Australia). Cultured C6/36 was incubated at 28°C in 5% CO2 humidified chamber, and was passaged every 2~3 days. At the time of virus multiplication, the serum concentration was reduced to 2% and temperature was increased to 33°C. DENV-2 was propagated using C6/36 cell line and harvested after CPE presentation on day five post-infection. Supernatants containing DENV-2 were collected, centrifuged at 4,000 xg for 10 minutes to clear cellular debris, and then were stored at -80°C until further use. The titer of viral stocks was measured by TCID (50) % using serial dilutions of 101 to 106 of the viral stocks inoculated into BHK-21 cells. The viral titer was calculated according with Reed and Munch [46]. Cell lines and virus were kindly provided by Shenzhen Center for Disease Control and Prevention (Shenzhen, China).
Adult mosquitoes of Aedes albopictus were obtained from the National Institute of Parasitic Diseases at Chinese Center for Disease Control and Prevention based in Shanghai and were raised at 26±2°C, 60~80% relative humidity, and a 12:12 light: dark cycle. The larvae were raised on a diet of rat food. Adults were provided with 10% (g/v) sucrose solution. Adult mosquitoes aged between three and five days post emergence from larvae were used as experiment objects.
The powdered ivermectin formulation was obtained from Sigma-Aldrich (St. Louis, MO). Ivermectin was diluted in dimethyl sulfoxide (DMSO) to 10 mg/ml and aliquots were frozen at −20°C. Frozen aliquots of ivermectin were thawed and serially diluted in phosphate buffered saline (PBS) prior to addition to human whole blood heated to 37°C prior to mixing. 10 μl of varied concentrations of ivermectin in PBS were added to 990 μl of human whole blood meal to reach 0, 2, 4, 8, 16, 32 and 64 ng/ml concentrations offered to mosquitoes.
Aedes albopictus aged between three and five days post emergence from larvae were fed together with human whole blood containing the same titer of DENV-2. After blood feeding, all fully engorged mosquitoes were gently transferred by aspiration to a new 3L cardboard cartons and held in an incubator at 26±2°C, 60~80% relative humidity, and a 12:12 light: dark cycle. Engorged mosquitoes were held for four days with access to human whole blood, and then were randomly divided into eight groups. Seven of them were held for six days with access to 0, 2, 4, 8, 16, 32 and 64 ng/ml ivermectin, respectively, and the last one was set as a historical control group. The mosquitoes in the historical control group were stored at -80°C until being detected at the same time with mosquitoes in the other groups. In this way, one parallel control group (0ng/ml), one historical control group and six treatment groups were set. Three replicates were performed for each group/concentration, with at least 20 mosquitoes per replicate being analyzed. The human whole blood was obtained from Jiangxi International Travel Healthcare Center, which provided healthy physical examination for community.
Mosquitoes treated as described above were collected, and DENV-2 RNA copies in each mosquito were detected by real-time RT-PCR at the same time, and the cycle threshold (CT) value of each mosquito was recorded. At least 60 mosquitoes were analyzed for each group.
After being frozen to death at −20°C, each mosquito was collected in a grinding tube with 350 μl lysis buffer and then was fully grinded by tissue grinded instrument. The DENV-2 RNA was isolated with RNeasy plus Mini Kit (250) (Qiagen, German), and quantitatively tested with the dengue virus 2 real-time fluorescent RT-PCR kit (Shanghai ZJ Bio-Tech, China). The Master Mix volume for each reaction was pipetted as follows: super mix 18 μl, enzyme mix 1 μl, internal control 1 μl, extraction RNA 5 μl. PCR reaction conditions were: one cycle of 45°C for 10 minutes and 95°C for 15 minutes, then 40 cycles of 95°C for 15 seconds and 60°C for 60 seconds, fluorescence measured at 60°C. During the bioassay, the standard curve between CT values and DENV-2 RNA concentrations (copies/ml) was also detected as described previously [47].
The standard curve between CT values and DENV-2 RNA concentrations (copies/ml) was analyzed by linear correlation regression with regression equation and the DENV-2 RNA concentration (copies/ml) in each mosquito were calculated by the CT value according to the regression equation. All the DENV-2 RNA concentration (copies/ml) in each group were presented by the key parameters, including the median, 75th percentile (P75), 25th percentile (P25), maximum (Max), minimum (Min) and inter-quartile range (Q). For the DENV-2 RNA copies, the differences among the eight groups were analyzed by Kruskal-Wallis test (K-W test), and then were further analyzed by the Turkey studentized range test to determine exactly which two groups had significant difference. According to the detection reagent protocol, when the CT value of mosquito was less than or equal to 40.00, the mosquito was judged to be positive with DENV-2, and the infection rate in each group was calculated. For the infection rates, Chi-squared test (χ2 test) was used to examine the statistical significances among the eight groups, and Duncan multiple range tests were used to determine pair-wise differences, and then linear correlation regression method was used to further analyze the correlation between the infection rates and ivermectin concentrations. P < 0.05 was considered to be significant.
The DENV-2 RNA concentration of positive control sample from the commercial kit was 10,000,000 copies/ml, which was serially diluted to 1,000,000, 100,000, 10,000, 1,000, 100 copies/ml. They were synchronously detected with mosquito samples. Three replicates were performed for each concentration. The relationship between CT values (X1) and log10 values of DENV-2 RNA concentrations (Y1) was expressed by the regression equation, which was obtained from the experimental data as Y1 = 12.70–0.28*X1 with R2 = 0.99. Thus, we got the concentration of DENV-2 RNA in each mosquito by the standard curve.
The infection rate in the mosquitoes fed with 0 ng/ml ivermectin (parallel control group) was 84.62%, which was not significantly higher than the infection rate (81.67%) in the historical control group (Table 1). And the mosquitoes fed with 0 ng/ml ivermectin were of higher DENV-2 RNA concentrations than mosquitoes in historical control group (Table 2), verifying the multiplication of DENV-2 inside Aedes albopictus when they were raised from 4 days to 10 days post infectious blood feeding without ivermectin intervention.
The average of infection rates in the seven groups treated with 0, 2, 4, 8, 16, 32 and 64 ng/ml ivermectin from 4 to 10 days post ingesting infectious blood were 84.62%, 85.29%, 82.54%, 74.24%, 63.33%, 54.29% and 42.62%, respectively, And the average of infection rates in historical control group was 81.67% (Table 1). Compared with the parallel control group or historical control group, infection rates in the mosquitoes fed with 2, 4, 8 ng/ml ivermectin were not significantly lowered; while infection rates in the mosquitoes fed with 16, 32, 64 ng/ml ivermectin were much lower (Table 1), with infection rate being reduced by as much as 49.63% (Fig 1). The regression equation between infection rates (Y2) and log2 values of ivermectin concentration (X2) was obtained as Y2 = 91.41–7.21*X2 with R2 = 0.89. (Table 1, Fig 2). What might confuse us here was that infection rate (85.29%) in mosquitoes fed with 2ng/ml ivermectin was seem to be higher than that in the historical control group (81.67%) or parallel control group (84.62%), but this differences were meaningless for being without statistical significance. In this part of experiment, antivirus effectiveness on DENV-2 in Aedes albopictus was observed in the ivermectin treatment groups at certain concentration, and the more ivermectin mosquito ingested, the lower the infection rate was.
Related parameters indicating the DENV-2 loads in mosquitoes, including Max, median, P75, P25, Min and Q in each group were presented in Table 2. Compared with mosquitoes fed with 0 ng/ml ivermectin, mosquitoes fed with 2, 4, 8 ng/ml ivermectin carried the same level of DENV-2 RNA concentrations (copies/ml), and mosquitoes fed with 16, 32, 64 ng/ml ivermectin exhibited much lower DENV-2 RNA concentrations (copies/ml) (Table 2), with Max, median, P75 and P25 of DENV-2 RNA concentrations (copies/ml) being reduced by up to 85.89%, 99.99%, 99.99% and 84.06%, respectively (Fig 3). On the other hand, compared with mosquitoes in historical control group, DENV-2 had well developed inside mosquitoes fed with 0, 2, 4, or 8 ng/ml ivermectin showing significantly higher DENV-2 RNA concentrations (copies/ml), and was effectively inhibited in mosquitoes fed with 16, 32, or 64 ng/ml ivermectin showing the same level of DENV-2 RNA concentrations (copies/ml). The evidences confirmed the observation of antivirus effectiveness that virus loads in Aedes albopictus were statistically reduced by treatment of ivermectin when concentration of ivermectin was more than 16ng/ml. (Table 2)
In the past decades, dengue fever was a neglected vector-borne tropical disease, with few of control efforts to reduce the burden of the disease at national or international levels [1]. With more outbreaks occurred every year around the world [48–56], people are being faced with the problem of difficulty in blocking the growing trend of dengue transmission [1]. Currently, it has been a consensus that vector control is a key component in the dengue control programs. However, the rapid rise and spread of insecticide resistance have become major threats to the efficiency of insecticide-based vector control activities [1, 3–8]. It is an urgent need to develop innovative control tools for dengue vector control. It was our first try in the laboratory to find out whether ivermectin was able to effectively inhibit the DENV-2 multiplication in Aedes albopictus (Tables 1 and 2, Fig 2). The results give us a hint that using ivermectin in some strategy (e.g. zooprophylaxis [45]) is potentially a new way to stop dengue epidemic through inhibiting DENV-2 in field Aedes mosquitoes.
Interestingly, both of quantitatively and qualitatively inhibiting effects of ivermectin on DENV-2 have been detected in this study. Generally speaking, without ivermectin intervention, DENV-2 was well developed in Aedes albopictus, whose virus loads were significantly higher when the fully engorged mosquitoes were held from 4 to 10 days post infectious blood feeding (Table 2). In contrast, with the treatment of ivermectin, the infection rate and the median of DENV-2 RNA concentrations (copies/ml) were reduced by up to 49.63% and 99.99% (Figs 1 and 3). The linear correlation regression was established between concentration of ivermectin and infection rate of mosquitoes, and we found that 88.5% reduction of infection rate was attributed to the antivirus effectiveness of ivermectin (Fig 2). But the inhibiting effort of ivermectin on the virus in mosquitoes depended on the ivermectin dose, only when the ivermectin concentration was high enough (e.g. over 16ng/ml) can effectively inhibit DENV-2 inside Aedes albopictus. Thus, it is a new need to find out the exactly effective concentration of ivermectin per bite by mosquito as well as action mechanism of ivermectin in the future research, so as to guide its actual application in zooprophylaxis [45].
This study does not attempt to explore the action mechanism of ivermectin towards DENV-2 in Aedes albopictus. In our opinion, several potential reasons are leading to the inhibiting effect on any of the three aspects, namely virus, vector and natural microbiome of mosquitoes. Ivermectin is of a wide range of bioactivity [57]. It has been initially used in livestock or pets to kill parasites (e.g. gastrointestinal and mite) since 1981. Subsequently, it was proved to be very effective in humankind for a variety of internal nematode infections (e.g. Onchocerciasis) [37]. The action mechanism is that ivermectin targets glutamate-gated chloride channels, which plays fundamental roles in nematodes and insects while not accessible in vertebrates, leading to flaccid paralysis [37]. Ivermectin may also interact with γ-aminobutyric acid-gated chloride channels [58]. Both of the two channels are absent in virus. The antiviral activity of ivermectin towards dengue virus had been reported repeatedly since 2012 [29, 30], and then was confirmed in 2016 [28], but all of the researches were carried out in vitro. Considering the existed evidences, the antiviral mechanisms of ivermectin inhibiting DENV-2 in Aedes albopictus can be assumed from the following six aspects: (i) by targeting virus NS3 helicase activity [30]; (ii) by inhibiting nuclear import with respect to virus NS5 polymerase proteins [28]; (iii) by altering some aspects of the mosquito physiology, e.g. reducing the thickness of the peritrophic matrix in Aedes aegypti [59], delaying blood ingesting in Anopheles gambiae [60]; (iv) by stimulating enhanced anti-pathogen innate immunity, e.g. helping the host’s own immune response being able to overcome the immature worms and so kill them [61]; (v) by interacting with glutamate-gated chloride channels or γ-aminobutyric acid-gated chloride channels in the mosquito, and then reducing the adaptability between mosquito and pathogen; (vi) by influencing the natural microbiome of mosquitoes, since the natural microbiome, like Wolbachia, is related with the DENV 2 infection in Aedes mosquitoes [21, 22]. After all, these complex interactions between the pathogen and vector make it possible for ivermectin to have the function of antivirus inside Aedes mosquitoes. All of these potential reasons are worth more deep and overall follow-up study. Moreover, there are still some other effects remain poorly understood. It is unclear that how ivermectin exerts its effect on microfilariae infection in human [62] and P. falciparum in Anopheles gambiae (Giles, 1902) [63].
Ivermectin can block the DENV 2 at any anatomical barrier, like midgut or salivary gland. It was a great pity that we did not study where virus was blocked, so we did not test the viral infection, dissemination and transmission rates, all of which are always different in the same group of mosquitoes. Infection of mosquitoes requires the navigation of several anatomical barriers (e.g. the midgut and salivary glands barriers), and last is excreted into saliva for transmission to a new host. Escape from the midgut or colonization of the hemolymph does not necessarily guarantee the infection of the salivary glands. All of these barriers to productive infection of mosquitoes affect the transmission of viruses. Thus, transmission rate is always lower than viral infection rate. In this study, we just chose viral infection rates as an outcome measure. Maybe transmission rate is a more direct indicator to reflect the significance of ivermectin for the dengue control program in terms of blocking the dengue transmission. Anyway, the results showed that virus infection rates were significantly decreased by ivermectin (Table 1, Figs 1 and 2), which could also largely illustrate the above-mentioned significance of ivermectin. As shown in Table 1 and Fig 1, the virus infection rate in Aedes albopictus mosquitoes fed with 0ng/ml ivermectin was 84.62% (55/65), which was only 42.62% (26/61) in the mosquitoes fed with 64ng/ml ivermectin. The reduction degree of virus infection rate in the treated mosquitoes was up to 49.63%, which meant that there were more negative mosquitoes without virus disseminating from midgut to salivary gland, or that there were less positive mosquitoes with virus transmitting from mosquitoes to a new host. In this sense, we concluded that ivermectin can be used as alternative tool for controlling dengue vectors.
The results of our study may be quite meaningful for the dengue control program in terms of blocking the dengue transmission by using ivermectin. On one hand, the inhibiting effect on dengue virus in vivo means ivermectin which has been proved to be safety in human [64] has the potential to be developed as a drug for curing dengue patients. On the other hand, its antiviral effect inside the dengue vectors may lead to stopping the epidemics of dengue transmission in the field. Moreover, apart from the observed antivirus effect, ivermectin also is of insecticidal action [60, 65, 66]. For an example, about 32.22% (29/90) of mortality was observed in the mosquitoes fed with 64ng/ml ivermectin, which was much higher than 5.79% (4/69) of mortality in the mosquitoes fed with 0ng/ml ivermectin (χ2 = 16.58, df = 1, P<0.0001). Thus, it is an ideal drug for zooprophylaxis and endectocides [45]. Both of the strategies have been used in combating with malaria elimination [15, 16, 67], and resulted in a decrease of malaria incidence and prevalence in Pakistan [67]. The data illuminates that these two strategies may be still suitable for the dengue control program. Because of the antivirus and insecticidal effect, ivermectin using in endectocides can not only kill a number of blood-sucking vectors, but also inhibit the development of the dengue virus in the survived vectors, playing an unexpected role in reversing dengue’s growing trend in the world. However, the exact antivirus effectiveness and eventually being used in blocking dengue transmission need to be further validated with field Aedes albopictus mosquitoes or even other three serotypes dengue virus. Moreover, the actual concentration for application in zooprophylaxis needs to be confirmed in the field trials.
In conclusion, our study shows for the first time that ivermectin can directly or indirectly inhibit DENV-2 multiplication in Aedes albopictus. While the exact antivirus effectiveness and eventually being used in blocking dengue transmission need to be further validated in the field trials with field Aedes albopictus mosquitoes or even other three serotypes dengue virus.
|
10.1371/journal.pgen.1004095 | Loss of Histone H3 Methylation at Lysine 4 Triggers Apoptosis in Saccharomyces cerevisiae | Monoubiquitination of histone H2B lysine 123 regulates methylation of histone H3 lysine 4 (H3K4) and 79 (H3K79) and the lack of H2B ubiquitination in Saccharomyces cerevisiae coincides with metacaspase-dependent apoptosis. Here, we discovered that loss of H3K4 methylation due to depletion of the methyltransferase Set1p (or the two COMPASS subunits Spp1p and Bre2p, respectively) leads to enhanced cell death during chronological aging and increased sensitivity to apoptosis induction. In contrast, loss of H3K79 methylation due to DOT1 disruption only slightly affects yeast survival. SET1 depleted cells accumulate DNA damage and co-disruption of Dot1p, the DNA damage adaptor protein Rad9p, the endonuclease Nuc1p, and the metacaspase Yca1p, respectively, impedes their early death. Furthermore, aged and dying wild-type cells lose H3K4 methylation, whereas depletion of the H3K4 demethylase Jhd2p improves survival, indicating that loss of H3K4 methylation is an important trigger for cell death in S. cerevisiae. Given the evolutionary conservation of H3K4 methylation this likely plays a role in apoptosis regulation in a wide range of organisms.
| Covalent histone modifications alter chromatin structure and DNA accessibility, which is playing important roles in a wide range of DNA-based processes, such as transcription regulation and DNA repair, but also cell division and apoptosis. Apoptosis is the most common form of programmed cell death and plays important roles in the development and cellular homeostasis of all metazoans. Deregulation of apoptosis contributes to the pathogenesis of multiple diseases including autoimmune, neoplastic and neurodegenerative disorders. The budding yeast Saccharomyces cerevisiae has progressively evolved as model to study the mechanisms of apoptotic regulation, and we study here the role of an evolutionary conserved trans-histone crosstalk, in particular histone methylation, in apoptotic signaling in yeast. We have identified a novel trigger for cell death in yeast and due to the strong evolutionary conservation our findings may apply to human cells and may be of importance for understanding the molecular mechanism underlying a specific subtype of acute leukemia.
| Apoptosis is the most common form of programmed cell death and plays important roles in the development and cellular homeostasis of all metazoans. Deregulation of apoptosis contributes to the pathogenesis of multiple diseases including autoimmune, neoplastic and neurodegenerative disorders [1]. The budding yeast Saccharomyces cerevisiae has progressively evolved as model to study the mechanisms of apoptotic regulation, as it had become evident that the extent of evolutionary conservation of the apoptotic core machinery makes it a suitable and attractive model system for apoptotic research. S. cerevisiae undergoes apoptosis when treated with various agents including hydrogen peroxide (H2O2), acetic acid and pheromone (reviewed in [2]). Physiological scenarios that trigger apoptosis in yeast are for example aging and failed mating, and chronological aging is in this respect the to date best-studied scenario [2], [3]. The chronological lifespan (CLS) is defined as the time a yeast cell can survive in a non-dividing, quiescence-like state [4], [5]. Genetic interventions with key yeast apoptotic regulators, such as Bir1p, Nma111p and Yca1p, have been described that influence the CLS of yeast cells and the appearance of the apoptotic features associated to it [6]–[10]. Particularly, disruption of the yeast metacaspase YCA1 gene delays cell death and the formation of an apoptotic phenotype during chronological aging [8].
The activation of apoptosis results in characteristic biochemical and morphological features outside and inside the cell nucleus [11] with chromatin condensation paralleled by DNA fragmentation being one of the most important nuclear events in cells undergoing apoptosis [12]. The mechanism by which chromosomes reorganize during apoptosis is still poorly understood, but evidence exists that histone modifications contribute critically to the nuclear changes experienced by apoptotic cells. Histone modifications that have been linked to apoptosis are phosphorylation of the histone variant H2A.X at serine 139 (S139) that occurs during the formation of DNA double strand breaks under various conditions, including apoptosis [13]. Phosphorylation of histone H2B at S14 has been associated with chromatin condensation and DNA fragmentation [14]–[16]. This modification is reciprocal and deacetylation of H2B at lysine 15 (K15) is necessary to allow H2BS14 phosphorylation [17]. A similar mechanism appears to exist in yeast. Here deacetylation of H2BK11, which is characteristic for exponentially growing yeast [18], is necessary to allow phosphorylation of H2BS10, an apoptotic mark [19], [20]. Therefore, the cis-crosstalk between H2B acetylation and phosphorylation appears evolutionary conserved in apoptosis. Phosphorylation of H2A at serine 129 is increasing in yeast cells undergoing H2O2-induced apoptosis and it is paralleled by a decrease in H3 tyrosine 45 phosphorylation [21], pinpointing to a trans-histone crosstalk related to apoptosis in yeast.
An evolutionary conserved trans-histone crosstalk, which thus far has not been linked to apoptosis, is the regulation of H3K4 and H3K79 methylation by H2BK123 ubiquitination [22]. This trans-histone crosstalk has gathered much attention in recent years, since H3K4 and H3K79 methylation have been implicated in many nuclear processes, such as transcription activation and repression, DNA replication, recombination and repair [22], [23]. The Set1p-containing complex COMPASS acts as H3K4 methyltransferase, and this methyl mark is important for transcriptional activation [24]–[27] as well as silencing at telomeres [27], [28] and rDNA loci [29]–[31]. Methylation of H3K79 is mediated by the histone methyltransferase Dot1p and is essential for efficient silencing near telomeres, rDNA loci, and the yeast mating type loci [28]. Moreover, H3K79 methylation is critical for proper DNA damage response (DDR) [32], [33], as it is prerequisite for Rad9p (53BP1) recruitment [34]. H2B ubiquitination, which is dependent on the ubiquitin conjugase Rad6p and the E3 ligase Bre1p [35]–[37], has been implicated in DNA repair and DDR [33], [38] and we have previously shown that lack of H2B ubiquitination causes metacaspase-dependent apoptosis in S. cerevisiae [39]. H2B ubiquitination is furthermore known to render chromatin resistant to nuclease digestion and its absence is consequently causing increased nuclease sensitivity [22], in keeping with the observed increase in apoptosis.
In this study we analyzed whether apoptosis sensitivity of cells that lack H2B ubiquitination is dependent on a lack of H3K4 and/or H3K79 methylation. We show that Δset1 cells are susceptible to Yca1p-dependent apoptosis, whereas DOT1 disruption affects apoptosis to a lesser extent. We moreover found that Dot1p along with the checkpoint kinase Rad9p is critical for cell death of Δset1 cells. Apoptosis sensitivity of Δset1 cells can be rescued by deleting the yeast homolog of endonuclease G, Nuc1p, suggesting that loss of H3K4 methylation in the presence of H3K79 methylation and the kinase Rad9p enhances chromatin accessibility to endonuclease digestion. Wild-type, but not dot1Δ cells, lose H3K4 methylation during chronological aging coinciding with a shorter lifespan, indicating that the loss of H3K4 methylation is an important trigger for apoptotic cell death.
Histone H3K4 methylation is mediated by the methyltransferase Set1p [27]. To test whether a lack of H3K4 methylation predisposes yeast to apoptotic stimuli, we analyzed the apoptosis sensitivity of Δset1 cells. Chronological aging is to date the best-studied physiological scenario of apoptosis induction in yeast and we therefore studied the effect of SET1 disruption on the chronological lifespan of yeast cells (see Material and Methods). We found that Δset1 cells showed an early onset of cell death during chronological aging when compared to wild-type cells (Figure 1). Almost 100% of Δset1 cells were dead after about 6 days in culture, whereas ∼30% of the wild-type cells were surviving for more than 10 days (Figure 1A). To quantify the difference in life span, we calculated the integral of the survival curve for wild-type and Δset1 cells, which allows to determine the survival differences for the two strains over the time course of the experiment [40]. The survival integral of Δset1 cells (integral 1.2) is significantly smaller than the integral for wild-type cells (integral 4.9) (Figure 1B). Next, we asked whether the death of SET1 disrupted cells is of apoptotic nature. Apoptosis (but also necrosis) is frequently accompanied by an accumulation of reactive oxygen species (ROS), which is an early step in the apoptotic process [41]. Staining with dihydroethidium (DHE) was used to visualize accumulation of ROS. DNA fragmentation was detected by using TUNEL staining, and combined Annexin V/propidium iodide (PI) staining was used to detect the cell surface exposure of phosphatidylserine, an early apoptotic event. PI staining further allows the discrimination between apoptotic (PI negative) and necrotic (PI positive) cell death. ROS accumulation was determined after 2 days in culture, when Δset1 cells showed survival of about 35% compared to ∼75% of wild-type cells (Figure 1A), as determined by clonogenicity. At this time point about 70% of Δset1 cells were DHE positive, but only ∼20% of wild-type cells (Figure 1C and D; Table 1). Consistently, Δset1 cells unlike wild-type cells show apoptotic DNA fragmentation, as determined by TUNEL staining (Figure 1C). Moreover, about 73% of Δset1 cells were stained positive for Annexin V, but negative for PI, compared to about 16% of wild-type cells (Figure 1C and E). Together our data demonstrate that cells lacking set1 have a reduced CLS and their death is predominantly of apoptotic nature.
Apoptosis in yeast can occur in a metacaspase-dependent or metacaspase-independent manner [42]. We thus asked whether the metacaspase Yca1p is required for the death of Δset1 cells. Therefore, Δset1Δyca1 double disruptants were generated and their survival was monitored during chronological aging. As shown in Figure 1A and B, deletion of YCA1 in the SET1-deleted background resulted in significantly better survival (integral 3.0) when compared to Δset1 cells (integral 1.2). The improved survival of the double mutant is accompanied by a significant reduction of ROS accumulation and phosphatidylserine exposure on the cell membrane (Figure 1C–E). Moreover, unlike Δset1 cells, but similar to Δyca1 cells, Δset1Δyca1 cells did not exhibit apoptotic DNA fragmentation as detected by TUNEL labeling (Figure 1C). Thus, the apoptotic death of Δset1 cells is in part Yca1p-dependent.
Histone H2B ubiquitination is not only a prerequisite for H3K4 methylation but also for H3K79 methylation. We next asked whether the lack of Dot1p and H3K79 methylation also influences apoptotic cell death of S. cerevisiae. DOT1 disrupted cells showed better survival during chronological aging when compared to wild-type cells (Figure 2A and B). The effect of dot1 disruption on cell survival is modest, but statistically relevant with survival integrals of 5.8 for Δdot1 cells versus 4.9 for the wild-type (Figure 2B). Consistently, the Δdot1 strain exhibited less ROS accumulation than wild-type cells, less apoptotic DNA fragmentation as detected by TUNEL labeling, and a slight decrease in the exposure of phosphatidylserine on the cell membrane (Figure 2C–E; Table 1). To further underline the statistical relevance of the survival advantage of Δdot1 cells as compared to wild-type cells, we normalized the survival of both strains to survival at day 2 to ensure that all yeast cells have reached stationary phase. Again, we found that the difference in cell survival between wild-type and dot1Δ cells is statistically relevant (Figure S2A and B). Together our data suggest that Dot1p in opposite to Set1p may protect against cell death.
Next, we asked whether Dot1p promoted cell death depends on Yca1p and generated a Δdot1Δyca1 double mutant to analyze its viability. A better survival of the double mutant as compared to the single mutant cells is expected, if Dot1p and Yca1p act independently. However, Δdot1Δyca1, Δdot1 and Δyca1 cells exhibited similar viability during chronological aging (Figure 2A and B) with similar survival integrals as wild-type cells (Figure 2B) and similar ROS accumulation (Figure 2D; Table 1). Together these data suggest that Dot1p and Yca1p act within the same apoptotic pathway as pro-apoptotic proteins.
The above experiments show that Dot1p-mediated H3K79 methylation supports cell death, while Set1p-mediated H3K4 methylation confers apoptosis resistance. Next, we asked if the death of cells lacking H3K4 methylation is dependent on H3K79 methylation and generated a Δset1Δdot1 double mutant, lacking histone H3K4 and K79 methylation. As shown in Figure 3A and B, we observed a significantly improved survival during chronological aging for the Δset1Δdot1 double mutant (integral 4.6) as compared to Δset1 cells (integral 2.7), with consistent ROS accumulation of only 47% for Δset1Δdot1 cells compared to 70% of Δset1 cells (Figure 3C and D; Table 1). These data confirm the pro-death role of Dot1p and suggest that H3K79 methylation is important for cell death of Δset1 cells.
Next, we asked whether or not Dot1p is required for Yca1p-dependent cell death of Δset1 cells. We therefore generated a triple mutant Δset1Δdot1Δyca1 strain and analyzed its survival during chronological aging. If Dot1p acts in an Yca1p-independent manner, a better survival of the triple mutant as compared to Δset1Δdot1 cells is expected. This, however, was not the case. The triple mutant Δset1Δdot1Δyca1 showed no better survival as compared to Δset1Δdot1 cells and a similar number of DHE-positive cells (Figure 3A–D; Table 1). Thus, Dot1p and Yca1p act together in Δset1 provoked cell death.
Dot1p and H3K79 methylation has been shown to confer yeast cells with resistance to DNA damaging agents and the loss of such histone modification causes defective DDR by impairing the function of Rad9p [32], [33]. Rad9p is an adaptor protein required for Rad53p activation [43], [44]. Interestingly, deletion of the RAD9 gene can partially suppress lethal effects of the apoptotic orc2-1 mutation in the origin recognition complex [45], suggesting that Rad9p-dependent checkpoint function is required for apoptosis induction in orc2-1 cells. Given that Dot1p is required for Rad9p-dependent checkpoint activation, Δdot1 cells might fail to activate apoptosis as a result of a defective checkpoint function. To test this hypothesis, we analyzed the survival of Δset1Δrad9 cells during chronological aging and found that the disruption of RAD9 in Δset1 cells significantly improved viability (Figure 4A and B). Consistently, DHE-detectable ROS accumulation was reduced in Δset1Δrad9 cells compared to Δset1 cells (Figure 4C; Table 1). Deletion of RAD9 in a wild-type background does not affect the survival of yeast cells and ROS production, respectively (Figure 4A–C). To test, if set1 depleted cells in fact accumulate DNA damage in a Rad9p-dependent manner, we assayed genome stability of these cells by measuring the mutation frequency in the CAN1 gene. Mutations in the CAN1 gene can be monitored by increased resistance of yeast cells to the toxic amino-acid analogue canavanine and has previously been linked to shorter CLS [46]. We found that SET1 deletion rapidly induced an increase in mutation frequency, which was further increased by a combined deletion of SET1 and RAD9 (Figure 4D), and maintained, but not further increased over time (Figure 4D). Interestingly, RAD9 deletion alone does not coincide with an increased mutation frequency. Together these data indicate that Dot1p is required for apoptosis of Δset1 cells in a Rad9p-dependent manner.
To confirm that Dot1p and Rad9p in fact act in the same apoptotic pathway, we tested the apoptosis sensitivity of a Δset1Δrad9Δdot1 triple mutant. Compared to the Δset1Δdot1 and Δset1Δrad9 double mutants, respectively, a decrease in cell death for the triple mutant is expected in the case DOT1 and RAD9 disruption confer apoptosis resistance independent of each other. The Δset1Δrad9Δdot1 triple mutant, however, exhibited similar survival curves, integrals and ROS accumulation as Δset1Δdot1 and Δset1Δrad9 cells (Figure 4A–C; Table 1). Therefore Dot1p and Rad9p act as pro-apoptotic proteins within the same pathway and the DDR machinery appears to be required for the activation of cell death of aged Δset1 cells.
Nuc1p (EndoG) is a mitochondrial nuclease that translocates into the nucleus upon apoptosis induction coinciding with DNA fragmentation [47]. As loss of H2B ubiquitination is accompanied by an increased sensitivity to nuclease digestion [22], we next asked whether the reduced viability and accelerated apoptosis of cells lacking H3K4 methylation is dependent on nuclease activity. As shown in Figure 5, deletion of NUC1 in Δset1 cells significantly enhanced survival and consequently the survival integrals (Figure 5 A and B; Table 1). Furthermore, deletion of NUC1 in Δset1 cells diminished ROS production during aging (Figure 5C and D; Table 1). Consistent with previously published data [47], aged Δnuc1 cells showed increased cell death compared to wild-type (Figure 5A and B), which likely is of non-apoptotic nature [47], but accompanied by increased ROS production (Figure 5C). In contrast to nuc1 disruption, deletion of the apoptosis-inducing factor AIF1, which also exhibits nuclease activity [7], does not rescue the apoptotic phenotype of Δset1 cells (Figure S1).
The reduced viability of Δset1 cells during chronological aging suggests that loss of H3K4 methylation accompanies cell death of aged wild-type yeast cells. To test this possibility, we carried out Western analysis to monitor H3K4 methylation in wild-type and Δdot1 cells during chronological aging (Figure 6A). A loss of Set1p-mediated H3K4 tri- and dimethylation was observed in aging wild-type cells after 6 days in culture, but not in Δdot1 cells. Dot1p-mediated H3K79 trimethylation remained unaltered in aged wild-type cells (Figure 6A). Quantification of the H3K4me3 and the H3K79me3 levels in wild-type cells normalized to phosphoglycerolkinase (PGK) revealed an about 5-fold reduction of H3K4me3 from day 1 to day 6 and 10, while H3K79me3 levels remain equal (Figure 6B). Similar to aged Δdot1 cells, H3K4 and H3K79 methylation remained unaffected in lymphocytes derived from Hutchison-Gilford progeria syndrome (HGPS) patients (Figure 6C). HGPS is a human premature aging disease, predominantly due to mutations in the gene encoding the intermediate filament protein lamin A, and these cells are thought to be in a senescence-like state [48]. Compared to unaffected control cells, cells derived from differently aged HGPS patients (5 years old, 9 years and 13 years) show a similar increase in H3K4me3 and H3K4me2 levels. These data indicate that loss of H3K4 methylation is not a general aging-related event, but rather specific for aged yeast cells undergoing apoptosis.
If loss of H3K4 methylation in fact acts as trigger for apoptosis, one would predict that abolishing demethylation of H3K4 protects aging yeast cells from cell death. Demethylation of H3K4 is mediated by the trimethyl demethylase Jhd2p [49], [50]. To test our hypothesis, we analyzed the viability of JHD2 deletion cells during chronological aging and found that Δjhd2 cells showed a slightly better survival as compared to wild-type cells with larger survival integrals and reduced ROS production (Figure 6D–G; Table 1). Moreover, the differences in survival of wild-type and Δjhd2 cells are statistically relevant when survival is normalized to day 2, when all cells have reached the postmitotic stage (Figure S2C and D). A double disruption of JHD2 and DOT1 has no additive effect on improvement of cell survival (data not shown), indicating that Jhd2p and Dot1p act in the same pathway. Consistent with the hypothesis that increased or stable H3K4me3 levels are advantageous for survival, H3K4me3 levels increase in aged Δjhd2 cells from day 1 to day 3 and remain stable until day 10 (Figure 6A). We observed, however, no or undetectable H3K4me2 in jhd2 deleted cells as in Δset1 cells. Similar to wild-type and Δset1 cells, H3K79me3 levels remain unaltered during aging of Δjhd2 cells (Figure 6A). Together, our data strongly support the notion that loss of H3K4 methylation, in particular reduced trimethylation, is the cause for apoptotic death of yeast cells.
To further strengthen the notion that loss of H3K4 trimethylation is causing the reduced viability of Δset1 cells during chronological aging, we next analyzed the viability of yeast cells lacking the two COMPASS subunits Spp1p and Bre2p, respectively. Both, Spp1p and Bre2p were previously shown to be required for proper H3K4 trimethylation [51]. Deletion of either SPP1 or BRE2 led to an early onset of cell death during chronological aging, similar to SET1 deleted cells, (Figure 7A and B; Table 1) and to a boosted production of ROS (Figure 7C and D; Table 1). These data strongly support the notion that loss of H3K4 trimethylation is correlated with an early onset of apoptosis. To furthermore reveal that in fact the lack of H3K4 trimethylation is accounting for the increase in apoptotic cell death, we next tested the consequence of a point mutation in H3K4, which prevents methylation of H3 at this site, on apoptosis. To do so, we analyzed the viability of the yeast strain H3K4A, which expressses a histone H3 variant containing a lysine-to-alanine substitution at lysine 4 [52], during chronological aging. We found that these cells showed an early onset of cell death (Figure 7 E and F), similar to Δset1, Δspp1, and Δbre2 cells. This increase in cell death of H3K4A cells coincided with enhanced ROS production (Figure 7 G and H). In contrast to H3K4A cells, H3K79A cells that lack methylation at lysine 79 showed an improved survival as compared to wild-type cells (Figure 7 E and F; Table 1). As for Δdot1 cells (Figure 2), the effect of the K79A substitution was remote, but statistically relevant (Figure 7F). ROS levels were similar to wild-type cells, albeit a bit increased (Figure 7 F and G). Thus, loss of H3K4 trimethylation directly triggers apoptotic cell death during chronological aging, whereas loss of H3K79 trimethylation moderately improves cell survival.
In order to rule out that the limited survival of Δset1 cells during chronological aging is due to acidification of the medium and/or metabolic effects [53] and to demonstrate the importance of H3K4 trimethlyation in apoptosis regulation in more general, we induced apoptosis in Δset1 cells using low concentration of hydrogen peroxide (H2O2). Whereas about 80% of wild-type and Δyca1 cells were recovered after treatment of cells with 0.6 mM H2O2 for 8 hours, less than 40% of set1 cells survived (Figure 8A). The reduced survival of set1 cells coincided with increased ROS production (Figure 8C and D, Table 1). As during chronological aging, disruption of YCA1 in the Δset1 cells conferred resistance to apoptosis induced by H2O2 (Figure 8A, C and D, Table 1). Similarly, double disruptants of SET1 and DOT1, RAD9, and NUC1, respectively, were less sensitive to H2O2, whereas deletion of AIF1 could not rescue the lethal effect of H2O2 on Δset1 cells (Figure 8B). Our data therefore suggest that loss of H3K4 methylation leads to increased apoptosis and sensitizes cells to apoptotic stimuli.
Covalent histone modifications alter chromatin structure and DNA accessibility, which is playing important roles in a wide range of DNA-based processes, such as transcription regulation and DNA repair, but also cell division and apoptosis. In this context, particular changes in phosphorylation and acetylation of histones have been associated with the apoptotic process [54]. Moreover, H2B ubiquitination is important for nucleosome stability and its loss sensitizes yeast to nucleases [55] and to metacaspase-dependent cell death [39]. Histone H2B ubiquitination is a prerequisite for histone H3 K4 and K79 methylation, and this highly conserved trans-histone crosstalk has gathered much attention in recent years, since H3K4 and H3K79 methylation have been implicated in a variety of nuclear processes, such as transcription regulation, DNA replication, recombination and repair [22], [23], [56]. To further explore our prior study, we asked here whether a lack of H3K4 and/or H3K79 methylation affects apoptotic death of yeast cells and uncover the loss of H3K4 methylation as a novel apoptotic trigger.
Methylation of histone H3 at lysine 4 and 79 is accomplished by the evolutionary conserved methyltransferases Set1p and Dot1p, respectively [57], [58]. Methylation of both lysine residues appears to be associated with yeast cell death as loss of H3K4 methylation due to SET1, SPP1 or BRE2 deletion accelerated apoptosis (Figure 1 and 7), while disruption of DOT1 and loss of H3K79 methylation delayed death of aged yeast cells (Figure 2). Similar results were obtained with the respective histone point mutants (Figure 7 E–H), indicating that H3K4 methylation is an anti-apoptotic mark, whereas H3K79 methylation is a pro-apoptotic mark. The loss of SET1 and H3K4 methylation becomes apoptotic only in the presence of Dot1p and H3K79 methylation and can be suppressed by co-disruption of DOT1 (Figure 3), which is likely linked to the DNA damage checkpoint. H3K79 methylation is important for the recruitment of the checkpoint adaptor protein Rad9p, the S. cerevisiae homolog of 53BP1, at damaged sites and for subsequent Rad53p phosphorylation to allow accurate DNA repair [32]. SET1 disruptants rapidly accumulate mutations (Figure 4D) indicative of genome instability (see also [56]) and accelerated DNA damage, which activates the apoptotic machinery after checkpoint activation and failed repair. In the absence of Dot1p and H3K79 methylation, Rad9p recruitment to damaged sites and Rad53p phosphorylation is impaired, the DNA damage checkpoint is not or insufficiently activated, and consequently apoptosis is not activated irrespective to the state of DNA damage. In keeping with this, the co-disruption of RAD9 in Δset1 cells rescued survival (Figure 4A and B), despite the fact that the Δset1Δrad9 cells exhibited a high mutation frequency (Figure 4D). Besides co-disruption of DOT1 and RAD9, respectively, also co-deletion of YCA1 consequently suppressed the lethality of Δset1 cells, at least in part (Figure 1). Yca1p is known as yeast metacaspase and numerous cell death scenarios depend on it [2], [8]. Yca1p likely acts downstream of the DNA damage checkpoint and insufficient DNA repair leads to its activation (Figure 9).
Another executioner of apoptosis in yeast is the endonuclease Nuc1p. Nuc1p can be activated independent of Yca1p and both proteins/pathways converge at the mitochondria [47]. Nuc1p translocates from mitochondria into the nucleus upon activation to degrade chromatin. Changes in chromatin structure due to loss of H3K4 methylation in the absence of Set1p may render yeast cells more sensitive to nuclease activity and consequently NUC1 disruption in the Δset1 background improved cell viability (Figure 5), in contrast to disruption of AIF1 (Figure S1), which also exhibits nuclease activity [7]. Together our data presented here therefore suggest that loss of Set1p-mediated H3K4 methylation causes changes in chromatin structure and genomic instability, which activates the Rad9p-mediated DNA damage checkpoint in dependency on H3K79 methylation. Accumulation of DNA damage and insufficient repair in turn leads to an apoptotic response of the cells, which is executed by Yca1p (in part) and Nuc1p (Figure 9). Deletion of Dot1p and the loss of H3K79 methylation blocks activation of the DNA damage checkpoint and subsequent apoptosis.
Apoptosis in yeast can be triggered exogenously and endogenously. Known endogenous triggers are, for example, defects in DNA damage response and replication, chromatin condensation, mRNA stability, or N-glycosylation [46], [59]–[62]. Chronological aging of yeast cells is the best-studied physiological scenario associated with apoptosis in S. cerevisiae and the lifespan of aged yeast cells can be prolonged or shortened in many ways [4], [63], [64]. Glucose and nutrients have a strong impact on the CLS of yeast [63], [65], whereas endogenous triggers, however, have remained largely unknown. Our data presented here suggest loss of H3K4 methylation as one such endogenous trigger. Wild-type yeast cells lost H3K4 tri- and dimethylation (Figure 6A) after 6 days of culturing, which coincided with a significant increase in cell death (Figure 1A). In contrast to that, H3K79 methylation is not altering during chronological aging. Preventing demethylation by either deleting DOT1 (Figure 6B) or deleting the trimethyl demethylase Jhd2p (Figure 6C and D) delayed against age-induced cell death, indicating that loss of H3K4 methylation is sufficient to drive yeast cells into apoptosis. Particularly, the loss of H3K4 trimethylation seems to promote apoptosis as Δjhd2 cells have low to no H3K4me2 levels (Figure 6A), similar to Δset1 cells. Loss of H3K4me3 is not only triggering apoptosis, but also sensitizes yeast cells to apoptotic stimuli such as exposure to H2O2 (Figure 8A–D), further underlying the importance of this histone modification in apoptosis regulation. H2B ubiquitination is required for H3K4 and H3K79 methylation and it remains to be seen if changes in H2B ubiquitination are the cause for the suppression of H3K4 demethylation upon disruption of DOT1 or if the recruitment of Jhd2p to H3K4 methylation is hindered in the absence of H3K79 methylation. This will be subject of future investigation.
Given the strong evolutionary conservation of H3K4 and H3K79 methylation by the Set1/COMPASS complex and Dot1, respectively, our findings pinpoint to a contribution of a deregulated apoptotic response to the pathology of acute myeloid leukemia (AML). AML is associated with chromosomal translocations involving the MLL gene, the human homolog of Set1p. MLL-associated leukemia are aggressive, characterized by a frustrating therapy outcome, and are DOT1L-dependent [66]. It will be interesting to see how much our findings described here apply to human cells, especially to hematopoietic cells.
BY4742 (MATα; his3Δ1; leu2Δ0; lys2Δ0; ura3Δ0) and its derivatives Δdot1, Δyca1, Δrad9, Δnuc1, Δjhd2, Δ aif1, Δspp1, and Δbre2 were obtained from Euroscarf. Δset1 was derived from BY4742, Δset1Δdot1 and Δset1Δdot1Δyca1 were derived from Δdot1, Δset1Δyca1 and Δdot1Δyca1 were derived from Δyca1, Δset1Δrad9 and Δset1Δrad9Δdot1 were derived from Δrad9, Δset1Δnuc1 were derived from Δnuc1, and Δset1Δaif1 were derived from Δaif1 strains. All derivate strains were constructed according to [67] PCR-based gene deletion. All strains are listed in Table 2.
Survival plating was conducted on YPAD (1% yeast extract, 2% peptone, and 2% glucose, 40 mg/ml adenine) media supplemented with 2% agar. For experiments testing the chronological lifespan, strains were grown in synthetic complete medium (SC) with 2% glucose [68].
Transformation of yeast cells was performed by the lithium acetate procedure, as described by [69].
Chronological aging experiments, hydrogen peroxide treatment, and apoptotic tests using DHE-staining and TUNEL-staining were performed as described previously [39]. All chronological aging experiments reported were conducted at least three times, with three replicates for each strain. Integrals of the life span curves were calculated by summing the trapezoids created by the viability time points as described previously [40]. For the calculation of integrals 100% survival was set as 1. P values were assigned by calculating the variance of integrals between biological replicates and comparing this to the integrals for wild-type cells using a T-test. Cells were viewed using a Leica TCS SP5 and a Zeiss LSM 710 confocal laser scanning microscope. Images were recorded using the microscope system software and processed using Image J and Adobe Photoshop. For quantification of DHE-staining using flow cytometry (FACS-Aria, BD), in each sample 10.000 cells were evaluated and processed using BD FACSDiva software. For Annexin V staining the Annexin-V-Fluos staining kit (Roche, Basel, Switzerland) was used, following the instructions of the manufacturer.
Spontaneous mutation frequency was determined based on the appearance of mutants able to form colonies on agar plates containing 60 mg l−1 L-canavanine sulfate according to [46]. Mutation rates were calculated per 106 living (colony forming on YPD) cells.
Cell extract were prepared by acid extraction using 10% trichloroacetic acid (TCA) according to [70]. In brief, 1.5 ml culture were pelleted by centrifugation at 4°C and frozen at −20°C. 150 µl TCA buffer (10 mM Tris, pH 8.0, 10% TCA, 25 mM ammonium-acetate, 1 mM EDTA) were added to the frozen pellet on ice. When thawed, half the volume glass beads were added and samples were vortex 5×1 min with 3 min intervals on ice in between. The cell lysates were then transferred into a fresh, pre-cooled microfuge tube on ice and centrifuged for 10 min at 16.000 g at 4°C. The supernatant was discarded, the pellet resuspended in 100 µl resuspension solution (0.1 M Tris, pH 11.0, 3%SDS), and boiled for 5 min. After cooling to room temperature, the samples were spun for 30 sec at 16.000 g to pellet the cell debris and 80 µl were transferred into a fresh microfuge. Protein concentrations were determined using the Bio-Rad DC protein assay (Bio-Rad, Munich, Germany) and 30 µg of proteins per well were loaded onto a 15% gel. After SDS-PAGE, proteins were transferred to a PVDF membrane and membranes were probed with the following rabbit polyclonal antibodies: anti-histone H3K4me3 (1∶1000 dilution; 39915, Active Motif), anti-histone H3K4me2 (1∶1000; 39141, Active Motif), anti-histone H3K79me3 (1∶1000; ab2621, Abcam), anti-histone H3 (1∶500; 9715; Cell Signaling), the mouse monoclonal anti-PGK antibody (1∶10.000; Invitrogen) and the respective alkaline-phosphatase conjugated secondary antibodies (1∶20.000; Sigma-Aldrich). Membranes were developed using the Western Lightning CDP-Star Chemiluminescence Reagent (Tropix) and X-ray films. The films were scanned and processed using Adobe Photoshop. Densitometric quantification was performed from three independent experiments using Image J.
Human lymphocyte cell lines were obtained from Coriell Institute (Coriell Institute, Camden, NJ, USA). Cell were grown in suspension in RPMI 1640 medium supplemented with 15% FBS and 2 mM L-glutamine. Cells were cultured at 37°C/5% CO2. Cells were harvested by centrifugation at 600× g for 5 min. The pelleted cells were washed in PBS, resuspended in lysis buffer containing 50 mM Tris-HCl, pH 7.8, 150 mM NaCl, 1% Nonidet P-40 and protease inhibitor cocktail tablets (Roche, Basel, Switzerland). 30 µg of proteins per well were loaded onto a 15% gel and SDS-PAGE and Western blotting was carried out as described above.
|
10.1371/journal.pcbi.1004068 | Prioritizing Therapeutics for Lung Cancer: An Integrative Meta-analysis of Cancer Gene Signatures and Chemogenomic Data | Repurposing FDA-approved drugs with the aid of gene signatures of disease can accelerate the development of new therapeutics. A major challenge to developing reliable drug predictions is heterogeneity. Different gene signatures of the same disease or drug treatment often show poor overlap across studies, as a consequence of both biological and technical variability, and this can affect the quality and reproducibility of computational drug predictions. Existing algorithms for signature-based drug repurposing use only individual signatures as input. But for many diseases, there are dozens of signatures in the public domain. Methods that exploit all available transcriptional knowledge on a disease should produce improved drug predictions. Here, we adapt an established meta-analysis framework to address the problem of drug repurposing using an ensemble of disease signatures. Our computational pipeline takes as input a collection of disease signatures, and outputs a list of drugs predicted to consistently reverse pathological gene changes. We apply our method to conduct the largest and most systematic repurposing study on lung cancer transcriptomes, using 21 signatures. We show that scaling up transcriptional knowledge significantly increases the reproducibility of top drug hits, from 44% to 78%. We extensively characterize drug hits in silico, demonstrating that they slow growth significantly in nine lung cancer cell lines from the NCI-60 collection, and identify CALM1 and PLA2G4A as promising drug targets for lung cancer. Our meta-analysis pipeline is general, and applicable to any disease context; it can be applied to improve the results of signature-based drug repurposing by leveraging the large number of disease signatures in the public domain.
| Computer algorithms that find new uses for known drugs can accelerate the development of new therapies for many diseases, including cancer. One promising strategy is to identify drugs that, at the transcriptional level, reverse the gene expression signature of a disease. A major difficulty with this strategy is variability: different gene expression signatures of the same disease or drug treatment can show poor overlap across studies. Since existing algorithms analyze one signature at a time, this means that the drug candidates they identify may reverse some signatures of a disease but not others. For many diseases, dozens of signatures from different labs are now available in online databases. Combining knowledge across all signatures should lead to better drug predictions. Here, we design a meta-analysis pipeline that takes in a large set of disease signatures and then identifies drugs that consistently reverse deleterious gene changes. We apply our method to find new drug candidates for lung cancer, using 21 signatures. We show that our meta-analysis pipeline increases the reproducibility of top drug hits, and then extensively characterize new lung cancer drug candidates in silico.
| Lung cancer accounts for the largest number of cancer-related deaths, and the 5-year survival rate (across all stages) is only 16% [1]; there is an urgent need for new therapeutics to help treat it. Over the past two decades, the application of high-throughput technologies has led to the rapid accumulation of comprehensive and diverse public datasets cataloguing genome-wide molecular alterations seen with lung cancer or with drug administration. Integrative computational methods that mine these data are fast, cheap, and can complement traditional methods of drug screening; complementary information in these distinct resources can be leveraged to develop comprehensive in silico screens for novel cancer therapeutics [2].
One such resource, the Connectivity Map (CMap), which is the focus of our analyses, catalogues the transcriptional responses to drug treatment in human cell lines for over a thousand small molecules [3]. CMap has been successfully applied to identify novel therapeutics for a diverse set of indications including various cancers [4,5], and most recently osteoarthritic pain [6] and muscle atrophy [7].
CMap was applied in three earlier studies to identify novel therapeutics for lung cancer. Wang et al. [8] combined two microarray data sets to create a single transcriptional signature of lung adenocarcinoma and screened it against CMap. They tested one of their drug hits (17-AAG) in vitro and found that it inhibited growth in two lung adenocarcinoma cell lines. Ebi et al. [9] constructed a transcriptional signature of survival in patients with lung adenocarcinoma; CMap analysis identified several drugs that might improve outcome. The authors experimentally confirmed the growth inhibitory activity of several drug hits, including rapamycin, LY-294002, prochlorperazine, and resveratrol. Jahchan et al. [10] combined two public datasets on small cell lung cancer into a single signature and screened it against the drug profiles in CMap. In vitro experiments confirmed the inhibitory activity of many of their top hits, and in vivo testing showed promising results for imipramine and promethazine.
Nearly every previous analysis using Connectivity Map data to link drugs to diseases has done so with the CMap online tool (http://broadinstitute.org/cmap/). The CMap tool takes as input a set of up-regulated probe sets and a set of down-regulated probe sets, and returns a list of drugs that reverts or mimics those gene expression changes. However, for most diseases, not one but many—often dozens—of distinct gene signatures are available. For example, the cancer-specific database Oncomine (version 4.4) currently stores mRNA data from 566 different studies [11]. As the CMap tool only deals with one gene signature at a time, the question of how best to take advantage of the information in a large collection of disease signatures remains an important open problem. Since different disease signatures can overlap poorly from study to study [2], combining information across many signatures has the potential to improve the performance of drug repurposing algorithms.
While a few studies have used multiple disease signatures in CMap analysis, e.g., [7,8] (though with one exception [12], they used only two or three signatures), they have all relied on essentially the same strategy of collapsing all disease signatures into a single meta-signature (by e.g., intersecting lists of significant genes from different studies, as in [7]) and querying the CMap data with this signature. Since each of the individual disease signatures was constructed using dozens or even hundreds of microarrays, there is fairly strong evidence for every gene in each signature. In contrast, the drug response data in CMap is noisy: the 1,309 drugs have each been tested only a median of 4 times (4 treatment microarrays). This noise has consequences: previous work has shown that even small changes in the input gene signature can lead to large changes in the list of drugs identified as significant by CMap analysis (with the sscMap program) [13,14].
Here we propose an alternative strategy for connecting a set of disease gene signatures to drugs, CMapBatch. Rather than collapsing all the gene signatures in the set into a single gene signature, we propose to screen each disease signature separately against CMap to produce a set of ranked lists of drug candidates. Next, we apply meta-analysis to identify which drugs are consistently ranked as the best candidates across all disease signatures. Thus, we perform the meta-analysis at a later step: our method combines lists of drugs rather than lists of genes. We show that this strategy returns more stable sets of top drug candidates compared to when individual gene signatures are used.
Next, we applied CMapBatch to lung cancer. We used three steps to identify and prioritize new lung cancer therapeutics. First, we conducted a meta-analysis using CMapBatch to identify drugs that reverse the transcriptional changes seen with lung cancer across 21 gene signatures (see Table 1). We identified 247 CMap drugs that consistently counter the gene changes that occur with lung cancer. Second, we performed in silico validation of drug candidates with the NCI-60 growth inhibition data. This validation supported our method: drug candidates identified by CMapBatch were significantly more likely to slow growth in nine lung cancer cell lines than other CMap drugs. Third, we implemented data integration for drug prioritization. We identified common protein targets of significant drugs, and used chemical structure similarity and drug-target relationships to prioritize candidate therapeutics.
Our CMapBatch meta-analysis pipeline comprises the following steps (Fig. 1): For each individual lung cancer signature (tumour vs. normal comparison), we calculate mean connectivity scores for 1,309 small molecules (as previously described [3]). Connectivity scores range between -1 and 1; a large, negative mean connectivity score indicates that drug treatment reverses many of the gene changes seen with lung cancer. We use the mean connectivity score to construct a ranked list of drugs for each signature. We combine the ranked lists of drugs into a single matrix, and identify drugs that were consistently highly ranked across all signatures using the Rank Product method [15] (see Materials and Methods).
Our analyses are based on 21 previously published gene expression signatures of lung cancer obtained from Oncomine [11] and CDIP, the Cancer Data Integration Portal (http://ophid.utoronto.ca/cdip/). The samples used to derive each signature have diverse histologies, and mRNA levels were measured on various commercial platforms.
Previous work has shown that CMap analysis of different gene signatures for the same disease can return very different lists of drug candidates [14]. This is undesirable, if perhaps unsurprising as gene signatures themselves can be highly variable [2]. Consistent with previous findings, we found that when we retrieved lists of the top 50 drugs for each of the 21 different gene signatures of lung cancer (using the CMap online tool), overlap was poor. The median number of drug candidates present in top 50 drug candidate lists from two different signatures was only 22 (Fig. 2 in blue). Repeating the same test using lung cancer signatures of the same type—10 adenocarcinoma signatures—did not lead to much improvement. For adenocarcinoma, the median number of drugs identified by two signatures was 26 (Fig. 2 in gray), but the difference is not statistically significant. We also tested whether the signatures were heterogeneous by computing, for each signature, the median number of drugs shared with all other signatures. For 19 signatures, the median number of shared drugs between any pair of them was similar, varying from 16–29. But there were two outliers: an adenocarcinoma signature [16] that shares zero drugs with any other signature, and a signature of carcinoid tumours [17] that shares a median of only three drugs with other signatures. Removing these two heterogeneous signatures from the signature set boosts the median number of drugs common to any pair of signatures to 24, but again this difference is not statistically significant.
Next, we sought to determine whether aggregating the information from a large set of signatures with CMapBatch would lead to a more stable list of top drug candidates. For this test, we randomly assigned the 21 lung cancer gene signatures to two groups, one with 10 and the other with 11 signatures. We ran CMapBatch separately on the two disjoint sets of signatures, and compared lists of the top 50 drugs identified for each set. We repeated this test 100 times. We found that CMapBatch consistently identifies the same drugs as combatting lung cancer, even when it is trained on completely different sets of lung cancer signatures. A median of 39 drugs were found to be common to both the lists of top 50 drugs identified from two disjoint sets of signatures (Fig. 2 in green), significantly more than are found with individual gene signatures (Wilcox test P << 0.01). This key finding is not sensitive to choice of threshold; using the top 25 or top 100 drugs as an alternative cut-off, CMapBatch again recovers a significantly higher number of drugs (P << 0.01; S1 Fig).
For the remainder of this paper, we focus on characterizing and prioritizing the full set of significant drugs identified by CMapBatch using all 21 gene signatures of lung cancer.
As an independent validation of our results, we used growth inhibition data from the NCI-60 collection [18] to determine whether the drug candidates we identified are better at slowing growth in lung cancer cell lines. For all our NCI-60 analyses we used the nine lung cancer cell lines in which over 100 Connectivity Map drugs were tested (see Methods). None of these nine cell lines were included in the CMap dataset, so they provide an independent test of the effectiveness of our predicted drugs for lung cancer.
The Tanimoto coefficient quantifies the chemical structure similarity between two molecules [23]; here, we call two molecules structurally similar if this number exceeds 0.8. We found that eleven drugs that reverse the transcriptional changes observed in lung cancer were structurally similar to one or more drugs in TOP (Fig. 4, right; S3 Table). These drugs were not evaluated as part of the NCI-60 project; furthermore, 9 of 11 appear in fewer than 20 Pubmed abstracts concerned with cancer. These are novel candidate anticancer therapeutics identified by our computational screen. Further cell-based screens and experimental characterization would be required to determine whether these structurally similar drugs show true anticancer activity.
We used drug-target data from DrugBank [24] and ChemBank [25] (as provided in MANTRA [26]) to construct a drug-drug interaction network on the set of CMap drugs; two drugs are linked by an edge if they share one or more protein targets (Fig. 5A). In total, 83 of the significant drugs were present in this network (the protein targets of many drugs are still unknown), including 9 TOP drugs. Thirty-eight significant drugs that were not tested in the NCI-60 collection share one or more protein targets with a TOP drug (S4 Table; Fig. 5A, purple and green nodes), indicating they may have a similar mode of action and may inhibit growth in lung cancer cell lines. However, since drug target databases do not systematically evaluate a range of drug concentrations and off-target effects, this evidence should only be considered preliminary.
Seven of these 38 drugs were also found to be structurally similar to TOP drugs (Fig. 5A, green nodes): prochlorperazine, promazine, trifluoperazine, fluspirilene, phenindione, vidarabine, and chlorpromazine. As these drugs are linked to TOP drugs by two separate lines of evidence, they are promising candidates for further experimental validation.
The largest connected component in the drug-target interaction network comprised 72 drugs, which is significantly larger (P << 0.01) than what would be expected by chance; random sets of 83 drugs in the drug-drug network yield largest connected components with a median size of only 42 drugs (Fig. 5B). This indicates that some gene targets are overrepresented among significant drugs; these genes may be valuable drug targets for lung cancer. We applied the hypergeometric test to each gene target of a significant drug and identified ten over-represented targets (P < 0.05; Table 2).
The top over-represented gene is Calmodulin 1 (CALM1), a gene involved in the cell cycle and in signal transduction; it’s a target of 9 CMap drugs, and we found that 8 of these reverse the transcriptional changes seen with lung cancer. Recent research suggests that CBP501, a drug currently in Phase II clinical trials for NSCLC, may sensitize tumors to the chemotherapeutic agents bleomycin and cisplatin by inhibiting CALM1 [27]. Thus, other significant drugs that target CALM1 may also enhance the effect of chemotherapy. The 8 drugs we identified are bepridil, felodipine, flunarizine, fluphenazine, loperamide, phenoxybenzamine, pimozide, and miconazole.
The second-most overrepresented gene is PLA2G4A, whose protein product is a member of the cytosolic phospholipase A2 family. Cytosolic phospholipase A2 (cPLA2) has been previously implicated in cancer progression and metastasis. Furthermore, in a mouse model of lung cancer, the inhibition of cPLA2 activity led to delayed tumour growth [28]. There are 4 drugs targeting PLA2G4A included in the CMap collection, and all 4 significantly reverse lung cancer gene changes in our analyses: flunisolide, fluocinonide, fluorometholone, and medrysone.
We used the CMap gene expression profiles from before and after drug treatment to calculate the number of genes differentially expressed in response to a drug, for each of the 1,309 drugs in the collection (see Materials and Methods). We found that significant drugs affect a median of 8.5 genes, while other CMap drugs affect only a median of 3 (Fig. 6; Wilcox test P << 0.01).
We investigated the top drugs that revert expression changes in different lung cancer subtypes by running CMapBatch on the two largest signature subsets in our collection, adenocarcinoma (10 signatures) and squamous cell carcinoma (6 signatures). We found a very high concordance among top drugs; 79 drugs are common to the top 100 drugs lists for adenocarcinoma and squamous cell carcinoma (Fig. 7). Furthermore, all 79 drugs are significant in the full 21-signature meta-analysis (FDR < 5%). This finding is consistent with previous work showing that a common transcriptional program contributes to the molecular signature of many diverse cancers [29], and that CMap predicts similar sets of drugs for some cancers originating from different tissues [12]. Many of the current FDA approved drugs for lung cancer are also approved for other cancers, e.g. methotrexate, cisplatin, etoposide, etc [30].
We selected the antipsychotic drug pimozide for further in vitro validation. Pimozide is both a member of the set of TOP drugs and an inhibitor of CALM1, the most overrepresented protein amongst all targets of significant drugs. We conducted experiments in four lung cancer cell lines that all overexpress CALM1, A549, H460, HCC4006, and H1437 (Cancer Cell Line Encyclopedia data [31]), to test the growth slowing effects of pimozide. Using the MTT assay, we found that pimozide showed significant anticancer activity in each of the four cell lines (P ≤ 0.05; Fig. 8). This validates our computational prediction that pimozide may help treat lung cancer. Since drug-target databases predict that pimozide inhibits CALM1, we assayed CALM1 expression before and after drug treatment in A549 and H460 cells to determine whether CALM1 inhibition might mediate the anticancer activity, but found no significant difference. We also tested whether pimozide was synergistic with cisplatin using the Chou-Talalay method[32,33] in all four cell lines, but our results were negative. Our experiments confirm that pimozide shows some initial promise as a lung cancer therapeutic, but the mechanism of its anticancer activity is unknown and appears to be CALM1-independent.
For many diseases, including several cancers, dozens of distinct transcriptional signatures are available. We developed CMapBatch to efficiently integrate these data with the Connectivity Map to automate drug repurposing and identify stable lists of candidate therapeutics. We applied it to perform the largest in silico drug screen on lung cancer transcriptomes. In total, we identified 247 candidate therapeutics, and for many of these we were able to obtain additional compelling evidence from high-throughput NCI-60 data and databases of known drug targets.
CMapBatch provides a principled approach to combining drug results across multiple gene signatures of disease. Several simple extensions may be appropriate in different applications. For example, weights could be incorporated so that some studies are weighted more highly than others. Also, instead of a meta-analysis across signatures, each of which incorporates multiple patient samples, CMapBatch could be extended to a meta-analysis across all individual samples. We anticipate that CMapBatch and similar methods that can take advantage of the full set of public data on disease will help speed the discovery and development of new medicines.
Code for all analyses was written in R 2.14.0. We converted gene names to HG-U133A probeset IDs for Connectivity Map analysis using the hgu133a.db (Bioconductor 2.8). The drug-target and mode of action networks were analyzed using igraph (Bioconductor 2.8) and visualized using NAViGaTOR 2.3.2 [34], and drug structures were visualized with PyMOL [35]. We calculated Tanimoto similarity for all pairs of 1,148 CMap drugs for which PubChem IDs were available using the PubChem Chemical Structure Clustering Tool [36].
We have made the CMapBatch meta-analysis workflow available as an R script from http://www.cs.utoronto.ca/~juris/data/cmapbatch.
We restricted our analyses to the NCI-60 GI50 (50% growth inhibition) data and to those lung cancer cell lines where at least 100 Connectivity Map drugs were tested (there were nine of these, all NSCLC: NCI-H23, NCI-H522, A549/ATCC, EKVX, NCI-H226, NCI-H322M, NCI-H460, HOP-62, HOP-92). As different GI50 thresholds were used to denote minimal activity in response to a drug for different concentration ranges, we filtered the data to make results comparable across drugs. We retained only those entries with an LCONC (maximum log10 concentration) of-4 and where the drug concentration was measured in units of molarity.
|
10.1371/journal.pgen.1007618 | The thirsty fly: Ion transport peptide (ITP) is a novel endocrine regulator of water homeostasis in Drosophila | Animals need to continuously adjust their water metabolism to the internal and external conditions. Homeostasis of body fluids thus requires tight regulation of water intake and excretion, and a balance between ingestion of water and solid food. Here, we investigated how these processes are coordinated in Drosophila melanogaster. We identified the first thirst-promoting and anti-diuretic hormone of Drosophila, encoded by the gene Ion transport peptide (ITP). This endocrine regulator belongs to the CHH (crustacean hyperglycemic hormone) family of peptide hormones. Using genetic gain- and loss-of-function experiments, we show that ITP signaling acts analogous to the human vasopressin and renin-angiotensin systems; expression of ITP is elevated by dehydration of the fly, and the peptide increases thirst while repressing excretion, promoting thus conservation of water resources. ITP responds to both osmotic and desiccation stress, and dysregulation of ITP signaling compromises the fly’s ability to cope with these stressors. In addition to the regulation of thirst and excretion, ITP also suppresses food intake. Altogether, our work identifies ITP as an important endocrine regulator of thirst and excretion, which integrates water homeostasis with feeding of Drosophila.
| Maintenance of energy and water balance is necessary for survival of all organisms. Even a mild dehydration triggers thirst, reduces appetite, and decreases diuresis (water excretion), thereby promoting conservation of water resources and survival under arid conditions. Homeostasis is regulated primarily by endocrine systems that utilize neuropeptides and peptide hormones. Whereas hormonal mechanisms that regulate the water balance in humans are relatively well understood, much less is known about these regulations in the fruit fly Drosophila melanogaster. Here, we describe the first thirst-promoting and anti-diuretic hormone of Drosophila, encoded by the gene Ion transport peptide (ITP). We show that ITP increases upon dehydration, and protects the animal from loss of body water by promoting thirst and repressing excretion. ITP also suppresses feeding, and can thus be considered as a master regulator integrating water and energy balance.
| Maintenance of homeostasis is based on ingestion and metabolism of water and nutrients in a manner that reflects the internal needs of the animal, but the precise regulatory mechanisms are incompletely understood [1]. Despite the strong evolutionary conservation of the main pathways underlying energy homeostasis [2–5], there is a considerable diversity in the strategies involved in the maintenance of water balance [6, 7]. In insects, this variability arises mainly from the diversity of their habitats and life history strategies. For example, some blood-sucking insects are able to ingest a blood meal that exceeds their body volume up to twelve-fold; their feeding is hence coupled to massive post-prandial diuresis of the excessive water and ions [8]. However, in most of the non-blood sucking terrestrial insects, water conservation is more important than water secretion [1, 9]. Studies on water balance in insects have historically focused mainly on the hormonal regulation of water excretion. These studies investigated the correlations between the hormone titers and diuresis, and analyzed the effects of injections or in vitro applications of the tested compounds (reviewed e.g. in [8–11]). These works contributed to a better understanding of water regulation at the level of fluid secretion by the Malpighian tubules and water reabsorption in the hindgut (reviewed e.g. in [8–11]). Later, development of genetic tools for Drosophila allowed analysis of diuretic hormones by direct genetic manipulations [12–14]. However, no anti-diuretic hormone has been identified in Drosophila until now.
Drosophila is under laboratory conditions raised on media that provide both nutrients and water, and flies therefore do not regulate food and water intake independently. Nevertheless, insects, including Drosophila, can sense water [15, 16] and exhibit hygrotactic behavior [17, 18]. If given the opportunity, flies differentiate between food and water sources, and are able to seek and drink free water [19, 20], or ingest media rich in water but devoid of nutrients [21]. Recently, a small group of neurons were identified in the Drosophila brain that antagonistically regulate thirst and hunger [22]. These neurons sense osmolarity cell-autonomously with the cation channel Nanchung, and internal nutrients indirectly via Adipokinetic hormone signaling [22]. Although several hormones have been shown to regulate feeding and satiety (reviewed in [23–27]), no endocrine regulator of thirst has been identified in Drosophila so far.
The mechanisms that orchestrate water sensing, water-seeking behavior and conservation of water remain unclear. We hypothesized that these processes are likely coordinated by endocrine signaling. Physiological roles of Drosophila hormones are mostly well characterized (reviewed e.g. in [23]); one of the few exceptions is Ion transport peptide (ITP), which belongs to the family of crustacean hyperglycemic hormones (CHH) [28, 29]. CHHs promote water reuptake and hence, act as an anti-diuretic hormones in crustaceans [30]. The locust homolog of ITP promotes water reabsorption by acting on chloride channels in the hindgut [31, 32]. Drosophila has a single ITP gene that gives rise to an amidated ITP hormone and to two longer forms called ITP-like peptides [28, 29]. The functions of Drosophila ITP have not been investigated so far, except for a study that has shown a role of ITP in modulation of evening activity by the circadian clock circuitry [33]. The findings from the crustacean [34] and locust [31, 32] members of the CHH family suggest that Drosophila ITP might be involved in the regulation of water balance as well. Here, we tested this hypothesis by investigating the effects of gain- and loss-of-function of ITP on key aspects of water homeostasis, such as body water content, desiccation and osmotic stress resistance, food and water intake, and excretion. Our work identified master regulatory roles of ITP in water homeostasis of Drosophila; ITP levels increase under desiccation stress and protect the fly from water loss by increasing thirst, reducing excretion rate, and promoting ingestion of water instead of food. Altogether, our work identifies the first anti-diuretic and drinking-promoting hormone in Drosophila, which also coordinates water balance with feeding behavior.
As the first step towards understanding the potential role of ITP in water homeostasis of Drosophila, we investigated whether expression of this gene reflects changes in the body water. We exposed standard (w1118) flies to a short-term (6 h) desiccation, which was sufficient to reduce body fluids (Fig 1A), and monitored expression of the ITP gene (CG13586) by quantitative PCR. Using a primer pair that covers all 5 known transcripts of the gene, we showed that desiccation stress increases expression of the ITP gene (Fig 1B), suggesting a role of ITP in water homeostasis. We confirmed that the transcriptional increase involves also the RE transcript (Fig 1C), the only transcript that gives rise to ITP (FlyBase FB2017_06), considered to be the only functional peptide produced by the ITP gene [28].
It has been shown that an ITP mutation is embryonically lethal [35], and RNAi driven by the ubiquitous daughterless-GAL4 (da-GAL4) also resulted in considerable developmental lethality (Fig 1D). Therefore, to investigate the role of ITP in water balance, we used the GeneSwitch system [36, 37], which allowed circumventing the developmental lethality of ITP and studying the gain- and loss-of function of ITP specifically during the adult stage. In addition, this system enabled investigation of genetically identical animals, thereby avoiding any confounding effects of genetic backgrounds. The system is switched on by feeding flies the drug RU-486, which in itself does not affect water balance (S1 Fig). The expression pattern of ITP is complex and involves several distinct neuron types in the central nervous system and periphery, but the hormone is supposed to be released into the hemolymph [28, 29]. Therefore, we used the ubiquitous daughterless-GeneSwitch (daGS) [38] driver for both RNAi (ITPi) and over-expression of ITP. The daGS-driven over-expression of ITP resulted in increased (Fig 1E), and RNAi in decreased water content (Fig 1F), demonstrating that ITP has anti-diuretic function. We reproduced this effect also using an independent RNAi line targeting an alternative part of the ITP transcript (S2 Fig), and confirmed that ITP has anti-diuretic activity in female flies as well (S3 Fig). However, despite their higher initial water content, animals with increased ITP levels were more sensitive to desiccation (Figs 1G and S4), with their survival reduced by over 30%. Interestingly, animals with reduced ITP levels had moderately increased sensitivity to desiccation as well (Figs 1H and S4), suggesting that survival under arid conditions depends on a tightly regulated expression of ITP.
Taken together, these experiments revealed that ITP codes for a hormone that is regulated by internal water content and has an anti-diuretic function.
Next, we asked if ITP regulates the response to desiccation, or whether it determines desiccation resistance only by influencing the initial water content prior to the desiccation. The daGS>ITPi manipulations from the experiments described above could not answer these questions, because they resulted in reduced body water already before the onset of desiccation. Thus, we looked for a weaker genetic manipulation of ITP, which would allow testing the desiccation resistance without affecting the initial water content. Using an ITP-specific antibody, we confirmed previous results [28, 29] showing that the gene is expressed in the neurosecretory cells of the brain termed ipc-1 and ipc-2, in the interneurons termed ipc-3 and ipc-4, in the abdominal ganglion cells (iag cells), and in the lateral bipolar dendrite neurons (LBD neurons) of abdominal segments A7/A8 (Figs 2A, 2B and S5). To achieve a weaker genetic manipulation of ITP, we used the Impl2-GAL4 driver, which targets only a subpopulation of the ITP-producing neurons: the neurosecretory neurons in the brain (ipc-1 cells and the ipc-2a cells) and the LBD neurons in the periphery (Figs 2A, 2B and S5). To avoid potential developmental effects, we took advantage of the TARGET switch (temporal and regional gene expression targeting, [39]), by which the temperature sensitive tubGAL80ts allows switching on the RNAi specifically in the adult flies. Although we did not test the RNAi efficiency in a cell-autonomous manner, the Impl2-based TARGET effectively decreased the global ITP mRNA (Fig 2C). Consistently, with targeting only a limited number of ITP-expressing neurons, the effect on the global ITP mRNA was approximately 20% weaker than the effect of the ubiquitous daGS-driven ITPi (Fig 2C). Importantly, the ITPi driven by Impl2-based TARGET was not sufficient to impair body fluids (Figs 2D and S6). Thus, this driver allowed us to disentangle the effect of ITP on water storage before the onset of desiccation from its role during the desiccation exposure. ITPi driven by the Impl2-based TARGET resulted in a reduced survival under desiccation (Figs 2E and S6), suggesting that ITP is required to cope with the desiccation stress via an additional mechanism, not only by regulating water storage prior to desiccation. An effect on desiccation survival, similar to Impl2-driven ITPi, was obtained also by daGS-driven ITPi, when the system was switched by a low RU-486 dose. This low dose (50 μM) was not sufficient to affect the body water content (Fig 2F), but was sufficient to reduce survival upon desiccation stress, although to a lower extent than the standard dose of 200 μM RU-486 (Fig 2G) used in the rest of the GeneSwitch-based experiments.
Thus, ITP regulates desiccation survival not only by accumulating proper levels of body water prior to the desiccation challenge, but it is also required to cope with the arid conditions.
Regulation of water balance is important especially under ionic stress. Therefore, we monitored ITP expression after feeding on a medium containing 4% NaCl, using a primer pair that covers expression of all 5 known transcripts of the gene (Fig 3A), and a primer pair specific for the ITP-RE transcript, the only transcript that gives rise to ITP [28]. Osmotic stress indeed increased expression of ITP-RE (Fig 3B). However, this treatment also reduced the amount of body water (Fig 3C) and hence we cannot differentiate whether the increase in ITP expression was driven by the changes in the osmolarity or the volume of body fluids. Genetic over-expression of ITP decreased survival during osmotic stress (Fig 3D), without affecting the osmolarity-induced changes in the body water (Fig 3E and S1 Table).
Similar to ITP over-expression, ITPi driven by the daGS and the Impl2-GAL4 lines resulted in a weak, but statistically significant reduction of osmotic resistance (Fig 3F and 3H), suggesting that both up-and down-regulations of ITP impair osmotic tolerance. The daGS-driven ITPi reduced water levels to an extent comparable to that seen under osmotic stress (Fig 3G). Subsequent exposure to osmotic stress did not decrease the body water of the daGS>ITPi flies any further (Fig 3G and S2 Table). The weaker Impl2-driven ITPi neither affected water content nor its reduction by osmotic stress (Fig 3I and S3 Table), suggesting that ITP is required to cope with osmotic stress independently of the regulations of water content.
Taken together, we show that survival under osmotic challenge requires tight regulation of ITP expression, as both up- and down-regulation of this gene resulted in a reduced survival on a food medium with a high salt content.
Next, we investigated the functional mechanism by which ITP regulates water balance. Under standard laboratory conditions, Drosophila obtains water from the food. Thus, we first asked whether ITP regulates food consumption. We tested whether ITP manipulations affect frequency of eating, measured as propensity to start spontaneous feeding. We transferred fed flies to fresh food supplemented with blue dye, which allows monitoring the time when animals initiate feeding (Fig 4A). Neither ITP over-expression nor ITP RNAi affected the propensity of flies to start spontaneous feeding (Fig 4B and 4C). Subsequently, we measured the total volume of food consumed (Fig 4D), using a modification of the capillary feeding (CAFE) assay [40, 41]. This assay revealed that ITP is an anorexigenic factor; an increase in ITP reduced the volume of consumed food (Fig 4E), whereas ITP RNAi increased the total food intake (Figs 4F and S7).
These experiments indicate that ITP is a negative regulator of food intake. Thus, increased water levels in the daGS>ITP and reduced levels in the daGS>ITPi animals suggest that ITP acts downstream of feeding to conserve body water.
The ureter of Drosophila feeds into the hindgut, and water that is not re-absorbed by the hindgut epithelium is excreted by the same route as the feces [9]. Thus, we investigated whether the ITP manipulations affect excretion. Since our previous experiments (Fig 4B and 4C) had shown that genetic manipulations of ITP do not affect the propensity to initiate feeding, we monitored the speed of food transit throughout the digestive tract as the time from initiation of feeding until excretion of the blue dye in the feces (Fig 5A). We transferred flies on the food with blue dye, and measured the time-dependent increase in the blue-dyed feces. The ITP gain-of-function reduced the speed of the food transition throughout the digestive tract (Fig 5B and S4 Table), whereas ITP RNAi increased it (Figs 5C and S8 and S5 Table).
Subsequently, we tested whether ITP regulates also the frequency of the defecation events. Thus, we continuously fed flies with the blue-dyed food for two days and observed defecation events under conditions when intake and excretion of the dye were at equilibrium. The frequency of defecation events was decreased by ITP over-expression (Fig 5D), and increased by ITPi (Fig 5E). Hence, deficiency for ITP leads to a phenotype reminiscent of human diarrhea. The size of individual feces was reduced by both manipulations of ITP (Fig 5F and 5G). Nevertheless, we were not able to detect significant differences in the color intensity of feces that might be indicative of differences in the water content (S9 Fig).
Altogether, the above experiments indicate that ITP regulates the rate of excretion. Deficiency in ITP results in a faster transit through the digestive tract and an increased number of defecation events, reminiscent of diarrhea, a common cause of dehydration in humans.
Under standard experimental conditions, flies obtain water from their food, and the classical food intake assays do not distinguish between thirst and hunger. To differentiate the role of ITP in water versus food intake, we modified a recent method by Lau et al. [19]. We reared flies on a medium poor in water (‘dry food’), and provided access to a separate, blue-dyed source of water (Fig 6A). Flies with increased ITP levels started to drink faster than controls (Fig 6B), and vice versa, ITPi resulted in a delayed time to the onset of water intake (Fig 6C). In order to test whether ITP also regulates the total volume of ingested water, we modified the CAFE assay monitors from the food intake experiment (Fig 4D); water was provided in microcapillaries in the presence of food poor in water (Fig 6D). Consistent with their increased propensity to start drinking, ITP over-expressing flies also drank more (Fig 6E), whereas ITPi flies drank less water than controls (Figs 6F and S10).
These experiments revealed that ITP is the first known hormonal regulator of thirst in Drosophila.
In summary, in this study we identified ITP as a neuroendocrine factor central to regulation of water homeostasis. ITP increases in response to hypovolemia, and triggers drinking, while repressing feeding and water excretion, promoting thus conservation of water resources and protection from dehydration (Fig 7).
With the colonization of dry land and evolution of terrestrial life, conservation, rather than elimination of water became the main challenge for the maintenance of water homeostasis [42]. Despite the differences in the organization of the endocrine systems, the main principles of fluid homeostasis are the same in vertebrates and invertebrates; these include thirst, compensation for the feeding-induced increase in osmolarity by water intake, and water re-absorption by the excretory systems [1, 9, 10, 42]. In humans, water homeostasis is regulated primarily by an osmostat located in the hypothalamus [43]. This osmostat increases water levels by triggering thirst, and reduces the water loss by inducing release of the anti-diuretic hormone vasopressin [43]. In addition to the regulation by osmolarity, thirst is also induced by the changes in the blood volume both via vasopressin [44, 45] and the renin-angiotensin system [42, 46]. Even though thirst and water retention are physiologically coupled, their regulation occurs independently [43, 47]. We show here that these regulations are simplified in Drosophila, where the same hormone promotes thirst, reduces appetite, and increases water storage. Thus, ITP acts as a functional analog of both vasopressin and renin-angiotensin. Interestingly, like the vasopressin [44, 45] and renin-angiotensin system [42, 46], also ITP is regulated by body water content.
Over-expression of ITP increases water content by 4.5%, whereas RNAi dehydrates the fly by 3.3%. The physiological consequences of such mild changes of water levels are not known in Drosophila, but for comparison, in human patients, loss of as little as 2% water significantly impairs cognitive abilities [48], and liquid overload and hypervolemia represent harmful conditions as well [49].
Our findings show that knockdown of ITP leads to increased water excretion similar to human disorders caused by defective water re-absorbance in kidney, such as diabetes insipidus [43, 50]. Conversely, ITP over-expression results in increased water retention reminiscent of the human syndrome of inappropriate anti-diuretic hormone secretion (SIADH) [43]. ITP manipulations may thus become useful tools to induce and study pathologies associated with these human disorders in Drosophila.
ITP is the first identified hormone that regulates drinking in Drosophila. Thus, it acts as a functional analog of the renin-angiotensin system of mammals. Similar to the renin-angiotensin system, ITP is most likely activated by hypovolemia. The neural circuits that control drinking and are regulated by ITP, however, remain to be investigated. Neurons that repress drinking in Drosophila have already been identified in the suboesophageal zone [22]. These neurons are regulated cell autonomously by an ion channel that senses osmolarity [22]. ITP-knockdown flies do not have the drive to drink despite their state of dehydration, whereas ITP over-expressing flies drink despite their excessive water content. Thus, unlike the Nanchung-expressing repressors of drinking [22], the ITP-regulated neurons are not regulated by the volume of body water, but rather by ITP itself.
In insects, primary urine is produced by the Malpighian tubules that are functional analogs of mammalian kidneys [9]. Water enters the lumen of these tubules by passive diffusion along the ionic gradient maintained by the vacuolar V-H+-ATPase [9]. The function of the Malpighian tubules is hormonally regulated by diuretic hormones [9], which in Drosophila include products of the genes capa [13, 51], DH31 [52], DH44 and leucokinin [14]. Urine then enters the hindgut, where it mixes with the gut contents. Importantly, considerable parts of the water and ions are subsequently re-absorbed in the ileum and rectum [9, 32, 53]. Here, we show that ITP reduces excretion of water by reducing the defection rate. Thus, it is likely that Drosophila ITP promotes water reabsorption in the hindgut similar to its homologs in the desert locust Schistocerca gregaria [31, 32] or in the European green crab Carcinus maenas [34]. It is noteworthy that ITP-expressing neurons in the abdominal ganglia innervate Drosophila hindgut [29], suggesting that in addition to the hormonal regulation [29], the hindgut may also be regulated by ITP in a paracrine fashion. In crabs and in the red flour beetle Tribolium castaneum¸ CHH- or ITP-producing endocrine cells, respectively, have even been detected in gut epithelia [34, 54]. Thus, whether produced in the neurosecretory cells or in the endocrine cells of the gut, the actions of CHHs and ITPs on the hindgut appear to be evolutionarily conserved.
In mammals, an increase in osmolarity due to food intake results in postprandial thirst, and conversely, dehydration inhibits feeding when water is not available [55] and this is likely also the case in Drosophila. Our findings of the ITP-driven positive regulation of water intake, concomitant with a negative regulation of feeding likely represents another level of regulation of thirst and hunger, acting in parallel to that of the four drink-repressing neurons in the suboesophageal zone [22].
Whereas many terrestrial arthropods frequently experience arid conditions, salt stress is not very common in non-blood feeding terrestrial insects. Nevertheless, desiccation and salt stress resistance have been traditional tests in the studies of Drosophila diuretic hormones. RNAi against diuretic hormones increases desiccation resistance, as shown for capa [13], DH44 [14] and leucokinin [12] genes. However, it remains unclear whether these hormones contribute to the natural response to the desiccation and osmotic stress. For example, desiccation does not change expression of diuretic hormones DH44 and leucokinin [14]. In contrast, ITP seems to be a natural component of the desiccation and osmotic stress responses, since both stressors trigger an increase in ITP expression. The role of ITP in thirst, hunger and excretion suggest that the ITP-regulated changes in behavior and physiology represent natural responses to cope with the reduction of body water. Consistently, knockdown of ITP reduces survival under desiccation and osmotic stress. However, it is unclear why over-expression of ITP reduces resistance to desiccation and osmotic stress. The UAS-GAL4 based manipulations may increase ITP levels far beyond the physiological range, which—although not lethal under standard feeding—might reduce survival under stressful conditions. Given the role of ITP in the ion transport across the hindgut epithelia of locusts [31, 32], it is tempting to speculate that a similar mechanism exists in Drosophila. In such a scenario, the non-physiological doses of ITP might considerably increase osmolarity of hemolymph. This would be toxic when feeding on a food medium with a high salt content, as well as under desiccation conditions (which further increase osmolarity).
Although ITP has been known for a long time [56], its function has remained enigmatic in Drosophila. Our pioneering work on its roles in Drosophila physiology suggests that ITP codes for a master regulator of water balance, which also integrates the water homeostasis with energy metabolism. Thus, our study not only shows that this member of the CHH family has an evolutionarily conserved anti-diuretic role in Drosophila as it has in other arthropods [34], but also reveals novel functions of this peptide family in food and water intake. It remains to be investigated to what extent these roles are conserved in other insect species or even in crustaceans, but the strong evolutionary conservation of the gene structure [30] suggests that this might be the case. It is possible that the fly ITP regulates, in addition to its here-described role in water balance, other processes that are known to be CHH-regulated in crustaceans [34]. For example, the high developmental lethality of ITP RNAi, together with the previously described lethality of ITP mutants [35] imply that Drosophila ITP plays a critical role during development, perhaps analogous to the role of CHHs in crustacean molting [34].
Although identification of the cellular sources of ITP that are responsible for the here-described functions of this hormone was beyond the scope of this manuscript, the expression pattern of the gene already provides some tempting hints. Previous in situ-hybridizations and immunohistochemistry experiments based on a locust anti-ITP antibody showed that Drosophila ITP is expressed in several neuronal types [28, 29]. Here, using an antibody specific to Drosophila ITP, we confirmed that these cells include ipc-1 and ipc-2a neurosecretory neurons in the brain, ipc-3 and ipc-4 interneurons, three pairs of iag cells in the abdominal ganglia, and the LBD neurons in abdominal segments A7 and A8. As described previously [28, 29], although ITP is expressed in several interneurons, the most prominent cells of the brain that express ITP are the neurosecretory protocerebral ipc-1 and the ipc-2a neurons, which send axons towards neurohemal release sites in the corpora cardiaca, corpora allata, and aorta. Our experiments based on the Impl2 driver showed that a proper response to desiccation and osmotic stress requires production of ITP in the ipc-1 neurons, ipc-2a neurons, or LBD neurons, or in their combination. The ITP production in these cells becomes nevertheless critical only under desiccation and osmotic stress. In contrast to the global manipulations, ITPi targeted to these neurons is not sufficient to impair water balance under standard conditions. Thus, water content is regulated either via ITP produced by cells outside of the Impl2 expression pattern, or the ITP-producing neurons are redundant in their ability to produce sufficient ITP to maintain water homeostasis under standard conditions. Altogether, additional cell type-specific manipulations are required to differentiate whether thirst, excretion and food intake are regulated by specific neurons, or whether different ITP-producing neurosecretory cells act redundantly to produce sufficient amount of the hormone to regulate physiology of the fly.
Another key step towards understanding the ITP actions is the identification of the hitherto unknown Drosophila ITP receptor. This will facilitate cell- and tissue-specific manipulations to unravel the neural circuit(s) responsible for the roles of ITP in the control of thirst and hunger, and allow more detailed studies of the peripheral roles of ITP in defecation and water excretion.
Flies were reared under a 12 h light–12 h dark cycle on a standard Drosophila medium consisting of 6 g agar, 50 g yeast, 100 g sugar, 5.43 mL propionic acid, and 1.3 g methyl 4-hydroxybenzoate per 1 L of medium. Adult flies were collected within 24 h after eclosion, flipped on fresh media, and housed in groups of around 50 females + 50 males per vial. Flies for the TARGET experiments developed at 18°C and on the third day after eclosion were transferred to 29°C for the RNAi induction. Flies for the GeneSwitch experiments developed at 25°C on standard medium, and were kept from the third day after adult eclosion on a standard medium supplemented with RU-486 and reared further at 25°C. All GeneSwitch experiments were conducted with 0 and 200 μM RU-486, and experiments described in Fig 2F and 2G were performed also with 50 μM RU-486. After the switch induction, both TARGET and GeneSwitch flies were flipped every second day onto fresh media. If not stated otherwise, male flies were used for experiments 6–7 days after the induction of the transgene expression. Controls for the non-GeneSwitch experiments were generated by crossing the UAS and GAL4 lines to the w1118 strain. Experiments on the desiccation and osmotic stress–induced changes in the ITP expression were performed on the w1118 strain. The list of used fly stocks is available in the S1 File.
Viability was expressed as egg-to-adult survival, i.e. as the percentage of eggs that gave rise to adult flies. Three independent egg collections (each at least 120 eggs) were tested for each genotype. Eggs were counted, allowed to develop at 25°C at 12 h light/12 h dark cycle on standard medium, and eclosed flies were collected and counted.
Water content was expressed as percentage of fresh body weight. Flies were weighed using a Mettler MT5 analytical microbalance (Mettler Toledo). Fresh weight was determined, then flies were desiccated for 2 days at 65°C and weighed again. The amount of water was calculated as the difference between the fresh and the dry weight, and expressed as % of the fresh body weight. At least 5 replicates (each consisting of 5 flies) were tested per treatment / genotype.
Desiccation resistance was estimated as survival of flies in empty vials without any water source. Experiments were done in 3–4 replicates. TARGET-based experiments took place at 29°C, GeneSwitch-based experiments took place at 25°C.
Osmotic stress resistance was determined as survival of flies on food medium containing 4% NaCl. Experiments were done in triplicates. TARGET-based experiments took place at 29°C, GeneSwitch-based experiments took place at 25°C. The food contained the same concentration of RU-486 (200 μM) or ethanol vehicle control as during the pre-feeding period.
The volume of ingested food was measured by a modification of the CAFE assay [40] in a feeder device constructed out of 24-well-plates, similar to the one described before [41]. Capillaries with food (Hirschmann minicaps, 5μ) were exchanged daily. Food intake of at least 15 animals per treatment was measured during 3 days, and corrected for the evaporation rate. The liquid food contained the same concentration of RU-486 (200 μM) or ethanol vehicle control as during the pre-feeding period.
Flies were transferred into a vial with a drop (approximately 0.2 mL) of food medium containing 0.5% Brilliant Blue (Sigma), and the proportion of flies that started feeding (blue dye was observable in their body after inspection under a stereomicroscope) was counted 1 h, 1.5 h and 3 h after the transfer. Flies were separated into the tested groups 1 day before the experiment to avoid potential interference of CO2 anesthesia with the food intake. Each time point was tested in 4 replicates, each consisting of at least 16 flies.
Flies were transferred into vials with a drop (approximately 0.2 mL) of food medium containing 0.5% Brilliant Blue (Sigma) and allowed to feed continuously. The cumulative numbers of feces that contained the blue dye were counted in the vial every hour, until 6 h after the switch to the blue-dyed medium. Feces were counted in three replicates, each vial containing 20 flies. For testing the statistical significance by two-way ANOVA, the number of new feces that were deposited within the given period was used.
Excretion rate was measured as the number of defecation events (number of feces) per fly per hour. Flies were fed for 48 h on standard food (with or without RU-486) with 0.5% Brilliant Blue (Sigma). Flies were subsequently transferred into a new vial with a small drop of colored food, and the number of feces produced per fly per vial was counted. Experiments were performed in three replicates, each consisting of 20 flies.
Flies were fed for 48 h on standard food (with or without RU-486) with 0.5% Brilliant Blue (Sigma). Subsequently, a new transparent plastic lid was put on top of the vials, and feces collected on this lid within 2.5 h and were photographed using Leica WILD M32 stereomicroscope with Leica DFC290 camera. The area and lightness were measured using the T.U.R.D. software [57].
Flies were separated into tested groups 1 day before the experiment to avoid potential effect of CO2 exposure on the water intake. Flies were transferred into vials containing approximately 2 mL of the water-deprived food, and after 30 min into new vials with 2 mL of the water-deprived food and a 0.2 mL of a water-rich agar droplet (0.6% agarose, 0.5% Brilliant Blue) and allowed to eat and drink. Flies that started to drink were identified based on the blue color in their abdomina after inspection under a stereomicroscope. The proportion of flies that started to drink was checked 1 h, 1.5 h and 3 h after transferring flies to the water source. Experiments were done in triplicates, and at each time point, at least 42 flies were tested. Water-deprived food medium contained 75% less water and agar than the standard medium, consisting of: 0.6 g agar, 20 g yeast, 40 g sugar, 0.54 mL propionic acid, 0.13 g methyl 4-hydroxybenzoate and 1 mL of 20 mM RU-486 or ethanol per 100 mL of medium.
The capillary drinking assay was performed in a device similar to the CAFE assay feeder, with the following modifications: the bottom of each chamber contained approximately 0.8 mL of water-deprived food with 200 μM RU-486 or ethanol as a vehicle control. Water-deprived food medium contained 75% less water and agar than the standard medium, as described above. Flies were allowed to drink water from the capillaries. To make the measurements of the ingested water easier, water was colored with 0.05% Brilliant Blue (Sigma). The volume of ingested water was measured over one day, and corrected for the evaporation rate. At least 18 flies were tested for each genetic manipulation.
Adult flies were dissected in ice-cold Drosophila Ca2+ free saline. After removing wings and legs, brain-thoracic/abdominal ganglia complexes were quickly excised from head and thorax. All preparations were fixed overnight in Zamboni's fixative overnight at room temperature, washed and treated as described in detail earlier [29]. The only modifications concerned the use of two different primary and secondary antibodies always at the same time of incubations. Primary antibodies were a polyclonal rabbit anti-DrmITP diluted 1:10,000 [33] and a monoclonal mouse anti-GFP (against Jelly fish GFP; Invitrogen) diluted 1:1,000. Secondary antibodies were goat anti-rabbit Alexa 546 and goat anti-mouse Alexa 488, respectively (Invitrogen), both diluted 1:1,000. Preparations were imaged with a Zeiss LSM 780 confocal microscope by use of 10× or 20× objectives. Confocal images were processed with Zeiss ZEN software, version 8.1 2012, for maximum intensity projections of z-stacks. Brightness and contrast was adjusted using Corel Photopaint X7 during plate-mounting using Corel Draw X7.
RNA was extracted using the Zymo Research QuickRNA MicroPrep kit according to the manufacturer’s instructions. cDNA was synthesized by the QuantiTect Reverse Transcription Kit (Qiagen) using 1 μg of the total RNA. Quantitative real-time PCR was performed using SensiFAST SYBR Hi-ROX Kit (Bioline) and StepOne Real-Time PCR System (Applied Biosystems). Expression levels were normalized to Actin 5C (Act5C). Information on the primers is available in the S1 File.
Measurement variables were analyzed by two-tailed Student’s t-test, one-way or two-way ANOVA. Nominal variables were analyzed by two-tailed Fischer’s exact test. Survival data were analyzed by log-rank test. P values are indicated by asterisk symbols (* P < 0.05, ** P < 0.01, *** P < 0.001). Error bars represent SEM. Data on the measurement variables were analyzed using Excel or PAST [58]: http://palaeo-electronica.org/2001_1/past/issue1_01.htm. Survival data were analyzed using PAST. Data on nominal variables were analyzed by Graphpad QuickCalcs (https://www.graphpad.com/quickcalcs/).
|
10.1371/journal.pcbi.1006813 | Predicting kinase inhibitors using bioactivity matrix derived informer sets | Prediction of compounds that are active against a desired biological target is a common step in drug discovery efforts. Virtual screening methods seek some active-enriched fraction of a library for experimental testing. Where data are too scarce to train supervised learning models for compound prioritization, initial screening must provide the necessary data. Commonly, such an initial library is selected on the basis of chemical diversity by some pseudo-random process (for example, the first few plates of a larger library) or by selecting an entire smaller library. These approaches may not produce a sufficient number or diversity of actives. An alternative approach is to select an informer set of screening compounds on the basis of chemogenomic information from previous testing of compounds against a large number of targets. We compare different ways of using chemogenomic data to choose a small informer set of compounds based on previously measured bioactivity data. We develop this Informer-Based-Ranking (IBR) approach using the Published Kinase Inhibitor Sets (PKIS) as the chemogenomic data to select the informer sets. We test the informer compounds on a target that is not part of the chemogenomic data, then predict the activity of the remaining compounds based on the experimental informer data and the chemogenomic data. Through new chemical screening experiments, we demonstrate the utility of IBR strategies in a prospective test on three kinase targets not included in the PKIS.
| In the early stages of drug discovery efforts, computational models are used to predict activity and prioritize compounds for experimental testing. New targets commonly lack the data necessary to build effective models, and the screening needed to generate that experimental data can be costly. We seek to improve the efficiency of the initial screening phase, and of the process of prioritizing compounds for subsequent screening.
We choose a small informer set of compounds based on publicly available prior screening data on distinct targets. We then collect experimental data on these informer compounds and use that data to predict the activity of other compounds in the set for the target of interest. Computational and statistical tools are needed to identify informer compounds and to prioritize other compounds for subsequent phases of screening. We find that selection of informer compounds on the basis of bioactivity data from previous screening efforts is superior to the traditional approach of selection of a chemically diverse subset of compounds. We demonstrate the success of this approach in retrospective tests on the Published Kinase Inhibitor Sets (PKIS) chemogenomic data and in prospective experimental screens against three additional non-human kinase targets.
| Early-stage drug discovery involves a search for pharmacologically active compounds (hits) that produce a desired response in an assay of protein function or disease-related phenotype. The active compounds serve as starting points for further structural optimization, with the ultimate goal of developing therapeutic agents. Virtual screening (VS) can be an effective strategy for prioritizing compounds that can lower high-throughput screening costs by reducing the experimental search to smaller, active-enriched compound subsets. This process can be cheaper and more effective than exhaustive, unguided testing of entire compound libraries [1]. VS may also allow us to evaluate much larger physical or virtual compound libraries. As on-demand synthetic capabilities expand, a VS-guided approach might obviate costs associated with purchasing and on-site storage/maintenance of large general libraries in favor of growing smaller, project-focused compound sets [2].
The choice of which VS methodology to deploy depends on the types of information available at the start of this effort [3]. Structure-based VS methods (such as docking) require specific, structurally-characterized biomolecular targets, but these target structures might only be approximated by homology models [4], or might not be available at all. Phenotypic endpoints like cell death or tumor shrinkage are not amenable to structure-based approaches because specific target structures and sites of action may not be known. Furthermore, structure-based VS performance varies substantially across targets, where failures are difficult to predict [4, 5]. Ligand-based VS approaches can provide more consistent levels of enrichment and are independent from any target structure, but they depend strongly on the quality and abundance of training data in the form of measured compound activities on the target of interest [6]. Such approaches, especially those using topological features for compound representations (such as graph-based fingerprints), may also suffer from high prediction uncertainty when presented with compounds whose chemotypes/scaffolds are outside the scope of the training set [6, 7]. The key issue, however, is that training data are usually scarce in early stages of the screening process, making it difficult to generate a predictive model.
For some well-studied target classes (for example, kinases or GPCRs), rich chemogenomic data are available in the form of compound activity profiles across many members of a target class. These data can be structured as a targets-by-compounds matrix of functional interactions, which we term the bioactivity matrix. Though sometimes sparse, incomplete, or limited in compound and target coverage, such matrices hold valuable information that can be leveraged to make predictions on new targets or compounds.
Predictions of compound activities are routinely made using machine learning algorithms to relate a selection of chemical features to the previously measured bioactivities of a training set of compounds. In many cases, these features are chemical fingerprints that describe the presence and proximity of chemical substructures in each compound [6, 7]. Alternative compound fingerprints have been developed on the basis of prior chemogenomic data [8–13]. In these cases, the bioactivity profile of a compound across a series of assays is used as a fingerprint, referred to as a “High Throughput Screening FingerPrint” (HTS-FP), based either on continuous bioactivity values or on a binary quantity representing activity/inactivity. HTS-FPs enable a useful expression of compound relationships through distances derived among standardized bioactivity profiles in much the same manner as chemical fingerprints. HTS-FPs have limited extensibility in that the wide array of assays/target responses that confers a rich pharmacological representation cannot be readily generated for new molecules. However, looking beyond compound representations, arrays of standardized bioactivity data, even when incomplete, can help to establish target relationships.
Given a new target with little or no prior structure–activity relationship information, building an effective ligand-based VS model requires training data acquired through preliminary screening. For the virtual screening model to be cost effective, the library subset providing training instances should be as small as possible. However, preliminary unguided screens constrained to only 100s to 1000s of compounds are likely to produce insufficient training data with few active training instances of limited potency and structural diversity.
Motivated by the need for batch selection strategies to enable effective iterative screening efforts, there has been significant recent effort in developing compound prioritization models from minimal data [14–18]. These methods prioritize additional compounds for testing based on an initial increment of screening data, but the selection of the initial subset of compounds to be screened is often random, pseudo-random, or based on chemical diversity. A recent effort by Paricharak et al. [18] uses an active learning process to select an informer set from the most active and least active compounds across a series of PubChem assays. Their work removes specific assay labels from the chemogenomic data to create a balanced data set, and selects compounds on the basis of uncertainty from previous predictive models. However, their optimal informer set is too large to be useful as an initial screening set in most HTS settings.
Our emphasis in this paper is on the selection of the informer set—the initial set of compounds to be assayed. The experimental data for these compounds may then be used to train initial models or to select additional compounds as the initial (0th iteration) set of compounds to be assayed in a multi-phase scheme. We refer to approaches based on informer sets as Informer-Based Ranking (IBR) methods. This is different than the focus of the studies cited above that focus on model-guided or heuristic selection of compounds for multiple phases of screening. Our approaches are analogous to earlier chemometric experimental design approaches like chemical cluster sampling [19], but leverage chemogenomic data instead. The proposed IBR methods each involve two steps; see Fig 1. In the first step, they select an informer set of compounds to evaluate experimentally for bioactivity on the new target. Importantly, this selection is guided by the bioactivity matrix. The second step involves prioritization of the compounds outside the informer set, according to their bioactivity against the new target. This prioritization may make use of both the bioactivity data on other targets as well as the new data obtained on the informer compounds on the new target. We describe algorithms both for the selection of the informer compounds and for the prioritization of other compounds after screening data are obtained for the informer compounds.
We propose three novel IBR strategies: Regression Selection (RS), Coding Selection (CS), and Adaptive Selection (AS). Each strategy consists of (i) an informer set selection method that chooses a small number of compounds to be tested, based only on characteristics of the bioactivity matrix, and (ii) a compound ranking method that leverages returned informer data to predict which of the untested compounds is active against a new target. Underlying all three strategies is the premise that targets may be naturally organized according to patterns in their bioactivity profiles across compounds. This organization leads to a clustering of targets as well as to the identification of informer compounds that are predictive of the cluster identity of a novel target. The strategies leverage advances in optimization and statistical analysis, and they differ in how patterns are recognized and computations are deployed.
We apply the proposed IBR strategies in the context of two public human kinase chemogenomics matrices: PKIS1 [20] and PKIS2 [21]. We demonstrate the strategies prospectively, by prioritizing PKIS1 and PKIS2 compounds for activity against three distinct protein kinase targets of potential therapeutic importance: Mycobacterium tuberculosis PknB [22], Epstein-Barr virus BGLF4, and Toxoplasma gondii ROP18 [23]. We also apply the strategies retrospectively, in a cross-validation study of each chemogenomics matrix, leaving out one target at a time and prioritizing compound activity against the left-out target.
The performance of each new IBR strategy was assessed prospectively by inspection of the successful activity predictions, and retrospectively using common VS metrics, including: Area Under the Receiver-Operator Characteristic Curve (ROCAUC) and enrichment factor (EF). We also assessed each strategy’s ability to retrieve structural diversity among active compounds by computing the fraction of active scaffolds identified in the top of the ranking. For benchmarking purposes, we compared the proposed IBR methods to a set of baseline models that make use of the compound structures, and include the commonly used diverse selection as well as a selection of the most frequently active compounds in the bioactivity matrix.
We describe IBR strategies that require experimental testing of some new target of interest on a small fraction of the compound library—the informer subset—with a view to effectively prioritizing the remaining compounds for subsequent testing for activity with the target. The complete IBR strategy thus has two parts: a scheme to identify the informer subset and a scheme to prioritize the remaining compounds after assay data have been obtained for the informer compounds. Initially, we may have no assay data on the new target, though we typically have some such chemogenomic data on related targets that populate a related sector of chemical space, in some sense. Ideally, a successful IBR strategy might be applied in target-agnostic drug development settings (for example, phenotypic targets or incompletely featurized targets), so we intentionally exclude from each IBR strategy target-specific features, such as protein sequences or structural information.
We described three novel IBR strategies that use statistical patterns in the bioactivity matrix that is available prior to informer-set assay testing. Regression Selection (RS), Coding Selection (CS), and Adaptive Selection (AS) all treat the target space as being partitioned into clusters of targets so that, within each cluster, there is some relevant similarity of the bioactivity profiles of the targets across the space of tested compounds. These three strategies also posit that a small number of compounds (the informer subset) have bioactivity profiles that are predictive of the cluster label appropriate to any target, including the novel target of interest. RS, CS, and AS differ in how they evaluate clusterings and potential informer subsets. For example, RS and AS involve kmeans clustering of targets followed by regularized multinomial regression to learn the relationship between compounds and cluster labels, but they differ in how the regression is regularized and how the informer compounds are identified. In contrast, CS forms a single objective function that simultaneously scores clustering strategies and potential informer compounds.
Computationally simpler baseline IBR strategies are useful to consider, as they may approximate practical experimental design scenarios. Baseline Chemometric strategies (BCs, BCl, and BCw) use chemical features for both informer selection and non-informer ranking. Three different chemometric ranking strategies are used for the non-informer ranking, as denoted by subscripts s, l, and w (described in detail in the Methods). Here, clustering is applied on the compound space using the known chemical structure (fingerprints) of the compounds (not used in RS, CS, or AS) in order to identify informer compounds. Then, prioritization of the non-informers makes use of various ways of ranking the chemical distance between bioactive informers and non-informers. Alternatively, a Baseline Frequent-hitters strategy simply takes as informer compounds those that show the highest rate of activity within the initial target set (BFs, BFl, BFw). Prioritization of non-informers uses chemical distance, as in the chemometric methods. To simplify, we only report baseline results for each of our top chemometric and frequent-hitters baseline strategies (BCw and BFw). Outcomes for the full set of baselines are available in the supplemental information.
Performance of the IBR strategies was evaluated using two virtual screening metrics that reflect successful prioritization of active compounds: ROCAUC and Normalized Enrichment Factor in top 10% of ranking (NEF10)). An additional metric Fraction of Active Scaffolds Retrieved (FASR10) assesses the diversity of the active chemical structures that were prioritized in the top 10% of the ranking. Also, standard classification metrics F1 score and MCC were applied.
We applied the IBR strategies on three novel kinase targets outside of the PKIS1 and PKIS2 target sets. These microbial targets are phylogenetically distant from most of the human protein kinases in the PKIS data sets, with relatively low kinase domain sequence identities to the nearest neighbors in the PKIS1/2 sets in comparison to kinase domain sequences (S1 Fig). For Mycobacterium tuberculosis kinase PknB (UniProt ID: P9WI81), the nearest neighbors were the human serine/threonine kinases MARK2 (16.1% kinase domain sequence identity) in PKIS1 and BRSK1 (16.1%) in PKIS2 (UniProt IDs: Q7KZI7 and Q8TDC3, respectively). For Epstein-Barr virus kinase BGLF4 (UniProt ID: I1YP37), the most similar kinase domain sequences were from human protein tyrosine kinase (PTK2 or FAK2) (13.8%) in PKIS1 and human serine/threonine-protein kinase (LRRK2) (14.2%) in PKIS2 (UniProt IDs: Q14289.2 and Q5S007). For Toxoplasma gondii ROP18 (UniProt ID: Q2PAY2), the most similar kinase domains are NEK7 (UniProt ID: Q8TDX7.1) (20.2%) in PKIS1 and aurora kinase C (AURKC, UniProt ID: Q9UQ89.1) (20.7%) in PKIS2.
To prioritize which PKIS compounds might be active on PknB, BGLF4, or ROP18, each IBR strategy selected 16 informer compounds from PKIS1 and 16 informer compounds from PKIS2. PknB and BGLF4 were obtained and screened in-house while ROP18 data were collected from an external collaborator [23]. The screening data were held separately from the IBR and baseline method operators prior to informer selection. After selecting PKIS1 and PKIS2 informer compounds, screening data only for those compounds were provided to each IBR and baseline method. Informer set selections by each IBR for PKIS1 are shown with their associated experimental bioactivity measurements in S1 Table. The assay results for the informer compounds selected by each of the IBR strategies were used to rank the remaining non-informers in PKIS1 or PKIS2. To evaluate the performance of the different methods, all of the available PKIS1 and PKIS2 compounds were assayed. Experimental active/inactive labels were assigned using μ + 2σ percent inhibition (activity) thresholds in PKIS1: PknB = 13.4%, BGLF4 = 20.2%, and ROP18 = 43.8% and PKIS2: PknB = 8.7%, BGLF4 = 12.5%, and ROP18 = 33.4% based on screening results from the PKIS compound sets.
The RS and CS approaches were the only methods that recovered multiple hits and active scaffolds in their top 10% of ranked compounds for all three kinase targets and both PKIS datasets (Table 1 and S2 Table). RS managed to recover actives for PknB even though it did not include any active compounds in its PKIS1 or PKIS2 informer sets. The RS method was also the best overall for BGLF4 on PKIS2 and tied as the best method for PknB on PKIS1. CS was the best approach for PknB on PKIS2.
AS and the three BF baseline methods (BFw shown in Table 1) struggled for PknB and BGLF4 with the PKIS2 compounds, each identifying only a single hit. However, AS was the best approach for BGLF4 on PKIS1 compounds and performed better on ROP18 with PKIS2 compounds. The three purely chemometric baseline approaches (BC) (BCw shown in Table 1) were the worst overall, in many cases failing to recover any hits. Nevertheless, BCs and BCl were the top methods on ROP18 with PKIS2 compounds (S2 Table). The best methods were the same when evaluated with NEF10 or FASR10, but varied slightly for ROCAUC (S3 Table).
The PknB, BGLF4, and ROP18 results demonstrate that the IBR methods perform reasonably well even in a challenging setting where the new targets have low kinase domain similarity with the targets used to construct the informer set. For a more comprehensive quantitative assessment of the IBR methods, we conducted retrospective leave-one-target-out (LOTO) analysis for each of the m = 224 targets in PKIS1. This involved m = 224 separate applications of all the IBR strategies applied to reduced chemogenomics matrices (m − 1 rows), again using an informer size of 16 compounds. Each time, the bioactivity profile of the left-out target was predicted in the sense that compounds were prioritized for activity against this one left-out target.
Results from PKIS1 LOTO cross validation are summarized in Table 2. With respect to the ROCAUC metric (Fig 2), the purely bioactivity-based RS model provides the best rankings with a median ROCAUC value of 0.92 ± 0.11 (± one standard deviation). RS and AS methods both had better performance than the top chemocentric and frequent-hitter baseline approaches, BCw (0.67 ± 0.22) and BFs (0.83 ± 0.14). The improvements in ROCAUC of RS and AS over BCw (p = 5.5E-31, 2.0E-23) and BFs (p = 1.3E-21, 4.9E-6) were statistically significant. All p-values were obtained from a 2-sided, pairwise Wilcoxon sign-rank test with S̆idák multiple comparison correction for 6 hypotheses (6 baselines). This correction increases the stringency of the statistical threshold applied on each of the 6 individual tests from α = 0.05 to α = 0.0085. The CS method also had statistically better ROCAUC performance than all baseline models except BFl (p = 0.037) and BFs (p = 0.032). A complete set of p-values from a pairwise comparison of the IBRs is available in S5 Table. The hybrid baseline approaches, which use compound bioactivity profiles to select the most broadly active compounds as informers, performed much better than the chemometric approaches that use chemical features for informer selection.
We also compared strategies using enrichment factor (EF) as an alternative VS metric that, like ROCAUC, reflects retrieval of active compounds (Fig 3). The maximal EF value that could be achieved on a target, however, depends on the active fraction in the set. To address the variation in the extent of the class imbalance across kinase targets (active fractions ranging from 0.01-0.12 in PKIS1) (S2 Fig), we apply the normalized EF metric NEF10. The EF cutoff was also extended from a typical 1% threshold out to 10%, due to the small number of compounds considered (n = 366). To simplify comparison with the ROCAUC metric, we scale NEF10 such that a value of 0.5 reflects a random classifier (equivalent to random ranking or no enrichment) and a value of 1.0 represents a perfect classifier, in which the top 10% has been maximally enriched. Over the 224 targets considered in PKIS1, the three bioactivity-based models (RS, CS, and AS) are statistically superior to all of the baseline approaches (all p <0.0085). The AS method had the strongest enrichment for active compounds with a median NEF10 of 0.85 ± 0.13. This was better than the top frequent hitters model, BFl, which had a median NEF10 of 0.74 ± 0.13 (p = 5.7E-14). The enrichment is even better compared to the chemometric models, the best of which is BCw, providing a median NEF10 of 0.60 ± 0.13 (p = 6.7E-28).
Another key characteristic of robust virtual screening performance is the recognition of diverse active compound structures rather than retrieval of only a subset of the active chemotypes. Because of the high rate of failure for hits in follow-up hit-to-lead or optimization efforts, we value methods that can retrieve as many active scaffolds as possible, even at some expense to predictive accuracy reflected by ROCAUC and NEF metrics. Across PKIS1 targets we assessed the diversity among the known active chemotypes prioritized by each model by monitoring the Fraction of Active Scaffolds Retrieved among the top ranking 10% of compounds (FASR10) (Fig 4). The bioactivity-based IBR methods outperform the top hybrid and chemocentric baseline models, according to this metric. The median FASR10 for the AS model 0.75 ± 0.23 exceeded the top hybrid model, BFl (0.52 ± 0.21, p = 2.1E-18), and chemocentric model, BCw (0.29 ± 0.21, p = 3.7E-32).
Although the IBRs were developed for compound ranking and not necessarily as classifiers, classifier metrics F1 score (F1) and Matthew’s Correlation Coefficient (MCC) were also evaluated for the methods across the PKIS1 targets. The scores/ranks returned by each method were converted to binary classifications using a threshold based on the median active fraction for PKIS1 (5.5%). Over the 224 targets considered in PKIS1, RS, CS, and AS are statistically superior to all of the baseline approaches (all p <0.0085). The full set of metrics evaluations on PKIS1 are provided in S4 Table with the corresponding p-values from pairwise comparisons in S5 Table.
To assess the robustness of IBR performance, we stratified the PKIS1 targets into four equi-sized subsets and compared IBR methods on all performance metrics separately on each subset. This stratification was based on target hit rate and was obtained by binning targets after ranking by hit rate. S7, S9 and S11 Figs stratify Figs 2–4, and indicate very little effect on performance of the target hit rate. To examine further, we used linear regression to decompose each target-method performance metric into a target effect and a method effect; S8, S10 and S12 Figs plot estimated method effects and multiplicity-adjusted 95% confidence intervals. AS, CS, and RS are all robust to the target hit rate, having quite similar performance in all strata. By contrast, BFw is relatively sensitive to the target hit rate. Considering that the hit rate of a novel target is unknown prior to testing, marginal features such as in Figs 2–4, reflect relevant operating characteristics of the proposed IBR methods.
All IBR strategies require choosing the number of elements nA to include in the informer set. Larger nA allows more information to be gleaned from intermediate screening data, and therefore improved prioritization of non-informer compounds. Marginal improvements in performance as a function of nA are expected to diminish as nA increases, because of redundancies in the information acquired as more activity data accrues. Larger nA also leads to higher assay costs. The experiments reported above used nA = 16, about 4% of the compounds in the chemogenomics matrix.
To examine the relationship of informer set size to prioritization performance, we applied IBR strategies on a range of informer set sizes. First, we considered AS, our best performing IBR strategy. S4 Fig shows ROCAUC and NEF10 metrics from the LOTO retrospective analysis of PKIS1 for nA varying from 9 to 28. Performance did not vary greatly over this range. We also tested a wider range of informer set sizes (nA = 1 to 48 compounds) on PKIS1 target predictions using LOTO cross validation, and examined ROCAUC and NEF10 using baseline IBR methods BCw (S5 Fig) and BFw (S6 Fig). Over this range, we observe performance degradation with diminishing informer set sizes. These experiments indicate that our preferred value nA = 16 strikes a reasonable balance between size and performance for this particular data set.
We set out to establish effective strategies to prioritize compounds for initial testing in iterative high-throughput screens in a drug discovery setting. Our approach is related to the cold-start problem in collaborative filtering (recommender systems) and involves informer-based ranking (IBR) strategies that identify a small subset of highly informative compounds to test in the initial screening round. Data obtained by testing the informers can be used to prioritize compounds for subsequent screening. As a proof of concept, we focused on kinases so that we could test methods using public kinase chemogenomic data matrices. Among the IBR strategies tested, we found that those leveraging bioactivity data from matrix targets (RS, CS, and AS) provided better initial sampling than baseline strategies that applied chemometric similarity methods (BC) or hybrid approaches (BF). The hybrid approaches used a “frequent hitters” heuristic for informer selection, based on matrix activities, and chemometric similarity for ranking.
We applied our chemogenomic IBR and baseline methods in prospective tests on three microbial kinases: PknB, BGLF4, and ROP18. An initial batch of just 16 informer compounds from each set (roughly 4% of the complete set of compounds) was selected for assays on these new targets. The methods were evaluated with regard to hit prioritization and diversity of active scaffolds prioritized, compared to the results of assays of the PKIS1 and PKIS2 compounds. Results from these prospective tests indicated that IBRs using bioactivity data and hybrid baseline IBRs outperformed baseline IBRs that use purely chemometric data for PknB and BGLF4. The baselines were superior on ROP18 but performed so poorly on PKIS2 compounds for PknB and BGLF4 that they would be risky to apply in practice on a new target.
For a more complete assessment of the IBRs, we performed a retrospective leave-one-target-out validation on the PKIS1 matrix (m = 224 targets by n = 366 compounds) using a batch selection of 16 informer compounds. We observed statistically better hit prioritization and active scaffold retrieval for the purely bioactivity-based IBRs (RS, CS, and AS) than for any of the baseline methods. The successful early hit and active scaffold retrieval in these small kinase datasets suggests that the IBRs could be a valuable approach for prioritizing compounds in larger libraries that cannot be exhaustively screened.
Chemogenomic assay data have been used through inductive transfer or transfer learning approaches to make successful predictions on compound-target interactions in several contexts [24, 25]. Reker et al. [16] and Cichonska et al. [26] placed chemogenomic predictions into 4 classes: (1) filling in missing elements within a relatively complete chemogenomic matrix (bioactivity imputation), (2) predicting interactions for a target on matrix compounds (virtual screening), (3) predicting interactions for a compound on matrix targets (drug re-purposing or off-target effects), and (4) predicting interactions for non-matrix compounds on a non-matrix target (virtual screening). Wasserman et al. [27] showed that simple kernel approaches using nearest proxy targets could be used to rank compounds effectively for a query target (class 2), as long as it was possible to identify proxy targets closely related to the query target. For kinase targets, Cichonska et al. [26] explored a wide range of ligand and target kernels to address class-1 and class-3 problems. For focused target sets (kinases and GPCRs), Janssen et al. [28] recently applied nearest-neighbor approaches to ligand and targets mapped on t-SNE projections to address class-2 and class-3 problems.
The methods we report differ from prior chemogenomic methods for addressing the class-2 problem by involving strategic but limited data acquisition on the query target. Determination of the responses of targets to key informer compounds shifts a relatively difficult class-2 problem into the more tractable class-1 problem of imputation. Unlike chemogenomic kernel-based approaches [26, 27], we did not use target features, focusing instead on target-agnostic strategies for compound ranking that could be used in the future for cell-based or phenotypic assays. Our focus on limited, strategic data acquisition on the target of interest frames the problem in a more practical context akin to compound prioritization in early, low-data stages of an iterative screening effort [14, 18, 29, 30]. Lack of active compound instances can stall implementation of supervised models for compound selections [15]. Our bioactivity-based IBR methods overlap hit expansion methods using chemogenomic data, as applied by groups at Novartis for guiding molecule selection in iterative screening [18, 29]. In agreement with their findings, IBR methods that use compound bioactivity profiles, rather than chemical features, provided broader active scaffold retrieval [29]. Previous implementations of HTSFP, however, define compounds by normalized bioactivity vectors from an independent reference assay set, whereas our IBRs use compound bioactivity profiles derived directly from the available chemogenomic matrix. We tested targets only from the same target class, namely, protein kinases. The IBR-based informer sets could be applied in the same way that Paricharak et al. used their Mechanism-of-Action Box (MoABox) of probe compounds for testing in “iteration zero” of their iterative screening procedure [29].
The IBR strategies described here could enable iterative screens either on orphan members of a target class or on targets on which very few compounds have been tested. Data returned on each screening iteration would then be used as new training instances to refine the model, potentially in an active-learning framework that also considers relevance of training instances for subsequent compound selection. To promote efficiency of an iterative approach, initial compound batches are often limited in size, with compounds are often being selected at random or to achieve chemical diversity. Initial screens chosen in this way are likely to return few active compounds, thus stalling effective implementation of a supervised activity prediction model. The IBR strategy reported here can be deployed for compound prioritization in early rounds of batch selection; the informer set could be tested to obtain preliminary compound rankings in the low-data phase of iterative screens. Due to class imbalance being skewed towards inactive compounds in drug discovery tasks, IBR methods could enable rapid identification of relatively rare but important active instances necessary for training the activity-prediction model until it can score compounds accurately for prioritization.
Moreover, the bioactivity-based IBR methods exhibited diverse active-scaffold recognition properties, yielding positive training instances with greater structural diversity for supervised compound prioritization models. The FASR10 results indicate that bioactivity-based IBR approaches generalize better over different compound structures than chemometric IBRs, so they should exhibit a greater tendency to scaffold hopping [29, 31, 32]. In contrast, all of the baseline IBR methods use Morgan fingerprint-derived distances to active informers, thus confining their perspective to those active regions of chemical space identified with the informer set. Different chemotypes, however, can exhibit strong activity on the same target. Plots of PKIS1 compounds projected into their three major principal components of chemical feature space (Morgan fingerprints) frequently show active regions that are non-adjacent (S3 Fig). While active compounds tend to cluster in specific regions of chemical space, many targets elicit multiple, sometimes distantly separated regions of active chemical space.
There are several potential uses for IBRs in drug discovery. This work demonstrates the possibility of effective prediction of activities for new targets within the same target class (kinases) from an extensive chemogenomics data matrix representing many targets within that class. A future direction of research is to quantify the amount of chemogenomic data needed to enable robust prediction within the same target class. It appears that low-rank structure in the chemogenomic matrix used in the IBR methods helps to enable reliable predictions of a target’s compound preferences. Statistical models that faithfully represent variation and dependence in bioactivity data also could be leveraged to guide the development of alternative IBR strategies beyond RS, CS, and AS.
Of greater interest is the development of a more general informer set from a broader collection of chemogenomic data. To investigate the generalizability of the methods, we plan to apply them to a wider range of novel targets (or held-out targets) using an expanded chemogenomic data set with broader target and compound coverage. We do not know how well IBRs will perform on new targets that are unrelated to those within the matrix. We are encouraged by the prospective predictive performance on query kinases (PknB, BGLF4, and ROP18) that are dissimilar from kinases in the chemogenomic data but note that these targets are still similar structure and chemical function as protein kinases. More comprehensive data matrices tend to be incomplete, with many missing data values, but they should be useful in testing whether these methods are effective in extended pharmacological spaces. The size of the informer set may well have a dramatic impact on overall performance.
It may be possible to use IBR methods for prioritizing non-matrix compounds on a new target (a class-4 problem). Chemogenomic matrices enable pharmacological mapping of a given new target (query) to matrix targets that exhibit similar bioactivity profiles (proxy targets). Associations between query targets and proxy targets can be made on the basis of full-compound bioactivity profile in the matrix, or potentially just informer assay results. Given that certain proxy targets are likely to be more extensively screened (tested with compounds outside the matrix set), it might be possible to use non-matrix screening data on proxy targets to infer activities for additional compounds and thus prioritize them for testing on some query target.
Most of the IBR strategies developed here leverage chemogenomics data matrices for activity predictions on compounds against selected kinase targets. The matrices were derived from two public human kinase chemogenomics data sets PKIS (PKIS1) [20, 33] and PKIS2 [21]. Prior to development and testing of methods, these sets were processed as described below. (Links to our processed PKIS datasets are provided below).
PKIS1. The original PKIS data set (PKIS1) was downloaded from https://www.ebi.ac.uk/chembldb/extra/PKIS/PKIS_screening_data.csv Each row in this data set contains an assay result on a specific compound. Each row lists several identifiers for each compound and the target, assay conditions, and the assay read-out (percent inhibition). For nearly every compound, kinase activity was tested independently at 0.1 μM and 1.0 μM concentrations. For this work, only the inhibition values obtained at 1.0 μM were used, in order to match the PKIS2 concentrations. PKIS1 contains 366 unique compounds with unique SMILES and ChEMBL IDs that were tested on 200 unique parent kinases having unique target ChEMBL IDs. When we include mutants/variants of the parent isoforms, there is a total of 224 targets with unique ChEMBL ASSAY IDs. Our processed PKIS1 data was therefore arranged as a matrix of 224 kinase targets by 366 compounds.
PKIS2 The original PKIS2 was downloaded from https://doi.org/10.1371/journal.pone.0181585.s004. This set comprises 641 unique compound SMILES and 406 target columns. However, only of these 415 compounds were available to us from the original set for testing. We included only these compounds from the PKIS2 data set, so our bioactivity matrix has 406 targets by 415 compounds. PKIS2 activity values represent percent inhibition values observed at inhibitor concentrations of 1 μM.
Mycobacterium tuberculosis PknB. Recombinant bacterial kinase (PknB) and bacterial substrate (GarA) were purified from E. coli following published procedures [22]. The kinase inhibition assay was done using the Kinase Glo(R) kit from Promega similar to published procedures. PknB was added to plated kinase inhibitor libraries (the available compounds from PKIS 1 and 2) and incubated at room temperature for 10 minutes, after which ATP and GarA (protein substrate) were added. The final concentrations were: PknB 0.25 μM, GarA 40 μM, ATP 100 μM, inhibitors 2 μM, DMSO 1% in a final volume of 5 μL. The kinase reaction proceeded at room temperature for 30 minutes and quenched by the addition of 5 μL of Kinase Glo(R) reagent. The plate was allowed to develop for 10 minutes and luminescence was detected on a BMG PheraStar multiplate reader. Luminescence was converted to μmol/minute of ATP consumed using a standard curve of ATP from 100 to 0 μM. A negative control (no inhibitor) was used to determine percent activity. A positive control (GSK690693) was used to ensure a baseline and compare plate-to-plate variation. Data were analyzed using CDD Vault (Collaborative Drug Discovery, Inc.) to determine plate Z′ > 0.5 and report percent inhibition for each compound.
Epstein-Barr virus BGLF4. Viral kinase BGLF4 was provided by the laboratory of Professsor Yongna Xing. BGLF4 was expressed with an N-terminal His8-MBP-dual-tag in insect cells, and purified over Ni2+-NTA resin (Qiagen) and then Maltose resin (Qiagen), followed by ion exchange chromatography (Source 15Q, GE Healthcare) and gel filtration chromatography (Superdex 200, GE Healthcare) to more than 95% homogeneity. The purified BGLF4 was then used for kinase inhibition assays using the C-terminal fragment peptide of retinoblastoma protein (RB) as substrate (Millipore Sigma cta# 12-439). The remaining assay parameters were the same as those applied for PknB except for the following changes. The final concentrations in the reaction medium were: BGLF4 0.004 μg/μL, RB 0.04 μg/μL, ATP 500 μM, inhibitors 3 μM, DMSO 0.3% in a final volume of 5 μL. As a positive control, K252a (5 μM) was used. The reaction proceeded at room temperature for 20 minutes and was then quenched by the addition of 5 μL of Kinase Glo(R) reagent. ADP depletion proceeded for 40 minutes, followed by addition of 10 μL of kinase detection reagent. The reactions were incubated for 1 hour prior to luminescence detection.
Toxoplasma gondii ROP18. Inhibition data for the PKIS compounds on the Toxoplasma gondii kinase ROP18 was provided by the University of North Carolina Structural Genomics Consortium and Professor L. David Sibley at Washington University in St. Louis. Their assay measured phosphorylation of a substrate peptide by purified ROP18 using microfluidic capillary electrophoresis [23].
Metrics. To evaluate model performance, we applied three different virtual screening metrics, ROCAUC, NEF10, and FASR10. Standard classification metrics, F1 score (F1) and Matthew’s Correlation Coefficent (MCC), were applied as well. ROCAUC and NEF10 measure the extent to which a model prioritizes the active compounds in its ranking. ROCAUC is a standard metric in virtual screening [41] and applied generally in machine learning to evaluate classifiers. Enrichment Factor (EF) (Eq 17) is another commonly used metric for assessing virtual screening performance. EF reflects the fold increase in active compounds over that expected from random compound selection, for a subset of a compound library taken from some top ranking portion of a prioritized compound list.
EF 10 i = ∑ j ∈ B z i , j | B | / ∑ j = 1 n z i , j n , (17)
where B is the set of compounds among the top 10% of those ranked by a method applied to target i, and zi,j is as in (1).
However, the number of active compounds for each left-out target i varies from target to target (S1 Fig). We apply a scaling scheme on EF at the top 10% (Eq 18), which enables better comparisons across targets exhibiting significant differences in active:inactive ratios.
NEF
10
i
=
1
+
EF
10
i
-
EFbase
EF10max
i
-
EFbase
2
,
(18)
where EFbase is 1, which corresponds to random guessing; EF10maxi is the maximum theoretical EF10i, which means all actives are ranked at the top and depends on the number of actives for each target. Our NEF metric returns a value between 0.5 and 1, where a NEF10i larger than 0.5 shows better ranking performance than random guessing–similar to ROCAUC. We selected the 10% threshold with consideration of the sizes of our informer (nA = 16) and full compound sets (n = 366 and n = 405). This threshold includes the 16 informers and 21 noninformer compounds in our PKIS1 evaluation.
For the ROCAUC and NEF10 metrics, experimental percent inhibition (activity) data were binarized using a target-specific μ+2σ threshold based on the activity distribution of the PKIS1 compounds for the kinase target. Actives were defined as compounds with greater than twice the standard deviations above the mean, as noted in (1). When applying the metrics, active informer compounds were counted as true positives, whereas inactive informers did not count against the models as false positives. It should be noted that the main purpose of the informer set is to facilitate accurate activity ranking on the non-informers. However, since informer compounds represent the highest priority compounds for testing, we reward models for retrieving active informers but refrain from penalizing models for choosing inactive informers. Some baseline models that rely upon binary compound labels occasionally failed to evaluate the noninformer compounds in cases where no active informers are returned. In such cases, metric scores reflecting random ranking were assigned to the model: ROCAUC and NEF10 of 0.5 and a FASR10 score of 0.0.
FASR10 assesses a model’s capacity to recognize different active chemotypes among the the top 10% of ranked compounds. The metric reflects the fraction of all active scaffolds identified on a given target within the compound set. Again, zi is the Boolean vector of compound binary activity labels on target i for compound set J. Let OJ be the vector of chemical scaffold identifiers for compounds in J. The scaffold identifiers are arbitrary integer scaffold indices assigned to each of the generic Bemis-Murcko scaffold presented in J, as obtained using the MurckoScaffold module in RDKit [39, 42]. Bemis-Murcko scaffolds were made generic by stripping hydrogens, converting all bonds to single, and setting all atom types to aliphatic carbon. The unique active scaffold identifiers are the set of all non-zero values in the Hadamard product vector:
C J = { z i ∘ O J } (19)
If we then let z i 10 and O J 10 be the binary activity labels and scaffold IDs for the top 10% ranked compounds, the subset of unique active scaffolds recognized just among the top 10% of compounds is:
C J 10 = { z i 10 ∘ O J 10 } (20)
The fraction of active scaffolds recognized in the top 10% is:
FASR 10 = | C J 10 | | C J | (21)
Note, active scaffolds were not considered retrieved unless an experimentally observed active member from that chemotype was in the top 10%. Cases arise where only inactive members of an active scaffold were obtained in the top 10% of the compound ranking. In such cases, the FASR10 metric does not count the chemotype as recognized.
Although the IBRs were developed for compound ranking and not necessarily as classifiers, standard classification metrics F1 and MCC were also applied for IBR performance evaluations. The scores/ranks returned by each method were converted to binary classifications using a fixed threshold across targets based on the median active fraction of the chemogenomic data: 5.5% for both PKIS1 and PKIS2. This amounts to assigning an active classification to the top 20 and top 23 scoring compounds in PKIS1 (366) and PKIS2 (415), respectively. As in the other metrics, inactive informers were not counted as false positives and were removed before metric calculations. Active informers, however, were counted as true positives.
Model evaluations. Performance of the models was evaluated in two stages. The first stage follows a retrospective leave-one-target-out (LOTO) evaluation scheme. Each of the 224 kinase targets in the PKIS1 target set is removed and treated as a new target of interest i. The PKIS1 compound activities are hidden for this target. An informer set Ai is selected for this new target, the activities are revealed for the informers, and then the model rank orders the remaining noninformers A i c using the informer data and in some cases data from the other 223 targets. The 9 models were evaluated in this stage using the 3 metrics described above. The second stage is a prospective evaluation of the 9 models as applied on three novel, non-human, kinase targets. In these evaluations, informer sets were generated twice for each model–once on each of the training matrices, PKIS1 and PKIS2. The remaining compounds (noninformers) from the corresponding matrix are then ranked on the two novel kinase targets using data returned for the informer sets and data within the corresponding PKIS1 or PKIS2 training matrix from which the informer set was selected. As in the retrospective PKIS1 LOTO evaluation, each model was assessed using the 3 metrics described above. However, in this prospective test on the new targets, each model was applied twice, using each of the PKIS data matrices, and therefore a total of 6 evaluations were performed on each model. We attempted to build a larger PKIS matrix by merging the PKIS1 and PKIS2 data matrices. The structure of the merged matrix, was however problematic in that the compound sets were nearly disjoint between PKIS1 and PKIS2. The resulting incomplete matrix lacks a structure that enables accurate imputation of the missing activity elements.
PknB and BGLF4 screening data obtained at the UW-Carbone Cancer Center’s Small Molecule Screening Facility, ROP18 data, formatted PKIS1 and PKIS2 datasets, and a Python implementation of the baseline IBR methods, evaluation metrics, and plotting procedures are available here: https://github.com/SpencerEricksen/informers. Matlab code and documentation involving the RS method is available here: https://github.com/leepei/informer. An R package for running CS and AS methods is available here: https://github.com/wiscstatman/esdd/tree/master/informRset.
|
10.1371/journal.pcbi.1003297 | Comparing Algorithms That Reconstruct Cell Lineage Trees Utilizing Information on Microsatellite Mutations | Organism cells proliferate and die to build, maintain, renew and repair it. The cellular history of an organism up to any point in time can be captured by a cell lineage tree in which vertices represent all organism cells, past and present, and directed edges represent progeny relations among them. The root represents the fertilized egg, and the leaves represent extant and dead cells. Somatic mutations accumulated during cell division endow each organism cell with a genomic signature that is unique with a very high probability. Distances between such genomic signatures can be used to reconstruct an organism's cell lineage tree. Cell populations possess unique features that are absent or rare in organism populations (e.g., the presence of stem cells and a small number of generations since the zygote) and do not undergo sexual reproduction, hence the reconstruction of cell lineage trees calls for careful examination and adaptation of the standard tools of population genetics. Our lab developed a method for reconstructing cell lineage trees by examining only mutations in highly variable microsatellite loci (MS, also called short tandem repeats, STR). In this study we use experimental data on somatic mutations in MS of individual cells in human and mice in order to validate and quantify the utility of known lineage tree reconstruction algorithms in this context. We employed extensive measurements of somatic mutations in individual cells which were isolated from healthy and diseased tissues of mice and humans. The validation was done by analyzing the ability to infer known and clear biological scenarios. In general, we found that if the biological scenario is simple, almost all algorithms tested can infer it. Another somewhat surprising conclusion is that the best algorithm among those tested is Neighbor Joining where the distance measure used is normalized absolute distance. We include our full dataset in Tables S1, S2, S3, S4, S5 to enable further analysis of this data by others.
| The history of an organism's cells, from a single cell until any particular moment in time, can be captured by a cell lineage tree. Many fundamental open questions in biology and medicine, such as which cells give rise to metastases, whether oocytes and beta cells renew, and what is the role of stem cells in brain development and maintenance, are in fact questions about the structure and dynamics of that tree. Random mutations that occur during cell division endow each organism cell with an almost unique genomic signature. Distances between signatures capture distances in the cell lineage tree, and can be used to reconstruct that tree. On this basis, our lab developed a method for cell lineage reconstruction utilizing a panel of about 120 microsatellites. In this work, we use a large dataset of microsatellite mutations from many cells that we collected in our lab in the last few years, in order to test the performance of different distance measures and tree reconstruction algorithms. We found that the best method is not the one that gives the most accurate estimates of the mean distance, but rather the one with the lowest variance.
| A multi-cellular organism develops from a single cell – the zygote, through cell division and cell death, and displays an astonishing complexity of trillions of cells of different types, residing in different tissues and expressing different genes. The development of an organism from a single cell until any moment in time can be captured by a mathematical entity called a cell lineage tree [1]–[4].
Uncovering the human or even the mouse cell lineage tree may help to resolve many open fundamental questions in biology and medicine, as illustrated by our earlier work [5]–[9].
In the past few years, our lab developed a method for reconstructing the lineage relations among cells of multi-cellular organisms 1,10 and applied it to various questions of biological and medical importance [5]–[9]. The method is based on the fact that cells accumulate mutations during mitosis in a way that, with a high probability, endow each cell with a unique genomic signature, and distances between genomic signatures of different cells can be used, in principle, to reconstruct the organism's cell lineage tree [1]. Instead of examining the whole genome of all cells of an organism, which is currently not feasible, our method uses Microsatellite (MS) loci which are repeated DNA sequences of 1–6 base pairs. Slippage mutations, in which repeated units are inserted or deleted, occur at relatively high rates (10−5 per locus per cell division in both wild type mice and humans [1], [11]), and thus provide high variation. These mutations are phenotypically neutral [11]–[13] and they are highly abundant in the genome (composing 3% of the genome). Importantly, Mismatch-Repair (MMR) deficient mice display an even higher mutation rate (10−2 per locus per cell division [14]) in MS and are available for experimentation and analysis [5]–[8], [10], [15], [16]. By comparison, SNPs have a mutation rate of the order 10−8 per site per generation [17], and thus about 10−10 per site per cell division.
Besides the use of MS to reconstruct cell lineage trees which was proposed also by others [3], [4], [18]–[20], several other retrospective methods to trace cell lineages in mammalians have been proposed. These methods include the genomic profiling of single nucleotide polymorphism (SNP) [21]–[23], copy number variations (CNV) [24] and DNA methylation [25]–[27]. A common feature of all these methods is the use of a genomic property that accumulates mutations during cell divisions, thus making it suitable to be used as a genomic signature.
While phylogenetic lineage tree reconstruction of cells is similar to that of organisms and species, it also has unique characteristics such as the existence of stem cells (that influence the shape of the tree), a sometimes very shallow tree (in the order of dozens of generations), a dramatic variation in the number of divisions the cells have undergone since the zygote (which is much larger than what exists in species with different evolutionary paces), and the fact that the cells have undergone binary cell divisions. Besides the last feature, which has been widely investigated in population genetics [28], these unique characteristics as well as the uncertainties about the exact nature of the mutational process in somatic cells require an assessment of the accuracy of known lineage reconstruction algorithms for this application. Another meaningful difference is that MS are usually used to define relationships between groups (species or populations) [29], and not between individuals. Thus the mathematical measures defining the distance are different.
The goal of this work is to test the existing algorithms on experimental cell data and to validate their use. In addition we want to test which of these methods (even though not developed for the purpose of cell lineage reconstruction) performs best.
In order to accomplish this goal, we took experimental data from clearly known biological expectations and examined whether the reconstructed trees present this knowledge. The data was obtained by isolating cells from different mice and humans, and extracting their genomic signatures (see Materials and Methods). Due to the fact that the real cell lineage tree is not known, we examined two aspects of the estimated tree. One aspect is the clustering of biologically distinct cell groups on the tree, and the second is the ability to distinguish between two groups of cells that are known to have different depths (number of divisions the cell has undergone since the zygote).
As mentioned earlier the goal of this work is to validate and quantify the ability to reconstruct a cell lineage tree utilizing genomic signatures of individual cells that record mutations in microsatellites. Since precise inference of tree topology cannot be accomplished using our current limited number of loci (See Table S6), we examined in this work whether certain aspects of the inferred tree reflect known biological scenarios. The first is the clustering of different cell groups. The basic assumption of this test is that if a statistically significant clustering on the cell lineage tree is consistent with a biological characteristic, then such clustering is very likely to reflect a real biological phenomenon, and therefore the more significant the clustering found by an algorithm, the better the algorithm. The simplest possible grouping of cells can be according to which individual they belong. Investigating the lineage relations among cells of different individuals is normally not done, however it is useful as a benchmark to test the validity of cell lineage reconstruction algorithms, as cells from different individuals clearly should be clustered separately. The second aspect we examined is the depth separation between different types of cells that are known to have different depths.
The tree reconstruction algorithms that we used are Neighbor Joining (NJ) [31], UPGMA [32], and a quartet-based method as implemented in the QMC tool [33]. The distance measures that we used are two versions of the Absolute genetic distance (regular and normalized), Euclidian distance, Equal or Not distance, and six versions of likelihood distances – assuming equal mutation rates for all loci, assuming two different mutation rates for mono-nucleotide and di-nucleotide repeats, and assuming length dependent mutation rates. These three mutation models were tested on both the Stepwise Mutation Model – SMM, and the Multistep Mutation Model – MMM (for more details regarding the reconstructing methods see Materials and Methods). In addition to distance-based algorithms, Bayesian methods can also be used to infer the cell lineage tree. Even though these Bayesian methods hold great promise, there are currently only a very limited amount of existing tools that can be used to analyze MS. In addition most of the existing tools (such as MrBayes [34], [35], Migrate [36] and Beast [37]) assume no linkage between the different loci. However, in the cell-lineage tree, the range of the linkage disequilibrium is infinite (i.e. the whole chromosome is fully linked, as the mitotic recombination rate is very small compared to the MS mutation rate multiplied by the chromosome length and the depth of the trees). Moreover, in the special case of cells inside multicellular organisms, since each cell has only one single parent cell from which all its chromosomes derive, all the MS loci share the same history, and hence all the loci are fully linked (including loci that are on different chromosomes).The only ML tool that we found to be applicable to our case is BATWING [38] which reads in multi-locus haplotype data, a model and prior distribution specifications. It may be worthwhile to test the performance of the other Likelihood\Bayesian algorithms, even though they assume that loci are not fully linked; however, these algorithms are highly computationally intensive, and therefore we could not test these.
In this section we checked the clustering quality of the tree reconstruction methods. An example for a case where cells from different individuals are clustered distinctively on the tree can be seen in Figure 1. We used all cell types (see Table S1 for the list of cell types of each individual) from three mice (Figure 1A) and from seven humans (Figure 1B), with the Normalized Absolute genetic distance and the Neighbor Joining tree reconstruction algorithm. It can be seen that the cells of each individual are clustered separately on the tree.
However, due to the many types of noise existing in the system, such a distinct clustering is not likely to happen in all cases, especially if the individuals are related to each other (as in the case of some of the experimental mice), and their zygotes are genetically close. In such cases, due to the small panel of MS used, cells from different individuals can randomly accumulate mutations that reduce the genetic distance between them, and may become closer to each other than to other cells from the same individual. This effect depends on the ratio between the genetic distance between the zygotes of the mice and the number of divisions the cells in each mouse underwent. We showed via computer simulation (Text S1 and Figures S1, S2, S3) that when this ratio is small it is very hard to distinguish which cells belong to which mouse. In addition, we showed that as the number of loci grows, the separation between the mice improves.
Our panel contains ∼120 loci but we prefer to ignore loci where there are allelic dropouts- i.e. where there is amplification failure of one of the two heterozygous alleles while the other allele successfully amplifies, which may often be misinterpreted and lead to errors in allele size determination. We thus used an average of about 80 loci per cell. In addition, those 80 loci can be different between distinct cells, so the actual number of loci used for the distance calculation can be even smaller (with a minimum limit of 25). We showed that when the ratio is 0.2 and the mutation rate is 1/100, a separation of 90% can be achieved using at least 200 loci. An example of such a case is shown in Figure 2 where we present the tree of five mice with all their cells. It can be seen that three of these mice (M1, M2 and M3) are separated quite well in the lineage tree, compared to the cells of the other two mice (M7 and M8) which are strongly mixed. It may be due to their possible family relations; however since we do not know the real relations but rather estimation, other types of noise can mix the cells, such as errors in the amplification of the genetic sequence or in the PCR reaction.
Nevertheless, not all the algorithms and distance measures suffer from this problem to the same degree. This may be due to the fact that some measures describe the mutational process more accurately, or due to some other robustness feature of the algorithm. An example for a different performance between different metrics is shown in Figure 3 where we applied two methods to the same dataset. It can be seen that the NJ- Normalized Absolute (Figure 3A) produced better cluster separation than the NJ- Equal or Not method (Figure 3B).
The difference between methods in the clustering separation of cell groups necessitates the quantification of their performance, in order to determine which method is the best (if any). In order to do so, we used three measures to quantify the clustering quality of distinct groups: the Quality of the Largest Cluster (QLC) from each group, the Tree Entropy (TE), and the probability of getting such a cluster under the assumption of hyper geometric sampling (HS). (A detailed explanation of the measures is given in the Materials and Methods).
We analyzed the performance of the methods on all the information that we have available: cells from nine mice and seven humans, each containing a few types of cells (Table S1). We combined two or more individuals into one larger dataset, using all their cell types or a single type. For each dataset, we reconstructed the cell lineage tree using all the methods we have. Then we quantified the separation performance of each method with the measures listed above, and determined the method with the best performance. In the following section, we present the results of this test and its variants.
As mentioned before, clustering is just one example of a feature of the tree in which we are interested. Another feature is the depth of specific types of cells. Different depths can indicate different biological scenarios, for example whether some types of cells divide only during the embryonic stage or also in the adult stage. In this section our goal is to quantify the performance of the different methods in identifying depth differences between groups of cells.
In order to obtain cases where the depth separation between the cell groups is known, we used the same type of cells from individuals with a substantial gap between their ages. The list of cell types of each individual is given in Table S1. The tree-reconstruction algorithm that was applied in this case is the NJ algorithm, since it is the only algorithm that allows different depths for different cells inside the same individual. Two examples of trees with depth difference are presented in Figure 6, one with a good depth separation (Figure 6B) and one with a poor separation (Figure 6A).
The depth of each group of cells as reconstructed by the NJ varies even if all the cells of this group actually have exactly the same number of divisions. Therefore for each group of cells, the depth is described by a distribution rather than by a single number. In order to quantify the performance of a method in separating between the two groups, quantities that differentiate between distributions are needed.
The most natural choice is the Kolmogorov-Smirnov (KS) test, which measures the similarities between two datasets. However, this test has some disadvantages for our purposes; the most significant one is the ability to determine that two datasets are different even if they have exactly the same average depth, in cases where their standard deviations are substantially different. Therefore in addition to the KS we added two other measures that focus on the separation between the two distributions. The first is the normalized distance between the mean of the two groups, and the second is the overlap percentage between the two distributions (see Materials and Methods for more details). The difference between them is that the overlap percentage is affected by the behavior of the extreme cells, while the normalized average distance captures the behavior of the bulk. Another minor difference is that the average normalized distance can distinguish between the separation qualities of methods even in the case of fully separated groups.
A summary of the depth separation tests' results is presented in Figure 7 and Figure S13 (full results are given in Table S5.4). In this case there is no one method which is superior over the others, but a few which are rather equally good: Normalized Absolute, Euclidean and SMM with length dependent mutational rates. This implies that for the depth separation there is no one tree which can be considered the correct one. The various inferred trees should be seen as approximate projections of the real tree, which cannot be inferred precisely as of yet, since the genetic identifier is not sufficiently informative. These results were validated by simulations in which lineage trees with different cell depths were reconstructed and the differences were evaluated using the depth measures described in Materials and Methods. In each iteration, two lineage trees were simulated and the difference between the depths was calculated, where we distinguished between cases in which the trees were relatively shallow, and cases in which the trees were relatively deep. The simulations show that when the trees are shallow, when using 100 loci, there is no one method which is uniformly better than the others, in accordance with our result obtained with real data. When using 50 loci, the Normalized-absolute has the worst performance, while with 500 loci; the Normalized-absolute performs best. When the trees are deep, with 500 loci there is no one method that is better than the others, whereas when using fewer loci, the Normalized-absolute has the worst performance (results are shown in Table S8 and in Figure S14).
Bootstrap analysis was used in order to evaluate the robustness and reproducibility of the estimated trees, the clustering of the tree and the depth separation according to cell type. We performed this analysis on several mice and human datasets which showed a good clustering or depth separation using the NJ-Normalized Absolute method.
The bootstrap values were obtained by generating 100 trees using MS values extracted from sampling with replacement of the loci from each dataset. The bootstrap showed that the robustness of any particular branch in the tree is low, but the robustness of the clustering results and depth separation according to cell type is high (see Table S9 for all the results).
In the preliminary stages of the cell lineage research conducted by our lab, small-scale investigations of the ability to reconstruct cell lineage trees were done. The large amount of information that was gathered during the last few years enabled us to conduct this investigation in a much more comprehensive way. The main outcome of this research is that even though currently only a limited amount of microsatellite loci are available, preventing the reconstruction of the accurate cell-lineage tree, many biological conclusions can still be confidentially drawn. By this we mean that apart from specific noisy cases, almost always we are able to identify the correct biological scenario from the reconstructed cell lineage if the proper tree reconstruction algorithm and distance measure are used.
Among the NJ methods, we have found that the NJ- Normalized Absolute method outperforms the other methods at inferring the clustering of distinct groups. Interestingly, this shows that clear cluster- separation is not necessarily correlated with the most precise description of the mutational process. A result with a similar spirit was obtained previously [29] for using MS allele frequency in order to infer the phylogenetic tree of different species/populations. They also found that the best method is not necessarily the one that described the mutational process in the most accurate way.
However, such a measure cannot describe accurately cell depth because depth information is eliminated in the normalization procedure. It is not unreasonable to assume that a Likelihood (or a Bayesian) method tailored towards cell lineage analysis that will make use of all the cells' information, without summarizing it into distance measure, will enable one to infer simultaneously both the topology and the depth in an accurate way. We hope to follow such a path in the future.
We expect that in the coming years next generation sequencing methods will provide us with a much richer genetic signature, and thus improve our ability to infer the cell lineage tree more accurately. This in turn will enable relying on even fine details of the cell lineage tree and not only its rough features.
All animal husbandry and euthanasia procedures were performed in accordance with the Institutional Animal Care and Use Committee (IACUC) of the Weizmann Institute of Science.
All human patients signed an informed consent; the study has received Helsinki authorization and was approved by the Rambam Hospital IRB committee and by the Bioethics Committee of the Weizmann Institute of Science.
Our aim is to quantify the performance of different tree reconstruction methods in inferring clustering and depth separation. Most of the methods we tested are distance-based algorithms which use a distance measure between the cells to iteratively join close samples together, such as the Neighbor-Joining algorithm (the full list of methods and distance measures is given below). We tested each method on two features:
The first distance based algorithm we used is UPGMA (Unweighted Pair Group Method with Arithmetic Mean algorithm) [32] which assumes that all lineages evolve at the same rate. This assumption limits us to reconstructing trees which contain cells that underwent the same (or very similar) number of divisions.
The second algorithm we used is NJ (Neighbor-Joining) [31]. In order to fit the algorithm to our problem, we corrected the branch lengths such that they will not be negative. When negative branches appear during the running of the algorithm, we set its length to 0, adjusting its sibling branch accordingly [40]. Note that this correction does not change the inference of the topology, since it depends only on the distance matrix, and is not affected by the branch lengths of the inferred tree. With the NJ algorithm, a rooted tree can be created by using an out-group, and the root can then effectively be placed on the point in the tree where the edges from the out-group connect. The root in our trees is usually a mix of a wide variety of cell types (a description of the root's determination is given below, in the Data description section).
The third algorithm we used is QMC (Quartet MaxCut) [33], [41]. Quartets-based methods were initially proposed to provide an alternative to maximum likelihood methods, which are computationally intensive. These methods first estimate a set of trees on many four-leaf subsets of the taxa, and then combine them into a tree on the full set of taxa. The QMC method is based on a recursive divide and conquer algorithm that seeks to maximize the ratio between satisfied and violated quartets at each step. The common implementations of the quartet method (including QMC) produce only a tree topology without any explicit information about the branches lengths. Even though it is possible to add depth estimation to the QMC, we limited ourselves to assessing the quality of existing methods without any new features added.
Apart from the distance based algorithms, we tested a Bayesian method for inferring the cell lineage tree. We used the computer software, BATWING [38] which reads in multi-locus haplotype data, a model and prior distribution specifications. This program uses a Markov Chain Monte Carlo (MCMC) method based on coalescent theory to generate approximate random samples from the posterior distributions of parameters such as mutation rates, effective population sizes and growth rates, and times of population splitting events. Even though there is currently no Likelihood or Bayesian tree reconstructing algorithm that uses microsatellites, which was developed to include the unique features of cell lineages, the growing population implementation of BATWING seemed most suited for our study. The priors we used are 1/100 for the mutation rate, and uniform distributions for the effective population size and the population growth rate per generation (on the intervals [10,000 10,000,000] and [0 2] respectively).
The distance-based methods require a distance measure between cells, which ideally should be linear with the actual number of divisions separating any two samples, and should provide the most robust tree reconstruction. We have tested several different distance functions. In these functions and are the number of repeats in the single allele of the and sampled cells, respectively, and is the set of alleles which were amplified for both samples and (for autosomal loci, both alleles were included, and for chromosome X loci, one allele was included in male samples):
We used three different measures to quantify the quality of the clustering separation ability:
All these measures evaluate the quality of the separation of the distinct groups on the tree, but they measure parameters that are slightly different. The QLC focuses on the existence of one large cluster and ignores the behavior of the rest of the cells. The TE on the other hand is determined by the number of distinct clusters of each cell type, and ignores their sizes. Hence, the TE focuses on the global behavior of the tree and not just on one sub-tree. The HS, like the QLC, focuses on the existence of a large cluster, but does not ignore the rest of the cells as it detects a statistically significant clustering of a group of cells on the lineage tree.
We used three different measures to quantify the quality of the depth separation ability:
We simulated trees similar to the ones reconstructed using the real data, with some trees containing 3 individuals with 5 different cell types for each individual, and other trees composed of a single individual. We simulated several kinds of topologies, which were different from each other in branch length. For example, in one topology the distance between the leaves and their MRCA was high, compared to the distance between the root and these MRCAs, and in another topology this distance between the leaves and their MRCA was much lower. We repeated the simulation 1000 times, where in each iteration, we built a random tree and randomly added MS mutations according to a given mutation rate (1/100, 1/1000 and 1/10000) using a binomial distribution. We then reconstructed the tree using all of our methods and compared it with the actual tree that was generated. The topology comparison between the inferred and the actual trees was done using Penny & Hendy's topological distance algorithm [47]. In this algorithm each internal edge confers a partitioning of the tree into two groups by removing the edge. We assigned a score equal to the ratio of equal partitions of the two trees to the total number of partitions. For each of the simulated trees we also calculated the clustering measures (mentioned above).
|
10.1371/journal.pntd.0001470 | Climate-Based Models for Understanding and Forecasting Dengue Epidemics | Dengue dynamics are driven by complex interactions between human-hosts, mosquito-vectors and viruses that are influenced by environmental and climatic factors. The objectives of this study were to analyze and model the relationships between climate, Aedes aegypti vectors and dengue outbreaks in Noumea (New Caledonia), and to provide an early warning system.
Epidemiological and meteorological data were analyzed from 1971 to 2010 in Noumea. Entomological surveillance indices were available from March 2000 to December 2009. During epidemic years, the distribution of dengue cases was highly seasonal. The epidemic peak (March–April) lagged the warmest temperature by 1–2 months and was in phase with maximum precipitations, relative humidity and entomological indices. Significant inter-annual correlations were observed between the risk of outbreak and summertime temperature, precipitations or relative humidity but not ENSO. Climate-based multivariate non-linear models were developed to estimate the yearly risk of dengue outbreak in Noumea. The best explicative meteorological variables were the number of days with maximal temperature exceeding 32°C during January–February–March and the number of days with maximal relative humidity exceeding 95% during January. The best predictive variables were the maximal temperature in December and maximal relative humidity during October–November–December of the previous year. For a probability of dengue outbreak above 65% in leave-one-out cross validation, the explicative model predicted 94% of the epidemic years and 79% of the non epidemic years, and the predictive model 79% and 65%, respectively.
The epidemic dynamics of dengue in Noumea were essentially driven by climate during the last forty years. Specific conditions based on maximal temperature and relative humidity thresholds were determinant in outbreaks occurrence. Their persistence was also crucial. An operational model that will enable health authorities to anticipate the outbreak risk was successfully developed. Similar models may be developed to improve dengue management in other countries.
| Dengue fever is a major public health problem in the tropics and subtropics. Since no vaccine exists, understanding and predicting outbreaks remain of crucial interest. Climate influences the mosquito-vector biology and the viral transmission cycle. Its impact on dengue dynamics is of growing interest. We analyzed the epidemiology of dengue in Noumea (New Caledonia) from 1971 to 2010 and its relationships with local and remote climate conditions using an original approach combining a comparison of epidemic and non epidemic years, bivariate and multivariate analyses. We found that the occurrence of outbreaks in Noumea was strongly influenced by climate during the last forty years. Efficient models were developed to estimate the yearly risk of outbreak as a function of two meteorological variables that were contemporaneous (explicative model) or prior (predictive model) to the outbreak onset. Local threshold values of maximal temperature and relative humidity were identified. Our results provide new insights to understand the link between climate and dengue outbreaks, and have a substantial impact on dengue management in New Caledonia since the health authorities have integrated these models into their decision making process and vector control policies. This raises the possibility to provide similar early warning systems in other countries.
| Dengue viruses are the most important arthropod-borne viruses affecting humans. During the past century, the four serotypes (DENV 1 - DENV 4) have spread to about a hundred countries in the tropical and subtropical world including Asia, Africa, the Americas and the Pacific. Each year, an estimated 50 million people contract dengue fever with at least 500,000 cases of dengue haemorrhagic fever or dengue shock syndrome leading to 25,000 deaths [1]. The spatial distribution of this emerging infectious disease largely reflects the distribution of its primary urban mosquito vector, Aedes aegypti [2]. As no effective vaccine and specific treatment exist, vector control currently represents the only resource to mitigate dengue outbreaks.
Epidemic dynamics of dengue, like those of other vector-borne diseases, are driven by complex interactions between hosts, vectors and viruses that are influenced by environmental and climatic factors. Several determinants in dengue fever emergence have been identified including human population growth, accelerated urbanization, increased international transport, weakened public health infrastructure as well as a lack of effective vector control and disease surveillance [3]–[6]. On the other hand, there is growing interest in the impact of climate change on the emergence or re-emergence of vector-borne infectious diseases such as dengue [7]–[10]. It has been shown that climate-induced variations in modelled A. aegypti populations were strongly correlated to reported historical dengue cases (1958–1995) at the global scale [11], and a potential increase in the latitudinal and altitudinal distribution of A. aegypti and dengue are expected under global warming [5], [12].
In a specific ecosystem, the required conditions for the occurrence of a dengue outbreak include i) the presence of a dengue virus, ii) the presence and a sufficient density of competent vectors, iii) a sufficient number of susceptible humans that is serotype-specific, and iv) favorable environmental and climatic conditions for dengue transmission. Despite evidence that climate can influence dengue like other vector-borne diseases (i.e. vector population size and distribution, vector-pathogen-host interactions, and pathogen replication [7], ), the relationships between climate, Aedes mosquitoes density and behaviour, human populations and dengue incidence are not well understood.
Previous studies have shown that temperature influences the lengths of the mosquito gonotrophic cycle and the extrinsic incubation period of the virus within the mosquito, the survival rate of adults, the mosquitoes population size and feeding behaviours and the speed of virus replication [7], [13], [15]–[19]. Water is necessary for eggs and larva development, mosquito breeding, and humidity affects adult mortality [16]–[17], [20]–[22]. Temperatures and precipitations have been identified as influencing incidence rates of dengue in several endemic areas in the world (i.e. Thailand [23]–[24], Taiwan [25]–[27], Singapore [28], and Puerto Rico [24], [29]). On a broader scale, it is plausible that El Niño-Southern Oscillation (ENSO) also influences patterns of dengue transmission [23]–[24], [30]–[31]. This coupled ocean-atmosphere phenomena results in warm waters displacement and changes in sea surface temperatures (SST) across the Pacific Ocean, and has a strong influence on regional climates, particularly in the Pacific. ENSO can induce large temperature, humidity and precipitation changes for months (see the websites of the International Research Institute for Climate and Society (IRI, www.iri.org), and the National Oceanic and Atmospheric Administration (NOAA, www.noaa.gov) for more details). Importantly, previous studies revealed a positive correlation between ENSO, as measured by the Southern Oscillation Index (SOI), and dengue outbreaks in the South Pacific islands [30]–[31].
Our study was conducted in New Caledonia where dengue represents a major public health problem like in many Pacific Islands Countries and Territories [32]. The first dengue outbreak in New Caledonia occurred in 1884–1885 [33]. Disease transmission increased after World War II, and successive waves of epidemics involving all four serotypes were reported. Since 2000, serotype 1 has been predominant [34] causing more than 6,000 cases during the 2003–2004 epidemics [35] and about one thousand of cases in 2008. Although the serotype 4 [36] was involved in a major outbreak in 2009 (8,456 cases), the serotype 1 is still circulating. New Caledonia has had an effective surveillance system for dengue and access to high quality meteorological data for many years. Since 2000, regular entomological surveillance is performed. This provides an opportunity to study the influence of climate variations on dengue dynamics.
We analyzed the epidemiology of dengue fever in Noumea, the capital of New Caledonia, from 1971 to 2010 together with local and remote climate influences. The objectives of this study were i) to improve our knowledge of the relationships between meteorological variables, entomological surveillance indices and dengue fever dynamics at seasonal to inter-annual time scales, ii) to identify suitable conditions for an epidemic occurrence, and iii) to develop a predictive model for dengue outbreaks that can be integrated in an early warning system in New Caledonia.
New Caledonia is a French overseas territory located in the subregion of Melanesia in the southwest Pacific, about 1,200 kilometres east of Australia and 1,500 kilometres northwest of New Zealand. It lies astride the Tropic of Capricorn, between 19° and 23° south latitude. Its climate is tropical.
This archipelago of 18,575 square kilometres is made up of a main mountainous island elongated northwest-southeast 400 kilometres in length and 50–70 kilometres wide, the Loyalty Islands (Mare, Lifou, and Ouvea), and several smaller islands (e.g. Isle of Pines). The population was estimated in January 2009 to be 245,580 [37]. Approximately half of inhabitants are concentrated in the southeast region of the main island around Noumea, the capital.
A. aegypti is the only mosquito vector of dengue in New Caledonia. The two others vectors of dengue present in the Pacific region, A. albopictus and A. polynesiensis, have never been detected in this archipelago [38]–[40]. In Noumea, most of A. aegypti breeding sites are outdoors and therefore rainfall dependent.
Bivariate and multivariate analyses were conducted using the R software package (R development Core Team version 2.9.1 [42]).
During the 1971–2010 period, a significant correlation was found between dengue incidence rates and mean annual mean Temp in Noumea (Spearman's coefficient rho = 0.426, p-value = 0.007, Figure 1) but there was no significant correlation with annual mean RH and Precip. Similar results were obtained with conserved trends and detrended data. Anomalies of annual means of mean Temp, Precip and mean RH were significantly correlated with ENSO, as measured by Niño 3.4 (rho = −0.365, p-value = 0.029; rho = −0.481, p-value = 0.003; rho −0.486, p-value = 0.003, respectively). During El Niño (positive value of Niño 3.4), the weather was cooler and drier. During La Niña (negative value of Niño 3.4), the weather was warmer and wetter. However, no direct correlation was found between ENSO and dengue incidence rates at the inter-annual scale (rho = −0.106, p-value = 0.539). Dengue outbreaks occurred during either El Niño, La Niña or neutral phases of ENSO.
During the 2000–2009 period, dengue incidence rates, meteorological and entomological data were analyzed in Noumea at a monthly scale. A strong seasonal distribution of HI, BI and API was observed (Figure 3), and significant correlations were found between monthly entomological surveillance indices and climate variables (data not shown). Although the highest dengue incidence rates and the highest values of HI, BI and API were observed during the same period of the year (from January to July), no significant time-lagged correlation has been found between monthly entomological indices and dengue incidence rates reported in Noumea over the 2000–2009 period (supporting Figure S1). We did not find relevant entomological patterns during dengue outbreaks. Accordingly, entomological surveillance indices were not used for the modelling of dengue outbreak risk.
Based on the tercile method, there were 13 epidemic years (dengue incidence rate in the upper tercile, i.e. >19.48 cases/10 000 inhabitants) and 13 non epidemic years (dengue incidence rate in the lower tercile, i.e. <4.13 cases/10 000 inhabitants). A detailed analysis was performed based on monthly and quarterly meteorological data measured from September (year y-1) to April (year y), i.e. four months before and after the outbreak onset.
Temperatures (min Temp, mean Temp and max Temp) were higher during epidemic years than during non epidemic years. The peak of max Temp, observed usually in February, preceded the epidemic peak of dengue with a lag of 1–2 months (Figure 4a). Analysis of daily data allowed identifying important temperature thresholds. It revealed that the number of days with max Temp exceeding 32°C, mean Temp exceeding 27°C, and min Temp exceeding 22°C were significantly higher during epidemic years than during non epidemic years. The most important and significant differences were observed during the first quarter of the year, principally in February for max Temp (p-value<0.01 using a t-test, Figure 4b).
By contrast, the relationships between Precip, mean RH and dengue dynamics were not clear, as shown in supporting Figure S2. Highest Precip and mean RH were observed in February–March–April during the epidemic phase of dengue. Using a t-test, Precip and mean RH were significantly lower in February during epidemic years than during non epidemic years (p-value<0.01 and = 0.04, respectively). Inversely, the ETP was significantly higher in February (p-value = 0.02). WF, HB, ENSO indices and entomological surveillance indices were not significantly different between epidemic and non epidemic years.
Meteorological variables showing strongest correlations with the epidemic years series, as defined in the Methods section, are presented for each family of variables in Table 1. Significant correlations were identified with several local meteorological variables (particularly Temp, Precip, RH, and ETP) but not with ENSO indices. No or poor correlation was found with WF and HB. In accordance with Figure 4 and supporting Figure S2, Temp were positively correlated with dengue outbreaks in Noumea, whereas Precip and RH measured in February were negatively correlated with dengue outbreaks. A positive correlation was found between the ETP measured in February and the occurrence of dengue outbreaks.
First, in order to produce an explicative model of dengue outbreak, we selected meteorological variables observed within the period of dengue outbreak onset, i.e. from January to April (Figure 2). The best SVM model based on the minimum AICc (−79.21) was obtained using two meteorological variables, i.e. the number of days with maximal temperature exceeding 32°C during the first quarter of the year (NOD_max Temp_32_JFM), and the number of days with maximal relative humidity exceeding 95% during January (NOD_max RH_95_January). The addition of a third meteorological variable did not improve the performance of the model. Results obtained in leave-one-out cross validation (Figure 5) were close to those obtained with the complete dataset (Figure S3) and were characterized by a high ROC-AUC value reaching 0.80 and 0.85, respectively. As indicated by the ROC curves, most of epidemic years were predicted correctly with high probability and few false alarms. Importantly, with bivariate analysis, NOD_max Temp_32_JFM was positively correlated with the occurrence of dengue outbreak (rho = 0.57, p-value = 0.002) whereas NOD_max RH_95_January did not appear to be a discriminatory meteorological variable (rho = −0.11, p-value = 0.58). With multivariate analysis, these two variables were highly informative and discriminatory. Scatter plots of epidemic and non epidemic years as a function of these two variables allowed the identification of three distinct groups (Figure 6): group A including years characterized by low NOD_max Temp_32_JFM (<12 days) and low NOD_max RH_95_January (<12 days), group B including years characterized by high NOD_max Temp_32_JFM (>12 days) and low NOD_max RH_95_January, and group C including years characterized by low NOD_max Temp_32_JFM and high NOD_max RH_95_January (>12 days). According to the tercile method of years classification, all non epidemic years belonged to group A whereas all epidemic years, except 1973 and 2003, belonged to either group B or group C. Similar results were obtained using the median method ensuring the inclusion of all years, preferable for the development of SVM models. Only four years (1978, 1979, 1985, and 2002) belonging to the middle tercile (dengue incidence rate ranging from 4.13 to 19.48 cases/10 000 inhabitants/year) were incorrectly classified using the median method. In 2002, although favorable climatic conditions for dengue outbreak were observed, the incidence rate (5.24 dengue cases/10 000 inhabitants/year) was close to the median (7.65 dengue cases/10 000 inhabitants/year). In 1978, 1979 and 1985, the low values of NOD_max Temp_32_JFM and NOD_max RH_95_January were not favorable for dengue outbreak. However, incidence rates (7.74, 10.63, and 11.24 dengue cases/10 000 inhabitants/year, respectively) were close to the median. Two years (1973 and 2003) belonging to epidemic years using either a tercile or a median method of classification were characterized by low NOD_max RH_95_January and intermediate NOD_max Temp_32_JFM, as members of group A (non epidemic years). However, dengue outbreaks occurred with high incidence rates (23.64 and 213.58 dengue cases/10 000 inhabitants/year in 1973 and 2003, respectively). These mismatches indicate that i) the model fails for years that are difficult to classify as their dengue incidence rates were close to the median and in the middle tercile and, ii) NOD_max Temp_32_JFM and NOD_max RH_95_January alone cannot account for all dengue outbreaks (Figure 6). It is likely that other climate events and other factors influencing dengue dynamics contribute to the epidemic spread of dengue viruses during these peculiar years. We were thus able to build an efficient explicative model of dengue epidemics based on meteorological variables contemporaneous to the outbreak.
Another challenge was to construct a predictive model for dengue epidemics using variables available prior to the outbreak onset, i.e. from September (year y-1) to December (year y-1). Accurate predictive skill (AICc = −66.64) was achieved with the SVM model built from the value of the two following variables: the quarterly mean of maximal relative humidity during October–November–December (max RH_OND), and the monthly mean of maximal temperature in December (max Temp_December) of the year y-1 with a ROC-AUC value of 0.83 (supporting Figure S4). Probabilities obtained in leave-one-out cross validation (Figure 7) and the corresponding ROC-AUC value reaching 0.69 illustrate the robustness of this predictive model. Importantly, max RH_OND and max Temp_December were not significantly correlated with the risk of dengue outbreak with bivariate analysis (rho = 0.24, p-value = 0.14; and rho = 0.25, p-value = 0.14, respectively).
Scatter plots of epidemic years and non epidemic years built from the combination of meteorological variables used for the SVM explicative model (Figure 8) and for the SVM predictive model development (Figure 9) show that dengue outbreaks occurred in distinct climatic conditions in Noumea. With the SVM predictive model, as noted with the SVM explicative model, epidemic years belonged to two different groups of data according to the value of max RH_OND and max Temp_December (see the two red kernels corresponding to high risk of dengue outbreak in Figure 9). Dengue outbreaks occurred following either years characterized by high max Temp_December and relatively low max RH_OND, or years characterized by high max RH_OND_December, and max Temp_December. To note, the high value of max Temp_December (31.2°C) and the relatively low value of max RH_OND (86.8%) measured in 2010 indicate a high risk (74%) of dengue outbreak for 2011.
A comparison of the results obtained with the explicative model and the predictive model was performed together with a detailed analysis of the relationships between meteorological variables used to build the explicative model (NOD_max Temp_32_JFM and NOD_max RH_95_January) and those used to build the predictive model (max RH_OND and max Temp_December). As shown in Figure S5, strong relationships exist between the values of max Temp and max RH measured at the end of the year y-1, and those measured at the beginning of the year y. Low max RH_OND and max Temp_December (year y-1) were predictive of low NOD_max Temp_32_JFM and NOD_max RH_95_January (years y, group A). High max RH_OND and max Temp_December (year y-1) were predictive of either high NOD_max Temp_32_JFM and low NOD_max RH_95_January (years y, group B), or low NOD_max Temp_32_JFM and high NOD_max RH_95_January (years y, group C). Results obtained with the predictive model were highly consistent with those obtained with the explicative model with similar probabilities of dengue outbreak risk obtained for 30 of the 40 studied years. Failures of the predictive model can be explained by a lack of correlation between these meteorological variables on a few occasions (e.g. 1982, 1983, 1995). For example, although the predictive model estimated a risk of dengue outbreak close to 5% in 1995, the explicative model estimated a risk over 90%, and a major outbreak occurred. The value of max RH_OND and max Temp_December measured in 1994 (87% and 27.6°C, respectively) were relatively low and therefore not predictive of outbreak risk. However, climatic conditions were favorable for a dengue outbreak occurrence (NOD_max Temp_32_JFM = 20 days, NOD_max RH_95_January = 0 day, group B). This suggests that other climate variables or meteorological processes may impact on the local value of NOD_max Temp_32_JFM and NOD_max RH_95_January.
The influence of climate on dengue dynamics in Noumea, the capital of New Caledonia, over the 1971–2010 period has been analyzed at different time scales using high quality and high resolution meteorological observation data, along with epidemiological and entomological surveillance data. During epidemic years, dengue outbreaks peaked around March–April at the end of summer season. The epidemic peak lagged the warmest temperature by 1–2 months and was in phase with maximum precipitations and maximum relative humidity. The seasonal evolution of entomological indices (e.g, Breteau, House and Adult productivity indices) matched the seasonality of dengue outbreaks.
No relationship was found between the inter-annual variations of dengue incidence rates and those of the entomological data. On the other hand, a number of meteorological indices developed from summertime temperature, precipitation or relative humidity showed a significant correlation with dengue occurrence.
New explicative and operational predictive models of dengue outbreak were developed. We used a multivariate SVM model to identify the best set of meteorological variables explaining dengue epidemics. We found that a non linear combination of two meteorological variables strongly outperforms a model based on a single variable or a linear approach, as commonly employed in the literature. We found the best explicative variables to be the number of days with max Temp exceeding 32°C during January–February–March (NOD_max Temp_32_JFM) and the number of days with max RH exceeding 95% during January (NOD_max RH_95_January). When the model gives a probability of dengue outbreak above 65%, these two variables explain 94% of the epidemic years and 79% of the non epidemic years (Figure 5). Most dengue outbreaks occurred within two kinds of distinct climatic conditions: high NOD_max Temp_32_JFM and low NOD_max RH_95_January, or low NOD_max Temp_32_JFM and high NOD_max RH_95_January. We were also able to build another SVM model based on two variables to predict dengue outbreaks in advance: the maximal temperature in December (max Temp_December) and maximal relative humidity during October–November–December (max RH_OND) of the year prior to the epidemics. For a probability of dengue outbreak above 65%, this model can predict 79% of the epidemic years and 65% of the non epidemic years (Figure 7).
Overall, the high performance of the climate-based models of dengue outbreak risk developed in our study suggest that dengue dynamics were essentially driven by climate during this 1971–2010 period in Noumea. The explicative model provides important and new information. We have shown that maximal values of temperature and relative humidity were determinant in dengue outbreaks occurrence and precise thresholds of their value were identified. Importantly, we found that the most relevant meteorological variables explaining dengue outbreaks were built using the number of days for which the variable was greater than a threshold value introducing the importance of the persistence of suitable climatic conditions. Our findings are compatible with the mosquito biology and viral transmission cycle.
The length of Aedes gonotrophic cycle is shorter at temperatures above 32°C and feeding frequency is more than twofold at 32°C as compared to 24°C; pupae development period reduced from four days at 22°C to less than one day at 32–34°C [16]–[17], [47]. Additionally, the experimental infection of A. aegypti with DENV-2 viruses showed that the extrinsic incubation period shortens from 12 days at 30°C to seven days at 32–35°C leading to an increasing risk of viral transmission from an infected mosquito to a susceptible host [15]. The influence of temperature on the rate of virus replication inside mosquitoes was also evidenced in the study of Watts et al. Temperatures may also influence the vector size and its biting rate [19], [21]. Consequently, it is likely that the increased level of viral transmission characterizing dengue outbreaks in Noumea at temperatures exceeding 32°C may be a consequence of shortening of the A. aegypti gonotrophic cycle and extrinsic incubation period, and of increased vector feeding frequency.
Mortality rate of larvae, pupae and adult mosquitoes as a function of temperature between 10 and 40°C can be represented by a wide-base ‘U’ graphical shape with lower mortality rate at temperature ranging from 15 to 30°C [16]–[20], [22]. Hence, A. aegypti mortality rate may be relatively constant at temperatures observed usually in Noumea, and the increasing mortality rate expected above 32°C is not likely to be an important limiting parameter in the spread of dengue viruses in this specific ecosystem.
Larval breeding places are mostly outdoors in Noumea and mosquito abundance increases during the rainy and humid season. Moreover, relative humidity may be determinant in A. aegypti egg development and adult population size that may itself be correlated with vectorial capacity [48]. High humidity shortens incubation and blood-feeding intervals; it favours adult mosquito longevity [20] and thus dengue transmission. This may explain why a sustained high RH during January is associated with a higher risk of dengue outbreak in Noumea.
On a broader scale, a growing number of studies have shown that ENSO may be associated with changes in the risk of mosquito borne diseases such as dengue [23]–[24]. By contrast, Hales et al. [31] further analyzed the relationships between the annual number of dengue cases in New Caledonia, ENSO, temperature and rainfall using global atmospheric reanalyses climate based data, and they did not find any significant correlation between SOI and dengue (Pearson's coefficient = 0.20). In accordance with this study, and with the advantage of observational and long term data, we found significant inter-annual correlations between ENSO and our local climate but not between ENSO and dengue (Table 1). Moreover, the selection process of multivariate models did not select any ENSO index neither in explicative mode nor in predictive mode. These findings suggest that, in New Caledonia, large-scale climate indices such as ENSO cannot account for the complexity of the local meteorological inter-annual situations. However, at a larger scale, Hales et al. showed that the number of dengue outbreaks in the South Pacific islands (aggregated data, 1970–1995) were positively correlated with the SOI [30], suggesting that La Niña may favour dengue outbreaks in this region of the world. The impact of ENSO on local weather in the South Pacific may strongly vary from one place to another. New Caledonia, located around 20° south latitude in the western Pacific is relatively far from the main centre of action of ENSO located in the equatorial central/eastern equatorial Pacific and its local weather is thus not only influenced by ENSO, but also by other climate modes such as the Madden-Julian Oscillation which strongly influences local meteorological parameters at intra-seasonal (30 to 90 days) time scales [49]. In contrast, ENSO influence may be stronger in islands located closer to the equator, the relationship between ENSO and dengue epidemics being therefore more straightforward [29].
Our long-term study also suggests an increasing risk of dengue outbreaks in New Caledonia in the context of global warming (Figure 1). Even though a global upward trend of dengue incidence rates was noted along the 1971–2010 period, and as surveillance methods and laboratory tests have evolved, it is difficult to know if the amplitude of dengue outbreaks is significantly growing.
Even though climate influenced the disease epidemiology in Noumea during this forty-year period, the reasons of dengue emergence in New Caledonia are multiple, including population growth (119,710 inhabitants in 1973 to 245,580 in 2009), accelerated urbanization particularly around Noumea, tourism development and increasing international and inter-islands traffic [50]. The emergence of dengue fever in other parts of the world, particularly South East Asia where dengue is endemic with a co-circulation of the four serotypes, represents an increasing source of virus introduction into New Caledonia. Indeed, multiple and repeated introductions of dengue viruses have been detected from several countries in Asia [34]. Moreover, the geographical distribution of A. aegypti has expanded during recent decades in New Caledonia (Paupy and Guillaumot, unpublished data).
Well known factors may have contributed to the epidemic dynamics such as the size of susceptible human hosts and vectors populations. In the absence of seroprevalence data, and due to the lack of long term entomological data, these variables were not included in the input dataset of the models. Nevertheless, as dengue is known to confer a prolonged serotype-specific immunity in the long term, herd immunity represents an important factor in understanding dengue dynamics [51]–[54]. In New Caledonia, successive waves of dengue outbreaks involving the same serotype were reported in 1980 and 1986 (DENV-4), 1989 and 1995 (DENV-3), 2003 and 2008 (DENV-1). This constant interval time between two epidemics involving the same serotype has already been observed in other South Pacific Islands [55]–[57]. Recently, a large molecular characterization of DENV-1 viruses collected regularly in French Polynesia between the 2001 and 2006 outbreaks revealed that the virus responsible for the severe 2001 outbreak was introduced from South-East Asia, and evolved under an endemic mode until its re-emergence under an epidemic mode five years later [56]. These findings suggest that 5–6 years may be necessary for the renewal of the susceptible population in these islands. In New Caledonia, at four occasions, dengue outbreaks were detected between January and July during two successive years: in 1976–1977 (DENV-1), 1995–1996 (DENV-3), 2003–2004 (DENV-1), and 2008–2009 (DENV-1 and DENV-4). This suggests that environmental conditions may be not favorable for dengue transmission all through the epidemic year, particularly during the second semester of the year characterized by lower values of entomological indices. It is likely that dengue re-emerged the following year when climatic conditions were favorable for dengue transmission (as suggested by the results of our explicative model in 1977, 1996, 2004 and 2009) and the size of the mosquito-vector and susceptible human populations were still sufficient for a large spread of dengue viruses. In these four examples of recurrent outbreaks during two consecutive years, it is more likely that the end of the epidemic was driven by limiting climatic factors and intricate entomological factors rather than by the depletion of the susceptible population.
The relationship between Aedes density and the intensity of dengue transmission remains unclear [47], [58]–[60]. Although dengue viruses cannot circulate if mosquito vectors are not present, the vector density of adult female A. aegypti necessary for dengue viruses to become endemic or epidemic remains unknown. In Noumea, entomological indices (HI, BI and API) were not correlated with the incidence rate of dengue, they were sometimes lower during epidemic than during non epidemic periods and lowest values were measured during the largest outbreak in 2009. The fact that these usual entomological surveillance indices (particularly API) are good indicators of adult density in Noumea suggests that the mosquito density threshold under which dengue viruses cannot spread widely may be very low and has never been reached up to now. Moreover, mosquito populations are influenced by human behaviours and meteorological variables alone cannot account for their geographical distribution and abundance [14], [61]. At the domestic level, A. aegypti populations are also influenced by global trends in urbanization, socioeconomic conditions, and vector control efforts. For instance, the outbreak predicted in 2002 with a probability close to 90% did not occur. A possible explanation is that strong vector control policies (e.g. increased efforts to reduce mosquito breeding sites and undertake human population education, development of perifocal spraying of insecticides) were undertaken in New Caledonia at the time of large dengue outbreaks in the other Pacific French overseas territories (French Polynesia in 2001, Wallis and Futuna in 2002). A relaxation in vector control efforts at the end of 2002 may have allowed the resurgence of dengue in the East coast and the spread of the virus through the archipelago during the next year.
Overall, our results suggest that the local climate had a major effect on dengue dynamics in Noumea during the last forty years. It is likely that other factors, not included in the input dataset of the models, had a lower influence on dengue epidemic dynamics. The introduction of dengue viruses may have been relatively constant, and the number of human hosts susceptible to a given serotype and of mosquito-vectors may have been always sufficient for an epidemic to occur when suitable climate conditions were met. It is likely that the susceptibility of human populations influenced the serotype involved in the outbreak and the epidemic magnitude. The variability of the length of the gonotrophic cycle, the extrinsic incubation period, and the life span of infected mosquitoes under climate change rather than the overall vector density may play a major role on the epidemic dynamics of dengue at the seasonal scale.
Although the meteorological variables contemporaneous to the epidemic season provide crucial information on local dengue dynamics as discussed above, prediction models are needed to anticipate the risk before the dengue outbreak onset and to make the model useful for health authorities in New Caledonia. In this study, we were able to build such a predictive model relying on maximal temperature and relative humidity measured in Noumea at the end of the previous year.
Biological interpretations about statistical associations between specific climatic conditions and the yearly risk of dengue outbreak in Noumea can be made in the frame of the explicative model as it uses relevant climatic variables that occur within the period of outbreak onset. The meteorological variables selected in the frame of the predictive model are tightly connected with the explicative meteorological variables (Figure S5).
As Noumea concentrates the majority of inhabitants and of dengue cases, as this city has been affected by all dengue outbreaks that occurred in New Caledonia during the last 40 years, and as dengue epidemics usually begin in Noumea, our predictive model is useful to anticipate the risk of dengue outbreak in New Caledonia. However, climatic conditions in Noumea can not account for dengue epidemics in other localities in New Caledonia that would not involve Noumea, even if this situation has never been observed in 40 years.
Depending on the user's objectives, different detection thresholds corresponding to a probability of dengue outbreak can be used. In the case of dengue, it is likely that decision makers would prefer to choose a detection threshold with high true positive rate and low false positive rate, as obtained with a detection threshold of 65% (Figure 7b). The model initialized in December 2009 indicated no risk of dengue outbreak for 2010 that was in accordance with the current epidemiological situation. To note, a high risk of dengue outbreak is predicted for 2011 (74%, Figure 9). Up to now, only a few cases of dengue fever have been reported. Only one case imported from the Philippines was possible to type and belonged to the serotype 1. It is likely that a significant part of the human population is immunized against the serotypes 1 and 4 involved in the largest dengue outbreaks reported in New Caledonia in 2008 and 2009 but the introduction of a new serotype (DENV-2 or DENV-3) may lead to another epidemic. However, several important confusing factors may interfere with dengue dynamics this year such as the massive rainfalls brought by the tropical cyclone Vania in middle January 2011 with its unknown effects on vector populations, the introduction and worrying local diffusion of Chikungunya viruses transmitted by the same mosquito and the subsequent enhancement of vector control policies.
In conclusion, the epidemic dynamics of dengue fever were strongly influenced by climate variability in Noumea during the 1971–2010 period. Local thresholds of maximal temperature and relative humidity have been identified with precision allowing the development of explicative and predictive climate-based models of dengue outbreak risk. The health authorities of New Caledonia have now integrated these models into their new decision making process in order to improve their management of dengue, in combination with clinical, laboratory (e.g. serotype determination), and entomological surveillance data. This work provides an example of the practical utility of research projects in operational public health fields and reinforces the need for a multidisciplinary approach in the understanding and management of vector-borne diseases. Our results provide also new insights for future experimental studies. It seems important now to study the impact of maximal temperatures exceeding 32°C and maximal relative humidity exceeding 95%, and the influence of their duration (more or less than 12 days) on the length of the extrinsic incubation period, feeding frequency and longevity of A. aegypti from New Caledonia.
The epidemic dynamics of dengue are driven by complex interactions between human-hosts, mosquito-vectors and viruses. These interactions are influenced by environmental and climatic factors that may have more or less burden according to the geographical localisation, the local climatic conditions, the vector characteristics (e.g. Aedes species and strains), the size and movements of human populations and the epidemiology of dengue. Consequently, our results can not be applied to other ecosystems. However, the methodology of analysis used in this study could be extended to other localities highly threatened by the emergence of dengue in the South Pacific, like in other tropical and subtropical countries. As global atmospheric reanalyses climate based data exist, there is hope for the development of local predictive models of dengue outbreak in countries where no reliable weather data are available.
|
10.1371/journal.ppat.1003820 | Systematic MicroRNA Analysis Identifies ATP6V0C as an Essential Host Factor for Human Cytomegalovirus Replication | Recent advances in microRNA target identification have greatly increased the number of putative targets of viral microRNAs. However, it is still unclear whether all targets identified are biologically relevant. Here, we use a combined approach of RISC immunoprecipitation and focused siRNA screening to identify targets of HCMV encoded human cytomegalovirus that play an important role in the biology of the virus. Using both a laboratory and clinical strain of human cytomegalovirus, we identify over 200 putative targets of human cytomegalovirus microRNAs following infection of fibroblast cells. By comparing RISC-IP profiles of miRNA knockout viruses, we have resolved specific interactions between human cytomegalovirus miRNAs and the top candidate target transcripts and validated regulation by western blot analysis and luciferase assay. Crucially we demonstrate that miRNA target genes play important roles in the biology of human cytomegalovirus as siRNA knockdown results in marked effects on virus replication. The most striking phenotype followed knockdown of the top target ATP6V0C, which is required for endosomal acidification. siRNA knockdown of ATP6V0C resulted in almost complete loss of infectious virus production, suggesting that an HCMV microRNA targets a crucial cellular factor required for virus replication. This study greatly increases the number of identified targets of human cytomegalovirus microRNAs and demonstrates the effective use of combined miRNA target identification and focused siRNA screening for identifying novel host virus interactions.
| Human cytomegalovirus is a prevalent pathogen. Like other herpesviruses, human cytomegalovirus expresses small regulatory RNAs called microRNAs. The focus of this study was to understand the role of these RNAs in the context of viral infection and to use this information to identify novel host factors involved in human cytomegalovirus biology. We used a biochemical approach that allowed us to systematically identify cellular genes targeted by virus microRNAs. Because the virus targets these genes, it is reasonable to propose that these genes play an important role during infection. We confirmed this hypothesis using a second screen in which we knocked down expression of a number of the identified targets of the virus microRNAs. Knock down of one of the targets, a cellular factor called ATP6V0C, resulted in an almost complete block in production of infectious virus. These data suggest that endosomal acidification is crucial to HCMV replication, and the virus targets this process by microRNA regulation.
| Human cytomegalovirus (HCMV) is a highly prevalent infectious disease, infecting greater than 30% of the population. Although normally asymptomatic in healthy individuals, HCMV infection is a significant cause of morbidity and mortality in immunocompromised populations, individuals with heart disease and recipients of solid organ and bone marrow transplants [1]–[8]. HCMV is also the leading cause of infectious congenital birth defects resulting from spread of the virus to the unborn fetus. Reactivation of virus from a latent infection, rather than primary infection, is often responsible for HCMV associated pathologies [9]–[13].
The capacity of HCMV to strictly regulate the expression of its own genes and to manipulate host gene expression is crucial to the virus's ability to replicate and its success in maintaining a persistent infection [14]. Studies in our lab and others have demonstrated that herpesviruses have evolved to encode microRNA (miRNA) genes, enabling regulation of the virus's gene expression profile as well as altering the host environment by targeting cellular transcripts. Recent reports have demonstrated roles for viral miRNAs in suppressing apoptosis, immune evasion and regulation of viral replication through targeting of both cellular and viral gene expression [15].
HCMV encodes at least 14 pre-miRNAs corresponding to a total of 27 mature miRNA species [16]–[20]. Clear functions have not been shown for the majority of HCMV miRNAs. However, these regulatory RNAs have been shown to target genes involved in viral latency, immune evasion, and cell cycle control [21]. We previously demonstrated that the HCMV miRNA, UL112-1, restricted viral acute replication through targeting of the major immediate early gene IE72, suggesting this miRNA may play a role in establishing and maintaining viral latency [22]. Others have since shown that targeting of immediate early genes by viral miRNAs may be a fundamental mechanism involved in herpesvirus latency regulation [23]–[26]. UL112-1 has also been shown to target the major histocompatibility complex class I-related chain B (MICB) resulting in reduced killing by NK cells [27].
Despite these advances, identification of miRNA targets remains challenging. Until we have a greater understanding of the rules governing miRNA target interaction, bioinformatic strategies alone continue to produce unreliable results, especially for viral miRNAs, which in most cases do not display significant evolutionary conservation.
Biochemical approaches have provided an alternative means for the identification of miRNA targets. One such approach, RISC immunoprecipitation (RISC-IP) has proved effective in identifying both cellular and viral targets [28]. Recently, we used a RISC-IP approach to identify multiple cellular targets of US25-1, an HCMV miRNA expressed at high levels during acute infection [29]. Here we use a combined approach of RISC-IP profiling in infected cells combined with focused siRNA screening to identify host targets of HCMV miRNAs that have significant effects on virus replication. Our results, using a laboratory strain as well as a clinical strain of virus greatly increases the number of identified and validated HCMV miRNA targets. Furthermore, the results show that the V-ATPase complex, involved in acidification of endosomal compartments, is essential for HCMV virus replication and is targeted by the HCMV miRNA US25-1.
RISC-IP techniques have recently been used to identify targets of viral miRNAs [29], [30]. The approach relies on the stable interaction of the miRNA associated RISC protein complex with the targeted transcript. Following lysis of cells the RISC complexes are immunoprecipitated using direct antibodies that recognize Argonaute 2. RNA is then isolated, labeled and analysed by microarray to identify transcripts which are significantly enriched due to miRNA targeting. In a previous study we used this technique to identify targets of a single miRNA, US25-1, in the context of HEK293 cells [29]. Here we used the same basic approach to identify targets in primary human fibroblast cells infected with either the laboratory adapted AD169 strain of HCMV or the clinical strain TR. In both cases cells were infected at a high multiplicity of infection (MOI) of three and cells harvested three days post infection. Following lysis and immunoprecipitation, RNA was isolated by trizol extraction and analysed by microarray using the Illumina HumanRef-8 platform which contains probes for approximately 24,000 well annotated genes (Figure 1A). In addition to uninfected cells, RNA was analysed from infected cells immunoprecipitated with pre-immune serum instead of anti-Argonaute 2 antibody. To determine the level of enrichment, lysate was sampled before immunoprecipitation to establish total levels of transcript expression. Enrichment was then calculated as the transcript level of the IP sample divided by the total RNA sample. To determine transcripts specifically targeted by HCMV miRNAs, the level of enrichment from infected samples was divided by the level of enrichment in uninfected samples. However, infection with HCMV results in significant perturbation of total levels of many cellular transcripts through mechanisms unrelated to miRNA expression. False positive enrichment attributed to viral miRNA targeting can therefore occur due to down regulation of total RNA levels in the infected sample, where IP background levels remain relatively unchanged. To overcome this, a correction calculation was introduced using the results from the control IP using the pre-immune serum pull down. As enrichment values in this sample would only occur through changes in total levels due to AD169 infection, rather than any effective enrichment through specific immunoprecipitation, false enrichment could be effectively subtracted from the data sets generated with anti-Argonaute 2 pull downs. Example calculations are shown in supplemental figure S1.
The results indicate that greater than 96% of transcripts showed little or no enrichment in infected cells compared to uninfected cells, as would be expected if virus miRNAs are targeting a specific subset of transcripts (Figure 1B). In cells infected with AD169, 686 transcripts were enriched two fold or more with enrichment levels as high as 28.9 fold for the top target ATP6V0C. Enrichment levels were slightly lower for TR infected cells with 442 genes enriched 2 fold or higher with the highest level of enrichment for COMMD10 at 19.8 fold (Figure 1C). The lower enrichment in TR infected cells was expected as the clinical strain replicates less efficiently than AD169 in primary fibroblast cells, resulting in lower levels of miRNA expression (data not shown). Given that HCMV miRNAs are completely conserved between TR and AD169, with the exception of miR-148D-1, which is deleted from AD169 due to genome rearrangements, a similar suite of enriched genes would be expected from each pull down experiment. Indeed, 222 of the 442 genes enriched by two fold or more in cells infected with TR, were also enriched in the AD169 sample. This is a highly significant level of correlation (P = 3.2×10−233 as determined by hypergeometric distribution analysis) (Figure 1D) that validates the biological reproducibility of the system. Of the top 30 most highly enriched transcripts from AD169 infected cells, 26 were also enriched at least two fold in pull downs from TR infected cells, giving a high level of confidence that these genes are specifically enriched due to HCMV miRNA targeting. Table 1 lists the top 30 most highly enriched genes from cells infected with AD169 with corresponding enrichment values for TR. The complete data sets are shown in supplemental tables S1 and S2.
As pull down experiments were performed in the context of viral infection, any one, or a combination of, HCMV encoded miRNAs could target the identified transcripts. It is also possible that transcripts could be enriched through targeting by an induced cellular miRNA. Target interaction between miRNAs and transcripts rely heavily on binding between the 5′ end of miRNAs, specifically nucleotides 1 through 8, known as the seed sequence. To define which HCMV miRNA has the potential to target the identified transcripts we predicted seed match sites using the online algorithm RNAHybrid for the 14 most abundant HCMV miRNAs. Stringent parameters of full Watson Crick base pairing with bases 1–7 or 2–8 were employed with the top 30 putative targets analysed. All but three of the transcripts contained at least one seed match to the major HCMV encoded miRNAs, with most transcripts containing targets sites for multiple HCMV miRNAs (Figure 2). The majority of target sites reside within the open reading frame of the transcripts with only 14 of the 30 transcripts containing predicted seed matches for HCMV miRNAs within the 3′UTR. Although there is evidence that cellular miRNA targeting is heavily constrained to the 3′UTR region of transcripts [31]–[33], a number of studies, including our own, suggest that these constraints may not always be applied to viral miRNA targeting. In fact targeting by US25-1 was shown to predominantly target sites within the 5′UTR [29]. Full analysis of transcripts is shown in Supplemental Table S3.
To delineate the miRNA target interactions, we compared the RISC-IP data from infected cells with the previous published study generated from cells transfected with US25-1 [29]. Figure 3A represents heat map analysis comparing enrichment profiles from the top 30 enriched transcripts of cells infected with AD169 or TR, with cells transfected with either a plasmid expressing US25-1 or immunoprecipitations using a synthetic US25-1 mimic containing a biotin moiety. The majority of transcripts show clear enrichment with AD169 and TR as would be expected. In addition, highly enriched genes from infected cells were also significantly enriched in cells only expressing US25-1, demonstrating that these transcripts are targeted by US25-1 in the context of viral infection. Six genes, including the top target from the previous study, CCNE2, and the top target from this study, ATP6V0C, were in the top 30 enriched genes from cells infected with AD169 or cells transfected with US25-1 (Figure 3B and C).
Although the combined data sets provide strong evidence that these six genes are targeted by US25-1 in the context of virus infection, it is possible that other viral miRNAs may also target these genes, potentially complicating further validation. To determine whether this was the case, additional RISC-IP analysis was carried out comparing wild type AD169 virus to two recombinant AD169 viruses in which either US25-1 had been deleted, or the entire US25 region, encoding both US25-1 and 2, was deleted. Enrichment levels for each of the six genes identified from the previous analysis was determined by direct RT-PCR using specific primer probe sets (Figure 4). To allow direct comparison, enrichment values for wild type pull downs were set at 100% (actual enrichment values are displayed above each bar for reference). All six targets showed significant enrichment in infected cells compared to uninfected cells, validating the results from the original microarray experiments. Four of the six genes, ATP6V0C, CCNE2, BCKDHA and LGALS3, also showed a near complete loss of enrichment from cells infected with either US25 knock out viruses, indicating that US25-1 is required for enrichment of these transcripts. The results were less clear-cut for NUCB2 and SGSH. Although the levels of enrichment were reduced, the reduction was not statistically significant, suggesting that other viral miRNAs may be involved in targeting these genes. Additional validation of genes from the top 30 most enriched transcripts also showed significant enrichment in infected samples compared to uninfected samples, again validating the original microarray data (Supplemental Figure S2). However, no other genes showed a complete loss of enrichment in the knock out viruses. Transcripts that were not predicted to be targeted by US25-1, such as LIN28B, showed enrichment in both wild type and knock out virus, confirming that successful enrichment from the knock out virus infected cells had occurred. In addition no enrichment was detected in control transcripts such as beta actin (data not shown). No significant enrichment was observed following transfection with mimics corresponding to US25-2-3p, US25-2-5p or a US25-1 mimic containing a mutated seed region, demonstrating that neighboring miRNAs do not play a role in the enrichment of the six identified targets and that the seed region of US25-1 is necessary for effective enrichment (Supplemental Figure S3A).
Although these results confirmed that US25-1 RISC complexes bind to the identified transcripts it remained necessary to determine whether these interactions were functional and resulted in effects on gene expression. In our previous study we demonstrated that targeting of CCNE2 by US25-1 resulted in reduced cyclin E2 expression and conversely deletion of US25-1 from the virus resulted in increased expression of cyclin E2 in the context of virus infection. Using the same approach the effect of US25-1 on the expression levels of all six identified genes was investigated. Primary human fibroblast cells were infected at high MOI with either wild type AD169 or US25-1 KO virus and protein levels for the six genes compared by western blot analysis (Figure 5A). Uninfected cell lysates were also included as well as cells transfected with a siRNA specific for the target gene, to confirm the specificity of the antibody. As has been shown before, CCNE2 levels were higher in the knock out virus infected cells compared to the wild type infected cells. In addition, ATP6V0C, BCKDHA, LGALS3 all showed increased levels of expression in cells infected with the US25-1 knock out virus, whereas NUCB2 and SGSH did not show significant difference between wild type infected cells and knockout infected cells. Representative western blots are shown in Figure 5A with direct quantitation from three biological repeats shown in Figure 5B. These results correspond well with the RISC-IP data, indicating that ATP6V0C, CCNE2, BCKDHA and LGALS3 are targeted by US25-1 and deletion of this miRNA results in near complete abrogation of the inhibitory effects, whereas NUCB2 and SGSH may be targeted by additional viral miRNAs, or other virally induced mechanisms. Further supporting the role of US25-1 in targeting of these cellular genes, transfection of fibroblast cells with US25-1 mimic RNA results in significant knockdown at RNA levels of the putative US25-1 targets. The exception to this is SGSH, where transfection of US25-1 reproducibly resulted in a two-fold increase in total RNA levels. Whether this is due to a direct effect of US25-1 binding to the SGSH transcript or secondary effects from regulation of other US25-1 targets is unclear and will require further study. In addition, transfection of US25-2-3p and US25-2-5p did not result in significant knockdown of ATP6V0C, again supporting the role of US25-1 alone in targeting these genes (Supplemental Figure S3B).
Our previous study showed that US25-1 predominantly targets the 5′UTR of transcripts and five of the six genes (including the previously identified CCNE2) have potential target sites within the 5′UTR for US25-1 (Supplemental Figure S5). However, ATP6V0C contains a 7mer target site for US25-1 downstream of the 5′UTR within the open reading frame. To determine whether the US25-1 target within the open reading frame is responsible for the observed knockdown in ATP6V0C protein expression we carried out luciferase assays using a construct containing the target site from ATPV0C cloned into the 3′UTR of the reporter construct psiCheck2 (Figure 6A). The construct was co-transfected into HEK293 cells with either US25-1 mimic, or a non-targeting control siRNA. A CCNE2 luciferase construct was included as a positive control. US25-1 mimic induced a significant reduction in luciferase expression, compared to the negative control siRNA for both constructs (Figure 6C). Furthermore, mutation of the ATP6V0C target seed region to a BamHI restriction site resulted in restoration of the luciferase activity, indicating that the target site identified in ATP6V0C is both sufficient and necessary for US25-1 specific inhibition of gene expression. Transfection of a US25-1 mimic with a mutated seed sequence that corresponds to the mutated ATP6V0C luciferase construct did not reduce expression of the wild type ATP6V0C luciferase construct, but did inhibit expression of the mutated ATP6V0C construct (Figure 6B and C).
Although it is clear that US25-1 targets and regulates the identified genes in the context of viral infection it is less clear whether these targets are functionally relevant. It is possible that viral miRNAs target many genes, but only a few are important to the virus while other targets represent fortuitous or irrelevant targets in the context of infection. Previous studies have established that cell cycle control and expression of cyclin E proteins are intimately involved in HCMV biology. However, the potential role of the remaining 5 targets is unclear, as they have not been reported as host factors involved in HCMV replication. To investigate their potential role, replication of HCMV was analysed following siRNA knockdown of each of the individual US25-1 targets in primary human fibroblast cells. Knock down of each gene was confirmed by RT-PCR (Supplemental figure S4A). Cells were infected, post siRNA transfection, at an MOI of 1 with the clinical strain TB40E, which expresses GFP fluorescence protein under control of the SV40 promoter. This allows continuous monitoring of virus levels through GFP fluorescence. As can be seen from Figure 7A, knockdown of SGSH resulted in a modest increase in virus replication. In contrast knockdown of ATP6V0C resulted in significant reduction in virus replication at all time points. To rule out the possibility that the effects on virus replication were caused by artifactual or non-targeting effects, the assay was repeated using three additional independent siRNAs targeting different regions of the ATP6V0C transcript (Figure 7B). All three siRNAs resulted in the same reduction of GFP fluorescence. To determine the effect on production of infectious virus, plaque assays were conducted following transfection of fibroblast cells with siRNA pools targeting ATP6V0C, SGSH or a negative control siRNA. The results support and confirm the GFP screen with a modest but statistically significant increase in replication in cells transfected with SGSH siRNA (Mann-Whitney U Test: p = 0.0039) and a more dramatic reduction in virus production in cells transfected with ATP6V0C siRNA (Figure 7C). In fact, knock down of ATP6V0C resulted in almost complete block in virus production, indicating that expression of ATP6V0C is essential for HCMV virus production and suggests that acidification of endosomal compartments is required for HCMV acute replication. Cell viability assays demonstrate that the reduction in virus replication was not due to cellular toxicity caused by ATP6V0C knockdown and transfection of the small RNAs did not induce an interferon response (Supplemental Figure S4B and C). However, previous reports have indicated that ATP6V0C may have functions independent of endosomal acidification [34]. To determine whether the observed inhibition of virus replication is due to a defect in endosomal acidification, fibroblast cells were transfected with siRNAs targeting ATP6V1A and ATP6V1H, components of the same vacuolar ATPase complex. Disruption of any of the essential components has been shown to be sufficient to destabilize the complex. Knockdown of either ATP6V1A or ATP6V1H resulted in a similar reduction in HCMV replication compared to cells in which ATP6V0C had been knocked down (Figure 7D). These results support the conclusion that acidification of the endosomal compartments by V- ATPase is essential for efficient HCMV replication and this gene is targeted by the HCMV miRNA US25-1.
Despite recent advances in our understanding of miRNA transcript interaction, identification of valid targets remains challenging. The nature of miRNA targeting, where functional effects may rely on multiple miRNAs targeting a single transcript or multiple genes within single pathways being targeted, requires a system wide approach to elucidate the functions of miRNAs. Recent studies have used such approaches to identify targets of gamma-herpesvirus miRNAs [30], [35]–[38]. However, no systematic screening approach has been presented for HCMV in the context of viral infection. Here we use a RISC-IP approach to identify putative targets of HCMV miRNAs in the context of viral infection, an important step towards generating a global understanding of the role these small regulatory RNAs play in the biology of HCMV and herpes viruses in general.
Using a laboratory strain of HCMV and clinical strain we identified a total of 906 transcripts that were enriched by at least two fold over immunoprecipitations from uninfected cells, 222 of which were enriched by both viruses. Relatively few cellular targets of HCMV miRNAs have been previously published [27], [39], [40]. Of those, BclAF1 and RANTES did not show significant enrichment in infected cells. In the case of RANTES and BclAF1 it is possible that the effects are cell type specific or the complex formed between the transcript and RISC is not stable and therefore does not result in enrichment. MICB was significantly enriched in both uninfected and infected cells correlating well with previous studies indicating that both cellular and viral miRNAs target this gene.
Many of the targets identified in this study have not previously been linked to HCMV, and in many cases, have not been linked to virus infections in general. Only a fraction of host genes have been investigated for potential roles in viral infections. Systematic analysis of viral miRNA targets can effectively exploit target identification for the discovery of novel host factors that play important roles in the biology of HCMV. Here we verify the effectiveness of this approach with the identification of at least two genes that have significant effects on HCMV replication. Knockdown of ATP6V0C resulted in attenuation of viral replication, while knockdown of SGSH resulted in an increase in viral replication.
The most highly enriched target identified in this study, ATP6V0C, is a component of the Vacuolar ATPase, which is responsible for acidification of endosomal compartments [41]. Knockdown of this gene resulted in striking inhibition of virus replication with almost no infectious virus detected during growth curve analysis. Acidification of endosomes has previously been shown to be required for HCMV entry into endothelial and epithelial cells through receptor mediated endocytosis. However, infection of fibroblast cells occurs through direct fusion with the plasma membrane and has been demonstrated to be pH independent [42]. The attenuation of HCMV replication through siRNA targeting of ATP6V0C is therefore unlikely to be due to a defect in viral entry. In support of this, although GFP levels were reduced in siRNA knockdown experiments, all cells were clearly GFP positive 24 hours post infection (Supplemental Figure S6). An alternative explanation could involve the marked reorganization of intracellular membranous organelles during the formation of HCMV assembly compartment [43]. A block in endosomal acidification may interfere with this process resulting in attenuation of virus replication and virion assembly. Interestingly, a previous report indicated that US25-1 expression has a negative effect on acute replication of HCMV [44]. This effect was not specific, as adenovirus replication was also inhibited, suggesting targeting of a cellular factor was responsible for the phenotypic effects. Our findings suggest this cellular factor may be ATP6V0C and acidification of endosomal compartments may be a necessary process for efficient replication of DNA viruses in general.
SGSH is involved in heparin sulphate degradation in the lysosomal compartment. Initial attachment of HCMV virions to target cells has been shown to occur through binding of viral glycoprotein B with heparin sulphate moieties on the cell surface [41]. However, it is unlikely that disruption of this pathway would result in higher levels of heparin sulphate on the cell surface. Western blot analysis in this study shows that infection with HCMV results in significant reduction in SGSH levels, and although this reduction appears to occur independently of US25-1, the result suggests targeting of this gene plays an important role in the replication of the virus.
The question remains as to why the virus would target a cellular gene, such as ATP6V0C, required for efficient replication. We previously demonstrated that UL112-1 attenuates HCMV replication through direct targeting of the immediate early gene IE72 and suggested that this represents a mechanism of establishing or maintaining viral latency [22]. Targeting of ATP6V0C may represent a similar mechanism, possibly blocking assembly and release of virions during latent infection. Alternatively, targeting by US25-1 may be unrelated to viral replication, but rather serve a different function such as immune evasion. Acidification has been shown to be required for efficient signaling by endosomal resident toll like receptors and for efficient MHC class II presentation [45], [46]. Blocking acidification of endosomes through targeting of ATP6V0C may be an effective way for the virus to interfere with both innate and adaptive immune response.
In conclusion, this study greatly increases the number of putative and validated targets of HCMV miRNAs. The use of systematic miRNA target analysis with focused siRNA screening is an effective strategy for the identification of novel host virus interactions. Finally the V-ATPase complex is an essential host factor in HCMV replication and is targeted by the HCMV miRNA US25-1.
Normal human dermal fibroblast (NHDF) cells (Clonetics) were cultured in Dulbecco's modified Eagle's medium supplemented with 10% fetal calf serum and penicillin-streptomycin-L-glutamine. HCMV strain AD169 was obtained from the American Type Culture Collection (Rockville, Md.). TR HCMV was obtained from Dr Jay Nelson. TB40E GFP was obtained from Dr Goodrum [47]. All HCMV strains were grown on primary fibroblast cells following infection at low MOI. Virus preps were purified over 10% sorbitol gradients.
RISC-IP analysis was carried as out previously described [28], [48]. In brief for systematic analysis of HCMV miRNA targets, primary human fibroblast cells were infected at a MOI of three with either AD169 or TR. Three days post infection cells were lysed, samples taken for total RNA and miRNP complexes immunoprecipitated using anti Ago2 antibody followed by streptavidin bead pull down. RNA was isolated using Trizol and analyzed for quality using an Agilent Bioanalyzer and transcript levels determined on the Illumina HumanRef-8 platform. Microarray data was analyzed using Gene sifter software. Enrichment of specific transcripts, through association with miRNP complexes was determined by dividing the immunoprecipitated levels of transcripts by the total levels. Analysis of specific genes by RT-PCR was conducted using the same protocol and parameters, except specific primer probe sets were used instead of microarray analysis. Primer probe sets were purchased from Lifetechnologies. For mimic RISC-IPs the same procedure was followed except 293T cells were transfected with 40 nM of mimic RNA and cells were harvested 48 hours post transfection.
Argonaute specific antibody was generated by immunization of rabbits with a peptide corresponding to the N terminal region of Argonaute 2 (5-MYSGAGPALAPPAPPPPIQGYAFKPPPRPD3′).
Transcript sequences were down loaded from NCBI using RefSeq ID's. Predicted binding between HCMV miRNAs and putative target transcripts were determined using the online algorithm RNAhybrid (http://bibiserv.techfak.uni-bielefeld.de/rnahybrid/) [49]. Parameters were selected to include Watson-Crick base pairing between either nucleotides 1 to 7 or 2 to 8. Full transcript data was searched for seed sequence matches using a Java based script program.
miR-US25-1 pre-miRNA coding region was deleted from AD169 BAC clone using BAC technology as previously described [50]. Briefly a PCR amplified cassette containing FRT flanked Kanamycin was recombined into AD169 BAC genome replacing the miR-US25-1 coding region using primers listed in Supplemental table S4.
Sequence in italics indicates regions homologous to FRT flanked Kanamycin cassette with remaining sequence homologous to recombination site in HCMV genome. The Kanamycin cassette was then removed by recombining the FRT sites through inducible FLIP recombinase. The resulting BAC was isolated and electroporated into human primary fibroblast cells to produce infectious virus. Schematic representations of recombination strategies are shown in Supplemental Figure S7.
Cells were transfected with small RNAs using RNAiMAX lipofectamine reagent (Life technologies) according to manufacturer's guidelines with the following modifications. Fibroblast cells were double transfected with 20 pmol (40 nM) of small RNA per 24 well 8 hours apart. Cells were either infected or harvested 24 hours post transfection for siRNAs or 48 hours post transfection for mimics. Control cells were transfected with a non-targeting negative control siRNA (Qiagen – cat 1027310). The sequence and siRNA IDs are listed in Supplemental Table S5.
Total RNA was harvested using Trizol with concentrations and RNA quality determined by nano-drop spectrophotometer analysis. 100 ng of total RNA was DNAse treated (Promega) then reverse transcribed using high capacity cDNA reverse transcription kit (ABI). Real time PCR was carried out using gene specific primer probe sets from ABI on a Rotor gene 3000 (Corbet Research). Relative expression levels were determined by delta delta Ct calculation with levels corrected to GAPDH levels.
Human primary fibroblast cells were grown in either 10% serum supplemented DMEM before infection at a multiplicity of 3 with either wild type AD169, miR-US25-1, or miR-US25-1/2 knock out virus. 72 hours post infection, cells were harvested using SDS sample loading buffer. 30 ul of protein sample were loaded and proteins were probed using primary antibodies to ATP6V0C (Aviva), BCKDHA (Cambridge Biosciences), CCNE2 (Abcam), LGALS3 (Cambridge Biosciences), NUCB2 (Sigma), and SGSH (Genetex) according to manufacturer's specifications. Protein loading was normalised to GAPDH (Sigma). IR800 or IR680 dye conjugated anti-rabbit IgG and anti-mouse IgG secondary antibodies were purchased from LiCor. Blots were imaged using infrared fluorescence of appropriately tagged secondary antibodies and quantified using a LiCOR Odyssey scanner and software.
ATP6V0C luciferase constructs were created using custom oligonucleotides corresponding to the genomic region between nucleotides 146 and 226 downstream of the transcriptional start site, flanking the bioinformatically predicted miR US25-1 target site (GAGCGGT starting at nucleotide 186). For the ATP6V0C mutant construct, the miR US25-1 target site was replaced with a BAMHI restriction site. These inserts were cloned downstream of the renilla luciferase reporter gene of the pSicheck 2 dual luciferase construct (Promega). Cloning oligonucleotides are shown in Supplemental table S4 Luciferase constructs were co-transfected with miR US25-1 mimic or control mimic (IDT) into HEK293 cells using Lipofectamine 2000 reagent according to the manufacturer's instructions. Cells were harvested 48 hours post transfection and luciferase levels measured using Promega's dual luciferase reporter kit. CCNE2 luciferase constructs were created and assays were performed as described previously [29].
For virus growth curve analysis by GFP fluorescence 96 well plates seeded with primary human fibroblast cells were transfected with siRNAs at a final concentration of 2 nM using RNAiMAX transfection reagent (Life Technologies). Specific siRNAs for ATP6V0C (S80), ATP6V1A, ATP6V1H BCKDHA, CCNE2, LGALS3, NUCB2, and SGSH were obtained from Life Technologies. 24 hours post transfection, cells were infected at a MOI of 1. The MOI was empirically determined to provide robust signal without inducing extensive cell death through CPE. Twenty-four hours post infection cells were washed three times and overlayed with fresh complete DMEM media without phenol red pH indicator (Lonza) and GFP levels monitored using Biotech Synergy HT plate reader. For plaque assays 24 well plates seeded with HCMV were transfected with pooled siRNAs for ATP6V0C (Life Technologies) and SGSH (Thermo Scientific). 24 hours post transfection, cells were infected at an MOI of 1. 24 hours post infection cells were washed three times and at indicated time points the cell monolayer was scraped into the media and the media and cells collected and frozen. Standard plaque assays were carried out on human primary fibroblast cells overlayed with carboxy methyl cellulose.
|
10.1371/journal.ppat.1003792 | Parental Transfer of the Antimicrobial Protein LBP/BPI Protects Biomphalaria glabrata Eggs against Oomycete Infections | Vertebrate females transfer antibodies via the placenta, colostrum and milk or via the egg yolk to protect their immunologically immature offspring against pathogens. This evolutionarily important transfer of immunity is poorly documented in invertebrates and basic questions remain regarding the nature and extent of parental protection of offspring. In this study, we show that a lipopolysaccharide binding protein/bactericidal permeability increasing protein family member from the invertebrate Biomphalaria glabrata (BgLBP/BPI1) is massively loaded into the eggs of this freshwater snail. Native and recombinant proteins displayed conserved LPS-binding, antibacterial and membrane permeabilizing activities. A broad screening of various pathogens revealed a previously unknown biocidal activity of the protein against pathogenic water molds (oomycetes), which is conserved in human BPI. RNAi-dependent silencing of LBP/BPI in the parent snails resulted in a significant reduction of reproductive success and extensive death of eggs through oomycete infections. This work provides the first functional evidence that a LBP/BPI is involved in the parental immune protection of invertebrate offspring and reveals a novel and conserved biocidal activity for LBP/BPI family members.
| Vertebrate immune systems not only protect adult organisms against infections but also increase survival of offspring through parental transfer of innate and adaptive immune factors via the placenta, colostrum and milk or via the egg yolk. This maternal transfer of immunity is critical for species survival as embryos and neonates are immunologically immature and unable to fight off infections at early life stages. Parental immune protection is poorly documented in invertebrates and how the estimated 1.3 million of invertebrate species protect their eggs against pathogens remains an intriguing question. Here, we show that a fresh-water snail, Biomphalaria glabrata massively loads its eggs with a lipopolysaccharide binding protein/bactericidal permeability increasing protein (LBP/BPI) displaying expected antibacterial activities. Remarkably, this snail LBP/BPI also displayed a strong biocidal activity against water molds (oomycetes). This yet unsuspected activity is conserved in human BPI. Gene expression knock-down resulted in the reduction of snail reproductive success and massive death of eggs after water mold infections. This work reveals a novel and conserved biocidal activity for LBP/BPI family members and demonstrates that the snail LBP/BPI represents a major fitness-related protein transferred from parents to their clutches and protecting them from widespread and lethal oomycete infections.
| The existence of complex immune systems implies that interactions with pathogens represent major selective forces shaping the evolution of animal and plant species [1]. Vertebrate immune systems not only protect the adult organism against infections but also increase reproductive success through parental transfer of innate and adaptive immune factors via the placenta, colostrum and milk or via the egg yolk [2]–[4]. This maternal transfer of immunity is critical for species survival as embryos and neonates are immunologically immature and unable to fight off infections at early life stages. Parental transfer of protection has also been found in invertebrates hosting mutualists and many vertically transmitted arthropod symbionts are able to protect offspring against specific infections [5], [6]. Despite the impressive advances recently made in characterizing invertebrate immune systems [7], [8], data on the nature of the symbiont-mediated or parentally transmitted protection across generations are scarce [9]–[11]. How the estimated 1.3 million of invertebrate species [12] protect their offspring against pathogens remains therefore an intriguing question.
The freshwater snail Biomphalaria glabrata is particularly well studied as it is the intermediate host of the human blood fluke Schistosoma mansoni, responsible for schistosomiasis affecting millions of people in developing countries [13]. Biomphalaria snails live in various resting water biotopes such as, ponds, marshes, irrigation channels or open sewer drains that are particularly rich in pathogenic organisms. Egg masses are laid on solid substrates under water where they remain for approximately a week before hatching [14]. In a proteomic study on the content of B. glabrata egg masses, 16 defense-related polypeptides were partially identified, among which a lipopolysaccharide binding protein/bactericidal permeability increasing protein (LBP/BPI) representing a major protein band [15]. LBP/BPIs are structurally related proteins belonging to the lipid transfer/binding protein (LT/LBP) family [16], which represent important components of the innate immune system against Gram-negative bacterial infections [17]. In mammals, LBPs and BPIs have been extensively studied due to their role in regulating transducing cellular signals from Lipopolysaccharide (LPS) [18], [19]. LBP functions as a carrier of LPS monomers onto CD14 and together with the TLR4-MD2 receptor complex, mediates the activation of monocytes and macrophages, which produce inflammatory mediators [20]. BPI is an antibacterial protein specifically active against Gram-negative bacteria that acts by damaging bacterial membranes [21]. BPI also enhances adaptive immune responses by promoting LPS uptake and presentation to dendritic cells [22]. Although these two proteins present similarities in sequence and activities, they exert different effects on interactions of the host with Gram-negative bacteria [23]. BPI neutralizes the inflammatory properties of LPS decreasing its uptake by LBP whereas LBP is an acute phase protein with LPS-dependent cell stimulatory activity [24], [25]. These antagonist functions efficiently regulate host response to bacterial invasion and allow the host immune system to return to its normal resting state.
The distinction between LBPs and BPIs has not been established in invertebrates. LBP/BPI family members have been reported only in a few invertebrate phyla such as annelids [26] and molluscs [27], [28]. To date, a single functional study of LBP/BPI has been performed, showing that the oyster Crassostrea gigas expresses a BPI-like protein endowed with the conserved LPS-binding and bacterial permeability increasing activity [27].
As B. glabrata snails apparently heavily invest in the production of LBP/BPI in their eggs [15], we investigated whether this protein showed the expected anti-bacterial activity and whether it could provide protection against other pathogens.
We first characterized the complete coding sequence of BgLBP/BPI1 (genbank accession number KC206037) from a partial transcript that we had previously identified in an albumen gland cDNA library [29], as this organ is known in gastropod snails to produce many components of the egg masses, among which the egg perivitelline fluid [30]. Interestingly, BLAST searches against non-redundant protein databases using the BLASTp program revealed that BgLBP/BPI1 corresponded to the “Developmentally regulated albumen gland protein” (partial sequence; genbank accession number AAB00448.1) previously identified as being over-expressed in Schistosoma mansoni resistant snails [31]. The sequence of BgLBP/BPI1 displayed the typical features of LBP/BPI family members, including a N-terminal LBP/BPI domain (pfam PF01273, Interpro IPR017942) containing conserved lysines involved in the interaction with LPS, a central proline-rich domain, and a C-terminal LBP/BPI domain (Figure S1) [17], [32]. Expression studies showed that the major site of expression for BgLBP/BPI1 was the albumen gland (Figure 1, A and B). We confirmed, using a specific antibody and western blot analysis followed by mass spectrometry, that the BgLBP/BPI1 gene product is the major protein found in egg masses of B. glabrata (Figure 1, C and D).
We purified the native BgLBP/BPI1 protein from fresh egg masses and estimated its physiologic concentration around 100 µg/ml of egg mass extract, representing 60% of the total protein dry weight. In order to control for trace contamination of the purified native protein by biologically active polypeptides, we produced a recombinant protein in Drosophila S2 cell culture and compared both the native and the recombinant proteins in our assays. Plasmon resonance analysis confirmed that both BgLBP/BPI1 bind LPS and the Lipid A region, which are common to all LPS's, with a range of affinity similar to that of human BPI protein, as shown by their dissociation constants (Figure 2) [33]. In addition, both proteins showed the typical membrane permeabilizing activity leading to bacterial death (Figure 3) that LBP/BPI proteins present toward short-LPS strains of E. coli [34], demonstrating that the biocidal activity of LBP/BPI proteins against Gram-negative bacteria is conserved in BgLBP/BPI1 [21], [35], [36].
Exposure of helminths (S. mansoni), Gram-positive bacteria (Micrococcus luteus, Bacillus cereus), Gram-negative bacteria (Citrobacter freundii, and Pseudomonas aeruginosa) and fungi (Candida albicans and Saccharomyces cerevisiae) to increasing concentrations of the recombinant BgLBP/BPI1 (up to the physiological concentration of 100 µg/ml) had no significant effect on the viability of the microorganisms (P>0.1, Figure S2 and S3). With the exception of helminths such as S. mansoni, information on natural pathogens of B. glabrata is scarce. Therefore we also wanted to investigate whether oomycetes are sensitive to BgLBP/BPI1. Oomycetes or water molds are a large group of eukaryotic microbes that can infect plants and animals and can cause devastating diseases in agriculture, aquaculture, and natural (aquatic) ecosystems [37], [38]. The motile zoospores (infective stage) and cysts (germinal stage) of both Saprolegnia parasitica and Saprolegnia diclina, two well-known pathogens of fresh water fish and their eggs [38], were exposed to increasing concentrations of BgLBP/BPI1 proteins for 30 min. No effect was observed on Saprolegnia cysts (not shown), but a strong biocidal activity was observed on the infective stage of these pathogens at all protein concentrations (Figure 4). The viability of S. parasitica or S. diclina zoospores was significantly reduced to 50.3% and 68.4% or 55.8% and 39% with 100 µg/ml of native BgLBP/BPI1 and recombinant BgLBP/BPI1, respectively (Figure 4). Interestingly, human BPI also strongly decreased the viability of both oomycete species (Figure 4), revealing a yet unsuspected activity of this well-studied human immune protein [39]. To better assess the LBP/BPI-dependent anti-oomycete activity, we also tested a plant pathogenic species, Phytophthora parasitica that has the ability to form biofilms [40] like Gram-negative bacteria. Biomphalaria proteins were tested on the three stages of biofilm formation; zoospores, cysts and microcolonies (Figure S4A). Similarly to the Saprolegnia species, only zoospores were affected by the BgLBP/BPI proteins. Both the snail and the human LBP/BPI proteins showed a strong biocidal activity of zoospores from P. parasitica in a dose dependent manner (Figure 5 and S4B). The effect of LBP/BPIs was observed as early as 10 min and resulted in 100% zoospore mortality after one or two hours of exposure to physiological concentrations of 100 µg/ml LBP/BPI proteins (Figure 5), confirming the strong biocidal activity of LBP/BPIs against oomycete zoospores.
In order to assess the role of BgLBP/BPI1 in vivo, we undertook to decrease its expression by using RNA-interference mediated knock-down. Following double strand dsRNA injections, BgLBP/BPI1 protein abundance was analyzed by western blotting and showed a significant decrease in the albumen gland and in the egg masses after 12 and 18 days, respectively (Figure 6). The number of eggs per clutch collected over 28 days from parents treated with dsRNA of BgLBP/BPI1 was significantly lower than in the control experiment, whereby dsRNA of the luciferase gene was injected. Furthermore, the egg masses of the BgLBP/BPI1 dsRNA-treated snails showed a significant decrease in fecundity (Table 1). Thereby confirming that the albumen gland, the site of expression of BgLBP/BPI1, is directly involved in egg mass production [41]. After exposure to zoospores of S. diclina, eggs from parents silenced for BgLBP/BPI1 expression suffered an important decrease in their hatching rate, when compared to snails injected with control dsRNA (Table 1).
The eggs from control-treated parents appeared healthy with a normal development of the embryos (Figure 7A), whereas the egg masses from parents treated with BgLBP/BPI1 dsRNA were covered by oomycete hyphae and the resulting infection impaired dramatically the survival of the snail embryos (Figure 7B).
Many invertebrate species lay fertilized eggs in nutritive egg masses that are highly suitable to the development of microorganisms [42], [43]. Although parental protection of eggs seems crucial to the survival of species, studies on the immune protection of invertebrate eggs are scarce. For example, an antibacterial activity was shown in eggs from 32 mollusk, 2 polychaete and 1 coral species, out of, respectively 34, 4 and 11 species tested [9], [42]. An antibacterial protein, the aplysianin-A was identified from eggs of the gastropod Aplysia kurodai [44], and an N-acetyl-galactosamine-binding lectin that agglutinates bacteria was identified in eggs from the pulmonate snail Helix pomatia [45]. Together with peptides of aplysianin, peptides of LBP/BPI proteins were identified in a proteomic study on Biomphalaria glabrata egg masses [15]. Here we characterized a B. glabrata LBP/BPI family member that is produced in the albumen gland and abundantly loaded into the egg masses. Consistent with the study on the BPI-like protein from Crassostrea gigas [27], we showed that the LPS and lipid A-binding activities against Gram-negative bacteria are conserved in BgLBP/BPI1. BgLBP/BPI1s did not exert any effect against the panel of microorganism tested. However, bacterial permeability activity was observed against E. coli SBS363, a mutant strain containing short-chain LPS. This is in agreement with previous studies reporting the resistance of Gram-negative bacteria harboring long lipopolysaccharide chain to the activity of hBPI [34], [46].
In addition to this expected anti-bacterial activity, we discovered a yet unsuspected anti-oomycete activity and demonstrated that BgLBP/BPI1 is a major fitness-related protein affecting both egg production under control conditions and offspring survival in the presence of pathogens. It is possible that the positive effect of BgLBP/BPI1 on the number of eggs produced is related to the glycoprotein nature of the molecule rather than to its antimicrobial activities. Interestingly, the glycoprotein HdAGP, identified from the snail Helisoma duryi albumen gland, was reported as the major nutritive glycoprotein secreted in the perivitelline fluid, and is also sharing sequence similarities with LBP/BPIs [43]. The content in BgLBP/BPI1 may therefore affect egg production as a major nutritive egg mass compound, independently of its antimicrobial action. However, once the egg masses are laid, we demonstrated that the biocidal activity of BgLBP/BPI1 affects offspring survival in the presence of oomycete pathogens.
LBP and BPIs are pleiotropic molecules, well characterized for their interactions with LPS from Gram-negative bacteria, but also reported to interact with other organisms such as Gram-positive bacteria and fungi [18], [47], [48]. A wide range of lipidic ligands have been reported for human LBP and BPI [18], [47], [48]. Our results further evidence the diversity of binding capabilities of LBP/BPIs as both BgLBP/BPI1 and hBPI can interact with an oomycete lipidic ligand that remains to be identified. Oomycetes do share physical characteristics with true fungi, including polarized hyphal extensions but they have a distinct evolutionary history and belong to the kingdom Stramenopila, which also includes brown algae and diatoms [49]. In contrast to fungi, they produce bi-flagellated swimming spores (zoospores) and the cell-wall of their cysts is composed of cellulose, β-glucans and hardly any chitin [49]. Oomycetes include some of the most devastating animal and plant pathogens. A few species cause Saprolegniosis in the aquaculture industry [38]. Saprolegnia species are endemic to freshwater habitat worldwide and are partly responsible for declining natural populations of salmonids and amphibians [50], [51]. Furthermore, the potato and tomato late-blight pathogen, Phytophthora infestans triggered the Irish Famine in the mid-1840s [37], [50], [52]. The potent oomycete killing activity of both BgLBP/BPI1 and hBPI was observed with three species belonging to two major oomycete orders, the Perenosporales and Saprolegniales [52]. Our observations demonstrate a conserved and broad-spectrum oomycete killing activity of BgLBP/BPI1, which may be of interest for both the agricultural and aquacultural sectors [53]. Interestingly, the specificity of this biocidal activity for the zoospore stage suggests that the ligand may be expressed specifically at this developmental stage. To date, despite the economic impact of oomycetes, there is no biochemical information on the membrane compounds of zoospores as studies have focused on identifying surface components of the cell wall of cysts [54], [55].
Collectively, our results significantly expand our knowledge of the multiple functions of LBP/BPI and highlight their importance in invertebrate biology. We demonstrated that LBP/BPI proteins display a conserved, potent and so far unexpected biocidal activity against zoospores from different oomycete orders. The precise binding and killing activity of the zoospores is unknown, but it is clear that BgLBP/BPI1 represents a major fitness-related protein transferred from parents to their clutches protecting snail eggs from widespread and lethal oomycete infections.
A partial cDNA sequence (EST GenBank accession number EB709540) was used to design specific primers and perform 5′- and 3′-RACE amplification (5′3′ RACE kit, 2nd generation - Roche) according to the manufacturer's instructions. PCR products were cloned into pCR4-TOPO vector (Invitrogen) for sequencing. Sequence similarity searches were carried out using NCBI's BLAST-X program [56] against non-redundant databases with default parameters. Global sequence alignments were performed with Clustal W software [57]. The protein domains and signal peptide were predicted with the SMART [58] and SignalP [59] softwares, respectively.
Adult Biomphalaria glabrata snails (albino strain) were raised in pond water and fed leaf lettuce ad libitum according to previously described procedures [60]. Bacterial and yeast strains used in this study were Micrococcus luteus (CIP A270), Pseudomonas aeruginosa (PA14) [61], Bacillus cereus (ATCC 11778), Citrobacter freundii (ATCC 8090), Candida albicans (a pathogenic strain isolated in patient no. 3 by Pr M. Koenig, CHU Strasbourg-Hautepierre) and Saccharomyces cerevisiae (Bioreference Laboratory – Institut Pasteur (Lille, France) as well as the E. coli SBS363, a Trp+ galU129 (truncated LPS) derivative of E. coli K12 strainD22 (gift from D. Destoumieux-Garzón, Université Montpellier 2). Bacterial strains were maintained in LB medium at 37°C and yeast in YPD medium at 28°C under standard conditions.
S. mansoni miracidia (swimming infective stage) were hatched from eggs axenically recovered from 50-days infected hamster livers according to previously described procedures [62]. Oomycete species used in the study were Saprolegnia parasitica, S. diclina and Phytophthora parasitica. Saprolegnia zoospores were obtained as described previously [63]. The average number of zoospores released was approximately 104 zoospores per ml. Phytophthora parasitica was grown in 90 mm-diameter Petri dishes on 20% V8 agar media (350 ml V8 juice, 5 g CaCO3, 3.5 g agar) at 25°C for 6–8 days under continuous light. To induce zoospore release, Phytophthora isolates were placed at 4°C for 30 min. Mycelial cultures were then flooded with 10 ml of warm sterile water and left at 28°C for 30 min [40]. The average number of zoospores obtained was approximately 106 zoospores per ml.
Snail organs or tissues, namely albumen gland, hepatopancreas, headfoot, digestive tract and gonads were dissected under a binocular microscope, pooled from 10 individuals and frozen in liquid nitrogen. Snail circulating hemocytes were recovered from hemolymph collected prior to tissue dissections according to previously described procedures [62]. Total RNA and protein were simultaneously isolated using TRIZOL LS Reagent (Invitrogen) according to the manufacturer's instructions. Total RNA was quantified using a NanoDrop Spectrophotometer ND-1000 (Thermo Scientific). For cDNA synthesis, 50 ng of RNA from dissected tissues and hemocytes were used for reverse transcription using iScript cDNA Synthesis kit (Bio-Rad) and the oligo(dT) 20 primer.
The relative expression of BgLBP/BPI1 was monitored using Quantitative Real-time PCR on a DNA engine opticon 2 system (Biorad). Primers specific for B. glabrata ribosomal protein S19 (Genbank accession number CK988928) [64], elongation factor EEF1- α (Genbank accession number ES482381.1) and BgLBP/BPI1, were designed with primer 3 software and used for amplification in triplicate assays. The PCR cycling procedure was as follows: initial denaturation at 95°C for 10 min, followed by 40 cycles of amplification 95°C for 30 s, 60°C for 30 s and 68°C for 30 s for signal collection in each cycle. To assess the specificity of the PCR amplification, a melting curve analysis of the amplicon was performed at the end of each reaction and a single peak was always observed.
To examine the distribution of BgLBP/BPI1 protein in snail dissected tissues or egg masses, an anti-BgLBP/BPI1 antiserum was produced in a rabbit using the LAKAHIEKNRLIPDLLSYD and AQDKPGAVLRLNQEALDYGSR peptides and the polyclonal sera were purified using a peptide linked resin column (Proteogenix). Total protein contents of tissues were first determined by the BCA method (BC assay kit, Uptima) using albumin as a standard. 15 ug of tissue or egg mass proteins were loaded onto 10% SDS-PAGE gels and either silver stained using standard procedures, or transferred to a PVDF membrane (0.2 mm pore size) using a semi-dry blotting system. Western blots were performed using the custom anti-BgLBP/BPI1 antisera (Proteogenix).
cDNA corresponding to the open-reading frame of BgLBP/BPI1 was ligated into the pMT/V5/His-A expression vector (Invitrogen). The Drosophila expression system with Schneider 2 (S2) cells (Invitrogen) was used to express recombinant C-terminally His-tagged full-length BgLBP/BPI1 (rBgLBP/BPI1) as described previously [65], [66]. Briefly, S2 cells were transiently transfected by calcium phosphate method with 1 µg of pMT/BgLBP/BPI1/V5/His-A vector and its expression was monitored by SDS-PAGE and western blotting after 3 days of induction with CuSO4 (500 mM). After confirmation of transient rBgLBP/BPI1 expression, stable cell lines were generated performing co-transfections along with 0.1 µg of PJL3 selection vector and 1 µg/ml puromycin. Establishment of stable cell lines and production of rBgLBP/BPI1 were carried out as described previously [66]. Nickel (II)-based immobilized metal affinity chromatography (Qiagen) in native conditions was performed to purify the recombinant BgLBP/BPI1 protein according to the manufacturer's protocol.
The native BgLBP/BPI1 protein (nBgLBP/BPI1) was purified from two-days old B. glabrata egg masses. Egg masses were homogenized in 20 mM acetate buffer, pH 4.5. The crude homogenate was centrifuged at 13000 rpm for 10 min to remove the gelatinous and solid debris. Supernatant containing nBgLBP/BPI1 was loaded onto SP Trisacryl M cation-exchange resin (BioSepra) equilibrated in 20 mM acetate buffer, pH 4.5. After washing 3 times with equilibration buffer, nBgLBP/BPI1 was eluted with 1 M NaCl, 20 mM acetate buffer, pH 4.5 and quantified by Bradford method. In order to analyze the sequence of the nBgLBP/BPI1, the purified protein was excised from a 10% SDS-PAGE gel and subjected to a MALDI TOF/TOF-MS analysis (Proteomic facility, University of Strasbourg, France). Protein identification was performed by subjecting the m/z values to Mascott software at an adjusted peptide mass tolerance of 50.000.000 ppm and/or 0.5 Da and at a fragment mass tolerance of 0.4 Da. For the subsequent activity assays, the purity of the purified nBgLBP/BPI1 protein was assessed after SDS-PAGE and silver staining. Assessments of both the egg mass volumes used for purification and the final concentration of the purified protein allowed to determine that the natural concentration of nBgLBP/BPI1 is in the range of 100 µg/ml of fresh 2 days old egg masses.
Binding of LPS or lipid A to rBgLBP/BPI1 and nBgLBP/BPI1 was assessed with a Biacore 3000 system (Biacore, GE Healthcare). RBgLBP/BPI1 and nBgLBP/BPI1 were immobilized at 7000 response units (RU) onto an activated CM5 sensor chip (Biacore) according to the manufacturer's instructions. Human BPI (hBPI - Athens Research, USA) and BSA proteins were immobilized using the same conditions as positive and negative control proteins, respectively. An activated and blocked flow-cell without immobilized ligand was used as a reference to evaluate nonspecific binding. HBS-EP running buffer (10 mM HEPES, 150 mM NaCl, 3 mM EDTA, and 0.005% Tween 20, pH 7.4) was used for sample dilution and analysis. Purified diphosphoryl lipid A from E. coli F583 Rd mutant and LPS from E. coli O26:B6 (Sigma) were sonicated 15 min at 25°C and injected at various concentrations. LPS was diluted at 50, 100, 250, 500, 1000 and 2000 nanograms and lipid A at 30, 60, 100, 150, 200, 300, 400, 500 nanograms and passed over the sensor chip at a flow rate of 50 µl/min. Regeneration was achieved with two washes of 20 mM NaOH for 5 min and 150 mM NaOH for 5 min for LPS and lipid A, respectively. Sensor chip was finally equilibrated with HBS-EP buffer for 2 min. All analyses were done at a constant temperature of 25°C. Data analysis was performed after subtraction of the uncoated flow-cell values by using BIAevaluation software version 4.1 (BIAcore). The association and dissociation phases of all sensor-grams were fitted globally. Kinetic parameters were then determined using a 1∶1 Langmuir binding model.
The effect of proteins on the permeability of bacterial membranes was determined by flow cytometry using E. coli SBS363 and the LIVE/DEAD BacLight Bacterial Viability Kit (Molecular probes). This kit enables assessment of bacterial viability based on membrane integrity by differentiating between bacteria with intact and damaged cytoplasmic membranes [67]. Bacterial culture in mid-logarithmic phase was adjusted to an optical density of A600 = 0.003 with poor-broth nutrient medium and treated with 10, 30, or 100 µg/ml of rBgLBP/BPI1 and nBgLBP/BPI1. BSA and hBPI were used at similar concentrations as a negative and positive control, respectively. Samples were incubated 1, 2 and 6 h at 28°C under vigorous shaking. Bacterial suspensions were stained with LIVE/DEAD BacLight staining reagent mixture (SYTO 9 and propidium iodide - PI) as described by the manufacturer. Staining was allowed for 5 min at room temperature in the dark. Flow cytometric measurements were performed on a FACSCanto II flow cytometer (BD Biosciences) with a 488 nm argon excitation laser. A total of 60,000 events were acquired and analyzed in each sample, using BD FACSDiVa software version 6.1.3 (BD Biosciences). Results are displayed as a percentage of permeabilized cells with respect to the negative control. The experiments were carried out three times independently.
The antibacterial activity was tested on Micrococcus luteus, Bacillus cereus, Citrobacter freundii, and Pseudomonas aeruginosa using the liquid growth inhibition method as previously described [68]. For antifungal activity assays, a similar liquid growth inhibition assay was performed using YPD medium. Microbial growth was controlled by measurement of the optical density at A600 after 6, 16 and 24 h incubation in proteins at 10, 50, 100 ug/ml. Fungal growth was additionally evaluated at 48 h. The percentage of growth (% growth) was deduced from the absorbance (OD) at 600 nm as previously described [69].
The effect on S. mansoni viability was tested on groups of 10–15 miracidia placed in 24-well plates. Miracidia were exposed to proteins at 10, 50, 100, 200 µg/ml at 28°C. Microscopic observations were performed at 30 min, 1, 2, 4, 6, 8 and 24 h of incubation.
The anti-oomycete assays were adjusted to the characteristics and life-cycles of the oomycete species. Zoospores of S. parasitica and S. diclina were adjusted to 10000 cells/ml and exposed to increasing concentrations (5, 10, 30, 100 µg/ml final concentration) of proteins. Because of the rapid encystment of zoospores into cysts, assessment of mortality was only performed at 30 min (prior to encystment of live zoospores). Zoospores of Phytophthora parasitica were adjusted to 500000 cells/ml and exposed to identical concentrations of proteins (5, 10, 30, 100 ug/ml final concentration). Microscopic observations of the number of live and dead zoospores were performed after 10, 30, 60, 120, 240 min of treatment. Zoospore-treated suspensions were stained immediately after incubation with the Live/Dead Cell Assay kit (Abcam) as described in the manufacturer's protocol. All conditions were tested in triplicate and assays were performed 3 times independently.
BgLBP/BPI1-specific primers containing a T7 promoter sequence were designed to amplify a 420-bp region of BgLBP/BPI1 used as a template for double stand RNA (dsRNA) synthesis according to manufacturer's instructions (MEGAScript T7 kit, Ambion). The firefly (Photinus pyralis) luciferase gene dsRNA (pGL3 vector, Promega) was produced and used as a non-relevant dsRNA control.
Each dsRNA (15 µg in 10 µl of sterile Chernin's Balanced Salt Solution - CBSS) was injected into the cardiac sinus of individual snails using a 10 µl Hamilton syringe with a 26 s needle [62]. A second injection was performed 12 days after the first dsRNA injection in order to optimize the knock-down efficiency. Groups of 10 snails were injected either with BgLBP/BPI1 or Luc dsRNA and were maintained under standard conditions. Egg masses produced during 28 days following the first dsRNA injection were scored, collected and observed under a stereoscopic microscope for assessment of the number of eggs in each egg mass. Egg masses laid from day 12 to 21 after the first dsRNA injection were either maintained under control conditions, or exposed to Saprolegnia diclina zoospores at a final concentration of 104cells/ml. Egg masses were microscopically observed during the following 10 days to assess the egg hatching rate. Results are shown as the percentage of eggs hatched. RNA interference experiments were performed three times independently.
All data were expressed as mean of three independent experiments plus or minus SE. Differences in relative BgLBP/BPI1 gene expression were tested for statistical significance by one-way ANOVA and the tukey-Kramer test (Software Prism v.5.0, GraphPad). Data from membrane-permeabilizing and antimicrobial (antibacterial, anti-oomycete) assays were analyzed by the chi-square test of independence [70] between treatments (rBgLBP/BPI1, nBgLBP/BPI1, hBPI and BSA) and the proportion of dead cells, using the computing environment R [71]. Schistosoma mansoni survival curves were analyzed by the Mantel-Cox log-rank test (Software Prism v.5.0, GraphPad). Results on snail fecundity and on egg viability were statistically analyzed by the likelihood ratio test between nested models [71]. Briefly, three variables were considered in these models; the mean number of egg per snail, the mean number of egg masses per snail and the mean number of eggs per egg mass. To test the effect of the BgLBP/BPI1 silencing on snail fecundity for each of the three variables, linear mixed models were fitted with the function lmr (LML 4 package of R software). In each case the model contains the BgLBP/BPI1 silencing as fixed effect and the time and the replicates as random effects. To normalize the data the mean number of eggs and egg mass per snail were log transformed. To assess the effect of the oomycete infection on the hatching rate of BgLBP/BPI1 silenced eggs, saturated binomial generalized linear mixed models were fitted. This model contained as fixed effects the BgLBP/BPI1 silencing, the oomycete infection, the time and all interactions among these variables; and as random effects, the replicates. To account for over-dispersion, individual level of variability was added. From this model variable selection of fixed effect was based on the AICc (dredge function of MuMIn package of R software) then the selected fixed effect was analyzed by the likelihood ratio test. A P value of <0.05 was considered statistically significant. Where indicated in figures: *P<0.05, **P<0.01, ***P<0.001.
|
10.1371/journal.pbio.2006507 | The evolution of the syrinx: An acoustic theory | The unique avian vocal organ, the syrinx, is located at the caudal end of the trachea. Although a larynx is also present at the opposite end, birds phonate only with the syrinx. Why only birds evolved a novel sound source at this location remains unknown, and hypotheses about its origin are largely untested. Here, we test the hypothesis that the syrinx constitutes a biomechanical advantage for sound production over the larynx with combined theoretical and experimental approaches. We investigated whether the position of a sound source within the respiratory tract affects acoustic features of the vocal output, including fundamental frequency and efficiency of conversion from aerodynamic energy to sound. Theoretical data and measurements in three bird species suggest that sound frequency is influenced by the interaction between sound source and vocal tract. A physical model and a computational simulation also indicate that a sound source in a syringeal position produces sound with greater efficiency. Interestingly, the interactions between sound source and vocal tract differed between species, suggesting that the syringeal sound source is optimized for its position in the respiratory tract. These results provide compelling evidence that strong selective pressures for high vocal efficiency may have been a major driving force in the evolution of the syrinx. The longer trachea of birds compared to other tetrapods made them likely predisposed for the evolution of a syrinx. A long vocal tract downstream from the sound source improves efficiency by facilitating the tuning between fundamental frequency and the first vocal tract resonance.
| The larynx is an important valve in the respiratory system of all air-breathing vertebrates that is located at the upper end of the trachea. In some amphibians, in nonavian reptiles, and in mammals, it has also assumed the function of a vocal organ. In contrast, birds have evolved a new and unique vocal organ, the syrinx, which is located at the lower end of the trachea. The selective forces that underlie the evolution of the syrinx as a novel organ have remained unclear. Among all air-breathing vertebrates, birds have the longest necks, and long necks require a long trachea. With a vocal organ at the base of the trachea, this long tube can act as vocal tract resonator and, therefore, can improve the conversion of aerodynamic energy into acoustic energy if fundamental frequency and a resonance frequency are matched. Here, we conducted experiments with simplified physical models, with real birds and a computational simulation in order to investigate the effect of the two different positions of a sound source within the respiratory tract. We find that sound is produced with greater efficiency by a sound source in syrinx position and that favorable interactions between sound source and vocal tract occur with syringeal position. The data provide support for the hypothesis that a selective pressure for high vocal efficiency may have contributed to the evolution of the syrinx in its unique location within the air tract.
| Evolutionary novelty in physiological and morphological features can often be traced to specific adaptations that allow organisms to exploit the fitness landscape successfully. The avian clade is characterized by a number of striking synapomorphies, which frequently have been linked to the evolution of active flight [1]. However, many of these avian features did not arise in the context of flight, and the selective regimes that led to their evolution are often poorly understood. The unique avian vocal organ, the syrinx, is such an example [2].
In all air-breathing vertebrates, the larynx regulates airflow and serves as a valve to protect the airways from food and water [3]. In most amphibians, reptiles, and mammals, the larynx has also evolved into a vocal organ [4]. Birds possess a laryngeal valve, but as a sound source, they have evolved a novel structure, the syrinx [5–7]. Even crocodilians, the closest extant relatives of birds, produce sound with a laryngeal source [8,9] and show no modifications at the tracheobronchial juncture, which could be interpreted as precursors of a syrinx [7]. Interestingly, the phonatory mechanisms of the syrinx and larynx are remarkably similar [10], i.e., airflow sets laterally positioned vocal folds into self-sustained vibration [11–15]. The archosaurian shift from producing sound with an organ located at the cranial end of the trachea to a novel structure near the tracheobronchial juncture must have conferred a selective advantage. The nature of the selective forces leading to the formation of a syrinx is still completely unknown.
Is the location of the syrinx at the tracheobronchial juncture linked to the specialized avian respiratory system? The unidirectional airflow through reptilian and avian lungs [16] is in birds associated with the longest tracheas among vertebrates [17,18]. The volume of the avian lung does not change markedly between respiratory phases, as it does in the tidally perfused mammalian lung. Instead, the air sac system functions as bellows and perfuses the parabronchi of the lung with a mostly unidirectional stream of oxygenated air during inspiration and expiration. This flow is thought to arise from two aerodynamic valves at critical conjunctions of the mesobronchi within the lung [19–22]. Although these differences in ventilation of the lung between birds and mammals might indicate that a second air-regulating valve at the tracheobronchial juncture in the interclavicular air sac is critical for regulating airflow, there are no experimental data to substantiate this hypothesis. Birds with experimentally deactivated syringeal muscles can breathe without difficulty in a lab setting, but of course, some more metabolically demanding dynamic behaviors such as flight have not been tested directly. Irrespective of whether an air-regulating valve evolved prior to the vocal function, the switch of vocal organ from a laryngeal to the syringeal position needs an explanation.
Selective pressures related to vocal production must therefore have been in play to cause this switch in vocal organ. In this study, we ask whether the syringeal sound source provides greater vocal efficiency than a larynx because the location of the sound source leads to differences in how self-sustained vocal fold vibrations interact with vocal tract resonances. Specifically, we predict that a sound source in syringeal position converts aerodynamic energy into acoustic energy more effectively because of more favorable interactions with the upper vocal tract. We tested this hypothesis through a series of experiments, which were designed to inform our understanding of the major difference between larynx and syrinx—their location within the respiratory system. Two modeling efforts were used to test whether positioning a sound source upstream (as in the syrinx) or downstream (as in the larynx) of an elongated tube (the trachea) has a significant effect on the efficiency of sound production. The modeling approaches cannot be replicated one-to-one in vivo because a syrinx or larynx, respectively, cannot be moved freely up and down the tracheal tube. However, a complementary experimental approach is the manipulation of tracheal length to investigate the effect of tracheal resonances on the syringeal sound source. The influence of acoustic airway pressures on vocal fold vibration in the syrinx was therefore studied in bird cadavers by changing the tracheal length above the sound source by tube extensions. The in situ experiments helped to investigate the role of an interaction between sound source and length of the vocal tract filter.
In the following section, we lay theoretical groundwork for exploring how the position of a sound source within the respiratory tract determines vocal output characteristics.
Direct visualization of the sound-producing larynx [23] and syrinx [12] and experiments with an excised larynx [24–27] and syrinx [11,13,15] confirm that both organs function as self-oscillating valves driven by airflow. The primary sound then travels along the respiratory tract above the source, which for laryngeal phonation consists of oral, nasal, and pharyngeal cavities [28–30] and in birds of the tracheal tube, larynx, oropharyngeal–esophageal cavity, and beak [31–34]. An obvious, major difference between the laryngeal and syringeal design is the relative length of the airway above and below the sound source. This difference has important consequences for how the vibrating tissue of the sound source and the air column in the vocal tract interact. The air columns above and below each sound source can affect the way energy is conveyed from the aerodynamic airflow to the vibrating tissue masses.
Titze [35] used a surface-wave model to show how energy from the airstream becomes coupled to the vocal folds. The driving force for tissue vibrations in the syrinx and larynx is lung pressure. The lowest driving pressure required for triggering tissue vibration is referred to as phonation threshold pressure (pth) and provides an important estimate of the energy conversion. Phonation threshold pressure was derived as
pth=kt(ρ2)(B2Llt)2,
(1)
in which kt is a dimensionless pressure coefficient (average of about 1.1 over a vibration cycle), ρ is the air density, B is vocal fold tissue damping, L is vocal fold length, and lt is the acoustic inertance of the downstream tube. Inertance is the sluggishness (or inertia) of the vocal tract air column. As the supraglottal column of air is driven forward and backward by airflow emerging from the glottis, the sluggishness of the air column creates an acoustic pressure that helps the vocal folds in their self-sustained oscillations [28, Chapter 4]. The phonation threshold pressure thus varies inversely with the square of inertance. Greater inertance lowers the phonation threshold pressure, making it easier to produce vocal fold oscillations. For a tube that is acoustically short (i.e., much less than a quarter of a wavelength), the inertance can be expressed simply as
lt=ρLtAt,
(2)
in which Lt is the length of the tube and At is the cross-sectional area. Eq 2 shows that a longer and narrower tube produces greater inertance. When the tube is lengthened much beyond a quarter wavelength, however, standing waves can be produced due to reflections from the distal end of the tube, and the inertance then becomes frequency dependent. In fact, for some frequencies, the vocal tract may become compliant rather than inertive, in which case supraglottal pressures hinder self-sustained oscillations of the vocal folds. To maximize energy transfer, it is then important that the bird develops an appropriate length–frequency combination that produces inertance at the input of the trachea, which passerine songbirds indeed do by matching fundamental frequency and first formant frequency [32]. Tissue oscillations are maintained much more easily, i.e., with lower subglottal/subsyringeal pressure, at frequencies for which the air column is inertive rather than compliant, which generally occurs at frequencies below a resonance frequency of the tube. Tracheal lengths that are slightly below one-fourth, three-fourths, one and one-fourth, one and three-fourths, etc. wavelengths are theoretically ideal for a tube closed at one end and open on the other end. The exact boundary conditions may differ, however, with variable glottal impedance and radiation impedance.
The effect of the vocal tract on self-sustained oscillation is reversed for a subglottal/subsyringeal airway system. Fletcher [36] expanded the theory of self-oscillating valves in a tube by including valves with both lateral and longitudinal degrees of freedom (relative to the airflow and an upstream acoustic tube). For the important lateral degree of freedom in vocal fold vibration, inertance below the larynx was not favorable to vocal fold vibration, raising the threshold pressure rather than lowering it. Titze [37] showed that combining a compliant system below (i.e., upstream) the sound source with an inertive system above (i.e., downstream) the sound source provides the best assistance to self-sustained oscillation. A second benefit obtained from an inertive tube is the delay of the peak acoustic airflow through the glottis relative to the peak excursion of the lateral vocal fold movement, generating a “skewing of the airflow waveform” toward a sawtooth shape. In human subjects, these interactions between source and filter have different effects on the voice during spontaneous vocalization, including an increase in overall intensity, an increase of the higher harmonic energy, or an increase in the probability of nonlinear phenomena [38].
In summary, theoretical acoustic analysis predicts that (1) subglottal/subsyringeal inertance raises the phonation threshold pressure but increases glottal waveform skewing, the benefit and costs of which can offset each other, and (2) that supraglottal/suprasyringeal inertance lowers oscillation threshold pressure and increases glottal waveform skewing, an additive beneficial effect. Thus, a sound source deeper in the airway would appear to have an advantage, assuming that the extra energy losses in the longer transmission system do not negate the additional energy converted.
We conducted three tests to investigate whether the position of the sound source within the respiratory tract affects the primary sound production.
All procedures using birds were approved by the Institutional Animal Care and Use Committee of the University of Utah (protocol number 16–03014). The protocols are in compliance with the Animal Welfare Act regulations and Public Health Service Policy. The university maintains accreditation by the Association for the Assessment and Accreditation of Laboratory Animal Care International.
In the first study, we used a model that consists of two vocal folds constructed from silicone [39]. The silicone model was a single-layer representation of the vocal folds or of labia and membranes in a syrinx [40,41]. The model cross section was uniform in the dorsoventral direction up to the point of intersection with the tube wall. Resonance properties of the model were determined by placing the model on a small shaker, which created small amplitude vibrations from 5 to 500 Hz. The beam of a laser Doppler vibrometer (Polytec Scanning Vibrometer; Polytec, Inc.) was positioned on one vocal fold in order to measure its frequency response. A Fourier analysis of the model vibration yielded a fundamental frequency of 78 Hz.
A simple physical model was used that neither includes detailed features of syringeal or laryngeal morphology nor imitates the respective respiratory anatomies. By not including features such as air sacs and accessory elastic tissue components or musculature, we limit the number of variables to sound source position and trachea length. This provides a clear and simple test of whether one location of a sound source confers an acoustic advantage over another. Physical models such as that used here have been tested in previous studies investigating different questions related to human voice production and biomechanics [26,42–44].
The two vocal folds were mounted inside a 1-inch inner-diameter PVC ring that could be coupled to a 1-inch inner-diameter PVC tube. Blowing compressed air through a tube containing the vocal folds initiated vocal fold vibration and created sound. The sound source (i.e., the ring containing the vocal folds) was placed either at the upper (larynx) or lower (syrinx) end of a PVC tube (Fig 1). The PVC tube length was varied from 0 to 248 cm in a stepwise fashion. For the syringeal position, the sound source was placed 2 cm above a y-shaped tube simulating the bronchial bifurcation. The laryngeal sound source was placed at the downstream end of the trachea and was equipped with a short 15-cm-long vocal tract above the source.
Air was supplied from a tank with compressed air through a 5-m Silastic tube, which was connected to both bronchial tubes. Average airflow was measured 3 m upstream from the tracheal bifurcation in the Silastic tubing with a rotameter mounted in series (KING instruments company, Garden Grove, CA; maximum flow rate: 4 SCFM) and in one bronchial tube (flow meter MC-5SLPM-D-15PSIA; Alicat Scientific).
Pressure was measured below each sound source (Fig 1) and, in the case of the laryngeal sound source, at the lower end of the trachea. The latter allowed us to monitor the pressure gradient along the tracheal tube, which never exceeded 20 Pa/m. Calibrated pressure measurements were made through small (1-mm inner diameter) stainless steel tubes mounted into the wall of the PVC tubing and connected through Silastic tubing to a pressure transducer (model FHM-02PGR-02; Fujikura, Tokyo, Japan).
A calibrated microphone (GRAS Sound and Vibration, Denmark; pressure microphone 40 AG, preamplifier 26 AK and 12 AD power module) was placed perpendicular to the vocal tract opening at a distance of 10 cm. Sound, airflow, and pressure signals were recorded through a multichannel AD-acquisition board (NI DAQ). Signals were digitized at 44.1 kHz sampling rate with Avisoft Recorder software (Avisoft, Berlin, Germany).
We measured driving pressure, tracheal airflow, sound pressure level, and fundamental frequency and estimated vocal efficiency as the ratio of radiated acoustic power (Pr) over aerodynamic power (Pa):
E=PrPa.
(3)
Radiated power is a measure of the amount of aerodynamic energy converted into acoustic energy and radiated into the air per second (in Watts). Assuming spherical radiation, it can be calculated from the sound pressure level (in dB) at a radius R from the opening of the vocal tract (lips or beak in animals; open tube end in our experimental setup):
Pr=4πR2×10(SPL−120)/10,
(4)
in which Pr is the radiated power, R is measured as distance between microphone and the opening of the vocal tract tube (here, 10 cm), and SPL is the sound pressure level (dB). Aerodynamic power is derived as the product of mean flow rate and subglottal/subsyringeal pressure:
Pa=psV,
(5)
in which Pa is the aerodynamic power (Watts, W), ps is the pressure below the sound source (Pascal, Pa), and V is the mean flow rate (cubic meters per second, m3/s).
The vocal fold model, either in syringeal or laryngeal position, was coupled with 18 different tracheal lengths. The segment lengths were chosen so that the first tracheal resonance is either lower, higher, or equal to the eigenfrequency of the physical model of 78 Hz. The first and second resonances are provided in S1 Table.
Studies linking fundamental frequency and body size [45] or body size and tracheal length [18] suggest that our vocal fold model would be that of a 2 to 30 kg bird with a tracheal length of approximately 40 cm. Therefore, the simulations with very short and very long tracheal lengths are representative for birds with extreme trachea morphologies [46]. It is the range around 40-cm trachea length that resembles most realistically a bird-like situation with average tracheal length.
In a second study, we investigated the interdependency between sound source position, glottal efficiency, and vocal tract length by computational simulation. In agreement with the physical model construct, we did not make the self-oscillating sound source specific to any species or gender nor did we include any layered tissue morphology. Rather, we used a simple generic self-oscillating tissue surface model. The “vocal folds” were defined by five serially coupled sections of a soft-wall tube (1.6 mm each section in the caudal–cranial direction), giving the vibrating tissue an overall thickness of 8.0 mm. The sections had elliptical cross sections. The minor diameters (also known as the prephonatory glottal widths) were 1.0, 0.8, 0.6, 0.6, and 1.0 mm, caudal to cranial, whereas the major diameters (also known as vocal fold length, ventral to dorsal) were all 10 mm. The viscoelastic properties of the wall were patterned after the two-section model of [47], with a Young’s modulus = 4.0 kPa, a shear modulus = 1.0 kPa, mass per unit area = 0.3 g/cm2, and a damping ratio of 0.1. This produced a natural tissue frequency of 130 Hz in each section.
Fluid flow and acoustic wave propagation in all airway sections (including the source sections) were calculated on the basis of conservation of momentum and mass transfer using the Navier–Stokes and continuity equations for nonsteady, compressible airflow [48,49]. Fluid pressures on the surfaces of the five source sections provided the driving forces for self-sustained oscillation. The tracheal length was varied from 22 cm to 154 cm in steps of 22 cm. In case of the syrinx position, an additional 1.0-cm section length was added between the source and the bronchial termination, while for the larynx position, an additional 1.0-cm length was added between the source and the mouth radiation. Radiation from the mouth was computed with the piston-in-a-spherical-baffle model [49]. The cross-sectional area of the tube was length dependent, as defined below. The viscoelastic wall properties for all sections except the vocal fold sections were chosen according to Titze and colleagues [49]: Young’s modulus = 9.62 kPa, shear modulus = 1.67 kPa, mass per unit area = 1.5 g/cm2, and damping ratio = 1.26.
The power calculations were as follows:
P=1N∑n=1N(pnUn)totalpower,
(6)
PDC=pDCUDCsteadyflow(DC)power,
(7)
PAC=P−PDCacoustic(AC)power,
(8)
in which n is the time sample index, N is the number of samples simulated (22,050 in a 0.5-s window), pn is the instantaneous pressure in a given section, Un is the instantaneous flow rate, pDC is the steady (DC) pressure, and UDC is the steady (DC) flow rate, computed as time averages over the 0.5-s window. The instantaneous powers pnUn in Eq 6 varied dramatically in the vocal fold sections where self-oscillation took place. Therefore, to get a representation of the mean values of intraglottal pressure, flow, and power, the calculations in Eq 6 were averaged over the five adjacent vocal fold sections.
The efficiency was computed as the acoustic power delivered to the mouth divided by the total input power. This was different from the efficiency calculation with Eq 3 for the physical model. Mouth power gave a more accurate difference calculation in Eq 8. Radiated power was many orders of magnitude lower than the input power when the tube was more than 1.0 m long. Thus, the magnitudes of the efficiency calculations between the physical model and the computational model are not comparable, but the variations with length and source position are comparable.
Unlike in the first experiment with the physical model, tracheal diameter was adjusted with changing tracheal length. This simulates conditions in avian archosaurs more realistically. As indicated in Eq 2, not only the length but also the diameter of the downstream airway affects tissue vibration characteristics. The relationship between tracheal length (TL) and tracheal diameter (TD) was estimated with Eq 9 following published empirical data [18]:
TD=(0.24*TL)+0.1.
(9)
We used tracheal diameter to estimate vocal fold length. We assume that vocal fold length is approximately equal to tracheal diameter [45]. According to Eq 9, a 1-cm tracheal diameter would suggest a tracheal length of 41 cm in a hypothetical bird, i.e., similar to the physical model. Tracheal lengths much shorter or longer could be representative of extreme tracheal morphologies.
In a third study, we investigated the effect of tracheal length on sound production by a syrinx in situ. The approach of the in situ syrinx has been proven effective [11,50–53]. It is important to perform these experiments in situ because excised syringeal preparations may reveal unnatural vibratory behavior of the labia [54].
If the length of the airway above a sound source is a critical factor determining the acoustic output of a bird, we expect systematic changes associated with tracheal length changes. The syrinx can be phonated by blowing air into the posterior thoracic or abdominal air sac. The experiments were performed in freshly killed birds. Five male specimens from each of three species (chicken, Gallus gallus; budgerigar, Melopsittacus undulatus; and zebra finch, Taeniopygia guttata) were used. Birds were euthanized with an overdose of Ketamine/Xylazine. Compressed humidified and warm air was injected into the right posterior thoracic or caudal air sac through a Silastic tube. A microphone was placed 10 cm downstream from the cranial opening of the trachea.
We measured fundamental frequency while phonating the syrinx. Subsyringeal air sac pressure, which is proxy for the driving pressure of the sound source, was measured in the right anterior thoracic air sac by inserting a flexible cannula through a small hole in the body wall (Silastic tubing; 1.65 mm o.d., 6 cm length). The free end of the tube was connected to a piezoresistive pressure transducer (model FHM-02PGR-02; Fujikura, Tokyo, Japan).
The numerical data used in all figures are included in S1 Data.
Phonation threshold pressure varied between 1.4 and 2 kPa for the syringeal and between 1.7 and 2.7 kPa for the laryngeal position of the physical model source. Phonation threshold pressure was lower for the syrinx for all tracheal lengths tested (Fig 2A). Very high pressures were required to trigger phonation in the larynx for tracheal lengths between 36 and 86 cm (Fig 2A) when the vocal fold eigenfrequency was located to the left of the first tracheal resonance.
For a very short trachea and for a tracheal length around 2 m, the laryngeal and syringeal sound sources generated comparable sound intensities, but in all other conditions, the syringeal source emitted louder sound (Fig 2B).
Tracheal resonances influenced the vibration rate of both sound sources (Fig 2C). Increasing tracheal length was first accompanied by increased fundamental frequency, which peaked at approximately 60 cm tracheal tube length and then decreased with longer tracheae.
Glottal efficiency of the syrinx was greater across almost all tracheal lengths (Fig 2D). The difference was most dramatic in the range between 40 and 80 cm. Efficiency of the laryngeal sound source was comparatively flat across all tracheal lengths.
In sum, the findings support the hypothesis that the position of the sound source within the respiratory tract critically affects vocal parameters. For a given body size (approximated at 20 to 30 kg body mass associated with a tracheal length of 40 cm), the syringeal position is more efficient in energy conversion.
Phonation threshold pressure varied between 0.4 kPa and 1.0 kPa for the syringeal position and between 0.5 and 1.5 kPa for the laryngeal position. This pressure was lower for the syrinx for all tracheal lengths tested (Fig 3A). Fundamental frequency was categorically lower for the syringeal position. This is in agreement with analytical predictions that greater supraglottal acoustic inertance lowers F0. Glottal efficiency was categorical greater for the syringeal position than for the laryngeal position. We also tested to what degree source position affects vocal parameters with the computational model. The critical length for the 130-Hz natural frequency of oscillation of the computational model was 67.5 cm, for which the tube resonated at a quarter wavelength at 130 Hz. For all lengths below this critical length, the vocal tract acoustic reactance is inertive, which means that there is a favorable source–vocal tract interaction [37]. This favorable interaction explains why in Fig 3, in the region below 67.5 cm, the phonation threshold pressure is lower and fundamental frequency is slightly lower. Vocal tract inertance adds effective mass to the coupled oscillator system and therefore lowers fundamental frequency. Furthermore, acoustic power at the mouth is higher, and vocal efficiency is higher than in the region above the critical length. Phonation threshold pressure was on the order of 0.5 kPa in the inertive region and rose to 1.2 kPa in the noninertive region for both the syrinx and the larynx position. It reached its minimum at a tracheal length of about 40 cm (Fig 3A).
All variations with tube length and source position in Fig 3 are not as large as in the physical model. This is attributable to four important differences: (1) the wall properties of the tube, (2) a more realistic simulation of tracheal diameter, (3) the difference in vocal fold geometry and material properties, and (4) the difference in radiation from the tube. The computational model used soft walls throughout, while the physical model used a hard-wall PVC tube. The acoustic energy levels in hard-wall tubes are much greater than in soft-wall tubes, increasing the degree of interaction between the source and the vocal tract. The interaction in the computational model was also lessened by the fact that the cross-sectional area of the tube increased with length. It is well-known that vocal tract pressures are scaled by the characteristic tube impedance ρc/A, where ρ is the air density, c is the speed of sound, and A is the cross-sectional area. Greater cross-sectional area lowers all vocal tract pressures. Furthermore, the way vocal folds respond to airflow is governed by factors such as geometry, layer structure, and viscoelastic properties. While the physical and computational vocal fold models shared similar characteristics, the effects of geometric and material property differences are not fully understood. Finally, the radiation from a tube without a baffle, which was the case in the physical model experiments, may differ from radiation from a piston in a spherical baffle, which was assumed in the computation.
A total of four chickens, five budgerigars, and five zebra finches were successfully phonated. The stepwise shortening of the trachea was accompanied by an increased fundamental frequency in chickens and budgerigars (Fig 4A and 4B). In zebra finches, fundamental frequency remained relatively constant (Fig 4C). The elongation of the trachea with tubes that fitted the respective trachea was associated with a decrease of fundamental frequency in chickens but not in budgerigars and zebra finches (Fig 4A–4C). In male zebra finches, we observed nonlinear phenomena in the phonations more frequently if tracheal length resonance was near the fundamental frequency.
The results presented here provide two findings that enhance our understanding of which selective pressures might have led to the evolution of the unique avian syrinx. First, the physical and the computational models of the sound source show that sound production with a source in the syringeal position can be dramatically more efficient than in the laryngeal position. Second, sound production by the in situ syrinx is affected by vocal tract resonance. In the following, we will discuss how these results might be important for shedding light on the evolution of the syrinx.
In our physical model, the use of an identical sound source in both positions provides strong support that the additional length of the vocal tract above the sound source plays a significant role in improving vocal efficiency. Importantly, these differences between larynx and syrinx emerged in both modeling approaches and without the need for incorporating other avian specializations of the respiratory system that likely improve vocal efficiency. Most notably, the syrinx resides in the interclavicular air sac, whose pressure conditions may affect sound production by pushing the vibrating tissue into the airstream and thus likely lower the phonation threshold pressure with increased adduction.
Although the influence of airway characteristics upstream and downstream from a sound source on the sound generation mechanism has been discussed before [37,55,56] and has been most clearly documented for the human voice [38,57–60], unequivocal experimental evidence for source–tract interactions has not been presented for birds. Our data from three species show that these interactions do play a role, but their prominence may differ substantially between species. The differences are either based on different types of interaction between sound source and vocal tract or on different strengths of one type of interaction. In two species, chicken and budgerigar, fundamental frequency was tightly linked to the first tracheal resonance (F1 in Fig 4). As the first tracheal resonance increased and moved away from the natural source frequency (trachea shortened), the fundamental frequency of self-sustained oscillation increased too, i.e., it was less bent downward by the inertive acoustic load of the first resonance. In the third species, the zebra finch, no such change was observed. Instead, there was an increased occurrence of nonlinear phenomena (e.g., subharmonics and frequency jumps) when the trachea was shortened.
This interpretation is further supported by experiments with a variety of approaches, which were undertaken to determine the mechanism of sound production in birds. In various studies using different types of physical models, the vibration frequency [61–64] showed strong coupling to the first resonance of the downstream vocal tract. In situ phonation experiments in bird cadavers further support our finding that species differ in the strength and/or nature of the coupling. Whereas fundamental frequency changed around resonance frequencies of the trachea in several species—Gallus gallus domesticus [50,11,65], Grus grus [11], Meleagris gallopavo [50], Anser anser [52]—it did not to the same extent in one Great-horned owl (Bubo virginianus) [51]. Furthermore, spontaneously singing passerine songbirds tested in a heliox atmosphere did not show marked changes in frequency [66], suggesting that, by altering resonance, vibration frequency is not markedly affected.
Our data and published evidence [63] in the budgerigar show an interesting additional detail. In intact birds singing in heliox, fundamental frequency changed very little. However, once the syringeal muscles were denervated, fundamental frequency increased dramatically and followed the first tracheal resonance, suggesting that the coupling of source and tract is under neural control. In our phonation experiments on this species, we also see effects of strong coupling on sound frequency in the absence of neural control.
In principle, tracheal length, i.e., the size of the air column above the sound source, can have an effect on the vibrating structures in two ways [37,67,68,69]. “Level 1” interaction affects the glottal airflow but has no effect on vocal fold movement, whereas a “Level 2” interaction affects both airflow and vocal fold movement (Fig 5). Whenever fundamental frequency and amplitude of vibration change with vocal tract adjustments, Level 2 interaction is indicated. If only the sound intensity and spectral content change without accompanying changes in vocal fold movement, Level 1 interaction is indicated. The output of a system with Level 1 interaction shows changes in the amplitude of various harmonics. The frequency output of a system with Level 2 interaction is characterized by first bending downward as it approaches the resonance frequency from below. It then locks onto the resonance frequency and follows it if the coupling is strong. With very strong interaction, occurrence of nonlinear phenomena (bifurcations) in the acoustic output can make predictions of the frequency change more difficult [59]. Nonlinear phenomena can be evident in the glottal flow signal (Level 1 interaction, Fig 5) as well as in the tissue movements of vocal folds (Level 2 interaction, Fig 5). The differentiation of the two types of interaction cannot be made accurately based on the acoustic output alone but requires additional evidence, i.e., direct visualization of the vocal folds [68], preferably in vivo.
Our results in chicken and budgerigar suggest Level 2 interaction because fundamental frequency follows the first resonance as the trachea becomes shorter. In other species that have been investigated, the interpretation is less clear because features of both types of interaction were found in the acoustic output (Fig 5). A possible explanation for these different source–tract interactions of different species may lie in morphological differences in the structure of the vocal folds. Vocal folds in birds vary substantially in design, ranging from thin membranes (e.g., in many Galloanseriformes and parrots [46]) to thick multilayered structures in Passeriformes [40,41]. The layer structure in passeriform vocal folds is a good predictor of a species’ fundamental frequency range used during singing [20].
Our approach has addressed the origin of the syrinx as opposed to its diversification. It is therefore imperative to assume a simple sound source, rather than the diverse morphologies found in extant birds. Once the relocation of the sound source had occurred in an as of yet unknown ancestor of Aves, the further diversification may have explored many different avenues for further increasing vocal output, such as two sound sources, different interactions with the upper vocal tract, etc. The different mechanisms of interaction suggest a possible, albeit speculative, scenario for the origin of the syringeal sound source within Aves. If strong interaction leads to less control or uncertainty in fundamental frequency because of a sparsity of vibration modes, then perhaps the original syrinx represented a vocal organ with Level 2 interaction. Typically, Level 1 interaction arises in conjunction with histologically more complex vibratory tissue [37]. In extant birds, more complex vocal folds evolved, i.e., in songbirds [20], while other groups, i.e., parrots or Galliformes, possess thin membranes but display variable degrees of muscular control of the syrinx [46]. The heliox data in intact and denervated budgerigars suggest that muscular control of the syrinx can modulate source–tract coupling.
This initial investigation of possible selective advantages of a syringeal location of the sound source also highlights that the evolutionary origin of novelty can be addressed with specific tests of hypotheses about selective scenarios. Our data show that one likely selective advantage of the syringeal position is increased efficiency. The ability to generate loud sounds is important for long-range acoustic communication and in the context of courtship and territory defense [70, 71]. Thus, both natural and sexual selective forces may have contributed to the evolution of the avian syrinx. To what degree an early syrinx may have coexisted with a laryngeal sound source remains to be determined.
The modeling and the experiments conducted here deliberately constitute a test of a limited and small set of parameters rather than a physical replica of the avian vocal organ, with all its complexity. While this minimalist approach is likely to inform about a possible selective advantage for the switch in source location, it does not include a thorough test of other selective scenarios and does not explore other likely adaptations for increased efficiency in extant birds. Therefore, future work will have to test whether the dramatic efficiency advantage of a syringeal position is maintained for various syrinx designs or if other variables emerge as the main targets of selection. Syrinx morphology shows remarkable diversity, including features such as multiple sound sources [53], multilayered vocal fold design [41], or changes in vocal tract design and motility [32]. All of these features affect efficiency, and we do not know how they are influenced by trade-offs between vocal efficiency and those other acoustic features. Nevertheless, our approach presents a first test and sets the stage for testing additional hypotheses related to syrinx origin and syrinx diversification.
The current study highlights vocal efficiency as an important selective force that may have played a role in the evolution of the syrinx as a vocal organ. The timing of the transition from larynx to syrinx in the theropod lineage leading to modern birds is unknown prior to 66–68 million years ago [7]. Whereas clarification awaits new fossil data, the results from this study allow some speculation about a possible scenario and therefore the timing of syrinx evolution.
The response curves in both the physical model and computational simulation (Fig 2 and Fig 3) demonstrate that the interactions between sound source and vocal tract are complex and nonlinear. There are regions that represent local optima and others that are unfavorable for sound production. For example, phonation threshold pressure is lowest in a region of tracheal length 75–150 cm (Fig 2A) and 40–70 cm (Fig 3A). This demonstrates that for a given sound source, a certain tracheal length range is optimal, and both achieved maxima that are higher for the syringeal than the laryngeal position.
There are also regions in which the rate of change reaches a maximum. For example, the region between 50 and 100 cm trachea length appears to be such an inflection point. Glottal efficiency begins to increase at about 100 cm and with smaller trachea lengths. Again, this is more dramatic for the syrinx than for the larynx (Fig 2D and Fig 3F). Our data therefore highlight the possibility that the evolution of a simple syrinx may be tied to a specific constellation of body size and vocal fold morphology. The theropod lineage leading to birds underwent sustained miniaturization of body size and rapid diversification [72], and in these processes, it is possible that combinations of body size–dependent vocal tract length and sound frequencies favored the evolution of a novel vocal organ. For the sound source used in our study, a tracheal length between 50 and 100 cm yields higher vocal efficiency for the syrinx than the larynx.
An important question for the evolution of the syrinx is to what degree the advantageous interaction between sound source and vocal tract resonance, shown here for a specific size, can be generalized to other body sizes? We postulate that this favorable interaction is not limited to a specific size; as long as the tracheal resonance remains above but close to the fundamental frequency, a favorable suprasyringeal inertive compliance will provide the best assistance to self-sustained oscillation [37]. While the one available study on avian tracheal length suggests that also in small birds, the trachea remains relatively longer than in similar sized other vertebrates [18], a broader and more systematic sampling of avian tracheal anatomy seems warranted.
However, even if the exceptional avian body size–trachea length relationship did not hold for smaller birds, many modern birds possess intrinsic syringeal musculature and are able to modulate vocal fold tension, i.e., fundamental frequency can be actively adjusted to remain close and below the tracheal resonance. Furthermore, small birds have three potential mechanisms for dynamically adjusting their vocal tract resonances: (1) tracheal length changes [73], (2) size changes of the laryngeal aperture [31], and (3) size changes of the oropharyngeal–esophageal cavity [32]. The long tracheal vocal tract in addition to dynamic filter components might allow birds of all sizes to easily maintain tuning of the vocal tract resonance to a quarter wavelength of the fundamental frequency.
If the syrinx is so much more efficient, why do other groups such as nonavian reptiles, frogs, and mammals continue to use the larynx as sound source? The avian trachea is acoustically longer than those of mammals, nonavian reptiles, and frogs. “Acoustically long” means that tube length is near the quarter wavelength of the fundamental frequency of the sound source (see Introduction, Eq 2). The first ancestral bird with a syrinx most likely produced a low fundamental frequency and covered only a small frequency range. The ancestral syrinx did probably not possess any intrinsic muscles if we assume it resembled that of ostrich, emu, or cassowary [46]. Substantial frequency modulation probably only arose once tension of the vibratory tissue could be adjusted by muscular control [45].
Avian archeosaurs tend to have relatively long or very long necks. While almost no mammal has evolved more than seven cervical vertebrae, in birds, cervical vertebrae are more numerous and often elongated [74,75]. Long necks contain long tracheas. Consequently, tracheal length (i.e., suprasyringeal tracheal airspace) of birds tends to be much closer to the quarter wavelength of their voice’s lowest fundamental frequency. Whereas tracheal length is greater in birds, tracheal diameter is not different between mammals and birds (Fig 6). Most importantly, tracheal length in birds is close to the quarter wavelength of the size-predicted lowest fundamental frequency, thus enabling a boost in vocal efficiency through the overlap of positive vocal tract reactance and fundamental frequency (Fig 6C and 6D). For the first bird with a syrinx, the lowest fundamental frequency may have overlapped with the high positive reactance range just to the left of the first formant, which coincides with the most dramatic supportive interaction between source and filter. This can boost vocal efficiency.
In contrast, the vocal tract of most mammals is acoustically short, i.e., the first resonance is much higher than the lowest fundamental frequency. The wavelength of the fundamental frequency is much longer than the tracheal length (Eq 2, Fig 6D). Most living mammalian species and therapsid ancestors had short necks and neck length did not vary much [74]. Consequently, tracheal length is not sufficient for facilitating a similar boost in vocal efficiency, as was possible for long-necked birds. For example, the average adult female human trachea is about 12 cm long, and the upper vocal tract is about 14 cm long. Even if the vocal folds were located at the tracheobronchial junction (12 cm + 14 cm = 26 cm), the quarter wavelength of the fundamental frequency of a human female’s average speaking voice (210 Hz) is much longer (close to 42 cm). The discrepancy is even more pronounced for males.
With the currently available data, we present the following testable model for the evolution of the avian syrinx. The unique avian respiratory system, with its unidirectional flow through the gas exchange tissue, made gas exchange very efficient. This freed the avian bauplan from an important constraint in the neck area by allowing for more dead space (i.e., a longer trachea without a simultaneous decrease in tracheal diameter). Consequently, respiratory needs were permissive of longer necks with longer tracheas. A longer trachea shifted the avian vocal system (i.e., sound source and vocal tract) into a range for which an overlap of fundamental frequency and first tracheal resonance was possible. At this point, it became advantageous to move the sound source upstream near the tracheobronchial juncture. This model indicates that a multitude of different interacting systems must generate a permissive scenario in which novel structures for particular functions can emerge. Perhaps the evolution of a novel structure for an already existing function, such as the switch of the sound source from larynx to syrinx, particularly requires coinciding, permissive interactions [2].
|
10.1371/journal.pgen.1008004 | The meiotic phosphatase GSP-2/PP1 promotes germline immortality and small RNA-mediated genome silencing | Germ cell immortality, or transgenerational maintenance of the germ line, could be promoted by mechanisms that could occur in either mitotic or meiotic germ cells. Here we report for the first time that the GSP-2 PP1/Glc7 phosphatase promotes germ cell immortality. Small RNA-induced genome silencing is known to promote germ cell immortality, and we identified a separation-of-function allele of C. elegans gsp-2 that is compromised for germ cell immortality and is also defective for small RNA-induced genome silencing and meiotic but not mitotic chromosome segregation. Previous work has shown that GSP-2 is recruited to meiotic chromosomes by LAB-1, which also promoted germ cell immortality. At the generation of sterility, gsp-2 and lab-1 mutant adults displayed germline degeneration, univalents, histone methylation and histone phosphorylation defects in oocytes, phenotypes that mirror those observed in sterile small RNA-mediated genome silencing mutants. Our data suggest that a meiosis-specific function of GSP-2 ties small RNA-mediated silencing of the epigenome to germ cell immortality. We also show that transgenerational epigenomic silencing at hemizygous genetic elements requires the GSP-2 phosphatase, suggesting a functional link to small RNAs. Given that LAB-1 localizes to the interface between homologous chromosomes during pachytene, we hypothesize that small localized discontinuities at this interface could promote genomic silencing in a manner that depends on small RNAs and the GSP-2 phosphatase.
| The germ line of an organism is considered immortal in its capacity to give rise to an unlimited number of future generations. To protect the integrity of the germ line, mechanisms act to suppress the accumulation of transgenerational damage to the genome or epigenome. Loss of germ cell immortality can result from mutations that disrupt small RNA-mediated genome silencing, which protects the germ line from foreign genetic elements such as transposons. Here we report for the first time that the C. elegans protein phosphatase GSP-2 that promotes core chromosome biology functions during meiosis is also required for germ cell immortality. Specifically, we identified a partial loss-of-function allele of gsp-2 that exhibits defects in meiotic chromosome segregation and that is also dysfunctional for transgenerational small RNA-mediated genome silencing. Our results are consistent with a known role of Drosophila Protein Phosphatase 1 in heterochromatin silencing, and point to a meiotic phosphatase function that ensures germ cell immortality by promoting genomic silencing in response to small RNAs.
| Animals, including humans, are comprised of two broad cell types: somatic cells and germ cells. Somatic cells consist of many diverse differentiated cell types, while germ cells are specialized to produce the next generation of offspring. An important difference between these two cell types is that somatic cells undergo aging phenomena while the germ line is effectively immortal and capable of creating new “young” offspring [1]. Understanding the basis of immortality in germ cells may provide insight into why organisms age.
In C. elegans, disruption of pathways that promote germ cell immortality results in initially fertile animals that become sterile after reproduction for a number of generations. Many such mortal germline (mrt) mutant strains are temperature-sensitive, becoming sterile at 25°C but remaining fertile indefinitely at 20°C [2]. Mutations that cause a Mrt phenotype have been reported in two distinct pathways: telomerase-mediated telomere maintenance [3,4] and small RNA-mediated nuclear silencing [5–9]. Mutations in the PIWI Argonaute protein cause immediate sterility in many species. However, disruption of the C. elegans Piwi orthologue PRG-1, which interacts with thousands of piRNAs to promote silencing of some genes and many transposons in germ cells, results in temperature-sensitive reductions in fertility and a Mrt phenotype [6–12]. Multiple members of a nuclear RNA interference (RNAi) pathway that promote the inheritance of transgene silencing also promote germ cell immortality and likely function downstream of PRG-1/Piwi and piRNAs [10,13]. One nuclear RNAi defective mutant, nrde-2, a number of heritable RNAi mutants, including hrde-1, and two RNAi defective mutants, rsd-2 and rsd-6, only become sterile after growth for multiple generations at the restrictive temperature of 25°C [10,12–16]. The reason for this temperature-sensitivity is not clear. These ‘small RNA-mediated genome silencing’ mutants fail to repress deleterious genomic loci as a consequence of deficiency for small RNA-mediated memory of ‘self’ vs ‘non-self’ segments of the genome [13,17,18]. The transgenerational fertility defects of such mutants could reflect a progressive desilencing of heterochromatin, which is modulated by histone modifications that occur in response to small RNAs, such as H3K4 demethylation and H3K9me2/3 [15,19].
The SPR-5 histone 3 lysine 4 demethylase promotes genomic silencing in the context of H3K9 methylation and represses transgenerational increases in sterility [20]. Deficiency for spr-5 also compromises germ cell immortality in a temperature-sensitive manner [21], similar to genome silencing mutants that are deficient for RNAi or RNAi inheritance [10,12–16]. However, thorough genetic screens for defects in RNAi inheritance failed to recover mutations in spr-5 [16], and a direct test confirmed that deficiency for spr-5 does not compromise RNAi inheritance [13]. It is therefore not clear if the role of SPR-5 and small RNA-mediated genome silencing proteins in maintenance of germ cell immortality is a consequence of deficiency for the same genomic silencing pathway. If this is the case, it is possible that deficiency for spr-5 leads to the upregulation of a compensatory RNAi inheritance mechanism that masks an overt role for SPR-5 in RNAi inheritance.
Pioneering studies in Neurospora demonstrated that unsuccessful pairing of whole chromosomes during meiotic prophase, as well as discrete ‘unpaired’ chromosomal regions within paired meiotic homologs, can trigger small RNA-mediated genome silencing [22]. Multigenerational transmission of hemizygous transgenes in C. elegans, which results in an ‘unpaired’ ~10 kb genomic segment within paired homologous chromosomes during meiosis, leads to transgene silencing in a manner that depends on small RNAs and the PRG-1/Piwi Argonaute protein [23]. Therefore, a conserved small RNA mechanism operates during meiosis to promote genomic silencing when either large (chromosome scale) or small (transgene scale) segments of the genome are not properly paired.
A central function of Piwi/piRNA-mediated genomic silencing is to protect the genome from foreign genetic elements like transposons and viruses [11]. Horizontal transfer of a transposon into the genome of a naïve species will result in a burst of transposition events that ends when the host mounts a small RNA-mediated genomic silencing response against the transposon. In this context, de novo transposon insertions that represent a threat to genomic integrity would create small ‘unpaired’ hemizygous discontinuities within paired homologous chromosomes during meiosis. The discrete ‘unpaired’ meiotic chromosome aberrations created by de novo transposon insertions are structurally analogous to hemizygous transgenes, which are the targets of a multigenerational small RNA-induced genome silencing process [23]. Small ‘unpaired’ meiotic discontinuities created by de novo transposon insertions are therefore likely to be important for shaping genomic and epigenomic evolution.
C. elegans chromosomes do not have a discrete centromere to maintain cohesion between chromosomes during meiosis. Therefore they utilize two domains, separated by a crossover, called the long and the short arms. These arms separate at distinct stages of meiosis to prevent premature separation, with the short arms separating in Meiosis I and the long arms separating in Meiosis II. The regulation of cohesion occurs through localization of GSP-2 to the long arms of meiotic chromosomes through binding to LAB-1, where it antagonizes AIR-2 (Aurora-B kinase) activity [24–26]. In addition, LAB-1 is also present on mitotic chromosomes where it likely antagonizes AIR-2 activity [27]. In C. elegans, LAB-1 and GSP-2 fulfills the roles played by Shugoshin and Protein Phosphatase 2A in many other organisms, by protecting meiotic chromosome cohesion on the long arms in Meiosis I [27–29]. Once recruited by LAB-1, GSP-2 keeps REC-8, a meiosis-specific cohesin subunit, dephosphorylated to protect it from premature degradation and chromatid separation [26,27]. Additionally, recent work has shown that HTP-1/2, HORMA-domain proteins are responsible for LAB-1 chromosomal recruitment and therefore GSP-2 phosphatase activity [30].
Here we report the identification of a hypomorphic allele of gsp-2, a PP1/Glc7 phosphatase, which fails to maintain germline immortality at 25°C. GSP-2 is one of four PP1 catalytic subunits in C. elegans [31,32]. PP1 phosphatase has roles in many cellular processes including mitosis, meiosis, apoptosis and protein synthesis [33]. Previously, GSP-2 has been shown to promote meiotic chromosome cohesion by restricting the activity of the Aurora B kinase ortholog AIR-2 to the short arms of C. elegans chromosomes during Meiosis I [26,27]. Here, we demonstrate that GSP-2 is likely to act during meiosis to promote germline immortality via a small RNA-mediated genome silencing pathway.
In a screen for mrt mutants [2], one mutation that displayed a Temperature-sensitive defect in germ cell immortality, yp14, was tightly linked to an X chromosome segregation defect manifesting as a High Incidence of Males (Him) phenotype, such that 3.9% of yp14 self-progeny were XO males, which was significantly greater than the 0.05% male self-progeny phenotype observed in wildtype animals at 20°C (Fig 1A, p <.0001). The yp14 mutation was mapped to Chromosome III, and whole genome sequencing revealed missense mutations in 6 genes within the yp14 interval (S1A and S1B Fig). Three-factor mapping of the yp14 Him and Mrt phenotypes suggested that yp14 might correspond to the missense mutation in gsp-2 (Fig 1C and 1D) or to a mutation in the G-protein coupled receptor gene srb-11 (S1A and S1B Fig).
To test whether the chromosome segregation defect of yp14 was due to a mutation in gsp-2, we performed a non-complementation test with a deletion mutation in gsp-2, tm301. yp14 / tm301 F1 heterozygous hermaphrodites gave rise to F2 male progeny at a frequency of 5.7% at 20°C, similar to the 3.8% male phenotype observed for yp14 homozygotes (S1C Fig). Thus, tm301 failed to complement gsp-2(yp14) for its Him phenotype. In contrast, neither gsp-2(tm301) / + nor gsp-2(yp14) / + control animals displayed a Him phenotype (S1C and S1F Fig).
Additionally, gsp-2(tm301) null mutants immediately exhibited high levels of embryonic lethality at 20°C with a few F3 embryos that survive until adulthood (Fig 1B), consistent with roles for PP1 in chromosome condensation and segregation during mitosis in several species [24,25,34]. High levels of embryonic lethality for F3 gsp-2(tm301) mutant embryos (97%), led to uniformly sterile uniformly sterile F3 adults that produced no F4 progeny [25] (Fig 1B). These very high levels of embryonic lethality contrast with the embryonic lethality observed for gsp-2(yp14) mutants, which was 6% at 20°C and 41.6% for F8 animals grown at 25°C (Fig 1B). Both the Emb and Him phenotypes were exacerbated at 25°C (Fig 1A and 1B), suggesting that gsp-2(yp14) has a chromosome segregation defect that may be mechanistically linked to its Mortal Germline phenotype (Fig 1A and 1E).
In gsp-2(yp14) mutants, the X chromosome non-disjunction defect was more pronounced at both temperatures than the embryonic lethality associated with non-disjunction of the five C. elegans autosomes (S1 Table). Mutations that cause chromosome non-disjunction during mitosis occasionally lead to loss of an X chromosome during germ cell development, which could result in the stochastic appearance of XX hermaphrodites with high numbers of XO male progeny [35]. However, jackpots of XO males did not occur when yp14 mutant hermaphrodites were isolated as single L4 larvae at 20°C or as L1 or L4 larvae at 25°C (Fig 1G, S1D and S1E Fig), implying that yp14 is a separation-of-function mutation that specifically compromises the meiotic chromosome segregation function of GSP-2, with little or no effect on mitotic chromosome segregation. It is formally possible that gsp-2(yp14) is deficient for a mitotic function of GSP-2 that is relevant to germ cell immortality that is either distinct from its role in mitotic chromosome segregation or so subtle that we could not detect it in our assays.
At 20°C, gsp-2(yp14) mutants remained fertile indefinitely, but at 25°C they exhibited sterility between generations F5 and F17 (Fig 1E and 1F). Given that LAB-1 promotes cohesion of the long arms of meiotic chromosomes via the GSP-2 phosphatase, we asked if LAB-1 is relevant to germ cell immortality by first outcrossing a lab-1 deletion with wildtype and re-isolating lab-1 homozygotes in an effort to eliminate epigenetic defects that could have accumulated in the parental lab-1 strain. Outcrossed lab-1 mutants displayed a Mortal Germline phenotype at 25°C (Fig 1E and 1F). We created lab-1; gsp-2 double mutants, which remained fertile indefinitely when grown at 20°C but displayed a slightly accelerated number of generations to sterility at 25°C in comparison with lab-1 mutants (Fig 1E and 1F). Together, these results suggest that a meiotic function of GSP-2 that is directed by LAB-1 promotes germ cell immortality. The small acceleration in the time to sterility in the double mutant animals suggests slight additivity between the mutations. Both the gsp-2 and lab-1 alleles are partial loss-of-function alleles that when combined could conceivably result in a stronger phenotype. Moreover, the weak Mortal Germline phenotype of lab-1 single mutants at 20°C was suppressed by gsp-2(yp14) (Log Rank Test, p = .001). One possible explanation for this very slight rescue at the permissive temperature is the loss of lab-1 alone results in GSP-2 being mis-localized and performing an ectopic function that is ablated when GSP-2 function is reduced. It is likely that this does not occur at 25°C because GSP-2 function is more severely compromised at the higher temperature.
Multiple genes that regulate small RNA-mediated epigenomic silencing promote germ cell immortality at high temperatures, like gsp-2(yp14) and lab-1 [10,12,16]. Three small RNA-mediated epigenomic silencing genes that are required for germ cell immortality promote a specific form of transcriptional gene silencing termed nuclear RNA interference, nrde-1, nrde-2 and nrde-4 [10,12,36]. The response to a dsRNA trigger that targets lin-26 is dependent on nuclear RNA interference [37]. Control wildtype and gsp-2(yp14) mutant animals displayed a completely penetrant Embryonic Lethality phenotype in response to lin-26 dsRNA, whereas nuclear RNAi defective mutant nrde-2 and the RNAi defective mutant rsd-6 did not (Fig 2A), indicating that nuclear RNAi within a single generation is normal in the gsp-2(yp14) mutant.
Small RNAs can trigger RNAi inheritance [10,13], where silencing of a gene in response to siRNAs can be transmitted for multiple generations. Transgenerational RNAi inheritance can occur when endogenous genes are targeted by dsRNA triggers [38], but this can also happen when GFP reporter transgenes are targeted by small RNAs derived from GFP [13,17,18]. We tested the transgene cpIs12 Pmex-5::GFP and found that it was silenced in response to GFP siRNAs and that silencing of this transgene was inherited for up to 4 generations after removal from the dsRNA trigger (Fig 2B, Results summarized S6 Table). In contrast, GFP expression in gsp-2(yp14); cpIs12 was initially silenced but silencing was not inherited over multiple generations (Fig 2B), indicating that gsp-2(yp14) promotes RNAi inheritance.
Propagation of GFP or mCherry transgenes in the hemizygous state for multiple generations elicits a strong transgene silencing response, which is thought to be due to persistent yet small ‘unpaired’ discontinuities in the structure of paired meiotic homologous chromosomes at the site of the transgene [23]. We found that hemizygosity for the transgene cpIs12 resulted in progressive transgene silencing in populations of animals over the course of several generations until cpIs12 became fully silenced by generation 5 (Fig 2C and 2D). In contrast, when cpIs12 was placed in a gsp-2(yp14) genetic background and propagated in a hemizygous state, we found that cpIs12 was initially weakly silenced but that genomic silencing never became fully penetrant (Fig 2C and 2D). Together, the above data indicate that gsp-2 promotes the silencing of unpaired hemizygous transgenes, which depends on small RNA-mediated genome silencing [23].
A central function of small RNA-mediated genomic silencing is to maintain silencing of repetitive elements and transposons in the germline, thereby protecting genomic integrity [15,19,39]. We previously reported that RNA expression of tandem repeat loci was upregulated in late-generation rsd-2 and rsd-6 mutants grown at 25°C [12]. Therefore, we asked if desilencing of tandem repeats occurred in gsp-2(yp14) mutants using RNA fluorescence in situ hybridization (FISH) to examine the expression of multiple repetitive elements. In wild-type controls grown at 25°C, we detected RNA from tandem repeat sequences using CeRep59 sense and anti-sense probes in embryos but not in the adult germline or somatic cells, consistent with previous observations (S2 Fig) [12]. However, in late-generation gsp-2(yp14) and rsd-6 mutants, robust expression of tandem repeats was observed throughout the soma and germline of adult animals, indicating that tandem repeats become desilenced in these strains (S2 Fig).
Given that small RNA-mediated genome silencing is dysfunctional in gsp-2(yp14) mutants, we asked if small RNA populations were perturbed by preparing RNA from early- and late-generation wildtype, gsp-2(yp14), rsd-6 and spr-5 mutants grown at either 20°C or 25°C. We examined rsd-6 and spr-5 mutants as they have known temperature sensitive germ cell immortality defects associated with loss genomic silencing as a consequence of small RNA or histone demethylation defects, respectively [12,21]. Small RNA libraries were prepared and subjected to high throughput sequencing, and we then examined levels of 22G RNAs that are 22 nucleotides in length beginning with a 5’ guanine, as 22G RNAs are the major effectors of genomic silencing in C. elegans [5,40]. 22G-RNAs in all late generation lines, normalized to total small RNA content showed a decrease relative to early generation N2 lines. The decrease was more pronounced in gsp-2 and rsd-6 mutants (p = 1.2e-7 and 4e-19, Wilcox paired test; S2 Table, S3 Fig) but not in spr-5 where the decrease was not significantly different from the difference in N2 (p = 0.13). Analysis of the 22G-small RNA data revealed that spr-5 and rsd-6 share some genes with reduced levels of 22G RNAs with increasing generations, but there are other genes that show dissimilar behavior for each individual mutant. This suggests that spr-5 may act both in conjunction with rsd-6 and in a separate pathway to promote germline immortality. In contrast, 22G RNAs from gsp-2(yp14) showed strong similarities to those of spr-5 mutants but showed little similarity to 22G RNA changes observed for rsd-6 mutants, suggesting that gsp-2(yp14) and spr-5 have similar effects on genome maintenance (S3 Fig). As a control, there is little coherent change in late-generation versus early generation N2 wildtype that overlaps with gsp-2(yp14) meaning that changes we see in gsp-2(yp14) are not due simply to passaging animals at 25°C (Fig 2E and 2F). As germ cell immortality is promoted in part by primary siRNAs termed piRNAs that interact with the Piwi Argonaute protein PRG-1 [8], we also examined piRNA populations, which are enriched for 21 nucleotide RNAs that begin with a 5’ uracil (21U RNAs) [6,7,9] and found that these were normal (Fig 2E and 2F). We also examined miRNAs, which have not previously been implicated in the Mortal Germline phenotype. Interestingly, miRNAs were significantly reduced in late generation spr-5 and gsp-2(yp14) mutants (p = 1.2e-20 and p = 2.05e-25 respectively; S3 Table, S3 Fig), but not in rsd-6 mutants. Since spr-5 does not show global decrease in 22G-RNAs this is unlikely to be a secondary consequence of disturbance of the total small RNA pool. The relevance of this finding to the Mortal Germline phenotype awaits further investigation. Together these results indicate that gsp-2(yp14) and spr-5 display common statistically significant changes in two classes of small RNAs, which implies that their genomic silencing defects may be more similar to one another than to those of rsd-6 mutants.
To study the relationship between gsp-2(yp14) and the small RNA genome silencing pathway, we created double mutants between gsp-2(yp14) and small RNA silencing mutants that display temperature-sensitive defects in germ cell immortality, hrde-1, nrde-2 and rsd-6. Because gsp-2(yp14) is a hypomorphic allele, we predicted that single and double mutants would display a similar number of generations to sterility if it were functioning in the small RNA silencing pathway. For gsp-2(yp14); hrde-1 and rsd-6; gsp-2(yp14), we saw a modest decrease in the number of generations to sterility suggesting a slight additive effect (Fig 3A and 3C, Log Rank test: p <.0001). In contrast, nrde-2; gsp-2(yp14) double mutants did not differ from the single mutants (Fig 3B, Log Rank test: p = .06). Together, these results indicate that there is a weak additive effect on transgenerational lifespan when gsp-2 is combined with hrde-1 or rsd-6, but not when it is combined with nrde-2. The modest acceleration observed for some small RNA genomic silencing pathway and gsp-2(yp14) double mutants may be consistent with a single genome silencing pathway, as many single mutants in this pathway that display similar germline phenotypes at sterility also display a consistent, slightly accelerated sterility as double mutants. There are a number of explanations for this, including transmission of epigenetic defects from germ cells of the grandparents that created these double mutants, or shared but non-equivalent functions in terms of which segments of the genome each protein silences [15].
We previously reported that sterile, late-generation small RNA genome silencing mutants display a wide range of germline sizes, including many with few or no germ cells [12,41]. Therefore to investigate the cellular cause of transgenerational sterility in gsp-2(yp14) and lab-1 mutants, we examined germline development in animals that became sterile after multiple generations. Most sterile generation L4 gsp-2(yp14) and lab-1 mutant germlines were normal in size, though a small minority had a reduction in total germline length, resulting in a weak but significant difference in germline profile compared to wild-type (Fig 4A–4E and 4H, S5 Table, Results summarized S6 Table). Differentiating germ cell nuclei in spermatogenesis were observed for sterile generation L4 larvae for all strains (Fig 4A and 4H). However, the germlines of two-day-old sterile gsp-2(yp14) and lab-1 mutant adults ranged in size from normal to a complete loss of germ cells (Fig 4B–4E and 4I), resulting in a significant difference when compared to wild-type controls (S5 Table p<1E-80). We studied small RNA genome silencing mutants and found that rsd-6, hrde-1 or nrde-2 mutant L4 larvae that were poised to become sterile displayed predominantly normal-sized germlines (Fig 4H). In contrast, sterile-generation rsd-6, hrde-1 and nrde-2 mutant adults had germline profiles that were similar to those of sterile gsp-2(yp14) mutant adults and markedly smaller than those of sterile generation L4 larvae (Fig 4I, S4 Table). lab-1(tm1791) displayed an increased frequency of germline tumors in comparison to other mutants, possibly due to a genetic modifier present in the tm1791 mutant background.
Lastly, we tested if sterile spr-5 mutants displayed similar germline phenotypes as those observed in small RNA mutants and gsp-2(yp14). We found that sterile spr-5 mutant adults displayed similar germline atrophy phenotypes, suggesting the resemblance to gsp-2(yp14) or lab-1 mutants (Fig 4H and 4I). Our previous work showed that mutations in the cell death genes ced-3 and ced-4 partially rescued the empty and atrophy phenotypes observed for germlines of rsd-2, rsd-6, and prg-1 small RNA genome silencing mutant adults [12,41], suggesting that apoptosis promotes germ cell atrophy as these animals develop from L4 larvae into adults.
To determine if acute loss of GSP-2 causes germline atrophy, we examined gsp-2(tm301) null mutants grown at 20°C and 25°C. gsp-2(tm301) homozygous F2 animals and their few surviving F3 progeny showed normal germlines, with no morphological defects in germline size or development for either L4 larvae or young adults, which significantly differed from the germline profiles of gsp-2(yp-14) animals (S4 and S5 Tables). Therefore, the late-generation sterility phenotype of yp14 mutants is distinct from the fertility defects that occur in response to acute loss of GSP-2 in maternally depleted F3 deletion homozygotes.
Mature C. elegans oocytes typically contain 6 bivalents (pairs of homologous chromosomes held together by crossovers), which can be scored as 6 DAPI-stained bodies. Defects in meiotic pairing, cohesion, synapsis, and crossing over can lead to the presence of univalents, which are observed as greater than 6 DAPI bodies per oocyte. We previously observed that small RNA nuclear silencing mrt mutants rsd-2 and rsd-6 displayed increased levels of univalents at sterility, which were not observed in either wildtype or in fertile rsd-2 or rsd-6 mutant late-generation animals grown at 25°C [12]. We measured the presence of oocyte univalents in N2 wildtype control worms grown at 20°C and 25°C, which almost always contained 6 DAPI bodies representing the 6 paired chromosomes (5 bodies are occasionally scored when bivalents that overlap spatially cannot be distinguished). However, when gsp-2(yp14) worms were passaged at 25°C until sterility, only 60% of oocytes contained 6 paired chromosomes with the other 40% contained 7 to 12 DAPI bodies (Fig 4J, Results summarized S6 Table). This increase in oocyte univalents was not present in fertile gsp-2(yp14) worms, at 20°C or even for fertile late-generation 25°C gsp-2(yp14) adults that were close to sterility (Fig 4J). In contrast, we found no univalents in the null gsp-2 allele tm301, either for F2 animals or for rare F3 escapers, consistent with previous observations [24,25].
LAB-1 has been previously implicated in the pairing of homologous chromosomes during meiosis [27]. To determine if homolog pairing is perturbed in gsp-2(yp14) mutants grown for two generations at 25°C, we examined the X chromosome pairing center protein HIM-8 in fertile 2 day old adults. When scored at pachytene only one spot was present in the majority of the nuclei suggesting pairing is occurring normally (Fig 4K, S4 Fig). In addition to gsp-2(yp14), we examined HIM-8 foci in fertile lab-1, rsd-6 and spr-5 mutants grown at 25°C for two generations and found that lab-1 mutants displayed decreased meiotic chromosome pairing consistent with previously reported data [27] but that pairing was relatively normal in the other mutants (Fig 4K, S4 Fig).
Given that LAB-1 and GSP-2 are known to promote meiotic chromosome cohesion, we tested the hypothesis that dysfunction of other factors that promote meiotic chromosome cohesion might be sufficient to elicit germline atrophy. Mutant strains with defects in cohesion, smc-3(t2553) and coh-3(gk112); coh-4(tm1857) double mutants [42–44] became sterile immediately and did not exhibit germline atrophy phenotypes observed in gsp-2(yp14) (S5 Fig, S4 Table). Therefore, the late-generation sterility phenotypes of gsp-2(yp14) and small RNA mutants are not due to acute loss of meiotic chromosome cohesion.
To further characterize the nature of the gsp-2(yp14) mutation, we examined the localization of LAB-1 and GSP-2 in pachytene nuclei of gsp-2(yp14), lab-1, rsd-6 and spr-5 animals. Decreased GSP-2 localization was observed in both gsp-2(yp14) and spr-5 mutants but not in lab-1 or rsd-6 mutants (Fig 5A). Similar defects in small RNA profiles of gsp-2 and spr-5 mutants are consistent with the localization of GSP-2 being normal in rsd-6 mutants but absent in gsp-2(yp14) and spr-5 mutants (Fig 5A), which supports the possibility that GSP-2 may promote genomic silencing in response to small RNAs. The presence of GSP-2 staining in the lab-1 deletion was surprising as animals treated with RNAi against lab-1 show decreased GSP-2 staining. However, as the tm1791 deletion is a non-null allele, it is possible that GSP-2 can still interact with LAB-1 to some degree. Additionally, we saw no change in LAB-1 localization in any strain except for the lab-1 deletion, which still exhibited some staining consistent with the tm1791 deletion being a non-null allele (Fig 5B). Lastly, we assessed LAB-1 localization at diakinesis to determine if LAB-1 localization on the long arms was altered in any of these mutants and we found that localization was relatively normal in gsp-2(yp14), rsd-6 and spr-5 mutants (S6 Fig). The localization of LAB-1 in gsp-2(yp14) along the long arms was abnormal looking but clearly did not localize to both the long and short arms of the chromosomes.
A previously identified phenotype of gsp-2 null mutants is an increase in Histone 3 Serine 10 (H3S10) phosphorylation due to expansion of the AIR-2-localizing domain [24,30]. In wildtype worms grown at 20°C and 25°C, H3S10 phosphorylation was visible on the condensed chromosomes in the -1 to -3 oocytes, which are defined relative to the spermatheca with the closest being called -1 (Fig 6A, Results summarized S6 Table). In both early- and late-generation gsp-2(yp14) mutant oocytes, H3S10 phosphorylation increased when compared with wildtype controls, with increased levels on chromosomes (Fig 6B and 6M). Late-generation gsp-2(yp14) mutant animals grown at 25°C had a small but significant increase in H3S10 phosphorylation levels compared to gsp-2(yp14) mutant controls grown at 20°C (Fig 6M). Furthermore, we observed increased levels of H3S10 phosphorylation in lab-1 mutants (Fig 6C and 6M), consistent with previous results [27]. By quantification of fluorescence intensity we measured significant increased levels of H3S10p in lab-1, rsd-6, and hrde-1 but not in nrde-2 mutants (Fig 6C–6F and 6M). The distinct phosphorylation levels in nrde-2 mutants could reflect its small RNA genome silencing function, where NRDE-2 works downstream of RSD-6 and HRDE-1 to promote accumulation of stalled RNA polymerase II at loci that are targeted by small RNAs [45]. This would suggest that the maintenance of histone marks occurs at the point in the pathway where RSD-6 and HRDE-1 function but not downstream at level of NRDE-2.
PP1 has been previously shown to dephosphorylate a number of histone amino acids, including Histone 3 Threonine 3 (H3T3) [46]. When we examined H3T3 phosphorylation in wildtype controls grown at 25°C, staining was visible in the -1 to -3 oocytes (Fig 4G and 4G’, Results summarized S6 Table). However, in sterile generation gsp-2(yp14) mutants, H3T3p staining was significantly brighter than controls when images were taken under the same conditions (Fig 6H and 6H’). Sterile generation lab-1 and the small RNA mutants hrde-1, rsd-6 and nrde-2 all exhibited increased H3T3 phosphorylation signal intensity in the -1 to -3 oocytes (Fig 6I, 6L and 6N). Furthermore, there was a significant increase in H3T3 phosphorylation in sterile generation gsp-2(yp14) mutant adults compared to the earlier, fertile generation animals suggesting transgenerational accumulation of H3T3 phosphorylation (Fig 6N). Together, our results suggest that an increase in phosphorylation of H3T3 consistently occurs in oocytes of gsp-2 and small RNA silencing mutants however, increased H3S10 phosphorylation occurs only in gsp-2(yp14), lab-1, rsd-6, and hrde-1 but not in nrde-2 mutants. This defect is sensitive to temperature, as observed for the meiotic chromosome segregation and germ cell immortality defects of gsp-2(yp14) (Fig 1E and 1F).
Finally, we examined histone marks that promote gene silencing or activation. H3K9 methylation can be deposited at silenced genomic loci, and H3K9me and H3S10p marks can function as a phospho-methyl switch where H3S10 phosphorylation can block some epigenetic regulators, such as HP1, from accessing the adjacent H3K9me mark [47–49]. In late-generation fertile gsp-2(yp14), lab-1, rsd-5 and spr-5 mutant animals grown at 25°C, we observed a significant decrease in H3K9me2 and H3K9me3 intensity in diakinesis oocytes (Fig 7A and 7B, Results summarized S6 Table). We also assessed the H3K4me3 transcriptional activation mark and found that it was significantly decreased in all mutant genotypes at diakinesis (Fig 7A and 7B). It is possible that excess H3T3 phosphorylation present in these mutant strains (Fig 6G–6N) could affect the activities of enzymes that modify histone H3, especially H3K4. Additionally, the presence of excess phosphorylation on adjacent amino acids could perturb the binding of the histone methylation antibodies, possibly disrupting our ability to assess methylation levels.
We demonstrate for the first time that gsp-2 and lab-1 are required for germ cell immortality at 25°C as strains deficient for these proteins become sterile when they are passaged for several generations (Fig 1C and 1D). Although PP1/GSP-2 is a general protein phosphatase with roles in a number of cellular processes including mitosis and meiosis [33], we identified a separation-of-function allele of gsp-2 that displayed an X chromosome non-disjunction phenotype that was specific for meiosis (Fig 1B and 1G, S1D and S1E Fig). The incidence of both X chromosome loss and inviable embryos, which are likely aneuploid for autosomes, was exacerbated at high temperature (Fig 1A and 1B), which is consistent with the temperature-sensitive defect in germ cell immortality observed for gsp-2(yp14) mutants. Stronger defects in segregation of the X chromosome of gsp-2(yp14) mutants during meiosis could be due to the fact that X chromosomes tend to have more central crossovers than the autosomes [50]. PP1/GSP-2 is known to be recruited to meiotic chromosomes by the C. elegans-specific protein LAB-1, and we found that deficiency for lab-1 elicited transgenerational sterility accompanied by adult germ cell degeneration phenotypes that were observed in sterile small RNA silencing mutants (Figs 1F and 4). Together, these results indicate that LAB-1 and GSP-2/PP1 are likely to define a critical step during meiosis that potentiates genomic silencing and germ cell immortality (Fig 8). Our data that GSP-2 acts in the context of hemizygous transgenes suggests that it may promote genomic silencing at a stage of germ cell development where homologous chromosomes physically interact.
We found that gsp-2(yp14) mutants were proficient for nuclear RNA interference and for the initial generation of silencing of a GFP transgene in response to an exogenous dsRNA trigger (Fig 2A and 2B). However, in subsequent (inheriting) generations, gsp-2(yp14) mutants failed to maintain GFP transgene silencing, indicating that gsp-2(yp14) is defective for RNAi inheritance (Fig 2B), a trait that is frequently associated with temperature-sensitive defects in germ cell immortality [16]. Consistently, propagation of an ‘unpaired’ hemizygous GFP transgene for multiple generations resulted in complete transgene silencing for wildtype controls, but only partial transgene silencing in a gsp-2(yp14) mutant background (Fig 2C). These independent tests indicate that gsp-2(yp14) is deficient for small RNA-mediated genomic silencing.
Hemizygous transgene silencing occurs in a manner that depends on prg-1/Piwi and associated piRNAs as well as downstream factors that promote second siRNA biogenesis [23]. However, we found that piRNA levels were normal in gsp-2(yp14) mutants, and also that late-generation gsp-2(yp14) strains displayed changes in 22G RNA levels that were similar to those of spr-5 histone H3K4 demethylase mutants but not to those of rsd-6 small RNA biogenesis mutants (Fig 2). Moreover, epistasis analysis indicated that there is a weak additive effect when gsp-2 is combined with the nuclear Argonaute hrde-1 or the small RNA biogenesis factor rsd-6, but no additive effect when gsp-2 is combined with nrde-2 (Fig 3) [45]. The parallels with spr-5 and nrde-2 mutants suggest that GSP-2 may help to integrate histone silencing modifications with the response to small RNAs (Fig 8). In this context, the GSP-2 phosphatase could directly modify histones or a component of the genome silencing machinery that responds to small RNAs. It is possible that the yp14 mutation compromises the ability of GSP-2 to interact with either LAB-1 or with small RNA genome silencing proteins in a manner that abrogates the process of small RNA-mediated genomic silencing.
Hemizygous transgenes cause persistent transgenerational discontinuities in the local pairing of small regions of DNA during meiosis, which promotes transgene silencing in a manner that depends on GSP-2 (Figs 2C and 8). Although deficiency for LAB-1 perturbs the pairing of homologous chromosomes during meiosis [26,30], we found that homolog pairing is normal for gsp-2(yp14) mutants. LAB-1 localizes to the interface between homologous chromosomes during pachytene, and LAB-1 recruits GSP-2 to nuclei during early stages of meiosis (Fig 8) [26,27,30]. We therefore suggest that LAB-1/GSP-2 may act at the interface between homologous meiotic chromosomes to promote small RNA-mediated epigenomic silencing (Fig 8). An intriguing possibility is that locally ‘unpaired’ hemizygous transgenes could create a structural discontinuity between paired homologous chromosomes that alters the normal meiotic function of LAB-1/GSP-2, creating an environment where the chromosome silencing machinery can respond to small RNAs (Fig 8). Alternatively, the presence of a homologous allele could provide protection from silencing [51–53].
In mammals, a wave of piRNA production occurs during the pachytene stage of meiosis [54,55]. Pachytene piRNAs are derived from intergenic regions, are depleted for transposons, and their functions are not well understood [56]. Given that LAB-1/GSP-2 localizes to the interface between homologous chromosomes during pachytene [26,30], we suggest that one purpose of pachytene piRNAs may be to detect and coordinate the response to ‘unpaired’ structural discontinuities that represent de novo transposition events that threaten genome integrity (Fig 8). Consistently, components of the C. elegans small RNA-mediated genome silencing machinery, such as the HRDE-1 and PRG-1/Piwi Argonaute proteins, are expressed throughout germ cell development and are present during meiotic prophase [6,10,13,18].
Consistent with our results, an allele of the Drosophila Protein Phosphatase 1 gene, Su var (3) 6, was identified as a suppressor of position-effect variegation, which relieves epigenetic silencing of a transcriptionally active gene that is placed adjacent to a segment of heterochromatin [34]. As position-effect variegation is promoted by small RNA-mediated genome silencing in animals, plants and fungi [57,58], we conclude that PP1 is likely to play a conserved role in this epigenomic silencing process. It has been suggested that the heterochromatin defect of Su var (3) 6 mutants could reflect a direct role of PP1 in dephosphorylation of H3S10p, a mark that results in dissociation of Heterochromatin Protein 1 from heterochromatin [46,59]. Moreover, human PP1 has been shown to dephosphorylate H3T3p, this function is also carried out by C. elegans GSP-2 during meiosis [46,59]. One or both of these silencing marks could be relevant to meiotic small RNA-mediated genome silencing.
We propose that the role of LAB-1/GSP-2 in genome silencing may be at a stage of meiosis, possibly pachytene, when LAB-1/GSP-2 are localized between paired homologs in a manner that might be capable of responding to small ‘unpaired’ discontinuities between homologs like hemizygous transgenes (Fig 8). This model raises questions about the significance of increased H3T3 and H3S10 phosphorylation levels in mature oocytes of gsp-2(yp14) mutants at diakinesis when homologous chromosomes are held together only by chiasma [26,27]. H3 phosphorylation defects were not observed at earlier stages of germ cell development, but similarly increased levels of H3T3 phosphorylation were observed at diakinesis for small RNA genome silencing mutants (Fig 7A and 7B). This could suggest that altered histone phosphorylation levels could be an indirect effect of dysregulation of heterochromatin, which could affect the activity of a protein that functions in the context of heterochromatin, such as the H3T3 kinase Haspin [36]. It is also possible that the diakinesis-specific phosphorylation defect that we observed reflects a fundamental property of how GSP-2 promotes genomic silencing in response to small RNAs. For example, the structure that triggers genomic silencing could occur at pachytene when homologous chromosomes are paired, but the role of GSP-2 in responding to small RNAs could occur at a later stage of germ cell development like diakinesis, potentially via H3 phosphorylation.
Work in several organisms, particularly in fungi and Drosophila, has shown that local regions of heterozygosity are prone to silencing during meiosis in a small RNA dependent manner (reviewed by [22]). Our study defines a meiotic process that links transgenerational small RNA-mediated genome silencing with the structure of paired homologous chromosomes during meiosis. Given that endogenous small RNAs promote germ line stem cell maintenance, oogenesis and meiosis itself [60,61], we suggest that small RNA pathways and germ cell development have evolved to become mutually reinforcing processes.
All strains were cultured at 20°C or 25°C on Nematode Growth Medium (NGM) plates seeded with E. coli OP50. Strains used include Bristol N2 wild type, gsp-2(tm301) III, gsp-2(yp14) III, lab-1(tm1791) I, cpIs12[Pmex-5::GFP::tbb-2 3'UTR + unc-119(+)] II, hrde-1(tm1200) III, rsd-6(yp11) I, nrde-2(gg95) II, rbr-2(tm1231) IV, smc-3(t2553) III, coh-4(tm1857) V, coh-3(gk112) V, air-2(or207) I, unc-32(e189) III, unc-13(e450) I, unc-24(e1172) IV. smc-3(t2553) is a temperature sensitive missense mutation, and coh-4(tm1857)/coh-3(gk112) are deletions.
Worms were assessed for the Mortal Germline phenotype using the assay previously described [2]. L1 or L2 larvae were transferred for all assays. After passaging plates that yielded no additional L1 animals were marked as sterile. Log-rank analysis was used to determine differences of transgenerational lifespan between strains.
DAPI staining was performed as previously described. L4 larvae were selected from sibling plates and sterile adults were singled as late L4s, observed 24 hours later for confirmed sterility, and then stained 48 hours after collection.
DNA oligonucleotide probes coupled with a 5′ Cy5 fluorophore were used to detect repetitive element expression. The four probes used in this study were as follows: tttctgaaggcagtaattct, CeRep59 on chromosome I (located at 4281435–4294595 nt); agaattactgccttcagaaa, antisense CeRep59 on chromosome I; caactgaatccagctcctca, chromosome V tandem repeat (located at 8699155–8702766 nt); and gcttaagttcagcgggtaat, 26S rRNA. The strains used for RNA FISH experiments were rsd-6(yp11), gsp-2(yp14), and N2 Bristol wild type. Staining was performed as described by Sakaguchi et al., 2014.
Adult hermaphrodites raised at 20°C or 25°C were dissected in M9 buffer and flash frozen on dry ice before fixation for 1 min in methanol at -20°C. After washing in PBS supplemented with 0.1% Tween-20 (PBST), primary antibody diluted in in PBST was used to immunostain overnight at 4 °C in a humid chamber. Primaries used were 1:500 pH3S10 (Millipore, 06570), 1:4000 pH3T3 (Cell Signaling, D5G1I, Rabbit) 1:50 GSP-2 antibody (Colaiacovo lab), 1:300 LAB-1 antibody (Colaiacovo lab), 1:200 HIM-8 antibody raised in guinea pig (Dernburg lab), 1:200 SYP-1 antibody raised in goat, 1:500 H3K9me3 (Abcam ab8898), 1:500 H3K9me2 (Milipore Upstate 07–441), 1:500 H3K4me3 (Active Motif 39159). Secondary antibody staining was performed by using an Cy3 donkey anti-mouse or Cy-5 donkey anti-rabbit overnight at 4°C. All images were obtained using a LSM 710 laser scanning confocal and were taken using same settings as control samples. Images processed using ImageJ and Icy (http://icy.bioimageanalysis.org/). Intensity quantification was done by measuring total fluorescence in individual condensed chromosomes and subtracting the background levels obtained from mitotic nuclei as nucleoplasm levels varied greatly. Histone methylation intensity measurements were measured without background subtraction since only very few background was present.
N2 wildtype, gsp-2, rsd-6 and nrde-2 animals were grown on lin-26 RNAi clones and the progeny of 10 worms each were scored for Embryonic Lethality.
cpIs12 and gsp-2; cpIs12 worms were scored for GFP expression on NGM plates and then transferred to RNAi plates targeting GFP. The next generation (that was laid on GFPi plates) were scored for GFP expression and their sisters were removed and transferred back to NGM plates. Worms were propagated for multiple generations on NGM and scored each time for GFP expression. Both GFP reporter gsp-2 doubles were created by marking with dpy-17.
cpIs12 was maintained as a heterozygote over dpy-10 unc-4 for gsp-2 strains that were heterozygous for cpIs12, gsp-2 remained mutant for the entire assay and cpIs12 was balanced over dpy-10 unc-4.
Paired sequence reads (2X100 nucleotide long) were mapped to the C. elegans reference genome version WS230 (www.wormbase.org) using the short-read aligner BWA [62]. The resulting alignment files were sorted and indexed with the help of the SAMtools toolbox [63]. The average sequencing depths for the mutant and wild-type N2 strains were 116x and 71x, respectively. Single-nucleotide variants (SNVs) were identified and filtered with SAMtools and annotated with a custom Perl script using gene information downloaded from WormBase WS230. Candidate SNVs in the mutant strain already present in the N2 strain were eliminated from further consideration.
The raw sequence data from this study have been submitted to the NCBI BioProject (http://www.ncbi.nlm.nih.gov/bioproject) under accession number PRJNA395732 and can be accessed from the Sequence Read Archive (SRA; https://www.ncbi.nlm.nih.gov/sra) with accession number SRP113543.
5’ independent small RNA sequencing was performed as described previously [13], using one repeat for each time-point of N2 wildtype, rsd-6 and spr-5 at 25°C. Custom Perl scripts were used to select different small RNA species from the library. To map small RNA sequences to genes, reads were aligned to the C. elegans ce6 genome using Bowtie, Version 0.12.7, requiring perfect matches [64]. Data was normalized to the total number of aligned reads and 1 was added to the number of reads mapping to each gene to avoid division by zero errors. To map 22G sequences to transposons and tandem repeats, direct alignment to the transposon consensus sequences, downloaded from Repbase (Ver 17.05) or repeats obtained from the ce6 genome (WS190) annotations downloaded from UCSC as above, was performed using Bowtie allowing up to two mismatches and reporting only the best match. Uncollapsed fasta files were used for these alignments to compensate for the problem of multiple identical matches. Data was normalized to the total library size and 1 was added to the number of reads mapping to each feature to avoid division by zero errors. In order to analyze data from rsd-2 mutants grown at 20°C [65], Fasta files were downloaded from the Gene Expression Omnibus and uncollapsed using a custom Perl script before aligning to transposons or tandem repeats as above. Analysis of data was carried out using the R statistical language [66].
The small RNA sequencing data from this study are available from GEO database accession number GSE126531.
|
10.1371/journal.pntd.0004319 | Ecological Drivers of Mansonella perstans Infection in Uganda and Patterns of Co-endemicity with Lymphatic Filariasis and Malaria | Mansonella perstans is a widespread, but relatively unknown human filarial parasite transmitted by Culicoides biting midges. Although it is found in many parts of sub-Saharan Africa, only few studies have been carried out to deepen the understanding of its ecology, epidemiology, and health consequences. Hence, knowledge about ecological drivers of the vector and parasite distribution, integral to develop spatially explicit models for disease prevention, control, and elimination strategies, is limited.
We analyzed data from a comprehensive nationwide survey of M. perstans infection conducted in 76 schools across Uganda in 2000–2003, to identify environmental drivers. A suite of Bayesian geostatistical regression models was fitted, and the best fitting model based on the deviance information criterion was utilized to predict M. perstans infection risk for all of Uganda. Additionally, we investigated co-infection rates and co-distribution with Wuchereria bancrofti and Plasmodium spp. infections observed at the same survey by mapping geographically overlapping areas.
Several bioclimatic factors were significantly associated with M. perstans infection levels. A spatial Bayesian regression model showed the best fit, with diurnal temperature range, normalized difference vegetation index, and cattle densities identified as significant covariates. This model was employed to predict M. perstans infection risk at non-sampled locations. The level of co-infection with W. bancrofti was low (0.3%), due to limited geographic overlap. However, where the two infections did overlap geographically, a positive association was found.
This study presents the first geostatistical risk map for M. perstans in Uganda. We confirmed a widespread distribution of M. perstans, and identified important potential drivers of risk. The results provide new insight about the ecologic preferences of this otherwise poorly known filarial parasite and its Culicoides vector species in Uganda, which might be relevant for other settings in sub-Saharan Africa.
| Mansonella perstans is a widespread, but relatively unknown human filarial parasite that occurs in many parts of Africa. In a nationwide survey carried out in Uganda in 2000–2003, the distribution of M. perstans was assessed by screening school children. Here, we studied the underlying environmental drivers and ecologic correlates of the observed M. perstans prevalence patterns, produced a predictive risk map, and investigated associations with Wuchereria bancrofti (causative agent of lymphatic filariasis) and Plasmodium (causative agent of malaria). Several Bayesian geostatistical logistic regression models with and without spatially structured random effects were fitted for comparison. The model that fitted the data best was used to predict M. perstans infection risk for all of Uganda. Positive associations with M. perstans infection status were observed with cattle densities, forested areas, and vegetation greenness, whereas negative associations were observed with land surface temperature. Only a small geographic overlap was observed with W. bancrofti, and the overall level of co-infection was low (0.3%). However, where the two infections overlapped, a positive association was found. Our study presents the first nationwide geostatistical risk map for M. perstans, and gives important clues about the ecologic preferences of the still unknown main Culicoides vector species in Uganda.
| The human filarial parasite Mansonella perstans has been considered as one of the most prevalent human parasites in Africa [1]. Despite the wide distribution, only very few studies have addressed its epidemiology and associated health consequences, and currently no effective drug therapy for treatment, control, and local elimination is available [2]. Indeed, M. perstans is viewed as one of the most neglected of the neglected tropical diseases (NTDs) [2].
On-going large-scale surveys and control programs for other filarial infections (e.g., lymphatic filariasis and onchocerciasis), considered to be of greater health importance, have largely ignored M. perstans infections, even though these filarial infections frequently co-occur. This lack of attention mainly stems from its predominance in poor rural communities, and from a paucity of a distinct and clearly recognizable clinical picture [2]. However, widespread co-occurrence with lymphatic filariasis and onchocerciasis could cause complications with regards to control program diagnosis and compliance assessment [2], and could potentially trigger adverse events during mass anti-filaricides administration [3]. It has also been suggested that there could be more subtle effects, as M. perstans might interfere with the host’s immune regulation and influence the susceptibility and effect of other, co-occurring pathogens such as Plasmodium spp. and HIV [2].
The geographic distribution and transmission of M. perstans is closely linked to its vectors, biting midges of the genus Culicoides, and their environmental requirements for breeding and feeding. Culicoides species are widespread throughout the world, and known to transmit a variety of pathogenic viruses, bacteria, protozoa and helminths to humans, and to domestic and wild animals [4–6]. Yet, they remain among the least studied of the Dipteran vectors [7]. As such, only a few studies have tried to incriminate the exact Culicoides species responsible for transmission of M. perstans in endemic areas in Africa [2]. An accurate understanding of the environmental drivers of both vector and parasite distribution is paramount for the development of spatially explicit risk models based on sound ecological principles, which can help optimize disease prevention planning, and control and elimination programs.
In 2000–2003 a national survey was conducted to map the distribution of M. perstans, concurrently with that of Wuchereria bancrofti [8] and Plasmodium parasites, in school-aged children in Uganda. While geostatistical risk and co-endemicity maps have been constructed for the two latter infections [9], M. perstans infections in Uganda have only crudely been mapped [10]. Furthermore, no risk factor analysis has been performed to identify the underlying environmental drivers of M. perstans infection, and the co-infection rates and geographic overlaps (co-distribution) between the three parasites have yet to be investigated. Delineating areas of geographic overlap, where co-infections might occur, is an important operational issue for integrated disease control planning and implementation [11].
The aim of the present study was to determine the underlying environmental drivers and ecological correlates of the observed prevalence patterns of M. perstans infection and to produce statistically robust prevalence estimates at non-sampled locations (smooth prevalence maps) across Uganda. We furthermore investigated the levels of co-infection and co-distribution with bancroftian filariasis and malaria.
The studies which contributed data used in this paper, received ethical clearance from the Uganda National Council for Science and Technology and were approved by the Central Scientific Ethical Committee of Denmark. Prior to each survey, meetings were held with school staff and village leaders, to explain the objectives and implications of the study. Written informed consent to participate was obtained from those examined (or from the parents/legal guardians of participants aged <15 years). At each study site, a clinical officer from a nearby health unit accompanied the team, examined all the children who were not feeling well and, if need be, either treated the children or referred them to a nearby clinic. For a full description, we refer the reader to prior publications [8, 10].
The surveys were carried out between October 2000 and April 2003 and included pupils aged 5–19 years from 76 Ugandan primary schools (12,207 pupils in total) covering the major topographical and ecological zones of the country (see S1 Appendix for a list of schools, with names, geographical coordinates and prevalence). Full details of the study design, data, and the procedures for selection of study sites and participants have been described elsewhere [8,10]. In brief, 100-μl blood sample was collected from each consenting child during the school day and used to prepare a thick film to examine for microfilaremia and Plasmodium parasites. After drying, the thick films were dehemoglobinized, fixed in methanol, stained with Giemsa, and examined under a microscope. All microfilariae observed were identified to species, using morphological criteria [12] and counted. Finger-prick samples of blood were also collected, and assayed for W. bancrofti specific circulating filarial antigens (CFA) by use of ICT cards [1]. Boys and girls were examined in approximately equal numbers.
We investigated a series of climatic and other environmental variables (Table 1) known to be of importance for the distribution of arthropod transmitted parasitic infections in the tropics, but also known ecological drivers of Culicoides species transmitting other parasites and viruses [7,13]. These included measures of temperature, known to influence parasite developmental rate and vectorial development rates, as well as habitat-related factors (i.e., vegetation and land use) and livestock densities that possibly influence the breeding and survival of the (unknown) Culicoides species believed to transmit M. perstans in Uganda.
The central longitude and latitude of each school obtained using a hand-held global positioning system (GPS; Garmin eTrex, Garmin, Olathe, KS, United States of America) was utilized to define an area of 1 km radius (representing the coarsest resolution of the environmental data) encompassing the community. Average values of each environmental layer were then extracted using ArcGIS 10.1 spatial analyst extension (ESRI; Redlands, CA, United States of America). Land cover variables, calculated as the number of pixels of each category of land use, was counted within the 1 km buffer zone (using the ‘Geospatial Modeling’ environment extension [14]), and the percentage of each category calculated. For a full description of these environmental variables, see Stensgaard and colleagues [15,16].
Initially, a non-spatial, frequentist bivariate logistic regression analysis was conducted in Stata version 13 (Stata Corporation; College Station, TX, United States of America) to assess the relation between various environmental and habitat-related predictors of M. perstans infection status. Significant candidate factors based on likelihood ratio test (LRT) with significance levels of 5% were selected as covariates in further multivariate analyses. To avoid over-parametrization and confounding arising from correlated environmental variables within the same “environmental theme”, these were ranked by the Akaike information criterion (AIC) [17], and strongly correlated variables (Spearman rank correlation r >0.75) excluded.
Next, Bayesian multivariate non-spatial and geostatistical logistic regression models were fitted in OpenBUGS version 3.1.1. (Imperial College and Medical Research Council; London, United Kingdom) via Markov chain Monte Carlo (MCMC) methods which provide higher flexibility in fitting complex models and avoid asymptotic inference than frequentist approaches, and overcome the computational challenges encountered in likelihood-based fitting [18]. Bayesian geostatistical modeling represents the current leading edge in spatial statistics, and makes it possible to incorporate both spatial dependence and covariates, but also enables full representation of uncertainty in model outputs [19] that can be visualized, for example, as maps of prediction errors.
The association between M. perstans, W. bancrofti, and Plasmodium spp. was assessed using multivariate regression models on a single parasite species with all other parasite species as covariates. Demographic and cluster effects were accounted for at the unit of the school.
We assumed that the M. perstans status Yij of child i at location si, which takes a value of 1 if the child was tested positive and 0 otherwise, follows a Bernoulli distribution Yij ~ Ber(pij), with pij measuring the infection risk at location si. The outcome can be related to its predictors via standard multivariate logistic regression analysis. This model is given by logit(pij)=β0+∑k=1pβkXij(k)+εi where βij = (β0, β1, β2, … βp) is the vector of regression coefficients and the intercept, and Xij=(Xij(1),Xij(2),…Xijp) are the model covariates (the fixed part of the model), and εij is a location-level exchangeable random effect that accounts for clustering of individuals in schools. They are assumed to be independent, arising from a normal distribution (∼N(0, τ2)) where τ2 accounts for the non-spatial variation in the infection risk data.
The spatial relationship often found among parasitemia survey locations was considered by introducing spatially correlated random effects ϕi at every sampled location si, which is the standard way of incorporating geographical dependence in the model. The underlying spatial process was modeled by the residuals using the geostatistical design described in Diggle et al. (1998) [18] via a multivariate normal distribution, ϕ = (ϕ1, …. ϕn)T with variance-covariance matrix Σ. Moreover, an isotropic spatial process was assumed, i.e., Σij = σ2exp(−ρdij), where dij is the Euclidean distance between locations i and j, σ2 is the spatial variation (known as the sill), and ρ is a smoothing parameter controlling the rate of correlation decay with increasing distance. For the exponential correlation function, the minimum distance at which the spatial correlation between locations is less than 5% (range of spatial process) is calculated by 3/ρ for the exponential correlation structure.
To complete Bayesian model specification, independent normal prior distributions was assumed for the regression coefficients, with mean 0 and variance 100. For σ2, τ2, and ρ inverse gamma distributions with mean 1 and variance equal to 100 were adopted. We ran a single chain sampler with a burn-in of 5,000 iterations, followed by 100,000 iterations. Convergence was assessed by inspection of ergodic averages of selected model parameters and convergence was successfully achieved before the 100,000th iteration. The strength of correlations and significance of the co-variates was assessed by inspecting the estimated odds ratios (ORs) and their Bayesian credible intervals (BCI).
For appraisal of the best fitting multivariate model, the deviance information criterion (DIC) was applied [20]. The smaller the DIC, the better the model fit. Bayesian kriging was applied to generate smooth risk maps for M. perstans prevalence based on the parameter estimates of the best fitting model [18].
Children with M. perstans microfilaremia were observed in 47 out of the 76 study sites (61.8%), with prevalence ranging from 0.4% to 72.8%. The highest prevalence was observed at sites south of Lake Albert and north-west of Lake Victoria with prevalence decreasing toward zero when moving to the north-eastern and southern sites (Fig 1).
Of the 12,207 children examined for M. perstans microfilaremia, 11,606 were examined concurrently for infection with W. bancrofti (CFA). Co-infections with M. perstans and W. bancrofti were observed in 33 individuals (0.3%) in six schools (Fig 2).
Four of these six schools were clustered together in the area north of Lake Kyoga that had relatively high W. bancrofti prevalence, but low M. perstans prevalence. However, the few children infected with M. perstans in these schools, also tested positive for W. bancrofti CFA. High levels of mono-infections with M. perstans were primarily found in the southern parts of Uganda, whereas mono-infections with W. bancrofti were restricted to the east-central northern areas of Uganda (Fig 2). Malaria has previously been found to be widespread in Uganda (see Stensgaard et al. (2011) [9] for more details).
Co-infections with M. perstans and Plasmodium spp were observed in 347 of 11,469 examined children (3.0%). Triple-infections with M. perstans, Plasmodium spp, and W. bancrofti were observed in only nine out of 11,267 (0.08%) examined children (in three schools). Co-infections and triple-infections were approximately equally distributed among age groups and sex.
The non-spatial bivariate logistic regression analyses revealed that most of the climatic and environmental predictors were significantly associated with M. perstans prevalence (Table 2).
Diurnal land surface temperature (LST) range (Tmax minus Tmin), which was negatively associated with M. perstans prevalence, showed the best fit among the temperature variables as measured by the AIC. Among the vegetation associated variables, NDVI and the percentage of forest cover around the schools showed positive associations with M. perstans infection status, with NDVI composited over the wet season showing the best fit to the data. Among the livestock density factors, only cattle densities showed a significant (and positive) association with M. perstans infection status. Furthermore, age was a significant risk factor, with three times as high odds of being infected in the oldest age group (14–19 years) as compared to the youngest age group (5–9 years). Sex, on the other hand, was not significant at the 5% significance level and thus not included in subsequent Bayesian models.
In the Bayesian multivariate regression analyses (Table 3), the introduction of exchangeable random effects (model B), improved model performance considerably based on DIC estimates (4,991 vs. 3,543).
The random effect had also an influence on the regression parameters of the covariates, but all covariates remained significant except forest cover. The introduction of location-specific random effect parameters into the model (model C) further increased model performance (DIC 3,543 vs. 3,389) suggesting that this is the best fitting model, while the covariate parameter estimates remained largely unchanged. Models 2 and 3 estimated approximately the same geographic variability σ2 (3.46 vs. 3.65). The estimated spatial range (above which spatial correlation drops below 5%) was 44.3, which is equivalent to about 7.5 km at the Equator.
A predictive M. perstans filariasis risk map for Uganda (Fig 3) was obtained based on the best fitting model, the spatial logistic regression model (model 3).
Highest risks were predicted in the central areas, below Lake Kyoga, with highest prevalence (>20%) south of Lake Albert, and at the northern areas and western shores of Lake Victoria. Intermediate levels (10–20%) were predicted in a belt stretching from south of Lake Albert to the north-eastern shores of Lake Victoria, but with pockets of high risk in the far north-western and south-eastern parts of the country. In contrast, low prevalence estimates (≤1%) were predominantly predicted in the north-east and central-south of Uganda. When interpreting the maps in Fig 3, it should be noted that these are based on model-predictions, and that areas with few survey points may have relatively high levels of associated prediction error. Furthermore, the predictions are based on data from school children only, and thus not necessarily representative for the adult Ugandan population infection levels, which may be considerably higher given the relationship between age and infection risk.
The smooth map of the predicted endemic areas (predicted prevalence >5%) of M. perstans from the present study was super-imposed with a map of predicted endemic lymphatic filariasis (>5% W. bancrofti CFA prevalence) and high risk malaria (>50% Plasmodium spp. prevalence) previously published [8] to delineate areas of co-endemicty (Fig 4).
Results from parasite-parasite association inferred from multivariate logistic regression models revealed a significant positive association between M. perstans microfilaremia and testing positive for W. bancrofti (CFA), when clustering at the unit of schools was accounted for (Table 4). No other significant associations were found.
The present study provided countrywide, model-based prevalence maps for M. perstans in Uganda, at a high spatial resolution. To our knowledge this is the first study to apply rigorous Bayesian geostatistical risk mapping to national survey data of this neglected human parasitic infection. The study furthermore identified risk factors and displayed high prevalence areas, and thus provides new insights into the ecological preferences of the unknown vector (Culicoides spp.). The resulting maps were finally combined with geostatistical risk maps previously developed for bancroftian filariasis and malaria [9], to delineate overlapping areas (co-distributions) and to investigate levels of co-infection and parasite-parasite associations. Overall, the investigations provide a deeper understanding of the zoogeographical patterns of this widespread, yet little studied parasitic infection, of importance for integrated disease control planning and implementation [11].
An increasing number of geospatial applications now analyze the relationship between parasitic infections and environmental factors, to generate predictive risk maps, including uncertainty estimates [21,22]. The majority of these studies have pertained to malaria risk [23–26], but more recently also to a number of NTDs, such as schistosomiasis [27–30], lymphatic filariasis [9], loiasis [31–32], and soil-transmitted helminthiasis [33,34]. Besides being useful for spatial targeting of control measures, surveillance, and measuring progress toward elimination, these studies can also give important new insights and clues about the ecology of the parasites and their vectors or intermediate hosts.
Because of the limited number of studies of M. perstans epidemiology in Africa, the current knowledge about the climatic and other environmental factors that help drive transmission of this filarial parasite is very scarce [2]. Our results indicated that high prevalence of M. perstans in Uganda was associated with cooler areas with little diurnal temperature variation, and with high NDVI values, a surrogate variable for soil moisture and correlated with vegetation biomass. A positive association was also observed with forested areas (although not significant in the final model). An association with forested ecosystems was also observed in Gabon [35,36] and Cameroon [37], and historically by Low (1903) [38], who noticed that high prevalence was associated with tropical forests alternating with swamps and other large, open ground areas. Other reports from West Africa have related the common occurrence of M. perstans along the forest fringes between the rain forest and open land, to the particular species and density of the vectors [39,40], while studies from central/southern Africa similarly indicated high prevalence in or near dense forest [41–44].
The association to forested areas, as well as banana plantations [45,46], has been linked to the importance of decomposing woody material, tree holes, and forest floor cover as breeding sites for the vector species, Culicoides spp. However, besides the affinity for moist substrates the some 1,400 described species of Culicoides show a highly diverse range of species habitat preferences, ranging from salt- and freshwater marshes, to animal dung, water logged pastures, and leaking irrigation pipes [5].
In Uganda, a total of 31 Culicoides species have been listed thus far [47], and several of them (e.g., C. grahamii), which have been identified as vectors of M. perstans in the Congo [48] and Cameroon [39], occur in areas predicted to be endemic for M. perstans [47]. Yet, no studies have been carried out to confirm the role of this or other Culicoides species in the transmission of M. perstans in Uganda.
Here, the identified environmental correlates of M. perstans give us important clues about the bionomics of the unknown vector species. Besides the climatic associations, areas of high prevalence of M. perstans were found to coincide with areas of high cattle densities (but not densities of other types of livestock). This is interesting and calls for further investigation, as M. perstans also has been shown to occur at high prevalence in Fulani nomads (cattle raising people) in northern Nigeria [49]. Possible explanations could be either a role for cattle in providing a steady source of blood meals for an opportunistic biting, or in creating habitats for larval development. Animal dung has been shown to provide important larval habitats for several Culicoides species [50,51] and, C. grahamii, for example, has been incriminated in blue tongue virus transmission among cattle in Kenya [52].
While models and maps of individual parasite infections are valuable, the distribution of these infections rarely occurs independently of each other. Multiple species are often found within populations (co-endemicity) and individuals (co-infection), and co-infections are increasingly being recognized to have important health consequences [53–56]. Concomitant infections with helminths have, for example, been shown to increase susceptibility to infection with P. falciparum [57,58]. Delineating areas of geographic overlap, where co-infections might occur, is thus an important operational issue for integrated disease control planning, implementation, and evaluation.
In Uganda, several NTDs have been reported to be co-endemic [59], and also to be co-endemic with malaria [9]. Here we found that while perstans filariasis was widely overlapping with areas of high malaria risk (Fig 3), there were no significant associations between infections with Plasmodium spp and M. perstans and/or W. bancrofti at the individual level or at the unit of the school. A similar result was found by Nielsen et al. (2006) [60], in a study from north-eastern Tanzania, whereas Kelly-Hope et al. (2006) found a negative spatial association between W. bancrofti and P. falciparum malaria prevalence in West Africa [61].
In contrast, we observed a distinct pattern of geographic segregation between M. perstans and W. bancrofti, another filarial parasite of human health importance in Uganda (Fig 2). While M. perstans was mainly predicted in the central-to-southern parts of the country, W. bancrofti dominated in the central-northern parts. This pattern has been noted previously [10], but this is the first time a co-distribution map based on rigorous geostatistical modeling of individual infections is presented.
This very likely reflects the ecological distinctiveness of the M. perstans Culicoides spp. vector compared to that of the Anopheles mosquitoes transmitting W. bancrofti in Uganda. It is noted, for instance, that while high M. perstans infection risk is related to forested and densely vegetated areas, the opposite seems to apply for W. bancrofti, which showed a negative association with NDVI [9]. Similar contrasting epidemiologies have been shown for other filarial infections in Africa, i.e., between onchocerciasis and loiasis in the Democratic Republic of the Congo (DRC) [62], although L. loa and M. perstans have been found to coexist with high prevalence geographically in some African countries [35,37].
The limited geographical overlap observed in Uganda, explains the relatively low levels of overall filarial co-infection (0.8%). Yet, where the two parasites did overlap in space, in a high prevalent W. bancrofti foci in central Uganda at the northern geographical range margin of M. perstans plus a location south of Lake Albert, with high M. perstans prevalence (Figs 2 and 3), a positive association was observed between the two species. This finding warrants further investigation of potential risk factors for co-infection at the school and individual level in Uganda, and indicates that special attention should be paid to children living in geographically overlapping areas, even if these areas may be few.
In conclusion, this study adds further to our knowledge about the distinct zoogeography of filarial parasites [11] in Africa. The observed correlation between M. perstans prevalence and cattle density warrants further scientific inquiry, particularly the role played by livestock as either opportunistic blood meals (resource) for the Culicoides midges and/or the role of dung as larval habitats. Finally, we urge further studies based on geographically stratified field-collections of Culicoides, to clarify the identity, bionomics, and behavior of the vector species of M. perstans in Uganda and elsewhere in Africa, as this is a vital piece of the puzzle toward a fuller understanding of the transmission cycle and epidemiology of M. perstans infections.
|
10.1371/journal.pntd.0005805 | Mitochondria and lipid raft-located FOF1-ATP synthase as major therapeutic targets in the antileishmanial and anticancer activities of ether lipid edelfosine | Leishmaniasis is the world’s second deadliest parasitic disease after malaria, and current treatment of the different forms of this disease is far from satisfactory. Alkylphospholipid analogs (APLs) are a family of anticancer drugs that show antileishmanial activity, including the first oral drug (miltefosine) for leishmaniasis and drugs in preclinical/clinical oncology trials, but their precise mechanism of action remains to be elucidated.
Here we show that the tumor cell apoptosis-inducer edelfosine was the most effective APL, as compared to miltefosine, perifosine and erucylphosphocholine, in killing Leishmania spp. promastigotes and amastigotes as well as tumor cells, as assessed by DNA breakdown determined by flow cytometry. In studies using animal models, we found that orally-administered edelfosine showed a potent in vivo antileishmanial activity and diminished macrophage pro-inflammatory responses. Edelfosine was also able to kill Leishmania axenic amastigotes. Edelfosine was taken up by host macrophages and killed intracellular Leishmania amastigotes in infected macrophages. Edelfosine accumulated in tumor cell mitochondria and Leishmania kinetoplast-mitochondrion, and led to mitochondrial transmembrane potential disruption, and to the successive breakdown of parasite mitochondrial and nuclear DNA. Ectopic expression of Bcl-XL inhibited edelfosine-induced cell death in both Leishmania parasites and tumor cells. We found that the cytotoxic activity of edelfosine against Leishmania parasites and tumor cells was associated with a dramatic recruitment of FOF1-ATP synthase into lipid rafts following edelfosine treatment in both parasites and cancer cells. Raft disruption and specific FOF1-ATP synthase inhibition hindered edelfosine-induced cell death in both Leishmania parasites and tumor cells. Genetic deletion of FOF1-ATP synthase led to edelfosine drug resistance in Saccharomyces cerevisiae yeast.
The present study shows that the antileishmanial and anticancer actions of edelfosine share some common signaling processes, with mitochondria and raft-located FOF1-ATP synthase being critical in the killing process, thus identifying novel druggable targets for the treatment of leishmaniasis.
| Leishmaniasis is a major health problem worldwide, and can result in loss of human life or a lifelong stigma because of bodily scars. According to World Health Organization, leishmaniasis is considered as an emerging and uncontrolled disease, and its current treatment is far from ideal, with only a few drugs available that could lead to drug resistance or cause serious side-effects. Here, we have found that mitochondria and raft-located FOF1-ATPase synthase are efficient druggable targets, through which an ether lipid named edelfosine exerts its antileishmanial action. Edelfosine effectively kills Leishmania spp. promastigotes and amastigotes. Our experimental animal models demonstrate that oral administration of edelfosine exerts a potent antileishmanial activity, while inhibits macrophage pro-inflammatory responses. Our results show that both Leishmania and tumor cells share mitochondria and raft-located FOF1-ATPase synthase as major druggable targets in leishmaniasis and cancer therapy. These data, showing a potent antileishmanial activity of edelfosine and unveiling its mechanism of action, together with the inhibition of the inflammatory responses elicited by macrophages, suggest that the ether lipid edelfosine is a promising oral drug for leishmaniasis, and highlight mitochondria and lipid raft-located FOF1-ATP synthase as major therapeutic targets for the treatment of this disease.
| Leishmaniasis, caused by several species of the protozoan Leishmania parasite, is one of the world’s most neglected diseases in terms of drug research and development, and for which current therapy is not satisfactory [1]. At present, about 350 million people in 98 countries worldwide are at risk of contracting leishmaniasis, and some 0.9–1.6 million new cases occur yearly, with about 0.7–1.2 million cases of self-healing cutaneous leishmaniasis, which can result in disfiguring skin lesions, and 0.2–0.4 million cases per year of life-threatening visceral leishmaniasis, which is a fatal disease if left untreated [1–3]. Leishmaniasis is the world’s second-deadliest parasitic disease after malaria, with a tentative estimate of 20,000 to 40,000 leishmaniasis deaths occurring annually [3], and has been classed as a category 1 disease (“emerging and uncontrolled”) by the World Health Organization (WHO). At present there are very few available antileishmanial drugs, being in general toxic, and the first line drugs are becoming ineffective due to emerging drug resistance [1, 2]. Thus, the development of novel therapeutic drugs is urgently needed. Leishmaniasis is transmitted by the bite of a female sandfly vector (Lutzomyia in the Americas and Phlebotomus elsewhere) infected with Leishmania parasites. Infection of humans and other animal hosts is initiated by flagellated promastigotes that develop within the digestive tract of the sandfly vector and are injected during a sandfly blood meal. Promastigotes are internalized into a number of phagocytic host cells, including neutrophils, dendritic cells, and macrophages, but proliferate only within the macrophage as aflagellate amastigotes [4, 5].
The so-called alkylphospholipid analogs (APLs) constitute a class of structurally-related antitumor compounds with multiple therapeutic indications, and include a number of clinically relevant and/or promising drugs, such as miltefosine (hexadecylphosphocholine), edelfosine (1-O-octadecyl-2-O-methyl-rac-glycero-3-phosphocholine), perifosine (octadecyl (1,1-dimethyl-piperidinio-4-yl)-phosphate) and erucylphosphocholine ((13Z)-docos-13-en-1-yl 2-(trimethylammonio)ethyl phosphate) (ErPC) (Fig 1) [6–8]. So far, miltefosine is the only APL that has entered the clinic, registered as Impavido, the first orally-effective treatment for visceral leishmaniasis, and as Miltex, a topical chemotherapy and palliative treatment in cutaneous metastases from breast cancer [7, 9, 10]. APLs induce an apoptosis-like cell death in Leishmania parasites [11, 12], but their antiparasitic mechanism of action remains unknown, although lipid metabolism [13] and dramatic increases in membrane dynamics [14] have been suggested to play a role.
The ether lipid edelfosine induces apoptosis in tumor cells, involving cholesterol-rich lipid rafts and mitochondria [15–20]. Lipid rafts are membrane microdomains highly enriched in cholesterol and sphingolipids, and recent findings in mammalian cells suggest that lipid rafts act as death-promoting scaffolds, where Fas/CD95 and downstream signaling molecules are recruited to tigger apoptosis [21–23]. Raft domains have also been described in Leishmania spp., although their biochemical and functional characterization remains incomplete [24].
Here, we analyzed whether our knowledge on processes involved in the anticancer activity of APLs could provide some insight into their antileishmanial mechanism of action. In addition, we tested the effect of APLs on intact and Leishmania-infected macrophages as the main host cells for Leishmania parasites, as well as the in vivo antileishmanial activity of edelfosine in animal models. In this study, we report the existence of common mechanisms that underlie the antileishmanial and antitumor activities of the APL edelfosine, involving mitochondria, lipid rafts and FOF1-ATP synthase (also named as FOF1-ATPase), which might open up new avenues for the development of novel targeted therapies.
Animal procedures were approved by the institutional research commission of the University of Salamanca, and were approved by the Ethics Committee of the University of Salamanca (protocol approval number 48531). Animal procedures complied with the Spanish (Real Decreto RD1201/05) and the European Union (European Directive 2010/63/EU) guidelines on animal experimentation for the protection and humane use of laboratory animals, and were conducted at the accredited Animal Experimentation Facility (Servicio de Experimentación Animal) of the University of Salamanca (Register number: PAE/SA/001).
Edelfosine was from R. Berchtold (Biochemisches Labor, Bern, Switzerland). Miltefosine was from Calbiochem (Cambridge, MA). Perifosine and erucylphosphocholine were from Zentaris (Frankfurt, Germany). Rotenone, malonate, antimycin A, azide, oligomycin and CCCP were from Sigma (St. Louis, MO).
The Leishmania strains used in this study were: L. donovani (MHOM/IN/80/DD8), L. major LV39 (MRHO/SU/59/P), L. panamensis (MHOM/CO/87/UA140), L. infantum (MCAN/ES/96/BCN/150, MON-1), kindly provided by Iván D. Vélez (Programa de Estudio y Control de Enfermedades Tropicales–PECET-, Medellin, Colombia), Ingrid Müller (Imperial College London, London, UK) and Antonio Jiménez-Ruiz (Universidad de Alcalá, Alcalá de Henares, Spain). To visualize Leishmania parasites inside macrophages we used GFP-L. panamensis promastigotes [25, 26]. Leishmania promastigotes were grown at 26°C in RPMI-1640 culture medium (Invitrogen, Carlsbad, CA), supplemented with 10% fetal bovine serum (FBS), 2 mM L-glutamine, 100 U/ml penicillin, and 100 μg/ml streptomycin. Promastigotes were treated at 26°C with the indicated compounds during their logarithmic growth phase (1.5 x 106 parasites/ml). Late stationary promastigotes were obtained after incubation of the parasites for more than 6 days with a starting inoculum of 1 x 106 parasites/ml. Leishmania axenic amastigotes were obtained as previously described [27]. Human acute T-cell leukemia Jurkat (American Type Culture Collection, Manassas, VA), myeloid leukemia HL-60 (American Type Culture Collection), multiple myeloma MM144 (provided by A. Pandiella, CIC, IBMCC, Salamanca, Spain), and cervical carcinoma HeLa (American Type Culture Collection) cell lines, as well as the mouse macrophage cell line J774 (American Type Culture Collection) were grown in RPMI-1640 culture medium supplemented with 10% FBS, 2 mM L-glutamine, 100 U/ml penicillin, and 100 μg/ml streptomycin at 37°C in humidified 95% air and 5% CO2.
Saccharomyces cerevisiae yeast (BY4741 strain: MATa his3Δ1 leu2Δ0 met15Δ0 ura3Δ0) was grown on standard synthetic complete medium (SDC), which consisted of synthetic minimal medium (SD; 0.17% yeast nitrogen base without amino acids, 2% glucose and supplements according to the requirements of the strains) with 0.079% complete supplement mixture (ForMedium, Norwich, UK). Yeast cultures were incubated at 30°C, and growth of cells untreated or treated with edelfosine was monitored by optical density at a wavelength of 595 nm (OD595). Cells were incubated for the indicated times and sample aliquots were taken to measure absorbance at 595 nm. Edelfosine was used at the concentrations indicated in the corresponding figure in liquid medium. The atp7Δ mutant was obtained from the EUROSCARF haploid deletion library in the BY4741 background [28]. This atp7Δ mutant was complemented with the corresponding wild-type gene expressed from a centromeric plasmid (pRS416), and yeast growth was determined as above.
L. infantum transfected with pX63-Neo or pX63-bcl-xL were kindly provided by Antonio Jiménez-Ruiz (Universidad de Alcalá, Alcalá de Henares, Spain) and grown in medium containing 100 μg/ml G418 (Sigma) [29]. HeLa cells (1–2 x 105) were transfected with 2 μg of pSFFV-bcl-xL or pSFFV-Neo expression vector [15], using Lipofectin reagent (Life Technologies, Carlsbad, CA). Transfected clones were selected by growth in the presence of 500 μg/ml G418, and monitored by Western blotting using the 2H12 anti-29 kDa Bcl-XL monoclonal antibody (BD Biosciences PharMingen, San Diego, CA).
Murine bone marrow cells were obtained by flushing out the femurs of mice from (C57BL/6 x BALB/c)F1 (CBF1) mice and cultured as previously described [30] in hydrophobic Teflon bags (Biofolie 25, Heraeus, Hanau, Germany) with DMEM culture medium containing 10% FBS, 5% horse serum, 2 mM L-glutamine, 60 μM 2-mercaptoethanol, 1 mM sodium pyruvate, 1% non-essential amino acids, 100 U/ml penicillin, 100 μg/ml streptomycin, and the supernatant of L929 fibroblasts at a final concentration of 15% (v/v) as a source of colony-stimulating factors which drive cells towards a >95% pure BMM [31].
CBF1 mice were treated orally with edelfosine (5 mg/kg body weight), daily for 13 days in 100 μl PBS, and then BMM were prepared as above. No weight loss or other visible side-effects were observed in mice treated with edelfosine.
Cell viability at the indicated times was measured by the WST-1 reduction to formazan method (Roche Diagnostics, Basel, Switzerland). 105 cells were incubated for 2 h at 37°C with 10 μl WST-1 solution in 0.2 ml DMEM culture medium supplemented with 10% FBS in a flat-bottom microtitre plate, and then absorbance was determined at 440 nm.
The production of superoxide anion (2 x 105 cells in 0.2 ml Hepes-DMEM without pH indicator and containing 125 μM lucigenin, 37°C) was initiated by addition of 50 μg zymosan, and measured as lucigenin-dependent chemiluminescence using a Microlumat LB96P (Berthold, Wildbad, Germany) [32].
NO end product nitrite was measured using the Griess reagent as previously described [32]. Culture supernatant was mixed with 100 μl of 1% sulphanilamide, 0.1% N-(1-naphthyl)ethylenediamine dihydrochloride and 2.5% H3PO4. Absorbance was measured at 540 nm in a microplate reader (Molecular Devices, Ismaning, Germany). LPS from S. abortus equi was kindly provided by Chris Galanos (Max-Planck-Institut, Freiburg, Germany).
IFN-γ was determined by a commercially available (Pharmingen) sandwich ELISA test according to the manufacturer’s protocol.
The ATP content was determined by the luciferin–luciferase method [33]. The assay is based on the requirement of luciferase for ATP in producing light (emission maximum 560 nm at pH 7.8). Briefly, cells (2 x 106) were harvested after treatment, resuspended in 1X PBS, and assayed for ATP using the Molecular Probes ATP determination kit (Thermo Fisher Scientific, Waltham, MA). The amount of ATP in each experimental sample was calculated from a standard curve prepared with ATP and expressed as percentage of the amount of ATP found in untreated control cells.
1.5 x 106 Leishmania spp. promastigotes or axenic amastigotes, and 106 Jurkat cells or other human cells were incubated in the absence or presence of the indicated concentrations of APLs for different incubation times, and then analyzed for DNA breakdown by flow cytometry, using a fluorescence-activated cell sorting (FACS)Calibur flow cytometer (Becton Dickinson, San Jose, CA), as previously described [16]. Quantitation of apoptotic-like cells was monitored following cell cycle analysis as the percentage of cells in the sub-G0/G1 region, representing hypodiploids or apoptotic-like cells [16].
2 x 106 Leishmania parasites and 106 Jurkat cells were pelleted by centrifugation, washed with PBS, incubated in 1 ml PBS containing 20 nM 3,3′-dihexyloxacarbocyanine-iodide (DiOC6(3), green fluorescence; Molecular Probes, Leiden, The Netherlands) and 2 μM dihydroethidine (HE, red fluorescence after oxidation; Sigma) at room temperature and darkness for 20 min, and then analyzed on a Becton Dickinson FACSCalibur flow cytometer as previously described [16].
Macrophages, cultured in RPMI 1640 culture medium containing 10% FBS, were incubated for 4 h with stationary-phase L. panamensis promastigotes at a 10:1 parasite-to-macrophage ratio. Then, cell monolayers were extensively washed and incubated in complete culture medium with or without edelfosine for 24 h. The intracellular parasite load was calculated by limiting dilution assay as previously reported [34]. Alternatively, macrophage monolayers infected with green fluorescent protein (GFP)-expressing p6.5-egfp-L. panamensis parasites were cultured in glass coverslips placed into culture vessels (Corning, Lowell, MA). After 24 h, coverslips were washed, and the rate of intracellular amastigotes and infected macrophages was visualized using a fluorescence microscope. Results are shown as the percentage of infected macrophages and as the parasite/macrophage ratio after counting 100 macrophages.
Four-week-old male Syrian golden hamsters (Mesocricetus auratus) (about 120 g) were obtained from Charles River Laboratories (Lyon, France) and maintained in a pathogen-free facility. Animals were handled according to institutional guidelines, complying with the Spanish legislation, in an animal room with 12-h light/dark cycle at a temperature of 22°C, and received a standard diet and water ad libitum. Hamsters were inoculated intradermally in the nose with 1 x 106 stationary-phase promastigotes in a volume of 50 μl PBS and treated with a daily oral administration of edelfosine (20 mg/kg in water), or an equal volume of vehicle (water) as previously described [26]. Nose swelling was evaluated through weekly caliper measurements, and compared with the nose size before inoculation and treatment. Evolution index of the lesion was calculated as the size (mm) of the lesion during treatment/size of the lesion before treatment. No loss in animal body weight and no sign of morbidity were detected during the 28-day drug treatment, and animals were killed, following institutional guidelines, 24 h after the last drug administration. Parasite burden in the infected tissues was calculated by limiting dilution assay as previously described [26].
Apoptosis-like cell death was also analyzed in situ by the TUNEL technique using the Fluorescein Apoptosis Detection System (Promega, Madison, WI), according to the manufacturer’s instructions. Parasites were fixed with 4% formaldehyde for 20 min on microscope slides, permeabilized with 0.2% Triton X-100, stained for fragmented DNA using the above kit, and then propidium iodide was added for 15 min to stain both apoptotic-like and intact cells as previously described [17, 35, 36]. Propidium iodide stained all cells in red, whereas fluoresecin-12-dUTP was incorporated at the 3’-OH ends of fragmented DNA, resulting in localized green fluorescence within the nucleus of apoptotic-like cells. Samples were analyzed with a Zeiss LSM 510 laser scan confocal microscope (Carl Zeiss AG, Jena, Germany).
L. panamensis and HeLa cells were treated for 1 h with 10 μM fluorescent edelfosine analog all-(E)-1-O-(15’-phenylpentadeca-8’,10’,12’,14’-tetraenyl)-2-O-methyl-rac-glycero-3-phosphocholine (PTE-ET) (Fig 1), kindly provided by F. Amat-Guerri and A.U. Acuña (Consejo Superior de Investigaciones Cientificas, Madrid, Spain) as described [17, 36, 37], and then incubated with 100 nM cell-permeant MitoTracker probe (Molecular Probes) for 20 min to label mitochondria. Colocalization was analyzed by excitation of the corresponding fluorochromes in the same section of samples, using a fluorescence microscope (Axioplan 2; Carl Zeiss MicroImaging, Inc., Oberkochen, Germany) and a digital camera (ORCA-ER-1394; Hamamatsu, Hamamatsu City, Japan).
Parasites (2 x 106/ml) were incubated in serum-free medium with 2.5 mg/ml methyl-β-cyclodextrin (MCD) for 40 min at 26°C, and then washed 3 times with PBS, and resuspended in complete culture before edelfosine addition. For cholesterol depletion in Jurkat cells, 2.5 x 105 cells/ml were incubated with 2.5 mg/ml MCD for 30 min at 37°C in serum-free medium, and then washed 3 times with PBS, and resuspended in complete culture before edelfosine addition.
Drug uptake was measured as previously described [15, 36], after incubating 2 x 106 parasites or 106 Jurkat cells with 10 nmol [3H]edelfosine (10 μM) (Amersham Buchler, Braunschweig, Germany) for 1 h in RPMI-1640, 10% FBS, and subsequent washing (six times) with PBS + 2% BSA. [3H]edelfosine (specific activity, 42 Ci/mmol) was synthesized by tritiation of the 9-octadecenyl derivative (Amersham Buchler, Braunschweig, Germany).
Lipid rafts were isolated from 1 x 108 Leishmania promastigotes or 8×107 Jurkat cells by using nonionic detergent lysis conditions and centrifugation on discontinuous sucrose gradients as previously reported [38, 39]. Twelve 1-ml fractions were collected from the top of the gradient, and 25 μl of each fraction were subjected to sodium dodecylsulfate (SDS)-polyacrylamide gel electrophoresis (PAGE) and assayed for the location of GM1-containing lipid rafts using the GM1-specific ligand cholera toxin (CTx) B subunit conjugated to horseradish peroxidase (Sigma, St. Louis, MO).
The proteomic analysis was performed in the proteomics facility of Centro de Investigación del Cáncer (CIC), Salamanca, Spain, which belongs to ProteoRed, PRB2-ISCIII. Samples (100 μg protein) from pooled fractions enriched in lipid rafts (fractions 3–6 from the sucrose gradient) were precipitated with methanol/chloroform, and then the pellets were resuspended in rehydration buffer (7 M urea, 2M thiourea, 4% CHAPS, 50 mM DTT, 5 mM TCEP, 15 mg DeStreak, 0.5% IPG buffer). Samples were applied to 13 cm IPG strips with a nonlinear pH gradient of 3 to 10 (Amersham Biosciences). Isoelectric focusing was performed at 50 V for 12 hours, 500 V for 1 h, 1000 V for 1 h, a voltage gradient ranging from 1000 to 8000 V for 30 min, and finally 5 h until the voltage reached 35000 V. Strips were treated with SDS equilibration buffer (375 mM Tris-HCl pH 8.8, 6 M urea, 20% glycerol, 2% SDS) plus 2% DTT for 15 min for protein denaturation, and then with equilibration buffer plus 2.5% iodoacetamide for protein alkylation. The second dimension electrophoresis was performed on 10% SDS-polyacrylamide gels. The protein spots were visualized with Sypro Ruby Protein Gel Staining (Invitrogen, Carlsbad, CA).
Spots of interest were automatically excised with Proteineer Spot Picker robotics workstation (Bruker Daltonics, Billerica, MA). The digestion was performed as previously described [40]. For MALDI-TOF peptide mass fingerprinting, a 0.5 μl aliquot of matrix solution (5 g/l 2,5-dihydroxybenzoic acid in 33% aqueous acetonitrile plus 0.1% trifluoroacetic acid) was manually loaded onto a 400 μm diameter AnchorChip Target plate (Bruker Analytic GmbH, Bremen, Germany) probe, and 1 μl of the above peptide extraction solution was added and allowed to dry at room temperature. Samples were analyzed on a Bruker Ultraflex MALDI-TOF mass spectrometer (Bruker-Franzen Analytic GmbH, Bremen, Germany). Each raw spectrum was opened in FlexAnalysis 3.0 (Bruker Daltonics) software and processed and analyzed using the following parameters: signal-to-noise threshold of 1, Savitzky-Golay algorithm for smoothing, tangential algorithm for baseline substraction, and centroid algorithm for monoisotopic peak assignment. In all cases, resolution was higher than 9000. The generated peaks were submitted to Mascot Server (version 2.2, February 2007) [41] using Bio Tools 3.1 (Bruker Daltonics) software, and searched against Uniprot database for human sequences and NCBI database for Leishmania sequences. Search parameters were set as follow: searches were restricted to all sequences for human searches and Other Eukaryota (69482 sequences) for Leishmania searches, up to one missed tryptic cleavage, mass accuracy of 100 ppm, MH+ monoisotopic masses, carbamidomethyl cysteine as fixed modification, and methionine oxidation as variable modification. Mowse scores with a value greater than 65 for human searches and 61 for Leishmania searches were considered as significant (p<0.05).
Data are shown as mean ± SD. Between-group statistical differences were assessed using the Student’s t test. A P-value of <0.05 was considered statistically significant.
First we analyzed the ability of different APLs (Fig 2A and 2B) in promoting apoptosis-like cell death in different Leishmania spp. promastigotes and human cancer cell lines, as assessed by DNA breakdown determined by flow cytometry. Our results showed that APLs ranked edelfosine > miltefosine ≥ perifosine > erucylphosphocholine (ErPC) for their leishmanicidal activity (Fig 2A), and edelfosine > perifosine > miltefosine ≅ erucylphosphocholine (ErPC) for their antitumor activity (Fig 2B), when incubated for 24 h at 10 μM with several Leishmania spp. promastigotes, including L. donovani (visceral leishmaniasis), L. panamensis (cutaneous and mucocutaneous leishmaniasis), and L. major (cutaneous leishmaniasis), or with human cancer cell lines, including myeloid leukemia HL-60 cells, multiple myeloma MM144 cells, and cervical cancer HeLa cells. This drug concentration (10 μM) corresponded to the pharmacologically relevant concentration range of edelfosine in plasma (10–20 μM), previously determined in a number of in vivo and pharmacokinetic studies [19, 42, 43]. We also found that edelfosine was very efficient in promoting cell death in additional human leukemic cell lines, including human T-cell acute lymphoblastic leukemia (T-ALL) cell lines Jurkat (53.4 ± 6.2% apoptosis) and CEM-C7H2 (58.2 ± 5.9% apoptosis). Edelfosine was equally effective against different Leishmania subgenera, including Leishmania Leishmania (L. donovani, L. major) and Leishmania Viannia (L. panamensis) (Fig 2A). The relative difference between the abilities to promote cell death of edelfosine vs. miltefosine was more evident using tumor cells than Leishmania spp. promastigotes, suggesting that processes involved in the mechanisms of action of both drugs are partially conserved, but not identical. For the ensuing studies, we focused our attention on the most effective compound, namely edelfosine, which has been considered as the APL prototype. Edelfosine induced DNA breakdown after 9-h incubation with L. panamensis promastigotes, and the percentage of parasites with a hypodiploid DNA content (sub-G0/G1 cell population) increased with the incubation time (Fig 2C), suggesting an apoptosis-like cell death in Leishmania parasites, similar to the apoptotic response triggered in cancer cells [15, 17, 35, 44]. Edelfosine (5 or 10 μM) also induced apoptosis-like cell death, as assessed by an increase in the sub-G0/G1 population, in L. panamensis axenic amastigotes (Fig 2D).
Because Leishmania are obligate intracellular parasites that infect macrophages within the mammalian host, we examined the location of edelfosine in L. panamensis-infected J774 macrophage-like cells. We have previously found that mouse J774 macrophages were rather resistant to edelfosine [26], and 10 μM edelfosine was unable to mount an apoptotic response after 24-h incubation (<2.5% apoptosis). Using the blue-emitting fluorescent edelfosine analog PTE-ET (Fig 1), a bona fide compound to explore the subcellular localization of edelfosine [17, 19, 36, 37, 45], we found that it was mainly located into the parasites inside the macrophage (Fig 2E), which were visualized by using infective GFP-L. panamensis parasites [25]. Edelfosine treatment highly diminished the amount of infected J774 macrophages and the number of parasites per macrophage (Fig 2F and 2G). Following limiting dilution experiments, we found that edelfosine was the most effective APL, when compared to miltefosine, perifosine and erucylphosphocholine (ErPC), in killing L. major amastigotes in infected mouse BMM (Fig 2H). Edelfosine was highly dependent on its molecular structure for its antileishmanial activity, since a structurally related compound, ET-18-OH (1-O-octadecyl-rac-glycero-3-phosphocholine) (Fig 1), containing a hydroxyl group instead of the methoxy group at the C2 position, was unable to kill Leishmania protozoa (Fig 2H), similarly to what has been found in cancer cells [15, 44].
We have previously shown the potent antitumor activity of orally-administered edelfosine in different xenograft animal models [19, 43, 46]. Recently, we have also found that edelfosine was effective in the treatment of leishmaniasis in different animal models when used at 26 mg/kg body weight [26]. Here, we found that oral treatment of edelfosine at a lower dose (20 mg/kg body weight) exerted a potent in vivo antileishmanial activity in L. panamensis–infected golden hamsters (Fig 2I–2K), an appropriate animal model for reproducing the pathological features of human leishmaniasis [47]. L. panamensis promastigotes were inoculated into the nose of 16 golden hamsters, and then animals were randomly distributed into two cohorts of eight hamsters. Each cohort received a daily oral administration of edelfosine or water vehicle (control) for 28 days. Disease progression was monitored by nasal swelling, determined by serial caliper measurements, and ulceration. Oral treatment with edelfosine led to a dramatic decrease in nasal swelling and parasite load at the site of infection (Fig 2I–2K), and ameliorated the signs of leishmaniasis, leading to an almost normal morphologic appearance (Fig 2K).
Edelfosine has been shown to be taken up preferentially by tumor cells, whereas normal non-malignant cells incorporated a relatively much lesser amount of the ether lipid [15, 17, 35]. Here, we found that normal mouse BMM took up large amounts of edelfosine, at even higher levels than the mouse RAW 309 Cr.1 tumor macrophage cell line (Fig 3A). However, edelfosine induced cell death in the transformed macrophage cell line, but spared BMM (Fig 3B). Interestingly, edelfosine blocked zymosan-induced respiratory burst in BMM (Fig 3C). Furthermore, BMM from mice that were orally treated with edelfosine (5 mg/kg body weight, daily) for two weeks showed a lower capacity to generate superoxide anion, NO and IL-12+IL-18-induced IFN-γ, when compared to BMM from mice treated with water vehicle (Fig 3D–3F). These results suggest that edelfosine treatment decreases macrophage pro-inflammatory responses. Our data, together with the above accumulation of edelfosine into the parasites in Leishmania-infected macrophages, suggest that edelfosine-induced killing of Leishmania is mediated by a direct action of the drug on the parasite, and not via generation of macrophage-derived antiparasitic molecules.
DNA breakdown induced by edelfosine treatment in Leishmania was further assessed by the TUNEL assay, staining all cells in red through the binding of propidium iodide to DNA, and only those cells with fragmented DNA and free 3’-OH ends in green. Interestingly, we detected first kinetoplast-mitochondrial DNA degradation, followed by nuclear DNA fragmentation upon treatment of L. panamensis promastigotes with edelfosine (Fig 4A). These results suggest that the death process induced by edelfosine in Leishmania spp. parasites starts at the mitochondrial level.
Next, we analyzed the subcellular localization of edelfosine in Leishmania promastigotes. The fluorescent edelfosine analog PTE-ET, which has been previously shown to fully mimic the antitumor [17, 19, 36, 37, 45, 48] and antileishmanial [49] actions of the parent drug edelfosine, accumulated mainly in the mitochondria of L. panamensis promastigotes, as indicated by colocalization with the specific mitochondrial marker MitoTracker (Fig 4B). PTE-ET also co-localized with MitoTracker-positive structures in human cervical carcinoma HeLa cells (Fig 4C).
We next examined the time-course of the effect of edelfosine on the following mitochondrial-related processes in L. panamensis promastigotes: a) ROS generation, through the conversion of non-fluorescent dihydroethidine (HE) into red fluorescent ethidium (Eth) after its oxidation via ROS; and b) changes in ΔΨm, through the accumulation of the fluorescent cationic probe DiOC6(3) (green fluorescence), which depends on the mitochondrial potential. As shown in Fig 5A, untreated parasites exhibited a high ΔΨm [(DiOC6(3))high], and the levels of intracellular ROS were undetectable [(HE → Eth)low]. Edelfosine induced first an increase in the percentage of cells with (HE → Eth)high, and then a loss in ΔΨm (Fig 5A). Changes in ROS generation and ΔΨm disruption apparently preceded DNA breakdown. Edelfosine induced Eth staining, i.e. ROS generation, in kinetoplasts, as assessed by using DNA staining to identify L. panamensis nuclei and kinetoplasts (mitochondrial DNA) (Fig 5B). The inhibitor of the mitochondrial permeability transition pore cyclosporin A [50], and the antioxidant agents glutathione (GSH) and N-acetylcysteine (NAC), inhibited edelfosine-induced cell death in L. panamensis promastigotes (Fig 5C). Likewise, cyclosporin A, GSH and NAC inhibited edelfosine-induced apoptosis in human T-cell leukemia Jurkat cells (Fig 5D). Taken together, our data suggest a critical role of mitochondria in the antileishmanial and antitumor activities of edelfosine, and that both ROS generation and ΔΨm disruption are involved in edelfosine-induced cell death in Leishmania parasites and human cancer cells.
We next examined whether the mitochondrial respiration chain was involved in edelfosine-induced ROS production, by using the following mitochondrial respiration inhibitors: rotenone, complex I inhibitor; malonate, complex II inhibitor; antimycin A, complex III inhibitor; azide, complex IV inhibitor; oligomycin, mitochondrial FOF1-ATP synthase inhibitor; and carbonyl cyanide m-chlorophenyl hydrazone (CCCP), one of the most frequently used uncouplers of oxidative phosphorylation [51]. Incubation of cells with CCCP, which disrupts the proton gradient by carrying protons across the mitochondrial membrane and causes mitochondrial depolarization, prompted the generation of ROS in L. panamensis and Jurkat cells, and increased edelfosine-induced ROS generation (Fig 5E and 5F). Incubation of L. panamensis and Jurkat cells for 9 h with rotenone, malonate, antimycin A or azide, affecting the electron transport at specific sites, neither promoted ROS generation nor affected edelfosine-mediated ROS production significantly (Fig 5E and 5F). In contrast, oligomycin, a specific inhibitor of the membranous proton channel (FO) of mitochondrial FOF1-ATP synthase [52], reduced ROS production levels induced by edelfosine in L. panamensis and Jurkat cells (Fig 5E and 5F). Lower concentrations of oligomycin in L. panamensis (1 μM) as compared to Jurkat cells (10 μM) were used, because Leishmania parasites were very sensitive to higher concentrations of oligomycin, resulting in cytotoxicity. Thus, our data suggest that FOF1-ATP synthase plays a role in edelfosine-mediated ROS production in both Leishmania and tumor cells, a process that eventually leads to cell death.
In order to generalize and further support the role of mitochondria in edelfosine-induced cell death in Leishmania parasites and tumor cells, L. infantum and HeLa cells stably transfected with the expression vectors pX63-Neo (Leishmania) and pSFFV-Neo (HeLa), containing the human bcl-xL open reading frame (pX63-bcl-xL and pSFFV-bcl-xL, respectively), were used. L. infantum and HeLa cells transfected with control empty pX63-Neo and pSFFV-Neo plasmids, respectively, were used as controls and behaved similarly to wild-type nontransfected Leishmania promastigotes and tumor cells, regarding edelfosine-induced cell death (Table 1). We found that Bcl-XL ectopic expression in L. infantum promastigotes and HeLa tumor cells inhibited the percentage of hypodiploid cells following edelfosine treatment (Table 1), further supporting the critical role of mitochondria in the induction of apoptosis-like cell death in Leishmania and tumor cells treated with edelfosine. The concentration of edelfosine was increased to 40 μM in the case of L. infantum promastigotes as they were rather resistant to APL treatment [26]. Taken together, these assays support a crucial role of mitochondria in edelfosine-induced cell death in both Leishmania spp. parasites and tumor cells. Thus, our results with two different species of Leishmania, causing distinct forms of disease, namely L. panamensis (cutaneous and mucocutaneous leishmaniasis) and L. infantum (visceral leishmaniasis), converge on the critical role of mitochondria in the killing activity of edelfosine on Leishmania parasites.
Because membrane rafts are a major target in the antitumor action of edelfosine [17–19, 36, 39], we analyzed, in a comparative way, the putative role of lipid rafts in the antileishmanial and anticancer activities of edelfosine. First, we found that raft disruption by preincubation of L. panamensis promastigotes with the cholesterol-depleting agent MCD [53], led to a partial, but statistically significant, inhibition of both edelfosine-induced cell death (Fig 6A) and edelfosine uptake (Fig 6B), suggesting a role for lipid rafts in Leishmania cell death. In addition, cholesterol depletion by MCD strongly inhibited edelfosine-induced cell death and drug uptake in the human T-cell leukemia Jurkat cells (Fig 6A and 6B). It is interesting to note that edelfosine uptake, assessed by the incorporation of [3H]edelfosine, was higher in Leishmania promastigotes than in tumor cells.
The anticancer activity of edelfosine has been shown to depend on the redistribution of lipid raft protein composition [17–19, 36, 38, 39, 54]. Because the above data indicated a remarkable parallelism between the mechanisms of action of edelfosine against Leishmania parasites and leukemic cells, we next isolated lipid rafts from untreated and edelfosine-treated L. panamensis promastigotes by fractionation of low-density detergent-insoluble membranes using discontinuous sucrose density gradient centrifugation. The position of lipid rafts in the sucrose gradients was determined by the presence of ganglioside GM1, detected using the GM1-specific ligand CTx B subunit (Fig 7A), which binds ganglioside GM1 [55], mainly found in rafts [56]. Following a proteomic study of the lipid raft fractions in L. panamensis promastigotes, we found a dramatic translocation to lipid rafts of mitochondrial FOF1-ATP synthase β subunit following drug incubation in L. panamensis promastigotes (Fig 7B and 7C). In addition, we found that oligomycin inhibited edelfosine-induced ΔΨm disruption and cell death in Leishmania (Fig 7D). These data suggest the involvement of FOF1-ATP synthase and its translocation to lipid rafts in the anti-Leishmania activity of edelfosine.
We also isolated lipid rafts from untreated and edelfosine-treated human T-cell leukemia Jurkat cells through sucrose gradient centrifugation, and the fractions enriched in lipid rafts were identified using the GM1-specific ligand CTx B subunit (Fig 8A). Similarly to L. panamensis parasites, a proteomic study of the lipid raft fractions from untreated and drug-treated cancer cells showed a major translocation of mitochondrial FOF1-ATP synthase β subunit to lipid rafts upon drug incubation in Jurkat cells (Fig 8B and 8C). Furthermore, oligomycin inhibited edelfosine-induced ΔΨm loss and cell death in Jurkat cells (Fig 8D).
Because the above data suggested that FOF1-ATP synthase could play a major role in the cytotoxic activity of edelfosine against Leishmania promastigotes and cancer cells, we next sought to obtain genetic evidence for the role of this enzyme in edelfosine cytotoxicity by using yeast as a eukaryotic model organism. We used Saccharomyces cerevisiae yeast atp7Δ mutant, with a deletion in the gene encoding for subunit d of the stator stalk of mitochondrial FOF1-ATP synthase, which is conserved in mammalian cells [57]. Since yeast is a facultative anaerobe and can survive with severely crippled mitochondrial function, we employed S. cerevisiae, which has been previously shown to be sensitive to edelfosine [58], as a good model for genetic ablation assays. We chose the yeast atp7Δ mutant because ATP7 is essential for FOF1-ATP synthase function, but is not essential for growth in yeast. ATP7 deletion leads to a “petite” phenotype that is slow-growing and unable to survive on nonfermentable carbon sources [57].
We found that edelfosine inhibited wild-type yeast growth at 30 and 60 μM (Fig 9A), but atp7Δ mutant strain was resistant at these drug concentrations (Fig 9B). This edelfosine-resistant phenotype was reverted by transformation of the atp7Δ mutant with a centromeric plasmid containing the atp7 wild-type gene (Fig 9C). Taken together, these results strongly support the involvement of FOF1-ATPase in the killing activity of edelfosine.
Leishmaniasis therapy is currently far from satisfactory and search for novel druggable targets and new therapeutic approaches is urgently needed. APLs, originally developed as anticancer agents, have proved to show antileishmanial activity, but their mechanisms of action remain to be fully elucidated. Our data reported here indicate that the APL edelfosine is a promising drug against Leishmania spp. parasites and tumor cells, and unveil common underlying processes in the killing activity of this APL on both Leishmania and cancer cells. Edelfosine killed both Leishmania promastigotes and amastigotes by an apoptosis-like process involving DNA breakdown, and edelfosine oral treatment exerted a potent in vivo antileishmanial activity. We found here that edelfosine killed intracellular Leishmania amastigotes within macrophages, but spared the host cells. Results reported here point out a number of remarkable actions of edelfosine on macrophages, namely: a) normal BMM take up edelfosine, but unlike cancer cells, they are spared; b) edelfosine accumulates in the Leishmania amastigotes inside macrophages; c) edelfosine treatment, in vitro and in vivo, reduces the pro-inflammatory capacity of macrophages. These results suggest that edelfosine kills Leishmania parasites by acting directly on the parasite. Edelfosine and other APLs have been shown to act rather selectively on a wide range of malignant cells, mainly due to their predominant uptake by tumor cells [15, 17–19, 43, 59]. The ability of macrophages to take up edelfosine constitutes the first evidence for the incorporation of edelfosine in a normal resting cell type at similar amounts as in cancer cells. This is of major importance in leishmaniasis because macrophages are the main host cells in Leishmania infection. We have previously found that the fluorescent edelfosine analog PTE-ET accumulated into naïve macrophages, especially around the nucleus, but once naïve macrophages were infected with Leishmania spp., an intense fluorescent signal was detected in the intracellular parasites [26]. These data, together with the findings reported here, indicate that edelfosine is taken up by naïve macrophages in significant amounts, and therefore it might affect some macrophage functions, such as the ones herein described, namely, a decrease in the generation of superoxide anion and nitric oxide, as well as in the production of IL-12+IL-18-induced IFN-γ. Interestingly, when Leishmania parasites enter the macrophage, a substantial amount of drug is translocated to the sites where the parasites are located and then the drug is incorporated into the protozoa [26]. Thus, edelfosine could affect both macrophage functions and Leishmania survival. The finding that edelfosine diminishes the capacity of macrophages to mount an inflammatory response might be relevant, as severe inflammation at the site of infection leads to tissue destruction in leishmaniasis [60]. In this regard, edelfosine has also been reported to display a potent anti-inflammatory action through L-selectin shedding in human neutrophils, thus preventing neutrophil extravasation [31], and recent in vivo evidence further supports the anti-inflammatory and immunomodulatory effect of edelfosine [61–63]. Furthermore, edelfosine promotes apoptosis in mitogen-activated T lymphocytes [64]. On these grounds, edelfosine can affect in different ways the major leukocyte types involved in inflammation, namely neutrophils, macrophages and lymphocytes, thus leading eventually to decreased inflammatory responses.
We have also found here that edelfosine accumulates in mitochondria in both Leishmania parasites and tumor cells, leading to ΔΨm loss and an apoptosis-like cell death. These results agree with recent reports showing a major location of different fluorescent edelfosine analogs in the mitochondria of cancer cells [37, 65]. Our data indicate that edelfosine induces firstly DNA fragmentation in the Leishmania kinetoplast-mitochondrion followed by nuclear DNA breakdown, while cell death in Leishmania parasites and tumor cells can be inhibited by protecting mitochondria through ectopic Bcl-XL expression. These results indicate a critical role of mitochondria in the edelfosine-induced cell killing mechanism in Leishmania parasites and tumor cells. Interestingly, the data reported here suggest that FOF1-ATP synthase plays a principal role in the edelfosine-induced killing activity in both Leishmania parasites and cancer cells. The involvement of the FOF1-ATP synthase in edelfosine cytotoxicity was further assessed through gene deletion experiments conducted in yeast, by showing that the lack of ATP7, which results in a defective FOF1-ATP synthase, inhibited edelfosine toxicity. Drug sensitivity was restored when atp7Δ mutant yeast were transformed with the cognate wild-type gene. Thus, the results shown here strongly indicate by genetic and biochemical approaches that FOF1-ATP synthase is involved in the killing activity of edelfosine in both Leishmania parasites and human tumor cells.
The major role of mitochondria in edelfosine-induced Leishmania killing was further assessed by the generation of ROS in the parasite mitochondrion and the involvement of ROS in edelfosine-induced Leishmania promastigote cell death. Interestingly, edelfosine-induced ROS generation in Leishmania promastigotes was inhibited by oligomycin, an inhibitor of the FO subunit of the mitochondrial FOF1-ATP synthase. Taken together, our data suggest a role for mitochondria and ROS generation in the execution of edelfosine-mediated apoptosis, and oligomycin is able to prevent edelfosine-induced ΔΨm collapse and DNA degradation in both Leishmania parasites and cancer cells. These data highlight a major role of the FO component of the FOF1-ATP synthase in the edelfosine-induced ΔΨm dissipation, ROS generation and cell death. In this regard, the involvement of FOF1-ATP synthase in the apoptotic response induced in glioblastoma cells by erucylphosphomocholine (ErPC3, Erufosine), another APL member, has been suggested [66]. Furthermore, oligomycin has also been reported to suppress TNF-induced apoptosis in human epithelioid carcinoma HeLa cells [67].
The mechanism by which FOF1-ATPase contributes to edelfosine-induced cell death remains to be established. FOF1-ATPase resides in the inner membrane of mitochondria and can pump protons in forward and reverse directions, either pumping protons into the mitochondrial matrix, flowing down their concentration gradient and leading to ATP generation, or pumping protons out of the mitochondrial matrix while hydrolyzing ATP. Because edelfosine affects membrane lipid organization, making membranes more fluid [68, 69], it might be suggested that edelfosine makes the outer membrane more porous, thus favoring the leakage of H+ ions from the outer-inner intermembrane space into the cytosol, which leads to the dissipation of the proton gradient. As a consequence, the FOF1-ATP synthase could run in reverse, that is, hydrolyzing ATP and alkalinizing the matrix by proton extrusion. Because matrix alkalinization has been shown to cause opening of the mitochondria permeability transition pore [70], the FOF1-ATP synthase could facilitate cell death by this mechanism. This explanation has been previously proposed for the effect of oligomycin in inhibiting Bax-induced apoptosis in yeast and mammalian cells [71]. In this regard, we have found that edelfosine treatment led to a reduction in the ATP content of L. panamensis promastigotes (Fig 10).
Furthermore, edelfosine has been reported to act through lipid rafts in human leukemic cancer cells [17–19, 22, 23, 36, 39], and recent evidence suggests a raft-mediated connection between the cell membrane and mitochondria in the action of edelfosine [20, 37, 38]. Here, we have found the involvement of lipid rafts in the antileishmanial activity of edelfosine, and edelfosine treatment led to a dramatic recruitment of mitochondrial FOF1-ATP synthase into rafts in both Leishmania promastigotes and cancer cells. These findings are in line with the identification of lipid rafts in Leishmania parasites [24], and with the impairment of miltefosine action against L. donovani by membrane sterol depletion [72]. The results reported here suggest a redistribution of the FOF1-ATP synthase within the mitochondria or to additional raft domains in other cellular membranes following edelfosine treatment, thus altering the normal function of the enzyme and affecting cell viability. It could also be envisaged that the action of edelfosine on lipid rafts and mitochondria might underlie the inhibition of superoxide anion production in edelfosine-treated macrophages, generated by the NADPH-oxidase located at the macrophage cell membrane [73], and the enhancement of mitochondria-dependent ROS generation in drug-sensitive cells.
The results reported here highlight a major role for mitochondria and lipid rafts in the mechanism of action of edelfosine as both antileishmanial and anticancer drug. Nevertheless, cancer cells seem to be more dependable on lipid rafts than parasites, as shown by the relatively higher inhibition observed in drug uptake and drug-induced cell death when rafts were disrupted by sterol depletion (Fig 6). This putative mechanism of action involving mitochondria, and briefly depicted in Fig 11, seems to be common to both Leishmania parasites and tumor cells. The fact that protecting mitochondria by Bcl-XL ectopic expression leads to an inhibition in drug-induced cell death, further supports the major role of mitochondria and mitochondrial-mediated pathways in the killing activity of edelfosine in both Leishmania parasites and human cancer cells. Previous data on human cancer cells have demonstrated the involvement of mitochondria in the pro-apoptotic activity of edelfosine as an antitumor drug [16, 18, 20, 37, 38, 46, 74], and the results reported here extrapolate this notion to its leishmanicidal activity. In addition, the present results pinpoint the major role of FOF1-ATPase in the killing activity of edelfosine against Leishmania parasites and tumor cells. We have previously found a link between lipid rafts and mitochondria in the mechanism of action of edelfosine [37, 38], suggesting an edelfosine-mediated redistribution of lipid rafts from the plasma membrane to mitochondria [37, 38]. The results reported here indicate that FOF1-ATPase is either translocated to cell surface lipid rafts or to raft domains present in mitochondria. A number of reports have shown the presence of raft-localized FOF1-ATP synthase at the cell surface of several cell types, having been proposed to act as a receptor for different ligands, a proton channel, or a modulator of extracellular ATP levels, involved in numerous biological processes through still unclear mechanisms [75–81]. Our results cannot discern between a cell surface and a mitochondrial localization for the raft-associated FOF1-ATP synthase following edelfosine treatment reported here. Thus, a putative translocation of the mitochondrial FOF1-ATP synthase to the cell surface cannot be ruled out at the moment, and additional experimental approaches should be applied to elucidate the prevailing localization of raft-located FOF1-ATP synthase. However, our present data indicating an accumulation of the ether lipid in the mitochondria of both Leishmania parasites and cancer cells, lead us to suggest that a plausible mechanism could involve the translocation of edelfosine from the plasma membrane to the mitochondria where it would ultimately exert its cytotoxic activity promoting the accumulation of FOF1-ATP synthase into mitochondrial rafts, and leading to the dissipation of the mitochondrial membrane potential, ROS generation, and eventually cell demise (Fig 11). The fact that kinetoplast-mitochondrion was the first organelle where ROS metabolites were generated and DNA was broken down, preceding nuclear DNA fragmentation, points out the critical role of mitochondria as a major target in the search for effective drugs to treat leishmaniasis. It is tempting to suggest that a link between lipid rafts and mitochondria could lead to interesting hints to unveil a novel framework in both Leishmania and cancer therapy. The present data also indicate that our insight on how edelfosine works as an antitumor drug can be of great aid to and give valuable information to uncover the mechanism of action of its leishmanicidal activity, which could be hypothetically extrapolated to other antileishmanial drugs, and might be of inspiration to further identify potential common therapeutic targets in cancer and leishmaniasis.
Taken together, our data indicate that the edelfosine antileishmanial and antitumor mechanisms of action share similar molecular processes, involving mitochondria, lipid rafts and FOF1-ATPase. This study provides a molecular explanation on how the antitumor drug edelfosine acts as an antileishmanial agent, and highlights that mitochondria, lipid rafts and FOF1-ATPase act as major players in cell death modulation, opening new avenues for therapeutic intervention in leishmaniasis and cancer. Our results show that the ether phospholipid edelfosine can be a promising orally administered therapeutic agent and a lead compound in the search for novel and much-needed antileishmanial agents, and identify lipid rafts, mitochondria and FOF1-ATPase as appealing new antileishmanial targets. Furthermore, the results shown here indicate that edelfosine is very effective in killing different species of Leishmania parasites, as well as in distinct developmental stages, such as promastigotes and amastigotes. Interestingly, recent data have shown an increasing rate of relapse against miltefosine and a decline in its efficacy [82–85], which could correspond to the readiness in acquiring experimental drug resistance to miltefosine in vitro [86–88]. We have previously shown that edelfosine is less prone to lead to drug resistance development than miltefosine, and displays a higher antileishmanial activity than miltefosine against a wide variety of Leishmania spp. [26]. Thus the potent leishmanicidal activity of edelfosine, together with its low toxicity profile [31], warrants further clinical evaluation for this ether lipid as a possible therapeutic agent against different forms of leishmaniasis.
|
10.1371/journal.pntd.0003644 | Structural Analysis of the Synthetic Duffy Binding Protein (DBP) Antigen DEKnull Relevant for Plasmodium vivax Malaria Vaccine Design | The Plasmodium vivax vaccine candidate Duffy Binding Protein (DBP) is a protein necessary for P. vivax invasion of reticulocytes. The polymorphic nature of DBP induces strain-specific immune responses that pose unique challenges for vaccine development. DEKnull is a synthetic DBP based antigen that has been engineered through mutation to enhance induction of blocking inhibitory antibodies. We determined the x-ray crystal structure of DEKnull to identify if any conformational changes had occurred upon mutation. Computational and experimental analyses assessed immunogenicity differences between DBP and DEKnull epitopes. Functional binding assays with monoclonal antibodies were used to interrogate the available epitopes in DEKnull. We demonstrate that DEKnull is structurally similar to the parental Sal1 DBP. The DEKnull mutations do not cause peptide backbone shifts within the polymorphic loop, or at either the DBP dimerization interface or DARC receptor binding pockets, two important structurally conserved protective epitope motifs. All B-cell epitopes, except for the mutated DEK motif, are conserved between DEKnull and DBP. The DEKnull protein retains binding to conformationally dependent inhibitory antibodies. DEKnull is an iterative improvement of DBP as a vaccine candidate. DEKnull has reduced immunogenicity to polymorphic regions responsible for strain-specific immunity while retaining conserved protein folds necessary for induction of strain-transcending blocking inhibitory antibodies.
| Plasmodium vivax is an oft neglected causative agent of human malaria. It inflicts tremendous burdens on public health infrastructures and causes significant detrimental effects on socio-economic growth throughout the world. P. vivax Duffy Binding Protein (DBP) is a surface protein that the parasite uses to invade host red blood cells and is a leading vaccine candidate. The variable nature of DBP poses unique challenges in creating an all-encompassing generalized vaccine. One method to circumvent this problem is to synthetically engineer a single artificial protein antigen that has reduced variability while maintaining conserved protective motifs to elicit strain-transcending protection. This synthetic antigen is termed DEKnull. Here, we provide structural and biochemical evidence that DEKnull was successfully engineered to eliminate polymorphic epitopes while retaining the overall fold of the protein, including conserved conformational protective epitopes. Our work presents validation for an improved iteration of the DBP P. vivax vaccine candidate, and provides evidence that protein engineering is successful in countering DBP polymorphisms. In doing so, we also lay down the foundation that engineering synthetic antigens is a viable approach and should be considered in future vaccine designs for pathogens.
| Plasmodium vivax is a causative agent of malaria, inflicting significant morbidity and impeding economic growth in highly endemic areas [1,2]. Increasing evidence indicates the severity of disease, economic impact, and burden of P. vivax has been severely underestimated [1,2]. Among the proposed methods for disease control, vaccines are appealing for a multitude of reasons. Vaccines are cost-effective, efficient, and have been historically successful in combating infectious diseases especially in resource poor environments [3]. Individuals living in regions with P. vivax develop naturally acquired protective immunity and antibodies isolated from those naturally immune have anti-DBP inhibitory effects that correlate with results from in vitro functional assays [4–6].
Establishment of a successful host infection necessitates specific receptor-ligand interactions between host red blood cells and Plasmodium parasites [7]. For P. vivax, the critical interaction is that between the merozoite Duffy binding protein (DBP) and the Duffy antigen receptor for chemokines (DARC) on reticulocytes. DARC-negative individuals are resistant to clinical P. vivax infection, and naturally immune individuals can possess anti-DBP antibodies that inhibit the DBP-DARC interaction and prevent parasite growth [6,8–12]. Additionally, polyclonal antibodies elicited by recombinant DBP exhibit similar protective and inhibitive effects to naturally acquired antibodies [6,11,13,14]. Certain isolates of P. vivax have been reported to invade Duffy-negative cells [15]. However, sequencing of these isolates identified a gene encoding a DBP paralog suggesting the increased copy number and/or expression of DBP may enable invasion into Duffy-negative cells [16]. Together, this highlights the central importance of the DBP-DARC interaction in P. vivax infection and presents DBP as a crucial parasite protein that can be developed as a vaccine target.
DBP is a member of the Duffy binding-like erythrocyte binding protein (DBL-EBP) family, and binds DARC through a conserved cysteine-rich DBL domain known as region II (DBP-II) [17–22]. DBP-II engages DARC through a multimeric assembly mechanism where two DBP-II domains initially bind one DARC to form a heterotrimer that rapidly recruits a second DARC to form a heterotetramer [23–26]. DBP-II amino acids F261-T266, L270-K289, and Q356-K367 form critical contacts with the DARC ectodomain during this process [23]. This receptor-induced ligand-dimerization model is conserved amongst other members of the DBL-EBP family and provides spatial orientation for DBL domains at the parasite-RBC membrane interface [24–30]. Residues that mediate multimeric assembly are important targets of protective immunity as the epitopes of naturally acquired anti-DBP-II antibodies that disrupt the DBP-DARC interaction localize to residues at the dimerization interface, DARC binding pockets, and the RBC proximal face of DBP-II [10]. However, clusters of highly polymorphic residues flank these protective epitopes, which is a pattern seen in pathogens undergoing selective pressure that results in an immune evasion where allelic variants can escape immunity elicited by a previous infection [10,21,26,31–37]. Therefore, polymorphic residues of DBP appear to have a high potential to be the basis of strain specific immune responses that misdirects immune responses away from conserved targets of broadly neutralizing protection. Although strain specific immunity can be protective these seemingly more immunogenic epitopes offer limited value because of the strain-limited nature of the immunity. Genetic analysis of DBP-II alleles reveal a high dN/dS ratio often seen when selection pressure drives allelic diversity as a mechanism for immune evasion [38–42]. In order to proceed with DBP as a P. vivax vaccine target, it is therefore critical to address the challenges presented by polymorphism and immune misdirection inherent in this allelic diversity.
Immunization with DBP-II elicits weakly reactive and allele specific immune responses, a far cry from the end objective of inducing strain-transcending protection [38]. The poor protectivity appears to be due in part to polymorphic non-functional residues diverting the immune response away from the more conserved, less immunogenic, critical receptor binding residues [10,38,43–45]. Consistent with this view, the most polymorphic region, identified as the DEK epitope, is positioned immediately adjacent to the conserved DARC-binding groove (Fig. 1) [10,23]. Antibodies to the DEK epitope can disrupt DBP function, but inhibition is strain limited. Therefore, we refer to DEK as a decoy epitope that distracts the immune response away for more conserved functional epitopes that could serve as basis of a broadly neutralizing protective immunity. To overcome this inherent deficiency of DBP as an immunogen, a novel synthetic DBP-II antigen termed DEKnull was engineered where the polymorphic residues that comprise the DEK epitope were mutated to amino acids not usually present (Fig. 1, S1 Fig) [38]. These proof of principle studies demonstrated the feasibility of redirecting the immune response to conserved, critical residues by eliminating polymorphic epitopes with the goal to create a vaccine that induces a greater percentage of protective antibodies to more conserved, less immunogenic epitopes. Indeed, anti-DEKnull sera lost reactivity towards the polymorphic patch as predicted, but still retained the ability to generate inhibitory antibodies, including epitopes reactive to naturally-occurring immune antibodies of persons infected with P. vivax [38,46]. DEKnull also induced strong anamnestic responses that were protective and cross-reactive against a panel of different DBP-II alleles [5]. Furthermore, DEKnull produced a more consistent inhibitory profile across variants [46].
However, mutation can alter the three-dimensional structure of a protein that in turn would alter the available epitopes presented in a synthetic antigen. This study presents the structure of a synthetic Plasmodium antigen and its implications for the future of vaccine design in targeting malaria. We determined the structure of DEKnull to identify if any shifts in fold and secondary structure or sub-domain rearrangements had occurred, and whether these changes affect DEKnull’s potential as a vaccine surrogate for native alleles [26]. The effects of mutating the DEK polymorphic patch on conserved protective epitopes was identified by comparison with the pre-existing Sal1 structure [26]. We examined and compared the epitope profile of DEKnull to DBP-II using computational approaches as well as through interrogation with a panel of DBP monoclonal antibodies [47]. Together these studies inform future efforts to guide the rational design of the next iteration of a synthetic DBP-II antigen to improve its immunogenicity and ability to mount a thoroughly protective response.
DEKnull was obtained by oxidative refolding. Inclusion bodies expressed in E. coli were solublized in 6 M guanidinium hydrochloride and refolded via rapid dilution in 400 mM L-arginine, 50 mM Tris pH 8.0, 10 mM EDTA, 0.1 mM PMSF, 2 mM reduced glutathione, and 0.2 mM oxidized glutathione. Refolded protein was captured on SP Sepharose Fast Flow resin (GE Healthcare), eluted with 50 mM MES pH 6.0, 700 mM NaCl, and dialyzed overnight in 50 mM MES pH 6.0, 100 mM NaCl. The protein was subsequently purified by sequential size exclusion chromatography (GF200) and ion exchange chromatography (HiTrapS). Protein was finally buffer exchanged into 10 mM HEPES pH 7.4, 100 mM NaCl with size exclusion chromatography. Sal1 DBP-II was purified similarly as DEKnull, but without overnight dialysis.
DEKnull crystals were grown by hanging-drop vapor diffusion. First, 1 μL of protein solution at 3–9 mg/mL was mixed with 1 μL of reservoir containing 0.2 M di-sodium tartrate, 20% PEG 3350 to create needle clusters. Crystals were shattered and microseeded into a mix of 1 μL of protein solution at 4 mg/mL and 1 μL of reservoir containing 0.2 M lithium chloride, 20% PEG 3350. Large needle rods of DEKnull grew within a week and were flash frozen in liquid nitrogen. Data was collected to a resolution of 2.1 Å at beamline 4.2.2 of the Advanced light Source, Lawrence Berkeley National Laboratory and processed with XDS [48].
The DEKnull structure was solved by molecular replacement in PHASER [49] using a single Sal1 DBP-II domain from 3RRC as a starting model. Manual rebuilding in COOT [50] and refinement in PHENIX led to a final refined model with final R-factor/R-free of 21.77%/25.88% with good geometry as reported by MOLPROBITY [50–52]. The MOLPROBITY score of 0.81 places this structure in the top 100th percentile of structures 1.85–2.35 Å. 98.22% of residues lie in favored, 1.78% of residues lie in additionally allowed, and 0% lie in disallowed regions of the Ramachandran plot. Atomic coordinates and structure factors have been deposited into the Protein Data Bank with accession code 4YFS.
The ELISAs were performed as previously described [28]. Briefly, BSA, Sal1 DBP-II, and DEKnull were coated on the plate overnight at 4°. The plates were washed with PBS/Tween-20 and then blocked with 2% BSA in PBS/Tween-20 for one hour at room temperature. The plates were washed with PBS/Tween-20 and then incubated with anti-DBP antibodies (2C6, 2D10, 2H2, 3C9, 2F12, 3D10) individually for one hour at room temperature. The plates were again washed with PBS/Tween-20 and then incubated with an anti-mouse secondary antibody conjugated to Alexafluro-488 for 30 minutes at room temperature. After a final wash step, the fluorescence was measured using a POLARstar Omega (BMG Labtech) plate reader.
We obtained the crystal structure of the DEKnull antigen to a resolution of 2.1 Å (Table 1). DEKnull maintains the overall fold and conserved disulfide bonding patterns of a DBL domain similar to that found in P. vivax DBP Sal1, from which DEKnull is derived [23,26]. The DBL fold is a conserved structural feature in other important Plasmodium adhesion proteins, including the P. falciparum EBA-175 and EBA-140, P. knowlesi α-DBP protein, and the NTS-DBL1α1, DBL6ε, and DBL3x domains of PfEMP-1 (Fig. 2A) [26,27,53–58]. DEKnull also retains the characteristic three sub-domain architecture of DBL domains with critical intra-domain disulfide bonding patterns (Fig. 2B). Sub-domain 1 (S1) contains residues K215 to L253 with two disulfide bonds, C217-C246 and C230-C237. Sub-domain 2 (S2) contains residues H262 to E386 and has a single disulfide bond C300-C377. Sub-domain 3 (S3) contains residues P387 to S508 and has three disulfide bonds: C415-C432, C427-C507, and C436-C505. All cysteines in DEKnull are involved in disulfide bonding and are structurally conserved with Sal1 DBP-II [26].
Alignment of DEKnull and Sal1 DBP-II structures shows minimal differences with an overall root-mean-square (r.m.s.) deviation of 0.435 Å (Fig. 2C), indicating there is minimal differences overall between the native and engineered domains. S1 alignment has a r.m.s. deviation of 0.308 Å and is not significantly different (Fig. 2D). S2 alignment has a r.m.s. deviation of 0.288 Å, and the only change is the region comprising K366 to I376, which is now structured in DEKnull as compared to Sal1 DBP-II (Fig. 2D). S3 alignment has a r.m.s. deviation of 0.310 Å and show shifts in loops G417 to D423 and K465 to T473, changes that can be attributed to solvent exposed flexible loops (Fig. 2D). Strikingly, the DEKAQQRRKQ polymorphic stretch within S2 overlaps well between DEKnull and Sal1 DBP-II. Alteration of these amino acids to ASTAATSRTS had no affect on the secondary structure nor do they shift peptide backbone Cαs (Fig. 3A, 3B).
The dimer interface and DARC binding residues play important roles in host-receptor binding [23,26]. These functional regions are recognized by naturally acquired antibodies that block the DBP-DARC interaction [10,23]. Any DBP-II based synthetic antigen must accurately replicate the three-dimensional conformation of these regions for antibody generation and epitope recognition. We therefore examined if the changes in DEKnull altered these important functional regions. The dimerization and DARC binding surfaces overlap well with the parental Sal1 DBP-II; there is no allosteric change to secondary structure or peptide backbone Cαs, retaining the conformational shape of protective targets (Fig. 3C, 3D). Furthermore, Define Secondary Structure of Proteins (DSSP) analysis assigns identical secondary structure elements between that of Sal1 DBP-II and DEKnull [59,60]. Together, these structural data demonstrate that the DEKnull conformation is not significantly different from that of the naturally occurring allele, except for the polymorphic DEK epitope, and supports the development of DEKnull as a DBP vaccine.
B-cell epitopes fall within two classes: linear and conformational. Linear epitopes are continuous stretches of amino acids in which the primary structure alone is responsible for immunogenicity and antibody recognition. Conformational epitopes can be continuous or discontinuous, but require a fold for immunogenicity and antibody binding. Ablation of the fold through the use of denaturants eliminates antibody recognition of conformational epitopes. While vaccines are able to induce either class, natively folded antigens tend to have a bias towards inducing conformational-dependent antibodies that are protective [61,62]. As a result, it is important to identify and characterize inhibitory and non-inhibitory epitopes on Sal1 DBP-II.
Bioinformatic B-cell epitope prediction methods for conformational epitopes are a powerful tool that can aid in the rational design and analysis of vaccine targets. DiscoTope is a widely used web-based computational algorithm that focuses on identifying potential discontinuous conformational epitopes based on available crystal structures [63]. DiscoTope analysis of Sal1 DBP-II identifies several distinct epitopes with the strongest signal located at the DEKAQQRRKQ polymorphic patch that is altered within DEKnull (Fig. 4A). The predicted residues are all solvent exposed and are spread across the entire surface of the protein, with no discernible predilection for certain sub-domains (Fig. 4B). DEKnull is predicted to have similar patches of epitopes, but lacks the signal at the DEK location induced by the mutational changes (Fig. 4A, 4C). Comparisons between the Sal1 DBP-II and DEKnull prediction results demonstrate only the DEKAQQRRKQ region is significantly different (Fig. 4A). An important concern of removing decoy-epitopes through mutation is the possibility of introducing novel epitopes caused by the amino acid changes. DiscoTope analysis determines that no new epitopes specific to DEKnull are introduced further demonstrating that DEKnull is a suitable surrogate antigen from native alleles of DBP-II.
The structural and computation approaches indicate that there are no signification changes to epitopes in DEKnull with the exception of the mutated DEKAQQRRKQ epitope. We sought to independently assess the DEKnull antigen retained recognizable epitopes by interrogation with a panel of conformationally dependent anti-Sal1 DBP-II antibodies [47] in ELISA assays. Two non-inhibitory and four inhibitory antibodies were probed; all six antibodies showed no difference in antigen recognition between that of Sal1 DBP-II and DEKnull (Fig. 5). This provides evidence that the DEKnull mutations have minimal effect on the overall structural fold of the protein, and are consistent with the antigenicity results seen in the DiscoTope B-cell epitope prediction (Fig. 4). It is interesting to note that two non-inhibitory antibodies, 3D10 and 2F12, bound to both DBP-II Sal1 and DEKnull equally well (Fig. 5). This suggests that DEKnull still retains at least one other immunogenic region that may continue to function in immune evasion, necessitating further development of DEKnull as a vaccine candidate.
The central role of P. vivax DBP and the necessity of DARC recognition in parasite invasion during the asexual red blood stage makes it an ideal vaccine target [8]. Anti-DBP antibodies isolated from naturally immune individuals and those generated through immunization are able to prevent DBP-DARC interactions and inhibit parasite growth [6]. However, the inherent polymorphic nature of DBP poses challenges that must be overcome in order to maximize its effectiveness as a vaccine [39,40]. Polymorphic immunodominant epitopes divert the immune system away from weakly immunogenic protective epitopes that are conserved across alleles, resulting in strain-specific responses as opposed to strain-transcending protection [43,45]. This is seen not only with DBP, but is an inherent problem observed with other Plasmodium vaccine candidates wherein single allele vaccinations often provide strain-specific inhibition but are yet susceptible to alternate alleles [4,10,64–72].
Currently two parallel strategies exist to enhance DBP as a vaccine candidate and to bypass the issue of polymorphism—a multi-allele vaccine composed of variants found in endemic areas, and a modified vaccine that directs immune responses towards conserved epitopes in order to impart broad protection [46,66,73,74]. The synthetic antigen DEKnull is the brainchild of the latter, an antigen in which a dominant variant B-cell epitope is mutated from the parent Sal1 allele [38]. Vaccination studies with DEKnull demonstrate early proof-of-concept success in manipulating the immune system towards protective responses [5,46]. Further iterations in design are expected to improve immunogenicity, protectivity, and cross-reactivity [5,46].
Here, we present the first structure of DEKnull, a synthetic Plasmodium vaccine candidate. These results demonstrate that the DEKnull antigen has insignificant structural changes relative to the parent Sal1 structure [26]. There are virtually no differences in overall DBL fold, orientations of sub-domains 1–3, disulfide bonding, or within the secondary structure and backbone of the mutated region itself (Fig. 2, Fig. 3). The conservation of DBL fold in DEKnull is confirmed with immunological assays examining reactivity against a panel of conformational dependent α-DBP-II(Sal1) antibodies [47]. Of the six antibodies tested, none had significant binding differences between Sal1 and DEKnull (Fig. 5).
The structure of DEKnull additionally allowed us to perform state-of-the-art bioinformatic B-cell epitope analysis through the use of DiscoTope [63]. The prediction results are significant for several reasons. First, the strong signal of the DEK polymorphic patch on the DBP-II Sal1 allele supports that it is strongly immunogenic and can divert immune responses away from conserved protective epitopes. Second, the loss of DEK antigenicity in DEKnull compared to Sal1 further reflects a success in synthetic antigen design in achieving the desired manipulation of epitopes. Third, the DEK mutation did not confound the design of the synthetic antigen by introducing novel epitopes. And finally, the conservation of the remaining epitopes between Sal1 and DEKnull indicates that the mutation does not change the protein’s overall epitope profile suggesting protective epitopes have been retained.
The results first and foremost reflect a success in the strategy of using a modified antigen to bypass DBP allele polymorphisms and poor-protectivity induced by strain-specific epitopes. This study demonstrate that antigen engineering to focus the immune response to conserved functional regions, such as the DARC binding residues and/or DBP dimer interface, is a viable and practical approach. The predicted dominant variant B-cell epitope was eliminated without affecting immunogenicity of the remaining epitopes. Furthermore, the results presented here build upon previous works to establish that protein engineering is a viable approach towards problematic multi-allelic vaccine targets and should guide future vaccine design in other pathogens [5,38].
It has been shown that preliminary immunogenicity studies with DEKnull elicited an immune response comparable to Sal1 DBP-II [5]. A next key step in evaluating DEKnull as a vaccine construct is to demonstrate that DEKnull is able to generate highly potent antibodies that are broadly protective across multiple strains. As a corollary, and one that is predicted in silico by DiscoTope results presented here (Fig. 4), DEKnull must also not generate DEKnull-specific antibodies that would be useless against natural alleles.
The ELISA data presented show that DEKnull still possess non-inhibitory epitopes (Mab 3D10 and Mab 2F12, Fig. 5). Characterizing these antibodies will give insight towards the design of future versions of DEKnull. A continual process of eliminating non-protective epitopes from this synthetic antigen will better focus immune responses towards protective targets. Future studies will examine further iterations of DEKnull to improve upon its overall immunogenicity, broad-spectrum inhibitory profile across different P. vivax DBP alleles, as well as to address the antigenicity of remaining non-protective epitopes.
|
10.1371/journal.ppat.1001335 | HYR1-Mediated Detoxification of Reactive Oxygen
Species Is Required for Full Virulence in the Rice Blast Fungus | During plant-pathogen interactions, the plant may mount several types of defense
responses to either block the pathogen completely or ameliorate the amount of
disease. Such responses include release of reactive oxygen species (ROS) to
attack the pathogen, as well as formation of cell wall appositions (CWAs) to
physically block pathogen penetration. A successful pathogen will likely have
its own ROS detoxification mechanisms to cope with this inhospitable
environment. Here, we report one such candidate mechanism in the rice blast
fungus, Magnaporthe oryzae, governed by a gene we refer to as
MoHYR1. This gene (MGG_07460) encodes a glutathione
peroxidase (GSHPx) domain, and its homologue in yeast was reported to
specifically detoxify phospholipid peroxides. To characterize this gene in
M. oryzae, we generated a deletion
mutantΔhyr1 which showed growth inhibition with
increased amounts of hydrogen peroxide (H2O2). Moreover,
we observed that the fungal mutants had a decreased ability to tolerate ROS
generated by a susceptible plant, including ROS found associated with CWAs.
Ultimately, this resulted in significantly smaller lesion sizes on both barley
and rice. In order to determine how this gene interacts with other (ROS)
scavenging-related genes in M. oryzae, we compared expression
levels of ten genes in mutant versus wild type with and without
H2O2. Our results indicated that the
HYR1 gene was important for allowing the fungus to tolerate
H2O2
in vitro and in planta and that this ability
was directly related to fungal virulence.
| Reactive oxygen species (ROS) are antimicrobial compounds and also serve as
stimulators and products of plant defense reactions. ROS appear to be active in
the critical zone where pathogens and plants come in contact. Therefore,
understanding the source, the role, and the destination of ROS in each
interacting partner will be crucial for understanding the pathogen-host
molecular battle. In this study, we focused on one potential fungal mechanism
for ameliorating effects of plant-produced ROS during the early stages of
infection. Characterizing the MoHYR1 gene from the rice blast
fungus Magnaporthe oryzae, suggested that
MoHYR1 was involved in overcoming plant defense-generated
ROS. The deletion of this gene caused a virulence defect in M.
oryzae, which we believe was associated with the mutant's
inability to detoxify plant-generated ROS. Together, our data suggested that
HYR1 is a virulence factor in the rice blast pathogen, and its role in virulence
was directly related to sensing and managing plant-generated ROS during early
infection events. HYR1 is part of a ROS scavenging and sensing pathway that is
well characterized in yeast, and our study is the first to examine this
important gene in filamentous fungi.
| Molecular oxygen, itself relatively nontoxic, is important to most living organisms
on this planet. However, its derivatives, reactive oxygen species (ROS), can lead to
oxidative destruction of cells [1]. For example, in mammals, ROS can accelerate aging by
making holes in membranes, or by stealing electrons from DNA, which may result in
cancer and other severe diseases [2]. However, animals, plants and fungi have all adapted to
use ROS as key signaling molecules [3]. In plants, ROS play a more
positive role as a defense mechanism against attacking pathogens, and are often
produced as a first line of defense [4]. In the plant-pathogenic fungus, Magnaporthe
oryzae, ROS regulation plays important roles in both development and
virulence. ROS itself has been shown to accumulate in the developing and mature
appressorium, or fungal penetration structure, while the two NADPH oxidases in
M. oryzae, NOX1 and NOX2 are
required for proper development of appressoria, as well as full virulence [5]. The catalase gene
family member, encoded by CATB, was shown to also be involved in
cell wall integrity as well as virulence, as deletion mutants were altered in
hyphal, spore and appressorial morphology [6]. Organisms, therefore, must
carefully balance the toxic effects of ROS and the need for ROS in cellular
signaling.
There are five major types of ROS in plants: superoxide
(O2−), hydrogen peroxide
(H2O2), hydroxyl radical (OH), nitric oxide (NO), and singlet
oxygen (1O2). In plant cells, organelles with an intense rate
of electron flow or high oxidizing metabolic activity are major sources of ROS
generation [7].
These organelles include mitochondria, chloroplasts and peroxisomes. ROS are also
generated via enzymatic sources, such as membrane-associated NADPH oxidases, cell
wall peroxidases and oxalate oxidases [8].
ROS play a crucial role during plant defense responses. Oxidative bursts have been
detected when plant cells are inoculated with biotrophic pathogens [9], hemi-biotrophic
pathogens [10],
necrotrophic pathogens [11], and pathogen elicitors [12]. More recent studies with
M. oryzae that causes rice blast disease, demonstrated that
rice produces H2O2 shortly after inoculation with a virulent
strain of the fungus [13], [14]. The toxic effects of ROS can directly kill pathogens,
and as a result, pathogens have developed counter measures [5]. The coexistence of hosts and
pathogens side-by-side determines that the increase of resistance in a host will be
balanced by the change of virulence in a pathogen, and vice versa. A metabolite
fingerprint study of three rice cultivars infected by M. oryzae
provided evidence for suppression of plant-associated ROS generation during
compatible interactions [9]. Fungal-produced catalase was secreted during infection,
and appeared to play a role in breaking down the plant-produced
H2O2, allowing the disease cycle to occur; in the absence
of catalase, infection was largely blocked by the plant's ROS [15].
ROS production and mitigation is a multifaceted process, incorporating many genes and
pathways [1]. One
mechanism of sensing and ultimate detoxification of ROS in yeast is via the
Hyr1 gene, formerly termed
Gpx3/Orp1; this gene, upon ROS induction,
activates its partner protein yAP1, which is a bZip transcription factor involved in
activating cellular thiol-redox pathways, and arguably one of the most studied
ROS-sensing proteins in yeast [16]. This AP1-like (activator protein) transcription factor
regulates H2O2 homeostasis in Saccharomyces
cerevisiae (S. cerevisiae), which in turn governs the
synthesis of glutathione [17]. Hyr1p plays a key role during the oxidative response in
S. cerevisiae
[18]; after being
directly oxidized by H2O2, it forms an intermolecular
disulfide bond with yAP1 [19]. A conserved cysteine residue at position 598 in Yap1p
becomes active by forming an inter-molecular disulfide bond with the Cys36 of Hyr1p.
This transient inter-molecular linkage is then resolved to a Yap1p intra-molecular
disulfide bond between the cysteines at positions C303-S-S-C598. During this
process, the Yap1 protein is released by Hyr1p in its active form, which is then
transported to the nucleus [20]. This conformational change shields its nuclear export
signal from the interacting protein Crm1p, allowing it to remain in the nucleus and
control a suite of antioxidant genes [21], [22]. Although YAP1
gene homologs have been analyzed in several plant pathogenic fungi such as
Aspergillus fumigatus, Alternaria alternata, Cocholiobolus
heterostrophus, Botrytis cinerea and Ustilago maydis
[16], [20], [23], [24], [25], [26],
HYR1 has yet to be studied in filamentous fungi.
In this study, we closely examined the HYR1 homolog in M.
oryzae as a candidate mechanism for coping with a ROS-intensive host
environment. We demonstrated that HYR1 was indeed involved in
detoxifying or preventing plant basal immune responses including plant-generated ROS
and callose deposits during initial stages of infection, which was correlated with
its role as a virulence factor.
As one of the key members during the oxidative stress response, the yeast
Saccharomyces cerevisiae Hyr1/YIR037W (formerly termed
Gpx3) was reported to be a glutathione-dependent phospholipid peroxidase (PhGpx)
that specifically detoxifies phospholipid peroxides [19]. In order to identify the
corresponding gene in M. oryzae, we performed a BlastP analysis
against the fully sequenced genomic database of M. oryzae
housed at the Broad Institute. Using an E-value of 1e-3 returned a single hit
located on Supercontig 20, with an accession number of MGG_07460.6. It is 1315
bp long including two introns, with an open reading frame of 783 bp, which
encodes a 172-amino acid protein. A sequence analysis was performed using
Prosite on the ExPASy Proteomics Server (http://ca.expasy.org/prosite/). Hits revealed a glutathione
peroxidase active site at amino acid positions 27–42, and a glutathione
peroxidase signature at amino acid positions 66–73 (Figure 1A). When a BlastP search was
performed against GenBank at NCBI, numerous hits were returned with high
similarity scores, from many organisms including fungi and bacteria. An
alignment indicates that the putative GSHPx domains of Hyr1 are highly conserved
across different organisms (Figure
1B). The MoHyr1 protein shares the highest amino acid conservation
with the model, non-pathogenic fungus, Neurospora crassa
(93% similarity and 73% identity), but shares between 81 and
90% similarity with eight other plant pathogenic filamentous fungi
examined (Table
S1 and Figure
1C). Secondary structure of the HYR1 protein was determined by PSIPRED
[27],
and consists of eight β-sheets (or strands) and four α-helices (Figure 2). As described in
Zhang et al [18], the ScHyr1p showed a typical ‘thioredoxin
fold’, also consisting of four β-sheets surrounded by three
α-helices [28]. We compared the crystal structure of ScHyr1p with
the predicted tertiary structure of MoHyr1 protein, generated with PyMOL
(http://www.pymol.org/). The MoHyr1 predicted structure appears
similar to a canonical thioredoxin fold, showing four β-sheets, with β1
and β2 running parallel and β3 and β4 running anti-parallel,
surrounded by three α-helices (Figure 2). We located three positionally conserved cysteines in our
HYR1 protein model compared to yeast, and these are marked in Figures 1B and 2. Two important cysteines,
Cys39 and Cys88, likely correspond with two active sites found in the yeast
Hyr1p, Cys36 and Cys82. Together, our in silico data suggest
that we have identified the structural homolog of the ScHyr1 from yeast, and
that this gene is highly conserved across filamentous fungi.
In order to functionally characterize the MoHYR1 gene, we
obtained the ATCC S. cerevisiae Δhyr1 mutant and its wild
type parent for complementation tests. Our hypothesis was that based on its
sequence and predicted tertiary structure, the MoHYR1 gene
would rescue the yeast mutant when grown on non-permissive concentrations of
hydrogen peroxide. As shown in Figure 3, the yeast mutant and the wild type strain both grow well
on 0 and 2 mM H2O2. However, growth of yeast
Δhyr1 was significantly hindered in 4 mM
H2O2. The wild type MoHYR1 gene was
transformed into the yeast mutant, which restored partial growth on this higher
concentration. To further support our hypothesis, we constructed mutations in
the two conserved cysteine residues at positions 39 and 88. Neither of the
mutations rescued the yeast phenotype on hydrogen peroxide (Figure 3).
To explore the biological role of the MoHyr1 protein in the development and
pathogenicity processes of M. oryzae, the deletion mutant
Δhyr1 was generated through homologous recombination of
the MoHYR1 open reading frame with a gene conferring hygromycin
resistance (hygromycin phosphotransferase; HPH) (Figure
S1A). A gene deletion fragment was generated by nested PCR amplification
of the 5′ flanking region of MoHYR1, the
HPH gene, and 3′ flanking region of
MoHYR1, using adapters to link the three pieces together.
This gene deletion fragment, which contained flanking regions homologous to the
MoHYR1 gene, was introduced into protoplasts of M.
oryzae via PEG-mediated fungal transformation. After PCR screening
of successful knockouts and ectopics using primer pairs outside the flanking
regions and inside the HPH gene, two Δhyr1 knockout mutants
(B25, B33) and two ectopic mutants (B40, B60) were identified (Figure S1B)
and confirmed with Southerns (Figure S1C). Real-time qRT-PCR was also
employed to confirm full deletion of the MoHYR1 gene and no
transcripts were detected. Deletion mutant Δhyr1 (B33) was
complemented with a full-length copy of the MoHYR1 gene linked
to the cerulean fluorescent protein (Figure S1D, see Materials and Methods).
HYR1p in yeast was reported to not only be a sensor of ROS, but to have
scavenging properties as well [19]. To investigate the role of MoHYR1 in scavenging
H2O2 during vegetative hyphal growth, we inoculated
the same amount of initial mycelia into complete media (CM) containing 0, 5 and
10 mM H2O2. No significant differences were detected among
wild type, the Δhyr1 knockout mutants and the ectopics when
growing in 0 mM H2O2. However, the mycelial growth of the
Δhyr1 knockout mutants was severely and significantly
affected at 10 mM H2O2 (Figure S2A and
B). By contrast, the wild type and ectopics did not display much
difference in mycelial growth at any concentration. The complemented mutant line
grew slightly better than wild type in all concentrations of
H2O2, and upon Southern analysis, we found that four
copies had inserted into the genome (Figure S1E). Together, these data indicated
that MoHYR1 was responsible for the H2O2
growth tolerance phenotype.
To determine the role of MoHYR1 in virulence, we drop-inoculated
detached leaves of three week-old blast-susceptible barley cultivars with
conidia from two independently generated Δhyr1 mutants, B25
and B33 (Figure 4A). The
mutants were still able to cause disease lesions, but there was a measurable and
significant reduction in lesion size compared to those produced by wild type,
ectopics, and the complemented line (Figure 4B). The complemented line,
hyr1- C, restored full virulence to the
Δhyr1 mutant, B33. All pathogenicity assays were
repeated on the susceptible rice cultivar Maratelli, with similar results (Figure 4C) using the
spray-inoculation technique. Disease was also quantified on rice using a
“lesion type” scoring assay [29] and error bars show that
while lesion types 1–3 do not differ between the mutants, ectopics and
wild type, lesion types 4 and 5 (severe, coalescing) did not form on
mutant-inoculated plants (Figure
4D) Interestingly, no other developmental phenotype examined was
compromised in the Δhyr1 mutant, including growth rate,
conidia production and shape, germ tube and appressorial formation (Table 1).
A fundamental question we wanted to assess was whether MoHYR1 was required for
infection-related activities in planta. The M.
oryzae's disease cycle is initiated when the conidium contacts
a hydrophobic surface, inducing it to germinate. The germinated conidium forms a
germ tube and appressorium that penetrates the plant surface via turgor pressure
and forms a thin penetration peg into the first plant cell [30]. Thus, we first examined
whether ROS was present during any of these processes, and if so whether MoHYR1
was involved in coping with it. We inoculated susceptible rice and barley
cultivars with the Δhyr1 mutants, ectopics and wild type.
ROS was detected using the indicator
2′,7′-dichlorodihydrofluorescein diacetate (H2DCFDA)
[31].
Conidia of wild type, ectopics and the Δhyr1 mutant all
elicited some degree of ROS when inoculated onto barley leaves (Figure 5A–C), whereas
ROS was undetectable under the same imaging conditions when non-inoculated
leaves were stained (data not shown). The Δhyr1 mutants
showed the strongest ROS signal 24 hours post inoculation (hpi) compared to the
others. The signal continued in this manner through 48 hours (data not shown).
These experiments were repeated six times and the results were consistent across
the two independent Δhyr1 mutant lines. ROS signals were
quantified via counting the number of ‘ROS haloes’ found around
appressoria and expressing this as a percentage of appressoria counted per
sample; a significant difference in signals was observed between the mutants,
wild type, and ectopics (Figure
5D). These results indicate that in the absence of the
MoHYR1 gene, the fungus can no longer manage the ROS that
is generated during initial infection events, or loses the ability to
effectively cope with it.
To better understand the reason for reduced virulence in the
Δhyr1 mutant, we wished to determine whether internal
fungal levels of ROS were altered in the absence of the gene. The deletion
mutant and wild type were grown on complete media and stained with nitroblue
tetrazolium (NBT) for production of superoxide anions (Figure S3).
Results showed little differences between mutant and wild type when examining
the entire colony (Figure S3E and F) or aerial hyphae (Figure
S3A–D).
Figure 5C suggested that
reactive oxygen species localized mainly around the appressoria. Upon closer
inspection, we observed that the ROS “haloes” around the appressoria
usually localized directly underneath the appressoria (Figure 6). Previous studies had demonstrated
that the rice blast fungus also generates internal ROS during infection-related
development, particularly during appressorial maturation and furthermore, that
ROS can be secreted from the fungus itself [5]. In order to identify the
source of the reactive oxygen species detected in our experiment, we inoculated
M. oryzae onto the hydrophobic side of gel-bond, which can
mimic the plant surface and induce ROS production in vitro
[32]. The result
shown in Figure 7 indicated
that first, M. oryzae does generate ROS during germ tube and
appressorial formation; second, the reactive oxygen species generated by
M. oryzae were mostly intracellular and did not appear to
be secreted or defused; and finally, that ROS were relatively weak in the fungal
structures by 24 hpi. These observations occurred in the wild type, ectopic and
mutant lines, indicating little difference in internal ROS levels regardless of
the presence of HYR1. Altogether, these results were different
from what we observed in planta, which was a strong ROS signal
from 24–48 hpi.
In order to identify the source of the ROS detected during susceptible
interactions, we used diamino-benzidine (DAB) to study the ROS distribution
pattern. Barley leaves were inoculated with Δhyr1 mutant
then stained with DAB and imaged using confocal reflected light signal to
visualize the DAB deposits from a top view of an interaction site (Figure 8A). The leaf samples
from this same interaction site was processed further and embedded in epoxy
resin to obtain a cross-section using a correlative microcopy approach. The
confocal images suggested that the dark region (DAB) was localized immediately
adjacent and inside the plant cell wall (Figure 8B) centered around the penetration
peg (arrowhead - Figure
8B).
The second piece of evidence resulted from scavenging for ROS with ascorbic acid,
an antioxidant that detoxifies hydrogen peroxide [33]. When 0.5 mM ascorbic
acid was mixed with Δhyr1 mutant conidia, inoculated onto
plants and stained with H2DCFDA, ROS haloes were clearly observed
(Figure 9A). However,
when barley leaves were pre-treated with ascorbic acid, then inoculated and
stained with H2DCFDA, almost no ROS haloes were detected (Figure 9B). This experiment
was repeated with another ROS-inhibitor called DPI (diphehyleneiodonium
chloride), with similar results (data not shown). Ascorbic acid-treated leaves
were also inoculated with mutant conidia and allowed to incubate in the growth
chamber for six days, after which time we observed wild type lesions (Figure 9C). This suggested
that the ROS haloes observed in this experiment are likely originated from the
plant.
Futhermore, we analyzed previously characterized nox1 and
nox2 mutants for ROS haloes; in M. oryzae,
NOX1 and NOX2 code for NAPDH oxidases, and
are largely responsible for producing internal ROS [5]. We hypothesized that if ROS
was emanating from the plant, than the loss of the NOX genes
should have no effect on haloes. Overall, haloes can still be produced when
either of the nox mutants, or its parental strain, Guy11 was
inoculated onto barley leaves (Figure S4A–F). While there was a slight
significant difference among the number of haloes observed when looking at the
individual mutants (nox1 made slightly more than
nox2), there was no significant difference between mutants
and wild type (20–30 appressoria were counted per strain, and the
percentage of those with haloes, reported; Figure
S4G).
Since our data strongly suggested that Δhyr1 mutants had a
lower capacity to eradicate plant-generated ROS during early stages of
infection. Our next goal was to determine whether this gene played a role in
fungal tolerance to ROS generated immediately following inoculation. In order to
carry out this experiment, we inoculated susceptible barley leaves with either
the Δhyr1 mutants or the wild type conidia, and imaged them
1 hpi. The ROS dye H2DCFDA was injected directly into the leaves, so
the result only showed the redox status inside the leaves, and not inside the
fungus, which might have skewed the results. Our data revealed that ROS was
detected 1 hpi, which indicated that the plant detected and responded to the
pathogen at an early time point (indicated by ROS fluorescence in the mesophyll
cells; Figure
S5A). A quantitative analysis of the signal intensities by ImageJ
(available at http://rsb.info.nih.gov/ij; developed by Wayne Rasband, National
Institutes of Health, Bethesda, MD) revealed no significant differences when
inoculated with the Δhyr1 mutants or with the wild type
conidia (Figure
S5B). We thus concluded that the MoHYR1 gene does
not play a role in ameliorating an early, or immediate, plant defense
response.
To test whether MoHYR1 had any impact on plant-produced ROS that
may occur later during infection, we inoculated Δhyr1
mutant conidia or wild type conidia onto barley leaves and stained with DAB at
24 hpi (Figure 10). Results indicated that the
Δhyr1 mutant was unable to block ROS produced at 24
hpi, where the ROS was both detected in an entire plant epidermal cell, as well
as in plant cells that were not in direct contact with the pathogen (Figure 10).
It has been documented that the presence of reactive oxygen species around CWAs
is a biochemical marker for non-penetrated cells during the interaction between
barley and barley powdery mildew, Blumeria graminis
[34]. To
determine whether the ROS observed during a susceptible barley-M.
oryzae was related to CWAs, we performed aniline blue staining on
inoculated leaves. At 24 hpi, we found callose deposits specifically localized
around the appressoria and penetration sites (Figure 11). Sequential correlative staining
with H2DCFDA for ROS followed by analine blue for callose, showed a
strong positional correlation between the two host responses when overlaid
(Figure 11C).
CWAs are believed to physically block pathogen penetration [34]. To further
characterize the CWAs formed during the barley- M. oryzae
interaction, we examined leaves that had been inoculated with M.
oryzae 24 and 40 hpi with either mutant or wild type conidia. The
result showed that classical CWAs were formed within 24 hpi in both strains and
no other differences in CWA morphology could be detected (Figure 12).
Given the fact that increased ROS accumulation occurs in the absence of MoHYR1,
we next tried to determine whether the ROS scavenging system was impaired in the
Δhyr1 mutants. We used real-time quantitative real time
reverse transcription PCR (real-time qRT-PCR) to compare the expression of
general antioxidant and redox control gene orthologs in both M.
oryzae wild type and Δhyr1 strains (Figure 13). Primer pairs for
the following genes were employed to examine gene expression:
YAP1 (MGG_12814.6), GSH1
(γ-glutamylcysteine synthetase; MGG_07317.6), GSH2
(glutathione synthetase; MGG_06454.6), GLR1 (glutathione
reductase; MGG_12749.6), GTT1 (glutathione transferase 1;
MGG_05677.6), SOD1 (Cu/Zn superoxide dismutase; MGG_03350.6),
CAT1 (catalase 1; MGG_10061.6), GTO1
(omega class glutathione transferase 1; MGG_05367.6), and cyt c
per (cytochrome c peroxidase; MGG_10368.6). The housekeeping
gene encoding Ubc (ubiquitin conjugating enzyme; MGG_04081.6)
was used as an internal control. We also included the gene MoHYR1 (MGG_07460.6)
in this experiment to confirm its deletion in the mutant lines. The expression
patterns of these ten genes were placed into two categories. The first category
(Figure 13A) is
comprised of four genes that show increased expression in the wild type strain
after induction with hydrogen peroxide, while expression in the mutant line is
low and unchanging. GTT1, GR and
GSH1 belong to this category, along with the
HYR1 partner protein YAP1;
YAP1 also shows slight but significant differences in
expression in the Δhyr1 mutant line with and without
H2O2, and has a higher expression level compared to
the wild type strain without ROS. The second category contains genes whose
expression does not significantly change, both in response to
H2O2, as well as in the presence of the
MoHYR1 gene. This category includes six genes: cyt
c per, CAT I, Cu/Zn SOD,
GTT I, GSHII and MoHYR1
(Figure 13B).
HYR1 shows no expression at all in the mutant line, which
was to be expected.
We evaluated the sub-cellular localization pattern of the MoHYR1 protein during
infection, conidia of a M. oryzae deletion line
(Δhyr1 B33) transformed with cerulean-MoHYR1 N-terminal
fusion (the same construct that was used for complementation), was inoculated
onto barley leaves and observed during the following time points: 1 hpi, 6 hpi,
12 hpi, 24 hpi and 72hpi. At 1 hpi, MoHYR1 was mainly localized in the conidial
vacuoles and with low levels in the cytoplasm. When the germ tube formed, the
protein was present throughout the germ tube (Figure 14A). At 6 hpi, the MoHYR1 protein
showed increased cytoplasmic localization in the appressorium and conidium and
at 12 hours, a concentration of HYR1 in the appressorial cytoplasm (Figures 14B and C). At the
later time point, 24 hpi, the protein appeared to be localized in the vacuoles
with reduced levels in the cytoplasm (Figure 14D), and a later, invasive stage time
point suggests the protein was again cytoplasmically localized (Figure 14E).
During the interaction between the pathogens and plants, plants mount defense
mechanisms to protect themselves from pathogens. The cellular environment within the
host can represent a major source of stress towards the invaders [16]. Pathogens, on the
other hand, must possess adaptive mechanisms in order to survive. In this study, we
hypothesized that the M. oryzae HYR1 protein defines one such
mechanism, the glutathione synthesis pathway, involved in coping with the oxidative
environment generated by plant defenses.
In M. oryzae, MoHYR1 is the only sequence homolog of the yeast
glutathione-dependent peroxidase, HYR1p, formerly termed
Gpx3
[35]. In yeast,
HYR1p senses H2O2 through two highly
conserved cysteines that are redox sensitive. Mutations in either of these two
cysteines leads to a non-functional HYR1 [18]. Indeed, we found that the
wild type MoHYR1, but not the MoHYR1 cysteine
mutants, was able to partially rescue the yeast HYR1p mutant on
non-permissive levels of H2O2. This result is similar to
Δyap1 yeast mutants complemented with homologs from two
pathogenic filamentous fungi, Cochliobolus heterostrophus and
Ustilago maydis, as both homologs partially complemented
the yeast mutation [20], [23]. These data suggested that MoHYR1
may function similarly during redox sensing and the subsequent signaling that
leads to ROS detoxification. This model was further supported by the presence of
ROS haloes located underneath appressoria during infection with a much greater
frequency in the Δhyr1 mutant compared to the wild type
strain.
The increase in ROS haloes in Δhyr1 mutants correlated with
significantly smaller lesions sizes when inoculated on susceptible rice and
barley plants, suggesting that ROS scavenging regulated by MoHYR1 was required
for full virulence. This was supported by a rescuing of the
Δhyr1 mutant phenotype to wild type lesions by
scavenging plant-derived ROS with ascorbic acid or disrupting plant-derived ROS
generation with DPI. These results were similar to a gene recently reported on
in the rice blast fungus called DES1 for
Defense Suppressor 1 [14].
DES1 was also involved in virulence and triggers a stronger
plant response upon infection, manifested by both an increase of the oxidative
burst, as well as expression of two plant defense genes. Intriguingly,
DES1 has no known functional domains and from sequence
analysis, its function cannot be predicted, although it is well-conserved
throughout fungi. It is also noteworthy that expression of
MoHYR1 was tested in the Δdes1 mutant,
and found to be slightly down-regulated. This could suggest that
HYR1 and DES1 represent two
semi-redundant, semi-dependent mechanisms evolved to cope with the plant defense
response. Equally interesting is a gene recently identified in the plant and
human fungal pathogens, Alternaria brassicicola and
Aspergillus fumigatus, respectively, called
tmpL
[16]. This
membrane-localized gene contains a FAD/NADP-binding domain and had not yet been
studied in fungi. A deletion of tmpL resulted in a severely
reduced virulence defect and hypersensitivity of exogenous oxidative stresses,
however when the YAP1 gene was over-expressed in the deletion
line, it rescued these and other mutant phenotypes, suggesting
tmpL, YAP1 and presumably
HYR1 may act in a concerted pathway to sense and trigger
ROS scavenging pathways.
A successful pathogen, which has the ability to detoxify ROS, will subsequently
have fewer barriers to overcome before reaching its ultimate goal, which are the
cell contents. Our results with the MoHYR1 gene suggest that
while there might be no effect of MoHYR1 on ROS that's
produced immediately by the plant (Figure S3), there is subsequent ROS
production which MoHYR1 clearly helps the fungus overcome
(Figure 10). Metabolic
profiling performed by Talbot and colleagues (2008) provides support for this
concept, revealing a M. oryzae-induced host metabolism
re-programming that suppressed or delayed plant-produced ROS during susceptible
interactions.
Although supporting evidence has shown that M. oryzae can
produce ROS during infection related development [5], through scavenging
experiments, the ROS observed in our studies appear to be largely
plant-generated. Internal fungal ROS was unaffected by the absence of the
MoHYR1 gene in vitro. Furthermore, ROS
haloes were not disrupted by the ROS scavenger, ascorbic acid, when applied only
to conidia, but were disrupted when ascorbic acid was specifically applied to
leaves. Several pathways for plant-generated ROS include cell wall-bound
peroxidases [1]. Plants defend themselves against pathogens by a
battery of cell wall-associated defense reactions, including generation of ROS
and cross-linking of lignin compounds [34]. During the
interaction between a French bean (Phaseolus vulgaris) and a
cell wall elicitor from Colletotrichum lindemuthianum, ROS
appears to originate from cell wall peroxidases [36]. Apoplastic alkalization
has been shown to be important in this process [34]. ROS generated from
cell wall peroxidases also serve as key molecules required for lignification and
cross-linking of cell walls [34]. In a study carried
out between barley and the powdery mildew fungus, barley cell wall localized
peroxidase HvRBOHA is responsible for generating
H2O2, which was only present in non-penetrated cells
[37].
Our results, particularly in Figure
8B, suggest ROS localized up against the plant cell wall. Further
investigations into M. oryzae-host interactions will include
analyzing plant defense-related genes, including the barley cell wall
peroxidase.
Callose and ROS are two plant defensive compounds known to be involved in cell
wall appositions, which are deposited during both compatible and incompatible
interactions [34]. H2O2 played an important role
in this process and enzymatic removal of H2O2 by catalase
significantly reduces the frequency of phenolic deposition [34]. Several components
were reported to be essential for this oxidative burst: peroxidases, a calcium
influx and K+ Cl− efflux, extracellular
alkalization, and post-Golgi vesicles [38]. ROS around the CWA areas
might function as signal compounds to gather the vesicles and components needed
for mature CWAs. We observed that ROS and callose deposits were positionally
related during attempted penetration by both wild type and
Δhyr1 mutants, immediately below the appressorium. From
this result, we hypothesize that ROS generated by plant defenses activates CWA
formation in a susceptible host and experiments to determine the timing of
deposition of ROS versus callose are currently underway.
A hypothesis that follows from these data is that when the
MoHYR1 gene is deleted, the plant responds as though
it's being challenged with an avirulent pathogen. As early as 12 hours post
inoculation, we observed that barley leaves inoculated with
Δhyr1 mutants showed higher ROS signals compared with
leaves inoculated with wild type. These data were consistent using two staining
methods, H2DCFDA and DAB. In leaves inoculated with wild type, ROS
was detected around appressoria but was mostly observed inside fungal
structures. However, ROS was seen both around appressoria and adjacent cells
when inoculated with the Δhyr1 mutants. Whole cells filled
with ROS were also observed when inoculated with Δhyr1
mutants, which was related with HR-type cell death. All these data indicated
that HYR1 might function to suppress later plant-generated ROS,
either by detoxifying it directly, or manipulating plant ROS secretion-related
gene expression.
While our data showed that HYR1 likely played an important role in
ROS-detoxification processes, our experiments did not preclude other ROS
tolerance mechanisms in the fungus, particularly since mutants were reduced in
virulence, but not completely non-pathogenic. Such mechanisms might involve the
aforementioned DES1 and tmpL genes. Currently,
we are characterizing the MoYAP1 homolog in M.
oryzae; our initial Δyap1 mutant data
suggested this gene was dispensable for pathogenicity, much like what has been
found in Botrytis cinerea, Aspergillus
fumigatus and Cochliobolus heterostrophus
[23], [25], [26].
Intriguingly, YAP1 did appear to be essential for virulence in
Ustilago maydis and Alternaria alternata
[20], [25], suggesting
that fungal lifestyle (i.e. necrotrophic vs. biotroph) had little to do with
this particular oxidative stress pathway, and further supporting redundant
pathways. Our real-time qRT-PCR data showed that YAP1 increases
in expression when wild type was challenged with H2O2 and
we also noted a decrease in YAP1 gene expression in the
Δhyr1 mutant background. One interpretation of this
result was that the fungal cell might be compensating for the absence of
HYR1, by boosting expression of its partner gene.
The glutathione pathway-related genes GLR1,
GTO1 and GSH1, all increased during
H2O2 challenge in the wild type however had extremely
decreased expression in the mutant line, regardless of ROS. This suggested that
these genes were reliant upon HYR1, which was not unexpected,
since the glutathione pathway was shown to be regulated YAP1,
which occurs after interacting with HYR1 [17]. Our results were also in
keeping with the C. heterostropus Yap1 homolog mutant
Δchap1, which showed extremely low levels of both
GLR1 and GSH1
[23].
Interestingly, we did not observe any of the other genes increasing in
expression in the mutant background; this suggested that at least for the genes
that we chose such as CAT1 and SOD1, they did
not provide compensatory mechanisms for a loss of HYR1. While
this is one hypothesis, it is also possible that these genes are regulated at
the protein level, as was found in the A. fumigatus mutant,
ΔAfyap1; both CAT1 and
SOD1 were among the proteins down-regulated in the mutant
[39], and
this could also hold true for the Δhyr1 mutant. Likewise,
catalase, SOD and peroxidase activities were measured in the A.
alternata mutant ΔAaAp1
[25]. A
transcriptomic study on the Δhyr1 deletion mutant would
answer many of these questions; further, such a study would uncover redundant
pathways of ROS detoxification masked by the presence of MoHYR1.
While numerous studies have examined localization of the Yap1p, we were unable to
find any studies on the localization of HYR1 either in yeast or filamentous
fungi. Our data revealed that the HYR1 protein mostly localized either to the
cytosol or to vacuoles, during early stage infection events on barley (germ
tube, early appressorial formation, appressorial maturation and penetration). At
one hpi, MoHYR1 was mainly moving through the germ tube, although it was
difficult to definitively ascertain which organelle it might be associated with.
At twelve hpi, the MoHYR1 protein shows cytoplasmic localization, mainly
expressed in the cytosol of the appressorium. We suspect that by twenty-four
hours, the fungus had penetrated and gained ingress to the first epidermal cell;
indeed cell biology studies on events following initial penetration suggested
that M. oryzae bulbous hyphae fill an entire rice leaf sheath
cell and were in the process of moving onto the next one by twenty-seven hours
post-inoculation [40]. Its vacuolar localization at this time-point could
reflect that fact that it was no longer needed by the fungus, which had
circumvented the plant's oxidative burst and at that point growing in the
first epidermal cell. We examined a later time-point at 72 hpi and found the
HYR1 gene to be once again cytoplasmically localized, perhaps indicating a
requirement for this pathway at the invasive growth stage.
In conclusion, we identified and characterized the MoHYR1 gene,
a functional homolog of the yeast Hyr1 (or
Gpx3) gene. Although MoHYR1 does not cause dramatic effects
in the disease phenotype, it nevertheless played an important role in virulence.
This effect appeared to be related to the deletion mutant's inability to
tolerate plant-generated ROS, or at least to do so in a timely and effective
manner to cause wild type levels of disease. Together, our results help to
define a mechanism that, while well-studied in yeast, has not yet been examined
in filamentous fungi; furthermore, our studies pose additional questions to be
answered regarding the role of the glutathione pathway in scavenging ROS in
filamentous fungi, how this aids in pathogenicity and what other underlying
redundant scavenging pathways exist.
Rice-infecting M. oryzae, strain 70–15 (Fungal Genetics
Stock Center 8958) was used as the wild type strain throughout this project, and
the strain from which mutants and transgenics were derived. All strains were
maintained at 25°C under constant fluorescent light on complete medium (CM 1
liter: 10 g sucrose, 6 g yeast extract, 6 g casamino acid, 1 ml trace element).
Oatmeal agar medium (OAM 1 liter: 50 g oatmeal and 15 g agar) was used for
sporulation. Conidia were harvested 10–12 days after plating.
Yeast strains BY4741 (wild type) and BY4741 YIR037W (Δhyr1 mutant) were
ordered from the American Type Culture Collection, grown out and maintained on
YPD medium. Constructs for transformation were built using standard PCR reaction
conditions and programs; briefly, pJS371 used overlapping primers to make an
intron-free version of the MoHYR1 gene in pJS318. Using the intron-free plasmid,
overlapping primers were used to make Cys39Ala and Cys88Ala mutant versions of
the coding sequence. These were cloned into pCRScript (pJS372 & pJS373,
respectively). The yeast HYR1 gene (ScHYR1) was then amplified from Sc46 and
cloned into pRS423, the His3 episomal plasmid, pJS374. These plasmids then form
the basis of the genes to be tested: MoHYR1 wild type, the 2 cysteine mutants of
MoHYR1 and the ScHYR1 gene. These four genes are under the same promoter and
terminator. Therefore ScHYR1 was engineered to have an NcoI site at the ATG and
a BamHI site at the beginning of the terminator (pJS375). Since the Magnaporthe
gene has a natural NcoI site at the ATG, the 3 genes of the MoHYR1 are cloned
into pJS379 as NcoI/BamHI fragments (pJS381, pJS382, pJS383).
For the complementation assays, five-microliter drops from serial dilutions from
cultures with anOD600 of 0.5 were spotted on plates with and without 0, 2 and 4
mM H2O2 and grown for 2 days at 30°C. This experiment was repeated 10 times.
In total, the following plasmids were used in this part of the study:
pSM387 ( = pRS423) HIS3 yeast episomal plasmid; pJS374
pSM387 + ScHYR1;
pJS381 ScHYR1-Pro::MoHYR1::ScHYR1Term;
pJS382 ScHYR1-Pro::MoHYR1_Cys36Ala::ScHYR1Term;
pJS383 ScHYR1-Pro::MoHYR1_Cys82Ala::ScHYR1Term.
Rice cultivar Maratelli (a gift from the Dean Lab; Raleigh, NC) and barley
cultivar Lacey (Johnny's Selected Seeds; Winslow, ME) were used throughout
this study, as both are susceptible to M. oryzae strain
70–15. Rice was grown in a growth chamber at 80% humidity, and 12
h:12 h day:night cycles, at 28°C. Barley was grown in a growth chamber at
60% humidity, and 12 h:12 h day:night cycles, at 24°C (day) and
22°C (night).
The targeted gene deletion was accomplished using the homologous recombination
method. We amplified 5′ and 3′ flanking regions of
Hyr1 using primer pairs #1 and 2 (Table S2).
Flanking regions were then linked via adaptor-mediated PCR to a 1.3 kb
HPH coding sequence, providing resistance to the antibiotic
hygromycin (Alexis Biochemicals, San Diego, CA). The entire length of the
deletion fragment was 3.7 kb. Fungal protoplasts of the wild type 70-15 were
directly transformed with the nested PCR product (primers used were forward
primer of primer pair #1 and reverse primer of primer pair #2). Protoplast
generation and subsequent transformation were conducted by following established
procedures [41]. To confirm the knockout mutant, the genomic DNA of
candidate strains was extracted and amplified with primer pairs #3, 4 and 5
(Table
S2).
Equal-sized pieces of mycelia were cut with #3 cork-borer tool (0.7 cm in
diameter), and immersed in 10 ml of liquid CM at 25°C in darkness. Colonies
were grown in CM containing H2O2 at concentrations of 0
mM, 5 mM and 10 mM. Colonies were removed from each well, vacuum filtered to
dryness, and measured on a scale one week post-immersion.
For point or drop inoculations, conidia were harvested from 12-day-old cultures
grown on OMA in 20 µl of a 0.2% gelatin (Acros organics, New
Jersey) suspension, for a final concentration of 1–5×105
conidia/ml. Point two percent gelatin was used as a non-inoculated control for
pathogenicity assays. For drop inoculations, three week old leaves of Maratelli
or Lacey were detached and laid flat in a humid chamber (90 mm Petri dish with
moist filter paper). Twenty microliters of conidial suspensions, or gelatin
alone, were dropped onto each leaf and kept in darkness overnight at
∼25°C. The next day, remaining water drops were wicked off and moved to
a growth chamber under constant fluorescent light. For spray inoculations,
conidial suspensions (10 ml; concentration as above) in 0.2% gelatin were
sprayed onto three week old Maratelli or Lacey seedlings. Inoculated plants were
placed in a dew chamber at 25°C for 24 hours in the dark, and then
transferred into the growth chamber with a photoperiod of 16 h:8 h light:dark
cycles. Disease severity was assessed seven days after inoculation.
Quantitative real time reverse transcription PCR (real-time qRT-PCR) was carried
out using primer pairs for the following genes: YAP1
(MGG_12814.6), GSH1 (MGG_07317.6), GSH2
(MGG_06454.6), GLR1 (MGG_12749.6), GTT
(MGG_06747.6), GTO1 (MGG_05677.6), GTT1
(MGG_09138.6), SOD1 (MGG_03350.6), CAT1
(MGG_10061.6) and cytochrome c peroxidase (MGG_10368.6). The
housekeeping gene encoding ubiquitin conjugating enzyme (MGG_00604.6) was used
as an internal control. We also included the gene MoHYR1
(MGG_07460.6) to confirm its deletion in the mutant lines. Primer pairs are
listed in Table
S3. Seventy-five nanograms of cDNA generated from mycelium grown as
per the H2O2 experiments described above (generated from
the 0 mM and 5 mM H2O2 samples), was used as templates for
each reaction. The mycelia were fragmented in a blender as per the protocol by
Mosquera et al [42], before being inoculated into liquid complete medium.
After 2–3 days, the mycelia were blended again to ensure the largest
amount of actively growing fungal tips. The H2O2
experiment was performed 24 hours after the 2nd blending, and RNA was
extracted. PCR reaction conditions were as follows for a 25 µl reaction:
13 µl H2O, 10 µl 5 Prime SYBR Green Master Mix (Fisher
Scientific, Waltham, MA), 0.5 µl Forward Primer (for a final concentration
of 2 µM; Integrated DNA Technologies, Coralville, IA), 0.5 µl
Reverse Primer (for a final concentration of 2 µM) and 1 µl template
DNA. Conditions for real-time quantitative RT-PCR conditions were as follows:
95°C for 2 min; 95°C for 15 sec, 58°C for 15 sec, 68°C for 20
sec (cycle 40 times); 95°C for 15 sec; 60°C for 15 sec (melting curve);
60°C –95°C for 20 min; 95°C for 15 sec; lid temperature
constant at 105°C. The 2−ΔΔCt method was used for
generating the data. ΔΔCt is defined as ΔCt treatment - ΔCt
calibrator. cDNA from the strain 70-15 in 0 mM H2O2 was
used as the calibrator for comparison of gene expression in 5 mM
H2O2 in both the Δhyr1 deletion
lines as well as the wild type For both the ΔCt treatment and ΔCt
calibrator, ΔCt is defined as Ct gene - Ct housekeeping-gene. For the
calibrator, which is 0 µM H2O2, this value would be
2−0 or 1. These experiments were repeated twice with
similar results.
A HYR1 N-terminal cerulean fusion construct was generated by fusion PCR. Briefly,
using M. oryzae genomic DNA as a template, a 1 kb promoter
region of HYR1 was amplified with primers 6 and 7 (Table S2).
Another set of primers, 8 and 9, were used to amplify the 2.4 kb
HYR1 open reading frame. Three resulting fragments, the 1
kb promoter fragment, the 1328 bp ORF (including 709 bp of terminator sequence)
and 740 bp cerulean fluorescent protein coding sequence [43], were mixed and
subjected to a second fusion PCR with primers 7 and 8. The resulting 3.1 kb PCR
product was generated with BamHI and NotI
restriction enzymes (New England Biolabs, Beverly, MA) and cloned into
pBlueScript II SK+. The construct was fully sequenced and found to be
correct, hence was co-transformed into the M. oryzae Δhyr1
knockout mutant protoplasts to make Cerulean-HYR1 fusion transformants.
Transformants with expected genetic integration events were identified by PCR
using primers pairs 6 and 10 (Table S2). Properly transformed
Δhyr1 mutants were also used as the complemented lines,
in Figures 3 and 4, designated as
“hyr1-C”.
Ten-fourteen day old rice and eight day old barley plants were used and collected
24 hours after being inoculated with 10–12 day old conidia (methods as
described above). All staining procedures were performed with both rice and
barley, however barley was best-suited for microscopy, hence all micrographs
shown in this study are of barley. For experiments with 29,79-dichlorofluorescin
diacetate (H2DCFDA) (Invitrogen, Carlsbad, CA), inoculated tissue
were collected and incubated for 60 min at room temperature in 5–20 mM
H2DCFDA dissolved in DMSO (less than 0.005% final
concentration), then washed with 0.1 mM KCl, 0.1 mM CaCl2 (pH 6.0)
and left for 60 min at 22°C before experimentation. Dye excitation was at
488 nm; emitted light was detected with a 500–550 band pass emission
filter. DAB staining was carried out using the protocol developed by Thordal
Christensen et al [44]. Briefly, leaves were cut at the base with a razor
blade and placed in a 1 mg/mL solution of DAB for 8 h under darkness at room
temperature. Leaves were decolorized by immersion in ethanol (96%) for 4
h followed by 2 hours in PBS buffer before imaging. A third method of ROS
detection was employed for examining ROS internal to, or secreted from, the
fungus. Nitroblue tetrazolium (Sigma-Aldrich, St. Louis) was used at 4 mg/mL (in
deionized water) and the staining performed for 5 min∼30 min at room
temperature prior to observation.
In order to eliminate the ROS generated by fungus, conidia of Δhyr1
(B25) and wild type (70-15) were mixed with 0.5 mM ascorbic acid
(AsA) and inoculated onto the leaf surface. Leaves were stained for ROS at 24
hpi. In order to eliminate ROS generated from the plant, leaves were first
treated with 0.5 mM ascorbic acid for 1 hour. To remove excess AsA, leaves were
then washed with 0.1 mM KCl, 0.1 mM CaCl2 (pH 6.0) buffer three times
for 5 minutes each. Finally, leaves were inoculated with conidia 1hpi and
stained for ROS 24 hpi. Additionally, barley leaves were injected with 5
µM DPI (diphenyleneiodonium; Sigma, St Louis), then washed and inoculated,
as above.
Calcofluor White M2R (Fluorescent brightener 28, F-6258, Sigma, St Louis) was
used for detection of the fungal cell wall. We made 10,000-fold dilutions from a
saturated Calcofluor White stock solution. For experiments involving conidia on
gel-bond (VWR, Arlington Heights, IL), Calcofluor White was applied 1, 4, 8, 12,
and 24 hours post inoculation, incubated for 15 minutes, then gently rinsed off
with 1X PBS buffer. For experiments involving inoculated plants, inoculated or
non-inoculated (control) leaf tissue was collected and immersed in working
solution for 15 minutes, then gently rinsed with 0.1 mM KCl, 0.1 mM
CaCl2 (pH 6.0).
For CWAs staining, we cleared inoculated or non-inoculated (control) leaves in
ethanol:acetic acid (6:1 v/v) overnight and washed them with water.
Subsequently, cleared leaves were incubated in 0.05% aniline blue (w/v)
in 0.067 M K2HP04 buffer at pH 9.2 overnight and rinsed
gently in sterilized deionized water for microscopy.
Inoculated barley leaves were stained using DAB and rinsed several times in PBS.
Thereafter, samples were fixed in 2% paraformaldehyde (Electron
Microscopy Sciences, Hatfield, PA) and 2% glutaraldehyde (Electron
Microscopy Sciences, Hatfield, PA) in sodium cacodylate (Electron Microscopy
Sciences, Hatfield, PA) buffer for 1 hour overnight. Samples were then rinsed
three times, 15 min each, in sodium cacodylate and post-fixed with 2%
OsO4 in sodium cacodylate for 3–5 hours on a rotator.
Again, samples were then rinsed three times for 15 min each, with water on a
rotator. Samples then underwent an ethanol dehydration series (25%,
50%, 80% ETOH; 20 min each) on a rotator. Samples were primed with
1% gamma-glycidoxylpropyl trimethoxysilane in 80% ETOH overnight
at room temperature and then washed three times for 15 min each in 100%
ETOH on a rotator. Samples then underwent a series of infiltrations on a rotator
as follows: 100% ETOH/n-BGE (Electron Microscopy Sciences, Hatfield, PA)
(1:1) for 30 min, 100% n-BGE for 30 min, n-BGE/Quetol-651 (Electron
Microscopy Sciences, Hatfield, PA) (1:3) for 1 hour, n-BGE/Quetol-651 (1:1) for
1 hour, n-BGE/Quetol-651 (3:1) for 1 hour, 100% Quetol-651 for 1 hour,
100% Quetol-651 for 1 hour, 100% Quetol-651 overnight and
100% Quetol-651 for 1 hour. Finally, samples were embedded and
polymerized in an oven at 60°C for about 24 hours.
BlastP analysis was done against the fully sequenced genomic database of
M. oryzae housed at the Broad Institute, using an e-value
of 1e-3. ClustalW (X2) was used to perform the full alignment and generate the
phylogenetic tree. The final tree image was generated with Tree Viewer. The HYR1
protein secondary structure was predicted using the PSIPRED protein structure
prediction server. The structural image of the HYR1 protein was created using
the PyMOL molecular viewer. All student t-tests were performed using JMP8 (SAS
Institute Inc. 2007. <Title>. Cary, NC: SAS Institute
Inc.).
Confocal images were taken with Zeiss LSM510 or Zeiss LSM5 DUO using a
C-Apochromat 40X (NA = 1.2) water immersion objective lens.
H2DCFDA ester was excited at 488 nm and fluorescence was detected
using a 505–550 nm band pass filter. Calcofluor white was excited at 405
nm and detected using 420–470 nm band pass filter. Cerulean was excited at
458 nm and detected using a 475 long pass filter. We also used transmitted light
and reflected light for some confocal experiments.
|
10.1371/journal.pgen.1007486 | Rad51 recruitment and exclusion of non-homologous end joining during homologous recombination at a Tus/Ter mammalian replication fork barrier | Classical non-homologous end joining (C-NHEJ) and homologous recombination (HR) compete to repair mammalian chromosomal double strand breaks (DSBs). However, C-NHEJ has no impact on HR induced by DNA nicking enzymes. In this case, the replication fork is thought to convert the DNA nick into a one-ended DSB, which lacks a readily available partner for C-NHEJ. Whether C-NHEJ competes with HR at a non-enzymatic mammalian replication fork barrier (RFB) remains unknown. We previously showed that conservative “short tract” gene conversion (STGC) induced by a chromosomal Tus/Ter RFB is a product of bidirectional replication fork stalling. This finding raises the possibility that Tus/Ter-induced STGC proceeds via a two-ended DSB intermediate. If so, Tus/Ter-induced STGC might be subject to competition by C-NHEJ. However, in contrast to the DSB response, where genetic ablation of C-NHEJ stimulates HR, we report here that Tus/Ter-induced HR is unaffected by deletion of either of two C-NHEJ genes, Xrcc4 or Ku70. These results show that Tus/Ter-induced HR does not entail the formation of a two-ended DSB to which C-NHEJ has competitive access. We found no evidence that the alternative end-joining factor, DNA polymerase θ, competes with Tus/Ter-induced HR. We used chromatin-immunoprecipitation to compare Rad51 recruitment to a Tus/Ter RFB and to a neighboring site-specific DSB. Rad51 accumulation at Tus/Ter was more intense and more sustained than at a DSB. In contrast to the DSB response, Rad51 accumulation at Tus/Ter was restricted to within a few hundred base pairs of the RFB. Taken together, these findings suggest that the major DNA structures that bind Rad51 at a Tus/Ter RFB are not conventional DSBs. We propose that Rad51 acts as an “early responder” at stalled forks, binding single stranded daughter strand gaps on the arrested lagging strand, and that Rad51-mediated fork remodeling generates HR intermediates that are incapable of Ku binding and therefore invisible to the C-NHEJ machinery.
| Genomic instability is a significant contributor to human disease, ranging from hereditary developmental disorders to cancer predisposition. Two major triggers to genomic instability are chromosomal double strand breaks (DSBs) and the stalling of replication forks during the DNA synthesis (S phase) of the cell cycle. The “rules” that govern mammalian DSB repair are increasingly well understood, and it is recognized that the two major DSB repair pathways—classical non-homologous end joining (C-NHEJ) and homologous recombination (HR)—compete to repair a mammalian DSB. In contrast, we do not yet have equivalent insight into the regulation of repair at sites of mammalian replication fork stalling. Here, we explore the relationship between C-NHEJ and HR at a defined chromosomal replication fork barrier in mammalian cells. We show that, in contrast to DSB repair, repair at stalled forks does not entail competition between C-NHEJ and HR. We find that Rad51, a key mediator of HR, accumulates in an intense and highly localized fashion at the stalled fork. Based upon these findings, we propose a model of HR initiation at the stalled fork in which a Rad51-mediated fork remodeling step prevents access of C-NHEJ to the stalled fork.
| The stalling of replication forks at sites of abnormal DNA structure, following collisions with transcription complexes or due to nucleotide pool depletion—collectively termed “replication stress”—is a significant contributor to genomic instability. Inherited mutations in genes that regulate the replication stress response cause a number of human diseases, ranging from developmental disorders to highly penetrant cancer predisposition syndromes [1–5]. Replication stress is thought to be a near-universal phenomenon in tumorigenesis and some of the molecules that act upon the stalled fork are considered promising targets for cancer therapy [6]. Replication fork stalling provokes a diverse set of cellular responses, including: stabilization of the stalled replisome; regulated replisome disassembly (“fork collapse”); protection of the fork from deleterious nucleolytic processing; remodeling of DNA structure at the stalled fork; and engagement of repair or “replication restart” [5, 7–15]. The S phase checkpoint and the homologous recombination (HR) systems are intimately involved in coordinating these responses, collaborating to suppress deleterious genome rearrangements at the stalled fork [2, 16–20]. However, the mechanisms governing this coordination remain poorly understood in mammalian cells.
DNA structure at the stalled fork is remodeled by topological stresses on the chromosome at the site of stalling and by the direct action of remodeling enzymes [5, 12, 21]. The fork can be reversed to form a Holliday junction, generating a solitary DNA end which is extensively single stranded due to accompanying nascent lagging strand resection [20, 22, 23]. Other forms of template switching can also occur in the vicinity of the stall site [18, 24, 25]. Endonuclease-mediated fork breakage—either scheduled or unscheduled—can generate double strand breaks (DSBs), which might be either one-ended or two-ended [5, 20]. The DNA structures generated by fork remodeling presumably limit the repair pathways that can be engaged. Two-ended DSBs can potentially be repaired by end joining mechanisms as well as by recombination [26, 27]. In contrast, a one-ended DSB or a solitary DNA end lacks a readily available ligation partner for end joining, and may preferentially engage break-induced replication [28, 29]. Consistent with this, HR induced by a two-ended chromosomal DSB is subject to competition by classical non-homologous end joining (C-NHEJ), whereas HR induced by a nicking enzyme (“nickase”)—in which the replication fork converts the nick into a one-ended DSB—is unaffected by deletion of C-NHEJ genes [30–32]. Thus, in mammalian cells, the susceptibility of HR to competition by C-NHEJ in a particular cellular context is a useful “probe” with which to analyze the DNA structural intermediates of HR. Since the stalled fork response entails the formation of diverse DNA structures and is not restricted to two-ended DSBs, repair pathway “choice” at a stalled fork may differ from that at a defined two-ended DSB.
Study of replication-coupled repair of a covalent DNA inter-strand crosslink (ICL) in Xenopus laevis egg extracts has revealed some of the fundamental steps of stalled fork processing and repair [2, 20]. The Fanconi anemia (FA)/BRCA pathway plays a key role in detecting and processing forks bidirectionally arrested at the ICL [33–35]. The FANCD2/FANCI heterodimer orchestrates dual incisions of one of the sister chromatids on either side of the ICL. Importantly, efficient incision of the bidirectionally arrested forks is suppressed until the two opposing forks have each stalled at the ICL [36]. The resulting two-ended DSB is repaired by HR-mediated sister chromatid recombination, in which the BRCA gene products play canonical roles in promoting Rad51 loading and strand exchange functions [37–41]. HR repair of such a two-ended DSB intermediate could, in principle, be subject to competition by C-NHEJ or other end joining systems. However, recent evidence of fork reversal during ICL repair suggests that at least one of the two DSB ends is extensively single stranded [23].
Competition between HR and C-NHEJ is not a major feature of DSB repair in yeast, since C-NHEJ is a relatively low-flux pathway. Additionally, the Fanconi anemia pathway in yeast is limited to evolutionarily conserved homologs of FANCM [42, 43], suggesting that the innovation of FANCD2/FANCI-coordinated incision of bidirectionally arrested forks occurred relatively recently in evolution. Thus, although certain “core” elements of DSB repair and stalled fork metabolism are conserved between yeast and vertebrates, there are likely significant inter-species differences that remain to be fully defined. Studies in yeast, using non-enzymatic, locus-specific replication fork barriers (RFBs), show that stalled fork HR can mediate both conservative and deleterious repair, the latter including gross chromosomal rearrangements and more localized copy number changes at the site of stalling [14, 18, 19, 24, 44–48]. In contrast to the above-noted X. laevis ICL repair model, HR at an RTS1 RFB in Schizosaccharomyces pombe is not accompanied by evidence of DSB formation [19, 24]. Processing of the stalled fork in S. pombe may also trigger an aberrant form of “replication restart”, a rad22Rad52-dependent process in which the restarted fork is prone to collapse [45]. This aberrant fork restart mechanism is reminiscent of break-induced replication (BIR) in Saccharomyces cerevisiae, which is characteristically unstable and mutation-prone [49, 50]. Indeed, current models of aberrant replication restart in S. pombe invoke a migrating bubble mechanism equivalent to the mechanism of BIR in S. cerevisiae [49]. Rad52-dependent pathways have also been implicated in stalled fork repair in mammalian cells [51–53].
To facilitate analysis of mammalian stalled fork metabolism and repair, we adapted the Escherichia coli Tus/Ter RFB for use in mammalian cells [54–58]. A chromosomally integrated array of six 23 bp Ter sites mediates Tus-dependent, locus-specific replication fork stalling and HR on a mammalian chromosome, enabling direct quantitation of the repair products of mammalian replication fork stalling. We showed that conservative “short tract” gene conversion (STGC) at Tus/Ter is positively regulated by BRCA1, BRCA2, Rad51 and the Fanconi anemia pathway—consistent with the idea that STGC represents a physiological HR response to fork stalling [56, 58]. In contrast, “long tract” gene conversion (LTGC)—an error-prone HR outcome in which a replicative mechanism copies several kilobases from the partner sister chromatid—is suppressed by BRCA1 and appears to be Rad51-independent. We recently identified a novel product of stalled fork repair in primary mouse cells lacking the hereditary breast/ovarian cancer predisposition gene, Brca1: the formation of small (2–6 kb) non-homologous or microhomology-mediated tandem duplications (TDs) [58]. Tus/Ter-induced TDs in Brca1 mutant cells are mediated by a replication restart-bypass mechanism, which is completed by Xrcc4-dependent C-NHEJ. This finding, together with previous observations, suggests that C-NHEJ can access DNA ends positioned close to the site of fork stalling [59, 60]. Notably, Tus/Ter-induced STGC is a product of bidirectional replication fork stalling [56]. By analogy with the processing of forks bidirectionally arrested at an ICL in X. laevis, Tus/Ter-induced STGC might entail the formation of a two-ended DSB intermediate and might therefore be subject to competition by C-NHEJ. To test this hypothesis, we have analyzed the impact of deletion of the C-NHEJ genes Xrcc4 and Ku70 on Tus/Ter-induced HR.
To determine whether C-NHEJ interacts with HR at Tus/Ter-stalled replication forks, we targeted a 6xTer-HR reporter as a single copy to the ROSA26 locus of mouse embryonic stem (mES) cells carrying biallelic conditional alleles of the C-NHEJ gene Xrcc4 (here termed “Xrcc4fl/fl”), as described in Materials and Methods [56, 61]. The 6xTer-HR reporter contains an I-SceI target site adjacent to the 6xTer array (Fig 1A). Thus, transfection of Tus enables analysis of HR in the stalled fork response, while transfection of I-SceI in parallel samples enables analysis of DSB-induced HR. The reporter also contains elements to distinguish short tract gene conversions (STGC) from long tract gene conversions (LTGC), the latter being rare HR products in wild type cells [62, 63]. Although HR by either STGC or LTGC converts the cell to GFP+, LTGC additionally converts the cell to RFP+, by replicative duplication of an RFP cassette within the reporter (Fig 1A) [56, 64].
We transduced a ROSA26-targeted Xrcc4fl/fl 6xTer-HR reporter clone with adenovirally-encoded Cre recombinase and screened for derivative clones that had either lost (Xrcc4Δ/Δ) or retained (Xrcc4fl/fl) Xrcc4. Xrcc4 loss or retention was detected by PCR on genomic (g)DNA and was confirmed in a subset of clones by western blotting (Fig 1B and 1C). We studied HR in five independent Cre-treated Xrcc4fl/fl 6xTer-HR reporter clones and five independent Cre-treated Xrcc4Δ/Δ 6xTer-HR reporter clones in response to either Tus or I-SceI—each transfected in parallel samples (see Materials and Methods). As expected, I-SceI-induced STGC and LTGC were elevated up to 4-fold in Xrcc4Δ/Δ cells in comparison to Xrcc4fl/fl cells (Fig 1D and 1E) [30]. Interestingly, deletion of Xrcc4 stimulated STGC more strongly than LTGC; as a result, the proportion of I-SceI-induced HR events that resolved as LTGC was reduced from ~5% in Xrcc4fl/fl cells to ~2–3% in Xrcc4Δ/Δ cells (Fig 1E). The impact of Xrcc4 deletion on Tus/Ter-induced HR was quite different. Tus/Ter-induced STGC was marginally reduced in Xrcc4Δ/Δ cells in comparison to Xrcc4fl/fl cells, while Tus/Ter-induced LTGC was unaffected by deletion of Xrcc4 (Fig 1D and 1E). These results suggest that the interaction between HR and C-NHEJ at a chromosomal DSB is not recapitulated in the regulation of HR at a stalled replication fork.
To determine whether the observed phenotypes are affected by re-expression of wtXrcc4, we used lentiviral transduction to express N-terminal influenza haemagglutinin (HA)-tagged wild type mouse (m)Xrcc4 in Xrcc4Δ/Δ 6xTer-HR reporter clones #11 and #13 and in Xrcc4fl/fl 6xTer-HR reporter clones #8 and #39. Briefly, we adapted the lentiviral vector pHIV-Zsgreen [65] by replacing the Zsgreen cDNA with a bicistronic cDNA encoding the enzyme nourseothricin (NTC) acetyl transferase (NAT) [66] fused via a self-cleaving T2A peptide to the human (h)CD52 antigen (S1A Fig) [67]. Transient expression of the empty pHIV-NAT-CD52 vector in mouse ES cells produced strong cell surface staining of hCD52, as revealed by immunostaining using an anti hCD52-specific monoclonal antibody [68] (S1B Fig). Transduction of mES cells with the empty pHIV-NAT-CD52 vector, followed by selection in NTC, generated pools of transduced cells that stained strongly and specifically with anti-hCD52, whereas transduction with pHIV-NAT (i.e., lacking hCD52 expression), followed by NTC selection, generated no CD52-specific cell surface signal (S1B Fig). CD52 expression levels in pHIV-NAT-CD52-mXrcc4-transduced, NTC-selected mES cells were lower than in control empty vector (pHIV-NAT-CD52)-transduced controls, possibly reflecting constraints imposed by Xrcc4 expression from the multicistronic lentiviral expression cassette. Nonetheless, exogenous wtXrcc4 was overexpressed in comparison to endogenous Xrcc4, as revealed by RT-qPCR and by western blotting in lentivirally transduced Xrcc4fl/fl cultures (Fig 2A and 2B). As expected, re-expression of wtXrcc4 complemented the sensitivity of Xrcc4Δ/Δ cells to the radiomimetic drug phleomycin (Fig 2C).
Xrcc4Δ/Δ 6xTer-HR reporter cells transduced with pHIV-NAT-CD52-Xrcc4 and selected in NTC revealed suppression of I-SceI-induced HR to levels equivalent to that observed in isogenic Xrcc4fl/fl 6xTer-HR reporter cells (Fig 2D). Indeed, I-SceI-induced STGC and LTGC were each restored to wild type levels and the ratio of LTGC:Total HR reverted from ~2% to ~4% in Xrcc4-transduced Xrcc4Δ/Δ cells. Parallel cultures transduced with pHIV-NAT-CD52 empty vector and selected in NTC retained the original Xrcc4Δ/Δ phenotype. These experiments confirm that Xrcc4 affects the balance between I-SceI-induced STGC and LTGC, suppressing STGC more strongly than LTGC. In contrast, all measures of Tus/Ter-induced HR were unaffected by re-expression of wtXrcc4 in Xrcc4Δ/Δ cells (Fig 2D). To confirm these findings, and to minimize opportunities for cellular adaptation during complementation with wtXrcc4, we used transient transfection to restore expression of wtXrcc4 in Xrcc4Δ/Δ cells. Consistent with the above-noted findings, transient Xrcc4 expression strongly suppressed I-SceI-induced HR in Xrcc4Δ/Δ 6xTer-HR reporter cells, but had no significant impact on Tus/Ter-induced STGC or LTGC in these cells (S2 Fig). Taken together, these experiments show that Xrcc4 status has no impact on Tus/Ter-induced HR in mouse ES cells.
We showed previously that STGC at Tus/Ter-stalled forks is controlled by the HR proteins BRCA1, CtIP, BRCA2 and Rad51 and by the structure-specific nuclease scaffold SLX4 [56, 58]. In contrast, Tus/Ter-induced LTGC is suppressed by BRCA1 and is independent of BRCA2 or Rad51. We found that these relationships were unaffected by Xrcc4 status (Fig 3A). In the regulation of I-SceI-induced HR, we previously noted a specific role for BRCA1 and CtIP in suppressing an HR bias towards LTGC [64]. In contrast, loss of BRCA2 or Rad51 had little impact on the LTGC/Total HR ratio in response to an I-SceI-induced DSB. We observed similar effects on I-SceI-induced HR in Xrcc4Δ/Δ 6xTer-HR reporter cells (Fig 3B). Thus, although Xrcc4 deletion affects the ratio of LTGC:total HR in response to I-SceI, the interactions between HR mediators in execution of HR appear to be largely unaffected by loss of C-NHEJ.
DNA polymerase θ, encoded by the POLQ gene, has been implicated in an alternative end joining (A-EJ) pathway and in the prevention of genomic instability at sites of replication fork stalling [69–72]. Polθ has also been found to suppress DSB-induced HR in some cell types [73, 74]. We therefore asked whether Polθ interacts with HR in mouse ES cells, either at a Tus/Ter RFB or in DSB repair. Interestingly, siRNA-mediated depletion of Polθ modestly suppressed Tus/Ter-induced STGC in multiple clones, but in each case the effect failed reach statistical significance (Fig 4). Depletion of Polθ had no impact on I-SceI-induced HR either in wild type or Xrcc4 null cells. These findings raise the possibility that Polθ supports conservative STGC at stalled forks. They also suggest that the previously reported competition between Polθ and HR in DSB repair is not a feature of mouse ES cells [73, 74].
The binding of the Ku70/Ku80 heterodimer to DNA ends is required for engagement of C-NHEJ [75]. Ku has also been implicated in modulation of repair functions at forks stalled by the action of Topoisomerase I inhibitors, where one-ended breaks are thought to predominate [76, 77]. To determine whether Ku DNA end binding activity can influence Tus/Ter-induced HR independent of later steps of the C-NHEJ pathway, we targeted a single copy of the 6xTer-HR reporter to the ROSA26 locus of Ku70–/–mES cells [78]. Nine independent ROSA26-targeted Ku70–/– 6xTer-HR reporter clones revealed wild type levels of Tus/Ter-induced HR but greatly elevated levels of I-SceI-induced HR (Fig 5). To complement this phenotype, we co-transfected either Tus or I-SceI expression vectors with either empty vector or with a vector for expression of wt human KU70. Transient expression of wtKU70 suppressed I-SceI-induced HR and complemented phleomycin sensitivity of Ku70–/–cells, as expected (Fig 6). In contrast, wtKU70 expression had no impact on Tus/Ter-induced HR (Fig 6).
In the processing of a conventional DSB, Ku binding to the DNA end is a barrier to DNA end resection. DNA end resection activity, initiated by CtIP and the Mre11 nuclease, can displace Ku from the DNA end, providing a mechanism by which the HR machinery can overcome the barrier formed by Ku DNA end binding [79]. To further search for evidence of Ku interaction with stalled fork HR, we determined the impact of siRNA-mediated CtIP depletion on HR in Ku70–/–cells either uncomplemented or transiently complemented with wtKU70. As previously reported, CtIP depletion reduced HR in response to Tus/Ter or to an I-SceI-mediated DSB [58], and this effect was observed in both uncomplemented and Ku70-complemented Ku70–/–cells (Fig 7A and 7B). However, the proportional impact of CtIP depletion appeared less pronounced in uncomplemented I-SceI-transfected Ku70–/–cells than in the same cells complemented with wtKU70 (Fig 7B). We quantified this effect by calculating, for each test group, the induced HR in cells that received siCtIP as a proportion of induced HR in cells that received the control siRNA directed to luciferase. Notably, for I-SceI-induced HR, this ratio was increased in uncomplemented Ku70–/–cells in comparison to wtKU70-complemented cells (Fig 7C and 7D). In contrast, for Tus/Ter-induced HR, this ratio was unaffected by Ku70 status. We interpret these results as follows: at a DSB, Ku binding creates a barrier to end resection and CtIP plays a significant role in displacing Ku. This Ku-displacing role of CtIP is not required in Ku70–/–cells, and the relative importance of CtIP in HR at a DSB in Ku70–/–cells is correspondingly less. In contrast, at a Tus/Ter RFB, CtIP plays a significant role in HR that is fully independent of Ku70. Taken together with the above findings with regard to Xrcc4, the data indicate that C-NHEJ does not compete with HR at a mammalian Tus/Ter RFB.
Rad51 loading onto ssDNA is a key step in HR. In contrast to a DSB, where ssDNA is exposed following canonical DNA end resection, the stalled fork might present ssDNA for Rad51 loading through a number of different mechanisms. To determine whether Rad51 accumulates at Tus/Ter-stalled forks, we used chromatin-immunoprecipitation to study Rad51 accumulation at the ROSA26 locus, in cells transfected with a DSB-inducing nuclease, Tus, or appropriate negative controls. To induce a DSB at ROSA26, we used either I-SceI or Cas9 targeted to the I-SceI target site by a sgRNA specific to the I-SceI site. As a negative control for I-SceI and Tus, we transfected empty expression vector. As a negative control for Cas9/I-SceI sgRNA, we co-transfected wtCas9 with a non-targeting sgRNA. The chromatin-immunoprecipitation method is further described in Materials and Methods. We assessed Rad51 recruitment at 24 and 48 hours following transfection, and assayed its enrichment near the 6xTer array or neighboring I-SceI site by quantitative real-time PCR, using primers at different positions within the ROSA26 gene (Fig 8A).
24 hours after transfection with either I-SceI or Cas9/I-SceI sgRNA, Rad51 was detected maximally at sites in close proximity to the I-SceI site, and this signal spread up to ~4 kb either side of the DSB (Fig 8B). By the 48 hour time-point, a specific DSB-induced Rad51 signal was no longer detectable (Fig 8C). The Rad51 response to a Tus/Ter RFB differed markedly. Notably, Rad51 accumulation at Tus/Ter was more intense than in the response to a DSB, even though Tus/Ter consistently induces lower HR frequencies than I-SceI in our experiments. A second striking difference was the distribution of Rad51. At the Tus/Ter RFB, Rad51 was strictly localized to within a few hundred base pairs of the RFB, with no spreading of the Rad51 signal detectable even 1.3 kb from the RFB. Third, the Rad51 signal remained detectable at Tus/Ter up to 48 hours after transfection, at a time when the DSB-induced Rad51 signal had subsided. These findings reveal that Rad51 accumulation at the Tus/Ter RFB is more intense, more sustained and more specifically localized than in the DSB response. Taken together, these findings suggest that the major DNA structures that bind Rad51 at a Tus/Ter RFB are not conventional DSBs.
In contrast to HR induced by a chromosomal DSB, where C-NHEJ competes to repair the two-ended break, we show here that HR induced by a Tus/Ter RFB in mammalian cells is unaffected by the status of the C-NHEJ genes Xrcc4 or Ku70. This shows that the fundamental mechanisms of repair pathway “choice” at a stalled replication fork and a chromosomal DSB differ markedly. The simplest explanation of these findings is that HR at Tus/Ter does not entail formation of a two-ended DSB intermediate. We recently used High Throughput Translocation Sequencing (HTGTS) to study translocation-competent DNA lesions at Tus/Ter [58]. In contrast to I-SceI-induced DSBs, where two-ended breaks predominate, the major lesions detected by HTGTS at Tus/Ter were solitary DNA ends. However, it is possible that two-ended DSB intermediates of STGC arise at Tus/Ter but are not readily detected by HTGTS. Indeed, in the X. laevis model of replication-coupled ICL repair, temporally coordinated dual incisions of one sister chromatid generate a two-ended DSB intermediate. Bidirectional replication fork stalling is a critical step in this repair process, the arrival of both forks being required for replisome disassembly, asymmetrical fork reversal, nascent lagging strand resection and FANCD2/FANCI-coordinated incisions flanking the ICL [20, 23, 36].
Significant parallels exist between Tus/Ter-induced STGC and the above-noted model of ICL repair, especially with regard to the role of bidirectional fork arrest. We previously used Southern blotting to show that Tus/Ter-induced STGC products are of a fixed size, identical to products of I-SceI-induced STGC [56]. In I-SceI-induced HR, where synthesis-dependent strand annealing (SDSA) is thought to be the dominant HR pathway, the fixed size of STGC products reflects the availability of a homologous second end of the two-ended break, which supports termination of gene conversion by annealing with the displaced nascent strand [26, 27]. Indeed, if I-SceI-induced STGC is denied a homologous second end, the STGC products retrieved are of variable size, reflecting termination of gene conversion at random sites within the reporter, without the assistance of homologous pairing/annealing [64]. These aberrant STGCs are likely completed by end joining with the non-homologous second end of the DSB [80]. In the case of Tus/Ter-induced HR, the stereotyped structure of the STGC products implies that a homologous second DNA end was available to enable termination of STGC by annealing. This second end, we believe, must originate from the second (opposing) fork that stalls at Tus/Ter [56]. In summary, the mechanism of STGC at Tus/Ter has paradoxical properties. The structure of Tus/Ter-induced STGC products and its dependency on the Fanconi/BRCA/HR pathway is suggestive of SDSA of a two-ended break. However, as shown here, C-NHEJ does not compete with Tus/Ter-induced HR. Several possible models could reconcile these paradoxical properties.
In one model, the processing of the stalled fork might entail production of a conventional DSB, but the ability of Ku to access the DNA ends productively might be impaired (Fig 9A). Indeed, unproductive binding of Ku to presumptive solitary DNA ends at Topoisomerase I inhibitor-induced DNA lesions has been reported [76, 77]. Notably, in these studies, DNA end binding by Ku was shown to modulate repair activity and to influence the requirement for early end resection activities regulated by CtIP and Mre11. In contrast, in our experiments, deletion of Ku70 had no impact on Tus/Ter-induced HR and we found no evidence of an interaction between CtIP and Ku70 in the regulation of Tus/Ter-induced HR. Thus, our findings do not fit readily with the idea that Ku binds unproductively to DSB intermediates during Tus/Ter-induced HR. In an alternative model, protein complexes at the stalled fork might deny Ku access to a conventional two-ended DSB intermediate by an as yet undefined steric exclusion mechanism. The process of V(D)J recombination in developing immune cells provides precedent for such a mechanism; the RAG protein recombination synapse both initiates incision of the recombining locus and helps to channel the DNA ends towards C-NHEJ, disfavoring engagement of alternative end joining pathways [81, 82]. However, none of our findings specifically support this model. Although inactivation of the Fanconi anemia pathway has been reported to promote C-NHEJ-mediated toxic chromosome rearrangements [59, 60], we have not yet found any genetic context in which an interaction between C-NHEJ and Tus/Ter-induced HR is “unmasked”.
A notable problem with the above-noted models, which invoke a conventional DSB intermediate, is their failure to account for the distinctive pattern of Rad51 accumulation we observe at Tus/Ter. We found that Rad51 accumulation at Tus/Ter is more intense, more sustained and more precisely localized than at a conventional DSB. These findings strongly suggest that the major DNA structures that recruit Rad51 to the Tus/Ter RFB are not conventional DSBs. We propose that Rad51 is recruited to non-DSB ssDNA structures at stalled forks and that the interaction of Rad51 with these structures accounts for the functional exclusion of C-NHEJ from stalled fork HR. A major trigger to Rad51 loading at Tus/Ter may be ssDNA gaps on the arrested lagging strand, present immediately adjacent to the Tus/Ter RFB (Fig 9B). Such ssDNA gaps would be present, albeit transiently, within a normally processive fork. However, fork stalling would render these same DNA structures abnormal, by virtue of their persistence. A static ssDNA signal at the site of stalling could provide a stable platform for the loading of Rad51. By this model, Rad51 might act as an “early responder” during stalled fork repair, as has been suggested previously [83]. If Rad51 deposition were a scheduled, early response to fork stalling, this might explain the intensity and localization of the Rad51 signal we observe at Tus/Ter.
Rad51 supports fork reversal in mammalian cells in response to a variety of DNA damaging agents [83]. Rad51-mediated template switching at the site of stalling could drive limited reversal of the collapsed fork. If initiated by Rad51-coated lagging strand gaps, this process would displace the unresected nascent leading strand as a 3’ ssDNA tail (Fig 9B). Rapid coating of the displaced ssDNA tail by RPA and Rad51 could render it inaccessible to binding by Ku and, hence, “invisible” to the C-NHEJ pathway. The hypothetical limited fork reversal intermediate envisioned by this model might be subject to further processing, leading to more extensive fork reversal and potentially enabling HR initiation without formation of a DSB. Alternatively, incision of the cruciate structure of the reversed fork could liberate a one-ended DSB with a long 3’ ssDNA tail formed by the displaced nascent leading strand. It is not yet clear whether Tus/Ter-induced HR entails the formation of such a DSB intermediate. In summary, a template switch/fork reversal model of HR initiation satisfies two of the key findings reported here: first, the intense, distinctively localized recruitment of Rad51 to the Tus/Ter RFB; second, the functional exclusion of C-NHEJ during Tus/Ter-induced HR. This hypothetical model makes a number of additional predictions, which it will be relevant to test in future studies.
An interesting feature of I-SceI-induced HR was revealed in this study. Specifically, although deletion of Xrcc4 elevated the frequencies of both STGC and LTGC, LTGC products as a proportion of all HR products were reduced from ~4% to ~2% in Xrcc4 null cells. Xrcc4 deletion did not perturb the fundamental relationships of I-SceI-induced HR control reported previously for BRCA1, CtIP, BRCA2 and Rad51 [64]. This suggests that Xrcc4 loss influences the balance between STGC and LTGC via an HR-independent mechanism. I-SceI-induced LTGCs, generated by the HR reporter used here, can be considered a type of gap repair [26]. Thus, I-SceI-induced LTGC might entail repair synthesis in one of two directions. The first would entail Rad51-mediated invasion of the misaligned GFP copy while the second would entail Rad51-mediated invasion of the correctly aligned, unbroken I-SceI site-containing GFP copy. (In the latter case, wtGFP would be generated by annealing at the point of SDSA termination.) In Xrcc4Δ/Δ cells, the loss of high flux error-free religation of I-SceI-induced DSBs might increase the proportion of cells in which I-SceI sites on both sister chromatids are broken simultaneously. In such a circumstance, the second mechanism of LTGC noted above would be suppressed. This, in turn, could lead to the observed reduction in the proportion of I-SceI-induced HR events that resolve as LTGCs in Xrcc4Δ/Δ cells.
The 6xTer-HR reporters used were assembled using standard cloning methods described previously for the 6xTer-HR reporter (REF). Stable Ter-containing plasmids were generated and manipulated in JJC33 (Tus–) mutant strains of E. coli. All primers for conventional and quantitative PCR were purchased from Life Technologies. All plasmids used for mouse embryonic stem (ES) cell transfection and 293T cell transfections were prepared by endotoxin-free maxiprep (QIAGEN Sciences, Maryland, MD). siRNA SMARTpools were purchased from GE Healthcare/Dharmacon.
Conditional Brca1 mutant mouse ES cell 1xGFP 6xTer reporters were previously described [58]. Conditional Xrcc4 mutant mouse ES cells (cells in which both Xrcc4 copies contained floxed Exon3 alleles) [61] or Ku70 mutant mouse ES cells (cells in which exon 4 and part of exon 5 is replaced with the neomycin resistance cassette [78] were thawed onto MEF feeders and subsequently maintained on gelatinized tissue culture plates in ES medium as described. 20 μg of Kpn I linearlized 6xTer/HR reporter ROSA26 targeting plasmid was introduced by electroporation of 2 x 107 cells. ES cells were plated onto 6-cm dishes containing Puromycin-resistant feeders and after 18 hours plates were supplemented with 4 μg/mL Puromycin for 24 hours. Individual colonies were picked for expansion between 9 and 14 days later. Multiple ROSA26 targeted lines were identified by PCR. HR cassette ROSA26 integration and overall structure was verified for targeted lines by Southern blotting. Multiple Xrcc4-deficient ES clones were generated by transient adenovirus-mediated Cre expression and excision of Xrcc4 Exon3. ROSA26 genotyping primers: ROSA26-sense-(CAT CAA GGA AAC CCT GGA CTA CTG); Ter-HR reporter antisense-(cct cgg cta ggt agg gga tc). KU70 status was verified by PCR: KU70 exon4 5’-sense-(CCA GTA AGA TCA TAA GCA GCG ATC G); KU70 exon5 3’-antisense-(CTC TTG TGA CTC ATC TTG AGC TGG); Exon 4/5-neo-deleted allele, KU70 3’- antisense-(GCC GAA TAG CCT CTC CAC CCA AGC G). Xrcc4 status was determined by PCR: Xrcc4 5’-sense-(ttc agc taa cca gca tca ata g); floxed allele, Xrcc4 3’-antisense-(gca cct ttg cct act aag cca tct cac); Exon 3-deleted allele, Xrcc4 3’- antisense-(taa gct att act cct gca tgg agc att atc acc). Exon3-deleted, Xrcc4-deficient mES cells were transduced with lentivirus expressing a single mRNA encoding nourseothricin acetyl transferase and human CD52 (the CAMPATH antigen), with or without wild type, hemagglutinin-epitope tagged mouse Xrcc4: pHIV-NAT-hCD52-EV (empty vector control) or pHIV-NAT-hCD52-mXrcc4. Stable cultures were selected and maintained in 100 μg/mL nourseothricin (Jenna Bioscience, AB-102L).
293T cells were propagated in standard DMEM media supplemented with 10% serum, glutamine and antibiotics. For lentivirus generation, 8 x 106 cells were seeded on 10 cm dishes and transfected 24 hours later with 5 μg pHIV, 4.45 μg psPAX2, and 0.55 μg pMD2G in antibiotic-free media using Lipofectamine 2000 (Invitrogen). Media was replaced 24 hours later, and supernatant harvested every 12 hours between 48 and 72 hours after transfection and stored at 4°C. Lentiviral particles were concentrated using Centricon Plus-70 filter devices (Millipore) per manufacturer’s instructions. 5 x 105 target mES cells were seeded per well in 6-well plate format, allowed to proliferate for 24 hours, transduced and placed under 100 μg/mL nourseothricin selection beginning 24 hours after transduction.
1.6 x 105 cells were co-transfected in suspension with 0.35 μg empty vector, pcDNA3β-myc NLS-Tus, or pcDNA3β-myc NLS-I-SceI, and 20 pmol ONTargetPlus-smartpool using Lipofectamine 2000 (Invitrogen). GFP+RFP–, GFP+RFP+ and GFP–RFP+ frequencies were scored 72 hours after transfection by flow cytometry using a Becton Dickinson 5 Laser LSRII or or Beckman Coulter CytoFlex LX in duplicate. For each duplicate sample condition, 3–6 x 105 total events were scored. Repair frequencies presented are corrected for background events and for transfection efficiency (50–85%). Transfection efficiency was measured by parallel transfection with 0.05 μg wild type GFP expression vector, 0.30 μg control vector and 20 pmol siRNA. For transient mXrcc4 rescue experiments, 1.6 x 105 cells were co-transfected in suspension with 0.4 μg empty vector, pcDNA3β-myc NLS-Tus [56], or pcDNA3β-myc NLS-I-SceI [62], and either 0.1 μg empty vector, or pcDNA3β-HA-Xrcc4 using Lipofectamine 2000. For transient hKU70 rescue experiments, 1.6 x 105 cells were co-transfected in suspension with 0.35 μg empty vector, pcDNA3β-myc NLS-Tus, or pcDNA3β-myc NLS-I-SceI, and either 0.15 μg empty vector, or pcDNA3β-hKU70 using Lipofectamine 2000. For transient hKU70 rescue experiments including siRNA treatment, 1.6 x 105 cells were co-transfected in suspension with 0.35 μg empty vector, pcDNA3β-myc NLS-Tus, or pcDNA3β-myc NLS-I-SceI, and either 0.15 μg empty vector, or pcDNA3β-hKU70 using Lipofectamine 2000 and 20 pmol siRNA.
RNA isolated from cells 48 hours after transfection was extracted using QIAGEN RNeasy Mini Kit (QIAGEN Sciences, Maryland, MD) 48 hours after transfection. All analyses of GAPDH and siRNA-targeted genes was performed using an Applied Biosystems 7300 Real time PCR System using Power SYBR Green RNA-to CTTM 1-Step Kit (Applied Biosystems, Foster City, CA). SYBR green RT-qPCR assays were performed using gene-specific primer sequences identified using the NIH NCBI Nucleotide utility for GAPDH, Slx4, Brca1, Brca2, CtIP, and Polq. Primers for RT-PCR: GAPDH-sense-(CGT CCC GTA GAC AAA ATG GT); GAPDH-antisense-(TCG TTG ATG GCA ACA ATC TC); Slx4-sense-(GTG GGA CGA CTG GAA TGA GG); Slx4-antisense-(GCA CCT TTT GGT GTC TCT GG); Brca1-sense-(ATG AGC TGG AGA GGA TGC TG); Brca1-antisense-(CTG GGC AGT TGC TGT CTT CT); Brca2-sense-(TCT GCC ACT GTG AAA AAT GC); Brca2-antisense-(TCA AGC TGG GCT GAA GAT T); CtIP-sense-(AGG AGA AGG AGG GGA CGC); CtIP-antisense-(TGA AAT ACC TCG GCG GGT G); Polq-sense-(TGC TTG GTC ACG TCT TGG AA); Polq-antisense-(CCT GAA ACA GAC TCT GGA GGT). mRNA was measured in triplicates. siRNA-target gene expression level was normalized to GAPDH and expressed as a fold difference from siLuciferase control treated samples analyzed in the same experiment (x = -2ΔΔCt, with ΔΔCt = [Ct target-CtGapdh]-[CtsiLUCIFERASE-CtsiGAPDH]). Error-bars represent the standard deviation of ΔCt (SDEV = √[SDEVTARGET2 + SDEVGAPDH2]). We used the Roche ProbeFinder utility based on Primer 3 software (Whitehead Institute, MIT) to generate gene-specific primer sequences for mouse Xrcc4 and human KU70: Xrcc4-sense-(AAA TGG CTC CAC AGG AGT TG); Xrcc4-antisense-(GGT GCT CTC CTC TTT CAA GG); KU70-sense-(ACA AGT ACA GGC GGT TTG CT); KU70-antisense-(TTC AGC AGT ACC AAC GGC TT). Xrcc4-specific primers mapped to exon 6 and the exon 6–7 boundary and hKU70-specific primers mapped to exon 7 and the exon 8, respectively. Xrcc4 gene expression level was normalized to GAPDH and expressed as a fold difference from a Xrcc4fl/fl reporter clone sample analyzed in the same experiment (x = -2ΔΔCt, with ΔΔCt = [CtXrcc4-CtGapdh]-[CtsiLUCIFERASE-CtsiGAPDH]). KU70 gene expression level was normalized to GAPDH and expressed as a fold difference from one Ku70Δ/Δ reporter clone sample analyzed in the same experiment (x = -2ΔΔCt, with ΔΔCt = [CtKu70-CtGapdh]-[CtsiLUCIFERASE-CtsiGAPDH]). Error-bars represent the standard deviation of the ΔCt value (SDEV = √[SDEVGene2 + SDEVGAPDH2]).
Cell lysates were prepared from cells 48 hours after transfection lysed in RIPA buffer (50mM Tris-HCl, pH 8.0, 250 mM NaCl, 0.1% sodium dodecyl sulfate, 1% NP-40 containing the protease inhibitors, PMSF, and Roche complete protease inhibitor tablet) and 10–30 μg resolved by 4–12% Bis-Tris SDS-PAGE (Invitrogen). Protein expression was analyzed by immunoblotting using the following antibodies; hRad51 (aliquot B32, 1:500), mXrcc4 (Abcam ab97351, 1:3,000), hKU70 (Thermofisher PA5-27538, 1:1000), mBrca1 (AB191042, 1:1000), HA (Abcam, ab18181, 1:500), beta-tubulin (Abcam ab6046, 1:4,000).
Live cells were prepared for measurement of cell surface expression of human CD52 as previously described. Cells were trypsinized and resuspended in FACS blocking buffer (PBS containing 1% BSA, 0.1% sodium-azide, and 5% heat-inactivated goat serum). Cells were stained for CD52 in blocking buffer: primary antibody, rat anti-hCD52 mAb YTH 34.5, 1:200 (Bio-Rad AbD Serotec Inc. MCA-1642); secondary antibody, Alexa-488 AffiniPure Goat anti-Rat IgG, 1:50 (Jackson Immunoresearch, 112-545-167). Stained cells were fixed in PBS containing 0.5% BSA, 0.05% sodium-azide, 1.5% paraformaldehyde, 1% sucrose prior to flow cytometric analysis. Cell staining was measured by flow cytometry using a Becton Dickinson 5 Laser LSRII or Beckman Coulter CytoFlex LX.
For Xrcc4 mutant cell competition experiments, 1.6 x 105 cells were co-transfected in suspension with 0.45 μg empty vector and either 50 ng empty vector or 50 ng GFP-expression plasmid using Lipofectamine 2000 (Invitrogen). For KU70 complementation cell competition experiments, 1.6 x 105 cells were co-transfected in suspension with 0.35 μg empty vector, 0.15 μg empty vector or hKU70-expression plasmid, and either 50 ng empty vector or 50 ng GFP-expression plasmid using Lipofectamine 2000 (Invitrogen). 18 hours after transfection, cells were counted, mixed 5:1, uncolored vs. GFP+ marked cells, and 5 x 104 cells plated in triplicate. 6 hours after cell plating growth medium was replaced with media containing phleomycin (Sigma-aldrich, P9564). After two days incubation, GFP+ frequencies were scored on a Beckman Coulter CytoFlex LX. Fold enrichment of cultures transiently co-transfected with GFP-expression plasmid normalized to 0 μg/mL phleomycin control. Plots represent the mean of triplicate samples from three independent experiments (n = 3).
24–48 parallel transfections of 1.6 x 105 cells were performed in suspension with 0.5 μg empty vector, pcDNA3β-myc NLS-Tus-F140A-3xHA, or pcDNA3β-myc NLS-I-SceI, or co-transfected with 0.45 μg spCas9 expression plasmid with either control (CAT CCT CGG CAC CGT CAC CC) or I-SceI-specific (GGA TAA CAG GGT AAT CAA GG) guide RNAs (in vitro transcribed, Engen sgRNA Synthesis kit, S. pyogenes, New England Biolabs E3322S, purified using RNA Clean and Concentrator Kit, Zymo Research, R1017, and quality assessed by denaturing 10% TBE-urea acrylamide gel run) using Lipofectamine 2000 (Invitrogen). 10 million 1xGFP 6xTer reporter cells [58] 24 or 48 hours after transfection were collected for chromatin immunoprecipitation (ChIP). Cells were fixed in serum free mES cell media containing 1% formaldehyde at room temperature, incubating for 15 min with gentle orbital shaking. Fixation was quenched by addition of glycine to 125 mM. Cells were lysed in lysis buffer (0.1% SDS, 20 mM EDTA, 50 mM Tris pH 8.1) containing protease inhibitor (PMSF supplemented with Roche protease inhibitor, Roche 13539320). All subsequent steps were performed in low DNA binding tubes (Fischer Scientific, 022431021). Chromatin shearing to 100–2000 bp was performed using Diagenode Bioruptor 300 with optional attached 4°C chiller. The predominant product size of ~500 bp as achieved by 20 sonication cycles, 15 seconds on and 30 seconds off. 100 μl lysate per ChIP reaction was precleared by addition of 10 μl Magna ChIP magnetic beads (Millipore Sigma, 16–663) in ChIP dilution buffer (1% Triton-X-100, 2mM EDTA, 150mM NaCl, 20mM Tris pH 8.1). Rad51 was immunoprecipitated by addition of 3 μg anti-Rad51 ChIP-grade antibody (Abcam, ab176458) and 12 hour incubation at 4°C on a Nutator mixer followed by addition of 10 μl Magna ChIP magnetic beads and additional 16 hour incubation at 4°C. Beads were washed six times using ice-cold ChIP RIPA buffer (50mM HEPES pH 7.6, 1mM EDTA, 7 mg/mL sodium deoxycholate, 1% NP-40). DNA was eluted in Elution buffer (1% SDS, 200mM sodium bicarbonate, 5.6 μg/mL RNAse A) and cross-links were reversed by 65°C overnight incubation. Protein was removed by proteinase K digest 30 min at 55°C. DNA purified by Qiagen PCR Purification column (Qiagen, 28106) was analyzed by qPCR using an ABI Prism 7300 sequence detection system and SYBR Green (Applied Biosystems, 4368702). Primers for qPCR: 79 bp CEN-sense-(CAA CAG CCA CAA CGT CTA TAT CAT G); 79 bp CEN-antisense-(ATG TTG TGG CGG ATC TTG AAG); 1.3 kp CEN-sense-(CAC CAC AAA TCG AGG CTG TA); 1.3 kp CEN-antisense-(GGA TCA AGG CAA AGG ATC AA); 4.1 kp CEN-sense-(TCC GGT GAA TAG GCA GAG TT); 4.1 kp CEN-antisense-(CAG GGA AAC CCA AAG AAG TG); 7.1 kp CEN-sense-(TGC AAA AAC CAT CCA AAC AA); 7.1 kp CEN-antisense-(GTG GAG GCT AGA AGC TGG TG); 165 bp TEL-sense-(TGG TGA GCA AGG GCG AGG AGC); 165 bp TEL-antisense-(TCG TGC TGC TTC ATG TGG TCG); 2.1 kp TEL-sense-(GGG AGG CTA ACT GAA ACA CG); 165 bp TEL-antisense-(GGT GGG GTA TCG ACA GAG TG); 3.1 kp TEL-sense-(GCA CGT TTC CGA CTT GAG TT); 165 bp TEL-antisense-(TCA GAG CGA CTT TGG GAG AG); 11 kp TEL-sense-(CAG GAA TTC TTT CCC CAC AA); 165 bp TEL-antisense-(TGC CAG GTC TCT AGG GCT TA). Data are presented as the mean calculated from three independent experiments (n = 3) normalized against untreated controls (empty vector or guide RNA controls) and control locus (beta-actin) using the 2-ΔΔCT method as described previously [84].
Data presented represents the arithmetic mean and error bars represent the standard error of the mean (s.e.m.) of between three (n = 3) and nine (n = 9) independent experiments (n values given in figure legends). Figure legends specify the number of replicates within each independent experiment (performed in duplicate) and the number of independent experiments (n) that were performed to generate the data presented. The arithmetic mean of samples collected for groups of independent experiments for repair frequency statistical analysis, was calculated and data points for each independent experiment used to calculate the mean and standard error of the mean (s.e.m.), calculated as standard deviation/√n, (n indicates the number of independent experiments). Differences between sample pairs repair frequencies were analyzed by Student’s two-tailed unpaired t-test, assuming unequal variance. One-way ANOVA statistical analysis of greater than three samples was performed when indicated. P-values are indicated in the figure legends. All statistics were performed using GraphPad Prism v6.0d software.
|
10.1371/journal.ppat.1002940 | Genome Analyses of an Aggressive and Invasive Lineage of the Irish Potato Famine Pathogen | Pest and pathogen losses jeopardise global food security and ever since the 19th century Irish famine, potato late blight has exemplified this threat. The causal oomycete pathogen, Phytophthora infestans, undergoes major population shifts in agricultural systems via the successive emergence and migration of asexual lineages. The phenotypic and genotypic bases of these selective sweeps are largely unknown but management strategies need to adapt to reflect the changing pathogen population. Here, we used molecular markers to document the emergence of a lineage, termed 13_A2, in the European P. infestans population, and its rapid displacement of other lineages to exceed 75% of the pathogen population across Great Britain in less than three years. We show that isolates of the 13_A2 lineage are among the most aggressive on cultivated potatoes, outcompete other aggressive lineages in the field, and overcome previously effective forms of plant host resistance. Genome analyses of a 13_A2 isolate revealed extensive genetic and expression polymorphisms particularly in effector genes. Copy number variations, gene gains and losses, amino-acid replacements and changes in expression patterns of disease effector genes within the 13_A2 isolate likely contribute to enhanced virulence and aggressiveness to drive this population displacement. Importantly, 13_A2 isolates carry intact and in planta induced Avrblb1, Avrblb2 and Avrvnt1 effector genes that trigger resistance in potato lines carrying the corresponding R immune receptor genes Rpi-blb1, Rpi-blb2, and Rpi-vnt1.1. These findings point towards a strategy for deploying genetic resistance to mitigate the impact of the 13_A2 lineage and illustrate how pathogen population monitoring, combined with genome analysis, informs the management of devastating disease epidemics.
| We have documented a dramatic shift in the population of the potato late blight pathogen Phytophthora infestans in northwest Europe in which an invasive and aggressive lineage called 13_A2 has emerged and rapidly displaced other genotypes. The genome of a 13_A2 isolate revealed a high rate of sequence polymorphism and a remarkable level of variation in gene expression during infection, particularly of effector genes with putative roles in pathogenicity. Collectively, these polymorphisms, in combination with an extended biotrophic phase, may explain the aggressiveness of 13_A2 and its ability to cause disease on previously resistant potato cultivars. The genome analysis identified conserved effectors that are sensed by potato resistance genes. These findings provide options for the strategic deployment of host resistance with a positive impact on crop yield and food security. This work stresses the benefits of a crop disease management strategy incorporating knowledge of the geographical structure, evolutionary dynamics, genome sequence diversity and in planta-induced effector complement of pathogen lineages.
| As the cause of potato late blight, Phytophthora infestans is one of the most destructive plant pathogens within this genus of fungus-like oomycetes and widely known as the Irish potato famine pathogen [1], [2]. P. infestans has migrated from Central or South America [3], [4], where it infects many native solanaceous hosts, and remains responsible for significant losses to key staple crops (potato, tomato and other solanaceous plants) worldwide [5], [6]. Potato late blight management relies on regular applications of a range of anti-oomycete ‘fungicides’. However, under optimal weather conditions the pathogen may complete several infection cycles a week on a susceptible host, with control failure leading to rapid epidemics and crop loss. Resistance breeding offers great potential but the durability of resistance conferred by R genes has been continually challenged by the evolution of new virulence traits within pathogen populations [7]. P. infestans is normally diploid with a heterothallic (i.e. outbreeding) mating system that requires co-infection of A1 and A2 mating types to form long-lived sexual oospores. A mixture of sexually compatible A1 and A2 mating types increases the opportunities for sexual reproduction, providing the pathogen with an evolutionary advantage via increased genetic diversity and oospores as a source of primary inoculum in the soil [8], [9]. In the absence of oospores, in temperate regions the pathogen can only survive as asexual clones in potato tubers (as seed, in discard piles or unharvested tubers). Mycelium from such infections generates sporangia that are carried by wind and rain-splash to a new host where they germinate directly or release multiple motile zoospores that infect, colonize and release new sporangia via host stomata. Many studies have demonstrated that, despite the theoretical advantages of sexual recombination [8], a succession of clonal lineages of P. infestans have dominated the population in many potato production regions [7], [10]. In Europe, the incursion of the A2 mating type occurred 135 years after the A1 type [11]. However until recently, the A2 type remained infrequent in most parts of Europe [10], [12], which limited the opportunities for sexual reproduction of the pathogen [10], [13], [14]. Conversely, in parts of Mexico and the Nordic regions of Europe, populations of P. infestans have more balanced A1:A2 mating type ratios and are genetically diverse, with sexually formed oospores that act as a source of primary inoculum [7], [15].
Effective management of potato late blight is aided by an understanding of the characteristics of the contemporary pathogen population. For example, the aggressive and metalaxyl resistant A2 mating type US-8 lineage replaced the US-1 lineage which resulted in significant potato crop losses across the USA from 1985–1995 [16]. Pathogen genetic diversity has been monitored using a range of genetic markers [17] of which simple sequence repeats (SSRs) have recently proved effective for defining multilocus genotypes (MLGs) [18]. Key adaptive traits such as the ability of sporangia or zoospores to infect and colonise host tissue (aggressiveness) combined with efficient dissemination and, in temperate regions, survival from season to season (fitness) determine the success of particular P. infestans MLGs. Lesion growth rate and the period from inoculation to sporulation (latent period) are important components of aggressiveness [19], [20]. Fitness, a measure of reproductive success [21], is best studied in the field over several disease cycles. In a polycyclic disease such as potato late blight, even minor differences in aggressiveness or fitness can have a significant effect on the relative success of an MLG in the population. Traits such as ability to overcome specific host resistance, fungicide resistance or altered response to environmental conditions [22] are also important determinants of evolutionary success in the pathogen population.
The sequenced genome of P. infestans strain T30-4 provides a ‘blueprint’ of the gene complement and genome architecture of this pathogen [23]. The assembly served as a reference sequence in this work. Recently, two additional isolates PIC99189 and 90128 were resequenced using 36 bp Illumina reads (10.4× and 17.1× coverage, respectively) [24]. These projects revealed that P. infestans possesses a ‘two-speed genome’ with gene dense and gene-sparse repeat-rich regions. Gene-sparse regions (GSRs) are enriched in genes that are induced in planta and genes showing presence/absence polymorphism, copy number variation (CNV) or high nonsynonymous over synonymous substitution rates [24]. Effectors and other pathogenicity factors [23] that reside in these GSRs have the potential to evolve rapidly [24], consistent with the pathogen's well-documented capacity to adapt to novel host resistance. These effectors include RXLRs, a class of host translocated proteins that carry an N-terminal signal peptide followed by an RXLR motif [23], [25]. All known effector genes with Avr (avirulence) activity are in planta-induced genes of the RXLR type [26]. The study of the RXLR repertoire in emerging P. infestans lineages provides insights into the molecular basis of the infection phenotype on plants carrying the cognate R genes.
In the present study, we investigated changes in the population of the late blight pathogen P. infestans in Great Britain (GB) and identified a major new lineage of P. infestans that first emerged in mainland Europe in 2004. We investigated the factors driving this population change, demonstrating that 13_A2 MLG was amongst the most aggressive and fit MLGs in laboratory and field studies and able to overcome an important, previously durable source of host resistance. We sequenced the genome of an isolate of the 13_A2 MLG and compared it to the reference genome strain T30-4. We identified genes unique to this MLG, signatures of positive selection and CNVs, in particular in the RXLR effector repertoire. We also studied patterns of gene expression during an infection time course and noted a remarkable extended biotrophic phase, with distinct sustained induction of genes including RXLR effectors in the 13_A2 MLG isolate compared to other reference isolates. Lastly, we evaluated the effectiveness of promising sources of R genes that recognise invariant Avr genes, demonstrating that they remain effective against a 13_A2 MLG isolate. Despite the differential expression of many RXLR effector genes, we present evidence of a common set of in planta-induced effectors which we consider ‘targets’ for durable late blight disease resistance breeding.
We collected and determined the simple sequence repeat (SSR)-based [18] multilocus genotypes (MLGs) of 4,654 P. infestans isolates from 1,100 late blight disease outbreaks in Great Britain (GB), sampled between 2003 and 2008 (Table S1 in Text S2, Figure S1 in Text S1) cross-referencing these to a sample of isolates (n = 537) collected in previous GB surveys from 1982–1998 [13], [27], [28]. These SSR markers yielded between 2 and 25 alleles per locus and proved an effective tool to discriminate isolates within the GB pathogen population (Figure S2 in Text S1, Table S2 in Text S2).
The P. infestans population was dominated by clonal lineages with fewer than seven MLGs comprising >82% of the isolates each year (Figure 1A, Table S3 in Text S2). The A2 mating type frequency increased and genetic diversity reduced markedly over the years 2005 to 2008 (Figure 1A and 1B, Figures S3, S4 in Text 1). A novel A2 mating type and metalaxyl resistant (Table S2A in Text S2) MLG, termed 13_A2, was first recorded in seven British potato crops from July 2005 and went on to rapidly displace other MLGs across the region (Figure 1C). In 2006 MLG 13_A2 was prevalent in England from late May but not sampled in Scottish crops until late August (Figure S5 in Text S1) which is consistent with a progressive crop-to-crop dispersal across the region in 2006 (Figure 1C). Variation within the more variable SSR loci (particularly G11 and D13) has allowed discrimination of minor variants amongst the 2,295 isolates of 13_A2 MLG in this study (Figure 1B, Table S2B in Text S2). P. infestans MLG 13_A2 was first detected in isolates collected from The Netherlands and Germany in 2004, which is corroborated by other reports of A2 metalaxyl resistant isolates in continental Europe and suggests a north-westward migration to Great Britain (GB) (Table S4 in Text S2) [29]–[31]. The ‘misc’ category of SSR genotypes is a composite of all the novel and rarely sampled MLGs representing diversity that is consistent with sexual recombination [15]. However, in contrast to some other regions of Europe where almost every isolate is genetically distinct [15], this ‘misc’ category was recovered in GB disease outbreaks at a frequency of below 5% of the population from 2003 to 2008 (Figure 1A) indicating that the population remained largely clonal over this period (Figure 1B and Figure S4 in Text S1).
We examined the selective forces behind the population displacement in extensive laboratory and field evaluations of the fitness of many isolates of P. infestans. Aggressiveness, ‘the quantity of disease induced by a pathogenic strain on a susceptible host’ [32], is a key component of pathogen fitness and was estimated by measuring lesion size and latent period (time elapsed from inoculation to spore production). Such adaptive traits contribute to the epidemiological success of this pathogen and closely correlate with spore production and infection frequency [19].
A detached leaflet laboratory screen of 26 P. infestans isolates on five contemporary potato cultivars varying in foliar late blight resistance (Tables S5 and S6 in Text S2) was conducted at 13°C and 18°C. The isolates comprised representatives of the 9 MLGs in the 2006 British survey and reference isolates from other years and other European countries. MLG 13_A2 isolates consistently ranked among the most aggressive, showing among the shortest latent periods and the largest lesions of the MLGs tested, on all potato cultivars (Figure 2, Figures S6, S7, S8 in Text S1). This effect was more pronounced at 13°C than at 18°C, suggesting that MLG 13_A2 is better adapted to cooler conditions. Consistent with its frequency in the population (Figure 1C), MLG 6_A1 was also shown to be aggressive in this test.
Measurements of the lesion size produced on two different potato cultivars by a 13_A2 MLG isolate (06_3928A) and by the reference genome strain T30-4 [23], showed that 06_3928A formed larger lesions, with a shorter latent period than T30-4 (Figure S9 in Text S1). Also, we observed marked differences in the pattern of induction of the Cdc14 gene in these two isolates during the biotrophic phase of infection on potato. This marker gene is associated with sporulation [33], and was induced earlier and more strongly in the biotrophic phase of infection by 06_3928A than by T30-4 which is consistent with the shorter latent period in 06_3928A (Figure S10 in Text S1).
The above experiments demonstrate that, in a single disease cycle, 13_A2 isolates tend to be more aggressive than other MLGs under laboratory conditions. We went further to examine the ability of MLG 13_A2 to compete directly with other MLGs over many disease cycles in a field epidemic via a ‘mark and recapture’ experiment. The central potato plant of each of 20 field plots (five cultivars) was inoculated with a mixture of five isolates: 13_A2 (isolate 06_3928A) and representatives of four other contemporary MLGs, including 6_A1 (Table S5 in Text S2). Infected leaves from the ensuing epidemic were sampled over 21 days and 716 blight lesions were fingerprinted using direct SSR analysis. 13_A2 was the most prevalent MLG recovered, being responsible for the disease in 93–100% of the lesions sampled (Figure 3A). This high frequency was noted on all five cultivars which supported the field survey data showing a high recovery rate of 13_A2 MLG isolates from the ten most sampled cultivars (Figure 3B and Figure S11 in Text S1). In accordance with our results on the aggressiveness of 13_A2 at 13°C, the cool and wet conditions during the field trial (Figure S12 in Text S1) may have favoured the spread of MLG 13_A2. Combined, these results provide strong evidence that isolates of 13_A2 MLG are more fit and aggressive than other MLGs on many host cultivars and under field and laboratory conditions, and are consistent with data on other P. infestans population displacements [34].
In field trials since 2006, significant levels of disease were observed on some cultivars known to be partially resistant to foliar blight since the 1990s, such as Stirling [35] and Lady Balfour, a cultivar used in organic production. This was supported in subsequent whole-plant resistance screens which indicated a collapse of Stirling's resistance (Figure 4). We examined the ability of many isolates of 13_A2 MLG to overcome foliar late blight resistance on eleven potato R differential plants that contain immune receptor genes derived from the Mexican species Solanum demissum. All isolates of 13_A2 were able to cause disease on all the differential plants, except R8 and R9 (Table S5 in Text S2 and Figure S13 in Text S1). This indicates that, in addition to being particularly aggressive on susceptible potato cultivars, isolates of 13_A2 caused more disease on a broader spectrum of late blight resistant potato cultivars than isolates belonging to other P. infestans MLGs.
In late blight resistant potato plants, hypersensitive cell death and resistance are triggered by recognition of specific pathogen RXLR effectors by matching R proteins [26]. Effectors are pathogen proteins delivered inside plant cells to promote host colonization, for instance by suppressing plant immunity [36]. RXLR proteins, encoded by ∼563 genes in the P. infestans T30-4 genome [23], are the main class of host translocated effectors. Some RXLR effectors are said to have an “avirulence” activity when acting as triggers of plant immunity. To determine the genetic features, in particular the effector gene repertoire, associated with the 13_A2 MLG phenotype, we generated ∼58-fold genome coverage Illumina paired-end reads of isolate 06_3928A (see details in Text S3). We processed the sequences first by aligning the reads to the reference genome of P. infestans strain T30-4 [23], and then by performing de novo assembly of unaligned reads. In total, 95.6% of the 06_3928A reads aligned to the T30-4 sequence (Table S7 in Text S2). We detected 18,106 coding sequences with an average breadth of coverage of 99.2% (Table S8 in Text S2). We optimized bioinformatic parameters for calling single nucleotide polymorphisms (SNPs) to reach 99.9% accuracy and 85.8% sensitivity (Figure S14 in Text S1). Using these parameters, we identified 22,433 SNPs in 5,879 coding sequences of 06_3928A (Tables S8 in Text S2 and Table S9). This is similar to the 20,637 and 21,370 SNPs reported for P. infestans isolates PIC99189 and 90128, respectively [24] (Table S8 in Text S2). Of the total SNPs discovered, 11,795 were unique to 06_3928A among the four examined strains, indicating a considerable degree of variation in the 13_A2 isolate (Table S9 and Figure S15 in Text S1).
To detect signatures of positive selection in the 13_A2 lineage, we calculated rates of synonymous (dS) and nonsynonymous (dN) substitutions for every gene (Table S10). Of the 22,523 coding sequence SNPs, 11,421 are nonsynonymous (51%) corresponding to an average dN/dS rate of 0.34. Secreted protein genes, particularly RXLR effector genes, show higher dN rates compared to other categories (Figure 5). Of the 405 SNPs detected in RXLR genes, 278 are non-synonymous (69%) corresponding to an average dN/dS rate of 0.53 (Table 1 and S11). RXLR effectors are modular proteins with N-termini involved in secretion and host-translocation while C-termini encode the effector biochemical activity [25], [37]. The C-terminal domains of RXLR effector genes are highly enriched in nonsynonymous substitutions as previously noted in other oomycete species (Figure 6) [38]. Several RXLR effector genes show high dN/dS ratios and multiple replacements in their C-terminal domain (Figure S16A–C in Text S1). In addition to RXLR effectors, other secreted proteins including a Kazal-like serine protease inhibitor show high dN/dS ratios (Figure S16D in Text S1). These amino acid polymorphisms could contribute to the enhanced aggressiveness and virulence phenotypes of this genotype.
To estimate copy number variation (CNV) in the resequenced genome of P. infestans 13_A2 isolate 06_3928A relative to T30-4, we used average read depth per gene and GC content correction (see Text S3). We detected 367 CNV events in 06_3928A genes, of which there are 320 duplications and 47 deletions (Tables S12, S13). In agreement with other studies [23], [24] genes showing deletions and duplications occur more frequently in the plastic gene sparse regions of the 06_3928A genome (Figure S17 in Text S1). RXLR effector genes show higher rates of CNV compared to other gene categories (Figure S18 in Text S1 and Table S13). We identified two RXLR effectors with ∼4–5 additional gene copies in the isolate 06_3928A and this was validated with a realtime PCR assay in 17 of 18 other isolates of 13_A2 MLG. Another 18 P. infestans MLGs had lower copy numbers suggesting the higher copy number duplications are unique to 13_A2 MLG isolates (Figure S19 in Text S1). Remarkably, 21% (10 out of 47) of the genes that are deleted in 06_3928A encode RXLR effectors (Table S14 in Text S2). 13_A2 MLG isolates are able to infect potatoes carrying the R1 gene (Figure S13 in Text S1) which is consistent with our finding of an ∼18 Kb deletion encompassing the region surrounding the Avr1 RXLR effector gene in the 06_3928A isolate (Figure S20 in Text S1) [26], [39].
To identify sequences that are unique to 06_3928A, we performed de novo assembly of the unmapped Illumina reads and identified a total of 2.77 Mb contigs that did not align to P. infestans T30-4 sequences. Ab initio and homology based gene calling in these 06_3928A-specific contigs revealed 6 candidate RXLR effector genes absent in the T30-4 reference genome strain (Table S14 in Text S2). All 6 RXLR genes were subsequently confirmed by PCR on genomic DNA to be present in the 06_3928A isolate and absent in T30-4 (see Text S3, Table S15 in Text S2). Among these, a highly divergent homolog of Avr2 evades recognition by the R2 resistance gene and explains the virulence of 06_3928A on R2 potatoes (Tables S14, S15 in Text S2 and Figure S13 in Text S1) [26], [40]. Interestingly, the PCR testing also showed that the six novel RXLR genes in the 06_3928A isolate of 13_A2 MLG are present in various combinations in other multilocus genotypes (MLGs) sampled from Great Britain. This illustrates the heterogeneity of the RXLR effector repertoire that can occur within the wider P. infestans population. These findings point to a series of genetic polymorphisms that collectively contribute to the aggressiveness and virulence phenotype of the 13_A2 MLG.
The phenotype of the 13_A2 MLG may not only result from changes in gene coding sequences as documented above, but also from changes in gene expression. An infection time course was performed by hybridizing NimbleGen microarrays with cDNA from potato leaves harvested at 2–4 days post inoculation (dpi) with P. infestans 06_3928A, the T30-4 reference genome strain, and a third strain, NL07434, collected in 2007 in The Netherlands (see Text S3). We observed frequent expression polymorphisms between the three strains with 1,123 genes specifically induced in 06_3928A, compared with 110 in T30-4 and 891 in NL07434 (Figure 7A, Table S16). Remarkably, only 398 out of 4,934 genes were induced in all three strains indicating distinct isolate-specific sets of genes induced during potato infection (Figure 7A). P. infestans effector genes are sharply induced during the biotrophic phase of infection, when the pathogen associates closely with living plant cells [23], [26]. We identified 104 RXLR effector genes that are induced during biotrophy in 06_3928A compared to only 79 and 68 in T30-4 and NL07434, respectively (Figure 7A, Table S11). Of these 104 RXLR genes, expression of 20 was specifically detected in the 06_3928A isolate but not in the other two (Figure 7A, Figure S21 in Text S1). In contrast, 18 RXLR effector genes are not induced in 06_3928A but are induced in at least one of the other two isolates (Figure 7A). One of these genes, Avr4 is recognized by the R4 resistance gene [26], [41]. The lack of induction of Avr4 in 06_3928A (Figure S21 in Text S1) is consistent with the virulence of 13_A2 isolates on plants containing R4 (Figure S13 in Text S1). The updated repertoire of RXLR effectors and their expression profiles presented in this study provides additional data for systems biology approaches to understanding the role of effectors in plant-microbe interactions [42].
We noted a distinct temporal pattern of in planta gene induction in 06_3928A. Most up-regulated genes in this isolate showed sustained induction over 2 and 3 dpi in contrast to T30-4 and NL07434, in which transcript abundance generally declines at 3 dpi (Figure 7B–C, Table S16) coinciding with the onset of host tissue necrosis [23]. These findings prompted us to determine the extent to which gene induction patterns and disease progression correlate in 06_3928A and these other isolates. Microscopic observations of lesions caused by 06_3928A revealed significantly larger biotrophic zones during infection (Figure 7D). The genes showing a sustained induction period in 06_3928A include putative virulence factors such as RXLR effectors, cell wall hydrolases, proteases and protease inhibitors (Table S16). The extended biotrophic phase of 06_3928A during host plant colonization, combined with expression of a range of effectors and other secreted virulence determinants, likely contribute to the enhanced aggressiveness (Figure 2) and field fitness of MLG 13_A2 isolates. However, additional work is required to determine exactly which genes contribute to MLG 13_A2 aggressiveness and fitness.
The genome analyses of MLG 13_A2 offers opportunities to identify useful forms of host resistance. The 45 “core” RXLR effectors showing in planta gene induction during biotrophy in all 3 examined strains include 5 known avirulence effector genes (Figure 7A). Whilst homologs of Avr2 [40] and Avr3a [43] in the 06_3928A isolate contain sequence polymorphisms and are known to evade recognition in plants carrying the corresponding R2 and R3a genes (Figure S13 in Text S1), Avrblb1 [44], Avrblb2 [45] and Avrvnt1 [46] occur as intact coding sequences that are induced during infection (Figure 8A). These three Avr effectors are therefore predicted to be recognized by their cognate immunoreceptors. To determine whether 13_A2 MLG can infect plants carrying the Rpi-blb1, Rpi-blb2 and Rpi-vnt1.1 resistance genes, we used isolate 06_3928A to inoculate stable transformant potato cv. Desiree lines expressing, independently, each of the three R genes. In each case, 06_3928A was unable to infect the R potatoes and triggered a typical hypersensitive response (Figure 8B) indicating that the three R genes are effective against this 13_A2 MLG isolate. Such sources of resistance will thus be an effective component of any integrated management system against late blight caused by genotype 13_A2.
We report the emergence of an aggressive clonal lineage of P. infestans, multilocus genotype (MLG) 13_A2, and its rapid displacement of other genotypes within the Great Britain (GB) population. MLG 13_A2 has overcome previously durable disease resistances in potato, such as in cultivar Stirling and is resistant to phenylamide fungicides. Late blight caused by this lineage has thus proved challenging to manage and its migration to other potato growing regions of the world poses a threat to sustainable crop production. Therefore, there is a need, when developing a strategy for deploying disease resistance, to identify and respond rapidly to dramatic changes, and new epidemics caused by emerging genotypes within the pathogen population. Genome analyses of the 13_A2 isolate 06_3928A revealed a high rate of sequence variation and a remarkable pattern of extended biotrophic growth, which may explain 13_A2's aggressiveness and ability to cause disease on previously resistant potato cultivars. The genome analysis proved valuable in identifying RXLR effectors sensed by potentially durable potato resistance genes. This stresses the benefits of a crop disease management strategy incorporating knowledge of the geographical structure and evolutionary dynamics of pathogen lineages combined with data on their genome sequence diversity (and in planta induced effector gene complement). Such data, when linked to the host R gene repertoire [47], offers options for strategic deployment of host resistance with a positive impact on crop yield and food security.
P. infestans isolates were obtained from more than 1,100 outbreaks of potato late blight across Great Britain (GB) from 2003 to 2008. The locations of 672 outbreaks sampled in 2006 to 2008 and further details on sampling and pathogen characterisation are available (Figure S1 in Text S1 and Text S3). The mating type of each of 4,654 isolates collected in this study was tested by pairing with known tester isolates on Rye A agar plates. After an initial screen of the new A2 mating type lineages using the RG57 [48] RFLP probe (Table S2A in Text S2), all isolates were genotyped using 11 SSR markers [18] in 3 multiplexed PCR assays using fluorescently labelled primers on an ABI 3730 capillary sequencer (Tables S2 and S3 in Text S2 and Text S3). The SSR data were used to define MLGs, explore the relatedness amongst the multilocus genotypes (MLGs) and to describe the population change. Due to the presence of three alleles in some isolates, we calculated clonal distance [49] using the infinite alleles mutation model, to quantify genetic distance between MLGs. This distance essentially counts the number of alleles that differ between individuals. Isolates with null alleles were included, but any isolates that were not genotyped at one or more loci were excluded. Distance among multilocus genotypes was calculated in GenoDive (Distributed by P. G. Meirmans at http://www.bentleydrummer.nl/software/software/Home.html). Minimum spanning networks were calculated by MINSPNET [50] and visualized using neato in the Graphviz package [51]. The numbers of isolates used to construct the trees were 748, 795, 1,072, and 892 for 2003–2005, 2006, 2007, and 2008, respectively (Figure 1B and Figure S4 in Text S1).
Representative isolates from the main MLGs from Great Britain plus a selection of reference isolates from other countries were used to examine two components of aggressiveness [19] (lesion size and latent period) on five contemporary potato cultivars (Tables S5 and S6 in Text S2) as follows. For each cultivar, leaflets of a similar age and size were placed in clear plastic boxes (26 leaflets per box) lined with moist tissue paper. After chilling to stimulate zoospore release, a droplet of 30 µl of inoculum (approx 420 sporangia) of each of the 26 isolates was applied to the centre of each leaflet. A total of 60 boxes of leaves were inoculated and 30 placed in a randomised block design with six replicate blocks in each of two adjoining illuminated walk-in growth rooms set at a constant 13°C or 18°C with 16/8 hours of light and dark. The 1,560 leaflets were scored daily for first symptoms (i.e. infection period, IP), and sporulation (i.e. latent period, LP) and at six days post inoculation (dpi), lesion size was measured in two orientations at right angles to each other using electronic calipers connected to a laptop computer.
A randomised block field trial comprising four replicate 25 plant plots of the five potato cultivars used in the laboratory assay was established. In mid-July an equal mixture of sporangia of 5 isolates (different MLGs) were used to infect the lower leaves of the central plant in each plot. Once the disease had spread from the central plant, single lesions were sampled from the epidemic over the following three weeks and direct SSR fingerprinting of P. infestans from lesions pressed onto FTA cards (Whatman, UK) was used to determine the MLG. For additional details see Text S3.
Genome sequencing of P. infestans 13_A2 isolate 06_3928A was performed in 2G GAs (Illumina Inc.) and alignments were obtained with Burrows-Wheeler Transform Alignment (BWA) software package v0.5.7 with a seed length (l) of 38 and a maximum of mismatches (M) allowed of 3 as parameters [52]. Unmapped reads of P. infestans 13_A2 isolate 06_3928A were assembled using VELVET software package v1.0.18 [53] and mapped to the reference genome using NUCmer program from MUMmer software package v3.2 (see details in Text S3) [54].
A False Discovery Rate (FDR) analysis was used to determine the performance of single nucleotide polymorphism (SNP) calling in the 06_3928A genome (Figure S14 in Text S1 and Text S3). Single nucleotide polymorphisms (SNPs) were called using a 90% consensus among reads calling a SNP with a minimum of 10× coverage (Figure S16 in Text S1). Rates of synonymous substitution (dS), non-synonymous substitution (dN) and omega (dN/dS) were calculated using the yn00 program of PAML [55] by implementing the Yang and Nielson method [56] for every coding gene predicted in 06_3928A in comparison to the homologous gene in the reference genome strain T30-4 (Figure 5, Table S10). Differences in frequencies of nonsynonymous minus synonymous SNPs were counted per 15 bp-long windows and sliding by 3 bp steps. Frequencies were calculated as the number of SNPs per bp per gene and averaged over 20 consecutive windows (Figure 6A). The 20 windows adjacent to the RXLR motif were considered for each of the domains. Numbers of SNPs in RXLR gene domains were counted per 15 bp-long windows and sliding by 3 bp steps (Figure 6B). A total of 118 RXLRs, 3,077 core orthologs and 2,442 gene-dense regions (GDR) genes that contain at least 1 SNP were analyzed (Figure 6).
The NimbleGen microarray data are available in GEO under accession number GSE14480 for P. infestans T30-4 [23] and GSE33240 for P. infestans 06_3928A and NL07434. Genes that are induced in planta were identified using a t-test (p value<0.05, >2 fold expression changes) and False Discovery Rate (FDR) analysis (q-value<0.05) [57] in samples from infected potato leaves relative to plate-grown in mycelia (see more details in Text S3).
|
10.1371/journal.pgen.1005422 | Retrohoming of a Mobile Group II Intron in Human Cells Suggests How Eukaryotes Limit Group II Intron Proliferation | Mobile bacterial group II introns are evolutionary ancestors of spliceosomal introns and retroelements in eukaryotes. They consist of an autocatalytic intron RNA (a “ribozyme”) and an intron-encoded reverse transcriptase, which function together to promote intron integration into new DNA sites by a mechanism termed “retrohoming”. Although mobile group II introns splice and retrohome efficiently in bacteria, all examined thus far function inefficiently in eukaryotes, where their ribozyme activity is limited by low Mg2+ concentrations, and intron-containing transcripts are subject to nonsense-mediated decay (NMD) and translational repression. Here, by using RNA polymerase II to express a humanized group II intron reverse transcriptase and T7 RNA polymerase to express intron transcripts resistant to NMD, we find that simply supplementing culture medium with Mg2+ induces the Lactococcus lactis Ll.LtrB intron to retrohome into plasmid and chromosomal sites, the latter at frequencies up to ~0.1%, in viable HEK-293 cells. Surprisingly, under these conditions, the Ll.LtrB intron reverse transcriptase is required for retrohoming but not for RNA splicing as in bacteria. By using a genetic assay for in vivo selections combined with deep sequencing, we identified intron RNA mutations that enhance retrohoming in human cells, but <4-fold and not without added Mg2+. Further, the selected mutations lie outside the ribozyme catalytic core, which appears not readily modified to function efficiently at low Mg2+ concentrations. Our results reveal differences between group II intron retrohoming in human cells and bacteria and suggest constraints on critical nucleotide residues of the ribozyme core that limit how much group II intron retrohoming in eukaryotes can be enhanced. These findings have implications for group II intron use for gene targeting in eukaryotes and suggest how differences in intracellular Mg2+ concentrations between bacteria and eukarya may have impacted the evolution of introns and gene expression mechanisms.
| Mobile group II introns are bacterial retrotransposons that are evolutionary ancestors of spliceosomal introns and retroelements in eukaryotes. They consist of an autocatalytic intron RNA (a ribozyme) and an intron-encoded reverse transcriptase, which together promote intron mobility to new DNA sites by a mechanism called retrohoming. Although found in bacteria, archaea and eukaryotic organelles, group II introns are absent from eukaryotic nuclear genomes, where host defenses impede their expression and lower intracellular Mg2+ concentrations limit their ribozyme activity. Here, we developed a mobile group II intron expression system that bypasses expression barriers and show that simply adding Mg2+ to culture medium enables group II intron retrohoming into plasmid and chromosomal target sites in human cells at appreciable frequencies. Genetic selections and deep sequencing identified intron RNA mutations that moderately enhance retrohoming in human cells, but not without added Mg2+. Thus, low Mg2+ concentrations in human cells are a natural barrier to efficient retrohoming that is not readily overcome by mutational variation and selection. Our results have implications for group II intron use for gene targeting in higher organisms and highlight the impact of different intracellular environments on intron evolution and gene expression mechanisms in bacteria and eukarya.
| Mobile group II introns are retrotransposons that also function as self-splicing introns [1]. They are found in bacteria, archaea, and in the bacterial endosymbiont-derived mitochondrial and chloroplast genomes of some eukaryotes, particularly fungi and plants [2]. Despite their prokaryotic origin, mobile group II introns are believed to have strongly impacted eukaryotic nuclear genomes as evolutionary ancestors of spliceosomal introns, the spliceosome, LINEs and other non-LTR retrotransposons, and telomerase [3,4]. Mobile group II introns insert into new DNA sites by a ribozyme-based site-specific DNA integration mechanism called retrohoming, which is thought to have enabled mobile group II introns or their close relatives to proliferate within the nuclear genomes of early eukaryotes before evolving into spliceosomal introns [4,5]. In addition to its evolutionary significance, retrohoming underlies the use of group II introns as gene targeting vectors (“targetrons”), which use intron RNA/DNA target site base-pairing interactions to achieve high and programmable DNA target specificity [6–8]. Targetrons are widely used for gene targeting in bacteria, where retrohoming frequencies are high enough to identify targeting events by colony PCR screening without using genetic markers [9]. By contrast, mobile group II introns and targetrons derived from them function inefficiently in eukaryotes [10–12], and group II introns appear to be completely absent from the nuclear genomes of present-day eukaryotes [13]. The reasons for the different behavior of group II introns in prokaryotes and eukaryotes and factors that dictated their conversion into spliceosomal introns and exclusion from eukaryotic nuclear genomes remain incompletely understood.
Mobile group II introns consist of a catalytically active intron RNA (a ribozyme) and an intron-encoded reverse transcriptase (RT), which function together to promote both RNA splicing and retrohoming [1]. The intron RNA catalyzes its own splicing from a precursor RNA via two sequential transesterification reactions that result in ligated exons and an excised intron lariat RNA, identical to the splicing reaction mechanism used by spliceosomal introns in higher organisms [4,14]. To catalyze splicing, the intron RNA folds into a conserved tertiary structure that consists of six interacting secondary structure domains (DI-DVI), with three distinct structural subclasses of group II introns, IIA, IIB, and IIC, distinguished by secondary and tertiary structure features [1]. This folded RNA forms a ribozyme active site that includes nucleotide residues highly conserved in all three group II intron subclasses and utilizes site-specifically bound Mg2+ ions to catalyze RNA splicing and reverse splicing reactions [15–17]. The group II intron RT contributes to splicing by binding to the intron RNA and promoting formation of this catalytically active RNA structure [18–20]. After splicing, the RT remains bound to the excised intron lariat RNA in a ribonucleoprotein particle (RNP) that initiates retrohoming by recognizing a DNA target site [21]. DNA target site recognition is primarily by base pairing of sequence elements within the intron RNA to DNA sequences spanning the intron-insertion site, with only a small contribution of the group II intron RT, which helps promote local DNA melting [22]. The intron RNA then uses its ribozyme activity to insert directly into the retrohoming site, where it is reverse transcribed by the intron-encoded RT into an intron cDNA that is integrated into the genome by host enzymes [5,23–26].
Early findings that group II introns use the same splicing reaction mechanism as spliceosomal introns and that some organellar group II introns have been fragmented by DNA recombination into two or three unlinked segments that reassociate to promote RNA splicing suggested an evolutionary relationship to spliceosomal introns and a possible evolutionary origin for present-day snRNAs [27]. Recently, these hypotheses have been strongly supported by group II intron RNA crystal structures and biochemical studies, which demonstrate striking structural and functional similarities between group II intron domains and three key snRNAs (U4, U5, and U6) that comprise the catalytic core of the spliceosome [17,28–31]. The similarities include identical RNA-catalyzed splicing reactions based on similarly positioned catalytic Mg2+ ions at the RNA active site [15,16,30,31]. Moreover, recent structural and bioinformatic studies indicate that the conserved spliceosomal core protein Prp8 was derived from a group II intron-like RT and functions similarly as a structural scaffold for an RNA catalytic core [32,33]. Considered together with the phylogenetic distribution of group II introns, these findings support a scenario in which mobile group II introns entered ancestral eukaryotes along with bacterial endosymbionts that gave rise to mitochondria, invaded the nucleus, proliferated to high copy number, and then degenerated into snRNAs [34]. Further, this proliferation of introns in eukaryotic nuclear genes is hypothesized to have been a major driving force for the evolution of eukaryotes themselves, including for features such as (i) the nuclear membrane to separate transcription and splicing from translation, thereby limiting mistranslation of intron-containing RNAs; (ii) nonsense-mediated decay (NMD) to degrade unspliced or misspliced intron-containing transcripts that escape to the cytosol; and (iii) large-scale alternative splicing, enabling greater organismal complexity within constraints on genome size [3].
Several factors have been identified that limit group II intron function and their ability to propagate in eukaryotes. First, studies in Saccharomyces cerevisiae showed that RNA polymerase II (Pol II) transcripts containing the Lactococcus lactis Ll.LtrB group II intron, which belongs to subgroup IIA, are subject to both NMD and translational repression, leading to their accumulation in cytoplasmic foci [11]. This translational repression appears to reflect strong intermolecular base-pairing interactions between the ligated-exon junction sequence in the spliced mRNA and the excised intron or intron-containing precursor RNAs, which may impede translating ribosomes and/or target the RNA for degradation [35]. A second factor affecting group II intron propagation in eukaryotes appears to be suboptimal intracellular Mg2+ concentrations, which limit group II intron ribozyme activity [10]. Group II intron splicing and retrohoming both require relatively high Mg2+ concentrations compared to other cellular processes, and Mg2+ concentrations appear to be significantly lower in eukaryotes than in bacteria [10,36–38]. Studies of S. cerevisiae mtDNA introns by Schweyen and coworkers showed that mutations in a mitochondrial Mg2+ transporter inhibit the splicing of all four group II introns, including both subgroup IIA and IIB introns, while having minimal effect on the transcription or splicing of group I introns, which use a different ribozyme-based splicing mechanism that is less sensitive to Mg2+ concentration [37]. Further, microinjection assays in Xenopus laevis oocyte nuclei or Drosophila and zebrafish embryos showed that in vitro reconstituted Ll.LtrB group II intron RNPs could retrohome efficiently into plasmid target sites only if additional Mg2+ was co-injected with the plasmid DNA [10]. An attempt to overcome this limitation in human cells by using an algal mitochondrial group IIB intron (Pl.LSU/2) that self-splices at physiological Mg2+ concentrations in vitro, was unsuccessful [12], perhaps because efficient self-splicing of this intron at low Mg2+ concentrations requires the presence of 1 M NH4Cl [39]. Recently, we selected variants of the Ll.LtrB group II intron with mutations in catalytic core domain V (DV) that retrohome 10- to 20-fold more efficiently than the wild-type intron in a Mg2+-deficient E. coli strain [36]. These findings suggested that it might be possible to overcome the high Mg2+ requirement that prevents efficient group II intron retrohoming in eukaryotes by mutations at a few critical sites within the intron RNA.
Here, we developed a mobile group II intron expression system for human cells that utilizes an Ll.LtrB group II intron RNA expressed by using T7 RNA polymerase (T7 RNAP) to overcome NMD and a separately expressed human codon-optimized group II intron RT. By using this expression system, we found that simply supplementing the cell culture medium with 20–80 mM Mg2+ enables the Ll.LtrB intron to retrohome into plasmid and genomic target sites, the latter at frequencies of up to ~ 0.1%, in viable human cells. Further, we performed multiple rounds of in vivo selection of the intron ribozyme, analyzed the fitness landscape using Pacific Biosciences deep sequencing, and identified positively selected mutations that were used for synthetic shuffling to generate Ll.LtrB variants that show enhanced retrohoming in human cells. However, the maximum enhancement was <4-fold and still required extra Mg2+ in the culture medium. These findings indicate that low Mg2+ concentrations constitute a natural barrier to efficient retrohoming in eukaryotes that is not readily overcome by mutational variation and selection, and they have implications for the use of group II introns for gene targeting in higher organisms and the evolution of introns and gene expression mechanisms.
The mobile group II intron expression system that we developed for human cells consists of three plasmids (Fig 1A). The first plasmid, denoted phLtrA, expresses the Ll.LtrB group II intron RT (denoted LtrA protein) with humanized codon usage and a C-terminal SV40 nuclear localization sequence (NLS) (NCBI Genbank, accession number KP851976)[40]. The humanized LtrA ORF is expressed from a constitutive RNA polymerase II (Pol II) promoter, the cytomegalovirus immediate early (CMV) promoter [41], and is followed by a polyadenylation signal (pA). An early version of this plasmid, phLtrA1, has a small artificial spliceosomal intron (IVS) inserted after the initiation codon [42] that was later found to be unnecessary for expression of hLtrA. The second plasmid, pLl.LtrB, uses a phage T7 promoter to express the Ll.LtrB intron with the LtrA ORF deleted (denoted Ll.LtrB-ΔORF) and short flanking 5’- and 3’-exon sequences (denoted E1 and E2, respectively). Finally, the third plasmid, pT7-NLS, expresses phage T7 RNAP with a fused N-terminal SV40 NLS driven by a CMV promoter. Previous work showed that T7 RNAP can produce high levels of uncapped, non-polyadenylated transcripts in human cells [43] and that its subcellular localization can be controlled, with nearly complete cytoplasmic or nuclear localization when expressed without or with an appended SV40 NLS, respectively [44]. The group II intron expression plasmids were not toxic when transfected by themselves or together into HEK-293 cells (Fig 1B).
We first compared expression of the LtrA protein with and without human optimized codons in HEK-293 cells. As shown in Fig 2A, the plasmid expressing the human-codon optimized LtrA ORF produced hLtrA protein that was readily detected by immunoblotting (lane 5 and 6), whereas an identical plasmid with a native non-codon-optimized LtrA ORF produced no detectable LtrA protein (lane 4). Further, nuclear lysates from HEK-293 cells transfected with the plasmid expressing hLtrA but not untransfected cells showed a high level of RT activity with a substrate that is efficiently used by purified LtrA protein (Ll.LtrB/E2+10 RNA; an Ll.LtrB intron-containing transcript with a DNA primer annealed downstream of the intron; Fig 2B). Immunofluorescence microscopy showed that the hLtrA expressed with a C-terminal NLS localized to the nucleus in HEK-293 cells and COS-7 cells, whereas hLtrA expressed without an added NLS (ΔNLS) localized to the cytoplasm (Fig 2C–2E). The requirements of LtrA for codon optimization and addition of an NLS to localize to the nucleus differ from recent findings for the Sinorhizobium meliloti RmInt1 group II intron RT, which does not require codon optimization and localizes to nucleoli in Arabidopsis thaliana protoplasts without an added NLS [46]. Together, our results show that optimization toward human codon usage overcomes a barrier to the expression of the Ll.LtrB group II intron RT in eukaryotes and that an appended NLS is required to localize this protein to the nucleus.
Previous studies in S. cerevisiae showed that RNA polymerase II (Pol II) transcripts containing the Ll.LtrB intron are subject to both NMD and translational repression [11]. To determine the effect of NMD on group II intron-containing transcripts in human cells, we constructed plasmids that use a CMV (Pol II) promoter to express blue fluorescent protein (BFP) with or without the Ll.LtrB-ΔORF intron and short flanking exon sequences (E1 and E2) inserted directly after the BFP start codon (Fig 3A). We transfected the plasmids into HeLa cells that were pre-treated with siRNAs targeted against UPF1 mRNA, which encodes an essential component of the NMD complex [48], or a scrambled siRNA control, and then quantified BFP transcript levels by RT-qPCR at 48 h after plasmid transfection. As shown in Fig 3B, transcript levels for the uninterrupted BFP ORF remained high in the presence of both the UPF1 and scrambled control siRNA, with little if any significant effect of NMD knockdown. By contrast, the inclusion of the Ll.LtrB intron in the BFP ORF led to a strong decrease in transcript level in the presence of the control siRNA, but not in the presence of the UPF1 siRNA to block NMD, irrespective of co-expression of the LtrA protein. UPF1 knockdown was confirmed by immunoblotting (Fig 3C). These findings indicate that the NMD pathway degrades Pol II transcripts containing the Ll.LtrB intron in human cells as it does in S. cerevisiae [11].
The finding that Pol II transcripts containing the Ll.LtrB intron are subject to NMD in human cells led us to test whether this barrier could be overcome by using T7 RNAP for Ll.LtrB expression. T7 RNAP transcripts are not capped, polyadenylated, or subject to pre-mRNA processing in the same way as Pol II transcripts and thus are not expected to be subject to NMD [43]. For these experiments, we constructed two T7-promoter-driven GFP expression plasmids, one denoted pGFP-Ll.LtrB containing the Ll.LtrB intron and short flanking exon sequences inserted within the GFP ORF, and the other denoted pGFP containing the ligated-exon sequences that would result from Ll.LtrB intron splicing inserted at the same location (Fig 4A). In both plasmids, the GFP ORF is preceded by an internal ribosome entry site (IRES) to enable GFP expression if the Ll.LtrB intron is spliced. Paralleling the protocol used for BFP-encoding Pol II transcripts in Fig 3, we transfected these GFP-encoding plasmids together with pT7-NLS, which expresses T7 RNAP, into HEK-293 cells that had been pre-treated with the UPF1 siRNA or a scrambled control siRNA, and we measured GFP transcript levels by RT-qPCR at 48 h after transfection of the plasmids. In this case, the T7-GFP ORF control and T7-Ll.LtrB-GFP transcript containing the Ll.LtrB intron were present at similar levels with either the scrambled or UPF1 siRNA, with UPF1 knockdown by the UPF1 siRNA again confirmed by immunobotting (Fig 4B). These findings indicate that a T7-transcript containing the Ll.LtrB intron is not subject to NMD pathway-related degradation in human cells.
The Pol II-transcripts with the Ll.LtrB intron inserted in BFP ORF described in the preceding sections were not spliced in human cells, and this was also the case for the T7 transcripts with the Ll.LtrB intron inserted in the GFP ORF. As we suspected that splicing of the Ll.LtrB-intron might be limited by low Mg2+ concentrations in human cells, we tested whether splicing of the T7 transcript containing the Ll.LtrB intron might be induced simply by growing cells in culture medium containing elevated concentrations of MgCl2 (Fig 4C). In these experiments, we transfected the three expression plasmids phLtrA, pT7-NLS, and pGFP-Ll.LtrB into HEK-293 cells in culture medium with or without added 80 mM MgCl2 and assayed Ll.LtrB intron splicing by RT-PCR of cellular RNAs at 48-h post-transfection. In standard culture medium, the GFP-Ll.LtrB transcript by itself showed no detectable splicing (lane 2), while co-expression of hLtrA led to low levels of splicing (lane 4). Surprisingly, the addition of MgCl2 to the culture medium by itself led to a large increase in splicing even in the absence of hLtrA (lane 3). Splicing levels in the presence of both exogenous MgCl2 and hLtrA appeared to be somewhat lower than with MgCl2 alone (lane 5). Accurate splicing was confirmed by sequencing of the ligated-exon junction in the PCR product. Notably, although the Ll.LtrB intron was spliced under these conditions, we detected no expression of GFP from the spliced transcript, whereas GFP was expressed efficiently from the control transcript containing the ligated-exon junction sequence inserted at the same location in the GFP ORF (Fig 4D). Together, these findings indicate that exogenous MgCl2 can by itself induce splicing of the Ll.LtrB intron in human cells, even in the absence of LtrA protein, which is required for Ll.LtrB intron splicing in bacteria [49]. However, T7 transcripts from which the Ll.LtrB intron had been spliced in human cells still appear to be subject to a translational block similar to what was found for RNAP II transcripts in S. cerevisiae [11].
In vitro, the LtrA protein can be reconstituted with excised intron lariat RNA to generate RNPs that are active in retrohoming [21]. Thus, we tested whether the excised intron RNA resulting from Ll.LtrB splicing in culture medium containing added MgCl2 (80 mM) could be combined with expressed hLtrA to promote retrohoming in human cells. To assess retrohoming in human cells, we used sensitive Taqman qPCR-based assays that quantify both the 5'- and 3'-integration junctions resulting from integration of the Ll.LtrB intron into the wild-type DNA target site (Fig 5A). We tested for retrohoming into a single genomic copy of the wild-type Ll.LtrB homing site in HEK-293 Flp-in cells and in the same cells after co-transfection of a recipient plasmid (pFRT) carrying the same target site. As the transfected plasmid is expected to be present in much higher copy number (~104) [50] than the genomic target site, this protocol enables direct comparison of plasmid and genomic targeting in parallel transfections of the same cells. In these experiments, a 24-h period of polyethylenimine (PEI)-mediated transfection of the expression plasmids was followed by an additional 24-h period in which cells were incubated in growth medium containing 80 mM MgCl2. After MgCl2 treatment, the cells (both adherent and non-adherent) were collected, and DNA was extracted for qPCR analysis.
Cells receiving all three expression plasmids and 80 mM MgCl2 showed significant retrohoming into both genomic and plasmid retrohoming assays (Fig 5B). In three separate experiments, the average retrohoming frequency and standard deviation for the genomic target site measured by qPCR of RNase-treated whole-cell DNA was 0.23 ± 0.02% for the 3'-integration junction and about 7-fold lower, 0.033 ± 0.002%, for the 5'-integration junction [note the different scales of the y-axis for 5’-junctions (blue bars) and 3’ junctions (red bars) in Fig 5B and 5C).] The retrohoming frequencies for cells co-transfected with the recipient plasmid, which is present at ~104 copies per cell and expected to be largely cytosolic [50,51], were substantially higher (1.4 ± 0.1% for the 3’-integration junction and 0.056 ± 0.004%, for the 5' junction). The lower frequency of 5’- than 3’-integration junctions for retrohoming of the Ll.LtrB intron into genomic and plasmid target sites may reflect that a high proportion of the retrohoming events result in the integration of 5’-truncated introns, similar to the situation for human LINE-1 elements where retrotransposition frequently results in 5’ truncations due to abortive reverse transcription [52]. Surprisingly, retrohoming efficiencies with both plasmid and genomic target sites were similar regardless of whether or not the expressed T7 RNAP contained an NLS (S1 Fig). This finding presumably reflects that RNPs resulting from transcription of pLl.LtrB that remains in the cytosol after transfection can still gain access to the genomic target site (S1 Fig; see Discussion).
For retrohoming into the plasmid target site, full-length intron integrations requiring all steps in retrohoming were confirmed by conventional PCR and sequencing of the integration junction in the above experiments (S2 Fig) and more extensively in genetic assays described below. For the genomic target site, the very low frequency of full-length intron integrations (0.033%) made it difficult to recover them from whole-cell DNA by conventional PCR. However, both the 5’- and 3’-integration junctions expected for full-length integrations were detected by Taqman qPCR assays of RNase-treated genomic DNA at levels well above background and with the same excess of 3’ junctions as found in the plasmid assay (Fig 5B and 5C). Additionally, unlike splicing of the Ll.LtrB intron in human cells, which is not dependent upon LtrA protein (see above), retrohoming of the Ll.LtrB intron into both plasmid and genomic DNA target site and the detection of both the 5’- and 3’-DNA integration junctions required the LtrA protein, which is needed for DNA target site recognition as well as reverse transcription (Fig 5B and 5C). Finally, in an important control, no significant retrohoming into the wild-type plasmid or genomic site was detected under any condition for an Ll.LtrB intron retargeted to insert into the CCR5 gene (Fig 5B). We confirmed that this CCR5 targetron retrohomes into a plasmid-borne CCR5 target in HEK-293 cells at frequencies of 0.24–0.27% for the 3’-integration junction, but could not detect integrations into the genomic CCR5 gene.
Although retrohoming frequencies in HEK-293 cells with added Mg2+ were relatively high, we observed that the addition of 80 mM MgCl2 to the culture medium to promote retrohoming resulted in cellular blebbing, a hallmark of apoptosis [53], with about half of the cells becoming non-adherent and unable to divide in fresh media. Inviable non-adherent cells could potentially have higher targeting rates due to enhanced Mg2+ influx due to more permeable cell membranes. Consistent with this possibility, we found that retrohoming frequencies at 80 mM MgCl2 were substantially higher in non-adherent cells (3’-integration junctions 0.4–1.0% and 0.4–1.5% for genomic and plasmid target sites, respectively) than in adherent cells (3’-integration junctions 0.01–0.08% and 0.13–0.50% for genomic and plasmid target sites, respectively) (Fig 5D). We tested whether lower MgCl2 concentrations, shorter targeting times, or different Mg2+ salts could alleviate the deleterious effects of added Mg2+, but found that all treatments that improved cell viability decreased retrohoming frequencies to unattractively low levels (S3 Fig). Cell populations in which the Ll.LtrB intron had integrated into the genomic site at 80 mM Mg2+ were viable and remained adherent in high MgCl2 growth medium indefinitely. Thus, these experiments indicate that the Ll.LtrB intron can retrohome into both plasmid and genomic target sites in viable human cells, the latter at frequencies as high as ~0.1% as measured by 3’-integration junctions, so long as extra Mg2+ is added to the culture medium.
The finding that retrohoming of the Ll.LtrB intron in human cells is limited by low Mg2+ concentrations led us to test whether we could select Ll.LtrB intron variants that could retrohome more efficiently at low Mg2+ concentrations in human cells. We previously selected Ll.LtrB variants with mutations in the distal stem of domain V (DV) that had 10- to 20-fold higher retrohoming efficiencies in a Mg2+-deficient E. coli mutant, as well as decreased Mg2+-dependence for RNA splicing and reverse splicing in vitro [36]. However, neither of the two best such variants had increased retrohoming efficiency into genomic or plasmid target sites in HEK-293 cells with or without 80 mM MgCl2 added to the culture medium (S4 Fig). We also tested an intron variant that was selected for enhanced retrohoming in Xenopus laevis oocyte nuclei, another environment in which low Mg2+ concentrations are stringently limiting for retrohoming [10,54]. Although this variant had ~4-fold higher retrohoming efficiency in X. laevis oocyte nuclei, it did not show higher retrohoming frequencies than wild-type Ll.LtrB in human cells in our assays (S5 Fig). A possible explanation is that these Ll.LtrB variants selected in E. coli or X. laevis are optimized for different intracellular environments and Mg2+ concentrations than those in human cells. Thus, we attempted to select Ll.LtrB variants with enhanced retrohoming directly in human cells.
For directed evolution in human cells, we adapted an E. coli plasmid-based genetic assay for retrohoming that avoids pitfalls of PCR amplification of low frequency intron-integration events [6]. In this assay, a group II intron carrying a phage T7 promoter retrohomes into a target site cloned on a recipient plasmid upstream of a promoterless tetracycline-resistance gene (tetR) resulting in a TetR plasmid that can be selected by transformation of human cell DNA preparations into E. coli (Fig 6A). The retrohoming efficiency of the Ll.LtrB variant containing the phage T7 promoter in HEK-293 cells supplemented with 80 mM MgCl2 was ~70% that of the wild-type intron as measured by Taqman qPCR in plasmid targeting assays (Fig 6B). For in vivo selections, the three Ll.LtrB intron expression plasmids were co-transfected with the recipient plasmid into HEK-293 cells, which were then incubated in culture medium with 80 mM MgCl2. After 24 h, plasmids were extracted from the HEK-293 cells by an alkaline-lysis procedure and electroporated into E. coli HMS174(λDE3) to select for TetR colonies, which were screened by colony PCR and sequencing of both 5’- and 3’-integration junctions to confirm retrohoming of the full-length Ll.LtrB intron into the DNA target site (S6 Fig). In controls, no retrohoming was detected by this assay in HEK-293 cells transfected with the same plasmids, but incubated in culture medium without 80 mM MgCl2. This control confirms that retrohoming detected in the assay occurred in human cells and not after transformation of the donor and recipient plasmids into E. coli, and it provides further evidence that addition of MgCl2 to the culture medium is needed to stimulate Ll.LtrB retrohoming in human cells.
We used the HEK-293 cell plasmid selection system to perform eight rounds of in vivo directed evolution in culture medium supplemented with 80 mM MgCl2 via an adaptive walk in which introns that retrohomed into the plasmid target site in each round were amplified by PCR at a relatively high mutagenesis frequency of 3 mutations per intron per round prior to re-cloning into the expression vector for the next round (Fig 6C). After eight rounds, we increased the stringency of the selection by reducing the MgCl2 concentration to 40 mM and performed four additional selection cycles without the addition of new mutations between cycles (rounds 9–12). The retrohoming efficiency of the selected pools relative to that of the wild-type intron assayed in parallel increased slowly from rounds 6 to 9 and somewhat more rapidly during rounds 10 to 12. After the 12 rounds of selection shown in the Figure, an additional three rounds of selection with and without mutagenesis gave no further improvement in retrohoming efficiency of the pools relative to the wild-type intron at 40 mM MgCl2. As described below, high-throughput sequencing indicated that this plateau in retrohoming efficiency reflected that a small number of mutations that moderately enhance retrohoming had overtaken the pool at round 12 and could not be substantially improved by other mutations that were positively selected at either 40 or 80 mM Mg2+.
Although the mutant pools were not increasing in activity at a rapid pace, the possibility remained that individual mutations or combinations of mutations in the pool had enhanced retrohoming. To investigate the mutational diversity of the evolution cycles, we used Pacific Biosciences single-molecule sequencing (PacBio RS), which provides long read lengths (1,000–15,000 nt), combined with circular consensus sequencing (CCS), which compensates for sequencing errors by using rolling-circle amplification to generate concatameric-sequencing reads of the same molecule [55]. An advantage of PacBio RS is that it reads single-molecules directly and thus alleviates problems stemming from formation of molecular hybrids during PCR, which can over-estimate the number of unique sequences in molecular diversity experiments [56,57]. We further avoided formation of PCR hybrids by preparing the sequencing libraries directly from TetR-positive recipient plasmids that contained integrated introns without PCR.
We first sequenced retrohomed introns from round 8 (NCBI SRA database, accession number SAMN03342363) and generated a fitness map that displays the degree of conservation of each nucleotide as a heat map on a secondary structure diagram of the Ll.LtrB intron (Fig 7). The degree of conservation of different nucleotides displayed a wide range and is shown with a scale ranging from dark to light blue for conserved sites (0–0.3% mutations) and from pink to red for mutable sites (>0.3–51% mutations) (Fig 7). On average, the round 8 mutant pool contained 4.4 mutations per intron. The majority of nucleotides (551 of 776) in the intron were conserved (dark or light blue) over eight cycles of directed evolution. Regions required for ribozyme activity (e.g., the catalytic triad in DV, J2/3, which interacts with DV to form the active site, the branch-point A residue in DVI, and the 5’ and 3’ ends of the intron) were invariant, with the exception of a few nucleotides previously shown to be less constrained within those regions (e.g., the dinucleotide bulge in DV). The most variable regions were DIVb, which lies outside the catalytic core, and the two terminal loops of DII. DIVa, which contains a high-affinity LtrA-binding site, showed strong conservation of most nucleotides found to be critical for LtrA binding (positions 557, 559, 561–564), but not position 556 [58,59]. A mutation at position 548 in an internal loop in DIVa was positively selected (green triangle) and could affect LtrA binding.
Although many of the nucleotide changes after 8 cycles of selection appear to be neutral, as they do not bias towards any specific nucleotide, mutations at 25 sites were positively selected (nucleotides within green triangles in Fig 7), meaning that >2% of the population had a mutation at that position of which >80% had the indicated base. Two of the positively selected mutations were within sequence elements involved in long-range tertiary interactions within the catalytic core (ζ and θ’), while six of the positively selected mutations disrupted or weakened base-pairing interactions. Mutations at two sites became highly prevalent in the population (>27%). The first was a G282A mutation in EBS1, which changes a UG to a UA base pair at position -4 of the EBS1/IBS1 interaction between the intron and 5’ exon and had been shown previously to result in an ~50% increase in the efficiency of reverse splicing into a DNA target site in vitro [60]. The second was intron position 642, which was mutated in 51% of the population at round 8 and 99% at round 12 (black arrow). At round 8, 63% of the mutations at position 642 were U to A and the other 37% were U to C. Position 642 is located two nucleotides upstream of the transcription start site of the T7 promoter inserted for selection purposes within DIVb. Although mutations at this position could in principle simply attenuate the T7 promoter [61], leading to less T7-induced toxicity in our E. coli assay, experiments below show that the selected mutations increase retrohoming efficiency in human cells in Taqman qPCR assays. The T7 promoter "TATA-box" region has been shown to interact with TFIID and Pol II in HeLa cell extracts [62,63], and mutations in this canonical “TATA box” could potentially decrease TFIID- and Pol II-binding, leading to increased production of full-length intron transcripts, or could affect retrohoming by some other mechanism. Finally, while the distal stem of DV was mutable, as previously shown in E. coli selections [36], it was not the site of mutations undergoing positive selection for retrohoming in human cells. This finding is in agreement with the results of S4 and S5 Figs, which show that mutations in the distal stem of DV that increased retrohoming efficiency in E. coli or X. laevis oocyte nuclei, did not increase retrohoming frequency in human cells.
To determine whether the mutations that were positively selected in HEK-293 cells at 80 mM Mg2+ (rounds 1–8) were enriched further after more stringent selection without mutagenesis at 40 mM Mg2+ (rounds 9–12 (Fig 6C), we sequenced retrohoming products from round 12 (NCBI SRA database, accession number SAMN03342364). In Fig 7, positions at which the mutation frequency increased or decreased by >2-fold from cycle 8 at 80 mM Mg2+ to cycle 12 at 40 mM Mg2+ are indicted by large green or red arrows, respectively. Surprisingly, over half (9 of 16) of the positively selected nucleotides that comprised >5% of the population in cycle 8 decreased to less than 0.3% of the population in cycle 12 (red arrows). Conversely, six of the eight positively selected mutations that comprised >4% of the population in cycle 12 (green arrows with indicated nucleotide) were not prevalent in the population at cycle 8 (<2%). Four of the eight mutations that were positively selected in round 12 weakened or disrupted base pairs in the intron secondary structure. Two positions that were under positive selection at both 80 and 40 mM Mg2+, the EBS1 mutation G282A and the DIVb mutations U642C and U642A, were present in 64 and 99% of the population, respectively, in cycle 12.
Finally, we identified the top sequencing reads present at highest frequency in cycles 8 and 12 (S1 Table). Many of these contained similar mutations that are candidates for increasing retrohoming activity in human cells. Combinations of these prevalent mutations were tested for linkage disequilibrium (S2 Table) to assess covariation between mutations. The majority of mutation pairs had D' values close to 0, indicating equilibrium, but three mutations in DIVb (U642A, G651A, and U652C) compared in pairwise combinations had D' values between 1 and 2.3, suggesting strong covariation. A number of Ll.LtrB variants that were most prevalent in the population and/or contained positively selected nucleotides were assayed for retrohoming in HEK-293 cells with 80 mM MgCl2 added to the culture medium. Ll.LtrB variants having only the mutations G282A (EBS1) or any of the DIVb mutations (U642A, G651A, U652C) had retrohoming efficiencies similar to or no greater than 50% better than wild type (Fig 8). However, the combinations of G282A (EBS1) and either U642C or U642A-G651A-U652C in DIVb had two- to three-fold higher retrohoming frequencies than the wild-type intron (Fig 8). These findings confirm that selections yielded beneficial mutations that increase retrohoming efficiency with added Mg2+ in human cells. However, all of the beneficial mutations identified lie outside the group II intron catalytic core, the most critical positions of which were invariant in the human cell selections.
While the PacBio deep sequencing identified some combinations of Ll.LtrB mutations that increase retrohoming frequency in human cells, separately testing every conceivable combination of mutations is an inefficient means of identifying the best variants for human cells. Instead, we turned to synthetic shuffling [64] of high frequency mutations identified from the fitness maps to screen many mutation combinations at once. Based on the sequencing of variants from rounds 8 and 12 (Fig 7), we generated a rationally designed synthetic shuffling mutagenesis library by assembly PCR [65]. The library was constructed to test combinations of mutations that showed positive selection and high penetrance during the initial directed evolution (>80% one nucleotide type present in >5% of the population; subsets of the nucleotides indicated by green triangles or green or black arrows in Fig 7). The library consisted of Ll.LtrB introns in which eighteen such positively selected nucleotides were doped at a 1:1 ratio of the selected to the wild-type nucleotide and position 642 in DIVb was randomized. The library was selected for four cycles of retrohoming in HEK-293 cells at either 80 or 40 mM MgCl2 and tested for retrohoming efficiency compared to the wild-type intron at both Mg2+-concentrations after each cycle. Both selections gave pools of Ll.LtrB variants with increased activity relative to the wild-type intron (S7 Fig), and we then performed PacBio sequencing of the fourth cycle pool for each of the selections (NCBI SRA database, accession numbers SAMN03342365 and SAMN03342366). The sequencing showed that specific mutations were selected at a number of positions, but these positively selected mutations differed for the selections done at the two different Mg2+-concentrations (Fig 9A).
To identify those variations associated with the highest retrohoming activity, we generated separate sequence logos for variants that appeared at least three times in the deep sequencing (Fig 9B). While the positions that were shifting towards the mutant nucleotide were shared between the total sequence reads versus just the highest prevalence sequence reads, the shifts were more pronounced in the latter. Both the EBS1 position 289 and DIVb position 642 mutations were present in 100% of the highest frequency variants. We assayed a number of these high prevalence variants for retrohoming in HEK-293 cells (Fig 9C–9E). All of the variants had 3–4 fold higher frequencies for retrohoming into the plasmid target site than did the wild-type intron. When we tested the best of these variants for retrohoming into the genomic target site, we found that variants 80–4 and 40–1 had about three-fold increased retrohoming frequencies. Although these variants were the best we found, they were only marginally better than the EBS1/DIVb mutation combinations tested in Fig 8. These findings suggest that the additional positively selected mutations outside EBS1 or DIVb contribute small fitness effects that together lead to increased retrohoming frequencies. The small contributions to enhanced retrohoming by these mutations is consistent with their relatively slow accumulation during the selections compared to the driving mutations in EBS1 and DIVb.
Here we show that a mobile group II intron, the L. lactis Ll.LtrB intron, can retrohome into a chromosomal DNA site in human cells. To do so, we developed a mobile group II intron expression system that overcomes barriers to group II intron proliferation in eukaryotic nuclear genomes, including suboptimal codon usage and translational repression of the intron-encoded RT, NMD of group II intron-containing RNAs, and suboptimal Mg2+ concentrations. NMD was overcome by using phage T7 RNAP rather than Pol II to express the group II intron RNA, while suboptimal codon usage and translational repression were overcome by separately expressing a human codon-optimized group II intron RT from a separate Pol II-transcript. The remaining barrier, suboptimal intracellular Mg2+ concentrations in eukaryotic cells, was overcome simply by adding 80 mM MgCl2 to the cell culture medium. Retrohoming in human cells was demonstrated by sensitive Taqman qPCR assays of both the 5’- and 3’-integration junctions for both plasmid and chromosomal DNA target sites and by conventional PCR and sequencing of recipient plasmids containing fully integrated intron with both of the expected integration junctions. The expression system workarounds enabled the Ll.LtrB intron to splice and retrohome into both plasmid and chromosomal target sites in viable human cells at frequencies up to ~0.5% and ~0.1%, respectively. However, in vivo selections and synthetic shuffling of positively selected mutations gave only modest further improvements in retrohoming efficiency that still required added Mg2+ in the cell culture medium. The latter findings suggest that low Mg2+ concentrations constitute an effective natural barrier to group II intron proliferation in human cells that is not readily overcome by selecting group II intron variants and may be a major factor in why mobile group II introns failed to persist as such in eukaryotic nuclear genes.
The finding that Pol II transcripts containing the Ll.LtrB intron are selectively degraded by NMD in human cells (Fig 3) extends previous findings for S. cerevisiae and suggests that this defense mechanism against mobile group II introns is used generally in eukaryotes [11]. The Ll.LtrB-intron contains multiple stop codons in all three reading frames and could be degraded either by the exon-junction complex (EJC)-dependent NMD pathway, if the Ll.LtrB-containing transcript contains cryptic spliceosomal splice sites, or by non-EJC-dependent NMD mechanisms, which are known to operate in mammalian cells [66]. By contrast, a T7 RNAP transcript containing the intron is not subject to NMD and accumulates to the same levels as a parallel control transcript lacking the intron (Fig 4). Although the T7 RNAP-synthesized Ll.LtrB transcript accumulates to levels sufficient to support retrohoming in human cells, it has a 5’-triphosphate and up-regulates interferon-response genes, such as RIG-I and IFIT1, which may lead to its sequestration or degradation [45]. Suppression of these innate immune responses could lead to higher levels of T7 RNAP transcripts and retrohoming in human cells than observed here.
The finding that supplementation of the cell culture medium with 80 mM Mg2+ was by itself sufficient to enable splicing and retrohoming of T7 transcripts containing the Ll.LtrB intron indicates that intracellular Mg2+ concentrations are limiting for these processes in human cells [67]. This finding extends previous work showing that group II intron RNPs microinjected into Xenopus laevis oocyte nuclei and Drosophila and zebrafish embryos could retrohome efficiently into plasmid target sites only when Mg2+ was injected in addition to the group II intron RNPs [10]. In contrast to yeast, where transcripts containing the Ll.LtrB group II intron RNA are spliced but not translated [11,35], we observed no detectable splicing of Ll.LtrB-transcripts in human cells without Mg2+ supplementation, even when intron RNA degradation by NMD was suppressed. The Pylaiella littoralis Pl.LSU/2 group II intron could also splice in yeast but not in a human cell line (HCT116 cells; [12]). Thus, the intracellular environment in human cells under normal growth conditions appears to be less amenable to group II intron splicing than it is in yeast.
Surprisingly, the Mg2+-stimulated splicing of the Ll.LtrB intron in human cells neither required nor was enhanced by the LtrA protein, which is needed for group II intron splicing in bacteria or in vitro [21,49]. This IEP-independent splicing could reflect either self-splicing of the Ll.LtrB intron or that human cellular proteins can replace LtrA to stabilize the active intron RNA structure. An intriguing possibility is that the Ll.LtrB intron can be spliced in human cells by a protein evolutionary related to LtrA, such as a LINE-1 or telomerase RT, or the spliceosomal protein Prp8, which evolved from a group II intron-like RT [32].
Although dispensable for splicing in human cells, the group II intron RT remains essential for retrohoming, where it contributes to DNA target-recognition and is required for target DNA-primed reverse transcription [22,68]. The expressed LtrA protein could in principle bind to the group II intron RNA either before or after splicing, the latter being analogous to the reconstitution of active group II intron RNPs in vitro by binding of purified LtrA to self-spliced intron lariat RNA [21]. The similar retrohoming efficiencies when T7 Pol was expressed with or without an NLS (S1 Fig) indicate that nuclear transcription and splicing of Ll.LtrB RNA to produce functional RNPs is not required for retrohoming and can also occur from transfected plasmids that remain in the cytosol. Free Mg2+ concentrations may be higher in the cytoplasm than the nucleus, where Mg2+ is sequestered by chelation to chromosomal DNA [69], thereby favoring group II intron RNA splicing and RNP assembly in that compartment rather than the nucleus. If so, group II intron RNPs may gain access to chromosomal DNA either passively during mitosis or by using a pre-existing RNP transport system. Both mechanisms have been suggested for LINE-1 and other non-LTR-retrotransposon RNPs, which are assembled in the cytoplasm but must gain access to the nucleus for retrotransposition [70–72].
Unlike retrohoming of the Ll.LtrB intron in bacteria, we found that retrohoming of the Ll.LtrB intron into both genomic and plasmid target sites in human cells yields an excess of 3’- over 5’-integration junctions detected by Taqman qPCR assays (7–49 fold; Figs 5B–5D and S1). This excess of 3’-integration junctions could reflect the integration of 5’-truncated introns similar to human LINE-1 elements, whose retrotransposition frequently results in the integration of 5’-truncated elements due to abortive reverse transcription [52]. For both group II introns and LINEs, a high frequency of 5’ truncations during retrotransposition could reflect a combination of barriers to reverse transcription, such as RNA-binding proteins, RNase cleavage of the intron or LINE RNA during or prior to cDNA synthesis, and the ability to ligate truncated cDNAs to upstream chromosomal DNA by non-homologous end-joining (NHEJ) mechanisms, which are not active in E. coli [73–75]. The excess of 3’-integration junctions for the Ll.LtrB intron could also reflect retrohoming of excised linear intron RNAs, which can carry out only the first step of reverse splicing, resulting in the attachment of the 3’ end of the intron RNA to the 3’ exon; TPRT would then yield a cDNA copy of all or part of the linear intron RNA that is ligated to the 5’ exon by NHEJ but could also potentially remain unattached [73,74]. Linear intron RNAs may be generated either by hydrolytic splicing induced by Mg2+ supplementation in the absence of LtrA protein or by debranching of lariat RNAs, possibly via the same enzyme (Dbr1) that functions in the debranching and turnover of spliceosomal intron lariats [76]. The latter could be yet another eukaryotic defense against the proliferation of mobile group II introns.
The newly developed mobile group II intron expression system enabled us to select directly for Ll.LtrB intron variants that could retrohome more efficiently in human cells. To do so, we used a plasmid-based mobility assay that enabled selection for low frequency retrohoming events via E. coli transformation and combined it with the long reads of the PacBio RS circular consensus sequencing to identify mutations under positive selection in the evolving populations. Selections at 80 and 40 mM Mg2+ showed that the majority of intron nucleotides were conserved and nucleotides that form the intron RNA’s active site were highly conserved or invariant. Variations were found mainly in terminal loops and at a few scattered positions within the intron. Two mutations, one strengthening the EBS1/IBS1 interaction between the intron and 5’ exon, and the other near the T7 promoter sequence inserted in DIVb, saturated the pool but gave only ~2-fold higher retrohoming efficiency, and other positively selected mutations did not confer substantial additional benefit, even in synthetic shuffling experiments to select for optimal combinations of mutations. Further, mutations selected at 80 mM Mg2+ differed from those selected at 40 mM Mg2, and Ll.LtrB intron variants selected for enhanced retrohoming in Mg2+-deficient E. coli [36] or X. laevis oocyte nuclei [54] did not show increased retrohoming frequencies in HEK-293 cells. The latter findings may reflect competing effects of altering Mg2+-binding at different sites on intron RNA folding, so that variants selected at one low Mg2+ concentration are not well suited to function at other low Mg2+ concentrations. Previous studies in which variants of the Azoarcus group I intron ribozyme were selected under different conditions showed that different combinations of mutations confer fitness for different environments [77,78].
It is possible that very rare mutations not sampled in our selections, different selections, selections with another group II intron, or rational redesign of the group II intron catalytic core based on X-ray crystal structures could yield group II intron variants that retrohome at high frequencies in eukaryotic cells. Until such time, our findings for the Ll.LtrB intron suggest that barriers to group II intron retrohoming in human cells are not readily overcome by mutational variation and selection, possibly reflecting that the group II intron catalytic core cannot be modified readily to function efficiently at lower Mg2+ concentrations. The latter could explain why group II introns failed to evolve into a form that could function in eukaryotes without fragmentation into spliceosomal introns and the spliceosome.
Although the Ll.LtrB intron works very efficiently for gene targeting in bacteria [9], its targeting efficiency via retrohoming in human cells is substantially lower than those for current methods using CRISPR/Cas9, zinc-finger nucleases or TALEN-based systems [79]. Additionally, retrohoming of the Ll.LtrB intron in human cells requires the addition of Mg2+ to the culture medium, which stresses the cells. Nevertheless, gene targeting efficiencies for the Ll.LtrB intron of near 0.1% might be sufficient for gene targeting applications and could potentially be increased substantially by stable rather than transient expression of the group II intron expression plasmids and/or by suppression of innate immune responses and lariat debranching enzyme. It also remains possible that other group II introns can be found that function more efficiently in human cells than does Ll.LtrB. Finally, as DNA target site recognition by mobile group II introns is not dependent upon ribozyme activity, the ability of group II intron RNPs to recognize a DNA target site in the human genome at appreciable frequency as found here suggests they could be used analogously to CRISPR/Cas9 nuclease-null mutants to localize group II intron RT fusion proteins or modified group II intron RNAs with different functionalities to desired chromosomal locations [80].
Mobile group II introns are thought to have evolved in bacteria where the intracellular Mg2+ concentrations are higher than in eukaryotes [1,36,81,82]. They are hypothesized to have entered an ancestral pre-eukaryote, likely an archaeon, with eubacterial endosymbionts that gave rise to mitochondria and chloroplasts, invaded the nucleus, proliferated as mobile elements, and then degenerated with group II intron domains evolving into snRNAs that reconstitute to form the catalytic core of the spliceosome [4,34]. Based on their discovery that Pol II transcripts containing the Ll.LtrB group II intron are subject to NMD and translational repression, Belfort and coworkers hypothesized that translational repression resulting from group II intron insertion into protein-coding genes contributed to group II intron loss from eukaryotic nuclear genomes and their evolution into spliceosomal introns [11,35].
Considered in the context of the above hypotheses, our results suggest that the ancestral eukaryote must have had relatively high intracellular Mg2+ concentrations that could support proliferation of group II introns in protein-coding genes by retrohoming and that lowering of intracellular Mg2+ concentration in eukaryotes may have been an evolutionary response to selective pressure to restrict group II intron proliferation. Mammals use an analogous defense mechanism based on iron limitation as part of an innate immune response to bacterial infections [83]. In this scenario, a decrease in intracellular Mg2+ concentrations in ancestral eukaryotes would have strongly inhibited group II intron splicing, thereby increasing selective pressure against retaining group II introns as such in protein-coding genes. The evolution of the nuclear membrane, itself hypothesized to be an evolutionary response to group II intron invasion [3], had the additional advantage of sequestering group II introns into a separate compartment where free Mg2+ concentrations are further decreased by chelation to DNA and chromatin, while enabling the cytosol to maintain higher Mg2+ concentrations for other cellular processes [36,67]. A lower free Mg2+ concentration in the eukaryotic nucleus would confer immunity from group II introns that are sporadically acquired by the integration of organellar DNA fragments into nuclear genomes [84] and could resolve the conundrum of why group II introns did not persist in non-coding regions of eukaryotic genomes, where they are not subject to selective pressures caused by translational repression and NMD [13]. Given the inability of multiple group II introns that had inserted into protein-coding genes in an ancestral eukaryote to be cleanly excised simultaneously or to mutate readily into a form that could splice efficiently at low Mg2+ concentration, the evolutionary response was their degeneration into relatively unstructured spliceosomal introns that maintain conserved splice site and branch-point sequences. Reflecting their evolutionary origin, these conserved sequences are recognized by a common splicing apparatus consisting of snRNAs derived from group II intron domains that can now with the aid of proteins promote splicing in the low Mg2+ environment of the eukaryotic nucleus. More generally, our results suggest that differences in intracellular environment had a profound impact on the evolution of introns and gene expression mechanisms in bacteria and eukarya.
Mammalian cells were grown in culture media supplemented with 10% fetal bovine serum (Gemini Biosystems), penicillin, and streptomycin at 37°C with 5% CO2 unless otherwise stated. HEK-293 (ATCC) and HEK-293 Flp-In cells (Invitrogen; Flp-In 293) were maintained in Dulbecco's Modified Eagle Medium (DMEM; Invitrogen) supplemented with glutaMAX (Invitrogen), and hygromycin B. HeLa cells were maintained in Eagle’s Minimum Essential Medium (EMEM; Invitrogen). COS-7 cells were maintained in DMEM. Antibiotics were added at the following concentrations: ampicillin (100 μg/ml), carbenicillin (150 μg/ml), hygromycin B (50–100 μg/ml), penicillin (1,000 U/ml), streptomycin (1,000 μg/ml), and tetracycline (15 μg/ml). Transfection reagents were: Fugene 6 (Roche), Lipofectamine 2000 (Life Technologies), Polyfect (Qiagen), and polyethylenimine (PEI; 40,000 linear molecular weight; Polysciences Inc).
E. coli HMS174(λDE3) (Novagen) was used for the selection of recipient plasmids after retrohoming of the Ll.LtrB intron into the plasmid target site in human cells. Electrocompetent HMS174(λDE3) were generated as described [10,85] and had a transformation efficiency of >2 x 1010 colony-forming units measured using pUC19 plasmid. E. coli strain DH5α was used for cloning.
Plasmid phLtrA is a derivative of pAAV (Stratagene) that expresses a human codon-optimized LtrA ORF (hLtrA; see below) with a 3X myc tag and SV40-NLS fused to its C-terminus. The hLtrA ORF is cloned behind a CMV promoter and followed by a human growth hormone polyadenylation signal. Plasmid phLtrA1 is an earlier hLtrA expression plasmid in which the human codon-optimized LtrA ORF with an SV40-NLS fused to its C-terminus is cloned behind a CMV promoter in a pIRES vector (Clontech). The LtrA ORF contains a small artificial spliceosomal intron, subsequently found to be unnecessary for hLtrA expression, inserted after the start codon and is followed by an SV40 polyadenylation signal. pLtrA is the same except with the native non-codon optimized LtrA ORF.
Plasmid pLl.LtrB contains an Ll.LtrB-ΔORF intron RNA (Ll.LtrB-ΔD4(B1-B3) [86]) cloned downstream of a T7 promoter in a TOPO2.1 vector (Invitrogen). Variants of this plasmid include pLl.LtrB-GFP in which Ll.LtrB intron and flanking exons interrupts the GFP ORF at position 386; pGFP, which contains a T7-driven GFP ORF with the 35-nt ligated exon sequence that would result from Ll.LtrB intron splicing inserted at position 386; and pLl.LtrB-HPRT and pLl.LtrB-CCR5 in which the wild-type Ll.LtrB-ΔD4(B1-B3) intron is replaced by one that has been retargeted to insert in the mouse hprt gene (position 115; [45]) or human CCR5 gene (position 332; [6]), respectively; pLl.LtrB-T7 is a derivative of Ll.LtrB-ΔD4(B1-B3) that contains a minimal T7 promoter in DIVb (positions 627–646); and pLl.LtrB-stuffer is a derivative that lacks the Ll.LtrB intron and was used for library construction.
Plasmid pCMV-BFP, pCMV-BFP-E1E2, and pCMV-BFP-Ll.LtrB contain the blue fluorescent protein (BFP) ORF without or with the ltrB exons 1 and 2 or the Ll.LtrB-ΔD4(B1-B3) intron flanked by ltrB exons 1 and 2 interrupting the ORF after the start codon cloned in pcDNA5FRT (Invitrogen).
Plasmid pT7-NLS contains the T7 RNA polymerase (T7 RNAP) ORF with an N-terminal SV40-NLS cloned behind a CMV promoter in pAAV vector (Agilent), and pT7 is the same plasmid containing the T7 RNAP ORF without a NLS.
Recipient plasmid pFRT contains a wild-type Ll.LtrB target site (positions -30 to + 15 from the intron-insertion site) inserted into the Flp-In recombinase site of pcDNA5/FRT (Life Technologies). The target site region is identical to that inserted into the HEK-293 Flp-In genome. Recipient plasmid pBRRQ is a derivative of pBRR-Tet [6] and contains a wild-type Ll.LtrB target site (positions -30 to +15 from the intron-insertion site) flanked by sequences with Tm values optimized for qPCR (S1 Table) cloned upstream of a promoter-less tetR gene. Recipient plasmid pBRR-CCR5 is identical to pBBRQ except for containing the CCR5 targetron insertion site (positions -30 to +15 from the intron insertion site). All recipient plasmids carry an ampR marker.
The human codon optimized LtrA sequence was generated from overlapping oligonucledotides by assembly PCR [65]. Oligonucleotides containing hLtrA sequence were synthesized by HHMI/Keck Oligonucleotide Synthesis Facility (Yale) and PCR reactions were carried out by using Vent DNA polymerase (New England Biolabs), high annealing temperatures (58–60°C), and manual hot start–i.e., adding Vent DNA polymerase after sample temperature reached 94°C). PCR products were gel-purified and digested with EcoRI and XbaI or HindIII and XbaI, then cloned into pKSBluescript (Agilent) to form pKS-hLtrA and confirmed by sequencing. The assembled ORF was re-cloned into a pIRES vector (Clontech) to generate phLtrA1.
HEK-293 cells were seeded at equal density into 96-well white plates (Corning), allowed to grow out, and transfected using Fugene 6 (Roche) according to manufacturer’s recommendations. After 48 h in culture, cytotoxicity analysis was carried out using the CellTiter-Glo direct lysis kit (Promega) according to manufacturer instructions. Luciferase activity was measured on a Mithras Multimode Platereader (Berthold). Trypan blue staining was performed by mixing 10 μl of cells with 10 μl of trypan blue solution (0.4%; Invitrogen) and then counting stained and unstained cells on a hemacytometer.
For immunoblotting, cells were collected and boiled in 1x Laemmli gel buffer for 5 min. After pelleting insoluble material by centrifugation in a microfuge for 2 min at top speed, the protein samples prepared from the same number of cells were run in 8% polyacrylamide/0.1% SDS gel, which was then blotted to a nitrocellulose membrane using a Hoefer SemiPhor blotter (Amersham). Anti-LtrA antibody [49] was used at 1:1,000 dilution, and goat anti-rabbit secondary antibody (Pierce) was used at 1:60,000 dilution, both at room temperature. After developing the immunoblot, the membrane was stained with AuroDye to confirm even loading.
For immunofluorescence, cells were washed twice with phosphate buffered saline (PBS) and then fixed in 2% paraformaldehyde for 30 min at room temperature. After three more washes with PBS, cells were permeabilized by incubating in 0.5% Triton X-100 in PBS for 15 min, followed by three washes with PBS containing 0.2% Tween 20 (PBST). Blocking was achieved by incubating the permeabilized cells with 10% normal goat serum and 1% BSA in PBST for 1 h. Primary antibody was pre-incubated with untransfected cell lysate (prepared by sonication) to deplete nonspecific antibodies and then incubated with cells at 1:5,000 dilution in blocking buffer for 1 h at 4°C. After four 5-min washes in PBST containing 0.1 M NaCl, cells were incubated with 1:100 dilution of goat anti-rabbit antibody conjugated with fluorescein in blocking buffer for 1 h, washed with PBST containing 0.1 M NaCl five times for 5 min each time, incubated with 2 μg/ml Hoechst dye for 10 min, and washed twice with PBS. Cells were mounted and observed under a fluorescence microscope (Olympus CKX41).
HEK-293 cells were grown to confluence, washed with PBS, blown off the dishes with ice-cold hypotonic buffer (10 mM HEPES, 10 mM KCl, 1 ml/100 mm dish), and incubated on ice for 15 min. Cells were broken by 15 strokes of a Dounce homogenizer. Nuclei were collected by centrifugation at 800 x g for 5 min at 4°C and then resuspended in the residual buffer in the same tube. After 3 cycles of freezing and thawing, chromosomal DNA was sheared by repeated pipetting, and 5 μl of the solution was used for each reaction. RT assays with Ll.LtrB/E2+10 substrate were carried out as described [47,49] in 10 μl of reaction medium containing 5 μl lysate, 40 nM Ll.LtrB template, 400 nM E2+10 primer, 450 mM NaCl, 5 mM MgCl2, 40 mM Tris-HCl, pH 7.5 plus 10 μCi [α-32P]dTTP (3,000 Ci/mmol; New England Nuclear) and 0.2 mM of each dNTP. The Ll.LtrB/E2+10 substrate consists of Ll.LtrB RNA (an in vitro transcript containing the Ll.LtrB-∆ORF intron and flanking exons) with a 20-mer DNA primer (E2+10) annealed to a position in the 3’ exon that corresponds to that of the cleaved bottom strand normally used as the primer for target DNA-primed reverse transcription of the intron RNA during retrohoming. Reactions were initiated by adding dNTPs and incubated at 30°C for 30 min. Incorporation of [α-32P]dTTP was measured by spotting onto DE81 paper (Whatman) and counting Cherenkov radiation in a scintillation counter (LS6500, Beckman).
UPF1 and scramble siRNAs (Dharmacon) were transfected into ~60% confluent HeLa or HEK-293 cells 24 h prior to transfection of BFP- or GFP-containing plasmids. UPF1 levels were measured in equivalent amount of proteins from crude cell lysates via SDS-PAGE (4–12% polyacrylamide gradient gel) and immunoblotting using a Trans-Blot Turbo system (Bio-Rad) to blot the gel to a nitrocellulose membrane, which was then probed with an anti-UPF1 antibody (ab10510; Abcam). Plasmid and siRNA transfections were carried out using Dharmafect as described [87].
For analysis of transcript levels and splicing via RT-qPCR and RT-PCR, respectively, RNA was purified from transfected cells using the ZR RNA Miniprep Kit (Zymo). 1 μg of each RNA sample was treated with DNase I (Invitrogen) at 37°C for 1 h to remove DNA and then converted to cDNA with a SuperScript III reverse transcriptase kit (Invitrogen) according to manufacturer’s recommendations. RT-PCR was carried out with GC-rich Phusion polymerase mastermix (New England Biolabs) under standard conditions, unless otherwise indicated. RT-qPCR was carried out using Power SYBR Green Master Mix (ABI) on an Applied Biosystems Viia7 system in 96-well format under standard conditions. For the CMV-BFP cassettes, the primers were pAAV MCSfw 5’ TCTTATCTTCCTCCCACAGCTCCT and GFP-L qPCRrev 5’ TCGTCCTTGAAGAAGATGGTG, and for the T7-GFP cassette, the primers were pTOPOsplicinginfw 5’ TGTCTTCTTGACGAGCATTCC and pTOPOsplicinginrev 5’ TAGGTCAGGGTGGTCACGA.
Retrohoming of the Ll.LtrB intron in mammalian cells was assayed by Taqman qPCR using an Applied Biosystems Viia7 system in 384-well format using Taqman probes (Life Technologies). Reactions were performed in technical triplicate in 10-μl volumes for 35 (plasmid) or 40 (genomic) cycles using Taqman PCR universal mastermix (Applied Biosystems) under standard conditions. Standard curves for quantitation used four 10-fold dilutions of either pBRRQ or pFRT plasmid containing an integrated Ll.LtrB intron and had >90% efficiency across the range of concentrations used. Standard curve plasmids were quantified using a Qubit system (Life Technologies). Standard curve dilutions were buffered with 10 ng/μl phage lambda DNA carrier. The primer/probe sets are shown in S3 Table.
HEK-293 Flp-In cells (Invitrogen) contain a FRT recombinase site in a decondensed region of the genome. A single copy of the wild-type Ll.LtrB insertion site (position -30 to +15 from the intron-insertion site) was recombined into the FRT site genomic locus according to manufacturer's recommendations. For retrohoming experiments, HEK-293 Flp-In cells containing the Ll.LtrB target site were seeded in multi-well culture plates (Corning) 24 h prior to transfection to reach a confluency of 60–80% on the day of transfection. Cells were dissociated using Stem Pro Accutase (Invitrogen), and cell counting was performed with a hemocytometer or using the Scepter system (Millipore).
For genomic targeting experiments, the Ll.LtrB intron expression plasmids, pLl.LtrB, pT7-NLS, and phLtrA were transfected at 276 ng each with 2.76 μg branched polyethyleneimine (PEI) (Polysciences, Inc) per well in a 12-well culture plate for 24 h. For plasmid targeting experiments, recipient plasmid pFRT or pBRRQ was included at 276 ng per well in addition to the above three plasmids. After 24 h, the media was removed and replaced with growth medium supplemented with MgCl2 or other Mg2+ salts for an additional 24 h unless otherwise specified. The next day, when the cells were typically 80–90% confluent, non-adherent cells were removed by vigorously rinsing with PBS three times, and adherent cells were collected into a 1.5-ml snap-tube unless otherwise specified. Total DNA was extracted from cell pellets with a Qiagen Blood and Tissue kit with an RNase step or the ZR-genomic miniprep kit (Zymo research) according to manufacturer's recommendations. In plasmid targeting experiments, plasmids were extracted from cells using alkaline lysis with the Wizard SV-miniprep system (Promega) or total DNA using the ZR-genomic miniprep kit (Zymo Research). Experiments typically used three wells that had been independently seeded and transfected in parallel for determination of SEMs. Biological replicates were performed on separate days and reported with SDs.
pLl.LtrB-T7 mutant libraries for each selection cycle were generated by PCR with Mutazyme II (Stratagene) according to the manufacturer’s recommendations for 3 mutations per kb. Approximately 200 ng Ll.LtrB DNA template was mutagenized in a 50-μl PCR with primers 309S 5’- CACATCCATAACGTGCGCC and 308A 5’- TAATTGCTAGCCGGCCGCATTAAAAATGATATG for 30 cycles, and then re-amplified to obtain a higher yield using Phusion polymerase (New England Biolabs). The PCR product was purified from an agarose gel stained with Sybr gold (Invitrogen) under blue-light illumination and then digested overnight with AatII and NheI-HF (New England Biolabs). After purification, 750 ng of the insert was ligated to 1 μg of linearized and dephosphorylated pLl.LtrB-stuffer for 2 h at room temperature in a volume of 400 μl using T4 DNA ligase (4,000 units; New England Biolabs). The ligation mix was purified and concentrated to a volume of 6 μl using a Zymo clean and concentrator column and then electroporated into 100 μl E. coli MegaXDH10B cells (Invitrogen) with total transformants typically reaching >2 x 108. The resulting library was purified by using an Endotoxin-free MiniKit II (Omega Biosciences) and transfected into HEK-293 Flp-In cells for both targeting and selection experiments.
In vivo selections in HEK-293 cells were done using a modification of a previously described E. coli plasmid-based retrohoming assay in which a group II intron with a phage T7 promoter inserted in DIVb integrates into a target site cloned in a recipient plasmid upstream of a promoterless tetR gene, thereby activating that gene [10,36]. HEK-293 cells were transfected with plasmids for the hybrid Pol II/T7 expression system (Fig 1), with pLl.LtrB replaced with pLl.LtrB-T7, which contains a minimal T7 promoter in DIVb, and pBRRQ, which contains an Ll.LtrB target site cloned upstream of a promoter-less tetR gene. After 24 h, plasmids were isolated from transfected cells by alkaline lysis using the Wizard SV plasmid miniprep kit (Promega). An aliquot was diluted and used for Taqman qPCR and the rest was concentrated to 6 μl using a Zymo clean and concentrator column. The concentrated plasmid was electroporated into 100 μl of electrocompetent E. coli HMS174(λDE3) cells, which were then plated onto LB-agar plates containing tetracycline (15 μg/ml) and grown for 2 days. The resulting colonies were pooled, and the TetR plasmids were isolated by alkaline lysis using a Wizard SV miniprep kit (Promega). Ll.LtrB introns that had successfully retrohomed into the TetR-recipient plasmids were PCR amplified by 21 cycles of PCR with or without mutagenesis as described above using primers that flank the integration site (primers 200S and 269A; S3 Table), and the PCR product was isolated from an agarose gel and used to generate a library for the next round of selection.
Assembly PCR was used to generate the synthetically shuffled library [65]. Briefly, multiple 80-120-mer oligonucleotides spanning the length of the intron and containing the randomized or doped positions of interest and complementary overlaps with a Tm of ~55°C were synthesized at the Center for Systems and Synthetic Biology at UT-Austin. For each intron library, the assembly PCR was done with a 500-ng equimolar mix of oligonucleotides for 25 cycles under standard conditions in 50 μl of Phusion PCR mastermix. A 5-μl aliquot was placed in 300 μl of Phusion PCR mix with forward and reverse primers that synthesize the full-length intron and run for an additional 25 cycles. The full-length product was purified by electrophoresis in an agarose gel and used to construct libraries in pLl.LtrB, as described above.
Libraries for Pacific Biosciences RS circular consensus sequencing (CCS) were generated according to manufacture's recommendations for A-tailed inserts, and sequencing was performed at the Johns Hopkins University Medical School deep sequencing and microarray core facility. Inserts for PacBio sequencing were generated directly from pooled TetR-positive plasmids isolated after directed evolution cycles by digesting >50 μg of plasmid DNA with AatII and EcoRI-HF (New England Biolabs) at sites 37-nt upstream and 16-nt downstream of the Ll.LtrB-integration site, respectively, and then purifying the resulting restriction fragment in a 1% agarose gel under blue light using Sybr Gold staining. To assess the sequencing error-rate for the PacBio CCS, we sequenced the wild-type intron and determined the number of substitutions, insertion, and deletion errors. With three rolling-circle sequencing passes of the intron, the substitution error rate was <0.01%. The insertion and deletion (indel) rates were 0.21 and 0.07% respectively, and these occurred predominantly at homopolymeric regions.
Sequence reads were filtered to remove reads that did not reach at least three circular passes. Raw sequence reads in the FastQ file format were aligned to the wild-type Ll.LtrB reference sequence using Mosaik Aligner 1.0 (https://code.google.com/p/mosaik-aligner/) and text files were extracted using the Tablet browser [88]. Insertion gaps were removed using a Perl script, Gapstreeze, available online at (http://www.hiv.lanl.gov/content/sequence/GAPSTREEZE/gap.html), and reads containing deletion-errors were removed. Aligned sequences were then analyzed for nucleotide variation using a Perl script courtesy of Dr. Scott Hunicke-Smith (UT-Austin). All other data analysis, including calculation of nucleotide frequencies and analysis of co-variations was performed using Unix shell scripts, including grep, cut, uniq, sort, and awk.
Standard linkage disequilibrium was calculated as D = (PAB x Pab)-(PAb x PaB), where PAB is the frequency at which the mutations occur together, PAb and PaB are the mutations occurring independently, and Pab the frequency at which neither occurred. The normalized linkage disequilibrium (D') was calculated by dividing positive D values by the theoretical maximum co-occurrence and negative D values by a theoretical minimum co-occurrence based on the observed individual frequencies in the population. The significance of these values was measured with the r2 value (the square of the correlation coefficient) calculated as r2 = D2/PaPbPAPB, and χ2 which is r2 multiplied by the number of sequences analyzed [77].
The Pacific Biosciences sequencing data are available at the NCBI SRA database (Biosample accession numbers: SAMN03342363, SAMN03342364, SAMN03342365 and SAMN03342366). The hLtrA sequence is available from NCBI Genbank (accession number KP851976). The primary data underlying the Figures are available in S1 Data.
|
10.1371/journal.pcbi.1006239 | A local uPAR-plasmin-TGFβ1 positive feedback loop in a qualitative computational model of angiogenic sprouting explains the in vitro effect of fibrinogen variants | In experimental assays of angiogenesis in three-dimensional fibrin matrices, a temporary scaffold formed during wound healing, the type and composition of fibrin impacts the level of sprouting. More sprouts form on high molecular weight (HMW) than on low molecular weight (LMW) fibrin. It is unclear what mechanisms regulate the number and the positions of the vascular-like structures in cell cultures. To address this question, we propose a mechanistic simulation model of endothelial cell migration and fibrin proteolysis by the plasmin system. The model is a hybrid, cell-based and continuum, computational model based on the cellular Potts model and sets of partial-differential equations. Based on the model results, we propose that a positive feedback mechanism between uPAR, plasmin and transforming growth factor β1 (TGFβ1) selects cells in the monolayer for matrix invasion. Invading cells releases TGFβ1 from the extracellular matrix through plasmin-mediated fibrin degradation. The activated TGFβ1 further stimulates fibrin degradation and keeps proteolysis active as the sprout invades the fibrin matrix. The binding capacity for TGFβ1 of LMW is reduced relative to that of HMW. This leads to reduced activation of proteolysis and, consequently, reduced cell ingrowth in LMW fibrin compared to HMW fibrin. Thus our model predicts that endothelial cells in LMW fibrin matrices compared to HMW matrices show reduced sprouting due to a lower bio-availability of TGFβ1.
| Therapies for a range of medical conditions, including cancer, wound healing and diabetic retinopathy can benefit from a better control over the growth of blood vessels. The chemical properties of fibrin, the material that forms scabs in wounds and can also occur in large concentrations in tumors, can regulate the degree of blood vessel growth (angiogenesis). Angiogenesis can be mimicked in cell cultures. These allow us to modulate the chemical properties of fibrin and study the effect on angiogenesis. Fibrin occurs in high molecular weight (HMW) and in low molecular weight (LMW) forms. Interestingly, there is more ingrowth of angiogenic-like structures into HMW than in LMW fibrin, but the mechanisms are poorly understood. To get more insight into these, we constructed a computational model. Using the model, we propose and analyse a hypothetical mechanism for sprouting that could explain the differences in endothelial cell sprouting in LMW and HMW fibrin matrices. Our model suggests that cells digest fibrin, thus creating space for ingrowth. At the same time, digestion frees growth factors bound to fibrin, that activates further secretion of digestive enzymes by the cells. We propose that the resulting positive feedback loop spontaneously selects cells in the endothelial monolayer for ingrowth and helps the blood vessel sprout move deeper into the fibrin. This could be a complementary mechanism to lateral-inhibition by Delta-Notch for the selection of leader cells, also called ‘tip cells’. Our model predicts that endothelial cells in LMW fibrin compared to HMW fibrin show reduced sprouting due to a lower bio-availability of TGFβ1.
| Tissues that are low in oxygen stimulate the outgrowth of side-branches from nearby blood vessels, in a process called neo-angiogenesis. A detailed understanding of angiogenesis is relevant for a range of physiological and pathological processes where obtaining a fine-level control of angiogenesis is of interest. Pathologies such as poor wound healing or diabetic retinopathy will benefit from simulating angiogenesis, whereas inhibition (or tempering) angiogenesis is required in the treatment of tumors. Tissue engineering of large organs will require the growth of a functioning blood vessel system.
In wounds and in some tumors, new blood vessels are formed within fibrin matrices. Fibrin is formed as a provisional scaffold by leakage and subsequent polymerization of fibrinogen within the tissue. To form a new blood vessel endothelial cells (ECs) from nearby blood vessels invade this fibrin matrix. A suitable in vitro model for angiogenesis within fibrin was introduced by Koolwijk et al. [1]. In this model, a monolayer of human microvascular endothelial cells (hMVECs) is seeded on top of a layer of polymerized fibrin. When stimulated with a pro-angiogenic factor, such as vascular endothelial growth factor (VEGF) and/or basic fibroblast growth factor (bFGF), in combination with the inflammatory mediator TNFα (tumor necrosis factor α), endothelial sprouts grow into the fibrin matrix.
This hMVEC-fibrin system is in wide use as an assay to screen for stimulators and inhibitors of angiogenesis. However, in absence of an exact understanding of how known molecular and cellular mechanisms interlock to produce the observed, dynamic angiogenesis-like behavior, it is difficult to go much beyond a ‘trial and error’ approach and use the model system to rationally design new strategies for interfering with angiogenesis. By ‘reconstructing’ a cell culture model in silico, mathematical modeling provides insight into how known mechanisms work together and interlink to produce the observed behavior. This paper introduces a mathematical modeling approach to analyse, (a) what mechanisms regulate the onset of an angiogenic sprout (or ‘ingrowth spot’) in an endothelial cell monolayer, and (b) what mechanisms consolidate the further invasion of an angiogenic sprout. Together, these two observables determine the level of angiogenesis, making them relevant targets for tissue engineering and design of medical therapies.
Fibrinogen occurs in three variants, whose relative abundance in fibrin affects its structure and pro-angiogenic capacity: (1) High molecular weight (HMW, MW 340 kDa, ∼70%, Fig 1A); (2) a low molecular weight form (LMW, MW 305 kDa, ∼26% of total fibrinogen, Fig 1B) that is formed after partial degradation of the carboxy-terminus of the fibrinogen Aα-chain; and (3) an alternative low-molecular weight form (LMW’) that is formed after degradation of both Aα-chains (MW 270 kDa, ∼4% of total fibrinogen) [2]. HMW fibrin has a more open matrix structure, with larger openings between the fibers compared to LMW fibrin (Fig 1A and 1B). LMW fibrin forms more complex networks with denser fibers. Fibrinogen composition is a key determinant of the number of ingrowth spots (Fig 1C–1E and S2 Fig) and the depth of sprouting [2–5]. hMVECs proliferate more and show more angiogenic ingrowth in HMW fibrin than in LMW or an unfractionated fibrin mixture [2].
The relative abundance of the three fibrinogen variants is changed in a number of pathologies. For example, the relative abundance of LMW and LMW’ has been found to be elevated in patients with vascular occlusion [6] and in patients with diabetes mellitus [7], possibly due to enhanced vascular leakage. In cancer patients the fibrinogen levels were elevated, but no changes in the HMW:LMW ratio were found [6]. In post-operative patients [6, 8] as well as after extensive acute myocardial infarction [8] the levels of HMW increased, followed by a delayed increase of LMW-fibrin [8]. In full-term newborns the levels of HMW have been found to be 25% lower than in adults [9].
Altogether, our in vitro evidence suggests that HMW-fibrinogen promotes angiogenesis more than LMW-fibrinogen. In vivo, increased levels of HMW are typically found in post-operative patients and after extensive myocardial infarction. It is unknown whether these changes in HMW:LMW ratios have clinical relevance, e.g., in stimulating angiogenesis (high HMW) in post-operative patients or in the inhibition (high LMW) of angiogenesis in diabetes mellitus patients.
During angiogenic ingrowth, the invading hMVEC proteolytically digest the fibrin matrix, suggesting that the low efficiency of in vitro angiogenesis in LMW fibrin is due to differential regulation of proteolysis. Cell-associated fibrinolysis is mostly performed by the trypsin-like protease plasmin [10–13]. Plasmin is the active conversion product of plasminogen, which is mainly produced by the liver and reaches fibrin scaffolds through the blood stream. Conversion of plasminogen into plasmin occurs by plasminogen activators and is highly regulated. Urokinase-type plasminogen activator (uPA) and tissue-type plasminogen activator (tPA) are secreted by ECs as inactive single-chain proteins. tPA is expressed in quiescent endothelium [14] and is primarily involved in clot dissolution [15], whereas uPA and its cellular receptor (uPAR) are expressed during angiogenesis and control pericellular proteolysis [14, 16]. ECs secrete inactive, single chain pro-uPA that binds to uPA receptors (uPARs) on the membrane of endothelial cells, and is subsequently converted into an active two-chained form. This active membrane-bound uPA-uPAR complex converts plasminogen into plasmin [11]. To balance fibrin degradation, ECs secrete plasminogen inhibitor type 1 (PAI-1) that binds to tPA and uPA for deactivation and the PAI-1-uPA-uPAR complex is internalized [10, 12]. Alongside plasmin, membrane-type 1 metalloproteinase (MT1-MMP) can perform cell-associated fibrinolysis [17], although its role is still poorly understood: the MT1-MMP inhibitor TIMP-1 had only minor effects on sprouting in a 100% fibrin matrix, but was inhibiting when a 90% fibrin-10% collagen matrix was used [18]. Altogether, based on the available evidence we assume that hMVEC-associated fibrinolysis [2] is primarily due to the plasminogen-plasmin degradation system.
To get more insight into a potential role of fibrinogen variants in regulating angiogenesis, here we ask, using mathematical modeling, what differences between HMW and LMW fibrinogen could explain the differences in angiogenic ingrowth that are observed in vitro. LMW fibrin has a reduced number of binding sites for growth factors, including latent-TGFβ1 [19]. TGFβ1 has a strong pro-angiogenic effect in hMVEC cultured on Matrigel [20] and is present in latent form in fibrin matrices. Thus, apart from the structural differences between fibrin variants discussed above, a possible difference between fibrin HMW and LMW matrices is their binding capacity of TGFβ1.
TGFβ1 upregulates PAI-1 and uPAR and is inhibited by TGFβ1 antagonist peptides. TGFβ1 also induces PAI-1 and uPAR expression in hepatic stellate cells [21] and uPA/PAI-1 levels in human tumor tissues [22]. LTBP1 (latent transforming growth factor β binding protein 1) potentially binds the C-terminus of this Aα-chain: LMW fibrinogen has a reduced number of C-termini of the Aα-chain compared to HMW fibrinogen. The level of LTBP1 is dramatically reduced in LMW fibrinogen fraction I-9, compared to commercially available fibrinogen and intact fibrinogen fraction I-2 [19]. LTBP1 sequesters latent-TGFβ1 in the plasma to fibrin, resulting in an inactive TGFβ1 reservoir within the fibrin matrix that can locally be activated and released by plasmin [23–25]. Endothelial cells also secrete TGFβ1 [25]; we here assume that this TGFβ1 fraction can be neglected relative to the high bio-availability of TGFβ1 in the matrix. Thus, the reduced number of LTBP1 binding sites in LMW fibrinogen compared to HMW fibrinogen can result in a lower bio-availability of TGFβ1, thereby reducing angiogenesis.
Based on the experimental data on cell-associated fibrinolysis and TGFβ1 that we have discussed above, we suggest that a local uPAR-plasmin-TGFβ1 positive feedback loop drives angiogenesis (see Fig 2). For simplicity, we assume that all cell-bound uPAR is active, i.e., it is bound to uPA. Cell-bound uPAR activates plasmin (Fig 2, arrow 1) and plasmin locally degrades fibrin and releases and activates TGFβ1 from its latent binding protein (see Fig 2, arrow 2). TGFβ1 stimulates the production/expression of uPAR in the protruding cell (see Fig 2, arrow 3), whereas nearby cells, which experience only low TGFβ1-dependent uPAR stimulation, are silenced by self-secreted PAI-1 (see Fig 2, arrow 4). The basic principle underlying this hypothesis is a reinforced random walk [26], as introduced to the problem of angiogenesis previously [27, 28]: (1) an external growth factor activates endothelial cells to enzymatically modify the ECM near the sprout, and (2) the endothelial cells move randomly, but with preference up gradient of the modified ECM.
More recent models have described specifically the hMVEC-fibrin culture system in silico [29, 30]. In both these previous models, the location of the novel capillary sprouts vascular ingrowths was specified a priori, prohibiting their use for analyzing the ‘degree’ of angiogenesis, usually measured as the number ingrowth spots in a cell culture [1]. Therefore, a detailed understanding and analysis of angiogenesis in the Koolwijk et al. [1] experimental 3D fibrin sprouting model requires two modifications of the assumptions in the previous work. Firstly, it is unpredictable which cells in the monolayer become sprout leaders (‘tip cells’). Thus we cannot pre-assign the location of the tip cells [29, 30], or assign the onset of angiogenesis by punching a hole in the basal lamina [31] or by initiating its local digestion [27, 28]. Also, previous models assumed that endothelial cells follow a gradient of VEGF. The Koolwijk et al. [1]in vitro model does not include growth factor gradients, so we have not included those in the present in silico model. This implies that both the location and the growth direction of sprouts in the present computational model emerge from local cell-cell and cell-matrix interactions. We hypothesize that such sprout initiation mechanisms may exist alongside the established role of the Dll4-Notch network in the selection of tip cells that lead the sprouts [32–35].
Altogether, to explore our hypothesis that the uPAR-plasmin-TGFβ1 positive feedback loop regulates spontaneous ingrowth, we model the plasminogen-plasmin degradation system in combination with a cell-based model of endothelial cell invasion. following previous continuum models [36–38]. We propose that a differential binding activity of TGFβ1 to HMW and LMW explains the higher ingrowth. We developed a computational model to evaluate if this sprouting mechanism can explain the reduced ingrowth in LMW compared to HMW; it is shown that it regulates the spacing of ingrowth spots and is also consistent with a number of additional observations.
To study how endothelial sprouting is reduced in LMW compared to HMW fibrin matrices, we developed a computational model that mimics the in vitro assay by Koolwijk et al. [1] and Weijers et al. [2]. Our hybrid model consists of a cell-based component to describe the endothelial cells and a continuum component to describe the plasminogen-plasmin system. The model represents a cross-section of the in vitro sprouting model (Fig 3), and is initialized with a monolayer of fifty endothelial cells on top of a fibrin matrix. Fibrin forms a physical obstruction for cells while, at the same time, offering support to the cells as they can adhere to fibrin. Using cell-based modeling, we explicitly describe cell shape, cell motility, cell-cell adhesion, and cell-fibrin adhesion. Each cell has a level of active uPAR, which homogeneously spread over the cell membrane, and each cell secretes PAI-1 into the extracellular space. PAI-1, and the other extracellular proteins (fibrin, latent-TGFβ1, active TGFβ1, plasminogen, and plasmin) are modeled as concentration fields. The extracellular proteins interact with one another and with the membrane-bound uPAR (Fig 2). uPAR activates plasminogen, forming plasmin that degrades fibrin and locally activates latent-TGFβ1 by releasing it from the fibrin matrix. The active TGFβ1 induces the production of uPAR in nearby cells. These reactions form a local positive feedback loop that keeps the invasion of the endothelial cells going.
To represent cells and their physical interactions with the fibrin matrix, the cellular Potts model (CPM) [40, 41] was used. For details see Section Fibrin invasion. Briefly, the CPM represents cells on a regular lattice as patches of connected lattice sites. Cells move by copying lattice sites inward or outward, representing the extension and retraction of pseudopodia. To model the physical obstruction imposed by high concentrations of fibrin, the extension probability of a pseudopodium is reduced if it attempts to invade a lattice site with high fibrin concentration. The concentration of fibrin, f ( x → ), is initialized at a uniform, non-dimensional concentration of f ( x → ) = 1 . 0 at all lattice sites x →. No fibrin is produced or added in the simulation, such that the concentration of fibrin will stay in the range f ( x → ) ∈ [ 0 , 1 ]. The invasion probability quickly drops for concentrations f ( x → ) > 0 . 5, while for f ( x → ) < 0 . 3 fibrin does not hinder invasion.
Fibrin is digested by the plasmin system, as illustrated in Fig 2. The concentration of uPAR within each cell is modeled by an ordinary differential equation (ODE). A concentration field for uPAR is projected on the CPM grid, with each lattice site that is occupied by a cell having the uPAR concentration of that cell. The concentration of uPAR moves along with the location of the cell after cell movement. A system of coupled partial differential equations (PDEs, see Section Methods) describes the reactions between fibrin, TGFβ1, plasminogen, plasmin, PAI-1 and all fibrin-bound forms. The equations for the plasmin system were based on the continuum model by Diamond et al. [36], which studies the penetration of uPA and tPA into a fibrin clot in the blood stream. To adopt this model to our problem, we added the uPAR-plasmin-TGFβ1 positive feedback loop, simplified the implementation of fibrinolysis, and deleted the convective terms.
Time steps in our model are measured as Monte Carlo step (MCS). One MCS is defined as the number of lattice site update attempts as there are sites in the lattice. It takes about 6000 MCS to simulate a 10 days long experimental assay, similar to the 3D-fibrin sprouting model of Koolwijk et al. [1]; so a MCS represents approximately 2.5 minutes in real time.
In summary, the model is based on the following mechanistic assumptions:
In the in vitro 3D-fibrin sprouting assay by Koolwijk et al. [1], uroplasmin (uPA) and its receptor uPAR were localized specifically at the invading endothelial cells that lead the sprouts [46]. The selection mechanism of these ‘uPAR-rich’ cells in the monolayer is not fully understood. We therefore asked if the uPAR-plasmin-TGFβ1 positive feedback mechanism is sufficient to confine uPAR-expression to the invasive cells.
We initialized the cells in our model with a uniform concentration of uPAR (Fig 4A). Random cell movements change the contact-level and contact-duration with fibrin, resulting in local, random differences in the levels of plasmin activation. Fibrin is degraded at sites with a high plasmin activity (Fig 4B), and TGFβ1 is released from the matrix (Fig 4C). The active TGFβ1 induces the expression of uPAR in nearby cells (Fig 4D). The expression of uPAR in more distant cells can also be induced by the released TGFβ1 to some degree, but such uPAR activity is counteracted by the self-secreted PAI-1. Due to stochasticity, only a few cells in the monolayer can trigger the positive feedback loop sufficiently to overcome inhibition by PAI-1 and gain high levels of uPAR to start ingrowth (Fig 4E). In absence of fibrin-bound latent-TGFβ1, none of the cells in the monolayer, in a hundred stochastic simulations, manage to gain high levels of uPAR due to the lack of TGFβ1-induced uPAR expression. Thus, our modeling results show that a uPAR-plasmin-TGFβ1 positive feedback loop suffices to select uPAR-rich cells in a monolayer of endothelial cells to form ingrowth spots.
Once uPAR-rich cells are selected spontaneously within the monolayer, the uPAR-plasmin-TGFβ1 positive feedback consolidates sprout progression in the model (see Fig 5). The cell leading the sprout, i.e., the tip cell, has the highest concentration of uPAR (see Fig 3E) in agreement with experimental observations [46].
In agreement with in vitro observations, in the in silico model sprouts branch spontaneously (see, e.g., simulation 3 for a constant uPAR production rate of 0.003 Relative Units (RU)/MCS in Fig 5A). This occurs when a cell adjacent to the tip cell moves into another direction, or when a cell higher up in the sprout manages to trigger the feedback loop and starts a branch. Sprouts are not formed in every simulation: due to the stochastic fluctuations in cell shape and movement, in some cases none of the cells activate the positive feedback loop sufficiently to overcome the inhibition of PAI-1. Similarly, ingrowth is not seen in every in vitro experiment, but is highly variable per cell donor and passage number of the cells within the same donor.
In vitro, TNFα is required to induce sprouting of human endothelial cells [1] and the mean tube length increases at higher doses of TNFα. TNFα increases uPA production and the level of cell-bound uPA [1]. To test if the model correctly reproduced this in vitro observation, we mimicked the effect of TNFα by increasing uPAR expression in the endothelial cells. Fig 5A shows a set of simulation results after ten days of sprouting for a uPAR expression level of 0.001, 0.002, 0.003, and 0.005 (RU/MCS). In simulations with higher uPAR expression levels, the number of ingrowth spots increases. For each parameter setting, four simulation results for the same parameter settings are shown; these demonstrate the stochasticity of ingrowth frequency and sprout morphology.
To quantify sprouting, we defined three measures: the angiogenesis level, the sprouting frequency and the fibrinolysis level. The angiogenesis level simultaneously reflects sprout depth and sprout count (see Section Methods for the quantification algorithm). The blue curve in Fig 5B represents the mean angiogenesis level for all simulations that formed sprouts (angiogenesis level>0). The sprouting frequency is the number of simulations out of a hundred simulations that formed sprouts (red curve in Fig 5B). The fibrinolysis level, defined as the mean percentage of initial fibrin lattice sites that are invaded by the endothelial cells in all 100 simulations, also increases for higher uPAR expression levels, as is expressed by the green curve in Fig 5B.
In summary, an increase of the basal uPAR-bound uPA activity in all cells increases the probability that the uPAR-plasmin-TGFβ1 positive feedback loop is triggered in one of the cells in the monolayer, leading to high uPAR expression and sprout initiation. As a consequence, sprouts form more frequently and more excessively at higher uPAR expression levels. Thus, the model explains mechanically how ubiquitous stimulation of uPAR-bound uPA activity by TNFα leads to confined uPA activity and sprouting.
A full quantitative validation of the model is not feasible at present, because only for a few parameters experimental estimates are available, leaving most other parameters as fitting parameters. To avoid overfitting, we have instead selected a set of ‘default’ parameter values for which the model qualitatively reproduces the fibrin culture system (see S2 Table). To validate the model, we then tested if qualitative shifts in the parameter values, corresponding with published experiments, qualitatively reproduce the outcome of three published in vitro experiments of the plasminogen-plasmin degradation system.
Firstly, Koolwijk et al. [1] reported that there was no angiogenic ingrowth and tubule formation in fibrin matrices that were made using plasminogen-depleted fibrinogen. In agreement with this observation, there is no ingrowth in our model for low initial level of fibrin-bound plasminogen (Fig 6A). The sprouting percentage, the fibrinolysis percentage, and the angiogenesis level all increased with the initial fibrin-bound plasminogen concentration.
Secondly, inhibition of uPAR-bound uPA activity by addition of uPA specific polyclonal antibodies, or prevention of the binding of uPA to uPAR by soluble uPAR or blocking antibodies inhibited capillary-like tube formation dose-dependently (see Refs. [1, 46] and Fig 6D (bt + trasylol and bt + H2)). We mimicked the inhibition of uPAR activity by increasing the decay rate of uPAR. Consistent with the experimental results, Fig 6B shows that this parameter change results in a decrease of the sprouting percentage, the fibrinolysis percentage, and the angiogenesis level.
Thirdly, experiments show that there is an optimum PAI-1 concentration for angiogenesis [47]: addition of PAI-1 to implants in wild-type mice enhanced angiogenesis up to 3-fold at low concentrations but inhibited angiogenesis nearly completely at high concentrations. In the 3D fibrin assay, addition of the anti-PAI-1 antibody MAI-2 shows a similar biphasic effect on angiogenesis (Fig 6D): Moderate inhibition enhances tube formation, whereas strong inhibition reduces tube formation. This is due to excessive fibrinolysis, which is incompatible with normal capillary formation [48, 49]. As for uPAR, we modeled the manipulation of PAI-1 activity by an increase of the decay rate of PAI-1. Fig 6C shows that the fibrinolysis percentage strongly increases when the decay rate of PAI-1 is increased. High decay rate of PAI results in low PAI-1 activity, and thus in excessive fibrinolysis; no sprouts are formed, but the entire monolayer lowers simultaneously. Low decay rates of PAI-1 result in high PAI-1 activity and sprouting is completely inhibited. Only for intermediate levels of PAI-1 activity we find sprouting, indicated by the peaks in Fig 6C for the sprouting percentage and the angiogenesis level.
In conclusion, the model can reproduce three essential validation experiments for the plasminogen-plasmin system. In absence of fibrin-bound latent-TGFβ1, no sprouts are formed in all 100 simulations with a parameter set for which sprouts formed well in presence of fibrin-bound latent-TGFβ1 in Figs 5B and 6 (constant uPAR production rate = 0.005 RU/MCS, initial fibrin-bound plasminogen concentration = 1 RU, PAI-1 decay rate = 0.01 MCS−1, and uPAR decay rate = 0.0095 MCS−1, using Relative Units, RU, and Monte Carlo Steps, MCS). This shows that initialization and consolidation of sprouts can be driven by activity of the proposed positive feedback loop formed by uPAR, plasmin, and TGFβ1.
Next we used our model to design new hypotheses about the mechanisms that reduce the level of angiogenic ingrowth in LMW fibrin matrices compared to HMW matrices. The level of LTBP1 is dramatically reduced in LMW fibrinogen fraction I-9, which lacks major parts of the C-termini of the Aα-chain, compared to commercially available fibrinogen and intact fibrinogen fraction I-2 [19]. As LTBP1 sequesters latent-TGFβ1 to fibrin, this could result in a lower level of fibrin-bound latent-TGFβ1. We hypothesize that this reduced level of fibrin-bound latent-TGFβ1, in combination with our suggested local uPAR-plasmin-TGFβ1 positive feedback, could cause the reduced level of endothelial sprouting in LMW compared to HMW fibrin matrices. If the levels of inactive TGFβ1 in the fibrin matrix are too low, cells are not able to induce a strong enough uPAR-plasmin-TGFβ1 positive feedback loop to overcome the inhibition of PAI-1 and thus will not form sprouts.
In line with this hypothesis, Fig 7A shows that the sprouting percentage, the fibrinolysis percentage, and the angiogenesis level decrease with lower initial concentrations of fibrin-bound latent-TGFβ1 in our model. In conclusion, our simulations results suggest that the angiogenic ingrowth is reduced in LMW fibrin matrices compared to HMW matrices due to a reduction in binding sites for LTBP1.
The addition of active TGFβ1 has a biphasic effect on in vitro sprouting [50]. Addition of active TGFβ1 to the assay stimulates sprouting at low doses and inhibits sprouting at high doses of TGFβ1. To test this biphasic effect in the model, we initialized the model with a homogeneously spread concentration of active TGFβ1. The medium containing TGFβ1 was refreshed every two days in vitro [50]. We similarly reset the TGFβ1 concentration to the initial value after every two days in the model. Fig 7B shows that TGFβ1 indeed has the reported biphasic effect on angiogenesis in the simulations. At low concentrations of added TGFβ1 (TGFβ1 = 0.5 RU and TGFβ1 = 10 RU in Fig 7B), more sprouts are formed than without addition of TGFβ1 (TGFβ1 = 0 in Fig 7B). The uPAR-bound uPA activity in all cells increases due to the overall addition of TGFβ1, allowing some cells to overcome the inhibitory PAI-1 threshold for triggering the uPAR-plasmin-TGFβ1 positive feedback loop. This is a similar effect as was seen for the stimulation with TNFα above. The upregulation of uPAR-bound uPA activity is too strong at high doses of TGFβ1, and consequently all cells degrade the matrix. This results in lowering of the complete endothelial cell monolayer, rather than in local sprouting (TGFβ1 = 1000 RU in Fig 7B). In this case, fibrin is quickly degraded and some cells loose contact with the fibrin. Once the cells loose contact with the fibrin layer, they are no longer stimulated to migrate along with the degrading matrix and form the ‘fingers’ show in Fig 7B. In some simulations stacks of cells hovering above the monolayer were left behind. This is of course a model artifact, so we did not take those into account while quantifying the degree of spouting.
We have developed a computational model to study what mechanisms cause angiogenic ingrowth and subsequent sprouting in an in vitro model [1, 2] of angiogenic-like tubule formation of endothelial cells in 3D-fibrin matrices. For this purpose, we asked what mechanisms cause a reduced level of endothelial sprouting in low molecular weight (LMW) compared to high molecular weight (HMW) fibrin matrices [2]. We propose that a uPAR-plasmin-TGFβ1 positive feedback loop selects ‘uPAR-rich’ cells and drives the further invasion of the sprout.
The mathematical model makes a number of plausible, mechanistic predictions on Koolwijk’s angiogenesis assay. Firstly, the model correctly predicts a reduced level of angiogenesis in LMW compared to HMW fibrin, suggesting that in LMW matrices the uPAR-plasmin-TGFβ1 positive feedback loop is not activated. This could be due to LMW fibrin’s reduced binding capacity for TGFβ1 [19]. Secondly, the model predicts that the uPAR-plasmin-TGFβ1 positive feedback loop is responsible for the spontaneous selection of uPAR-expressing cells in the monolayer. Random cell movement activates the feedback loop more strongly in some cells than in others, resulting in random selection of sprout leader cells (tip cells) in the monolayer, and a large variability in the number of sprouts that are formed. Similarly, there is a large variation in the success of sprouting in vitro. Thirdly, and in line with this prediction, the base expression level of uPAR regulates the density of ‘uPAR-rich’ cells and sprouts. A possible candidate for the induction of uPAR-bound uPA activity is TNFα [1]. Addition of TNFα is required in the in vitro experiment with human endothelial cells to induce sprouting. Thus, our simulations provide an explanation for how upregulation of uPAR-bound uPA activity by TNFα can induce sprouting.
Endothelial cells in the in vitro assay of Koolwijk et al. [1] secrete PAI-1, but it is unknown if all cells, or only the quiescent cells in the monolayer or perhaps only the invading uPAR-rich cells secrete PAI-1. Interestingly, the uPAR-plasmin-TGFβ1 positive feedback loop resembles reaction-diffusion systems with activator-inhibitor dynamics [51–53]. Activator-inhibitor systems produce periodic patters, so such dynamics could be responsible for regular placement of tip cells in the in vitro assay. Most conditions for activator-inhibitor dynamics are met: The positive feedback loops stimulates local activation of uPAR-bound uPA, and the inhibitor PAI-1 diffuses faster than the “activator” uPAR, which is expressed intracellularly. A missing element for such activator-inhibitor dynamics is that the inhibitor (PAI-1) must be produced locally, whereas we currently assumed that all cells secrete PAI-1. We are unsure of this assumption: TGFβ1 induces production of uPAR as well as PAI-1 in MVEC cultured on Matrigel [20], raising the possibility that uPAR-rich cells secrete most PAI-1 and that all conditions for activator-inhibitor dynamics are met. Thus future work should determine the localization of PAI-1 secretion in the 3D-fibrin sprouting assay [1].
Besides the activator-inhibitor dynamics, the closely related substrate-depletion model [52] is a well-studied theoretical model for pattern formation. In our model, plasminogen is the substrate for plasmin production. Conversion of plasminogen at sites of matrix invasion results in depletion of plasminogen in surrounding regions through diffusion. Indeed, plasminogen is a limiting factor for endothelial sprouting in the fibrin assay [1]. Plasminogen depletion has low impact in the current simulations, because we have initialized them with a high, homogeneous concentration of immobile, fibrin-bound plasminogen. However, plasminogen binds fibrin reversibly and can bind to ECs, so this mechanism might regulate the location of ingrowth spots for lower levels of fibrin-bound plasminogen. Interestingly, there is a delay in sprout initiation when the model is initialized with unbound plasminogen. It takes some time to reach high enough concentrations of fibrin-bound plasminogen, which is then converted to plasmin by uPAR for matrix degradation.
A key patterning mechanism that is involved in angiogenesis is lateral inhibition by Delta-Notch signaling [32–35]. Cells that have high levels of Delta ligands on their membrane differentiate into so called ‘tip cells’, which are the leaders of sprouts, and cells with low levels of Delta become ‘stalk cells’ [35]. Lateral inhibition occurs by interaction of Delta ligands with the Notch receptor of neighboring cells, resulting in the suppression of Delta production in those neighbors [32–35]. Lateral inhibition creates a pepper-and-salt pattern of tip and stalk cells, with tip cells surrounded by a rosette of stalk cells in monolayers in silico [54, 55]. Thus, Delta-Notch signaling alone cannot account for the more widely spaced pattern of uPAR-rich leader cells in a monolayer as observed in vitro [46]. Possibly other regulation mechanisms, e.g., the proposed uPAR-plasmin-TGFβ1 positive feedback loop, act alongside the Delta-Notch mechanism to distribute tip cells more sparsely. Notably, gene expression levels of Dll4 and Notch4 are significantly higher in endothelial cells cultured in LMW matrices than in HMW matrices [2]. The Dll4 and Notch4 expression differences by themselves cannot explain the reduced ingrowth in LMW fibrin matrices, as specific inhibition of Dll4-Notch was unable to induce recovery of tube formation in LMW.
Inclusion of Delta-Notch signaling will likely affect sprout morphology. In simulations of our current model, cells adjacent to the tip cell are also activated by the released TGFβ1, and they contribute to sprouting. This results in fairly wide, sometimes cyst-like sprouts. In our previous model [29] of the fibrin assay, narrow sprouts formed if only the tip cell secreted proteolytic enzymes for matrix degradation, and cyst-like sprouts formed when the stalk cells contributed to fibrin degradation as well. In this light, Delta-Notch signaling could repress proteolytic activity in cells adjacent to the tip cell, such that thinner sprouts will form.
Our model explains differences in ingrowth between LMW and HMW fibrin based on the binding capacity of latent-TGFβ1. An alternative explanation for the increased ingrowth in HMW fibrin compared to LMW fibrin could be that the ECs can invade the open matrix structure of HMW fibrin more easily. In absence of proteolysis, differences in matrix porosity can explain cell migration speed and persistence [56]; however, with small pore sizes of fibrin (order 1 μm; see Fig 1) and the importance of fibrinolysis for angiogenic ingrowth, small differences in pore size are unlikely to contribute to differences in ingrowth. An alternative, or complementary explanation could lie in differences in the bulk mechanical properties of HMW and LMW fibrin. Indeed mechanical cell-cell communication [57, 58] through strain-stiffening materials such as fibrin [59] suffices for generating vascular-like patterns [60]. In addition, individual fiber architecture, including fiber thickness and fiber density also affects cell spreading behavior on fibrin substrates independently of the bulk mechanical properties [61], suggesting that fiber architecture differences (Fig 1) could also contribute to differences in angiogenesis level on HMW and LMW fibrin matrices.
The present model reproduces angiogenic sprouting by means of cell-fibrin adhesion and cell-division. A limitation is that the addition of TNFα in the in vitro model inhibits cell division [1]. The general cell invasion mechanism proposed here does not depend on cell division. Alongside the fibrinolysis-driven sprouting mechanisms proposed here, many alternative mechanisms of cell migration during angiogenic sprouting have been proposed that could act alongside or instead of cell division to replenish cells in growing sprouts. A range of models have shown that mutual attraction of endothelial cells suffices for the formation of vascular networks, e.g., via a chemoattractant [62–70], via mechanical forces [71, 72] or via mechanically induced durotaxis [60], and preferential attraction to elongated structures [73]. Our model could be extended with such sprouting and cell migration mechanisms to replace cell division.
A detailed description of the plasminogen-plasmin system is included in our model, but still some simplifications were made. For instance, we did not take into account interactions with matrix metalloproteinases (MMPs). Membrane-type 1 metalloproteinase (MT1-MMP) can perform cell-associated fibrinolysis [17], but only plays a minor role in Koolwijk’s assay [18]. Furthermore, we neglected the low proteolytic activity of pro-uPA [11], and only modeled active uPAR-bound uPA. Interactions between pro-uPA and plasmin could give interesting dynamics. Venkatraman et al. [37] considered a positive feedback loop in which the initial cleavage of plasminogen into plasmin is more efficient by uPA than pro-uPA, and the conversion of pro-uPA to uPA is driven by plasmin. By the use of a continuum model, they predict that uPA-plasmin activation is bistable in the presence of this positive feedback loop in combination with substrate competition for plasmin.
A further limitation of the present model of the plasmin system, is that the numerical method cannot describe the advection of chemical species due to displacement of fibrin. This approximation is reasonable in the low Péclet number regime simulated here; i.e., cell movement (and the resulting advection of chemicals due to movement of fibrin) is much slower than the movement of chemicals relative to the ECM due to diffusion and fibrin degradation. Because cells cannot ‘push’ fibrin, but only grow over it if fibrin is sufficiently degraded, the low Péclet number regime is ensured for fibrin and all fibrin bound growth factors. Also, cell movement is slower than the diffusive spread of the unbound growth factors, further justifying our approximation. A suitable method for modeling advective transport in the CPM due to cell movement for higher Péclet number cases has been proposed elsewhere [74], and can be applied in future extensions of our model.
It could be argued that the present two-dimensional approximation in silico does not represent Koolwijk’s three-dimensional cell culture model well, because in two-dimensional cell cultures the cellular micro-environment is usually not well represented [75]. However, note that in two-dimensional cross-section the cellular micro-environment of the endothelial cells corresponds with those in the three-dimensional cell culture. The leading cell is flanked by other endothelial cells and by the fibrin matrix (see, e.g., the uPAR-rich cell in Fig 4D), whereas the following endothelial cells are flanked by fibrin, culture fluid and cells (see, e.g., Fig 5A). Thus the two-dimensional cross-section in silico suffices as an approximation of the three-dimensional model in vitro. Nevertheless, the model will run in 3D with some adjustments, through appropriate scaling of the cell volume constraint and the adhesion parameters [76]. The positive feedback loop hypothesis, and the mechanisms involved, will both work in the same qualitative way in 3D as in 2D, since the reaction-diffusion equations have the same form.
In conclusion, our model predicts that the reduced level of endothelial sprouting in LMW compared to HMW fibrin matrices can, at least in part, be explained by a reduced level of fibrin-bound latent-TGFβ1 in LMW fibrin. To validate this hypothesis experimentally, we propose to check if there is indeed a reduced level of fibrin-bound latent-TGFβ1 in the experimental setup [1, 2]. As a second experiment, we propose to validate whether sprouting can be reduced in HMW fibrin matrices by addition of TGFβ1-antagonists. These validation experiments can bring us closer to an understanding of the mechanisms of selection of leader or ‘tip cells’ in the monolayer and sprouting in the in vitro setup.
We developed a hybrid, cell-based and continuum, computational model of angiogenic sprouting to represent the in vitro 3D-fibrin sprouting assay of Koolwijk et al. [1] (Fig 3). The model includes a uPAR-plasmin-TGFβ1 positive feedback loop that drives sprouting and is used to explain the reduced ingrowth on LMW compared to HMW. Cells and their physical interaction with fibrin are modeled with the cellular Potts model. The CPM is coupled to concentration fields to model the uPAR-plasmin-TGFβ1 positive feedback loop. Each cell has a concentration of uPAR, homogeneously spread on its membrane, modeled by an ordinary differential equation (ODE). A system of partial differential equations (PDEs) describes the interactions between fibrin, plasminogen, plasmin, PAI-1 and TGFβ1.
The shape and motility of endothelial cells are modeled with the cellular Potts model (CPM) [40, 41]. The model domain is a two-dimensional regular lattice Λ ⊂ Z 2, with x → ∈ Λ the coordinates of the lattice sites. Cells and extracellular materials are projected onto the grid as patches of (usually connected) lattice sites, marked with the same unique identifier σ ( x → ). Thus a generalized cell (e.g., a cell or ECM material) s is defined as the set of lattice sites marked with the same identifier σ ( x → ), C ( s ) = { x → ∈ Λ | σ ( x → ) = s }. Each identifier is further associated with a type τ(σ). Here τ(σ) ∈ {cell, fibrin, cell, patch, border, medium}; its function is simply to define parameters and properties for categories of Potts domains, not for all domains individually.
Cells move by extending or retracting pseudopodia, which include lamellipodia, filopodia and invadopodia. Pseudopodia movement is modeled by attempting to copy the state (σ ( x → )) of a randomly selected lattice site x → into a lattice site x → ′ selected at random from the eight, first- and second-order neighbors. We then calculate the change, ΔH, of the Hamiltonian H = Hcontact + Hsize, which defines the force resulting from cell behaviors and properties in the model. An additional energy H0 is added to ΔH at the time of copying to represent dissipative energies (or other copy biases), including those associated with physical obstruction by the fibrin matrix. The components of the Hamiltonian and H0 are described in more detail below.
As in Hamiltonian systems F → ∝ ∇ → H, any copy attempt for which ΔH + H0 < 0 represents a passive force (e.g., due to adhesion or pressure differences) that is sufficiently large to overcome the local, dissipative energies. These copy attempts are always accepted. In addition, cells exert active forces on their environment due to random membrane fluctuations; we assume these fluctuations are distributed according to the Boltzmann probability function,
P Boltzmann ( Δ H , H 0 ) = e - Δ H + H 0 μ , (1)
with μ, the rate of active random membrane fluctuations (a.k.a. cellular temperature).
The model includes a number of “static” cells, τ(σ) ∈ {cell patch, border}. Any copy attempt from or to the static states is ignored (i.e., updates are applied only if τ ( σ ( x → ) ) ∈ { medium , cell } ∧ τ ( σ ( x ′ → ) ) ∈ { medium , cell }). Copy attempts from fibrin sites (τ ( σ ( x → ) ) = fibrin) are also ignored; copy attempts into fibrin (τ ( σ ( x → ′ ) ) = fibrin) are a special case (see Section Fibrin invasion).
The plasminogen-plasmin system in this model is based on the continuum model by Diamond et al. [36]. We made some changes to make it suitable for our system and, most importantly, we included the uPAR-plasmin-TGFβ1 positive feedback, simplified the implementation of fibrinolysis, and removed convective terms. Fig 8 shows an overview of the binding and conversion reactions of plasminogen and latent-TGFβ1 in relation to fibrin that are included in our model. In this section we will discuss the reactions in Fig 8 to explain the PDE system that describes the plasminogen-plasmin system and the uPAR-plasmin-TGFβ1 positive feedback loop.
|
10.1371/journal.pbio.1001451 | The Mechanism of Toxicity in HET-S/HET-s Prion Incompatibility | The HET-s protein from the filamentous fungus Podospora anserina is a prion involved in a cell death reaction termed heterokaryon incompatibility. This reaction is observed at the point of contact between two genetically distinct strains when one harbors a HET-s prion (in the form of amyloid aggregates) and the other expresses a soluble HET-S protein (96% identical to HET-s). How the HET-s prion interaction with HET-S brings about cell death remains unknown; however, it was recently shown that this interaction leads to a relocalization of HET-S from the cytoplasm to the cell periphery and that this change is associated with cell death. Here, we present detailed insights into this mechanism in which a non-toxic HET-s prion converts a soluble HET-S protein into an integral membrane protein that destabilizes membranes. We observed liposomal membrane defects of approximately 10 up to 60 nm in size in transmission electron microscopy images of freeze-fractured proteoliposomes that were formed in mixtures of HET-S and HET-s amyloids. In liposome leakage assays, HET-S has an innate ability to associate with and disrupt lipid membranes and that this activity is greatly enhanced when HET-S is exposed to HET-s amyloids. Solid-state nuclear magnetic resonance (NMR) analyses revealed that HET-s induces the prion-forming domain of HET-S to adopt the β-solenoid fold (previously observed in HET-s) and this change disrupts the globular HeLo domain. These data indicate that upon interaction with a HET-s prion, the HET-S HeLo domain partially unfolds, thereby exposing a previously buried ∼34-residue N-terminal transmembrane segment. The liberation of this segment targets HET-S to the membrane where it further oligomerizes, leading to a loss of membrane integrity. HET-S thus appears to display features that are reminiscent of pore-forming toxins.
| Filamentous fungi have the potential for genetically distinct individuals to fuse, resulting in a cell with multiple nuclei known as a heterokaryon. This fusion event is controlled by genetic variants that determine the compatibility of the individuals, such that the fusion of incompatible genotypes triggers a cell death reaction in the heterokaryon. We have investigated the molecular mechanism of toxicity in the HET-S/HET-s incompatibility system in the fungus P. anserina. HET-s is an infectious yet non-toxic protein (prion) whose interaction with the almost identical protein HET-S has been shown to re-localize HET-S to the cell periphery, an event that is associated with the death of heterokaryons that simultaneously contain both proteins. We find that the HET-s prion converts soluble HET-S into a protein that binds to and destabilizes lipid membranes. Furthermore, we identify a potential transmembrane helix that is normally buried within the soluble fold of HET-S and show that its presence is associated with toxicity. We conclude that upon interaction with a HET-s prion, the HET-S globular domain partially unfolds, exposing a previously buried transmembrane segment that targets HET-S to the membrane. Once there, it further oligomerizes into a structure that causes a loss of membrane integrity, reminiscent of the mode of action of pore-forming toxins.
| Amyloids have long been associated with dozens of diseases including Alzheimer, Parkinson, and prion diseases [1]. However, there are also amyloids with normal biological activities termed “functional amyloids” [2] of which the HET-s prion of the filamentous fungus P. anserina is an interesting example. This prion protein controls an allorecognition process known as heterokaryon incompatibility. Filamentous fungi have developed self/non-self recognition systems that prevent the vegetative fusion of individuals that differ in specific loci termed het-loci [3],[4]. The fusion of incompatible strains results in the rapid destruction of the fusion cell or heterokaryon in a process known as heterokaryon incompatibility. This limitation of cytoplasmic exchanges between unlike strains could act as a protection against the transmission of parasites [5] such as mycoviruses, senescence plasmids, or parasitic nuclei [6]. The number of het-incompatibility loci varies from species to species and although a number of het-loci have been characterized over the years in P. anserina, Neurospora crassa, and Cryphonectria parasitica, the mechanistic modalities of cell death remain largely elusive [6],[7]. The het-s locus is one of nine het-loci in P. anserina and the only one involving a prion protein [8]. The het-s locus exists as two incompatible allelic variants termed het-s and het-S. A strain that expresses the HET-s prion in its soluble form has the [Het-s*] phenotype (phenotypes are denoted throughout with square brackets) and is compatible with het-S strains. However, the HET-s protein can spontaneously aggregate in vivo to form an infectious amyloid yielding the [Het-s] phenotype that is incompatible with [Het-S] strains (those that harbor the HET-S protein) [9]. Because HET-s is a prion, when a [Het-s*] strain fuses with a [Het-s] strain, the amyloid form spreads from one to the other and in the end they both display the [Het-s] compatibility phenotype.
The molecular details of the infectivity of the HET-s prion have been studied in detail. The HET-s and HET-S proteins have two domains: an N-terminal globular domain (HeLo) and a C-terminal prion-forming domain (PFD) [10]. HET-s can function in vivo without its HeLo domain, i.e., its PFD (fused to GFP) is able to carry out the functions of infectivity and incompatibility [10]–[12]. HET-S, although 96% identical to HET-s [13], cannot form a prion, is not infectious, and does not readily form amyloids in vitro, yet it still requires both its HeLo domain and an intact PFD for its function in incompatibility [10],[14]. The HeLo domain of HET-S has also been shown to inhibit in cis the in vivo prion formation of its own PFD [10] as well as in trans the in vivo propagation of [Het-s] [15] and the in vitro aggregation of the HET-s [14]. The 3D structures of the HeLo domains of HET-s and HET-S [14] as well as of the infectious amyloid form of HET-s [16]–[19] have shed some light on the molecular details of infectivity and have confirmed the domain boundaries first identified by protease protection of the soluble and amyloid forms of HET-s [10],[20]. In solution, the HeLo domain comprising residues 1–227 is a mostly helical domain followed by a C-terminal disordered PFD domain comprising residues 228–289. Upon aggregation into the infectious state, the PFD of HET-s adopts a highly ordered β-solenoid structure [17] that is the infectious entity of the HET-s prion [10],[11],[16],[19]. The ten amino acid residue overlap of the two domains (218–227) means that the HeLo domain must at least partially unfold during the β-solenoid formation of the PFD, a statement that is supported by solid-state nuclear magnetic resonance (NMR) data [19].
In contrast to the mechanism of infectivity, little is known about the mechanism of toxicity in the HET-S/HET-s heterokaryon incompatibility reaction. In vivo both the HET-s prion and HET-S are required to produce toxicity, whereas the co-presence of soluble HET-s and HET-S proteins does not affect fungal viability [9],[21]. A recent report demonstrated that a relocalization of HET-S to the membrane periphery is associated with toxicity and that the cytoplasmic co-aggregates of HET-S and HET-s are not toxic [12]. The authors also show that the mechanism of HET-s prion-mediated HET-S toxicity functions in yeast as well as in filamentous fungi. Herein, we present insights into the origin of toxicity of HET-S/HET-s mixtures in which the HeLo domain of HET-S is directly involved in effecting membrane damage. Our data indicate that HET-S binds to the HET-s amyloid through its PFD domain thereby forming a unit of the β-solenoid structure. The beta structuring of the HET-S PFD leads to a destabilization and partial unfolding of the HET-S HeLo domain upon which an N-terminal segment of ∼30 amino acid residues is expelled from the HeLo fold. This segment, which is predicted to be a transmembrane (TM) helix, is found to anchor the protein in the membrane and to convert HET-S into an integral membrane protein. In the lipid bilayer, HET-S oligomerizes into complexes that perforate the membrane thus leading to toxicity via membrane leakage.
Recent results indicated that the HET-s prion triggers HET-S to relocalize at the cell membrane (in Podospora and yeast), and that this event is associated with toxicity [12]. In order to further probe the determinants of toxicity, we sought to reproduce the observed phenomenon in vitro with purified proteins and liposomes composed of Escherichia coli-derived polar lipids. Freshly extruded 100-nm diameter liposomes were incubated for 1 h at 4°C with HET-S with or without approximately equimolar amyloid seeds (the prion form) of the HET-s PFD, HET-s(218–289). We use the latter protein construct as it has been shown to be sufficient in vivo for HET-s infectivity and as a GFP-fusion sufficient for incompatibility [10],[11]. The liposome samples were then frozen in liquid ethane and visualized by freeze-fracture electron microscopy. Transmission electron microscopy (TEM) images of samples that had been incubated at 4°C in the absence of protein revealed liposomes with smooth surfaces (Figure 1A). In the combined presence of HET-S and HET-s(218–289) amyloid seeds, the liposomal membranes are severely disrupted, with more than half of the liposomes displaying defects of 10 to 60 nm in size (white arrows in Figure 1B). In addition, the space between liposomes in the protein-containing sample is filled with small aggregates (indicated by a black arrow in Figure 1B) similar in size to those documented in liposome-free co-aggregation assays of HET-S with HET-s(218–289) [14]. In another experiment, the extruded 100 nm liposomes were incubated for 1 h at 4°C with HET-S only or HET-S with approximately equimolar HET-s(218–289) amyloid seeds. The results (Figure S1) show that at 4°C only the combination of HET-S with amyloid seeds leads to defects in the liposome membrane.
The size of the defects found in the TEM images suggest that they should lead to a significant loss of liposome integrity, so we employed a dye leakage assay to observe the effects of HET-S on liposomal membranes under various conditions. Using liposomes filled with 60 mM calcein, a concentration at which significant self-quenching occurs, we could simultaneously measure kinetics for the efflux of liposomal contents for many samples with a fluorescence plate reader [22]. As the calcein leaves the interior of the liposome, it is diluted more than 10,000-fold leading to a fluorescence increase that is monitored at 520 nm with excitation at 485 nm. Triplicate measurements were made for samples with a range of protein concentration: 125 nM–8 µM HET-S and 125 nM–4 µM HET-s(218–289) seeds (monomer-equivalent concentration). We performed the experiments at 10°C in order to minimize the volume change due to evaporation during the 6-h-long measurements. Additionally, the liposomes were more stable at 10°C than at room temperature, leading to a lower background leakage. The results, plotted in Figure 1C and 1D, demonstrate that significant liposome leakage is only observed for the samples that contain both HET-S and HET-s(218–289) amyloid seeds. In the range of concentrations tested, the individual proteins did not give rise to significant leakage in the time course of the experiment at 10°C. Comparison of the kinetic curves in Figure 1C and 1D indicates that even though the amyloid seeds are required to get HET-S dependent leakage, the rate of leakage and maximum leakage is determined by the HET-S concentration and essentially independent of the seed concentration in the range tested. This finding indicates that the leakage induction requires an interaction between HET-S and a HET-s(218–289) prion but that very few seeds of HET-s(218–289) (at least 1,000× less than HET-S) are required to get the maximal effect.
We have long observed that upon induction of HET-S expression in E. coli, but not for HET-s induction, the host undergoes a growth arrest during expression at 37°C (Figure 2). Low temperature induction can partially overcome this toxicity, and we routinely use 18–25°C to minimize the amount of insoluble HET-S and maximize overall yield. In contrast, overexpressed HET-s is not toxic to E. coli. While HET-s is usually found in inclusion bodies (short inductions can yield some soluble protein), HET-S was always found to partition between a soluble and insoluble fraction, with longer induction times leading to a lower yield of soluble protein and more proteolytic degradation (unpublished data). In light of its membrane disruption activity (Figure 1), we wanted to investigate whether HET-S exerts its toxicity in E. coli by interacting with the host membrane. We found that upon induction at 37°C in E. coli, the insoluble fraction of HET-S was associated with the bacterial inner membrane. First, we observed that HET-S remains in the insoluble fraction when washing the membranes with 1 M urea but is solubilized by N-lauroylsarcosine. This detergent is known to not solubilize the outer membrane as exemplified by the observation that the outer membrane protein goes into the pellet (Figure 2C). We also fractionated E. coli cell lysates via isopycnic centrifugation using a 25%–55% (w/w) sucrose gradient, conditions known to resolve the two membranes of different densities [23]. SDS-PAGE analysis of the fractions revealed that the majority of HET-S was soluble while a significant fraction migrated with the membranes, having a density distribution that more closely matched the less dense inner membrane (visualized by NADH oxidase activity) than the outer membrane (visualized by OMP-F migration) (Figure 2D). In addition, there was a dense fraction that migrated to the bottom of the tube that, based on its high density, must either be lipid-free or low-lipid content aggregates. The identity of the HET-S and OmpF bands on the gel were confirmed by tryptic digest followed by MALDI-TOF analysis (unpublished data) and the migration of HET-S observed by western blot with an anti-His-tag antibody. Thus, in addition to its soluble form, HET-S occurs in a membrane-associated state and as insoluble protein aggregates when overexpressed in E. coli. The membrane association and toxicity of HET-S in E. coli suggests a mechanism of action that is similar to what has been seen in yeast and Podospora [12].
In contrast to Podospora, where the HET-s prion is required for toxicity, HET-S is toxic in E. coli in the absence of a HET-s prion seed. Our previous report on the mechanism of prion inhibition by the HET-S HeLo domain found that a lower thermodynamic stability of the HeLo domain of HET-S compared to that of HET-s is associated with HET-S activity. Our data indicated that prion inhibition of HET-s by HET-S involves a complex with a destabilized HET-S HeLo domain, and we hypothesized that it may also be the precursor to the toxic entity [14]. Therefore we attempted to learn if at a temperature closer to the observed unfolding transition of the HET-S HeLo domain (48°C) prion-independent liposome disruption and leakage could be observed. The results depicted in Figure 3A show that at 30°C, HET-S can induce the same hole-like structures in liposomes as seen previously at 4°C in the presence of prion seeds, while the negative control with liposomes in absence of HET-S did not show such defects (unpublished data). Also, at 30°C HET-S can induce a low but significant level of liposome leakage without the HET-s prion (Figure 3B). Apparently, in vitro at elevated temperature HET-S is capable of some level of self-activation in the absence of HET-s seeds. Yet, this seed-independent liposome leakage activity of HET-S is not observed for the isolated HET-S HeLo domain (Figure 3B), further evidence of a role for the PFD region in HET-S activation. An analogous auto-activation of HET-S has not been observed even at elevated growth temperatures (37°C) in either P. anserina or yeast (unpublished data). It may be that the extent of auto-activation is too low to lead to a detectable in vivo toxicity or else that the chaperone machinery suppresses the toxic effect of this auto-activation.
Several reports have identified single amino-acid substitutions in HET-S that convert it to a protein that gives the [Het-s] phenotype in vivo [21],[24]. Such a high degree of structural and sequence similarity between two proteins that differ in their in vivo localization and activities is suggestive of a loss (or gain) of function mutation. In light of the membrane leakage data and EM images (Figures 1 and 3), the fact that all of the inter-converting mutations (HET-S↔HET-s) reside in the HeLo domain suggests that the HeLo domain of HET-S has a membrane disrupting function. In principle, this function should require a membrane-binding activity as well as a pore-forming activity. Since membrane binding may be more a permissive activity than a pore-forming activity, we set up an assay to measure liposome binding in order to see if the two activities are distinct. Liposomes were prepared as for the TEM analyses and proteins were incubated either at 4°C in the presence or absence of amyloid seeds (prion activation) or at 30°C without seeds (thermodynamic activation). The samples were then centrifuged at 150,000 g, and the pellets and supernatants analyzed by SDS-PAGE. The assay for the seeded activation of liposome binding suffers from the inability to separate the liposome-bound protein from simple prion-induced, lipid-independent aggregates; however, there is still significant and reproducible prion-activated binding of HET-S to the liposomes (unpublished data). The results for the thermodynamic activation are clearer and show that HET-S does but HET-s does not bind to liposomes (Figure 3C and 3D). We also tested the HeLo domain of HET-S alone using the construct HET-S(1–227), and it is able to associate with liposomes, whereas HET-s(1–227) is not. Finally we tested HET-S[E86K], a mutant that has lost its incompatibility with [Het-s], thus converting HET-S to a HET-s like protein [21]. However, unlike HET-s, HET-S[E86K] retains membrane binding activity. This means that membrane binding, as demonstrated by HET-S(1–227) and HET-S[E86K] is not sufficient for heterokaryon incompatibility. These same proteins were also subjected to the liposome leakage assay and only the full-length HET-S protein gives rise to leakage (Figure 3B). Thus, the [Het-S] phenotype of a protein is directly correlated with liposomal membrane leakage and not just with liposome binding.
A protein that can exist in both a soluble and a membrane-associated form is likely to undergo a structural change when moving from the more polar to the more hydrophobic environment. Faced with the finding that both thermal destabilization and the interaction with a prion can induce HET-S to adopt a membrane-associated form, we wanted to investigate the structure of the HET-S HeLo domain when it is in complex with the HET-s prion. Therefore co-aggregates of [13C, 15N]-labeled HET-S with non-labeled (and therefore NMR silent) HET-s or HET-s(218–289), formed by mixing them at a 1∶1 molar ratio as well as co-aggregates obtained by seeding with non-labelled HET-s(218–289) amyloid seeds, were investigated by solid-state NMR. Inspection of the NCA 2D spectra (Figure 4A–4C) immediately indicated that the spectra of HET-S contain the previously assigned narrow resonances of HET-s(218–289) [18],[25]. Upon sequential assignment of the spectra using the 3D correlation experiments NCACB, NCOCX, and CANCO [26], a good correlation between the chemical shifts of previously reported HET-s(218–289) aggregates and HET-S in the various co-aggregate samples was observed (Figure 4A–4C), indicating a very similar backbone and sidechain conformation. Only for residues near K235 (which is E235 in HET-s) and for residues A237, A248, and N279 were considerable chemical-shift differences observed (Figure 4G), however barely larger than one ppm, which is typically regarded as a value for significant structural differences. Thus, it is concluded that the PFD of HET-S assumes the HET-s β-solenoid fold [17] when it is co-aggregated with HET-s(218–289) or HET-s.
Since the conformational change of the PFD of HET-S appears to be triggered by the presence of HET-s(218–289) it is likely that the two proteins interact with each other in the β-solenoid structure in the same way as do the individual molecules within HET-s(218–289) amyloid. In order to confirm this type of inter-molecular interaction in the HET-S/HET-s(218–289) co-aggregates, an aggregated sample from 13C-labeled HET-S and 15N-labeled HET-s(218–289) was prepared. The presence of cross peaks in a 13C-15N polarization transfer experiment (in the present case PAIN) [27],[28] is a direct proof of molecular-level contacts between HET-S and HET-s(218–289), thus indicative of an intimate mixing of the two proteins in the fibrils. The PAIN spectrum of the mixed sample of 13C-labelled HET-S and 15N-labelled Het-s(218–289) given in Figure 4D, allowed us to assign a number of hetero-intermolecular backbone-backbone contacts. Almost all peaks are explained by in-register contacts (residue peaks (i→i+36) between HET-S and HET-s(218–289) [17], establishing that the intermolecular interface between the residues involved in the β-solenoid is the same between HET-S and HET-s(218–289) as between two HET-s(218–289) monomers in pure fibrils. All the peaks observed are also present in the mixed sample of HET-s(218–289) (Figure 4D, grey contours), although the signal-to-noise in the HET-S/HET-s(218–289) spectrum is smaller due to the larger molecular mass of HET-S and hence fewer molecules in the sample. Further support for the hetero-intermolecular interaction within the β-solenoid is the observation that G271 of HET-S shows a resonance doubling. This indication of a structural heterogeneity is expected for a mixed amyloid of HET-S and HET-s(218–289) because in the β-solenoid structure, G271 is adjacent to residue 235 in the neighboring molecule (G235 in HET-s and K235 in HET-S, depicted in Figure 4F). Comparison of the peak-intensities for G271 and G271′ in the 1∶1 HET-S:HET-s(218–289) co-aggregates indicates that the HET-S/HET-s and HET-S/HET-S interfaces are roughly equally abundant, indicating a random mixing of the two monomers.
While the PFD segment of HET-S in these coaggregates with HET-s(218–289) displays narrow NMR signals attributed to the formation of the β-solenoid structure, the HeLo domain (residues 1–220) has very broad peaks (Figures 4A–4C and S2). An analogous observation has been documented for the HeLo domain in full-length HET-s fibrils [19] and points to a loss of the well-defined tertiary structure in the globular HeLo domain of HET-s induced by the β-solenoid fibril formation of its PFD. The similarity of the spectra obtained for HET-S and HET-s aggregates indicates that both HeLo domains undergo a similar structural rearrangement. Since the HeLo domain in its soluble form and the PFD in its amyloid form both include residues 218–227, it is rationalized that this domain overlap induces a global loss of tertiary structure in the HeLo domain upon amyloid formation of the PFD. This is consistent with the finding that decoupling the HeLo domain from the PFD β-solenoid conformation, by a construct that fuses the HET-S globular domain (1–227) to the PFD region (218–289) with a six amino-acid linker between the two domains, results in a loss of the HET-S heterokaryon incompatibility activity [14].
To probe the structural differences between the soluble versus lipid-associated forms of HET-S we chose limited proteolysis with proteinase K (PK) since it has already proven useful for the characterization of the domain boundaries of the soluble and amyloid forms of HET-s [20]. For this experiment, HET-S was pre-incubated for 1 h with and without liposomes at 37°C followed by proteolysis with varying concentrations of PK. The liposome sample was centrifuged at 150,000 g for 20 min to collect the digested protein fragments that remain associated with liposomes and this fraction was analyzed side by side with the liposome-free sample by SDS-PAGE. Significant differences in the protection pattern as well as with the overall sensitivity of HET-S to proteolysis were observed. The results in Figure 4G show that HET-S becomes more sensitive to PK digestion when bound to liposomes. Several of the digested bands could be analyzed successfully by N-terminal Edman degradation, giving insights into the parts of the protein that remain associated to the lipids after proteolysis. Of particular interest is that the fragments that remained associated with the liposome and that could be unambiguously identified (the most abundant) all contain at least 24 of the first 34 residues (Figure 4G), while a major band in both the non-liposomal HET-S digest and in the supernatant of the liposome digest has residue 35 as its N-terminus. In summary, the limited proteolysis data indicate that the liposome association occurs via an N-terminal segment and that it induces an increase in protease sensitivity. An increased proteolytic sensitivity indicates a loss of tertiary structure, consistent with what was observed in the NMR experiments when HET-S was exposed to HET-s(218–289) prion seeds (see above).
During routine bioinformatics analyses of HET-S and HET-s we discovered a “hidden” TM helix in the HeLo domain. Several algorithms predict with varying degrees of certainty that an N-terminal stretch of residues in the HeLo domain is a TM helix. HMMTOP [29] predicts one in both HET-S (residues 7–25) and HET-s (residues 5–22). TOPPRED [30] also predicts a TM helix but with more certainty for HET-S (residues 5–25) than HET-s (residues 2–22), the latter scored as “putative.” TMpred, which is based on the TMbase [31] database of membrane proteins, also predicted similar regions to be TM helices with a higher score for HET-S (residues 7–25) than HET-s (residues 4–22). While it is clear that there is some TM character to the HeLo domain, another prediction algorithm, TMHMM [32], appears more discerning. The output of the TMHMM algorithm includes a per residue probability that is used to calculate the overall TM score for a stretch of amino acids. While the prediction algorithm is more complex than these simple scores imply, TM helices in membrane proteins generally have per residue probabilities above 0.7 and total scores (the sum of the per residue probabilities within a single stretch) of at least 17. While TMHMM predicts neither HET-S nor HET-s to be membrane proteins, there is a considerable signal in the HET-S sequence (TM score = 9) that is absent in the HET-s sequence (Tm score = 0.5). Therefore we used TMHMM for all further predictions of the HET-S/s variants that interconvert the protein activities. We also found that shorter sequences gave higher scores, so that when residues 1–45 instead of 1–289 are submitted to the TMHMM-2.0 server, HET-S is predicted to have a TM helix (TM score = 18) while HET-s is still not (TM score = 4). Experimental support for these TM predictions is that the N-terminus of HET-S remains bound to HET-S/liposome mixtures treated with PK (Figure 4G).
To further test the relationship between the TM prediction, the in vitro observations of membrane leakage and the in vivo incompatibility phenotype, we initiated a study of some of the well-known variants of HET-s and HET-S. Of the 13 amino acid differences in the sequences HET-s and HET-S it is sufficient to replace a single residue in HET-S (H33P) to convert the protein to a HET-s-like protein that has the [Het-s] compatibility phenotype in P. anserina. Conversely, minimally two substitutions are required in HET-s (D23A and P33H) in order to switch the phenotype from [Het-s] to [Het-S] [24]. In other words, in vivo HET-s[D23A, P33H] leads to toxicity upon confrontation with a strain that carries the HET-s prion, while a P. anserina host that carries the HET-S[H33P] variant is compatible upon fusion with a strain bearing a HET-s prion. In fact these HET-S variants confer all the qualities of [Het-s] including infectivity, prion formation, and incompatibility with HET-S bearing strains. The results of the liposome leakage assay with these variants are depicted in Figure 5. There is a near perfect correlation between the TMHMM predictions, the leakage data, and the observed phenotype. The only exception to the correlation is the positive TM prediction for HET-S(H33P), as this variant actually gives a [Het-s] phenotype and does not cause liposome leakage. Thus, while the TM prediction is a good indicator of phenotype, the correlation between liposome leakage and HET-S-like toxicity is 100% (Figure 5; Table S1).
If, as our data suggest, HET-S exerts a toxic activity by damaging the fungal membrane through the formation of an integral membrane protein, then this raises the question of how a membrane protein with a single TM helix is able to induce liposome leakage and the large observed defects in the liposomal membrane (Figure 1). In light of the observation that HET-S(1–227) is able to bind to liposomes without causing liposome leakage, a simple explanation of the mechanism of toxicity is that HET-s amyloid seeds not only trigger HET-S to form a membrane active complex, but also support its oligomerization into an aggregate species that then is both membrane active and toxic. Oligomerization-based mechanisms of membrane disruption are in fact common for many pore-forming toxins. Consistent with this type of mechanism is the finding that the HET-s-like variant HET-S[E86K], which has lost its ability to dimerize in solution, can still bind to the liposomes like HET-S(1–227) without causing liposome leakage however. In order to find support for the suggested oligomer-based mechanism we attempted to monitor the homo-oligomerization of HET-S in a membrane-like environment by size-exclusion chromatography with multi-angle light scattering (MALS), refraction index (RI), and ultraviolet (UV) detection. This triple detection scheme with MALS measurements allows for a model-free molecular weight determination of the protein and detergent components in a protein-detergent complex (PDC) [33]. For these measurements, recombinant HET-S was incubated at room temperature in 0.4% foscholine-12 (FC-12) or extracted by 0.4% FC-12 from liposomes with which it had been pre-incubated at 37°C (conditions shown to cause HET-S membrane binding and damage, Figure 6). When starting from soluble protein, the addition of FC-12 led to a time-dependent evolution of HET-S from a monomeric species to an oligomeric species of intermediate size and finally to a larger aggregate (Figure 6A). HET-S extracted from liposomes had a similar distribution of species as the 16-h incubation of soluble HET-S (Figure 6A). The mass of the protein component in the intermediate species was in the range of 100–500 kDa while the larger aggregates were >1 MDa (a precise measurement was not possible because the larger aggregates eluted in the void volume of the column and we therefore only measure a weight-average mass of any co-eluting species). Similar time-dependent oligomerization was observed with another detergent, n-Dodecyl-β-D-maltoside (unpublished data). To test the role of the PFD in this oligomer formation, HET-S(1–227) was treated with the same detergents. The HeLo domain alone remains monomeric for a longer time and does not form the intermediate-sized oligomers, rather going directly into >1 MDa aggregates (Figure 6A). We followed the same time-dependent processes by circular dichroism (CD) spectroscopy and found that HET-S continuously loses alpha-helical content on the same time scale as the aggregation, reaching a stable point at about 50% of the original signal (Figure 6B). This loss of structure is consistent with the solid-state NMR measurements of the HET-S aggregates (Figures 4 and S2) and the increase in PK sensitivity (Figure 4G).
Here we show that HET-S and HET-s variants that give the [Het-S] phenotype can be activated in vitro by the HET-s prion to form a membrane-permeabilizing complex (Figures 1, 2, and 7). Our findings give detailed insights into prion-mediated heterokaryon incompatibility and support a new model for the underlying mechanism. Building upon our previous model for the HET-S prion-inhibition mechanism, our model for the mechanism of HET-S toxicity is depicted in Figure 7. HET-S interaction with HET-s prion aggregates leads to a template driven folding of the HET-S PFD into the same β-solenoid structure as the HET-s PFD (Figure 4). Due to the structural overlap of the HeLo and PFD domains, this folding of the PFD leads to a destabilization of the HeLo fold (Figures 4 and S2). This destabilization activates the prion-inhibitory activity as well as the toxic activity of HET-S. The destabilized HeLo domain of HET-S undergoes a restructuring that exposes the previously buried amino-acid residues of its TM helix, targeting the activated HET-S to the cell membrane where it assembles into a membrane-disrupting complex. We propose that HET-S activity is dependent on two of its functions: (i) membrane binding as demonstrated by the liposome pulldown assay (Figure 1), and (ii) oligomerization into the toxic entity as demonstrated by the MALS data with HET-S in membrane-mimicking detergents (Figure 6) and as suggested by the size of the defects in the liposome membranes (Figure 1). Interfering with either of these functions converts HET-S to a HET-s-like protein. The D23A mutation in HET-S eliminates the predicted TM helix and converts it to a more HET-s-like protein (Figure 5). The E86K mutation disrupts a dimerization interface that we previously identified in the HeLo domain [14] and therefore affects the oligomerization properties of HET-S and correspondingly its liposome leakage potential (Figure 3B). The H33P mutation does not eliminate the predicted TM helix (Figure S3); however, the proline may restrict the conformation of the helix so that it cannot oligomerize into the toxic entity. Other HET-s-like variants of HET-S although not studied here can also be classified into the type of function that they have likely lost. For example, HET-S[F25S] loses the predicted TM helix and becomes like HET-s [21]. A construct that fuses the HeLo domain to the PFD via a six-residue linker (relieving the domain overlap) has been shown to be HET-s-like [14], and this can be explained by the uncoupling of the HeLo domain structure from the aggregation of the PFD (Figures 4 and S2). Finally, all mutations in the PFD that disrupt the β-solenoid structure are null mutants. That this is true for both HET-S and HET-s further validates the model that the activation of HET-S toxicity involves a template-directed structuring of its PFD by the HET-s PFD.
The proposed model in which HET-S is activated by a HET-s prion in order to carry out its toxic membrane-permeabilizing activity (Figure 1) may seem at odds with our results showing that HET-S can bind to and cause leakage from liposomes in a prion-independent manner (Figure 3) and that expression of HET-S alone leads to a growth arrest in E. coli (Figure 2). Taken at face value, these results contradict our model; however, it is easy to see how an in vitro or non-native in vivo system could lead to such discrepancies as discussed in the detail in Text S1. The most important factor described therein is that the mechanism of interest is dependent on changes in the thermodynamic stability of the HeLo domain, an event that in vivo is carefully orchestrated by the proteostasis including chaperones. While these apparent discrepancies can only be rationalized by arguments, it must be noted that there is a perfect correlation of the phenotype to the in vitro activity of the proteins (whether prion-induced or thermodynamically controlled) indicating the discussed activity is biologically relevant.
Initially described in the context of human disease, amyloids were later found to be able to ensure a variety of functional roles for instance as cell surface structures in microbes or as storage and release devices of peptide hormones [2],[34],[35]. In the present model, the β-solenoid fold of HET-s is used as a specific conformational switch (Figure 7) for the activation of the pore-forming HeLo domain. This mode of activation is specific and simple as it does not involve a covalent modification such as a proteolytic cleavage or other type of post-translational modification. The folding energy associated with the acquisition of the amyloid structure drives the conformational conversion of the HeLo domain (Figures 4 and S2). This mode of regulation appears highly efficient, as very little HET-s seeds are required for conversion. Our study reveals that an amyloid fold can have a regulatory role in controlling the activity of another protein or domain, thus expanding the repertoire of functional tasks that can be assigned to amyloids. It appears that the inherent templating ability of amyloids is specifically exploited in this system not only to ensure maintenance of the [Het-s] state but also to allow for activation of HET-S by HET-s.
Since, the HET-s/HET-S system involves a prion amyloid in the context of a cell-death reaction, it was at first tempting to imagine that the mechanisms involved might be related in some way to the amyloid toxicity seen in human diseases. The present study now clarifies that this is not the case. It is not that amyloid toxicity is exploited to control a purposeful cell death reaction, but instead that the amyloid prion fold has a distinct functional role as an activation trigger of a toxic domain (Figure 7). The β-solenoid fold of HET-s is the trigger that converts the HET-S protoxin into a toxin (Figures 1C and 1D, 4, and 7). Our observations illustrate how both components of this cell death system can be harmless on their own but lethal when combined (Figure 1C and 1D). HET-s can constitutively harbor the β-solenoid fold because conversion of its own HeLo domain does not lead to toxicity while HET-S is stable in vivo as a protoxin while awaiting activation.
The HET-s/S proteins, the yeast prion proteins Ure2p and Sup35 and the mammalian prion proteins (PrPs) are not homologous proteins and have no sequence similarities. What they have in common is the ability to transmit a phenotype via propagation of structural rearrangements in an infectious manner. However, they also share the structural motif of a globular folded domain attached to a highly flexible region that is able to convert to an amyloid-like entity [36]. In contrast to HET-s with a folded N-terminal HeLo domain, the folded domain of PrP (residues 124–227) is C-terminal to its flexible domain (residues 23–123). Further similarities are that for both HET-s and PrP, the prion forming domains (residues 218–289 in HET-s and ∼23–140 in PrP) overlap with the folded domain, thus requiring a conformational change of the latter upon prion formation. In addition, the globular domains of HET-S and PrP have a predicted TM helix (residues 1–34 in HET-S and 113–128 in human PrP) that is integrated into the soluble fold (Figures S3 and S4 for HET-S). These similarities that occur in unrelated prion systems are intriguing and let us speculate that like in the HET-s/S system (Figure 7), conversion of PrPC to the prion form PrPSc, may induce a structural transition in its globular domain that exposes a TM segment, thus forming an integral membrane protein that exerts its toxicity through membrane interaction. This hypothesis is supported by some experimental data showing that PrPC (including in particular familial variants thereof) can be an integral membrane protein under certain circumstances associated with toxicity [37].
While we have observed that the HET-S HeLo domain itself is toxic when overexpressed in E. coli (Figure 2), the PFD is also required for the HeLo domain to function in vivo. Therefore we cannot exclude that the amyloid structure of the PFD (Figure 4) is not intimately involved in the membrane disruption activity (i.e., more than as an oligomerization scaffold). There are many examples of amyloid proteins that can disrupt membrane integrity though channel or pore formation [38], most notably Aβ, which has been shown to form non-specific ion channels [39]. The propensity of amyloids to associate with lipid bilayers has led to the channel hypothesis of amyloid toxicity (reviewed in [40]). However, the similarity between HET-s and other pore-forming amyloids does not go beyond the fact that they are both amyloids. Our data currently support a pore-forming activity in the HeLo domain with a mode of activation that more resembles larger non-amyloid cytotoxins than pure amyloid peptide-based toxicity (Figure 7).
The mechanism of HET-S toxicity is general enough to be organism-independent (observed in bacteria [Figure 2], yeast, and filamentous fungi) and involves a membrane association. As we demonstrate, HET-S forms holes in membranes and causes the leakage of liposomal contents (Figure 1). HET-S should therefore be considered a pore-forming toxin. Pore-forming toxins are a large class of virulence factors employed by a wide range of bacterial pathogens (such as diphtheria toxin and anthrax toxin) but are also found in eukaryotes where they can be involved in immune defense [41]. As implied by their name, this class of toxins forms pores in the membranes of target cells that result either in membrane leakage or in the delivery of toxic components through the pores. They are a diverse class of proteins that are further classified as α- or β-pore-forming toxins, depending on the type of secondary structure they acquire when inserted into the membrane. However, despite the wide diversity in sequences and structures of pore-forming toxins, they typically act by a common mechanism. One important characteristic of pore-forming toxins is their transformation from a soluble, monomeric protein to an integral membrane protein oligomer [42],[43]. For this transformation the toxin must undergo a conformational change in which hydrophobic segments that are buried in the soluble fold become exposed to the hydrophobic environment of the membrane lipids. In this manner, many toxins can be secreted in a soluble form that diffuses to the membrane of the target cell. Oligomerization of the toxin initiated by another protein/antigen/small molecule to a prepore is necessary for stable binding to the target cell and for membrane integration [44],[45]. HET-S therefore bears many of the typical characteristics of pore-forming toxins. In addition to the toxicity induced by its membrane association and the generation of pore-like entities in liposomes (Figure 1), it also has a protein partner that triggers the conformational change from a soluble to an integral membrane protein (i.e., the HET-s prion) (Figures 1, 2, and 4). While the structure of the membrane active complex has yet to be determined for HET-S, the fact that it has a predicted TM helix and the observation that the corresponding peptide fragment in detergents is helical as determined by CD (unpublished data) makes it likely to be an α-pore forming toxin.
Whether as a mechanism of defense or attack, production of pore-forming toxins is an ancient and conserved mode of warfare between cells. The majority of the currently characterized pore-forming toxins are of bacterial or animal origin [46]. Yet, the existence of pore-forming toxins has also been reported in the fungal kingdom where they function in pathogenicity, host-defense, and inter-organismal competition [47]–[50]. It is striking that the cell death process occurring in heterokaryon incompatibility also appears to rely on that basic and ancient mode of toxicity based on membrane disruption.
This study represents the first insight into the mechanistic aspects of cell death by incompatibility in any fungal system. It appears that in this system induction of cell death occurs by a direct and universal toxicity mechanism and relies on the formation of a protein structure related to pore-forming toxins. Pore-forming toxins are typically used in prokaryotes and eukaryotes both as attack and defense molecules in various pathogenic or competitive biotic interactions. One may find it surprising that an organism should use this mode of toxicity against itself by targeting its own membrane as occurs in HET-s/HET-S incompatibility. It has been proposed that genes controlling heterokaryon incompatibility in Podospora may derive by exaptation from genes involved in host-defense against heterospecific non-self [51]. The HET-S/HET-s system could exquisitely illustrate this concept. Indeed, the mechanism of HET-S toxicity induction postulated here might also occur in a different context. Recently, a potential additional functional partner of HET-S was identified. A search for proteins displaying homology to the HET-s PFD has identified a protein termed NWD2 encoded by the gene immediately adjacent to het-S in the genome of Podospora [52]. NWD2 belongs to the NWD-gene family comprising other het-genes. NWD proteins are STAND proteins resembling Nod-like receptors and are believed to represent the fungal counterparts of pathogen recognition receptors described in plants and metazoans. STAND proteins are signal-transducing NTPases that undergo ligand-induced oligomerization and typically display three domains: a central nucleotide-binding oligomerization domain (NOD) flanked by a C-terminal ligand-binding domain and an N-terminal effector domain. NWD2 lacks a defined effector domain and in place displays at the N-terminal end a short region of homology with the elementary repeat motif of the HET-s PFD. A model postulating the existence of a functional interaction between NWD2 and HET-S was proposed. In that model, NWD2 recognizes a non-self ligand via its C-terminal WD-40 repeat domain and oligomerizes in response to this binding. This oligomerization step would put the N-terminal extensions of NWD2 proteins into close proximity and allow their cooperative folding into the β-solenoid fold. Once formed this fold would template the HET-S PFD thus activating the HeLo toxicity domain much in the same way as the HET-s prion does. In this mode of activation, it might be that the membrane-disrupting activity of HET-S is used as part of a host-defense or attack mechanism and directed to the membrane of a biotic interactant of Podospora. This second proposed mode of activation appears more widespread and evolutionary ancient than the HET-s-induced activation, because the het-S nwd2 gene pair is conserved in a number of distant species while there is no evidence for the existence of het-s/het-S incompatibility system outside of Podospora. Since a simple loss of function mutation in the HeLo domain (for instance mutation H33P) will turn HET-S into a non-toxic prion protein that now acquires the ability to trigger activation of HET-S, it is easy to see how the HET-s/HET-S might have evolved by exaptation starting from het-S.
There are to date more than 100 HeLo domain sequences retrieved with an iterative PSI-BLAST search, all of which belong to the genomes of filamentous fungi (Pfam accession: PF14479). Among those are putative het-S-homologs with an associated PFD, all of which are adjacent in their respective genomes to a gene encoding a STAND protein with a consensus PFD-like motif at its N-terminus [52]. A number of HeLo domains occur as N-terminal domains of STAND proteins. This domain organization can be seen as the “all-in-one” counterpart of the het-S nwd2 gene pair architecture [52] and is analogous to the three domain architecture found in other PCD inducing het genes from Podospora, in which the HeLo domain is replaced by another type of cell-death–inducing domain termed the HET domain (Figure 8B). Since the HET domain has been shown to mediate PCD [53] and HeLo and HET domains both occur as the N-terminal domain of numerous STAND proteins, we suggest it is likely that the HeLo domain has a general role as a cell death-inducing domain. Our data indicate that the HET-S HeLo domain mediates PCD through its action as a pore-forming toxin upon activation through oligomerization of its C-terminal PFD. This finding is also consistent with the fact that in plant and animal kingdoms, STAND proteins have been shown to oligomerize, leading to activation of their N-terminal effector domain [54],[55], such as the death domain, the death effector domain, or the caspase-activation and recruitment domain [56].
In addition to the STAND proteins, the HeLo domain appears in many proteins with uncharacterized domains (domains that are automatically generated by ADDA [57]). Interestingly, the HeLo domain is found almost exclusively at the N-termini of the proteins and the few exceptions could be artifacts of the automatic ORF prediction and annotation of the sequence databases. That the HeLo domain appears as the first domain in multi-domain proteins further suggests that it has a conserved function involving the insertion of its own N-terminal segment into a lipid membrane. With it being an N-terminal domain there are no topological restrictions on its insertion in to the membrane, whereas a centrally located HeLo domain would need either an extra TM segment or the ability to translocate the domains N-terminal to it, across the membrane. Furthermore, the TM segment is well conserved in HeLo domain proteins as highlighted in Figure 8. The sequences in this alignment are the non-redundant output of a PSI-BLAST search with residues 4–33 of HET-S (carried to convergence at an E-value threshold of 0.005). Therefore, these 35 sequences are a subset of HeLo domains that are more similar to HET-S in their TM region. The output of the TMHMM algorithm, represented graphically in Figure 8, indicates that 21 of the 35 sequences have a TM helix (when only residues 1–38 are analyzed, see Figure 8 legend) while another algorithm, TMPRED, predicts all of them to have TM helices except for HET-s, which is classified as “putative.” The prediction of a high conservation of the TM helix is not simply an artifact of our having performed the PSI-BLAST search with the TM region of HET-S. TMPRED identifies the same TM helix in nearly all of the more than 100 HeLo domains in the sequence databases. We therefore conclude that the conservation of the TM helix as well as the HeLo domain's context and its N-terminal position in the sequence of multidomain proteins indicates that HeLo domain proteins are a widespread (in filamentous fungi) class of pore-forming toxins.
Constructs with N-terminal histidine tags were expressed from genes that were sub-cloned into the pRSET vector, yielding a protein product with the N-terminal sequence MRGSHHHHHHGLVPRG/S directly preceding the coding region. Thus cleavage of the protein product with thrombin (recognition site underlined, cleavage indicated by a backslash) yielded the protein of interest with an N-terminal serine. The C-terminally histidine-tagged constructs were expressed from pET21 or pET24 vectors with the tag directly following the last residue of the protein (no cleavage site).
Cultures of E. coli BL21*(DE3)pLysS were prepared from a single fresh transformant in the following manner: a single colony was streaked onto a new LB plate, grown overnight at 37°C, and then the entire plate was used to inoculate M9 minimal medium at a starting OD600 of 0.1–0.2. When the cell density reached OD600 = 0.8, the culture was transferred to 18°C and protein expression was induced with 0.5 mM IPTG for 12 h. All the mutants were expressed similarly, but for 6 h at 27°C. The cells were collected by centrifugation and re-suspended in the lysis buffer (50 mM Tris pH 8.0, 200 mM NaCl, 20 mM imidazole, 2 mM DTT, 10% v/v glycerol, 1 mM EDTA) with 1 complete, EDTA-free Protease-Inhibitor Cocktail tablet (Roche) per 50 ml of Lysis buffer and 0.5 mg/ml lysozyme (Sigma) at 4°C. Lysis was achieved by three passes over a microfluidizer at processing pressure of 18,000 psi (M-110S, Microfluidics) and the insoluble material was removed by centrifugation at 100,000 g for 35 min. The supernatant was passed over a nickel FF-sepharose column (GE Healthcare) that was then washed with ten column volumes lysis buffer followed by elution buffer (50 mM Tris pH 8.0, 200 mM NaCl, 200 mM imidazole, 2 mM DTT, 10% v/v glycerol). The histidine-tag was removed by thrombin (bovine plasma, Sigma) treatment at room temperature overnight in the elution buffer. Thrombin was removed on a benzamidine FF column (GE healthcare). The protein was further purified on a Superdex 200 gel filtration column (GE Healthcare) pre-equilibrated with 10 mM NaPO4 (pH 7.4), 200 mM NaCl, 10% glycerol, 1 mM DTT. The protein was either used immediately or frozen in small aliquots in liquid N2 and stored at −80°C.
HET-s(218–289) and HET-s variants were expressed as described for other HET-s constructs [10] and therefore the purification procedure is only described briefly. The bacterial pellet was lysed in lysis buffer, passed over a microfluidizer and the insoluble material was collected as described above. The pellet was then solubilized in 50 mM Tris (pH 8.0), 150 mM NaCl, 6 M guanidinium hydrochloride overnight at 60°C. After centrifugation for at least 2 h at 150,000 g, the protein was purified from the supernatant using a nickel FF-sepharose column and then desalted into 200 mM acetic acid. Monomeric protein was immediately lyophilized and stored at −80°C. Full-length HET-S with a non-cleavable C-terminal His-tag was also expressed and purified as follows: one plate of transformed cells was pooled into a starter culture of LB medium and grown at 37°C to an OD600 of 0.6, pelleted at 5,000 g, and resuspended in 1 l minimal medium with a starting OD600 of ∼0.1. At an OD600 of 0.6, the culture was induced with 0.5 mM IPTG for 3–4 h at 37°C. This protein construct was purified as described above from the soluble fraction, but it was also partially present in the membrane fraction of E. coli. For the membrane fractionation, cells were lysed by three passes over a microfluidizer at processing pressure of 18,000 psi. The protein lysate was centrifuged at 8,000 g for 15 min to collect non-lysed cells and large debris. The supernatant was immediately subjected to a 2-h centrifugation at 100,000 g. The pellet containing the membranes was resuspended in PBS containing 1 M Urea. Washed membranes were harvested subsequently by a second round of centrifugation. This crude membrane fraction was treated overnight with 1% N-laurylsarcosine in PBS and then centrifuged again at 100,000 g.
Cultures of E. coli BL21*(DE3)pLysS were prepared from transformants of an entire plate. All colonies were resuspended in 10 ml of LB media, the OD was measured, and cells were diluted in 100 ml to a starting OD600 of 0.1–0.2. The cultures were incubated at 37°C with shaking, and when the cell density reached OD600 = 0.5, they were induced with 0.5 mM IPTG for 7 h. During the entire growth, the OD600 was measured every 30 min.
The separation of inner and outer membranes of E. coli was adopted from a protocol described by Osborn et al. [23]. Briefly, cells from a 4 h induction of HET-S (C-terminal HisTag) in LB medium at 37°C were resuspended at 0.2 g wet-cell-weight per ml of lysis buffer (including protease inhibitor and lysozyme) and sonicated in a volume of 2 ml with a Bandelin Sonopuls UW2070 microtip at 70% power for 1 min on ice. The lysates were centrifuged for 5 min at 5,000 g at 4°C to remove unbroken cells. The supernatant (1.2 ml) was mixed with 50% (w/w) sucrose (0.8 ml) so that the resulting mixture was 25% (w/w) sucrose, and 1 ml of this was loaded onto the top of the sucrose gradient. The sucrose gradient was prepared in a 12-ml centrifugation tube (Beckman, REF 331372) by layering 2.4 ml of 50%, 1.8 ml each of 45%, 40%, 35%, and 2.3 ml 30% over a 0.5-ml 55% sucrose cushion. All sucrose solutions contained 10 mM Tisi, X mM EDTA (pH 8). The gradients were centrifuged in a SW41 rotor at 30,000 rpm for 16 h at 4°C. Fractions of 1 ml each were collected though a needle by inserting it from the top so that the tip was 5 mm from the bottom of the tube. The bottom 200 µl (Figure 3, fraction 1) were removed afterwards with a pipette so as to retrieve any material that sedimented through the 55% cushion onto the bottom of the tube. DTT (2 mM) was immediately added to the fractions and they were separated by SDS-PAGE and visualized by both Coomassie staining and Western blot with an anti-HisTag antibody (Abcam 18184) using a standard ECL kit (GE Healthcare). The NADH oxidase activity of each fraction was measured in triplicate by mixing 50 µl with 350 µl water and then 400 µl NADH buffer (100 mM Tris pH 7.5, 0.2 mM DTT, 0.12 mM NADH). The rate of decrease of absorbance at 340 nm during the first 5 min was taken as the relative activity of NADH oxidase.
Lyophilized monomeric HET-s(218–289) was solubilized by an appropriate amount of 1.5 M Tris (pH 8) to reach (pH 7) followed by immediate mixing with HET-S (10 mM NaPO4, 200 mM NaCl, 1 mM DTT) at a 1∶1 molar ratio. The co-aggregating mixtures were incubated with slow rotation in 50 ml tubes for at least 7 d at room temperature. The fibrils for solid-state NMR were centrifuged directly into a 3.2-mm Bruker MAS rotor at 200,000 g using a custom-made tool [58].
E. coli –derived polar lipids were purchased from Avanti Polar Lipids. Chloroform-containing lipid solutions were dried in round bottom flasks (8 cm in diameter) under a gentle nitrogen stream in a fume hood. Residual chloroform was removed under vacuum. For the calcein release assay, the dry lipid film resulting from evaporation was resuspended in aqueous 60 mM calcein (Sigma), which was prepared in 10 mM Tris, 150 mM NaCl. The pH of the calcein solution was adjusted with concentrated 1 M Tris to (pH 8) and the solution was filtered through a 0.2-µm membrane before it was added to the dried lipids. The resuspended solution at a final lipid concentration of 20 mg/ml was incubated for 1 h at room temperature with occasional vortexing to allow lipid hydration and vesicle formation. After 1 h, the lipid suspensions were vigorously vortexed to allow complete detachment of hydrated lipids. The resulting suspension was subjected to five freeze-thaw cycles of 10 min each: freezing in liquid nitrogen followed by 10 min thawing at 37°C. Calcein-loaded vesicles were frozen at a concentration of 20 mg/ml and in 100 µl aliquots at −20°C. For the preparation of liposomes for cryo-EM, PK digestion, and liposome binding assays, the lipid film was resuspended in buffer L (20 mM Tris, 150 \mM NaCl, pH 7.5) without calcein and then processed as in the calcein-containing lipids.
Liposomes for the calcein leakage assay were prepared starting with 1 ml of calcein-loaded vesicles that were diluted to a lipid concentration of 2 mg/ml and extruded with 20 passes through a 400-nm porous membrane (Whatman), followed by 20 passes through a 100-nm membrane. The calcein that remained outside of the liposomes was separated from extruded liposomes by two sequential passes over Sephadex G-25 PD10 columns (GE Healthcare) pre-equilibrated with buffer L. The calcein-loaded liposomes were then diluted 100-fold for the calcein leakage assay and used within 2 d. Liposomes were mixed with wild-type HET-S and HET-s or their variants as indicated in the text at final concentrations in a range between 8 µM to 80 nM. The assay was carried out using a plate reader instrument (Pherastar, BMG) and 96-well plates (Greiner Bio, 655900) allowing for simultaneous data acquisition of 32 reactions in triplicate (including the control samples). The release of calcein from liposomes was monitored over time as an increase of calcein fluorescence intensity (λex = 485 nm; λem = 520 nm). Leakage was calculated from average values of the triplicate measurements following fthe equation , where and are the emission intensities in buffer controls and in liposome solutions with the addition of 1% SDS, respectively.
The standard deviation of the leakage was calculated in three steps. First, the standard deviation of each measured intensity was calculated from , where is the value, is the average value, and the number of values, here. Second, the standard deviations of the blank-subtracted values and were calculated from standard deviations of the sample and blank intensitiesstandard deviations: . Third, the standard deviation of the leakage was calculated: where the angled brackets indicate average values.
For the cryo-EM studies, 100 nm extruded liposomes (10 mg/ml) were incubated with and without 20 µM protein at 37°C or 4°C for 90 min. Droplets of the liposome solution were placed on an Au-grid between two copper blades used as sample holders and then frozen in liquid propane cooled to −180°. Freeze fracturing of the samples was performed in a Balzers 400T apparatus cooling the specimen stage at −160°C. Etching occurred at −110°C for 8 min, then Pt/C shadowing was applied in a 45° angle (with respect to the specimen stage) and pure carbon at 90° onto the sample. After thorough cleaning of the replicas in water, EtOH and Acetone, Au-grids were analyzed in a FEI Morgagni TEM electron microscope at 100 kV with nominal magnification of 8,000×, 16,000×, 32,000×. The freeze fractures were also analyzed by a Zeiss 1530 Gemini at 5 kV with nominal magnifications up to 200,000× at −120°C.
Proteins were mixed with 20 µl of 20 mg/ml E. coli polar liposomes or buffer L. The volume of all samples was adjusted with buffer to 50 µl and samples were incubated at 30°C for 2 h or overnight or at 4°C overnight. After incubation liposomes were pelleted by a 20-min centrifugation at 180,000 g. Control samples containing no liposomes were treated alike as a control for liposome-independent protein precipitation. The supernatant was immediately collected and mixed with SDS-PAGE-loading buffer. The pellet was resuspended in 50 µl PBS containing 0.1% DDM and 1× SDS-loading buffer. Samples were directly subjected to SDS-PAGE or immediately frozen at −20°C.
50 µg of HET-S protein were digested on ice for 30 min with 20, 2, 0.2, and 0.02 µg of PK in a volume of 250 µl after incubation at 37°C for 1 h in the presence and absence of 16 mg/ml liposomes. Reactions were stopped by mixing equal volumes of the reaction and 30 mM phenylmethylsulfonylfluorid (PMSF). Liposomes were separated from the supernatant by centrifugation at 180,000 g for 10 min. 40 µl of the supernatant was mixed with a 6× SDS loading buffer and boiled (final 2× SDS loading buffer) for 5 min. The pellet was first resuspended in 40 µl of buffer and then brought to 2× SDS loading buffer in the same manner. 20 µl of each reaction were immediately analyzed by SDS-PAGE followed by Coomassie Blue staining. For N-terminal sequencing, an unstained SDS-PAGE copy was transferred onto a PVDF membrane and analyzed by the Functional Genomics Center Zurich.
Concentrated protein constructs were diluted to 20 µM in 0.4% FC-12, PBS, 10% glycerol for analysis on a Jasco J-815 CD spectrometer. Far UV CD spectra (190–240 nm) were measured at 27°C with a 20-nm/min scan rate in a 0.1-cm path-length cuvette. The loss of structure in HET-S was observed for 16 h by collecting spectra every hour.
Recombinant HET-S and HET-S 227 were mixed with 0.4% of FC-12, PBS, 10% glycerol to a final concentration of 20 µM protein. The protein solutions were separated by size with a G4000PWXL gel filtration column (Tosoh Bioscience) in the same buffer on an Agilent 1200 series HPLC system injected by the autosampler (20 µl) at a flow rate of 0.5 ml/min. The column eluate was simultaneously monitored by the 1,200 series diode array detector, a TREOS light-scattering detector (Wyatt Technology), and the 1,200 series refractive index detector. The signals from the detectors were analyzed by protein conjugate analysis in the Astra V program (Wyatt Technology) using an average protein dn/dc value of 0.187 ml/g, the literature value for FC-12 (0.1398 ml/g) (www.wyatt.eu) and the protein extinction coefficient as calculated by the ProtParam tool (http://web.expasy.org/protparam/).
All solid-state NMR spectra were recorded on a Bruker Avance II+ 850 spectrometer operating at a static magnetic field of 20.0 T using a Bruker 3.2 mm triple-resonance low-E (LLC) magic-angle spinning (MAS) probe. The experiments were carried out at MAS frequencies of 17.5 or 18 kHz at sample temperatures of 3–10°C. All experimental parameters are given in Table S2.
|
10.1371/journal.pbio.1001237 | A Modular Library of Small Molecule Signals Regulates Social Behaviors in Caenorhabditis elegans | The nematode C. elegans is an important model for the study of social behaviors. Recent investigations have shown that a family of small molecule signals, the ascarosides, controls population density sensing and mating behavior. However, despite extensive studies of C. elegans aggregation behaviors, no intraspecific signals promoting attraction or aggregation of wild-type hermaphrodites have been identified. Using comparative metabolomics, we show that the known ascarosides are accompanied by a series of derivatives featuring a tryptophan-derived indole moiety. Behavioral assays demonstrate that these indole ascarosides serve as potent intraspecific attraction and aggregation signals for hermaphrodites, in contrast to ascarosides lacking the indole group, which are repulsive. Hermaphrodite attraction to indole ascarosides depends on the ASK amphid sensory neurons. Downstream of the ASK sensory neuron, the interneuron AIA is required for mediating attraction to indole ascarosides instead of the RMG interneurons, which previous studies have shown to integrate attraction and aggregation signals from ASK and other sensory neurons. The role of the RMG interneuron in mediating aggregation and attraction is thought to depend on the neuropeptide Y-like receptor NPR-1, because solitary and social C. elegans strains are distinguished by different npr-1 variants. We show that indole ascarosides promote attraction and aggregation in both solitary and social C. elegans strains. The identification of indole ascarosides as aggregation signals reveals unexpected complexity of social signaling in C. elegans, which appears to be based on a modular library of ascarosides integrating building blocks derived from lipid β-oxidation and amino-acid metabolism. Variation of modules results in strongly altered signaling content, as addition of a tryptophan-derived indole unit to repellent ascarosides produces strongly attractive indole ascarosides. Our findings show that the library of ascarosides represents a highly developed chemical language integrating different neurophysiological pathways to mediate social communication in C. elegans.
| Chemical signaling is an ancient form of inter-organismal communication. The nematode Caenorhabditis elegans exhibits a wide range of social behaviors, including mutual attraction and aggregation, and has served as a useful model towards investigating the signaling pathways that regulate these behaviors. Recent investigations showed that other C. elegans behaviors, like population density sensing and mating, are regulated by small molecule signals called ascarosides. However, it remained unclear whether C. elegans uses small molecules to promote intraspecific attraction and aggregation, despite the presence of extensive neural circuitry regulating these behaviors. In this study, we show that C. elegans uses a specifically modified variant of the ascarosides including an indole unit as a highly potent aggregation pheromone. These indole ascarosides integrate input from two major metabolic pathways, amino acid catabolism and lipid beta-oxidation, suggesting that C. elegans communicates metabolic status via a modular code of small-molecule signals. Our study thus provides evidence for use of a multilayered chemical language for inter-organismal signaling by a model organism. Understanding of chemical signaling in nematodes may aid the development of new treatment approaches for parasitic nematodes, which remain among the most prevalent human disease agents.
| Communication among individuals of a species relies on a number of different sensory inputs including chemical, mechanical, auditory, or visual cues [1]. Chemical signaling is perhaps the most ancient form of interorganismal communication [1],[2], and analysis of the chemical signals and the behaviors they mediate is of great significance for understanding the ecological and evolutionary dynamics of intra- and inter-specific interactions. The free-living nematode C. elegans is used extensively as a model system for social behaviors such as foraging, population density sensing, mating, and aggregation (http://www.wormbook.org; [3]). Recent investigations have shown that a family of small molecules, the ascarosides, play important roles as chemical signals regulating several different aspects of C. elegans behavior (Figure 1A) [4]–[8]. The ascarosides ascr#1, ascr#2, and ascr#3 were originally identified as major components of the dauer pheromone, a population-density signal that promotes entry into an alternate larval stage, the non-feeding and highly persistent dauer stage [4]–[7]. Additional work showed that at concentrations far below those required for dauer formation, synergistic mixtures of ascarosides act as strong male-specific attractants, and that male attraction to ascarosides requires the amphid sensory neurons ASK and the cephalic sensory neurons CEM [6],[7]. Wild-type (N2) hermaphrodites do not respond to low concentrations of ascarosides and show repulsion at dauer-inducing concentrations [4]. However, a recent study showed that mutation of the neuropeptide-Y receptor homolog NPR-1 strongly affects hermaphrodite response to ascarosides [9]. The strong loss-of-function mutant npr-1(ad609) showed attraction or reduced repulsion to specific combinations of ascarosides, in contrast to wild type (N2) worms that express a high-activity variant of NPR-1 [10],[11]. The interneuron RMG, the central site of action of NPR-1, is proposed to serve as a central hub computing aggregation and attraction signals originating from several different sensory neurons, including the ascaroside-sensing ASK neurons [9].
These findings suggested that mutual attraction and aggregation in C. elegans are mediated primarily by signaling via NPR-1, and that strains carrying the high-activity form of NPR-1 including wild-type (N2) hermaphrodites may not rely on small molecule signaling to promote aggregation. Nonetheless, wild-type (N2) hermaphrodites also display aggregation behaviors, for example, in response to environmental cues such as limited food availability [12] or perturbations of transforming growth factor–β (TGF–β) signaling [13]–[15]. Given the existence of small molecules that serve as social cues for population density sensing and mate-finding, and the complicated neural circuitry implicated in aggregation behavior, we hypothesized that structurally distinct small molecules might exist that serve as aggregation signals in C. elegans. Here we show that C. elegans aggregation behavior is regulated by a dedicated set of highly potent signaling molecules, the indole ascarosides, which form part of a modular chemical language that elicits structure-specific behaviors via several distinct neurophysiological pathways. Our findings provide evidence for multi-layered social signaling in C. elegans.
All currently known small-molecule pheromones in C. elegans are derived from peroxisomal β-oxidation of long-chained fatty acids via DAF-22, a protein with strong homology to human sterol carrier protein SCPx [6],[16]. We hypothesized that putative aggregation pheromones may be derived from the same pathway, suggesting that daf-22 mutants would not produce them. In this case, a spectroscopic comparison of the wild-type metabolome with that obtained from daf-22 mutant worms should reveal candidate compounds for attraction or aggregation signals.
In a previous study, we had used an NMR spectroscopy-based technique termed Differential Analysis of NMR spectra (“DANS”) to compare the wild-type metabolome with that of daf-22 mutant worms [6]. This comparison had led to identification of ascr#6–8, of which ascr#8 is a major component of the male-attracting signal [6]. Based on NMR spectra with improved signal-to-noise ratio, we conducted a more detailed comparison of wild type and daf-22-mutant metabolomes, which revealed several indole-containing compounds in the wild-type metabolome that were not produced by daf-22 worms (Figure 1B,C). The established role of DAF-22 in pheromone biosynthesis [6],[16],[17] suggested that these indole derivatives may represent a previously unrecognized family of signaling molecules.
To clarify the structures and biological roles of the daf-22-dependent indole derivatives, we pursued their complete identification via NMR spectroscopy-guided fractionation of the wild-type metabolome. Reverse-phase chromatography produced eight metabolite fractions, which were analyzed by two-dimensional NMR spectroscopy. The NMR spectra revealed the presence of daf-22-dependent indole-derivatives in two fractions, which were selected for additional NMR-spectroscopic and mass spectrometric studies. These analyses indicated that the most abundant daf-22-dependent indole derivative consists of an indole carboxy unit linked to ascarylose bearing a 9-carbon unsaturated side-chain identical to that found in the known ascr#3 (see Supporting Information for NMR and MS data) [7]. Based on its structural relationship to the known ascr#3, we named the newly identified metabolite indole carboxy ascaroside “icas#3” (Figure 1E).
Next we asked whether some of the other daf-22-dependent indole compounds we had detected by DANS also represent indole ascarosides. For this purpose, we employed a mass spectrometric (MS) approach, because analysis of the mass spectra of icas#3 had revealed a characteristic MS fragmentation pattern (loss of the indole-3-carboxy moiety, Figure S1) that enabled a screen for related compounds. MS screening for compounds with similar fragmentation profiles indicated that icas#3 is a member of a larger series of indole ascarosides featuring side chains with five to nine carbons (Figure 1D,E). The most abundant components of this family of indole ascarosides are icas#3, icas#9, and icas#10, which are accompanied by smaller amounts of icas#1 and icas#7 (Figure 1E). All of these compounds represent new metabolites, except for icas#9, which recently has been reported to possess moderate dauer-inducing activity and is unique among known dauer pheromones in producing a bell-shaped response curve [18]. We also detected two new non-indole ascarosides: ascr#9, which features a saturated 5-carbon side chain, and ascr#10, which features a saturated 9-carbon side chain, thus representing the saturated analog of the known ascr#3 (Figure 1F).
The MS analyses further revealed that the indole ascarosides' quantitative distribution is distinctly different from that of the corresponding non-indole ascarosides, suggesting that incorporation of the indole unit is strongly regulated. Notably, the most abundant indole ascaroside, icas#3, is accompanied by 10–40-fold larger amounts of the corresponding non-indole ascaroside, ascr#3, whereas icas#9 is more abundant than the corresponding ascr#9 (Figure 1G). To determine the biosynthetic origin of the indole ascarosides and to exclude the possibility that they are produced by the E. coli food source, we established axenic (bacteria-free) in vitro cultures of C. elegans (N2) using the chemically defined CeMM medium [19],[20]. HPLC-MS analysis of the axenic cultures revealed the presence of icas#1, icas#3, icas#9, and icas#10, thus indicating that indole ascarosides are produced by C. elegans without participation of dietary bacteria. Use of a 1∶1 mixture of L-[2,4,5,6,7-D5]-tryptophan and L-tryptophan in the axenic medium resulted in production of [D5]-icas#1, [D5]-icas#3, [D5]-icas#9, and [D5]-icas#10, along with equivalent amounts of the unlabelled compounds (Figure S2). In conclusion, our biochemical studies established the tryptophan origin of the indole-3-carboxy moiety in the indole ascarosides and indicate that these compounds are products of a strongly regulated biosynthetic pathway.
The addition of an indole-3-carboxy moiety to the ascarosides represents a significant structural change, and we hypothesized that this chemical difference may indicate signaling functions for these compounds distinct from those of their non-indole cognates. Using synthetic samples (see Supporting Information), we tested three indole ascarosides of varying side-chain lengths, icas#1, icas#3, and icas#9, in the spot attraction assay we had used previously to demonstrate social functions of small molecules (Figure 2A) [6],[7]. We found that all three tested indole ascarosides, icas#1, icas#3, and icas#9, attract both males and hermaphrodites at high concentrations (Figure 2C). Testing the most abundant indole ascaroside, icas#3, over a broader range of concentrations, we observed that at low concentrations icas#3 was strongly attractive to hermaphrodites, whereas males were no longer attracted (Figure 2D, Movies S1, S2, S3). Similarly, hermaphrodites, but not males, are strongly attracted to low concentrations of icas#9 (). We further investigated hermaphrodite attraction to icas#3 using a quadrant chemotaxis bioassay as described previously (Figure 2B) [9],[21]. In contrast to the spot attraction assay, which measures attraction to a point source of compounds, the quadrant chemotaxis assay measures aggregation of hermaphrodites on plate sections with well-defined compound concentration [9],[21]. We found that concentrations as low as 1 pM icas#3 result in strong attraction of hermaphrodites (Figure 2E), both in the presence and absence of food (Figure S3B).
The biological role of icas#3 thus starkly differs from that of the corresponding non-indole ascaroside ascr#3, which strongly attracts males but repels hermaphrodites [6],[7]. Our results show that simply by attaching an indole-3-carboxy group to the 4-position of the ascarylose, the strongly male-attracting ascr#3 is converted into a signal that primarily attracts hermaphrodites. The difference in the amounts at which ascr#3 and icas#3 are produced by the worms corresponds to their relative potency: the male-attracting ascr#3, which is of much lower potency than icas#3, is produced in much higher concentrations than the highly potent hermaphrodite attractant icas#3 (Figure 1G).
The results from the spot attraction and quadrant chemotaxis assays indicate that hermaphrodites are strongly attracted to indole ascarosides, suggesting that these compounds regulate C. elegans aggregation behavior. C. elegans exhibits natural variation in its foraging behavior with some strains (e.g., the common laboratory strain N2) dispersing individually on a bacterial lawn, whereas most wild-type strains (e.g., RC301 and CB4856 (Hawaii)) accumulate and aggregate where bacteria are the most abundant [10],[22]. These variants are referred to as “solitary” and “social,” respectively [10],[11]. These differences in foraging and aggregation behavior are associated with two different alleles of the neuropeptide Y-like receptor NPR-1 [10],[11], which differ at a single amino acid position: solitary strains such as N2 express a high-activity variant of NPR-1 (215-valine), whereas aggregating strains such as CB4856 express a low-activity variant of NPR-1 (215-phenylalanine) [10],[11]. The strong loss-of-function mutants npr-1(ad609) and npr-1(ky13), which were generated in the N2 background, also show a high tendency to aggregate [10],[22].
A previous study showed that loss of function of npr-1 affects hermaphrodite response to non-indole ascarosides [9]. Whereas wild-type (N2) worms expressing the high-activity variant of NPR-1 are repulsed by non-indole ascarosides, npr-1(ad609) mutants showed attraction to a near-physiological mixture of the most abundant non-indole ascarosides, ascr#2, ascr#3, and ascr#5 [9]. We confirmed attraction of npr-1(ad609) hermaphrodites to ascr#2/3/5 mixtures using both the quadrant chemotaxis and spot attraction assays, but found that hermaphrodites of the two tested social wild-type strains (RC301 and CB4856) show no attraction in either assay (Figure 3A–B). In contrast, both social wild-type strains (RC301 and CB4856) as well as npr-1(ad609) hermaphrodites were strongly attracted to icas#3, in both the quadrant chemotaxis and spot-attraction assays (Figures 3B–D, S3B–C). These results indicate that icas#3 functions as a hermaphrodite attractant in both solitary and social C. elegans strains.
We next tested how a constant background concentration of indole ascarosides affects hermaphrodite behavior. We measured aggregation of solitary N2 worms and several social strains (including the social wild-type strain CB4856 and two npr-1 loss-of-function mutants) in response to icas#3 using two different conditions: “high worm density,” with 120 worms per 5 cm plate, and “low worm density,” with 20 worms per 5 cm plate. At low worm density, we observed a very strong increase in aggregation at concentrations as low as 10 fM (femtomolar) icas#3 for both solitary and social hermaphrodites (Figure 4A, 4E). Aggregation of N2 hermaphrodites increased as much as 4-fold at 1 pM icas#3, with higher icas#3 concentrations producing less aggregation. Similarly, the naturally occurring social strain CB4856 displayed a bell-shaped response curve with maximal aggregation at 1 pM of icas#3 and lower aggregation not significantly different from control at 1 nM of icas#3 (Figure 4A). In contrast, icas#3 increased aggregation of npr-1(ad609) hermaphrodites over the entire tested concentration range, without a drop-off at higher concentrations (Figure 4A). At high worm density, we observed up to a 3-fold increase in aggregation of N2 hermaphrodites on icas#3 plates (Figure 4B,F), whereas hermaphrodites from all three tested social strains showed nearly complete aggregation even in the absence of icas#3, which precluded detection of any additional aggregation-promoting effect of icas#3 (Figure 4B). These results show that icas#3 increases hermaphrodite aggregation even in the absence of a concentration gradient of this compound, and that solitary and social strains are similarly affected. Similarly, the second-most abundant indole ascaroside, icas#9, increased aggregation of both solitary and social hermaphrodites (Figure S4A). We also investigated the effect of icas#3 on aggregation of males, which generally tend to aggregate in the absence of hermaphrodites [23]. We found that aggregation of him-5 males on icas#3 plates was significantly increased (Figure S4B).
These results show that indole ascarosides promote aggregation behavior even in the absence of a concentration gradient, suggesting that sensing of icas#3 and icas#9 affects response to other aggregation-promoting (chemical or other) signals or conditions. For example, secretion of additional indole ascarosides by the worms on plates containing exogenous icas#3 could contribute to the observed increase in aggregation. To investigate this possibility, we tested daf-22 hermaphrodites in the aggregation assay. daf-22 hermaphrodites do not produce indole ascarosides but respond to icas#3 in both the spot attraction and quadrant chemotaxis assay as strongly as N2 worms (Figures 3B, S3C). We found that daf-22 hermaphrodites show less aggregation than N2 worms at 1 pM icas#3 but not at 10 pM icas#3 (Figure 4C). These results suggest that secretion of additional indole ascarosides or other daf-22-dependent compounds by the worms may contribute to aggregation on icas#3 plates, but that other factors, for example low oxygen levels or contact with other worms [12],[13],[24], are more important. Furthermore, changes in locomotory behavior on icas#3 plates could affect the level of aggregation [12]. Using an automated machine-vision system to track worm movement [25], we found that aggregation-inducing concentrations of icas#3 strongly increase mean stopped duration and affect other locomotory parameters (Figures 4D, S4D, S4E). These changes in worm locomotion, in conjunction with other aggregation-mediating factors, may contribute to the observed increase in aggregation on icas plates.
The amphid single-ciliated sensory neurons type K (ASK) play an important role in mediating C. elegans behaviors, and previous work has shown that the ASK neurons are required for behavioral responses of males and hermaphrodites to non-indole ascarosides [7],[9]. ASK sensory neurons are connected via anatomical gap-junctions to the RMG interneuron, which has been shown to act as a central hub regulating aggregation and related behaviors based on input from ASK and other sensory neurons (Figure 5A) [9],[26]. To investigate the neural circuitry required for icas#3-mediated hermaphrodite attraction and aggregation, we first tested whether the ASK neurons are required for these behaviors. For this purpose, we used worms lacking the ASK neurons due to cell-specific expression of mammalian caspase in the developing neurons (Tokumitsu Wakabayashi, Iwate University Japan, personal communication). We found that ablation of ASK sensory neurons resulted in a near complete loss of attraction to icas#3 (Figure 5B). In contrast, ablation of the ASI neurons, which like the ASK neurons partake in dauer pheromone sensing, had no significant effect on icas#3 mediated attraction in hermaphrodites (Figure 5B). Further, ablation of both ASI and ASK neurons did not result in a more significant loss of attraction compared to ASK ablations alone, suggesting that the ASK sensory neurons are required for sensing icas#3 (Figure 5B). Next we tested whether the ASK neurons are required for icas#3 mediated aggregation. We found that hermaphrodites lacking the ASK neurons do not aggregate in response to icas#3 at any of the tested concentrations (Figure 5C). Locomotory analysis of ASK-ablated hermaphrodites on icas#3 plates showed neither increased reversal frequency nor decreased velocity, as we had observed for wild-type worms (Figure S5A,B).
Next we tested whether the RMG interneuron is required for icas#3-mediated behaviors. We identified the cell position of the RMG interneuron in wild-type worms using DIC microscopy [27] and in a transgenic strain expressing ncs-1::gfp (a gift from the Bargmann Lab). This transgene expresses GFP in the RMG interneuron and a few other sensory neurons [9]. We found that ablation of the RMG interneuron in both wild-type and ncs-1::gfp worms did not affect icas#3-response in the spot attraction assay (Figure 5B). These results indicate that the RMG interneuron is not required for transduction of icas#3-derived attraction signals from the ASK sensory neurons, in contrast to the behavioral effects of non-indole ascarosides, which require both the ASK sensory neurons and the RMG interneuron [9]. Given this observation, we sought to understand which interneuron downstream of ASK is required for response to icas#3. According to the wiring diagram of C. elegans, the primary synaptic output of the ASK neuron is the AIA interneuron [26]. To test whether this neuron is required for sensing icas#3, we used a transgenic line expressing a hyperactive form of MEC-4 in the AIA interneuron (a kind gift from the Ishihara lab, Japan) [28]. Expression of MEC-4, a DEG/ENaC sodium channel, causes neuronal toxicity in C. elegans, thereby resulting in the loss of the AIA neuron [29]. These AIA-deficient worms did not show any attraction to icas#3, suggesting that the AIA interneurons are required for icas#3 response. Hence the neural circuitry required for attraction to icas#3 is different from that of the non-indole ascarosides.
Since behavioral assays showed that the ASK and AIA neurons participate in sensing icas#3, we asked whether icas#3 elicits calcium flux in these neurons. To measure Ca2+ flux, we used transgenic lines expressing the genetically encoded calcium sensors (GCaMP) in these neurons [9]. We used the “Olfactory chip” to restrain the worms and applied ON and OFF steps of icas#3 while imaging from these neurons [30]. We were not able to detect Ca2+ transients in ASK neurons even when applying a wide range of concentration ranging from 1 pM to 1 µM. We then monitored calcium responses in the AIA interneuron, which is the primary synaptic target of the ASK neuron [26]. We found that icas#3 elicited significantly increased G-CaMP fluorescence in the AIA neurons (Figure 5D,E, Movie S4), similar to the results reported by Macosko et al. for stimulation of AIA interneurons with a mixture of three non-indole ascarosides [9]. These results show that the ASK sensory neurons are required for icas response and that this response is transduced via the AIA interneuron.
Previous studies have shown that high, dauer-inducing concentrations of ascr#3 strongly deter both social and solitary hermaphrodites [7],[9]. To investigate whether addition of ascr#3 would affect icas#3-mediated attraction of hermaphrodites, we tested mixtures containing these two compounds in a near-physiological ratio of 12∶1 (ascr#3∶icas#3) in a modified spot attraction assay, in which we scored N2 hermaphrodite attraction to three concentric zones A–C (Figure 2A). We found that at the lower of the two concentrations tested (120 fmol ascr#3 and 10 fmol icas#3), the presence of ascr#3 did not interfere with icas#3-mediated attraction, whereas higher concentrations of ascr#3 resulted in strong repulsion, even in the presence of proportionally increased icas#3 concentrations (12 pmol ascr#3 and 1 pmol icas#3, Figure 6A). Following retreat from the outer edge of zone A, many worms remained “trapped” in a circular zone B surrounding central zone A, repulsed by the high concentration of icas#3/ascr#3-blend inside zone A, but attracted by the lower concentrations of the icas#3/ascr#3 blend that diffused into zone B (see Movie S5 for a visual record of this behavior). These results show that at high concentrations of physiological icas#3/ascr#3 mixtures the repulsive effect of ascr#3 prevails, whereas at lower concentrations attraction by icas#3 dominates.
Indole ascarosides are the first C. elegans pheromones that strongly attract wild-type hermaphrodites and promote aggregation. The indole ascarosides fit the broad definition of aggregation pheromones in that they attract and/or arrest conspecifics to the region of release irrespective of sex [1],[31],[32]. In promoting these behaviors, the indole ascarosides are active at such low (femtomolar) concentrations that the worm's behavioral response must result from sensing of only a few molecules. For example, at an icas#3 concentration of 10 fM there are only about 20 icas#3 molecules contained in a cylinder corresponding to length and diameter of an adult hermaphrodite. Given their high specific activity, it is not surprising that indole ascarosides (icas') are of much lower abundance than non-indole ascarosides (ascr's).
The indole ascarosides' strongly attractive properties suggest that these compounds serve to attract conspecifics to desirable environments such as food sources. However, the results from our competition experiments indicate that attraction of hermaphrodites by icas#3 can be counteracted by high concentrations of ascr#3, which are repulsive to hermaphrodites [7]. The competition experiments further showed that at low concentrations of a physiological blend of icas#3 and ascr#3, the attractive properties of icas#3 dominate, whereas at high concentrations of the same blend the repulsion by ascr#3 becomes dominant (Figure 6A, Movie S5). These findings suggest that under dauer-inducing conditions with high population density, the associated high concentrations of ascr#3 promote dispersal [7], whereas low population density and correspondingly lower concentrations of ascr#3 result in attraction mediated by icas#3. Therefore, icas' and ascr's could represent opposing stimuli regulating population density and level of aggregation. In turn, population density, food availability, and other environmental factors may affect relative rates of the biosyntheses of ascr's and icas' as part of a regulatory circuit.
Indole ascarosides affect aggregation behavior even in the absence of a concentration gradient: very low background concentrations (fM-pM) of icas#3 and icas#9 strongly increase the propensity of hermaphrodites (and males) to aggregate. This finding suggests that sensing of icas#3 and icas#9 increases susceptibility for aggregation-promoting (chemical or other) signals or conditions, for example additional quantities of icas' secreted by the worms on the plate.
Aggregation in C. elegans is known to depend on a diverse set of genetic factors and environmental conditions, including food availability and oxygen concentration, suggesting the existence of neuronal circuitry that integrates inputs from different sources [10],[33]–[36]. Aggregation and attraction signals originating from several different sensory neurons, including the oxygen-sensing URX-neurons and the ascr-sensing ASK neurons, have recently been shown to converge on the RMG interneuron, which is proposed to act as a central hub coordinating these behaviors [9]. The RMG interneuron is the central site of action of the neuropeptide-Y receptor homolog NPR-1, which distinguishes solitary strains (high NPR-1 activity) from social strains (low NPR-1 activity) [10],[11]. In social npr-1(lf) mutant hermaphrodites, oxygen-sensing URX neurons promote aggregation at the edges of the bacterial lawn, whereas solitary N2 hermaphrodites do not respond to oxygen gradients. Similarly, repulsion by ascr's depends on NPR-1, as solitary hermaphrodites are repelled by ascr's, whereas social npr-1(lf) hermaphrodites display either greatly diminished repulsion or weak attraction [9]. In contrast, we show that icas#3 promotes hermaphrodite attraction and aggregation in both social and solitary strains. Icas#3 attracts solitary N2 as well as social npr-1(lf) hermaphrodites and increases hermaphrodite aggregation in the solitary strain N2, the social wild-type strains RC301 and CB4856 (Hawaii) carrying a low-activity variant of NPR-1, and the two tested npr-1 null alleles. The finding that icas#3-mediated attraction and aggregation is not reduced by high NPR-1 activity suggests that these icas#3-mediated behaviors rely on signaling pathways distinct from those controlling aggregation responses to other types of stimuli, for example low oxygen levels. This hypothesis is supported by our observation that hermaphrodites lacking the RMG interneuron, which coordinates other aggregation responses via NPR-1, are still attracted to icas#3. Furthermore, icas#3-mediated aggregation differs from NPR-1-dependent aggregation behavior in that aggregation of worms on icas#3 plates is more dynamic and not restricted to the edge of the bacterial lawn where oxygen is limited (Animations S1, S2). Worm velocity is not significantly reduced at the icas#3 concentrations that induce maximal aggregation (1–10 pM, Figure S4D), and icas#3-mediated aggregation is associated with less clumping (average clump size 3-5 worms) than found for aggregating NPR-1 mutant worms (average clump size 6–16 worms) [12]. These observations show that icas#3-mediated aggregation is phenotypically distinct from aggregation behaviors controlled by NPR-1 and the RMG interneuron.
Icas#3-mediated attraction and aggregation depend on the ASK neurons, similar to hermaphrodite repulsion and male attraction by ascr's [7], confirming the central role of this pair of sensory neurons for perception of different types of pheromones in C. elegans (Figure 5). We further show that icas#3 responses are dependent on the AIA interneurons and do not require the RMG interneuron. Therefore, it appears that the sensory neuron ASK participates in perception of two different types of pheromones, ascr's and icas', and that these signals are transduced via two different neurophysiological pathways, as part of a complex neural and genetic circuitry integrating a structurally diverse array of pheromone signals.
Calcium transients have been recorded from amphid sensory neurons in response to non-indole ascarosides; however, the reported changes in G-CaMP fluorescence were relatively small (on the order of about 20%) [9],[37]. Recently, it was reported that the non-indole ascaroside ascr#5 does not elicit calcium transients in the ASI sensory neurons, although the ASI neurons function as sensors of ascr#5 and express the ascr#5-receptors srg-36 and srg-37 [38]. Similarly, we were unable to detect significant Ca2+ transients in the ASK neurons in response to a wide range of concentrations of icas#3 (unpublished data). It is possible that any icas#3-elicited Ca2+ signals in this neuron are even weaker than those of non-indole ascarosides, as icas#3 is active at extremely low concentrations (femtomolar to low picomolar). Additionally, we cannot rule out involvement of additional neurons in icas#3 signaling, given that the ASK neurons are postsynaptic to a number of other sensory neurons [26]. Notably, icas#3 elicited significant changes in G-CaMP fluorescence in the AIA interneurons, which are the primary postsynaptic targets of the ASK sensory neurons (Figure 5D,E, Movie S4).
The identification of indole ascarosides as aggregation signals reveals unexpected complexity of social signaling in C. elegans. Our results indicate that ascarylose-derived small molecules (icas' and ascr's) serve at least three distinct functions in C. elegans: dauer induction, male attraction, and hermaphrodite social signaling (Figure 6B). Previous studies have shown that ascr's often have more than one function; ascr#3, for example, plays significant roles for both dauer signaling and male attraction [4],[7]. Our study demonstrates that specific structural variants of ascarylose-derived small molecules are associated with specific functions (Figure 6C). We show that addition of an indole carboxy group to ascr's changes the signaling properties such that the indole-modified compounds can have signaling effects very different from those of the unmodified compounds: icas#3 strongly attracts hermaphrodites and promotes aggregation, whereas ascr#3 repulses hermaphrodites and attracts males. In addition to structural variation, distinct signaling functions are associated with different concentration windows: whereas for dauer formation, high nanomolar concentrations of ascr's are required, low nanomolar to high picomolar concentrations of ascr's promote male attraction, and picomolar to femtomolar concentrations of icas' promote hermaphrodite attraction and aggregation (Figure 6B).
Social signaling in C. elegans thus appears to be based on a modular language of small molecules, derived from combinatorial assembly of several structurally distinct building blocks (Figure 6C). Different combinations of these building blocks serve different, occasionally overlapping signaling functions. Our results for the relative abundances of ascr's and icas' with identical side chains (Figure 1G) indicate that integration of the different building blocks is carefully controlled. Biochemically, the building blocks are derived from three basic metabolic pathways: carbohydrate metabolism, peroxisomal fatty-acid β-oxidation, and amino acid metabolism. These structural observations raise the possibility that social signaling via small molecules transduces input from the overall metabolic state of the organism. Food availability and nutrient content in conjunction with other environmental factors may control ascr and icas biosynthesis pathways to generate specific pheromone blends that differentially regulate aggregation, mate attraction, and developmental timing. The expansive vocabulary of a modular chemical language would make it possible for different nematodes to signal conspecifically as well as interspecifically, but it is not known whether nematode species other than C. elegans rely on ascarylose-based small molecules for chemical communication. However, lipid-derived glycosides of ascarylose have been identified from several other nematode species [39], suggesting that many nematodes have the ability to produce ascr- or icas-like compounds.
The identification of indole ascarosides as attraction and aggregation signals demonstrates that C. elegans aggregation behavior depends on dedicated chemical signals produced by conspecifics and not just shared preference for specific environmental conditions. C. elegans social signaling thus appears to be significantly more highly evolved than previously suspected.
NMR spectra were recorded on a Varian INOVA 600 NMR (600 MHz for 1H, 151 MHz for 13C). NMR-spectra were processed using Varian VNMR and MestreLabs MestReC software packages. Additional processing of bitmaps derived from NMR spectra was performed using Adobe Photoshop CS3 as described [6]. HPLC–MS was performed using an Agilent 1100 Series HPLC system equipped with a diode array detector and connected to a Quattro II spectrometer (Micromass/Waters). Data acquisition and processing was controlled by MassLynx software. Flash chromatography was performed using a Teledyne ISCO CombiFlash system.
All strains were maintained at 20°C unless mentioned otherwise on NGM agar plates, made with Bacto agar (BD Biosciences), and seeded with OP50 bacteria grown overnight. For the attraction bioassays and the automated tracker experiments, we used C. elegans var. N2 Bristol and males from the him-5(e1490) strain CB1490. The him-5(e1490) mutant segregates XO male progeny by X chromosome nondisjunction during meiosis [40]. For genetic ablation of the ASK neuron, we used the transgenic strain PS6025 qrIs2[sra-9::mCasp1], which expresses mammalian caspase in the ASK neuron under the influence of the sra-9 promoter (this strain is a kind gift of Tokumitsu Wakabayashi, Iwate University). Other strains used are as follows: CB4856, C. elegans Hawaii isolate [22]; RC301, C. elegans Freiburg isolate [10],[22]; DA609 npr-1(ad609); CX4148 npr-1(ky13) [10]; CX9740 C. elegans (N2); kyEx2144 [ncs-1::GFP] [9]; N2;Ex(gcy-28::dp::mec-4D) [28]; CX10981 kyEx2866 [“ASK::GCaMP2.2b” sra-9::GCaMP2.2b SL2 GFP, ofm-1::GFP] (ASK imaging line); CX11073 kyEx2916 [“AIA::GCaMP2.2b” T01A4.1::GCaMP2.2b SL2 GFP, ofm-1::GFP] (AIA imaging line) [9]; DR476 daf-22 (m130) [17]; and daf-22 (ok693) [16].
All newly identified ascarosides are named with four letter “SMID”s (Small Molecule IDentifiers)—e.g., “icas#3” or “ascr#10.” The SMID database (www.smid-db.org) is an electronic resource maintained by Frank Schroeder and Lukas Mueller at the Boyce Thompson Institute in collaboration with Paul Sternberg and WormBase (www.wormbase.org). This database catalogues newly identified C. elegans small molecules, assigns a unique four-letter SMID (a searchable, gene-style Small Molecule IDentifier), and for each compound includes a list of other names and abbreviations used in the literature.
Metabolite extracts were prepared according to a previously described method [6], which was modified as follows. Worms (N2 or daf-22) from three 10 cm NGM plates were washed using M9-medium into a 100 mL S-medium pre-culture where they were grown for 5 d at 22°C on a rotary shaker. Concentrated OP50 derived from 1 L of bacterial culture (grown for 16 h in LB media) was added as food at days 1 and 3. Subsequently, the pre-culture was divided equally into four 1 L Erlenmeyer flask containing 400 mL of S-medium for a combined volume of 425 mL of S-medium, which was then grown for an additional 10 d at 22°C on a rotary shaker. Concentrated OP50 derived from 1 L of bacterial culture was added as food every day from days 1 to 9. Subsequently, the cultures were centrifuged and the supernatant media and worm pellet were lyophilized separately. The lyophilized materials were extracted with 95% ethanol (250 mL 2 times) at room temperature for 12 h. The resulting yellow suspensions were filtered and the filtrate evaporated in vacuo at room temperature, producing media and worm pellet metabolite extracts.
The media metabolite extract from two cultures was adsorbed on 6 g of octadecyl-functionalized silica gel and dry loaded into an empty 25 g RediSep Rf sample loading cartridge. The adsorbed material was then fractionated via a reversed-phase RediSep Rf GOLD 30 g HP C18 column using a water-methanol solvent system, starting with 100% water for 4 min, followed by a linear increase of methanol content up to 100% methanol at 42 min, which was continued up until 55 min. The eight fractions generated from this fractionation were evaporated in vacuo. The residue was analyzed by HPLC-MS and 2D-NMR spectroscopy.
Worm media extracts or metabolite fractions derived from the chromatographic fractionation were resuspended in 1.5 ml methanol, centrifuged at 2,000 g for 5 min, and the supernatant submitted to HPLC-MS analyses. HPLC was performed using an Agilent 1100 Series HPLC system equipped with an Agilent Eclipse XDB-C18 column (9.4×250 mm, 5 µm particle diameter). A 0.1% acetic acid–acetonitrile solvent gradient was used, starting with an acetonitrile content of 5% for 5 min, which was increased to 100% over a period of 40 min. Mass spectrometry was performed with a Quattro II spectrometer (Micromass/Waters) using electrospray ionization in either negative or positive ion mode.
Axenic in vitro cultures of C. elegans (N2, Bristol) were established as described by Nass & Hamza [20], using the C. elegans Maintenance Medium (CeMM, [19]) with cholesterol (5 mg/l) instead of sitosterol and nucleoside-5-phosphates. After 21 d the cultures were centrifuged and the supernatant media and worm pellet were lyophilized separately. The lyophilized worm pellets (1.2–2.0 mg) were extracted with 2 ml methanol, filtered, and concentrated in vacuo. The lyophilized worm media were extracted with ethyl acetate–methanol (95∶5, 100 mL 2 times), filtered, and concentrated in vacuo. Residues were taken up in 150 µl methanol and investigated by HPLC-ESI-MS. For the application experiment 50 ml CeMM medium was supplemented with 9.2 mg L-[2,4,5,6,7-D5]-tryptophan (from Cambridge Isotope Laboratories).
These assays were done as previously described [6],[7]. For both C. elegans hermaphrodites and males, we harvested 50–60 worms daily at the fourth larval stage (L4) and stored them segregated by sex at 20°C overnight to be used as young adults the following day. For the competition experiments we used 120 nM ascr#3 and 10 nM icas#3 (Condition 1), or 12 µM ascr#3 and 1 µM icas#3 (Condition 2) in water containing 10% ethanol. Aliquots were stored at −20°C in 20 µL tubes. 10% ethanol in water was used as control.
Chemotaxis to both non-indole and indole ascarosides was assessed on 10 cm four-quadrant Petri plates [21]. Each quadrant was separated from adjacent ones by plastic spacers (Figure 2B). Pairs of opposite quadrants were filled with nematode growth medium (NGM) agar containing either indole ascarosides or non-indole ascarosides at different concentrations. Animals were washed gently in a S-basal buffer and placed in the center of a four-quadrant plate with ascarosides in alternating quadrants, and scored after 15 min and 30 min. A chemotaxis index was calculated as (the number of animals on ascaroside quadrants minus the number of animals on buffer quadrants)/(total number of animals).
Reversal frequency and velocity were measured using an automated worm-tracking system as previously described [6],[7].
We measured aggregation behavior of worms using assays described previously [10]. Aggregation assays were conducted on standard NGM plates. Plates containing indole ascarosides were made by adding the indole ascaroside stock solution to the NGM media before they were poured onto the plates. These plates were dried at room temperature for 2–3 d. Control plates were treated similarly except that instead of icas solutions ethanol solutions were added to the plates, corresponding to the amount of ethanol introduced via the icas solutions. Final ethanol concentrations of the plates were below 0.1% for all conditions. After drying, both control plates and plates containing indole ascarosides were seeded with 150 µl of an overnight culture of E. coli OP50 using a micropipette and allowed to dry for 2 d at room temperature. For “low worm density” experiments, we placed 20 worms onto the lawn and left them at 20°C for 3 h. For “high worm density” experiments we placed approximately 120 worms onto the bacterial lawn and left them at 20°C for 3 h. Aggregation behavior was quantified as the number of animals that were in touch with two or more animals at >50% of their body length.
For calcium imaging we used transgenic lines that express the genetically encoded Ca2+ sensor in ASK (kyEx2866) and AIA (kyEx2916) [9]. Young adults or adult worms were inserted into an “Olfactory chip” microfluidic device. [30]. Dilutions of icas#3 were done with S-basal buffer (with no cholesterol). As stock solutions of icas#3 contained small amounts of ethanol, equivalent amounts of ethanol were added to the S-basal control flow. Imaging was done using an inverted Zeiss microscope equipped with an Andor camera. Exposure time for image acquisition was 300 ms. Before imaging the ASK neuron, the worm was exposed to blue light for 3 min since ASK responds to the blue light itself. This step is necessary so that the neuron adapts to the blue light that is used for Ca2+ measurements. The movies were analyzed using custom-made Matlab scripts. For calculating the average change in fluorescence upon exposure to either buffer or icas#3, we chose the first peak of fluorescence immediately after exposure to buffer or icas#3. The value for this maximum was then subtracted from the mean fluorescence during the 5 s before the delivery of icas#3/buffer (corresponding to the region between 5 s to 10 s in Figure 5D).
Figures 2C,D, 3A,D, 6A, S3A, and S4C: We used unpaired Student's t tests with Welch's correction for comparing attraction of hermaphrodites and males on indole ascarosides *p<0.01, **p<0.001, ***p<0.0001. Figures 2E, 3B,C: For comparing the quadrant chemotaxis indices of the various strains, we used one-factor ANOVA followed by Dunnett's post-test, *p<0.05, **p<0.01. Figures 4A–C, S3C, S4A,B: For comparing aggregation of solitary, social worms and Cel-daf-22 on plates containing indole ascarosides, we used one-factor ANOVA followed by Dunnett's post-test, *p<0.05, **p<0.01. Figure 4D: To compare stopped duration of worms on plates with indole ascarosides, we used one-factor ANOVA followed by Dunnett's post-test, *p<0.05, **p<0.01. Figure S4D,E: To compare velocities and reversal frequencies on plates with indole ascarosides, we used one-factor ANOVA followed by Dunnett's post-test, *p<0.05, **p<0.01. Figure S5A,B: To compare reversals between unablated and ASK ablated lines, we used Student's t tests with Welch's correction, *p<0.01, **p<0.001. Figure 5B: To compare the attraction of wild-type worms to the genetically ablated lines for ASK and AIA as well as the ASI and RMG neuron ablations, we used unpaired Student's t test with Welch's correction, ***p<0.0001. Figure 5E: For comparing G-CaMP fluorescence changes to buffer and icas#3, we used unpaired Student's t test with Welch's correction, **p<0.001. All error bars indicate standard error of mean (S.E.M).
Samples of indole ascarosides icas#1, icas#7, icas#3, and icas#9 for use in biological assays and as standards for HPLC-MS were prepared via chemical synthesis. Detailed procedures and NMR-spectroscopic data are contained in Text S1.
|
10.1371/journal.pgen.1001144 | Gene–Environment Interactions at Nucleotide Resolution | Interactions among genes and the environment are a common source of phenotypic variation. To characterize the interplay between genetics and the environment at single nucleotide resolution, we quantified the genetic and environmental interactions of four quantitative trait nucleotides (QTN) that govern yeast sporulation efficiency. We first constructed a panel of strains that together carry all 32 possible combinations of the 4 QTN genotypes in 2 distinct genetic backgrounds. We then measured the sporulation efficiencies of these 32 strains across 8 controlled environments. This dataset shows that variation in sporulation efficiency is shaped largely by genetic and environmental interactions. We find clear examples of QTN:environment, QTN: background, and environment:background interactions. However, we find no QTN:QTN interactions that occur consistently across the entire dataset. Instead, interactions between QTN only occur under specific combinations of environment and genetic background. Thus, what might appear to be a QTN:QTN interaction in one background and environment becomes a more complex QTN:QTN:environment:background interaction when we consider the entire dataset as a whole. As a result, the phenotypic impact of a set of QTN alleles cannot be predicted from genotype alone. Our results instead demonstrate that the effects of QTN and their interactions are inextricably linked both to genetic background and to environmental variation.
| Phenotypic variation among individuals is caused by naturally occurring genetic differences, or alleles. The relationship between an allele and the phenotype is extremely complex; for example, the effect of an allele often depends upon both the environment and the individual's genetic background. To better understand these complex relationships, we examined the effects of four quantitative trait nucleotides (QTN) in three genes that cause variation in sporulation efficiency between vineyard and oak tree strains of yeast. We measured the effects of the QTN while varying both the genetic makeup of the strains and their growth environments. We found that the effects of each of the four QTN alleles depended upon the genotypes at the other QTN, the growth environment, and whether the strain carried the oak or vineyard parent genome. There were no simple rules that describe the effects of the alleles across all environments; instead, detailed models were needed to account for environmental and genetic variation in order to predict the effects of alleles in specific individuals.
| As we identify more genetic loci that underlie complex traits, the challenge remains to understand and predict the effects of the causal genetic variants upon individuals' phenotypes. The relationship between genotype and phenotype is rarely simple. The effect of an allele often depends upon the environment, resulting in gene-environment interactions (GxE). GxE is a well-documented occurrence in many species, including humans [1]–[5]. Gene-gene interactions also take place that render the effect of one locus dependent upon the genotype at another locus. Genetic interactions can occur between characterized loci (epistasis) [6], [7], or between one known locus and other unknown loci (genetic background effects) [8]. If individuals vary in their environmental exposure and genetic makeup, as they almost always do in nature, then GxE and genetic interactions will create differences in the effects of alleles among individuals.
Therefore, to understand allelic effects, we must also understand the scope and prevalence of genetic and environmental interactions. However, standard approaches for the identification of causative loci, such as association analysis and linkage mapping [9], measure the average effects of alleles in populations. Without very large sample sizes, population averages cannot account for potential individual-to-individual variation created by complex interactions [10]. Some study of interactions on an individual-to-individual basis has occurred through the use of near isogenic lines [11], but there are still few examples that illustrate the impact of interactions from one individual to the next at the resolution of single-nucleotides [7], [12].
To better understand the effects of GxE and genetic interactions at the resolution of single nucleotides, we took advantage of four naturally occurring quantitative trait nucleotides (QTN) known to cause variation in yeast sporulation efficiency [13], [14]. We engineered allele replacement strains that carry all possible combinations of these QTN in two genetic backgrounds, and we then systematically measured the phenotypes of these strains in eight environments. Our results provide a detailed picture of how segregating QTN, environmental variation, and genetic background all combine to shape variation in a quantitative trait through complex relationships.
Our phenotype of interest, yeast sporulation, is a cell fate decision executed by diploid yeast cells in response to a shift from fermentative to respiratory conditions [15]. Yeast cells switch to primarily aerobic respiration when faced with only a non-fermentable carbon source. When this environmental change is accompanied by a reduction in a critical nutrient such as nitrogen, a fraction of yeast cells in a culture will initiate meiosis and enclose the meiotic products in a protective spore wall.
Our QTN all affect the proportion of cells in a culture that initiate meiosis (the sporulation efficiency) after a shift from glucose (fermentable) to acetate (non-fermentable) media. The QTN include a coding polymorphism in RSF1 (a positive regulator of respiration) [16], both coding and non-coding polymorphisms in IME1 (the master regulator of sporulation) [17], and a non-coding polymorphism in RME1 (a direct repressor of IME1) [18], [19]. Each of these genes encodes a transcription factor.
Each QTN has two alleles: a reference allele found in the wild oak tree isolate YPS606, and an allele that reduces sporulation efficiency from the vineyard isolate UCD2120 [14]. (In the rest of this article, we denote the QTN with the labels: rsf1, rme1, ime_coding, and ime_nc.) Both the patterns of phenotypic variation in sporulation efficiency and the sequence variation of the causal genes indicate that sporulation efficiency is subject to purifying selection in oak strains and disruptive selection in vineyard strains [14], [20]. The change in phenotype caused by the QTN therefore represents genotype-phenotype variation that has occurred due to a shift in selective pressures between two habitats.
To broadly test for GxE effects, we first measured the sporulation efficiency of each parent strain genetic background (designated oak and vineyard) carrying two QTN genotype combinations: either all the QTN alleles of the oak parent, or all the QTN alleles of the vineyard parent. We generated environmental variation by growing the strains in eight different fermentative media conditions (Table 1) prior to the induction of sporulation in acetate (see Methods). In all eight environments and across both genetic backgrounds, the oak QTN alleles collectively increase sporulation efficiency, and the vineyard QTN alleles collectively decrease sporulation efficiency (Table 2). However, the environments vary with respect to the proportion of the phenotype that the QTN explain (Table 3). For example, in grape juice, we can explain 99% of the phenotypic difference between the parents by placing the vineyard QTN alleles into the oak background. However, in raffinose, the same allele replacement explains only 55% of the parental difference. The phenotypic difference explained by the QTN also depends on the genetic background. For example, placing the oak QTN alleles into the vineyard background explains 90% of the difference between the parent strains in raffinose, but we explain only 55% of the difference between the parents if we conduct the reciprocal experiment that places the vineyard QTN alleles into the oak background. Because the phenotypic difference created by the QTN varies across both the environments and genetic backgrounds, our results imply genetic interactions among the QTN, the environmental treatments, and uncharacterized loci in the two parent genetic backgrounds.
To further investigate the extent of these interactions, we measured the phenotypes of strains with all 16 possible QTN genotype combinations in both genetic backgrounds (32 total strains). We calculated a correlation matrix of the eight environments from their effects on the phenotype rank-order of the 32 strains so that we can broadly compare the QTN effects across environmental treatments and genetic backgrounds (Table 4). Because the QTN alleles always act in the same direction regardless of condition, all the environments were positively correlated (Spearman's ρ = 0.69 to 0.99). The differences in correlations therefore reflect changes in the rank order (and therefore relative magnitude) of QTN effects.
We used hierarchical clustering to construct a dendrogram that reflects the correlations between environments (Figure 1). Sucrose, fructose, and glucose are the most highly correlated environments (Spearman's ρ>0.99 for all pair wise comparisons) and cluster closely. We did not detect significant differences between these three environments in either genetic background. Their values were therefore pooled and averaged as “glucose-like” (YGlu) for all subsequent analyses.
Maltose and raffinose cluster separately from YGlu and are slightly less correlated with glucose (ρ = 0.93, 0.96, respectively). Both the oak and vineyard genetic backgrounds sporulate more efficiently in raffinose than in YGlu. (Figure 2A) The effect of maltose, however, depends upon the genetic background (Figure 2B). Sporulation efficiency of the vineyard background is similar in maltose and YGlu, but the oak background sporulates more efficiently in maltose. Therefore, there is an interaction between the genetic background and maltose.
Galactose also shows a background:environment interaction. The oak background sporulates similarly in galactose and YGlu, but the vineyard background sporulates more efficiently in galactose (Figure 2C). In Figure 1, galactose clusters distinctly from all other environments. When we run the clustering algorithm separately for each genetic background, this separation disappears (Figures S1, S2). Therefore, the disparity of galactose relative to the other environments appears to result from the background:environment interaction.
Synthetic oak exudate and grape juice also cluster distinctly from the other environments, but these two conditions are highly correlated with each other (ρ = 0.97). The sporulation efficiency of both genetic backgrounds tends to be lower in exudate and grape juice than in the other environments (Figure 1, Tables S1, S2). There are also QTN:environment interactions that occur in both exudate and grape juice relative to YGlu. For example, in the oak background, the rsf1 vineyard allele has a much larger effect in exudate and grape juice than it does in YGlu (Figure 3). To quantify this difference in QTN effect, we constructed a linear model of sporulation efficiency in the oak background that incorporates the main effects of single QTN, the effects of the environments, and the QTN:environment interactions (see Methods). In this model, the differential effect of rsf1 is manifested as a QTN:environment interaction in exudate and grape juice relative to YGlu (exudate: effect = −34±2%, t-test, P<2e-16; grape juice: effect = −25±2%, t-test P<2e-16; all errors reported in the text are the standard errors of coefficient estimates). The effect of rsf1 is the largest of any single QTN in both grape juice and exudate (Tukey's HSD, maximum adjusted P = 0.002). However, in YGlu the rsf1 QTN does not even have a significant main effect in the oak background (effect = −1.8±1%, t-test P = 0.08).
The effect of rsf1 in the vineyard background reveals a different story. In the vineyard background, the rsf1:environment interaction is not significant in exudate or grape juice (exudate: effect = 4.8±2.8%, t-test P = 0.1; grape juice: effect = −0.3±0.03%, t-test, P = 0.92). Instead, rsf1 has a large main effect in YGlu as well as exudate and grape juice (Figure 3). Therefore, the effect of rsf1 can be best explained as an environment:rsf1:background interaction that reduces the effect of rsf1 in YGlu relative to exudate and grape juice, but only in the oak strain background.
How do the effects of these environment and background interactions compare with the role played by QTN:QTN interactions? We previously demonstrated significant QTN:QTN epistasis in the oak background and the glucose environment [14]. In that context, epistasis appears to play a large role in shaping phenotypic variation. However, the differences in rsf1's effect across backgrounds and environments imply that the QTN:QTN interactions might occur only in certain environments or backgrounds. We therefore tested for all possible QTN:QTN interactions across all eight environments and both genetic backgrounds.
To do so, we modeled variation in sporulation efficiency in a standard linear framework using all phenotypic measurements across QTN genotypes, genetic backgrounds, and environments (see Methods). A completely saturated model that incorporates all possible effects and interactions between environment, background, and QTN has an adjusted R2 of 0.99. All of the parameters in the model are controlled variables, so this R2 indicates that 1% of the variation in our experiment is due to experimental error. We then constructed a reduced model that explains most of the variation, but with fewer parameters and only two and three-way interactions (Figure 4, adjusted R2 = 0.963, see Methods). This model (the global model) captures the predominant interactions in the data (Table S3). For example, it contains a significant positive interaction term between galactose and the vineyard background (effect = 25.6±2.4%, t-test P<2e-16). This term is expected given the higher sporulation efficiencies we observe in the vineyard background in galactose (Figure 2C). There are also significant interactions between rsf1 and both exudate (effect = −28±2.8%, t-test P<2e-16) and grape juice (effect = −13.8±2.8%, t-test P = 9e-7), which are expected due to the larger effect of rsf1 in these two conditions (Figure 3).
The most striking result from the global model is the lack of two-way QTN:QTN interactions. Three QTN:QTN interaction terms were left in the model after stepwise regression (Table S3). Only one of these, a negative interaction between rsf1 and ime1_coding (effect = −7.6±1.8%, P = 3.9e-5), passed either Bonferonni correction or permutation testing. This result stands in contrast to what we observe within a single condition. In line with our previous data in glucose [14], we find abundant QTN:QTN interactions when YGlu is modeled alone (Table S4). For example, the rme1:ime1_coding interaction is large (effect = −29.4±2.3%, t-test P<2e-16). However, when all the environments and both backgrounds are analyzed together in the global model, the same rme1:ime1_coding interaction is small and only marginally significant (effect = −3.9±1.6%, t-test P = 0.02).
In the place of QTN:QTN interactions, the global model contains several significant three-way QTN:QTN:environment and QTN:QTN:background interactions. This suggests that significant QTN:QTN interactions cause variation in sporulation efficiency, but the interactions only occur in particular environments and genetic backgrounds. To examine this possibility further, we modeled each environment-background combination separately and observed that QTN:QTN interactions varied widely. For example, the rme1:ime1_coding interaction that is strong in YGlu in the oak strain is marginal in exudate (Figure 5, YGlu effect = −29.4±2.3%, t-test P<2e-16 ; exudate effect = −4.2±2.1% , t-test P = 0.052). This interaction is present in maltose (effect = −29±3.9%, t-test P = 1.6e-8), but not in the vineyard background (Figure 6, effect = 0.00±2.1%, t-test P = 0.23). Taken together, these results show that the vineyard alleles of rme1 and ime1_coding act synergistically in the oak strain and specifically in YGlu and maltose, as the combination of two vineyard QTN produces a larger change in phenotype than could be expected from their individual effects. However, in exudate, or in the vineyard background, the effects of these same QTN alleles remain independent. Therefore, the synergistic interaction between the vineyard alleles is not intrinsic to the alleles themselves, but instead depends upon the specific context of the environment and genetic background.
Our measurements of sporulation efficiency therefore indicate that QTN:QTN interactions are not widespread, but QTN:QTN:environment and QTN:QTN:background interactions are common. In a linear model that ignores genetic background and environment, no interactions between the QTN are significant (adjusted R2 = 0.4). This QTN-only model correctly identifies that individuals with all vineyard alleles tend to sporulate poorly, but it does not provide the ability to accurately predict the phenotypes of individuals with intermediate genotypes (Figure 7). Ultimately, the effects of the QTN and their interactions are shaped by the environmental and genomic context in which they occur. Knowledge of the environment and genetic background is therefore crucial to accurately predict the effects of QTN across individuals (compare Figure 4 to Figure 7).
In this set of experiments, we measured sporulation efficiency in a variety of isogenic strains that differed with respect to QTN genotypes, genetic background, and growth environment. Overall, our results show that a complex set of genotype:environment:background interactions shape variation in sporulation efficiency. Our results also shed light on the general effects of environment on sporulation efficiency in the context of natural variation. We found that carbon sources with similar effects on yeast catabolite repression tended to have similar effects on sporulation efficiency. For example, glucose and fructose both cause strong catabolite repression in yeast [21], and their effects on sporulation efficiency are highly correlated (Table 4). Sucrose, a disaccharide composed of glucose and fructose, is likewise highly correlated with glucose. Raffinose and galactose, which cause weaker catabolite repression [22], cluster less closely with glucose. One surprising result is the GxE we observed in maltose relative to glucose (Figure 2B). Since maltose is composed of two glucose molecules, one might expect the effect of maltose to be as similar to glucose as that of sucrose or fructose. One possible explanation for the GxE in maltose arises from the fact that maltose catabolism genes commonly display copy number variation among yeast isolates [23]–[25]. We observed a slow growth phenotype of the oak strain in maltose and mapped this phenotype to the MAL1 multigene locus (K. Lorenz and B. Cohen, unpublished results). We suspect that this locus is responsible for the maltose:background interaction we observe for sporulation efficiency, and it may also modulate the QTN effects and QTN:QTN interactions in maltose, but confirmation of this hypothesis awaits the cloning of the causative polymorphism.
Exudate and grape juice produce lower sporulation efficiencies than the other environments. This result occurs in spite of the fact that exudate is composed of exactly the same ingredients as YGlu, but with reduced concentrations of peptone and yeast extract. This reduction of nutrient concentrations not only reduces sporulation efficiency in both genetic backgrounds, but it also alters the effect of rsf1 in the oak background relative to the other QTN (Figure 3). The fact that exudate consists of the same ingredients as YGlu but produces different effects on sporulation efficiency suggests that QTN effects are shaped not only by nutrient type, but also by nutrient concentrations. Drops in nitrogen concentration are well-known to strengthen the signal to sporulate, so the difference in peptone concentration between exudate and rich media may explain some the differences in sporulation efficiency through nitrogen sensing.
Across multiple environments, the unknown polymorphisms in the genetic background not only interact with the environment but also alter the effects of the known QTN. The known QTN used in this study were mapped in glucose and explain ∼90% of the segregating variation in that condition [14]. The interactions we observe here suggest that the remaining unmapped loci may have stronger effects (and be easier to map) in non-YGlu environments. For example, the known QTN only explain half of the phenotypic difference in the vineyard background in grape juice (Table 3). Presumably, the remaining unknown polymorphisms that regulate sporulation efficiency have larger effects in this environment-background combination than they do in YGlu. An attractive experiment to identify new QTN governing sporulation efficiency would therefore be to map the phenotype in grape juice using a cross of the original oak parent with a new version of the vineyard parent strain that is fixed for all four known oak QTN. It is possible, however, that the new polymorphisms uncovered by this experiment would not reside in the sporulation pathway per se, but would instead be metabolic factors specific to grape juice catabolism.
Despite the fluctuations in QTN effects across environments and backgrounds, the direction of QTN effects remain consistent. Vineyard alleles always decrease sporulation efficiency relative to oak alleles. Without accounting for changes in the environment or differences in genetic background, we can therefore safely predict that a strain with all four vineyard alleles will sporulate poorly relative to a strain carrying all oak alleles. However, because the effect magnitudes of the QTN change across environments and backgrounds, we cannot predict the sporulation efficiency of intermediate allelic combinations (Figure 7). This case reminds us of the situation unfolding in human association studies, where it appears that high-risk individuals can be identified as carriers of collections of disease associated polymorphisms, even though it is more difficult to predict the actual phenotypic outcome of a particular individual with intermediate sets of alleles [26]. In this case of yeast sporulation efficiency, complexity occurs because the relative importance of particular alleles and their interactions are not constant across individuals, but instead vary with the individuals' genetic background and environment.
If context dependencies on allelic effects are common, how can we achieve better predictive power when environment and background are unknown? Environment and genetic background presumably influence the phenotype just as all genetic changes must: through effects on cell physiology. It might be possible to account for the physiological effects of environment and background using a biomarker or physiological indicator that is correlated with, but upstream of, the phenotype of interest. Biochemical markers are used in medicine to inform calculations of disease risk and diagnosis [27]. Inclusion of a physiological marker into the genetic model may condition the model to unknown parameters and therefore increase the accuracy of genotype-phenotype predictions.
Although such a model could improve predictive power, it still does not increase our understanding of how various physiological forces in the cell combine to quantitatively alter phenotype. Perhaps improved understanding could arise from interpreting QTN effects through a framework rooted in cell biology and biochemistry, rather than through an abstract linear model. Biochemical and gene regulatory pathways have long been theorized to naturally generate non-linear effects through the basic thermodynamic properties of proteins and DNA [28], [29]. We have modeled sporulation efficiency in glucose through a thermodynamic framework, and this method shows promise in revealing the molecular basis of genetic interactions [30]. However, thermodynamic modeling requires detailed knowledge of molecular mechanism of the proteins involved, and this information is not available for most traits. Also, the challenge of applying this approach to multiple environments is nontrivial [31].
A more traditional method to deal with statistical interactions is to eliminate them through data transformations. We experimented with a number of scale transformations for our dataset, but found that the best transformation for reducing the complexity of the interaction terms varied from one environment:background combination to the next. Furthermore, data transformations that reduced the number of interaction terms sometimes had undesirable effects, such as increasing the dependence of the variance upon the mean. More importantly, scale transformations that worked well on some subsets of the data still required numerous interaction terms to provide a global model. None of the data transformations we tried improved the three-way interaction fit obtained on the natural scale (Figure S3). Although data transformations may be appropriate to obtain simpler predictive models in single background:environment combinations, they do not account for the non-linear dynamics that create complexity across conditions and backgrounds.
Regardless of the approach taken in the future, our results clearly show that the genetic architecture of sporulation efficiency is environment-dependent. QTN effects cannot be understood without taking into account contextual factors such as the environment's influence on cell physiology. We expect that quantitative biochemical measurements will be required to illuminate what is happening inside the cell and bridge the missing link between genotype and phenotype.
Each of the 32 strains were grown for 15 hours in growth media (except for grape juice, in which we instead grew the yeast for 54 hours). After the growth period, we diluted each culture 1∶50 into 1% potassium acetate to induce sporulation. We tested three replicates of each QTN genotype - environment - genetic background combination. One exception is the strain carrying only the ime_nc vineyard QTN allele in the vineyard background grown in sucrose, for which there were only two measurements due to a sample failure. The experimental design is balanced such that the genotype frequencies of the four QTN do not vary across environments or backgrounds, so any significant interactions between QTN reflect physiological effects rather than differences in allele frequency [32]. Sporulation efficiency was calculated by flow cytometry on samples of 15,000 cells per replicate using methods we have described elsewhere [20]. The raw data of sporulation efficiencies for each replicate is available as a supplementary data file (Dataset S1).
Each of the eight environmental treatments was composed of a different growth medium prior to the induction of sporulation in acetate (Table 1). Six of the environments consisted of rich yeast media (1% yeast extract, 2% peptone) supplemented with 2% of a sugar or polysaccharide: glucose, fructose, sucrose, maltose, galactose, or raffinose. The other two environments were synthetic oak exudate and chardonnay grape juice. Synthetic oak exudate is composed of the same nutrients as rich media, but contains yeast extract and peptone at ten-fold reduced concentrations (Table 1). Exudate also contains a mixture of fructose, sucrose, and glucose at a total concentration of 2% [33]. After each environmental treatment, sporulation was induced for 30 hours in 1% potassium acetate, which provides a non-fermentable carbon source but no source of nitrogen.
First, we created allele replacement strains in each parental background that carry single QTN alleles from the opposite parent [34]. These strains were created by backcrosses of initial haploid ura3− allele replacement transformants with their prototrophic diploid parents. Ura3+ progeny from the backcross of each allele replacement were then intercrossed to generate strains carrying multiple QTN alleles from the opposite parent. Each cross was performed in triplicate. We confirmed after each cross that the QTN co-segregated with variation in sporulation efficiency in glucose, and we also ensured that the phenotypes resulting from replicate crosses were identical. This assured us that no new mutations governing sporulation efficiency had arisen elsewhere in the genome during the crossing scheme. Once a strain with the desired QTN alleles from the opposite parent was created, this strain was backcrossed once more to its original wild type parent strain. Individual homothallic diploid progeny from this final cross were isolated and genotyped until we obtained three replicates of every possible QTN allele combination. Genotyping was based on the restriction digest of PCR amplicons [14]. The selected strains were arrayed in a 96-well plate such that all the strains from both genetic backgrounds can be assayed in a single block.
We generated a matrix of the Spearman rank correlations of the means of each of the 32 strains across each environment. A distance matrix was then defined as 1−ρ, where ρ is the matrix of pair wise Spearman rank correlations. We carried out hierarchal cluster analysis with the complete linkage clustering method as implemented in the hclust function in the statistical package R. We also split the data by genetic background, then calculated rank correlations and clustered separately for the oak and vineyard genetic backgrounds.
All statistical analyses were performed in R. In all linear models, the strain with all oak QTN alleles was treated as the intercept, so the additive effects represent the effect of a single vineyard QTN placed into a strain with oak QTN alleles at all other loci. We chose this reference point because the oak strain probably best resembles the genotype of the common ancestor of the two parent strains [14], [35]. To compare the effects of QTN:QTN interactions in single environment-background combinations, we created linear models of QTN effects including all possible interaction terms within each condition, and significant coefficients were calculated by t-tests of the coefficient's estimated effect versus its standard error. All interaction terms reported in the text are significant by Bonferonni correction (P = 0.05/N, where N is the number of coefficients in the model).
To analyze QTN effects across all environments and both backgrounds, we constructed a linear model in which the oak genetic background, oak QTN alleles, and the glucose environment are treated as intercepts. Therefore, coefficients in the model represent the effects of the vineyard genetic background, vineyard QTN alleles, and non-glucose environments. The simplest additive model therefore takes the following form:Where EFF is sporulation efficiency, Oak is the oak strain phenotype in glucose (the y-intercept), BG is the effect of genetic background, RME, RSF, IMEC, and IMENC are the effects of the vineyard QTN alleles, ENV is the effect of non-glucose environments, and e represents the error across the multiple replicates of each combination of strain and environmental treatment. Sucrose and fructose were not significantly different from glucose, so these three conditions were pooled into a single treatment. We found that models with increasing levels of interaction terms were often significant, but very little improvement to the fit or explanatory power of the model was gained by adding four-way interactions (Table S5). We therefore limited our analysis to models with three-way interaction terms to reduce saturation without much sacrifice of explanatory power. To select a specific model with a subset of the three-way terms, we used stepwise regression as implemented in the stepAIC function in R. We then took the output from stepwise regression and manually removed terms from the model if their treatment contrast P-values did not pass a model-wide Bonferonni correction. Table S3 displays the coefficients in our final model and the P-values of each coefficient. The significance of the QTN:QTN interactions in this model were also tested by creating 10,000 null linear models from random permutations of the entire dataset. The critical P-value from these permutations was P = 0.016.
Probit and Logit transformations, which are common used for frequency data, provide good fits with fewer interaction terms in some individual conditions. However, we chose to model the data on the raw scale. The Probit and Logit transformations obtain a fit by weighting the explanatory power at extremely high and low values of sporulation efficiency at the expense of intermediate values (Figure S3). For example, under the Probit transformation, a difference in sporulation efficiency from 1% to 2% is as great in magnitude as a raw difference of 40% to 50%. No transformation eliminated interactions altogether, and transformations did not improve the overall fit of the model across multiple environments. The raw scale allows more intuitive interpretation of the model coefficients, and our reduced model performs well on values of sporulation efficiency between ∼5 and 95%. Some extreme values are fit below zero or above 100%. However, with one exception (a data point at 92%), all data points fitted to higher than 100% have actual values greater than 96%. All data points predicted to be below zero have actual values less than 5%. We therefore simply bounded all predicted values between 0% and 100%.
To model the specific QTN:environment interactions in exudate and grape juice, we conducted an analysis of variance on only the additive effects (no QTN:QTN interactions) of the four vineyard QTN separately in each genetic background. This model focused on the additive effects because the phenotypes of vineyard background strains carrying multiple vineyard QTL approach zero in a non-linear fashion (Tables S1, S2). The model took the form:Where EFF is sporulation efficiency, GEN is the genotype across the four QTN alleles, ENV is the environment, and e is the error. To confirm significant differences in the rank order of QTN effects in different environments, we took the estimated QTN effects from an analysis of variance in each environment separately and computed Tukey's Honest Significant Difference to determine the rank-order of the QTN within each environment. The reported P value is the largest adjusted P value among all the possible comparisons between the effect of rsf1 and the effects of other QTN.
|
10.1371/journal.pgen.1003509 | Imputation-Based Meta-Analysis of Severe Malaria in Three African Populations | Combining data from genome-wide association studies (GWAS) conducted at different locations, using genotype imputation and fixed-effects meta-analysis, has been a powerful approach for dissecting complex disease genetics in populations of European ancestry. Here we investigate the feasibility of applying the same approach in Africa, where genetic diversity, both within and between populations, is far more extensive. We analyse genome-wide data from approximately 5,000 individuals with severe malaria and 7,000 population controls from three different locations in Africa. Our results show that the standard approach is well powered to detect known malaria susceptibility loci when sample sizes are large, and that modern methods for association analysis can control the potential confounding effects of population structure. We show that pattern of association around the haemoglobin S allele differs substantially across populations due to differences in haplotype structure. Motivated by these observations we consider new approaches to association analysis that might prove valuable for multicentre GWAS in Africa: we relax the assumptions of SNP–based fixed effect analysis; we apply Bayesian approaches to allow for heterogeneity in the effect of an allele on risk across studies; and we introduce a region-based test to allow for heterogeneity in the location of causal alleles.
| Malaria kills nearly a million people every year, most of whom are young children in Africa. The risk of developing severe malaria is known to be affected by genetics, but so far only a handful of genetic risk factors for malaria have been identified. We studied over a million DNA variants in over 5,000 individuals with severe malaria from the Gambia, Malawi, and Kenya, and about 7,000 healthy individuals from the same countries. Because the populations of Africa are far more genetically diverse than those in Europe, it is necessary to use statistical models that can account for both broad differences between countries and subtler differences between ethnic groups within the same community. We identified known associations at the genes ABO (which affects blood type) and HBB (which causes sickle cell disease), and showed that the latter is heterogeneous across populations. We used these findings to guide the development of statistical tests for association that take this heterogeneity into account, by modelling differences in the strength and genomic location of effect across and within African populations.
| Severe malaria, meaning life-threatening complications of Plasmodium falciparum infection, kills on the order of a million African children each year [1]. However this represents only a small proportion of the total number of infected individuals, the majority of whom recover without life-threatening complications. Understanding the genetic basis of resistance to severe malaria could provide valuable insights into molecular mechanisms of pathogenesis and protective immunity that will aid the development of treatments and vaccines. It might also identify selective pressures that have shaped human physiology and susceptibility to other common diseases, because of the historical impact of malaria as a major cause of mortality in ancestral human populations.
Genome-wide association studies (GWAS) have identified thousands of genetic variants which predispose individuals to particular disease phenotypes. However, the vast majority of these studies are of non-communicable disease in collections of individuals with European ancestry. The challenges of applying these approaches to studying disease in Africa are well documented [2]; the long ancestral history of African populations has two consequences. Firstly it has led to an overall reduction in the correlation (linkage disequilibrium) between alleles at neighbouring loci. Secondly it has given rise to differences in the combinations of alleles along chromosome (haplotypes) both between, and within, geographically defined populations.
The first of these complications is problematic because GWAS rely on the correlations between causal mutations and genotyped markers to identify susceptibility variants. From a statistical perspective, unless the causal marker is typed directly, the reduced linkage disequilibrium acts to dilute association signals [3], making it hard to distinguish real effects on disease risk from apparent effects that arise from sampling. In theory this loss of power can be overcome simply by increasing the sample size or the number of typed markers [3]. Another way in which GWAS critically rely on correlation among nearby variants is via imputation based meta-analysis, which has proven a powerful tool for combining information across collections of individuals with similar ancestry. These approaches work by first obtaining genotypes at a common set of loci and then combining the statistical evidence at each locus, across collections, often by assuming the alleles to have a consistent, or fixed, effect on susceptibility. However, the differences in haplotype structure in Africa means that the correlation between any given marker locus and the causal allele may vary from one collection to the next, so the apparent effect on risk may be heterogeneous. The utility of applying the same methodology for meta-analysis in African populations is therefore unclear.
Here we describe the first imputation based meta-analysis approach in Africa to study severe malaria susceptibility across multiple African populations. We use data from large samples of individuals collected from Kenya, Malawi and Gambia which together included 5425 individuals with severe malaria and 6891 population controls. As each of the collections was typed on a different Illumina genotyping array, we used imputation (using the program IMPUTE2 [4]) to obtain genotypes at a common set of 1.3 million SNPs. We use these data to investigate the accuracy of the imputation, assess genetic structure within and between populations, describe association patterns at loci known to influence risk of severe malaria, and investigate methods for identifying new susceptibility loci.
MalariaGEN partners in Gambia, Malawi and Kenya collected blood samples from children diagnosed with severe malaria, including cerebral malaria and severe malarial anaemia. As population controls, cord blood samples were collected from the same geographic areas as the cases. Ethical approval was obtained from each ethics committees at each of the partner study sites and institutions (Table S1). DNA was extracted from the blood samples and assayed at SNPs genome-wide on Illumina arrays. There are challenges specific to collecting blood from children in Africa, particularly if they are ill with severe malaria, making it hard to obtain large quantities of DNA. We performed whole-genome amplification (WGA) on a small quantity of DNA before array-based genotyping to preserve samples. The process of WGA can lead to additional experimental noise, potentially leading to genotype calling errors. To produce a robust set of genotype calls, we used three different calling algorithms to process intensity data from the Illumina arrays, separately in each of the three cohorts. A set of consensus calls was obtained by treating as missing any genotype that was discordant among algorithms (see Methods).
As high levels of missing (or discordant) genotypes are indicative of poor genotype calling (either due to poor assay performance or sporadic genotype clustering errors) we excluded SNPs with >2.5% missing genotype calls. Using the three-way calling in this way provided a robust set of SNPs for analysis and very little additional filtering of SNPs was required (Table S3).
Prior to imputation we excluded samples with outlying levels of missing or heterozygous genotypes as well as one of each pair of duplicate samples (see Table S2). Preliminary analysis of the data revealed a subset of control samples in the Malawi cohort which showed sporadic assay failure at a small number of SNPs across the genome. This type of error is hard to identify prior to analysis, and we provide a description of our observations in Text S1 and Figure S1 in case it is helpful to readers with similar data.
Imputation is now a well-established strategy for exploiting densely genotyped reference panels to infer genotypes at SNPs not assayed directly in a given study. For each study collection we thinned the set of SNPs to just those that passed quality control filters and were present in the HapMap3 haplotype panel [5] made available for use with the imputation program IMPUTE2 (http://mathgen.stats.ox.ac.uk/impute/impute_v2.html). We ran IMPUTE2 separately on each collection using the entire HapMap3 reference panel to obtain genotypes at a common set of 1.3 million SNPs (see Methods).
The accuracy of the genotype imputation typically depends on the correlation between typed and untyped SNPs and the similarity of haplotypes in available reference panels to those in the study samples. In both regards imputation in Africa is more challenging than in non-African populations [6]. Our study provides an opportunity to quantify the utility of imputation in this setting, and illustrates a number of issues that are relevant to other imputation-based studies in African populations.
We assessed the accuracy of imputation by comparing the genotype calls obtained from the three-way calling described above to the probabilistic estimates for those same SNPs produced by IMPUTE2 (type 2 r2 in IMPUTE2). Figure 1 shows per-individual imputation accuracies broken down by country. While less accurate than typically achieved using similar genotyping arrays in European populations [7], imputation still captures the majority of common variation in these three populations (a mean type 2 r2 of 0.93 in Malawi, 0.92 in Kenya and 0.87 in Gambia). Common SNPs were better imputed than low-frequency SNPs, suggesting that this analysis, much like similar experiments in Europeans, will be relatively less well powered to detect associations at low frequency SNPs.
Two different types of inaccuracy are possible in imputed data: overconfidence in predicting an incorrect genotype, or underconfidence in predicting a correct genotype. We therefore evaluated the calibration of the confidence of IMPUTE2 (measured by the info score) against its actual performance at genotyped SNPs. The calibration of confidence was high across our three samples (r2 between predicted and true accuracy: 0.93 in Malawi, 0.92 in Kenya, 0.96 in Gambia) but, like overall accuracy, on average worse than in European samples (0.96). Quality scores were less well calibrated for low-frequency variation, but still remained relatively high across all three populations (89%, 84% and 92% respectively for variants with MAF<0.05). We included only SNPs with info score>0.75 for downstream analyses, leaving a high quality set with mean r2>0.9 in all samples, and less than 1% of either very overconfident (predicted r2>0.75, actual<0.6) or very underconfident (predicted<0.75, actual>0.9) SNPs. Taken together, these results suggest the underlying model of IMPUTE2, combined with a diverse reference panel provided by the HapMap project, is generally applicable to samples from African populations.
Despite the high performance of imputation overall, there were a number of factors that influenced relative imputation performance, including (i) genotyping platform, (ii) ethnic matching of target GWAS samples to the imputation reference panel, and (iii) homogeneity of individual GWAS collections. Our Gambian samples (typed on the Illumina 650Y array) show much poorer imputation quality (Figure 1) than our Kenyan and Malawian samples (typed on Illumina chips with >1 million SNPs). While genotyping array represents the single most important factor to imputation accuracy, two aspects of population genetics are also critical: good matching between reference and target samples (here achieved by using a cosmopolitan reference panel, likely to be improved by future reference samples of African diversity [8]) and homogeneity within a GWAS sample (illustrated by a small number of samples of differential ancestry in Kenya with poorer imputation quality, Figures S12 and S13).
Genetic diversity in Africa is extensive [7] and our collections derive from locations separated by thousands of miles and include individuals from several distinct ethnic groups. To characterise population structure we performed principal component analysis (PCA) across our three African collections and a set of African individuals genotyped as part of the HapMap 3 project. For this purpose we selected a set of 121029 SNPs with accurate (MAF>1%, IMPUTE2 info score>0.9) genotypes in all three study collections, and then thinned the data to reduce the correlation between neighbouring SNPs (see Methods).
To summarise the relatedness structure within our data, we similarly produced a thinned list of SNPs with good calls separately in each collection, and calculated allele sharing between all pairs of individuals at the thinned set of SNPs. The distribution of the degree of similarity between each individual and their closest relative within each study is shown in Figure S2. High levels of relatedness between individuals can violate the assumptions of standard tests of association and can dominate attempts to characterise population structure. For these reasons we iteratively removed closely related individuals and those with atypical ancestry as described in Methods. We refer to the remaining set of individuals as the “filtered set” and use them for analyses which rely on the use of principal components (PCs).
The projection of a subset of study and HapMap individuals onto the first two PCs is shown in Figure 2. Some care is needed in interpreting PCA of genetic data [9]; however, the analysis has the property that the distance between any two individuals on the plot is proportional to the genome-wide similarity in their genotypes. The relationships among our samples broadly reflect the geography and peopling of Africa; we see that East African study samples from Kenya and Malawi cluster near one another and are relatively close to HapMap Luhye individuals who are also from Kenya (LWK); and the Gambian samples cluster closer to Yoruban individuals from Nigeria (YRI), representing individuals from West Africa. The Kenyan study samples are recruited from the coastal region of Kilifi and our data confirm they are genetically distinct from the Maasai (MKK).
To characterise the genetic diversity within collections we performed PCA on each study separately and plotted all individuals on the first two principal components (Figure 2). In Gambia, the first PC separate the Fula from the rest of the sample with subsequent PC (Figure S14) stratifying further ethnic groups, as previously shown [10]. A similar relationship between genetic diversity and self-reported ancestry is seen in Kenya.
It is well known that subtle differences in the patterns of relatedness between case and control individuals can potentially lead to spurious signals of association [11]. We observed significant correlation between case control status and the principal components in each of studies (Table S4), and the probability that the closest relative to each sample had the same case status was significantly greater than an expected by chance (P<1×10−4 in all three cohorts).
We took two approaches to controlling for the potential bias that can result from population structure. Firstly we restricted analysis to the filtered set described above and included five PCs as covariates in a logistic regression analysis as implemented in the program SNPTEST (https://mathgen.stats.ox.ac.uk/genetics_software/snptest/snptest.html). Secondly we included all individuals passing quality control filters by modelling the covariance in case status due to relatedness between samples as a random effect in a linear model approximation, as implemented in the program MMM [12]. The latter of these approaches potentially retains more power by increasing the sample size, and provides additional robustness to population structure by modelling relatedness at all levels (or equivalently using all PCs) [13]. Empirically, the evidence for association at each SNP is similar between the two methods (Figures S3 and S4) and both reduce the overall inflation in test statistic to acceptable levels (Table S6). Throughout we assume an additive model of association, estimating a single parameter which determines the effect each copy of risk allele has on the log-odds of being a case individual, using the mixed model approach, unless otherwise stated.
To combine the evidence of association across studies we use an inverse-variance weighted fixed effects approach to calculate an estimate of the log of the odds ratio and its standard error combined across the three studies [14]. This approach has become the standard method for meta-analysis of case-control studies. We return to a discussion of the meta-analysis below.
The results of the autosomal genome-wide association analysis are shown in Figure S5. Two regions of the genome show compelling evidence of association (P<5×10−8). These include SNPs near established malaria susceptibility loci; in the beta globin (HBB) gene on chromosome 11 and in the ABO blood group gene (ABO) on chromosome 9. Several other regions show strong but not conclusive association (P<1×10−6) and are detailed in Table S7. Additional analysis using dominant, recessive or two-parameter models did not reveal any other convincing regions of association showing consistent evidence across collections, nor did direct analysis of the genotype data (Figures S9 and S10).
The strong signals of association at HBB and ABO demonstrate that, despite the additional challenges of genetic analysis in Africa, the standard approach to imputation based meta-analysis can identify loci with convincing levels of evidence when sample sizes are sufficiently large.
The non-synonymous variant rs334 in the HBB gene, whose derived allele (HbS) causes sickle-cell anaemia in homozygote individuals and is known to be strongly protective against malaria in heterozygotes, is perhaps the best known case of balancing selection in the genome [15]. The pattern of association around this region across the three cohorts is shown in Figure 3. The SNP rs334 is neither genotyped nor imputed in our genome-wide data. In its absence, the strongest signal of association is seen over 400 kb upstream of HBB in the Kenya and Malawi studies. The combined evidence at this locus drives the main fixed-effect meta-analysis signal. Strikingly there is very little signal of association at this position in the Gambian data, although strong evidence for association is seen closer to HBB gene at 200 kb upstream and, to a lesser extent, 200 kb downstream.
Direct genotyping of rs334 across the MalariaGEN data (see Methods) allows us to measure the correlation (allelic r2, estimated using EM algorithm) between the HbS allele and the SNPs around HBB in our genome-wide data. The strength of the correlation is indicated by the colour of the points in Figure 3. Association analysis of the region including genotype at rs334 as a covariate completely removes the signal (P>10−4,see Figure S6). Together these observations confirm that the heterogeneous signals of association in the three cohorts are driven by their different patterns of correlation with the causal allele rs334, probably because it arose on different haplotypes in different ancestral populations [16]. The lowest meta-analysis P value across the region is at rs12788102 (P = 1.9×10−16), which was imputed in Gambia, whereas at the directly typed rs334 the meta-analysis P = 2.8e−36. We note that the lowest SNPTEST meta-analysis P value in this region is 5.7×10−13 at rs17325567.
In contrast to the complex patterns of association at HBB, the strong meta-analysis evidence for association at ABO derives from a combination of signals at the same set of SNPs across the three cohorts (Figure 3). In each cohort the strength of association is moderate with P>1e-6, so under the assumption that malaria susceptibility loci of modest effect are rare in the genome, none of these signals are convincing in any one study. However, when combined via meta-analysis they reach established level of significance (P<5×10−8) for genome-wide analysis.
The determinants of the ABO blood group are SNPs within the ABO gene, and have also been typed across the MalariaGEN data, including the deletion at rs8176719 which determines AB/O groups. The correlation between rs8176719 and neighbouring SNPs is also shown in Figure 3. The pattern of association across the cohorts is similar to those typically seen in studies of European populations where the correlations between alleles at tag (marker) SNPs and the causal allele are consistent across different samples. The strongest meta-analysis signal at the locus is at rs8176722 with P = 8.9×10−10. Conditioning on rs8176719 also removes any other signals of association (P>10−4) from the region (Figure S6).
The differences in ancestries of the three study samples can lead to a causal SNP being differentially tagged, as observed at the HBB locus in our data. As a consequence the effect sizes at directly typed or imputed loci can vary between samples, even when the risk at the causal locus is the same. Moreover, variation across studies in imputation accuracy can also lead to differential levels of effect size underestimation at SNPs not genotyped directly. These effects are likely to be more important in studies across African populations and motivate approaches which relax the assumption of the same or “fixed” effect.
To investigate the impact of non-fixed effect approaches on the evidence for association we used a normal approximation to the logistic regression likelihood suggested by Wakefield [17]. One way of thinking about the approach is that it uses the study-wise estimated log-odds ratio (beta) and its standard error as summary statistics of the data (See Text S2). For each model of association we assume a prior on the log odds ratio which is normally distributed around zero with a standard deviation of 0.2 (see [18] for a discussion). By changing the prior on the covariance (or correlation) in effect sizes between studies we can formally compare models where:
(See Text S2 for details). For each model we can obtain a Bayes factor (BF) for association by comparing it with the null model where all the prior weight is on an effect size of zero. These models are similar in spirit to those employed to look at shared effects across sub-phenotypes (rather than populations) in a study of ischemic stroke [19] or at heterogeneity between studies [20].
The genome-wide Bayes factors for the fixed effects and structured effects models are plotted against each other in Figure 4. From this plot we can see that the fixed effect BF is larger for SNPs at ABO, while at HBB, there is more evidence for association when effect sizes are allowed to vary more extensively between East and West African collections (structured effects). The posterior probability on each of the models at the SNPs in these regions is shown in Figure S17. Similar results are seen when the prior on the effect size is increased (standard deviation of 0.75; data not shown).
Motivated by these observations at known malaria susceptibility loci we performed a genome-wide scan (Figure 5) to look for regions with strong evidence of association (log10(BF)>4) under model 2, 3 or 4 above. These are listed in Table S7. In large case-control samples, like those analysed here, then for common SNP (minor allele frequency >5%) the fixed effect meta-analysis P values are highly correlated with the fixed effect BF (see Figure S7). Nonetheless, power to detect new regions of association is highest when the prior distribution of the effect sizes across cohorts is close to the truth. We therefore advocate this approach as a way of accounting for our uncertainty in the correct meta-analysis model in terms of similarity of effect sizes between cohorts. Two regions showing over twice as much evidence for association under the structured effects model compared to the fixed effects model were on chromosome 16 in the large gene CDH13, where the signal of association is strongest in East Africa (Kenya and Malawi), and a region of association on chromosome 14, where the association signal is largely confined to West Africa (Gambia). The structured effects log10(BF) for these regions is 4.84 and 4.03 respectively (see Table S7).
Another possible approach to identifying genetic associations across populations, where the most associated SNPs at a locus are not necessarily the same, is to base a test on all SNPs within a region [21], [22]. One way to formulate this test is to consider the causal SNP as a random effect, which is not observed, but is assumed to have a correlation structure across individuals dictated by the pattern of relatedness (or allele sharing) within the region of interest. A test for association can then be constructed by asking whether the random effect explains any of the covariance in the phenotype (case-control status), after accounting for population structure, which can be captured by including PCs as fixed effects in the model (See Text S2). As the model includes only fixed effects and a single additional random effect it can be computed using the MMM software [11]. To assess the evidence for association we used a score test statistic which has a complex, but computable, distribution under the null to calculate a P value. We also computed Bayes factors by specifying priors on the model parameters, in particular on the proportion of the variance explained by the region.
To exploit a region-based test we constructed allele sharing matrices (as defined in [12]) for all SNPs within 50 kb of each of approximately 20,000 genes in the genome (where there were at least 5 SNPs in each cohort). We then tested for association as described above to obtain a P value for each gene in each of the three study collections using the filtered data sets. To check that population structure was sufficiently accounted for by the inclusion of five PCs, we calculated the genomic inflation factor of the P values in each study and found them to all be less than 1.07 (after removing HBB and ABO regions). We combined evidence across cohorts using Fisher's method [23]. The meta-analysis P values had an acceptable genomic inflation factor of 1.071. See Figure S8. We also multiplied Bayes factors across studies to obtain a study-wide Bayes factor. We note that, unlike the fixed-effect SNP analysis, these approaches to meta-analysis do not assume that the same allele, or combination of alleles, determine susceptibility in each cohort.
As a proof of principle we applied this test to all SNPs within the HBB region (4.6 Mb to 5.5 Mb of chromosome 11), which covers the two peaks of association in the single SNP analysis (Figure 3). The region-based test showed evidence for association (P<5×10−5) in each cohort and had a meta-analysis P = 3.5×10−17, whereas the lowest SNPTEST meta-analysis P value in the region (which also uses 5 PCs as covariates) is 5.7×10−13. Although this region would have been discovered via either approach, this additional boost in power highlights the potential benefit of region based tests.
In the gene-based analysis the most associated region was a gene overlapping the region of strongest association at HBB in Malawi and Gambia (OR51F1 from 4.73 Mb to 4.8.3 Mb, containing approximately 70 SNPs), it had a meta-analysis P value = 4.4×10−11. In contrast the gene-based signal at the ABO locus (P = 1.6×10−5) was significantly less than the SNP-based analysis. It is likely that the inclusion of multiple SNPs from the region and removal of assumptions about the direction of effect across cohort reduces the signal of association at this locus. However, the region-based analysis focuses attention on the approximately 20,000 annotated genes; since consensus does not yet exist on interpreting gene-based P values, BFs are useful in evaluating the evidence for association in the gene-based tests. For instance, if we assume that there are 20 annotated genes which contain SNPs within 50 Kb that influence malaria susceptibility then the prior odds of association are roughly 1 in 1000. In comparison, the prior probability associated with any given SNP is much lower, perhaps 1 in 100,000 [18]. Thus, a log10(BF) of 2 (BF = 100) for the region-based analysis gives the same posterior probability of association to a log10(BF) of 4 in the single-SNP analysis. Plots of the empirical distribution of the estimated proportion of the phenotypic variance explained by the regions, and a comparison of Bayesian and frequentist tests, are shown in Figure S19.
Outside the ABO and HBB regions, five regions contained genes with BF greater than 100. Although this analysis is gene focussed, it does not necessarily directly implicate specific genes but regions of the genome (see Table S7), but here we refer by gene name to the regions with the most evidence. These include the regions of the genes BET1L (telomere chromosome 1, log10(BF) = 2.504), C10orf57 (chromosome 10, log10(BF) = 2.387), MYOT (chromosome 5, log10(BF) = 2.051)), SMARCA5 (chromosome 4, log10(BF) = 2.04) and ATP2B4 (chromosome 1, log10(BF) = 2.015). Interestingly, we note that the SMARCA5 region is 250 kb upstream of the GYPE/A/B gene cluster which encodes the M blood group antigens, and that the variants in the BET1L gene have been associated with platelet volume in Europeans [24], [25]. A recent study [26] identified malaria-associated variants in ATP2B4.
Another benefit of a test which averages over SNPs within a region to obtain a single P value or BF is that it is possible to look for consistent association in collections of genes (or regions) of interest. We hypothesised that loci either previously implicated in auto-immune disease [27] (referred to below as “Immunochip regions”), associated with measurable properties of red and white blood cells and platelets [24], [28], [29], or known determinants of blood groups (obtained from the HUGO database and excluding the ABO types) might be candidates for malaria susceptibility variants. To investigate, we also calculated region-based Bayes factors for these regions and ranked them against the results from the gene-based analysis. Table S8 shows that only the Immunochip loci showed a nominally significant (P = 0.001) excess of high-ranking BFs (those in the top 5% of the empirical distribution) of genes. We note that, other than the BET1L locus, the two highest ranking (in the top 0.1% of the empirical distribution) regions include the RUNX3 locus, implicated in ankylosing spondylitis [30] (empirical P = 0.0060) and the region containing the IL12A gene (empirical P = 0.0096) implicated in Celiac disease [31] and multiple sclerosis [31], [32].
The empirical observations described above, including the heterogeneity of signal at the HBB locus, and the ability of the region-based test to detect the recently identified association at the ATP2B4 locus [26], motivates further investigation of the new approaches to association analysis. To further assess the utility of these methods we used HAPGEN [33], [34] to generate a series of simulated case-control meta-analysis datasets in ten randomly chosen genomic regions, using samples from three African populations collected as part of the HapMap project. We conducted two sets of simulation, designed to test two distinct association scenarios (see Methods). In the first set, the three populations were assigned the same underlying causal variant (with an odds ratios of 1.2 per allele), but the causal variant was assumed to be untyped. In the second set, each of the three populations had a different causal variant. We also carried out a series of null simulations (with no causal variants present at all) in order to quantify false-positive rates. We ranked all the simulations according to either the strongest evidence of association under different single SNP analyses or the region-based approach, and plotted true positive and false positive rates for all methods (Figure S18). We found that if the causal SNP is the same across populations then the fixed effect approaches, and Bayesian approaches that assume highly correlated effects, perform best. In contrast, when the causal SNP is population-specific the region-based approaches have the highest power for a fixed false positive rate. Importantly, the single SNP correlated effects Bayesian approach performs well under both scenarios, highlighting the utility of this approach when the assumption of homogeneous effect sizes across populations does not hold.
The purpose of our analysis was to assess the utility of imputation based meta-analysis for combining data from individuals, typed on different genotyping arrays, and sampled from different African populations, for studying Malaria susceptibility. Until recently [26], no single genome-wide analysis of malaria has revealed evidence of association strong enough to overcome the implicit low prior probability that any given SNP affects susceptibility. We show that by increasing sample size, through meta-analysis, it is possible to identify such polymorphism from diverse African populations. This reinforces the utility of applying the GWAS approach in this setting.
The two loci we identified with P values<5×10−8 have been known to influence malaria susceptibility for more than 25 years and are likely to exhibit some of the strongest effects on risk for common alleles. Extensive analysis of other phenotypes, both communicable and non-communicable, suggests that these loci are the tip of the iceberg, with many smaller effects left to be found. Identifying these regions requires additional statistical power. Primarily this can be achieved by increasing the size of the data sets by collecting more individuals or obtaining data at a denser set of SNPs, perhaps initially through imputation [6], but ultimately by whole-genome sequencing of study individuals. By typing more SNPs the difference in patterns of linkage disequilibrium between different ancestral groups, which can complicate analysis, may be mitigated as it becomes more likely that the causal variant is typed directly. Nonetheless, a combination of approaches such as the Bayesian random effect models and region-based tests outlined here may still provide additional power by relaxing the assumptions of standard SNP-by-SNP fixed effect analysis. For example, when the genetic effects are modified by the environment (such as the parasite or mosquito sub-species), or the clinical criteria for inclusion as a case individual varies between cohorts, or when different mutations arise at the same locus in different ancestral populations, even typing the causal variant may still result in effect heterogeneity. We note that the application of these new methods requires additional care because they are potentially less robust to sporadic genotyping errors in one or more cohorts. The ultimate decision about which of the approaches we have explored will be most appropriate for other researchers working on GWAS in complex populations will depend on the circumstances of individual studies. We view comparison of the models to be informative, and suggest averaging across models where a single summary of the evidence for association is preferable. Prior information on the likelihood of real, or apparent, effect heterogeneity can easily be incorporated in this approach.
Although the methods described in this paper do not confidently identify any new malaria susceptibility loci, they do highlight a set of potential candidates. For example, variation in the large gene CDH13 has recently been conclusively associated with adiponectin levels [35] and other metabolic traits [36]. The chromosome 4 gene-based association may further implicate glycophorins A, B and E which encode the M blood group antigens and are potential receptors for the malaria parasite P. falciparum. The signal of association at ATP2BA coincides with the findings of a recent GWAS study in Ghana [26] and is of potential functional significance as it encodes the major calcium-transporting ATPase on the erythrocyte plasma membrane. It is also just upstream of LAX1, a transmembrane protein expressed in peripheral blood lymphocytes and implicated in T and B cell responsiveness to stimulation [37]. Further data will be required to confirm if these replicate in other collections, which specific genes are involved, and how genetic variation in the region influences severe malaria susceptibility.
There has been a long standing hypothesis that high mortality from infectious disease in ancestral populations has led to selection pressures which have had an impact on human physiology. For the first time in human genetics we are in a position to test such hypotheses. The random effects models and the region-based test described here provide a statistically principled approach to looking systematically for shared, antagonistic or pleiotrophic effects across phenotypes. The identification of new malaria susceptibility loci that will result from larger studies will empower investigations of this kind as well as providing desperately needed insights into the aetiology of malaria infection and the host's response.
The studies and sample sets described in this manuscript form part of a larger ongoing project within the Malaria Genomic Epidemiology Network (www.malariagen.net). Here we describe partner projects from the MRC laboratories in Fajara, Gambia, The KEMRI-Wellcome Trust Unit in Kilif, Kenya and The Blantyre-Wellcome Trust Project in Blantyre, Malawi (Table S1). At each study site, cases of severe malaria were recruited on admission to hospital while controls (cord blood samples) for the cases were sampled from the same populations by recruiting mothers giving birth at local maternity units. This was usually done as part of a larger programme of clinical research on malaria, designed and led by local investigators. Further details can be found by visiting the MalariaGEN website (www.malariagen.net). We define a case of severe malaria as an individual admitted to a hospital or clinic with P. falciparum parasites in the blood film and with clinical features of severe malaria as defined by WHO criteria [38], [39].
Study sites worked with the MalariaGEN resource centre to define best practices for ethical conduct of these genetic studies in the local setting, including the development of guidelines for obtaining informed consent [40]–[42]. Further information on policies, research and the consent process may be found on the MalariaGEN website (http://www.malariagen.net/community/ethics-governance). All research was reviewed and approvals granted by local Research Boards and Ethics committees in The Gambia: The Gambia Government/MRC Unit Joint Ethics Committee (SCC1029 and SCC670/630); Kenya: Research Ethics Committee from the KEMRI-Wellcome Research Programme, Kilifi, Kenya (SCC1192); Malawi: College of Medicine Research Ethics committee, University of Malawi and the Blantyre Malaria Project with the Malawi-Liverpool-Wellcome Trust Programme, Blantyre, Malawi (P.05/06/422) and Oxford; Oxford University Tropical Research Ethics committee (OXTREC), Oxford, United Kingdom (OXTREC 020-006). This paper is published with the permission of the Director of KEMRI.
Sample collection and DNA extraction were undertaken at partner study sites and at the MalariaGEN Resource Centre as previously described [4]. All samples submitted to the MalariaGEN Resource Centre underwent a standard set of procedures that included quantification using picogreen, genotyping of 65 polymorphisms (including HbS - rs334, and 3 gender-typing SNPs) on the Sequenom iPLEX MassArray platform and matching to baseline clinical data (e,g gender, ethnic group and case-control status) as described in [4]. Samples meeting DNA concentration and genotyping criteria with appropriate clinical data were selected for GWAS. However, due to restrictions on the total amounts of blood and DNA collected, it was necessary to first whole-genome amplify all gDNA samples by multiple-displacement-amplification as previously described [10]. Briefly gDNA was whole-genome amplified using the REPLI-g kit (Qiagen, Crawley, UK) with the modification for increased sample volumes). All final reaction DNA concentrations were measured using PicoGreen reagent (Invitrogen, Paisley, UK) and adjusted to 100 ng/ul with low TE (10 mM Tris-HCL pH 8, 0.1 mM EDTA-Na2).
Twelve percent of samples were assessed for amplicon size range using the Agilent 2100 bioanalyser (Agilent Technologies, Stockport, UK) according to manufacturer's instructions and for genotyping efficiency using the SNP set described above. Whole-genome amplified material was submitted to the Wellcome Trust Sanger Institute for genotyping on as part of the ongoing MalariaGEN consortial Project 1 (http://www.malariagen.net/node/128). Details of the 3 datasets are described in Table S1.
Our processing pipeline first uses the AutoConvert function in Illumina Beeline software to convert raw read data from Illumina BeadArray (idat) files into binary genotype call (gtc) files using cluster positions and normalisation information in the (egt) files. We used the gtc files to extract the calls made by the algorithm GenCall [5], to extract the raw intensities for GenoSNP [6] and the normalised intensities for the program Illuminus [7]. We wrote custom software to split the data into chunks, and modified Illuminus and GenoSNP to accept the new format. This allowed us to parallelise the genotype calling. We then included all the three sets of genotypes, along with normalised intensity data into a single vcf file (see http://www.1000genomes.org/node/101).
Three genotype calling algorithms were specifically chosen to utilise different information in the intensity data; GenoSNP which independently calls genotypes in each individual by clustering probe intensities across SNPs; Illuminus which independently calls genotypes at each SNP by clustering probe intensities across study individuals; and Gencall which uses predetermined probe intensity information to infer genotypes at each SNP in each individual.
Each of the three algorithms can make one of four calls for a given individual at a given SNP: 0,1,2 or missing. To merge the calls we took a simple consensus approach to generate a single call for each genotype. The rules were as follows:
The above rule is strict in the sense that only complete agreement between algorithms that made a call leads to a genotype call in the merged data. Analysis of trio data demonstrated that this retains a relatively high fraction of calls relative to anyone calling approach and had the lowest number of Mendelian errors in terms of absolute errors made and the fraction of SNPs with one or more errors (data not shown).
Prior to imputation we applied quality control (QC) metrics to each cohort as follows. All QC was performed on the “consensus” genotype call, as defined above. First, we aligned each dataset to the forward strand of the reference genome using the Illumina-supplied chip manifest, and restricted attention to the set of SNPs in the HapMap 3 reference panel (obtained from the IMPUTE website). We excluded samples with missingness >10% or heterozygosity outside the range 0.25–0.35. We then proceeded to filter out SNPs based on missingness (or discordance of the consensus genotype call), minor allele frequency, and HWE P value. We also examined differential missingness between cases and controls in each cohort, but did not apply exclusions based on this criterion. Next, we computed pair-wise concordance between samples using a thinned set of approximately 20,000 SNPs chosen to be no closer than 100 kb apart. For each pair of highly concordant samples we removed the sample with higher missingness.
Imputation was performed with IMPUTE 2.12, using the phased release #2 of HapMap3 from the Impute website (http://mathgen.stats.ox.ac.uk/impute/). All HapMap3 haplotypes from all populations (African and non-African) were used. The genome was split up into segments which are either 5 Mb, or have 20 000 reference SNPs (whichever is smaller), with an additional 500 kb buffer on either side of the segment. We used imputation parameter settings of k = 80 and Ne = 14000. Imputation was performed in parallel for each segment, and segments were reconstructed into chromosomes once all imputations had finished. For each cohort we examined diagnostic metrics to assess imputation performance (see Figure S11).
Accuracy of imputation was measured using IMPUTE2's “type 2 r2”, which is the squared correlation coefficient between actual genotypes in our GWAS dataset (discrete values of 0, 1, or 2 measured by the genotyping algorithms described above) and the expected genotype, or “dosage”, predicted by IMPUTE2 (a continuous value from 0–2). As long as variants present in the GWAS are not biased towards being easier or more difficult to impute than typical variants in HapMap3, this metric is a good representation of the accuracy of our imputed genotypes at all sites [5].
To investigate population structure, and for genome-wide scans performed using SNPTEST and other tools that do not directly model relatedness, we further restricted the set of individuals as follows. For each cohort we computed a matrix (denoted R) of genome-wide allele sharing, using approximately 100,000 SNPs thinned to be at least 0.01 cM apart using the HapMap combined recombination map. Using this matrix we excluded one of each pair of individuals that were closely related. Using the same matrix R we also computed the projection of samples onto the principal components (given by the eigenvectors of R). To investigate the potential for population structure to generate false signals of association we calculated the correlation between each PC and case control status using logistic regression. The resulting P values are shown in Table S4 and show that case control status is significantly associated with the principal components.
To ensure individuals of unusual ancestry did not dominate analyses, we iteratively excluded individuals that were extreme outliers (>10 standard deviations away from the mean) in any of the first ten PCA components, and re-computed PCs; this resulted in a small number of further exclusions (see Table S5). Projections of samples onto the principal components are depicted in Figure 2 and Figures S14, S15, S16.
Relatedness matrix and PCA computations were performed using QCTOOL (http://www.well.ox.ac.uk/~gav/qctool) and the generateR program included with the MMM program [12] . Association tests were performed using SNPTEST v2.3.0, using maximum likelihood estimation taking into account uncertainty in imputed genotypes, and including 5 PCAs to control for population structure. Mixed model scans and region based tests were performed using MMM, with mean-centering of genotypes, imputation of missing and uncertain genotypes. For region-based tests we computed relatedness matrices using all SNPs within the region that passed QC filters (info>0.75, MAF>0.001) and used this relatedness matrix as a random effect in the MMM program. Five PCs were included to control for population structure. For Bayesian tests, we specified the prior on the h parameter (see Text S2) as a beta distribution with parameters 1.5 and 100; intuitively this corresponds to a belief that the regional relatedness matrix explains relatively little (around 1%) of the overall residual variance. Pairwise LD computations were performed using QCTOOL, which uses the EM algorithm to estimate the phase of individuals heterozygous at both markers. Meta-analysis was performed using fixed-effect inverse variance weighting and Fisher's method using custom software written in C and R [43]. Bayes factors for meta-analysis models were also calculated in R. For technical details on novel methods see Text S2.
For the gene based analysis regions were obtained by taking the transcript start and end positions from the refGene table of the UCSC genome browser database [44]; where a gene had multiple transcripts we used the longest transcript. This resulted in 22903 gene regions, of which 21908 were on autosomes. We then added 50 kb to start and the each end of the region before applying the regional association test and only included regions with more than 5 SNPs in each of the three cohorts.
For empirical investigation of regional association test statistics we used several lists of regions. A list of blood group antigen genes was obtained from the HUGO Gene Nomenclature Committee website (http://www.genenames.org/genefamilies/blood-group) and these were extracted from the gene based analysis described above by matching gene names. To define lists of genes associated with blood cell phenotypes, we took association signals from [28] (red blood cell traits comprising Hemoglobin concentration (Hgb), hematocrit (Hct), mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC) and red blood cell count (RBC)) [24], (platelet traits comprising mean platelet volume (MPV) and/or platelet count (PLT)), and [29] (white blood cell traits comprising total white blood cell count (WBC), and counts of Neutrophils, Basophils, Lymphocytes and Monocytes.) For each of the three overall phenotypes (red or white blood cell, or platelet traits) we recorded the most-associated SNP across sub-phenotypes for each reported locus from the relevant study. For each such locus, we then defined a region by finding the furthest SNP upstream and downstream of the locus having r2>0.5 with the associated SNP, and then moving out to the nearest recombination hotspot. If the region contained no genes we further extended it by 25 kb in each direction. This procedure is the same as those implemented in GRAIL [45]. To define a list of regions associated with auto-immune disorders we used fine-mapping regions from the Immunochip platform [27].
Ten 100 Kb regions were chosen uniformly from across the autosomal genome. For each region we used the program HAPGEN (v2.1.2, with default settings used [34]) to simulate a total of 1,000 meta-analysis datasets, each consisting of 1000 cases and 1000 controls from each of the three African populations (YRI, LWK and MKK) from HapMap3 (release 2). We repeated this simulation under three scenarios of association for a total of 30,000 datasets. The scenarios considered were:
Causal variants were picked at random from among those with combined MAF>0.05, and assumed to act additively on the log odds scale, with odds ratio of 1.2.
For each dataset we tested each SNP in the region for association, separately in each population, using SNPTEST. We combined effect size estimates and standard errors across populations using frequentist fixed-effect and Bayesian meta-analyses (Methods and Text S2). For Bayesian meta-analysis, prior variance on effect sizes was set to 0.22 and we used between-population prior correlations of 1 (fixed effect), 0.9, 0.5, and 0 (independent effect). For the single causal variant scenario, the chosen causal variant was masked from association testing.
For each dataset and population we also computed a Pvalue and Bayes factor for the regional association test (Methods and Text S2). For the single causal variant scenario, the causal variant was masked from computation of the covariance matrix. We combined P values using Fisher meta-analysis and multiplied Bayes factors across populations to produce a single P value and a single Bayes factor for the dataset.
For each of the two scenarios of association and each method of detecting association across the region (regional test with Fisher meta-analysis, regional test with Bayesian meta-analysis, best single-SNP frequentist meta-analysis, best single-SNP Bayes factor for each of the four choices of correlation parameter) we produced ROC curves (Figure S18) by combining all datasets simulated under that scenario and the null scenario, ranking by the chosen P value or Bayes factor, and computing empirical true and false positive rates.
For information on access to project data see www.malariagen.net.
|
10.1371/journal.pcbi.1003773 | Unveiling Time in Dose-Response Models to Infer Host Susceptibility to Pathogens | The biological effects of interventions to control infectious diseases typically depend on the intensity of pathogen challenge. As much as the levels of natural pathogen circulation vary over time and geographical location, the development of invariant efficacy measures is of major importance, even if only indirectly inferrable. Here a method is introduced to assess host susceptibility to pathogens, and applied to a detailed dataset generated by challenging groups of insect hosts (Drosophila melanogaster) with a range of pathogen (Drosophila C Virus) doses and recording survival over time. The experiment was replicated for flies carrying the Wolbachia symbiont, which is known to reduce host susceptibility to viral infections. The entire dataset is fitted by a novel quantitative framework that significantly extends classical methods for microbial risk assessment and provides accurate distributions of symbiont-induced protection. More generally, our data-driven modeling procedure provides novel insights for study design and analyses to assess interventions.
| While control options for plant, animal, and human pathogens are emerging rapidly, reliable assessment of the effect of interventions in biological systems presents many challenges. A major question is how to connect laboratory experiments and measurements with the relevant process in natural settings, where hosts are subject to pathogen exposures that vary in time and geographical location. With this aim, measures of protection that are invariant under varying exposure intensity need to be developed and integrated with mathematical models. In this article, we introduce a method to assess host susceptibility to pathogens, and apply it to survival of Drosophila melanogaster challenged with different doses of Drosophila C virus. By replicating the procedure in groups of flies that carry the symbiont Wolbachia, we are able to estimate how the viral protection induced by this intracellular bacterium is distributed in the host population. Our results disentangle host infection status from observed mortality, accounting naturally for time since exposure. The multiple-dose design proposed challenges traditional study designs to assess interventions.
| Hosts exposed to disease-causing agents respond in accordance to the challenge dose. Therefore dose-response curves contain information about disease processes that can be extracted by suitable analytic frameworks. Early examples concerning microbial risk assessment include counting lesions caused by tobacco mosaic virus on plant leaves [1], as well as human responders to experimental challenge with polio viruses [2], Vibrio cholerae [3] and Streptococcus pneumoniae [4], for escalating challenge doses. Dose-response models have been in use for analyses and extrapolation of experimental datasets [5].
Models that account for the sigmoidal shape in log-linear scale of the typical dose-response curve have been derived mechanistically, based on the assumption that each individual pathogen has a probability of infection independent of others, the so-called independent action hypothesis [6]. This results in a one-parameter exponential-function model [7]. The frequent observation of shallower-than-exponential, or overdispersed, relationships has then prompted the implementation of heterogeneity in the probability of infection of individual hosts [8]–[10].
In the 1960s, Furumoto and Mickey [9] developed a dose-response model that could accommodate both shallow and steep increases in the response by considering the probability of infection of individual hosts described by a Beta-distribution. Although a mechanistic justification for this specific distribution has not been given, the model has been widely applied in microbial risk assessment due to its ability to outperform the simple exponential model [5].
Susceptibility distributions other than Beta have also been considered and are more commonly used in frailty models adopted in survival analysis [11], where the data consist of survivor counts over time in host groups that are constantly subject to a hazard [12], [13]. These frailty models appeared in the 1980s and have since been adapted to infection hazards, where surviving signifies remaining uninfected [14]–[16]. While most informative when the exposure is continued or repeated over time, these formalisms would be inadequate for estimating distributions of susceptibility to infection from instantaneous challenge protocols.
The importance of accounting for time between challenge and observable toxicity responses to pathogens or other agents has been recognized. Recent models in ecotoxicology [17], [18], consider explicit kinetics within exposed organisms. Also in microbial risk analysis, previous studies [19], [20] have included time postinoculation as an additional parameter in classic dose-response models, although using an approach that conceptually allows for a different susceptibility distribution at each time point. Here we present a schema to infer a distribution of host susceptibilities to infection that holds consistently across dose and time. We introduce an experimental design and inference framework that enables such inferences by analyzing simultaneously a collection of survival curves, each representing a different challenge dose. The resulting Beta distributions are compared against those obtained by classic dose-response models based on single day measurements.
Recent evidence for symbiotic interactions that reduce host susceptibility to pathogens has stimulated the development of quantitative frameworks to assess the levels of individual and population protection attributable to specific symbionts. The intracellular bacterium Wolbachia, found among many arthropod species including Drosophila melanogaster, is one such symbiont [21], [22]. To analyze the protection conferred by Wolbachia to D. melanogaster, we apply our inference framework simultaneously to two sets of time-dependent dose-response data: in one set the flies carry the symbiont bacterium Wolbachia (Wolb+); while in the other they do not (Wolb−). In this instance we extract the Beta distribution that best describes individual protection attributable to Wolbachia, as well as population statistics valid across entire dose ranges.
We used virus free D. melanogaster lines with DrosDel w1118 background, with or without the endogenous Wolbachia strain wMelCS [21], [23], [24]. Flies were reared in standard food at 25°C. To assure that potential for heterogeneities are minimized by the experimental procedure, we used fifty 3–6 days old adult males per group, 10 per replicate and 5 replicates. To study the response to viral infection, we anesthetized with CO2 and pricked flies with different doses of Drosophila C virus (DCV). We used tenfold serial dilutions – from 1010 TCID50/ml to 104 TCID50/ml – in Tris-HCl buffer, pH 7.5. Controls were pricked with buffer solution only. We used the pricking protocol described in [24], produced and titrated virus as in [21]. After pricking, we kept flies at 18°C and checked daily survival until day 80 and twice a week until the end of the experiment. Food was changed every 5 days. We summarized the data in 16 dose-response curves (8 per group, including control) from day 0 after treatment until day 139 (Dataset S1).
Starting from established models, we refine the occurrence of mortality from infection, i.e. the response, as a function of the concentration of infectious units given to hosts, i.e. the dose. We present a step-by-step derivation of descriptions that integrate dimensions that are usually treated separately as well as the motivations for doing so.
Assuming independent action of infectious units, each unit has probability p of causing an infection, while for d infectious units infection occurs with a probability described by . Given further considerations about the distribution of infectious units in a homogeneous solution (see [9] for a complete derivation of the expression), the number of units causing infection can be described by a Poisson distribution, resulting in the exponential dose-response model [7], that describes the probability of infection in a host challenged with pathogen dose d:(1)This most basic formulation is hereafter referred to as the homogeneous dose-response model.
Furumoto and Mickey [9] expanded this formulation by allowing the probability of infection to be described by a parametric distribution, specifically the Beta distribution. To facilitate normalization across datasets, here we maintain the probability p fixed across individual hosts (as in [25]), and introduce a multiplicative parameter, the susceptibility factor , to describe any natural or induced effect that decreases susceptibility. We assume that susceptibility to infection is Beta-distributed so as to describe the variation of susceptibility in the host population. Thus, we obtain the probability that a host contracts infection as(2)where and B is the Beta function. We refer to this formulation as the heterogeneous dose-response model.
At last we introduce a small parameter ε to account for a small probability of ineffective challenge, such that is the random variable representing the number of infected hosts, in a group of n hosts challenged with a given dose. Assuming that an ineffectively challenged host behaves like a control host with regard to death rates, the probability that m hosts are dead a number of days after challenge is then(3)where is either (1) or (2) depending on which dose-response model is adopted.
The parameters to be estimated for this dose-response model are the maximum probability of infection per infectious unit (p), the shape parameters for the Beta distribution that describes the susceptibility factor (a, b), and the probability of ineffective challenge ().
These models require a choice of how many days post-challenge cumulative mortality should be measured, which is difficult to establish for host-pathogen systems where times to death from infection or other causes overlap significantly. To overcome this difficulty, we develop a model that integrates an explicit representation of time to death with the dose-response process for infection just described. It should, however, be noted that time is introduced with the main purpose of enabling the use of survival curves to obtain robust estimates for probabilities of infection given different challenge intensities and consistently infer susceptibility to infection. From this perspective, parameters defined from now on should be regarded as auxiliary and will be implemented as simply as possible.
We first consider a survival model for a control group of flies pricked with buffer solution only (no DCV), subject to two hazards: , an age-dependent death hazard rate; and , a background age-independent death hazard rate. The overall death hazard rate for uninfected hosts is therefore(4)Denoting the random variable representing time to death of control hosts, we have(5)where and are the times to death from and , respectively. Their corresponding distributions are assumed to be and , where is the background mortality rate, is the mean time to death, and is the shape parameter for the Gamma distribution of day of death from aging.
Hosts challenged with pathogen can become infected or remain uninfected and this infection status is hidden. If uninfected, they are subject to the age-dependent hazard rate that affects control hosts, ; if infected, they are subject to an infection hazard rate, , and the age-independent background mortality. Thus the overall hazard rate of infected hosts is(6)Now let be the random variable representing the number of hosts infected by challenge with a given pathogen dose. Then the probability that i hosts are infected after n hosts were challenged is(7)where is either (1) or (2) depending on which dose-response model is adopted.
Let T be the random variable representing the time to death of hosts challenged by a given pathogen dose. The probability density of observing a death event at time t given that i hosts are infected is(8)where denotes the distribution of time to death of infected hosts, given by(9)and is the distribution of times to death from the infection hazard rate . This distribution is assumed to follow , where is the mean time to death of infected hosts, and is the shape parameter for the Gamma distribution of day of death from infection.
In setting the priors for parameter estimation we note that background mortality is small and therefore is kept small by setting to be much greater than the last day of the experiment. To enforce that deaths due to infection occur earlier than deaths due to aging, we constrain the mean time to infection death to be lower than old-age death, i.e. , and the probability of dying before the end of the study to be greater for infected hosts, i.e. , where is the last day of the experiment.
To construct the likelihood to be maximized by the parameter estimation procedure, we let be the random variable denoting the day fly died and the random number of survivors up to . Then the likelihood of observing the actual number of survivors and the times of death , for a given dose is(10)Since the observations for each dose are independent, taking the product of the likelihoods over the different doses yields the global expression for the likelihood of the entire dataset.
In this time-dependent dose-response model, the parameters to be estimated are the maximum probability of infection per infectious unit (p) used for normalization purposes, the Beta distribution shape parameters to describe variation in susceptibility factor (a,b), the parameters that control death due to aging (, ), infection (, ), and background mortality (), as well as probability of ineffective challenge (). Parameters and are typically small and were introduced to improve performance of the likelihood.
Model parameters were estimated using Markov chain Monte Carlo sampling implemented with the PyMC package [26] (code available from [27]). The prior distributions considered are listed in Table 1. Initial values were chosen so as to start with a non-zero likelihood. Using Metropolis-Hastings algorithm, we ran two separate chains for 252,000 iterations. The first 27,000 iterations were discarded. The recording interval was set to 250 so that the autocorrelation between samples was negligible. Convergence was assessed by inspection of the trace plots. All analyses were performed on the pooled samples from the two replicate chains.
Groups of Wolbachia-negative (Wolb−) and positive (Wolb+) D. melanogaster flies were challenged with a range of DCV doses and survival curves were traced as shown in Figure 1. This dataset was analyzed by applying the models introduced in Methods.
To emphasize the importance of day selection to infer distributions of susceptibility to infection by classic dose-response models [5] we have applied these procedures to mortality data observed by two specific days (30 and 50). Parameter estimates from these models are listed in Table 2. The model fits to the mortality data at the selected days are shown in Figure 2, as well as the associated distribution of Wolb+ susceptibilities and the posterior samples for the Beta distribution shape parameters. For simplicity we have adopted the homogeneous model for Wolb− and focus on comparing susceptibility distributions of Wolb+ inferred at different days. Mean protection conferred by Wolbachia in this illustration is estimated as 79% and 56%, based on mortality measurements at day 30 and 50, respectively. Moreover, the distributions have fundamentally different shapes, with the appearance of a high susceptibility group as time progresses. This sensitivity to the day by which mortality data are collected is a concern that raises the need to disentangle infection status from the associated time-dependent mortality. In the following sections, infection and mortality are estimated explicitly using the integrated time-dependent model described in Methods. The procedure is illustrated in Figure 3.
Control curves from Wolb− and Wolb+ flies pricked with buffer solution (no DCV) were compared with the Kaplan-Meier method using the log-rank test and no significant difference was found (with a p-value of 0.47). By fitting the uninfected time-dependent model (4–6) to the control survival curves (Figure 1) we estimated the parameters describing aging () and background () mortality (Table 3).
For each group of flies (Wolb− and Wolb+), the time-dependent dose-response model constructed in Methods was fitted simultaneously to the entire dataset of survival curves (one for each DCV challenge dose), fixing across doses the distribution of times to death from infection (, ) and aging (), while estimating the susceptibility parameters (p, a, b) that govern the dependence of response on challenge dose according to the adopted dose-response model. The estimated parameter values are listed in Table 4. The deviance information criterion (DIC) [28] favored the homogeneous model for the Wolb− group and the heterogeneous model for Wolb+ (Text S1). Mean time to death from infection is 9 and 14 days in the Wolb− and Wolb+ groups, respectively. The variance in time to death from infection is lower for Wolb−, with a standard deviation of 2 days, compared to 6 days in the Wolb+. Figure 4 compares fitted with observed survival curves.
The fitted dose-response curves that result from this analysis are shown in Figure 5A, while the inferred distribution of Wolb+ susceptibilities normalized by the Wolb− measure is displayed in Figure 5B and the corresponding posterior distribution of the Beta shape parameters is in Figure 5C. Given the homogeneity in the Wolb− group, the distribution of susceptibility in Wolb+ provides a direct indication of how antiviral protection conferred by Wolbachia is distributed among its carriers. Typically defined as , where RR is the risk reduction attributed to the susceptibility modifier (Wolbachia in this case), we determine the mean protection conferred by the symbiont to its host as 85% (with a 95% HPD of 60–93%).
To assess the best possible performance of classic methods [5] in the inference of susceptibility distributions (for Wolb+ in the case) we must have previously reduced the set of survival curves to a set of effectively infected proportions - one entry per challenge dose. To search for a range of days in which absolute mortality might provide an approximate indication of infection, we compare the estimated proportions effectively infected by each challenge dose with the mortality proportion measured at each day. Using a normalized Euclidean distance between these two measures, a day-selection score is provided by the red curve in Figure 6. We identify day 30 as optimal and 17–46 as the interval of days in which the score is at least 95% of the optimal. Reassuringly, the optimal day appears to coincide with the saturation of infection-induced mortality (see position of vertical dash-dotted gray line in relation to the Gamma distributions). We now recall Figure 2 and Table 2 for the inferences based on day 30 mortality data to confirm that classic dose-response models can in principle infer susceptibility distributions that are consistent with those obtained under our extended model (Figure 5). A major issue, however, is that results are sensitive to a day-selection criterion that relies on having previously carried out the entire procedure. The appearance of a high susceptibility group in distributions inferred at later days are an artifact due to the accumulation of background mortality that should be factored out. These results highlight the importance of adequately representing the time dimension in the analysis.
Dose-response models have become standard quantitative frameworks in microbial risk assessment. Less recognized is their ability to estimate host trait distributions. Here we illustrate the concept by extracting host susceptibility distributions from mortality measured as a function of pathogen challenge dose, but similar procedures can be developed for measures of infection or infectiousness (instead of mortality), and can be made a function of other environmental variables such as temperature or humidity (instead of dose). Understanding how to detach host trait distributions from environmental variables is crucial for the formulation of measures that can be transported between laboratory and natural conditions [29], [30].
We address this problem with an experimental design and inference framework that enables the estimation of distributions of host susceptibility to infection by analyzing simultaneously a collection of survival curves, each representing a different challenge dose (Figure 3). The procedure is illustrated on a specifically collected dataset where two distinct groups of hosts (D. melanogaster) were experimentally challenged by viruses (DCV): one group consists of isogenic flies where no significant variability in susceptibility to infection is found; and another with the same genetic background but now carrying the symbiont bacterium Wolbachia known to reduce susceptibility to DCV [21], [22].
Our inferences indicate that Wolbachia confers on average 85% DCV protection to D. melanogaster under the specified laboratory conditions, and suggest significant variability in this effect. This variance in susceptibility is induced by the symbiont, since model selection criteria did not support heterogeneity in the susceptibility of flies not carrying Wolbachia. Since the Drosophila and Wolbachia populations used in this study are isogenic, the heterogeneity in susceptibility of Wolbachia-carrying flies uncovered here indicates variation in the host-microorganism interaction that lacks a genetic basis. A simple hypothesis is that variance in Wolbachia levels at the individual host level leads to variance in resistance to viruses. Although several lines of evidence support this hypothesis [31]–[34], further experiments are required to discriminate whether heterogeneity in resistance is directly linked to variance in Wolbachia levels or, alternatively, a result of another environmental/physiological variance that is only expressed in the presence of Wolbachia.
Previous estimates of protection were based on survival analysis or viral titres in a dose-specific manner [21], [22], [24]. To our knowledge, the experimental design and analysis presented here provides the first estimation of protection in way that is detached from challenge dose. Future developments might consider: estimation of alternative distributions to compare with the shapes suggested by the Beta family; extension of the adopted experimental design to measure responses other that mortality; and move towards host populations and environmental conditions that are closer to natural systems.
The parameters estimated here should not be seen as isolated from the relevant ecological context. On the contrary, they are intended as a first step to inform the construction of ecological and epidemiological models where Wolbachia, other symbionts, or interventions that modify host susceptibility to infection, are introduced to induce desired transitions in populations. The introduction of Wolbachia into Aedes aegypti and other arthropod vectors is being considered as a promising strategy to control dengue and other infectious diseases of humans (see [35] and references therein). The inference frameworks presented can be readily adapted to provide accurate quantification of Wolbachia-induced protection and integrated in population models of public health importance.
The challenge of considering the time dependence of processes leading to observable ecotoxicity responses has also been addressed in toxicology where the so-called General Unified Model of Survival (GUTS) has been proposed [18]. These models simulate the time-course of external and internal processes leading to toxic effects on organisms to generate an output that can be fitted to mortality over time. While those studies tend prioritize the mechanistic descriptions of the toxicokinetic and toxicodynamic processes that damage the organisms, we have chosen to adopt a phenomenological approach and focus on the inference and interpretation of how susceptibility to infection is distributed in a population.
In epidemiological systems, the baseline transmission intensity is often not directly measurable but indirectly inferred in a model-based manner. Dose-response models, on the other hand, can account for experimentally controlled patterns of exposure [36], [37]. Variation in host susceptibility to pathogens is one component of both classes of systems that mostly influences estimates of intervention impacts [29]. Therefore, building on the methods developed here furthers our potential to accurately evaluate the burden of infectious diseases and design effective interventions.
|
10.1371/journal.pgen.1007394 | Whole exome sequencing reveals HSPA1L as a genetic risk factor for spontaneous preterm birth | Preterm birth is a leading cause of morbidity and mortality in infants. Genetic and environmental factors play a role in the susceptibility to preterm birth, but despite many investigations, the genetic basis for preterm birth remain largely unknown. Our objective was to identify rare, possibly damaging, nucleotide variants in mothers from families with recurrent spontaneous preterm births (SPTB). DNA samples from 17 Finnish mothers who delivered at least one infant preterm were subjected to whole exome sequencing. All mothers were of northern Finnish origin and were from seven multiplex families. Additional replication samples of European origin consisted of 93 Danish sister pairs (and two sister triads), all with a history of a preterm delivery. Rare exonic variants (frequency <1%) were analyzed to identify genes and pathways likely to affect SPTB susceptibility. We identified rare, possibly damaging, variants in genes that were common to multiple affected individuals. The glucocorticoid receptor signaling pathway was the most significant (p<1.7e-8) with genes containing these variants in a subgroup of ten Finnish mothers, each having had 2–4 SPTBs. This pathway was replicated among the Danish sister pairs. A gene in this pathway, heat shock protein family A (Hsp70) member 1 like (HSPA1L), contains two likely damaging missense alleles that were found in four different Finnish families. One of the variants (rs34620296) had a higher frequency in cases compared to controls (0.0025 vs. 0.0010, p = 0.002) in a large preterm birth genome-wide association study (GWAS) consisting of mothers of general European ancestry. Sister pairs in replication samples also shared rare, likely damaging HSPA1L variants. Furthermore, in silico analysis predicted an additional phosphorylation site generated by rs34620296 that could potentially affect chaperone activity or HSPA1L protein stability. Finally, in vitro functional experiment showed a link between HSPA1L activity and decidualization. In conclusion, rare, likely damaging, variants in HSPA1L were observed in multiple families with recurrent SPTB.
| Preterm birth is the leading cause of infant mortality, and prematurity is further associated with serious morbidities in later life. Genetic and environmental risk factors play a role in the susceptibility to preterm birth. Despite numerous studies, the genetic basis for preterm birth remains poorly defined. We investigated the presence of rare, possibly risk associated nucleotide variants in mothers with spontaneous preterm births (SPTB). The first set of mothers with family history of recurrent preterm births was of northern Finnish origin. An additional set of mothers (sister pairs, both giving birth preterm) of European origin was also studied. Whole exome sequencing identified multiple rare, likely damaging HSPA1L variants in several families affected by SPTB, and this gene was associated with the glucocorticoid receptor signaling pathway. Potential involvement of one of the HSPA1L variants in SPTB was further supported by large GWAS dataset. In addition, this variant alters protein post-translational modification potential, and thus may affect protein stability and its function as a chaperone.
| Preterm birth (PTB), defined as birth before 37 completed weeks of gestation, is a major global public health concern. Worldwide, over 15 million infants (more than one in ten babies) are born preterm and of those, more than one million die from complications related to preterm birth each year [1]. Preterm birth and its complications are the leading cause of neonatal deaths and have become the major cause of death among children under five years old [2]. Moreover, preterm infants are at increased risk, not only of short-term complications but also of life-long disabilities, such as respiratory and cognitive disorders [1]. Preterm birth also increases the risk of adult-onset disorders, such as obesity, diabetes and cardiovascular diseases [3, 4]. Currently, there is no generally effective method for prevention of preterm delivery.
The majority (~70%) of preterm births occur after spontaneous onset of labor, with or without preterm prelabor rupture of the membranes (PPROM) [5]. Most spontaneous preterm births (SPTBs) are idiopathic [1, 5]; however, recurrence of preterm birth among mothers and within families indicates that genetic factors may be important. Genetic factors are estimated to account for 25–40% of the variation in birth timing [6], with the maternal genome playing the major, but not only, role in predisposition to preterm birth [7–11]. Despite many studies of the genetics of SPTB [6, 12, 13], only a few variants have been robustly associated with this outcome [14], and their functional implications are unclear.
Previous genome-wide association studies (GWAS) of SPTB have involved common variants, but they explain only a small portion of the genetic risk. The role of rare variants in SPTB has been essentially unexplored. Whole exome sequencing (WES) in families offers a comprehensive method to identify rare variant associations with disease, including almost complete coverage of the protein coding regions of the genome. Even though studies of rare variants underlying Mendelian disorders have revealed novel genes [15, 16], using WES to study complex multifactorial syndromes remains a challenge [17]. Previous sequencing studies of PTB [18] or PPROM [19, 20] have focused only on a set of candidate gene regions and, consequently, have missed the majority of the coding regions of the genome. In contrast to whole genome sequencing, WES is more cost effective and has the advantage of providing more easily interpreted results.
We performed a WES study using families under the hypothesis that familial recurrence is influenced by rare variants with large individual effects on SPTB susceptibility. Such an approach has the potential of identifying genes containing rare variants shared in these multiplex families, as well as genes in pathways common across families. This method applies a hypothesis-free testing approach to identify potentially novel candidate genes for SPTB.
Seventeen mothers from seven northern Finnish multiplex families (Discovery cohort) and an additional 192 mothers from 95 Danish families (Replication cohort) were sequenced using WES. The pedigrees of the multiplex Finnish families are shown in S1 Fig. For the Discovery cohort, all samples (except one that was excluded from subsequent analyses) passed the quality control parameters used for the clinical exome sequencing at the CMH; quality control cutoffs were 85% reads aligned, 80% aligned with alignment quality of 20 or greater. For these samples, the mean and median heterozygous/homozygous variant ratios were 0.765 and 0.748, respectively. Prior to variant filtering in Ingenuity, the mean/median numbers of nucleotide variant calls per individual were 318,767/326,474 (Discovery cohort) and 221,682/184,381 (Replication cohort). This difference in variant calls between the Discovery and Replication populations is likely due the fact that populations were sequenced using different Next Generation Sequencing platforms, Illumina for Discovery cohort and Complete Genomics for Replication cohort, and their respective primary quality control measures and variant calling methods were thus different. Mean transition/transversion (Ti/Tv) ratios were 2.2 and 2.0 for Discovery and Replication exomes, respectively. An overview of the WES workflow is presented in Fig 1.
Three software programs (Ingenuity Variant Analysis, Varseq and the CMH Variant Warehouse) were used to assess common shared (by affected mothers per family) rare variants. Only those variants that passed the prioritizing steps with at least two of the annotating software tools were considered valid and are described below (summarized in S4 Table). The benefits of comparing data obtained from multiple software is that it minimized the possibility of picking up falsely called variants that passed quality control filters only by one software. This approach resulted in a total of 844 variants in the Discovery population. For the Replication population, we combined and compared the shared rare variants passing the annotation and prioritizing steps of Ingenuity Variant Analysis and Varseq; a total of 8431 variants passed the filters of both software tools. The CMH Variant Warehouse was not available for the Replication set. For both populations, variants were categorized as loss of function, moderate, or other, according to their predicted consequences, i.e. pathogenicity (S5 Table). We further compared the list of variants resulting from the family-based analyses (as described above) between the Discovery and Replication populations. Numbers of common genes and variants for both populations are shown in S2 Fig. There were 72 rare variants that were found in both populations in 72 genes (S6 Table).
Rare single nucleotide variants from HSPA1L [heat shock protein family A (Hsp70) member 1 like], identified by the Discovery Ingenuity pathway analysis, were further investigated using imputed GWAS data that also included variants with MAF <1%. In the Discovery set, variants in AR, NCOA3 and NCOR2 were either CAG repeat length polymorphisms, in-frame deletions or insertions, respectively, and were, therefore, not investigated in the GWAS datasets. Three independent GWAS datasets were used, one of general European ancestry containing more than 40,000 mothers of live births (23andMe dataset) and two from Northern Europe containing 4,600 and 600 mothers (Nordic and northern Finnish datasets, respectively). In the large 23andMe preterm birth GWAS dataset, the minor allele of rs34620296 in HSPA1L, which is in the glucocorticoid receptor signaling pathway, was found to be more common in cases than in controls (case frequency 0.0025 vs. control frequency 0.0010, p = 0.002; Table 3). This association was also significant for gestational age as a continuous trait (gestational age as weeks; p = 0.0016, effect -0.8238, standard error 0.2608). The HSPA1L variants from the Discovery (rs34620296 and rs150472288) and the Replication (rs482145, rs139193421) analyses are listed in detail in Table 3. In the two smaller GWAS datasets, however, these four HSPA1L variants were absent or not significant. Lack of significance may be due to smaller numbers of individuals, especially in cases.
Sanger sequencing confirmed the genotypes of the two rare HSPA1L missense variants (rs34620296 and rs150472288) in the samples from the Discovery cohort. The rare HSPA1L variants were observed in a total of six mothers from four unrelated families. Additional family members with available DNA were sequenced for these variants. Interestingly, in two of the families, female carriers of the maternally inherited rs34620296 minor T-allele were born preterm, whereas in the other two unrelated families the male carriers of maternally inherited rs150472288 minor T-allele were born preterm. However, numbers of minor allele carriers are too small for any definite gender related conclusions.
Pathogenicity predictions for rs34620296 and rs150472288 derived from the Discovery cohort as well as for rs482145 and rs139193421 from the Replication cohort were assessed using in silico tools SIFT and PolyPhen-2, and all these variants were predicted as damaging and probably/possibly damaging, respectively (Table 4). In addition, MutationTaster and MutationAssessor predicted all four variants as disease causing and predicted functional (high), respectively. According to the Combined Annotation Dependent Depletion (CADD) score (>20), all of these variants, except for rs139193421, are among the top 1% of deleterious variants in human genome (Table 4).
To assess potential consequences of these variants on transcriptional activity, we evaluated them for evidence of histone modification or DNase I hypersensitivity. In silico tools HaploReg 4.1 and/or RegulomeDB showed that all four variants were in regions that had histone marks, as well as strong transcriptional regulatory signatures in various cells of the immune system, especially in T lymphocytes from peripheral blood (Table 4). Evidence of an active transcription start site was predicted in the HeLa-S3 Cervical Carcinoma Cell Line for rs34620296 and in foreskin fibroblast primary cells for rs482145 (Table 4). Further evidence of active DNA accessibility (DNAse) was found in ovarian tissue for rs150472288, and in psoas muscle tissue for rs139193421 (Table 4). There was also evidence of a transcriptional effect of rs34620296 and rs150472288 (Discovery) in ovary and fetal adrenal gland. Together these results from HaploReg 4.1 and RegulomeDB provide evidence for the potential involvement of HSPA1L variants in the endocrine system, as well as in the adaptive immune cells. These variants could, therefore, have a role in the etiology of SPTB.
We further investigated putative effects of HSPA1L rs34620296 on protein structure. This variant was selected due to its association with SPTB in the large 23andMe GWAS dataset. This variant causes an amino acid change from Alanine to Threonine at position 268 (Ala268Thr). According to the NetPhos 3.1 in silico prediction, Ala268Thr generates an additional phosphorylation site next to an existing phosphorylation site (T267-p) (S3 Fig). Furthermore, Ala268Thr is near an adenosine triphosphate (ATP) nucleotide-binding site located downstream at position 270−277 (Fig 2A). Gain of phosphorylation may cause changes in binding energy, modulate physio-chemical properties or stability kinetics and dynamics of the protein functions such as strength of protein-protein interactions [22].
To investigate the possible effects of the missense variant on protein structure, the reference HSPA1L protein structure and a structure including the Ala268Thr variant were compared simultaneously using UCSF Chimera. There was not a visible change in the overlaid protein structures (Fig 2B). Instead, there was a slight change in the chemical bond lengths (≥0.002Å) of the adenosine diphosphate (ADP)-ligand binding amino acid side chains at positions Glu270, Arg274 and Asp368, shown in the 3D model of the HSPA1L (Fig 2C). This may be due to the change from a small size, and hydrophobic, (Ala) to medium size, and polar, (Thr) residue. Such a change in the amino acid side chains could affect the binding efficiency of the ADP molecule.
To further explore possible underlying biological functionality, we investigated the tissue expression established via HSPA1L, along with AR, NCOA3 and NCOR2, using HumanBase (http://hb.flatironinstitute.org). HSPA1L was expressed in placental tissue with reasonable confidence (0.65), and in ovarian (0.57) and fetal tissues (0.48) as well as in uterus (0.29). For HSPA1L, AR, NCOA3 and NCOR2 together, the average expression confidence was high in placenta (0.74), ovary (0.70), fetus (0.65), and moderate in uterus (0.29), indicating high confidence for expression in female reproductive system overall (S4 Fig).
To determine whether the HSPA1L Ala268Thr (rs34620296) variant alters activity of the GR signaling pathway, we analyzed the consequences of glucocorticoid exposure during decidualization. Human endometrial stromal fibroblasts were transfected with plasmids containing either WT or Ala268Thr cDNA, or with empty vector serving as control. The cells were treated with decidualization media for 72h in a presence of glucocorticoids (100nM dexamethasone) as a surrogate of stress. Protein levels of HSPA1L and GR, as well as mRNA levels of Wnt Family Member 4 (WNT4) were measured. Cells transfected with the WT HSPA1L-pcDNA3.1 trended to greater increases in cytosolic HSPA1L protein content than those transfected with the Ala268Thr HSPA1L-pcDNA3.1 (mean ± SEM; 1.272 ± 0.142 vs. 0.893 ± 0.146, respectively, p = 0.09) (Fig 3). Furthermore, the Western blot analysis showed that the relative cytosolic protein levels of GR differed significantly between the WT and Ala268Thr groups with more GR present in the WT group than in the Ala268Thr group (mean ± SEM; 1.309 ± 0.099 vs. 0.993 ± 0.096, respectively, p = 0.04) (Fig 3; numerical data available in S7 Table).
Next, we determined the relative gene expression of WNT4 by qPCR. WNT4 is a critical decidualization target found in the recent GWA study [14] associated with gestational length. Increased expression of WNT4 was observed in the WT group, whereas, the Ala268Thr group was less able to activate the WNT4-signaling pathway leading to a lower expression of WNT4 (p = 0.04).
To move beyond traditional case-control GWAS and family-based linkage studies, we performed a case-only whole exome sequencing study designed to investigate the burden of rare variants in families with recurrent SPTB. Whole exome sequencing enables the discovery of rare, putatively functional variants associated with the etiology of complex disease on a gene-by-gene or a pathway-by-pathway basis, and enrichment in multiplex families provides a means to filter large-scale sequencing data.
Comparisons of mothers with recurrent preterm deliveries identified the glucocorticoid receptor signaling pathway as a candidate for mediating the risk of SPTB. Specifically, within this pathway, likely pathogenic missense variations in HSPA1L were found among four unrelated Finnish families (rs34620296 and rs150472288), and within Danish sister pairs (rs482145 and rs139193421). Notably, the rs34620296 minor allele variant was observed at a higher frequency in cases than controls in a very large 23andMe GWAS set. These variants were also identified via bioinformatics analyses as likely affecting either protein function or expression. Further functional evidence linked HSPA1L activity and decidualization.
HSPA1L is a member of the Hsp70 superfamily and is near HSPA1A and HSPA1B within the major histocompatibility complex class III region on chromosome 6. The HSPA1L protein (also known as Hsp70-hom) is ~90% identical to HSPA1A and HSPA1B, also known as Hsp70-1 and Hsp70-2, respectively [23, 24]. Heat shock proteins (HSPs) are highly conserved cellular defense mechanisms for cell survival and are present in all cell types in all organisms. Some HSPs are expressed constitutively, while others are stress-induced (e.g. heat, hypoxia, oxidative stress, infection and inflammation) [25, 26]. Intracellular HSPs act as molecular chaperones and, together with co-chaperones, stabilize existing proteins against aggregation, mediate folding of newly translated proteins, and assist in protein translocation across intracellular membranes [25, 27]. HSPs are categorized into families according to their approximate molecular weight; of which Hsp70 (a group of proteins sized approximately 70 kDa) is the best characterized. Potential involvement of stress-induced HSPA1A in adverse pregnancy outcomes, including preeclampsia and PTB, has previously been suggested [28, 29]. Although, studies of the role of HSPA1L and HSPA1L in pregnancy are lacking, there is some evidence of involvement in adverse pregnancy outcomes such as preeclampsia [30].
The rare HSPA1L missense variants observed in our study, are in the nucleotide-binding domain (NBD), except the rs482145, which is in the substrate-binding domain (SBD) (Fig 2). ATP binds to the NBD, which is followed by the exchange from low-binding affinity ATP state to high-binding affinity ADP state [29, 31, 32]. We showed that the non-synonymous variant rs34620296 (Ala268Thr) generates an additional phosphorylation site near the nucleotide-binding site. It showed a modest change in the binding efficiency at this site, which could affect the interaction with ADP or HSPA1L stability itself, as suggested by our transfection studies. In agreement with our findings, a previous study of Caucasian patients with inflammatory bowel disease found that rare mutations in HSPA1L were significantly enriched in patients but absent in healthy controls [33]. Interestingly, one of the associated rare variants was Ala268Thr, and further in vitro biochemical assays of the recombinant HSPA1L showed reduced chaperone activity with this variant [33]. There is also evidence that possibly connects inflammatory bowel disease to adverse perinatal outcomes [34]. Additionally, a previous SPTB study in African Americans found a common nonsynonymous HSPA1L variant, rs2075800, to associate with SPTB [35]. Furthermore, a meta-analysis of previously PTB associated genes linked HSPA1L and SPTB using Ingenuity Pathway Analysis [36].
Due to a very low incidence of the rare HSPA1L variants associating with SPTB in our study, the anticipated attributable risk in the population level is probably small. However, the identification of the damaging alleles may facilitate the identification of causative pathways. For instance, interaction between Hsp70 and Hsp90 chaperones as well as their co-chaperones is essential in the maturation and inactivation of nuclear hormone receptors (e.g. glucocorticoid, androgen, estrogen and progesterone receptors) [37, 38]. In the absence of its ligand, glucocorticoid receptor (GR) is bound to a complex constituting of Hsp40, Hsp70 and Hsp90 chaperones; this complex keeps the GR in a ligand-receptive conformation but remaining transcriptionally inactive until ligand binding [38]. As shown previously [33], rare HSPA1L variants can cause partial loss of HSPA1L chaperone activity, and therefore, altered function or expression. Altered function of the chaperones can compromise the stability of the GR complex, leading to an accumulation of partially unfolded proteins that are prone for aggregation and degradation events [37].
Glucocorticoids, steroid hormones that mainly signal through the GR, have anti-inflammatory and immunosuppressive actions. Glucocorticoid signaling communicates with estrogen signaling pathways to tightly regulate the pro- and anti-inflammatory milieu in reproductive tissues [39], and progesterone signaling, via nuclear GR, mediates anti-inflammatory and immunosuppressive effects in genital tract during pregnancy [40, 41]. Sustaining a pregnancy is a complex interplay and balance between the innate and adaptive immune cells in the reproductive tissues and at the maternal-fetal interface. Imbalance between the inflammatory cells can cause a breakdown of maternal-fetal tolerance leading to activation of labor (both term and preterm). An untimely stimulus (e.g. stress, infection or inflammation) together with impairments in the glucocorticoid receptor signaling pathway could impose an inadequate response against inflammation or stress. This can elicit a shift from an anti-inflammatory to pro-inflammatory microenvironment, causing a premature activation of labor initiating signals resulting in preterm birth [42, 43].
Possible limitations of our study are that the Discovery and Replication populations were sequenced using different Next Generation Sequencing platforms, and primary quality control measures and variant calling methods were thus different. In addition, Next Generation Sequencing generates an enormous amount of data, which could lead to many sequencing artifacts that may be misidentified as variants. We attempted to minimize these artifacts by applying a variety of quality control filters and using a large internal control population to detect potential sequencing or annotation errors. We also compared the results of variant annotation and prioritizing filters from three different software tools to ensure reproducible results. Furthermore, reported variants were confirmed by Sanger sequencing. Another possible limitation was that our study did not include unrelated control samples. This limitation has been partly overcome with the use of additional large GWAS datasets including control samples.
In conclusion, whole exome sequencing of families with recurrent occurrence of SPTB enables identification of rare alleles influencing the predisposition to SPTB. Among the individual genes, two minor alleles of HSPA1L had a strong association to SPTB in multiplex Finnish families and the association of a specific minor allele was confirmed in a large GWAS set. Furthermore, this variant was associated with altered modification and function of the protein. Overall, our data suggest the need for precise regulation of steroid signaling in mediating birth timing.
Written informed consent was obtained from all individuals participating in this study, and the study was approved by the Ethics committees of the participating centers: Oulu University Hospital (78/2003, 73/2013), University of Southern Denmark (NVK#1302824), and University of Iowa (IRB#200608748). Individuals in the large European American GWAS were research participants of 23andMe, Inc., a personal genetics company. All 23andMe participants provided informed consent and participated in the research online, under a protocol approved by the external AAHRPP-accredited IRB, Ethical & Independent Review Services (E&I Review).
DNA samples of the 17 individuals from Discovery population were extracted from whole blood and saliva samples using standard methods [45]. Although, using DNA from both blood and saliva samples, there were no major difference in the overall sequencing metrics (alignment metrics or total number of variants) between the sample types. DNA samples were subjected to exon specific next generation sequencing performed at the Center for Pediatric Genomic Medicine, Children’s Mercy Hospital (CMH; Kansas City, MO). Exome samples were prepared with the Illumina Nextera Rapid Capture Exome kit according to the manufacturer’s protocols as described previously [47]. Sequencing was performed on Illumina HiSeq 2500 instruments utilizing v4 chemistry with 2 x 125 nucleotide sequences. Sequence data were generated with Illumina RTA 1.18.64.0 and bcl2fastq-1.8.4, and aligned against the reference human genome (GRCh37.p5) using bwa-mem [48], and variant calls were made using the Genome Analysis Toolkit (GATK) [49] version 3.2–2 using previously described methods [50]. Duplicate reads were identified and flagged with the Picard MarkDuplicates tool. Realignment of reads around known indels was performed with the RealignerTargetCreator and IndelRealigner, and variants were called on individual samples using the HaplotypeCaller modules of the GATK.
In addition, whole exome sequencing was performed on 192 affected individuals from 95 Danish families (Replication set). Exome capture of the samples were carried out with the BGI Exon Kit following manufacturer’s protocols (BGI, Shenzhen, China). DNA libraries were generated using combinatorial Probe Anchor Ligation (cPAL) technology, and 35 base paired end reads were generated from 500 bp genomic fragments. Whole exome sequencing was performed using the Complete Genomics platform (BGI) and using the manufacturer’s pipeline. Reads were aligned against the National Center for Biotechnology Information (NCBI) build 37 reference human genome.
The variant call files (VCF), containing the variant call results, generated by CMH and BGI were analyzed using Ingenuity Variant Analysis Software (Qiagen, Germany) and Golden Helix VarSeq Software v.1.2.1 (Bozeman, MT) for both the Discovery and Replication population sets. Variants were filtered based on variant quality control measurements, frequency and predicted pathogenicity, as well as a dominant inheritance model.
In the Ingenuity Variant Analysis, low quality variants (read depth <15 and call quality <20) were removed. Furthermore, we only included rare variants (i.e. MAF <1% in the 1000 Genomes Project, ExAC or in European American population in NHLBI ESP exomes) and variants that would likely have functional effect (i.e. variants that are predicted by SIFT or PolyPhen-2 as damaging or likely damaging, listed in Human Gene Mutation Database, or associated with gain or loss of function of a gene).
In VarSeq, variants with read depth <15 and genotype quality score <20 were excluded. Only rare variants in Europeans (MAF <1% or absent; 1kG Phase 3: Variant frequencies 5, GHI Jan 2015) and missense or loss-of-function variants were included for analyses. Further filtering was applied to the data obtained from VarSeq. Variant quality was enhanced by applying range criteria (0.3–0.85) for alternative allele frequency (i.e. ratio of alternate allele read depth / alternate allele read depth + reference allele read depth); variants outside this range were excluded. Allele frequencies were searched from the Sequencing Initiative Suomi (SISu) database (www.sisuproject.fi) for variants originating from the Finnish mother data, and from the Exome Aggregation Consortium (ExAC) database (http://exac.broadinstitute.org/) for Danish sister pair data. For these variants, a MAF cut-of value <1% in Finnish general population (SISu) or European “non-Finnish” (ExAC) population was used for Finnish or Danish mothers, respectively. The SISu database was used for Finnish mothers to exclude rare variants that are enriched in Finnish general population compared to the rest of the Europeans.
For the Discovery samples, we also used allele frequency calculations derived from Center for Pediatric Genomic Medicine’s CMH Variant Warehouse database (http://warehouse.cmh.edu) including ~3900 individuals previously sequenced at the center [50]. Pathogenicity was categorized according to the American College of Medical Genetics [21] as 1; previously reported to be disease-causing, 2; expected to be pathogenic (loss of initiation, premature stop codon, disruption of stop codon, whole-gene deletion, frame shifting indel, and disruption of splicing), and 3; unknown significance but potentially disease-causing (nonsynonymous substitution, in-frame indel, disruption of polypyrimidine tract, overlap with 5' exonic, 5' flank, or 3' exonic splice contexts). Only variants that fit one of these criteria (1−3) were included for analyses. Rare and novel variants with relatively high frequency in this internal control population were also excluded as they were thought to be technical artifacts.
In Ingenuity Variant Analysis, a dominant inheritance model (including gain of function variants, and all heterozygous, compound heterozygous, haploinsufficient, hemizygous, and het-ambiguous variants) was used to investigate predisposing variants that are inherited in the families. When analyzing affected mothers as a group, rare variants in genes that were common for a proportion of all cases were investigated. Whereas in family specific analyses, only variants that were shared by the affected individuals within each family were included.
Ingenuity Variant Analysis provides a list of most significant pathways calculated specifically for each filtering output. P-values were calculated according to Fisher’s exact test assessing overlap enrichment of dataset-variant genes relative to known phenotype-implicated genes. Here, only pathways with p<0.01 were included for further analyses.
To investigate rare variants in genes arising from the whole exome data in a larger population setting including controls, we used available sources of preterm birth GWAS data. GWAS data from a large cohort, identified among 23andMe’s research participants, included 43,568 mothers of general European ancestry [14] and meta-analysis data including 4,632 mothers from three independent Nordic (Finnish, Danish, and Norwegian) birth cohorts of European ancestry [51]. In addition, a set of GWAS data from a total of 608 mothers passing quality control measures was available. This set included mothers with spontaneous preterm deliveries and mothers with term deliveries originating exclusively from northern Finland. Genotyping was performed with Illumina Human CoreExome chip, followed by prephasing and imputation procedures with ShapeIT2 [52] and IMPUTE2 [53]. Association analysis was performed using SNPTEST v. 2.5.2 [54].
Since WES methodologies are associated with significant false positive rates, the presence of interesting variant findings from WES analyses was confirmed using Sanger sequencing. Samples were sequenced using capillary electrophoresis with ABI3500xL Genetic Analyzer (Applied Biosystems, CA) in Biocenter Oulu Sequencing Center, University of Oulu, Oulu, Finland. Details of the PCR primers and reaction conditions are available upon request.
The possible functional effect of the rare HSPA1L variants (rs34620296 and rs150472288 from Discovery analyses as well as rs482145 and rs139193421 from Replication analyses) were investigated using in silico prediction tools such as SIFT, PolyPhen-2, MutationTaster and MutationAssessor. These pathogenicity predictions were annotated using Varseq, whereas CADD scores to identify pathogenic and deleterious variants were obtained from Ingenuity. Variants with CADD score >20 are amongst top 1% of deleterious variants in human genome [55]. RegulomeDB (http://www.regulomedb.org/) was used to investigate variant locations for e.g. chromatin state activity. In addition, we used HaploReg v4.1 (http://archive.broadinstitute.org/mammals/haploreg/haploreg.php) to assess whether variants are located within regions that show evidence for promoter or enhancer activity (i.e. presence of histone modification marks H3K4me1 and H3K27ac that are associated with enhancer regions, or H3K4me3 and H3K9ca that are associated with promoter regions), as well as for DNase I hypersensitivity in human tissues and cell line samples.
We further investigated the potential effects that missense variation rs34620296 (Ala268Thr) could have on protein sequence or structure. We used NetPhos 3.1 [56] to investigate possible changes in phosphorylation events in HSPA1L sequence. NetPhos 3.1 predicts serine, threonine or tyrosine phosphorylation sites in amino acid sequences of eukaryotic proteins. Evidence of being a phosphorylation site is given when the score is above the threshold (0.5). To investigate the possible effects of missense variant in protein structure, the reference protein structure and the modified protein structure, including the missense variant, were compared. Original (UniProtKB: P34931) and modified (Ala268Thr) amino acid sequences were submitted to SWISS-MODEL (https://swissmodel.expasy.org/) for protein modeling. Resulting protein models were compared simultaneously using UCSF Chimera. Molecular graphics and analyses were performed with the Chimera-1.11.2. package (https://www.cgl.ucsf.edu/chimera/). Chimera was developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco (supported by NIGMS P41-GM103311) [57].
|
10.1371/journal.pcbi.1000567 | A Threading-Based Method for the Prediction of DNA-Binding Proteins with Application to the Human Genome | Diverse mechanisms for DNA-protein recognition have been elucidated in numerous atomic complex structures from various protein families. These structural data provide an invaluable knowledge base not only for understanding DNA-protein interactions, but also for developing specialized methods that predict the DNA-binding function from protein structure. While such methods are useful, a major limitation is that they require an experimental structure of the target as input. To overcome this obstacle, we develop a threading-based method, DNA-Binding-Domain-Threader (DBD-Threader), for the prediction of DNA-binding domains and associated DNA-binding protein residues. Our method, which uses a template library composed of DNA-protein complex structures, requires only the target protein's sequence. In our approach, fold similarity and DNA-binding propensity are employed as two functional discriminating properties. In benchmark tests on 179 DNA-binding and 3,797 non-DNA-binding proteins, using templates whose sequence identity is less than 30% to the target, DBD-Threader achieves a sensitivity/precision of 56%/86%. This performance is considerably better than the standard sequence comparison method PSI-BLAST and is comparable to DBD-Hunter, which requires an experimental structure as input. Moreover, for over 70% of predicted DNA-binding domains, the backbone Root Mean Square Deviations (RMSDs) of the top-ranked structural models are within 6.5 Å of their experimental structures, with their associated DNA-binding sites identified at satisfactory accuracy. Additionally, DBD-Threader correctly assigned the SCOP superfamily for most predicted domains. To demonstrate that DBD-Threader is useful for automatic function annotation on a large-scale, DBD-Threader was applied to 18,631 protein sequences from the human genome; 1,654 proteins are predicted to have DNA-binding function. Comparison with existing Gene Ontology (GO) annotations suggests that ∼30% of our predictions are new. Finally, we present some interesting predictions in detail. In particular, it is estimated that ∼20% of classic zinc finger domains play a functional role not related to direct DNA-binding.
| DNA-binding proteins represent only a small fraction of proteins encoded in genomes, yet they play a critical role in a variety of biological activities. Identifying these proteins and understanding how they function are important issues. The structures of solved DNA protein complexes of different protein families provide an invaluable knowledge base not only for understanding DNA-protein interactions, but also for developing methods that predict whether or not a protein binds DNA. While such methods are useful, they require an experimental structure as input. To overcome this obstacle, we have developed a threading-based method for the prediction of DNA-binding domains and associated DNA-binding protein residues from protein sequence. The method has higher accuracy in large scale benchmarking than methods based on sequence similarity alone. Application to the human proteome identified potential targets of not only previously unknown DNA-binding proteins, but also of biologically interesting ones that are related to, yet evolved from, DNA-binding proteins.
| The past decade has witnessed tremendous progress in genome sequencing [1]–[5]. According to the Genomes On Line Database, the complete sequenced genomes of almost 1,000 cellular organisms have been released, and about 5,000 active genome sequencing projects are on the way [6]. The unprecedented amount of genetic information has provided hundreds of thousands of protein sequences [7]. This poses a challenging problem to elucidate their functions, as experimental characterization of all newly sequenced proteins is obviously impractical. Fortunately, many of them are homologous to proteins that have been experimentally studied. Consequently, it would be highly desirable to develop computational approaches that automatically annotate a new protein sequence through its functionally characterized homologs [8]–[10]. The key component of such approaches is the ability to detect homologous relationships between un-characterized and characterized proteins. For this purpose, many sequence and structural similarity comparison methods have been developed [11]–[15]. While sequence-based methods are powerful and widely adopted for function inference [16]–[18], structure-based methods are more sensitive in detecting homologs with low or no sequence similarity [19]–[21]. However, significant sequence or structural similarity does not necessarily lead to identical function, since the functional roles of related proteins can diverge during the course of evolution [22],[23]. To address this problem, it is often necessary to examine the conservation of functionally discriminating residues when predicting enzymatic functions [17], or to evaluate the interaction energy when predicting protein-protein [24] or protein-DNA interactions [19].
DNA-binding function is a key characteristic of many proteins involved in various essential biological activities; these include DNA transcription, replication, packaging, repair and rearrangement. These DNA-binding proteins have a diversified classification according to their structures and the way they interact with DNA [25],[26]. Due to the importance of DNA-binding proteins, a few dedicated computational approaches have recently been proposed for the prediction of DNA-binding function from protein structure [19], [27]–[31]. These methods can be classified into two groups: structure template-based and template-free, depending on how (or if) they use the information from the known structures of DNA-binding proteins. Template-based methods utilize a structural comparison protocol to detect significant structural similarity between the query and a template known to bind DNA at either the domain or the structural motif level and use a statistical or electrostatic potential to assess the DNA-binding preference of the target sequence [19],[29]. The latter assessment reduces the number of false positives, which is important for the success of these methods. Template-free methods do not perform direct structural comparison, but typically follow a machine-learning framework and use features such as sequence composition and biophysical properties of surface patches [27],[28],[30],[31]. Although they can potentially detect a novel DNA-binding protein fold, template-free methods generally have lower accuracy than template-based methods, which perform well on large-scale datasets and have been applied to structural genomics targets [19].
In addition to DNA-binding function, it is also of interest to predict the amino acids that directly participate in DNA-binding. This is often straightforward for a template-based approach, as one can infer the binding residues directly from the identified template [19]. By comparison, in a template-free approach, one needs to design a new prediction protocol [32]–[36]. Recently, a new approach has been developed to predict DNA-binding residues through DNA-protein docking [37]. This approach, which takes the advantage of the non-specific DNA-binding ability of DNA-binding proteins, provides a coarse model of the DNA-protein complex in addition to the prediction of DNA-binding sites.
Although structural information is helpful for predicting DNA-binding function, it can also limit the scope of application because less than 1% of all proteins have an experimentally determined structure [38]. To overcome this limitation, we introduce a threading-based method, DBD-Threader, for the prediction of DNA-binding domains and associated functional sites. Threading-based approaches, which require only sequence as query input, have been successfully applied to the prediction of protein-protein interactions [24],[39] and protein-ligand interactions [40]. Below, we first describe the framework of our approach, and then compare its performance with three established methods, including the standard sequence alignment tool PSI-BLAST [11], the threading method PROSPECTOR [41], and the experimental structure-based DNA-binding prediction method DBD-Hunter [19]. Finally, we present the application of DBD-Threader to the human genome, for which DBD-Threader detected ∼7,000 DNA-binding domains in 59 SCOP superfamilies. We also predict that ∼20% of classic zinc finger domains play a functional role not related to direct DNA-binding.
We briefly review the general strategy of DBD-Threader (see Methods for details). Fold similarity and DNA-binding propensity are two properties employed for inferring function. Fold similarity is evaluated by a threading procedure, and the DNA-binding propensity is calculated using a statistical DNA-protein pair potential. Given the sequence of a target protein, the method first threads the sequence against a template library composed of DNA-binding protein domains whose structures have been experimentally determined in complex with DNA. Significant template hits obtained through threading, if any, are further evaluated by the DNA-protein interaction energy, calculated using the target/template alignment and the corresponding DNA structure complexed with the template protein. If a target protein has at least one significant template that satisfies both the specified Z-score and energy threshold conditions, the protein is predicted as DNA-binding and as non-DNA-binding otherwise. The threshold conditions are optimized through benchmark tests. For predicted DNA-binding proteins, DBD-Threader further assigns the SCOP superfamily to identified DNA-binding domains, provides structural models, and infers the DNA-binding protein residues according to the top-ranked template. A web-server implementation of the method is available at http://cssb.biology.gatech.edu/skolnick/webservice/DBD-Threader/.
DBD-Threader uses fold similarity evaluated by the threading Z-score, and DNA-binding propensity, evaluated by DNA-protein interaction energy, as two properties to discriminate DNA-binding proteins from non-DNA-binding proteins. The effectiveness of these two properties are demonstrated through an analysis of 179 DNA-binding proteins (DB179) and 3797 non-DNA-binding proteins (NB3797); two non-redundant datasets collected previously [19]. The sequences of these ∼4,000 proteins were used as input. For each target, we excluded from the library any template whose sequence is more than 30% identical to the target, since we are mostly interested in detecting homologs at low sequence identity.
A significant threading Z-score for a pair of target/template proteins typically suggests a high level of structural similarity. Since two proteins with similar structures more likely share the same function than those in different structures, the threading Z-score can serve as a good indicator not only for structure similarity, but also for function similarity. As shown in Figure 1A, 70% (126) of the proteins in DB179 hit at least one template from the DNA-binding domain library with a significant Z-score>6; 25% (44) hit with a high Z-score>20. By contrast, only 3.9% (149) proteins of NB3797 hit at least one template with a Z-score>6, and only two targets from NB3797 hit a template with a high Z-score>20. These results suggest that one can utilize threading to filter out the vast majority of non-DNA-binding proteins, while keeping many homologs with DNA-binding function. However, since the numbers of proteins with a significant hit are about the same in the DNA-binding and the non-DNA-binding protein sets, about half of the predictions would be incorrect if one chooses a Z-score of 6 as the threshold to determine the DNA-binding function. One can raise the threshold to a high Z-score of 20, which would greatly improve the precision of the predictions to 96%. But, it would also reduce dramatically the sensitivity (coverage) of the predictions to only 25%. Thus, use of threading alone has a limited accuracy when applied for functional inference.
To further improve the precision without seriously compromising the sensitivity of the predictions, we introduce a DNA-protein statistical pair potential to assess DNA-binding propensity [19]. It has been shown that this term can be used to differentiate DNA-binding protein residues from non-DNA-binding residues, independent on the specific DNA substrates involved [19],[37]. If a pair of target/template proteins has similar structure, then the target protein might favorably interact with the template DNA in a similar way as the template protein. This assumption is generally valid, as shown in the distributions of the DNA-protein interaction energy of targets with at least one significant (Z-score>6) template (Figure 1B). For each target, the lowest energy is shown if more than one significant hit is identified. 94 of 126 targets from DB179 have attractive DNA-protein interaction energy values <−5, whereas only 28 of 149 targets from NB3797 have an energy value <−5. The analysis suggests that a functional relationship between remote homologs can be established at quite high precision through a combination of threading and interaction energy calculations, which is the strategy adopted by DBD-Threader.
To benchmark the performance of our approach, DBD-Threader is compared with three methods: PSI-BLAST [11], PROSPECTOR [41], and DBD-Hunter [19]. Two sequence libraries from NCBI and from UniProt were used to derive the position specific sequence profile for PSI-BLAST, respectively. Details of the assessment procedures are given in Methods. Figure 2A shows the precision-recall (PR) and Figure 2B shows the Receiver Operator Characteristic (ROC) curves for benchmark tests on DB179 and NB3797. DBD-Threader generally performs better than PROSPECTOR and PSI-BLAST, especially at a precision higher than 0.75 and at False Positive Rate (FPR) lower than 0.01, the regime relevant to practical applications. Correspondingly, the sensitivity obtained by DBD-Threader can be higher than 0.55 within this regime. If one considers only fold similarity suggested by the threading Z-score or sequence similarity measured by the PSI-BLAST E-value, one obtains an inferior precision/FPR at the same level of sensitivity. For example, at a sensitivity value of 0.55, the precision/FPR for DBD-Threader, PROSPECTOR, and PSI-BLAST (NCBI), and PSI-BLAST (UniProt) is 0.85/0.004, 0.69/0.012, 0.24/0.085, and 0.24/0.081, respectively. Therefore, the results suggest that the quality of predictions by DBD-Threader is significantly improved when both threading Z-score and protein-DNA interaction propensity are taken into account. We also note that threading itself (PROSPECTOR) typically performs better than PSI-BLAST.
The comprehensive performance of these methods can be assessed by the Matthews Correlation Coefficient (MCC) [42]. A perfect prediction at 100% accuracy yields a MCC of one, whereas a random prediction gives a MCC of zero. The best MCCs of these four methods are provided in Table 1. The highest MCC of DBD-Threader is 0.680, corresponding to a sensitivity of 0.56 and a precision of 0.86, whereas the best MCCs of PROSPECTOR, PSI-BLAST (NCBI), and PSI-BLAST (UniProt) are 0.609, 0.540, and 0.553, both as shown in Table 1 at lower sensitivity and precision than DBD-Threader. Moreover, the best performance of DBD-Threader is only slightly lower than that (MCC 0.681) of DBD-Hunter, which requires the structure of the target as input. Note that the previous results of DBD-Hunter were obtained on a smaller template library [19]. The results reported here are based on the updated template library employed by all methods. Direct structural comparison allows DBD-Hunter to detect homology between a pair of template/target proteins with no sequence similarity, resulting in the highest sensitivity of 0.61 among all four methods at the same precision of 0.79. Nevertheless, the performance of DBD-Threader is comparable to that of DBD-Hunter in terms of its MCC. The optimal thresholds corresponding to the best performance of DBD-Threader were adopted in the application to the human genome below.
The contributions by threading and by energy to the optimal performance of DBD-Threader are further dissected through an analysis of DNA-binding and non-DNA-binding proteins that share common structural folds. Here, we use the Structural Classification of Proteins (SCOP) to classify structural folds [43]. Table 2 shows the numbers of proteins (and their relevant domains) that belong to the same SCOP folds across two benchmark sets DB179/NB3797. In total, there are 109/599 proteins that contain 127/646 domains from 24 common SCOP folds. The vast majority of non-DNA-binding proteins were filtered out after the threading procedure, resulting in a 90% reduction in non-DNA-binding proteins but only a 31% reduction in DNA-binding proteins to 75/58 DB/NB proteins. After applying the optimal energy criteria, the number of DNA-binding proteins is reduced by 23% to 58, whereas the number of non-DNA-binding dramatically decreases again by 86% to 8. We note that in some sparsely populated (number of DB targets ≤4) folds, successive filtering by threading and energy left no true positive from the DB set. This is mainly due to the absence of a suitable template under the specified sequence identity cutoff of 30%. By ignoring these folds, one still obtains about 75% and 80% reduction rates on non-DNA-binding proteins through threading and energy filtering, respectively, while the majority of DNA-binding proteins are kept. Overall, the analysis shows that both threading and energy calculations significantly contribute to the ability to distinguish the DNA-binding function among proteins with similar folds.
There are 91 non-DNA-binding proteins with at least one significant template hit (threading Z-score>6), but they are from other SCOP folds that lack any known DNA-binding protein. These non-DNA-binding proteins may contain structural fragments similar to their significant template hits or may be falsely identified by threading. By applying the energy criteria, 82 of these 91 proteins were correctly filtered out as non-DNA-binding proteins. The energy calculations, therefore, serve to reduce the number of potential false positives generated by threading.
The contribution of energy filtering can be illustrated through two examples from the NB3797. The top ranked template hits by these two targets are significant with Z-scores over 20, but these templates did not satisfy the energy criteria because of their high repulsive DNA-protein interaction energies. Both proteins are classified as non-DNA-binding. The first example is an inositol polyphosphate 5-phosphatase (PDB 1i9yA), which hits a DNA repair protein APE1 (1dewB). They are evolutionarily related and belong to the same SCOP superfamily. However, they have very different selectivity for substrate, as the inositol polyphosphate 5-phosphatase is not known to bind DNA. The second example is λ lysozyme (1am7A), which hit an endonuclease (2fldA) with a high Z-score. This seems to be a false positive by threading, since the target/template pair shares no apparent structural similarity and are not related. Nevertheless, the template did not pass energy screening.
Overall, by applying energy filtration, the number of true/false positives decreases from 131/149 (after threading) to 100/17, the numbers including results from all targets with official SCOP classification, as well as those unclassified. Thus, the filtration by energy improves the precision from 47% to 86% without dramatically compromising the sensitivity.
In addition to function prediction, DBD-Threader also predicts structural models of DNA-binding domains from templates that provide the structural basis for function prediction. Furthermore, one may infer the functional sites directly from the template, once the functional and structural similarity between the template and the target is established. To demonstrate this point, we implemented a simple procedure in DBD-Threader that predicts DNA-binding protein residues from the top ranked template by those residues in the target aligned to DNA-binding residues in the template. In benchmark tests on DB179, this procedure was conducted on 124 domains from 100 DNA-binding proteins predicted as positives by DBD-Threader at the optimal thresholds. The value of the MCC, which measures the degree of overlap between predicted binding residues and the true binding residues observed in the native (experimental) complex structures, is used to assess the accuracy of functional site prediction. As shown in Figure 3A, DBD-Threader performs well on both structural and functional site prediction. The mean Template Modeling score (TM-score) of the top-ranked structural models of the 124 DNA-binding domains with respect to their native structures is 0.65, and 92% of these domains have a TM-score higher than 0.4, which indicates significant structural similarity [15]. Similarly, 70% of these domain models have a backbone Cα RMSD of less than 6.5 Å from their native structures. Accordingly, the mean MCC of binding site predictions is generally satisfactory, being about 0.52 for all predicted DNA-binding domains and 0.54 for domains with a TM-score higher than 0.4. As one may expect, the accuracy of binding site prediction is correlated with model quality. High quality models with a TM-score higher than 0.6 generally provide a high accuracy binding site prediction, yielding a mean MCC of 0.57, whereas low quality models with a TM-score lower than 0.4 provide inferior binding site predictions with MCCs lower than 0.4.
We further analyzed the performance according to the SCOP superfamily association of these predicted domains, as shown in Table 3. The analysis considers 84 predicted DNA-binding proteins that have an official SCOP assignment, which includes 100 domains detected by DBD-Threader and an additional 10 domains missed by DBD-Threader (see SCOP superfamily prediction below). According to their SCOP classifications, the 100 detected domains are from 31 SCOP superfamilies. The performance of DBD-Threader is generally good across various SCOP superfamilies. 24 of 31 superfamilies have a mean TM-score/MCC higher than 0.4. It appears that members of the winged helix superfamily have rather diverse DNA-binding sites. This is indicated by the mean MCC of 0.38, despite the high quality of models that are obtained (mean TM-score of 0.65).
The performance measures, sensitivity, specificity, accuracy and precision, were also calculated for each of 100 proteins including all DNA-binding domains. As shown in Figure 3B, for 61% of predicted DNA-binding proteins, good functional site predictions were obtained at a MCC higher than 0.50. On average, a MCC of 0.53, a sensitivity of 0.60, a specificity of 0.93, an accuracy of 0.86 and a precision of 0.64 were obtained. The results imply that DNA binding residues were identified with satisfactory accuracy in most cases.
The homologous relationship between the target/template pairs identified by DBD-Threader was further validated using their SCOP superfamily classifications [43]. Here, we test the idea of inferring the SCOP superfamily identity of a predicted DNA-binding domain from its templates. Among 100 predicted DNA-binding proteins, we only consider those whose SCOP assignments have been officially assigned. The consideration leads to 84 proteins composed of 110 true domain assignments, which are then compared with the predictions by DBD-Threader. The predictions can be classified into four groups, as shown in Table 3. The first group is 73 SCOP superfamily predictions that are consistent (C) with the official SCOP assignments. These correctly predicted domains are from 21 different SCOP superfamilies, including two of the most populated superfamilies, homeodomain-like and winged helix domains, with 16 and 11 correct predictions, respectively. These 27 domains consist of diverse members from 16 different SCOP families. The second group of predictions is 20 DNA-binding domains correctly identified as DNA-binding, but their SCOP superfamily classifications were un-annotated (U) because the corresponding templates have no official SCOP assignment. In 16/20 un-annotated cases, significant structural similarities between target/template pairs were found at a TM-score>0.5, implying that most of these pairs likely belong to the same superfamily. The 16 cases that are un-annotated combined with the 73 consistent SCOP predictions lead to 89 cases, or 81% of 110 domains, that may be considered correct. The third group is comprised of ten missed (M) DNA-binding domains, which are from proteins with multiple DNA-binding domains. In these cases, DNA-binding function can be successfully predicted by identifying some but not all of its DNA-binding domains. The fourth group of predictions are from the seven cases where the SCOP superfamily predictions are inconsistent (I) with the true SCOP assignment. Inspection of these predictions suggests potential functional homology in 5/7 cases. Two are presented in detail below.
In the first example, the target protein is the DNA-binding domain of PhoB, a transcription activator from E. coli [44]. According to SCOP, this domain belongs to the superfamily named C-terminal effector domain of the bipartite response regulators. DBD-Threader predicts that the domain belongs to the superfamily of winged helix DNA-binding domains based on its top ranked template, the Zα domain of an enzyme ADAR1 (Adenosine Deaminase Acting on RNA) from human [45]. Although ADAR1 is best known as an RNA binding protein, it is also known to bind Z-DNA with its Zα domain, as shown in multiple crystal structures of ADAR1/DNA complexes [45],[46]. In addition, the DNA-binding ability of Zα has been used to detect stable Z-DNA segments in the human genome [47], and has been linked to a new functional role of ADAR1 as a sensor of immunoreactive DNA [48]. Despite the difference in SCOP superfamily classification, the target and the template share a similar structural motif, with a high TM-score of 0.69, as shown in Figure 4A. In fact, both structures are members of the same superfamily of winged helix domains according to CATH, a hierarchical classification of protein domain structures [49]. In addition, both DNA-binding domains have similar DNA-binding sites, which include six residues from a α helix and a β hairpin (Figure 4A). The significant structural similarity and the overlap of the DNA-binding sites suggest that these two domains might have remote homology, despite the lack of sequence similarity. Thus, we have an interesting case of the PhoB domain being correctly assigned as DNA binding through the matching to an RNA binding protein that is also known to bind DNA.
In the second example, the target is the N-terminal domain from a eukaryotic DNA polymerase, Pol β [50]. The target hits a significant template from an archaeal endonuclease XPF, whose structure is composed of two heterogeneous domains [51]. As shown in Figure 4B, the target domain from Pol β was aligned to the N-terminal domain of XPF with significant structural similarity, having a TM-score of 0.48, and considerable overlap of DNA-binding residues, despite the fact that the two domains have different superfamily classifications in SCOP. The structural and functional site analysis suggests that the two domains may have a remote relationship.
To demonstrate that DBD-Threader is a useful tool for automatic function annotation, we applied DBD-Threader to 18,621 unique protein sequences from the human genome. The method made positive predictions for 1,654 (8.9%) proteins (see Methods for availability). Our predictions are compared to the GO annotations for the human genome [52] in Figure 5. According to the GO molecular function annotations, all human proteins can be classified into four sets: DB−1,744 proteins annotated as DNA-binding, UB−1,573 proteins not explicitly annotated as DNA-binding but annotated with a molecular function likely implicating DNA-binding, such as transcription factor activity, NB−10,616 proteins with at least one molecular function annotation and not in either DB or UB, and UK−4,688 proteins with unknown molecular function. While the vast majority of entries in DB are classified based on electronic annotations, we collected a small subset of DB, named DB EXP, in which the DNA-binding function has been verified for each member in direct experimental assay. This DB EXP set consists of 69 sequences. DBD-Threader detected at least one significant structural template for 56 of them and correctly predicted 54 as DNA-binding. Similarly, when applied to the DB set, DBD-Threader found at least one significant template for 1,235 sequences and predicted 1,179 (95%) of them as DNA-binding proteins. Notably, when applied to the UB set, DBD-Threader predicted 325 DNA-binding proteins. Among these UB positives, 256 and 51 have transcription factor activity and RNA-binding activity according to their GO annotations, respectively. These proteins likely possess DNA-binding function as well. While 89% of the positives are from either DB or UB, very few positives, 72 (0.68%), are from the NB set. This result is expected, since the chance that a protein has both DNA-binding and an unrelated molecular function is small. Despite the fact that 298 targets from NB hit a significant structural template that binds DNA with Z-score>6, 75% of them are filtered out by the interaction energy criterion. These negatives likely possess a fold similar to a DNA-binding domain, but they do not carry out the same function. Furthermore, DBD-Threader predicts 78 DNA-binding proteins among previously uncharacterized sequences. These predictions provide potentially interesting targets for further experimental validation.
A total of 6,896 DNA-binding domains from 59 SCOP superfamilies were located by DBD-Threader in the sequences of 1,654 positives. The top twenty most populated SCOP superfamilies of predicted DNA-binding domains are listed in Table 4. Notably, zinc-fingers appear in about 41% (674) of predicted DNA-binding proteins, and this particular superfamily dominates the domain predictions at 80% of the total (5,504). The second and third most common SCOP superfamilies are homeodomain-like and winged helix domains, which are found in 263 and 143 sequences, respectively. Many DNA-binding proteins, particularly zinc-fingers, contain two or more DNA-binding domains. Moreover, it is not uncommon that a sequence encodes DNA-binding domains from different SCOP superfamilies. In our annotations, we found 175 such cases.
Our predictions are compared with Pfam predictions, which are based on Hidden Markov Models (HMMs) [53] in Table 4. The results of Pfam predictions were obtained from the UniProt knowledge base. For an objective comparison, we consider Pfam families defined for DNA-binding proteins from our template library. The Pfam definitions of these template structures were initially obtained from the PDB. These were then manually curated to ensure that the definitions correspond to the DNA-binding domains. This led to 179 Pfam families that likely include all DNA-binding proteins with known atomic complex structures, but not those with unknown structures or with only DNA-unbound structural forms. Using the SCOP definitions of the templates, we are able to assign these Pfam families to 69 SCOP superfamilies. Overall, Pfam found 7,162 significant domain matches in 1,591 proteins from the same sequence set scanned by DBD-Threader. The numbers are consistent with the 6,896/1,654 domains/proteins predicted independently by DBD-Threader.
As shown in Table 4, the results of the top three most populated SCOP superfamilies are comparable between these two methods. About 80%/85%, 4.5%/4.3%, and 2.1%/1.4% of predicted domains are zinc finger, homeodomain, and winged helix proteins by DBD-Threader/Pfam, respectively. The zinc finger proteins dominate both predictions, and over 95% of predicted zinc finger proteins were positively hit by both methods. Despite the similarity, however, about 26% of zinc finger domains detected by Pfam are predicted as negatives by DBD-Threader. One interesting question is whether these domains have evolved their function from DNA-binding to have other roles not involving DNA-binding. Although the vast majority of these proteins have not been experimentally studied, we found a few potential examples of such zinc finger domains with experimental evidence from the literature (see Case Studies below). Moreover, we noticed that DBD-Threader predictions are more diverse in terms of the number of SCOP superfamilies detected (59 vs. 48) and are about three times more sensitive than Pfam in assigning putative DNA binding ability to functionally uncharacterized proteins (78 positive hits vs. 20).
Predictions of DBD-Threader are further compared with PSI-BLAST results in Figure 6. For each target, we identified the lowest PSI-BLAST E-value of all sequence alignments with all templates. In Figure 6A, the distributions of the lowest PSI-BLAST E-values are given for both positives and negatives predicted by DBD-Threader. One can immediately recognize that most positives share significant sequence similarity with a known DNA-binding domain. About 79% (1,314) of positives hit a significant template at a PSI-BLAST E-value<10−20. In contrast to positives, only 0.3% (53) of negatives fall into this significant E-value regime. Using the GO annotations, we found that 78% of the 1,314 positives belong to the DB set, while only 49% of the 53 predicted negatives belong to DB. On the other hand, the overwhelmingly majority of negatives (16,642) are found within the regime where the E-value is higher than 10−3. However, DBD-Threader managed to predict 136 positives in this regime, despite low/no sequence similarity. Analysis of their GO annotations found that 13% (18) of positives belong to DB, the ratio is over four times 2.9% (476), the rate of negatives classified as DB in the same E-value regime. The comparison suggests that DBD-Threader considerably enriches the predictions of true positives compared to PSI-BLAST.
DBD-Threader can make a strong prediction without apparent sequence similarity. This is illustrated through an application to the origin recognition complex subunit 6 (Orc6), which is a component of the heterohexameric origin recognition complex (ORC). The main function of ORC is to initiate DNA replication, which necessitates DNA-protein interactions [54]. It has been shown experimentally that Orc6 of Drosophila melanogaster binds to DNA [55]. Human Orc6 has a statistically significant sequence similarity to Drosophila Orc6 (PSI-BLAST E-value = 10−24), though the global sequence identity is relatively low at 30% over ∼240 AAs. It is not clear, however, whether human Orc6 has a similar DNA binding function [55],[56]. The sequence of human Orc6 was assessed by DBD-Threader, which predicted two DNA-binding domains in the N-terminal region (residues 1–202), based on a significant hit to the transcription factor TFIIB at a Z-score of 26 and an energy value of −9.3. By contrast, neither PSI-BLAST nor Pfam can detect a significant template from our library, which is not surprising given that there is no apparent sequence similarity between TFIIB and Orc6. Although the structure of Orc6 has not been experimentally solved, our prediction agrees with a structural model of Drosophila Orc6 that was recently proposed [57]. In addition, point mutations of Ser72 and Lys76, two residues located within a putative DNA-binding helix-turn-helix motif and conserved between human and Drosophila, abolish the DNA-binding ability of Drosophila Orc6 [55].
It is well-known that function inference based on sequence or structural comparison, even at a statistically significant level of similarity, can be misleading [8],[20]. By applying the energy based filter, DBD-Threader can reduce false positives generated from structural or sequence similarity comparison. This is illustrated through a second example, the barrier-to-autointegration factor-like (BAF-L) protein, whose sequence is about 40% identical to that of BAF, a known DNA-binding protein [58]. The homologous relationship was detected by PSI-BLAST (E-value<10−46), Pfam (E-value<10−48), and DBD-Threader (Z-score = 35). The GO annotations of BAF-L include DNA-binding function, probably inferred from BAL based on sequence similarity. However, using the energy filter, DBD-Threader predicts that BAF-L is not a DNA-binding protein due to its repulsive DNA-protein interaction energy. The prediction is supported by an experimental study which suggests that the functional role of BAF-L is not DNA-binding [59]. Instead, it is proposed to be a regulator of BAF through dimerization with BAF. The prediction is also supported by the fact that most residues involving DNA-binding of BAL are not conserved in BAF-L.
Particularly interesting are the classic (C2H2/C2HC type) zinc finger domains found in 41% of predicted DNA-binding proteins. The classic zinc finger domain is one of most abundant protein domains encoded in the human genome. According to the domain annotations in the UniProt knowledge base, these are 6,873 C2H2/C2HC zinc finger domain matches in 751 protein sequences of the 18,621 sequences scanned by DBD-Threader. The vast majority (92%) of these sequences contain multiple zinc finger domains. An interesting question is what functions these domains perform. If one assigns DNA-binding function according to sequence similarity detected by PSI-BLAST or Pfam, all zinc finger domains detected would be assigned as DNA-binding. Although the classic zinc finger domains originally discovered are DNA-binding domains of many transcription factors, recent studies have demonstrated that they can play a functional role through protein-protein interactions (see reviews, [60],[61]). While DNA-protein and protein-protein interactions are not necessarily mutually exclusive, it is possible that some zinc finger domains play a role involving only protein-protein interactions. About 91% of zinc finger domains annotated in UniProt were detected during threading, and 22% of these significant threading hits were assessed as negatives according to the energy calculations by DBD-Threader. Although there are inevitably false positives/negatives, we speculate that most of these negatives have acquired a functional role that does not involve DNA-binding but other biological interactions, such as protein-protein interactions.
To further examine our hypothesis, we compiled from the review in [60] a list of 18 zinc finger domains likely involved only in protein-protein interactions, as shown in Table 5. These domains, collected from six human sequences, are all experimentally well characterized. Note that we excluded domains with known DNA-binding function from these sequences. If the predictions by DBD-Threader were random, one would expect that a true negative is predicted at a success rate of 22%. Assuming that all 18 domains we collected are true negatives, we further test the null hypothesis that DBD-Threader predicts non-DNA-binding zinc finger domains at a success rate of 22% or less. Among the 18 domains, DBD-Threader predicts 4 positives and 14 negatives, which yields a significant p-value (7.6×10−7) in a one-tailed binomial test. Therefore, we rejected the null hypothesis. The result suggests that the predictions by DBD-Threader are statistically highly significant.
Lastly, we examine an intricate example from Table 5 in the transcription factor OAZ (Olf1/EBF-associated zinc finger protein, also known as ZNF423). This is a 1284 AA long sequence composed of 30 zinc-fingers distributed in several clusters (Figure 7). The homology of OAZ to other well-characterized zinc finger proteins, such as Zif268 and TFIIIA, were readily established by both PSI-BLAST and DBD-Threader. Significant hits with PSI-BLAST E-values<10−20 and threading Z-scores>15 cover virtually all zinc-finger repeats of OAZ. However, evaluation of the DNA-protein interaction energy by DBD-Threader suggests that only fingers 2 to 6 are DNA-binding, whereas other fingers do not carry out this function due to their highly repulsive energy values (typically E>10). The prediction is in agreement with two independent experimental studies [62],[63]. In the former study, the protein was partitioned into six clusters, and the DNA-binding activity of each was assessed with SELEX. Only the cluster containing fingers 2-5 was found to be DNA-binding [62]. The second study was performed on rat OAZ, the ortholog nearly identical (∼96%) to its human counterpart. Consistently, the DNA-binding region was mapped within the first seven fingers of OAZ [63]. In addition, both studies identified the same consensus DNA sequence recognized by these fingers. Among other zinc fingers, it was suggested that the three C-terminal zinc-fingers are essential for the interactions between OAZ and another transcription factor Olf-1/EBF, which regulates olfactory gene expression in rat [63],[64]. Another study reported that zinc-fingers 14 to 19 mediate the interaction with transcription factors Smad1 and Smad4, and that zinc-fingers 9 to 13 bind BMP (bone morphogenetic proteins) target gene promoters together with Smads [65]. The latter result that fingers 9–13 bind DNA apparently disagrees with the prediction by DBD-Threader. One possible explanation for the discrepancy is that zinc-fingers 9–13 of OAZ may adopt an atypical DNA-binding mode not present in our template library. This is supported by the observation that zinc-fingers 9–13 have unusually long (>15 AAs) linkers between them (Figure 7), whereas other structurally known DNA-binding zinc-finger proteins have shorter linkers, typically six residues, connecting their fingers. In summary, OAZ plays a central role in two distinct processes involving BMP signaling and olfactory neurogenesis, and its multi-functional role is fulfilled by different zinc fingers. While sequence or structure similarity alone cannot distinguish the functional roles of zinc-fingers, which may interact with DNA or other proteins, DBD-Threader provides a means to assess the DNA-binding preference of individual zinc-finger domains.
Previously, threading-based methods were proposed for predicting protein-protein and protein-ligand interactions [39],[40]. In this study, DBD-Threader extends this idea to the prediction of DNA-binding function. The method employs two key functional discriminating features: fold similarity and DNA-binding propensity. Given a target, sequence threading is used to identify a template that has a similar fold to the target. Compared with standard sequence comparison methods, such as PSI-BLAST, threading is more sensitive in detecting homology, especially when the sequence identity is lower than 30% [41]. However, since threading itself does not differentiate functional roles among sequences with a similar fold, this can give rise to a considerable number of false positives. To reduce the number of false positives, the DNA-protein interaction energy is calculated to assess whether the target preferentially interacts with DNA. In our approach, DBD-Threader uses a statistical pair potential, which has been successfully implemented in our previous application (DBD-Hunter) in predicting DNA-binding function given the native protein's structure [19]. Overall, DBD-Threader achieves better performance than approaches using only sequence homology. In benchmark tests on ∼4000 proteins, DBD-Threader is about 15% to 25% higher in sensitivity than PSI-BLAST at the same false positive rate of less than 1%, using templates that share less than 30% sequence identity with the targets. The optimal performance of DBD-Threader has a MCC of 0.68, better than the MCC of 0.61 of PROSPECTOR and 0.55 of PSI-BLAST, and is comparable to the performance of DBD-Hunter where the experimental structure of the target is required.
There exist quite a few template-free methods for predicting DNA-binding function [27],[28],[30],[31] or DNA-binding protein residues [32]–[34],[36],[37], the latter class of methods require the information that the protein is known to be DNA-binding. Most of these methods use machine-learning techniques, which provide no structural and limited biological insights. While these template-free approaches have the potential to predict the DNA-binding sites of a novel fold, their accuracy is generally lower than template-based methods [19],[37], and their performance has not been tested in large-scale benchmarks. DBD-Threader, as a template-based method, provides not only function prediction, but also structural insights into the predicted function by identifying the DNA-binding domains and associated DNA-contacting protein residues. In benchmark tests using templates with less than 30% sequence identity to the target, the backbone RMSDs of the top-ranked structural models are within 6.5 Å of their native structures for 70% of predicted DNA-binding domains. In addition, the mean sensitivity and specificity of binding site predictions is 60% and 93% among predicted DNA-binding proteins, whose DNA-binding domains have been correctly identified in terms of SCOP superfamily in most cases. The main disadvantage of a template-based approach is that it cannot predict DNA-binding function/sites for structures not present in the template library. In addition, one should generally not expect a high-level of detailed binding-site conservation between a template/target pair at low sequence identity, though the success of DBD-Threader suggests that it tends to identify functionally related template/target pairs, whose DNA-binding sites are significantly similar in most cases.
In the post-genomic era, there is a pressing need for accurate, automatic function annotation tools. DBD-Threader, implemented as a fully automated method, contributes to such a task. This is illustrated in the application of DBD-Threader to the human genome. The method predicts 1,654 DNA-binding proteins among ∼19,000 unique sequences from human. Comparing the results of DBD-Threader to their existing GO annotations, about 68% of the positives by DBD-Threader agree. Most of the remaining predictions have a GO annotation related to DNA-binding, such as transcription factor activity. Therefore, they very likely play a DNA-binding role. Moreover, DBD-Threader predicts a few protein sequences among uncharacterized sequences as DNA-binding. These can serve as candidates for further experimental examination.
The predicted DNA-binding proteins from the human genome contain 6,896 DNA-binding domains from 59 SCOP superfamilies. The vast majority of these predicted DNA-binding domains are cross-validated by other sequence annotation methods, such as Pfam annotations. The largest population of DNA-binding proteins is the zinc-finger proteins, which are about 41% of predicted DNA-binding proteins. Interestingly, 22% of detected zinc finger domains yield negative results based on the DNA-protein interaction energy assessment. Case studies of these zinc finger domains suggest that they likely perform other biological functions, such as protein-protein interactions, but not direct DNA-binding.
Function prediction from protein sequences is a challenging problem. Since proteins are evolving, they can acquire new functions and/or lose old ones. With respect to DNA-binding, a possible scenario is that it evolves to become a regulator of DNA-binding through interactions with other DNA-binding proteins, instead of directly participating in DNA-binding. While such evolution is biologically very interesting, it creates problems for approaches to function inference based on sequence similarity alone, such as those based on PSI-BLAST or HMMs. By assessing DNA-binding propensity through use of the DNA-protein interaction energy, DBD-Threader can help to discriminate DNA-binding from other functional roles, thus improving the overall quality of the predictions. Application of the method generates not only potentially interesting positives, but also negatives evolved from direct DNA-binding. Through this study, we identified 22% of zinc finger domains annotated in the human genome as such negatives with DBD-Threader.
All datasets listed below, the statistical potential parameters, prediction results on the human genome, and a web-server implementation of DBD-Threader are freely available at http://cssb.biology.gatech.edu/skolnick/files/.
The method DBD-Threader has three main modules: sequence threading, domain partition, and function prediction. Sequence threading was conducted using the in-house program PROSPECTOR [41]. The purpose of threading is to examine whether a target sequence encodes a structural fold similar to any structurally known DNA-binding domain. Specifically, the target sequence is threaded sequentially against two template libraries. The first library is a regular template library composed of ∼8000 protein structures, which share less than 35% global sequence identity among each other; the second library is composed of DNA-binding domains described above. The target is first threaded against the regular template library, generating statistically more robust mean and standard deviation of threading scores than threading directly on the much smaller library of DNA-binding domains. Then, the mean and standard deviation are used to calculate the Z-scores when threading templates of the DNA-binding domain library are used. Note that in the benchmark test on DB179/NB3797, we excluded all templates with more than 30% sequence identity from both threading libraries for any target. In the application to the human genome, the exclusion rule was eliminated.
For each pair of target/template proteins, a corresponding Z-score is calculated as where S is the score associated with the best alignment between the pair, and quantity in angle brackets denotes the mean of the quantity over all entries in the regular template library. Based on our benchmark results, we consider templates with Z-scores>6 as significant hits, which are then ranked according to their Z-scores.
Since most DNA-binding proteins are composed of multiple domains, it is necessary to locate the domain(s) that directly fulfill DNA-binding function. To this end, an iterative clustering procedure was implemented to partition domains of the target sequence based on significant template hits. Clustering is required because a DNA-binding protein may contain multiple DNA-binding domains, which can hit different sets of templates. Initially, the top Z-score-ranked template is chosen as the clustering seed, and all significant templates having more than 50% overlap with respect to the seed are moved to this cluster, and excluded from subsequent clustering. After this process, if there is any template left, the highest ranked template remaining is used as a new clustering seed, and this clustering procedure is repeated until no template is left. The clustering is used to consolidate redundant templates that hit the same sequence region, and a domain can then be defined according to the alignment of a seed to the target. The threading and partition procedures are iterated for any sequence region without a significant hit that is longer than 40 amino acids, until no new domain is found. This iterative procedure can reduce missing hits to domain repeats, e.g., zinc finger clusters, because threading itself only returns the most significant alignment from each template in each round.
For function prediction, we evaluate the DNA-protein interaction energy and use it to assess DNA-binding propensity. Here, we consider only significant templates hits whose DNA-protein contacts have been obtained beforehand using the experimentally determined DNA-protein complex structures. The contacts between the target and a corresponding template DNA are inferred by replacing original template protein residues with aligned target residues. The protein-DNA interaction energy is then calculated using these contacts and a statistical pairwise potential developed previously [19]. Negative and positive energy values indicate attractive and repulsive interactions, respectively. A target is predicted to be a DNA-binding protein if at least one template yields an energy value below a specified threshold, and non-DNA-binding if no template satisfies the energy criterion. Finally, the SCOP superfamily domain assignment is inferred from the highest Z-score-ranked template that satisfies the energy criteria, and corresponding DNA-binding residues are also transferred from this template. The SCOP superfamily prediction will be skipped if the top template does not have official SCOP classification.
The optimal energy threshold values determined in benchmark tests on DB179/NB3797 are shown in Table 6. Depending on the threading Z-score of their templates, the targets fall into two regimes: Medium (20≥Z-score>6) and Easy (Z-score>20). In each regime, we select an optimal energy threshold that gives the highest MCC of predictions on DB179 and NB3797. As one expected, a more permissive energy value is obtained for the Easy targets. In the benchmarks, the ROC and PR curves of DBD-Threader were obtained by varying the energy threshold for templates in the Medium regime, and use the optimal energy threshold for templates in the Easy regime. The optimized values were adopted in the application to the human genome.
DBD-Threader was compared with three alternative approaches: DBD-Hunter [19], PROSPECTOR [41], and PSI-BLAST [11]. To ensure fair comparison, the same template library and benchmark sets DB179/NB3797 were employed. In case of DBD-Hunter, structures of targets were used as input, and the results obtained with optimized parameters are reported. When applying PROSPECTOR, the threading Z-score was used as the criterion for predictions. A target protein is classified as DNA-binding if it hits a template with a Z-score higher than a specified threshold and as a non-DNA-binding otherwise. When applying PSI-BLAST (version 2.2.17), two position specific sequence profiles were derived separately for each target using two libraries: the NCBI-NR protein sequence library (the Jul 2007 release), and the UniProt sequence library (UniRef100 version 15.5). Each profile was obtained using up to four rounds of scanning the respective libraries. We tested up to twenty rounds of iterations for profile derivation and found that four rounds gave the best performance in our benchmark tests. An inclusion E-value threshold of 0.001 and default values for other arguments were employed. For each profile generated, a final PSI-BLAST run was performed on the sequence library of the DNA-binding protein templates. If a target hits a template with an E-value higher than the specified threshold, then the target is classified as being a DNA-binding protein; otherwise, it is classified as a non-DNA-binding protein. For each target in the benchmark tests, its homologs with global sequence identity >30% were excluded from the template library of DNA-binding proteins. Note that the exclusion rule was not applied during the derivation of the PSI-BLAST profiles, and we allow all sequence hits for building the profiles.
In each prediction scenario, the numbers of true positives, false positives, true negatives and false negatives are designated as TP, FP, TN, and FN, respectively. In case of DNA-binding function prediction, a TP refers to a protein sequence correctly predicted as DNA-binding protein; in case of DNA-binding site prediction, TP refers to an amino acid correctly assigned as a DNA-binding residue. Performance measures are defined as the following:
|
10.1371/journal.pntd.0000529 | Mosquito Infection Responses to Developing Filarial Worms | Human lymphatic filariasis is a mosquito-vectored disease caused by the nematode parasites Wuchereria bancrofti, Brugia malayi and Brugia timori. These are relatively large roundworms that can cause considerable damage in compatible mosquito vectors. In order to assess how mosquitoes respond to infection in compatible mosquito-filarial worm associations, microarray analysis was used to evaluate transcriptome changes in Aedes aegypti at various times during B. malayi development. Changes in transcript abundance in response to the different stages of B. malayi infection were diverse. At the early stages of midgut and thoracic muscle cell penetration, a greater number of genes were repressed compared to those that were induced (20 vs. 8). The non-feeding, intracellular first-stage larvae elicited few differences, with 4 transcripts showing an increased and 9 a decreased abundance relative to controls. Several cecropin transcripts increased in abundance after parasites molted to second-stage larvae. However, the greatest number of transcripts changed in abundance after larvae molted to third-stage larvae and migrated to the head and proboscis (120 induced, 38 repressed), including a large number of putative, immunity-related genes (∼13% of genes with predicted functions). To test whether the innate immune system of mosquitoes was capable of modulating permissiveness to the parasite, we activated the Toll and Imd pathway controlled rel family transcription factors Rel1 and Rel2 (by RNA interference knockdown of the pathway's negative regulators Cactus and Caspar) during the early stages of infection with B. malayi. The activation of either of these immune signaling pathways, or knockdown of the Toll pathway, did not affect B. malayi in Ae. aegypti. The possibility of LF parasites evading mosquito immune responses during successful development is discussed.
| Filarial worms that cause human lymphatic filariasis (LF) are transmitted by many species of mosquitoes. Within susceptible mosquitoes, Brugia malayi develop from microfilariae (mf) to infective-stage larvae (L3s), in approximately eight days. These nematodes develop as intracellular parasites within mosquito flight muscle cells, in which they ingest cellular material and eventually cause cell death when L3s migrate to the mosquito's proboscis. We examined the effects of B. malayi parasitism on Aedes aegypti by analyzing changes in mosquito gene expression at different stages of parasite development. We found that a few genes were differentially expressed at the RNA level relative to non-infected controls. The majority of changes occurred at two time periods, when the filarial worms began feeding and when the L3s were in the head and proboscis. Many transcriptional changes in the later group concur with documented descriptions of tissue damage, clean-up and repair that occurs in mosquitoes infected with filarial worms. In addition, we activated two innate immunity signaling pathways and observed the effects on filarial worm development. B. malayi seems to be capable of evading these immune responses, because its development was not impeded by the activation of either the Toll or Imd signal pathways in Ae. aegypti.
| It is estimated that 120 million people are infected with Wuchereria bancrofti, Brugia malayi, or B. timori, the mosquito-transmitted, parasitic nematodes that cause human lymphatic filariasis (LF). In approximately 40% of cases, the disease is manifested by lymphedema of the extremities or hydrocoele. Although human LF does not increase mortality in endemic areas, morbidity causes major economic losses and often leads to psychosocial and psychosexual conditions in infected individuals [1]. Recent efforts by the Global Program for the Elimination of Lymphatic Filariasis (GPELF) have decreased the numbers of individuals infected with, and at risk for, this parasitic disease [2].
Several different mosquito species within the genera Culex, Anopheles, Aedes and Mansonia can serve as primary vectors of LF parasites. The geographical location and habitat type influence which mosquito species function as vectors in any particular endemic area. Biological transmission of filarial worms is termed cyclodevelopmental, i.e., the parasite undergoes development within the vector to become infective to the vertebrate host, but does not multiply. In competent vectors, microfilariae (mf), produced by adult female worms and found circulating in the peripheral blood, are ingested with a blood meal and will quickly (within 2 hr) penetrate the midgut epithelium to access the hemocoel [3]. Mf migrate in the mosquito's hemolymph to reach the thoracic musculature and from there penetrate into the indirect flight muscles. This tissue is the site of development, where mf undergo two molts and emerge as infective-stage larvae (L3s). Approximately eight days after exposure, L3s migrate to the head and proboscis from where they escape by penetrating the labellum of the proboscis when the mosquito takes a blood meal. Within the human host, the parasites undergo two additional molts and grow as they migrate to lymphatic vessels where adult male and female worms mate and females give birth to mf. Mf then make their way into the circulating blood from where they can be ingested by another blood feeding mosquito.
LF parasites grow nearly seven times in length (B. malayi grow from ∼200 to ∼1,350 µm in length, from mf to L3s respectively) during the extrinsic developmental period within the mosquito [4]. As parasites develop, the mosquito must tolerate a series of insults due to parasite activities, e.g., migrating mf damage both midgut [5] and muscle cells as they penetrate through or into them; second-stage larvae (L2s) actively ingest mosquito cellular components; L3s are very large and migrate out of the thoracic muscles and through the body cavity to reach the head and proboscis. Ultrastructural studies of Aedes mosquitoes infected with Brugia parasites have revealed that nuclear enlargement (a sign of a putative repair response) occurs in both Brugia-infected and neighboring non-infected muscle cells, and that complete degeneration of infected muscle cells occurs once L3s exit the flight muscles [6]. Other studies have shown that mosquito flight muscle cells become devoid of glycogen granules following infection with Brugia parasites [7],[8]. Considering the amount of tissue damage observed in muscle cells, it is not surprising that Brugia-infected mosquitoes are known to have decreased flight activity and longevity [9],[10]. But the successful development (and subsequent transmission) of LF parasites depends on the ability of competent mosquito vectors to survive infection.
Some mosquitoes are able to limit or prevent filarial worm infections with various refractory or resistance mechanisms. For example, mf can be damaged during ingestion by an armed pharyngeal and/or cibarial pump (often found in Anopheles spp.), inhibiting them from penetrating the midgut wall [11],[12]. Mf that successfully penetrate the midgut and enter the hemocoel come in contact with hemolymph components, including circulating blood cells called hemocytes. Melanotic encapsulation is a hemocyte-mediated, innate immune response that can be very specific and robust, and can limit or prevent parasite development in some mosquito species [13]. In contrast, the involvement of a humoral immune response is not well understood in many compatible filarial worm-mosquito systems. Some parasites are able to evade or suppress a host's immune system in order to survive, but it is unknown if such interactions occur between LF parasites and mosquitoes [14].
Previous studies have investigated the effect of an activated mosquito immune response on filarial worm development, but the results remain inconclusive. When bacteria are inoculated into the mosquito hemocoel, to induce the expression of infection responsive immune factors prior to filarial worm exposure, a reduced B. malayi prevalence and mean intensity in Ae. aegypti was observed as compared to non-inoculated controls [15]. However, when the same bacterial strains and inoculation procedures were used to pre-activate the immune response of Culex pipiens prior to W. bancrofti exposure, there was no difference in prevalence or mean intensity between bacteria-inoculated and control mosquitoes [16]. Further investigation is needed to assess the effects of an activated mosquito immune response on a LF parasite infection.
The role of Toll and Imd signaling pathways in the immune recognition, modulation, and response of mosquitoes to LF parasites has yet to be examined. Recently, Xi et al. [17] developed methods to manipulate these immune signaling pathways in Ae. aegypti by (1) gene silencing of Cactus, a negative regulator of the Toll pathway, (2) gene silencing of MyD88, an adaptor required for endogenous Toll pathway signal transduction, and (3) gene silencing of Caspar, a negative regulator of the Imd pathway. Using these tools, the Toll and Imd pathways are transiently relieved from endogenous suppression (1 and 3 above) or made unresponsive to detected stimuli (2 above).
In this study we assess transcriptome changes associated with the development of B. malayi in Ae. aegypti and investigate the effect of the immune signaling pathways, Toll and Imd, on parasite development.
All animals were handled in strict accordance with good animal practice as defined by the University of Wisconsin-Madison School of Veterinary Medicine/University of Wisconsin Research Animal Resource Center, and all animal work was approved by these entities.
Aedes aegypti black-eyed, Liverpool (LVP) strain used in this study were maintained at the University of Wisconsin-Madison as previously described [18]. Briefly, mosquitoes were maintained on 0.3 M sucrose in an environmental chamber at 26.5±1°C, 75±10% RH, and with a 16 hr light and 8 hr dark photoperiod with a 90 minute crepuscular period at the beginning and end of each light period. Ae. aegypti LVP was originally selected for susceptibility to Brugia malayi by Macdonald in 1962. This strain supports the development of mf to L3s. Four- to five-day-old mosquitoes were sucrose starved for 14 to 16 hours prior to blood feeding.
Mosquitoes were exposed to B. malayi (originally obtained from the University of Georgia NIH/NIAD Filariasis Research Reagent Repository Center) by feeding on ketamine/xylazine anesthetized, dark-clawed Mongolian gerbils, Meriones unguiculatus. The same animals were used for all three biological replicates. Microfilaremias were determined, using blood from orbital punctures, immediately before each feeding and ranged from 50–150 mf per 20 µl of blood. Control mosquitoes were exposed to anesthetized, uninfected gerbils. Mosquitoes that fed to repletion were separated into cartons and maintained on 0.3 M sucrose in the laboratory.
In early stages of development (1 h to 3 d post-infection [PI]), individual mosquitoes were separated into midgut, thorax and abdomen (with midgut removed) and dissected in Aedes saline [19]. Tissue dissections were cover-slipped and parasites were observed with a compound microscope using phase-contrast optics. The same procedure was used for dissections at 5–6 d PI, except only thoraces were dissected and examined. At 8–9 d PI, the thorax was processed as described above, but the abdomen and head & proboscis were dissected in separate drops of Aedes saline to observe L3s using a dissection microscope. Individual mosquitoes were dissected in a drop of Aedes saline for the recovery of L3s at 13–14 d PI. Images of B. malayi developmental stages were captured and processed as previously described, with the addition of Nomarski optics [20].
Five sample groups were created to study the transcriptional response of mosquitoes to B. malayi development. In each group, 20 mosquitoes were collected for RNA extraction. These sample groups are defined by the time after the blood meal and represent significantly different stages of parasite development. Briefly, Group 1 consisted of mosquitoes collected at 1, 6, 12 and 24 h PI. At these early time points, mf are penetrating the mosquito midgut, migrating through the hemocoel and penetrating thoracic muscle cells. Group 2 was collected at 2–3 d PI, a time when mf have differentiated into intracellular first-stage larvae (L1s). At 5–6 d PI, B. malayi complete the molt to second-stage larvae (L2s) and actively feed on mosquito muscle tissue (Group 3). In Group 4, at 8–9 d PI, parasite development is complete with a second molt to the L3s. Tissue damage continues as L3s break out of the thoracic muscles and migrate to the mosquito's head and proboscis. The final collection (Group 5), made at 13–14 d PI, occurs when the majority of L3s are located in the head and proboscis (see [4],[21]). Five mosquitoes, from both B. malayi-infected and uninfected blood meals, were collected at 1, 6, 12, and 24 h PI and ten mosquitoes at 2, 3, 5, 6, 8, 9, 13 and 14 d PI for transcriptional analysis. Mosquitoes were pooled (5 mosquitoes/tube), flash frozen in liquid nitrogen and stored at −80°C prior to total RNA extraction using RNeasy (QIAGEN). In addition, five B. malayi-infected mosquitoes were dissected to verify filarial worm infection and to determine the stage of parasite development at each time point. Three biological replications were completed.
Microarray assays were conducted and analyzed as reported previously [17],[22]. A full genome microarray platform (Agilent; 4×44k) was used with the probe sequences identical to the previous version (1×22k) [22]. In brief, 2–3 µg total RNA was used for probe synthesis of Cy3- and Cy5-labeled dCTP. Hybridizations were conducted with an Agilent Technologies In Situ Hybridization kit at 60°C according to the manufacturer's instructions. Three independent biological replicate assays were performed. Hybridization intensities were determined with an Axon GenePix 4200AL scanner, and images were analyzed with Gene Pix software. To produce the expression data, the background-subtracted median fluorescent values were normalized according to a LOWESS normalization method to reduce dye-specific biases, and Cy5/Cy3 ratios from replicate assays were subjected to t-tests at a significance level of p<0.05 using TIGR MIDAS and MeV software [23]. Expression data from all replicate assays were averaged with the GEPAS microarray preprocessing software prior to logarithm (base 2) transformation. Self-self hybridizations have been used to determine the cut-off value for the significance of gene abundance on these microarrays to 0.8 in log2 scale, which corresponds to 1.74- fold regulation [22]. For genes with P<0.05, the average ratio was used as the final fold change; for genes with P>0.05, the inconsistent probes (with distance to the median of replicate probe ratios larger than 0.8 log2) were removed, and only the value from a gene with at least two replicates was further averaged. The robustness of these microarray gene expression assays were validated through qPCR (Text S1).
Real time PCR assays were conducted as previously described to validate gene silencing efficiency and microarray expression data for selected genes [24]. Briefly, RNA samples were treated with Turbo DNAse (Ambion, Austin, Texas, United States) and reverse-transcribed using Superscript III (Invitrogen, Carlsbad, California, United States) with random hexamers. Transcript relative quantification was performed using the QuantiTect SYBR Green PCR kit (Qiagen) and ABI Detection System ABI Prism 7300 (Applied Biosystems, Foster City, California, United States). qRT-PCR reactions were conducted using a 10 minute step at 94°C and 40 cycles of 15 seconds at 94°C, 15 seconds at 55°C and 15 seconds at 72°C. Three independent biological replicates were conducted and all PCR reactions were performed in triplicate. Transcript abundance was normalized against the ribosomal protein S7 gene. All primers used for qPCR assay are presented in Table S1. The primer sequences used to verify gene knockdown efficiency included: S7 (AAEL009496-RA) forward: 5′-GGGACAAATCGGCCAGGCTATC-3′, reverse: 5′- TCGTGGACGCTTCTGCTTGTTG-3′; Caspar (AAEL003579-RA) forward: 5′-GAATCCGAGCGAGCCGATGC-3′, reverse: 5′-CGTAGTCCAGCGTTGTGAGGTC-3′; Cactus (AAEL000709-RA) forward: 5′-AGACAGCCGCACCTTCGATTCC-3′, reverse: 5′-CGCTTCGGTAGCCTCGTGGATC-3′; MyD88 (AAEL007768) forward: 5′-CATCCCATTCAGTTTCTCAGC-3′, reverse: 5′-ACCGGTTGGAAGTTCTGATG-3′. A complete list of PCR primer sequences is presented in Table S1.
RNAi was conducted by intrathoracic injection of dsRNA using described methodology [24],[25]. Mosquitoes were three to four days old at the time of blood feeding and dsRNA was injected either 48 h before or after parasite exposure, i.e., dsRNA injections were performed on non-blood fed one- to two-day-old mosquitoes and on blood fed, five- to six-day-old mosquitoes. Approximately 0.5 µl of dsRNAs (1.0 or 0.5 µg/µl) were injected into the thorax of cold-anesthetized mosquitoes. The primers used to synthesize Cactus, Caspar and MyD88 dsRNA have been published previously [17]. To synthesize GFP dsRNA, methods described by Bartholomay et al. [25] were used with minor changes. The following sequences were annealed by heating at 95°C for 5 min and slow cooling: GFP_F 5′-TAGTACAACTACAACAGCCACAACGTCTATATCATGGCCGACAAGCAGA AGAACGGCATCAAGGTGAACTTCAAGATCCGCCACAACA-3′ and GFP_R: 5′-TCGATGTTGTGGCGGATCTTGAAGTTCACCTTGATGCCGTTCTTCTGCTTGTCGGCCATGATATAGACGTTGTGGCTGTTGTAGTTGTA-3′ (Integrated DNA Technologies, Inc., Coralville, IA). This dsDNA was ligated into pBlueScript KS+ (Stratagene) at XbaI (T7) and SalI (T3) sites. A complete list of PCR primer sequences is presented in Table S1.
B. malayi exposures were performed as described above and microfilaremias ranged from 35–162 mf per 20 µl of blood. Mosquitoes injected with GFP dsRNA were blood fed on each infected gerbil and used as a control for parasite infections. Mosquito mortality was observed every 24 h and mosquitoes were dissected at 6 d or 12–13 d PI to observe parasite development. To verify gene knockdown, five mosquitoes were collected 48 h post dsRNA injection, placed in a microcentrifuge tube, flash frozen and stored at −80°C until RNA extraction. Real-time PCR was used to quantify gene silencing efficiency. Silencing of Cactus, Caspar and MyD88 resulted in a reduction of mRNA levels by 60%, 84% and 27%, respectively. For each exposure, the prevalence and mean intensity of infection was calculated. Comparisons of mean intensities and mosquito mortality curves were done with the Mann-Whitney and Log-Rank tests, respectively, using GraphPad Prism 5 (GraphPad Software, Inc., La Jolla, CA). Results were considered significant at P≤0.05.
The development of B. malayi was observed at 1 h to 14 d PI for each biological replicate and is summarized in Table S2. Worms were recovered from 166 of the 180 mosquitoes examined for an overall infection prevalence of 92%. Mf were recovered from 1 h to 3 d PI, but after 24 h PI mf were no longer the most abundant developmental stage recovered. At 2 d PI, mf recovery began to decrease and almost all worms had differentiated into intracellular L1s by 3 d PI. Parasites molted to L2s in the thoracic musculature by 5 d PI, and were the only developmental stage identified in transcriptional group 3. The molt from L2 to L3 occurred at 8–9 d PI. At 8 d PI, L2s and L3s were recovered in the thorax, and only 4% of the worms were located in the head and proboscis. In contrast, at 9 d PI, both L2s and L3s were observed in the thorax region, but the majority were L3s and 47% of all recovered worms were located in the head and proboscis. By 13–14 d PI, all parasites had developed into L3s. Images of B. malayi development, from mf to L3s, are presented in Figure 1. The prevalence of L3s (for all three biological replicates at 13–14 d PI) was 80% (n = 30) and the mean intensity was 6.9±5.7.
The Ae. aegypti global transcript responses to the successful development of B. malayi were determined using a genome microarray expression approach. These transcriptome infection-response patterns differed significantly, in both the number of regulated genes and their direction of regulation, at the different parasite development stages (Fig. 2A). A general suppression of transcription was evident during the early stages of infection (group 1); 20 genes were down-regulated while only 8 genes were up-regulated. However, this transcriptional suppression was reduced when parasites developed into L1s (group 2), and reversed when they became L2s (group 3); 59 genes were up-regulated and only 5 genes were down-regulated at this stage. Parasite infection at later stages of infection mainly caused transcriptional up-regulation (groups 4 and 5). Strikingly, 120 genes were up-regulated in group 5 that represented mosquitoes in which L3s had migrated to the head and proboscis from where they can be transmitted to a host upon blood feeding. As many as 158 mosquito genes were differentially expressed at this late stage. The second largest number of infection-responsive genes was observed when L2s were present (group 3).
Interestingly, the transcripts that were over represented in the groups that displayed the most prominent transcriptional regulation (3 and 5), were highly enriched with putative immune genes (Fig. 2). Specifically, several antimicrobial peptide effector genes were strongly induced in group 3 and a number of putative pattern recognition receptor and signal modulator genes were up-regulated in group 5 (Table 1). Group 5 also contained several serine protease cascade components and putative melanization –related factor transcripts in increased abundance (Tables 1 and 2).
To investigate whether the mosquito's two major immune signaling pathways, Toll and Imd, had any effect on B. malayi infection we used an established gene silencing approach to either simulate the activation of the Toll and Imd pathway, by depleting their negative regulators Cactus and Caspar, respectively, or inhibit the Toll pathway through the depletion of the MyD88 factor [17],[26].
Caspar and MyD88 gene knockdown did not change the mortality rate in either pre- or post-bloodfed dsRNA-injected mosquito groups compared to their respective GFP dsRNA controls (Log-rank tests, P-values ranging from 0.23 to 0.41; see Fig. 3 and 4). In contrast, Cactus gene knockdown, that results in the activation of the Rel1 factor, resulted in a significantly increased mosquito mortality in both pre- and post-blood feeding Cactus dsRNA injected groups (Log-rank test, P<0.001; data not shown). This increased mortality was not related to infection status (Fig. 5A and B).
There was no significant difference in B. malayi mean intensities between Caspar, Cactus or MyD88 silenced mosquitoes compared to the GFP dsRNA injected controls that had fed on the same microfilaremic gerbil. Likewise, there was no difference in prevalence or mean intensity of L3s between Caspar and MyD88 depleted mosquitoes as compared to their respective GFP dsRNA injected controls before (Fig. 3B and 4B) or after (Fig. 3D and 4D) blood feeding. Although knockdown of Cactus increased the mortality rate of Ae. aegypti, parasites that were recovered from live mosquitoes at six and 12 d PI had developed normally and there was no difference in the infection prevalence or mean intensity compared to the controls (Figure 5).
In this study, we provide insights into the interactions between filarial worms that cause human LF and their compatible mosquito vectors. As Brugia and Wuchereria parasites develop, the mosquito experiences a series of insults that include: (1) the penetration of cells and tissues by mf, (2) the consumption of cellular material by developing larvae, and (3) the migration of L3s through the body cavity.
The infection response of mosquitoes is surprisingly diverse during the course of nematode development, as different gene transcripts and regulatory trends were observed in each of the five different developmental time points examined. By infection response we refer to the overall transcriptional and physiological change that occurs in the mosquito as a result of parasite infection, and it includes a vast array of distinct types of responses (e.g., repair, immune, metabolic, reproductive, behavioral, etc.). Overall, the response to filarial worm infection in this compatible system is mainly comprised of molecules involved in cellular signaling, proteolysis, stress response, transcriptional regulation, and repair (see Table S3). And these include several genes that have traditionally been classified as immunity related (Table 1). Very few transcriptional changes were observed until L2s were present, and the most profound transcriptional changes were observed in mosquitoes that harbored infective-stage parasites for 4–5 days. A large proportion of the regulated transcripts represented genes of unknown function (32.9%), and genes that have multiple or diverse functions (40%).
As expected, the transcriptomic profiles of B. malayi-infected Ae. aegypti are very different than those previously described in B. malayi-infected Armigeres subalbatus (see Table S4) [27]. In this non-compatible relationship, B. malayi development does not occur due to the rapid recognition and melanization of mf in the hemocoel of Ar. subalbatus [28]. The different transcriptional changes following infection of mosquitoes that support parasite development and those that do not can provide clues to the molecular mechanisms that determine compatible versus incompatible mosquito-filarial worm associations. Such comparisons have been made between Cx. pipiens and Ae. aegypti infected with W. bancrofti [29], but transcriptional responses that occur in these mosquitoes may not represent genes that are used to deter filarial worm infection in an incompatible system, i.e., it is quite possible that differences in gene transcription of mosquitoes in different genera could represent unique strategies for overcoming damage caused by filarial worms and therefore do not represent anti-filarial worm responses. Similarly, identification of immune-responsive genes activated in response to filarial worm infection does not indicate that the mosquito is/has mounted an immune response against the parasites itself [30]. It is possible that the observed response could be an indirect effect caused by the infection, i.e., mf midgut penetration or muscle cell damage that occurs later in development in compatible mosquitoes. The current study provides transcriptional data from a strain of Ae. aegypti that is highly compatible with B. malayi, and can help guide the planning of future studies measuring transcriptional changes in a strain of Ae. aegypti that does not support the development of B. malayi.
The differences in parasite size (Fig. 1) and behavior among developmental stages provide a foundation for discussing the response of Ae. aegypti to a successful B. malayi infection. In mosquitoes sampled from 1–24 h PI, mf are in the process of penetrating the midgut epithelium and migrating through the hemocoel to penetrate the indirect flight muscle cells. Differentiation to L1s begins as soon as mf become intracellular parasites. At this early stage of infection, transcriptional profiles suggest the presence of B. malayi may alter blood digestion/proteolysis (four serine proteases; sterol trafficking), chitin-related interactions (two transcripts contain the chitin-binding Peritrophin-A domain; IPR002557), and immune function (DEF D; AAEL003857).
B. malayi-infected mosquitoes sampled at 2–3 d PI harbor L1s, a stage when parasites have a markedly decreased mobility within the indirect flight muscle cells and are in the process of developing a digestive track. The few infection-responsive transcripts (13 genes) during infection with L1s include three down-regulated immunity-related genes: CEC F (AAEL000625), CEC A (AAEL000627), and a hypothetical protein (AAEL003843) which is a putative knottin with an interesting genomic location; just upstream and on the opposite strand from DEF A (AAEL003841). The regulation of cecropin transcripts in response to pathogens is complex [31], and the interpretation of their decreased transcriptional abundance therefore remains speculative.
Mosquitoes sampled at 5–6 d PI contain worms that have molted to L2s. At this stage in development, the parasites are actively ingesting cellular material, have developed a digestive system (with an open mouth but an anus still closed) and have grown four times in length compared to mf. Even though filarial worms remain intracellular until the molt to L3s, these internally damaged cells are likely to provide the necessary stimuli for the mosquito's infection response. Genes involved in cellular signaling (e.g., G-protein coupled receptors, Spaetzle 5, DSCAM) and transcriptional regulation (i.e., changing patterns in transcription factors) were identified as components of the infection response at this time interval (Table S3). Another component of the infection response to L2s is the increased abundance of six cecropin transcripts (Table 1).
At 8–9 d PI, worms have molted to L3s and begun migrating from the thoracic musculature to the mosquito head and proboscis. In Ae. aegypti, the onset of severe muscle damage occurs when L3s have exited the infected muscle cells [6]. It has been noted that this muscle damage, and the subsequent damage repair response, also occurs in mosquitoes that are mechanically damaged by external thoracic punctures [32]. Considering these ultrastructural observations by Beckett et al. [6], [32]–[34], it is interesting to find that the infection response in this group involved only modest changes in transcript abundance. Of the 26 infection responsive transcripts identified during L3 migration, the majority are induced and putatively involved in stress response (n = 4) and transcriptional regulation (n = 4; see Table S3).
The most profound transcriptomic changes, in response to infection, occurred at 13–14 d PI, an infection stage when mosquitoes have harbored L3s for approximately 4–5 days. Many components of intra- and extracellular signaling pathways (n = 12; see Table S3) were differentially transcribed in these infected mosquitoes, again lending support to the fact that detection and communication of stimuli is key to mounting a response and repairing tissues. L3-infected mosquitoes seem to be responding to tissue damage with 15 genes identified with possible functions in proteolysis and seven insect cuticle protein transcripts in increased abundance. There are multiple sources of tissue damage at this point in infection: (1) the degrading muscle cells that supported the development of the parasites [6], (2) L3 migration throughout the mosquito body cavity [35], and (3) the ability of L3s to penetrate through the cuticular surface [36]. The data also show that infected mosquitoes are responding to the stressful conditions of harboring L3s (9 stress response-related transcripts).
Studies have shown that mosquito behavior can be modified by filarial worm infection, and occurs in an intensity-dependent manner [9]. Comparisons between an earlier study of spontaneous flight behavior changes in Brugia-infected Ae. aegypti [9] and transcriptional changes seen in the present study could be made for three of our five groups. The estimated mean intensities of Brugia infections in mosquitoes collected for transcriptional analysis fit within the categories of low (1–10 parasites) to moderate (11–20 parasites) intensities created by Berry et al. [9]. Changes in the transcriptome of Brugia-infected mosquitoes are associated with an increase in flight behavior during the time L2s are feeding (Group 3; 5–6 d PI), a marked decrease followed by recovery of flight when L3s emerge and migrate from infected muscle cells (Group 4; 8–9 d PI), and up to a 60% decrease in spontaneous flight activity when mosquitoes harbor Brugia L3s (Group 5; 13 d PI). The data from ultrastructural [6], [32]–[34], behavioral [9], and transcriptional observations (presented herein) all support the fact that filarial worm development is not a benign infection to mosquitoes.
Certain mosquito-borne pathogens are known to be controlled by vector immune responses, which are regulated by intracellular signaling pathways, such as Toll and Imd. For example, the Toll and Imd pathways in An. gambiae regulates infection with malaria parasites (Plasmodium berghei and P. falciparum) and is required for antibacterial defenses [26],[37]. Innate immune response in tsetse flies has also been implicated in regulating the intensity of trypanosome infection [38]. An immune response is considered a mechanism by which a host attempts to eliminate or reduce an infection. A host's immune response to parasitism may however not always lead to an elimination of parasites because of the latter's capacity to evade the immune defense mechanisms [39]. Previous studies in our laboratory suggest that LF parasites either elicit an immune response, e.g., melanotic encapsulation, or go undetected and therefore unmolested by an immune response in certain mosquitoes [28],[40]. Although the interactions between these nematodes and the mosquito immune system are mechanistically undescribed, there is potential for LF parasites to evade and/or suppress the mosquito immune system [41]–[44]. In this study, we manipulated the mosquito immune system in an effort to activate immune response pathways to determine what effects, if any, they might have on parasite infection and development.
We used a RNAi–mediated gene knockdown approach to transiently activate the two major immune signaling pathways, Toll and Imd, by targeting their negative regulators, Cactus and Caspar, respectively. Post-transcriptional silencing of these pathway regulators leads to pathway-specific immune responses. Previous studies on the effect of mosquito immune responses on filarial worm development utilized bacterial challenges to activate the immune system [15],[16]. We selected time points for the activation of immune pathways that would specifically target the parasites early in development, i.e., when they might be most vulnerable to the mosquito immune system; when microfilariae migrate to the thoracic musculature and when parasites undergo the first molt (first- to second-stage larvae). The activation of these immune pathways had no detectable effect on B. malayi development in Ae. aegypti (Fig. 3–5). The lack of an anti-parasite effect as a result of activating Toll and Imd pathways suggest that the parasite limiting mechanism, that was observed in bacterial-challenged Ae. aegypti, was not attributed to a Rel1 or Rel2 nuclear translocation. This may imply that the bacteria challenge induced some other defense system, independently of the Toll or Imd pathways, or that the bacteria exerted a direct anti-parasitic effect on the filarial worms.
Infection of Ae. aegypti with a compatible filarial parasite, B. malayi, resulted in fairly few changes in the mosquito transcriptome; however, these infection responses were diverse and differed vastly between the different infection stages. The majority of these transcriptional infection responses are most likely a reflection of the mosquito's attempt to repair tissue damage resulting from nematode development. We have also shown that removing the inhibitors of Rel1 and Rel2 activation did not affect the permissiveness of this mosquito to B. malayi infection. This observation may indicate a resistance, immune evasion and/or suppression stategy(ies) by the parasite, whereby it remains inert to destruction by the mosquito's immune system.
Not all mosquitoes respond the same to LF parasite infection, and differences between natural and artificial systems should be carefully considered. Aedes aegypti is a common laboratory vector that has been genetically selected for susceptibility to many pathogens, including B. malayi [45], but is not a natural vector of LF parasites [46]. Investigations of flight muscle cell damage caused by developing B. malayi in natural and artificial vectors have concluded that tissue damage is more severe in Ae. aegypti compared to Mansonia uniformis, which is a natural vector [6],[34]. This increased pathology occurs when L3s migrate out of the flight muscle cells, and is reflected by a spike in Ae. aegypti mortality [47]. It is apparent that Ae. aegypti may utilize different mechanism(s) for surviving infection, and future studies comparing the infection response of natural mosquito-LF parasite systems would allow a better assessment of these differences.
As advancements are made within the field of lymphatic filariasis parasite-host interactions, it will be interesting to compare the infection responses of both the vertebrate and invertebrate hosts. Mf and L3s are the developmental stages transmitted between hosts, and are known to elicit a vertebrate immune response [48]–[50]. The short time interval between L3s escaping the mosquito and infecting the vertebrate host exemplifies the link between the two host environments. For example, the unknown mechanism(s) employed by L3s to suppress the infection response of vertebrates [51] might be functional before L3 escape and may therefore also act on the innate immune response of the mosquito.
|
10.1371/journal.pgen.1004144 | Relationship Estimation from Whole-Genome Sequence Data | The determination of the relationship between a pair of individuals is a fundamental application of genetics. Previously, we and others have demonstrated that identity-by-descent (IBD) information generated from high-density single-nucleotide polymorphism (SNP) data can greatly improve the power and accuracy of genetic relationship detection. Whole-genome sequencing (WGS) marks the final step in increasing genetic marker density by assaying all single-nucleotide variants (SNVs), and thus has the potential to further improve relationship detection by enabling more accurate detection of IBD segments and more precise resolution of IBD segment boundaries. However, WGS introduces new complexities that must be addressed in order to achieve these improvements in relationship detection. To evaluate these complexities, we estimated genetic relationships from WGS data for 1490 known pairwise relationships among 258 individuals in 30 families along with 46 population samples as controls. We identified several genomic regions with excess pairwise IBD in both the pedigree and control datasets using three established IBD methods: GERMLINE, fastIBD, and ISCA. These spurious IBD segments produced a 10-fold increase in the rate of detected false-positive relationships among controls compared to high-density microarray datasets. To address this issue, we developed a new method to identify and mask genomic regions with excess IBD. This method, implemented in ERSA 2.0, fully resolved the inflated cryptic relationship detection rates while improving relationship estimation accuracy. ERSA 2.0 detected all 1st through 6th degree relationships, and 55% of 9th through 11th degree relationships in the 30 families. We estimate that WGS data provides a 5% to 15% increase in relationship detection power relative to high-density microarray data for distant relationships. Our results identify regions of the genome that are highly problematic for IBD mapping and introduce new software to accurately detect 1st through 9th degree relationships from whole-genome sequence data.
| The determination of the relationship between a pair of individuals is a fundamental application of genetics. The most accurate methods for relationship estimation rely on precise, localized estimates of genetic sharing between individuals. Earlier methods have generated these estimates from high-density genetic marker data. We performed relationship estimation using whole-genome sequence data for 1490 known pairwise relationships among 258 individuals in 30 families along with 46 population samples as controls. Our results demonstrate that complexities specific to whole-genome sequencing result in regions of the genome that are prone to false-positive estimates of genetic sharing. We provide a map of these spurious IBD regions and introduce new methods, implemented in the software package ERSA 2.0, to control for spurious IBD. We show that ERSA 2.0 provides a 5% to 15% increase in relationship detection power for distant relationships with whole-genome sequence data relative to high-density genetic marker data.
| The identification of related individuals from genetic data has a broad range of applications. The validation of known relationships in familial disease-gene studies ensures that pedigree errors or sample switches do not adversely affect power [1]. In case-control studies, the removal of related individuals is a standard quality control step to avoid spurious associations [2]. Population genetics studies typically must either explicitly account for familial relationships [3], or else exclude related individuals from analyses that rely on random mating and representative sampling assumptions [4]. Genetic relationship identification is also widely used in a number of forensic applications, including criminal investigations, identification of missing persons and victims of mass disasters [5], [6].
Methods applicable to the detection of close relationships have been available for decades [1], [7]. These methods typically rely on either genome-wide estimates of identity-by-descent (IBD) [8] or joint inference of IBD and relationships using sparse genetic markers [9]. With approximately 1,000 highly polymorphic markers, such methods are well powered to accurately identify relationships as distant as 3rd-degree relatives [9], but these methods do not benefit from further increases in marker density [10]. With the introduction of single-nucleotide polymorphism (SNP) microarrays, increased marker density enabled the accurate detection of local IBD segments. Newer relationship estimation methods take advantage of local IBD segment data to increase the range of detectable relationships [10], [11]. The relationship estimation software that we previously developed, Estimation of Recent Shared Ancestry (ERSA), has high power to detect relationships as distant as 8th-degree relatives (e.g., 3rd cousins once removed) from high-density SNP microarray data [10].
Whole-genome sequence (WGS) should represent the final step in increasing marker density, and thus, improved relationship detection accuracy. However, with the current complexity of WGS, which is based on high-throughput short-reads mapped to a legacy reference genome, a number of technical challenges must be overcome before potential improvements in relationship detection accuracy can be realized. To assess these challenges, we analyzed WGS data for 1490 distinct pairwise relationships from 258 individuals in 30 families (see Table 1). Our results highlight new issues specific to relationship estimation from WGS data and introduce new methods in ERSA 2.0 to mitigate these issues.
To evaluate relationship-estimation accuracy on WGS data, we first inferred IBD segments between each pair of individuals with three different methods: Genetic Error-tolerant Regional Matching with Linear-time Extension (GERMLINE), Beagle's fastIBD, and Inheritance State Consistency Analysis (ISCA) [12]–[14]. We then applied ERSA separately to each of the three resulting IBD-segment datasets. For our initial analysis of control genomes from putatively unrelated individuals of European ancestry, we set the chance of falsely detecting a relationship between unrelated individuals to 0.1% (α = 0.001). With this threshold, we detected a significant relationship of 9th-degree or closer using GERMLINE in approximately 10% of all pairs of individuals. The estimated level of cryptic relatedness was 10-fold higher than we previously observed from high-density microarray data in this population [10], and thus was a strong indication of an elevated false-positive rate (Table 2). After further investigation, we identified several regions of the genome that were detected to be IBD far more often than would be expected by chance among pairs of controls (see Materials and Methods). Table 3 shows 14 regions of the genome greater than 5 cM with detected pairwise IBD identified in GERMLINE that exceeds the expected pairwise IBD by at least 4-fold between European controls. The regions of spurious IBD were largely consistent between the three IBD methods and among European, East Asian, and Mexican American populations (Figures 1 and S7, Tables 3 and S2, S3), which is a strong indication that the IBD segments in these regions are artifactual. To account for these spurious IBD segments, we developed a procedure within ERSA 2.0 to identify and mask regions of the genome with excess IBD in controls (see Materials and Methods). After applying this procedure, the rate of detected relationships among the European controls decreased from 10% to 1% at α = 0.001 using ERSA 2.0 and GERMLINE, which is the rate of cryptic relationships that we previously observed in this population [10]. In addition to region masking, we also implemented new models in ERSA 2.0 that improve the accuracy of relationship estimates for closely related individuals (see Materials and Methods). Although fastIBD detected many of the same regions as GERMLINE and ISCA, the rate of spurious IBD detection was generally much lower (Table 1 and Figure 1). For this reason, the rate of detected relationships among European controls was less than 0.002 at α = 0.001 using ERSA 2.0 and fastIBD, even without masking spurious IBD segments.
Figure 2 summarizes the ERSA 2.0 results from the 30 pedigrees (see also Table S1 and Figure S3). ERSA 2.0 detected all 1st through 6th degree relationships and 55% of 9th through 11th degree relationships in the 30 pedigrees. The performance of ERSA was very similar across the three IBD detection methods, with approximately a 5% difference in exact relationship prediction accuracy.
Although the 30 pedigrees included 1490 documented pairwise relationships, only 28 of these relationships were more distant than 6th degree. To evaluate performance of ERSA 2.0 and IBD detection methods for more-distant relationships, we simulated WGS data in 15-generation pedigrees (See Materials and Methods; Figures S1). ERSA 2.0 performed well with all three IBD detection methods (Figures 3, S5, and S10). For each method, we observed greater than 95% power to detect relationships as distant as 5th degree and greater than 50% power to identify relationships as distant as 8th degree (α = 0.001). We also performed IBD estimation using subsets of the data to represent SNP microarray data (using the set of positions from the Affymetrix 6.0 array) and whole-exome data. The increase in marker density from SNP microarray data to WGS data resulted in a 5% to 15% increase in power for distant relationships between 7th and 11th degree (Figure 3). Restricting markers to exonic regions reduced power relative to WGS data, with a 10% to 60% decrease in power for GERMLINE-ERSA and ISCA-ERSA and a 5% to 10% decrease in power for fastIBD-ERSA with 5th through 12th degree relationships (Figure 3). With exonic markers, we observed a modest increase in the rate of detected relationships among control populations of between 0.2 to 0.5% (Table 2). However, for exonic markers in simulated families, the power to detect distant relationships (10th–15th degree) increased by as much as 5%. This increase in power is very likely to be an artifact and is probably an indication that the increased difficulty of detecting IBD data from exonic markers may lead to improperly calibrated Type I error in ERSA 2.0 for some whole-exome datasets.
To compare ERSA to an approach that does not rely on local IBD segment estimates, we also estimated pairwise relationships using RELPAIR, a method that jointly estimates IBD and relationships using sparse marker data. RELPAIR's performance was similar to ERSA for 1st and 2nd degree relationships. Both approaches accurately differentiate between parent-offspring and full-sibling relationships in over 96% of comparisons. RELPAIR had no ability to differentiate between 3rd through 5th degree relationships and had low power to detect relationships more distant than 5th degree (Figures S4 and S6), as previously reported [10].
Our results demonstrate that several regions of the genome exhibit an excess of detected IBD with state-of-the-art WGS and IBD detection methods. These suspect IBD regions were typically characterized by an increase in departures from Hardy-Weinberg Equilibrium and were often near centromeric regions. Gaps in the recombination map and human reference assembly were also overrepresented. For example, although the regions in Table 3 represent less than 5% of the human genome, they represent 13% of the centromeric regions and 47% of the unmappable heterochromatic regions of the genome (“Gap” tracks in the UCSC Genome Browser). Notably, the IBD regions were not enriched for repetitive segments of the genome [15]. Because many of the regions were identified using three distinct IBD detection methods, the regions we identified with spurious IBD are unlikely to be the result of IBD detection algorithm errors. Although strong recent positive selection can produce this effect on a population scale, positive selection is unlikely to explain this result because the regions we identified were typically detected among Europeans, East Asians, and Mexican Americans and were far larger than any previous reported genomic signal of positive selection in humans (Figure 1 and Table 3). In addition, we observed very little overlap between the regions identified in Table 3 and a genome-wide search for genomic regions influenced by positive selection based on signals of excess IBD (Table S4) [16]. The regions identified using WGS data usually exhibited excess IBD in Affymetrix high-density microarray data as well, although at lower magnitudes and with smaller segment sizes (Table 3), suggesting that the excess IBD is not simply due to artifacts specific to high-throughput short-read resequencing. One potential explanation is that errors in published genetic maps in these regions overestimate the size of the IBD segments when measured by genetic distance. This hypothesis is supported by the gaps in the published recombination maps and relatively sparse high-density microarray marker density in these regions. Gaps in the human reference assembly may be another contributing factor, both directly due to the absence of markers and indirectly as a general indicator of mapping difficulty in flanking regions. The increased rate of deviations from Hardy-Weinberg equilibrium could also provide a partial explanation, given that erroneous heterozygote calls can result in false inferences of IBD segments. Some of the regions we have identified may be the result of long-range haplotypes with limited recombination between haplotypes [17], [18]. Of the 14 regions identified in Table 3, seven overlapped with regions previously identified in studies of long-range linkage disequilibrium (Tables S5, S6) [17], [18]. One of these regions, at position 10.5 to 13.5 Mb on chromosome 8, overlaps with a known inversion polymorphism that suppresses recombination between haplotypes [19].
Our analysis focused on three complementary pairwise IBD detection methods, GERMLINE, fastIBD, and ISCA (Figure 4). GERMLINE accepts phased genotype data and employs a haplotype hashing algorithm to reduce computation time [12]. Although GERMLINE is capable of analyzing unphased data, in our experience IBD segment identification and subsequent relationship estimation accuracy are both greatly reduced. Beagle fastIBD employs a similar approach to GERMLINE, but obtains multiple estimates of haplotype phase internally and evaluates each of these haplotypes [13]. The rate of spurious IBD detection in fastIBD was substantially lower than GERMLINE and ISCA, and we did not observe an excess of detected relationships among control populations with fastIBD and ERSA, even in the absence of masking (Table 2). However, the power of ERSA 2.0 to detect relationships was slightly reduced with fastIBD relative to the other two methods (Figure 3). Both GERMLINE and fastIBD are well optimized for large sample sizes, but neither distinguishes between haploid-identical regions (IBD1) and diploid-identical regions (IBD2). We originally described ISCA as a method for simultaneous detection of all blocks of identity throughout a pedigree [20], [21]. ISCA also performs well for detecting both IBD1 and IBD2 segments between pairs of individuals with an unknown relationship. ISCA employs a Hidden Markov Model that identifies both IBD1 and IBD2 segments [20]. Because ISCA is optimized for whole-genome data, the algorithm suppresses noise from segments of the genome that give rise to false positive IBD1 and IBD2 regions, such as compressions, centromeres, hemizygous regions, CNVs, reference gaps, and other irregularities [20]. Unlike GERMLINE and fastIBD, ISCA does not require phased data or population controls. However, because ISCA's execution time scales linearly with the number of individual pairs, it is slower than both GERMLINE and fastIBD for large sample sizes.
All of the datasets we evaluated included complete documentation of missing genotypes (i.e. no-calls). In our experience, missing genotype data are essential to accurate IBD estimation. Variant call data that does not report missing genotypes should not be used for relationship estimation.
WGS data present new challenges for IBD detection and relationship estimation. Using existing approaches, we observed a major increase in the detection of spurious IBD segments and false-positive relationships from WGS data of population controls. We provide a map of spurious IBD regions in the human reference sequence and present methods implemented in ERSA 2.0 that mask these regions to accurately detect pairwise relationships from IBD segment data. ERSA 2.0 also incorporates additional refinements to improve relationship detection accuracy for 1st- and 2nd-degree relationships. When error-prone IBD regions are masked, the relationship estimation methods in ERSA 2.0 perform well for a variety of IBD detection methods, including GERMLINE, fastIBD, and ISCA. Compared to high-density microarray data, WGS data provide a 5% to 15% increase in relationship detection power for 7th through 12th-degree relationships. Whole-exome data perform substantially worse than high-density microarray data for this purpose. Our results demonstrate that ERSA 2.0 can detect relationships as distant as 12th degree and has high power to detect relationships as distant as 8th degree from whole-genome sequence data.
We included 258 individuals from 30 families and 46 unrelated individuals (34 Europeans, 4 Mexican-Americans, and 8 East Asians) in this study. We evaluated population structure for each unrelated individual and for one member of each family by performing principal components analysis (PCA) that incorporated HapMap population samples [22]. Of the 30 pedigrees, 25 clustered with Europeans, 3 with Mexican-Americans, and 2 with East Asians (Figure S2). The 30 pedigrees include 1490 documented pairwise relationships (see Table 1). One of these pedigrees was CEPH Pedigree #1463, which consists of genomes of a seventeen-member, three-generation pedigree, with publically available data (ftp://ftp2.completegenomics.com/Pedigree_1463/). Complete Genomics performed all whole-genome sequencing.
With the exception of the publicly available CEPH Pedigree #1463, all other pedigree datasets are protected by human subjects protocols approved by the Western Institutional Review Board. Procedures followed were in accordance with institutional and national ethical standards of human experimentation. Proper informed consent was obtained. During subject recruitment, relationships were determined by interview and recorded.
We simulated non-founder whole-genome data from fifteen-generation families (Figure S1), selecting founders randomly from the unrelated individuals of European ancestry described above. The whole genomes of two offspring were simulated in each generation. Genotypes of non-founders were obtained by simulating meiosis (recombination points were randomly selected based on the recombination rate map in [23]) and de-novo mutation with an expected rate of 1e-7. Sequencing errors were added to all non-founder genomes with an error rate of 0.001 per polymorphic site. There were 1035 pairs of individuals in each family, containing 330 unrelated pairs, 75 first-degree relationships (60 parent-offspring and fifteen full sibling pairs), 84 second-degree relationships, 78 third-degree relationships, 72 fourth-degree relationships, and 66 fifth-degree relationships.
We used ISCA to infer pairwise IBD1 and IBD2 segment estimates from unphased SNV data. We used Beagle and fastIBD to compute IBD estimates from unphased SNV data separately for each population. Each population combined European, Mexican-American, or East Asian control individuals with the pedigrees that clustered with those populations in PCA (Figure S2). We chose sequenced European genomes to serve as founders for each of the simulated pedigrees. The simulated pedigree genomes were phased with the European controls. Per the authors' recommendations, we ran fastIBD 10 times in each population and merged all segments within one megabase that overlapped between any of the 10 output files [13]; this additional step proved necessary for accurate relationship estimation in ERSA (Note that in our previous evaluation of fastIBD in ERSA we did not perform this step [10]). For GERMLINE, we first applied the grouping criteria above in three population analyses to phase each pedigree and each of the three control populations using Beagle [24], and then analyzed the phased data in GERMLINE.
We applied identical procedures for subsets of SNVs that lie within protein-coding exon boundaries or are Affymetrix 6.0 markers (Figure 3). For GERMLINE, we pruned the WGS datasets prior to phasing in Beagle. After generating IBD segments, we evaluated GERMLINE and fastIBD in ERSA 2.0. We estimated relationships for every pair of individuals within the pedigrees, using the appropriate control population identified in Figure S2.
ERSA models the distribution of IBD segments between two individuals in a maximum likelihood framework. The null model assumes that the size and number of IBD segments follow an empirical distribution approximated from the control population. Under the alternative model, some IBD segments may follow the control population distribution, but one or more segments follow a theoretical distribution derived according to a hypothesized recent relationship. Let a equal the number of shared ancestors and d equal the total number of meioses that separate the two individuals for the proposed relationship. For each pair of individuals, ERSA calculates the maximum likelihood for each possible relationship to identify the most likely relationship for that pair. We use the chi-square approximation to the maximum likelihood ratio to establish confidence intervals and to test for significance. This test has two degrees of freedom. One degree of freedom results from a parameter describing the number of segments that are attributable to hypothesized relationship for the pair of individuals (the remaining segments are attributed to the population distribution). A second degree of freedom results from the parameters d and a, which act approximately as a single parameter for most values of d. For direct ancestor-descendant relationships, a = 0. In ERSA 1.0, we assume that, for most relationships, the length l of an IBD segment inherited from the proposed relationship is exponentially distributed with mean equal to(1)in cM [25]. This approximation assumes that only recombination can break up an IBD segment. Because IBD segments are also broken at chromosomal boundaries,(2)where c is the number of autosomes and r is the expected number of recombination events per generation (r∼ = 35 in humans [23]). As d increases, Eq. 2 approaches Eq. 1, and thus Eq. 1 is a close approximation for distant relationships but is less accurate for close relationships. ERSA 2.0 uses Eq. 2 when a is equal to 1 or 0, resulting in an improvement in accuracy for closely related individuals (Figure S8). Empirically, we observed that Eq. 2 slightly reduced relationship estimation accuracy when a is equal to 2, perhaps due to minor biases in the estimated IBD lengths in GERMLINE and ISCA. Thus, we continue to use Eq. 1 for models where a is equal to 2. Both versions assume that the number of IBD segments, n, is Poisson distributed with mean equal to(3)
Modifications of these formulas for specific relationships are described below. Hill and White have very recently employed simulations to derive precise estimates for the joint distribution of the number and length of shared segments for a wide range of relationships [26]. These distributions are likely to provide further improvement in relationship detection accuracy in the future.
Although related individuals may sometimes share short IBD segments, such segments can be difficult to distinguish from more distant IBD segments that would be shared by unrelated members of the population. Thus, we typically set a minimum IBD segment length, t, and ignore all segments smaller than this length. By default, t is equal to 2.5. Whenever t is greater than 0, all formulas are adjusted to condition on the probability that IBD segment lengths are greater than or equal to t [10].
All ERSA results were with confidence level = 0.999 and α = 0.001. Reported power estimates were the percentage of related pairs that were correctly predicted to be related at α = 0.001.
To mask genomic regions potentially prone to false-positive IBD, we first evaluate the distribution of IBD segments in a control population. Genomic regions are masked from the analysis if the ratio of observed to expected total IBD segment length exceeds a specified threshold, h. By default, h is equal to 4, but we observed similar results for values of h between 2 and 6 (Figure S9). The total IBD segment length equals the sum of all pairwise IBD segments that overlap a genomic region, with the segments truncated at the region boundaries. The expected total IBD segment length is calculated under the assumption that pairwise IBD segments in the population are distributed uniformly across the genome. Let m equal the summed length of all masked genomic regions, in cM. We subtract m/100 from r in all models to account for recombination events that cannot be observed. For each IBD segment, the length of the masked region is subtracted from the length of the IBD segment if the IBD segment wholly contains the region and extends at least b base pairs past the beginning and end of the region. By default, b is equal to 1 Mb. All other IBD segments that cross a masked region are truncated at the region boundary. Genomic region masking is an optional parameter in ERSA 2.0 (mask_common_shared_regions) that is inactive by default.
Because parents and offspring are IBD1 throughout the entire genome, there is no stochasticity in the number and lengths of IBD segments. Therefore, for both versions of ERSA, parent-offspring is reported as the most likely relationship if the total IBD segment length is at least z standard deviations above the expected total segment length of a full-sibling relationship (0.75). By default, z is equal to 2.33.
Other than parent-offspring, direct ancestor-descendant relationships (e.g. grandparent-grandchild) were not explicitly modeled in ERSA 1.0. The primary difference in IBD segment distribution between an ancestor-descendant relationship and a relationship with a shared ancestor is that recombination events in the first generation cannot be detected in a pairwise comparison unless complete phase information is available. ERSA 2.0 accounts for this difference with the following equations:(4)and(5)
Because GERMLINE and fastIBD do not differentiate between IBD1 and IBD2, regions of IBD2 among full-siblings are merged with their flanking IBD1 segments and are reported as a single, larger IBD segment. Multiple IBD2 segments can be joined together in this manner. Let k equal the number IBD1 segments that have been bioinformatically merged. Conditioned on k and ignoring chromosomal boundaries, l follows a gamma distribution with shape parameter equal to k+1. ERSA 1.0 approximated the distribution of l using the maximum likelihood estimate of k with a single gamma distribution, which introduced an additional free parameter in the sibling model relative to other relationship models. To eliminate this free parameter, ERSA 2.0 assumes that l is distributed according to a mixture of gammas by summing over possible values of k. The likelihood of l is equal to:(6)
The expected number of IBD segments for full sibling relationships in both versions is:(7)
ERSA 2.0 now has the option of evaluating full-sibling models using IBD2 segment data. ERSA assumes that all overlapping IBD1 segments are merged into a single segment, and that each IBD2 segment is reported separately as an additional overlapping segment, which matches the output format that we generated from ISCA. Conditioned on the total length of IBD1 segments, T, the expected number of IBD2 segments under the null model is equal to the unconditional number of expected IBD1 segments multiplied by T/100r. The expected length of an IBD2 segment under the null model follows the empirical distribution for IBD1 segments. Under the alternative model, the number of IBD2 segments, n2, is approximately Poisson distributed with mean equal to(8)and the expected length of an IBD2 segment is approximately exponentially distributed with mean equal to 25 cM. The IBD2 option (‘use_ibd2_siblings’) was used for all ISCA analyses.
All segments that are IBD2 between siblings must be IBD1 in an avuncular relationship involving one sibling and the offspring of the other sibling. For each such segment, an additional IBD1 segment in the siblings may be inherited by the offspring with probability of 0.5 if the following two events occur: 1) a recombination event occurs within the segment in the offspring (probability of approximately 0.5) and 2) an IBD1 segment does not flank a second IBD2 segment (probability of 0.5). All IBD1 segments in the siblings that are not part of an IBD2 segment are broken into two segments by recombination in the offspring with probability of approximately 0.5, in which case one of the segments are shared, and are otherwise are inherited in the offspring with probability 0.5. This leads to the following expression:(9)
Thus, expected number of segments in an avuncular relationship is equal to the expected number in a full-sibling relationship (Eq. 7). This correction was implemented in ERSA 2.0 (ERSA 1.0 erroneously applied Eq 3 to avuncular models). Both versions approximate the distribution of IBD segment lengths in an avuncular relationship using Eq. 1 and assuming an exponential distribution.
We used RELPAIR to estimate relationships for all the pairs of individuals to which we applied ERSA 2.0. From the whole-genome data sets, we extracted 9999 well-spaced, relatively independent biallelic SNP loci with minor allele frequency >20%. Allele frequencies and linkage disequilibrium for all loci were assessed in the 34 unrelated CEU individuals using PLINK [27]. Linkage disequilibrium between loci was minimized by pruning correlated SNP loci (PLINK –indep, variance inflation factor up to 1.5 allowed, analysis conducted in windows of 400 SNPs, step size 10 SNPs.) Remaining closely-spaced SNP loci were removed until the target of 9999 SNPs was reached.
The relationship identification methods described above are implemented in the software package ERSA 2.0, which is freely available for academic use (www.hufflab.org). The software for ISCA is available at http://familygenomics.systemsbiology.net/software. The pedigree simulation programs are available in http://caballero.github.io/FakeFamily/.
|
10.1371/journal.pntd.0002551 | Venezuelan Equine Encephalitis Viruses (VEEV) in Argentina: Serological Evidence of Human Infection | Venezuelan equine encephalitis viruses (VEEV) are responsible for human diseases in the Americas, producing severe or mild illness with symptoms indistinguishable from dengue and other arboviral diseases. For this reason, many cases remain without certain diagnosis. Seroprevalence studies for VEEV subtypes IAB, ID, IF (Mosso das Pedras virus; MDPV), IV (Pixuna virus; PIXV) and VI (Rio Negro virus; RNV) were conducted in persons from Northern provinces of Argentina: Salta, Chaco and Corrientes, using plaque reduction neutralization test (PRNT). RNV was detected in all studied provinces. Chaco presented the highest prevalence of this virus (14.1%). Antibodies against VEEV IAB and -for the first time- against MDPV and PIXV were also detected in Chaco province. In Corrientes, seroprevalence against RNV was 1.3% in the pediatric population, indicating recent infections. In Salta, this was the first investigation of VEEV members, and antibodies against RNV and PIXV were detected. These results provide evidence of circulation of many VEE viruses in Northern Argentina, showing that surveillance of these infectious agents should be intensified.
| Venezuelan equine encephalitis viruses (VEEV) are responsible for human diseases in the Americas. They produce severe or mild illnesses with symptoms indistinguishable from dengue and other arboviral diseases; for this reason, many cases remain undiagnosed. We detected neutralizing antibodies (NTAbs) against VEEV IAB, VEEV ID, MDPV (VEEV subtype IF), PIXV (VEEV subtype IV) and RNV (VEEV subtype VI) in human serum samples of Northern provinces of Argentina. Chaco province showed presence of NTAbs against VEEV IAB, MDPV, PIXV and RNV. In Corrientes province, we detected NTAbs against RNV in a pediatric population. NTAbs against PIXV and RNV were also detected in Salta province. These findings demonstrated the circulation of many VEEV strains in Northern Argentina and underscore the need for surveillance of dengue like illness in this region.
| Venezuelan equine encephalitis (VEE) is a reemerging mosquito-borne viral disease that is severely debilitating and sometimes fatal to humans [1]. The etiological agent, VEE virus (VEEV), belongs to the VEE complex (Togaviridae: Alphavirus), one of the major alphavirus serogroups found in the New World [2]. Members of the VEE complex are distributed throughout America and have been originally classified in subtypes based on their serology; however, they are now considered different virus species. Only subtypes IAB and IC are considered epidemic/epizootic varieties since they have been responsible for outbreaks involving equine and human cases [3]. These subtypes undergo an amplification cycle that involves equids, which develop high titer viremia, and mosquitoes [1]. Enzootic strains (subtype I varieties ID, IE, IF and subtypes II to VI) are not associated with equine disease, producing low titer viremia, with the exception of VEEV IE. Interestingly, strains in this subtype have been responsible for epizootics in Mexico and appear to be equine neurovirulent, but are not known to produce high titer viremia in equids [4]. Enzootic strains circulate in forested or swamp habitats, where rodents serve as reservoir hosts and Culex mosquitoes -mainly in the subgenus Melanoconion- act as vectors [5]. However, these viruses have also been detected in urban areas [6], [7], [8]. Human infection by any of these strains can be completely asymptomatic or present with a mild disease, with symptoms similar to dengue or influenza, although a fatal human case caused by enzootic VEEV ID was reported in Panama in 1961 [1]. Enzootic VEEV are increasingly recognized as important endemic pathogens of people who live near the enzootic transmission foci and/or enter the habitats where enzootic circulation occurs [1]. Some of these enzootic viruses are postulated to be progenitors of epizootic strains [4].
In Argentina, the circulation of Rio Negro Virus (VEEV subtype VI; RNV) is well known; it was isolated for the first time in 1980 by Mitchell et al. from mosquitoes of Chaco province [9]. In 1989, Contigiani et al. reported an outbreak of acute febrile illness in humans from General Belgrano Island (Formosa province) associated to RNV, with symptoms indistinguishable from dengue [10]. Subsequent serological studies carried out in the same area showed the presence of human antibodies not only against RNV, but also against subtype IAB (TC83 vaccine strain) [11]. Recent investigations have reported the molecular detection of RNV and Pixuna Virus (VEEV subtype IV; PIXV) in Chaco and Tucumán provinces [7], [8], demonstrating that more than one VEEV is currently active in Argentina. RNV has also been detected in Córdoba Province [12].
Because epidemics of arboviruses often receive notice only when they are acute and massive, the public loses sight of ongoing transmission, which has a significant daily impact on the life of people living in endemic countries. These diseases are often ignored and neglected because they have not yet impacted the lives of those living in affluent areas. They are understudied and go unnoticed until outbreaks occur [13]. The epidemiological characteristics and geographic range for many endemic arboviruses in South America are poorly understood [14]. This is the case for endemic VEE, which is underreported in many parts of the continent, where enzootic circulation occurs and surveillance of febrile illness is limited, such as in Argentina. To begin to address this gap, we proposed to determine the occurrence of VEEV infection in humans of the North part of the country -where circulation of RNV and PIXV is well known-, and investigated the presence of neutralizing antibodies (NTAbs) against VEEV IAB, VEEV ID, Mosso das Pedras virus (VEEV subtype IF; MDPV), PIXV and RNV in human sera obtained during the period 2006–2011.
This study was designed as a non-associated, anonymous survey: data registered were only number of sample, date of sampling, age of the patient (years), gender and address (street and neighborhood). It was approved by the ethics committee of the Faculty of Medicine, National University of Northeast (UNNE) and conducted within the project N°FBBI11/10.
All the studied locations are indicated in Figure 1. The sera were obtained from people without symptoms who attended health centers to perform routine or other analysis within the project “Ecoepidemiology of arboviruses in Argentina”.
Sera were inactivated at 56°C during 25 minutes, then centrifuged at 11.400 g for 30 minutes to clarify; the supernatant was stored at −20°C until assayed. Samples were analyzed for NTAb's against VEEV subtypes IAB, ID, IF, IV and VI by PRNT using Vero cells as described by Early et al. (1967) [18]. Serum samples were initially tested at a dilution of 1∶10. Those that neutralized at least 80% of inoculated viral plaque forming units (pfu) were considered positive, and in order to determine the end-point titer, they were further titrated with the same technique, using 2-fold serial dilutions.
Viruses used in this study were: a) VEEV IAB strain TC83 [19], b) VEEV ID strain 3880 -isolated for the first time in 1961 in Panama- [2], c) MDPV (VEEV subtype IF) strain 78V3531 -first isolated in 1978 in Brazil- [2], d) PIXV (VEEV subtype IV) strain BeAr35645 [20], and e) RNV (VEEV subtype VI) strain AG80-663 [9]. Viral suspensions were prepared with a 10% dilution of infected suckling mice brain in Minimum Essential Medium (MEM) 10% fetal bovine serum (FBS) and 1% antibiotics (gentamicin), centrifuged at 11.400 g for 30 minutes.
The 149 specimens analyzed were obtained from patients aged 0–20 years old. They were tested against VEEV subtype IAB and RNV. Two of them presented NTAbs against both RNV and TC83 strain, and only 1 tested positive against RNV (Table 4). Sample 313 corresponded to a resident of Chaco Province, for this reason, it was excluded of the analysis. The other 2 positive samples belonged to patients who lived in Corrientes city. Seroprevalence of RNV in pediatric population was 1.3%.
The 197 serum samples from Orán (Salta) were tested against VEEV ID, PIXV and RNV. The last two viruses were included because of previous reports about their molecular detection in other Northern regions. Subtype ID was included due to its recent detection in Bolivia (bordering country in Salta province) (Figure 1), where it has been associated with human disease with symptoms similar to dengue [6].
Five samples tested positive (age range: 21–67 years, Table 5). One specimen presented NTAbs only against RNV (sample 422); the others presented NTAbs against both RNV and PIXV, with titers varying in favor of one or the other virus. Sample 419 showed titer of NTAbs against PIXV 4 times greater, determining that this was the infecting virus; sample 619 presented a greater difference in favor of RNV (16 times). Titers obtained in samples 234 and 413 showed no significant differences (Table 5). Neutralizing antibodies against VEEV ID were not detected.
Over the past few decades, a global resurgence of arboviruses has been detected worldwide [16]. Despite the public health relevance, the geographic range, relative impact and epidemiologic characteristics associated with arbovirus infection are poorly described in many regions of America, including Argentina. Surveillance plays an important role in the prevention of these diseases -particularly VEE-, knowing circulating viral species, susceptible population and epidemiology. Within this surveillance, studies of serological prevalence have served for a long time -and are still useful- to indicate immune status of a given population against an infectious agent, providing information of its circulation in a region. For arboviruses, PRNT is the gold standard technique due to its high sensitivity and specificity. In Argentina, there is serological evidence of the presence of VEEV since 1950's decade in humans and rodents [24], [21], [11]. Our studies detected current VEEV activity in the North part of the country.
In Chaco Province prevalence for RNV was high. Detections of this virus in individuals aged 3–5 years provide evidence of its recent circulation. Titers of NTAbs obtained in the period 2007 were, in many cases, very high (>640 and 1280), agreeing with secondary infections. In Pampa del Indio -where there are no prior registers of VEEV circulation- titers obtained against RNV showed profiles compatible with primary infections (only one sample presented titer >640). This could suggest that Pampa del Indio is an area of more recent circulation of RNV, compared to other places of Chaco province, like Resistencia city.
This is the first search of antibodies against PIXV in our country, with positive results in samples from Chaco in both studied periods. Some of these positive samples could be consequence of a serological cross reaction with RNV, while others could represent true positive results. In 2007, despite the fact that the titers obtained against RNV were four-fold higher than those obtained against PIXV and VEEV IAB, we cannot discard that, as it has been documented for some flaviviruses [22], the observed heterotypic immunological response could be the result of sequential infections by PIXV and/or VEEV IAB, especially in those samples in which the titers against PIXV were of 80 or more. For this reason, some of these individuals may have suffered infection by two or more viruses. Samples with similar titers of NTAbs against PIXV and another virus may correspond to double or triple infections, positioning Chaco province as a hyper endemic area. Pixuna virus infections are supported by previously reported molecular detections of this virus in the same region [7]. Detection of NTAbs against PIXV in Pampa del Indio represents not only another evidence of its presence in Chaco province (and Argentina), but also the wide distribution of this virus in our country.
We can assume that VEEV IAB positive specimens might be the result of antigenic cross reactivity with PIXV (in many cases titers against PIXV were 4 times greater than against VEEV IAB); however, results obtained with sample 432 in 2007 (Table 1) and sample 66 in 2011 (Table 3) could indicate activity of subtype I, since it was demonstrated that VEEV IAB does not present serological cross reactivity with RNV. This should be considered cautiously, since there is no evidence of epidemic/epizootic viral activity in our region, but the circulation of an enzootic variety of other subtype I might exist. This hypothesis was first proposed by Monath in 1985, and is supported by previous detections of antibodies against VEEV IAB in animals during the 80's [21] and in humans in 1991 [11], although no isolations of members of this subtype have been reported up to date. This leaves an open door to suspect that these antibodies previously detected could correspond to antibodies against PIXV captured by TC83 strain in the PRNT.
Many of the samples analyzed in Pampa del Indio had a positive result to MDPV. Some of them could be due to serological cross reactions with VEEV IAB and PIXV, since MDPV captures NTAbs generated against these viruses [23]. Other specimens presented titers against MDPV that were equal or higher than other viruses. These cases could represent evidence of MDPV circulation in Pampa del Indio and for the first time, in Argentina. This should be considered cautiously because, as we previously mentioned, there are no molecular detections of this or other enzootic subtype I in this country. Further studies with emphasis on the search of MDPV (or other enzootic VEEV subtype I) are needed, both by molecular detection as well as by surveillance of undifferentiated febrile human cases.
In Corrientes, the prevalence of 1.3% for RNV in a pediatric population indicates recent circulation of the virus. The low value may show less viral activity in this city compared to Chaco (with a prevalence of 14.1% in 2009) or could be due to the fact that the detections were made in a pediatric population; this percentage may be higher in the adult or general population. The high titer of NTAbs against RNV observed in sample 326 (≥640), could indicate current infection. Our results demonstrate that VEEV strains have a silent and endemic circulation in this area and highlight the need of constant surveillance.
This is the first investigation of VEEV in Salta Province. Four out of the 5 positive results exhibited serological profiles compatible with a heterotypic serological response. In two of them RNV appears to be the infecting virus, and in one PIXV. In the other two samples it was impossible to determine the etiologic agent. All positive samples were obtained from adult patients, and in consequence, it was not possible to define whether they were recent infections or not. These are the first detections of RNV and PIXV in Salta Province, showing the wide distribution of these viruses in Northern Argentina. In Orán city, some outbreaks of other arboviruses such as DENV, which share symptoms with RNV, have occurred. Previous reports in other countries of America have shown a sub-estimation of VEEV cases in regions with co-circulation with DENV or other arboviruses [4]. As this could be the case of Orán city, it is relevant to acknowledge the circulation of VEEVs in this area and in all the province, emphasizing the investigation of acute febrile cases reported as probable dengue. The fact that all the samples tested negative against VEEV ID indicates no presence of this virus in our country so far. However, since intense commercial activity is developed between Argentina and Bolivia in this area, the introduction of VEEV ID in our territory should not be discarded; therefore, surveillance should be active in this region with the aim to detect cases early.
These serological findings lead us to postulate the hypothesis that RNV is expanding into new regions, probably due to climate changes -since the climate may influence the ecology of microbial systems [25] -as well as to an increase in the commercial activities of the area. They also constitute an evidence of enzootic VEEV circulation in Northern regions of Argentina, although the role of these viruses in the production of human diseases and their impact on public health is still unknown. While detections of VEEV NTAbs in our study all belong to enzootic types, genetic studies have demonstrated that enzootic and epizootic subtypes are closely related, and a modest number of nucleotide changes can alter the viral phenotype dramatically, converting an enzootic strain to an epizootic strain [14], [26]. For this reason, it is important to perform complementary molecular studies in order to provide information about the variability of local VEEV strains.
This report is an approach to recognize which VEEV strains circulate in Northern areas of Argentina. In light of the lack of a distinctive clinical presentation and the diversity of the etiologic agents circulating in the studied area, more investigations that focus on arboviral transmission patterns, phylogenetic relationships between the strains and occurrence of clinical cases produced by RNV and a potential VEEV subtype I virus, are needed to achieve a better understanding of the impact of these viruses on human health.
|
10.1371/journal.pcbi.1000923 | Correlated Mutations: A Hallmark of Phenotypic Amino Acid Substitutions | Point mutations resulting in the substitution of a single amino acid can cause severe functional consequences, but can also be completely harmless. Understanding what determines the phenotypical impact is important both for planning targeted mutation experiments in the laboratory and for analyzing naturally occurring mutations found in patients. Common wisdom suggests using the extent of evolutionary conservation of a residue or a sequence motif as an indicator of its functional importance and thus vulnerability in case of mutation. In this work, we put forward the hypothesis that in addition to conservation, co-evolution of residues in a protein influences the likelihood of a residue to be functionally important and thus associated with disease. While the basic idea of a relation between co-evolution and functional sites has been explored before, we have conducted the first systematic and comprehensive analysis of point mutations causing disease in humans with respect to correlated mutations. We included 14,211 distinct positions with known disease-causing point mutations in 1,153 human proteins in our analysis. Our data show that (1) correlated positions are significantly more likely to be disease-associated than expected by chance, and that (2) this signal cannot be explained by conservation patterns of individual sequence positions. Although correlated residues have primarily been used to predict contact sites, our data are in agreement with previous observations that (3) many such correlations do not relate to physical contacts between amino acid residues. Access to our analysis results are provided at http://webclu.bio.wzw.tum.de/~pagel/supplements/correlated-positions/.
| Point mutations (i.e., changes of a single sequence element) can have a severe impact on protein function. Many diseases are caused by such minute defects. On the other hand, the majority of such mutations does not lead to noticeable effects. Although previous research has revealed important aspects that influence or predict the chance of a mutation to cause disease, much remains to be learned before we fully understand this complex problem. In our work, we use the observation that sometimes certain positions in a protein mutate in an apparently correlated fashion and analyze this correlation with respect to mutation vulnerability. Our results show that positions exhibiting evolutionary correlation are significantly more likely to be vulnerable to mutation than average positions. On one hand, our data further support the concept of correlated positions to not only be associated with protein contacts but also functional sites and/or disease positions (as introduced by others). On the other hand, this could be useful to further improve the understanding and prediction of the consequences of mutations. Our work is the first to attempt a large-scale quantitation of this relationship.
| Most of the missense mutations do not lead to an appreciable phenotype when they occur in nature or are introduced experimentally. There are, however, numerous counterexamples where even a subtle change of the primary protein sequence results in severe phenotypical effects – i.e. genetic disease. Understanding the underlying mechanisms which determine the link between genotype and phenotype is the key issue in developing strategies for diagnosis and treatment of hereditary diseases.
Databases such as OMIM (Online Mendelian Inheritance in Man) or HGMD (Human Gene Mutation Database) provide a wealth of information [1], [2] about phenotypes associated with thousands of known human mutations. These databases assist researchers in analyzing the molecular basis of human disease. With the current quest for the “1000 Dollar Genome”, there is no doubt that entire patient genomes will be available in the near future which will substantially accelerate the discovery of new mutations of unknown significance. In such a situation the question “What does this mutation mean for a patient's health?” will become more and more practical for the affected individuals and physicians.
What rules determine the spectrum of allowed mutations, and why do mutations cause disease in some genes while other genes appear to be more tolerant to substitutions? Much effort has been invested in answering such questions and promoting our understanding of the underlying mechanisms which rule the complex network of factors contributing to human disease. In particular, we would like to understand what properties or features are shared by disease-associated genes.
In addition to numerous careful experiments on individual genes and proteins, with the advent of high-throughput technologies such as genomics, proteomics and, more recently, metabolomics large bodies of experimental data have been analyzed towards this end. It has been shown that genes and proteins, which are known to be involved in a large variety of diseases and syndromes, differ from genes without such association in many aspects. Disease genes have a broader phylogenetic distribution, tend to be longer on average, and more of them have homologs in other mammals compared to the average human gene [3]. Disease-related proteins have been found to be better conserved and their synonymous substitution rates are significantly higher than expected [4], [5]. Further contributions demonstrated that disease proteins have less designable folds, tend to have isoelectric points closer to neutrality, contain more alternating hydrophilic/hydrophobic stretches compared to the average human protein and have a higher tendency to aggregate [6]. “Disease genes” are highly expressed in a small number of tissues, and their encoded proteins are more likely to be secreted and mutated in genetic diseases with Mendelian inheritance [7]. Finally, genes associated with inherited disease mutations are less likely to be essential and display an intermediate level of connectivity on protein interaction networks [8].
Knowing what genes and proteins are involved in disease is only one part of the challenge. Clearly, not every site of a protein is equally vulnerable when hit by mutations. While some parts of the molecule will remain functional even after substantial changes of the primary sequence, other positions cannot be changed at all without serious consequences. Evolutionary conserved and functionally important residues, such as those in active centers of enzymes, as well as residues important for preservation of the protein's overall stability, in particular those located in buried positions, have been shown to be frequent targets of disease-associated mutations [9], [10], and multiple prediction techniques based on both structural and sequence features of proteins have been suggested to distinguish benign mutations from those implicated in inherited disease. Notable tools to combine multiple lines of evidence to produce more reliable prediction include SIFT and PolyPhen [11], [12] (see [13] for an excellent review).
The phenomenon of correlated mutational behavior between columns of a multiple sequence alignment has been described for many years for both DNA/RNA and protein sequences [14]–[16]. For proteins, the initially hypothesized notion of the underlying biological event was, that an unfavorable amino acid change in a structural contact site may go without negative consequences if its direct binding partner is simultaneously mutated in such a way that the original interaction is salvaged (compensatory mutation).
Accordingly, the analysis of such correlated mutations has been traditionally employed for the identification of residue contact pairs within or between different protein chains. The first approach to detect co-evolving residues in a multiple sequence alignment was proposed in 1994 [17]. Many other methods have been reported since then and evaluated with respect to their potential of predicting residue-residue contacts [18]–[24]. However, despite significant progress in method development, comparative studies have shown that prediction accuracies for structural contacts hardly exceed 20% with any of these methods [25], strongly limiting the application of the predicted contacts as structural constraints in ab initio structure prediction.
While some authors have explained the low contact prediction accuracies with the difficulty of differentiating correlation signal from random noise [26], [27], recent studies indicate that co-evolution of amino acids in fact may originate not only from structural contacts but from a much broader range of biological reasons. Using Statistical Coupling Analysis Ranganathan et al. (2005) detected correlation rules in the WW domain which describe aspects of the fold architecture going beyond simple protein contacts. They introduced the concept of a correlation backbone in the fold which they claimed was nearly sufficient to describe the architecture without additional information [28]. They impressively demonstrated the power of this idea by synthesizing artificial WW domains solely based on the previously derived correlation model and showing that a substantial percentage of these designed polypeptides were able to fold into functional WW domains in vitro [29].
In addition, further contributions have demonstrated that correlated mutations may also occur due to reasons related to protein function. Gloor et al. analyzed 12 mutations of the ATP synthase subunit and 7 mis-sense mutations of the homeodomain and came to the conclusion that certain co-evolving residues are more likely to be functional sites and thus possibly more likely to be related to disease [18]. Within a study on the Hsp70-Hop-Hsp90 system, regions previously known to be functionally important could be identified based on residue co-evolution [30]. Additionally, the authors pointed out that co-evolving amino acids were often found to be in close proximity to functionally important sites. Similar results were obtained in an analysis of correlated mutations within the cytochrome c oxidase subunit I where many co-evolving residues were found adjacent to hypothesized proton pumping channels [31]. In a recent publication Lee et al. provided further evidence for the hypothesis that correlated mutation may be related to functional importance in an analysis of 44 selected protein families [32].
All together, these results indicate that co-evolving residues may be both structurally or functionally important positions within protein folds and therefore could be likely targets for disease-associated point mutations. Here, we present the first comprehensive analysis of human disease mutations with respect to co-evolving residues using all known point mutations and proteins currently available in the Human Gene Mutation Database (HGMD). Our data confirm that correlated mutations go well beyond contact prediction and are a hallmark of amino acid positions leading to disease when affected by mutation.
For our analysis, we used all human proteins known to be affected by at least one disease causing point mutation according to HGMD annotation and for which at least 30 orthologous proteins of sufficient sequence diversity were available for building a multiple sequence alignment. 1153 proteins fulfilled all requirements and were analyzed for correlated mutations using the OMES algorithm. In addition, we repeated all analyses on a more rigorous dataset using a cutoff of proteins per ortholog cluster which left us with 855 human disease proteins. Using these two data sets, we identified 62 365 and 46 022 residues as correlated with other positions, respectively. A total of 14211 and 10508 positions were found to be disease-related in these two sets.
As stated above, our work is motivated by the observation that co-mutation of residues over the course of evolution may not to be restricted to protein contacts but rather be the result of other types of functional association among residues. Accordingly, our first goal was to test the hypothesis that point mutations affecting correlated residues are more likely to result in disease than expected by chance (i.e. compared to random positions).
Using all residues represented in the datasets described above, we produced contingency tables of correlatedness vs. known disease-mutations. Based on these tables, we computed the background rates of disease mutations to be 0.019% for random positions and 0.032% for correlated positions (0.0195 and 0.0325 for the clusters ). In other words, correlated residues were found to be times more likely to be known disease positions than expected by chance translating to a log odds value (LOD, ) of 0.73.
For the clusters , the relative increase was found to be 1.66 (). Fisher's exact test for count data confirmed that the observed difference is highly significant in both cases (; see Table 1 for summary). As the stringent data did not yield a substantial gain over the less strict set, we are reporting the results of the latter () data in the subsequent text. All results for the stringent set are reported in the Text S1.
Figure 1 shows the empirical background distribution of LOD values generated by 1000-fold permutation of correlation scores in comparison to the observed LOD. In addition, we show the bootstrap distribution of the observed LOD generated by 1000fold resampling of individual positions from all multiple sequence alignments (same as 8.1 in Text S1). Clearly, the observed value is far outside the background distribution.
In order to cross-check our findings, we also computed the LOD for a set of positions that are highly unlikely to be associated with disease because of evolutionary accepted mutations (see Materials and Methods). In this data we find an LOD of −1.26 indicating that these positions are clearly underrepresented in the set of correlated positions.
After having shown that correlated positions are in general significantly more likely to be hit by disease-causing point mutations we sought to investigate the implications of this finding for individual proteins. We repeated the above analysis for all 1153 proteins separately. As both the number of known disease-mutations and the degree of correlation varies among proteins, one would expect that for some proteins, correlation is strongly associated with disease-susceptibility while in others no such signal can be detected. In fact, analysis of individual proteins also yields an arithmetic annoyance: proteins with a very low number of known disease mutations have a very large chance that none of them is located in a correlated position simply because of the small sample size, resulting in an . These cases were excluded from the following analysis as no valid statistical analyses could be carried out. In total, an LOD score could be calculated for 524 proteins of the data set and 629 proteins obtained no score. The analysis of the proteins for which our approach failed shows that 50% (315) of these proteins have only one known disease mutation in HGMD and for only 3.8% (24) more than 9 disease-related substitutions were available.
The LOD distribution for individual proteins is depicted in Figure 2 (Alignment threshold ; see Text S1 for threshold ). Only a small fraction of proteins (10% in Figure 2a) in our data set had LOD values and a clear majority of proteins had at least slightly positive LODs. In some cases, we observed LOD scores which represents an increase of 1500% over expectation. Taken together, these numbers indicate that, except for cases with very few known disease mutations, the global result applies to the majority of individual proteins.
Detection of correlated residue pairs is not entirely independent of the degree of conservation of the respective positions. Depending on the algorithm used, substantial crosstalk between conservation and correlation can be observed [25]. Although the OMES method is reputed to be among the more robust approaches with respect to interactions with conservation, we investigated the degree to which sequence conservation affects our results. This is especially important as evolutionary conservation of a sequence region is generally taken as an indication of functional importance and thus would represent a bias in favor of the hypothesis under test.
We calculated two different measures of conservation for each position in the multiple sequence alignments. The first method () computes fractional identity to the human (reference) residues in each column of the MSA. The second procedure () computes a conservation score based on the BLOSUM62 amino acid substitution-matrix [33] for each column.
To evaluate the interaction between correlation and conservation we applied stepwise filtering on conservation. Starting from the full dataset, we decremented the conservation threshold in small steps thus removing more and more of the top conserved columns from the MSA. The background frequency as well as the observed frequency of all positions was re-computed for all of these filtered datasets and the corresponding LOD values were computed. Figure 3(A,B) shows LOD values plotted against the respective conservation cutoff . E.g. indicates that for the calculation of the global LOD score only residues with were taken into account.
For comparison, we also computed the LOD score of conservation with respect to disease mutations. Our initial intuition was that conservation would probably be a much more potent indicator of functional importance than correlation and thus yield substantially higher LOD values for being affected by disease mutations. We performed this analysis at different conservation cutoffs to get an impression of the degree of conservation required to detect functional importance.
As expected, we found well conserved positions to be clearly enriched in disease mutations, and we observed a correlation between the degree of conservation and the LOD score (Figure 3C,D). Depending on the conservation measure and cutoff, well conserved positions were times as likely to be affected by known disease mutations as random positions (). As expected, conservation clearly outperforms correlation. Given the obvious link between conservation and functional importance, the numbers for correlated positions are surprisingly high. Furthermore, the LOD values for correlation remain remarkably stable over a fairly wide range of conservation thresholds indicating that the correlation signal is not merely an artifact caused by relatively well conserved positions which happen to also correlate. Taken together, these results suggest that evolutionary conservation is a useful measure for the assessment of disease-susceptibility and thus functional significance of amino acid positions in a protein.
Other groups have previously demonstrated that correlated positions without physical contact do occur in protein structures [28], [34] and, in a recent study, Noivirt-Brik et al. have demonstrated the emergence of long-range interactions in lattice models of proteins [35].
In the dataset analyzed in our own work, we found that only 2714 out of 16555 (16.4%) of correlated pairs had a distance of less than 5.5Å which would imply physical contact. If the distance threshold is relaxed to a generous 8.0Å, still only 17.4% are in proximity. Thus, even applying a very permissive threshold, the majority of correlations is observed between residues which are not in direct contact – an observation compatible with the hypothesis of functional correlations.
Of course, many positions correlate with more than one other residue and accordingly, some of these correlations coincide with contacts while others do not. In our data, 29.6% of all correlated positions had at least one contact correlation. For the positions which were found to be both correlated and relevant for disease 31% had at least one contact Å.
Finally, we wanted to determine if non-contact correlations are enriched in disease mutations. Out of 4960 correlated residues without a single contact, 252 (5.0%) were also disease positions. This corresponds to 1.3 fold increase over the expectation (LOD = 0.39). Due to the small sample size the latter is not statistically significant. Nevertheless, the trend is encouraging and points in the direction of the hypothesis of functionally relevant non-contact correlations. Table 2 summarizes the results for correlation, conservation and residue contacts.
The most plausible explanation for a connection between correlated mutations and disease mutations is that correlated mutations indicate functional relevance of the respective residues. So on one hand, many of the disease positions are probably functionally active themselves or correlate with functionally important positions in the protein. That would imply that we should find significant enrichment of active sites in correlated position, too.
We used the SwissProt feature annotation to test this hypothesis and found that 3.7% of functional sites are annotated with at least one disease associated point-mutation in HGMD. That means that functional sites are roughly twice as likely to host disease mutations than expected. We carried out an analysis of enrichment in correlated positions with the functional site data and found it even stronger than for the disease mutation data (LOD = 1.04, ). When active sites and HGMD mutations are combined into a single set the resulting enrichment lies between the results for disease and functional sites (LOD = 0.88, ). So both types of information show the same trend to lie in correlated positions.
Next, we asked the question if a position would be more likely to be involved in disease if it was correlated with another disease site or a known functional residue. We found that residues which are correlated with a disease position or a functional site are 5.2 and 2.9 times more likely to be disease positions themselves than randomly chosen residues, respectively. These numbers indicate that not only is correlated mutation itself an indicator of disease-relevant sites but apparently functional/disease positions seem to be preferentially correlated with each other. As for the disease positions analyzed above, we found that only a modest fraction (18%; 5.5 Å threshold) of functional positions which show significant correlation are involved in physical contact and thus the majority are non-contact correlations.
So far, correlation was the only variable taken into account in our analysis, but other structural features may also contribute to the potential of a correlated residue to be associated with disease. In order to test if basic structural features of a position affect our results, we investigated the accessibility of a residue in this context. This analysis was carried out in the subset of proteins with known structure which is much smaller than the full data set. The global LOD for disease in correlated positions in this smaller subset is 0.39. Interestingly, correlated residues that were exposed showed an LOD of 0.73 while partially and fully buried correlated residues had LOD of only 0.42 and 0.1, respectively.
Another obvious structural condition is the local secondary structure. Our analysis revealed that correlated positions are most likely to be associated with disease when they are embedded in an -helix (LOD = 0.47) followed by turns (LOD = 0.37) and much weaker in in -sheets (LOD = 0.09). This ranking is in part connected to the different accessibility found in these secondary structures: helices and turns are known to exhibit a much larger fraction of exposed residues than beta sheets and our sample is in accordance with this. Furthermore, in our data, the probability of a position being associated with disease is significantly negatively correlated with accessibility and the same is true for the probability of showing co-evolution with another residue. In a logistic regression model of the disease probability vs. accessibility and secondary structure, the secondary structure just barely achieved significance () while accessibility was highly significant (). For the probability of a residue to co-evolve with another one, secondary structure was not a significant predictor () while accessibility was highly significant again ().
As we have seen above, the enrichment of correlated disease residues is more pronounced in exposed sites. This observation appears plausible because much of the functional features of a protein are located on its surface: regulatory modifications, interactions with other proteins and binding sites for substrates need to be accessible. As we also find co-evolving residues to favor accessible locations, the above numbers come as no surprise.
Phenotypic consequences of mutations are determined by many different factors like functional role of the protein, specific amino acids involved, structural properties etc. Accordingly, no single feature can be expected to be a strong predictor on its own and correlation is no exception. Tools like SIFT [36] or PolyPhen [12] combine different signals into an integrated prediction of phenotypic effects of mutations. Currently, correlated mutation is not among the features analyzed by these programs. Based on our results presented above, we would expect that adding the correlation signal to these tools would further improve their predictions. A full assessment of this expectation would require access to the source code of these tools, which we did not have. On the other hand, we can analyze the overlap between predictions by integrated tools and the correlation data in order to estimate if the correlation data is fully included in these predictions or not. We compared predictions by SIFT and PolyPhen to the correlation signal in a set of 813 proteins featuring 8838 known disease mutations and 5730 substitutions unlikely to cause disease (see Materials and Methods). While the majority of the correlated positions in the disease mutation set were predicted to be damaging by either or both of the prediction programs, roughly 20% of the correlated positions were not. These numbers suggest that the correlation data is at least partially complementary to the results produced by both tools and thus has the potential to improve predictions (Figure 4). Furthermore, we carried out a multiple logistic regression, modeling the probability of disease association by the predictions by SIFT, PolyPhen and correlation and found all three terms to make highly significant contributions to the model ().
In the previous sections we analyzed the dependence between co-evolving positions and known disease mutations. We were able to demonstrate that correlated mutations are associated with disease significantly more frequently than expected and found similar results for sites known to be functionally active. In the following case studies we present two examples of proteins with long-distance correlations and discuss the functional aspects of the residues involved.
The human protein PEPD (UniProt: PEPD_HUMAN) is a proline dipeptidase which plays an important role in the recycling of proline during the final stages of degradation of collagen and dietary proteins. The enzyme hydrolyzes dipeptides with a prolyl or hydroxyprolyl residue in the C-terminal position. For catalytic activity, binding of 2 manganese ions per subunit is required as a co-factor [37], [38]. Swissprot annotation marks residues D276, D287, H370, E412 and E452 as responsible for manganese binding [39], [40].
Mutations of the PEPD protein have been identified as the cause of autosomal recessive prolidase deficiency (PD). E.g. Ledoux et al. have characterized several disease-causing point mutations in the PEPD protein [41]. Their data show the different extent of enzyme inhibition by these mutations. For instance, the R184Q mutations resulted in a residual activity of 7.4% compared to the wild type enzyme. The G278D and G448R mutations caused complete abrogation of peptidase activity. Other sources have found residues D276, S202 and E412 to cause the same disorder when hit by point mutations [42]–[44].
The phenotypic consequences of mutations involving positions 276 and 412 are easily explained by the fact that these are directly involved in metal binding. Positions 278 and 448 are in close proximity to these functional sites so it does not come as a surprise that they are critical.
R184 and S202 on the other hand, are situated far away from the metal binding sites in the primary sequence. While S202 gets close to the metal binding region in the three dimensional structure, R184 is located quite distant from this area. So why do point mutations at these sites cause disease? Of course, one possibility is that the enzyme function is destroyed by mechanisms totally unrelated to metal binding, but constructing a link to the important functional sites is another. We find that position 184 shows a strong co-evolution connection to positions 453 and 277 which are both in direct proximity of the metal binding residues (Figure 5a). Position 202 also shows a correlation with positions 277 and 184. We thus suggest that both R184Q and mutations of S202 do in fact inhibit manganese binding mediated by a non-contact interaction between these residues.
In summary, all known PD causing point mutations are either in close proximity to the critical residues or a correlated mutation link to such residues can be found. Figure 5b shows the spacial relations of the metal binding residues and some of the correlations found in this region.
Upon close inspection of the co-evolution connections depicted in Figure 5a, it is easy to see that the correlation links are not distributed evenly across the protein. Some residues or regions are connected to others by multiple arcs. Also, many positions are not only connected to each other but also share common neighbors hinting at a network of correlated positions. This observation seems to hold for the entire group of metal binding sites and disease positions discussed above: all of these positions seem to be part of a small correlation network.
Our second example is AK1, the human adenylate kinase isoenzyme 1. AK1 catalyzes the reversible transfer of the terminal phosphate group between ATP and AMP that is essential for cell maintenance and growth.
Point mutations of AK1 cause hemolytic anemia due to adenylate kinase deficiency (OMIM: 612631). AK1 catalyzes the reaction of ATP and AMP to two ADPs that is done by two nucleotide bindings regions. According to SwissProt annotation, the ATP binding site is located at residue 15–23 and the AMP binding site comprises residues 39 and 94–101.
Several different positions of the protein have been found to result in an altered phenotype upon mutation. Residue G40 is in direct neighbors of the nucleotide binding T39 while Y164 is somewhat further away. Some other positions (G64 and R128) are located in the neighborhood of these active sites in the 3D structure (Figure 6).
We find that residues implicated in disease phenotypes of AK1 are located at or near correlated positions which link to or close to the nucleotide bindings sites. As for the PEPD example, there appears to be a network of correlated positions. Residue G64 is directly linked to residue G40 that is a disease position itself and is immediately adjacent to one of the binding site T39. In addition, residue Y164 is directly linked to the disease residue G40. We found correlated links between both nucleotide binding regions and a direct link of G64 and G40 with residue G22 that is located in the ATP binding site. Obviously, disease relevant mutations can be “explained” by co-mutation links to binding sites or they are not involved in correlations themselves but are located in their close neighborhood (e.g. R128).
Both case studies illustrate that correlation networks may connect residues located elsewhere on the protein structure to a given disease associated mutation site. Disease affected positions and their immediate neighborhood tend to be connected in the network with other functionally important residues or regions such as metal binding sites or binding sites. In some other cases, disease-affected residues are located next to correlated residues or in well-connected regions of the network. These linkages could provide hints on the effects of disease-associated mutations as the corresponding networks could be used to transport the effects of point-mutations to functionally important regions. Accordingly, correlation networks provide a novel basis for selecting promising target residues for mutation studies or estimating the potential effects of yet uncharacterized naturally occurring mutations.
Above we mentioned small networks of correlated positions. From a graph perspective, positions with a higher degree (number of edges) should be more important than those with a low degree. In order to test if this also applies to our correlation network with respect to disease we analyzed the correlation graphs and found a clear association between the degree of a position and its probability to be associated with disease (Figure 7). Positions that are correlated with increasing numbers of other positions are increasingly more likely to be associated with disease (, ).
In the past, analysis of residue co-evolution in proteins has been applied to various problems mainly centered around the idea of compensatory amino acid substitutions on protein contact surfaces. Correlations not readily explained by contacts have been discussed in the field and were simply labeled “noise” by many groups [27], [45], [46]. New ideas in this field have been introduced by many researchers and concepts such as a correlation backbone as an element of protein structure or mapping functional sites to correlation hotspots have been explored and illustrated by various examples. First discussions of disease relevance have only recently entered the literature and were restricted to selected proteins or protein families [18], [32]. To our knowledge, the data presented above represents the first attempt at a comprehensive analysis of evolutionary residue co-mutation in the light of disease associated point mutations. Our data indicate that residues highly correlated with others are indeed more likely to be associated with disease than expected. Surprisingly, as little as 30 orthologous sequences sufficed to detect a significant difference and considering only orthologous groups with at least 125 proteins did not yield a substantial difference.
Of course, a single parameter such as correlated evolution cannot be expected to yield a high positive predictive value when used in isolation but our findings clearly show that it is one property to look for when judging the functional significance and/or potential to cause a disease phenotype upon mutation. A fair assessment of the value of such a measure is probably the direct comparison to the most popular approach of using sequence conservation as an indicator of functional importance. As our analysis indicates, correlation seems to be associated with disease less than conservation but appears to be on the same order of magnitude.
This has clear practical implications. Some diseases are caused by large numbers of different point mutations in seemingly random locations in the protein. Analysis of co-evolution could serve as an interesting tool to explain some of these cases. Given the rapidly decreasing cost of sequencing, the hunt for SNPs and the trend towards personalized medicine, more and more data on variations of unknown physiological consequences will be gathered. Analysis of correlated mutations could prove to be a useful complement to data on conservation, structure and other protein features in the attempt to understand functional relevance of mutations.
We found that the correlated mutation detection was largely independent of conservation signal over a wide range and that the majority of correlations did not coincide with contacts in the subset of proteins for which protein structure data was available. Accordingly, we believe that the data indicate a genuine signal of co-evolution among functionally linked positions which are vulnerable to mutations. These findings are in line with the work on structural determination of protein domains based on the co-mutation signal [28] and provide good evidence for the concept of long-distance associations within proteins.
Another interesting observation which we have not yet analyzed systematically, is the role of “near-miss correlations”, i.e. correlations between residues in the direct neighborhood of functionally essential sites and/or known disease-associated positions. This concept is similar to our observation that sometimes point mutations causing disease are located next to critical residues without actually destroying them. In some cases one may argue that hitting the critical residue itself would be too drastic a defect to be viable, in others a simple mechanistic point of view arguing that local changes are likely to influence their direct neighborhood may suffice as an explanation.
Many conceptually different algorithms have been developed to detect residue co-mutation. Fodor et al. have shown that these algorithms have a preferred level of conservation to extract significantly correlated pairs [25]. In our work, we used the OMES algorithm, as it was found to be among the best performers and most robust methods for contact prediction. But that does not necessarily imply that OMES is the single best choice for the analysis of non-contact correlation or disease mutations under each and every condition. We have preliminary data which shows that other methods produce comparable LOD scores for the global analysis, but differ in their performance depending on the degree of conservation. In future work, we are planning to study this in more detail and take advantage of differential preferences.
Correlated behavior of contact residues is a plausible and well accepted concept but what mechanisms produce long-range correlations? Lapedes et al. have suggested subsequent pairs of contact residues which extend the co-variation along a chain of contact-correlations [47]. This model is in good agreement with the concept of networks of correlated positions like the ones seen in our case studies and other preliminary data.
Another possible explanation for long-range correlations is the impact on protein folds. It is conceivable that amino acid substitutions may change the orientation of neighboring secondary structure elements by a few degrees. Such a change could easily impair protein function. It is not too difficult to imagine another change on the other side of such our helix to compensate this structural change and thus keep the rest of the fold largely unaffected.
Russ et al. introduced the concept of correlated sites as an architectural backbone of protein domains [29]. Many of the correlations used in their model related to positions which are clearly not in contact with each other and thus forming a correlation backbone. This work suggest that co-evolving residues within a protein contain more information than the mere potential to be in physical contact.
Methods like OMES that solely operate on multiple sequence alignments in analyzing correlations have previously been criticized for being “tree agnostic” – i.e. ignoring the underlying phylogenetic tree in their assessment. Others have argued that no substantial gain is achieved by tree-aware methods [48]. In order to test if using the phylogenetic information substantially affects our results, we repeated all major analyses with the algorithm of Noivirt et al. [49]. Although this method yielded a slightly lower LOD than OMES (0.63 vs. 0.73), all of our conclusions could be confirmed in this re-analysis.
In summary, we believe that researchers should not only look at conservation in their judgment of functional significance of residues in the protein sequence. Correlation patterns between residues clearly provide additional evidence which should not be ignored.
Disease mutation data was obtained from the HGMD database [2], a comprehensive collection of mutations underlying human inherited disease. Access to the full HGMD release 7.3 was licensed from Biobase, Germany (http://www.biobase-international.com).
Out of a total of 73 411 mutation entries 41628 referred to point mutations, while the rest represented other types of mutation. From this initial set of 2253 human proteins, we eliminated all entries describing mutations leading to stop codons and thus truncation of the full length protein. The final data set comprised 32 923 disease-related amino acid substitutions affecting 27 522 unique positions in 2067 different proteins.
Groups of orthologous proteins were downloaded from the STRING database (release 7.0) [50] where these were generated for their COG-interaction mode. Because some of these computationally derived orthologous groups contain a very broad range of sequence homology which in some cases may go beyond orthologs, we removed all sequences which failed to cover at least 80% of the human reference sequence in the multiple alignments according to STRING.
Often, such ortholog sets contain many sequences from very closely related species resulting in an excessive apparent sequence conservation and thus undue weight of almost identical sequences which does not reflect the evolutionary situation but the bias in protein selection. To overcome this bias, we iteratively removed near-identical sequences from the ortholog sets. We calculated the sequence identity of all pairs in global alignments. If two sequences were over 90% identical, one of them was picked at random.
Very small ortholog sets cannot reasonably be expected to allow valid conclusions about evolutionary correlation, because the resulting multiple alignments simply do not contain enough sequences. In the literature, a wide range of minimum ortholog clusters sizes have been applied. While some used as little as 15 proteins [21], [51], [52], other required more than 125 different orthologs [18], [53], [54]. In our study, we initially used all clusters with at least 30 orthologous proteins, which is a very common cutoff [53], [55]. In addition, we also performed our analysis on a more stringent data set by using a threshold of . The former threshold yields 1153 human proteins with their orthologs, while the more strict cutoff still leaves us with 855 such clusters.
We used the feature annotation from SWISSPROT [39], [40] to identify functionally relevant positions in each of the disease associated proteins with a sufficient number of orthologs. We included the following feature tags in our analysis: CA_BIND (calcium-binding), DNA_BIND (DNA binding), NP_BIND (nucleotide phosphate-binding), ACT_SITE (involved in enzyme activity), METAL (metal binding), BINDING (binding of unspecified chemical group), MOD_RES (posttranslational modification) and LIPID (lipid binding). In total, we obtained 12021 functional residues in 745 proteins.
Alignments of orthologous proteins were carried out using MUSCLE 3.6 [56] with default parameters. To reduce computational requirements we limited each ortholog set to a maximum of 300 sequences, after filtering, by selecting a random sample of 299 sequences plus the human reference sequence, which always needs to be present for analysis.
Correlated mutations were analyzed using the OMES (Observed Minus Expected Squared) algorithm. The OMES method is based on the goodness-of-fit test and compares the observed co-occurrence of amino acid at position and amino acid at position to the expected co-occurrence at positions and [20], [25], [57]. In this work we use the OMES variant defined by Fodor et al. [25]. We computed OMES correlation scores for all combinations of positions in each protein based on the multiple sequence alignments described above. Following the previously described approach we selected the top co-evolving residue pairs where is the length of the respective protein and is a constant which is often set to 5 [55], [58]. To assess the influence of the constant we evaluated our findings over a range of . Based on this analysis, the commonly used value of appears to be a good choice for our application: For values of from 0 to 5 the observed LOD shows a steep increase. Somewhere around or 10 the curve adopts a much more moderate slope (Figure 8). A choice of takes advantage of the initial LOD improvement without including excessive numbers of positions in each protein.
For further analysis, a sequence position was called correlated if it had at least one significant correlation with another position according to the above criteria.
Two different measures of sequence conservation were computed for each position of a human protein. The BLOSUM conservation-score for each human residue was calculated by summing over the BLOSUM-scores for each residue pair between the human amino acid and all ortholog residues in the column of the MSA and normalizing to the maximum score for the given residue:where is the th residue of sequence in the MSA; is the total number of sequences; is the number of gaps in column ; refers to the human reference sequence and is the score for amino acids and according to the BLOSUM62 scoring matrix [33]. As BLOSUM scores can be negative,
The second approach simply computes the fraction of residues identical to the reference sequence for column of the MSA:
Protein structure analysis was performed for all proteins of the filtered set for which at least a partial crystal structure was available from the PDB database [59]. 238 proteins fulfilled this criterion. The spatial distance between correlated residue pairs was calculated taking into account all non-hydrogen side chain atoms of both amino acids. Two residues were considered to be in contact with each other, if the smallest distance between any pair of their non-hydrogen atoms was Å. This represents a commonly applied threshold – other groups have used distance cutoff in the range 5.0Å–8.0Å using the closest non-hydrogen or C atoms, respectively [20], [53], [55]. In order to exclude local contacts in secondary structure elements, such as in -helices, we only considered residue pairs outside a window of 10 positions up- and downstream of a given position.
Some studies on correlated mutations have computed distances solely based on atoms (Glycine: ) and a 8.0Å cutoff [25]. We provide additional data using this definition in the Text S1.
The accessible surface area (ASA) was computed with DSSP [60] and converted to a relative solvent accessibility (RAS) by dividing by the maximum possible ASA of the respective amino acid. The data was discretized into three accessibility states: buried (); intermediate () and exposed ().
Amino acid substitutions with a low chance of causing harm were identified with a strategy used by [61]. The rationale of this approach is that amino acid substitutions tolerated during evolution are very unlikely to cause disease when observed in humans, at least for closely related species. We selected all human proteins from our set for which we were able to identify mammalian orthologs (from the ortholog clusters provided by the STRING database [50]) with at least 95% identity to the human sequence. Orthologs were only considered if they covered at least 80% of the human sequence in a pairwise alignment. 813 proteins satisfied both criteria. Amino acid substitutions found in the orthologs were considered non-damaging.
Statistical significance of enrichment of disease mutations in correlated positions was assessed by two different means. First we used Fisher's exact test for count data comparing the proportions of disease-annotated residues in correlated vs. non-correlated positions. In addition, we performed a 1000-fold permutation test in which we shuffled the disease/non-disease tags of the entire data set in order to obtain the empirical density function of the log odds (Figure 1). In order to assess the robustness of the observed LOD value we performed a 1000-fold bootstrapping of columns of the multiple sequence alignments. I.e. in each bootstrap we sampled, with replacement, from the columns of each of the multiple sequence alignments in the data set and then re-computed the LOD value for the entire re-sampled data set.
All analysis programs for this work were written in Python, except for the program for correlated mutation analysis which was implemented in Java. Final data analysis and statistics was performed with the R statistical language [62]. Visualizations of correlations in the linear sequence were created with Circos [63]. Protein structure images were made with PyMOL (http://pymol.sourceforge.net/).
|
10.1371/journal.pntd.0001521 | Estimating the Non-Monetary Burden of Neurocysticercosis in Mexico | Neurocysticercosis (NCC) is a major public health problem in many developing countries where health education, sanitation, and meat inspection infrastructure are insufficient. The condition occurs when humans ingest eggs of the pork tapeworm Taenia solium, which then develop into larvae in the central nervous system. Although NCC is endemic in many areas of the world and is associated with considerable socio-economic losses, the burden of NCC remains largely unknown. This study provides the first estimate of disability adjusted life years (DALYs) associated with NCC in Mexico.
DALYs lost for symptomatic cases of NCC in Mexico were estimated by incorporating morbidity and mortality due to NCC-associated epilepsy, and morbidity due to NCC-associated severe chronic headaches. Latin hypercube sampling methods were employed to sample the distributions of uncertain parameters and to estimate 95% credible regions (95% CRs).
In Mexico, 144,433 and 98,520 individuals are estimated to suffer from NCC-associated epilepsy and NCC-associated severe chronic headaches, respectively. A total of 25,341 (95% CR: 12,569–46,640) DALYs were estimated to be lost due to these clinical manifestations, with 0.25 (95% CR: 0.12–0.46) DALY lost per 1,000 person-years of which 90% was due to NCC-associated epilepsy.
This is the first estimate of DALYs associated with NCC in Mexico. However, this value is likely to be underestimated since only the clinical manifestations of epilepsy and severe chronic headaches were included. In addition, due to limited country specific data, some parameters used in the analysis were based on systematic reviews of the literature or primary research from other geographic locations. Even with these limitations, our estimates suggest that healthy years of life are being lost due to NCC in Mexico.
| Neurocysticercosis (NCC) is a major public health problem caused by the larvae of the parasite Taenia solium. The condition occurs when humans ingest eggs of the pork tapeworm Taenia solium, which then develop into larvae in the central nervous system. The disease is predominantly found and considered important in Latin American, Asian, and African countries and is associated with a large social and economic burden. Very few studies have been conducted to evaluate the burden of NCC and there are no estimates from Mexico. We estimated the disability adjusted life years (DALYs) lost due to NCC in Mexico incorporating morbidity and mortality due to NCC-associated epilepsy, and morbidity due to NCC-associated severe chronic headaches. NCC-associated epilepsy and severe chronic headaches were estimated to cause a loss of approximately 0.25 healthy year of life per 1,000 persons annually in Mexico. This is the first estimate of DALYs associated with NCC in Mexico. However, this value is likely to be underestimated since only the clinical manifestations of epilepsy and severe chronic headaches were included.
| Neurocysticercosis (NCC) is a major public health problem caused by the larvae of the zoonotic cestode Taenia solium. Humans are the definitive hosts of T. solium and become infected with the intestinal adult tapeworm (taeniasis) by ingesting undercooked pork containing cysticerci. Humans can also become accidental intermediate hosts after ingesting T. solium eggs leading to cysticercosis and/or NCC, which occurs when larvae develop in the central nervous system. A recent meta-analysis of published studies on the frequency of NCC estimated that 29% (95% CI: 23%–36%) of epilepsy cases in NCC-endemic areas exhibit NCC lesions as identified by brain neuroimaging [1]. NCC may also manifest as migraine-type headaches and stroke, among others [2], [3].
NCC is common in many developing countries where health education, sanitation, and meat inspection infrastructure are insufficient [4]. This disease is predominantly found and considered endemic in Latin American, Asian, and Sub-Saharan African countries [5], [6], [7]. In Mexico, NCC is one of the main causes of late onset epilepsy [8]. Very few studies have been conducted to evaluate the burden of NCC [9], [10] and there are no estimates from Mexico. The disability adjusted life year (DALY), developed for the Global Burden of Disease (GBD) Study, is the most common metric used to measure disease burden. It combines years of life lost due to premature mortality (YLL) and years of life lost due to time lived in a disability state (YLD). One DALY is considered the equivalent of one year of healthy life lost [11].
Two previous studies, both conducted in Africa, have evaluated the burden of cysticercosis [9], [10]. A study in Cameroon revealed that the average number of DALYs lost due to NCC was 9.0 per 1,000 person-years (95% CR: 2.8–20.4) and the monetary burden per case of cysticercosis amounted to 194 Euro (95% CR: 147–253) [10]. This estimate only accounted for the disease burden due to NCC-associated epilepsy and used serology for the diagnosis of NCC. Another study conducted in South Africa estimated that the monetary burden of NCC varied from US$ 632 to US$ 844 per NCC-associated epilepsy case, indicating high financial losses associated with this condition [9].
Studies are needed to estimate the burden of NCC in endemic countries, such as Mexico, to facilitate international comparison of disease burden and identify priorities for control. The research presented here provides the first estimate of DALYs associated with NCC in Mexico, incorporating two common clinical manifestations of patients with NCC, epilepsy and severe chronic headaches [12].
This study was conducted in Mexico, which is the third largest country in Latin America, with a 2005 population of almost 103 million and an annual population growth rate of 1.1% [13]. Traditional pig rearing practices in NCC endemic areas allow pigs to have access to human feces in open fields facilitating the completion of the T. solium life cycle [14].
Literature published through June 2011 was searched in PubMed using the key words expressions “epilepsy AND prevalence AND Mexico” and “epilepsy AND incidence AND Mexico”. Articles that were referenced in the literature identified through the original search were also obtained and reviewed (Figure 1). Community-based studies that employed door-to-door surveys of households or surveys of school children were eligible for inclusion if they were carried out by trained personnel using standardized previously validated questionnaires. In addition, data from systematic reviews and meta-analyses on epilepsy frequency were included as appropriate. The identified literature was further evaluated for epilepsy frequency data reported separately for adults and children and for urban and rural areas. For the purpose of this study, epilepsy was defined as the occurrence of at least two unprovoked seizures separated by at least 24 hours [15].
The epilepsy prevalence estimates used in the current study were based on estimates from Quet et al. (2011) [16]. This study reported epilepsy prevalence for males and females in different age groups (<15 years, 15–40 years, 41–60 years and >60 years) in a rural community in the Puebla State of Mexico in 2007. For analysis purposes, the prevalence estimates were applied to GBD age groups as appropriate. Uncertainty was modeled with uniform distributions using the upper and lower confidence intervals values from the Quet et al. (2011) study. A systematic review of published literature on epilepsy conducted in Latin America in 2005 reported that the prevalence of epilepsy in urban and rural areas was not significantly different [17]. Since there were no comparable data available specifically for Mexico, it was assumed that the prevalence of epilepsy in urban and rural areas of Mexico is also not significantly different.
A flowchart depicting how incidence of treated and untreated NCC-associated epilepsy was determined is shown in Figure 2. The number of cases of epilepsy was estimated by multiplying the age and rural/urban stratified population size from the 2005 census [13] by the epilepsy prevalence estimates described above [16]. While the prevalence of epilepsy in rural and urban areas of Mexico was not assumed to be different, the proportion of NCC-associated epilepsy cases in urban and rural areas was assumed to be different. This assumption was based on the only identified study to look at the prevalence of NCC among patients with active epilepsy in rural versus urban clusters [18]. Although the identified study was conducted in the Vellore district of India, the difference between urban and rural areas likely reflects a difference in T. solium endemicity, which would also be found in Mexico. For example, pig rearing practices used in the rural areas of Mexico facilitate the completion of the T. solium life cycle.
The estimated proportion of epilepsy cases with NCC lesions from a meta-analysis of NCC frequency data [1] was used for rural areas of Mexico (Table 1). This value was chosen because the meta-analysis used data from primarily rural endemic communities in countries such as Bolivia, Ecuador, Honduras, Burkina Faso, and India to calculate the proportion of epilepsy cases with NCC lesions. The Vellore, India study reported the prevalence of NCC among patients with acute epilepsy in urban areas to be approximately 0.52 times that in rural areas [18]. Since such information is not available from Mexico, the lower and upper confidence bounds of the proportion of children and adults with NCC-associated epilepsy from the meta-analysis [1] were multiplied by 0.52 to obtain the corresponding lower and upper confidence bounds of the proportion of children and adults with NCC-associated epilepsy in urban areas. The estimated numbers of adults and children with epilepsy in rural and urban areas were then multiplied by the respective proportion of people with epilepsy with NCC lesions to obtain the number of NCC-associated epilepsy cases in adults and children in rural and urban areas.
A literature review was conducted to identify information on the epilepsy treatment gap in Mexico. Literature published through June 2011 was searched in PubMed using the key words expression “epilepsy AND treatment gap AND Mexico”. When this search resulted in no usable data, a broader search using the key words expression “epilepsy AND treatment gap” was conducted. No literature was found that directly reported an estimate of the epilepsy treatment gap in Mexico. However, a 2010 systematic review of the epilepsy treatment gap literature worldwide was identified [22]. The authors of this systematic review conducted a meta-analysis of treatment gap data based on country income level (low, low middle, upper middle, and high) as defined by the World Bank [19], as well as urban versus rural location. Urban versus rural designation was based on the site description in the methods sections of the manuscripts included in the review. Based on data presented graphically in the epilepsy treatment gap meta-analysis, it was estimated that 77% (95% CI: 67%–87%) and 38% (95% CI: 27%–50%) of people with epilepsy in rural and urban areas, respectively, do not receive treatment [20]. These values were used given the unavailability of such data from Mexico. The number of NCC-associated epilepsy cases not receiving treatment was estimated by multiplying the number of NCC-associated epilepsy cases in rural and urban areas by the respective percentage not receiving treatment according to the epilepsy treatment gap meta-analysis [20]. Incidence of NCC-associated epilepsy was estimated by dividing the prevalence of NCC-associated epilepsy by the reported duration of epilepsy for different age groups from the GBD study (Table 2) [21].
A recent systematic review and meta-analysis of the frequency of the main clinical manifestations associated with NCC, using documents published from January 1, 1990 to June 1, 2008, was used to estimate the proportion of children and adults with symptomatic NCC seen at neurological clinics that have epilepsy as a clinical manifestation. This meta-analysis reported that 79% (95% CI 70%–86%) of children (0–19 years old) and 63% (95% CI 52%–74%) of adults with symptomatic NCC seen in neurological clinics have epilepsy [22]. The total number of symptomatic cases of NCC receiving care in neurology clinics was calculated by dividing the number of NCC-associated epilepsy cases who seek treatment (estimated in the previous section) by the respective proportion of people with NCC that have epilepsy as a clinical manifestation. This requires the assumption that all NCC patients seeking treatment do so in neurology clinics.
A flowchart depicting how incidence of treated and untreated NCC-associated severe chronic headaches was determined is shown in Figure 3. Severe chronic headaches were defined as severe headaches that last for more than three continuous days. The number of people with NCC-associated severe chronic headaches who seek care was estimated by multiplying the total estimated rural/urban stratified numbers of NCC cases who go to neurology clinics (see previous section), by the proportion of NCC cases attending neurology clinics with headaches based on the systematic review of clinical manifestations associated with NCC [22]. The total number of people with NCC-associated severe chronic headaches in urban and rural areas was then calculated by dividing the total number of NCC-associated severe chronic headaches cases in neurological clinics by the respective proportion of NCC cases with severe chronic headaches [22].
Two scenarios were initially considered for the treatment gap for severe chronic headaches. The first scenario assumed that the epilepsy and severe chronic headaches treatment gaps were the same and the second assumed a 10% difference in treatment gaps. However, in the final analysis, the treatment gap for severe chronic headaches was assumed to be 10% higher than that of epilepsy due to the generally greater clinical severity of epilepsy. This estimate is consistent with treatment gaps reported in other countries. For example, studies conducted in the United Kingdom reported that the epilepsy treatment gap was 2%, whereas the migraine treatment gap was 14% [23], [24]. Morillo et al. (2005) estimated that 48% of the urban population of Mexico with severe headaches did not receive treatment [25]. This estimate is similar to our estimate (49%). The number of NCC-associated severe chronic headaches cases not receiving treatment in Mexico was estimated by multiplying the number of NCC-associated severe chronic headaches cases in rural and urban areas of Mexico by the respective estimated treatment gap.
The incidence of NCC-associated severe chronic headaches was estimated by dividing the prevalence estimated above by the mean duration of NCC-associated severe chronic headaches obtained from chart reviews conducted at two referral neurological hospitals in Mexico City [12]. Chart reviews captured information on duration of NCC-associated severe chronic headaches in adults 18 years of age and older from the time of diagnosis to the date when data were abstracted. The mean duration was calculated for each GBD age group (Table 2). Since chart review data were not available for children, a search of the literature was conducted to identify NCC-associated headache duration data for this age group. Literature published through June 2011 was searched in PubMed using the key words expression “neurocysticercosis AND headache AND duration”. However, no information on duration of NCC-associated headaches in children could be identified. Therefore, duration of severe chronic headaches used for the 15–44 years age group was also applied to the 0–15 years age group (Table 2).
The number of DALYs lost was calculated by adding the number of years lived with a disability (YLD) to the number of years of life lost due to mortality (YLL). The formulas used for the calculation of YLD and YLL are shown in Eqs. 1 and 2 respectively:(1)where I = age and sex specific estimates of incidence, DW = disability weight, D = average duration of disability(2)where N = number of deaths per age-sex group, L = remaining life expectancy at age of death [26].
The calculation of years of life lost due to time lived in a disability state (YLD), requires the use of disability weights. Disability weights for NCC were not included in the original GBD Study or its subsequent updates. Therefore, disability weights for epilepsy provided by GBD studies were used for NCC-associated epilepsy [21]. Disability weights for severe chronic headaches were also not included in previous GBD estimates, so published disability weights for migraine were used as a surrogate [27]. Disability weights for epilepsy and severe chronic headaches used in this study are listed in Table 2.
Years of life lost due to premature mortality (YLL) were estimated using standard life expectancies (life expectancy of 82.5 years at birth for women and a life expectancy of 80.0 years at birth for men) [26]. Three percent discounting and non uniform age weighting (β = 0.04 and C = 0.1658) were applied [26]. These standard values were chosen to compare our estimates to DALYs calculated for other conditions in Mexico. World Health Organization (WHO) 2004 mortality rates for epilepsy in Mexico were used to calculate the years of life lost due to premature NCC-associated epilepsy mortality (YLL) [28]. The numbers of deaths of adults and children with epilepsy in Mexico were multiplied by the respective proportion of people with epilepsy with NCC lesions to obtain NCC-associated epilepsy deaths [1]. A review of the literature was conducted to identify data on NCC-associated headache deaths. Literature published through June 2011 was searched in PubMed using the key words expressions “neurocysticercosis AND headache AND mortality” and “neurocysticercosis AND headache AND death”. However, no publications could be identified that evaluated deaths attributable to NCC-associated headaches. Therefore, due to a lack of available data, we did not feel comfortable including an estimate for NCC-associated headaches deaths.
The number of DALYs lost due to NCC, with its 95% credible region (95% CR), was estimated using @Risk (Palisade Corporation, Ithaca, NY, version 4.5). Latin Hypercube sampling was used for uncertain parameters (distributions shown in Tables 1–2). The model was run for 30,000 iterations to achieve convergence. Uncertain epidemiological parameters were modeled using uniform distributions, while disability weight uncertainty was modeled using beta distributions. Regression sensitivity analysis was conducted in @ Risk by varying the value of each parameter to estimate its correlation to the total DALYs estimate. The relative values of the regression coefficients indicate which parameters had the greatest impact on the total DALYs estimate.
Approximately 0.14% of the total population of Mexico was estimated to have NCC-associated epilepsy and 0.08% was estimated to have NCC-associated severe chronic headaches. The estimated numbers of people in urban and rural areas with NCC-associated epilepsy and severe chronic headaches, along with 95% CRs, are reported in Table 3. Annual incident cases of NCC-associated epilepsy and severe chronic headaches were 0.05 and 0.02 per 100 person-years, respectively (Table 4).
The total number of DALYs lost due to NCC-associated epilepsy and severe chronic headaches in Mexico was estimated at 23,020 (95% CR: 11,283–43,276) and 2,321 (95% CR: 198–8,754), respectively, with 0.25 (95% CR: 0.12–0.46) DALY lost per 1,000 person-years (Figure 4). Twenty-eight percent of DALYs lost due to NCC was attributed to YLL and the remaining 72% was due to YLD (Table 5).
Based on the regression sensitivity analysis, the epilepsy disability weights for untreated and for treated NCC in people older than 5 years of age and the prevalence of epilepsy in 15–44 year-old males and females were the four parameters with the greatest effect on the total DALYs estimate (Figure 5).
This study represents only the second study to estimate the burden of NCC using DALYs. The first, which was conducted in Cameroon, estimated human NCC burden based on epilepsy alone [10]. The estimated number of DALYs lost per 1,000 person-years was higher in Cameroon (9.0) compared to Mexico (0.25). One difference between the two studies is that all of our data were stratified by urban/rural areas, age groups, and gender. Such stratification was not used in the Cameroon study. Since the majority of the Mexican population is urban, and the proportion of epilepsy cases attributable to NCC is lower in urban areas, the overall burden per person is expected to be lower in Mexico. In addition, NCC-associated epilepsy patients were four times more likely to receive treatment in Mexico than in Cameroon. Because the disability weight for treated epilepsy is much lower than that for untreated epilepsy, this results in fewer DALYS per 1,000 person-years.
Similarly, annual number of deaths due to NCC-associated epilepsy was estimated to be higher (6.9% of the total annual incident cases) in Cameroon compared to Mexico (0.5% of the total annual incident cases). When the current model for Mexico was run using the mortality rate from the Cameroon study, 1.08 (95% CR: 0.8–1.4) DALYs per 1,000 person-years were projected to be lost compared to 0.25 (95% CR: 0.12–0.46) DALYs per 1,000 person-years. This suggests that the high mortality associated with NCC-associated epilepsy in Cameroon had a significant impact on disease burden in that country. Finally, the estimated 9 DALYs lost per 1,000 person-years [10] due to NCC-associated epilepsy in Cameroon is three times higher than the 2004 GBD estimate of 2.45 DALYs per 1,000 person-years due to all cases of epilepsy in Cameroon [28]. This suggests that the authors may have overestimated the burden of NCC associated epilepsy in that country or else that the GBD estimates for epilepsy were highly conservative.
According to 2004 GBD estimates, 1.7 DALYs per 1,000 person-years were estimated to be lost due to epilepsy in Mexico, with approximately the same number of DALYs lost due to migraine [28]. Our estimates for the number of DALYs lost per 1,000 person-years due to NCC was higher than such estimates for other helminthic infections in Mexico (ascariasis-0.05, trichuriasis-0.10, hookworm-0.03) due to the severity of clinical manifestations associated with NCC and because NCC not only causes morbidity, but also mortality in humans [28].
Our study has some limitations. The total estimated number of DALYs lost was most likely underestimated since only the NCC-associated clinical manifestations of epilepsy and severe chronic headaches were included. There are many other clinical manifestations of NCC [22] which could not be included largely due to lack of information on frequency and disability weights. Since data on the incidence of NCC-associated epilepsy and severe chronic headaches were not available, the prevalence was divided by the duration of symptoms to obtain an incidence value. In addition, we assumed that the duration of epilepsy and severe chronic headaches was the same among treated and untreated cases, which is unlikely to be accurate. The mean duration of NCC-associated severe chronic headaches was estimated from the time of diagnosis to the end of symptoms based on the review of medical charts of patients seeking care in tertiary hospitals of Mexico City [12]. This may overestimate the true duration of NCC-associated severe chronic headaches if only the most severe cases are seen in tertiary hospitals. On the other hand, it could also lead to an underestimation of the duration since people may wait a long time before seeking care. Due to limited country-specific data, parameters for the proportion of NCC patients with epilepsy and severe chronic headaches and the epilepsy treatment gap were based on systematic reviews of the literature [1], [20], [22].
Based on the regression sensitivity analysis, the disability weight used for individuals with epilepsy who were greater than 4 years of age was by far the most influential parameter. More precise values of disability weights in future versions of the GBD should reduce the uncertainty. The next most influential values were linked to the prevalence of epilepsy in Mexico. The estimates used here were based on a single study that may not fully reflect the variation of prevalence among the whole country. Better knowledge of the actual prevalence of epilepsy in Mexico would also improve our estimates.
It should be noted that DALY estimates only incorporate human health losses. However, Taenia solium cysticercosis not only causes losses to human health, but also to pig farmers and their communities. Therefore, the total societal burden is higher than that estimated by the number of DALYs lost. An analysis of the monetary burden of NCC in Mexico is currently underway and will be presented in a later publication.
In conclusion, this is the first estimation of the non-monetary burden of NCC in Mexico using the DALY. These estimates suggest that healthy years of life continue to be lost annually in Mexico, with a continued effort needed to control this parasitic disease in endemic regions.
|
10.1371/journal.pgen.1005140 | Spatio-temporal Remodeling of Functional Membrane Microdomains Organizes the Signaling Networks of a Bacterium | Lipid rafts are membrane microdomains specialized in the regulation of numerous cellular processes related to membrane organization, as diverse as signal transduction, protein sorting, membrane trafficking or pathogen invasion. It has been proposed that this functional diversity would require a heterogeneous population of raft domains with varying compositions. However, a mechanism for such diversification is not known. We recently discovered that bacterial membranes organize their signal transduction pathways in functional membrane microdomains (FMMs) that are structurally and functionally similar to the eukaryotic lipid rafts. In this report, we took advantage of the tractability of the prokaryotic model Bacillus subtilis to provide evidence for the coexistence of two distinct families of FMMs in bacterial membranes, displaying a distinctive distribution of proteins specialized in different biological processes. One family of microdomains harbors the scaffolding flotillin protein FloA that selectively tethers proteins specialized in regulating cell envelope turnover and primary metabolism. A second population of microdomains containing the two scaffolding flotillins, FloA and FloT, arises exclusively at later stages of cell growth and specializes in adaptation of cells to stationary phase. Importantly, the diversification of membrane microdomains does not occur arbitrarily. We discovered that bacterial cells control the spatio-temporal remodeling of microdomains by restricting the activation of FloT expression to stationary phase. This regulation ensures a sequential assembly of functionally specialized membrane microdomains to strategically organize signaling networks at the right time during the lifespan of a bacterium.
| Cellular membranes organize proteins related to signal transduction, protein sorting and membrane trafficking into the so-called lipid rafts. It has been proposed that the functional diversity of lipid rafts would require a heterogeneous population of raft domains with varying compositions. However, a mechanism for such diversification is not known due in part to the complexity that entails the manipulation of eukaryotic cells. The recent discovery that bacteria organize many cellular processes in membrane microdomains (FMMs), functionally similar to the eukaryotic lipid rafts, prompted us to explore FMMs diversity in the bacterial model Bacillus subtilis. We show that diversification of FMMs occurs in cells and gives rise to functionally distinct microdomains, which compartmentalize distinct signal transduction pathways and regulate the expression of different genetic programs. We discovered that FMMs diversification does not occur randomly. Cells sequentially regulate the specialization of the FMMs during cell growth to ensure an effective and diverse activation of signaling processes.
| Cells typically compartmentalize their cellular processes into subcellular structures (e.g. organelles) to optimize their efficiency and improve their activity. One of the most interesting concepts in cellular compartmentalization is the proposed existence of lipid rafts in the membranes of eukaryotic cells [1]. Eukaryotic membranes organize a large number of proteins related to signal transduction, protein sorting and membrane trafficking into discrete nano-scale domains termed lipid rafts [1,2]. The functional diversity of lipid rafts is currently attributed to a different lipid and protein composition, as compelling evidence suggests that a heterogeneous population of lipid rafts could exist on a given cell [3–5]. Yet, the molecular mechanisms by which cells generate and regulate raft heterogeneity are still unclear. In eukaryotic systems, it is known that the integrity of lipid rafts requires the activity of two different raft-associated proteins termed flotillins (FLO-1 and FLO-2) [6,7]. Flotillins are scaffolding proteins, which may redundantly act as chaperones in recruiting the protein cargo to lipid rafts and interact with the recruited proteins that activate the signal transduction processes [8–10]. Consequently, the perturbation of the activity of flotillins causes serious defects in several signal transduction and membrane trafficking processes, which seems to be intimately related to the occurrence of severe human diseases, such as Alzheimer’s disease, Parkinson’s disease or muscular dystrophy (reviewed in [11]).
The spatial organization of signaling networks in lipid rafts has been considered a hallmark in cellular complexity because their existence is exclusively associated with eukaryotic cells. However, we recently discovered that bacteria organize many proteins related to signal transduction in functional membrane microdomains (FMMs) that are structurally and functionally similar to the lipid rafts of eukaryotic cells [12]. Bacterial flotillins are important components for the organization and the maintenance of the architecture of FMMs. Similar to the eukaryotic flotillins, bacterial flotillins probably act as scaffolding proteins in tethering protein components to the FMMs, thereby facilitating their efficient interaction and oligomerization and to mediate the efficient activation of signal transduction pathways harbored in FMMs. Consequently, mutants lacking flotillins show a severe defect in FMM-localized signaling pathways concomitantly with a severe dysfunction of diverse physiological processes, such as biofilm formation, natural competence or sporulation [12–17].
The FMMs of the bacterial model Bacillus subtilis contain two different flotillin-like proteins, FloA and FloT [12]. FloA and FloT flotillins physically interact [13] and presumably play a redundant role because the dysfunction of specific FMM-associated physiological processes, like biofilm formation, only occurs in the ΔfloA ΔfloT defective mutant and is not observed in either of the ΔfloA or ΔfloT single mutants [17]. Likewise, the overexpression of both floA and floT causes pleiotropic effects in cell division and cell differentiation but this effect is not observed in cells that overexpress one single flotillin gene [16]. In this respect, bacterial flotillins seem to behave similarly to human flotillins FLO-1 and FLO-2, given that both FLO-1 and FLO-2 are associated with each other in hetero-oligomeric complexes and have a strong regulatory correlation [18–20]. These experimental evidences led to the general assumption that both flotillins play a redundant function in both eukaryotic lipid rafts and bacterial FMMs.
In this report, we provide evidence that a heterogeneous population of membrane microdomains coexists on bacterial cells. We show that FloA and FloT are two functionally different flotillins that physically interact but unevenly distribute within the FMMs of bacterial cells. FloA and FloT act as specific scaffold proteins that tether a defined group of FMMs-associated proteins. This generates functionally distinct microdomains, which compartmentalize distinct signal transduction pathways and regulate different genetic programs. Importantly, we show that cells sequentially regulate the functional specialization of the FMMs during cell growth. Cells restrict the expression of the floT gene to stationary phase to ensure an effective activation of signaling processes at specific times during the lifespan of the bacterium.
While exploring flotillin redundancy in the FMMs of B. subtilis, we discovered that the expression of FloA and FloT is controlled by different genetic programs, which could indicate that these are two functionally different flotillins. We came across this finding by examining the expression profiles of floA and floT genes in the 249 different growing conditions that are published in [21,22], and are available in SubtiExpress (http://subtiwiki.uni-goettingen.de/apps/expression). By doing this, we consistently found high expression of floA in all the growing conditions tested, including LB and MSgg growth media, the two growth media that we normally used in the laboratory to grow B. subtilis (S1A Fig) [23]. However, the expression of floT showed more variability among the growth conditions tested and exhibited an important difference in gene expression between LB (lower expression of floT) and MSgg (higher expression of floT). To test this in the laboratory, we constructed B. subtilis strains harboring the PfloA-yfp and PfloT-yfp transcriptional fusions (YFP is yellow fluorescence protein) and grew them in LB and MSgg media [23]. Our laboratory uses the chemically-defined medium MSgg to induce sporulation and the formation of robust biofilms in B. subtilis cultures and differs to LB medium in which B. subtilis did not show any of the developmental characteristics of MSgg [24]. By growing B. subtilis cells in these two growth conditions, we detected an activation of floT expression in MSgg (S1A Fig), while LB medium showed poor activation of floT expression (S1B Fig). In contrast, floA was equally expressed in both MSgg and LB media. Furthermore, we generated strains labeled with the FloA-GFP and FloT-GFP translational fusions (GFP is green fluorescence protein) to visualize and quantify flotillin protein production using flow cytometry. The FloT-GFP labeled strain showed a reduction of the fluorescence signal when grown in LB medium while FloA-GFP was equally expressed in both MSgg and LB media (S1C Fig).
To investigate whether the differential production of FloA and FloT is a cell-regulated process, the strains labeled with PfloA-yfp and PfloT-yfp transcriptional fusions were used to systematically inactivate regulatory genes of B. subtilis and search for mutants capable of altering the expression of floT in MSgg medium (S1D Fig). We detected a uniform expression of floA in all mutants tested. However, we discovered that cells lacking the abrB gene showed increased expression of floT. Additionally, we found inhibition of floT expression in cells when the spo0A gene was deleted. Importantly, spo0A and abrB belong to the same signaling pathway. AbrB is a repressor of biofilm formation among other processes [25] and its expression is negatively regulated by Spo0A [26]. Spo0A is a master regulatory protein necessary for the activation of many physiological processes related to stationary phase [27]. Therefore, this provides epistatic evidence that Spo0A positively regulates floT expression at stationary phase via inhibition of abrB and that this genetic cascade does not affect the expression of floA. To test this hypothesis, we deleted spo0A and/or abrB genes in FloA-GFP and FloT-GFP labeled strains and monitored the subcellular distribution pattern of flotillins using fluorescence microscopy and applying a deconvolution algorithm to eliminate out-of-focus signal and to improve their correct visualization (see material and methods section) (Fig 1). Indeed, Δspo0A or Δspo0A ΔabrB mutants showed no variation in the distribution pattern of fluorescence foci that were generated by FloA (Fig 1A and 1B). In contrast, Δspo0A mutant showed a severe reduction of fluorescence foci that were generated by FloT, which could be reconstituted in the Δspo0A ΔabrB double mutant (Fig 1C and 1D).
Activation of Spo0A (Spo0A~P) occurs at stationary phase due in part to the activation of the histidine kinase C (KinC) [28,29], which is driven by the action of the self-produced signaling molecule surfactin. Thus, FloA-GFP and FloT-GFP labeled strains were grown in LB medium and complemented with exogenously added surfactin (5 μM) (S2A and S2C Fig). FloA-GFP labeled cells showed no alteration of the fluorescence signal but FloT-GFP labeled cells showed an increase in the number of foci (S2A–S2D Fig). This is a Spo0A-depedent effect because the spo0A deficient strain showed no recovery of FloT expression upon addition of surfactin (S2E and S2F Fig). Altogether, these results show an upregulation of FloT production at stationary phase in a Spo0A-dependent manner likely via AbrB. In contrast, the production of FloA is not influenced by this regulatory cascade.
The distinct regulatory programs for FloA and FloT production led us to hypothesize that FloA and FloT may play different roles in B. subtilis cells. To investigate this hypothesis, we first explored whether FloA and FloT show any structural difference. FloT is a larger protein (509 aa) that has an extended C-terminal region compared to FloA (331 aa) (Figs 2A and S3A). To determine if these structural differences are associated with a different subcellular distribution pattern, we used strains labeled with the FloA-GFP and FloT-GFP translational fusions to visualize and quantify the number of fluorescent foci (n = 400) using fluorescence microscopy. On average, FloA distributed in 13 foci per cell while FloT distributed approximately in 6 foci (Figs 2B, 2C and S3B). These results are consistent with the number of foci that we detected in the genetic analysis that are shown in Fig 1. However, to validate that these results were not a consequence of clustering artifacts [30], we compared their distribution pattern using non-dimerizing monomeric red fluorescence protein mCherry (mCh) in a total of 400 cells. Likewise, FloA distributed in 13 foci per cell while FloT distributed in 6 foci, as previously observed (Figs 2B, 2C and S3B). Importantly, the subcellular localization of flotillins consistently showed that FloA distributed in more foci per cell than FloT. To gain more insight about the differential distribution pattern of flotillins, we performed co-localization experiments using FloA-GFP, FloT-mCh and FloA-mCh, FloT-GFP double-labeled strains. Co-localization of both signals was detected by fluorescence microscopy, showing colocalization of FloT with FloA in all cells examined (Fig 2D), which adds to the previous notion that FloA and FloT physically interact [12,13,17]. However, the obvious differences in the number of foci between FloA and FloT resulted in the colocalization of both FloA and FloT signals only in a subset of foci (Pearson’s correlation coefficient R2 = 0.81). This diversified the pool of FMMs of a given cell into two different types of microdomains: one family of microdomains that contains solely FloA signal and a second type of microdomains in which both FloA and FloT signals converge. We performed time-lapse fluorescence microscopy experiments using a FloA-GFP FloT-mCherry double-labeled strain to investigate the dynamics of the subcellular co-localization. A series of images were taken at one-second time interval (Fig 2E). The fluorescence signal attributable to FloA and FloT reorganized dynamically within the membrane and consistently showed co-localization of signals from FloA and FloT.
The asymmetrical distribution pattern of FloA and FloT was further examined at higher resolution using super-resolution imaging by PALM [31,32]. To this end, FloA and FloT were labeled with the photoactivatable monomeric protein mEOS2 and expressed in B. subtilis cells. FloA-mEOS2 and FloT-mEOS2 proteins were activated by low intensity irradiation at 405 nm. Photoactivated proteins were excited at 568 nm, imaged and bleached before the next cycle of photoactivation. Individual protein positions were determined (localized) in each image frame and used to reconstruct a high-resolution PALM image (Fig 3A and 3B). Clusters candidates were defined by either one connected pixel area in image-based analysis or by a cloud of scattered localizations with spatial coherence in localization based analysis. Spatial coherence implies that the increase local density of localizations follows a Gaussian distribution within the cluster, which is indicative of the nonrandom distribution of localizations. Using the raw localization data and the corresponding super-resolved image, we generated a mask to define possible cluster candidates and separate them from the localization pseudo background. By using this technique, we confirmed that FloA assembled in 13 small clusters per cell (Diameter = 46.73 ± 1.35 nm). FloT however, assembled in approximately 6 larger clusters per cell (Diameter = 63.39 ± 2.28 nm) with a higher content of proteins (Fig 3C–3F). To validate these results, we also monitored the distribution pattern of FloA and FloT when fused to photoactivatable monomeric PAmCherry using PALM (S4A and S4B Fig). The statistical analysis of the signals detected by PALM and further validation by western blot analysis suggested that FloT is more abundant than FloA in cells and yet, based on our results, is concentrated in a lower number of foci (S4C and S4D Fig).
The molecular basis of the asymmetrical distribution of FloA and FloT was explored by monitoring the intra- and inter-specific interactions that occur between FloA and FloT flotillins. To do this, we used a bacterial two-hybrid (BTH) assay, in which FloA and FloT were tagged to T25 or T18 catalytic domains of an adenylate cyclase that reconstitute the enzyme upon interaction of two proteins [33]. A fully active adenylate cyclase produces cAMP, which accumulates in the cytoplasm and triggers the expression of a cAMP-inducible lacZ reporter gene [33]. Using this assay, we detected a strong interaction signal with FloA alone (Fig 4A) (Instructions of the manufacturer define a positive signal if above the threshold of 700 Miller Units [33]). Likewise, a strong interaction signal was detected with FloT (Fig 4A). This is indicative of the capacity of FloA and FloT to form homo-oligomers. However, when we assayed the interactions between FloA and FloT, the interaction signal was less prominent in comparison to the FloA-FloA and FloT-FloT interactions, suggesting that flotillins are prone to form homo-oligomers while hetero-oligomerization occurs to a lesser extent (Fig 4A). The propensity to form homo-oligomers suggests different interaction properties between FloA and FloT, which is probably a determinant in the generation of distinct subcellular distribution patterns.
Both FloA and FloT have a N-terminal region that anchors the protein to the membrane and the SPFH domain that is characteristic of this protein family (for stomatin, prohibitin, flotillin and HflK/C) [34,35]. However, the C-terminal region, which is the most variable region between FloA and FloT, contains four glutamate-alanine repeats (EA repeats) that are responsible for the oligomerization of human FLO-1 and FLO-2 (Figs 2A and S3A) [36] and are probably important in determining the interactions between FloA and FloT. We performed site-directed mutagenesis of the C-terminal region of each flotillin, which generated several variants of FloA and FloT, in which each one of the four EA repeats was replaced (EA→GL) (S5B Fig). We assayed the interaction properties of each one of the resultant variants using a BTH approach. FloA-FloA and FloT-FloT interactions did not occur when we altered the EA2 or EA4 repeats (≤ 700 Miller Units). Additionally, FloA-FloA interaction was abrogated when EA1 was mutated (≤ 700 Miller Units) while EA3 seemed to minimally affect the homo-oligomerization of both FloA and FloT (≥ 700 Miller Units) (Fig 4B). Moreover, the localization pattern of GFP-labeled variants was examined. Variants with EA2 and EA4 altered repeats showed poor aggregation and a severe decrease in the number of foci (Fig 4C and 4D). Alterations in EA1 affected severely the oligomerization of FloA while the variants with altered EA3 showed mild alterations in their distribution pattern (Fig 4C and 4D). None of the distribution patterns were appreciably altered in the absence of the alternative flotillin, suggesting that additional interaction motifs may exist to facilitate hetero-oligomerization (S5C Fig). Since the expression of the altered variants was still detected by western blot analysis (S5D Fig), it is possible that they become dispersed throughout the cellular membrane. Thus, we constructed a mEOS2-tagged version of FloA(EA4) and FloT(EA2) to study their subcellular distribution pattern using PALM microscopy. By using this approach, we detected a large number of single fluorescent proteins randomly dispersed across the cellular membrane (Fig 4E and 4F) rather than organized in foci.
The abovementioned results suggest that FloA and FloT display distinct subcellular distribution pattern, due in part to their different oligomerization affinities, which are determined by the specific interactions that occurred at the C-terminal region of each flotillin. We confirmed these observations by generating a chimeric version of FloA that contains the C-terminal region of FloT (FloAT) and a chimeric version of FloT that contains the C-terminal region of FloA (FloTA). GFP-fused versions of these proteins were generated to examine their subcellular distribution pattern (Fig 5). Using this approach, we consistently observed that the distribution pattern of FloAT resulted different from wild-type FloA and resembled the distribution pattern of wild-type FloT. Likewise, the distribution pattern of FloTA resulted very different from the wild-type FloT pattern, showing approximately 13 smaller foci per cell, which is similar to the distribution of wild-type FloA. These results are in agreement with what is shown in Fig 4 and confirmed that the c-termini regions confer specific oligomerization properties to each flotillin.
We were interested in exploring the biological significance of cells expressing two different flotillins with distinct spatio-temporal distribution patterns. We hypothesized that this may occur because these are two functionally different flotillins and therefore, they serve as scaffold in tethering the components of distinct signal transduction pathways in B. subtilis. We explored this hypothesis by first identifying the proteins that distinctively bind to either FloA or FloT. To do this, His6-tagged versions of FloA and FloT were expressed in B. subtilis cells. The membrane fraction was resolved by blue-native PAGE (BN-PAGE) to allow the separation of the membrane protein complexes in their natural oligomeric states [37]. Our BN-PAGE assays used a polyacrylamide gradient of 4%–20%, which allows the resolution of membrane-bound protein complexes with a molecular weight between 100 kDa and 1000 kDa. BN-PAGE coupled to immunoblotting, using antibodies against the His6 tag, was used to identify a number of membrane-associated protein complexes that exclusively interacted with FloA or FloT (Fig 6A and S3 Table). The corresponding bands were identified by mass spectrometry (MS) and validated as components of the protein cargo of the FMMs previously identified in analyses of the DRM fraction [12,13,17] (S3 Table).
MS analysis identified nine membrane proteins exclusively associated with FloA (Fig 6B). Their functional classification suggested their active participation in processes related to cell envelope regulation and cell division regulation. Those include the cytoskeletal-associated proteins MreC and PBP1A/1B or proteins related to cell wall remodeling, such as TagU and PhoR [38] (Fig 6B). We were particularly interested in the PhoR-FloA interaction, as this is a signaling kinase that activates a cascade that is related to cell wall organization [39] and is probably representative of the contribution of FloA to the FMMs. Using a BTH assay, we confirmed a specific interaction between PhoR and FloA (≥ 700 Miller Units) that was not observed between PhoR and FloT (≤ 700 Miller Units) (Fig 6C). In contrast, a total number of sixteen proteins were identified in exclusive association with FloT and their functional classification suggested an important role in adaptation to stationary phase (Fig 6B). This is the case for YclQ, YhfQ or YfiY proteins involved in siderophore uptake (reviewed in [40]); the protein secretion components SecA, SecDF and YacD, which have been correlated to FloT in previous studies [13] and the membrane-bound sensor kinase ResE, required for antibiotic, siderophore production and adaptation to oxygen-limiting conditions [41]. BTH analysis confirmed the interaction of ResE and FloT (≥ 700 Miller Units) that was not observed between ResE and FloA (≤ 700 Miller Units) (Fig 6C). We also identified a group of twenty-six proteins that interacted with both FloA and FloT (Fig 6B). The functional classification of this group is more diverse but generally related to cell differentiation processes. This includes the metalloprotease FtsH, required for the activation of Spo0A and thus, biofilm formation and sporulation [42] and known to interact with FloA and FloT from previous studies [13,16,17] and the OppABCDF oligopeptide permease, responsible for importing peptidic signals to activate biofilm formation or natural competence [43].
To investigate in more detail the interactions between FloA-PhoR and FloT-ResE that we discovered in the BN-PAGE and the bacterial two-hybrid analysis, we performed co-localization experiments using FloA-mCherry, PhoR-GFP and FloT-mCherry, ResE-GFP double-labeled strains. We confirmed by RT-PCR analyses that PhoR-GFP and ResE-GFP translational fusions complemented ΔphoR and ΔresE mutants respectively, which suggested that the translational fusions were functional (S6 Fig). Co-localization of FloA and PhoR signals was detected by fluorescence microscopy. Likewise, we also detected co-localization of the FloT and ResE signals (Fig 6D). Colocalization of PhoR and ResE with their respective flotillin was detected in all cells examined (Pearson’s correlation coefficients R2 = 0.82 and R2 = 0.85, respectively). These results suggest that FloA-PhoR and FloT-ResE are spatially correlated and support our hypothesis that FloA-PhoR and FloT-ResE physically interact. The specific interaction detected between PhoR and ResE sensor kinases and their respective flotillins was explored in further experiments to better understand how the scaffold activity of bacterial flotillins physically influences the activity of their signaling partners. The most direct hypothesis is that scaffold proteins facilitate signal transduction through tethering of signaling partners, because they enforce proximity and increase the likelihood of their interaction [44]. Thus, we investigated the effect of increasing concentrations of the scaffold flotillins on the interaction and activity of PhoR and ResE. PhoR and ResE belong to the PhoPR and ResDE two-component systems (TCS), which comprise a receptor histidine kinase and their cognate response regulator (PhoP and ResD). Histidine kinases are activated by forming homodimers, autophosphorylate and generate a phosphotransfer reaction to their response regulators. First, we generated a BTH assay to quantitatively monitor the homo-dimerization of PhoR and ResE (Fig 7A). This assay was complemented with a pSEVA modulable vector system [45], to generate different strains that produced lower, medium and higher levels of their respective flotillins that were further validated by immunoblotting (Fig 7A). These strains were used to quantitatively monitor the homo-dimerization efficiency of PhoR and ResE kinases with different concentrations of FloA and FloT, respectively. Both PhoR and ResE kinases responded similarly to increasing concentrations of their respective flotillins. A slight improvement in their interaction efficiency was observed with lower concentration of flotillins, which improved with medium concentration of the flotillins. Importantly, the BTH assay that produced higher concentration of the flotillins showed a decrease in the interaction efficiency of both kinases. This is consistent with the typical limitation of scaffold proteins, in that higher concentrations of the scaffold titrate signaling partners into separate complexes, thus inhibiting their interaction [46] (Fig 7B), as it has been experimentally shown in the scaffold protein Ste5 in yeast [47] and the JIP1 scaffold human cells [48]. This suggests that bacterial flotillins act as scaffold proteins to specifically facilitate signal transduction through tethering of signaling partners.
To investigate the influence of flotillins in the activation of PhoPR and ResDE TCS, we performed qRT-PCR analysis to quantify the transcription of genes which expression is strongly controlled by PhoP and ResD regulators (Fig 8A and 8B). We detected that the expression of the PhoP-regulated genes glpQ and tuaB involved in cell envelope metabolism [49,50] were reduced in a strain lacking the kinase PhoR and a strain lacking FloA. Likewise, the expression of the ResD-regulated gene sboX, responsible for the production of the antibiotic subtilosin [51], and yclJ, a gene that encodes for a regulatory protein [52] was reduced in a strain lacking the kinase ResE and a strain lacking FloT (Fig 8A and 8B). Control strains producing tagged versions of the cognate regulators (PhoP-3xFlag and ResD-3xFlag) showed comparable level of the regulators among the different strains, suggesting that the deletion of the respective flotillin specifically affects the activity of each cognate regulator, which in turn inhibits the expression of regulated genes.
Activation of the cognate regulators promotes a conformational change that impacts gene expression. Thus, the protein-protein interaction experiments were coupled to an in-depth analysis of the transcriptional profile of B. subtilis cells lacking floA or floT genes. The ΔfloA and ΔfloT mutants were grown to stationary phase. Total RNA was purified and used to perform microarray analysis using whole-genome B. subtilis genechips. Experiments were performed in triplicate and genes were considered differentially expressed when ≥2 fold in expression was detected in all replicates. Our microarray analysis indicated 123 genes to be differentially expressed (S4–S6 Tables and GEO database accession number GSE47918). 77 of these genes belong to different signaling regulons of B. subtilis, which were organized in a Voronoi treemap (Fig 8C). Each sector of the Voronoi treemap represents a gene and is labeled with the name of the gene that it represents. Each section is labeled in a two-color code to denote upregulated genes (in green) and downregulated genes (in red). There is no biological significance associated with the different shapes that are assigned to each sector. The magnitude of the fold change can be examined in supplemental S4–S6 Tables. This categorization revealed a group of genes whose expression depended on floA expression and a second group whose expression depended on floT expression. For instance, cells lacking floA showed induction of a large number of genes related to cell envelope metabolism, represented by sigM and yhdl, yhdK, yfml and csbB and ytrGABCDEF sigM-induced genes [53]. Additional genes related to cell wall reorganization were also detected (ytgP, dnaA, scpA and scpB), including tagAB and tagDEFGH operons, which are known of being repressed by PhoPR. Cells lacking floT displayed a strong inhibition of the genes that constitute the ResDE regulon (qcrABC, ykuNOP, dhbABCEF, hmp, nasDE and sboXA-albABCDEFG) [54]. Their expression is particularly prominent at stationary phase, when the production of the antibiotic subtilosin (sboXA-albABCDEFG) [51] and the siderophore bacillibactin (dhbABCEF) [55] is necessary. To validate the results obtained by microarray analysis, we performed qRT-PCR gene expression analysis on several genes that belong to the different regulons that are represented in the Voronoi treemap. qRT-PCR analysis showed comparable results to microarray analysis (Fig 8D).
The differential regulation of gene expression that is caused by the activity of FloA and FloT was manifested at the physiological level. We detected phenotypic differences in the ΔfloA and the ΔfloT mutants that may be related to the different expression of the controlled genes. For instance, when mutants were grown in Fe2+-containing growth medium, only the ΔfloA mutant accumulated the extracellular red pigment pulcherrimin (Fig 9A), resulting from the condensation of Fe2+ with the dipeptide pulcherriminic acid (Leu-Leu) (abs 420 nm) [56]. Pulcherriminic acid accumulates and is released into the medium in response to an excess of amino acid residues that decorate peptidoglycan precursors of bacterial cell wall synthesis, which is usually indicative of a defective cell wall metabolism [56–59]. This suggest that ΔfloA mutant is defective in cell wall turnover and is consistent to our proteomic and transcriptomic analyses, suggesting that FloA plays a role in the regulation of cell wall metabolism. Moreover, the ΔfloA mutant showed reduced sensitivity to the antibiotic vancomycin (Fig 9B), similar to other cases in which reduced sensitivity to vancomycin has been observed in cell-wall deficient strains. Vancomycin binds to the C-terminal D-Ala-D-Ala sequence of the pentapeptide peptidoglycan, thereby preventing the integration of peptidoglycan subunits into the cell wall. Cells that show a defective peptidoglycan turnover also show a reduced number of targets to the action of vancomycin and therefore, reduced sensitivity to the action of this antibiotic [60–62]. However, a defective cell wall often implies a less efficient barrier against the diffusion of other antibiotics [63–65], as is the case of the membrane pore-former sublancin [66]. Accordingly, the ΔfloA mutant shows a higher sensitivity to the glycopeptide sublancin [67].
Likewise, we tested the capacity of the ΔfloA and ΔfloT mutants to adapt to stress-related conditions that are typically associated with cultures that undergo stationary phase. When we grew the ΔfloA and ΔfloT mutants under oxygen-limiting conditions, only the ΔfloT mutant displayed a defective growth (Fig 9C). In contrast, the ΔfloA mutant was able to grow at similar rate to the wild-type strain. The incapacity of the ΔfloT mutant to adapt to oxygen-limiting conditions could be attributed to a defective activation of the ResDE regulon, as the activation of this regulon is necessary to allow nitrate respiration and thus, cell growth in oxygen-limiting conditions. Our data shows that this mechanism seemed defective only in the ΔfloT mutant, which grew poorly in oxygen-limiting conditions (Fig 9D). This is consistent with the role that FloT plays in the regulation of stationary phase and stress-related cellular processes, including the activation of the ResDE regulon, which we have detected in our proteomic and transcriptomic data. Taken together, Fig 9E shows a tentative model that integrates our proteomic, transcriptomic and physiological data. This model shows how FloA and FloT scaffold tether distinct signal transduction pathways, which ultimately control different cellular processes in B. subtilis. Furthermore, this model illustrates how functionally different FMMs regulate different genetic networks in a bacterial cell, which leads to the activation of different physiological processes.
There is growing recognition of the importance of eukaryotic lipid rafts in numerous cellular processes as diverse as protein sorting, membrane trafficking, compartmentalizing signaling cascades or pathogen entry [2,68]. This functional diversity is currently attributed to a different lipid and protein composition of lipid rafts, as it is hypothesized that a heterogeneous population of lipid rafts could exist on cellular membranes specialized on different biological processes [3–5]. Yet, the molecular mechanisms by which cells generate and regulate raft heterogeneity are still unclear. Nevertheless, it is assumed that cells likely regulate the process of raft diversification, to avoid the assembly of membrane signaling platforms that could simultaneously send distinct and conflicting signals to the cell. Here we use a bacterial model to show that B. subtilis cells are able to diversify FMMs into distinct families of signaling platforms, which are specialized in regulating distinct cellular processes, supporting the current hypothesis that a heterogeneous population of functionally specialized microdomains could exist on cellular membranes.
The discovery of the existence of FMMs adds to other examples of compartmentalization of macromolecules in bacteria, which demonstrate that bacteria are sophisticated organisms with an intricate cellular organization [69,70]. The biological significance of bacterial FMMs could be similar to the role of lipid rafts in eukaryotic cells. One possible function of FMMs could be the generation of a specific microenvironment to protect certain biological processes from inadequate conditions and non-specific interactions. For instance, spatial separation of signal transduction pathways may benefit their interaction specificity. Another plausible role for FMMs is to serve as platforms that control the assembly of membrane-bound protein complexes. By accumulating functionally related proteins in subcellular compartments, the likelihood of interaction increases and thus protein-protein interactions can be efficiently organized in space and time [2,11]. This phenomenon is facilitated by the activity of flotillins, which are FMMs-localized scaffold proteins that coordinate the physical assembly of protein interaction partners [44].
FloA and FloT seem to behave like other scaffold proteins that were described in eukaryotic cells, by specifically tethering signaling partners at lower concentrations or titrating, and thereby inhibiting their interaction at higher concentrations [44,46–48]. We show in this report that FloA and FloT self-interact and distinctively distribute within the FMMs of B. subtilis. Furthermore, FloA and FloT bind to and facilitate the interaction of different protein components and thus, activate different signal transduction cascades. The main force involved in generating raft heterogeneity is the uneven spatio-temporal distribution of two distinct flotillins FloA and FloT, Similarly, there are two flotillin paralogs in metazoans, FLO-1 and FLO-2, which show differential expression in distinct tissues, suggesting that these proteins may display certain level of specialization in scaffolding distinct cellular processes [71]. Based on this, it is possible that distinct families of lipid rafts may exist in the membrane of eukaryotic cells as well, yet this hypothesis still needs to be experimentally addressed.
Why do cells need or use different rafts? Cells may use this strategy to deliberately activate diverse cellular processes in time to ultimately dictate cell fate [3,72]. Here we show an example in which FMM remodeling occurs during bacterial growth using differential regulatory programs for flotillin expression. While FloA is constitutively expressed, the expression of FloT is restricted to stationary phase. Bacteria could use this mechanism to restrict the assembly and activation of particular protein components to stationary phase. Furthermore, the expression of a different scaffolding protein at stationary phase could help to rapidly adapt the signal transduction networks to face new environmental conditions. Bacteria possibly use this strategy to deliberately activate diverse cellular processes in time to ultimately ensure an effective activation of signaling processes during the lifespan of a bacterium [3,72].
Cells control the expression of each flotillin to restrict their expression to the growth stage in which their functionality is necessary. FloA preferentially tethers protein components associated with cell wall turnover and primary metabolism. Consequently, the ΔfloA mutant shows a defect in cell wall turnover. In contrast to FloA, FloT is responsible for tethering protein components that are related to adaptation to stationary phase, such as production of siderophores and antibiotics. In addition to this, we found several proteins associated with the FMMs that interact with both FloA and FloT and are related to biofilm formation and sporulation (see Fig 4). An example of this is the membrane-bound protease FtsH that is required for biofilm formation and sporulation [42], which has been shown to interact with FloA and FloT [13,16,17], as we confirmed in this report. Based on these results, it is likely that the ΔfloA ΔfloT double mutant shows additional and more pleiotropic defects in signal transduction than the ΔfloA and ΔfloT single mutants [17]. Likewise, a pleiotropic defect in cell division and biofilm formation has been associated with the overproduction of both FloA and FloT, which is not observable with the overproduction of either FloA or FloT separately [16].
The differential distribution of flotillin within lipid rafts opens additional questions as to whether other structural components of the lipid rafts, like for instance the constituent lipids, show a different spatio-temporal distribution pattern and thus, may also contribute to raft heterogeneity. All these questions were hindered by the difficulty to characterize subcellular structures in the past. However, the development of recent technologies is changing our knowledge about the structure and function of subcellular structures, including lipid rafts [73,74]. The development of super-resolution microscopes and corresponding data analysis methods may well ease the study of bacteria and offer a tractable model to study the role of membrane microdomains, which is rather complicated in their eukaryotic counterparts. The finding that bacteria organize membrane microdomains functionally and structurally equivalent to lipid rafts represents a remarkable level of sophistication in the organization of bacterial signaling networks that allow prokaryotes to amplify and integrate diverse stimuli. Overall, the spatio-temporal organization of signaling networks in bacteria evidences that bacteria are more complex organisms than previously appreciated.
Bacillus subtilis undomesticated wild type NCIB 3610 was used as parental strain in this study [23]. Escherichia coli DH5α and B. subtilis 168 strains were used for standard cloning and transformation procedures. A full strain list is shown in S1 Table. Selective LB agar was supplemented with antibiotics at final concentrations of: ampicillin 100 μg/ml; spectinomycin 100 μg/ml; erythromycin 2 μg/ml and lincomycin 25 μg/ml, tetracycline 5 μg/ml; chloramphenicol 3 μg/ml; kanamycin 50 μg/ml. When required, surfactin (Sigma, USA) was added from a stock solution to a final concentration of 5 μM. To maintain B. subtilis cells at exponential phase, cells were grown in shaking liquid LB cultures at 37° C overnight. Liquid LB medium was inoculated with 1:100 volume of the overnight culture and grown to OD600nm = 0.3 with vigorous shaking (200 rpm). To prolong growth at exponential phase, cells were repeatedly passed to fresh LB medium. Passaging was performed when cells reached OD600nm = 0.3. We repeated this procedure as described in [24] for approximately 20 generations prior to cell examination. To search for regulatory proteins that control the expression of floA and floT genes, the collection of mutants harboring the PfloA-yfp and PfloT-yfp transcriptional reporters were grown overnight in LB medium at 37°C with continuous agitation (200 rpm). After this, 2 μl of the overnight LB culture was spotted on MSgg agar plates and colonies were allowed to grow at 30°C for 72 h. Images were taken on a Nikon SMZ 1500 Zoom Stereomicroscope equipped with an AxioCam color (Zeiss, Germany). To monitor gene expression, YFP reporter signals were detected using a 520/20 excitation and BP535/30 emission filter. The excitation time was set to 5 s. Unlabeled wild type strain was used as negative control to determine the background.
Deletion mutants were generated using long flanking homology PCR [75] (using the primers listed in S3 Table). Markerless gene deletions were used to generate the ∆floA, ∆floT and ∆floA ∆floT mutants. Upstream and downstream regions of the floA and floT genes were joined by long flanking homology PCR [75] and cloned into the vector pMAD [76]. Gene deletion occurs via a sequential process of double recombination. Isolation of the mutants was achieved by counterselection, as described in [76]. The strains harboring the PfloA-GFP and PfloT-GFP transcriptional reporters were generated by cloning the promoter region of floA and floT into the vector pKM003 containing the gfp gene and integrating the constructs into the bacterial genome at the amyE locus. The vector pKM003 was kindly provided by Dr. David Rudner (Harvard Medical School, USA).
Translational fusions were constructed by long flanking homology PCR and subsequently cloned into pDR183 or pKM003. The vector pDR183 was kindly provided by Dr. David Rudner. These plasmids allowed the integration of the constructs into the bacterial genome at the lacA and amyE locus, respectively. Unless specified in the body of the paper, the translational fusions were expressed under the control of their natural promoters. When overexpression of FloA or FloT was necessary (e. g. BN-PAGE), floA and floT genes were cloned in the pDR111 plasmid under the expression control of an IPTG-inducible promoter Php [77–79]. The constructs were integrated into the bacterial genome at the amyE locus. Linearized vectors were added to B. subtilis 168 cells grown in competence inducing conditions. Double recombination occurred at the amyE locus when using the plasmids pDR111 and pKM003 or the lacA locus when using the plasmid pDR183. Cells were plated on corresponding selective media and colonies were checked for integration of constructed fusions by colony PCR. Utilizing the same strategy, GFP translational fusions of the kinases ResE and PhoR were generated using the vector pSG1154 and placed under the expression control of a constitutive promoter. SPP1 phage transduction was used to transfer constructs from B. subtilis 168 to wild type NCIB 3610, according to [80].
Site-directed mutagenesis of the EA C-terminal repeats of FloA and FloT was performed by using an overlap extension PCR. We used an adaption of the protocol that is published in [81]. Complementary primers that harbored the desired mutation were generated and used to amplify floA and floT genes in combination with outer primers (S2 Table). Two DNA fragments resulted from each gene were subsequently joined to one single fragment using long flanking homology PCR. The resulting gene was further sequenced to confirm the presence of the mutation. Mutations replaced glutamic acid by leucine and alanine by glycine in each specified EA repeat. The resultant variants were fused to GFP or mEOS2 and cloned into pDR183 under the expression control of their own promoter. This allowed the integration of the constructs into the bacterial genome at lacA locus by a single event of double recombination.
Cells were collected from the cultures by centrifugation, resuspended in 500 μl paraformaldehyde (4%) and incubated 7 min at room temperature to effect fixation. Samples were then subjected to three washing steps and resuspended in PBS buffer. Samples were finally mounted on microscope slides with thin agarose pads (0.8% agarose in PBS). Variations of growth conditions and preparation methods are specified in figure legends. Images were taken on a Leica DMI6000B inverted microscope. The microscope is equipped with a Leica CRT6000 illumination system, a HCX PL APO oil immersion objective with 100 x 1.47 magnification, a Leica DFC630FX color camera and an environment control system. The following filters were used to detect fluorescence signals: BP480/40 excitation filter and BP527/30 emission filter to detect GFP, BP546/40 excitation filter and BP600/40 emission filter to detect mCherry. GFP and mCherry fluorescence was observed by applying excitation times between 100 and 200 ms, while transmitted light images were taken at 36 ms exposure. Leica Application Suite Advanced Fluorescence V3.7 was used to process raw data and fluorescence signals were deconvoluted using AutoQuant software (MediaCybernetics). Further processing of images and calculation of Pearson’s correlation coefficient were performed using ImageJ. To calculate Pearson’s correlation coefficient, we selected 200 cells that simultaneously expressed both kinase and flotillin signals. Pearson’s correlation takes into consideration PhoR/ResE clusters and estimate whether flotillins clusters colocalize with them.
PALM was performed as described elsewhere [32]. Briefly, we used an inverted microscope (Olympus IX-71) equipped with an oil-immersion objective (60x, NA 1.45; Olympus) [82]. A 405 nm diode laser (Cube 405–100C, Coherent, USA) was used for converting mEOS2 from the green to the red fluorescent state, and a 568 nm laser (Sapphire 568 LP; Coherent, USA) was used for excitation of the converted state. A dichroic mirror (FF580-FDi01-25x36, Semrock, USA) in the excitation path and two emission filters (ET 575 LP, Chroma and FF01-630/92, Semrock, USA) in the detection path were used to image the fluorescence light with an electronmultiplying CCD camera (EMCCD; Ixon DU897, Andor, USA). A pixel size of 106 nm was achieved by using additional lenses. About 20000 frames were recorded with a frame rate of 10 or 20Hz at an excitation intensity of 5kW/cm² (568nm) until no mEOS2 signals could be detected any further. For photoconversion, the 405nm laser was pulsed with a frequency matching the frame rate, with a pulse duration between 1 and 50ms at irradiation intensities of <1 kW/cm². The PALM image stacks were analyzed with the open source software rapidSTORM [83,84], version 2.21. Only single spot events with more than 250 photons were used for image reconstruction. mEOS2 protein fluorescent in consecutive frames was summarized with the "track emissions" filter of rapidSTORM in order to be localized only once and improve localization precision.
Cluster analyses were performed by an in house written python routine (python 2.7.3, Python Software Foundation). The position of one mEOS2 fluorophore was determined as a single localization according to our software rapidSTORM by fitting a Gaussian function to the Point Spread Function. Clusters were defined by either one connected pixel area in image-based analysis or by a cloud of scattered localizations with spatial coherence in localization. Spatial coherence implies that the increased local density of localizations follows a Gaussian distribution within the cluster, which is indicative of the nonrandom distribution of localizations. Using the untracked localization raw data set and the corresponding super-resolved image, a mask was generated to define possible cluster candidates and separate them from the localization pseudo background. A nearest neighbor based global density threshold was applied to assist the separation process, i.e. all localizations exhibiting a nearest neighbor distance above 50 nm were pre-discarded. According to the mask, the tracked localization data set was then filtered by cropping single cluster candidates and rejecting those with just two or less remaining tracked localizations. Cluster diameters where determined by calculating the standard deviation of the localization cloud from its center of mass. The stated cluster diameters represent the FWHM, which was derived from the standard deviation.
B. subtilis strains harboring translational fusions were grown overnight in LB agar. 3 ml liquid MSgg was inoculated with cells from the overnight culture and grown to stationary phase (OD600nm = 3.5). Next, cells from 1 ml MSgg culture were collected by centrifugation and resuspended in PBS buffer. To disperse cells, sonication was applied in three series of 10 pulses (power output 0.72 / cycle 50%). Finally, cells were diluted 1:200 in PBS buffer and used for analysis. Flow cytometry experiments were conducted using a benchtop MACSQuant Analyzer (Miltenyi Biotech, Germany). Single cells were detected in two scatter channels (FSC, SSC) and one fluorescence channel (B1). Cells were excited by the blue laser (488 nm) coupled to a 488/10 nm filter and detected as size (FSC) and granularity (SSC) signals. GFP fluorescence (B1) was detected by excitation of cells with the blue laser (488 nm) coupled to a 525/50 nm filter. The voltage intensity of channels was set as follows: forward scatter channel (FSC) 265 V, sideward scatter channel (SSC) 410 V and B1 channel 450 V. The number of events measured per sample was 50,000. We used a flow rate of 1,500 to 3,000 events per second. No gates were selected in any experiment. Flow cytometry data was processed with FlowJo 9.4.3 software.
To compare the differential transcript levels in the ΔfloA, ΔfloT single and ΔfloT ΔfloA double mutants, cells were grown in MSgg medium until the late exponential phase and their transcriptome was compared to that of the wild type strain NCIB 3610. Up- and downregulated genes are listed in S3–S5 Tables (Bayes p value ≤ 1.0 x 10–3). The microarray data has been validated for various genes using quantitative RT-PCR experiments (Fig 8). The isolation of total RNA, cDNA synthesis, hybridization, scanning, data normalization has been performed as described previously using three independent biological replicates for each strain [85]. Briefly, pellets were frozen in liquid nitrogen and stored at -80°C. RNA extraction was performed with the Macaloid/Roche protocol [85,86], RNA concentration and purity was measured using NanoDrop ND-1000 Spectrophotometer. RNA samples were reverse transcribed into cDNA using the Superscript III reverse transcriptase kit (Invitrogen, USA) and labeled with Cy3 or Cy5 monoreactive dye (GE Healthcare, The Netherlands). Labeled and purified cDNA samples (Nucleospin Extract II, Biokè, The Netherlands) were hybridized in Ambion Slidehyb #1 buffer (Ambion Europe Ltd) at 48°C for 16 h. The arrays were constructed according to [87]. Briefly, specific oligonucleotides for all 4,107 open reading frames of B. subtilis 168 were spotted in triplicate onto aldehyde-coated slides (Cell Associates) and further handled using standard protocols for aldehyde slides. Due to the array design, the transcript levels of the plasmid-encoded genes of B. subtilis 3610 are not determined. Slide spotting, slide treatment after spotting and slide quality control were done as before [88]. After hybridization, slides were washed for 5 min in 2x SSC with 0.5% SDS, 2 times 5 min in 1x SSC with 0.25% SDS, 5 min in 1x SSC 0.1% SDS, dried by centrifugation (2 min, 2.000 rpm) and scanned in GenePix 4200AL (Axon Instruments, USA). Fluorescent signals were quantified using ArrayPro 4.5 (Media Cybernetics, USA) and further processed and normalized with MicroPrep [89]. CyberT [90] was used to perform statistical analysis. Genes with a Bayes P-value of ≤ 1.0 x 10–4 were considered significantly affected.
Quantitative PCR experiments were performed as described before [91]. Gene classification was adapted from [92,93]. Data processing in Voronoi treemap was performed with TreeMap software (Macrofocus GMbH, Switzerland). RNA samples obtained as described above for the microarray experiments were treated with RNase-free DNase I (Thermo Fisher Scientific, Germany) for 60 min at 37°C. Reverse transcription was performed with 50 pmol random nonamers on 2 μg of total RNA using RevertAidTM H Minus M-MuLV Reverse Transcriptase (Thermo Fisher Scientific, Germany). Quantification of cDNA was performed on an iQ5 Real-Time PCR System (BioRad, USA) using Maxima SYBR Green qPCR Master Mix (Thermo Fisher Scientific, Germany). We performed 8 replicates reactions per gene analyzed. The primers used are listed in S2 Table. The amount of target cDNA was normalized to the level of girB cDNA [94].
To overexpress the His-tagged version of flotillins (S1 Table), cells from a freshly streaked LB agar plate were used to inoculate 100 ml liquid MSgg medium supplemented with 1 mM IPTG. Cultures were incubated overnight at 37°C with agitation (200 rpm). Cells were collected by centrifugation, resuspended in buffer H [95] containing 1 mM PMSF and lysed in a French pressure cell at 10,000 psi. Cell debris was removed by standard centrifugation at 12,000 x g for 10 min. Membranes were isolated from the supernatant by ultracentrifugation at 100,000 x g, 4°C for 1 h. Pellets containing membranes were carefully resuspended in solubilization buffer A, supplemented with 10% glycerol and 1 mM PMSF protease inhibitor [37]. Samples were subjected to one step of shock freezing in liquid nitrogen and thawed on ice. Membranes were solubilized using 1% dodecyl maltoside (DDM—Glycon Biochemicals, Germany) and prepared for blue native PAGE as described by [37]. To separate protein complexes, samples were mounted on a 4–20% Roti-PAGE gradient gel (Carl Roth, Germany) and blue native PAGE was run for 3 h at 15 mA. Native gels were used for standard immunoblotting procedures without further processing. After blotting, PVDF membranes were destained to eliminate Coomassie staining and washed with TBS-T. His-tagged flotillins were detected using a polyclonal anti-His antibody (MicroMol, Germany). Bands of native complexes that contained the proteins of interest were cut and analyzed by mass spectrometry (LC/MS).
Peptides identified by LC/MS were aligned to the B. subtilis proteome by using MASCOT Peptide Mass Fingerprint software http://www.matrixscience.com. The protein libraries used for peptide alignment were Uniprot-Swissprot http://www.uniprot.org/, NCBInr http://www.ncbi.nlm.nih.gov/protein, EST-EMBL http://www.ebi.ac.uk/ena/, and Subtilist http://genolist.pasteur.fr/SubtiList/. Alignment conditions were restricted to significance threshold p < 0.05 and ions score cut-off = 15. Protein mass was unrestricted, peptide mass tolerance was 10 ppm and the fragment mass tolerance was 0.02 Da. We systematically discarded proteins that were identified with less than 20% of amino acid coverage or MudPIT score below 700. Even if peptides match randomly, it is possible to obtain multiple matches to a single protein. MudPIT score is a Poisson distribution that defines a threshold to discard random matches based on the ratio between the number of spectra and the number of entries in the database. For MudPIT results, the score for each protein is the amount of peptides (peptide abundance) that are above the threshold.
floA and floT genes as well as the versions of these genes with altered EA-repeats and kinase genes phoR and resE were PCR-amplified (using primers specified in S1 Table) and cloned into the pKNT25 or pUT18 plasmids (EuroMedex, France). Each one of these plasmids contains a gene that encodes for one of the two catalytic domains of the adenylate cyclase from Bordetella pertussis (referred to as T25 and T18 catalytic domains). The genes of interest were cloned and C-terminally fused to the T25 and T18 encoding genes. Plasmids were propagated in the E. coli BTH101 strain (S2 Table). Positive control was the two oligomers of the leucine zipper GCN4, which are fused into the pKT25-zip and pUT18C-zip plasmids respectively. This was provided by EuroMedex. An E. coli strain harboring pKNT25 and pUT18 plasmids was used as a negative control (S2 Table). Protein-interaction assays were performed following the protocol previously described by Karimova et al. [33]. Experiments that required LB plates 100 μg/ml Ampicillin, 50 μg/ml Kanamycin and 40 μg/ml X-Gal were incubated 48h at 30°C. The appearance of blue product indicated protein interaction that can be monitored and quantified. Quantification of Miller units was performed to monitor the efficiency of protein interactions according to [96].
To assay the scaffold activity of FloA and FloT, the kinase genes phoR and resE were PCR-amplified and cloned into the pKNT25 and pUT18 plasmids. phoR and resE were C-terminally fused to the T15 and T18 encoding genes and propagated in E. coli BTH101 strain. Protein-interaction assays were performed following the protocol previously described by Karimova et al. [33] to determine the interaction efficiency between PhoR-PhoR and ResE-ResE. These strains were subsequently used to clone pSEVA modulable plasmids [45] that produce different levels of FloA and FloT. We specifically used pSEVA-621, pSEVA-631 and pSEVA-641 plasmids to produce FloA and FloT at different concentrations. These plasmids contain distinct replication origins and propagate in E. coli at low, medium and high copy number, respectively. This generates low, medium and high concentration of FloA and FloT in the in bacterial two-hybrid E. coli strains in which the plasmids are propagated. Experiments that required the propagation of pSEVA vectors were performed in LB medium with 100 μg/ml Ampicillin, 50 μg/ml Kanamycin and 10 μg/ml Gentamicin. Quantification of the Miller units was performed to monitor the efficiency of protein interactions, as it is described in [96].
Pulcherriminic acid was estimated by using a method adapted from [56,97]. Cell samples containing pulcherrimin were washed twice with methanol and once with distilled water before extraction with 2M NaOH. The amount of pulcherriminic acid that was converted to the sodium salt turned yellow and could be determined spectrophotometrically as a specific peak in absorbance at 410 nm. To assess the sensitivity of wild-type and flotillin mutant strains to vancomycin, MSgg was supplemented with 0.2 μg/ml Vancomycin. Similarly, strains were grown in LB medium supplemented with 0.2 μg/ml Vancomycin in a Tecan Infinite 200 Pro Microplate Reader at 37°C with agitation (200 rpm). Growth was measured at OD = 595nm. The growth curves shown are a mean of three independent experiments. To grow B. subtilis in anaerobic conditions (i.e. nitrate respiration), MSgg medium was modified by replacing glutamate with sodium nitrate (0,2%) and glycerol with glucose (1%). Liquid cultures were grown in 2.5 L seal chambers containing Oxoid CampyGen (5% O2) or anaerobic atsmosphere generation bags (Oxoid, UK).
Gene Expression Omnibus (GEO): Accession database number GSE47918
http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?token=fxkbdeiukswcgxs&acc=GSE47918
|
10.1371/journal.pgen.1005819 | Strong Components of Epigenetic Memory in Cultured Human Fibroblasts Related to Site of Origin and Donor Age | Differentiating pluripotent cells from fibroblast progenitors is a potentially transformative tool in personalized medicine. We previously identified relatively greater success culturing dura-derived fibroblasts than scalp-derived fibroblasts from postmortem tissue. We hypothesized that these differences in culture success were related to epigenetic differences between the cultured fibroblasts by sampling location, and therefore generated genome-wide DNA methylation and transcriptome data on 11 intrinsically matched pairs of dural and scalp fibroblasts from donors across the lifespan (infant to 85 years). While these cultured fibroblasts were several generations removed from the primary tissue and morphologically indistinguishable, we found widespread epigenetic differences by sampling location at the single CpG (N = 101,989), region (N = 697), “block” (N = 243), and global spatial scales suggesting a strong epigenetic memory of original fibroblast location. Furthermore, many of these epigenetic differences manifested in the transcriptome, particularly at the region-level. We further identified 7,265 CpGs and 11 regions showing significant epigenetic memory related to the age of the donor, as well as an overall increased epigenetic variability, preferentially in scalp-derived fibroblasts—83% of loci were more variable in scalp, hypothesized to result from cumulative exposure to environmental stimuli in the primary tissue. By integrating publicly available DNA methylation datasets on individual cell populations in blood and brain, we identified significantly increased inter-individual variability in our scalp- and other skin-derived fibroblasts on a similar scale as epigenetic differences between different lineages of blood cells. Lastly, these epigenetic differences did not appear to be driven by somatic mutation—while we identified 64 probable de-novo variants across the 11 subjects, there was no association between mutation burden and age of the donor (p = 0.71). These results depict a strong component of epigenetic memory in cell culture from primary tissue, even after several generations of daughter cells, related to cell state and donor age.
| Regenerative medicine specialists have been using a type of cell commonly found in the skin called the fibroblast because it is easily obtained from skin samples, grows well in culture, and can be manipulated in the laboratory to de-differentiate into a primordial state known as the induced pluripotent stem cell. These primitive stem cells can then be transformed into mature tissues, such as liver or pancreas cells. Here we show that fibroblasts, coming from different locations in the same individual, vary significantly in epigenetic marks called DNA methylation, which are involved in the regulation of gene expression. In addition to location-specific patterns of DNA methylation, we also find that fibroblasts from different anatomical locations respond differently in epigenetic patterns related to aging. As the field of regenerative medicine advances, our study demonstrates that deciding upon the source of fibroblasts from an individual to generate new tissues and organs may be an important consideration.
| DNA methylation (DNAm) at CpG dinucleotides plays an important role in the epigenetic regulation of the human genome, contributing to diverse cellular phenotypes from the same underlying genetic sequence. For example, DNAm levels at particular genomic loci can accurately classify different tissues [1] and even underlying cell types within tissues [2]. These stable cell type- and tissue-discriminating loci appear to represent only a subset of "dynamic" CpGs, approximately 21.8%, actively involved in regulation of gene expression [3]. Changes in these epigenetic patterns across aging have been extensively studied [4], particularly in large studies of whole blood [5–7], but subsets of these age-associated CpGs appear tissue-independent [8].
These epigenetic barcodes also play an important role in cellular reprogramming (the conversion of somatic cells to pluripotent stem cells), a powerful and promising experimental system in biology, genetics and personalized medicine [9]. This epigenetic reprogramming of somatic cells to induced pluripotent stem cells (iPSCs) induces demethylation [10] followed by specific patterns of subsequent DNA methylation that can reflect the original somatic tissue [11]. Fibroblasts are one of the most popular cell types for generating iPSCs [12], particularly from skin, given the relative ease of access to these cells, although other skin-derived cell types such as keratinocytes from the same individual generate similar iPSC lines [13]. Skin, however, is perhaps the most susceptible tissue source in the body to environmentally induced insult, particularly through sunlight and chemical exposures, which can induce changes in epigenetic patterns [14]. The epigenetic “memory” of source tissue for iPSC characterization has been well characterized [11].
In our previous work, we successfully cultured fibroblast lines from the dura mater of postmortem human donors, a source location largely protected from environmental insult with slowly dividing cells [15]. We compared these cultured fibroblast lines to those derived from scalp samples from the same individuals, and found that the rate of culture success was higher for dura-derived fibroblasts; in some cases only the dura fibroblasts from an individual would culture. While the resulting cultured cells from these two sampling locations were largely morphologically indistinguishable (see Figure 1 in Bliss et al, 2012 [15]), we hypothesized that increased culture success might have a strong epigenetic component. Previous research has shown that dermal fibroblasts from different locations in the body have distinct gene expression profiles [16], including compared to some non-dermal sources [17], and previous reports have indicated that cultured cells have largely stable epigenomes, with the exception of a small number of loci [18]. We therefore sought to characterize the methylomes and transcriptomes of fibroblasts from these two sampling locations–scalp and dura–from donors across the lifespan.
Here we identify several components of epigenetic “memory” in cultured fibroblasts after multiple passages (i.e. splitting and continuing to grow) where primary tissue originated from two locations in the body. The strongest epigenetic memory was related to sampling location in the body, as we identified widespread DNAm differences at local and regional spatial scales preserved through identical culturing processes. We further find increased stochastic epigenetic variability in cultured fibroblasts from the scalp compared to dura. This increased variability manifested in significant increased quantitative pairwise methylome-wide distances in a combined analysis with publicly available DNAm data on skin fibroblasts [19], pure cell populations from peripheral blood [20], and cells from the dorsolateral prefrontal cortex [21]. Another component of epigenetic memory was related to the age of the donor, including a subset of CpGs that displayed location-dependent changes through aging. The epigenetic differences between these fibroblasts appear to occur largely through epigenetic-dependent mechanisms, as there were few differences in coding sequence across the fibroblasts from the two locations within the same individual. These results demonstrate the effect of epigenetic memory in cultured fibroblasts by sampling location and donor age in morphologically indistinguishable cells.
We measured DNA methylation (DNAm) levels from scalp- and dura-derived cultured fibroblasts in 11 postmortem donors (22 samples) from across the lifespan, ranging from early infancy to 85 years (S1 Fig, S1 Table), using the Illumina HumanMethylation450 microarray (Illumina 450k) [22]. After data processing, normalization, and quality control with the minfi package [23], we obtained normalized data on 21 samples (one dura sample with lower quality was removed prior to across-sample normalization) across 456,513 probes (probes with single nucleotide polymorphisms, SNPs, at the target CpGs or single base extension sites were removed, as were probes on the sex chromosomes, see Methods).
We first characterized differences in DNAm levels from cultured fibroblasts derived from different locations (scalp versus dura). Many probes, targeting individual CpGs, were differentially methylated between scalp- and dura-derived fibroblasts– 101,989 (22%) at genome-wide significance (false discovery rate, FDR < 5%, see Methods). These significant DNAm differences between cultured fibroblasts from the scalp and dura were large in magnitude, with 57,704 probes having differences in DNAm levels greater than 10%, and 23,752 with differences greater than 20% (Fig 1A). The directionality of these DNAm differences was balanced, with approximately equal proportions of CpGs showing increased versus decreased methylation in cultured fibroblasts from scalp compared to dura. These differentially methylated probes (DMPs) were widely distributed across the genome, as 18,551 genes (defined by UCSC knownGene database) had at least one DMP within 5 kilobases (kb), as did 33,247 transcripts (see Methods). These widespread single CpG differences manifest as the largest component of variability in the entire dataset, as the first principal component (Fig 1B, explaining 38% and 62.3% of the variability before and after surrogate variable analysis, SVA [24]) represents the sampling location of these cultured fibroblasts, suggesting a strong epigenetic memory of original cell location.
Since these differentially methylated CpGs tended to cluster in a smaller number of genes, we further identified 697 differentially methylated regions (DMRs) at stringent genome-wide significance (family-wise error rate, FWER < 10%)–these regions were identified based on adjacent probes showing directionally-consistent differences in DNAm > 10% between groups [25] (see Methods). For example, we identified a region of 24 contiguous probes hypermethylated in scalp-derived fibroblasts within the gene RUNX3 –a tumor suppressor that plays an integral role in regulating cell proliferation and the rate of apoptosis [26] (Fig 1C, see S2 Fig and S2 Table for all significant DMRs). Regional differences, particularly in CpG island shores, previously have been shown to better distinguish tissues and cell types [1] and correlate with neighboring gene expression levels [23] than individual CpGs. Unlike at the single CpG level, which had balanced directionality of differential methylation, the majority of DMRs had higher DNAm levels in fibroblasts derived from scalp compared to those derived from dura (N = 414, 59.4%). Using gene sets defined by biological processes [27], these neighboring genes (within 5 kb) were strongly enriched for morphogenesis (including morphogenesis of the epithelium), developmental processes, cell differentiation, and epithelium and connective tissue development, among other more general gene sets (all p < 10−8, S3 Table).
In addition to the extensive differential methylation at both the CpG and regional level, we identified 243 long-range regions with consistent significant methylation change (FWER < 10%), called “blocks” [28], using an algorithm adapted from whole genome bisulfite sequencing (WGBS) data to Illumina 450k [23]. A representative significant block is shown in Fig 1D (see S3 Fig for all significant blocks at FWER < 10%). Blocks have now been identified across many cancer types [29], and tend to associate with higher order chromatin structure including nuclear lamin-associated domains (LADs) [30] and large organized chromatin K9 modification (LOCKs) [28]. The 243 significant blocks in our data represent 41 Mb of sequence and contain 298 annotated genes. These blocks contain 41 of the significant DMRs that differentiate sampling location of the fibroblasts, and more interestingly, every block overlaps at least one “dynamic” cell/tissue DMR identified using WGBS data from Ziller et al (2013) [3].
While these cultured fibroblasts were several generations/passages removed from the primary tissue and morphologically indistinguishable, we nevertheless found widespread epigenetic differences by sampling location of the primary fibroblasts at varying spatial scales, suggesting a strong epigenetic memory of the original cell location.
We next sought to determine the functional correlates of the widespread epigenetic differences identified between scalp- and dura-derived fibroblasts by performing RNA sequencing (RNA-seq) on polyadenylated (polyA+) mRNA from the same cultured samples (see Methods). Briefly, we aligned the reads to the transcriptome using TopHat [31] and generated normalized gene counts (as fragments per kb per million mapped reads, FPKM) based on the Illumina iGenome hg19 annotation using the featureCounts software [32]. A median of 88.0% (interquartile range, IQR: 85.5%– 88.8%) of reads mapped to the genome, of which a median of 84.7% (IQR: 84.4%–85.5%) mapped to the annotated transcriptome (see S1 Table for sample-specific percentages). We identified 11,218 expressed genes with average FPKM expression greater than 1.0. Initial clustering of the gene FPKM values separated the fibroblast samples by location in the first principal component (PC), which explained 35.4% of the variance (S4 Fig), mirroring the first principal component of the DNAm data (Fig 1B). We could further cluster our samples by sampling location using a set of 337 genes (of which 210 were in our dataset) that were previously identified by Rinn et al [17] to group largely dermal fibroblasts by their anatomical sites of origin (S5 Fig)–these genes better clustered the samples by sampling location than random sets of 210 genes (p<0.001, see Methods). Differential expression analysis of the RNA-seq data, independent of the results from the epigenetic analyses above, identified many genes that differed by the source of the primary fibroblast– 5,830 genes at FDR < 5%. Both scalp- and dura-derived fibroblasts expressed high levels of Fibroblast Specific Protein-1 (FSP-1) and this gene was more highly expressed scalp-derived fibroblasts (fold change = 5.5, FDR = 5.6x10-6) in line with increased higher proliferation rates in the scalp-derived versus dura-derived fibroblasts [15]. The differentially expressed genes were strongly enriched for signaling and cell communication, cell proliferation, apoptotic processes, and epithelium development and morphogenesis via gene ontology (GO) analysis (all p < 10−8, S4 Table)–these gene sets were similar, and much more significant, to those identified comparing gene expression profiles across positional-identity genes in dermal fibroblasts [17].
We next used the gene expression data as a functional readout of the differentially methylated loci identified between fibroblasts cultured from scalp and dura. The majority of significant DMPs (76,971/101,989, 75.47%) were inside or near (within 5kb of) a UCSC annotated gene, and 28.2% (21,742/76,971) were significantly associated with gene expression levels (at p < 0.05). This percentage of DMPs with significant expression readout was elevated (34.9%) among those DMPs with larger DNAm differences by sampling location (greater than 10% difference in DNAm levels). These DMPs were strongly significantly enriched among the CpG sites that associated with expression levels at the p < 0.05 (48,062 probes within 5kb of genes, odds ratio, OR = 3.99, p < 2.2x10-16) and FDR < 0.05 (6,559 probes within 5kb of genes, OR = 19.54, p < 2.2x10-16) significance thresholds. Surprisingly, we found that the DNAm levels at the majority of these expression-associated differentially methylated CpGs tended to be positively associated with gene expression, regardless of overall methylation levels (un-, partially-, or highly-methylated) or their location in the gene (islands, shores and shelves)–these biases towards positive associations were statistically significant for many of these comparisons (see S5 Table, panels A and B). We hypothesize these positive correlations could be due to the probe design of the Illumina 450 (the majority of probes are in lowly methylated regions) combined with the majority of genes having low expression (38.75% had mean FPKMs < 1).
We identified similar associations using transcript-level expression data using the Sailfish program [33] (see Methods) on the above transcriptome– 76.5% (77,981/101,989) of the DMPs were within 5 kb of a transcript, and 30.4% of them (23,672/77,981) correlated with expression (at p < 0.05). 33,247 unique transcripts overlapped or were within 5 kb of DMPs, and of them, 27.0% (8,981/33,247) exhibited significant correlation between DNAm and expression (at p < 0.05). The 33,247 transcripts proximal to the DMRs corresponded to 18,699 genes, the majority of which (84.3%, 15,761/18,699) contained more than one transcript. Interestingly, these associations often appear in a transcript-specific manner—6,190 genes (39.3%) had ≥ 1 transcript with significant correlation between DNAm and expression (at p < 0.05), with ≥ 1 transcripts that were not associated with nearby CpG levels. These results suggest that genes, and their underlying transcripts, can functionally validate many of the differentially methylated CpGs for sampling location.
Moving beyond individual CpGs, 587/697 (84.2%) DMRs were in or near (<5kb) genes, and many had DNAm levels that were significantly associated with gene expression levels (306/587, 52.1% at p < 0.05). For instance, a DMR overlapping an intronic sequence of the SIM1 gene (Fig 2A) was unmethylated with low corresponding expression of the gene in the cultured fibroblasts from dura, and highly methylated with corresponding high expression levels of the gene in the scalp-derived fibroblasts (Fig 2B and S2 Table). This is in line with previous reports suggesting that gene body methylation levels positively associate with local gene expression [34], unlike CpG island shore methylation that tends to be negatively associated with gene expression levels [1]. Of the 478 unique genes in or within 5kb of DMRs, the expression of 235 (49.2%) of them was significantly correlated with DNAm (p < 0.05). These 235 unique genes tended to exhibit stronger differential expression between the scalp- and dura-derived fibroblasts (median fold change = 1.59, IQR = 1.23–2.68) than individual CpG results, in line with previously published findings [23]. GO analysis on expression-associated genes proximal to DMRs revealed enrichment for multiple important biological processes such as connective tissue development, epithelium morphogenesis and development, cell differentiation (specifically including epithelial cell differentiation), and cell proliferation (including epithelial cell proliferation), among other more general sets (all p < 10−8, see S6 Table). Unlike at the single CpG-level, we found that the majority of DMRs in and around the transcriptional start sites of genes (CpGs islands and shores) were negatively correlated with gene expression (S5 Table), in line with previous research [1]. We observed similar methylation-expression associations using transcript-level expression measurements– 312/599 DMRs (52.1%) near ≥ 1 transcripts associated with expression, and like at the single CpG level, found evidence for transcript-specific epigenetic regulation of expression (among 28.9% of genes containing multiple transcripts and associated with DNAm levels within the DMRs).
Lastly, we found that the majority of differentially methylated blocks contained at least one gene and transcript differentially expressed between scalp- and dura-derived fibroblasts. The majority of blocks contained at least one gene (N = 188/243, 77.4%); 63.8% (N = 120/188) had at least one gene and 66.66% (N = 124/186) at least one transcript that was differentially expressed (at p < 0.05).
As a representative example, one of the blocks, hypermethylated in scalp-derived fibroblasts, overlaps the HOXB gene cluster (Fig 3A), which has previously been shown to be play a role in the position identities of fibroblasts [17]. In this block, expression levels of the HOXB genes are significantly greater in fibroblasts cultured from scalp than those from dura (Fig 3B), which contrasted previous microarray-based data showing these genes were not expressed in dermal samples taken from the head [17] highlighting the improved precision of RNA-sequencing data to quantify expression levels. Similarly, the 188 significant blocks contained 298 unique genes, and 126 of them (42.3%) were differentially expressed (at FDR < 0.05) which is a higher proportion than the rest of the transcriptome (0.42 vs. 0.32, p = 3.79x10-9).
Given the strong association between DNAm levels and local expression levels, we sought to more fully examine the epigenetic states of these sampling location-associated DNAm differences. We downloaded chromatin state data (18 states) from the NIH Roadmap Epigenomics Consortium on the four available fibroblast samples (2 primary foreskin, 1 adult dermal, and 1 lung) [35], and mapped our DMPs, DMRs, and blocks for fibroblast sampling location onto these states (S7 Table). The CpGs differentially methylated by sampling location were largely enriched for enhancer chromatin states, including preferential enrichment of genic (EnhG2) and active (EnhA1) enhancer states and depleted for active transcriptional start site (TSS) states (TssA). At the region level, DMRs were largely enriched for bivalent TSS (EnhBiv) and repressive polycomb (ReprPC) states and depleted for transcription (Tx) genic enhancer (EnhG2) states, and blocks were strongly enriched for quiescent (Quies) and heterochromatin (Het) states and depleted for transcriptional states. These enrichments were relatively conserved across the four Roadmap fibroblast samples, further suggesting distinct epigenetic states in scalp- compared to dura-derived fibroblasts. These results suggest that epigenetic memory related to original cell location manifests in genomic state differences and largely reads out in the transcriptome, particularly among regional changes in DNAm related to fibroblast sampling location.
We hypothesized that scalp-derived fibroblasts might have more variable levels of DNAm than dura-derived fibroblasts, given the chronic exposure to environmental factors (e.g. sunlight, chemicals) in the primary tissue across the lifespan. At the individual CpG level, we tested for differences in variance between the scalp- and dura-derived fibroblasts independent of the underlying mean methylation levels [36] (see Methods section). While only two probes reached genome-wide significance (at FDR < 0.05) for differences in variance, at marginal levels of significance (p < 0.05), fibroblasts cultured from scalp had more variable DNAm levels than fibroblasts cultured from dura (N = 13,169/16,330, 80.6%).
We next sought to characterize methylome-wide patterns of DNAm across these fibroblasts in the context of other diverse cell types. After downloading and normalizing Illumina 450k data from sorted blood [20] and frontal cortex [21], as well as skin-derived fibroblasts [19] and melanoma samples (SKCM) from the Cancer Genome Atlas (TCGA) [37], we computed methylome-wide Euclidean distances between and across each of the 11 cell types (see Methods section). We noted that these cell types largely cluster by tissue source (brain, blood, and fibroblasts in the first two principal components and largest dendrogram splits, S6 Fig).
The inter-individual epigenomic distances, and their variability, were much greater in the scalp-derived (as well as skin-derived) fibroblasts than dura-derived fibroblasts (p = 1.34x10-9 and p = 1.77x10-14 respectively, see Fig 4). The distances within scalp- and skin-derived fibroblasts were significantly larger than those calculated within pure blood and cortex cell types (p-values range from 1.04x10-21 to <10−100). Interestingly, the inter-individual distances between fibroblasts cultured from scalp samples were greater than the distances between different cell types within a blood cell lineage (e.g. natural killer cells versus CD4+ T-cells) which were previously suggested for different dermal fibroblasts [16] and instead more similar to distances across lineage (e.g. natural kill cells versus monocytes). Note that comparing inter-individual distance between two cell types (e.g. scalp- versus dura-derived fibroblasts) reflects the extensive differential methylation between these two cell types (see Fig 1)—the inter-individual distances are large but the variability in distances was low.
As another example, the distances across scalp-derived fibroblasts were lower than the inter-individual variability between neurons and non-neurons (via NeuN+ sorting), which reflects the extensive methylation differences between these two cell types [21]. As expected, we found the greatest methylome-wide distances and largest inter-individual variability in the melanoma samples [28,36], which highlights the relative scale of these methylome-wide distances (ranging from pure cell types to cancer). These increased epigenomic distances may relate to the rate of cell division, which is non-existent in neuronal cells [38] and infrequent in T-lymphocytes at the population level [39]. The increased epigenetic variability in the scalp samples was further not associated with differences in donor age (p > 0.05, S7 Fig), suggesting increased epigenetic stochastic variability in scalp- (and skin-) derived fibroblasts.
We hypothesized that a subset of this increased variability might result from age-related divergence in DNAm at individual loci that were differential by sampling location, such that young donors would have lesser difference in DNAm levels, and older donors would have larger differences in DNAm. By fitting linear models on the difference in DNAm levels across sampling location as a function of donor age (see Methods), we identified 7,265 CpGs associated with diverging DNAm levels across aging (at FDR < 10%, S8 Fig). These loci appeared to be clustered into representative patterns of their age-related changes (Fig 5). The majority of these CpGs had significant age-related changes in fibroblasts derived from the scalp (64.0%), but not dura (17.4%), and the magnitude of change across age was larger in scalp-derived fibroblasts–the average change in percent DNAm per decade of life was 3.13% (IQR = 1.81%-4.29%) in fibroblasts derived from scalp compared to 1.13% (IQR = 0.295%-1.61%) in those from the dura mater.
A subset of these CpGs showing sampling location-dependent age-related changes associated with nearby gene expression levels. Most of the probes (N = 5,185/7,265, 71.4%) were annotated to 3,553 unique genes (within 5kb) and 21.8% of these (N = 775/3,555) showed significant correlation between DNAm and gene expression (p < 0.05). These DNAm associated genes were enriched for multiple general developmental processes including cell development, morphogenesis, and differentiation (all p<10−8, S8 Table). Several of the age-related CpGs showing expression association were within genes that are involved in cell proliferation and apoptosis. For instance, DNAm levels at two significant probes inside the gene TEAD1, which regulates notochord development and cell proliferation [40], were significantly associated with gene expression levels (p = 8.60x10-4 and 0.045, respectively). Another significant DNAm-expression pair (p = 0.02) involved AVEN, a gene shown to inhibit Caspase activation in apoptosis [41]. Interestingly, while we identified a large number of age-related CpGs, “DNA methylation ages” [8] were very similar to the chronological ages of the samples (see Methods and S9 Fig)–these associations did not differ by sampling location (p = 0.72) and there was further no association between “DNA methylation age” and sampling location alone (p = 0.96). The age-associated CpGs identified here therefore suggest that altered regulation of DNAm levels across aging occurs primarily in fibroblasts derived from scalp but not from dura, perhaps through altered cell proliferation and apoptosis, and possibly reflecting greater exposure to environmental agents that can affect the methylome.
Lastly, we characterized the expressed sequences of the scalp- and dura-derived fibroblasts within each individual to examine the extent of genetic mosaicism, which may contribute to differences in DNAm through changing the underlying genetic sequence in the fibroblasts taken from scalp. De novo variants were called directly from the RNAseq data, and after filtering by many quality metrics (see Methods) we identified 64 high-confidence candidate variants that were discordant by sampling location in at least a single individual (S9 Table), including 22 annotated coding variants (13 synonymous and 9 non- synonymous) [42]. We found no association between coding variant burden and subject age (p = 0.71, S10 Fig). These results suggest that many of the location- and age-associated DNAm differences are not due to somatic mosaicism and likely arise through epigenetic mechanisms that are maintained through cell culture and multiple passages.
Here we interrogated the methylomes and transcriptomes of pairs of fibroblasts cultured from scalp and dura mater taken from the same individual, in a subject cohort that ranges in age across the human lifespan. These cultured fibroblasts, generations removed from the primary tissue of origin, and with indistinguishable morphology, still maintained strong components of epigenetic “memory” related to sampling location (scalp versus dura) and differential changes in DNAm levels across aging. The widespread differences in DNAm levels by sampling location were identified at many spatial scales, including single CpGs, differentially methylated regions, blocks, and globally. Furthermore, many of these differences in DNAm levels manifested in the transcriptome, showing significant corresponding differences in expression for genes most proximal to these epigenetic changes. The genes with differences in expression and DNAm levels by sampling location were previously implicated in processes relating to cell proliferation and apoptosis, which likely relate to the function of the fibroblasts in the primary tissue. One might have predicted this outcome, as fibroblasts in the scalp, including those that are cultured, turnover much more rapidly than those in the dura mater [15], which we confirmed here with increased FSP-1 expression in the scalp-derived fibroblasts.
Another component of epigenetic memory in these cultured fibroblasts was related to ages of the donors, where age-related changes occurred differentially by sampling location. These age-associated loci can be clustered into general patterns of epigenetic changes by age and location, all showing significant interaction between donor age and sampling location. While some patterns were expected, such as divergence in DNAm levels from similar levels at birth (clusters 1, 4, 5, and 7), several other clusters showed an unexpected convergence in DNAm across aging (clusters 2 and 3). We do note that the elderly donor (age 85) is influential in both the statistical discovery at individual loci and in some of the subsequent clusters–larger sample sizes can hopefully further define and replicate these observations. Also, while the fibroblasts were analyzed from some subjects with psychiatric disorders, almost all comparisons between scalp and dura sampling locations, and differential changes with age were naturally matched within an individual, reducing the potential impact of diagnostic confounding. Furthermore, a larger sample size would likely identify significant age-related divergence in DNAm at the region level–while we found 7,265 individual CpGs, we found very few DMRs at global significance (6 and 11 DMRs at FWER ≤ 10% and 20% respectively). The region-finding approach has been shown to be statistically conservative [25] and the identification of these differential age-related changes by sampling location was based on number of donors (N = 10), not the number of observations (N = 21). Lastly, while proliferation rates were not measured for these particular fibroblast samples, analyses in a much larger skin biopsy sample (N = 298) showed no association between proliferation rates and donor age [43], which was our sampling location with the greater number of age-related changes in DNAm levels.
These age-related changes in cultured fibroblasts are one of the first examples, to our knowledge, of genome-wide significant age-related changes in a pure cell population that is many mitoses and passages from the original donor cells. Many papers have identified widespread age-related changes in heterogeneous cell populations, like blood [5,7], brain [44], and other tissue types [8], which may result in false positives when the underlying cellular composition changes across aging [4]. Other papers have used individual cell populations to validate age-associated loci identified in homogenate tissue at marginal significance [45] or have identified age-related changes in targeted approaches at limited numbers of loci [46].
Similarly, these fibroblasts cultured from the scalp and dura mater were the first example, again to our knowledge, of morphologically indistinguishable cells with vastly different epigenomic profiles. Using epigenomic distances, these two cohorts of fibroblasts were more different in their DNAm patterns than different lineages of blood cells, while less different that neuronal versus non-neuronal cells from the frontal cortex (Fig 4); the cells underlying each comparison have very different morphologies and cellular function. Furthermore, the majority of differences in DNAm levels between scalp- and dura-derived cultured fibroblasts appeared to be determined early in development, prior to early infancy in this sample, and remained stable throughout the lifespan. Of the 101,989 significant DMPs for sampling location, 98,461 (96.5%) were not associated with differential age-related changes. These findings demonstrate strong components of epigenetic memory related to cell location and aging in fibroblasts cultured from the scalp and dura mater from postmortem human donors.
There are important implications from this study for the field of regenerative medicine. If fibroblasts are going to be the source for iPSCs, and ultimately differentiated tissues, the source of these fibroblasts, and their epigenetic characteristics, may be an important consideration. For example, these differences in cellular states in cultured fibroblasts may relate to the number of cell divisions, as skin and scalp fibroblasts have a much quicker turnover than fibroblasts in the dura [15]. The extent of cell division could relate to the epigenomic distances between and across the diverse cell types we have analyzed. Analyses in larger samples of skin biopsy-derived fibroblasts suggest that while donor age does not appear to associate with proliferation rates of fibroblasts, the cultured cells derived from younger donors reprogrammed more readily [43], which presumably has a strong epigenetic component. Further research may better determine the extent of epigenetic memory of cell state of fibroblasts cultured from different locations after the generation of iPSCs and subsequent differentiation into new cell types. As the field of regenerative medicine advances, our study demonstrates that deciding upon the source of fibroblasts from an individual to generate new tissues and organs may be an important consideration. While it was shown that transcriptional variability by tissue of origin was low in iPSCs (13), it was also demonstrated that the DNAm landscape in iPSCs differs greatly by tissue or origin, and this phenomenon may explain the propensity of iPSCs derived from different somatic tissues to differentiate into different lineages (11).
Human dural and scalp fibroblasts on which the methylation and gene expression studies were performed were obtained from fibroblast lines derived from human post mortem scalp and dura mater tissues. For this study, tissues from 11 individuals were used, with the ages of individuals ranging from 0.1 to 85 years of age (see S1 Table for additional demographics). The post-mortem tissues from 2 of the subjects were collected by the Lieber Institute for Brain Development (LIBD) and the tissues from the remaining 9 subjects were collected by National Institute for Mental Health (NIMH) (Clinical Brain Disorders Branch (CBDB), Division of Intramural Research Programs (DIRP)). The NIMH tissues were collected from two medical examiners (Washington, DC office and Commonwealth of Virginia, Northern District office); the LIBD tissues were obtained the Office of the Chief Medical Examiner (Baltimore, MD). A preliminary neurological or psychiatric diagnosis was given to each case after demographic, medical, and clinical histories were gathered via a telephone screening on the day of donation. For each case, the postmortem interval (PMI) (the time (in hours) elapsed between death and tissue freezing) was recorded. (See S1 Table for PMIs and demographics for every subject used in this study). Every case underwent neuropathological examinations to screen for neurological pathology. Additionally, the medical examiner’s office performed toxicology analysis of every subject’s blood to screen for drugs.
Dura and scalp tissue were collected at the time of autopsy. From the autopsy room, the tissues were transported in separate bags: one containing cerebral dura mater and the other a 1 in X 1 in scalp segment with hair attached. Both bags were transported on wet ice to the lab, where the culture procedure was immediately started.
The dura culture medium was prepared out of 1X DMEM (Ref#11960–044, GIBCO) with 10% by volume fetal bovine serum, 2% by volume 100X GlutaMAX (Cat#: 35050, GIBCO), 1% by volume Penicillin-Streptomycin/Amphotercin solution (Ref# 15140–122, GIBCO), and 1% by volume Gentamicin solution (Cat# 17105–041, Quality Biological). This culture medium was used in all subsequent steps of the dura culturing procedure. The scalp culture medium used for all subsequent steps of the scalp culturing procedure was made the same way except without the 1% Gentamycin. A rinsing solution was prepared out of 1X PBS (pH 7.2) (Ref# 21-040-CV, Corning Life Sciences), 1% by volume Penicillin-Streptomycin/Amphotericin solution (Ref# 15140–122, GIBCO), and 1% by volume Gentamicin (Cat# 17105–041, Quality Biological).
The dissected scalp sample was washed with the rinsing solution three times, the fat tissues were cut away, and all hair was plucked out with forceps. The scalp sample was then placed epidermis side down on a dish and floated with Dispase II enzyme solution (2.4 units of the Dispase II enzyme per mL of PBS, Dispase II enzyme: Cat#17105–041, GIBCO). (Dispase II enzyme is a proteolytic enzyme used to separate the dermis from the epidermis by cleaving the zone of the basement membrane.) The dish was covered with parafilm and foil, and placed in a 37°C incubator for 24 hours. After the 24-hour period, the epidermis was peeled away from the dermis. The dermis was washed with the rinsing solution, dried, and cut into 2–3 mm2 pieces. The pieces were placed in a Falcon Easy Grip tissue culture 35×10 mm dish and one drop of scalp culture medium was added to each piece of scalp. The dish was placed in the incubator at 37°C and 5% CO2 for culturing.
A similar procedure was followed for the dura samples. Dura samples were washed with the rinsing solution three times. Then, a few 2–3 mm2 pieces were cut from the dura mater and placed together in an Easy Grip cell culture 35×10 mm dish. One drop of dura culture medium was added to each dura piece. The culture dish was then placed in an incubator (at 37°C and 5% CO2) for culturing. The medium of each culture was changed to fresh medium 2–3 times per week to promote growth of the fibroblasts. On average, fibroblast cells started to proliferate at 7–14 days, however some samples took longer (up to 3 weeks).
The dura and scalp tissue cultures were monitored under a phase-contrast microscope. When the fibroblast growth reached 90–95% confluence, 1 mL of a 0.25% trypsin solution (Cat#T4049, Sigma) was added to each culture dish, and the cells were incubated for 5 to 8 min. Then, 1mL of media was added to each dish stop the enzymatic reaction. Next, the contents of each culture dish were transferred into separate 15 mL Falcon conical tubes and 8mL of media was added to each tube. The conical tubes were centrifuged for 5 min at 1100 rpm. The supernatant was discarded, 5mL of fresh media was added to each conical tube, and the contents of the tubes were transferred onto separate 25 cm3 cell culture Easy Flasks (Thermo Scientific, Cat# 156367), where they were kept in cultures for 3–5 days in an incubator (at 37°C and 5% CO2). When the cells reached 90–95% confluence, the cells from each 25 cm3 flask were transferred onto two 75 cm3 cell culture easy flasks (Thermo Scientific, Cat# 156499) and kept in cultures for continued growth. When the cells reached 90–95% confluence, they were incubated with 3 mL of 0.25% trypsin solution for 5 to 8 min, after which 3mL of fresh culture media was added to stop the enzymatic reaction. Then, the contents of the flasks were transferred into separate 15 mL Falcon conical tube and 4mL of media was added to each tube. The tubes were centrifuged (5 min, 1100 rpm), the supernatant was discarded and the pellets containing the fibroblasts were removed from the centrifuge tubes and transferred to cryoTube vials (Cat#375418, Thermo Scientific). 0.5 mL of recovery cell culture freezing medium (Cat#12648–010, GIBCO) was added to each vial, after which the vials were insulated with Styrofoam and placed into a -80°C freezer. Later, the tubes were transferred to a -152°C liquid nitrogen freezer.
These frozen dura and scalp fibroblast cells were then used generate DNA methylation and gene expression levels. Genomic DNA was extracted from approximately 3 million cultured human fibroblast cells using the AllPrep DNA/RNA/miRNA Universal Kit (Qiagen). Bisulfite conversion was performed on 600 ng genomic DNA was done with the EZ DNA methylation kit (Zymo Research).
DNA methylation landscapes of the dura- and scalp-derived fibroblasts were analyzed using the Illumina HumanMethylation 450 BeadChip array (“450k”). The 450k array interrogates >485,000 DNA methylation sites (probes) and measures the proportion DNA methylation at each target site (the 450k array interrogates both CpG and CH sites). The microarray preparation and scanning were performed in accordance with the manufacturer’s protocols. The resulting data from the 450k consists of R(ed) and G(reen) intensities using two different probe chemistries [22], which we converted to M(ethylated) and U(nmethylated) intensities using the minfi Bioconductor package [23], version 1.14.0 using with R version3.2. One dura sample had lower median probe intensities and was removed prior to normalization and downstream analyses. After quality control (QC), the M and U intensities were normalized separately across samples using stratified quantile normalization [23]. Probes containing common SNPs (based on dbSNP 142) at the target CpG or single base extension site, and probes on the sex chromosomes were removed, leaving 456,513 probes on 21 samples for analysis.
We determined differential methylation using linear modeling on the normalized DNAm levels, using the model:
yij=αi+βiLocj+ζiSVsj+εij
(1)
where yij is the normalized proportion methylation at probe i and sample j, αi is the proportion methylation in the fibroblasts sampled from the dura mater, βj is the difference in methylation in the scalp-derived fibroblast, and Locj is the sampling location represented by a binary variable (Dura = 0, Scalp = 1). These statistical models were adjusted for surrogate variables (6 SVs) estimated using surrogate variable analysis (SVA) [24].
Differentially methylation probes (DMPs) were identified by fitting Eq 1 to each probe, and obtaining the corresponding moderate t-statistic and p-value using the limma package [47]. P-values were adjusted for multiple testing using the false discovery rate (FDR) [48] and significant probes were called were FDR < 0.05. Principal component analysis (PCA) was performed after regressing out the surrogate variables from the DNAm levels of each probe, preserving the effect of fibroblast sampling location. Finding differentially methylated regions (DMRs) involves identifying contiguous probes where β ≠ 0 using the bumphunter Bioconductor package (version 1.6.0) [25], here requiring |β| > 0.1 (argument: cutoff = 0.1) and assessing statistical significance using linear modeling bootstrapping with 1000 iterations (argument: nullMethod = ‘bootstrap’ and B = 1000). DMRs were called statistically significant when the family wise error rate (FWER) ≤ 0.1. We identified blocks using the same model as above using the blockFinder function in the minfi package [23], which collapses nearby CpGs into a single measurement per sample, and then fits Eq.1 above, only here j represents probe group, not probe. Here we again required at least a 10% change in DNAm between groups and assessed statistical significance using the FWER based on 1000 iterations of the linear model bootstrap.
RNA was extracted from the cultured dura and scalp fibroblasts with the RNeasy kit (Qiagen), in accordance with the manufacturer’s protocol. RNA molecules were treated with DNase, polyadenylated (polyA+) RNA was isolated, and resulting sequencing libraries were constructed using the Illumina TruSeq RNA Sample Preparation Kit (v2) and sequenced on an Illumina HiSeq 2000. We note that while all samples were run on the same flow cell, the samples were somewhat imbalanced by lane–however, the first PC of the expression data did separate perfectly by sampling location. Sample-specific information on reads and alignments are available in S1 Table.
Resulting reads were mapped to the genome using TopHat2 [31] using the paired-ends procedure (we used the option—library-type fr-firststrand). Gene counts relative to the UCSC hg19 knownGene annotation were calculated using the featureCounts script of the Subread package (version 1.4.6) [32]. There were 23,710 genes in this annotation, and we dropped 305 genes that were annotated to more than 1 chromosome. Of the remaining 23,405 genes, 18,316 genes had non-zero expression counts in at least one sample. Counts were converted to FPKM (fragments per kilobase per million reads mapped) values to allow comparisons across genes with different lengths and libraries sequenced to different depths. These FPKMs were transformed prior to statistical analysis: log2(FPKM + 1). The log transformed FPKM values were used in all subsequent gene-level analyses.
Next, we used the Sailfish software (33), version 0.7.6, to quantify isoforms from our RNA-seq reads. As a result, we obtained TPM (transcripts per million) values for each isoform, which we log transformed: log2(TPM + 1). The log transformed TPM values were used in all subsequent transcript-level analyses.
Differential expression for sampling location was identified using Eq 1 above, where yij represents transformed expression (rather than DNAm) levels, and different SVs (N = 4) were calculated from the expression data.
To test whether we could use a subset of genes to cluster our fibroblasts by sampling location, like reported by Rinn et all [17], we took the 337 genes published by the authors, which they found to group fibroblasts by anatomical location. Of these 337, we used only 210 genes, since a subset of the tabulated genes did not contain gene symbols, another subset was not interrogated by our RNAseq, and yet another subset was not expressed in any of our samples. We then perfumed Euclidean distance computations and clustering analysis by first using these 210 genes and then repeating the analysis 1000 times using 210 randomly chosen genes.
We carried out gene ontology analysis on the differentially expressed genes with the GOstats package [49]. Transformed FPKMs were next used to assess functional significance of differentially methylated features. We mapped the DMPs to genes in the UCSC knownGenes (hg19) and determined which DMPs exhibit correlation between DNAm and gene expression with the MatrixEQTL package [50]. We used Pearson's Chi-squared test with Yates' continuity correction to examine whether DMPs are more likely to exhibit correlations between DNAm and gene expression than non-DMPs. We then mapped significant DMRs to genes expressed in the RNA-seq data (e.g. showing non-zero expression levels in ≥ 1 samples), and correlated the average DNAm level within the DMR to the transformed expression level. When multiple genes were within or near a DMR, we retained the gene (and its correlation) with the largest absolute correlation. We carried out gene ontology analysis for the genes proximal to DMRs with the GOstats package. For each significant block, we found the UCSC annotated gene(s) containing within the block and their evidence for differential expression as calculated above. We used Pearson's Chi-squared test with Yates' continuity correction to test whether differentially expressed genes were enriched in blocks compared to the rest of the transcriptome. Finally, we analyzed the directionality of DNAm—expression correlations for DMPs and DMRs, as a function of DMR/DMP positions relative to genes. We used the binomial test to access the significance of distributions between positive and negative correlations of DNAm and gene expression.
In addition to gene-level analysis, we studied transcript-level expression and its correlation with DNAm. We carried out the same analysis for isoform expression as for gene-level expression, with the exception that here we used relative isoform abundance values that we obtained with the Sailfish software (see above).
The 18-chromatin state data, derived using hidden Markov models (HMMs), was obtained for 4 fibroblast samples: samples E055 and E056 (foreskin primary fibroblasts), E126 (adult dermal fibroblast), and E128 (lung fibroblsts) in the Epigenome Roadmap project22 (http://egg2.wustl.edu/roadmap/web_portal/chr_state_learning.html). The chromatin states overlapping DMPs, DMRs, and blocks were obtained, and compared to a background of all 450k probes, considered probe groups, and collapsed probe groups respectively. Overlap was assessed based on the total coverage (in base pairs) of the chromatin states. Fold changes for enrichment of > 1.5 fold were highlighted. Prior to carrying out the enrichment analysis, the sex chromosomes and the mitochondrial chromosome were dropped.
We performed a second larger data processing and normalization procedure on our scalp- and dura-derived fibroblasts after adding data from skin fibroblasts (GSE52025) [19], pure populations of blood [20] and prefrontal cortex cells [21] from the FlowSorted.Blood.450k and FlowSorted.DLPFC.450k Bioconductor packages respectively, and then melanoma data from TCGA [37]. The M and U channels were combined across all experiments and then normalized with stratified quantile normalization as described above. We then dropped the probes on the sex chromosomes as well as probes that are common SNPs (based on dbSNP 142) as described above. Within the normalized data, we then calculated all pairwise Euclidean distances on the proportion methylation scale, and selected specific comparisons to display in Fig 4.
We calculated differential variability between scalp and dura CpG DNAm levels using the Levene test [51] and subsequent p-values were adjusted for multiple testing using the FDR. We filtered out the 101,989 genome-wide significant probes showing mean methylation differences by sampling location, as there is a strong mean-variation relationship in DNAm data due to being constrained within 0 and 1 (e.g. gaining methylation from an unmethylated state or losing methylation from a fully methylated state increases variance).
We tested for probes that showed differential age-related divergence in DNAm by fibroblast sampling location. First, we calculated the difference in DNAm between scalp- and dura-derived fibroblasts from the same individual at every probe (creating a 456,513 probe by 10 individual matrix). We then computed 3 surrogate variables (the number estimated by the SVA algorithm) for a statistical model with donor age, and fit the following linear model:
Δyij=γi+δiAgej+ζSVsj1+εij
(2)
where Δyji is the difference in DNAm levels between scalp and dura for probe i and individual j, γi is the difference in DNAm levels at birth, Agei is the age of the donor, and δi is the change in the difference of DNAm per year of life. We then generated a Wald statistic and corresponding p-value for δi and adjusted for multiple testing via the FDR. Post hoc age-related changes, e.g. the change in DNAm levels per year of life, were calculated within the scalp and dura samples. We then associated expression of nearby genes (within 5kb) with the DNAm levels at the probes showing significant age-by-location effects and performed gene ontology on the significant genes with the GOstats package [49]. We lastly computed the “DNAm age” of our scalp and dura samples using the R code published by S. Horvath, (available at https://labs.genetics.ucla.edu/horvath/dnamage/) and fit a linear model containing main effects of biological age and sampling location, and an interaction term between these two variables on “DNAm age”.
We called expression variants directly from the RNA sequencing alignments using samtools (version 1.1) and mpileup across all samples [52]. We then filtered variants in the resulting variant call format (VCF) file based on coverage (<20), variant distance bias (p<0.05), read position bias (p<0.05), mapping quality bias (p<0.05), base quality bias (p<0.05), inbreeding coefficient binomial test (p<0.05), and homozygote bias (p>0.05). The resulting 64 high quality variants were annotated with SeattleSeq138 [42].
For every subject from whom the post-mortem tissues were collected, informed consent was obtained verbally from the legal next-of-kin using a telephone script, and was both witnessed and audiotaped, in accordance with the IRB approved NIMH protocol 90-M-0142 and the Department of Health and Human Services for the State of Maryland (protocol # 12–24).
DNA methylation data in both raw and processed forms are available on the Gene Expression Omnibus (GEO): GSE77136. RNA sequencing reads (raw data) are available on the Sequencing Read Archive (SRA): SRP068304 (BioProject: PRJNA286856) and the genes and transcript counts (processed data) are available on GEO at the above accession number (GSE77136).
|
10.1371/journal.pcbi.1003557 | Influence of Wiring Cost on the Large-Scale Architecture of Human Cortical Connectivity | In the past two decades some fundamental properties of cortical connectivity have been discovered: small-world structure, pronounced hierarchical and modular organisation, and strong core and rich-club structures. A common assumption when interpreting results of this kind is that the observed structural properties are present to enable the brain's function. However, the brain is also embedded into the limited space of the skull and its wiring has associated developmental and metabolic costs. These basic physical and economic aspects place separate, often conflicting, constraints on the brain's connectivity, which must be characterized in order to understand the true relationship between brain structure and function. To address this challenge, here we ask which, and to what extent, aspects of the structural organisation of the brain are conserved if we preserve specific spatial and topological properties of the brain but otherwise randomise its connectivity. We perform a comparative analysis of a connectivity map of the cortical connectome both on high- and low-resolutions utilising three different types of surrogate networks: spatially unconstrained (‘random’), connection length preserving (‘spatial’), and connection length optimised (‘reduced’) surrogates. We find that unconstrained randomisation markedly diminishes all investigated architectural properties of cortical connectivity. By contrast, spatial and reduced surrogates largely preserve most properties and, interestingly, often more so in the reduced surrogates. Specifically, our results suggest that the cortical network is less tightly integrated than its spatial constraints would allow, but more strongly segregated than its spatial constraints would necessitate. We additionally find that hierarchical organisation and rich-club structure of the cortical connectivity are largely preserved in spatial and reduced surrogates and hence may be partially attributable to cortical wiring constraints. In contrast, the high modularity and strong s-core of the high-resolution cortical network are significantly stronger than in the surrogates, underlining their potential functional relevance in the brain.
| Macroscopic regions in the grey matter of the human brain are intricately connected by white-matter pathways, forming the extremely complex network of the brain. Analysing this brain network may provide us insights on how anatomy enables brain function and, ultimately, cognition and consciousness. Various important principles of organization have indeed been consistently identified in the brain's structural connectivity, such as a small-world and modular architecture. However, it is currently unclear which of these principles are functionally relevant, and which are merely the consequence of more basic constraints of the brain, such as its three-dimensional spatial embedding into the limited volume of the skull or the high metabolic cost of long-range connections. In this paper, we model what aspects of the structural organization of the brain are affected by its wiring constraints by assessing how far these aspects are preserved in brain-like networks with varying spatial wiring constraints. We find that all investigated features of brain organization also appear in spatially constrained networks, but we also discover that several of the features are more pronounced in the brain than its wiring constraints alone would necessitate. These findings suggest the functional relevance of the ‘over-expressed’ properties of brain architecture.
| The physical brain is a network of extraordinary complexity on multiple spatial scales. On the macroscopic scale, regions are connected by a large number of white-matter projections that form an intricate system: the connectome [1]. Understanding the principles of the large-scale architecture of the brain, how this architecture shapes brain dynamics to in turn support brain function and human behaviour, is a central challenge for contemporary neuroscience [2], [3].
Recent advances in non-invasive anatomical [4]–[6] and functional [7] imaging techniques, along with the development of automated, high throughput post-processing methods [8] now allow the application of complex network science as a principled and systematic framework for studying the connectome [2], [3]. Accordingly, numerous principles of organisation in the large-scale structural anatomy of the brain have been characterized, including small-world properties [9], hierarchical architecture [10], modular structure [11], the existence of a strong structural core [12] and a so-called ‘rich-club’ organisation [13]. Exposing both the structural origin and functional relevance of these properties of the human connectome is an essential, but difficult step towards a deeper understanding of the large-scale organisation of the brain.
A common approach to evaluating the significance of a particular network property, observed in a particular network, is via surrogate or null-hypothesis comparison [14], [15]. In this approach, a set of surrogate networks represents a null-hypothesis for the target network property by preserving some a priori chosen properties of the network under investigation, while randomizing other network properties. Quantitative comparison of the original network with the ensemble of surrogate networks allows drawing conclusions on the significance of the target property of the network with respect to those properties preserved in the ensemble. Therefore, in its essence, surrogate network comparison allows testing if some, usually very elementary, properties of the target network induce, or at least contribute to, the expression of some of its more global and complex network properties.
When choosing appropriate surrogate networks, the most widely used null-hypothesis properties are size (number of nodes), connection density (number of edges) and degree distribution (the number of connections of each node). This approach – which we term the ‘random surrogate’ approach – has illuminated the topological investigation of many abstract, spatially-unembedded networks, including the World Wide Web, semantic networks, food-webs, and gene-regulatory and metabolic networks [14], [16]. It is also routinely applied in the analysis of brain networks in order to demonstrate that global, ‘higher order’ network property of brain maps, such as modularity or ‘small-worldness’, cannot be attributed solely to these basic network properties [10], [11], [17].
Physical networks like the brain are, however, embedded into physical space and are therefore subject to additional constraints deriving from the costs of developing and maintaining connections [18] which are not conserved by random surrogates. Random surrogates therefore represent a rather loosely constrained null-hypothesis set for physical networks. Specifically, they tend to possess a large number of long-range connections because they ‘smooth’ local inhomogeneities of physical networks. They thus form highly and rather homogeneously integrated networks, while at the same time lacking the high topological segregation (locally dense, globally sparse inter-connectivity) associated with predominantly local connectivity, which is one of the most prominent features of brain networks [19]. When compared against random surrogates, then, certain properties of brain networks may appear to be highly distinctive even though they can be attributed to the spatial constraints of its embedding into the physical world (wiring cost) and/or of the economic pressure of minimising the number of the energetically expensive long-range connections (metabolic cost) [18].
To address this problem, so-called ‘lattice surrogates’ have been introduced [15], [20]–[22] to preserve (or rather increase) the high segregation of brain networks. The motivation behind lattice surrogates, originating from the Watts–Strogatz notion of ‘small-worldness’ [23], was to represent a lattice-like, topologically over-segregated (and thus under-integrated) surrogate network type, the opposite of random surrogates in a sense, and to compare the target network with these two extremes. This is reflected in the rule commonly used to generate lattice surrogates from the connectivity of a brain network (during a ‘random’ network rewiring process, edge swaps are only made if the nonzero entries of the resulting connectivity matrix are located closer to the main diagonal [15], [20]), which is only indirectly linked to physical distance through some arbitrary spatial ordering of the network nodes. For this reason, lattice surrogates are only partially appropriate as a null-hypothesis network set for physical wiring constraints of brain networks. Furthermore, lattice surrogates are designed to reduce, rather than preserve, network connection lengths thus further undermining their utility in assessing the effects of wiring constraints on cortical network properties.
In this paper, we introduce two new classes of surrogates, spatial surrogates and reduced surrogates, Like random surrogates, spatial surrogates preserve network size, connection density, and degree distribution, but (unlike random surrogates) they also preserve the wiring length distribution of the target network. Reduced surrogates are like spatial surrogates with the difference that they do not preserve but actually reduce overall network wiring, in similar way to traditional lattice surrogates, but in a spatially well-defined and controlled manner. We reasoned that in virtue of these properties, these surrogates provide improved baselines by which to assess the extent to which a target network property can be attributed to cortical wiring constraints [18].
This approach enables us to evaluate a number of prominent findings regarding the structural properties of the connectome (see Figure 1) with respect to the extent to which these properties are preserved in the novel spatial surrogates as compared to random and connection length optimised (reduced) surrogates. To ensure robustness we perform these analyses on both weighted and unweighted (binary), and on the full resolution (998 regions) as well as on a lowered resolution (66 regions) version of the cortical structural connectivity data set provided by Hagmann et al. [12]. Overall, the method allows us to distinguish those significant network properties of the connectome that are derivable from its predominantly local, spatially segregated connectivity (as indicated when both the cortical network and its spatial and reduced surrogates differ from random surrogates) from those that are the consequences of some other, primarily not (or not only) spatial, but potentially more functionally relevant organisation principle of cortical connectivity (as indicated when the cortical network differs from all of its surrogate groups).
Specifically, during the evaluation of each specific network property, the logic of our surrogate analysis is the following (see Table 1). We measure the expression of the network property in the cortical network and every surrogate group by an appropriate complex network metric. If all surrogate groups exhibit similar metric indices to that of the cortex, then the basic network properties preserved in all surrogates (the number of regions, number of white-matter projections and regional degree distribution of the cortical network) appear to be sufficient for the observed expression of the investigated network property. If, however, all spatially constrained networks (cortical network, spatial and reduced surrogates) exhibit similar values, but differ from random surrogates, we reason that cortical wiring constraints may account for the level of expression of that network property in the cortical network. Additionally, if the cortical network is more similar to spatial than to reduced surrogates, we reason that solely the presence of long-range connections in the cortex may facilitate the network property, irrespectively of the specific arrangement of these connections in the cortex. If, however, the cortical network is more similar to reduced than to spatial surrogates, then we reason that the predominantly local (short-range) connectivity of the cortex can account for the expression of the network property even in the absence of long-range cortical connections (as indicated by the similarity between the cortical network and reduced surrogates). In addition, this case also indicates that the particular arrangement of long-range cortical connections appears to be such that it does not interfere with (strengthen or hinder) the expression of the network property (as indicated by spatial surrogates, with randomised long-range connections, being different from both the cortical network and reduced surrogates). Finally, if the cortical network differs from every surrogate ensemble, we reason that the network property is specific to the particular connectivity of the cortex, it cannot fully be attributed to the topological properties and wiring constraints that are conserved in the surrogates, but instead may be a more functionally relevant organisation feature of cortical connectivity.
We use the cortical connectivity network of Hagmann et al. [12] (Figure 2). This data was obtained by non-invasive tracing of white-matter projections linking pairs of cortical sites in the brains of five human subjects, combining magnetic resonance imaging (MRI) and diffusion spectrum imaging (DSI) techniques, semi-automated brain parcellation, diffusion tractography and appropriate post-processing methods. The individual connectivity networks of the five subjects were aggregated into a single network in order to reduce the impact of inter-subject variability. The resulting dataset is a compact network representation of cortical grey matter regions as network nodes, and their connecting white-matter fibre bundles as edges. For a detailed description of the acquisition procedure and validation test results of the procedure, see the original paper and [8].
By the nature of its processing pipeline, the network consists of a two-level hierarchical parcellation of the cortex: it is composed of 66 anatomical regions at the higher level, and of 998 regions of interest (ROIs) at the lower level. Each node on the level of ROIs represents an area of the cortical surface of approximately 1.5 cm2 size (region), and there are a total of 17,865 undirected weighted connections between these regions. These figures result in a fairly sparse, 3.6% connection density network on the high-resolution cortical parcellation (i.e., on the lower hierarchical level of the segmentation). For the low-resolution network, similarly to [23], we calculate the strength of the connection between every two-region pair by summing the weights of all the high-resolution connections linking the ROIs that compose the two cortical regions. This method results in 574 aggregated white-matter fibre bundles between the 66 regions on the low-resolution parcellation, which increases the connection density of the low-resolution cortical connectivity to 26.8%.
While a few studies on high-resolution structural connectivity networks have appeared recently [e.g.], [ 12], [24,25], many earlier results, in particular those based on the data set used here, have relied on low-resolution data [e.g.], [ 26], [27,28]. Although focussing on low-resolution data allows comparing to earlier low-resolution studies on other brain networks [e.g.], [ 11,20], utilizing the information afforded by the available higher resolution connectivity may influence the outcome of complex network analysis [29], [30] and has the benefit of maximizing usage of the available information. Here, we primarily analyse the high-resolution, 998-node anatomical connectivity network (see Figure 2), but we also compare to lower resolution results where appropriate.
We employ three types of null-hypothesis networks, namely random, spatial and reduced surrogate networks. All three surrogate types preserve the size (number of nodes), connection density (number of edges) and degree distribution (the number of connections of each node) of the cortical network, and differ from each other only in their physical wiring constraints: random surrogates are spatially non-constrained, spatial surrogates preserve the total wiring length of the cortical regions (and thus that of the entire cortical network globally), and reduced surrogates possess reduced wiring lengths.
All three types of surrogate networks were generated by the widely applied iterative rewiring algorithm [14], [31], the basic version of which proceeds as follows: Starting from the original cortical network, in each iteration two connections, (r1, r2) and (r3, r4), are randomly chosen (where ri refers to region i). After ensuring that no self-connections or parallel links (multiple connections between two regions) would be created, the two original connections are swapped to (r1, r3) and (r2, r4).
The above basic rewiring algorithm is sufficient to generate random surrogate networks. For the spatially constrained surrogate network sets, we incorporated the following additional rewiring conditions: each rewiring step is only executed if the resulting total connection length of every region (i) does not exceed that of the region in the original cortical network (for spatial surrogates), or (ii) is reduced in every step (for reduced surrogates). Because the complex curving trajectories of pathways cannot be preserved during rewiring, connection lengths are approximated by Euclidean distances between the positions of the region-pairs, for both cortical and surrogate networks. In the case of random and spatial surrogates, the procedure is terminated when each connection has been rewired 20 times on average (20 * ne/2 = 178650 connection swaps). For the most constrained reduced surrogates this stopping criterion is too severe because, as the algorithm progresses, progressively fewer rewiring operations with connection length reductions can be found. As a compromise, for this surrogate we chose to rewire each connection only once on average (ne/2 = 8932 connection swaps), resulting in a reasonably diverse (i.e., not overly self-similar) and yet well-optimised set of reduced surrogate networks (see Results). On both resolutions, we generated n = 20 networks for all three surrogate types.
To assess the topological similarity between the cortical connectivity network and its surrogates, we calculated the overlap between the set of connections of the cortical network and the surrogate networks, both in binary and weighted fashion. Specifically, we calculated the binary and weighted overlap between the cortical network C and each of its surrogate S using a modified version of the Sørensen similarity quotient QS [32], which measures the similarity or relative overlap between two sets by the quotient of their intersection and union. We define the binary version of the similarity measure QSb as:(1)where N is the (identical) set of all nodes in networks C and S, and Cb (Sb) is the binarized connectivity matrix of C (S) with Cbij (Sbij) being 1 if there is a link between nodes i and j in C (S) and 0 otherwise. Note that the number of connections in C and S, |Cb| and |Sb|, are equal, and that the product Cbij Sbij is 1 if there is a connection between node i and j both in C and S, and 0 otherwise.
Similarly, we define the weighted similarity quotient QSw as:(2)where Cij and Sij are the connection weights between regions i and j in networks C and S, respectively (0 if the two regions are not connected).
QSb and QSw measure the relative similarity between the connection sets of two networks C and S that are defined on the same set of regions. Both QSb and QSw are normalised similarity quotients taking the value 0 if the two networks share no common connection (minimal similarity), and 1 if the networks are equivalent, that is, they are composed of exactly the same set of binary/weighted connections (maximal similarity). We use both measures because they assess network similarity of two networks in a complementary manner: the overlap in the binary layout of the two networks can only be assessed faithfully by QSb (if the networks are different only in a small number of very high weight links, QSw is already low, despite the high binary overlap), while QSw accounts for the importance (weight) of the connections (if the networks are different only in a number of very low weight links, QSb is lower, despite the high weighted overlap).
We assess spatial network similarity between a network and its surrogates as the average spatial replacement of the connections of each region r, that is, the average change in the positions of all topologically adjacent (linked) regions of r (its topological neighbourhood) in the original and surrogate networks. The theoretically optimal solution for measuring such spatial displacement of the connections would require finding the ‘best matching pairing’ between the original and the rewired neighbour sets of r, i.e., the pairing in which the sum of distances between the (original, rewired) region-pairs is minimal. An exhaustive search for this optimal pairing is however computationally infeasible (given that the regions on average possess 35 connections, a lower estimate on the average number of pairings to check per region is 35!≈1040), therefore we developed and utilized the following algorithm to find an approximation of the optimal pairing.
Given the set of the original topological neighbours of region r in the cortical connectivity, L = [l1, l2, …], and the set of r's rewired neighbours in the surrogate network, M = [m1, m2, …], we calculate the pair-wise distances D(L,M) = [d(l1,m1), d(l1,m2), …, d(l2,m1), d(l2,m2), …] between all element-pairs of the two sets. Then we sort D(L,M) ascending (from the closest to the farthest original-rewired neighbour pairs), and, while iteratively going through the region-pairs of this sorted list, we put the current (li, mj) pair into pairing list P if and only if neither li nor mj is currently in P.
Although the resultant pairing P provided by the ‘greedy algorithm’ above is not guaranteed to be the optimal pairing Popt between L and M, i.e., the one having the lowest sum of (original, rewired) pair-wise distances, it is expected to provide a reasonable estimate on Popt given the close to homogeneous spatial distribution of the regions of the cortical network on the spheroid surface of the cortex [12].
Having obtained P for every cortical region r, we calculate the global relative spatial displacement D between the cortical connectivity C and its surrogate network S as:(3)where N is the set of all regions in the networks (identical in C and S), Dr(C,S) is the average displacement of r's neighbours in C and S, Pr(C,S) contains the (original, rewired) neighbour-pairs of r for C and S, and d(a,b) is the spatial distance between regions a and b. With the above definition, D measures the distances between the original and the rewired neighbours of r (connection displacement) normalised by the distance of the original (cortical) neighbour from r, averaged over all connections and all cortical regions. D = 0 if there is no spatial displacement between the two networks, meaning that they are (both topologically and spatially) identical. A low D value indicates that there is only minor spatial displacement in the neighbour sets of the regions on average, while higher D values indicate a greater neighbourhood displacement, hence a larger difference in the spatial layout between the cortical connectivity and its surrogate network. Generally, the upper limit of D depends on the particular spatial distribution of the nodes and edges of the original network as well as of the wiring constraints of the rewired network in a complex manner. As a simplifying rule for the sparsely and predominantly locally connected (high-resolution) cortical network, however, we can regard D values on the order of 1 as indicators of substantial spatial neighbourhood displacement.
A basic measure of network integration, global network efficiency [33] is the average of the inverse of the shortest path lengths dij between a node i and every other network node j, averaged over all network nodes:(4)where Ei is the efficiency of node i, n is the number of nodes, and dwij is the weighted shortest path length between nodes i and j (the minimal of the weighted sums of constituent edges along each path between i and j, where connection weights are the reciprocal of their strength). High global efficiency implies that, on average, nodes require fewer intermediate steps along stronger (higher weight) edges to reach other nodes; therefore, networks with higher global efficiency possess greater potential for efficient internal information exchange and integration. The advantage of efficiency as a measure for integration over the more traditional measure of the mean shortest path length [15] is that efficiency can be computed for networks with multiple components, and generally is a more balanced measure due to the fact that the mean shortest path length can be strongly biased by the presence of only a few, very long paths [34].
A basic metric of network segregation, the clustering coefficient [23] is the fraction of triangles around a node (the proportion of the node's topological neighbour pairs that are connected with each other), averaged over all network nodes. The weighted clustering coefficient [35], which we use in this study on weighted networks, is defined as follows:(5)where Ci is the clustering coefficient of node i, ki is the degree of i, tiw is the (weighted) geometric mean of triangles around i, wij is the (normalised) connection weight between regions i and j (0 if i and j are not linked). The clustering coefficient of a node is high (1) if many (all) of its neighbours are also directly connected pair-wise (by strength 1 connections in the weighted version of the measure), and it is 0 if none of its neighbour-pairs are directly connected. The clustering coefficient hence measures the (topologically) local density of connectivity of a network.
Informally, a small-world network is a highly segregated (i.e., preferentially locally connected) and yet relatively highly integrated (i.e., easily traversable) network [23]. For the quantitative assessment of small-worldness, the network's high integration is usually translated to relatively short path lengths, while strong segregation is measured by a high level of clustering [36]. Among the several formulae developed to assess the degree of small-worldness of complex networks (e.g. [33], [37]), we chose an altered version of the Humphries–Gurney small-worldness index [37], modified in the following way:(6)where C and Crnd are the clustering coefficient of the network and its random surrogates, while E and Ernd are their global efficiencies, respectively [15]. We note that Humphries and Gurney in [37] use average shortest path lengths instead of efficiency; however we prefer efficiency for the reasons stated above. A network is then said to be small-world if its clustering coefficient is larger than those of its random surrogates (C≫Crnd), while their efficiencies (shortest path lengths) are comparable (E≈Ernd), resulting in SW≫1 [37].
Using the intuition that high degree nodes should occupy a topologically central position in a hierarchical network as a starting point, Ravasz and Barabási introduced the simple but elegant hierarchy coefficient β for assessing hierarchical architecture in scale-free networks [38]. Noticing a distinctively exponential relationship between node degrees and clustering coefficients for various synthetic and real-world scale-free networks, they proposed that the exponent β of this relationship quantifies the tendency of high degree nodes to be linked to a large but sparsely intra-connected neighbour set (hence exhibiting low clustering) and thus effectively serving as connector nodes between segregated parts of the network [38].
Unfortunately, the human cortical network under study, and therefore also its degree-distribution preserving surrogates, exhibit an exponential, rather than scale-free degree distribution [12], and the node degree – clustering relationship does not show a clear exponential shape, so that the β index of Ravasz and Barabási [38] cannot be applied directly. However, their basic idea remains valid irrespective of the specific shape of the functional degree to clustering relationship. Therefore, we here characterize hierarchical organisation by directly observing the degree to clustering relationship in the cortical network and in its surrogates. Specifically, in sparsely connected and locally highly clustered networks, (of the sort studied here, see Results), high degree nodes of a network that possess a lower than average clustering coefficient are typically in a position to connect segregated parts of the network, suggesting a hierarchical element of the architecture with these high degree nodes in its centre (see Figure 1B). In contrast, equal or higher than average clustering coefficients of high degree nodes indicate more homogeneous architectures and the lack of the hierarchical organisation pattern investigated in [38]. We note that the specific kind of topological organisation described above is of course not the only conceivable network architecture that exhibits hierarchical attributes. It is nevertheless the one that has previously been discovered in many sparsely connected, but highly clustered and modular real-world networks [38], making it a good candidate to test for here.
The modularity index Q, proposed by Newman [39], has proved to be a highly accurate and powerful indicator of the modularity strength of a given partitioning of a complex network [16], [40]. Given a set of node groups (modules or communities) M, that fully partition the network without overlaps, the modularity index Q of that partition is given by(7)where Qu is the modularity index of module u, euv is the proportion of all weighted edges wij between modules u and v in the network, lw is the sum of all weights in the network, and kiw is the sum of all connection weights of node i.
Numerous algorithms have been developed to recover the modular structure of complex networks utilising Q as a ‘fitness’ measure to be optimised by some means (for reviews, see [16], [40]). In this study, we use the simple and elegant spectral algorithm developed by Newman [41]. Starting from the entire network as a single module, this algorithm iteratively splits each module into two, at each step finding the optimal bipartition by utilising a so-called ‘modularity matrix’ derived from the network's connectivity matrix. The leading eigenvector of the modularity matrix determines the node composition of the two sub-modules of each module to be split. The algorithm stops when no more increase in the global modularity index Q can be achieved by any additional split [41]. Along with its high accuracy, Newman's module detection procedure has the additional advantages of being a divisive, deterministic and generalisable method with low computational cost. See [41] for a detailed description of its implementation.
We measured the consistency of the cortical module partition in surrogate networks with the scaled inclusivity index [42]. Application of this measure capitalized on the fact that the cortical network and its surrogates are defined on the same set of nodes (cortical regions) and differ only in their connection sets. Additionally, scaled inclusivity has the advantage of making no assumptions on the investigated partitions, and is thus generically applicable even on partition-pairs which differ in the number and sizes of modules they contain. For other pair-counting, cluster-matching, and information-theoretic techniques applied to compare module (community) structures of different networks, see [43]–[45].
The calculation of the scaled inclusivity index proceeded as follows. First, the individual module partitions of the cortical and surrogate networks were identified independently by Newman's spectral algorithm (introduced above). Then, the cortical module partition QC, composed of m modules, was taken as a reference partition, and its match with the partition QSi of each network i of surrogate group S, composed of n modules, was assessed by calculating the n×m module-by-module similarity matrix XiC, which (p,q)-th element is calculated as:(8)where QSi(p) is the set of nodes (regions) belonging to the p-th module in QSi and QC(q) is the set of nodes belonging to the q-th module in QC. The resulting values range from 0 to 1, where XiC(p,q) = 0 indicates zero overlap between the modules p and q (i.e., they do not share any node), and 1 indicates that the two modules are identical (i.e., they are composed of the same set of nodes).
After calculating the matrix X for all networks in a surrogate group, the scaled inclusivity index SI of each cortical region is calculated as the mean of the similarity indices XiC(p,q) between all modules QSi(p) and QC(q) that contain the region, averaged over all surrogate networks i. Thus, scaled inclusivity measures how consistently a region is classified in each surrogate group, based on how well its cortical modules match with its surrogate modules, on average. We stress that SI is intended as a generically applicable metric to measure the degree of similarity between the module classification of network nodes, and it does not aim to accurately measure the actual magnitude of ‘overlap’ between the partitions (see [42] and Eq. 8 above).
The ‘core’ of a network is usually determined by an iterative peeling algorithm. These algorithms, at each step, remove (‘peel off’) a set of ‘shell’ or ‘crust’ nodes, in order to progressively focus on the more ‘centralised’ nodes. Centralisation in these procedures is assessed by a specific ‘coreness condition’, as described below.
To find the core structures of binary and weighted networks, we used the k-core and s-core decomposition methods, respectively. The k-core of the network [46], for a given degree k, is the maximal set of nodes that are connected to at least k other nodes in the core. The k-coreness index of a node is then the highest degree k for which the node is still a member of the k-core. Similarly, the weighted variant of the k-core, the s-core of the network [12] is the group of nodes in which each node has a summed connection strength of at least s towards the rest of the s-core (i.e., the sum of the weights of its intra-core connections is not less than s). For increasing s (k), the s-core (k-core) shrinks progressively and the tightest or innermost s-core (k-core) of the network [simply s-core (k-core) from here on] is the set of remaining nodes in the last non-empty s-core (k-core).
The so-called rich-club phenomenon is the tendency of high degree nodes to be preferentially connected to each other [47], [48]. The degree of ‘rich-clubness’ is usually measured by the k-density function φ(k) of the network, which is the internal connection density among all nodes with degree larger than k. There is a basic difference between k-core/s-core and rich-club properties: while k-core and s-core nodes are selected by their connections within the subnetwork formed by the core, rich-club nodes are chosen simply and solely on the basis of their global degree in the entire network. (Of course the ‘rich-clubness’ of this subnetwork does then depend on its internal connectivity.)
A possible weighted variant of the rich-club measure, as introduced in [49], evaluates the tendency of the highest connection weights to be distributed among high degree (‘rich’) nodes. However, this variant, due to normalisation by the number of edges, is a connection density-independent index of weight centralisation and thus loses the ability of the unweighted rich-club index to measure edge centralisation among high degree nodes. Here we propose a novel weighted version of rich-clubness, which is sensitive to both properties, connection density and weight centralisation, and may hence be a more appropriate generalisation of the unweighted rich-club index to weighted networks.
We define weighted rich-clubness as the internal weighted connection density φw(k) of the set of nodes with degrees larger than k, N>k, which is the ratio between the sum of connection weights W>k among the nodes in N>k and the maximum of their possible weight sum, Wmax>k:(9)where is the maximum possible number of edges among the nodes in N>k, and wrankedl is the weight of the lth strongest (highest weight) edges in the network.
φw(k) defines a normalised measure of coreness, which takes a value in [0,1] for each degree k. φw(k) is 1 only in the extreme case where N>k is fully connected by exactly the strongest connections of the network. In general, φw(k) measures the fraction of total interconnection strength within N>k relative to this theoretical maximum (as defined by the connection weights present in the network).
Note that in Eq. 9 the denominator is not calculable if Emax>k is greater than the number of edges, E, in the network. This condition renders the interpretation domain of φw dependent on the connection density of the investigated network, implying that φw is meaningful for weighted rich-clubness measurements only for that fraction of the highest degree nodes N>kmin. Specifically, for undirected graphs, the number of these nodes |N>kmin| cannot be larger than the real solution of the quadratic equation(10)Eq. 10 specifies the largest number of nodes x that can still be fully interconnected by the existing number of edges E in the network. The cortical network under study has E = 17865 connections, hence we obtain |N>kmin| = 188 nodes as the largest weighted rich-club size that can be assessed by our measure. This corresponds to 18.8% of the nodes of the entire network, and gives φw (k) the domain of k∈[kmin, kmax], where kmax = 97 is the largest node degree in the network, and kmin = 49 is the degree of the 188th node in the degree rank ordered node list. We note that, apart from of this interpretation limit of the measure, when applied to unweighted (binary) networks φw gives the same result as the traditional rich-club metric, underlining that φw can be interpreted as a generalisation of this traditional measure for weighted networks.
Degree assortativity is a global measure of the tendency of nodes to be preferentially connected to other nodes with similar degree [50]. Degree assortativity is thus closely related to the phenomenon of rich-club formation, although while the latter only accounts for high degree nodes, the former measures preferential connectedness across nodes of all degrees. The assortativity coefficient r of a network is formally defined as:(11)where ji, ki are the degrees of the nod es at the ends of edge i, and M is the number of edges [50]. Degree assortativity is a normalised measure (−1≤r≤1), so that a network has positive r assortativity values if its edges tend to connect nodes of similar degree, while negative assortativity values indicate the tendency for nodes with different degrees to be linked. A network with r≈0 expresses neither of these trends, and is non-assortative.
In the analyses presented below we used the structural connectivity network of the human cortex obtained by Hagmann et al. [12] comprising 998 regions of interest and 17,865 undirected and weighted connections (Figure 2), see Methods. Unlike many previous studies [e.g.], [ 26], [27,28], we performed analysis on both the full maximal resolution and on a low-resolution sub-sampling of the data set and surrogate networks of the same size, and on both weighted and unweighted (binary) versions of these networks. Additionally, we repeated the analysis on a single cortical hemisphere of the high-resolution network, in order to test for any artefacts arising from the features of inter-hemispheric connections (see Single hemisphere analysis). In the following, we present the results with a focus on the high-resolution weighted connectivity type (as it contains the most information), and discuss the findings on the other network types at the end of the section (Results on low-resolution and binary connectivity types). In the following we first describe validation of the three surrogate sets. We then compare standard topological integration and segregation properties of cortical and surrogate networks, and finally report analysis of more complex network properties such as small-worldness, hierarchy, modularity and core formation.
The high-resolution weighted cortical connectivity matrix and averaged connectivity matrices of the three surrogate sets are illustrated in Figure 3A–D. To allow meaningful comparisons, surrogate networks need to be sufficiently randomised. The rewiring algorithms, as outlined in Methods, are constrained by several factors during the randomisation of cortical connectivity. In order to assess that sufficient randomisation has been achieved in spite of these constraints, we quantified the degree of similarity between each ensemble of surrogate networks and the cortical network, and we also examined the similarity within each surrogate ensemble.
To examine topological similarity, we calculated the mean binary and weighted similarity quotients, QSb and QSw (Eqs. 1 and 2) of the networks in the three surrogate sets to the cortical network. For random surrogates, QSb(C, Srnd) = 0.054±0.002 and QSw(C, Srnd) = 0.047±0.002, indicating that their connections are almost entirely different from those of the cortical network. For spatial surrogates, we obtained intermediate similarity quotient values QSb(C, SS) = 0.494±0.002 and QSw(C, SS) = 0.483±0.002, and for reduced surrogates higher similarity quotients QSb(C, SR) = 0.670±0.001 and QSw(C, SR) = 0.700±0.001. These results confirm that, as expected, conserving and, even more significantly, further decreasing the already short connection lengths of the cortical connectivity network limits the achievable topological randomisation of the spatial and reduced surrogate networks.
The similarity quotient values described above exhibit only very small deviations around their respective means. This could reflect the combined consequence of a sufficiently extended connection shuffling process together with the relatively large size of the networks, following the law of large numbers. But it could also indicate an undesirably low diversity in the generated surrogate sets, i.e., each set might be composed of highly similar networks. To test for this possibility we calculated the similarity quotient between every pair of surrogate networks in each of the surrogate sets. The resulting mean intra-group values and their standard deviations are QSb(Srnd, Srnd) = 0.053±0.001 and QSw(Srnd, Srnd) = 0.045±0.001 for random surrogates; QSb(SS, SS) = 0.474±0.003 and QSw(SS, SS) = 0.483±0.002 for spatial surrogates, and QSb(SR, SR) = 0.873±0.002 and QSw(SR, SR) = 0.861±0.002 for reduced surrogates. Together, these results indicate that topological differences among surrogate ensembles, although decreasing with stricter spatial constraints, are nevertheless significantly nonzero.
Interestingly, the low intra-group variance of the similarity values within every surrogate set suggests that in each such set S there is a ‘characteristic similarity’, QS(S,S), between any two members of that set. In addition, the similarity of the cortical network to its surrogate networks is comparable to these characteristic intra-group similarities in the case of random and spatial surrogates (QS(C, SS)≈QS(SS, SS) and QS(C, Srnd)≈QS(Srnd, Srnd)). This suggests that the cortical network is a generic member of the random and spatial surrogate sets in terms of its basic region-to-region connectivity, as measured by QS. This further supports the use of random and spatial surrogates as suitable null-hypothesis networks with respect to the preserved basic properties of the cortical connectivity defined by each surrogate type.
By contrast, reduced surrogates appear to form a separate class of networks that are more similar to each other than to the cortical network (QS(C, SR)≪QS(SR, SR)). This is expected given the restrictive form of spatial constraint applied during their generation (strictly decreasing total connection length in every rewiring step), which is likely to make them collectively drift away from their cortical origin, converging towards the (hypothetical) single, minimal connection length surrogate network.
The QS values illustrate well the highly optimised wiring of the cortical network in terms of connection length. While random surrogate networks share only 5.4% of their connections with other random surrogates and with the cortical network, this ratio increases to 49.4% for spatial surrogates, and each reduced surrogate is only able to substitute about one third of the long-range cortical connections with shorter ones. Furthermore, as shown in Figure 3A–C, these pair-wise overlaps translate into a ‘core’ set of connections collectively shared between the cortical network and its spatial and reduced surrogates. This ‘skeleton connectivity’ is primarily located along the main diagonal of the connectivity matrices, where most of the potential short-distance connections can be placed (due to the spatial ordering of the brain regions in the connectivity matrices, explained in detail in the caption of Figure 3).
We note, however, that Figure 3B–D show the averages of the connectivity matrices of the surrogate network groups and therefore exaggerate the pair-wise overlap of the networks in each group. This is a consequence of the relatively small set of potential short-range connections in cortical space (compared to the number of all possible connections), a number of which are inevitably shared by many reduced and spatial surrogates. For example, to examine the most extreme case of shared connectivity, we can determine the connections that are present in all network instances of each surrogate group. As expected, there are no such collectively shared connections among random surrogates. On the other hand, the highly optimised, and hence self-similar, reduced surrogates collectively share as many as 65.0% of their connections, while the ‘intermediately’ constrained spatial surrogates have only 7.6% of their connections shared among all of them, rendering the latter surrogate group relatively diverse. Furthermore, all shared connections of reduced and spatial surrogates are also present in the cortical network. These findings, in accordance with the ones on QS above, indicate that the cortical network is indeed a generic member of its spatial (and random) surrogates in terms of the basic properties of its connectivity, adding some topological credibility to our surrogate analysis.
Having assessed the topological similarity of the surrogate ensembles to the cortical network, we now investigate the other relevant aspect of surrogate creation, namely to what degree the spatial layout and wiring properties of the cortical network have been changed in the surrogate ensembles. Although topological and spatial similarity are related, they do not specify each other. For example, low topological similarity between the cortical network and its surrogates in itself does not exclude that connections of the cortical network may only have been displaced by a short distance, leaving the spatial layout of the network largely unaffected by the randomisation procedure.
In order to assess the impact of the randomisation procedure on the spatial layout of the cortical network, we calculated the relative spatial displacement D between the high-resolution cortical network and its surrogate groups (see Assessing spatial similarity in Methods). We obtained a D(C,Srnd) = 4.04±3.43 mean displacement value for random surrogates, indicating that on average a neighbour l of each region r in the cortical network is replaced by a new neighbour m in random surrogates, which is about four times further away from the original cortical neighbour l than the length of the original cortical connection (r,l). In spatial and reduced surrogates, we measured D(C,SS) = 0.50±0.62 and D(C,SR) = 0.29±0.43, respectively, indicating a necessarily lower mean spatial displacement of the regions' neighbourhoods in the topologically more similar spatially constrained surrogates. However, because a significant number of connections is shared by the cortical network and its surrogates (see Topological similarity of surrogate networks) and hence have zero displacement, the high standard deviation in D(C,SS) and D(C,SR) indicates that those connections that have actually been rewired are displaced to a location that is substantially distant from their original target region in the cortical network. This is indeed what we see if we exclude the overlap of the connectivities and calculate the spatial displacement Dr of the replaced connections only: Dr(C,SS) = 0.97±0.57 and Dr(C,SR) = 0.88±0.30, which indicates that the average displacement of rewired connections is almost as large as the length of the original connection.
The connection length distribution and total connection length of each region (sum of distances to all neighbours) in the high-resolution cortical network and its surrogates are shown on Figure 3E and F. Consistent with the predominantly local connectivity of the cortical network (mean connection length per region: CLC = 27.625 mm), random rewiring of cortical connections nearly tripled the average connection length (mean ± standard deviation of random surrogate network means: CLrnd = 75.971±0.164 mm). For this reason, that is, due to the natural tendency of random connection swapping to increase the length of originally short cortical connections, the simple condition applied during spatial surrogate generation (i.e., ‘not to exceed the original total connection length of the cortical network’) was sufficient to actually achieve conservation of connection lengths (CLS = 27.507±0.120 mm), and resulted in a slightly narrower connection length distribution (standard deviation of connection lengths: cortical network: σlC = 22.146 mm→spatial surrogates: σlS = 18.589 mm) originating from a somewhat shorter tail of the distribution (see Figure 3E).
Wiring length optimisation in reduced surrogates of the high-resolution weighted network successfully reduced the mean cortical connection length by 29.6% (CLR = 19.433±0.013), effectively substituting long-range cortico-cortical projections with shorter, local ones. This also led to a much narrower distribution of connection lengths (standard deviation of connection length: cortical network: σlC = 22.146 mm→reduced surrogates: σlR = 7.382 mm). As a result of the above, the total connection lengths of individual cortical regions were preserved in spatial surrogates (cortical network – spatial surrogates mean difference: −2.4±5.5%, Wilcoxon rank-sum test for identical distribution: p = 0.898), while reduced and random surrogates had significantly decreased (−24.6±17.0%) and increased (+227.6±114.0%) regional connection lengths, respectively (p<10−4 in both cases).
Several earlier studies investigated spatially minimally wired surrogates of various neural and brain connectivity networks [51]–[54]. In order to investigate how much excess wiring length cortical connectivity has over its theoretical minimum, as well as to assess how the reduced surrogates compare to ‘bottom-up’ constructed, minimally wired models, we assembled two such models.
For the first, unconstrained minimally wired network model, which we call absolute minimal (AM) network, we took the 998 cortical regions without their connections and simply placed links between the 17865 spatially closest region-pairs. This created a network with minimal total wiring length given the spatial arrangement of the cortical regions and the total number of connections in the cortical connectivity. The resulting AM network is composed of a single component (no disconnected regions or groups of regions). The sum of its connection lengths is 62.9% of that of the cortical network, which, importantly, is only 10.6±0.1% less than the total connection lengths of the reduced surrogate networks. Importantly, the degree distribution of the original cortical network has been completely lost in the AM network (mean relative deviation of regional degrees between cortical network and the AM network: 52.5±130.7%). This means that the reduced surrogates were able to achieve highly optimised wiring lengths while fully preserving the cortical network's degree distribution, thus providing a powerful topological baseline to assess the significance of the cortex's high level network properties. Both the cortical network and its reduced surrogates share a large number of their connections with the AM network (binary similarity quotient: QSb(C,AM) = 0.621, QSb(SR,AM) = 0.760±0.0009), showing once again the remarkably conservative wiring of the cortex: 62.1% of the cortical connections are among the theoretically shortest possible links in the cortical network.
We devised a second ‘bottom-up’ constructed minimally wired connectivity model with the additional constraint of approximating the degree distribution of the cortical network. We construct this network, which we call the degree preserving minimal (DPM) network, in the following way. As with the AM model we start with the 998 cortical regions without any connections, and, by going through the list of potential connections (region-pairs) ordered from shortest to longest, we add each connection to the DPM network only if the current degrees of both corresponding regions in the DPM network are less than their original degrees in the cortical network. By this simple strategy we are able to create a network with 17799 connections (66 connections [0.4%] less than the cortical network) that closely approximates the degree distribution of the cortical connectivity (mean percentage deviation in regional degrees between cortical network and the DPM network: 0.2±1.8%). Due to the similarity in degrees, the DPM network shares an even larger number of connections with both the cortical network and the reduced surrogates than the AM network (binary similarity quotient: QSb(C,DPM) = 0.653, QSb(SR,DPM) = 0.855±0.002). The sum of connection lengths in the DPM network is 72.1% of that of the cortical network, which is on average 2.4±0.1% more than those of the reduced surrogates, despite the fact that the DPM network has slightly less connections than the reduced surrogates. This demonstrates that simple ‘bottom-up’ algorithms are not guaranteed to be more successful in constructing minimally wired (surrogate) networks than the rewiring methods used in the current study.
We conclude that the spatial surrogates effectively preserved the wiring length properties of the cortex, both globally and at the level of individual regions, and that the reduced surrogates significantly decreased wiring length by substituting long-range connections with shorter ones. These properties render spatial and reduced surrogates suitable for representing a wiring-length-matching and wiring-length-optimised null-hypothesis network set of the cortical connectivity, respectively. The results so far demonstrate that, as opposed to the highly unrestricted nature of random surrogates, the presence of strict wiring constraints necessarily limits the form of potential connectivities of the cortex at the basic level of region-region connections, as shown by elevated similarity between cortical network and its spatial and reduced surrogates as compared to random surrogates.
In the remainder of the paper, we go beyond these basic properties, to examine which other, network-level properties of the cortical connectivity these wiring constraints preserve. We measure the degree of expression of these properties by a series of complex network metrics, in each case applying the interpretations detailed in the Introduction (see also Table 1).
The need for the simultaneous presence of functional integration and segregation imposes conflicting constraints on network architecture [55], reflected in properties collectively known as ‘small-world’ characteristics. Small-world properties have been found in many real-world complex networks [23], including various brain networks [10], [56]–[58].
We measured the global integration and segregation potential of the cortical network compared to its surrogates using the quantities efficiency E and clustering coefficient C (see Methods). As shown in Figure 4A, the cortical network is more similar to its reduced surrogates than to its other two surrogate sets (high-resolution weighted cortical network: EC = 0.174, CC = 0.271, reduced: ER = 0.162±0.001, CR = 0.289±0.002, spatial: ES = 0.214±0.001, CS = 0.169±0.002, random: Ernd = 0.260±0.001, Crnd = 0.024±0.001). Considering that the total connection length of each region in the cortical network is the same as in its spatial surrogates, and that long-range connections are largely absent in reduced surrogates, the efficiency results indicate that the long-range cortico-cortical connections are distributed in a topologically sub-optimal way for enhancing tight functional integration (efficiency) in the cortical network. Furthermore, the clustering coefficient indices demonstrate a prevalence of topologically segregated neighbourhoods of groups of regions, beyond what would be expected from the wiring constraints of its individual regions (CC is significantly larger than CS and much closer to CR than to CS). Therefore, not only when comparing against the necessarily more highly integrated and less segregated random surrogates, but also when taking into account the total length of the connections of each cortical region in the spatial surrogates, the cortical network appears to strongly favour topological segregation over integration (efficiency).
In order to assess the effect of wiring constraints on its small-world attributes, we calculated the small-world index SW of the cortical network and its spatial and reduced surrogates (see Methods), using random surrogates as reference networks (random surrogates hence have SWrnd = 1 by definition). First we note that all three investigated network types (cortical network, spatial and reduced surrogates) satisfy the basic small-worldness condition [37], that is, their clustering coefficient is larger than those of its random surrogates (C≫Crnd) while their efficiencies (average shortest path lengths), while being lower (higher), are still comparable to those of their random surrogates (E≈Ernd). In case of the cortical network, this results in the relatively high small-world index SWC = 7.478 (see Figure 4A), indicating a well-expressed small-world organisation of the cortex. In comparison, we obtain on average SWS = 5.746±0.031 for spatial surrogates, and SWR = 7.419±0.031 for reduced surrogates, both much closer to SWC than the random surrogates (recall SWrnd = 1), indicating that the small-world architecture of the cortex can be attributed to a great extent to its wiring constraints. However, considering that SWC is significantly higher than SWS, the cortical network appears to exhibit the small-world property beyond what would be implied by its local connectivity alone. Furthermore, this excess level of cortical small-world organisation does not necessitate any particular arrangement, or even the presence, of the long-range cortical connections, as indicated by SWR not being significantly different from SWC. Therefore, the highly segregated connectivity of the cortical network, also found in reduced surrogates, but not in spatial surrogates (see above), appear to contribute more to the small-world organisation of the cortex than the mere existence or particular arrangement of cortical long-range connections.
In their seminal work, Ravasz and Barabási [38] detected well-expressed hierarchical structure in all investigated non-spatial (non-geographical), real-world networks, but not in spatial examples (e.g. the power grid network and the Internet). They reasoned that the high cost of establishing physically long connections substantially limits the type of topology spatial networks can exhibit, potentially excluding strongly hierarchical forms. However, in a study of a 104-region structural network of the human cortex Bassett et al. [10] did find hierarchical properties in the brain among multimodal cortical regions, but not within unimodal and transmodal regions.
Following Ravasz and Barabási [38] (see Methods), we calculated the average clustering coefficients of groups of cortical regions with similar degrees, relative to the global clustering coefficient of the cortical network (see Figure 4B). We observe that the cortical network exhibits a steep decline in its mean clustering – degree relation, indicating that the cortex exhibits the type of hierarchical organisation illustrated in Figure 1B. This finding supports the general notion of a hierarchically organised brain [59], which is quite remarkable given the tendency of spatially embedded, physical networks not to develop hierarchical features due to the basic spatial (geographical) constraints acting on them [38]. Furthermore, there are highly similar tendencies for spatial and reduced surrogates, but not for random surrogates, in which clustering actually increases with region degree. The remarkably high consistency of the clustering – degree relationship across the cortical network and its spatial and reduced (but not random) surrogates indicates that the individual wiring lengths and positioning of high degree regions in the cortex by itself entails a global hierarchical organisation.
Many real world networks have a characteristic topology that allows them to be separated into relatively densely intra-connected and weakly inter-connected subgroups [16], [60]. These subgroups are usually referred to as the modules (or clusters, communities) of the network. One possible functional advantage of modularity is reduced systemic risk during development and evolution [61], [62]. Another is that modular architectures are related to potentially useful dynamical properties such as high dynamical complexity [21] and metastability [63], as well as limited sustained network activity [64].
Recent studies have reported a highly modular architecture of the human brain in its structural [12], [13], [65] as well as in its resting state functional connectivity (rsFC) [66]–[68]. Furthermore, studying the effect of ageing on the brain's modular structure, Meunier et al. [69] found marked differences in the composition and putative topological roles between the modules in the rsFC of younger and older human subjects. These results suggest that modular ‘decomposability’ is a prominent feature of the brain, which is continuously shaped during its development, maturing and ageing. In line with these results, recent theories regard the brain's modular structure as the main facilitator of regional specialisation and segregated functional processing [18].
We investigated the modular structure of the cortical network and its surrogates by utilising Newman's module detection algorithm [41] (see Methods). In order to assess the strength of modular organization, that is, the magnitude of the Q modularity index, we use the modularity of the random surrogates as a baseline value (representing the modularity index of a non-modular network with size and connection density matching that of the cortical network). These random surrogates, as expected due to their quasi-zero segregation, express almost no modularity (mean modularity index: Qrnd = 0.087±0.003, number of modules: Nrnd = 23.25±1.95). In contrast, the cortical network has a strongly modular architecture (QC = 0.558) composed of NC = 13, spatially compact and hemispherically symmetric modules (Figure 5A). Interestingly, reduced surrogates, in spite of their lack in long-range (thus mostly inter-module) connections, do not exhibit a significantly higher modularity index (QR = 0.567±0.015, NR = 15.55±0.87, one-tail t-test assuming normal distribution: p = 0.274), but spatial surrogates do possess a significantly lowered level of modularity (QS = 0.477±0.020, NS = 11.55±0.87, p<10−4) than the cortical network. These results show that while the physically constrained length of cortico-cortical white-matter connections are a fundamental factor in shaping the high strength (QC) and granularity (NC) of the global modular architecture of the cortex, the cortical network nevertheless has a stronger modular organisation than these wiring constraints by themselves would suggest, indicating the functional relevance of the cortex's modular structure.
The strength of the modular organisation of the cortical network can be illustrated by its inter- versus intra-modular connection distributions (Figure 5B). The NC = 13 identified modules contain 63.2% (n = 11294) of the total number of projections internally, meaning that only 36.8% (n = 6571) of the connections cross module boundaries. This results in an average 25.6% intra-module and 1.4% inter-module connection density, indicating that while more than every fourth intra-module region-pair is linked, this ratio falls to 1∶70 for region-pairs from different modules. For comparison, the global average connection density of the entire network is 3.6%.
The cortical connectivity matrix ordered by the recovered module partitioning is shown in Figure 5B. To compare this partitioning with an ‘average’ partitioning for each surrogate group, we calculated the frequency with which every region-pair (ni, nj) can be classified into a single module (m(ni) = m(nj)) in each of the three surrogate network groups. The resulting matrices are shown in Figure 5C–E. The high concentration of frequent co-partitioning of region groups along the main diagonal of the matrices is apparent in the case of reduced and spatial surrogates, indicating that the corresponding cortical modules are reasonably preserved across these surrogate networks. Furthermore, there is also a tendency for the formation of off-diagonal blocks in Figure 5C and 5D which suggests that parts of some of the cortical modules are frequently merged into single surrogate modules, and therefore they are at least partially preserved in reduced and spatial surrogates.
Motivated by these findings, we quantitatively tested the robustness of the modular partitioning of the cortical network against the rewiring applied to its surrogate groups by assessing the consistency of the cortical partition in the surrogate groups. To do this, we used the obtained cortical modules as a reference partition and measured the scaled inclusivity index SI of each cortical region in the surrogate groups (see Methods). Among the three surrogate sets, reduced surrogates exhibited the highest mean SI index, indicating the highest overall conservation of cortical modules in reduced networks, although with high variations across the individual cortical regions (mean ± std: reduced surrogates: SIRC = 0.235±0.182, spatial: SISC = 0.202±0.145, random: SIrndC = 0.007±0.002).
The SI values for the individual cortical regions, and for each surrogate group, are illustrated in Figure 5F–H. We found elevated robustness of the cortical modules in both reduced and spatial surrogates at specific cortical sites, including the entire pre-central and post-central cortices (composing cortical modules M2 and M6 on Figure 5A), large areas of the temporal lobe (M3 and M5) and some frontal (M4 and M9), and superio-parietal and limbic areas (M10). The high SI of these specific areas indicates that their modular structure exhibits greater robustness against spatially constrained rewiring, as opposed to the low SI of, and thus higher variance in, the module formations in other parts of the cortical network.
The results so far, regarding the small-world, hierarchical and modular architecture of the cortex, suggest the existence of specific cortical areas that are topologically centrally positioned in the modular structure of the cortical network. This ‘core formation hypothesis’ has been the topic of several studies recently (see below), and we next test its significance against the wiring constraints of the cortex by again analysing the surrogate ensembles.
Intuitively, the core of a network, illustrated in Figure 1C, is a set of ‘elite’ nodes that are topologically centrally positioned, forming a highly intra- and inter-connected global centre [28]. The existence of a single, but strong core formation in the topology of a network typically suggests that the network exhibits a pronounced global core-periphery structure [70]–[72] and indicates the presence of centralisation in the network's dynamics and functional operation, which is fundamentally different from that of a homogeneous network architecture composed of distributed, identically segregated units (e.g., Figure 1A).
Prior studies have identified and investigated a core structure in various brain networks, including the rich-club structure of the cat thalamo-cortical complex [17], , the k-core of the macaque brain [75], the s-core of the human cortex [12], and the rich-club of the entire human brain [13], [25]. We here compare s-core and rich-club properties of the cortical network and also assess the extent of their dependence on, and emergence given, different wiring constraints using the three surrogate types.
S-core analysis assesses the extent to which a network exhibits a densely intra-connected inner core, by measuring the size of, and overall connection strength within, the most strongly intra-connected group of nodes. We identify the s-core of the cortical network through a ‘peeling’ procedure that iteratively removes less connected regions from a candidate s-core (see Methods). Examining the evolution of the s-core decomposition of the high-resolution cortical network and those of its surrogates (Figure 6A) during the peeling procedure, we can identify two characteristic phases. A longer, rather stable early phase of ‘crust peeling’ transitions into an unstable phase for s>11, in which the s-cores of both random and spatial surrogates diminish rapidly and then abruptly vanish. The cortical network, on the other hand, closely follows the trend of its reduced surrogates and continues to sustain a substantial s-core of n = 100 regions (10.0%) for much longer. This s-core eventually collapses at a significantly higher strength threshold (sC = 13.095) than its counterparts in the random (srnd = 12.055±0.078) or spatial surrogates (sS = 11.433±0.124), within the range of the s-cores of reduced surrogates (sR = 13.027±0.143), but with a somewhat larger size (s-core size of cortical network SC = 100, reduced surrogates: SR = 74.500±17.119, see Figure 6A inset). Considering that the connectivity of reduced surrogates is spatially more concentrated than that of the cortex, which is a property that favours the formation of a strong s-core, the above finding suggests that cortical connectivity may be optimised towards the formation of a global s-core, which is much stronger and larger than its connection length constraints alone would suggest.
An alternative measure of core formation in a network is the assessment of its rich-club index [47], [48]. The weighted variant of a rich-club index, φw(k), measures the tendency of high degree nodes to be both densely and strongly inter-connected (see Methods). Examining the evolution of φw(k) with increasing k in the cortical network and in its surrogates (Figure 6B), the cortical network demonstrates a rich-club of significant strength (weighted k-density at n = 100 regions: φwC(100) = 0.164) compared to its random surrogates (φwrnd(100) = 0.106±0.002). However, the cortical network does not possess a significantly stronger rich-club structure than its reduced surrogates (φwR(100) = 0.164±0.001, one-sample t-test: p = 0.23) or its spatial surrogates (φwS(100) = 0.163±0.003, one-sample t-test: p = 0.34).
Previous studies [13], [17], [25] used only random surrogates as null-hypothesis baselines for assessing the rich-club property of brain networks, a comparison in which the cortical networks we study here also express a highly developed rich-club (Figure 6B, compare blue and magenta lines). However, we show here that this property is equally, or even more, expressed in spatial and reduced surrogates. Closer inspection reveals that the relatively low variance in the spatial locations of highly connected regions (Figure 6D and F), in combination with the highly clustered, local connectivity of the cortex, naturally results in a tendency for strong rich-club formation.
The wiring-constraint-dependent rich-club formation tendency of the cortex is further supported by the assortativity coefficients r of the network and its surrogates (see Methods). We found significantly positive assortativity coefficients for the high-resolution cortical network (rC = 0.288) and its spatial (rS = 0.283±0.004) and reduced surrogates (rR = 0.326±0.002), indicating their tendency to connect nodes of similar degree, whereas almost no degree assortativity is found in random surrogates (rrnd = 0.051±0.006). This preferentially mutual connectedness of high degree regions suggests that the rich-club patterning of the cortical network naturally arises from the physical location of cortical hubs and the cortical wiring constraints.
The s-core and rich-club regions selected by the two methods (Figure 6C–F), are largely consistent with earlier findings [12], [13]. Furthermore, the s-core (n = 100 regions in final, non-empty core) and rich-club regions (n = 100 highest degree regions) exhibit a considerable, exactly 50% (n = 50 regions) overlap in the cortical network. There are, however, marked differences in the anatomical composition and spatial dispersion of the two structures. The s-core of the cortical network encapsulates the caudal part of the cortical midline, formed by the precuneus, the cingulate cortex and the superior part of the occipital lobe (cuneus, lingual gyrus and pericalcarine cortex). This centralisation is also present, though much less pronounced, in the cortical network rich-club, since about one third of it extends to the lateral and frontal parts of the cortex. The spread of arborisation of the two cores also exhibits this difference (see Figure 6C–F): while the more numerous (n = 5662 [31.7%] connections) and rather externally projected connections (20.6% internal connection density) of the rich-club establish direct connectivity with almost the whole remainder of the cortex (n = 795 [88.5%] regions), the s-core possesses a smaller (n = 3921 [21.9%] connections), as well as more internally projected connection set (37.7% internal connection density), which connects it directly with only one third (n = 294 [32.7%] regions) of the rest of the network. These differences, originating from the definitions of the s-core and rich-club structures, demonstrate the more distributed nature of the cortex's rich-club, as opposed to the rather encapsulated, but spatially and topologically central position of the s-core.
Along with the analysis on the high-resolution weighted version of the cortical connectivity dataset presented above, we also performed our surrogate analysis on four ‘subsets’ of the full dataset, namely: on the binarized (unweighted) version of the high-resolution cortical network, on the weighted and the binarized versions of a lower resolution (down-sampled) cortical network (see Methods), and on a single hemisphere extracted from the high-resolution weighted cortical network (discussed in detail in the following section).
Similarly to the analysis of the high-resolution weighted cortical network, we first tested the surrogates of the three cortical networks considered here with the topological similarity measure QS and the measure of mean connection lengths per region, CL. Our surrogate test results on the three cortical networks showed the same pattern that we described for the weighted high-resolution cortical network (see Figure 7), albeit with an overall lower level of randomisation (higher topological similarity) for the low-resolution networks, due to the higher connection density of these networks (high-resolution: 3.6% connection density, low-resolution: 26.8%), as well as a slightly (but significantly) reduced connection length in low-resolution spatial surrogate networks, likely due to the limitations of re-wiring algorithms on smaller networks.
Next we assessed the global integration and segregation potential of these cortical networks by calculating their clustering coefficient C and efficiency E, respectively. In accordance with the high-resolution weighted results, we found the same pattern of higher similarity of each cortical network to its reduced than to its spatial surrogates consistently across all analysed cortical networks, to the extent that for low-resolution networks there is no significant difference between them (see Figure 7). Therefore, as with the high-resolution weighted cortical network, these networks also demonstrate a small-world index more similar to their reduced than spatial surrogates (see Figure 7). Surrogate analysis of the modularity strength Q of these cortical networks also yield highly consistent results with those of the high-resolution weighted cortical network (see Figure 7). Taken together, these findings are consistent with our results on the high-resolution weighted cortical network; they indicate that the functional segregation potential and the small-world and modular organisation of the connectome, even when observed on lower resolutions, are significantly stronger than its wiring constraints alone can account for.
We next evaluated the core formation tendencies of the three cortical networks. Results on the k-core (unweighted s-core) of the high-resolution binary connectivity are in agreement with the high-resolution weighted network results discussed above. On low network resolution, however, we observe different characteristics (see Figure 7). Specifically, the significantly strong k-core and s-core structures of the binary and weighted high-resolution cortical networks seem to weaken to a weighted low-resolution cortical s-core of comparable strength to its wiring-constrained surrogate ensembles, and further diminish to a binary low-resolution k-core with a strength significantly weaker than any of the surrogates. When investigating the binary low-resolution cortical k-core more closely, we discovered that it contains as much as 80.3% (53 regions) of the entire network, and any subsequent peeling step (see Methods) destroys the whole structure. This is in stark contrast to the k-core of the low-resolution spatial surrogates, which are on average composed of only 52.3% of the network, or with the k-core (s-core) of the binary (weighted) high-resolution cortical network, which contains only 11.2% (10.0%) of the 998 regions. The difference between network resolutions may largely be attributable to the high degree of spatial concentration of the high-resolution s-core (and k-core) regions (Figure 6C and E). This concentration results in the collapse of large parts of the high-resolution core structure into single low-resolution regions of the cortex (specifically into the precuneus, the cingulate cortex and superior areas of the occipital lobe), the extremely dense internal connectivity of which is not accounted for during low-resolution analysis. Consequently, even the weighted, but especially the binary, cortical network, as observed on lower resolution, appear to exhibit a more distributed, homogeneous connectivity, with highly inhomogeneous intra-region connection densities, that are only accounted for at the higher resolution analysis. More generally, these findings underline the importance of multi-resolution analysis in cortical connectivity research in order to obtain a more complete and accurate picture on the inherently multi-level organisation of the connectome.
In conclusion, our surrogate analysis results extend those of Hagmann et al [12] by showing that the core structure of the high-resolution cortical network is both topologically and spatially significant, as measured by both k-core and s-core analysis. Furthermore, our findings on the low-resolution connectivity also indicate that this result depends on high-resolution analysis because the cortical connectivity becomes increasingly sparse and centralised at higher resolutions.
We next evaluated the tendency of the three additional cortical networks for the formation of the other putative ‘core’ structure, the rich-club. In line with the results on the weighted high-resolution connectivity, we obtained cortical network rich-clubs in the low- and high-resolution binary connectivities with strengths comparable to those of their spatial surrogates, and even somewhat weaker than those of their reduced surrogates, assessed by the traditional (unweighted) rich-club measure (see Figure 7).
In contrast to these results, we found a rich-club in the weighted low-resolution cortical connectivity that is statistically stronger than those of its spatial surrogates (one-sample t-test: p = 0.02, see Figure 7). Originating from its agglomerative construction from the high-resolution cortical network (see Methods), this finding may reflect the highly non-uniform (exponential-like) connection weight distribution of the weighted low-resolution cortical network. In essence, the surrogate rewiring process in the random and spatial surrogates of this cortical network, but not in its reduced surrogates, was effective in relocating the few very short-range, but extremely strong cortical connections to random positions in the network, resulting in a highly variable, but on average lowered weighted rich-club strength in these random and spatial surrogates (Figure 7, second row). (We note that by the nature of their definition, the rest of the weighted metrics investigated in this study, including the s-core structure, are largely immune to this kind of variation in the specific location of these few, extreme strength connections in the low-resolution weighted cortical network.) Nevertheless, the results indicate that the low-resolution weighted cortical network, in agreement with the other three connectivity types, demonstrates a significantly strong rich-club structure, by comparison with traditional random surrogates (one-sample t-test, p<10−4). Contrary to the other three connectivity types, however, the strength of the rich-club in the low-resolution weighted connectivity does not seem to be fully attributable to the spatial constraints of the cortex, as indicated by spatial surrogate comparison.
An analysis approximating fibre length by the Euclidean distance of the connected regions (see Methods) may disproportionately underestimate the length of the longer curved inter-hemispheric fibres, particularly those connecting homotopic regions around the cortical midline [8]. This, in turn, may result in an increase in the number of inter-hemispheric connections with underestimated lengths in the wiring constrained surrogate networks. Indeed, evaluating the proportion of intra- and inter-hemispheric connections in the cortical network and in the surrogate networks shows that while only 11.5% of the high-resolution cortical connections run between the hemispheres, this ratio increases to 13.2% for reduced, 18.1% for spatial and 50.2% for random surrogates. Some, although certainly not all, of these (surrogate) inter-hemispheric connections are likely to cause a corresponding underestimation in the connection length of reduced and spatial surrogates compared to that of the cortical network. This concern, however, is greatly eased by noting that the regions of the cortex along its midline are already highly intra-connected (see Figure 6), leaving only few potential places where such new connections can be formed. Indeed, calculating the mean (Euclidean) distance between inter-hemispherically connected region pairs DIH on high network resolution, we found an increase, rather than a decrease, in the DIH of spatial surrogates compared to that of the cortical network (DIHctx = 26.2 mm, DIHS = 38.8 mm). In comparison, we found, as expected, that the mean distance between the inter-hemispherically connected region pairs is somewhat lowered in reduced surrogates (DIHR = 22.2 mm) and greatly increased in random surrogates (DIHrnd = 86.2 mm). These results indicate that the newly created inter-hemispheric connections in spatial surrogates are predominantly between relatively distant regions, therefore suffer less from the disproportionate underestimation of connection length, as approximated by Euclidean distance, between homotopic regions along the cortical midline.
Nevertheless, in order to test our results against potential artefacts originating from the different degree of inter-hemispheric connectedness in the cortex and its surrogates, we repeated the analyses using a single cortical hemisphere. Specifically, we extracted the right hemisphere of the weighted high-resolution dataset, generated n = 20 surrogate networks for each of the three surrogate network types using the same method as before, and measured the complex network metrics assessed in the paper.
The results of single hemispheric analysis (Figure 7, third row) are largely in agreement with the previous bi-hemispheric analysis. The main differences are that the (hemi)-cortical network has an increased small-world index compared to reduced surrogates, and its rich-club is slightly but not significantly weaker than those of spatial surrogates (one-sample t-test: p = 0.1), and stronger than those of reduced surrogates. We note that if there was a significant bias in the full cortex surrogate networks to form an excess number of inter-hemispheric connections between homotopic midline regions, we would expect single hemisphere surrogate analysis to detect a consistent increase, rather than decrease, in the strength of s-core and rich-club structures, given the highly central positioning of these structures along the cortical midline in the full cortical network (see Figure 6). Due the fact that we observe such an increase in only one out of the four possible cases (the rich-club of reduced surrogates), we conclude that the single-hemisphere analysis validates the Euclidean approximation on fibre lengths for our surrogate analysis, and our main conclusions on the bi-hemispheric cortical network appear to largely apply to the uni-hemispheric cortical connectivity as well.
Standard models of complex network science in conjunction with the fundamentals of neuroscience shape the techniques we use for the analysis of brain networks. For example, the abstract concept of small-worldness has traditionally been defined in relation to random and lattice networks [23]. Thus, the diffuse nervous systems of coelenterates (such as Cnidaria) have long been recognised to exhibit a characteristically regular, lattice-like pattern [76]. These and other findings have contributed to the wide application of random and lattice surrogate techniques in brain network analysis. In this paper we have investigated how the use of more constrained null-hypothesis models, incorporating not only basic topological but also spatial properties of the human connectome, will help us better understand the structural organisation and functional operation of the inherently spatially and economically constrained brain.
We analysed a dataset representing the large-scale anatomical connectivity of the human cortex in order to confirm previously reported topological organisation patterns (network properties), such as small-worldness, modularity, hierarchy and core formation (see Figure 1), at both high- and low-resolution representations of cortical connectivity, and to then analyse the relationship of these patterns to the wiring constraints of the brain. To do so we devised two novel surrogate types, ‘spatial’ and ‘reduced’ surrogates that conserve the total length of connections from each region (spatial) or decrease it (reduced). For each network property, our analysis adopted the reasoning detailed in the Introduction (see also Table 1).
First, by comparing the cortical network and the spatially constrained surrogates to random surrogates, we found that cortical wiring constraints seem to contribute strongly to its relatively low potential for functional integration (as measured by global efficiency) and very high potential for functional segregation (as measured by clustering coefficient), and thus highly, although not fully (see below), account for the known small-world cortical organisation [57], [58]. In addition, comparison of the cortical connectivity network to the new surrogates also showed a relatively low level of global efficiency in the cortical network, closer to its reduced than to its spatial surrogates. Efficiency is a measure of functional integration potential in the network [15] and is generally most effectively increased by adding sparse long-range connections [18]. Because reduced surrogates to a great extent lack these long-range connections, our findings suggest that long-range cortico-cortical connections are in fact sub-optimally placed for maximising efficiency, and therefore, to the extent that brain structure determines function, they may not contribute to tight functional integration in the cortex as much as they could. In line with this, the cortical network was also found to be more similar to its reduced than to its spatial surrogates in its very high clustering coefficient. Functional segregation, facilitated by high structural clustering coefficient [15], is widely acknowledged to be a fundamental characteristic of the cortex [77]. Taken together, our findings indicate that the cortical network may possess an excess level of segregation and a relatively reduced level of functional integration potential over the extent that its wiring constraints alone can account for. Furthermore, spatial surrogates exhibited significantly weaker small-worldness compared to the cortical network, while reduced surrogates exhibited comparably high small-worldness to the cortical nework. These findings suggest that high cortical segregation combined with the concentrated spatial distribution of high degree regions (see Figure 6D) may suffice to ensure the strong small-world organisation of the cortical connectivity, even in the absence of long-range cortical connections.
Hierarchical organisation is believed to be a central architectural feature of various complex social networks and the World Wide Web [38] (Figure 1B). Hierarchical aspects of network architectures can fundamentally affect their evolution, development, adaptability and efficiency on multiple scales [61], [62]. The structural connectivity of the cortex is generally regarded to have a hierarchical organisation [59]. However, neither the degree and extent of hierarchical organisation, nor the constraints governing its expression, have yet been analysed in large-scale whole-brain networks as comprehensively as for instance the concepts of modularity or regional centrality [59]. This may be due to a lack of a consensus on the formal definition and assessment of this rather informal notion, in combination with currently available data being insufficiently detailed to evaluate and characterise the exact nature of this organisation pattern on a global scale [1], [78]. Here, we utilised the mean clustering coefficient as a function of degree, as a simple model for detecting hierarchical features in complex networks. The results indicated the presence of hierarchical organisation in the cortical network and in both spatially constrained surrogates, but not in random surrogates. One interpretation of this finding is that the predominantly local connectivity of the cortex and the central positioning of high degree regions give rise to the observed hierarchical structure. However, we cannot exclude an alternative explanation, namely that it is in fact the strong evolutionary pressure favouring the presumably functionally beneficial hierarchical organization, that led to the observed spatial embedding of cortical network nodes. Nevertheless, as both pressures, economical to conserve wiring cost and adaptive to achieve brain function, appear to benefit from a hierarchical organisation [18], [59], it seems most likely that their joint, mutual presence resulted in the observed hierarchical pattern in the structural connectivity of the cortex.
The brain's modular architecture is organised around spatially compact modules and their predominantly short, intra-module connections [77]. This feature of cortical connectivity is believed not only to keep global wiring costs low (economic pressure), but also to improve local communication efficiency within its structurally segregated and functionally specialised modular units (functional pressure) [18]. Our modularity analysis revealed that all spatially constrained networks indeed exhibit a strong and spatially compact modular architecture compared to random surrogates, indicating that basic wiring constraints of cortical regions naturally result in a tendency for cortical module formation. On the other hand, the high strength and granularity of the modular organisation of the cortex is more akin to its reduced surrogates, than to its relatively less modular spatial surrogates. This suggests that the long-range cortico-cortical projections may be more optimally placed towards a highly modular cortical architecture, than wiring constraints alone would suggest, supporting the widely acknowledged notion of high functional importance of cortical modules [21], [63], [64].
Furthermore, while the module partitions of the cortical network and its surrogates exhibit considerable differences, we found a set of cortical areas with modules that are highly preserved both in reduced and spatial surrogates. According to our analysis, the highly robust topological encapsulation of these predominantly lateral modules against the applied spatially constrained rewiring indicates that their existence can largely be explained by cortical wiring constraints. Additionally, however, the natural emergence of these module formations may enable them to provide a consistent base or ‘backbone’ to the cortex's modular structure both across individual variation and through development and ageing processes [24]. Such a modular ‘backbone’ structure could provide the structural basis for some relatively invariant, recurring components of the continuously reconfiguring functional networks of the brain [77].
While the exponential degree distribution [12] and hierarchical organisation already suggested a centralised organisation of cortical topology, we explicitly examined which, if any, parts of the cortex are located in its topological centre. Surrogate comparison revealed that the s-core of the high-resolution cortical network is stronger and larger than those of its spatial surrogates, and similar to those of its reduced surrogates. Furthermore, confirming previous results [12], the s-core of the cortical network was found to be spatially encapsulated at a medial-caudal location, composed by the precuneus, the cingulate cortex and the superior part of the occipital lobe. The cortical network, when observed on high-resolution (but not on low-resolution, see below), therefore appears to have a spatially compact, central s-core, the strength of which is significantly higher than its wiring constraints alone would suggest. One could interpret these findings to suggest that the cortical network s-core is not a by-product of wiring constraints but may instead be relevant for the brain's function; it might even serve the purpose of a putative central, global integrator substructure among the otherwise separate, functionally more specialised areas of the brain [79].
The other candidate central structure, the rich-club of the cortical network, also exhibits a significantly denser than random intra-connectedness, which is in agreement with previous studies detecting a well-expressed cortical rich-club structure [13], [17], [25]. However, in contrast to our results on the cortical s-core, we found rich-club structures of similar strength in the reduced and spatial surrogates. Thus, the rich-club formation of the cortical network appears to strongly correlate with its wiring constraints and the spatial distribution of the cortical hub regions (one of the ‘basic’ network property preserved in all surrogate ensembles). These findings extend earlier studies consistently discovering the brain's strong rich-club structure [13], [17], [25] by pointing to a plausible relationship between the remarkably dense inter-connectedness of high degree cortical regions and cortical wiring constraints. It is important to note, however, that similarly to the case of the hierarchy analysis, our method does not provide information with respect to the direction of causation between these network properties. Thus, it remains to be seen whether the economical pressure to conserve connection length is in fact the primary driving factor in the spatial arrangement of hub nodes, or the functional pressure for rich-club formation necessitates the specific spatial distribution of hub nodes in the cortex in the first place, and thus ultimately the formation of the cortical rich-club patterning.
Furthermore, compared to the s-core, the rich-club of the high-resolution cortical network was found to be internally relatively loosely coupled and formed by a spatially and topologically rather dispersed set of regions. These findings render the even spatially highly significant, well-confined and more tightly intra-connected cortical s-core a more appropriate candidate for a putative central cortical core [12], while the rich-club seems to be more suited for fulfilling the role of a ‘dynamic router’ [25], a set of distributed cortical hub regions predominantly connecting their local neighbourhoods with distant parts and the s-core of the cortex. Nevertheless, the large (50%) overlap between the s-core and rich-club regions suggests a great extent of functional cooperation between these highly intertwined, both topologically and spatially central cortical structures.
In line with these results, areas in the overlap between the s-core and rich-club structures of the cortex, the precuneus, the cingulate cortex and parts of the primary visual cortex (BA 17, 18), have also been repeatedly identified as global functional hubs of the human brain [80], [81], and found to functionally mediate between cortical areas that are structurally not directly connected [82]. Furthermore, some of the regions that belong to both the s-core and rich-club structures, most notably the precuneus, have also been highlighted as prominent areas of the default mode network of the brain [12], [83]. These findings suggest that the regions shared between the cortical network's s-core and rich-club, are not only topologically central, but also play a functionally pivotal role in coordinating, integrating or routing the activity of distant cortical regions in both resting and task-evoked states of the brain [25], [79], [83].
Figure 7 summarises the results on the investigated properties of the cortical network with respect to the three surrogate groups, at both network resolutions (998 regions at high-resolution and 66 regions at lower resolution), both for binary and weighted networks on each resolutions, as well as for the high-resolution weighted single hemisphere analysis (500 regions). First, comparing the metric values of the cortical network with those of its surrogates, we note that the cortical network tends to exhibit more similar values to its reduced than to its spatial surrogates for several network measures. One could argue that this may simply originate from the fact that the spatial surrogates are in general more randomised, and hence less similar to the cortical network, than reduced surrogates (see QS in Figure 7). However, while similarities in the expression of higher level network properties are certainly expected to be related to the extent of similarity on the lowest level of single connections, considering solely the overlap in the connection sets does not satisfactorily explain all observed tendencies. Indeed, as we showed in Results/Topological similarity, spatial surrogates are equally different from each other in their connection sets than from the cortical network, and yet their network properties are highly similar, but significantly different from that of the cortex. The overlap QS between connection sets alone is therefore not a good predictor of the obtained results, supporting our reasoning about the observed differences being attributable to the particular connectivity of the cortex – to its predominantly local connectivity and the specific arrangement of its long-range connections (see Table 1).
Secondly, Figure 7 assesses the consistency of our analyses across all investigated cortical network types (the five main rows of Figure 7). We start by noting that the results for several measures, most notably clustering coefficient, efficiency, small-worldness and modularity, are highly consistent across all investigated cortical network types. There is, however, some disagreement in the results of other complex network measures, specifically the k-core/s-core and rich-club metrics, across the various cortical networks. Generally, these disagreements indicate that the significance of the corresponding network properties (in terms of their relationship to the corresponding surrogate ensembles) may depend on the resolution the cortical network is observed at, or on the inclusion/exclusion of connection strengths (estimated number of fibres constituting the fibre bundles linking the regions), see detailed discussion in Results. Most notably, at the s-core/k-core metric, the strength of the cortical core only becomes visible in the high-resolution network, indicating a change in the organisation of the cortical connectivity at the different observable network resolutions and underlining the importance of multi-resolution approaches in connectome research. Specifically, on low resolution we found that the relatively weak cortical k-core is composed of 80% of the entire cortex, suggesting a more ‘homogeneous’ (non-centralised) connectivity between larger cortical regions on low network resolution. In contrast, on high-resolution the cortical network demonstrates a relatively small (10%), highly localised and significantly strong core structure, indicating a rather centralised organisation at the finer connectivity of the cortex. These findings are largely consistent with previous results on the s-core of the low-resolution [13] and high-resolution [12] cortical connectivity, and support the notion that, as we map the brain's network on increasingly higher resolutions, observed connectivity necessarily becomes sparser, leading in turn to the observation of fundamentally different organisation features at the various resolutions [78].
In this study, we focused on two distinguishable, supposedly competing factors that shape brain structure: economic pressure and functional pressure [18]. We note, however, that there are other important factors, such as evolutionary or developmental processes, that are likely to impose certain basic constraints on brain architecture [18]. Future extensions of this study may need to incorporate certain aspects of these further constraints, for example by generating surrogate networks via some neurobiologically informed developmental model [84]. It is also important to consider the accuracy of the cortical connectivity dataset used here. It is well known (and indeed increasingly articulated) that diffusion magnetic resonance imaging (dMRI) based tractography techniques suffer from certain biases and constraints, such as limitations in the ability to track fibre crossings and wide angular changes along the trajectory of the fibre tract [85], [86]. Therefore, in the current absence of comprehensive tract-tracing data in the human brain, it will be important that the hypotheses and computational findings of our study are tested against the increasingly complete and accurate maps dMRI techniques will be delivering in the future. Relatedly, it is likely that the spatially constrained surrogate analysis introduced in this study may give insights into the relative significance and potential origin of certain properties of the brain networks of other species, such as the cat [17] or the macaque [75], for which tract-tracing data is available.
Being a real complex network with a diverse and extraordinarily complex set of functions to carry out, it is not surprising that the cortex adopts and takes advantage of several functionally beneficial organisation patterns, even given the additional constraints imposed by wiring constraints [18]. Small-world architecture has been shown to naturally foster high dynamical complexity [9], [87], which is one of the hallmarks of brain activity [88] and has been associated with conscious states involving the efficient coordination of multiple sensorimotor modalities in generating flexible behaviour [89]. Modularity is widely acknowledged to promote network robustness and evolvability by minimising dependencies and isolating the effect of local mutations and disturbances [2]. It also has been shown to increase dynamical metastability [63] thus hindering the pathological cases of prolonged synchronisation and seizures [90] and again supporting functional flexibility [91]. Hierarchically modular organisation has been found to facilitate limited sustained network activity [92], it hence may serve a crucial role in maintaining the critical functional range in which the human brain operates [93]. Furthermore, the strong central core as well as the distributed and yet densely inter- and intra-connected rich-club structure may play a central role in facilitating efficient global functional integration and information flow in the cortex [13], [25], [74] hence providing the structural basis of various cognitive integration processes, from sensorimotor integration through attention to higher cognition and consciousness [77], [94]. Combining all these findings into a single description of the structural connectivity of the human cortex, our results outline a hybrid, reasonably centralised and hierarchical, but nevertheless strongly modular anatomical architecture, with a remarkably strong central network core.
Consistent discovery of characteristic network properties of the human connectome in this and previous studies emphasises a fundamental question: What factors contribute to the small-world, modular, hierarchical and centralised features of the cortical connectivity? Our results, extending those of earlier studies [51], [95], [96], support the notion that the emergence of these network properties is shaped by a complex interaction involving economic pressures (towards minimising wiring and running cost of the brain) and functional pressures (towards stable, reliable and adaptive operation of the brain) [18]. In this study we characterised how much each specific network property depended on the first of these factors, economic pressures, and we found that the level of dependency differed for different properties. Our results suggest that the more independent properties, such as the small-world, modular and core structure of the cortex, may be more related to the function of the brain than the more dependent ones, such as hierarchical organisation and rich-club patterning, which may be primarily driven by economic pressures. These results motivate further computational and experimental research to uncover the specific ways in which economic and functional pressures complement, reinforce or counteract each other in shaping the large-scale architecture of the human brain.
|
10.1371/journal.pntd.0003608 | Prevalence and Diversity of Small Mammal-Associated Bartonella Species in Rural and Urban Kenya | Several rodent-associated Bartonella species are human pathogens but little is known about their epidemiology. We trapped rodents and shrews around human habitations at two sites in Kenya (rural Asembo and urban Kibera) to determine the prevalence of Bartonella infection. Bartonella were detected by culture in five of seven host species. In Kibera, 60% of Rattus rattus were positive, as compared to 13% in Asembo. Bartonella were also detected in C. olivieri (7%), Lemniscomys striatus (50%), Mastomys natalensis (43%) and R. norvegicus (50%). Partial sequencing of the citrate synthase (gltA) gene of isolates showed that Kibera strains were similar to reference isolates from Rattus trapped in Asia, America, and Europe, but that most strains from Asembo were less similar. Host species and trapping location were associated with differences in infection status but there was no evidence of associations between host age or sex and infection status. Acute febrile illness occurs at high incidence in both Asembo and Kibera but the etiology of many of these illnesses is unknown. Bartonella similar to known human pathogens were detected in small mammals at both sites and investigation of the ecological determinants of host infection status and of the public health significance of Bartonella infections at these locations is warranted.
| Bartonella are bacteria that infect many different mammal species and can cause illness in people. Several Bartonella species carried by rodents cause disease in humans but little is known about their distribution or the importance of bartonellosis as a cause of human illness. Data from Africa are particularly scarce. This study involved trapping of rodents and other small mammals at two sites in Kenya: Asembo, a rural area in Western Kenya, and Kibera, an informal urban settlement in Nairobi. Blood samples were collected from trapped animals to detect and characterize the types of Bartonella carried. At the Kibera site over half of the trapped rats were infected with Bartonella very similar to human pathogenic strains isolated from rats from other global regions. In Asembo, Bartonella were detected in four of the five animal species trapped and these Bartonella were less similar to previously identified isolates. All of the small mammals included in this study were trapped in or around human habitations. The data from this study show that Bartonella that can cause human illness are carried by the small mammals at these two sites and indicate that the public health impacts of human bartonellosis should be investigated.
| Bartonella species are Gram-negative haemotrophic bacteria that infect mammalian erythrocytes and are transmitted between hosts by blood-sucking arthropods. Over 30 species of Bartonella have been described and members of this genus infect a broad range of mammalian hosts including rodents, bats, carnivores and ruminants [1]. Arthropod vectors including fleas, sandflies, lice, ticks, bat flies and ked flies are implicated in the transmission of these pathogens [2–4]. The genus Bartonella has a global distribution. The Bartonella elizabethae complex includes several Bartonella genotypes and strains (including B. elizabethae, B. tribocorum, B. rattimassiliensis and B. queenslandensis) that have been isolated from Rattus and Bandicota species around the world [1]. Recent analyses indicate that this complex has south-east Asian origins and has been globally dispersed by Rattus species [5].
Several Bartonella species are recognized as human pathogens that cause diverse clinical presentations [6]. Among rodent-associated Bartonella species, B. elizabethae is a known cause of human endocarditis [7]. Other rodent-associated species including B. tribocorum, B. vinsonii subsp. arupensis, B. washoensis and B. alsatica have been associated with a range of symptoms in humans including fatigue, muscle and joint pain, and serious complications, such as endocarditis and neurological signs, particularly in immunocompromised patients [8,9].
Bartonella species have been identified as important causes of febrile illness in some settings. In two studies conducted in Thailand, 15% of febrile patients were diagnosed with confirmed Bartonella infection based on a four-fold rise in antibody titres, and six different Bartonella species were identified by culture from blood clots collected from febrile patients [10,11]. Non-specific clinical signs and difficulties in culturing the organism present substantial challenges to the diagnosis of bartonellosis. Consequently, Bartonella species may well be under-recognized as a cause of human disease [12]. This is particularly true for Africa, where very few data on the etiology of febrile illness are currently available [13].
In the Democratic Republic of Congo (DRC), a seroprevalence study identified IgG antibodies against Bartonella (B. henselae, B. quintana or B. clarridgeiae) in 4.5% of febrile patients [14]. Bartonella bacteraemia was detected by PCR in 10% of HIV-positive patients in South Africa [15]. Apart from these studies however, there is little information on the impact of Bartonella on human health on the African continent.
A variety of Bartonella species have been detected in animal and ectoparasite populations in Africa. Considering rodents and small mammals specifically, B. elizabethae and two other Bartonella lineages were detected in Namaqua rock mice sampled in South Africa, where 44% of the 100 individuals sampled were positive by PCR for Bartonella species [16]. B. elizabethae, B. tribocorum and a Bartonella species with intermediate species classification based on sequence data were detected in 28% of rodents and hedgehogs (n = 75) sampled in Algeria [17]. B. elizabethae, B. tribocorum and novel Bartonella species were also detected in rodents sampled in the Democratic Republic of Congo (DRC) and Tanzania [18]. Small mammals trapped in Ethiopia, had an overall Bartonella infection prevalence of 34% (n = 529) and were infected with multiple genotypes including genotypes very closely related to B. elizabethae [19]. B. elizabethae has also been detected in invasive and indigenous rodents sampled in Uganda [20] and genotypes related to B. rochalimae, B. grahamii and B. elizabethae have been detected in Mearn’s pouched mice studied in Kenya [21]. Bartonella have also been detected in fleas collected in Egypt, Morocco, DRC and Uganda [20,22–24].
The first objective of this study was to determine the presence and prevalence of Bartonella infections in small mammals trapped at rural and urban locations in Kenya. We also aimed to characterize the Bartonella isolates obtained using partial sequences of the citrate synthase (gltA) gene and to compare the Bartonella genotypes detected in these distinct Kenyan populations with each other and with Bartonella detected in small mammals in other parts of the world.
Cross-sectional rodent trapping surveys were conducted within two locations: Asembo, a rural area on the northern shore of Lake Victoria in Nyanza Province western Kenya (Latitude-0.1443, longitude 34.3468) and Kibera, an urban informal settlement in Nairobi City (Latitude-1.3156, longitude 36.7820, Fig. 1). These locations are the study sites for ongoing population-based human health surveillance [25]. In Asembo, subsistence farming is the primary occupation for 65% of household heads, 13% work in the informal economy and 5% are salaried [25]. Households are clustered into compounds of closely related family units. Livestock ownership is common: 44% of Asembo households own cattle and 43% own at least one sheep or goat. In contrast, in urban Kibera, 53% of heads of household are salaried and 43% work in the informal sector [25]. Ownership of large livestock species in Kibera is very rare and prohibited by City Council law.
In Asembo, trapping was conducted over the period of July—August 2009. Traps were placed at 50 compounds that were a randomly selected subset of livestock-owning compounds enrolled in a larger study of zoonoses epidemiology [26]. Within each selected compound, five or six medium-sized foldable Sherman traps (H.B. Sherman Traps Inc., Tallahassee, FL) were placed for three or four nights. Traps were placed in three categories of habitat: within occupied dwellings; within outbuildings, which included unoccupied dwellings, stores, latrines or kitchens separate from the main dwelling; and outside, in areas within the compound yard. In Kibera, trapping was conducted over the period of September—November 2008. The overall study site was divided for this study into five trapping zones of similar area and within each zone a 50m x 50m trapping area was defined (Fig. 2). Within each of the five trapping zones, medium-sized foldable Sherman traps were placed for a minimum of two consecutive nights and a maximum of six nights with the aim of trapping approximately 50 rodents per zone. In Kibera, all traps were placed indoors at 270 occupied dwellings.
All trapped animals from both locations were euthanized by overdose of the inhalant anesthetic halothane and whole blood was collected by cardiocentesis using aseptic technique. Blood samples were processed to remove serum and the remaining blood clots frozen at -80°C prior to testing. Blood clots were shipped on dry ice to the Bartonella laboratory at the Centers for Disease Control and Prevention, Fort Collins, Colorado for laboratory testing. Morphometric data were collected from each trapped animal for species identification at the National Museums of Kenya. The Asembo small mammals were submitted for archiving under accession numbers NMK 171860—NMK 171922. The Kibera rodent population included in this study is as described previously [27].
Culture was performed using previously described techniques [28]. Briefly, blood clots were re-suspended 1:4 in brain heart infusion broth supplemented with 5% amphotericin B, then plated onto agar supplemented with 5% sterile rabbit blood and incubated at 35°C in an aerobic atmosphere of 5% carbon dioxide for up to 30 days. Bacterial colonies were presumptively identified as Bartonella based on their morphology. Subcultures of Bartonella colonies from the original agar plate were streaked onto secondary agar plates and incubated at the same conditions until sufficient growth was observed, usually between 5 and 7 days. Pure cultures were harvested and stored in 10% ethanol.
The identity of presumptive Bartonella isolates was confirmed by PCR amplification and sequencing of a specific fragment of the Bartonella citrate synthase (gltA) gene. Crude DNA extracts were obtained from bacterial cultures by heating a heavy suspension of the microorganisms. Two oligonucleotides (BhCS.781.p and BhCS.1137.n) were used as PCR primers to generate a 379-bp amplicon of the Bartonella gltA gene [29]. PCR products were separated by 1.5% agarose gel electrophoresis and visualized by ethidium bromide staining. Sequencing reactions were carried out in a PTC 200 Peltier Thermal cycler (Applied Biosystems; Foster City, California) using the same primers as the initial PCR assay.
Sequences were analysed using Lasergene 12 Core Suite (DNASTAR, Madison, WI) to determine sequence consensus for the gltA amplicons. Unique gltA sequences generated through this study were submitted to GenBank (accession numbers KM233484—KM233492). The Clustal V program within the MegAlign module of Lasergene was used to compare homologous Bartonella gltA sequences generated in this study with others available from the GenBank database. Phylogenetic trees were constructed using the neighbor-joining method with the Kimura’s 2-parameter distance model and bootstrap calculations were carried out with 1000 replicates. B. tamiae was used as the outgroup. A criterion of >96% homology was used to define similarity of study sequences to known Bartonella species [30].
Generalized linear models were used to examine associations between individual Bartonella infection status (culture positive or negative for Bartonella) and host and environmental variables in R (Version 3.0.3) [31]. Binomial family models with a logit link function were used and p values ≤ 0.05 were considered statistically significant. Variables examined included host species, sex, mass and trapping location. Data from the Asembo (S1 Table) and Kibera (S2 Table) sites were analysed separately.
Written informed consent for trapping was obtained from representatives of the study households. The protocols and consent forms were reviewed and approved by the Animal Care and Use and Ethical Review Boards of the Kenya Medical Research Institute (#1191). The study protocols were also approved by the Institutional Animal Care and Use Committee and Institutional Review Board of the U.S. Centers for Disease Control and Prevention (#5410) and complied with the Public Health Service Policy on Humane Care and Use of Laboratory Animals.
A total of 49 small mammals trapped at 29 compounds in Asembo and 220 rodents trapped at 143 households in Kibera were included in this study. The small mammals trapped in Asembo included Crocidura olivieri (n = 16, African giant shrew), and rodents of the species Lemniscomys striatus (n = 2, striped grass mouse), Mastomys natalensis (n = 14, Natal multimammate mouse), Mus minutoides (n = 1, pygmy mouse), and Rattus rattus (n = 16, black rat). All of the rodents trapped in Kibera were Mus musculus (n = 178, house mouse), Rattus norvegicus (n = 10, brown rat) or Rattus rattus (n = 32) (Table 1).
Ten of the 49 (21%) animals trapped in Asembo were culture-positive for Bartonella, including: Crocidura olivieri (n = 1, 7%); Lemniscomys striatus (n = 1, 50%); Mastomys natalensis (n = 6, 43%); and Rattus rattus (n = 2, 13%). Overall, 24 of the 220 (11%) animals trapped in Kibera were culture positive: including Rattus norvegicus (n = 5, 50%) and R. rattus (n = 19, 60%). None of the 178 samples collected from Mus musculus in Kibera were positive (Table 2). Culture-positive Rattus species were trapped in four of the five trapping grids established at the Kibera site (Fig. 2).
Information on gltA sequences was obtained from all 34 culture-positive animals. The phylogenetic relationships between the isolates obtained in this study and previously described Bartonella species are shown in Figs 3 and 4. Bartonella detected in one M. natalensis and two R. rattus trapped in Asembo belong to the B. elizabethae species complex based on the similarity of the gltA sequences (Fig. 3 & Table 2). Five additional sequences detected in M. natalensis and one detected in L. striatus were most closely related to B. tribocorum. None of the strains obtained from M. natalensis or L. striatus were identical (≥ 96% sequence identity) to reference strains of the Bartonella species described previously in Rattus species trapped elsewhere. The strain of Bartonella cultured from a C. olivieri was not very similar to any previously described Bartonella reference species but has 93.8% similarity B. birtlesii. Three pathogenic Bartonella species (B. elizabethae, B. tribocorum and B. queenslandensis) were detected in the two Rattus species sampled at the Kibera site. The gltA sequences for all Bartonella strains from Kibera rodents were identical (≥ 96% sequence identity) to reference isolates that are typical of Bartonella detected in Rattus populations globally (Table 2 & Fig. 4).
At the Asembo location, there was a weak association between individual infection status and host species (likelihood ratio test p = 0.053) where infection probability was higher in M. natalensis individuals than in the reference species C. olivieri (OR = 11.25, 95% CI = 1.15–110.47, p = 0.038). Approximately half (24/49) of the small mammals trapped in Asembo were trapped outside (Table 1). None of the Bartonella positive animals trapped in Asembo were trapped within occupied dwellings. Two positive R. rattus were trapped in outbuildings but all other positive animals (one C. olivieri, one Lemniscomys striatus and six M. natalensis) were trapped outside. There were no statistically significant associations between the probability of Bartonella infection and small mammal sex or mass within the Asembo population. At the Kibera location, there was a clear influence of genus upon infection probability. None of the 178 Mus trapped were Bartonella positive but 24/42 Rattus were positive indicating much higher infection probability in Rattus (OR = Infinite). Considering the data for Rattus individuals only, there were no statistically significant associations between the probability of Bartonella infection and rodent species, sex, or mass. The proportion of infected Rattus and proportion of infected rodents overall varied by trapping zone in Kibera (Fig. 2). There was no statistically significant difference in the probability of Bartonella infection in Rattus from different trapping zones. However the sample size for this analysis was small and the existence of zones where no positive individuals were trapped complicate this analyses and its interpretation. Descriptively, the trapping data from Kibera fall into two groups. In zones A, B and D few Rattus were trapped (Fig. 2), the rodent populations in these zones were dominated by Mus musculus and only four Bartonella isolates were identified in the combined rodent populations from these three zones (Fig. 2). In contrast, in trapping zones C and E, Rattus made up larger proportions of the total trapped population (51% in zone C and 40% in zone E) and more Bartonella isolates of several species were identified in these populations.
This study reports isolation of Bartonella strains from rodent and shrew species in Asembo and Kibera, Kenya. Bartonella strains were found in several small mammal species with variation observed in the infection prevalence and in the strains of Bartonella detected between host species and study sites. The majority of Bartonella isolates obtained from these Kenyan mammals are genetically similar to reference strains of known human pathogens.
Several recent studies indicate that the prevalence of Bartonella infection in Rattus in Africa may be low in contrast to the frequently high prevalences observed in Asian Rattus populations [18–20,32]. It has been argued that this pattern of lower prevalence in African Rattus populations could be attributed to host escape during colonization [19,20], a phenomenon where relatively small founding populations of invading species can leave their parasites behind when colonizing new areas [33]. Consistent with this, a relatively low prevalence was seen in R. rattus from Asembo (13%). However, the high infection prevalence observed in Rattus trapped at the Kibera site (e.g. 50% R. norvegicus and 60% R. rattus) is more similar to prevalence values observed in studies of Asian Rattus populations than to other African populations [19]. There are multiple possible explanations for the difference in the prevalence observed in Rattus at these two sites. Phylogenetic analyses indicate that B. elizabethae complex strains originated in Southeast Asia and have been disseminated throughout Asia, Europe, Africa, Australia and the Americas through multiple dispersal events of commensal Rattus species [5]. The Kibera study site is near the centre of Nairobi, the Kenyan capital, and is likely to have greater international connectivity (in terms of international rodent movement through trade etc.) than the Asembo site, which is more rural. The higher prevalence observed at the Kibera site could therefore be explained by repeated introduction of Rattus and their associated Bartonella species to this site [19,33]. Further analyses would however be needed to elucidate the colonization history of Rattus and their associated Bartonella at these sites specifically. The number of species trapped in Kibera was smaller than the number trapped in Asembo, indicating a simpler species composition at this site and these data could also suggest a possible dilution effect of the increased community complexity in Asembo on Bartonella prevalence [34]. Finally, temporal dynamics in host and ectoparasite population structure are known to affect Bartonella infection prevalence [21,35]. This study involved cross-sectional trapping surveys conducted at different times of year in the two study locations. There are few data on the seasonal variation in the abundance or diversity of the rodent populations at these sites but it is likely that there are seasonal influences upon rodent abundance and diversity with differences between the urban Kibera site and the more rural Asembo site in the seasonal population dynamics observed [36]. It is therefore possible that differences in the sampling time may have contributed to the differences in infection prevalence seen between these two Rattus populations.
Notably, no bartonellae were detected in Mus musculus trapped in Kibera despite a large number of tested animals and high infection prevalence observed in Rattus trapped in the same locations. Low-level Bartonella infection has been reported from Mus trapped in Ethiopia but the absence of Bartonella in Mus was also reported in a small-scale study from Nigeria [19,37].
All of the Kibera rodents were trapped within residents’ homes. In contrast, although nearly half of the animals trapped in Asembo were trapped indoors or in outbuildings, none of the positive animals at this site were trapped indoors, and only two positive animals were trapped in outbuildings. Approximately one third of the animals trapped outside in Asembo however were culture positive for Bartonella. The two culture positive animals trapped in outbuildings were R. rattus and they were carrying Bartonella similar to the B. elizabethae reference strain (Fig. 2). This species was most commonly found indoors or in outbuildings (14/15 records) and therefore may pose a risk due to closer human contact, even though only 2/16 were positive.
Many of the Bartonella detected at in this study (except the birtlesii-like isolates from Crocidura) belong to the Bartonella elizabethae complex and many of the strains identified in invasive Rattus hosts particularly are closely related to known human pathogens. All of the Bartonella strains isolated from Kibera rodents have ≥ 96% sequence identity with strains that are common in Rattus species sampled in Asia and on several other continents [5]. In contrast, several strains isolated from Asembo rodents and shrews were less similar to the international reference strains from Rattus and were more similar to isolates gathered previously from Ugandan rodents, suggesting a longer history of circulation of these strains within these species at the Asembo site. The identification of similar B. tribocorum sequences in Mastomys and Lemniscomys individuals trapped in Asembo suggests an absence of strong host-pathogen associations in these populations.
There is a high incidence of acute febrile illness in people in both Asembo and Kibera [25]. A variety of pathogens are known to account for a proportion of febrile illness in Asembo and Kibera but considerable proportions remain unexplained [38–41]. Bartonella species have been identified as important causes of human febrile illness in several global settings but there has been little investigation of the impact of bartonellosis upon human health in Africa particularly and it is conceivable that Bartonella may be an important cause of febrile illness in these study populations.
The data presented from the Asembo and Kibera sites indicate clear differences in: the prevalence of Bartonella infection in the same host (Rattus species) at the two sites; the prevalence of infection in different hosts trapped at the same sites; the abundance of different infected hosts between the two locations and also between trapping zones in Kibera; the strains of Bartonella detected and finally in the locations within communities where rodents overall and Bartonella infected rodents were trapped. The impact of this variation in rodent host community composition, infection prevalence, ectoparasite vector preferences, and other ecological factors need to be understood to evaluate human Bartonella infection risks at these sites. The data presented here suggest that investigations of the multi-host infection dynamics of Bartonella and the public health significance of Bartonella infections at these Kenyan locations and others where there are close associations between people and small mammals are warranted.
|
10.1371/journal.pgen.1006205 | Integrity of Narrow Epithelial Tubes in the C. elegans Excretory System Requires a Transient Luminal Matrix | Most epithelial cells secrete a glycoprotein-rich apical extracellular matrix that can have diverse but still poorly understood roles in development and physiology. Zona Pellucida (ZP) domain glycoproteins are common constituents of these matrices, and their loss in humans is associated with a number of diseases. Understanding of the functions, organization and regulation of apical matrices has been hampered by difficulties in imaging them both in vivo and ex vivo. We identified the PAN-Apple, mucin and ZP domain glycoprotein LET-653 as an early and transient apical matrix component that shapes developing epithelia in C. elegans. LET-653 has modest effects on shaping of the vulva and epidermis, but is essential to prevent lumen fragmentation in the very narrow, unicellular excretory duct tube. We were able to image the transient LET-653 matrix by both live confocal imaging and transmission electron microscopy. Structure/function and fluorescence recovery after photobleaching studies revealed that LET-653 exists in two separate luminal matrix pools, a loose fibrillar matrix in the central core of the lumen, to which it binds dynamically via its PAN domains, and an apical-membrane-associated matrix, to which it binds stably via its ZP domain. The PAN domains are both necessary and sufficient to confer a cyclic pattern of duct lumen localization that precedes each molt, while the ZP domain is required for lumen integrity. Ectopic expression of full-length LET-653, but not the PAN domains alone, could expand lumen diameter in the developing gut tube, where LET-653 is not normally expressed. Together, these data support a model in which the PAN domains regulate the ability of the LET-653 ZP domain to interact with other factors at the apical membrane, and this ZP domain interaction promotes expansion and maintenance of lumen diameter. These data identify a transient apical matrix component present prior to cuticle secretion in C. elegans, demonstrate critical roles for this matrix component in supporting lumen integrity within narrow bore tubes such as those found in the mammalian microvasculature, and reveal functional importance of the evolutionarily conserved ZP domain in this tube protecting activity.
| Most organs in the body are made up of networks of tubes that transport fluids or gases. These tubes come in many different sizes and shapes, with some narrow capillaries being only one cell in diameter. As tubes develop and take their final shapes, they secrete various glycoproteins into their hollow interior or lumen. The functions of these luminal proteins are not well understood, but there is increasing evidence that they are important for lumen shaping and that their loss can contribute to diseases such as cardiovascular disease and chronic kidney disease. Through studies of the nematode C. elegans, we identified a luminal glycoprotein, LET-653, that is transiently expressed in multiple developing tube types but is particularly critical to maintain integrity of the narrowest, unicellular tubes. We identified protein domains that direct LET-653 to specific apical matrix compartments and mediate its oscillatory pattern of lumen localization. Furthermore, we showed that the LET-653 tube-protecting activity depends on a Zona Pellucida (ZP) domain similar to that found in the mammalian egg-coat and in many other luminal or sensory matrix proteins involved in human disease.
| Most epithelial and endothelial tube cells secrete an apical extracellular matrix (aECM) or glycocalyx that lines the tube lumen and consists of a complex mix of gel-forming and fibril-forming glycoproteins, including both secreted and transmembrane proteoglycans, mucins, and zona pellucida (ZP) domain proteins [1–4]. There is increasing evidence that this aECM plays important developmental roles in shaping epithelial tubes [5–8]. For example, CD34/podocalyxin-family sialomucins are among the earliest markers of apical identity on developing lumens in vertebrates [9], and are important for vascular, lymphatic, kidney and gut tube integrity in mice [10–12], where they appear to play both anti-adhesive and signaling roles [13–17]. Proteoglycans or mucin-type glycoproteins expand lumen diameter in the sea urchin archenteron [18], C. elegans vulva [19] and in the Drosophila hindgut [20] and retina [21], consistent with a model in which hydration-mediated expansion of a gel-like aECM generates intraluminal pressure to drive circumferential lumen enlargement. A transient luminal cable consisting of the carbohydrate chitin and various ZP-domain proteins is required for uniform lumen expansion and/or tube integrity in developing Drosophila tracheal tubes [22–24], and has been proposed to function as a scaffold that connects to the apical membrane and resists and evenly distributes expansion forces generated by secretion-driven membrane growth [25]. ZP-domain proteins are also found within luminal matrices of the vertebrate vascular system, kidney and gut, and are mutated in diseases affecting the integrity of these tubes [26–29]. While these examples highlight the diverse and important roles that specific aECM components can play, we still have a very limited understanding of aECM composition in most tubes, or of how different aECM components interact and work together to influence tube shaping and maintenance.
We are using the C. elegans excretory system as a model to identify factors important for shaping and maintaining small, unicellular tubes. Unicellular tubes have an intracellular lumen and can be seamed (sealed by an autocellular junction) or seamless (lacking junction along the lumen) [30]. Unicellular tubes are prevalent in capillary beds of the mammalian microvasculature [31], and defects in their organization or maintenance may be associated with cardiovascular disease and stroke [32–35]. Narrower tubes may be particularly sensitive to defects in the aECM, or may have unique aECM components or organization that is tailored to their unique biophysical requirements [10,24,32,36].
C. elegans unicellular excretory tubes develop in the context of a luminal aECM that has been visualized by electron microscopy but characterized to only a limited extent [37,38]. This early aECM appears to differ molecularly between the largest and most internal excretory tube (the excretory canal cell) vs. the two smaller external excretory tubes (the excretory duct and pore cells). The duct and pore tubes, like all other external epithelia, eventually become lined with a collagenous cuticle that forms at the end of embryogenesis [39,40]. Prior to and during cuticle formation, external epithelia express the apically-localized transmembrane proteins LET-4 and EGG-6, members of the extracellular leucine-rich repeat only (eLRRon) family, which are required to maintain duct and pore tube integrity and for the barrier functions of the cuticle and/or eggshell [38]. The similarity of LET-4 and EGG-6 to mammalian small leucine-rich proteoglycans (SLRPs), which bind collagen [41], combined with phenotypes of the mutants, suggests that these proteins are components or regulators of the pre-cuticular and cuticular aECM, and that this aECM is critical for duct/pore integrity. However, other luminal aECM components in these tubes are largely unknown.
One proposed component of the excretory canal tube aECM is the ZP domain and mucin-like protein LET-653 [37,42]. Here we show that LET-653 instead is a component of a transient aECM that precedes cuticle secretion and is essential for excretory duct and pore tube integrity.
The C. elegans excretory system is a simple, renal-like organ that consists of three tandemly-connected unicellular tubes: the larger excretory canal cell and the smaller duct and pore cells that connect the canal cell to the outside environment for excretion [39,43] (Fig 1A). The canal and duct are seamless tubes, whereas the pore tube is sealed by an autocellular junction (AJ). The three tube cells also connect to each other and the external epidermis via ring-shaped intercellular junctions.
The excretory system is essential for osmoregulation [44], and fluid appears to flow from the canal cell through the duct and pore for excretion [43]. Luminal discontinuity within the pore or duct tubes leads to fluid retention and a characteristic “rod-like” larval lethal phenotype that first manifests as a dilation within the upstream duct or canal lumen [38,45,46] (Fig 1B). To find additional genes important for tube development or maintenance, we conducted an EMS mutagenesis screen for recessive, rod-like lethal mutants with this phenotype, using a strain carrying fluorescent markers for apical junctions (AJM-1::GFP) and for the duct and pore cell bodies (dct-5pro::mCherry) in order to visualize duct and pore morphology (Fig 1C and 1D, Materials and Methods). This screen identified 85 lethal mutants with excretory abnormalities, including two new alleles of let-653 (Fig 1B, 1D and 1E), a gene that had previously been proposed to encode a component of the canal cell glycocalyx [37,42].
let-653 encodes several protein isoforms that each contain a signal peptide, two PAN-Apple domains, and a C-terminal ZP domain. PAN-Apple domains are putative protein or carbohydrate interaction domains found in plasminogen and clotting factor XI, as well as in various invertebrate cuticle proteins [47,48]. ZP domains are polymerization-competent domains found in many apical matrix proteins, including invertebrate cuticle proteins [3,48–50]. The LET-653 ZP domain consists of a ZP-N subdomain with the typical four cysteines, and a ZP-C subdomain with eight cysteines, including six that are conserved among many ZP proteins and two that are unique to LET-653 (Fig 1E, S1 Fig). A large linker region between ZP-N and ZP-C varies in size and sequence between LET-653 splice isoforms and resembles mucins in that it is rich in proline, serine and threonine residues and predicted to be highly O-glycosylated [42] (Fig 1E, S1 Fig). LET-653 is also N-glycosylated at multiple sites [51] (S1 Fig). LET-653 does not have a transmembrane domain or GPI anchor as assessed by SMART, TMpred and big-Pi [52,53], but it does have a consensus cleavage site (CCS) C-terminal to the ZP domain, similar to other membrane-anchored ZP proteins (Fig 1E). Cleavage at a CCS is thought to be a prerequisite for ZP polymerization and fibril formation [54–57]. In summary, LET-653 contains multiple different domains that are typical of apical ECM factors in all animals.
let-653(cs178) is a nonsense mutation at codon 54 that is predicted to truncate all LET-653 proteins within the first PAN-Apple domain (Fig 1E); this allele is likely to be null. let-653(cs204) is a splice donor mutation, and the original reference allele s1733 is a nonsense mutation at codon 250; these are predicted to truncate LET-653 proteins near the beginning of the ZP domain (Fig 1E). All three let-653 alleles are recessive and cause 100% penetrant larval lethality (Fig 1F). This lethality could be rescued by cDNAs corresponding to the shortest let-653 isoform, let-653b (Fig 1F); therefore we used that isoform throughout this study. Mutants for each of the three let-653 alleles have very similar excretory phenotypes, with luminal dilation occurring in both the excretory canal cell and duct, variable absence of the pore AJ, and detachment of the duct and pore cells (Fig 1D; see below), showing that LET-653 is essential for shaping all three tubes in the excretory system.
let-653 reporter analyses and tissue-specific rescue experiments indicated that let-653 functions within the excretory duct and pore rather than in the canal cell (Fig 1F, S2 Fig). A let-653 promoter::GFP transcriptional reporter was expressed in external (cuticle-producing) epithelial cells, including the epidermis, vulva, rectum and excretory duct and pore, but was mostly excluded from internal epithelia such as the pharynx, intestine and excretory canal cell (S2 Fig). We did occasionally observe let-653pro::GFP expression in the canal cells of embryos, but this expression disappeared in later stages (S2 Fig) and we never observed LET-653 fusion proteins in the canal cell (see below). The lethality and excretory defects of let-653 mutants were efficiently rescued by let-653 transgenes expressed in the excretory duct (lin-48 promoter), the excretory pore (dpy-7 promoter) or both cells (lpr-1 promoter), but very inefficiently or not at all rescued by transgenes expressed in the excretory canal cell (glt-3 promoter) or body muscle (unc-54 promoter) (Fig 1F). Weak rescue using the canal promoter transgene could be attributed to movement of LET-653 protein from the canal cell into the duct and pore (S2 Fig, see below). We conclude that let-653 acts locally within the excretory duct and pore, and that the canal cell defects described previously [37,42] are a secondary consequence of duct and pore defects.
To better understand the requirements for let-653 in the excretory duct and pore, we examined apical junctions (marked with AJM-1::GFP) and the apical membrane (marked with RDY-2::GFP; [58]) or apical actin (marked with dct-5pro::VAB-10ABD::GFP; [59]) in staged embryos and larvae (Fig 2). let-653 mutants appeared normal through the early 3-fold stage (Fig 2A and 2G), indicating that the duct and pore had successfully wrapped to form polarized tubes and that let-653 is not required to establish cell junctions or to form an initial lumen. Lumen dilations first appeared in the duct between the early-and mid 3-fold stages (Fig 2A–2C), several hours before the bulk of cuticle secretion [40]. Much later, around the time of hatching (and after duct dilations had become quite severe), the pore AJ disappeared and the pore detached from either the duct or its ventral epidermal partner G2, while remaining attached to the other (Fig 2A and 2K–2N).
We focused our analyses on the duct, since mutant defects first appeared in this cell. Examination of wild-type embryos revealed two phases of duct lumen morphogenesis (Fig 2A and 2D–2F). During phase I, between the 1.5-fold and early 3-fold stages, the duct lumen increased in both length and diameter. During phase II, between the early and late 3-fold stages, the duct lumen elongated even further, but slightly narrowed in diameter, suggesting apical constriction. By the time of hatch, the duct lumen measured ~14 microns long but was less than 400 nm in diameter.
In let-653 mutants, the duct lumen did initially elongate and widen to some degree, but lumen diameter appeared increasingly irregular during the phase II of elongation (Fig 2A and 2G–2H). Lumen dilation always initiated near the intercellular junction between the duct and canal cells (Fig 2C and 2H), and coincided with a very thin and/or fragmented appearance of the lumen in the more distal part of the duct (Fig 2H and 2I), where a very narrow cellular process, only slightly wider than the lumen, normally connects the duct to the pore (Fig 2F). By L1, this narrow process often disappeared completely and the duct and pore appeared detached (Fig 2L–2N). The continued presence of RDY-2::GFP and F-actin at the remaining proximal (now dilated) duct apical membrane, as in WT, indicated that cell polarity remained intact (Fig 2D–2K). These observations suggested that lumen dilations occur as a result of fluid accumulation due to lumen blockage.
Transmission electron microscopy (TEM) analysis of serial sections in one let-653(cs178) 3-fold embryo and two let-653(s1733) L1 larvae confirmed duct lumen discontinuities and revealed multiple, small membrane-bound compartments (3/3 animals) instead of a single lumen (Fig 3). In mutant L1s, remaining segments of duct lumen were lined by abnormally small or abnormally large rings of cuticle-like material (Fig 3E and 3F), indicating that cuticle secretion had proceeded after lumen fragmentation and dilation. The pore lumen remained intact in the embryo and one of the L1 mutants, and had a normal diameter (Fig 3C and 3G). The canal lumen was greatly dilated immediately upstream of the duct-canal junction (Fig 3D), but otherwise intact and fairly normal further upstream of this connection point. It therefore appears that the duct tube is particularly sensitive to let-653 loss, and that duct and canal lumen dilation occurs as a secondary consequence when excretory fluid flowing from the canal cell backs up behind a distal duct luminal discontinuity, as has been suggested for other mutants [38,45]. We conclude that LET-653 is needed to maintain duct apical membrane integrity and a continuous, open lumen during elongation and narrowing.
Consistent with the observation that let-653 is widely expressed in external epithelia (S2 Fig), let-653 mutants exhibited additional defects in shaping of the epidermis and vulva. The epidermis and its overlying aECM normally undergo circumferential apical constriction to shape the embryo into an elongated worm (Fig 4A) and to form specialized cuticular ridges, termed alae, that run along its lateral surfaces (Fig 4C) [50,60,61]. When rescued to viability with a heterologous duct promoter construct (Fig 1G, S2 Fig), let-653 mutants had a moderately short and fat (Dumpy) body morphology (Fig 4B). Furthermore, TEM analysis revealed that the L1 alae of let-653 mutants were flatter and wider than in WT (Fig 4C and 4D). The internal striated layer of the cuticle was present but thinner than normal, and the space between the cuticle and the epidermis was expanded (Fig 4C’ and 4D'). The Dumpy and flat alae defects are similar to (but less severe than) those reported for mutants in ZP domain cuticulins [50], and suggest a defect in apical constriction of the lateral epidermis, possibly due to failure to appropriately connect the epidermis to the cuticle. Despite these defects in body shaping and cuticle morphology, the barrier function of the cuticle remained largely intact in let-653 L1 larvae (Fig 4C).
Expression of let-653 was also observed in the vulva (S2 Fig), which is a large multi-cellular tube used for egg-laying. Shaping of the vulva lumen requires both an expansive force generated by chondroitin proteoglycans [19,62], and a constrictive force generated by circumferential actomyosin contraction [63]. Duct-rescued let-653 mutants exhibited variable abnormalities in the shape of the vulval lumen, including asymmetries and general under-expansion along the dorsal-ventral axis (Fig 4F–4H). These defects were incompletely penetrant and much milder than those reported for either chondroitin biosynthesis mutants such as sqv-5 (Fig 4I) [62] or Rho kinase mutants [63], but suggest a role for LET-653 in expanding vulval lumen dimensions.
The non-transgenic progeny of duct-rescued let-653 animals arrested as L1 larvae with the typical let-653 excretory phenotype. Therefore, let-653 is not required for other chondroitin- or aECM-dependent processes such as eggshell formation [64], cytokinesis [65], or pharyngeal-epidermal attachment [66].
To visualize the LET-653 protein, we used the let-653 promoter to express LET-653 either N- or C-terminally tagged with Superfolder (Sf) GFP, which has been reported to fold stably in oxidizing extracellular environments [67]. Both fusions rescued let-653 mutant lethality (Fig 5A and 5B), indicating that the tagged proteins are functional.
To test if LET-653 is cleaved as described for other ZP proteins, we examined the fusion proteins via Western blotting. Western analyses of lysates from transgenic embryos failed to detect the full-length fusion proteins, but did consistently detect ZP domain fusions lacking the PAN domains as well as release of the C-terminal, but not the N-terminal, SfGFP tags (Fig 6). These data confirm that LET-653 can be cleaved, likely at the CCS. Furthermore, as expected if significantly O-glycosylated, the N-terminally tagged LET-653(ZP) fusion ran at ~120kD despite a predicted molecular weight of only ~80kD. Surprisingly, both N- and C-terminally tagged LET-653 fusions showed very similar patterns of localization in vivo that were different from the localization of secreted (ss) SfGFP alone (Fig 5). Therefore, either a significant proportion of LET-653 must remain uncleaved, or the cleaved portion must remain non-covalently associated with the mature form. The latter model is consistent with in vitro studies of mammalian ZP3, which showed very slow dissociation of the C-terminal region after ZP3 cleavage [68].
LET-653 translational fusions were apically secreted and accumulated transiently in cycles that preceded cuticle deposition (Fig 5). Beginning at ventral enclosure, LET-653 fusions accumulated throughout the region between the embryo and the inner layer of the eggshell (S2 Fig), indicating apical secretion prior to and during formation of the embryonic sheath, the first known layer of the cuticle [60,69]. By the 1.5-fold stage, LET-653 fusions (but not ssSfGFP alone) also accumulated within the newly-formed lumen of the excretory duct and pore (Fig 5D and 5E). LET-653 remained extra-embryonic and luminal for the next 3–4 hours, but then disappeared prior to the bulk of cuticle secretion (Fig 5A and 5B). This temporal pattern indicates that LET-653 is present throughout phase I of duct lumen growth, but disappears during phase II, as the lumen narrows. Therefore, LET-653 may act during the first phase of lumen growth to protect against lumen fragmentation in the second phase.
LET-653 reappeared transiently within the duct and pore lumen in the latter part of each subsequent larval stage (Fig 5A, 5B, 5G and 5H). It did not detectably incorporate into the body cuticle, but accumulated in the space between the new and old cuticles during molting (S2 Fig), and within the vulval luminal space during the mid-L4 period when the epidermal cuticle separates that luminal space from the outside environment (Fig 5J and 5K). In the vulva, where resolution was highest due to the large size of the luminal cavity, both LET-653 fusions were enriched near apical membranes and also present in wispy fibrils within the central lumen cavity; these fibrils extended along the ventral border of the lumen to contact the apical membrane of the vulB cells (Fig 5J and 5K). ssSfGFP alone appeared uniformly distributed (Fig 5L), indicating that the LET-653 pattern is specific. We conclude that LET-653 is part of an early and transient pre-cuticular aECM.
We compared the LET-653 localization pattern in the vulva to archival TEM images of the vulva lumen (Fig 7A). TEM reveals a prominent matrix layer that lines the luminal membrane; this layer is thickest at the dorsal apex of the vulva, near the vulE and vulF cells, but continuous along the entire luminal membrane, and has a discrete, darkly staining outer border. A loosely organized fibrillar matrix, which is variably preserved in different sections, fills the ventral portion of the lumen and abuts the apical matrix near the vulA-C cells. The LET-653 localization pattern closely matches these two matrix compartments observed by TEM (Fig 7B and 7C). The excretory duct and pore tubes are too narrow to resolve different matrix compartments by confocal imaging, but our previous TEM data are consistent with a similar two-compartment organization of the pre-cuticular aECM in those tubes [38]. In summary, in the vulva and likely in the excretory duct and pore tubes, LET-653 localizes transiently to two distinct pre-cuticular aECM compartments.
The oscillatory pattern of LET-653 accumulation in the duct and pore lumen may result in part from transcriptional regulation, since published RNAseq data show that let-653 transcripts oscillate with the molt cycle during larval development, with peak expression in the inter-molt period prior to cuticle collagen expression [70]. However, LET-653::SfGFP driven by the constitutively active lin-48 (duct) promoter also showed an oscillatory pattern of accumulation within the duct and pore lumen (S2 Fig), showing that post-transcriptional regulation also occurs. Furthermore, LET-653::SfGFP driven by the constitutively active glt-3 (canal) promoter accumulated within the canal lumen (and sometimes within the duct and pore lumen; S2 Fig) only during embryogenesis, and was rarely if ever detected in larvae, possibly explaining the minimal rescue activity seen with glt-3 promoter transgenes (Fig 1G). Finally, secreted ssSfGFP alone, when driven by the let-653 promoter, never accumulated to detectable levels in the duct and pore lumen (Fig 5C, 5F and 5I). These observations suggest that LET-653 luminal retention depends upon the correct site of synthesis and requires interactions between LET-653 and some anchoring factor(s).
To identify domains of LET-653 important for function, localization and cycling, we analyzed N-terminally tagged fusion proteins lacking the PAN, ZP or C-terminal domains (Fig 8). A LET-653 variant truncated at the CCS (LET-653∆C) was non-functional and retained within cells (Fig 8A, 8D, 8G and 8J), consistent with prior studies showing that the proper trafficking and apical secretion of ZP proteins requires hydrophobic sequences C-terminal to the cleavage site [55,57], which form an integral part of the non-polymerized ZP-C fold [56,68]. Indeed, shorter variants lacking the entire ZP-C subdomain were normally secreted (Fig 8B, 8E, 8H and 8K, S3 Fig), indicating that the C-terminus is essential for secretion only in the presence of the ZP-C domain. These truncated LET-653 constructs were non-functional in let-653(cs178) rescue assays (Fig 8B and S3 Fig). In contrast, a LET-653 variant lacking the PAN domains (LET-653(ZP)) was still partly functional (Fig 8C). We conclude that the ZP domain is critical for function, but the PAN domains are at least partly dispensable.
LET-653(ZP) and LET-653(PAN) fusions showed distinct patterns of spatial and temporal localization. LET-653(ZP) accumulated in the duct lumen of embryos and then was downregulated normally, but it was not detectably retained within the duct lumen of L1 larvae (Fig 8C, 8F and 8I). In the vulva, LET-653(ZP) specifically associated with the dorsal portion of the apical membrane (Fig 8L), a small subset of the locales normally occupied by the full-length protein (Figs 5J, 5K and 7B). Conversely, LET-653(PAN) showed a normal cyclic pattern of accumulation in the duct lumen of both embryos and larvae (Fig 8B, 8E and 8H), and labeled vulval luminal fibrils but was not enriched near apical membranes (Fig 8K). Similar localization was observed for constructs also retaining the mucin and/or ZP-N domains (S3 Fig), indicating that the C-terminal portion of the ZP domain is required for membrane-proximal localization.
We draw several conclusions from these data. First, the main role of the PAN domains is to help anchor the functional ZP domain in the correct location, perhaps preventing its flow-induced displacement. This role is less critical in the embryonic duct, where flow is presumably weaker, but is both necessary and sufficient for cyclic duct lumen retention in larvae. Second, the PAN and ZP domains interact with different binding partners in different locations–the PAN domain binds to a partner in the luminal core, while the ZP domain binds to a partner near the apical membrane. Third, ZP-C-dependent interactions at the apical membrane during tube morphogenesis correlate with LET-653 duct lumen protective activity. Finally, LET-653 reappearance at later stages is apparently dispensable.
The above data indicate that LET-653 is present during periods of lumen expansion and is required for proper lumen dimensions. To test if LET-653 is sufficient to expand lumen diameter, we used a heat shock promoter to express LET-653 in the gut tube, where it is not normally present. The gut tube is not cuticle-lined and therefore relevant LET-653 partners also may be absent. Full-length LET-653 could expand gut lumen diameter more than two-fold (which equates to a four-fold increase in tube volume), whereas LET-653(PAN) and the other truncated constructs lacking the ZP-C domain had either no effect or a very modest effect (Fig 9). Lumen expansion was irregular along the gut tube, suggesting localized action rather than a uniform hydrostatic force. We conclude that LET-653 has a ZP-dependent lumen-expanding activity, and this activity cannot be explained solely by the presence of the mucin-like linker or by effects on cuticle organization.
ZP-containing fibrillar matrices typically are quite stable [25]. To test the stability of LET-653-mediated interactions, we assessed the mobility of LET-653 fusions using fluorescence recovery after photobleaching (FRAP) experiments. We performed these experiments in the duct lumen at two different developmental stages (1.5-fold and late L1) (Fig 10, S4 Fig), and also in the L4 vulva, where we could compare apical membrane-localized vs. centrally-located pools of LET-653 (Fig 11). In the duct lumen, mobility of the full-length ssSfGFP::LET-653 fusion decreased between the 1.5-fold and the L1 stages (Fig 10D and 10G), and at L1, the mobility of this fusion was significantly lower than that of ssSfGFP::LET-653(PAN) (Fig 10E–10G). These data indicate that PAN-mediated interactions are relatively dynamic. Furthermore, although the ZP domain is not sufficient to confer duct localization at the L1 stage, it does contribute to more stable interactions. In the vulva, the apical membrane-associated pool of full-length LET-653 was significantly less mobile than the centrally-located pool of either full-length or LET-653(PAN) (Fig 11A–11D and 11G–11I), again suggesting that ZP-mediated interactions are quite stable, while PAN-mediated interactions are more dynamic.
Because cleavage at the CCS and dissociation of the C-terminal region has been proposed to be essential for ZP fibril formation, we compared the above results to that of a C-terminally tagged LET-653(ZP) fusion to ask if C-terminus retention affects LET-653 mobility. Consistent with earlier results demonstrating that N- and C- terminally tagged LET-653 localized similarly (Fig 5), LET-653(ZP)::SfGFP also localized to the vulva apical membrane and behaved identically to the apical membrane pool of full-length ssSfGFP::LET-653 in FRAP experiments (Fig 11E and 11G–11I).
The let-653 duct lumen fragmentation and G1 pore junction loss phenotypes are very similar to those caused by loss of the apical eLRRon protein LET-4 [38] or the secreted lipocalin LPR-1 [45]. However, lumen abnormalities appeared somewhat later in the other mutants compared to let-653 (Fig 12A). Loss of let-4 or lpr-1 did not perturb LET-653::SfGFP apical secretion or duct luminal retention in embryos (Fig 12B and 12C).
To test the functional relationship between let-653 and let-4 or lpr-1, we analyzed double mutants (Fig 12D). At an early timepoint where each single mutant had lumen defects at low to moderate penetrance, let-653; let-4 or lpr-1; let-4 double mutants were not significantly worse than single mutants. In contrast, lpr-1; let-653 double mutants displayed a significantly more penetrant lumen phenotype. These data suggest that let-653 and lpr-1 act in parallel and cooperate to maintain lumen integrity at early stages, while let-4 may act at a later step of cuticle matrix assembly.
There is increasing recognition of the importance and disease relevance of the aECM or glycocalyx, which lines the apical surfaces of epithelia and the luminal surfaces of cellular tubes. Nevertheless, the aECM has been much less studied than the basal ECM, in part due to challenges in visualizing the aECM both in vivo and ex vivo. Relatively little is known about the mechanisms underlying aECM functions, how different components of the aECM interact with each other and with the apical membrane, and how the aECM is assembled and cleared. This work identifies the PAN-Apple, mucin and ZP domain glycoprotein LET-653 as a component of a transient apical glycocalyx that precedes cuticle secretion in C. elegans. Although LET-653 is broadly expressed and affects apical domain shaping of multiple epithelial tissues, including the epidermis and vulva, it is particularly critical to maintain patency and integrity of the very narrow unicellular excretory duct and pore tubes. Different domains within LET-653 localize the protein to distinct matrix pools, and ZP-dependent interactions near the apical membrane are critical for lumen shaping and integrity during a period of dramatic tube elongation and narrowing.
The cuticle or exoskeleton of invertebrates is a particularly prominent example of apical ECM. In C. elegans, the mature cuticle consists primarily of collagens, but also includes various other insoluble proteins, termed cuticulins, several of which contain a ZP domain [40,50]. The mature cuticle is a multi-layered structure that is built via sequential secretion of different matrix components, and it has been presumed that the earliest secreted layers end up on the most apical surface [69,71]. The earliest observed layer, the embryonic sheath, forms between the bean and 1.5-fold stages [60]. We showed that LET-653 is secreted apically prior to formation of the embryonic sheath, but then disappears prior to the bulk of cuticle secretion. We found no evidence that LET-653 persists as part of the epicuticle. However, LET-653 absence does affect epidermal cuticle structure and alae morphology, consistent with the idea that the early glycocalyx affects deposition or organization of subsequent matrix layers.
We were able to visualize the early aECM in the vulva by both live confocal imaging and TEM. In this large, multicellular tube, LET-653 localizes to two distinct pre-cuticular aECM compartments (Fig 13A). LET-653 interacts dynamically via its PAN domains with a loose, fibrillar matrix in the center of the lumen, which corresponds roughly to the region where chondroitin proteoglycans (CPGs) are most abundant [72]. LET-653 interacts more stably via its ZP domain with a matrix compartment along the apical membrane, where cuticle assembly will later occur. This compartment may be equivalent to the embryonic sheath in the epidermis, although it is much thicker than the sheath in some regions. At the base of the vulva lumen, the two matrix compartments are connected via LET-653(PAN)-decorated fibrils. The specific partners that LET-653 interacts with in these two compartments remain to be identified, but LET-653 domain-specific reporters will be excellent tools for testing candidate partners.
TEM data [38] and our PAN- and ZP-domain localization data are consistent with a similar two-compartment aECM organization during early excretory duct and pore tube development. The LET-653 PAN domain is both necessary and sufficient to confer a cyclic pattern of localization to the duct/pore lumen, but the ZP domain is not, suggesting that a PAN-mediated interaction with the central luminal matrix captures LET-653, preventing its outflow from the lumen and allowing other ZP-mediated interactions to occur (Fig 13A). The cyclic reappearance and disappearance of LET-653 in the larval duct/pore lumen is apparently not required to maintain tube integrity, but likely reflects other dynamic changes in the matrix environment that occur surrounding each molt. Disappearance of LET-653 from the duct lumen could result from clearance of its relevant binding partners or severing of its connections to those partners.
Although LET-653 has often been cited in the literature as an example of mucin-dependent tube shaping (for example, see [5]), we found that the mucin-like region was not sufficient to rescue duct/pore integrity defects in let-653 mutants, even when tethered in the lumen by the PAN domains. Furthermore, the mucin-like region did not affect LET-653 matrix localization or gut lumen expansion. While these observations do not exclude a relevant role for the mucin-like region, they do show that tube protective activity and lumen expansion require other portions of the LET-653 ZP domain.
The ability of the LET-653 ZP domain to localize to the membrane-proximal matrix compartment correlates absolutely with duct/pore protective activity. Since the duct lumen elongates significantly over the period that LET-653 is present, the LET-653 matrix must be quite flexible or capable of rapid remodeling to accommodate this elongation. Glycosylated mucin domains can form rod-like structures that extend over long distances [73,74], so one possibility is that progressive glycosylation of this domain allows it to stretch. The mechanisms driving duct elongation are not known, but the duct is likely subjected to an external stretching force as it has apical junctions to partners (the canal and pore tubes) that move apart as the embryo elongates, gains motility and begins to flex its body. Such stretching forces and/or additional constrictive forces within the duct cell could contribute to narrowing of the duct lumen. A LET-653 ZP-dependent expansion activity is needed to counteract such narrowing and maintain a patent lumen with a uniform diameter. In let-653 mutants, the lumen still forms and expands to some degree, but it is irregular in diameter and fragments during the narrowing phase (Fig 13B). Because LET-653 is already disappearing as lumen narrowing occurs (Fig 13B), we propose that the membrane-anchored LET-653 matrix not only directly expands and constrains lumen dimensions during early tube morphogenesis, but also serves as a scaffold for assembly of later cuticle matrix layers that are important for continued lumen integrity.
ZP domains typically confer the ability of proteins to homo- or hetero-polymerize into mesh-like fibrils [49]. The mechanisms that regulate ZP matrix assembly are not well understood, but interactions among the ZP-N, ZP-C and C-terminal subdomains appear to hold the protein in a polymerization incompetent conformation to allow trafficking through the secretory pathway [55–57,68,75]. Once at the plasma membrane, cleavage at a CCS to remove the C-terminus is thought to eliminate those inhibitory interactions to allow polymerization [54,55,57]. Prior to or concomitant with cleavage, interactions between the C-terminus and other factors may also facilitate proper matrix assembly [75].
Our structure/function data indicate that the LET-653 C-terminus is required for proper secretion, C-terminal cleavage does occur, and the region encompassing the ZP-C and C-terminal domains is essential for apical localization and duct protective activity. Therefore, a reasonable model is that the stable pool of LET-653 near the apical membrane consists of ZP-dependent polymers whose localization and anchoring depend on unknown ZP-C- or C-terminal domain interactors (Fig 13A). We note, however, that there is also evidence for non-polymerizing, signaling roles of some luminal ZP-domain proteins, such as mammalian Endoglin and Betaglycan [26,29], as well as for mucins [15–17,76]. Furthermore, LET-653 lacks otherwise conserved aromatic residues in the ZP-N domain that are thought to be important for polymerization (S1 Fig) [77], and the C-terminal domain remains stably associated with the membrane-proximal LET-653 matrix rather than dissociating as expected (Fig 12E). Therefore, other potential models are that LET-653 affects assembly of other matrix proteins or interacts with transmembrane factor(s) to affect signaling and/or cytoskeletal organization.
Why is LET-653 essential for duct/pore integrity, whereas it has only modest effects on shaping of larger tubes such as the vulva? The distinction could merely reflect redundancy, if other aECM factors, such as chondroitin proteoglycans, compensate for the absence of LET-653 in some tissues but not others. However, other observations also suggest that narrow tubes may be particularly sensitive to disruptions in ZP-domain components of the aECM. In the Drosophila trachea, loss of the ZP proteins Piopio or Dumpy causes junctions between smaller tubes to break, whereas larger tubes generally remain intact but over-elongate and change shape [24,25,78]. The mammalian microvasculature contains many unicellular capillaries [31], and loss of Endoglin, a luminal ZP protein of endothelial cells, causes capillary ruptures and hemorrhage in both mice and human patients [79,80]. Endoglin phenotypes have been attributed to defects in TGF-beta or BMP signaling, although additional direct effects on physical integrity have not been ruled out [80]. ZP-dependent apical shaping roles could be particularly important in narrow tubes, where apical membranes otherwise could easily come in contact with each other, leading to membrane damage or lumen collapse.
LET-653 is just one of a set of transmembrane or secreted proteins that have similar requirements in the maintenance of the excretory duct and pore tubes. Others include the apically-localized transmembrane eLRRon proteins LET-4 and EGG-6 [38] and the secreted lipocalin LPR-1 [45]. Our data suggest that LET-653 and LPR-1 act in parallel to promote duct lumen integrity during morphogenesis; cooperative but distinct mechanisms of action are consistent with prior evidence that LPR-1 can protect lumen integrity even when provided from outside of the excretory system [45]. LET-4 appears to affect a later step of aECM organization, with a later onset of expression and broader effects on cuticle organization and barrier function that are not seen in let-653 mutants [38].
Possible functional relationships among ZP proteins, eLRRon proteins and lipocalins have been found in other systems. In Drosophila, the PAN and ZP domain protein NOMP-A and the eLRRon protein Artichoke are both constituents of the dendritic cap matrix that links ciliated neurons to glial structures within mechanosensory and chemosensory organs [81,82]. Mutations in the ZP proteins Piopio and Dumpy and in the eLRRon proteins Convoluted, Capricious and Tartan all affect tracheal tube shape and/or cell connectivity [24,83,84]. In the mammalian kidney, the ZP domain protein uromodulin (UMOD, also known as Tamm-Horsfall protein) and Lipocalin2 (Lcn2, also known as NGAL or 24p3) are among the most abundant luminal components in nephron tubules, where they appear to play protective functions [27,85]. Mutations in UMOD cause inherited forms of chronic kidney disease (CKD) that may be associated with structural damage to nephron tubules [27,86]. Further studies of ZP, eLRRon and lipocalin proteins in the C. elegans excretory system should better our understanding of the molecular and cellular roles of these families of apical factors, and how disruptions of aECM organization can contribute to disease.
All strains were grown at 20˚C under standard conditions [87] unless otherwise noted. Allele let-653(s1733) [42] is linked to unc-22(s7) and unc-31(e169). Alleles let-653(cs178) and let-653(cs204) are linked to jcIs1. Mutants were obtained from mothers heterozygous for the balancer nT1 [qIs51] [88] or rescued with a let-653(+) transgene. Transgenes used included: jcIs1 (AJM-1::GFP, rol-6d) IV [89], qnEx59 (dct-5pro::mCherry, unc-119+) [46], sEx10642 (let-653pro::GFP, dpy-5+) [90]. See S1 and S2 Tables for a complete list of strains and newly generated transgenes used in this study.
Strain UP2214 [unc-119; jcIs1; qnEx59] was mutagenized with ethylmethanesulfonate (EMS) as described [87], 2239 F1 animals were picked to individual plates and F2 progeny were screened for rod-like lethality. A total of 85 recessive rod-like lethal mutations were identified, of which 61 had obvious duct or pore junction or lumen abnormalities. Of 24 mutants with initially WT junction patterns as embryos, two mapped to chromosome IV and failed to complement let-653(s1733) and each other. Lesions were identified by whole genome sequencing followed by bioinformatics analysis with Cloudmap [91] and/or by Sanger sequencing of PCR-amplified genomic fragments. cs178 changes codon 54 from CAA to TAA. cs204 changes the intron 4 splice acceptor from CAG to CAA. s1733 changes codon 250 from TGT to TGA. All three let-653 alleles also contain a common polymorphism (a CCG to CTG change at codon 189, leading to a P189L substitution) that differs from the reference N2 and cDNA sequence.
For tissue-specific rescue experiments, let-653b cDNA was PCR-amplified from yk1667d04 with primers oCP10 (GGGGCTAGCAAAATGCGACATCCACTAATTTCTCTAC) and oCP2 (GGGGGTACCTCAGATGTTTCCAGTTCGAAC) and cloned into derivatives of vector pPD49.26 (Addgene) containing different tissue-specific promoters. lpr-1pro drives expression in the duct, pore and hypodermis beginning at the 1.5-fold stage [59]. lin-48pro drives expression in the duct cell beginning at the 2-fold stage [92]. dpy-7pro drives expression in the pore and hypodermis beginning at the 1.5-fold stage [93]. glt-3pro drives expression in the canal cell beginning at the early 3-fold stage [94]. unc-54pro drives expression in the body muscle beginning at the 1.5-fold stage [95]. Plasmids were co-injected with marker pIM175 unc-119pro::GFP (see S1 Table).
For localization and structure/function experiments, a 2.2 kb fragment containing the let-653 promoter and upstream regulatory region was PCR-amplified from the let-653 TransgeneOme clone [96] with primers oJAF4 (GGGAAGCTTGTCTATGTGAACCAGTCAATG) and oJAF5 (GGGGGATCCGTGGATGTCGGATTTACTGAAGAGAGCAG) and cloned into pPD49.26, upstream of tagged let-653b cDNAs. See S2 Table for a complete list of LET-653 full-length fusions and structure/function constructs. SfGFP was obtained from pCW11 (a kind gift from Max Heiman, Harvard U.). Plasmids were co-injected with marker pHS4 (lin-48pro::mRFP) (see S1 Table). Animals transgenic for the TransgeneOme LET-653::GFP reporter [96] showed a similar pattern of GFP expression and localization to that reported in Fig 5, but expression appeared much fainter.
Embryos were selected at the 1.5-fold stage and then incubated for the indicated number of hours before mounting for imaging. Larvae were staged by hours after egg-lay, with five hours corresponding to the 1.5-fold stage and thirteen hours corresponding to hatch.
Fluorescent and DIC images were captured on a compound Zeiss Axioskop fitted with a Leica DFC360 FX camera or with a Leica TCS SP8 confocal microscope. Images were processed and merged using ImageJ and Adobe Photoshop. Note that there is bleed-through of the red channel into the green channel in some Axioskop images, so many images are displayed as inverted grayscale with both channels visible. Duct lumen dimensions were measured using the line tool in ImageJ. For each specimen, measurements were made in triplicate and then averaged.
For TEM, the let-653(s1733) L1 specimens were cut open with a razor blade before fixation in buffered glutaraldehyde, rinsed and fixed again in buffered osmium tetroxide, then washed, en bloc stained with uranyl acetate, dehydrated and embedded into plastic resin [37,97]. The wild-type L1 specimen (called “L1C”) shown in Figs 3B, 3C, and 4C and the wild-type L4 specimen (“N2_L4_vulva”) shown in Fig 7A did not have a primary fix in glutaraldehyde, but were just fixed in buffered osmium tetroxide before dehydration and embedment. The normal (lin-17) and let-653(cs178) embryo specimens in Fig 3A and 3D were prepared by high pressure freezing and freeze substitution into osmium tetroxide in acetone, then rinsed and embedded into LX112 resin [98]. Images were collected on a Jeol-1010 transmission electron microscope, processed in ImageJ and pseudocolored in Adobe Illustrator.
Worms were grown to near-confluence on 60mm plates. Two 60mm plates were bleached and embryos were allowed to develop in M9 for 4–5 hours until they reached 2-fold stage. Worms were pelleted and transferred to Laemmli buffer (BioRad, 161–0737) with 1x Protease Inhibitor Cocktail (Sigma, 2714) and 1:20 B-mercaptoethanol. Samples were boiled for 5 minutes and then transferred to -80C until ready for use. Samples were boiled for 10 minutes before loading into Mini-Protean TGX gradient gels (BioRad, 456–1084). Electrophoresis was performed at 0.03A-0.04A under 1x electrophoresis buffer (BioRad, 161–0732). Protein was transferred onto 0.2um nitrocellulose membrane (BioRad, 162–0147) overnight in 1x transfer buffer (20% ethanol, 0.58% Tris Base, 2.9% Glycine, 0.01% SDS). Membranes were washed in PBS + 0.02% Triton (PBST) and blocked for 1 hour at room temperature in PBST + 1% dry, nonfat milk. Membranes were probed for GFP for 1 hour at room temperature with PBST + 1% milk + 1:500 anti-GFP antibody (Rockland Immunochemicals, 600-101-215). Membranes were washed in PBST and then probed with PBST + 1% milk + 1:2000 anti-Goat-HRP antibody (Rockland Immunochemicals, 605–4302) for 1 hour at room temperature. Membranes were washed with PBST before detection using SuperSignal West Femto Maximum Sensitivity Substrate (Pierce 34095) under a CCD camera (AV Imaging Systems). For the loading control, membranes were stripped for 15 minutes in Restore Western Blot Stripping Buffer (Thermo Scientific, 21059) and then washed in PBST. The same protocol was used for the loading control as for the primary blot. Anti-UNC-15/paramyosin (R224, kindly provided by Dr. Jeff Hardin, Univ. of Wisconsin) diluted 1:2000 was used as the primary antibody and anti-rabbit IgG (GE Healthcare, NA934V) diluted 1:10,000 was used as the secondary antibody.
Specimens were mounted on 10% agarose pads containing 20mM sodium azide and 10mM levamisole in M9. FRAP was performed using Leica Application Suite X software FRAP module on a Leica TCS SP8 MP confocal microscope. For 1.5-fold embryos and L1 larvae, a 1 μm X 1 μm bleach ROI was defined within the wizard, and mean fluorescence intensity measurements within the ROI were taken at specified intervals. The following experimental time-course was used: 20 pre-bleach frames every 0.4 seconds, 5 bleach frames every 0.4 seconds, and 90 post-bleach frames every 2.0 seconds. Laser intensity during bleach was set to 70.0%, while pre- and post-bleach laser intensity varied from 0.25% to 1.50% on a specimen-by-specimen basis. A pinhole size of 3.0 (units) was used for all FRAP experiments. For L4 larvae, bleach laser intensity was increased to 100%, the number of bleach frames was increased to 10, and the ROI size was increased to 3 μm X 3 μm.
FRAP plots were created and analyzed in Prism, where one-phase association curves derived from the model Y = Y0 + (Plateau—Y0)*(1 –e^(-Kx)) were fitted to the data. For statistical tests, mobile fractions and recovery half-times were derived from one-phase association curves fitted to individual experiments. Mobile fraction = Plateau-Y0; t1/2 = ln(2)/K, where K is the recovery rate constant.
Full-length or partial let-653b cDNAs, tagged with SfGFP or ssSfGFP were cloned into the hsp16.41 promoter-containing vector pPD49.78 (Addgene) (S2 Table). Transgenic embryos were heat shocked at 34°C for 30 minutes 3 hours after egg lay and imaged 2.5 hours later. Gut lumen width measurements were taken at the widest point using ImageJ freehand line tool and Plot Profile function.
|
10.1371/journal.ppat.1006574 | The role of the C-terminal D0 domain of flagellin in activation of Toll like receptor 5 | Flagellin is a wide-spread bacterial virulence factor sensed by the membrane-bound Toll-like receptor 5 (TLR5) and by the intracellular NAIP5/NLRC4 inflammasome receptor. TLR5 recognizes a conserved region within the D1 domain of flagellin, crucial for the interaction between subunits in the flagellum and for bacterial motility. While it is known that a deletion of the D0 domain of flagellin, which lines the interior of flagella, also completely abrogates activation of TLR5, its functional role remains unknown. Using a protein fusion strategy, we propose a role for the D0 domain in the stabilization of an active dimeric signaling complex of flagellin-TLR5 at a 2:2 stoichiometric ratio. Alanine-scanning mutagenesis of flagellin revealed a previously unidentified region of flagellin, the C-terminal D0 domain, to play a crucial role in TLR5 activation. Interestingly, we show that TLR5 recognizes the same hydrophobic motif of the D0 domain of flagellin as the intracellular NAIP5/NLRC4 inflammasome receptor. Further, we show that residues within the D0 domain play a previously unrecognized role in the evasion of TLR5 recognition by Helicobacter pylori. These findings demonstrate that TLR5 is able to simultaneously sense several spatially separated sites of flagellin that are essential for its functionality, hindering bacterial evasion of immune recognition. Our findings significantly contribute to the understanding of the mechanism of TLR5 activation, which plays an important role in host defense against several pathogens, but also in several diseases, such as Crohn’s disease, cystic fibrosis and rheumatoid arthritis.
| Receptors of the innate immune system typically recognize conserved microbial patterns, crucial for pathogen fitness and survival. Flagellin, the main structural protein of bacterial flagella, is recognized by two receptors of the innate immune system, the intracellular inflammasome receptor NAIP5/NLRC4 and the membrane-bound Toll-like receptor 5. Ligand-induced dimerization is a crucial step in Toll-like receptor 5 activation. A crystal structure of segments of the ligand-bound receptor revealed binding interfaces on the ligand and receptor, but failed to fully clarify the activation mechanism, since the D0 domain of flagellin, which is crucial for receptor activation, is missing in the structure. We propose a role for the D0 domain in receptor dimerization and pinpoint specific amino-acid residues within the D0 domain, which contribute to Toll-like receptor 5 activation. We show that Toll-like receptor 5 recognizes the same protein motif detected by the intracellular NAIP5/NLRC4 receptor. Our work represents an important advance in the understanding of the mechanism of activation of Toll-like receptor 5.
| Toll-like receptors (TLRs) belong to a family of germ-line encoded innate immune receptors able to sense pathogen-associated molecular patterns (PAMPs) [1]. Upon ligand binding, TLRs dimerize, recruiting adaptor molecules that bind to the intracellular TIR domain dimer and induce downstream signaling, resulting in the synthesis of pro-inflammatory cytokines and other immune response effectors [2,3]. Despite a conserved global fold of the structures of TLR receptors, the diversity of ligands they are able to recognize is very broad, ranging from small molecules such as nucleoside analogues to larger molecules such as nucleic acids and proteins. The ability to respond to such a wide array of agonists lies in the distinct recognition mode specific for each TLR and its ligands [4]. Recognition of the characteristic molecular features of microbial ligands (PAMPs) that are essential for the microbial survival and virulence makes it difficult for the pathogen to modify the structure of PAMP in order to circumvent immune recognition without losing its functionality.
Toll-like receptor 5 (TLR5) recognizes flagellin, the main structural protein of bacterial flagella, which exhibits a remarkable level of conservation among bacterial species, thus representing an attractive target for innate immune recognition [5]. Flagellin is composed of multiple structural domains, D0-D3. Flagellin monomers are stacked into a helical filament with the conserved D0 and D1 domains facing inward into the filament core channel, through which flagellin molecules are transported during flagella formation, while the variable domains D2 and D3 protrude outward from the core and are solvent exposed [6]. Functional flagella represent a virulence factor for several important human pathogens [7–9]. Flagellins of β- and γ-proteobacteria, such as Serratia marcescens or Salmonella typhimurium, are efficiently detected in their monomeric form by TLR5 at pico-molar concentrations, while some bacterial species, such as the Epsilonproteobacteria gastric pathogen Helicobacter pylori or the food-borne pathogen Campylobacter jejuni, have evolved their flagellin to evade TLR5 recognition while retaining bacterial motility. Several amino acid changes that contribute to TLR5 evasion in H. pylori FlaA flagellin have been identified within the D1 domain [10]. This evolutionary adaptation involves large restructuring of packing of flagellin monomers into filaments, which comprise 7 molecules of the FlaA of H. pylori per turn, in comparison to the 11-fold symmetry in flagellin FliC of S. typhimurium [11].
Mutational and structural studies have identified conserved regions on the TLR5 ectodomain that interact with amino acid residues within the conserved D1 domain of flagellin [12–14]. The crystal structure of the N-terminal fragment of the zebrafish TLR5-N14VLR comprising approximately two thirds of the TLR5 ectodomain in complex with a fragment of Salmonella flagellin, lacking the conserved D0 domain, provides a detailed characterization of this interaction. The structure reveals a primary binding interface for α-helices of the D1 domain of flagellin on the ascending lateral surface of TLR5, stretching from leucine-rich repeat LRRNT to LRR10, leading to formation of a 1:1 TLR5-flagellin complex. A secondary but relatively small binding interface mediates the interaction of the αND1b helix and the subsequent β-hairpin region of flagellin with the convex side of LRR12/13 on the opposite TLR5 receptor and is thus proposed to guide the dimerization of the complex in a stoichiometry of 2:2, the physiological relevance of which has also been confirmed in human cells [14,15]. However, the identified interactions do not appear to be sufficient for the formation of an active signaling receptor dimer, since the truncated form of flagellin lacking the D0 domain is insufficient for TLR5 activation [12,16]. Moreover, the 2:2 complex observed in the crystal was not detected in solution, despite binding of flagellin to TLR5 in a 1:1 stoichiometric ratio [14]. A 2:2 complex required for receptor activation was also not observed in a subsequent report on the crystal structure of B. subtilis flagellin in complex with a fragment of the TLR5 ectodomain [17]. This suggests, in agreement with previous studies, that the truncated fragment of the receptor lacks part of the binding site for flagellin [18]. The D0 domain is essential for functional flagella formation and signaling, yet its deletion only slightly impairs binding to the TLR5 monomer; therefore, the mechanism of its functional role in TLR5 signaling remains unknown [14,16,19].
The motivation to clarify the mechanism of TLR5 signaling and, on the ligand side, to assess the contribution of the D0 region to the evasion of host sensing led us to analyze the role of the D0 domain in receptor activation. We propose a role for the D0 domain in crosslinking the two TLR5 receptor monomers and hence stabilizing the functional signaling complex. While flagellin lacking the D0 domain was incompetent in triggering TLR5 activation, tethering two inactive flagellin D0 deletion variants into a covalent dimer restored activation, suggesting a role for the D0 domain in receptor dimerization. A structure-guided mutagenesis study identified amino acid residues at the C-terminal segment of the D0 domain involved in receptor activation. Further, we pinpointed amino acid residues within the D0 domain that enabled evasion of immune recognition by the H. pylori flagellin. Together, this shows that the multipartite recognition of flagellin by TLR5 hinders an easy evasion of immune recognition by single point mutations through targeting conserved segments of flagellin, which are essential for flagellar self-assembly.
To assess the role of the D0 domain of flagellin in TLR5 activation, we used a synthetic biology strategy by constructing a chimeric protein, composed of the D0 and D1 domains of S. typhimurium flagellin fliC, which we named short flagellin (SF), fused via a flexible peptide linker to the N-terminus of human TLR5 (SF-TLR5). This chimeric protein, combining selected ligand domains and the full-length receptor in a single molecule, exhibited constitutive activity when expressed in HEK293 cells, while a fusion protein lacking the D0 domain (SFΔD0-TLR5) was inactive [15]. Most post-translational modifications of flagellin are located in the variable D2 and D3 domains and it has been shown previously that bacteria-specific post-translational modification of flagellin is not required for TLR5-based recognition [12,20]. The D0 domain is composed of two discontinuous epitopes spanning the N- and C-terminal regions (hereafter referred to as ND0 and CD0), comprising amino acid residues 1–42 and 455–494 of SaTy, respectively (Fig 1A). We aimed to assess the contribution of distinct regions of the D0 domain to TLR5 activation. HEK293 cells were transiently transfected with plasmids for the chimeric proteins and activation of NF-kB reporter was determined through a dual luciferase reporter assay (Fig 1). SF-TLR5 was used as a positive and SFΔD0-TLR5 as a negative control. A fusion protein variant comprising short flagellin lacking the ND0 domain (SFΔND0-TLR5) retained the constitutive activity of the positive control, SF-TLR5 (Fig 1B). On the other hand, deletion of the C-terminal D0 domain (SFΔCD0-TLR5) completely abrogated NF-κB activation, as did the negative control, a chimeric receptor lacking the complete D0 domain (SFΔD0-TLR5) (Fig 1C). One explanation for this effect might be that the truncation of the CD0 domain in the chimeric construct shortens the distance between flagellin and TLR5 and could thus lead to a steric hindrance in binding. To rule out this possibility and confirm that the truncation of the C-terminal flagellin segment on its own is the cause of the impaired signaling, we tested constructs with two longer linker lengths, 27 and 57 aa (Fig 1A), which resulted in the same level of activation (Fig 1C). To rule out inactivation of the truncated constructs due to protein misfolding, a cotransfection assay with wtTLR5 was performed. The results showed a decrease in wild-type TLR5 activation by flagellin upon addition of increasing levels of the inactive SFΔD0-TLR5, suggesting an interaction between wt-TLR5 and TLR5-SFΔD0, which is inactive but can still bind TLR5 and can therefore inhibit TLR5 activation (Fig 1D). Western blot analysis confirmed that the expression level of wtTLR5 remains constant, even upon expression of increasing levels of the fusion receptor SFΔD0-TLR5 (S1B Fig). The decrease in signaling is therefore not due to a limited cell capacity for overexpression of recombinant proteins.
Further, we prepared deletion constructs lacking the C-terminal 10 (SFΔC10D0-TLR5) or 20 (SFΔC20D0-TLR5) amino acids of the CD0 domain (Fig 1A). The NF-κB luciferase reporter assay showed a decrease in signaling proportional to the size of the deletion, although complete abrogation of signaling was observed only upon deletion of the complete C-terminal D0 domain (Fig 1E). All proteins where detected in cell lysates at comparable levels, confirming that the decrease in activation is not a consequence of altered protein expression (S1A Fig). These results underline a prominent role for the conserved C-terminal D0 domain of flagellin in TLR5 activation.
Based on the alignment of flagellins of different bacterial species, amino acid residues were selected for alanine mutagenesis. The selection criteria included amino acid residues of the Salmonella flagellin FliC, which differ from their H. pylori FlaA counterparts, but also residues that are conserved among the different clades (Fig 2A). Alanine point mutations were introduced, with the exception of a substitution of the terminal arginine R494 to glutamic acid since an alanine mutant at this site was highly prone to proteolysis in bacterial overexpression. In addition to single alanine point mutations in the CD0 domain, a substitution of the hydrophobic motif in the region 489 to 493 (VLSLL) by five alanine residues was introduced (VLSLL_A) (Fig 2A). Recombinant flagellins were produced and isolated via affinity chromatography (S2A Fig) and tested for TLR5 activation potential in two reporter cell lines: HEK293 cells transiently expressing human TLR5 (Fig 2B and 2C) and the human epithelial cell line A549, which endogenously expresses TLR5 (S2B–S2E Fig). Mutations D457A and Y458A in the conserved spoke region showed a strong negative effect on TLR5 activation. A mutation in the central region of CD0, R467A, had a profound effect on activation, while other mutations in the central region had a low or no apparent effect on the activation of the NF-κB signaling pathway. Substitution of the hydrophobic amino acid residues at the very tip of the protein VLSLL_A completely abrogated TLR5 signaling, while a single point mutation of serine (S491A) within this motif had no effect, thus attributing the effect of protein VLSLL_A entirely to the hydrophobic residues. Mutation of the polar residue N488A, located at the tip of the CD0 domain preceding the hydrophobic motif VLSLL also significantly impaired signaling. Additionally, regions 460–463 (TEVS) and 472–474 (QQA) within the CD0 domain, which differ between S. typhimurium and H. pylori flagellin (SaTy and HePy, respectively), were mutated from SaTy to the corresponding residues of HePy (TEVS_EESA and QQA_VGS). The isolated recombinant proteins were tested for TLR5 activation (S2F Fig). However, these substitutions showed no effect on TLR5 activation, suggesting that the residues in the spoke region and at the C-terminus are crucial for TLR5 activation.
The mutant protein VLSLL_A was also tested for activation of the intracellular NAIP5/NLRC4 inflammasome. Wild type or NLRP3-deficient macrophages were primed with LPS and stimulated with wt or mutant flagellin and IL-1β secretion was measured. Mutation of the terminal hydrophobic residues resulted in decreased NAIP5/NLRC4 inflammasome activation, in agreement with a previous report (S2H and S2I Fig) [21].
Co-immunoprecipitation studies showed a comparable binding intensity of recombinant mutated flagellins to the TLR5 ectodomain, regardless of their activation potential, which is in agreement with previous studies where a deletion of the D0 domain completely abrogated signaling but hardly affected the binding efficiency of a truncated flagellin molecule to TLR5 [14] (Fig 2D).
These results suggest that the role of the D0 domain in TLR5 activation is largely defined by amino acid residues in the conserved C-terminal spoke region and in the C-terminal hydrophobic tip of flagellin. However, these point mutations do not significantly affect the binding of flagellin to hTLR5, as the primary binding site is located within the D1 region, as shown by previous structural studies [14].
In innate immune recognition, a correlation is often observed between function of the PAMP or its structural moiety for the microbe and recognition by TLRs, as structural or functional restrictions often hinder modifications of PAMPs that would enable immune evasion. Previous reports suggested a correlation between TLR5 activation and effects on the motility in the conserved D1 region of flagellin [10,22]. Amino acids in the C-terminal region of flagellin are also evolutionarily conserved among bacterial species since they tile the inner channel of the flagellum and participate in packing of neighboring flagellin chains. To assess whether there is a correlation between TLR5 recognition and motility for this region of flagellin, we tested mutated flagellins for their effect on functional flagella formation using a swarming motility assay (Fig 3). The effect on bacterial motility could be grouped into three categories: mutations exerting low motility (left panel), mutations with a moderate effect on motility (middle panel), and mutations which increased bacterial motility with respect to wild-type flagellin (right panel). Among the mutations with a significant effect on TLR5 activation, Y458A in the spoke region and VLSLL_A at the C-terminus had the most profound effect on motility while mutation R467A had a moderate effect. These results show that, while the majority of mutations that impair TLR5 activation also have a negative effect on motility, there is no direct correlation between TLR5 stimulation and function, at least not at the amino acid level.
For the formation of an active signaling complex, two flagellin molecules must bind to two TLR5 ectodomains [15], forming an active complex in which a dimer of the intracellular TIR domains initiates recruitment of the signaling adapter MyD88. Flagellin binds to TLR5 via a two-partite primary binding site and to the opposing TLR5 through a secondary site, both encompassed in the D1 domain of flagellin [14]. Superposition of full-length flagellin from the assembled flagellum (PDB code 1UCU) to the truncated FliC-ΔD0 from the crystal structure TLR5-N14VLR/FliC-ΔD0 demonstrates that in the extended form, the D0 domain of flagellin would clash with the cell membrane (S3 Fig). We reasoned that, upon binding of flagellin to TLR5, the D0 domain must reorient itself relative to the conformation in the assembled flagella. Taking into account this information and the necessity of TLR5 ectodomain dimerization for activation but not for flagellin binding, we hypothesized that the D0 domain might have a role in dimer formation by binding to the opposite TLR5 ectodomains and stabilizing an active complex by bringing the two 1:1 complexes of flagellin:TLR5 into sufficient proximity for signal transduction mediation by the dimerized cytosolic TIR domains. To test this hypothesis, we prepared a recombinant truncated flagellin molecule lacking the D0 and variable D2 and D3 domains (SFΔD0) and a recombinant protein in which the two SFΔD0 domains are tethered into a single polypeptide chain by a 27 amino acid peptide linker (dimSFΔD0) (Fig 4A). SDS PAGE (Fig 4B) and particle size analysis determined by dynamic light scattering (DLS) confirmed the expected size of dimSFΔD0 (6.0 nm) as being approximately twice the size of the monomeric SFΔD0 (2.3–2.5 nm) and comparable to the size of the full-length SaTy flagellin (6.1 nm). The monomeric SFΔD0 was unable to stimulate TLR5 (Fig 4C and 4D A549 cells, S4A and S4B Fig hTLR5-transfected HEK293 cells) in agreement with previous findings [14]. If the role of D0 is indeed to crosslink the two TLR5 ectodomains into a functional signaling dimer, tethering of the truncated flagellin fragment should improve activation. Indeed, in contrast to SFΔD0, dimSFΔD0 was active at concentrations as low as 50 ng/ml (Fig 4A and 4D). Tethering the two inactive short flagellins therefore substantially restores activation, albeit not to the full extent of wild type flagellin. We propose that the linker to a certain extent substitutes the protein-protein interactions mediated by the D0 domain in the full-length flagellin-TLR5 heterodimeric complex. Together, our results suggest a role for the D0 domain in crosslinking two 1:1 TLR5:flagellin complexes into an active 2:2 signaling complex.
Despite a high level of sequence conservation in the D1 and D0 domains of flagellin, the quaternary filament structures differ between the two proteobacteria clades [6,11]. H. pylori belongs to the clade of ε-proteobacteria that form a distinct flagellar filament assembly composed of 7 rather than of 11 protofilaments in the flagellum and are able to evade immune recognition by TLR5. A study by Andersen-Nissen et al. (2005) identified mutations in the primary binding site located within the D1 domain that could contribute to the evasion of immune recognition at the cost of impaired motility and compensatory mutations, which restored functional flagella formation and mobility. However, the D0 domain of flagellin also plays a crucial role in flagellar filament assembly by forming contacts with other monomers within the inner channel. To assess the role of the D0 region in the evasion of TLR5 detection by H. pylori, chimeric flagellins were constructed by exchanging the D0 domain of SaTy with the C-terminal D0 domain of HePy (SaTy-CD0(HePy)) or both the C- and N-terminal D0 domains of HePy (SaTy-D0(HePy)) (Fig 5A). All chimeric flagellin variants were produced in a bacterial expression system and purified via affinity chromatography (S5A Fig). These chimeric proteins comprised both the primary and secondary binding sites of SaTy. Therefore, any difference in the ability of these variants to activate TLR5 depends on the differences between the SaTy and HePy D0 domains. Circular dichroism analysis of chimeric flagellins demonstrated comparable secondary structure content to that of the wild type flagellin, indicating no deleterious effects on protein folding (Fig 5B). Both chimeric flagellins, SaTy-D0(HePy) and SaTy-CD0(HePy) were unable to activate TLR5 either at the endogenous level or ectopic TLR5 expression (Fig 5C human lung epithelial cell line A549, S5B Fig hTLR5-transfected HEK293 cells), similar to a protein completely lacking the D0 domain (SFΔD0), demonstrating the crucial role of the CD0 domain of flagellin in TLR5 activation. Chimeric flagellin of H. pylori with the C- and N-terminal D0 domains of SaTy also failed to activate TLR5 (Fig 5C, S5C Fig). The chimeric proteins SaTy-CD0(HePy) and SaTy-D0(HePy) had a similar secondary structure to SaTy, as demonstrated by the analysis of the circular dichroism spectra of the isolated proteins. This indicates that the evolutionary adaptations that allowed for the immune evasion of H. pylori are distributed across the whole length of flagellin and are not restricted to the primary TLR5 recognition surface located within the D1 region.
The amino acid sequences of HePy and SaTy flagellin are highly conserved in the D0 region, pointing to a high level of evolutionary constraints in this area. However, several residues represent more pronounced differences in charge or hydrophobicity. Therefore, we tested a combination of five substitutions from HePy to SaTy counterparts to identify the role of these residues in the evasion of TLR5 activation (Fig 5A). Isolated chimeric protein SaTy-CD0(HePy)mut with the selected counterpart mutations showed a significant recovery of the TLR5 activation potential in comparison to the chimeric protein SaTy-CD0(HePy), suggesting that these particular amino acid differences contribute to immune evasion by H. pylori (Fig 5C, S5D Fig).
A crucial step in the detection of PAMPS and active complex formation of TLRs is receptor ectodomain dimerization. Despite the important insight into the mechanism of ligand binding by TLR5 from the crystal structure of the complex of the ligand and receptor fragments, important aspects of TLR5 activation by flagellin remain unknown. A distributed binding site on the concave and lateral surfaces of TLR5, extending from LRRNT to LRR10, directs the primary binding of flagellin, enabling formation of a TLR5:flagellin 1:1 complex. A secondary binding site between the D1 domain of flagellin and LRR12-13 of the opposing TLR5 contributes to the formation of a 2:2 complex. These observed interactions are however not sufficient for receptor activation in vivo, as the truncated flagellin lacking the D0 domain is not able to trigger signaling. Furthermore, a subsequent study reported weak binding of flagellin to the ectodomain region beyond LRR17, not included in the crystal structure [23]. The unresolved issues concerning the TLR5 activation mechanism motivated us to investigate the role of the D0 domain in TLR5 activation in greater detail.
Results of a flagellin subdomain deletion revealed a key role of the C-terminal segment of the D0 domain in TLR5 activation. Further, a detailed alanine-scanning mutagenesis of this region revealed the contribution of several amino acid residues to TLR5 activation. Substitutions of two amino acid residues in the spoke region significantly affected receptor activation. The most conserved residues across all flagellins are those in the spoke region. The spoke region is crucial in filament formation for maintaining the integrity of the inner and outer tubes of the filament, and it forms inter-subunit interactions that enable tight packing of flagellin monomers [5]. A pronounced role in TLR5 activation was also demonstrated for amino acids at the very tip of flagellin, including polar asparagine at position 488 and the hydrophobic motif VLSLL from positions 489 to 493, excluding Ser491. While this region has not been previously reported to have a role in TLR5 activation, the same terminal hydrophobic motif is also crucial for recognition by the intracellular NAIP5/NLRC4 inflammasome [21]. Our results therefore suggest a dual mechanism of sensing the same region of bacterial flagellin through two distinct receptors of innate immunity. Although the molecular mechanism of flagellin recognition by NAIP5 is not known and is likely to differ from the mechanism of recognition by TLR5, selection of the same segment as the target for the innate immune receptors is likely due to the functional importance of this region for the assembly of functional flagella.
In the crystal structure study, Yoon et al. demonstrated that the contribution of the D0 domain to the formation of the 1:1 complex of flagellin:TLR5 is minimal [14]. Therefore, unless there is a third unknown co-factor involved in the TLR5-flagellin signaling event, the contribution of the D0 domain is most likely pertained to the formation of the 2:2 complex, required for activation. In line with these findings, we showed that, while point mutations within the C-terminal D0 domain strongly decreased signaling, binding of flagellin to TLR5 was not affected. Superposition of full-length flagellin to the TLR5 dimer based on the crystal structure TLR5-N14VLR/FliC-ΔD0 demonstrates that the structure of TLR5-bound flagellin must differ from the structure of flagellin in filaments. The spoke region connecting D0 and D1 is highly conserved among bacterial species and represents a flexible subunit of the flagellin molecule between domains D0 and D1, which are mostly α-helical in flagellar filaments, whereas the D0 domain is disordered in monomeric flagellin [6,24]. We propose that the flexible spoke region connecting the D0 and D1 domains functions as a hinge, enabling the reorientation of the D0 domain toward the opposing TLR5 ectodomain and thus stabilizing the 2:2 complex. This is supported by the recovered activity of a completely inactive form of flagellin without the D0 domain (SFΔD0) by forced dimerization in a covalently tethered SFΔD0 dimer. One might argue that increased activation of the dimeric construct could be simply due to increased local concentration of the ligand, although we believe this to be less likely, since the monomeric form of flagellin lacking the D0 domain is essentially inactive, even in a higher concentration range. Our results therefore identify a previously unrecognized segment of flagellin to play an essential role in TLR5 activation and suggest a role for the D0 domain in the process of receptor dimerization most likely through binding to the opposing ectodomain in the active complex. We note that these results provide indirect evidence to the mechanism of TLR5 dimerization and that direct proof of this concept would have to be supported by a structural study of the full length form of the ligand-receptor complex. However, structural studies of the D0 domain in the monomeric form of flagellin have, at least do date, been unsuccessful, most likely due to the disordered structural conformation of the terminal regions of flagellin [6,14,17,24].
Inter- and intra-subunit interactions of the D1 and D0 domains enable the tight packing of flagellin subunits into functional filaments [6]. The high level of sequence conservation in these regions illustrates their functional importance, and single point mutations can disrupt the proper quaternary structure and therefore bacterial motility. Indeed, bacterial motility was significantly reduced for several flagellin mutants. Inter-subunit connections between the D0 domain in the filament core are mostly hydrophobic [6,25], and therefore it is not surprising that of the mutants that affected TLR5 activation, the two hydrophobic mutations, Y458A and VLSLL_A, located in the spoke region and at the C-terminal tip, profoundly impaired motility, exhibiting a link between structure and function as an ideal target for immune recognition. These residues are directly involved in packing interactions with neighboring flagellin molecules in the filament. Main-chain and side-chain atoms from residues at the C-terminus of flagellin, including Q484, N488, S491, and R494, constitute the hydrophilic surface of the inner channel, which is important for the transport of monomeric flagellin through the channel in the process of filament formation [6,26]. While mutations at positions S491 and R494 weakly impaired motility, mutations at positions N488 and Q484 even increased motility with respect to wt flagellin. Retained and even increased bacterial motility might be explained by the position of residues N488 and Q484, which are oriented towards the center of the hydrophilic channel, where they do not participate in filament packing. Mutation of asparagine into alanine is expected to increase the size of the channel, without significantly affecting its polarity due to the small size of the alanine side chain, presumably facilitating transport as a result of increased channel size.
The distribution of the TLR5 binding sites across the length of the flagellin molecule includes several conserved segments of flagellin within the D1 and D0 domains required for flagellin self-assembly [12], thereby rendering the evasion of immune recognition rather difficult, requiring multiple coordinated mutations. The extent of the required change is illustrated by radically different packing of flagellar filaments in ε-proteobacteria, such as H. pylori. In addition to the previously recognized contribution of the D1 domain, we demonstrated a role of the D0 domain in the evasion of TLR5 recognition based on the exchange of the D0 domain of a potent TLR5 activator, S. typhimurium flagellin (FliC), with the D0 domain of the H. pylori flagellin (FlaA), which is unable to trigger TLR5-based immune activation. Further, specific amino acid residues have been identified, which had to be altered in the evolution of H. pylori to enable this evasion. We may speculate that the complex rearrangement of flagellin sequences that modify the supramolecular structure of flagella and evade TLR5 recognition may be easier in H. pylori due to its ability to recombine the genetic material from several strains within the same organism and therefore simultaneously combine multiple point mutations within the same molecule [27].
In conclusion, we suggest a functional role of the D0 domain of flagellin in TLR5 activation through the promotion of receptor dimerization. We propose a mechanism for the formation of a fully active TLR5:flagellin complex, in which the contribution of at least three distinct sites on flagellin is required; the primary binding site within the D1 domain that guides the formation of the flagellin:TLR5 heterodimer, and the secondary binding site that promotes interaction between flagellin and the opposite TLR5 ectodomain [14], which needs to be supported by an additional third interaction between the D0 domain of flagellin and the opposing TLR5 ectodomain. The multiple interaction surfaces identified in this and previous studies [12–14] contribute to the high affinity of binding and underlay the evolutionary robustness of TLR5 recognition. The importance of the D0 domain in the self-assembly of flagella makes it an excellent choice for a recognition target and it is not surprising that two completely different types of receptors of the innate immune system target the same region of the molecule. In fact, the membrane TLR4 and cytosolic caspase 11 receptors also recognize a very similar structural pattern of the LPS molecule [28], demonstrating the convergent evolution of the innate immune system to the functionally most relevant microbial targets.
Human embryonic kidney cell lines HEK293 (ATCC CRL-157) and HEK293T (ATCC CRL-3216) and A549-Dual adherent epithelial cells (Invivogen) were cultured in complete media (DMEM; 1 g/l glucose, 2 mM L-glutamine, 10% heat-inactivated FBS (Gibco)) in 5% CO2 at 37°C. The human A549 lung carcinoma cell line expresses a secreted embryonic alkaline phosphatase (SEAP) reporter under the control of the IFN-β minimal promoter fused to five NF-κB binding sites, and a secreted luciferase under the control of an ISG54 minimal promoter in conjunction with five IFN-stimulated response elements (Invivogen). The plasmids used include pUNO-hTLR5 encoding human TLR5 (InvivoGen) and pcDNA3 (Invitrogen). S. typhimurium flagellin, chimeric flagellins, and flagellin point mutants were cloned into the pET19b expression vector (Novagen). For the bacterial motility assay, flagellin mutants were cloned into the pRP4 plasmid expressing wild-type flagellin (courtesy of E. Miao, Institute for Systems Biology, Seattle). Further, chimeric proteins SFΔD0, dimSFΔD0, and chimeras of SaTy and HePy flagellin were cloned into the pET19b expression vector (Novagen). Fusions of short flagellins with TLR5, SF-TLR5, SFΔD0-TLR5 [15], SFΔCD0-TLR5, SFΔCD0-l57-TLR5, SFΔND0-TLR5, SFΔC10D0-TLR5, and SFΔC20D0-TLR5 were cloned into the pFLAG-CMV3 expression vector (Sigma-Aldrich). The chimeric constructs prepared for this study are described in detail in S1 Table.
Escherichia coli BL21 (DE3) pLysS cells transformed with the pET19b plasmid expressing wild-type or mutant flagellin were cultivated at 37°C in Luria-Bertani (LB) medium, containing 50 μg/ml ampicillin. Overnight cultures were transferred to fresh media, grown to an optical density of ~0.8 at 600 nm, and supplemented with 1 mM Isopropyl β-D-thiogalactoside (IPTG). Cells were grown at 37°C for 4 hours, harvested, and lysed in buffer (10 mM TRIS pH 7.5, 1 mM EDTA, 0.1% DOC) containing a protease inhibitor cocktail (Sigma P8849), followed by sonication (pulse 1 s on, 2 s off, 10–15 min) and centrifugation (12000 rpm for 30 min). N-terminally His10-tagged recombinant proteins were purified on Ni-NTA affinity agarose (Qiagen) and dialyzed against 20 mM HEPES buffer. Flagellin point mutants prone to protease degradation (D455A, D457A, Y458A, and T476A) were additionally purified via a Strep-tag on the C-terminus on a Strep-Tactin Sepharose column (Iba), according to the manufacturer’s guidelines. Protein concentration was determined with the BCA assay (Pierce), and purity was confirmed using SDS-PAGE and immunoblotting.
For the dual-luciferase assays, HEK293 cells were seeded in 96-well plates (Corning) at 2–3 × 104 cells per well (0.1 ml). The next day, the cells were transiently transfected with plasmids expressing wtTLR5 or chimeric constructs, pELAM-1 (C. Kirschning, University of Duisburg-Essen, Germany) expressing NF-κB-dependent firefly luciferase (50 ng per well), and phRL-TK (Promega) constitutively expressing Renilla luciferase (5 ng per well) using the jetPEI transfection reagent (Polyplus Transfection). The total amount of DNA for each transfection was kept constant by adding appropriate amounts of control plasmid pcDNA3 (Invitrogen). After 24 h, the cells were either lysed or the medium was changed and the cells were stimulated with purified recombinant flagellin (10 μl) for an additional 18 h before lysis. The cells were lysed in Passive Lysis 5x Buffer (Promega) and analyzed for reporter gene activities using a dual-luciferase reporter assay. Each experiment was repeated at least three times, and each measurement was performed in at least four biological parallels. A student’s unpaired two-tailed t-test was used for statistical comparison.
A549 epithelial cells express endogenous TLR5 and an NF-κB-inducible SEAP reporter and do not express the intercellular reporter for flagellin, NLRC4 [29]. Cells were seeded in 96-well plates (Corning) at a density of 5 × 104 cells per well (0.1 ml). Immediately after seeding, the A549 cells were stimulated with flagellin. After 10 h, the supernatants were collected, heated for 1 h at 65°C, and the NF-κB-dependent SEAP activity was determined using Quanti Blue reagent according to the manufacturer’s instructions (Invivogen). Each experiment was repeated at least three times, and each measurement was performed in at least five biological parallels. A student’s unpaired two-tailed t-test was used for statistical comparison.
HEK293T cells were seeded in a 6-well plate (Techno Plastic Products) at a density of 5–7 × 105 cells per well. The next day, the cells were transiently transfected with 2 μg of plasmids expressing TLR5 or TLR5 fusion constructs using the Lipofectamine transfection reagent (Thermo Fisher Scientific). In the case of transfection of varied amounts of plasmids, the overall amount of transfected DNA was equalized using control plasmid pcDNA3. Forty-eight h post transfection, the cells were lysed in lysis buffer (50 mM Tris-HCl (pH 8), 1 mM EDTA, 1 mM EGTA, 137 mM NaCl, 1% Triton X-100, 1% sodium deoxycholate (DOC), 10% glycerol, 1 mM Na3VO4, and 50 mM NaF) containing a cocktail of protease inhibitors (Roche). Cell debris was removed by centrifugation at 13200 rpm for 15 min. The total protein concentration in the supernatant was determined using the BCA assay. Further, proteins from the supernatant were separated by SDS-PAGE and transferred to a Hybond ECL nitrocellulose membrane (GE Healthcare). The membrane was washed (1 × PBS buffer) and incubated in blocking buffer (1 × PBS, 0.1% Tween 20, 0.2% I-Block (Tropix)) overnight at 4°C. The membranes were incubated with primary antibodies diluted in blocking buffer for 90 min, washed (1 × PBS, 0.1% Tween 20), and incubated with secondary antibodies for 45 min at room temperature. Secondary antibodies were detected with the ECL Western blotting detection reagent (GE Healthcare), according to the manufacturer’s protocol. The primary antibodies were rabbit anti-FLAG (F7425, Sigma), rabbit anti-AU1 (ab3401, Abcam;) and mouse β-aktin (Cell Signaling techn., 3700;), all diluted 1:1000. Secondary antibodies were horseradish peroxidase-conjugated goat anti-rabbit IgG (ab6721, Abcam) and HRP-conjugated goat anti-mouse IgG (Santa Cruz, sc-2005), both diluted 1:4000. Further, the purity of the recombinant flagellins isolated from E. coli was analyzed by SDS-PAGE and western blot analysis using mouse Tetra-His antibodies diluted 1:2000 (34670, Qiagen) and horseradish peroxidase-conjugated goat anti-mouse IgG diluted 1:4000 (sc-2005, Santa Cruz) as secondary antibodies.
For the co-immunoprecipitation studies, a soluble hTLR5 ectodomain fused to the Fc region of human IgG1 (Invivogen) was bound to protein A-coupled Dynabeads (Thermo Fisher Scientific), according to the manufacturer’s protocol. A total of 10 μg of hTLR5-Fc diluted in MQ water or just MQ water as a control was incubated with 40 μl of beads for 1 h at room temperature and washed 3 times with wash buffer (1 × PBS, 0.02% Tween 20, pH 7.5). Twenty μg of wt or mutant flagellin was added per sample, incubated for 1 h at room temperature, and washed 3 times with wash buffer. The samples were eluted in 0.1% SDS at 95°C for 5 min. The flagellin in the samples was detected with Western blot analysis using anti-His antibodies.
The impact of flagellin mutations on bacterial motility was tested using an immobile bacterial strain, S. typhimurium FliC FljB ATCC 14028s (courtesy of E. Miao, Institute for Systems Biology, Seattle), transformed with a pRP4 plasmid expressing wild-type or mutant flagellin. The bacterial cells were transformed using electroporation at 2.5 kV, 200 Ω, and 25 μF in 10% glycerol. A single colony of transformed bacteria grown on LB agar plates containing 50 μg/ml ampicillin was selected and transferred to the motility test plates (LB media containing 0.3% agar, 1 mM IPTG, and 50 μg/ml ampicillin). The cultures were incubated overnight in an upright position at room temperature. Each plate was inoculated with a single colony of S. typhimurium expressing mutated flagellin and bacteria expressing wild-type SaTy as a control. The motility of the strains expressing the mutated flagellin was compared to the motility of the control strain transformed with wild-type SaTy.
Immortalized wild type BMDMs from C57BL/6 mice and NLRP3-deficient mice (both gift of K. A. Fitzgerald; University of Massachusetts Medical school, Worcester, MA, USA) were cultured in DMEM supplemented with 10% FBS. NAIP5/NLRC4 activation assays were performed in serum-free DMEM. Cells were seeded at 1.5 x 105 cells per well on 96 well plates and primed with ultra-pure LPS (100 ng/mL) for 6 hours for the stimulation of pro-IL-1β expression. Further, the growth medium was removed and wild type or mutant flagellin (3 μg/ml) mixed with DOTAP (1:5) in DMEM was added for 4 h. The concentration of secreted IL-1β was measured by ELISA (e-Bioscience) according to manufacturer’s instructions.
DLS data of the soluble flagellin species were acquired using a Zetasizer Nanoseries instrument (Malvern). The samples were centrifuged at 13000 rpm for 30 min to remove protein aggregates. Measurements were made at 20°C using automated settings, and 3 independent acquisitions of 10 measurements each were analyzed using the associated DTS nanoparticle-sizing software.
CD measurements were used to determine the secondary structure of soluble flagellins. The CD spectra were taken between 195 and 280 nm on a Chirascan CD spectrometer (Applied Photophysics) fused with nitrogen gas and equipped with a temperature controlled cuvette holder. A cell path length of 1 mm was used with concentrations of flagellins in the range of 0.1–0.5 mg/ml. All samples were dissolved in demi-water, and the results are the average of 3 spectra measured at 20°C.
Flagellin amino acid sequences of S. typhimurium (SaTy, UniProt id. P06179), S. marcescens (SeMa, UniProt id. P13713), S. dublin (SaDu, UniProt id. Q06971), H. pylori (HePy, UniProt id. P0A0S1), and C. jejuni (CaJe, UniProt id. P22252) were aligned using ClustalW (http://embnet.vital-it.ch/software/ClustalW.html).
UCSF Chimera 1.6.2 software was used to generate the structural figures and to determine the distances among atoms (http://www.cgl.ucsf.edu/chimera/) [30]. Graphs were prepared with Origin 8.1 software (http://www.originlab.com/), and GraphPad Prism 5 (http://www.graphpad.com/) was used for statistics. A students’ unpaired two-tailed t-test was used for statistical comparison of the data.
|
10.1371/journal.pntd.0003866 | A Systematic Review of the Mortality from Untreated Leptospirosis | Leptospirosis occurs worldwide, but the global incidence of human disease and its mortality are not well understood. Many patients are undiagnosed and untreated due to its non-specific symptoms and a lack of access to diagnostics. This study systematically reviews the literature to clarify the mortality from untreated leptospirosis. Results will help quantify the global burden of disease and guide health policies.
A comprehensive literature search was performed to identify untreated patient series. Included patients were symptomatic, but asymptomatic patients and those who had received antibiotics, dialysis or who were treated on Intensive Care Units were excluded. Included patients had a confirmed laboratory diagnosis by culture, PCR, or serological tests. Data was extracted and individual patient series were assessed for bias. Thirty-five studies, comprising 41 patient series and 3,390 patients, were included in the study. A high degree of bias within studies was shown due to limitations in study design, diagnostic tests and missing data. Median series mortality was 2.2% (Range 0.0 – 39.7%), but mortality was high in jaundiced patients (19.1%) (Range 0.0 – 39.7%), those with renal failure 12.1% (Range 0-25.0%) and in patients aged over 60 (60%) (Range 33.3-60%), but low in anicteric patients (0%) (Range 0-1.7%).
This systematic review contributes to our understanding of the mortality of untreated leptospirosis and provides data for the estimation of DALYs attributable to this disease. We show that mortality is significantly higher in older patients with icteric disease or renal failure but is lower in younger, anicteric patients. Increased surveillance and accurate point-of-care diagnostics are required to better understand the incidence and improve diagnosis of disease. Empirical treatment strategies should prioritize early treatment to improve outcomes from leptospirosis.
| Leptospirosis is a common cause of fever in the developing world but often goes undiagnosed and untreated due to its non-specific clinical features and the limited availability of point-of-care diagnostics. This review systematically evaluated available literature to clarify the mortality from untreated leptospirosis. Untreated patients were defined as patients not receiving antibiotics, dialysis, or treatment on an Intensive Care Unit. All patients had a confirmed laboratory diagnosis of leptospirosis through culture, PCR or serological tests. Results showed that mortality from untreated leptospirosis is significant in older patients and those who develop complications such as jaundice and renal failure, but mortality is low in younger patients and those with anicteric disease. There was a high degree of bias within studies due to limitations in diagnostics and missing data. The data presented in this review, when coupled with improved understanding of the true incidence of the disease, will help estimate the burden of disease from leptospirosis. Increased surveillance and accurate point-of-care diagnostics are required to better understand the incidence of disease and outcomes from leptospirosis. Empirical treatment strategies of undifferentiated fever should focus on early treatment of fever to reduce mortality from leptospirosis.
| Leptospirosis is a bacterial zoonosis caused by pathogenic Leptospira species which are transmitted to humans by exposure to water containing the urine of infected mammals, predominantly rodents [1]. The disease occurs worldwide and over 853,000 cases and 48,000 deaths are estimated to occur each year [2]. Incidence is highest in tropical regions, including the Asia-Pacific, Latin America, and the Caribbean, where there is an estimated incidence of >10 cases per 100,000 population per year [3] Around one billion people are thought to reside in urban slum areas where frequent outbreaks occur following heavy seasonal rains [4], most notably the recent epidemics in Nicaragua in 2007 and the Philippines in 2009 [5]. Despite the wide availability of effective antibiotic treatment, leptospirosis remains under recognized, mainly due to its non-specific clinical manifestations within a wide differential diagnostic spectrum. The current gold standard diagnostic techniques: culture and the microscopic agglutination test (MAT), are expensive, not useful for early diagnosis, require considerable expertise, and are impractical in resource poor settings. To date there is no widely deployable and reliable point-of-care test, meaning a large proportion of patients are never diagnosed or treated [6]. Outbreaks are often confused with viral infections, such as dengue fever, leading to delays in treatment and increased mortality [6–8].
Knowledge of the mortality from untreated leptospirosis is important for our understanding of the global burden of disease and the calculation of Disability Adjusted Life Years (DALYs), and will inform empirical fever treatment strategies and economic analyses [9]. The DALYs from leptospirosis are currently unknown and our understanding of the untreated mortality from leptospirosis remains limited. Mortality is thought to depend on host factors such as age, and bacterial factors such as serotype or inoculum size [4,6] but current estimates of mortality vary widely according to the clinical presentation, from 0% in patients with non-severe disease [10] to over 50% in those with Severe Pulmonary Haemorrhagic Syndrome (SPHS) [4,11]. This review aims to improve these estimates and better define the untreated mortality from leptospirosis, based on a comprehensive systematic review of previously published literature.
This review included all studies that contained untreated patients with leptospirosis. Patients were defined as untreated if they had received no leptospirosis-effective antibiotic treatment, were not treated with convalescent serum or admitted to an Intensive Care Unit (ICU), and did not receive dialysis. Included patients were of all ages and presented with symptoms consistent with leptospirosis and a confirmed diagnosis through either identification of leptospira by culture or direct microscopy, diagnosis by Polymerase Chain reaction (PCR) or serology through MAT. Patients admitted to hospital for supportive treatment (including IV fluids) were included in the analysis. All study designs and articles in all languages were included in the search. Studies were excluded if the diagnosis was on a clinical basis alone, there was no confirmed laboratory diagnosis for all patients in a clinical series, or if patients had asymptomatic infection. Studies with fewer than 10 patients were excluded to reduce patient selection bias. Inclusion and exclusion criteria are summarized in S1 Table. The primary outcome of the analysis was mortality from leptospirosis. Secondary outcomes were total days of fever, clinical signs and symptoms, and laboratory results for liver and kidney function where available. This review followed the PRISMA statement for systematic reviews (S1 Checklist).
Studies were identified through electronic resources, by scanning reference lists of relevant articles, and from library index catalogues, resulting in a comprehensive collection of published and peer reviewed full text articles. The electronic search was performed using Ovid MEDLINE (1946–Present), Embase Classic (1947–Present), and Global Health (1910–Present) on 28th July 2014 (S1–S3 Figs) and results reviewed manually. The search term used were: “Leptospirosis or leptospira or leptospir*; Weil's disease; Weil’s Syndrome; Swamp Fever; Mud fever; Autumn fever; Akiyami disease; Swineherds disease; Rice field fever; Cane cutters fever; Haemorrhagic Jaundice; Stuttgart disease; Canicola fever; Fort Bragg fever; icterohemorrhagic fever; seven day fever; dairy farm fever” and “Mortality or death”. Authors were not contacted regarding further information, no unpublished or grey literature was obtained, and studies were excluded if the full text was not available. Duplicate articles were removed using the reference manager “Mendeley” (2008–14 Mendeley Ltd, Version 1.12.1). One author (AT) reviewed the title and abstract, and papers were excluded if they did not fit the eligibility criteria. If there was doubt as to whether a paper was appropriate for inclusion then the whole paper was acquired and reviewed for eligibility.
One author (AT) extracted all available data and used Google Translate to translate non-English articles. Several articles contained more than one patient series, which were extracted separately, and all series were reviewed to prevent duplication of patient series. Treated patient subgroups were separated from untreated patient subgroups and excluded, and the whole patient series was excluded if separate outcomes were not clearly defined. There is no standardised method for assessing bias in non-interventional observational studies, and an existing data extraction sheet [12] was modified to create standardized criteria to assess bias within each study. Bias was graded according to patient selection and study design, diagnostic criteria, missing data and missing outcomes (S2 Table). To grade diagnostic certainty, we adapted diagnostic criteria from the WHO [13] and a grading system from Phommasone et al. [14]. Grade I diagnosis was identification of spirochetes through microscopy or culture, PCR, or a 4-fold rise in MAT titre, Grade II a single high MAT titre of ≥1:400 in an endemic region or ≥1:100 in a non-endemic region, and Grade III a single high titre MAT with no specified titre or a confirmed diagnosis but no record of the diagnostic method. Due to differences in methodology, inclusion criteria, missing data, and bias across patient series, a statistical meta-analysis was not performed. Each patient series was defined as a separate population and the median and range were used to summarize outcomes across patient series. The primary outcomes of the review was measured as the median mortality across all patient series and termed the “median series mortality”. Secondary outcomes were measured as the median value across patient series. For graphs, 95% Confidence Intervals (CI) of mortality were estimated using the Wilson score method [15]. Data was mapped using an image from NASA—Visible Earth.
A total of 35 studies, comprising 41 patient series, and containing a total of 4,247 patients were identified for inclusion in the review (Table 1). Within the included patient series, 857 patients had an unknown outcome, were treated, or had no laboratory diagnosis and were excluded (Table 1) and 3,390 patients were included in the final review. Six studies were excluded as the full article was not obtained and 3 excluded as the article could not be translated into English (S3 Table). Details of excluded articles with reasons for exclusion are displayed in Fig 1, with further information for each excluded article given in S4 Table.
Twenty-five articles were in English, 3 in Dutch, 3 in French, 2 in German, 1 in Spanish and 1 in Danish. The 41 patient series consisted of 25 retrospective patient series, 8 prospective patient series, 3 randomised control trials (RCTs), 3 non-randomised control trials (NRCTs) and 2 summaries of case reports, and were published between 1917 and 1984. Information on the location of the patient series was present for all case series (Fig 2). Eighteen series were located in Europe, 15 in Asia, 7 in the Americas and 1 in Africa. The median (range) number of patients in each series was 32 (10–459). All 41 patient series were designed to assess the clinical symptoms and outcome of leptospirosis. Each patient series was assessed for bias, and a summary of methodological quality across each criterion is displayed in S4 Fig, with further details for each series displayed in S5 and S6 Tables. Many patient series were limited by non-standardized study design and incomplete data. Diagnostic tests were at high risk of bias in 44% (18/41) of studies due to use of a single high admission titre with no confirmed cut-off titres, or no confirmed method of diagnosis for patients.
Mortality data was available for all (41/41) patient series, for a total of 3,390 patients. Median series mortality was 2.2% (range 0–39.7%) with a total of 314/3,390 deaths, and a wide variation in mortality across series. No deaths were reported in 16/41 patient series, but a mortality of 20% or more was reported in 7/41 patient series (Fig 3). Data for demographics, secondary outcomes and laboratory results are displayed in Table 2, but were not available from all patient series. Information on secondary outcomes were described in between 2–38 (4.8–92.7%) studies for each outcome, but often had heterogeneous definitions or were not reported numerically, meaning that data could not be extracted from many articles. Available data showed that fever, headache and myalgia occurred in nearly all patients, while conjunctival suffusion was reported in over half of the patients. Haemorrhagic symptoms ranged from epistaxis to more severe bleeds, but details of haemorrhage were not always specified so it is not possible to report on the incidence of severe haemorrhage or SPHS.
Mortality varied according to year and location of patient series and study design, with information present for all patient series (41/41). There was a wide range in mortality by year of study but no evident trend across time (S5 Fig). Between locations there was a wide range in mortality with median series mortality in Africa 9% (range n/a), 0.0% (range 0–39.7%) in the Americas, 1.0% (range 0–20%) in Asia, and 8.8% (range 0–26.3%) in Europe (S6 Fig). Median series mortality was high in 25 retrospective case series (8.0% (0.0–39.7)), and 2 patient series summarizing case reports (22.4% (20.0–24.8)) but lower in 6 controlled trials (0.0% (0–8.7%)) and 8 prospective case series (2.7% (0.0–19.6)) (S7 Table).
Median series mortality varied according to diagnostic grade and serovars. Mortality was lowest in patients with a grade I diagnosis and highest for those with a grade III diagnosis (Table 3). Across series with a grade II and III diagnosis the majority of patients were male and had a similar median age, while jaundice was highest in those with a grade III diagnosis, and lowest in those with a grade I diagnosis. Median series mortality was highest for serovar Icterohaemorrhagiae at 13.6% (0–34.3%), compared to 0.0% (0–50.0%) for Canicola and other serovars. Infections with serovar Icterohaemorrhagiae had a higher frequency of jaundice compared to other serovars (Table 4).
Data on mortality by age was available in 13/41 studies for 838 patients. Median series mortality was 0% (0–25%) in 7 series containing 51 patients aged 0–15, 16.3% (0–34.1%) in 11 series containing 308 patients aged 16–45, 36.7% (16.7–66.7%) in 6 series containing 70 patients aged 45–59, and 60.0% (33.3–60) in 3 series containing 23 patients aged over 60. Two large patient series, which were not incorporated due to differences in age stratification also showed a low mortality in children and a high mortality in older age groups. Smith [44] demonstrated a low mortality of 1% in 105 children aged 0–20 years compared to 50% in 18 patients aged 51 years or over, while Walch-Sorgdrager, [49] using the same cohort as Schuffner [41], demonstrated a mortality of 60% in 15 patients aged 60 years and over compared to 7.1% in 210 patients aged 10–40 years. Data on mortality by sex of patient was available in 21/41 patient series. Across 8 patient series the median series mortality for 227 female patients was 0% (range 0–40%), compared to 8.7% (range 0.0%- 39.7%) in 21 series containing 1077 male patients. No data was available on the untreated mortality of leptospirosis in pregnant women.
Frequency of jaundice was reported in 37/41 patient series, ranged from 0% to 100%, and was associated with increased mortality (Fig 4). In 8 patient series where the incidence of jaundice was 0%, median series mortality was 0% (range 0–1.7%) with 0.3% (1/348) overall mortality; while in 9 patient series where incidence of jaundice was 100%, the median series mortality was 19.1% (range 0.0–39.7%), with 21.6% (143/662) overall mortality. Data on renal function was reported in 12/41 studies and mortality increased with higher frequency of renal failure. In 4 patient series, with a total of 137 patients, where 29.8% (0.0–44.9%) patients had renal failure, median series mortality was 0% (0–3.4%), while in 8 patient series, with a total of 349 patients, where 80.5% (52.0–100%) patients had renal failure; the median series mortality was 12.1% (range 0–25.0%). Data on mortality in patients with meningitis was present in 12/41 studies for 188 patients with 4 deaths reported overall. Median series mortality in patients with meningitis was 0% (range 0–25%), with a low median incidence of jaundice in these patients of 12.2% (0–100%).
This systematic review is the first to comprehensively evaluate available literature to define the untreated mortality from leptospirosis. The median mortality of leptospirosis across all patient series was 2.2%. The mortality range, however, was broad (0–39.7%), reflecting the wide clinical spectrum of disease severity alongside the heterogeneity of study designs, inclusion criteria, and diagnostic methods used. Median series mortality was lower than previous reports, which often cite series with a high mortality, such as those summarizing case reports. This is highlighted by the high median series mortality of 22.4% across series summarizing case reports in this review. Compiled results showed that leptospirosis normally causes uncomplicated, anicteric disease with around 10 days of fever and a low mortality of less than 1%. Uncomplicated disease is therefore a major cause of morbidity, an important contributor to DALYs, and a significant burden to local health resources, but is not usually fatal. In a minority of cases however, especially the older population, more severe disease complicated by jaundice, renal failure, meningitis and death can occur.
Mortality was associated with host factors; the median series mortality in series including jaundiced patients approached 20% and in those with a high incidence of renal failure was over 12%, although data on renal failure was less reliable. These findings are consistent with previous findings [4,6,51,52], that show severe complications like jaundice and renal failure are associated with a significantly higher mortality. Interestingly, the median series mortality in patients with meningitis was found to be low at 0%, with 2.1% (4/188) overall mortality. This contrasts with a recent study which suggested that meningitis is associated with a significant mortality [53].
Mortality increased with age and was highest in those aged over 60 (60.0% (33.3–60)) but was negligible in children under 15 years of age at 0% (0–25%), although patient numbers were small in these age groups. Previous studies have reported a low mortality in untreated children and an increased mortality in untreated patients over 40 years, and it is likely that co-morbidities that occur with age such as diabetes and renal failure, alongside immunosenescence, increase mortality, although these factors were not accounted for in these patients [8,54–56]. Recent research is not suggestive of significant differences in mortality between sexes and the fact that the majority of patients were male makes it hard to show any reliable difference in untreated mortality between sexes [8,56]. The predominance of male patients may be explained by occupational risk factors such as farming and military work, which are more common in the male population. No information on mortality in pregnant women was available in this review, but the relative immunosuppression of pregnancy is thought to increase complications and mortality from leptospirosis [57].
Mortality is thought to be affected by pathogen factors and in this study mortality varied according to infective serovar and location, although there was a wide variation within each continent and region. Mortality was higher in Europe and North America than in Asia, which may reflect the predominance of serovar Icterohaemorrhagiae in studies from these regions, as serovar Icterohaemorrhagiae had a higher mortality of 13.1% (range 0.0–34.3) and an increased frequency of jaundice 62.9% (range 0–100) compared to other serovars. These facts underline the importance of understanding the local epidemiology of disease to predict disease outcomes, although identification of serovar is often inaccurate, with one recent study showing that MAT only correctly identified serovar in 33% of cases [58]. Furthermore, current opinion is that specific serovars are not associated with particular disease types [6]. It is also possible that mortality is affected by the endemicity of leptospirosis, as a previous study has shown that morbidity and mortality are lower in pre-exposed populations compared to populations with no previous exposure [59].
There are several limitations to this study. Study selection bias is likely due to the limitations of electronic searches in old literature, the use of reference lists to identify articles, the exclusion of unobtainable studies, and the under representation of modern studies. Furthermore, some studies were excluded because data on treatment or diagnosis was incomplete, meaning that results from this study may not be representative of outcomes in all untreated patients. The use of only one author and of “Google Translate” for data extraction may have introduced inaccuracies, although all extracted data was checked for errors. Data extraction and analysis were hindered by the age of studies with baseline patient characteristics often absent, outcomes imprecisely reported or missing, and laboratory tests not performed, while extraction of untreated patients from a larger cohort may have led to further inaccuracies. Analysis was performed by patient series, rather than by study, which may have overestimated findings from studies containing more than one patient series.
The characteristics of study populations are important when interpreting results. This review excluded patients with asymptomatic disease and those with mild self-limiting disease who did not seek medical attention. It has been shown that a large proportion of people in endemic communities are seropositive for leptospirosis [59] but have not sought medical attention, meaning that many mild cases go undiagnosed and unrecorded. The lack of access to healthcare and diagnostics, especially in remote and resource-limited settings may also prevent patients with severe disease from obtaining medical care and lead to an underrepresentation of severe cases. Both of these factors are likely to influence the final estimate of the mortality from untreated disease. The increased life expectancy of modern populations and the fact that the large majority of patients included in this review were male (83.2% (2140/2571)), must also be taken into account when applying results to a wider context. Retrospective case series were more likely to include patients with severe disease and thus overestimate mortality, while 5 controlled trials published after 1950 excluded patients with severe disease and reported a mortality of 0%, which is likely to reflect their patient selection criteria, as it would have been unethical to include untreated patients with severe disease after the introduction of antibiotics in the 1940s. Diagnostic methods mean that included patients may not represent the overall population. Culture methods are likely to have a high specificity but low sensitivity and thus miss many diagnoses of leptospirosis, while serological tests have been shown to lack both sensitivity and specificity and therefore may not accurately diagnose leptospirosis [60].
Leptospirosis remains a major, under appreciated and under recognized infection whose burden of disease falls disproportionately on those in poor and developing regions of the world [61]. This review clarifies the untreated mortality from leptospirosis and shows that despite wide variation, it is of high significance in elderly, jaundiced patients and/or those with renal failure, but much lower in younger, anicteric patients. The results from this study will support the quantification of DALYs from leptospirosis and may be used to guide empirical treatment strategies. A greater understanding of the true incidence of disease through increased surveillance should be encouraged to understand the pathogenicity of local serovars and guide local empirical treatment strategies, while improved genotypic based tests are required to more accurately predict the virulence of local strains [62]. The development of accurate and inexpensive point-of-care antigen based diagnostic tests for the diagnosis of leptospirosis early in its disease course would prevent the development of complications or death in this easily treatable disease [60,63]. Strategies for managing the disease should stress the importance of early empirical treatment of fever with effective antibiotics such as penicillin and doxycycline, which are cheap and widely available.
|
10.1371/journal.ppat.1007124 | Flap endonuclease 1 is involved in cccDNA formation in the hepatitis B virus | Hepatitis B virus (HBV) is one of the major etiological pathogens for liver cirrhosis and hepatocellular carcinoma. Chronic HBV infection is a key factor in these severe liver diseases. During infection, HBV forms a nuclear viral episome in the form of covalently closed circular DNA (cccDNA). Current therapies are not able to efficiently eliminate cccDNA from infected hepatocytes. cccDNA is a master template for viral replication that is formed by the conversion of its precursor, relaxed circular DNA (rcDNA). However, the host factors critical for cccDNA formation remain to be determined. Here, we assessed whether one potential host factor, flap structure-specific endonuclease 1 (FEN1), is involved in cleavage of the flap-like structure in rcDNA. In a cell culture HBV model (Hep38.7-Tet), expression and activity of FEN1 were reduced by siRNA, shRNA, CRISPR/Cas9-mediated genome editing, and a FEN1 inhibitor. These reductions in FEN1 expression and activity did not affect nucleocapsid DNA (NC-DNA) production, but did reduce cccDNA levels in Hep38.7-Tet cells. Exogenous overexpression of wild-type FEN1 rescued the reduced cccDNA production in FEN1-depleted Hep38.7-Tet cells. Anti-FEN1 immunoprecipitation revealed the binding of FEN1 to HBV DNA. An in vitro FEN activity assay demonstrated cleavage of 5′-flap from a synthesized HBV DNA substrate. Furthermore, cccDNA was generated in vitro when purified rcDNA was incubated with recombinant FEN1, DNA polymerase, and DNA ligase. Importantly, FEN1 was required for the in vitro cccDNA formation assay. These results demonstrate that FEN1 is involved in HBV cccDNA formation in cell culture system, and that FEN1, DNA polymerase, and ligase activities are sufficient to convert rcDNA into cccDNA in vitro.
| Hepatitis B virus (HBV) infection remains a worldwide health problem that affects more than 350 million people. HBV is one of the major etiological pathogens for liver cirrhosis and hepatocellular carcinoma. HBV covalently closed circular DNA (cccDNA) is a key viral intermediate for persistent infection. However, the molecular mechanism of cccDNA formation has not been clarified. Here, we found that the host factor flap-endonuclease 1 (FEN1) is pivotal in cccDNA formation. We developed a novel cccDNA formation assay by the incubation of purified viral DNA with recombinant FEN1, DNA polymerase, and DNA ligase. This study provides new insights into the molecular mechanisms of cccDNA formation and proposes FEN1 as a potential anti-HBV drug target.
| Hepatitis B virus (HBV) is a major pathogenic cause of human cirrhosis and hepatocellular carcinoma [1]. Infectious HBV particles contain relaxed circular DNA (rcDNA) encapsidated by core proteins [2]. After entering the host hepatocyte, rcDNA is converted into covalently closed circular DNA (cccDNA), which is stably maintained as an episome in the nucleus. cccDNA serves as the template for all HBV transcripts, including pregenomic RNA (pgRNA), a viral replicative intermediate [2–4]. pgRNA, viral reverse transcriptase P protein, and core proteins assemble into a nucleocapsid, where pgRNA undergoes reverse transcription by the P protein to produce rcDNA. The mature nucleocapsid is further assembled with surface proteins to allow secretion as an infectious virion. Alternatively, the rcDNA containing nucleocapsid is recycled back to the nucleus to maintain the pool of cccDNA [5].
Reverse-transcriptase inhibitors are the major medical intervention for controlling HBV infection. These inhibitors can effectively shut down viral replication, but are unable to eliminate cccDNA from infected hepatocytes; this inability often leads to viral rebound upon therapy withdrawal [2, 3, 6]. New therapeutic approaches are needed to target the mechanisms of cccDNA maintenance and generation. However, a lack of comprehensive knowledge on the molecular mechanisms of cccDNA formation and maintenance has hampered the effective development of such approaches.
The cccDNA precursor rcDNA has unique structural features that are absent from cccDNA. These include a P protein-linked sequence approximately 10 nucleotides in length, known as terminal redundancy (r), which is located at the 5′ end of the minus-strand DNA, and a small RNA oligomer attached at the 5′ end of the plus strand [2, 6]. The first step in cccDNA conversion from rcDNA is removal of the P protein and RNA oligomer linkage from the 5′ ends. Resulting protein-free rcDNA or deproteinized rcDNA is proposed to be a direct precursor to cccDNA [7, 8]. In addition to removing the r sequence and RNA oligomer from rcDNA, filling-in the single-stranded region and ligation of nicks in both DNA strands are required for cccDNA formation.
Flap endonuclease 1 (FEN1) is a flap structure-specific endonuclease. FEN1 plays a role in removing 5′-flap structures formed during Okazaki fragment maturation and long-patch base excision repair (LP-BER) [9, 10]. Because the r sequence and RNA oligomer at the 5′ end of rcDNA may form a 5′-flap structure, we examined the possible involvement of FEN1 in the removal of 5′-flap structures from rcDNA and its subsequent conversion to cccDNA.
To determine whether FEN1 protein removes the r sequence from rcDNA, we designed a synthetic DNA substrate that mimics the 5′-flap structure of the r sequence (S1A Fig) by modifying an established FEN assay [11]. Human wild-type (wt) and catalytic mutant FEN1 proteins were prepared by immunoprecipitation (S1B Fig) and used for the HBV-FEN assay. Cleavage of the r sequence was determined by fluorescence intensity (S1C Fig) and polyacrylamide gel electrophoresis (PAGE) (S1D Fig). Incubation with immunoprecipitated FEN1 protein caused an increase in cleavage of the synthetic r sequence over time when compared with that of the mock-precipitated protein (S1C Fig). Conversely, two catalytic mutant FEN1 proteins [12] lost their cleavage activity (S1D Fig). Previous studies demonstrated the inhibition of flap endonuclease activity by the FEN1 inhibitor 3-hydroxy-5-methyl-1-phenylthieno[2,3-d]pyrimidine-2,4(1H,3H)-dione (PTPD) [11, 13]. In the current study, we examined whether PTPD could inhibit FEN activity of the immunoprecipitated FEN1 protein by using the HBV FEN assay. PTPD addition strongly inhibited FEN1 cleavage activity (S1E Fig).
We used this inhibitor to explore the possible involvement of FEN1 in cccDNA formation in a cell culture system. Hep38.7-Tet cells replicate HBV and accumulate cccDNA after removing tetracycline from the culture medium [14, 15]. Using Hep38.7-Tet cells, the production of viral HBV intermediates was determined (Fig 1A–1E). Hirt extraction of cccDNA was followed by T5 exonuclease treatment to digest non-cccDNA molecules. T5 exonuclease removes nucleotides from 5′ termini, at gaps and nicks of linear or circular double-stranded DNA. The levels of cccDNA were determined by cccDNA-selective qPCR, which targets the gap region in rcDNA [16, 17] (Fig 1C). To demonstrate selective detection of cccDNA by the cccDNA-selective qPCR, a control experiment was performed. The Hirt extracted DNA and secreted HBV DNA were prepared from Hep38.7-Tet cells. The same copy number of the Hirt extracted DNA and secreted HBV DNA was applied to the cccDNA-selective qPCR. Our cccDNA-selective qPCR quantitatively detected cccDNA only from the Hirt-extracted HBV DNA, but not from secreted HBV DNA (S2 Fig).
To characterize the effect of PTPD treatment in Hep38.7-Tet cells, the effect of a reverse-transcriptase inhibitor, 3TC, was compared with that of PTPD. 3TC suppressed secreted and cytoplasmic nucleocapsid-associated DNA (cytoplasmic NC-DNA) and cccDNA levels (Fig 1A–1C). These results were expected, as HBV NC-DNA and cccDNA generation were completely dependent on reverse transcription in Hep38.7-Tet cells (S3 Fig), and 3TC was simultaneously added when Tet-CMV promoter was activated by removal of tetracycline from culture medium. Pre-C mRNA is transcribed from cccDNA, but not from the HBV transgene chromosomally integrated in cellular genome in the Hep38.7-Tet cells [18]. Consistent with the decreasing cccDNA levels in 3TC-treated Hep38.7-Tet cells, 3TC also reduced pre-C mRNA levels (Fig 1C and 1E). Conversely, PTPD significantly decreased both cccDNA and pre-C mRNA levels (Fig 1C and 1E). Importantly, PTPD did not affect the levels of secreted and cytoplasmic NC-DNAs (Fig 1A and 1B), confirming that transcription of pgRNA from the chromosomal copy was not affected by PTPD. Treating Hep38.7-Tet cells with PTPD (5 μM) for 5 days did not affect cellular proliferation (S4 Fig). These results suggest that PTPD blocked a step of cccDNA formation but did not inhibit reverse-transcription of pgRNA and cellular proliferation, as well as transcription from the HBV transgene.
The FEN1 inhibitor experiments suggested that FEN1 is involved in cccDNA formation, but not rcDNA formation. However, the observed reduction of cccDNA level by PTPD was moderate, compared to that of 3TC (Fig 1C). To confirm this result by different approaches, we performed small interfering RNA (siRNA)-based knockdown and genome editing. Two siRNAs designed against FEN1 mRNA and control siRNA were transfected into Hep38.7-Tet cells. Transfection of FEN1 siRNA reduced FEN1 mRNA and protein expression by more than half of the control siRNA levels (Fig 2A). Consistent with the result in Fig 1, knockdown of FEN1 expression reduced, albeit moderately, the cccDNA levels without affecting the cytoplasmic NC-DNA level generated from the HBV transgene (Fig 2B and 2C).
To further confirm the requirement for FEN1 in cccDNA formation, CRISPR/Cas9-mediated genome editing was applied to Hep38.7-Tet cells. We obtained two independent lines of FEN1+/− Hep38.7-Tet cells; each had one base (T) insertion in exon 2 of the FEN1 gene. The one-base insertion caused a frame shift and premature stop codon at amino acid position 102, immediately after insertion (S5 Fig). RT-qPCR and Western blot analyses demonstrated reduced FEN1 expression up to approximately half of the parental Hep38.7-Tet cells (Fig 2E). Consistent with the results obtained from the knockdown experiments, FEN1+/− Hep38.7-Tet cells produced cccDNA at approximately half the level of parental Hep38.7-Tet cells (Fig 2F). Southern blotting also showed moderately reduced cccDNA levels (Fig 2G, right), while intact cytoplasmic rcDNA production was observed in FEN1+/− Hep38.7-Tet cells (Fig 2G, left). Taken together, the knockdown and genome editing results clearly demonstrated that reduced FEN1 expression decreased cccDNA levels without reducing cytoplasmic rcDNA levels in Hep38.7-Tet cells.
Recent studies documented successful HBV infection in NTCP-expressing HepG2 cells [19, 20]. Thus, we examined the involvement of FEN1 in cccDNA formation using NTCP-expressing HepG2 (HepG2-hNTCP-C4) cells [19]. HepG2-hNTCP-C4 cells were pretreated with PTPD for 1 day, and subsequently infected with HBV. HBV-infected HepG2-hNTCP-C4 cells were cultivated for 3 days in the presence of PTPD, and cccDNA levels were determined by Southern blotting. As indicated in Fig 3A, the cccDNA level was mildly reduced by PTPD treatment (51.3% of control cccDNA level). Exposure of PTPD for 7 days in this cell line did not affect cellular proliferation (S6 Fig). To further confirm the results of infected HepG2-hNTCP-C4, we used PXB primary human hepatocytes derived from liver-humanized mice [21]. As shown in Fig 3B, PTPD treatment both inhibited secretion of HBV DNA and reduced HBV RNA levels in HBV-infected PXB cells. On the other hand, 3TC suppressed HBV DNA secretion but did not reduce HBV RNA levels. Importantly, in the infection model, rcDNA in the inoculum is first converted into cccDNA in the nucleus, and the newly produced cccDNA is transcribed into pgRNA, resulting in encapsidation and reverse-transcription in the nucleocapsid, yielding the mature virion (S3 Fig). Therefore, it is reasonable that PTPD treatment reduced both cccDNA formation and HBV DNA secretion in HBV-infected cells (Fig 3A and 3B). These results indicated that the FEN1 inhibitor blocks cccDNA formation following viral replication in infected NTCP-HepG2 cells and human primary hepatocytes.
It was previously reported that a point mutation (D181A) of FEN1 results in a loss of nuclease activity, while deletion of 20 amino acids from the C-terminus (ΔC) of FEN1 results in a loss of binding to the telomere maintenance protein, WRN, and truncation of nuclear localization signal (Fig 4A) [22–24]. We first tested the FEN activity of these mutants with the HBV-FEN assay used in Fig 1. As demonstrated in Fig 4B, FEN1 wt protein cleaved the r sequence, while FEN1 ΔC showed lower cleavage activity and D181A could not cleave r. Importantly, all protein levels were comparable (S7 Fig).
To assess the requirement of nuclease activity and the C-terminus of FEN1 for cccDNA formation, we knocked-down endogenous FEN1 expression and simultaneously overexpressed either FLAG-tagged-wt, D181A, or ΔC FEN1 protein using the pResQ vector [25]. The pResQ lentiviral expression vector simultaneously expresses both short hairpin RNA (shRNA) targeting the 3′-untranslated region (UTR) of FEN1 mRNA (shFEN1) and exogenous wt or mutant FEN1 protein (S8A Fig). As shown in S8B Fig, endogenous FEN1 expression in shFEN1-transduced cells (mock, wt, D181A, ΔC) was significantly lower than in cells transduced with control shRNA (shCtrl). Furthermore, overall FEN1 expression levels in shFEN1-expressing wt, D181A, and ΔC transfectants were substantially higher than in shCtrl- and shFEN1-expressing mock transfectants, due to the exogenous expression of FEN1 protein. Western blotting confirmed the reduction of endogenous FEN1 and the expression of exogenous FEN1 protein, although endogenous FEN1 protein is visible in knockdown cells (S8C Fig). Cytoplasmic NC-DNA and cccDNA levels were also determined in these transfectants. All cells produced cytoplasmic NC-DNA at similar levels, while shFEN1-mock transfectants exhibited lower levels of cccDNA that were restored by wt FEN1 expression (Fig 4C). In addition, D181A and ΔC mutant transfectants tended to exhibit lower cccDNA levels than that of wt, although this tendency was not statistically significant (Fig 4C). Although these experiments do not conclusively demonstrate cccDNA formation roles for the FEN1 protein catalytic site and C terminus, they do show that FEN1 expression is required for cccDNA formation.
Since the reduction of the cccDNA level in FEN1+/− Hep38.7-Tet cells was moderate (Fig 2G), we further examined the additive effect of genome editing and shRNA knockdown on cccDNA formation. pResQ shFEN1-mock lentiviral vector was transduced into two independent clones (#1 and #2) of FEN1+/− Hep38.7-Tet cells. Endogenous FEN1 protein was effectively reduced in these FEN1+/− shFEN1 cells (S8D Fig). Southern blotting analysis (Fig 4D) showed that FEN1+/− shFEN1 cells produced cytoplasmic NC-DNA as effectively as shRNA control cells (shCtrl). Meanwhile, the cccDNA level was clearly reduced in FEN1+/− shFEN1 cells compared with shRNA control cells (39.7 and 25.9% of the control cccDNA level, respectively). The additive effect of cccDNA reduction with the combination of FEN1+/−and FEN1 shRNA clearly indicates the requirement of FEN1 for cccDNA formation. It has been reported that rcDNA in Hirt extraction was reduced upon inhibition of cccDNA formation by knock-out of DNA ligases LIG1 and LIG3 [26]. It was proposed that nicked cccDNA behaves similar to rcDNA during electrophoresis, and a concurrent decrease of rcDNA may be due to a decrease in cccDNA formation. Reduction of rcDNA in Hirt DNA is also observed in our study (Figs 2G and 4D).
Subcellular localization of FEN1 protein was examined in HBV-replicating Hep38.7-Tet cells. As expected [25], wt FEN1 protein localized to the nucleus, which was disrupted by the ΔC mutation (Fig 5A). We next utilized immunoprecipitation in order to determine whether FEN1 can associate with HBV DNA. c-Myc-tagged-wt or ΔC FEN1-expressing Hep38.7-Tet cells (Fig 5B) were treated with formaldehyde to cross-link protein and DNA, and FEN1 proteins were immunoprecipitated using a c-Myc antibody. The cross linkage and fragmentation of DNA that were necessary for this approach can make it difficult to judge which viral DNA forms are precipitated during FEN1 immunoprecipitation. However, the ability of FEN1 to associate with any of the viral DNAs can be estimated by comparison of the immunoprecipitated HBV DNA levels in the FEN1 with the control IgG conditions. As predicted, a significantly higher level of HBV DNA was detected in the FEN1 wt precipitate compared with that in the control as well as with the ΔC mutant precipitation (Fig 5C). Importantly, the ΔC mutant, missing its nuclear targeting ability, exhibited decreased HBV DNA binding, relative to the wild type FEN1 protein. This finding suggests that wild type FEN1 localizes to the nucleus and associates with nuclear HBV DNA, such as nuclear rcDNA, either directly or indirectly.
Because FEN1 protein can remove the HBV r sequence in vitro (Fig 1A) and cellular experiments suggest a role of FEN1 in cccDNA formation (Figs 1–4), we assessed whether FEN1 can participate in any process of conversion of rcDNA to cccDNA in vitro. First, the FEN activity of recombinant FEN1 protein was reconfirmed by the HBV-FEN assay. Consistent with the results in Fig 1, recombinant FEN1 protein cleaved the r sequence from the synthetic HBV substrate in a dose-dependent manner (S9 Fig). Next, we determined whether recombinant enzymes, including FEN1, could convert the purified rcDNA into cccDNA. The purified NC-DNA from Hep38.7-Tet cells was incubated with recombinant FEN1, DNA polymerase, and DNA ligase, and cccDNA formation was determined by cccDNA selective-qPCR, rolling circle amplification (RCA), and Southern blot (Fig 6A–6D). RCA is able to speficically amplify closed circular DNA. The combination of FEN1, DNA polymerase, and DNA ligase led to the significant production of cccDNA (Fig 6B–6D). Meanwhile, incubation with two enzymes (DNA polymerase and DNA ligase) did not support efficient cccDNA formation (Fig 6B–6D). DNA sequencing of the closed circular DNA produced by three enzymes confirmed that the rcDNA gap region was precisely filled and did not have any mutations (S10 Fig). Furthermore, replication competency of in vitro-generated cccDNA was confirmed by transfecting the self-circularized RCA product into HepG2 cells (Fig 6E and 6F). These results indicate that the circular DNA generated by incubating with FEN1, DNA polymerase, and DNA ligase is functional HBV cccDNA.
Host DNA repair factors are expected to be involved in cccDNA conversion because the virus genome does not encode the responsible DNA modifiers [2, 3, 6, 27]. The TDP2 enzyme has been proposed to remove P protein [28], although another study reported that TDP2 is not required for cccDNA formation in vivo [29]. We previously showed that the host DNA repair enzyme UNG removes uracil residues from deaminated duck HBV (DHBV) cccDNA (or its precursor), thus changing its mutation frequency [30]. FEN1 plays a role in various DNA metabolic pathways, including Okazaki fragment maturation and LP-BER. During lagging strand DNA synthesis, Polδ/Polε use the strand exchange activity to produce the 5′-flap structure, and then FEN1 cleaves the 5′-flap and generates a ligatable end to facilitate lagging strand synthesis. In LP-BER, FEN1 removes the 5′-flap structure containing a damaged sugar and generates a ligatable 5′ end to facilitate its repair process [9, 10]. However, it remains unknown whether other back-up systems can substitute for absence of FEN1 activity.
During cccDNA formation, rcDNA-specific structures have to be removed. We assume that some of these rcDNA-specific structures form the 5′-flap structure. FEN1 is a good candidate to remove them. To test this, we utilized FEN1 loss-of-functional approaches in HBV-replicating cells, including a FEN1 inhibitor (Figs 1 and 3), siRNA and shRNA knockdown (Figs 2 and 4), and CRISPR/Cas9-mediated genome editing (Fig 2). The four different approaches of FEN1 loss of function showed the same trend, that is, a moderate reduction in cccDNA levels. We also utilized a combined approach of the genome editing and shRNA knockdown (Fig 4D). This method resulted in FEN1+/−-FEN1 shRNA Hep38.7-Tet cells with a clearly reduced cccDNA level. This reduction seemed to be specific to cccDNA because NC-DNA production was not reduced in Hep38.7-Tet cells, which could produce NC-DNA from genome integrated HBV transgene. We interpreted this reduced level of cccDNA as a specific phenotype of FEN1 loss-of-function, rather than off-target effects from each approach. Infection experiments showed that inhibition of FEN1 activity reduced HBV DNA secretion in NTCP-expressing cells and primary human hepatocytes (Fig 3). FEN1 protein could cleave the r sequence in vitro (Figs 4 and S1) and convert purified rcDNA into cccDNA along with DNA polymerase and ligase in vitro (Fig 6). Altogether, these results demonstrate, for the first time, that the host DNA repair factor FEN1 is involved in HBV cccDNA formation, at least in the experimental models used in this study.
The cccDNA was not eliminated completely, even when the combination approach for FEN1 loss-of-function was employed, suggesting the possibility of other redundant enzymes. FEN1 is a member of the RAD2/XPG structure-specific 5′-nuclease family [31, 32]. Among them, exonuclease 1 (Exo1) is another candidate to remove the HBV r sequence from rcDNA, because it has both 5′ to 3′ exonuclease activity and endonuclease activity of the 5′-flap structure. Moreover, yeast genetic studies suggested that Exo1 and FEN1 activities may have a redundant role [33, 34]. The human genome encodes another flap-structure specific endonuclease designated DNA2. DNA2 plays a role to resolve a flap structure during Okazaki fragment maturation in yeast [35]. Further studies are needed to determine the host players other than FEN1 that remove the flap structure from rcDNA.
Inhibition of FEN1 activity did not lead to an obvious reduction in proliferation, at least in the experimental conditions used in this study (S4 and S6 Figs). Consistent with our observation, it was reported that the PTPD inhibitor showed little or no effect on cell growth of the T24 bladder cell line, but increased sensitivity to a DNA damage agent, i.e. methyl methanesulfonate [13]. Moreover, FEN1 mutations that abrogate nuclease activity have been detected in lung cancers and corresponding knock-in mice are viable with autoimmune, chronic inflammatory, and cancer phenotypes [36]. Meanwhile another knock-in mice of FEN1 mutant (F343A and F344A) that lose ability to bind PCNA but retain nuclease activity, die at birth [37]. On the other hand, FEN1−/−mice, which have a complete knock-out of FEN1, have a lethality phenotype as early as embryonic day 3.5 [38]. Recent biochemical and genetic approaches identified more than 30 FEN1 associating proteins [32] including proteins involving in DNA replication, such as PCNA, apoptosis, telomere stability, post-transcriptional modification, and DNA repair. It is likely that complete loss of FEN1 protein in mammal manifests as a disturbance of cellular survival because both nuclease-dependent and -independent functions of FEN1 are disrupted. Meanwhile, inhibition of nuclease activity of FEN1 may not result in immediate disturbance of cellular proliferation. Consistent with this idea, the FEN1−/− cell line was not established in this study, even by CRISPR/Cas9-mediated genome editing.
The HBV-FEN assay revealed that FEN1 could remove the r sequence from a synthetic HBV DNA flap substrate. Moreover, the combination of FEN1, DNA polymerase, and DNA ligase was sufficient to convert cccDNA from purified rcDNA in vitro. It has been reported that Polδ and Lig I cooperate with FEN1 in Okazaki fragment maturation, and Polβ and Lig III cooperate with FEN1 in LP-BER [10, 32]. However, the specific polymerase and ligase involved in HBV cccDNA formation remain unknown. Interestingly, the T5 exonuclease-resistant cccDNA (Fig 6C, top) migrated at approximately 3.4 kbp which is much higher position than that of cccDNA formed in infected hepatocytes, suggesting that its topology was different from cellular cccDNA. It is also possible for other additional factors such as topoisomerase and gyrase to be involved in cccDNA formation in vivo. Thus, further studies need to determine other host factors responsible for cccDNA formation.
In summary, we demonstrate that reduced FEN1 expression and activity decreases cccDNA levels, and that FEN1 protein can bind and cleave the 5′-flap structure of HBV rcDNA in vitro to facilitate cccDNA conversion. The data implicate FEN1 as a critical enzyme involved in HBV cccDNA formation.
Hep38.7-Tet cells derived from the HepAD38 cell line (obtained from Dr. Christoph Seeger at Fox Chase Cancer Center, Philadelphia) [15], and HepG2-hNTCP-C4 cells derived from HepG2 cells (obtained from the JCRB Cell Bank) [19] were cultured as described previously. Hep38.7-Tet cells were cultured with 0.3 μg/ml tetracycline to terminate HBV transcription. HBV production was induced in the cells by incubation in a tetracycline-free medium. 293FT cells (purchased from Invitrogen) were cultured as described previously [39]. PXB primary human hepatocytes were derived from liver-humanized mice [21]. The culture medium was purchased from PhoenixBio. For FEN1 inhibition experiments, PTPD (3-hydroxy-5-methyl-1-phenylthieno[2,3-d]pyrimidine-2,4(1H,3H)-dione; Glixx Laboratories) [13] was added to the culture medium.
The HBV-FEN assay was performed as described previously [11] with minor modifications. Wild-type (wt) and mutant human FEN1 proteins were produced by transfecting FEN1 expression vectors [12] (S1 Table) into 293FT cells and enriched by immunoprecipitation with anti-c-Myc agarose affinity gel (A7470; Sigma-Aldrich), as described below. The DNA substrate was prepared by annealing of “flap,” “quencher,” and “template” oligonucleotides containing the HBV sequences listed in S2 Table (also see S1A Fig). Since 5-carboxytetramethylrhodamine (TAMRA) is attached to the 5′ end of the r sequence corresponding to the flap oligonucleotide, cleavage of the flap oligonucleotide by FEN activity can be measured as increasing fluorescence. DNA substrates were incubated with either FEN1 immunoprecipitants at room temperature or recombinant protein Thermostable FEN1 (Thermococcus species 9°N origin; New England Biolabs) at 65°C. Kinetic fluorescence data were collected on PowerScan (DS Pharma Biomedical). Cleavage of the labeled, “flap” oligonucleotide was confirmed with 6 M urea/20% polyacrylamide gel electrophoresis (S1D Fig).
Purification of HBV DNA (supernatant, cytoplasmic NC-DNA, and cccDNA) and total RNA were described previously [30] with minor modification. HBV DNA in culture supernatant was extracted using a NucleoSpin kit (Takara) according to the manufacturer’s protocol. The purified HBV DNA from this fraction is designated as secreted HBV in this study. Viral DNAs from enveloped virions and naked capsids are included in this fraction. For cytoplasmic NC-DNA, the cells were lysed with buffer [10 mM Tris-HCl (pH 8.0), 1 mM EDTA, 1% NP-40, 8% sucrose, proteinase inhibitor cocktail (Roche)]. After centrifugation, supernatants were collected and further treated with DNase I and RNase A. NCs were then digested with proteinase K and sodium dodecyl sulfate (SDS). The cccDNA purification was performed using a modified Hirt extraction procedure [30]. The Hirt-extracted DNA was purified and treated with T5 exonuclease (New England Biolabs) to digest linear and open circular DNA according to the manufacturer’s instructions. Total RNA was treated with amplification grade DNase I (Thermo Fisher Scientific) and reverse transcribed using an oligo (dT) primer and the SuperScript III kit (Thermo Fisher Scientific). qPCR analysis of resulting cDNA was performed using SYBR green ROX (Toyobo) with MX3000 (Stratagene) as described previously [40]. For cccDNA quantification, TaqMan probe and cccDNA-selective primers spanning the gap region of rcDNA were used [16, 17]. Validation of selective amplification for cccDNA but not rcDNA, is shown in S2 Fig. Primers and probe sequences are listed in S2 Table.
RCA was performed as described previously [30, 41] using purified DNA from the Hirt extraction. In brief, the DNA (T5 exonuclease treated) was mixed with 8 HBV-specific primers (S2 Table), denatured at 95°C for 3 min; cooled sequentially at 50°C for 15 s, 37°C for 15 s, and room temperature, and reacted with the phi29 DNA polymerase (New England Biolabs) at 37°C for 16 h. RCA products were digested with EcoRI, which cuts HBV cccDNA once. The digested products were analyzed by gel electrophoresis and ethidium bromide staining.
Immunoprecipitation and Western blotting were performed as described previously [30, 40]. To analyze FEN1-HBV DNA binding, cells were fixed with 1% formaldehyde for 10 min at room temperature, quenched with 125 mM glycine, resuspended in lysis buffer (50 mM Tris-HCl pH 8.0, 5 mM EDTA, 150 mM NaCl, 1% Nonidet P-40, 0.1% sodium dodecyl sulfate [SDS], protease inhibitor), sonicated in a Bioruptor sonication device (Diagenode) for 10 min using pulses of 30 s, and immunoprecipitated with anti-c-Myc antibody (9E10, sc-40; Santa Cruz Biotechnology) and protein G Sepharose (GE Healthcare) overnight at 4°C. Following proteinase K/SDS digestion, DNA was extracted with phenol/chloroform and precipitated with ethanol. Target DNA fragments were analyzed by qPCR as described above. The antibodies used for Western blotting were: mouse anti-FEN1 (4E7, GTX70185, GeneTex), rabbit anti-glyceraldehyde-3-phosphate dehydrogenase (GAPDH) (G9545; Sigma-Aldrich), mouse anti-FLAG (F3165, Sigma-Aldrich), mouse anti-c-Myc (9E10, sc-40; Santa Cruz Biotechnology), anti-rabbit Igs-horseradish peroxidase (HRP) (ALI3404; Biosource), and anti-rabbit/anti-mouse IgG-HRP TrueBlot (18–8816 and 18–8877; eBioscience).
Southern blotting was performed as described previously [30]. HBV DNAs were detected using a probe spanning the entire viral genome. Rhamda DNA probe was also simulateiously added to hybridzation buffer to visualize DNA size marker. Probe labeling and signal development was performed using the AlkPhos direct labeling system (Amersham), and the signals were detected using the LAS1000 imager system (Fuji Film). The Hirt-extracted DNAs were heat denatured at 95°C for 10 min and then subjected to EcoRI digestion to linearize DNAs. For the in vitro cccDNA formation (Fig 6C), HBV DNAs were treated with T5 exonuclease (New England Biolabs) to eliminate any DNAs, except double-stranded closed circular DNA. After phenol-chloroform extraction, DNAs were digested with EcoRI, and then agarose gel electrophoresis was performed.
Cell viability was evaluated using the Premix WST-1 Cell Proliferation Assay System (Takara) according to the manufacturer's instructions. The cell lines used for the WST-1 assay were sensitive to puromycin; hence, puromycin was used as a control.
Two FEN1-specific siRNAs and control siRNA were purchased from Santa Cruz Biotechnology (sc-37795, sc-37007) and Sigma-Aldrich (SASI_Hs02_00336939). Lipofectamine 3000 (Thermo Fisher Scientific) was used to perform transfections with these siRNAs according to the manufacturer’s instructions. Cells and viruses were analyzed 4 days after transfection.
Human FEN1-targeting oligonucleotides (target sequence with the protospacer adjacent motif is in exon 2: 5′-AGCTGGCCAAACGCAGTGAGCGG-3′) were designed and cloned into the BbsI site of the pX330-U6-Chimeric_BB-CBh-hSpCas9 vector (a gift from Feng Zhang, Addgene plasmid # 42230) [42]. The resulting pX330-FEN1 vector was co-transfected into Hep38.7-Tet cells with pIRES-GFP-bsd, a blasticidin-resistant gene expression vector (S1 Table). Transfected clones were then selected using limiting dilution in the presence of blasticidin, and genome editing was confirmed by direct sequencing of the targeted region (oligonucleotides are listed in S2 Table).
Lentivirus-mediated gene transduction was performed as described previously [40], using pResQ shFEN3 3XF-FEN1 wt, pResQ shFEN3 3XF-FEN1 D181A, and pResQ shFEN3 3XF-FEN1 ΔC (gifts from Sheila Stewart, Addgene plasmid # 17752, 17753, and 17754, respectively) [25]. Construction of pResQ vectors is described in S7A Fig and S1 Table.
HBV infection was performed as described previously [19, 21]. Briefly, HBV (genotype D) was prepared from the culture supernatant of Hep38.7-Tet cells and concentrated with PEG8000 precipitation. The amount of HBV DNA was quantified by qPCR as described above. HepG2-hNTCP-C4 cells and PXB cells were seeded in collagen-coated plates, and the medium was replaced with fresh medium containing 4% PEG8000 and the prepared HBV (15,000 GE/cell for HepG2-hNTCP-C4 infection, 100 GE/cell for PXB infection). Twenty-four hours post-infection, the infected cells were washed three times with phosphate buffered saline and switched to fresh medium with PTPD or lamivudine (3TC). The cells and culture supernatants were collected at the indicated days post infection (d.p.i.).
Purified NC-DNA from Hep38.7-Tet cells was used as substrate DNA. NC-DNA (108 copies) was incubated with 32 units (U) of Thermostable FEN1 in ThermoPol Buffer (New England Biolabs) at 65°C for 10 min, followed by incubation with 8 U of Bst DNA polymerase, 40 U of Taq DNA ligase, 100 μM dNTPs, and NAD+ (all from New England Biolabs). After further incubation at 37°C for 20 min, DNA was purified by phenol/chloroform extraction and ethanol precipitation, and subjected to cccDNA-selective qPCR or RCA, as described above. The sequence corresponding to gap region of rcDNA was confirmed by direct sequencing (oligonucleotide is listed in S2 Table). To verify the replication competence of resulting products, EcoRI-digested RCA products (3.2 kb) were extracted from the gel; 50 ng of these fragments were subjected to self-circularization by T4 DNA ligase (Takara). For the negative control, HBV plasmid (pPB [30]) was amplified with RCA and then digested with PstI. Because the 5.4-kb PstI fragment has a partial HBV sequence, it was used as a replication-defective control. Self-circularized DNAs were transfected into HepG2 cells. Three days after transfection, HBV DNAs were analyzed by qPCR.
Statistical analyses were performed using GraphPad Prism (GraphPad Software). Significance between two groups was determined using a Student’s t-test, while significance between three or more groups was determined using a one-way ANOVA with Dunnett's post-hoc test. P-values < 0.05 were considered statistically significant.
|
10.1371/journal.pgen.1000416 | The Complete Genome and Proteome of Laribacter hongkongensis Reveal Potential Mechanisms for Adaptations to Different Temperatures and Habitats | Laribacter hongkongensis is a newly discovered Gram-negative bacillus of the Neisseriaceae family associated with freshwater fish–borne gastroenteritis and traveler's diarrhea. The complete genome sequence of L. hongkongensis HLHK9, recovered from an immunocompetent patient with severe gastroenteritis, consists of a 3,169-kb chromosome with G+C content of 62.35%. Genome analysis reveals different mechanisms potentially important for its adaptation to diverse habitats of human and freshwater fish intestines and freshwater environments. The gene contents support its phenotypic properties and suggest that amino acids and fatty acids can be used as carbon sources. The extensive variety of transporters, including multidrug efflux and heavy metal transporters as well as genes involved in chemotaxis, may enable L. hongkongensis to survive in different environmental niches. Genes encoding urease, bile salts efflux pump, adhesin, catalase, superoxide dismutase, and other putative virulence factors—such as hemolysins, RTX toxins, patatin-like proteins, phospholipase A1, and collagenases—are present. Proteomes of L. hongkongensis HLHK9 cultured at 37°C (human body temperature) and 20°C (freshwater habitat temperature) showed differential gene expression, including two homologous copies of argB, argB-20, and argB-37, which encode two isoenzymes of N-acetyl-L-glutamate kinase (NAGK)—NAGK-20 and NAGK-37—in the arginine biosynthesis pathway. NAGK-20 showed higher expression at 20°C, whereas NAGK-37 showed higher expression at 37°C. NAGK-20 also had a lower optimal temperature for enzymatic activities and was inhibited by arginine probably as negative-feedback control. Similar duplicated copies of argB are also observed in bacteria from hot springs such as Thermus thermophilus, Deinococcus geothermalis, Deinococcus radiodurans, and Roseiflexus castenholzii, suggesting that similar mechanisms for temperature adaptation may be employed by other bacteria. Genome and proteome analysis of L. hongkongensis revealed novel mechanisms for adaptations to survival at different temperatures and habitats.
| Laribacter hongkongensis is a recently discovered bacterium associated with gastroenteritis and traveler's diarrhea. Freshwater fish is the reservoir of L. hongkongensis. In order to achieve a rapid understanding on the mechanisms by which the bacterium adapts to different habitats and its potential virulence factors, we sequenced the complete genome of L. hongkongensis, compared its gene contents with other bacteria, and compared its gene expression at 37°C (human body temperature) and 20°C (freshwater habitat temperature). We found that the gene contents of L. hongkongensis enable it to adapt to its diverse habitats of human and freshwater fish intestines and freshwater environments. Genes encoding proteins responsible for survival in the intestinal environments, adhesion to intestinal cells, evasion from host immune systems, and putative virulence factors similar to those observed in other pathogens are present. We also observed, in gene expression studies, that L. hongkongensis may be using different pathways for arginine synthesis regulated at different temperatures. Phylogenetic analysis suggested that such mechanisms for temperature adaptation may also be used in bacteria found in extreme temperatures.
| Laribacter hongkongensis is a recently discovered, Gram-negative, facultative anaerobic, motile, seagull or S-shaped, asaccharolytic, urease-positive bacillus that belongs to the Neisseriaceae family of β-proteobacteria [1]. It was first isolated from the blood and thoracic empyema of an alcoholic liver cirrhosis patient in Hong Kong [2]. In a prospective study, L. hongkongensis was shown to be associated with community acquired gastroenteritis and traveler's diarrhea [3],[4]. L. hongkongensis is likely to be globally distributed, as travel histories from patients suggested its presence in at least four continents: Asia, Europe, Africa and Central America [4]–[6]. L. hongkongensis has been found in up to 60% of the intestines of commonly consumed freshwater fish, such as grass carp and bighead carp [4],[7],[8]. It has also been isolated from drinking water reservoirs in Hong Kong [9]. Pulsed-field gel electrophoresis and multilocus sequence typing showed that the fish and patient isolates fell into separate clusters, suggesting that some clones could be more virulent or adapted to human [8],[10]. These data strongly suggest that this bacterium is a potential diarrheal pathogen that warrants further investigations.
Compared to other families such as Enterobacteriaceae, Vibrionaceae, Streptococcaceae, genomes of bacteria in the Neisseriaceae family have been relatively under-studied. Within this family, Neisseria meningitidis, Neisseria gonorrhoeae and Chromobacterium violaceum are the only species with completely sequenced genomes [11]–[13]. In view of its potential clinical importance, distinct phylogenetic position, interesting phenotypic characteristics and the availability of genetic manipulation systems [14]–[17], we sequenced and annotated the complete genome of a strain (HLHK9) of L. hongkongensis recovered from a 36-year old previously healthy Chinese patient with profuse diarrhea, vomiting and abdominal pain [4]. Proteomes of L. hongkongensis growing at 37°C (body temperature of human) and 20°C (average temperature of freshwater habitat in fall and winter) [9] were also compared.
The complete genome of L. hongkongensis is a single circular chromosome of 3,169,329 bp with a G+C content of 62.35% (Figure 1). In terms of genome size and number of predicted coding sequences (CDSs), rRNA operons and tRNA genes (Table 1), L. hongkongensis falls into a position intermediate between C. violaceum and the pathogenic Neisseria species. A similar intermediate status was also observed when the CDSs were classified into Cluster of Orthologous Groups (COG) functional categories, except for genes of RNA processing and modification (COG A), cell cycle control, mitosis and meiosis (COG D), replication, recombination and repair (COG L) and extracellular structures (COG W), of which all four bacteria have similar number of genes (Figure 2). This is in line with the life cycles and growth requirements of the bacteria. C. violaceum is a highly versatile, facultative anaerobic, soil- and water-borne free-living bacterium and therefore requires the largest genome size and gene number. The pathogenic Neisseria species are strictly aerobic bacteria with human as the only host and therefore require the smallest genome size and gene number. L. hongkongensis is a facultative anaerobic bacterium that can survive in human, freshwater fish and 0–2% NaCl but not in marine fish or ≥3% NaCl and therefore requires an intermediate genome size and gene number.
The L. hongkongensis genome lacks a complete set of enzymes for glycolysis, with orthologues of glucokinase, 6-phosphofructokinase and pyruvate kinase being absent (Table S1). This is compatible with its asaccharolytic phenotype and is consistent with other asaccharolytic bacteria, such as Campylobacter jejuni, Bordetella pertussis, Bordetella parapertussis and Bordetella bronchiseptica, in that glucokinase and 6-phosphofructokinase are also absent from their genomes [18],[19]. On the other hand, the L. hongkongensis genome encodes the complete sets of enzymes for gluconeogenesis, the pentose phosphate pathway and the glyoxylate cycle (Table S1). Similar to C. jejuni, the L. hongkongensis genome encodes a number of extracellular proteases and amino acid transporters. These amino acids can be used as carbon source for the bacterium. The genome encodes enzymes for biosynthesis of the 21 genetically encoded amino acids and for biosynthesis and β-oxidation of saturated fatty acids (Tables S2 and S3). The L. hongkongensis genome encodes a variety of dehydrogenases (LHK_00527–00540, LHK_01219–01224, LHK_02418–02421, LHK_00801–00803, LHK_01861, LHK_02912–02913 and LHK_00934) that enable it to utilize a variety of substrates as electron donors, such as NADH, succinate, formate, proline, acyl-CoA and D-amino acids. The presence of three terminal cytochrome oxidases may allow L. hongkongensis to carry out respiration using oxygen as the electron acceptor under both aerobic conditions [type aa3 oxidase (LHK_00169–00170, LHK_00173)] and conditions with reduced oxygen tension [type cbb3 (LHK_00995–00996, LHK_00998) and type bd (LHK_02252–02253) oxidases]. The genome also encodes a number of reductases [fumarate reductase (LHK_02340–02342), nitrate reductase (LHK_02079–02085), dimethylsulfoxide (DMSO) reductase (LHK_02496–02498) and tetrathionate reductase (LHK_01476–01478)], which may help carry out respiration with alternative electron acceptors to oxygen (fumarate, nitrate, DMSO and tetrathionate) under anaerobic conditions. This is supported by the enhanced growth of L. hongkongensis under anaerobic conditions in the presence of nitrate (data not shown). Further studies are required to confirm if the bacterium can utilize other potential electron acceptors.
There were 441 transport-related proteins (13.6% of all CDSs) in the L. hongkongensis genome, comprising an extensive variety of transporters, which may reflect its ability to adapt to the freshwater fish and human intestines, and freshwater environments. According to the Transporter Classification Database (TCDB) (http://www.tcdb.org/), all seven major categories of transporters are present in L. hongkongensis. Primary active transporters (class 3 transporters) were the most abundant class of transporters, accounting for 43.3% (191 CDSs) of all annotated CDSs related to transport, among which 104 belong to the ATP-binding cassette (ABC) transporter superfamily and 41 were oxidoreduction-driven transporters. Electrochemical potential-driven transporters (class 2 transporters) were the second most abundant class of transporters, accounting for 27.9% (123 CDSs) of all annotated CDSs related to transport, most of which (117 CDSs) are various kinds of porters including major facilitator superfamily (MFS) (19 CDSs), resistance-nodulation-cell division (RND) superfamily (22 CDSs), amino acid-polyamine-organocation family (8 CDSs), dicarboxylate/amino acid∶cation symporter (DAACS) family (5 CDSs) and monovalent cation∶proton antiporter-2 family (3 CDSs), and various heavy metal transporters which may be involved in detoxification and resistance against environmental hazards. Three different types of class 2 transporters, belonging to the DAACS, tripartite ATP-independent periplasmic transporter and C4-dicarboxylate uptake C family, are likely involved in the transport of malate, which can be used as the sole carbon source for L. hongkongensis in minimal medium [unpublished data]. The remaining class 2 transporters were ion-gradient-driven energizers belonging to the TonB family (6 CDSs). The third most abundant class of transporters was the channels and pores (class 1), with 39 CDSs including 12 α-type channels, 26 β-barrel porins. Among the 12 α-type channels, four were mechanosensitive channels which are important for mediating resistance to mechanophysical changes. The remaining transporters belong to four other classes, namely group translocators (class 4, 9 CDSs), transport electron carriers (class 5, 16 CDSs), accessory factors involved in transport (class 8, 9 CDSs) and incompletely characterized transport system (class 9, 54 CDSs). In line with their asaccharolytic nature, the genomes of L. hongkongensis and C. jejuni do not contain genes that encode a complete phosphotransferase system. The five families of multidrug efflux transporters, including MFS (6 CDSs), RND (8 CDSs), small multidrug resistance family (2 CDSs), multidrug and toxic compound extrusion family (2 CDSs) and ABC transporter superfamily (5 CDSs), were all present in L. hongkongensis, which may reflect its ability to withstand toxic substances in different habitats [20]. 20 CDSs were related to iron metabolism, including hemin transporters, ABC transporters of the metal type and ferrous iron, iron-storage proteins and the Fur protein responsible for iron uptake regulation. In contrast to C. violaceum which produces siderophores for iron acquisition, but similar to the pathogenic Neisseria species, proteins related to siderophore formation are not found in L. hongkongensis genome. In addition to a TonB-dependent siderophore receptor (LHK_00497), a set of genes (LHK_01190, LHK_01193, LHK_01427–1428) related to the transport of hemin were present, suggesting that L. hongkongensis is able to utilize exogenous siderophores or host proteins for iron acquisition, which may be important for survival in different environments and hosts.
Except the first strain of L. hongkongensis isolated from the blood and empyema pus of a patient which represented a non-motile variant, all L. hongkongensis strains, whether from human diarrheal stool, fish intestine or environmental water, are motile with polar flagella. The ability to sense and respond to environmental signals is important for survival in changing ecological niches. A total of 47 CDSs are related to chemotaxis, of which 27 encode methyl-accepting chemotaxis proteins (MCPs) and 20 encode chemosensory transducer proteins. While most MCPs are scattered throughout the genome, the transducer proteins are mostly arranged in three gene clusters (Figure S1). At least 38 genes, in six gene clusters, are involved in the biosynthesis of flagella (Figure S2).
Enteric bacteria use several quorum-sensing mechanisms, including the LuxR-I, LuxS/AI-2, and AI-3/epinephrine/norepinephrine systems, to recognize the host environment and communicate across species. Unlike the genomes of C. violaceum and the pathogenic Neisseria species which encode genes involved in LuxR-I and LuxS/AI-2 systems respectively, the L. hongkongensis genome does not encode genes of these 2 systems. Instead, the AI-3/epinephrine/norepinephrine system, which is involved in inter-kingdom cross-signaling and regulation of virulence gene transcription and motility, best characterized in enterohemorrhagic E. coli [21],[22], is likely the predominant quorum-sensing mechanism used by L. hongkongensis. Several human enteric commensals or pathogens, including E. coli, Shigella, and Salmonella, produce AI-3 [23]. A two-component system, QseB/C, of which QseC is the sensor kinase and QseB the response regulator, has been found to be involved in sensing AI-3 from bacteria and epinephrine/norepinephrine from host, and activation of the flagellar regulon transcription [21]. While the biosynthetic pathway of AI-3 has not been discovered, two sets of genes, LHK_00329/LHK_00328 and LHK_01812/LHK_01813, homologous to QseB/QseC were identified in the L. hongkongensis genome, suggesting that the bacterium may regulate its motility upon recognition of its host environment. The presence of two sets of QseB/QseC, one most similar to those of C. violaceum and the other most homologous to Azoarcus sp. strain BH72, is intriguing, as the latter is the only bacterium, with complete genome sequence available, that possesses two copies of such genes.
Before reaching the human intestine, L. hongkongensis has to pass through the highly acidic environment of the stomach. In the L. hongkongensis genome, a cluster of genes, spanning a 12-kb region, related to acid resistance, is present. Similar to Helicobacter pylori, the L. hongkongensis genome contains a complete urease gene cluster (LHK_01035–LHK_01037, LHK_01040–LHK_01044), in line with the bacterium's urease activity. Phylogenetically, all 8 genes in the urease cassette are most closely related to the corresponding homologues in Brucella species (α-proteobacteria), Yersinia species (γ-proteobacteria) and Photorhabdus luminescens (γ-proteobacteria), instead of those in other members of β-proteobacteria, indicating that L. hongkongensis has probably acquired the genes through horizontal gene transfer after its evolution into a distinct species (Figure S3). Upstream and downstream to the urease cassette, adi (LHK_01034) and hdeA (LHK_01046) were found respectively. Their activities will raise the cytoplasmic pH and prevents proteins in the periplasmic space from aggregation during acid shock respectively [24],[25]. In addition to the acid resistance gene cluster, the L. hongkongensis genome contains two arc gene clusters [arcA (LHK_02729 and LHK_02734), arcB (LHK_02728 and LHK_02733), arcC (LHK_02727 and LHK_02732) and arcD (LHK_02730 and LHK_02731)] of the arginine deiminase pathway which converts L-arginine to carbon dioxide, ATP, and ammonia. The production of ammonia increases the pH of the local environment [26],[27].
Similar to other pathogenic bacteria of the gastrointestinal tract, the genome of L. hongkongensis encodes genes for bile resistance. These include three complete copies of acrAB (LHK_01425–01426, LHK_02129–02130 and LHK_02929–02930), encoding the best studied efflux pump for bile salts, and two pairs of genes (LHK_01373–01374 and LHK_03132–03133) that encode putative efflux pumps homologous to that encoded by emrAB in E. coli [28]. Furthermore, five genes [tolQ (LHK_00053), tolR (LHK_03174), tolA (LHK_03173), tolB (LHK_03172) and pal (LHK_03171)] that encode the Tol proteins, important in maintaining the integrity of the outer membrane and for bile resistance, are also present [29].
In the L. hongkongensis genome, a putative adhesin (LHK_01901) for colonization of the intestinal mucosa, most closely related to the adhesins of diffusely adherent E. coli (DAEC) and enterotoxigenic E. coli (ETEC), encoded by aidA and tibA respectively, was observed (Figure S4) [30],[31]. aidA and tibA encode proteins of the autotransporter family, type V protein secretion system of Gram-negative bacteria. All the three domains (an N-terminal signal sequence, a passenger domain and a translocation domain) present in proteins of this family are found in the putative adhesin in L. hongkongensis. Moreover, a putative heptosyltransferase (LHK_01902), with 52% amino acid identity to the TibC heptosyltransferase of ETEC, responsible for addition of heptose to the passenger domain, was present upstream to the putative adhesin gene in the L. hongkongensis genome (Figure S4). In addition to host cell adhesion, the passenger domains of autotransporters may also confer various virulence functions, including autoaggregation, invasion, biofilm formation and cytotoxicity. The L. hongkongensis genome encodes a putative superoxide dismutase (LHK_01716) and catalases (LHK_01264, LHK_01300 and LHK_02436), which may play a role in resistance to superoxide radicals and hydrogen peroxide generated by neutrophils.
The same set of genes that encode enzymes for synthesis of lipid A (endotoxin), the two Kdo units and the heptose units of lipopolysaccharide (LPS) are present in the genomes of L. hongkongensis, C. violaceum, N. meningitidis, N. gonorrhoeae and E. coli. Moreover, 9 genes [rfbA (LHK_02995), rfbB (LHK_02997), rfbC (LHK_02994), rfbD (LHK_02996), wbmF (LHK_02799), wbmG (LHK_02800), wbmH (LHK_02801), wbmI (LHK_02790) and wbmK (LHK_02792)] that encode putative enzymes for biosynthesis of the polysaccharide side chains are present in the L. hongkongensis genome. In addition to genes for synthesizing LPS, a number of CDSs that encode putative cytotoxins are present, including cytotoxins that act on the cell surface [hemolysins (LHK_00956 and LHK_03166) and RTX toxins (LHK_02735 and LHK_02918)] and those that act intracellularly [patatin-like proteins (LHK_00116, LHK_01938, and LHK_03113)] [32],[33]. Furthermore, a number of CDSs that encode putative outer membrane phospholipase A1 (LHK_00790) and collagenases (LHK_00305–00306, LHK_00451, and LHK_02651) for possible bacterial invasion are present.
To better understand how L. hongkongensis adapts to human body and freshwater habitat temperatures at the molecular level, the types and quantities of proteins expressed in L. hongkongensis HLHK9 cultured at 37°C and 20°C were compared. Since initial 2D gel electrophoresis analysis of L. hongkongensis HLHK9 proteins under a broad range of pI and molecular weight conditions revealed that the majority of the proteins reside on the weakly acidic to neutral portion, with a minority on the weak basic portion, consistent with the median pI value of 6.63 calculated for all putative proteins in the genome of L. hongkongensis HLHK9, we therefore focused on IPG strips of pH 4–7 and 7–10. Comparison of the 2D gel electrophoresis patterns from L. hongkongensis HLHK9 cells grown at 20°C and 37°C revealed 12 differentially expressed protein spots, with 7 being more highly expressed at 20°C than at 37°C and 5 being more highly expressed at 37°C than at 20°C (Table 2, Figure 3). The identified proteins were involved in various functions (Table 2). Of note, spot 8 [N-acetyl-L-glutamate kinase (NAGK)-37, encoded by argB-37] was up-regulated at 37°C, whereas spot 1 (NAGK-20, encoded by argB-20), was up-regulated at 20°C (Figures 3, 4A and 4B). These two homologous copies of argB encode two isoenzymes of NAGK [NAGK-20 (LHK_02829) and NAGK-37 (LHK_02337)], which catalyze the second step of the arginine biosynthesis pathway.
The transcription levels of argB-20 and argB-37 at 20°C and 37°C were quantified by real time RT-PCR. Results showed that the mRNA level of argB-20 at 20°C was significantly higher that at 37°C and the mRNA level of argB-37 at 37°C was significantly higher that at 20°C (Figure 4C and 4D), suggesting that their expressions, similar to most other bacterial genes, were controlled at the transcription level. When argB-20 and argB-37 were cloned, expressed and the corresponding proteins NAGK-20 and NAGK-37 purified for enzyme assays, their highest enzymatic activities were observed at 37–45°C and 45–50°C respectively (Figure 4E). Moreover, NAGK-20, but not NAGK-37, was inhibited by 0.25–10 mM of arginine (Figure 4F).
L. hongkongensis probably regulates arginine biosynthesis at temperatures of different habitats using two pathways with two isoenzymes of NAGK. L. hongkongensis and wild type E. coli ATCC 25922, but not E. coli JW5553-1 (argB deletion mutant), grew in minimal medium without arginine, indicating that L. hongkongensis contains a functional arginine biosynthesis pathway. NAGK-20 is expressed at higher level at 20°C than 37°C, whereas NAGK-37 is expressed at higher level at 37°C than 20°C. Bacteria use either of two different pathways, linear and cyclic, for arginine biosynthesis. Similar to NAGK-20 of L. hongkongensis, NAGK of Pseudomonas aeruginosa and Thermotoga maritima, which employ the cyclic pathway, can be inhibited by arginine as the rate-limiting enzyme for negative feedback control [34]–[37]. On the other hand, similar to NAGK-37 of L. hongkongensis, NAGK of E. coli, which employs the linear pathway, is not inhibited by arginine [35],[36]. We speculate that L. hongkongensis can use different pathways with the two NAGK isoenzymes with differential importance at different temperatures of different habitats.
Phylogenetic analysis of NAGK-20 and NAGK-37 showed that they were more closely related to each other than to homologues in other bacteria (Figure 5). The topology of the phylogenetic tree constructed using NAGK was similar to that constructed using 16S rRNA gene sequences (data not shown). This suggested that the evolution of argB genes in general paralleled the evolution of the corresponding bacteria, and argB gene duplication has probably occurred after the evolution of L. hongkongensis into a separate species. The requirement to adapt to different temperatures and habitats may have provided the driving force for subsequent evolution to 2 homologous proteins that serve in different environments. Notably, among all 465 bacterial species with complete genome sequences available, only Thermus thermophilus, Deinococcus geothermalis, Deinococcus radiodurans, Roseiflexus castenholzii and Roseiflexus sp. RS-1 possessed two copies of argB, whereas Anaeromyxobacter sp. Fw109-5 and Anaeromyxobacter dehalogenans 2CP-C possessed one copy of argB and another fused with argJ (Figure 5). The clustering of argB in two separate groups in these bacteria suggests that argB gene duplication has probably occurred in their ancestor, before the divergence into separate species. The prevalence of T. thermophilus, Deinococcus species and Roseiflexus species in hot springs suggested that this novel mechanism of temperature adaptation may also be important for survival at different temperatures in other bacteria. Further experiments on differential expression of the two isoenzymes at different temperatures in these bacteria will verify our speculations.
Traditionally, complete genomes of bacteria with medical, biological, phylogenetic or industrial interests were sequenced only after profound phenotypic and genotypic characterization of the bacteria had been performed. With the advance in technology and bioinformatics tools, complete genome sequences of bacteria can be obtained with greater ease. In this study, we sequenced and analyzed the complete genome of L. hongkongensis, a newly discovered bacterium of emerging medical and phylogenetic interest, and performed differential proteomics and downstream characterization of important pathways. In addition, putative virulence factors and a putative novel mechanism of arginine biosynthesis regulation at different temperatures were discovered, further characterization of which will lead to better understanding of their contributions to the survival and virulence of L. hongkongensis, the Neisseriaceae family and other bacteria. A similar “reverse genomics” approach can be used for the study of other newly discovered important bacteria.
The genome sequence of L. hongkongensis HLHK9 was determined with the whole-genome shotgun method. Three shotgun libraries were generated: one small-insert (2–4 kb) library and one medium-insert (5–6 kb) library in pcDNA2.1, and a large-insert (35–45 kb) fosmid library in pCC2FOS. DNA sequencing was performed using dye-terminator chemistries on ABI3700 sequencers. Shotgun sequences were assembled with Phrap. Fosmid end sequences were mapped onto the assembly using BACCardI [38] for validation and support of gap closing. Sequences of all large repeat elements (rRNA operons and prophages) were confirmed by primer walking of fosmid clones.
The nucleotide sequence for the complete genome sequence of L. hongkongensis HLHK9 was submitted to Genbank under accession number CP001154.
Gene prediction was performed by Glimmer [39] version 3.02, and results post-processed using TICO [40] for improving predictions of translation initiation sites. Automated annotation of the finished sequence was performed by a modified version of AutoFACT [41], supplemented by analysis by InterProScan [42]. Manual curation of annotation results was done with support from the software tool GenDB [43]. In addition, annotation of membrane transport proteins was done by performing BLAST search of all predicted genes against the curated TCDB [44]. Ribosomal RNA genes were annotated using the online RNAmmer service [45]. Putative prophage sequences were identified using Prophage Finder [46]. Frameshift errors were predicted using ProFED [47]. CRISPRs (Clustered Regularly Interspaced Short Palindromic Repeats) were searched by using PILER-CR [48], CRISPRFinder [49] and CRT (CRISPR recognition tool) [50].
Single colony of L. hongkongensis HLHK9 was inoculated into brain heart infusion (BHI) medium for 16 h. The bacterial cultures were diluted 1∶100 in BHI medium and growth was continued at 20°C for 20 h and 37°C for 6 h, respectively, with shaking to OD600 of 0.6. After centrifugation at 6,500×g for 15 min, cells were lysed in a sample buffer containing 7 M urea, 2 M thiourea and 4% CHAPS. The crude cell homogenate was sonicated and centrifuged at 16,000×g for 20 min. Immobilized pH gradient (IPG) strips (Bio-Rad Laboratories) (17 cm) with pH 4–7 and 7–10 were hydrated overnight in rehydration buffer containing 7 M urea, 2 M thiourea, 4% CHAPS, 1% IPG buffer pH 4–7 (IPG strip of pH 4–7) and pH 6–11 (IPG strip of pH 7–10) (GE Healthcare) and 60 mM DTT with 60 µg of total protein. The first dimension, isoelectric focusing (IEF), was carried out in a Protean IEF cell electrophoresis unit (Bio-Rad Laboratories) for about 100,000 volt-hours. Protein separation in the second dimension was performed in 12% SDS-PAGE utilizing the Bio-Rad Protean II xi unit (Bio-Rad Laboratories). 2D gels were stained with silver and colloidal Coomassie blue G-250 respectively for qualitative and quantitative analysis, and scanned with ImageScanner (GE Healthcare). ImageMaster 2D Platinum 6.0 (GE Healthcare) was used for image analysis. For MALDI-TOF MS analysis, protein spots were manually excised from gels and subjected to in-situ digestion with trypsin, and peptides generated were analyzed using a 4800 Plus MALDI TOF/TOF Analyzer (Applied Biosystems). Proteins were identified by peptide mass fingerprinting using the MS-Fit software (http://prospector.ucsf.edu) and an in-house sequence database of L. hongkongensis HLHK9 proteins generated using the information obtained from the complete genome sequence and annotation. Only spots with at least two-fold difference in their spot volume between 20°C and 37°C and those uniquely detected at either temperature were subjected to protein identification by MALDI-TOF MS analysis. Three independent experiments for each growth condition were performed.
L. hongkongensis HLHK9 cells were grown in minimal medium M63 [51] supplemented with 20 mM L-malate as carbon source and 19 mM potassium nitrate as nitrogen source, and 1 mM each of vitamin B1 and vitamin B12. The pH of all media was adjusted to 7.0 with KOH. Essentiality of arginine for growth of L. hongkongensis HLHK9 was determined by transferring the bacterial cells to the modified M63 medium with or without 100 mM of L-arginine. Escherichia coli ATCC 25922 and JW5553-1 (argB deletion mutant) [52] were used as positive and negative controls respectively. All cultures were incubated at 37°C with shaking for 5 days. Growth in each medium was determined by measuring absorbance spectrophotometrically at OD600. The experiment was performed in duplicate.
mRNA levels of argB-20 and argB-37 in L. hongkongensis HLHK9 cells grown in 20°C and 37°C were compared. Total RNA was extracted from culture of L. hongkongensis HLHK9 (OD600 of 0.6) grown in conditions described in proteomic analysis by using RNeasy kit (Qiagen) in combination with RNAprotect Bacteria Reagent (Qiagen) as described by the manufacturer. Genomic DNA was removed by DNase digestion using RNase-free DNase I (Roche). The total nucleic acid concentration and purity were estimated using A260/A280 values measured by NanoDrop ND-1000 spectrophotometer (NanoDrop Technologies). Bacteria were harvested from three independent replicate cultures. cDNA was synthesized by RT using random hexamers and SuperScript III kit (Invitrogen) as described previously [53],[54]. cDNA was amplified by TaqMan PCR Core Reagent kit (Applied Biosystems) in an ABI Prism 7000 Sequence Detection System (Applied Biosystems). Briefly, 2 µl of cDNA was amplified in a 25 µl reaction containing 2.5 µl of 10× TaqMan buffer A, 5.5 mM of MgCl2, 0.2 mM of each deoxynucleoside triphosphates (dNTPs), 0.8 µM of each primer, 0.8 µM of gene-specific TaqMan probe with a 5′-[6-carboxyfluorescein (6-FAM)] reporter dye and a 3′-[6-carboxytetramethylrhodamine (TAMRA)] quencher dye, 2.5 U of AmpErase Uracil N-glycosylase (UNG) and 0.625 U AmpliTaq Gold polymerase (Applied Biosystems). Primers and TaqMan probes were designed using Primer Express software, version 2.0 (Applied Biosystems) (Table S4). Reactions were first incubated at 50°C for 2 min, followed by 95°C for 10 min in duplicate wells. Reactions were then thermal-cycled in 40 cycles of 95°C for 15 s and 60°C for 1 min. Absolute standard curve method was used for determination of transcript level for each gene. Standard curves were made by using serial dilutions from plasmids containing the target sequences with known quantities. Housekeeping gene RNA polymerase beta subunit, rpoB, was used as an internal control. Triplicate assays using RNAs extracted in three independent experiments confirmed that transcript levels of rpoB were not significantly different (P>0.05) at 20°C compared with 37°C (data not shown). The transcript levels of argB-20 and argB-37 were then normalized to that of rpoB. Triplicate assays using RNAs extracted in three independent experiments were performed for each target gene.
The phylogenetic relationships among NAGK-20 and NAGK-37 of L. hongkongensis HLHK9 and their homologues in other bacteria with complete genomes available were analyzed. Phylogenetic tree was constructed by the neighbor-joining method using Kimura's two-parameter correction with ClustalX 1.83. Three hundred and eleven positions were included in the analysis.
Cloning and purification of (His)6-tagged recombinant NAGK proteins of L. hongkongensis HLHK9 was performed according to our previous publications, with modifications [53],[55]. To produce plasmids for protein purification, primers (5′- GGAATTCCATATGCTGCTTGCAGACGCCC -3′ and 5′- GGAATTCCATATGTCAGGCTGCGCGGATCAT -3′ for argB-20 and 5′- GGAATTCCATATGGTTATTCAATCTGAAGT -3′ and 5′- GGAATTCCATATGTCAGAGCGTGGTACAGAT -3′ for argB-37) were used to amplify the genes encoding NAGK-20 and NAGK-37, respectively, by PCR. The sequence coding for amino acid residues of the complete NAGK-20 and NAGK-37 was amplified and cloned, respectively, into the NdeI site of expression vector pET-28b(+) (Novagen) in frame and downstream of the series of six histidine residues. The two recombinant NAGK proteins were expressed and purified using the Ni2+-loaded HiTrap Chelating System according to the manufacturer's instructions (GE Healthcare).
Purified NAGK-20 and NAGK-37 were assayed for N-acetyl-L-glutamate kinase activity using Haas and Leisinger's method [56], with modifications. The reaction mixtures contained 400 mM NH2OH⋅HCl, 400 mM Tris⋅HCl, 40 mM N-acetyl-L-glutamate, 20 mM MgCl2, 10 mM ATP and 2 µg of enzyme in a final volume of 1.0 ml at pH 7.0. After incubation at 25°C, 30°C, 37°C, 45°C, 50°C, 55°C or 60°C for 30 min, the reaction was terminated by adding 1.0 ml of a stop solution containing 5% (w/v) FeCl3⋅6H2O, 8% (w/v) trichloroacetic acid and 0.3 M HCl. The absorbance of the hydroxamate⋅Fe3+ complex was measured with a spectrophotometer at A540 [57]. Inhibition of the kinase activities of NAGK-20 and NAGK-37 were examined with and without 0.25, 0.5, 0.75, 1, 2.5, 5, 10, and 20 mM of L-arginine and incubated at 37°C for 30 min. One unit of N-acetyl-L-glutamate kinase is defined as the amount of enzyme required to catalyze the formation of 1 µmol of product per min under the assay conditions used. Each assay was performed in duplicate. Results were presented as means and standard deviations of three independent experiments.
|
10.1371/journal.ppat.1005862 | MIF-Mediated Hemodilution Promotes Pathogenic Anemia in Experimental African Trypanosomosis | Animal African trypanosomosis is a major threat to the economic development and human health in sub-Saharan Africa. Trypanosoma congolense infections represent the major constraint in livestock production, with anemia as the major pathogenic lethal feature. The mechanisms underlying anemia development are ill defined, which hampers the development of an effective therapy. Here, the contribution of the erythropoietic and erythrophagocytic potential as well as of hemodilution to the development of T. congolense-induced anemia were addressed in a mouse model of low virulence relevant for bovine trypanosomosis. We show that in infected mice, splenic extramedullary erythropoiesis could compensate for the chronic low-grade type I inflammation-induced phagocytosis of senescent red blood cells (RBCs) in spleen and liver myeloid cells, as well as for the impaired maturation of RBCs occurring in the bone marrow and spleen. Rather, anemia resulted from hemodilution. Our data also suggest that the heme catabolism subsequent to sustained erythrophagocytosis resulted in iron accumulation in tissue and hyperbilirubinemia. Moreover, hypoalbuminemia, potentially resulting from hemodilution and liver injury in infected mice, impaired the elimination of toxic circulating molecules like bilirubin. Hemodilutional thrombocytopenia also coincided with impaired coagulation. Combined, these effects could elicit multiple organ failure and uncontrolled bleeding thus reduce the survival of infected mice. MIF (macrophage migrating inhibitory factor), a potential pathogenic molecule in African trypanosomosis, was found herein to promote erythrophagocytosis, to block extramedullary erythropoiesis and RBC maturation, and to trigger hemodilution. Hence, these data prompt considering MIF as a potential target for treatment of natural bovine trypanosomosis.
| Bovine African trypanosomosis is a parasitic disease of veterinary importance that adversely affects the public health and economic development of sub-Saharan Africa. Anemia is a major cause of death associated with this disease. Yet, the mechanisms underlying anemia development are not elucidated, which hampers the design of effective therapeutic strategies. We show here that in a Trypanosoma congolense infection mouse model relevant for bovine trypanosomosis, red blood cells (RBCs) are generated in the spleen. This compensates for the impaired maturation of RBCs occurring in the bone marrow, the normal site of RBC generation, and for the destruction of RBCs taking place in the liver and the spleen. Instead, anemia results from an increase in blood volume (hemodilution). The immune molecule Macrophage Migration Inhibitory Factor (MIF) was found to drive RBC destruction, to block RBC maturation, as well as to trigger hemodilution. Iron accumulation in tissue due to sustained RBC destruction and hemodilution causes tissue damage, which culminates in the release of toxic molecules like bilirubin, in impaired production of blood detoxifying molecules like albumin, and in defective coagulation. Combined, these effects initiate multiple organ failure that can reduce the survival of infected mice. Given the unmet medical need for this parasite infection, our findings offer promise for improved treatment protocols in the field.
| African trypanosomosis (AT) is a neglected tropical disease of medical and veterinary importance that adversely affects human health and welfare, as well as the economic development in sub-Saharan Africa [1,2]. AT is caused by blood-borne hemoflagellated protozoan parasites from the Trypanosoma genus that are transmitted by the hematophagous tsetse fly (Glossina spp.) vector [3]. Trypanosoma-induced diseases in mammals include sleeping sickness in humans or nagana in domestic livestock, with fatal consequences unless treated [4,5]. Given that so far no effective vaccine is available, that certain trypanosome strains have become resistant to curative and preventive treatments, and that eradication of tsetse flies remains impossible in most regions [1,6], strategies focusing at reducing AT-associated pathogenicity might be an alternative approach to reduce the economic losses in cattle production.
In case of bovine trypanosomosis, the main difference between so-called trypano-susceptible and -tolerant animals relies in their capacity to control anemia development, the major cause of death associated with the disease [7,8], and hereby to remain productive [9]. Differences in erythropoietic potential have been suggested as a contributing factor to anemia development [8]. Yet, the mechanisms underlying this phenomenon remain poorly understood, which hampers the design of effective therapeutic strategies. Given the similarities between the anemic responses of cattle and C57Bl/6 mice upon trypanosome infection [8], the underlying mechanisms were mainly scrutinized in murine models [10,11]. The data collectively suggest that the pro-inflammatory type I immune response, involving TNF, IFN-γ and M1-type (classically activated) myeloid cells, contributes to pathogenicity in general and anemia in particular, and in combination with impaired B-cell functionality, results in reduced survival of the mice [10,12,13]. Thus, identification of gene-products regulating pro-inflammatory signals during the course of the disease might pave the way to develop novel intervention strategies. In this context, we previously identified macrophage migration inhibitory factor (MIF) as a potential susceptibility marker for African trypanosomosis. This ubiquitously produced cytokine is a prominent inducer of systemic inflammation in many inflammatory diseases [14,15] that acts by recruiting and activating myeloid cells towards M1-type cells to the site of inflammation [16,17], and by suppressing apoptosis of inflammatory cells [18]. We have shown using a trypanosusceptible model based on C57Bl/6 mice infected with T. brucei brucei, that MIF contributes to tissue pathogenicity by sustaining throughout infection a persistent type I pro-inflammatory chemokine (CXCL1, CCL2) and cytokine (IFN-γ, TNF, IL-6) response, and by enhancing the recruitment of Ly6C+ monocytes and neutrophils (PMNs) in the liver with concomitant hepatomegaly [19]. Moreover, PMN- but not monocyte-derived MIF was mainly responsible for liver damage. In addition, MIF promotes the development of so-called anemia of inflammation in trypanosusceptible mice by enhancing red blood cell (RBC) clearance from the blood, and by triggering the storage of iron in liver myeloid cells that deprives iron from erythropoiesis and impairs the generation of mature RBCs. However, despite reduced liver injury and anemia levels in T. b. brucei-infected Mif-/- mice or mice treated with a neutralizing anti-MIF antibody, the host survival time was not affected [19].
Compared to T. b. brucei, T. congolense infection in C57Bl/6 mice is considered a trypanotolerant model more relevant for bovine trypanosomosis [20,21]. In contrast to T. b. brucei, T. congolense causes a chronic infection (3–4 months versus 1 month), due to the capacity of T. congolense-infected mice to restrain the type I immune response and to switch to an IL-10-mediated tissue protective anti-inflammatory response [21,22]. This model could thus allow a thorough analysis of the mechanisms underlying anemia development in an immune environment different from that of anemia of inflammation. Here, we evaluated the role of MIF in T. congolense infection-associated anemia development, by focusing on the modulation of the erythropoietic and erythrophagocytic potential in tissues including the bone marrow, the spleen and the liver. Additionally, the contribution of hemodilution to anemia was addressed.
The systemic production of MIF increased progressively during the course of T. congolense infection (S1 Fig). Hence, the potential role of MIF in the outcome of i.p. infection was evaluated by comparing wild type (WT) and MIF-deficient (Mif-/-) C57Bl/6 mice. Although parasitemia development was similar in WT and Mif-/- mice, a prolongation of median survival time occurred in Mif-/- mice (Fig 1A and 1B). Similar observations were obtained using a natural route, tsetse fly-mediated infection model (S2A and S2B Fig). Considering the similar capacity of WT and Mif-/- mice to control parasite growth, the increased survival of Mif-/- mice could result from lower tissue pathogenicity [23,24]. In agreement, as compared to WT mice, Mif-/- mice exhibited reduced serum AST (alanine aminotransferase, reflecting systemic tissue injury) and ALT (aspartate aminotransferase, reflecting liver injury) levels (Fig 1C and 1D), as well as reduced hepato- and splenomegaly that coincided with a lower increase in the number of white blood cells (WBC) in the liver and the spleen of Mif-/- mice (Figs 1E and 1F and S3). No WBC accumulation was observed in the bone marrow of infected WT and Mif-/- mice (Fig 1F).
The differences in pathogenicity between WT and Mif-/- infected mice were clearly established from 3 months post infection (p.i.) (Figs 1C–1F and S3), hence we focused at this time point. At 3 months p.i., MIF secretion was enhanced in the liver, spleen and bone marrow of infected WT mice (Fig 2), mirroring the increased MIF level measured in the blood. In agreement with the reduced tissue injury in infected Mif-/- mice, the levels of neutrophil (CXCL1) and monocyte (CCL2) chemoattractants, as well as of pro-inflammatory cytokines documented to contribute to T. congolense-induced tissue destruction (TNF, IL-6, IFN-γ) [25,26], were increased to a lesser degree in the liver, spleen, bone marrow and blood of Mif-/- than WT mice (Fig 2A–2D). This was also true for IL-12p70 in the blood (Fig 2D) [27]. However, the systemic and tissue levels of IL-10, which increases upon T. congolense infection and is crucial to limit tissue destruction [28], did not differ between WT and Mif-/- mice (Fig 2).
In addition to differences in pro-inflammatory cytokine production, the decreased tissue pathogenicity and increased survival of Mif-/- mice also could be due to a superior ability of the Mif-/- mice to mount a parasite-specific antibody response [13]. As shown in Figs 3A and S4, similar serum levels of parasite-specific IgG antibodies were recorded in WT and Mif-/- mice until 1.5 months p.i.; thereafter and from 3 months p.i., the IgG levels declined in WT mice while they remained elevated in Mif-/- mice. The drop in IgG levels in infected WT mice did not correlate with a decrease in the number of total (B220+) or germinal center (GL-7+Fas+B220+) splenic B-cells (Figs 3B and 3C, S5A and S5B), but could be due to an increase in B-cell apoptosis (Fig 3D). The increased IgG levels observed in infected Mif-/- mice as compared to WT mice was associated with an increased number of total and germinal center B-cells as well as with lower B-cell apoptosis (Fig 4A–4D).
Collectively, in T. congolense-infected mice, the absence of MIF results in a reduced pro-inflammatory immune response, which in turn could contribute to a lower hepato-splenomegaly and an enhanced B-cell response that collectively could enhance the survival time.
Anemia is the prominent pathogenic feature of a natural T. congolense infection, which is mediated by hematopoietic cells, but not by T lymphocytes or antibodies [10,29]. Since MIF (i) contributes to the accumulation of WBC (including phagocytes; Fig 1F) in tissues that are potentially involved in erythrophagocytosis and extramedullary erythropoiesis, and (ii) stimulates the production of the erythroid lineage development-blocking cytokine IL-6 [30] in the two main erythropoietic tissues (bone marrow, spleen; Fig 2B and 2C), we investigated MIF’s role in anemia development during T. congolense infection.
Anemia is characterized by two distinct phases: (i) a rapid decline in RBC levels followed by partial recovery during the early phase of infection and (ii) a more progressive decline in RBC levels during the chronic phase of infection (Fig 4A). During the early phase of infection (i.e. day 5–10 p.i.), the RBC percentages initially dropped to about 50% of non-infected mice in both WT and Mif-/- mice (Fig 4A). Between day 10–14 p.i., a partial recovery that reaches about 75% of the RBC level in non-infected mice occurred in WT mice, while in Mif-/- mice this recovery reached about 95% (Fig 4A). Subsequently, during the chronic phase of infection, the RBC levels declined progressively and remained significantly lower in WT than Mif-/- mice. Mif-/- mice also exhibited reduced anemia in a natural tsetse fly-mediated infection (S2C Fig). The serum hemoglobin and iron levels were reduced in WT mice as compared to non-infected animals during the course of infection (Fig 4B and 4C). In infected Mif-/- mice, these reductions were less pronounced when compared to infected WT mice. In both groups of mice, the serum hemoglobin and iron levels reached nadir levels at 3 months p.i., the time point when the differences in tissue pathogenicity were established (Figs 1C and 1D and 4A).
Collectively, during the chronic stage of T. congolense infection, MIF partially impaired recovery from early stage anemia and contributed to the decline in serum hemoglobin and iron levels.
Typically, in response to chronic anemia, splenomegaly and increased blood reticulocyte content are indicative of inefficient erythropoiesis in the bone marrow and extramedullary erythropoiesis in the spleen [31]. Compared to non-infected mice, the numbers of splenic Ter119+ RBCs increased more prominently in T. congolense-infected WT than Mif-/- mice from 3 months p.i. (Fig 4D). Remarkably, the largest increase in cell number in the spleens of infected mice was found in the RBC and not the WBC compartment (Fig 4D), showing the expansion of the erythroid compartment as a main cause for T. congolense-associated splenomegaly.
The relative abundance of Ter119+CD71+ reticulocytes and mature Ter119+CD71- RBCs (identified as described in S6A Fig) was quantified in the blood of WT and Mif-/- mice at 3 months p.i., which corresponds with a time point when maximal cell numbers, level of hepato-splenomegaly and tissue pathogenicity were reached in both groups of mice (Figs 1C–1F and S3). The concentration of reticulocytes increased in the blood of infected WT mice when compared to non-infected animals, and this increase was less pronounced in Mif-/- mice (Fig 5A). Concomitantly, the reduction in the concentration of mature RBCs was more pronounced in the blood of WT than Mif-/- mice. Within the erythropoietic tissues of infected WT mice, reticulocyte and mature RBC accumulation was not affected in the bone marrow, but was dramatically increased in the spleen (Fig 5B and 5C), suggesting extramedullary erythropoiesis. A comparison of infected WT and Mif-/- mice revealed a detrimental contribution of MIF to mature RBC accumulation in the bone marrow, and to reticulocyte accumulation in the spleen (Fig 5B and 5C).
Next, we assessed the stage at which erythropoiesis could be affected (from nucleated erythroblasts (I) until enucleated erythrocytes (VI)), by gating for Ter119+ RBCs in a CD44 versus FSC-A plot [32] (S6B Fig). In the bone marrow of infected WT mice, a blockade in the two last steps of RBC differentiation resulted in decreased percentage (S6C Fig) and numbers (Fig 5D) of nucleated reticulocytes (stage V) and enucleated reticulocyte/erythrocyte (stage VI) cells. In the spleen, infected WT mice exhibited increased accumulation of the RBC differentiation stage III polychromatic erythroblasts to stage VI mature RBCs (Fig 5E), confirming extramedullary erythropoiesis in this organ. The erythropoiesis efficacy in the bone marrow and the spleen was improved in infected Mif-/- as compared to WT mice. Indeed, Mif-/- mice exhibited a more efficient RBC maturation mainly during the transition from orthochromatic erythroblasts and nucleated reticulocytes (stage IV and V) to the last step of differentiation i.e. enucleation of erythrocytes (stage VI, Fig 5D and 5E). Of note, the gene expression levels of Vcam1 (Vascular cell adhesion molecule-1) and Maea (EMP: erythroblast-macrophage protein), two molecules crucial for erythroblast—macrophage interaction in the terminal stage of erythropoiesis, i.e. during the erythroblast enucleation [33,34], were higher in Mif-/- than in WT mice (Fig 5F), suggesting a negative effect of MIF on the reticulocyte terminal maturation.
Collectively, these data indicated that mice exhibited reticulocytosis during the chronic stage of T. congolense infection, which was likely due to an increased production of RBCs, mainly in the spleen, to overcome the chronic loss of mature RBCs observed in the blood. Moreover, RBCs were impaired in their terminal stages of maturation both in the bone marrow and spleen of infected mice. Finally, MIF contributed to the reticulocytosis and to the impairment of the terminal differentiation of RBCs from the orthochromatic erythroblast to the enucleated reticulocyte/erythrocyte stage.
The percentage of blood Annexin-V+ RBCs increased in T. congolense-infected WT mice, and this increase was lower in Mif-/- than in WT mice (Fig 6A). Phosphatidylserine exposure, which forms the basis of the Annexin-V staining assay, is an “eat-me” signal observed during apoptosis of senescent cells. Thus, we investigated whether an increased RBC elimination through phagocytosis in the liver and the spleen contributed to anemia in T. congolense- infected mice. An assay consisting of i.v. injection of pHrodo-labeled RBCs in infected WT or Mif-/- mice followed by analysis of the appearance of a fluorescent signal in phagocytes from the liver and the spleen that have engulfed labeled RBCs [35], was used. In infected WT mice, PMNs, monocytes and macrophages (gated as in S7A Fig) exhibited RBC phagocytic activity. This activity was more pronounced and MIF-dependent for PMNs in the liver and for monocytes and macrophages in the spleen (Fig 6B and 6C). In parallel, a higher number of phagocytic PMNs, monocytes and macrophages was observed in the liver of infected WT mice when compared to Mif-/- mice (Fig 6B and 6C). The same held true for macrophages in the spleen. Together, these data suggest that the reduced anemia in T. congolense-infected Mif-/- mice is a combined effect of a reduced apoptosis/senescence of RBCs, a reduced number of phagocytic cells and a reduced phagocytic capacity of these cells.
Erythrophagocytosis results in the release of hemoglobin within liver and spleen phagocytes, where the heme is catabolized to iron, carbon monoxide and bilirubin. The latter is then released in the circulation and coupled to albumin to be transported to hepatocytes [36]. Accordingly, histological analyses revealed iron accumulation mainly in the splenic tissue of infected WT mice (Fig 7A and 7B). Furthermore, the observed erythrophagocytosis in T. congolense-infected WT mice coincided with a progressive increase in total bilirubin and decrease in albumin serum levels (Fig 7C and 7D). The decreased anemia and erythrophagocytosis observed in infected Mif-/- mice correlated with lower intracellular iron retention, lower bilirubinemia and higher albuminemia as compared to WT mice (Fig 7A–7D).
Hemodilution can also contribute to anemia [37]. Moreover, the hypoalbuminemia observed in T. congolense-infected mice may also reflect hemodilution. APC-labelled hydroxyethyl starch (HES) is used to monitor hemodilution and is not affected by differences in RBC numbers. Hence, HES was injected i.v. in WT and Mif-/- mice at different time points post T. congolense infection. After 5–10 minutes, the blood was collected and the blood volume and concentration of HES were evaluated. As compared to non-infected mice, the blood HES concentration dropped in infected WT mice while the volume of the blood collected increased, whereby there was about a 3-fold change from 3 months p.i. (Fig 8A and 8B). These data suggest that hemodilution was occurring in T. congolense-infected mice during the later stage of infection. In agreement with observations in T. congolense-infected cattle [38], the packed cell volume (PCV) of the collected blood was reduced in WT mice as compared to non-infected animals (Fig 8C). However, when taking into account the blood volume in the whole animal, it appeared that the total PCV, i.e. the total amount of RBCs, was not affected by the infection while the plasma volume was increased (Fig 8D). When compared to infected WT mice, the blood and plasma volume were lower, and the HES concentration and total PCV were higher in infected Mif-/- mice (Fig 8A–8D). Together, these data indicated that anemia resulted from MIF-dependent hemodilution and not from lower production of RBCs in T. congolense-infected mice.
Hemodilution can also give rise to a reduction in platelet concentration, which in turn contributes to inefficient blood coagulation. As shown in Fig 8E, the concentration of FSClo/SSClo CD41+ platelets (gated as in S7B Fig) declined in T. congolense-infected WT mice and to a lesser extent also in Mif-/- mice. However, when taking into consideration the total blood volume, the number of platelets per animals was not affected by the infection (Fig 8F). Furthermore, the platelet dilution was associated with a drastic increase in clotting time in infected WT mice. Indeed, while the clotting time of non-infected mice was about 2 minutes, >70% of the infected WT mice continued to bleed 15 minutes after the tail cut (time at which the wound was sealed following ethical guideline to avoid otherwise lethal hemorrhage) (Fig 8G). The coagulation time in infected Mif-/- mice was higher than in non-infected mice but remained lower than in infected WT mice, with all mice coagulating in about 10 minutes.
Collectively, these data suggest that MIF contributed to hemodilution that coincided with decreased blood platelet concentration. Combined, these effects can result in inefficient coagulation during T. congolense infection.
To further assess whether MIF triggered hemodilutional anemia, Mif-/- mice at 3 months p.i. were treated with recombinant mouse MIF (rMIF) every second day for 1 week and tested for anemia-associated parameters. rMIF-treated Mif-/- mice exhibited a more severe anemia than untreated Mif-/- mice, reaching a percentage of RBC levels close to that of infected WT mice (Fig 9A). Moreover, the number of erythrocytes and reticulocytes—albeit to a non-significant extent, decreased in the blood of rMIF-treated Mif-/- mice (Fig 9B). This effect coincided with a drop in the terminal stage VI of erythroid development in the spleen, which paralleled the results of WT mice (Fig 9C), thereby strengthening the notion that MIF affected the reticulocyte enucleation process. In addition, rMIF treatment increased the level of annexin-V+ apoptotic RBCs (Fig 9D). Finally, infected rMIF-treated Mif-/- mice exhibited an increased plasma volume recovered from the whole animal, increased splenomegaly, and a decreased concentration but not number of platelets (Fig 9E–9H). Collectively, these data showed that rMIF treatment in T. congolense-infected Mif-/- mice could partially recapitulate the pathogenic features associated with anemia and hemodilution development in infected WT mice. These rMIF treatment data further support the conclusion that MIF exerted negative effects on anemia and hemodilution development during the later stage of T. congolense infection.
We have recently reported that MIF, an upstream regulator of the inflammatory response, contributed to anemia in trypanosusceptible T. b. brucei-infected C57Bl/6 mice [19]. As compared to T. b. brucei-infected mice that are locked in a type I inflammatory immune response, T. congolense-infected C57Bl/6 mice are trypanotolerant due to their ability to restrict the type I driven immune response and to mount a tissue-protective IL-10-mediated immune response in the chronic phase of infection [10,26]. Based on the observation that MIF was produced in the chronic phase of T. congolense infection, we hypothesized that this less virulent model of African trypanosomosis could allow a refined analysis of the MIF-dependent pathogenic mechanisms at play during infection-induced tissue damage and anemia.
As in trypanosusceptible mice, in the absence of MIF, the systemic and tissue-restricted production of pathogenic chemokines (CXCL1, CCL2) and cytokines (IFN-γ, TNF, IL-6, IL-12p70) was impaired during the chronic stage of T. congolense infection. MIF deficiency also limited hepato-splenomegaly and tissue destruction, including anemia. In both T. b. brucei- and T. congolense-infected mice, the reduced anemia in Mif-/- mice coincided with a partial recovery of serum hemoglobin and iron levels. The higher iron bioavailability partially restored erythropoiesis, which was reflected by a decreased concentration of reticulocytes and increased concentration of RBCs in the blood and the spleen of infected Mif-/- mice. The absence of MIF also improved the terminal stage of erythroid development, i.e. the differentiation from nucleated reticulocytes to enucleated RBCs. The latter effect can result from the reduced circulating IL-6 levels in infected Mif-/- mice [30]. The reduced anemia observed in T. b. brucei- and T. congolense-infected Mif-/- mice also could result in part from the better RBC recovery during the early stage of infection (up to 15 days p.i.), which in turn could require a lower erythropoietic demand during the latter stages of infection.
In cattle, anemia development can result from the extravascular destruction of RBCs due to massive erythrophagocytosis by activated macrophages in the spleen and the liver [39]. We found that enhanced erythrophagocytosis indeed occurred in both the spleen and the liver of T. congolense-infected WT mice. This phenomenon was less pronounced in Mif-/- mice despite their better IgG response, and likely due to the reduced number of phagocytic hepatic macrophages, Ly6C+ monocytes and PMNs, and of splenic macrophages. Because of the increased erythrophagocytic activity throughout T. congolense infection, a compensatory demand for increased production of RBCs was evidenced in the spleen but not in the bone marrow. This extramedullary erythropoiesis led to a massive generation and accumulation of reticulocytes and mature RBCs that could account for the splenomegaly observed in infected mice. In the absence of MIF, an increased maturation of reticulocytes to mature RBCs occurred and coincided with reduced splenomegaly and anemia. One week of rMIF treatment in infected Mif-/- mice in turn recapitulated anemia development, including splenomegaly, increased apoptosis/senescence of RBCs and impaired maturation of reticulocytes.
Our accumulated evidence argues for hemodilution, and not erythrophagocytosis, as the main contributor to the chronic anemia developing in T. congolense-infected mice in a MIF-dependent manner. Despite a study showing no increase in the blood volume in T. congolense-infected calves [40], others found a marked hypervolemia in T. congolense-infected sheep and calves [41–45]. In line with the latter observations, the blood and plasma volume was augmented in infected WT mice. However, the total number of RBCs in the blood of infected WT mice, calculated on the basis of the blood volume and the PCV, did not differ from that of non-infected mice. These data suggested that hypervolemic hemodilution developed in infected mice. They also suggested that, in infected WT mice, the increased erythropoiesis and maturation of RBCs in the spleen could compensate for both the impaired maturation of RBCs in the bone marrow and the increased clearance of RBCs through erythrophagocytosis in the liver and the spleen. Remarkably, in anemic T. b. brucei-infected WT mice, the blood and plasma volumes were not affected and the PCV decreased (S8 Fig), in line with observation in T. b. brucei-infected calves [46]. Collectively, these data support the view that a predominantly inflammatory anemia, i.e. anemia of inflammation, develops in a type I immune environment in T. b. brucei-infected mice [10,47], while a predominantly hemodilutional anemia occurs in a type II environment in T. congolense-infected mice. rMIF treatment in infected Mif-/- mice phenocopied the hemodilution development, confirming a MIF-dependent mechanism in infected WT mice. Hemodilution in rMIF-treated Mif-/- mice however did not reach the level observed in WT mice. Although this result could argue for the occurrence of MIF-independent mechanisms of hemodilution, we cannot exclude the possibility that rMIF treatment for a period longer than one week may be necessary.
Hemodilution could also account for the reduced concentration of platelets in the circulation of T. congolense-infected WT mice. This hemodilutional thrombocytopenia concurred with a delayed blot clotting time in infected mice that could lead to lethal haemorrhage when the tail cut for blood sampling was not sealed. Of note, thrombocytopenia was reported to occur in T. congolense-infected cattle and sheep [41,48]. Whether hemodilution of the clotting factors also accounts for the inefficient coagulation in infected animals deserves further investigation.
Despite similarly decreasing tissue pathogenicity, a difference in the pathogenic role of MIF between trypanosusceptible and trypanotolerant mice nevertheless was observed. In T. b. brucei-infected mice, MIF deficiency resulted in increased production of IL-10, the main anti-pathogenic cytokine in experimental African trypanosomosis, but did not affect the survival time of the infected hosts [19]. In T. congolense-infected mice, the absence of MIF had no effect on IL-10 production but resulted in prolonged survival time. Although MIF-independent mechanisms could determine the survival of T. b. brucei-infected mice, we could not exclude that these mice are not sufficiently responsive to IL-10 [49,50]. Alternatively, the virulence of T. b. brucei due to its tissue-invading capacity could be higher than that of T. congolense, which remains strictly in the blood vessels [51]. It is postulated that T. b. brucei- and T. congolense-infected mice die from inflammation-mediated multiple organ failure [12], but the cause of death remains unclear and may differ between the two parasite species. In this respect, a refined mechanism for the death of T. congolense-infected WT mice could be envisaged based on the data reported herein. Our results support the interpretation that in these animals, the continual and month-lasting low-grade inflammatory response drives erythrophagocytosis and that the ensuing catabolism of hemoglobin resulted in iron accumulation mainly in the spleen, followed by the enhanced release of bilirubin in the blood circulation. Importantly, hyperbilirubinemia could favour the externalisation of phosphatidylserine on RBCs observed herein and thus further contribute to erythrophagocytosis or eryptosis during T. congolense infection [36,52]. The hyperbilirubinemia and the hypoalbuminemia—with the latter resulting most likely from the hemodilution and liver damage in infected mice, could contribute to a greater degree of systemic tissue destruction, including not only the hepatic tissue but also the heart and the brain [53,54]. The severe hemodilutional anemia could also reduce cerebral oxygen delivery and further promote cerebral damage [55,56], although T. congolense is not a blood brain barrier penetrating parasite. Combined, the hyperbilirubinemia, hypoalbuminemia, and hemodilution could thus contribute importantly to the increased mortality of T. congolense-infected WT mice. Thrombocytopenia and delayed coagulation, as correlates of hemodilution, could also negatively impact the survival of infected WT mice. MIF deficiency partially alleviated iron accumulation in tissues, hyperbilirubinemia, hypoalbuminemia and defective coagulation, most likely because T. congolense-infected Mif-/- mice exhibited reduced erythrophagocytosis combined with reduced hemodilution and liver injury. Each of these effects may be contributory to the enhanced survival of Mif-/- mice. Of course, our data do not exclude the reduced inflammatory cytokine response and the better parasite-specific antibody response, both vital for the control of African trypanosomosis [13,28,57], as reasons for the enhanced survival of T. congolense-infected Mif-/- mice.
Collectively, our results suggest that during the chronic phase of T. congolense infection, anemia did not result from the impaired production of mature RBCs. MIF-induced splenic extramedullary erythropoiesis could compensate for the impaired differentiation of erythroblasts in the bone marrow and for the enhanced erythrophagocytosis in the liver and the spleen (Fig 10). In contrast, anemia induced by T. congolense mainly occurred through MIF-dependent hemodilution. The heme catabolism ensuing erythrophagocytosis could lead to iron accumulation in tissue and to hyperbilirubinemia. Hypoalbuminemia resulting from hemodilution in infected mice impaired the elimination of toxic circulating molecules, including bilirubin. Hemodilution with thrombocytopenia as a consequence could also account for impaired coagulation in infected mice. Combined, these effects could trigger multiple organ failure and uncontrolled bleeding hereby reducing the survival time of infected mice. Together, this study suggests that interfering with MIF signaling could represent an approach to limit inflammation-associated anemia complications during natural T. congolense trypanosomosis. Given that polymorphisms in the human MIF gene contribute to differences in susceptibility in several inflammatory diseases [58,59], it could be interesting to assess whether differences between trypanosusceptible and trypanotolerant cattle associates with genetically-predetermined differences in MIF expression.
All experiments, maintenance and care of the mice complied with the European Convention for the Protection of Vertebrate Animals used for Experimental and Other Scientific Purposes guidelines (CETS n° 123) and were approved by the Ethical Committee for Animal Experiments (ECAE) at the Vrije Universiteit Brussel (Permit Numbers: 14-220-4 and 14-220-6). Infection of tsetse flies with T. congolense parasites was performed in compliance with the regulations for bio-safety and under approval from the Environmental administration of the Flemish government (Permit number: BM2012-6).
Clonal T. congolense parasites (Tc13) were kindly provided by Dr. Henry Tabel (University of Saskatchewan, Saskatoon) and stored at -80°C. Wild type (WT) C57Bl/6 mice were obtained from Janvier. MIF deficient (Mif-/-) C57Bl/6 mice generated as described in [60] were bred in our animal facility. Female mice (7–8 weeks old) were infected intraperitoneally (i.p.) with 2 x 103 Tc13 trypanosomes. Alternatively, tsetse fly infection was achieved by feeding teneral flies on T. congolense-infected mice as described in [19]. When required, infected mice were treated i.p. with 200 ng of mouse recombinant MIF (rMIF, Abcam, ab191658) every second day for 1 week. Parasite and red blood cell (RBC) numbers in blood were determined via hemocytometer by cold tail-cut (2.5 μl blood in 500 μl RPMI). Anemia was expressed as the percentage of reduction in RBC counts compared to non-infected animals. Packed cell volume (PCV) was measured following collection of anti-coagulated blood in heparinized capillaries and centrifugation at 9500g for 7 min. using a micro-centrifuge (Fisher BioBlock Scientific).
Liver cell isolation was performed as described by Stijlemans et al. [61]. Briefly, livers from CO2 euthanized mice were perfused with 30 ml heparinized saline (10 units/ml; Leo Pharma) containing 0.05% collagenase type II (Clostridium histolyticum; Sigma-Aldrich), excised and rinsed in saline. Following mincing in 10 ml digestive media (0.05% collagenase type II in Hanks' Balanced Salt Solution (HBSS) without calcium or magnesium; Invitrogen) and incubation at 37°C for 30 min., the digested liver was homogenized and filtered (40 μm pore filter). The cell suspension was centrifuged (7 min., 300×g, 4°C) and the pellet treated with erythrocyte-lysis buffer. Following centrifugation (7 min., 300×g, 4°C) the pellet was resuspended in 2–5 ml RPMI/5% FCS medium, cells counted and adjusted at 107 cells/ml for flow-cytometric analysis and cell culturing.
Spleen and bone marrow (tibia and femur) cells were obtained by homogenizing the organs in 10 ml RPMI/5%FCS medium, passing the suspension through a 40 μm pore filter and centrifugation (7 min., 300×g, 4°C). Cells were counted and brought at 107 cells/ml in RPMI/5% FCS medium for RBC analysis via flow cytometry. Remaining cells were pelleted (7 min., 300×g, 4°C) and subsequently treated with erythrocyte-lysis buffer and processed as described for the liver (see above) for analysis of white blood cells (WBCs).
Cells were diluted at 2x106 cells/ml in complete medium (RPMI-1640 medium, 10% FBS, 1% sodium pyruvate (Gibco), 1% non-essential amino acids (Gibco), 1% glutamate, 1% penicillin-streptomycin). Next, 500 μl of cell suspension/well were cultured (36–48 hours, 37°C, 5% CO2) in 48 well plates (Nunc) and the supernatant was tested in ELISA.
To analyze the RBC composition and platelet counts, the blood, spleen and bone marrow cells were analysed omitting RBC lysis. Briefly, total blood (2.5 μl diluted in 500 μl RPMI/5% FCS) and 106 spleen or bone marrow cells (in 100 μl) were incubated (15 min., 4°C) with Fc-gamma blocking antibody (2.4G2, BD Biosciences), and subsequently stained with labelled antibodies (S1 Table) and matching control antibodies. Samples were washed with PBS, measured on FACSCanto II (BD Bioscience) and results were analysed using FlowJo software by excluding CD45+ cells and gating on Ter-119+ or CD41+ cells.
The WBC composition within the bone marrow, spleen and liver cells and the B cell compartment in the spleen (106 cells/100 μl) were analyzed after RBC lysis as described above using labelled antibodies (S1 Table). The results were analysed after selection of CD45+ cells, followed by exclusion of aggregated and death cells (7AAD+, BD Pharmingen). Annexin-V staining was performed as described by the suppliers (TACS Annexin-V FITC Apoptosis Detection kit, R&D Systems).
Blood was collected from CO2 euthanized mice via cardiac puncture, centrifuged (15 minutes, 10.000xg, 25°C), and serum was kept at -20°C. Serum levels of IFN-γ, IL-6, IL-10, IL-12p70, TNF and KC (CXCL1) were determined using the V-PLEX Custom Mouse Cytokine assay (Meso Scale Discovery, Maryland, USA). Serum MIF levels were measured by ELISA as recommended by the suppliers (R&D Systems). Alternatively, culture medium concentrations of MIF, TNF, CCL2 and KC (R&D Systems) as well as IFN-γ, IL-6 and IL-10 (Pharmingen) were determined by ELISA as recommended by the suppliers. Parasite-specific IgG responses were determined using soluble lysate freshly prepared from DEAE-purified T. congolense parasites recovered from WT mice at the peak of infection (around day 7). Lysate was coated overnight at 10 μg/ml PBS in 96-well Maxisorp plates (NUNC). Plates were washed (0.1% Tween 20 in PBS) and blocked (1% BSA in PBS) for 1 hour. Next, plates were washed and the sera (100 μl) serially diluted starting from 1/100 in blocking buffer were added. The ELISA was subsequently performed as described by the suppliers (SBA Clonotyping system-HRP kit (SouthernBiotech, USA)). As negative controls, blood samples incubated on lysate-free plates were used. The OD450nm recorded on lysate-free plates was subtracted from the OD450nm recorded on lysate-coated plates.
One μg of total RNA prepared from 107 cells (RNeasy plus mini kit, Qiagen) was reverse-transcribed using oligo(dT) and Superscript II Reverse Transcription following the manufacturer's recommendations (Roche Molecular Systems). RT-QPCR was performed in an iCycler iQ, with iQ SYBR Green Supermix (Bio-Rad). Primer sequences were Vcam-1-F: 5’-CTCTCCCAGGAATACAACGA-3’, Vcam-1-R: 5’-CACGTCAGAACAACCGAATC-3’ and Maea-F: 5’-GAGTGGTCTCCTCTCAACAG-3’, Maea-R: 5’-AGCTACCATCTGTC TGGATG-3’. PCR cycles consisted of 1-minute denaturation at 94°C, 45-second annealing at 55°C, and 1-minute extension at 72°C. Fold change in gene expression was expressed as compared to non-infected animals after normalization against the Ct value of the ribosomal S12 (Mrps12) protein as household gene.
The pHrodo-labeling of red blood cells (RBCs) was described in [35]. 109 pHrodo-labelled RBCs isolated from non-infected WT mice were injected i.v. in WT or Mif-/- mice. 18 h later, mice were sacrificed, liver and spleen CD11b+Ly6CintLy6G+ PMNs, CD11b+Ly6ChighLy6G- monocytes and CD11b+Ly6C-Ly6G-F4/80+ macrophages were tested for delta median fluorescent intensity (MFI) of the intracellular pHrodo signal determined by subtracting the PE signal of cells from mice receiving unlabeled RBCs from the PE signal of cells from mice receiving pHrodo-labeled RBCs.
Serum AST and ALT levels were determined as described by the suppliers (Boehringer Mannheim Diagnostics).
100 μl of APC-labelled hydroxyethyl starch (APC-HES (130/0.4) at 100 μg/ml in 0.9% NaCl) was injected i.v. (as described in [62]). After 5–10 min., blood was collected and the APC signal measured via cytofluorimeter (OD660nm, Ultra Microplate reader, ELx808, Bio-Tek instruments.inc). A standard curve consisting of a serial dilution of APC-HES (starting from 200 μg/ml) diluted in blood from non-infected mice was used to calculate the concentration of APC-HES in collected blood. The OD660nm from blood of non-infected mice was subtracted from all samples.
For hemoglobin quantification, 2 μl of blood collected via cold tail cut was diluted in 200 μl distilled water in a 96 well round bottom plate (Falcon). After incubation for 30 min. at 37°C and centrifugation (600xg, 10 min.), the supernatant was collected and the OD550nm measured. The hemoglobin concentration was calculated using a standard (Sigma) curve. Total iron (IRON FZ kit, Chema Diagnostics), bilirubin and albumin (Chema Diagnostica, Italy and Sigma Aldrich, respectively) were measured as recommended by the suppliers.
Bleeding times of mice were obtained by using the tail-cut model [63]. Briefly, anesthetized animals were transected at the 5-mm mark from the tip of the tail and incubated in warm saline (37°C). The time for cessation of bleeding was recorded. The experiment was terminated after 15 min. to avoid lethality, whereby the tail was cauterized and the bleeding time was taken as 15 min.
Total spleens and the largest lobe of the liver were embedded in Tissue-Tek O.C.T. compound (Sakura Belgium B.V.B.A.) and kept at -80°C. Next, 5 μm cryosections were cut using a Leica microtome, fixed in cold acetone for 15 min. and washed shortly in distilled water (2–3 changes). The sections were stained for 20 min. in equal volumes of warm HCl 4% and K4[Fe(CN)6]·3H2O 4% at 45°C. After washing shortly 3x with distilled water, the slides were counterstained with NFR (Nuclear Fast Red, Sigma-Aldrich) for 5 min. After rinsing 3x with distilled water, the slides were dehydrated (short in EtOH 96%, short in EtOH 100% (2x), 3 min. in xylol (2x)) and mounted with DPX mounting medium (Sigma-Aldrich). Images were obtained using an OLYMPUS BX41 fluorescent microscope. The CellSens Dimension 1.9 software was used for quantification. For each sample, an average quantification of 5 representative images was determined. The results are expressed as % stained area within the region of interest.
The GraphPad Prism software was used for statistical analyses (Two-way ANOVA or student t-test). Values are expressed as mean ± SEM. Values of p≤ 0.05 are considered significant.
|
10.1371/journal.pbio.1000080 | Impaired Antibody Response Causes Persistence of Prototypic T Cell–Contained Virus | CD8 T cells are recognized key players in control of persistent virus infections, but increasing evidence suggests that assistance from other immune mediators is also needed. Here, we investigated whether specific antibody responses contribute to control of lymphocytic choriomeningitis virus (LCMV), a prototypic mouse model of systemic persistent infection. Mice expressing transgenic B cell receptors of LCMV-unrelated specificity, and mice unable to produce soluble immunoglobulin M (IgM) exhibited protracted viremia or failed to resolve LCMV. Virus control depended on immunoglobulin class switch, but neither on complement cascades nor on Fc receptor γ chain or Fc γ receptor IIB. Cessation of viremia concurred with the emergence of viral envelope-specific antibodies, rather than with neutralizing serum activity, and even early nonneutralizing IgM impeded viral persistence. This important role for virus-specific antibodies may be similarly underappreciated in other primarily T cell–controlled infections such as HIV and hepatitis C virus, and we suggest this contribution of antibodies be given consideration in future strategies for vaccination and immunotherapy.
| Persistent viruses such as hepatitis C virus (HCV) or HIV can defeat the body's defense system and cause devastating epidemics worldwide. Recent attempts at vaccinating against HIV have relied on the induction of specific antiviral killer T lymphocytes but have failed to confer protection on the host. Better knowledge about how a successful defense should operate is therefore essential for developing and refining new vaccines. Here, we have used a prototypic mouse model to investigate basic defense mechanisms required to eliminate persisting viruses. Experiments in several genetically engineered mouse models show that contrary to common belief, not only antiviral killer T cells, but also antibodies (produced by B cells), are needed to prevent a virus from persisting in its host. These findings suggest that induction of antibodies, along with antiviral killer T lymphocytes, should be envisaged when devising new strategies for vaccinating against HIV or HCV.
| Infections associated with persistent viremia include human immunodeficiency virus (HIV) and the hepatitis B and C viruses (HBV, HCV), which affect more than 500 million people worldwide. However, available options to prevent and treat particularly HIV and HCV are unsatisfactory. To refine existing strategies aimed at combating these devastating epidemics, and to help direct future efforts, a better understanding of the immune effector pathways preventing viral persistence is of particular importance.
For almost a century, lymphocytic choriomeningitis virus (LCMV) infection of mice has served as a primary model to study basic mechanisms of the virus–host relationship in persistent infection [1]. It has led to the discovery of several essential concepts [2], including MHC restriction of T cells, viral mutational escape from CD8 cytotoxic T cells (CTL), CTL dysfunction in persistent infection and MHC linkage of virus control. LCMV neutralizing antibody (nAb) responses typically appear late and remain relatively weak [1]. Accordingly, the key role of CTL in controlling and resolving systemic persistent infections has initially been described for LCMV [3–5] with subsequent extension of the concept to important human pathogens such as HIV and HCV. Declining viremia in HIV coincides with the appearance of antiviral CD8 T cells [6,7], and the concept of CTL-mediated HIV control was further strengthened by the association of “protective” HLA molecules with long-term nonprogression in many so-called “elite controllers” [8]. In addition, experimental depletion of CD8 T cells in simian immunodeficiency virus (SIV)-infected macaques also underlined the importance of CTLs in the control of acute, as well as long-term infection [9–11]. Analogous observations were made in HBV- and HCV-infected monkeys [12,13].
Apart from the virtually undisputed contribution of CTLs, evidence has accumulated to suggest that other mechanisms of immune defense are also needed to contain or resolve systemic persistent virus infection. For instance, “protective” HLA alleles are also found in up to one third of individuals with poor or undetectable immune control of HIV infection [14,15], suggesting that even potent CD8 T cell responses are insufficient for HIV control. Conversely, many “elite controllers” lack any of the known “protective” alleles [15]. Moreover, the recent failure of the CD8 T cell–based Merck “STEP” vaccine trial in human HIV infection has alerted the community and has sparked renewed interest in complementary mechanisms that may aid immune defenses against persistent viral disease [16].
Antibodies are among the obvious candidates to complement CTL-mediated control. However, their contribution to the resolution of primary virus infections in general, and persisting ones in particular, has remained controversial. Rapid mutational escape of persisting viruses from antibody neutralization represents a major obstacle to efficient antibody-mediated control [17–21]. Moreover, observations that patients with Bruton's agammaglobulinemia can control acute viral diseases [22] helped create a generally held notion that, unlike what applies for protection against reinfection, primary viral infections were predominantly controlled by cell-mediated immunity [22]. Experiments in mice, monkeys, and man had shown that passive administration of potent nAbs or transgenic expression of a virus-neutralizing B cell receptor (BCR) can prevent infection [23,24], augment virus control during infection [25–27], or prevent the establishment of persistence [28,29]. Still, these experimental observations did not challenge the above dogma since the experimental conditions chosen did not mimic the kinetics and magnitude of the host's spontaneous nAb response (delayed and weak).
Similarly, it seemed unlikely that antibodies could influence LCMV control and persistence, until B cell–deficient mice were found to control the infection only transiently, or not at all. B cell–deficient mice showed vanishing CD8 T cell function and viral recrudescence [30,31], but the conclusions became doubtful when the mice were shown to have a distorted splenic microarchitecture and intrinsically defective CD4 T cell responses [32–34]. As CD4 T cells are essential to the maintenance of effective antiviral CD8 T cell responses [35], the shortcomings in viral resistance were concluded to result from defective T help, rather than from the lack of antibody [34].
Given the outlined uncertainties, combined with the importance of such fundamental knowledge in order to refine preventive and therapeutic strategies in humans, we have readdressed the role of specific antibody responses to the control and resolution of persistent infection. We used the LCMV model to establish viral infection in genetically engineered mice that support the development of B cells, but do so only with restricted diversity and predominantly LCMV-unrelated specificity. In addition, we infected B cell–sufficient mouse models, unable to mount either serum immunoglobulin M (IgM) or immunoglobulin G (IgG) responses. Our studies reveal that virus-specific antibodies, including early adaptive IgM responses, play an essential role in reducing viral loads and ultimately determine viral clearance or persistence.
Using the murine model of LCMV infection, we aimed here at investigating the contribution of specific antibody to prevent persistent infection. To overcome the limitations intrinsic to B cell–deficient mouse models (i.e., distorted splenic microarchitecture with resulting defects in CD4 T cell responses), we first exploited two genetically engineered mouse models with a severely narrowed, predefined BCR repertoire of LCMV-unrelated specificity. T11μMT [36] carry an immunoglobulin (Ig) heavy chain transgene in an IgM heavy chain–deficient background, whereas VI10YEN [37] combine an Ig light chain transgene with a knockin at the endogenous Ig heavy chain locus. Both constructs render the respective B cells specific for vesicular stomatitis virus (VSV) that is antigenically unrelated to LCMV (for a more detailed description of these strains, including their residual ability of generating antibody repertoire diversity, see Text S1). Unlike B cell–deficient mice, these animals exhibited a normal splenic microarchitecture in immunohistochemistry, and mounted unimpaired CD4+ T cell responses against LCMV, as determined by intracellular staining of interferon γ (IFNγ) upon peptide stimulation (Figure S1 and Text S1). We infected B cell–deficient μMT mice [38] (targeted deletion of the IgM transmembrane domain), BCR-restricted T11μMT and VI10YEN mice, and C57BL/6 control mice with 106 plaque-forming units (PFU) of LCMV intravenously (i.v.) (Figure 1). Unlike C57BL/6 mice that resolved viremia within 12 d, T11μMT mice exhibiting the lowest degree of BCR diversity failed to contain the infection and—like B cell–deficient μMT mice—remained viremic throughout the observation period of 96 d (Figure 1A). Similar, albeit less-pronounced, effects were seen in VI10YEN mice displaying a more diverse BCR repertoire than T11μMT mice. Seven of ten VI10YEN mice tested in three independent experiments exhibited protracted viremia as compared to C57BL/6 wild-type mice (Figure 1A and unpublished data). Even more pronounced was the impact of BCR diversity on the control of the more invasive Clone 13 strain of LCMV (Figure 1B). Only C57BL/6 mice succeeded in resolving viremia, whereas BCR-restricted VI10YEN and T11μMT mice, and B cell–deficient JHT [39] mice (targeted deletion of the immunoglobulin JH locus; JHT and μMT mice were used likewise in this study) remained viremic throughout the observation period of 123 d. Thus, BCR diversity was essential for efficient resolution of LCMV infection. Further support for this notion came from experiments in “quasimonoclonal” (QM) mice [40] with a predefined nitrophenyl-specific B cell repertoire owing to knockin of a rearranged immunoglobulin heavy chain gene in combination with an immunoglobulin light chain transgene (Figure S2). Interestingly also, the requirements for BCR diversity became apparently more stringent as the infection was prone to persistence. That is, VI10YEN mice were able to clear LCMV strain WE (LCMV-WE), albeit with some delay, but they failed at resolving chronic infection with LCMV strain Clone 13. The above patterns of virus control or persistence correlated only to a limited extent with the ability of the respective mouse strains to mount a late virus-neutralizing antibody response (Figure 1C and 1D). In LCMV Clone 13 infection, the appearance of neutralizing serum activity around day 45 after infection coincided with viral clearance. In contrast, a clear rise in LCMV-WE-nAb occurred only between 50 and 74 d after infection, i.e., more than 1 mo after viral clearance from the blood. In C57BL/6 mice, this response was consistently measured although the titers varied considerably between individual animals. With further delay and barely above the detection limit of our assays, nAbs were also measured in some VI10YEN mice (Figure 1C, not statistically significant), providing only partial correlation with this mouse strain's ability to control LCMV-WE infection. In contrast, nAbs always remained below detection levels in viremic T11μMT mice. The lack of temporal association, at least in LCMV-WE infection, between the appearance of nAb and clearance, prompted us to study nonneutralizing antibody (non-nAb) responses. The glycoprotein (GP) is the only surface determinant on LCMV particles. It is synthesized as a precursor protein and is posttranslationally cleaved into GP1 and GP2 subunits that remain noncovalently associated [41]. GP1 makes up an outer globular domain, whereas GP2 forms a membrane-anchored stalk [41]. Hence, GP1 is accessible on the infectious virion surface, rendering this antibody specificity of particular interest. Here, we exploited recently developed ELISA techniques [42] for measuring LCMV-WE GP1-specific antibodies. By day 12 after infection, LCMV-WE evoked a GP1-specific IgG response in C57BL/6 mice and at lower titers also in V10YEN mice, but not in T11μMT mice, correlating with virus control (Figure 1E, B cell–deficient μMT mice shown as negative controls). Thus, the timing of the GP1-binding antibody response as well as the differential magnitude in C57BL/6, VI10YEN and T11μMT mice matched best the pattern of virus clearance.
Next, we assessed the individual contribution of IgM and IgG responses to virus control. All monoclonal LCMV nAbs characterized today are of an IgG isotype, and so is the late nAb response observed in the course of natural infection [27]. Hence, any potential role of antibodies in resolution of LCMV infection had previously been accredited to IgG. To test for the role of class switch-dependent isotypes including IgG, we used gene-targeted mice lacking activation-induced cytidine deaminase (AID−/−) [43]. AID−/− mice are unable to undergo class-switch recombination and somatic hypermutation, and in our experiments, could not resolve LCMV-WE infection during the observation period of 96 d (Figure 2A). As expected, AID−/− mice displayed a complete absence of nAbs and GP1-specific serum IgG (Figure 2C and 2E), suggesting that immunoglobulin class-switch recombination and IgG production together with somatic hypermutation are essential steps in the resolution of LCMV infection.
To assess a potential role of IgM antibodies we exploited the sIgM−/− mouse model [44]. sIgM−/− B cells express IgM as their surface receptor and secrete IgG upon class-switch recombination but are unable to secrete the early IgM isotype. Experiments were carried out to confirm that according to expectations and unlike B cell–deficient μMT mice, B cell–competent sIgM−/− mice display a normal lymphoid microarchitecture and mount unimpaired CD4+ T cell responses (Figure S3). Surprisingly, however, LCMV-WE infection resulted in substantially prolonged viremia in sIgM−/− mice as compared to C57BL/6 control mice (Figure 2A), suggesting that contrary to expectations, an antibody response of IgM isotype contributed to virus control. More strikingly even, nine of ten sIgM−/− mice failed to resolve LCMV Clone 13 infection for a period of at least 100 d, whereas all nine C57BL/6 mice had cleared viremia within 42 d after infection (Figure 2B and unpublished data).
The analysis of nAb responses (measuring both IgM and IgG, Figure 2C and 2D), confirmed that the kinetics and magnitude of the nAb response were indistinguishable in sIgM−/− and C57BL/6 controls, and therefore, likely were of IgG isotype as previously reported. As expected, also LCMV-WE-GP1–binding IgG responses showed normal kinetics in sIgM−/− mice. Somewhat higher GP1-specific IgG peak titers in sIgM−/− mice as compared to C57BL/6 control mice were likely the result of prolonged viremia with an increased antigen burden (Figure 2E). In support of this notion, differences in antibody titers became particularly apparent between day 12 and 20 when C57BL/6, but not sIgM−/−, mice had cleared the infection. The above experiments had suggested that differences in virus loads of sIgM−/− and C57BL/6 mice were manifest as early as 1 wk after infection (p < 0.01 for LCMV Clone 13, Figure 2B). Additional experiments corroborated this difference in early virus loads also for LCMV-WE infection (Figure 3A, p < 0.01). As a likely mediator of this difference, ELISA assays detected GP1-specific IgM responses in day 8 LCMV-WE–infected C57BL/6, but not sIgM−/−, mice (Figure 3B), antibodies that were absent from naive C57BL/6 mouse serum (Figure 3B). Importantly also, the GP1-specific IgM responses measured here were confirmed to be entirely antigen-specific since total serum IgM (Figure 3C), unlike serum IgG [45], remained largely unaltered after LCMV infection. A time-course analysis revealed that GP1-specific IgM was highest on day 4 and 7 after infection, followed by a continuous decline of this isotype concomitant with class switch and appearance of GP1-specific IgG (Figure 3D). These assays were, however, performed with unseparated serum, and competition between IgG and IgM in ELISA may have resulted in an underestimation of IgM levels at later time points.
sIgM−/− mice not only lack adaptive IgM responses but also natural IgM, which contributes to control of other viral infections [46]. For dissecting the role of preexisting natural antibodies, we reconstituted sIgM−/− mice with naive C57BL/6 mouse serum (Figure 3E). Despite reaching total serum IgM levels at least equivalent to normal C57BL/6 mice, reconstitution of natural IgM in sIgM−/− mice failed to restore virus control (Figure 3F). Taken together, these data demonstrated that adaptive IgM as well as IgG responses both played essential roles in the efficient resolution of LCMV infection. Interestingly, unaltered nAb kinetics in sIgM−/− and C57BL/6 control mice suggested that antiviral IgM mediated its effects by mechanisms other than classical virus neutralization.
Next, we studied whether antibody therapy could restore virus control in BCR-restricted LCMV noncontroller mice. For this purpose, we infected T11μMT mice with LCMV-WE and treated them on day 4 and day 7 with either normal serum (negative control) or with normal serum reconstituted with GP1-specific monoclonal antibody (Figure 4A). T11μMT mice treated with GP1-specific antibody eliminated LCMV as efficiently as did C57BL/6 wild-type mice, whereas control-treated T11μMT mice remained viremic, as expected (compare Figure 1A). The same antibody treatment that was successful in T11μMT mice failed, however, to exert a detectable effect on virus loads when administered to TCRβ−/−δ−/− mice [47] lacking T cells owing to homozygous deletion of the T cell receptor β and δ chain loci (Figure 4B). These data were compatible with the interpretation that T cells of T11μMT mice could control LCMV infection if appropriately aided by specific antibodies, whereas neither T cells nor antibody therapy was sufficient to control LCMV on its own.
Exhaustion of CD8 T cell responses as a result of continued antigen exposure is a common observation in persistent viral infection [48,49]. Hence, we investigated whether antibody therapy could prevent CD8 T cell exhaustion in T11μMT mice. The initial LCMV-specific CD8 T cell response of T11μMT mice not only was of normal frequency and was functional in terms of IFNγ secretion (Figure S1D) but also displayed an unimpaired capacity for killing antigenic cells in vivo, irrespective of antibody treatment (Figure 4C). Of note, virus loads were still similar in all groups when these tests were performed on day 7 (Figure 4D). On the contrary, defective cytolytic activity was observed in control-treated T11μMT mice on day 35 during the chronic phase of infection. Prevention of viral persistence by antibody therapy (Figure 4F) restored in vivo cytotoxicity of T11μMT mice to normal levels on day 35 (Figure 4E). This lent further support to the interpretation that the CD8 T cell response of T11μMT mice was intrinsically normal, and that its decline during chronic infection was merely the result of viral persistence rather than the cause thereof. Albeit less likely, a subtle intrinsic CD8 T cell deficiency of T11μMT mice cannot, however, be formally excluded. Irrespective thereof, antibody therapy may help preserve the antiviral CD8 T cell response.
To evaluate the role of the classical and alternative complement cascades as major effector pathways of antibody-mediated immunity, we studied clearance of LCMV-WE in mice lacking complement components C3 and C4 (C3−/−C4−/− mice; see Materials and Methods). C3−/−C4−/− mice resolved viremia as efficiently as wild-type control mice (Figure 5A), whereas B cell–deficient JHT control mice remained viremic throughout. An analogous experiment was carried out in mice lacking Fc γ receptors I, III, and IV owing to deletion of the common γ chain (FcRγ−/− [50]). FcRγ−/− mice cleared LCMV-WE infection as efficiently as did C57BL/6 wild-type mice, whereas B cell–deficient JHT and T11μMT mice both showed unchecked viremia throughout the observation period (Figure 5B). Unlike Fc γ receptors I, III, and IV, Fc γ receptor IIB (FcγRIIB) expression does not depend on the common γ chain. To study the contribution of this receptor, we infected FcγRIIB -deficient mice (FcγRIIB−/−; see Materials and Methods) with LCMV, but found unimpaired virus control (Figure 5C). Taken together, these data excluded an essential individual contribution of classical and alternative complement cascades, of Fc γ receptors I, III, and IV, and of FcγRIIB, respectively, in mediating protective antibody effects in the natural course of LCMV infection.
The present data show that virus-specific antibody responses, including early IgM, play an unexpected key role in preventing viral chronicity in the CTL-controlled murine model of LCMV infection. These observations are compatible with the rapid escape from antibody recognition seen in other primarily CTL-controlled infections, including HIV and HCV [19–21], and indicate that specific antibody responses represent a level of antiviral pressure that tends to be underappreciated.
The observed antiviral effects can only be partially accredited to antibody-mediated virus neutralization. Albeit clearance of LCMV Clone 13 in C57BL/6 mice did coincide with the appearance of nAbs (compare Figure 2B and 2D), IgM effects on LCMV Clone 13 and LCMV-WE titers were evident already on day 7/8 after infection (Figures 2A, 2B, and 3A) at a time when nAbs were undetectable even if using virtually undiluted serum for the assays (unpublished data). Similarly, LCMV-WE was cleared weeks before nAbs could be detected (compare Figure 2A and 2C). Obviously, “absence of proof” for early nAb does not equate “proof of absence,” and we recognize that “nAb consumption” during viremia or subsequent phases of protracted clearance from tissues [51] would provide an explanation for our inability to detect nAbs. However, we favor the idea that the delay in LCMV nAb detection, relative to the antiviral effects observed, rather results from the need for time-intensive affinity maturation [42].
The protective capacity of nAbs is classically explained by “virion occupancy,” i.e., sterical hindrance interfering with cell-surface receptor binding [52]. Non-nAbs, on the other hand, may mediate protection via a number of mechanisms, including: (1) virion occupancy by complement C1q binding to a virion-bound antibody [53], (2) complement cascade activation, leading to further virion occupancy through covalent opsonization, (3) complement-mediated virion lysis, (4) Fc-receptor–mediated virion phagocytosis and destruction, (5) Fc-receptor–mediated stimulation of the innate immune system, (6) immune complex formation and resulting modification of tissue distribution and cellular tropism, (7) antibody-dependent cellular cytotoxicity (ADCC), via antibody binding to viral surface proteins on infected cells, (8) impaired virus production, through antibody-mediated cross-linking of cell surface–expressed viral envelope protein [54], or (9) destruction of target protein or host cells, through antibody-mediated reactive oxygen catalysis [55]. Although a full assessment of the individual contribution of each of these potential pathways lies outside the scope and intention of the present study, we do present data ruling out a major individual contribution for covalent complement opsonization and lysis, mediated through C3 and/or C4 activation, as well as for FcRγ- and FcγRIIB-facilitated mechanisms [50,56] (Figure 5). It remains possible that another mechanism not yet experimentally addressed here may account for most of the antibody effects observed, e.g., Fc α/μ receptor–mediated clearance [57] could explain the observed IgM effects (Figures 2A, 2B, and 3A). However, substantial redundancy in these multiple mechanisms may render it difficult to work out the contribution of individual mechanisms including the ones we have tested here, i.e., in the absence of a specific pathway, compensation by the remaining ones may suffice for virus clearance.
The present findings are of considerable importance for our understanding of virus–host relationship in persistent infection and for refining preventive and therapeutic strategies: The success of antibody therapy in T11μMT mice, but not in TCRβ−/−δ−/− animals (Figure 4A and 4B), suggests a synergistic effect of cellular and humoral immune defense, at least for LCMV. Antibody therapy can apparently help preserve T cell function, and hence early administration may be most promising. Albeit our experimental therapy was administered during a phase of infection in which IgM predominates (compare Figure 3D), IgG was efficient. This may be of practical importance since both vaccination and immunotherapy typically rely on IgG rather than on IgM.
Owing to structural reasons, potent nAb responses against persisting viruses are generally difficult to elicit through vaccination [20,58], but non-nAbs represent an attainable goal. The present data from LCMV infection in mice strongly suggest that non-nAbs, alongside antiviral CTL responses and nAbs, can determine clearance or persistence. We suggest that non-nAbs operate by blunting the infection and thereby strengthening the efficacy of other immune mediators such as CTLs and nAbs, but also NK cells [59,60]. In the context of the cited literature, our data support the idea that antibodies should be considered anew in vaccination strategies aimed at combating persistent viral disease, and that aside from nAbs as a vaccine goal, non-nAbs also should be induced and assessed.
It has recently been shown that non-nAbs specific for LCMV GP-1 can mediate protective effects when expressed in a transgenic context [27]. In nontransgenic wild-type mice, we now show that virus-specific antibody responses, including GP-1–binding IgM, are not only generated rapidly (see also Figure 2G), but also exert significant antiviral pressure (compare Figure 2B and 2F; p < 0.01) in the days before nAbs become detectable. Although we focused in our assays on GP-1 binding antibodies, it remains entirely possible that additional non-nAbs of alternative specificities may also contribute to the observed protective effects. Defining characteristics and specificities of “protective” and “nonprotective” non-nAbs may therefore represent an important next step in the direction of exploiting the protective capacity of non-nAbs for vaccination and immunotherapy.
Failure of the HIV AIDSVax trial eliciting mostly non-nAbs in the absence of cell-mediated immunity has somewhat dampened the hope that non-nAbs could help containing persistent infections [61,62]. Albeit non-nAbs are apparently unable to protect on their own, studies have correlated ADCC with HIV nonprogression, suggesting that non-nAbs may indeed contribute to long-term control of HIV [63]. However, much remains unclear about the overall importance of non-nAbs, and antibodies in general, to the natural course of HIV infection [64]. Moreover, the available data emphasize that the mechanisms of antibody-mediated protection do not always follow the traditional way of thinking. For instance, even a broadly HIV-neutralizing monoclonal antibody was shown to protect primarily via Fc-receptor–dependent mechanisms [65]. Of note in this context, the HIV envelope displays defective glycoproteins in great abundance [66]. Albeit unable to mediate cell entry, such defective glycoproteins are highly immunogenic and may represent efficient targets for non-nAbs. Of further importance here, non-nAbs have a relatively broad spectrum of activity against both autologous and heterologous HIV strains [67].
Taken together, the results from this study show that CD8 T cells, even if firmly established as the predominant mechanism of antiviral immune defense, need support from specific antibodies to prevail and prevent viral persistence. Given the relative ease of induction of non-nAbs (relative to nAbs), combined with the observed protective effects, our findings may provide new impetus for inclusion of antibody targets in vaccines against persistent viral diseases.
C57BL/6 wild-type mice, μMT−/− [38], JHT−/− [39], T11μMT [36], VI10YEN [37], QM [40], AID−/− [43], sIgM−/− [44], C3−/−C4−/− double-deficient mice (a crossbreed of C3−/− [68] and C4−/− [69] mice), TCRβ−/−δ−/− [47], and FcRγ−/− [50] were bred at the Institute of Laboratory Animal Science, University of Zurich, and were housed under specific pathogen-free (SPF) conditions throughout. FcγRIIB−/− mice on a pure C57BL/6 background, in which exons 4 and 5, encoding the ligand-binding EC2 and transmembrane (TM) region, have been deleted by gene targeting in Bruce4 ES cells (C57BL/6 background), were generated in the laboratory of Sjef Verbeek. Absence of functional FcγRIIB was confirmed both in functional in vivo and in vitro assays and at the protein level, as will be described elsewhere in more detail. Experiments with FcγRIIB−/− mice and controls were performed in a conventional mouse facility. Animal experiments were carried out at the University of Geneva and the University of Zurich with authorization by the respective cantonal authorities and in accordance with the Swiss law for animal protection.
LCMV-WE was originally obtained from F. Lehmann-Grube (Heinrich-Pette Institut, Hamburg, Germany) and was propagated on L929 cells. LCMV Clone 13 was obtained originally from R. Ahmed (Emory University, Atlanta, Georgia, United States) and was grown on BHK-21 cells. Infections were performed at a standard dose of 106 PFU by the intravenous route. For therapy of T11μMT and TCRβ−/−δ−/− mice, GP1-specific monoclonal antibody KL25 [70] was administered intraperitoneally on day 4 (100 μg) and on day 7 (1 mg), reconstituted in 400 μl of normal (nonimmunized and uninfected) C57BL/6 serum. Control animals were given 400 μl of normal serum.
LCMV virus stocks and blood samples were titrated by standard immunofocus assays on MC57G cells [71]. nAbs against LCMV-WE and LCMV Clone13 were measured in an immunofocus reduction assay using the respective homologous virus as described [58].
GP1-specific IgM and IgG responses were measured by ELISA using a GP1-Fc fusion construct produced in an eukaryotic system as described [42]. In the GP1-Fc construct, amino acids 1–265 (i.e., the GP1 domain [42]) of the LCMV-WE glycoprotein gene are fused to human Fc. As sole modification to the published method, anti-mouse IgM monoclonal antibody coupled to HRP (Sigma) was used instead of anti-mouse IgG when detecting GP1-specific IgM. Total serum IgM titers were measured in ELISA as described previously [45]. Titers displayed represent the serum dilution yielding twice background optical density values.
Single-cell suspensions of splenocytes were used for intracellular cytokine assays as described [72]. Restimulation of virus-specific cells was performed for 5–6 h in the presence of the following synthetic peptides at 10−6 M concentration: KAVYNFATC (GP33, CD8+ T cells), GPDIYKGVYQFKSVEFD (GP64, CD4+ T cells), and SGEGWPYIACRTSVVGRAWE (NP309, CD4+ T cells).
Cytotoxic activity of CD8+ T cells was measured in an in vivo CTL assay as previously described [73]. In brief, syngeneic C57BL/6 splenocytes were labeled with the fluorescent dye carboxyfluorescein diacetate succinimidyl ester (CFSE) at two different concentrations (CFSEhigh or CFSElow). In addition, CFSEhigh cells were pulsed with GP33 peptide at 10−6 M concentration for recognition by antiviral CTLs. 3 × 107 cells of each population were cotransferred into virus-infected recipient mice and into naive C57BL/6 mice (control). Five hours later, the percentage of CFSEhigh and CFSElow donor cells in peripheral blood mononuclear cells was determined by flow cytometry. Specific killing was calculated as: 100 − ([(% CFSEhigh in test animal / % CFSElow in test animal) / (% CFSEhigh in naive / % CFSElow in naive)] × 100).
Histological analyses were performed on snap-frozen tissue. Sections were stained with rat monoclonal antibodies against murine B220 (Pharmingen), F4/80, MOMA1, and ERTR9 (all from BMA Biomedicals). Bound antibody was detected using a goat anti-rat antibody (Caltag Laboratories) and an alkaline phosphatase–coupled donkey anti-goat antibody (Jackson ImmunoResearch Laboratories) with naphthol AS-BI (6-bromo-2-hydroxy-3-naphtholic acid 2-methoxy anilide) phosphate and new fuchsin as a substrate. The sections were counterstained with hemalum. For tissues of VI10YEN mice carrying a light-chain transgene with rat constant domains, reaction of anti-rat monoclonal antibody with the transgenic light chain was prevented by using an alkaline phosphatase conjugated Fc γ fragment–specific goat anti-rat IgG antibody as a secondary antibody (Jackson ImmunoResearch Laboratories).
One-way analysis of variance (ANOVA) with the Least Significant Difference (LSD) post test was used for the comparison of individual values from multiple groups. Two-way ANOVA was performed to compare antibody responses over time. ANOVA was performed with SPSS version 13.0. Differences in individual values between two groups were analyzed by t-tests (unpaired, two-tailed), and virus clearance kinetics were compared in log-rank tests using GraphPad Prism software vs. 4.0b. Viral titers were log-transformed for statistical analysis, and viral clearance kinetics were compared in a Kaplan-Meier format. p-Values < 0.05 were considered statistically significant; p-values < 0.01 were considered highly significant.
|
10.1371/journal.pcbi.1006794 | A complete statistical model for calibration of RNA-seq counts using external spike-ins and maximum likelihood theory | A fundamental assumption, common to the vast majority of high-throughput transcriptome analyses, is that the expression of most genes is unchanged among samples and that total cellular RNA remains constant. As the number of analyzed experimental systems increases however, different independent studies demonstrate that this assumption is often violated. We present a calibration method using RNA spike-ins that allows for the measurement of absolute cellular abundance of RNA molecules. We apply the method to pooled RNA from cell populations of known sizes. For each transcript, we compute a nominal abundance that can be converted to absolute by dividing by a scale factor determined in separate experiments: the yield coefficient of the transcript relative to that of a reference spike-in measured with the same protocol. The method is derived by maximum likelihood theory in the context of a complete statistical model for sequencing counts contributed by cellular RNA and spike-ins. The counts are based on a sample from a fixed number of cells to which a fixed population of spike-in molecules has been added. We illustrate and evaluate the method with applications to two global expression data sets, one from the model eukaryote Saccharomyces cerevisiae, proliferating at different growth rates, and differentiating cardiopharyngeal cell lineages in the chordate Ciona robusta. We tested the method in a technical replicate dilution study, and in a k-fold validation study.
| We present a complete statistical model for the analysis of RNA-seq data from a population of cells using external RNA spike-ins and a maximum-likelihood method for genome-wide estimation of transcripts per cell. The model includes biological variability of cellular transcript number and sampling noise. We derive an unbiased estimator of transcripts per cell for every transcript, given by simply multiplying the count by a library-dependent, but transcript-independent, scale factor. This is a nominal estimate that can be converted to an absolute estimate by dividing by the transcript’s relative yield coefficient, measured in a separate experiment. A negative binomial probability mass function with novel normalization (size) factors allows for parametric testing of hypotheses concerning dependence of the absolute abundance of each transcript on experimental condition. Our method integrates information from every RNA-seq experiment across all replicates and experimental conditions to determine the calibration constants. We test the method with a dilution study and a k-fold cross-validation study. We illustrate our method with applications to two independent data sets from yeast and the sea squirt that were derived by different library preparation protocols. We show that our methods detect genome-wide amplification of expression, and we compare our method to others.
| Accurate transcriptome measurements are central to understanding the fundamental mechanisms of gene expression. A main challenge presented by the RNA-seq method for digitizing information about cellular RNA content—both its composition and abundance—is correcting noise, errors, and biases introduced in the process of making the measurement. An important step in typical library preparation for sequencing is random fragmentation of the molecules to be sequenced. The actual unit that is being digitized during RNA-seq is therefore not the RNA molecules directly, but their fragments, whose number depends on transcript length and RNA abundance. However, it is the molecular abundance of a transcript that is of interest, rather than the distribution of fragments over its gene model.
The introduction of Reads per Kilobase of exon model per Million mapped reads (RPKM) as the unit measurement by Mortazavi et al. [1] addressed this problem. The work of Mortazavi et al. [1] and of Tarazona et al. [2] both addressed the problem of varyng sample size (sequencing depth). Other researchers looked more carefully at fragment biases and developed a maximum likelihood algorithm to estimate the true differences [3]. The adoption of “read-counts” overlapping each transcript [4] instead of FPKMs (Fragments per Kilobase of exon model per Million mapped reads) [5] has allowed a more intuitive interpretation of the data.
Among the models that have been proposed for RNA-seq count data [6–9], the intuitively appealing negative binomial model [10] has become the most popular. The negative binomial probability mass function can be thought of as a mixture distribution arising from concatenating biological noise in transcript abundance (described by a gamma probability density function) and sampling noise (described by a Poisson distribution) in the compilation of corresponding sequencing counts. Applications of the negative binomial distribution and methods of hypothesis testing have been reviewed recently in [11]. Widely used normalization, statistical modeling, and hypothesis-testing methods are implemented in widely used R packages, edgeR [12], EDASeq [13], and DESeq2 [14]. Methods that focus particularly on the removal of unwanted variation from unspecified extraneous, nuisance sources are implemented in the R package RUVSeq [15]. A common assumption among all these methods is that, while the proportions of some transcripts vary across conditions/treatments, most transcripts do not vary between experimental conditions, and the total abundance of cellular RNA remain more-or-less fixed.
Lovén et al. [16] were the first to demonstrate an experimental system in which the central assumption of transcriptome equivalence among conditions is not satisfied. The researchers discovered that in cells overexpressing the oncogene cMyc, 90% of all transcripts are also overexpressed. To overcome the problem and allow comparison of expression levels between normal and cMyc overexpressing cells, the group incorporated external spike-ins in their samples, which were then used as a de facto invariant pool of RNAs. The spike-in approach had been previously used successfully in microarray experiments and its use in RNA-seq was facilitated by the development of external RNA spike-in mixes by the External RNA Controls Consortium (ERCC) [17, 18]. The need to normalize high-throughput RNA and DNA counts, in general, by the use of spike-in standards was recently explained and validated in the wide-sweeping paper of [19]. Even more recently, external RNA spike-ins were used in single cell RNA-seq experiments [20].
The ERCC external RNA spike-in mix 1 (Ambion) that was used in this study, consists of 92 different synthetic RNAs at 22 different concentrations spanning six orders of magnitude (30,000–0.01 amol/μL). Instead of relying on normalization methods that aim to match the cellular and spike-in RNA read-count distributions, we take advantage of the digital nature of the RNA-seq output and we use the spike-ins as calibrators of known absolute abundance in the samples. We demonstrate that our calibration/normalization model is applicable in two different model organisms (the unicellular eukaryote Saccharomyces cerevisiae and the multicellular chordate Ciona robusta, three experimental setups, a growth rate regulation and a dilution study in yeast, as well as an embryonic differentiation and cell lineage specification study in Ciona), and two library preparation protocols. We perform hypothesis testing and detect global amplification of gene expression in both organisms.
We used three distinct experimental setups, two representing different cases in which the assumption of constant transcriptome sizes is violated. We added a fixed, known amount of external RNA spike-ins to the sample of cells. In the rest of the paper, we refer to spike-in abundance per cell to mean the ratio of spike-in amount (molecules) added to the sample divided by the number of cells in the sample.
In all libraries we avoided the use of poly-dT for reverse transcription as it has been previously shown to be incompatible with quantitation through external spike-ins for RNA-seq [28]. In all cases the filtered aligned reads were converted to counts using the function featureCounts from the package Rsubread (R, Bioconductor) and strand information.
GR samples and samples for the dilution study were prepared essentially as described in the Borodina et al. [29] directional RNA-seq protocol. We modified the protocol by using UMI adaptors [30, 31] that were used to eliminate PCR duplicates from the results. RNA was extracted from ten million cells after the addition of 2 μl of 1:20 dilution of ERCC spike-in Mix 1 stock [32] in the lysis buffer. All samples were distributed in two lanes and sequenced on an Illumina HiSeq 2000, with 100 nt-long, single end reads. The RNA-seq data (fastq files) were first filtered for residual rRNA reads. The data were aligned to the latest version of the yeast genome (sacCer3) and spike-in sequences, and filtered for mapping quality using Bowtie with optimized parameters. The aligned reads were then processed with a custom script that removes PCR duplicates based on the combination of mapping coordinates and UMI adaptor barcodes.
For the samples of three different in vivo Ciona cell lineages, 800 cells were directly sorted into lysis buffer from RNAqueous-Micro Total RNA Isolation Kit (Ambion) containing a fixed amount of ERCC spike-in mix 1. Total RNA extraction was performed according to the manufacturer’s instructions, followed by depletion of rRNA in the samples. The quality and quantity of total RNA in all stages were measured using Agilent RNA 6000 Pico Kit (Agilent) on Agilent 2100 Bioanalyzer. cDNA were synthesized using the SMART-Seq v4 Ultra Low Input RNA Kit (Clontech). RNA-Seq Libraries were prepared and barcoded using Ovation Ultralow System V2 1-16 (NuGen). The samples were pooled in one lane and sequenced on an Illumina HiSeq 2500, with 50 nt-long, single end reads. The RNA-seq reads were mapped to the Ciona genome (v.2008, ghost database) using Tophat2 with default parameters. The mapped reads were assigned to Ciona KH gene models (v.2013).
All computer programming, including data analysis, Monte Carlo simulations, hypothesis testing, and figure preparation were done in the R programming language and environment. For pairwise testing for differential gene expression, we used DESeq function in the DESeq2 package [14]. In some cases hypothesis testing followed the application of an RUV normalization method in a suite of 3 R functions, RUVr, RUVs and RUVg, in the RUVseq package [15]. For maximum likelihood estimation of parameters we used the R function nlm, a general R function that minimizes a supplied objective function over its parameters. We used it to minimize minus log likelihood of the observed data in the context of a model. The first argument of the nlm function is the name of the name of the R function that computes minus log likelihood for the problem at hand. We wrote such a function for each maximum likelihood problem we considered.
All symbols, variables, parameters, and statistics used throughout this paper are listed in Table 1.
We use external RNA spike-ins as a calibration tool to normalize RNA-seq counts by introducing the variable relative yield (Fig 1). We parametrize a multinomial model for sampling noise, conditioned on native RNA abundances, with library size, relative yield coefficients, and known absolute abundances of spike-in molecules. In the context of this model, we derive a maximum likelihood estimation of nominal RNA abundance, proportional to absolute abundance, for each endogenous transcript in our sample. In the remainder of this paper, we often omit the qualifier “nominal” when we mean nominal abundance, and use the phrase “absolute abundance” when we mean molecules or attomoles (per cell or per sample).
In S1 Appendix, we show that the maximum likelihood estimator of RNA abundance is a library dependent scaling of counts by a factor that is proportional to total spike-in counts; it also depends on the known abundance of a reference spike-in, and fraction of overall spike-in counts contributed by this reference spike-in.
The expected proportion of counts for a given molecule (native RNA or spike-in) i represented in library j depends on the product of its abundance (attomoles or molecules per cell) in the original sample ni,j and its relative yield coefficient αi; (Fig 1B(ii)). For spike-in molecules, ni,j are known. We define the relative yield coefficient of a molecule (spike-in or RNA) to be the ratio of its yield coefficient to that of a reference spike-in. By yield coefficient of a molecule we mean the expected number of fragments per molecule contributed by that molecule to a total RNA-seq library of fixed size and prepared according to a fixed protocol. The relative yield coefficient captures specific properties of an RNA molecule such as transcript length and GC content. By convention, we assign the index 1 to the reference spike-in; consequently α1 = 1, by definition (Fig 1B(i)). For the sake of generality, we do not assume that relative yield coefficient is proportional to transcript length as in [33]. The relative yield coefficient of a spike-in is related to its FPKM within an RNA-seq spike-in library as follows. The relative yield coefficient multiplied by abundance in the original sample, and divided by length is proportional to FPKM.
In S1 Appendix, we express the expected proportion of counts for each molecule, indexed by i, in library j in terms of all zi,j = αini,j, in a multinomial joint distribution of counts, and then solve for the maximum likelihood estimator of zi,j. Because we do not estimate the relative yield coefficients, αi for cellular RNA molecules in the present paper, we cannot disentangle here their relative yield coefficients αi and absolute molecular abundances, ni,j (see Discussion). Consequently, we refer to zi,j as a nominal abundance. The corresponding terms for spike-in molecules do not depend on library index j when the same amounts of spike-ins are used in each sample; so, for spike-ins, we can write more simply zi = αini. For sake of clarity, we mention that, by our convention, the indices for the s spike-in molecules are i = 1, 2, …, s, and the molecule indices for the q detected native RNA molecules are i = s + 1, s + 2, …, s + q.
As shown in S1 Appendix, the derived maximum likelihood values of abundances, zi,j for the RNA molecule i in library j are given by
z i , j = y i , j ν j for i = s + 1 , s + 2 , … , s + q , and j = 1 , 2 , … , r (1)
where yi,j is the count (sequencing reads for RNA molecule i in library j and νj is the maximum likelihood calibration constant for library j. The calibration constant νj is given by
νj=deff1LjSIn1, (2)
where f1 is the proportion of spike-in counts across all libraries contributed by the reference spike-in, n1 is the attomoles or molecules per cell, depending on one’s choice of units, for the reference spike-in, and L j SI is the size (total counts) of spike-in library j.
In S1 Appendix we extend the estimation of zi,j to a full statistical model, including biological variation, linking cellular RNA abundance to RNA-seq counts.
The νj calibration constant in Eq (1) is qualitatively like the dimensionless “technical” (library) size factor sj of [34], but with an explicit relationship to absolute abundance, because 1/νj is on the scale of attomoles or molecules per cell. The relationship between the two factors is discussed thoroughly in S8 Appendix. The numerator on the right-hand side of Eq (2), according to our statistical model, is the expected number of counts from the reference spike-in, in replicate j, given the spike-in library size L j SI. As shown in S2 Fig, the “expected” counts given by the model for the reference spike-in, closely approximate the actual number. Therefore, Eq (1) says that the inferred abundance of RNA transcript i in replicate j is given by the counts for this transcript multiplied by a scale factor that is the attomoles or molecules per cell, per count of the reference spike-in. If it were known that, in fact, RNA transcript i on average yields twice as many aligned counts per amol as the reference spike-in, the abundance of RNA transcript i in the 107 cells (from which our sample came) would be given by zi,j/2.
The normalization in Eq (2) is reminiscent of RPM (reads per million mapped reads) normalization, but the denominator involves the total number of reads in the spike-in library only. It makes intuitive sense, because read depth scales spike-in counts and endogenous RNA counts the same way [17]. Therefore, dividing an endogenous RNA count by spike-in counts derived from a fixed number of molecules in the original biological sample simultaneously normalizes for read depth and provides a measure proportional to the molecular abundance of the endogenous RNA in question. [17] applied this sort of normalization to ERCC spike-in counts, but also divided by spike-in length to obtain FPKM (fragments per kilobase per million mapped reads). Under the assumption that count scales with molecular length, FPKM would be proportional to spike-in abundance, in the absence of any molecular biases, e.g., GC content (see Fig 2 in [17]).
As shown in S1 Appendix, the derived maximum likelihood values of abundances for the s spike-in molecules are given by
z i = n 1 f 1 f i for i = 1 , … s , (3)
where fi is the empirical fraction of total spike-in counts, across all libraries, that is accounted for by spike-in molecule i, and f1 is that of the reference spike-in.
Because the absolute abundances of the spike-ins, ni are known, Eq (3) and the definition zi = αini, allows us to estimate the spike-in relative yield coefficients as
α i = z i n i = n 1 n i f i f 1 (4)
We chose as the reference spike-in the one that contributes the largest fraction of overall spike-in counts. Any of the top few spike-ins would do just as well.
We compute the spike-in relative yield coefficients, αi, describe their statistical properties and model the spike-in relative yield coefficients in terms of their biophysical properties in S6 Appendix. However, we do not yet have a large enough repertoire of spike-in molecules nor accurate enough biophysical models to use the model relative yield coefficients for spike-ins to estimate those of native RNA molecules. Thus, the computed spike-in αi terms in the present paper are not participating in the estimation of native RNA abundances. In the Discussion we analyze how αi could be used in the estimation approach.
Following the recommendation of [35], we prepared diagnostic relative-log-expression (RLE) plots for all three of our data sets (dilution study, yeast GR study, Ciona embryonic differentiation study) to help in evaluating our maximum likelihood (νj) calibration method. We found unwanted variation in the total inferred RNA abundance within conditions for all three data sets, and the variations are similar across data sets. We ascribe this variation to technical errors in one or more steps in the preparation of samples to be sequenced: variation in RNA extraction efficiency, error in cell count, dilution and/or volume errors in preparation of the spike-ins added to the cellular RNA. We refer to these errors collectively as library preparation errors. Details of results and analyses are presented in S1 Fig and S2 Appendix. Accordingly, we derived a library-specific scale factor, δj (S2 Appendix), much like the total RNA correction factor, ξj, in the single-cell RNA-seq study of [36] (S8 Appendix). The corrected nominal abundance values are computed as zi,j/δj, and we performed statistical analyses and hypothesis testing on these corrected values. In S8 Appendix we discuss, and in S5 Fig. we illustrate, the similarities and differences of this noise reduction to a removal of unwanted variation (RUV) method RUVr in the RUVSeq R package [15] that is based on residuals in a generalized linear model.
For this dilution study we prepared 3 replicate libraries with a “high” spike-in aliquot dilution, and 3 with a“low” dilution as described in Materials and Methods. In the absence of any library preparation noise, the molar amount of spike-ins are 4.44 times larger in samples that were added to low dilution spike-in aliquots compared to those with the high-dilution aliquots. The generalization of the νj normalization in this case is to simply add the subscript j (indexing library) to nref to give ν j = f 1 L j SI / n 1 , j. Perfect performance of our method, would give identical mean abundances for libraries prepared with the high- and low-volume aliquots. In Fig 3 the MA plot shows less than ideal performance in that the ordinates in the scatter plot are offset a bit (by 0.28) from zero. This corresponds to a mean fold difference of 1.2 rather than 1. This discrepancy might be due, at least in part, to sample preparation handling errors.
Modeling RNA abundances within a condition as gamma-distributed random variables results in counts with the familiar negative binomial distribution, as shown in S1 Appendix. Previous applications of the negative binomial distribution for modeling RNA sequencing counts and various methods of hypothesis testing have recently been reviewed in [11]. Our model for sequencing counts is formally equivalent to that in [34], but with an important distinction. The library-specific size factor of [34] (based on genes), written according to our notation, is
s j = median i y i , j (∏ k = 1 r y i , k ) 1 / r . (5)
The size factors sj is a dimensionless constant that carries with it an implicit assumption of fixed total amount of cellular RNA. Furthermore, in practice, the median of sj across libraries is of order 1. In contrast, although our calibration constants νj can be thought of as “size factors,” they are proportional to the total spike-in library size L j SI in Eq (2), and 1/νj has dimensions of attomoles, or molecules per cell, depending on the units one chooses to use for the spike-in abundances ni (i ∈ {1, 2, …s}).
Our νj calibration factor is closely related to the extension of [34], by [36] to estimate cellular RNA abundance with the use of a “technical size factor” resembling that in Eq (5), but computed using spike-in counts only. Hypothesis testing methods currently in the literature that are based on the negative binomial distribution of counts with library size factors sj, such as DESeq [34], can be used following our νj normalization in the manner described in S4 Appendix.
In this paper, we described a detailed statistical model for cellular RNA and exogenous spike-ins in a sample prepared from a fixed number of cells to which a population of spike-in molecules of known numbers has been added. In the context of this model, we derived by maximum likelihood arguments, a calibration method for RNA-seq data that estimates cellular molecular abundance of RNA. Although our molecular abundance z-values are nominal, they are only one step away from absolute molecular abundance. Once the relative yield coefficient for transcript i, αi, is measured in separate experiments, the absolute molecular abundance in library j, ni,j will be known via the equation: ni,j = zi,j/αi.
Our method employs an explicit statistical model for spike-ins, the simplest sensible one, namely that the spike-in counts for a given library are sampled from a joint multinomial distribution with fixed proportion parameter for each spike-in molecule across all libraries/conditions for a fixed protocol. As a consequence, the counts within each spike-in library, regardless of condition, represent a technical replicate. We evaluated the spike-in model quantitatively in a number of ways (Fig 2 and S2 Fig). We found that the spike-in molecules adhere closely to the multinomial model provided the spike-in library exceeds roughly 250,000 reads. In other words, our results support those of [19]: the spike-in molecules contribute to spike-in counts within a spike-in library, embedded in an overall RNA-seq library, in a manner which is independent of the native RNA. A caveat is that we don’t know for sure if deviations of spike-in counts from the multinomial model that we observed are a consequence of some sort of poorly understood noise that is particularly prominent in spike-in libraries of low size, or if the unaccounted for noise was unrelated to library size per se.
We adopted a multinomial mode for spike-in noise, but our model could be extended with a more accurate model. Technical noise in spike-in counts has been studied and modeled recently [36], and we present similar analysis and modeling in S2 Fig and S5 Appendix. Although the proper experimental technique was followed in our study to minimize these errors, pipetting and dilution errors can not be completely eliminated. Pipetting, dilution, and cell number errors may have been sources contributing to the very high variation between experiments that was observed in previous attempts to incorporate spike-ins as normalization standards [42]. [20] however demonstrate technical robustness in the performance of spike-ins in sensitive single cell RNA-seq experiments. Our data agree with the assessment of [20].
We have shown however that our method, especially when supplemented with RUVr [15] correction or our own δj correction, is able to compensate for this source of unavoidable technical variability. Our model could be extended and improved in the future by incorporating a different model for spike-ins. Nevertheless, our model allows for powerful, genome-wide, parametric testing of hypotheses of various sorts concerning nominal RNA abundances, z-values that are explicitly related to absolute cellular molecular abundance (transcripts per cell or attomoles).
We applied our method, to quantify RNA abundance and to test for differential gene expression, using data from two studies with different library preparation protocols, and in species from different kingdoms: a growth rate study in yeast, and a low cell count differentiation study in Ciona. We found global changes in gene expression in both systems: a global increase in transcript abundance with growth rate in yeast, and a global decrease in the FgfrDN embryonic cell type in Ciona. Reanalysis of the raw data with other algorithms that hold the assumption of equivalent transcriptome sizes, as expected, were not able to reveal these global transcriptome trends.
Our focus in this paper is on deriving a nominal cellular molecular abundance that can be converted to absolute abundance by the transcript’s relative yield coefficient, which could be measured in separate experiments. In this study however, we do not attempt to measure the relative yield coefficient values, or estimate the absolute number of molecules per cell for each transcript within a condition. The current work allows us to say, that, for example, RNA transcript A has x times more molecules per cell, on average, in condition 1 compared to condition 2, even if the corresponding RNA-seq libraries were prepared in different batteries of experiments, different studies, or even prepared in different laboratories. Such a conclusion about what might be called, an absolute ratio of abundances, can be drawn without knowing the relative yield coefficient of transcript A. In the section that follows, we discuss the links between our work and methods by which these relative yield coefficients might be measured.
In this manuscript we offer RNA abundance estimates that are proportional to absolute transcript abundance. For this we assign a (relative) yield coefficient value of 1 to a reference spike-in, arbitrarily chosen from among those that contribute a sizable fraction of total spike-in counts. Our nominal abundance of an RNA molecule is based on the temporary assumption that this molecule has the same yield coefficient as the reference spike-in. If our calibration method is supplemented with additional data on the effect that a broad range of transcript physicochemical characteristics has on library preparation and sequencing, a more realistic relative yield coefficient could be assigned to each RNA molecule of interest.
A technical statement of the outstanding problem is that our inferred nominal abundances zi,j do not disentangle true absolute molecular abundance, ni,j, and the corresponding relative yield coefficient, αi; because, by definition, zi,j = αi ni,j. However, once one measures absolute cellular abundance of transcript i in a preparation of cells from which library j was derived (ni,j), the relative yield coefficient becomes known, at, least in the idealized situation ignoring various sorts of noise, because αi = zi,j/ni,j. For example, ni,j might be measured by single-cell Fluorescence In Situ Hybridization (FISH) methods, performed on a large population of cells from which library j was derived.
Statistical methods taking into account biological noise and technical noise could be used to compute a confidence interval for αi, provided ni,j could be estimated. Likelihood methods could be used to integrate data across several libraries in the estimation of αi. In principle, once αi is estimated from one or more libraries and a population of cells from which those libraries were derived, this estimate could be used for other libraries (prepared using the same protocol), past, present, and future, to allow the determination of absolute cellular molecular abundances of transcript i.
Modeling, like that presented in S6 Appendix and S2 Fig, and like that of [17] could also play a vital role in estimating relative yield coefficients, especially if a wider array of synthetic spike-ins covering a large gamut of physical properties were designed and utilized. Our methods have the potential of facilitating statistical modeling of RNA counts because of the explicit relationship between our nominal abundances and absolute molecular, cellular abundances of RNA. In principle, variation in counts as a consequence of true biological variation in random attomoles, N, and variation in counts due to variation in relative yield coefficient across transcripts with nearly identical mean abundances, μN, could be disentangled.
Our approach lays the groundwork for investigating, testing, and modeling how the physical properties—e.g., length, GC content, folding energy—determine the relative yield coefficient of spike-ins and native RNA transcripts alike. Empirical measurements of relative yield coefficients, as we have defined them, and biophysical modeling could facilitate progress in making the connection between sequencing counts and the underling molecular cellular abundances of the corresponding transcripts.
Our work follows up on and extends the work of [15, 16, 36, 43, 44]. Our inference method is linear and global for each library, like that of [19], [36] and [45]. We showed that our global (library specific) νj calibration constants are closely related to the Anders and Huber-like “technical” size factors of [36], which are based on spike-in counts. We called their normalization constants s j SI, and we showed that they are proportional to our νj normalization constants in the cases of 2 of our data sets with large library sizes, as predicted by theory (S8 Appendix). An important difference is that the s j SI calibration constants are on a dimensionless scale, on the order of 1, and do not allow one to infer absolute abundances of transcripts once their relative yield coefficients become known.
[16] applied loess normalization to ERCC spike-in counts to derive a normalization function that they then applied to the counts corresponding to native RNA. Our analysis and rigorous testing of our theory and methods suggest that a local nonlinear transformation, such as loess normalization of the count data is not needed for our RNA-seq data. It seems likely that any local nonlinear fitting of counts to make replicate spike-in libraries as similar as possible would involve overfitting the data.
Our work has some important features in common with the HTN method of [46], particularly, the assumptions underlying their Eq (1) and our Eqs S1 Appendix (2) and (3). These equations explicitly allow for differences in total RNA abundance across conditions. In addition, both normalization methods are global and linear. However the HTN method of [46]: relies on having de facto housekeeping genes rather than experimentally-added spike-ins; does not include a model for biological noise; assumes that relative yield is simply proportional to transcript length; is focused primarily on testing for differential gene expression; and does not provide estimates of absolute RNA abundance. Their global scale factor for a given library is determined by minimizing the sum over spike-ins of the square differences between the spike-in counts in that library and those of a library chosen to be the reference library. That scale factor is then used for the native RNA counts within the same non-reference library. It can be shown that this library-by-library normalization procedure, in the limit as library size (native RNA and spike-ins) approaches infinity, will give an abundance measure that is proportional to our z abundance values based on νj normalization.
A quite different suite of normalization methods, called RUV (removal of unwanted variation), was introduced by [15, 36, 43, 44] and applied with great effect to many different data sets. The methods involve singular value decomposition (SVD) variant of factor analysis to compute a factor matrix W, which is used to model nuisance sources of variation that are unrelated to the experimental design. The factor matrix W is included, in addition to a design matrix, in a generalized linear model for normalized counts. One qualitative way of thinking about the W matrix is that is adds columns to the original design matrix for explanatory variables that one didn’t originally know about. Although this method is widely effective at reducing unwanted variation in RNA-seq data, it does not allow one to infer absolute cellular molecular RNA abundances, even if the factor matrix is computed based on spike-ins or an invariant gene set (S8 Appendix), as the authors are well aware. The simple reason is that proportion of spike-in count is tightly correlated with the biological phenomenon of interest the change of total RNA abundance with condition. However, we showed that results of our maximum likelihood normalization method can be improved, with respect to clustering and detection of differential gene expression, by applying an an RUV method based on residual, RUVr (RUVSeq package [15]) after νj normalization. We obtained closely similar results by a simpler method involving a correction factor δj for each library that was based on our discovery in a dilution study with technical replicates that we seem to have some noise in the actual overall amount of spike-ins added to the cellular RNA. We tentatively ascribed these to dilution/volume errors in handling the stock spike-in mixture. This finding highlights the importance of replacing pipetting methods for handling the spike-ins with more accurate robotic methods.
The continuing discovery of examples in which there are gross transcriptome differences between cellular states, has established a need for spike-in controls in RNA-seq experiments [19]. Despite some criticisms [15], external RNA spike-ins have been adopted in several recent studies alongside methods developed to use them for RNA-seq quantitation [16, 19, 36, 46, 47].
The model presented in this work lends itself for both absolute and relative RNA quantitation, dependent on the experimental ability to accurately isolate a fixed number cells for library preparation. In both cases, we offer evidence that our approach provides reproducible results in a wide variety of conditions and has a strong predictive power. In conclusion, the presented model allows for improved unbiased RNA-seq quantitation in any experimental setup using external RNA spike-ins.
|
10.1371/journal.pcbi.1004527 | Reputation Effects in Public and Private Interactions | We study the evolution of cooperation in a model of indirect reciprocity where people interact in public and private situations. Public interactions have a high chance to be observed by others and always affect reputation. Private interactions have a lower chance to be observed and only occasionally affect reputation. We explore all second order social norms and study conditions for evolutionary stability of action rules. We observe the competition between “honest” and “hypocritical” strategies. The former cooperate both in public and in private. The later cooperate in public, where many others are watching, but try to get away with defection in private situations. The hypocritical idea is that in private situations it does not pay-off to cooperate, because there is a good chance that nobody will notice it. We find simple and intuitive conditions for the evolution of honest strategies.
| We study the evolution of cooperation based on reputation. This mechanism is called indirect reciprocity. In a world of binary reputations, people help a good individual but do not help a bad one. They also monitor their own reputation to receive reciprocation from others. We propose a novel model of indirect reciprocity where two types of interactions exist. In a public interaction your behavior is always observed by others. In a private interaction, your behavior is less likely to be observed. We study the competition between honest and hypocritical strategies. The former always help good individuals, whereas the latter do so only in private interactions. We describe conditions for the evolution of honest strategies.
| Most human interactions occur in situations where repetition is possible and reputation is at stake. Repeated interactions in a group of players facilitate evolution of cooperation via indirect reciprocity [1, 2]: here players use conditional strategies that depend on what has happened between others. Cooperation is costly but can establish a good reputation. Others might preferentially cooperate with those who have a good reputation. Many studies explore theoretical [3–49] and empirical [50–70] aspects of indirect reciprocity. Experiments reveal that people help those who help others [50, 52–57, 59, 60, 62–67, 70]. Reputation is a strong driving force of prosocial behavior [54, 61, 71–80]. Internet commerce is based to a large extent on reputation systems: buyers are sensitive to sellers’ reputation [76, 81–87].
The standard framework of indirect reciprocity assumes that each interaction consists of a donor and a recipient. The donor can choose between cooperation and defection. If the donor cooperates, her cost is c and the benefit for the recipient is b. If the donor defects, there is no cost and no benefit. A crucial parameter is the benefit-to-cost ratio, b/c.
Many studies of indirect reciprocity so far assumed that social interactions are public, which means that everyone is informed about the outcome of these interactions. In reality, however, social information is often incomplete. A previous study [6] showed that if donors remember the reputation of their recipient only with probability q, cooperation evolves if b/c > 1/q, suggesting that incomplete information hinders indirect reciprocity.
But in this paper we study another source of incomplete information; the absence of observers. Engelmann & Fischbacher [63] found that donors helped recipients substantially more when their reputation score was seen by others than when it was not publicly announced. Accumulating evidence suggests that humans are very sensitive to cues of being observed, and change their social behavior accordingly [72, 73, 75, 77, 79]. These facts motivate us to study strategies in indirect reciprocity when two types of interactions differing in observability are mixed. In public interactions the donor’s action always affects his reputation, but in private ones it affects his reputation only with probability q and otherwise his reputation is unchanged. In our framework people can use different strategies depending on whether they are in private or in public situations. Moreover, observers can evaluate private and public interactions with different assessment rules. For example, they could be indifferent to displays of public cooperation, but very much reward private cooperation, if they hear about it, or vice versa. Crucial questions such as those have not yet been studied in the context of indirect reciprocity.
We have constructed a theoretical model of indirect reciprocity to study the effect of the mixture of public and private interactions. We consider an infinitely large population of players who are engaged in indirect reciprocity interactions. In each time step, they are randomly matched to form a pair that consists of a donor and a recipient. The donor can choose between cooperation and defection. If the donor cooperates, her cost is c and the benefit for the recipient is b. If the donor defects, there is no cost and no benefit.
In addition, each interaction within a pair is public with probability, p, and private with probability 1 − p. For simplicity we assume that a public interaction is always observed, while a private interaction is observed only with probability q. Therefore, on average, an interaction is observed with probability, q ¯ = p + ( 1 - p ) q.
We consider a world with binary reputation scores: each player has either a good or a bad reputation, depending on his previous actions. We assume that everyone knows and agrees on others’ reputation scores. A strategy in indirect reciprocity games is called an action rule. We have 16 different action rules. Each action rule specifies whether to cooperate or to defect given that the situation is either public or private and given that the reputation of the recipient is either good or bad (Fig 1). For example, the action rule CDCD (honest) means: cooperate with good recipients and defect with bad ones, no matter whether the situation is public or private. In contrast, CDDD (hypocrite) means: in public situations the actor will cooperate with good recipients and defect with bad ones, but in private situations the actor will always defect. There is also unconditional cooperation, CCCC, and unconditional defection, DDDD.
If the other people in the population are informed about an interaction, then they update the reputation of the actor by using their social norm. We consider second order social norms [14, 15], which depend on the action of the donor, the reputation of the recipient and whether the interaction was public or private. Thus, there are 256 possible social norms (Fig 1). We assume that all people in the population use the same social norm. After a single game interaction, players leave the pair and go back to the population pool to wait for another recruitment. After a sufficiently large number of interactions, natural selection acts on action rules according to their payoffs. A goal of our analysis is to identify which of the 16 strategies (action rules) are evolutionarily stable for each of the 256 norms.
Because of the assumption of an infinitely large population, no two players can ever meet more than once, so the chance of any forms of directly reciprocity is excluded from the model. We further assume that the time scale of natural selection is much slower than that of social interactions and reputation updates. Throughout the analysis, therefore, we can assume that the fraction of time that one has a good reputation is at an equilibrium level. The model presented here can further be extended by incorporating the effect of errors and finite time horizon. Details of the present model as well as those extensions are described in S1 Text.
Our analysis uses the method of dynamic programming [18, 27]. Details are provided in S1 Text, but to grasp the idea, here we show one example. Consider the social norm that regards cooperation as good and defection as bad irrespective of the other factors. Such a norm is called “Scoring” [11, 15]. We ask, for example, if the action rule CDCD (honest) is an evolutionarily stable strategy (ESS) under this norm. For that purpose, we temporarily assume that everyone adopts CDCD.
Dynamic programming allows us to calculate the value of a good reputation, v, which reflects the advantage of possessing a good reputation over a bad one. In our example it is calculated as v = b / q ¯ > 0 (see S1 Text for the derivation), suggesting that keeping a good reputation is advantageous, but it is qualitatively a trivial consequence of everyone using CDCD.
With the value v, we check if each digit of the action rule CDCD follows the optimal behavior, because otherwise it is not an ESS. In our example, cooperation in public interactions costs c to the donor but makes his reputation good, by which he obtains the future benefit, v. Therefore, the relative size of c and v matters. If v < c then cooperation in public interactions does not pay, so the first digit of CDCD is not optimal. If v > c then defection in public interactions does not pay, so the second digit of CDCD is not optimal. In either way, CDCD does not follow the optimal behavior, so it is not ESS. An analysis of this kind is systematically repeated for each of the 256 social norms and for each of the 16 action to find all ESS.
We find that notorious DDDD is always ESS. The question is which other strategies are ESS and for what conditions. We note that DDDD is ESS for any social norm, while all cooperative strategies require specific social norms to be ESS. In Fig 2, we show social norms that allow CDDD (hypocrite) and CDCD (honest), respectively, to be ESS.
We obtain the following results. If b/c is less than both q ¯ / q and q ¯ / p then the only ESS is DDDD. If b / c > q ¯ / p then CDDD is ESS. CDDD achieves only partial cooperation. If b / c > q ¯ / q then CDCD is ESS; this is the crucial condition for the evolution of an honest strategy. Social norms that support the evolutionary stability of CDCD turn out to be the combinations of “Simple-Standing”, “Kandori” (also known as “Stern-Judging”), and “Shunning” social norms [15, 19, 21] (see Fig 2). There is also the possibility that more lenient strategies can evolve if stronger conditions are fulfilled. If b / c > q ¯ / [ ( 1 - p ) q ] then CCCD is ESS. If b / c > q ¯ / ( p q ) then CDCC is ESS. All three strategies, CDCD, CCCD and CDCC, achieve full cooperation for appropriate social norms. S1 Text describes all the ESS we found in our analysis. Note that our model allows multiple ESS. In Fig 3, we show ESS that achieve the highest level of cooperation for the given parameters, because either group competition [17, 19, 22, 30] or intergroup contingent movement by players [27] is likely to favor such ESS. Fig 3 also distinguishes between the two cases, b/c > 2 and 1 < b/c < 2. Quite interestingly, in the latter case we find a pocket inside our parameter space (shown in gray in Fig 3) where no combination of p and q allows the evolutionary stability of action rules other than DDDD. This result is an unexpected consequence of the interplay between public and private interactions; the reputation that one acquires in a public interaction affects one’s private interactions, and vice versa.
Our results have simple intuitive justifications. Consider the crucial condition, b / c > q ¯ / q, that needs to hold for the honest strategy, CDCD, to be the most cooperative ESS. The most dangerous invader is CDDD. Let us now examine the situation where you have a good reputation and meet another person with a good reputation in a private situation. If you cooperate then you lose c, but maintain a good reputation. If you defect then you save c, but obtain a bad reputation with probability q. Once it occurs you will not receive cooperation (losing b) until you recover a good reputation. To regain a good reputation your help must be observed by a third party, which occurs with probability q ¯ per interaction. Therefore you must wait 1 / q ¯ rounds. Thus, cooperation costs c, whereas defection costs q b / q ¯. Cooperation is less costly if c < q b / q ¯ yielding the desired condition. Similar intuitions exist for other conditions and are described in S1 Text.
When comparing with previous models of incomplete information, one may wonder why the evolutionary condition of the honest strategy, CDCD, is now more relaxed. Note that b / c > q ¯ / q requires a lower benefit-to-cost ratio than the previously known condition, b/c > 1/q. The explanation is as follows. In previous models, observers are always present and your reputation as a donor is always updated, but people sometimes fail to remember your reputation. In contrast, our current model assumes that observers are sometimes not present. The absence of observers could be good news for defectors, but not necessarily; once you obtain a bad reputation, it carries over to your future interactions until your next interaction is observed. Therefore, defectors experience more hardship in the current model.
In Fig 4 we compare our analytical results with numerical simulations of evolutionary dynamics. We examine a particular social norm where defection against a ‘good’ recipient leads to bad reputation in public and private situations (when they are observed), while other actions lead to good reputation. The pairwise invasion plots confirm our evolutionary stability analysis for the various parameter regions. We also study convergence to equilibrium points starting from various initial conditions. The system behaves in accordance with our analytical results. Additional numerical tests are performed in S1 Text.
We have studied indirect reciprocity in public and private situations. Our system has three parameters: the benefit-to-cost ratio, b/c, the frequency of public interactions, p, and the probability that private interactions are revealed, q. Of fundamental interest is the competition between the honest strategy, CDCD, which cooperates with deserving recipients both in public and private and the hypocritical strategy, CDDD, which only cooperates in public but never in private. We find that for reasonably small q values CDCD can prevail over CDDD. The critical q for CDCD to be the most cooperative ESS is a declining function of b/c and an increasing function of p. Specifically, we can derive the following prediction: helping in a private interaction is suppressed (i) if the observability q is low AND if (ii) private interactions are rare (large p); see Fig 3. Empirical studies are needed that examine a mixture of public and private interactions in a laboratory setting or field study to test this prediction.
One is tempted to associate the two key strategies with different motives: CDCD behaves ‘properly’ irrespective of the probability of being observed, while CDDD cooperates (with deserving recipients) only when there is certainty that others will notice the good deed; in contrast it tries to get away with defection when no one is watching. Thus CDCD seems to have ‘higher morals’, while CDDD represents a more utilitarian approach. There is, however, also a cynical interpretation of CDCD: as we have noted this strategy is only stable if the probability to be observed in private interactions is sufficiently high; therefore a CDCD player cooperates in private situations, because on average it does pay-off to do so. Our theoretical results suggest that strategic reputation building [15, 57, 63] is a strong driving force for the evolution of action rules in indirect reciprocity.
|
10.1371/journal.pcbi.1003951 | The HIV Mutation Browser: A Resource for Human Immunodeficiency Virus Mutagenesis and Polymorphism Data | Huge research effort has been invested over many years to determine the phenotypes of natural or artificial mutations in HIV proteins—interpretation of mutation phenotypes is an invaluable source of new knowledge. The results of this research effort are recorded in the scientific literature, but it is difficult for virologists to rapidly find it. Manually locating data on phenotypic variation within the approximately 270,000 available HIV-related research articles, or the further 1,500 articles that are published each month is a daunting task. Accordingly, the HIV research community would benefit from a resource cataloguing the available HIV mutation literature. We have applied computational text-mining techniques to parse and map mutagenesis and polymorphism information from the HIV literature, have enriched the data with ancillary information and have developed a public, web-based interface through which it can be intuitively explored: the HIV mutation browser. The current release of the HIV mutation browser describes the phenotypes of 7,608 unique mutations at 2,520 sites in the HIV proteome, resulting from the analysis of 120,899 papers. The mutation information for each protein is organised in a residue-centric manner and each residue is linked to the relevant experimental literature. The importance of HIV as a global health burden advocates extensive effort to maximise the efficiency of HIV research. The HIV mutation browser provides a valuable new resource for the research community. The HIV mutation browser is available at: http://hivmut.org.
| Naturally occurring mutations within the HIV proteome are of therapeutic interest as they can affect the virulence of the virus or result in drug resistance. Furthermore, directed mutagenesis of specific residues is a common method to investigate the function and mechanism of the viral proteins. We have developed novel computational text-mining tools to analyse over 120,000 HIV research articles, identify data on mutations and work out which amino-acid in which protein has been mutated. We have organised these data and made them available in an online resource—The HIV mutation browser. The resource allows HIV researchers to efficiently access previously completed research related to their region of interest in the HIV proteome. The HIV Mutation Browser complements currently available manually curated HIV resources and is a valuable tool for HIV researchers.
| Human immunodeficiency virus (HIV), the causative agent of acquired immunodeficiency syndrome (AIDS), infects millions of people worldwide and, to date, has been responsible for over 25 million deaths [1]. The clinical importance of the virus has prompted substantial funding of HIV/AIDS research across many diverse clinical, therapeutic (drug design, vaccine production) and basic research fields. This research has produced an extensive catalogue of HIV literature and consequently finding literature pertinent to a particular topic is a difficult task. Researchers are often interested in the phenotypic variation resulting from naturally occurring single nucleotide polymorphism or directed mutagenesis in the HIV genome. Traditionally, mutation data for a particular protein or region must be manually collected by trawling literature repositories such as PubMed using author names, protein/gene names, keywords or a mixture of all three. The scale of the HIV literature (over 270,000 articles) makes such an approach inadequate. Several valuable online resources have provided mutation data to researchers by manually curating polymorphism and mutagenesis data from HIV studies. These include the Stanford Drug Resistance database [2], which curates mutations related to drug resistance, the UniProt knowledgebase [3], which manually annotates articles describing mutagenesis of HIV proteins and the Los Alamos HIV Database, which annotates various sources of HIV data including epitope variants and escape mutations (http://www.hiv.lanl.gov/). However, these resources are limited in scope because manual curation cannot feasibly be carried out on all of the available literature.
The technology exists to quickly computationally scan, annotate and organise scientific literature and these techniques should be applied to facilitate the work of HIV researchers [4], [5]. Consequently, it is surprising that so few resources are available to access the available literature in an organised and structured way. This incongruity can partly be explained by the strict licensing agreements with scientific publishers that prohibit the bulk download and computational processing of scientific research literature. Fortunately, recent pressure from government and scientific bodies and the rise of open access publishing has softened the stance of publishers and many are now receptive to waiving these restrictions. Such advances will pave the way for many large-scale literature text-mining projects and will likely change the way we access scientific literature.
Here we have applied text-mining techniques to extract data on polymorphisms and mutations from the available HIV literature. We have organised this data in a protein and residue-centric way and have made it available through an online resource, the HIV mutation browser (http://hivmut.org). This publicly available resource will simplify the task of virologists attempting to identify the relevant literature for their research, thereby aiding experimental design and reducing replication of efforts.
Creation of the HIV mutation browser required a number of steps (Figure 1). First, we obtained permission from publishers, identified, and accessed the relevant literature. Second, we established and applied text-mining techniques to retrieve data on mutagenesis and polymorphism from the HIV literature. Third, we associated the mutation data to the appropriate residues within the HIV proteome. Finally, we developed a browser through which the data can be accessed in an intuitive and informative way.
We identified ∼270,000 articles containing the search term “HIV” or “Human Immunodeficiency Virus” indexed in the PubMed database (from a total of ∼23 million publications). We retained 120,899 of these articles, published across 2,614 journals, representing approximately 45% of the total (see materials and methods, Figure 1). For the remaining ∼150,000 citations, permission for computational processing of articles was not obtained from the publisher. The 120,899 articles from participating publishers were text-mined for mutagenesis or polymorphism information, and the mutations were mapped to particular residues within the HIV proteome. This required the development of a method to retrieve the text of these articles, scan the articles for patterns that are widely used to describe directed mutations in mutagenesis experiments or polymorphisms, and to map these mutations to the correct position in the correct protein (see materials and methods). A total of 7,608 distinct mutations (a unique non-wildtype amino acid at a given residue in a given protein) were collected. As each mutation can be described in multiple articles and each article can describe multiple mutations, the 7,608 distinct mutations were defined by 43,264 unique references to 5,267 articles.
The identified mutations shed light on the nature of the HIV research effort of the last decades. On the one hand it has been broad in scope: 2,520 of the 3,118 residues in the HIV proteome have one or more associated references to a mutation in the repository. On the other hand it has been narrow in focus: the coverage is far from uniform and certain regions such as the catalytic sites of the protease and reverse transcriptase, as well as host interaction interfaces, are much more highly studied (Figure 2).
The above analysis resulted in a database within which each reference to a mutagenesis experiment or polymorphism in a citation is indexed using three pieces of information: the protein in which the mutation is present, the position in the protein which has been mutated, and the non-wildtype amino acid to which the wildtype residue has been mutated. To make this data accessible to virologists in a simple, intuitive and informative manner, we designed the HIV Mutation Browser, as a web-interface that acts as a front end for the database. The browser presents the data in a hierarchically organised manner. The user selects a gene of interest, then a position of interest, and the citations relating to this position are presented to the user grouped by non-wildtype amino acid.
The web interface is organised in three panels: the navigation panel at the top; the protein panel in the middle; and the residue panel at the bottom (Figure 3).
The available mutagenesis and polymorphism data for a residue can be downloaded in both tab delimited text and Excel formats directly from the web interface.
HIV is an important therapeutic target and has been the subject of a major research effort as evidenced by the large catalogue of HIV experimental literature. Appropriate organisation and categorisation of the available HIV literature is necessary to allow efficient and intuitive access to relevant data. In this paper, we have presented the HIV Mutation Browser, a residue-centric resource of HIV mutagenesis and polymorphism literature designed for use by those carrying out basic and applied HIV research. The HIV Mutation Browser is one of the first resources to computationally text-mine mutagenesis and polymorphism data [7], [8], [9], and the first to apply such methods to the extensive corpus of HIV literature. As such the HIV Mutation Browser will complement the available manually annotated and curated HIV resources such as the Stanford Drug Resistance database [2], the UniProt knowledgebase [3] the Los Alamos HIV Database (http://www.hiv.lanl.gov/). In the coming years, we expect this method or similar methods to be applied to other viral or cellular systems.
The resource will continue to evolve in the following ways. Firstly, HIV literature is produced continuously at a rate of approximately 1,500 articles a month and consequently the HIV Mutation Browser resource will be updated on a quarterly basis. Secondly, while the resource does contain the majority of important HIV and general interest journals, it is still incomplete, as we did not receive permission from all publishers to text-mine their HIV related articles. Journals from additional publishers will be added when possible. Thirdly, not all mutations can be correctly identified and assigned by the text-mining methods. There are various reasons for this. Many mutations are annotated in an article using non-standard patterns that are not widely used to describe directed mutations in mutagenesis experiments or polymorphisms. For example, consider the following excerpt taken from an article by Mitchell et al., “The phenotype of the combination mutant VpuD51A-S52/56N was indistinguishable from that of Vpu-D51A and Vpu-S52/56N” [10]. The pattern “S52/56N” is a non-canonical construct for describing a mutagenesis experiment and currently will not be discovered by the text-mining method. Furthermore, the position of a mutation in a paper can be ambiguous and as a result mapping of the mutation information to the correct residue and protein can be a difficult task. For example, when multiple proteins are referenced and only a single mutation is discussed (more than one possible mapping can be possible), when unconventional numbering is used (particularly when describing mutations in Gag or Env as both are translated as polypeptide chains and subsequently cleaved) or when unusual strains with insertions and deletions are used (this shifts the numbering of residues). We will continue to improve the methods for text-mining and assignment. We request that members of the community utilise the feedback system for misannotated mutations in the resource and contact us about mutation data that should be in the resource yet is not present. This community input will improve the quality of the annotated data and will pinpoint parts of the text-mining method that require improvement.
In summary, the HIV Mutation Browser is a valuable addition to the currently available HIV resources that will allow researchers to quickly and intuitively access data on mutagenesis and phenotypic variation. We expect the database to aid the process of experimental design and be a key resource for the HIV community.
A list of HIV-related articles was programmatically retrieved from PubMed using the search terms “HIV” and “Human Immunodeficiency Virus”. A list of target journals was constructed based on the number of published HIV-related articles. The licensing agreements of the majority of scientific journals prohibit a licensee from (1) downloading articles in bulk and (2) computationally processing the text of an article. Permission to waive these aspects of the licensing agreement was requested and received from the majority of virology and general interest scientific journals. The text of all HIV-related articles from the participating journals was retrieved programmatically from the publisher's websites to create a HIV literature dataset. An up to date list of the participating journals and publishers is available on the HIV Mutation Browsers website.
There is no globally applied nomenclature to define directed mutations in mutagenesis experiments or polymorphisms [5], [11], [12]. A set of templates that define phrases and shorthand widely used to describe directed mutations in mutagenesis experiments or polymorphisms was created based on the work of Caporaso et al. [5] (Figure 4A, see Table S1 for full list). Each article in the HIV literature dataset was converted to plain text and scanned using this set of templates. These templates consist of 3 pieces of information: the position in the protein which has been mutated, the amino acid present in that position in the wildtype sequence of the isolate, and the non-wildtype amino acid to which the residue has been mutated. For example, consider the sentence “As reported previously, S52A and S56A mutations of Vpu had no effect on virus release” [13]. S52A and S56A refer to the experimental mutagenesis of a serine to an alanine at position 52 and 56 in the Vpu protein.
The annotation of a mutation text-mined from papers in the HIV literature dataset requires three piece of information: the sequence of the isolate used in the study; the protein containing the mutation; and the position of the mutation within the protein. This information is sufficient to map a mutation to a reference HIV-1 proteome, but cannot always be directly extracted from the text-mined paper. The nomenclature for describing isolates, genes, proteins, chains and domains have not been standardised. Therefore, mapping dictionaries for HIV isolates and HIV proteins were constructed. The isolate mapping dictionary was constructed from isolate names and their synonyms retrieved from HIV data within the UniProt [3] and Allie [14] resources (Table S2). The protein mapping dictionary was constructed from synonyms for genes, proteins, cleavage products, chains and domain names from the HIV data, also retrieved from the UniProt [3] and Allie [14] resources (Table S3). The highly-studied HIV group M subtype B HXB2 isolate was selected as the reference proteome and all HIV genes, proteins, cleavage products, chains and domain names, and their synonyms, were mapped onto the 9 proteins of the isolate (Gag, Gag-Pol, Env, Tat, Nef, Rev, Vif, Vpr, Vpu). This mapping included normalised start positions to correct the inconsistent numbering schemes of cleavage products, chains and domain (Figure 4B). Both dictionaries were further manually curated to improve upon the computationally retrieved mapping.
Several different experimental isolates are commonly used in HIV experiments. Each paper in the HIV literature dataset was scanned using the contents of the isolate mapping dictionary to identify the experimental isolate used in the study (Table S2). If no isolate information was retrieved, the Human immunodeficiency virus type 1 group M subtype B HXB2 isolate was set as the experimental isolate for the paper. The numbering of a mutation in the HIV literature can refer to the numbering of a protein, chain, domain or cleavage product, consequently, for a defined mutation numbering a conclusive mapping may not possible. Inconclusively mapped mutations text-mined from the HIV literature dataset were mapped to the HIV proteome using a co-occurrence based approach (Figure 4C). The co-occurrence based approach utilised the Reflect tool for automated tagging of biological entities to scan mutation-containing sentences for protein identifying terms from the protein mapping dictionary [4]. Each mutation's position was normalised to the protein-numbering scheme of the full-length protein based on the co-occurring protein identifying terms. If the mutation wildtype amino acid matched the amino acid at the normalised mutation position in the experimental isolate, the mutation was retained as a mapped mutation. In the cases where no information relating to the mutated protein was available, all HIV HXB2 proteins were scanned at their full-length protein, chain and domain levels. In the case of chain and domain, a displacement factor was applied to adjust the mutation's position and map the mutation to all possible positions in the proteome. A mutation mapping score (see below) was calculated for each putative mutation mapping and the top scoring mapping was retained as the mapped position of the mutation. In the cases where no matches to the experimental isolate proteome were found, the search was expanded to other commonly studied HIV isolates (Table S2).
For each mutation mapped using the above approach, a mutation mapping score, S, is calculated. The score is the function of three parameters: the probability of a match by chance; the number of mapped mutations in the paper; and the displacement from the reference protein position numbering scheme. The score ranges from 0 to 1, with values closer to 1 representing high confidence mapping of a mutation. The top-scoring mapping was retained as the mapped position of the mutation.
The score, S, is calculated as:where M is the number of mapped mutations in the paper, N is the total number of mutations mentioned in the paper, d is the distance between the defined mutation position and the mapped position, L is the sequence length of the protein, the values of a, b, and c are constants to weight the contribution of each parameter in the equation (a = 0.7, b = 0.15 and c = 0.15) and P is the probability that the mapped mutation would map to the protein by chance and is calculated as:where p is 0.05, the probability of matching an amino acid by chance given a 20 amino acid alphabet and assuming an equal frequency for each amino acid in the HIV proteome, and r is the number of mutations that have been mapped unambiguously to the protein.
The HIV Mutation Browser interface integrates information from several resources to increase the ease of interpretation of the available HIV mutation and mutagenesis data. Conservation information is displayed using multiple sequence alignments (aligned using the MAAFT algorithm [15]) retrieved from the HIV Subtype Reference Protein sequences from the Los Alamos National Laboratory (http://www.hiv.lanl.gov/). Structural information is displayed using structures of HIV proteins retrieved from the RCSB Protein Data Bank (PDB) [6]. Intrinsic disorder predictions for the proteins are calculated using the IUPred algorithm [16]. Enzymatic active sites, sites of post-translational moiety addition, sites of proteolytic cleavage and other sites of functional importance are retrieved from the UniProt resource [3]. Short linear motif interaction interfaces are retrieved from the ELM databases [17]. An up to date list of the ancillary information used and displayed is available on the HIV Mutation Browsers website.
|
10.1371/journal.ppat.1005042 | Human Non-neutralizing HIV-1 Envelope Monoclonal Antibodies Limit the Number of Founder Viruses during SHIV Mucosal Infection in Rhesus Macaques | HIV-1 mucosal transmission begins with virus or virus-infected cells moving through mucus across mucosal epithelium to infect CD4+ T cells. Although broadly neutralizing antibodies (bnAbs) are the type of HIV-1 antibodies that are most likely protective, they are not induced with current vaccine candidates. In contrast, antibodies that do not neutralize primary HIV-1 strains in the TZM-bl infection assay are readily induced by current vaccine candidates and have also been implicated as secondary correlates of decreased HIV-1 risk in the RV144 vaccine efficacy trial. Here, we have studied the capacity of anti-Env monoclonal antibodies (mAbs) against either the immunodominant region of gp41 (7B2 IgG1), the first constant region of gp120 (A32 IgG1), or the third variable loop (V3) of gp120 (CH22 IgG1) to modulate in vivo rectal mucosal transmission of a high-dose simian-human immunodeficiency virus (SHIV-BaL) in rhesus macaques. 7B2 IgG1 or A32 IgG1, each containing mutations to enhance Fc function, was administered passively to rhesus macaques but afforded no protection against productive clinical infection while the positive control antibody CH22 IgG1 prevented infection in 4 of 6 animals. Enumeration of transmitted/founder (T/F) viruses revealed that passive infusion of each of the three antibodies significantly reduced the number of T/F genomes. Thus, some antibodies that bind HIV-1 Env but fail to neutralize virus in traditional neutralization assays may limit the number of T/F viruses involved in transmission without leading to enhancement of viral infection. For one of these mAbs, gp41 mAb 7B2, we provide the first co-crystal structure in complex with a common cyclical loop motif demonstrated to be critical for infection by other retroviruses.
| Antibodies specifically recognize antigenic sites on pathogens and can mediate multiple antiviral functions through engagement of effector cells via their Fc region. Current HIV-1 vaccine candidates induce polyclonal antibody responses with multiple antiviral functions, but do not induce broadly neutralizing antibodies. An improved understanding of whether certain types of non-neutralizing HIV-1 specific antibodies can individually protect against HIV-1 infection may facilitate vaccine development. Here, we test whether non-neutralizing antibodies with multiple antiviral functions mediated through FcR engagement and recognition of virus particles or virus-infected cells can limit infection, despite lacking classical virus neutralization activity. In a passive antibody infusion-rhesus macaque challenge model, we tested the ability of non-neutralizing monoclonal antibodies to limit virus acquisition. We demonstrate that two different types of non-neutralizing antibodies, one that recognizes both virus particles and infected cells (7B2) and another that recognizes only infected cells (A32) were capable of decreasing the number of transmitted founder viruses. Further, we provide the structure of 7B2 in complex with the gp41 cyclical loop motif, a motif critical for entry. These findings provide insights into the role that antibodies with antiviral properties, including virion capture and FcR mediated effector function, may play in protecting against HIV-1 acquisition.
| The induction of HIV-1 broadly reactive neutralizing antibodies (bnAbs) by experimental vaccines is a critical goal of HIV-1 vaccine development efforts. However, bnAbs cannot be induced by existing HIV-1 vaccine candidates [1]. The RV144 ALVAC/AIDSVAX B/E HIV-1 vaccine efficacy trial demonstrated 31.2% estimated vaccine efficacy 42 months after the vaccination regimen was initiated [2]. Antibodies that mediated antibody dependent cell-mediated cytotoxicity (ADCC), or Tier 1 neutralizing antibodies in the presence of low envelope IgA antibodies, were identified as correlates of decreased transmission risk [3–6]. Thus, there is considerable interest in determining if commonly elicited ADCC-mediating, but non-broadly neutralizing antibodies against HIV-1 envelope have potential for protection against transmission [7,8].
Holl et al. have described non-neutralizing antibodies that bind to the immunodominant region (aa 579–613) of gp41 that can prevent HIV-1 infection of macrophages in vitro [9]. Others have demonstrated that these types of gp41 immunodominant antibodies bind to virions [10–12] and mediate ADCC [13,14]. Recently the HIV-1 gp41 immunodominant loop structure was determined for the first time in the context of the pre-fusion viral spike [15]. In that structure, the loop was disulfide bonded and buried under the trimer gp120 head groups and other elements of the observed pre-fusion gp41 fold [15]. However, to date, the only antibody to the immunodominant loop with its structure determined is that of unliganded mAb 3D6 [16,17].
Ferrari et al. [18], Guan et al. [19], Bonsignori et al. [20] and Veillette et al. [21,22] have described non-neutralizing gp120 antibodies that bind to a conformational epitope on Env in the first constant (C1) region of primary virus-infected CD4+ T cells were potent mediators of ADCC. The crystal structures of the C1 conformational A32-like antibodies N5-i5 and 2.2c were found to recognize overlapping epitopes formed by mobile layers 1 and 2 of the gp120 inner domain, including the C1 and C2 regions, but bind gp120 at different angles via juxtaposed VH and VL contact surfaces [23].
Mucosal transmission involves a series of events wherein HIV-1 or HIV-1-infected cells traverse genital tract mucus and epithelia, and subsequently infect epithelial or sub-mucosal CD4+ T cells [24,25]. Theoretically, antibodies that can bind virions at the initial stages of the transmission event [26] might be able to prevent virus movement across mucus and epithelium, thus preventing infection. Moreover, antibodies that mediate ADCC may be able to sensitize infected CD4+ T cells in the submucosa for natural killer (NK) cell killing [18,19], and abort transmission events. Alternatively, some studies have raised the issue that non-neutralizing antibodies might enhance infection [27–30].
BnAbs passively administered to rhesus macaques provide protection from SHIV challenges at mucosal surfaces [31]. Fc receptor interactions with natural killer (NK) cells are important for CD4 binding site bnAb neutralization in the macaque model [32]. However, non-neutralizing Env antibodies, when administered systemically or locally, have been reported to have only limited or no protective effect against vaginal SHIV challenge in rhesus macaques [30]. Moog et al. have demonstrated that local administration of gp41 immunodominant region non-neutralizing mAbs did not prevent infection with SHIV but did slow the onset of viremia in one of six animals and blunted the peak viremia in two others [13].
Enumeration of transmitted/founder (T/F) genomes has proved to be an important method of discerning the number of infecting viruses in various clinical settings [33,34], and has been used to monitor infection in macaques following SHIV challenge [30]. In this latter study, the mean number of T/F viruses was approximately 1 in most groups, thus, the contribution of non-neutralizing antibodies in limiting founder viruses compared to the control antibody was difficult to analyze [30]. Instead, this study showed a statistically significant enhancement in the number of founder viruses resulting in productive clinical infection in animals treated with the non-neutralizing CD4bs mAb, b6 [30]. Thus, the critical question remains as to whether antibodies that have an effector profile including recognition of virus particles and/or engaging FcR on effector cells can prevent HIV-1 transmission to any degree, or conversely, do such antibodies enhance infection?
In this study, we used T/F virus enumeration as a measure of relative protection from infection along with VL set point and CD4 preservation. We studied the activity of two human antibodies with characteristics of antibodies commonly induced by HIV-1 vaccine candidates, mAb 7B2 IgG1_AAA directed against the envelope gp41 immunodominant region [11,35] and the gp120 C1 mAb A32 IgG1_AAA, in the setting of high dose SHIV-BaL mucosal challenge. We show that while these non-neutralizing mAbs were unable to reduce the rate of productive infection in the high dose SHIV-BaL rhesus macaque challenge model, they reduced the number of T/F viruses involved in transmission events. Importantly, the challenge studies did not show evidence of antibody-mediated enhancement of virus infection.
The epitope specificities of two of the three HIV-1 Env mAbs examined in the rhesus macaque challenge model (i.e. A32 IgG [18] and CH22 IgG [36]) have already been previously described. Thus, here we characterized the epitope specificity of 7B2 IgG1. 7B2 IgG1 mAb recognizes a linear epitope in the gp41 immunodominant region with cross-clade reactivity (clades A, B, C, D, CRF1 and CRF2) as measured by peptide microarray [37] (Fig 1A). Specific interacting residues within this gp41596-606 11mer were examined with peptide alanine substituted mutants and surface plasmon resonance assays (Fig 1B). The footprint from epitope mapping demonstrated that residues in the 7B2 epitope were Cys598, Gly600, Leu602, Ile603, and Cys604. Ala substitution of Cys604 resulted in a peptide that bound with a higher peak response, yet had an off-rate approximately 2.8 times that of the Cys604 peptide. From these data we postulated that Cys604 mediated cyclization of the peptide that would be critical for the generation of a stable 7B2-peptide complex. Therefore we repeated the binding experiments of 7B2 using both longer and minimal epitope-containing wild-type gp41 peptides in standard (oxidizing) and reducing conditions (Fig 1C). In reducing conditions, the steady state binding of 7B2 mAb to both the longer and shorter peptides was at a relatively lower level with 100 to 150 fold increased off rates. These data suggested two primary factors for antibody binding, first, a disulfide-bonded, cyclical structure, and second, an induced fit is a likely factor in binding.
The co-crystal structure of 7B2 Fab bound to a gp41596-606 peptide was solved by molecular replacement and refined to a resolution of 2.7 Å (Table 1). The most prominent finding in the structure occurred in the central segment of the epitope-bearing peptide where a closed loop was displayed by virtue of the disulfide bond between Cys598 and Cys604. The cyclical conformation of the gp41 peptide was fostered by Gly600, where any side chain would diminish the backbone's ability to adopt the closed conformation and would clash with the side chain of Tyr32 in CDR-H1. Overall, the majority of the contacts between 7B2 antibody and the gp41 nominal peptide occurred through the base of CDR-H3 and the cyclical portion of the gp41 peptide (Figs 1D and S1). We concluded that the cysteine-cysteine bond and resulting loop conformation of the gp41 immunodominant region was indispensable for 7B2 binding.
The disulfide-linkage of the gp41 immunodominant loop in the 7B2 complex structure was similar to those seen in the chain reversal regions of other viruses [40–45]. However, the only component of the chain reversal motif that HIV-1 has in common with filoviruses and most retroviruses is the core, disulfide-linked motif, where it likely served the same function as in other retroviruses, specifically stabilization of the chain reversal region and a role in the transition state and formation of the 6HB [40,42,46].
The structure of the unliganded, wild-type gp41 immunodominant loop has been determined in solution by NMR [38,39], however a superposition of the solution structure upon that seen in our complex did not suggest a compatible conformation for strict docking (Fig 1E). Similar comparisons were seen between the coordinates of the gp41 peptide in our crystal structure and the coordinates of the SIV gp41 and HIV-1 gp41 solution structures [47,48]. The structure of the gp41 immunodominant loop has also been determined in the context of the Env pre-fusion trimer [15]. The immunodominant loop as seen in the BG505 SOSIP.664 structure resembled neither the conformation in the 7B2 complex structure nor the NMR solution peptide structures (Fig 1E). A radically different conformation of the polypeptide backbone atoms within the loop resulted in different positions for several key residues. The conformational variability of the immunodominant loop is evident in comparing these structures. Thus, we concluded that the wild-type 7B2 epitope-containing peptide was likely subject to induced fit though we could not rule out conformational selection of the disulfide-bonded structure as observed in other published cases of antibody-antigen recognition [49–53]. Moreover, the immunodominant loop appeared in an ordered and disulfide-bonded state in the SOSIP structure, but it was buried beneath glycoprotein. Thus, the immunodominant loop is inaccessible on pre-fusion spikes though it is present on Env stumps [54–59], so 7B2 and antibodies like it may capture virus and bind to virus-infected cells but fail to neutralize because they bind stumps and post-fusion structures.
We constructed 7B2 IgG1 optimized for binding to human FcRIII (CD16) by introducing alanine substitutions at positions 298, 333 and 334 (S298A, E333A and K334A) [60]. The ability of 7B2 IgG1_AAA to mediate antiviral function depends on both antigen recognition as well as engagement of FcγR on effector cells. We determined the binding of mAb 7B2 IgG_AAA to human FcγR1 (CD64), FcγRII (CD32) and FcγRIIIa (CD16) by SPR measurements (Fig 2A). 7B2 IgG1_AAA mAb bound with high affinity to FcγRI/CD64 (Kd = 49 nM) and with lower affinities to FcγRII/CD32a (Kd = 0.17 μM) and FcγRIIIa/CD16 (Kd = 1.1 μM) proteins. Among the three FcγR proteins, binding to FcγRIIIa (CD16) displayed fast kinetics, with an off rate that was almost two orders of magnitude faster when compared to FcγRI binding. The range of the observed differences in binding Kd and kinetics of 7B2 IgG_AAA mAb was consistent with previous report of IgG binding to the three classes of human Fc receptors [61–63]. As expected, the 7B2 IgG1_AAA Fab fragment control did not bind any FcRs. (Fig 2A). We next tested the binding of 7B2 IgG1 mAb to rhesus FcR (FcγRIIIa-1/-2 and FcγRIIIa-3). 7B2 IgG1mAb bound most avidly to rhesus macaque FcγR3A (Fig 2B). 7B2_AAA bound to rhesus FcγR3A (both allelic variants) with higher affinities (due to slower off-rates) than 7B2_SEK, which contains Fc region aa S298, E333 and K334 (Table 2). CH22 and A32 mAbs bound to both human FcγR3A-1 and FcγR3A-3 with similar affinities (Fig 2C and 2D, Table 2).
To confirm that 7B2 IgG1_AAA Fc bound to rhesus FcRs on monocytes, we coated the SP400 gp41 immunodominant region peptide on sheep red blood cells (SRBC), then coated the SP400- SRBC with 7B2 IgG1_AAA or a negative control antibody. Next we compared the abilities of human and rhesus monocytes to phagocytose the 7B2 IgG1_AAA-opsonized SRBC. Rhesus monocytes were able to phagocytose 7B2 IgG1_AAA mAb-coated SRBC equally as well as human monocytes (55 +/- 10% of rhesus monocytes with ≥2 internalized SRBC; 57 +/- 7% of human monocytes with ≥2 internalized SRBC).
It was important to determine if the 7B2 IgG1_AAA mAb could capture infectious HIV-1 virions using an assay that could differentiate between infectious and noninfectious virus using both viral RNA and an infectious virus readout [11,64]. We previously reported that whereas mAb 2G12 captured the majority of infectious virus, mAb 7B2 IgG1_AAA captured both non-infectious as well as infectious virions [11,12,64]. However, mAb 7B2 IgG1_AAA could not, in any experiment, capture all infectious virions. Thus, mAb 7B2 IgG1_AAA captured a subset of both infectious and non-infectious virus particles. Here we tested whether mAb 7B2 IgG1_AAA could capture the infectious CCR5- tropic SHIV virions (SHIV-BaL and SHIV-SF162P3). As with HIV-1, 7B2 IgG1_AAA was able to capture a portion of infectious virions of both SHIVs (Fig 3A and 3B). In contrast, A32 mAb did not significantly capture any virions (Fig 3C and 3D).
We next asked whether recognition of HIV-1 virions by 7B2 IgG1_AAA mAb was dependent on CD4 binding. Using a virion capture assay that measures capture of virus particles without an infectious readout [65] we determined whether sCD4 binding to virus makes it more susceptible to capture by 7B2 IgG1_AAA mAb. We found that mAb 7B2 IgG1_AAA capture of HIV-1 SF162.B, BG1168.B, 6535.B, 6846.B and CAP 45.C virions, was augmented in the presence of sCD4 (S1 Table). Similarly, mAb 7B2 IgG1_AAA effectively captured parental HIV-1 SF162.B and BaL.B as well as SHIV-SF162P3 and SHIV-BaL (Fig 3C and 3D) in the presence of sCD4.
The ability of 7B2 IgG1_AAA mAb to prevent infection of human colorectal tissue was assessed using an established ex-vivo rectal explant model [66]. 7B2 IgG1_AAA mAb had no direct impact on infection of colorectal explants (Fig 3E). 7B2 IgG_AAA IgG1 mAb did reduce dissemination of infection from dendritic cells that emigrate from the tissue during the first 24 hours of culture on incubation with CD4+ indicator T cells (Fig 3E). However, 7B2 and A32 IgG1 mAbs did not inhibit infection of monocyte derived DC or DC mediated trans-infection of co-culture with T cells (S2 Fig).
MAb 7B2 IgG1_AAA when expressed in both the SEK and AAA IgG1 backbones effectively neutralized HIV-1 BaL with ID50s of 0.05 μg/ml in peripheral blood monocytes differentiated into macrophages (Fig 4A). The negative control, anti-respiratory syncytial virus mAb, palivizumab, had no inhibitory effect at 100 μg/mL. 7B2 IgG1 mAb also neutralized SF162 (0.2 IC90 for both the SEK and _AAA mAbs) and TV-1 (1 IC90 for SEK and 0.5 for _AAA) (Fig 4B). In another assay, 7B2 IgG1 mAb was confirmed to mediate virus inhibition of infection in macrophages [9] (Fig 4C). The negative control IgG and A32 mAb at 50 μg/ml did not block infection (107%, 109% of control infection), respectively. Moreover, 7B2 IgG1 mAb could mediate virus inhibition in tissue derived peritoneal macrophages (Fig 4D) and in antibody-dependent cell mediated virus inhibition assays (ADCVI) (S2 Fig).
To determine if 7B2 IgG1_AAA mAbs could coat virus infected targets and arm natural killer (NK) cell effectors for antibody dependent cellular cytotoxicity (ADCC), we first confirmed the ability of 7B2 IgG1 SEK, 7B2 IgG1_AAA and A32 IgG1 mAb to bind to the surface of HIV-1 B.BaL infected CD4+ T cells (Fig 5A and 5B). Palivizumab and A32_AAA were included as negative and positive controls, respectively. Both the SEK and _AAA versions of mAb 7B2 IgG1 had similar ability to bind the surface of HIV-1 B.BaL infected cells. Since our challenge stock was SHIV-BaL, we next assayed for the ability of mAb 7B2 IgG1_SEK (and 7B2 IgG1_AAA to mediate ADCC against HIV-1 B.BaL-infected CD4 T cells. We found that 7B2 IgG1_AAA as well as A32_AAA IgG1 mAb, and the neutralizing V3-region specific mAb CH22 IgG1_AAA could indeed mediate killing of HIV-1 B.BaL-infected CD4+ T cells. In contrast, when mutation of the S298A as well as E333A and K334A in the AAA form was reverted back to SEK in the Fc domain of the mAb, it reduced the ability of the 7B2 IgG1 mAb to mediate ADCC (Fig 5C). Therefore we used the 7B2 IgG1_ AAA mutant for passive infusion into macaques to determine the ability to protect against SHIV-BaL intrarectal challenge.
To determine if the human HIV-1 specific mAbs can engage the rhesus FcR on NK cells, we examined binding of the mAbs to CD16 on rhesus NK cells by flow cytometry (Fig 6). Peripheral blood mononuclear cells (PBMCs) were isolated at a pre-infusion time point from each of the rhesus macaques enrolled in the passive infusion study (both palivizumab and HIV-1 mAb). Binding of 7B2 mAb, A32 mAb and CH22 mAb to these cells was measured to ensure that there was intact FcR-Ab engagement in these animals. In all rhesus macaques tested, there was substantial binding of the infused mAb to rhesus NK cells.
In vivo PK studies were performed prior to passive protection studies for all antibodies to determine the concentrations and the half-lives of the antibodies in circulation and at the mucosal sites (Table 3; Fig 7). Two rhesus monkeys that were infused once with 7B2 IgG1_SEK IgG1 at 30 mg/kg had ~10 μg/ml of 7B2 IgG1_SEK in the rectal secretions. For the PK study using 7B2 IgG1_AAA IgG1, the Ab was administered to three rhesus monkeys twice at 50 mg/kg at 0 and 48 hours [67]. This resulted in a peak concentration of 30 μg/ml of 7B2 IgG1_AAA in the rectal secretions after the first infusion and ~90 μg/ml after the second infusion.
We also tested the non-neutralizing, A32 IgG1_AAA mAb that is one of the more potent of the ADCC-mediating antibodies, and is known to bind the surface of virus infected CD4+ T cells [18,20,65], but does not bind to Env on virions, and thus is unable to capture infectious virions. To address the question of whether the commercially prepared anti-RSV control antibody, palivizumab, may, in some way have affected transmission, we produced another control antibody, the influenza neutralizing anti-hemagglutinin IgG1 mAb CH65 [68], using the exact same protocol and methods as used in the production of 7B2 IgG1_ AAA and A32 IgG1_ AAA. Finally as a positive control, we produced CH22, an anti-V3 mAb derived from the RV144 HIV-1 vaccine efficacy trial [36] that neutralized SHIV-BaL in vitro in the TZM-bl neutralization assay at an IC50 of 1.9 μg/ml, in contrast to 7B2 IgG1_AAA and A32 IgG1_AAA that did not neutralize SHIV-BaL (>50 μg/ml). With these new reagents, we completed PK studies to determine the time of peak mAb concentration at the mucosal sites to perform the challenge studies.
In the passive protection trial 7B2 IgG1_ AAA was administered at 50 mg/kg at 0 and 48 hours in six Indian-origin rhesus monkeys and the monkeys were challenged with 1 ml of SHIV-BaL (2×105 TCID50/ml) via the intrarectal route at time 56 hours, the time of the peak antibody concentration post-second infusion. Another six rhesus monkeys received the control antibody palivizumab at the same dose and times as 7B2 IgG1_ AAA and were challenged at the same time based on the antibody concentration post-infusion in PK studies. We found that infusion of mAb 7B2 IgG1_AAA had no effect on peak viral load or on viral load at day 42 post-challenge (Fig 8A). Moreover, there was no significant impact on CD4 counts (Fig 8B). Thus, infusion of 7B2 IgG1_AAA resulting in peak mucosal antibody levels of 90 μg/ml at the time of challenge (Table 3, Fig 7), yet had no protective effect on SHIV-1 acquisition or control of viremia following high dose challenge with virus.
We then carried out studies using four new groups of Indian-origin rhesus macaques with 6 animals in each group, and each group infused with 50 mg/kg antibody at times 0 and 48 hours. The first group was infused with A32 IgG1_AAA ADCC antibody, the second group with the positive control CH22-IgG1_AAA V3 neutralizing antibody, the third group with negative control antibody, palivizumab, and the fourth group with new negative control anti-influenza neutralizing antibody, CH65. In order to insure the SHIV-BaL challenge coincided with peak plasma levels of mAb, monkeys infused with A32 IgG1_ AAA mAb and palivizumab control antibody were challenged intra-rectally with SHIV-BaL immediately following the first infusion whereas, monkeys infused with CH22 mAb and CH65 control mAbs were challenged at 60 hours using the same virus and same route. SHIV-BaL challenges were performed at the time of peak concentration of mAbs at the mucosal sites for each experimental mAb based on the findings from the PK studies.
Infusion of antibody A32 IgG1_ AAA into rhesus macaques had no effect on clinical acquisition of SHIV-BaL infection with 6 of 6 animals infected or on viral load or CD4 counts (Fig 8C and 8D). In contrast, infusion of the positive control V3-loop neutralizing antibody, CH22 IgG1 mAb, resulted in prevention of infection in 4 of 6 monkeys (Fig 8E). Similarly infusion of the negative control antibodies palivizumab and CH65 IgG1 (influenza neutralizing antibody) did not prevent infection in any of the monkeys (Fig 8C–8E).
Keele and others have shown that in ~80% of cases of primary HIV-1 infection, one T/F viral genome (range 1–6) established productive clinical infection [33,69,70]. In men who have sex with men (MSM), one T/F genome (median 1, range 1–12) accounted for approximately 60% of cases, and in injection drug users this proportion fell to about 40% of cases (median number of T/F viruses 3, range 1–16) [34,71,72]. The env diversity present in the SHIV-BaL challenge stock (mean 0.3%, range 0–0.7%) is substantially less than the HIV-1 env diversity that is typically found in chronically infected humans (>>1%) [73], but it is nonetheless sufficient for distinguishing discrete T/F genomes (S3 Fig). We estimated the minimum number of T/F genomes responsible for productive SHIV-BaL infection (see Materials and Methods) in the 18 control animals to range from 1–27 with a median of 6.5 (Table 4). In contrast, after infusion of mAb 7B2 IgG1_AAA, the estimated numbers of T/F viruses was reduced 58% from the control Ab median of 6.0 to a median of 2.5 with a range of 1–5. Given that 60 sequences per animal were used to estimate numbers of T/F variants in the 7B2 IgG1_AAA treated and matched control animals, we could be 95% confident of sampling every variant that was present at >5% prevalence in each animal. Thus, the results revealed a significant reduction in the numbers of T/F viruses in the 7B2 treated animals compared with the six matched control animals (Mann-Whitney Rank Sum Test, p = 0.01) and when compared with the larger control group of 18 matched and unmatched control animals (p = 0.001) (Table 4).
We next evaluated the number of T/F viruses in control mAb and A32 IgG1_AAA and CH22 infused monkeys (Table 4). The median number of T/F viruses in the first palivizumab control group performed with the 7B2 IgG1_AAA trial was 6, and the median number of T/F viruses in the second palivizumab control group was also 6. The median number of T/F viruses in the third control group infused with the new control antibody CH65 IgG1 was 10 with a range of 1 to 27 (Fig 9). Thus, among the 18 control animals, the median number of T/F variants was 6.5, and the difference in median T/F numbers between palivizumab and CH65 control antibody treated animals was not significant (p = 0.12, Mann-Whitney rank-sum test).
The median number of T/F viruses with the non-neutralizing ADCC-mediating A32 IgG1_AAA mAb was 3, a 50% reduction compared with the palivizumab control group where the median number of T/F viruses was 6 (p = 0.03, Mann-Whitney rank sum test) (Table 4). In the positive control CH22 antibody group, there were a median of 2 T/F viruses in each of the two infected animals, compared to a median of 10 founder viruses in the control group (p = 0.01 Mann-Whitney rank sum test) (Fig 9). Given a mean of 37 sequences per animal in the A32 IgG1_AAA study and a mean of 40 sequences per animal in the CH22 IgG1_AAA study, we are 95% confident that variants >8% prevalence in each animal are represented in the T/F enumeration. There was no evidence of selection pressure on the breakthrough viruses nor impact on the neutralization sensitivity of these viruses in animals treated with CH22 (S4 and S5 Figs). Additionally, for A32 IgG1_AAA and 7B2 IgG1_AAA passive infusion, there were no phylogenetically corrected signatures that were significant with a q value < = 0.3 that might suggest a selective pressure or sieve effect on breakthrough viruses. The lack of selection pressure on the antibody contact residues are shown in S6 Fig. The lack of evidence of virus escape from these mAbs is not surprising, since mutation in the highly conserved residues recognized by the A32 and 7B2 mAbs is likely destabilizing to the HIV-1 envelope glycoprotein and would alter virus fitness, and in the case of the A32 epitope, might decrease CD4 and co-receptor binding [74].
Thus, the high dose SHIV-BaL intra-rectal challenge model of rhesus macaque with an infection dose of ~105 TCID50 was able to show protection by a V3 loop neutralizing antibody that neutralized the SHIV-BaL challenge stock. Using the same challenge model, both of the ADCC mediating antibodies 7B2 IgG1_ AAA and A32 IgG1_ AAA were not able to prevent productive clinical infection nor reduce viral load, but both mAbs were able to reduce the number of founder variants by ~50%. Importantly, none of the anti-HIV Env mAbs tested showed any evidence of enhancement of viral transmission.
In this study we have shown that two non-neutralizing antibodies, the gp41 targeted, ADCC-mediating, virus-capturing antibody 7B2 IgG1_AAA and the gp120 ADCC-mediating antibody A32 IgG1_AAA, each with mutations that enhance FcR binding [60], were able to limit the number of T/F viruses of SHIV-BaL in a high dose intra-rectal challenge model in rhesus macaques. While both antibodies did not reduce the frequency of clinical infection by SHIV-BaL, they did reduce the number of founder viruses by ~50%. In 7B2 IgG1_AAA treated animals, the median number of T/F viruses was reduced from 6 to 2.5 (p = 0.01). In A32 IgG1_AAA treated animals, the median number of T/F viruses was reduced from 6 to 3 (p = 0.03). In animals treated with the positive control V3 loop antibody CH22, virus transmission was eliminated altogether in 4 of 6 animals and the numbers of T/F viruses in the remaining two animals reduced to two compared with a median of 10 in the control group (p = 0.01).
There are caveats to the interpretation of these findings. First, the numbers of animals in treated and control groups were small (n = 6 for each). Second, the diversity in the SHIV-BaL challenge stock is limited (median env diversity 0.3%; range 0–0.7%), so distinguishing between discrete T/F genomes and genomes that acquire shared mutations post-transmission can be problematic. Third, the in vivo error rate of the SIV reverse transcriptase and the number of virus generations from infection to sampling must be estimated in distinguishing distinct T/F viruses from evolved viruses; we used previous empirical data and mathematical modeling of early viral replication dynamics [33,73] to estimate the frequency of mutations that might be expected in the first 10–21 days of infection to guide the identification of T/F genomes (see Methods). Fourth, APOBEC-mediated G-to-A hypermutation can confound T/F lineage analysis [33]; in the analysis presented we deleted G-to-A hypermutated sequences from our analyses [30,33,73]. Of note, the results regarding reduction in the number of transmitted strains in the presence of the non-neutralizing antibodies were supported when all G-to-A APOBEC motifs were removed from the analysis [33,75]. Fifth, our measurements of T/F genomes are point measurements that depend on depth of sampling for their sensitivity. They are not corrected for variants that might be present but not observed, and they are not expressed as point estimates with confidence limits. Alternative approaches to estimating numbers of T/F genomes that account for these limitations are in development (L. Blair, B. Korber, T. Bhattacharya).
SHIV-BaL is a CCR5-tropic tier 1 virus that is relatively easy to neutralize and is not a pathogenic SHIV. This SHIV was used to establish a sensitive model for non-neutralizing antibodies and to determine if any protection was present compared to previous studies using SHIV-BaL that have shown full protection with bnAb infusions [67]. While the challenge dose was high, it was the least amount of virus that caused infection of 100% of challenged animals (2.0 x105 TCID50). Previous work has demonstrated that non-neutralizing antibodies administered locally or systemically were not as potent in protecting against SHIVs in rhesus macaques as antibodies that neutralize the challenge SHIVs. The present study sought to define the protective capacity of antibodies that are not able to neutralize HIV-1 in the TZM-bl assay but can mediate ADCC and other types of anti-HIV-1 immune effector functions. One caveat of this study is that the amount of virus present in the challenge stock, 2.9 x 109 copies/ml (2.0 x 105 TCID50) far exceeds the average amount of virus in semen of untreated HIV-1 infected individuals which is 0.426 x 104 copies/ml (range 0.01–6.9 x 104) [76]. If this seminal viral load reflects the amount of virus responsible for natural infection in the RV144 clinical trial, in which ADCC responses correlated with lower risk of infection in vaccinees with low anti-Env IgA responses, then we used a SHIV challenge dose that was 5 orders of magnitude greater than what may be responsible for natural HIV-1 infection. Therefore, it is possible that, in order to achieve infection in 100% of the control animals, we were considerably above the challenge threshold for the amount of virus that can be controlled by these non-neutralizing Abs. Thus, transmitted/founder virus enumeration was a more sensitive approach to study the impact of these non-neutralizing antibodies on transmission in this high dose challenge model.
A vaccine induced antibody response is unlikely to induce a single antibody specificity at the plasma concentrations present during the time of challenge (1,324–2,034 μg/ml); however, the mucosal antibody concentrations present in this study at the lower end of the range (0.96–27.4 μg/ml) are likely concentrations to be induced by vaccination. Whether or not the level of mucosal antibody concentrations present in individual animals in this study played a role in protection or reduction of founder viruses is unknown, but is worth further study to better understand the antibody concentrations needed for protection.
Antibodies that neutralize HIV-1 in conventional neutralization assays protect in passive protection trials. However, these neutralizing antibodies must have breadth and potency to be effective when passively administered. Administration of bnAbs clearly demonstrates the relevance of mAb breadth and ability to neutralize the few R5 SHIVs that are available for testing as challenge isolates. Hessell et al. have shown that the combination of conventional neutralization activity and IgG FcR-mediated activity such as ADCVI provides optimal protection in the setting of passive protection trials [32]. However, they have also demonstrated that nonfucosylated antibodies with better ADCC function were not better at protection[77]. Thus, additional studies are needed to determine whether the most likely attributes of protective IgG antibodies are to have conventional neutralizing activity with sufficient breadth to be clinically relevant and to potentially be able to mediate FcR dependent anti-HIV-1 activities. Moreover, additional passive infusion studies utilizing antibodies engineered to have optimal FcRn binding [78] to improve the antibody half-life can provide a way to examine the role of different non-broadly neutralizing antibody specificities in protection in a low dose mucosal challenge study.
MAb 7B2 IgG1_AAA and similar gp41 antibodies capture a subset of infectious virions in a CD4–dependent manner, mediate ADCC, and neutralize BaL in macrophage cultures. In contrast, mAb A32 IgG1_AAA does not capture virons, does not mediate macrophage neutralization, but is a very potent mediator of ADCC [18]. There is some controversy on the role of macrophages either as a virus reservoir or as part of the initial foci of infection after rectal or vaginal challenge. However, we broadly tested Fc-mediated activity of these mAbs, including macrophage neutralization and phagocytosis to characterize their potential effector function in vivo. That both of these antibodies (but not control antibodies) limited the number of founder SHIV-BaL viruses suggests that if viruses with traits like the tier 1 SHIV-BaL are indeed involved in HIV-1 transmission, then these common types of dominant gp41 and gp120 antibodies may play a role in protecting against HIV-1 transmission. Burton et al. found a suggestion of protection in 2 of 5 challenged animals using the gp41 non-neutralizing mAb F240, but the result was not statistically significant [30]. The authors counted T/F viruses in this study but there was no reduction in T/F viruses by F240 IgG mAb in infected animals [30]. Recently, in a live SIV vaccine (SIVmac239delta Nef) model, gp41 reactive antibodies, with properties similar to F240 mAb were shown to correlate with protection [79]. Also in the Burton study, the CD4 binding site non-neutralizing Ab b6 IgG mAb appeared to result in enhancement in the numbers of founder viruses compared with control animals [30]. This was not seen in the present study for 7b2 or A32 mAbs.
Others have demonstrated that the ability of an antibody to block virion transcytosis through epithelia is a predictor of protection both in the setting of subjects who are exposed and uninfected, and in the setting of vaccination [80]. The 7B2 IgG1_AAA antibody has recently been reported to block transcytosis in vitro [81], although in our mucosal explant models, neither A32 IgG1_AAA nor 7B2 IgG1_AAA blocked HIV-1 infection in the explant model in vitro.
It has previously been recognized that the 7B2 IgG1_AAA mAb can capture both infectious and non-infectious virions [11] and that the ability to capture virions does not correlate with ability of an antibody to neutralize HIV-1 [64]. MAb 7B2 IgG1_AAA most likely recognizes gp41 stumps from which gp120 has been shed [54] or some other non-native form of envelope on the virion. Since virions can contain a mixture of native and non-native Env forms [54–57], it was conceivable that an antibody like 7B2 IgG1_AAA could provide some measure of protection by binding to non-native Env forms on infectious virions. We do not know whether virus capture, ADCC, both or neither was involved in limiting founder viruses by 7B2 IgG1_AAA. However, A32 IgG1_AAA also limited founder viruses. Thus, we hypothesize that the effector mechanism was ADCC or a mechanism other than virus capture and retardation of virus transport across epithelia. One reason for this limited effect on prevention of founder virus number by 7B2 IgG1_AAA may be that although 7B2 IgG1_AAA does bind to infectious virions, it does not bind to all infectious virions (Fig 3). We recently demonstrated that although mAb 7B2 IgG1 cannot capture all infectious virus, the combination of two mAbs, 7B2 IgG1 and the V2 antibody CH58 IgG1 mAb, increased the capture of infectious virions [12].
It was of interest that our positive control, CH22 mAb, a V3 mAb isolated from an RV144 vaccinee [36] did neutralize SHIV-BaL (1.9 μg/mL IC50/10.4 μg/mL IC80), mediated ADCC (peak activity at 50 μg/mL), and protected 4 of 6 macaques from SHIV-BaL infection. V3 specific antibodies, in the context low specific Env IgA responses, were associated with a reduced risk of infection in the RV144 trial [82]. However, the infecting viruses in RV144 were Tier-2, while CH22 mAb only neutralized tier 1 viruses. Thus the protective capacity of CH22 mAb for the RV144 setting remains unclear. Moreover, in contrast to the protection of 4/6 animals by the linear V3 specific CH22 antibody, PGT121, a glycan V3 bnAb protected all of the animals [83]. The difference in the protective capacities of PG121 and CH22 are likely due to the differences in their epitope recognition and breadth/ potency for primary isolates (i.e. PGT121 recognizes glycans in the V3 region and is a broadly neutralizing antibody, unlike CH22 that is against a linear V3 and neutralizes only Tier 1 viruses).
We also show the co-crystal structure of gp41 immunodominant region mAb 7B2 Fab with its gp41 nominal epitope. The most prominent finding in the structure was the presence of a cysteine-cysteine loop with intact disulfide bond in the center of the gp41 Env contact region. Most recently, the BG505 SOSIP.664 trimer structure has showed partial structural detail of the pre-fusion gp41, in particular an ordered, disulfide-bonded immunodominant loop structure that was buried inside the glycoprotein complex [15]. This disulfide-bonded immunodominant loop structure of HIV-1 gp41 is involved in gp41-gp120 interactions [47,84–86] and is homologous to similar regions in other enveloped viruses and key structural elements in those virus infection mechanisms [48,86]. The extended conformation and hydrophobic nature of the residues at the tip of 7B2 CDR-H3 are reminiscent of the properties of gp41 membrane proximal broadly neutralizing antibodies, specifically that they tend to bear long, highly mutated CDR-H3s [1,87,88]. However, 7B2 was not polyreactive [88], does not have its epitope in the membrane proximal external region [89], and the CDR-H3 is not exceptionally long [1,87,90]. The data in this report demonstrate an exposed gp41 structure at amino acids 596–606 of Env that is exposed both on gp41 on HIV-1 virions and on virus-infected cells. This is the first structure of an antibody complex that shows binding at the functionally important disulfide bond in the chain reversal region of a retroviral entry protein. The cyclical loop motif has been demonstrated to be critical for function in other retroviruses, in particular to help constrain the local structure of the chain reversal region and to participate in formation of the hairpin structures during the fusion process.
Vaccines that induce immunodominant gp41 antibody responses, targeting the same epitope as 7B2 IgG1_AAA mAb, are in HIV-1 clinical trials [91]. It is important to note that these vaccines induce this antibody specificity in the context of a polyclonal antibody response. A recent non-efficacious HIV-1 vaccine trial (HVTN 505) [92] used a gp140 that had the C-C immunodominant loop region bound by 7B2 IgG1_AAA deleted, and thus could not induce 7B2 IgG1_AAA-like mAbs. For those vaccine regimens that do induce antibody specificities similar to 7B2 mAb [91] (Seaton et al., in preparation), our study can shed light on potential antiviral functions induced by HIV-1 vaccination. In our current study, non-neutralizing antibodies with defined specificities and functional characteristics were tested individually and some restriction in the number of founder viruses was observed. Future proof of concept studies in NHP can combine different Abs such as A32 IgG1_AAA (conformational C1), 7B2 IgG1_AAA (linear gp41) and CH58 IgG1_AAA (V2) [65] to determine whether a well-defined polyclonal mixture of antibodies can improve protection against mucosal infection.
Rhesus macaques (Macaca mulatta) were housed at BIOQUAL, Inc. (Rockville, MD), in accordance with the standards of the American Association for Accreditation of Laboratory Animal Care. The protocol was approved by the BIOQUAL's Institutional Animal care and Use Committee under OLAW Assurance Number A-3086-01. Bioqual is IAAALAC accredited. This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health (NIH) and with the recommendations of the Weatherall report; “The use of non-human primates in research”. All procedures were performed under anesthesia using ketamine hydrochloride, and all efforts were made to minimize stress, improve housing conditions, and to provide enrichment opportunities (e.g., social housing when possible, objects to manipulate in cage, varied food supplements, foraging and task-oriented feeding methods, interaction with caregivers and research staff). Animals were euthanized by sodium pentobarbital injection in accordance with the recommendations of the panel on Euthanasia of the American Veterinary Medical Association.
Human peripheral blood mononuclear cells and peritoneal macrophages from HIV-1 negative individuals were collected with IRB approval by the Duke Medicine Institutional Review Board for Clinical Investigations (Protocol Pro00006526, Pro00000873, Pro00009459) and from healthy human subjects enrolled in the UC Irvine Normal Blood Donors Program (HS #2002–2430). All subjects were consented following 45 CFR 46 and written informed consent was obtained by all participants. No minors were recruited into this study. Additionally, monocytes were purified from blood packs purchased from the blood bank, with written informed consent from the donors. The approval to collect and store tissue was issued by the Imperial Tissue Bank with reference number Med_RS_11_014. Approval to use the tissue was given under Project number R11021. Approval for this project was granted by the Tissue management Committee at Imperial College Healthcare Trust in July 2011, and ethics thus conveyed through this process by MREC Wales, reference number 07/MRE09/54.
The 7B2 IgG monoclonal antibody was isolated from a HIV-1 chronically infected subject using Epstein-Barr (EB) virus B cell transformation and heterohybridoma production [35]. To produce recombinant 7B2 IgG1 antibody, the variable regions of immunoglubulin heavy and light chain (VHDJH and VLJL) genes of 7B2 IgG1 were isolated from the 7B2 IgG1 cell line by RT/PCR using the primers and methods as described [93], where the VH and VL gene rearrangements expressed by 7B2 IgG1 were determined by sequence analysis and annotated using the IMGT database.
For structural studies, after affinity capture the 7B2 Fab was further purified via gel filtration chromatography using a HiLoad 26/60 Superdex 200pg 26/60 column at 1 ml/min with a buffer of 10 mM Hepes pH 7.2, 50 mM NaCl, 0.02% NaN3. Peak fractions of Fab were pooled and exchanged into 50 mM Hepes pH 7.5 via five dilute/concentrate cycles in an UltraFree 4 ml 10K MWCO, then run over a cation exchange column (Mono S 5/50 GL) at 1 ml/min. At pH 7.5, the Fab passes through the column and excess light chain binds. The excess light chain is later eluted with 50 mM Hepes pH 7.5, 1 M NaCl. At lower pH values, the Fab binds the column and can be eluted with a gentle salt gradient, e.g. 0–15%. Peak fractions of Fab were pooled and exchanged into 10 mM Na Hepes pH 7.5, 50 mM NaCl, 0.02% NaN3 via four cycles of dilution-concentration and brought to a concentration >20 mg/ml for subsequent dilution.
Two forms of recombinant IgG1 heavy chain Fc were produced, one with the wild type IgG1 sequence termed as 7B2 IgG1_ SEK, and the other termed 7B2 IgG1_AAA. 7B2 IgG1_SEK contains Fc region aa of S298, E333 and K334 and 7B2 IgG1_AAA contains the Fc region aa of S298A as well as E333A and K334A, amino acid mutations previously reported to augment antibody ability to bind to FcRIIIa and to augment antibody ADCC activity [60]. The 7B2 IgG1_AAA mAb was expressed in CHO cells (Catalent, Somerset, NJ).
The A32 IgG1 monoclonal antibody was isolated from a chronically infected HIV-1 infected patient using Epstein-Barr virus B cell transformation and heterohybridoma production [94]. To produce recombinant A32 IgG1 antibody, the A32 IgG1 VHDJH and VLJLgenes were isolated from the cloned A32 IgG1 cell line by RT/PCR as described [93]. The VHDJH and VLJL gene rearrangements of A32 IgG1 were determined by sequence analysis. The recombinant A32 antibody was expressed in CHO cells (Catalent, Somerset, NJ) as IgG1_AAA to be optimized for binding to FcRIIIa [60]. Palivizumab, a commercial anti-respiratory syncytial virus antibody (Medimmune, Inc, Quakertown, PA), as well as recombinant CH65, an anti-influenza hemagglutinin mAb [68], were used as negative controls in the passive protection studies. Finally, CH22 neutralizing mAb reactive with the V3 region of HIV-1[36] was produced as a recombinant CH22 IgG1_AAA mAb in CHO cells (Catalent, Somerset, NJ). As a control for CH22, the influenza mAb CH65 [68] was produced as recombinant IgG1_AAA in CHO cells using the same technologies for 7B2_AAA, A32_AAA and CH22_AAA mAbs.
MAb binding 7B2 IgG_AAA to peptides and gp140 proteins was performed by ELISA, by binding antibody multiplex assays [26,95] and by surface plasmon resonance (SPR) assays [96,97]. Epitope mapping of mAbs was performed by peptide microarray microarray [37,98]. Gp41596-606 peptides with N-terminal biotin tags were commercially synthesized for SPR experiments. They were the wild type peptide and each residue substituted with Ala in turn. An additional gp41596-606 peptide was synthesized with an acetylated N-terminus and amidated C-terminus for crystallography. SP400 peptide (gp41579-622) was synthesized solely for SPR [26].
Surface plasmon resonance (SPR). SPR binding assays were performed on a BIAcore 3000 (BIAcore Inc, Piscataway, NJ) instrument at 25°C and subsequent epitope mapping was carried out using a BIAcore 4000 instrument. Data analyses were performed using the BIAevaluation 4.1 software (BIAcore) as previously described [88]. Each residue of the 11mer WGCSGKLICTT (gp41596-606) was mutated to Ala and the peptides were made with N-terminal Biotin tag to facilitate their coupling to a BIAcore streptavidin chip. The biotinylated 11mer and its alanine-substituted variants were screened for binding by 7B2 (at 20 μg/ml). Binding responses of an irrelevant antibody Synagis was used to subtract out responses due to non-specific interactions.
For the gp41-binding experiments, both 7B2 and the negative control Synagis at a 50 μg/ml concentration were flowed at a 50 μl/min rate over SP400 tetramer (~4000 RU) and recombinant gp41 MN (~5700 RU) immobilized on a CM5 chip. The Gp41MN is a commercially available protein (Immunodiagnostics, Inc. product number 1091) and SP400 is a gp41-derived peptide with sequence RVLAVERYLRDQQLLGIWGCSGKLICTTAVPWNASWSNKSLNKI which was commercially synthesized (CPC Scientific) and tetramerized with streptavidin tags. For the experiments testing reducing conditions, biotinylated peptides were coupled to a BIAcore streptavidin chip. The peptides were initially screened for binding by 7B2 (at 10 μg/ml) at a flow rate of 20 μl/min with a PBS running buffer. Binding response to biotinylated SP62WT (a peptide containing gp41 MPER sequence) was used to subtract out non-specific interactions. The BIAcore 3000 was then primed with PBS, 20 mM DTT running buffer. The peptides were screened for 7B2 (at 10 μg/ml) at 20 μl/min with the PBS, 20mM DTT running buffer. Non-specific interactions were accounted for using the previously stated method.
Peptides of sequence WGCSGKLICTT (gp41596-606) were commercially synthesized with acetylated N-termini and amidated C-termini in both reduced and disulfide-bonded forms (CPC Scientific). Lyophilized peptides were solubilized to 100 mg/ml in DMSO with no further dilution in aqueous solution prior to mixing with 7B2 Fab in a 1:3 Fab:peptide molar ratio. The complex was then diluted with 10 mM Na Hepes pH 7.5, 50 mM NaCl buffer to a total protein concentration of 12.5 mg/ml. Fab-peptide complexes were screened against various sparse matrix screens. Plates were incubated at 20°C. Crystals of the Fab with cyclical peptide were observed within two weeks in the Hampton Research PegRX screen condition H9, a solution of 5% 2-propanol, 0.1 M citric acid pH 3.5, 6% PEG 20K. Crystals were replicated in a fine matrix screen about the hit condition, the optimal of which was 4–6% 2-propanol, 0.1 M citric acid pH 3.5, 8% PEG 20K.
Crystals were cryoprotected with reservoir solution supplemented with 30% ethylene glycol immediately prior to cryocooling. Data were collected at SER-CAT BM at 1 Å wavelength. Data were reduced in HKL-2000 [99] in space group P21 (P1211). Matthews analysis suggested two Fab-peptide complexes of approximate molecular weight 49 kD each in the asymmetric unit corresponding to a solvent content of 66% and a Matthews coefficient of 3.6 [100]. The structure was phased by molecular replacement in PHENIX [101]. Source models were the light chain of humanized antibody CC49 [102] (86% identity to the light chain of 7B2) and the heavy chain of an antibody to neuropilin [103] (77% identity to the heavy chain Fab fragment of 7B2) with its CDR-H3 deleted. Conformations of the CDRs were rebuilt and the peptide antigen was constructed de novo as features of the electron density map improved with refinement. Rebuilding and real-space refinements were done in Coot [104] with reciprocal space refinements in PHENIX [105] and validations in MolProbity [106]. Non-crystallographic symmetry was employed throughout refinement. Coordinates and structure factors have been deposited into the Protein Data Bank (www.rcsb.org) with accession code 4YDV.
Antibodies tested for binding to FcR were the same lots as used for the passive infusion studies. Detection reagents were made as described [20,107]. All experiments included aliquots of the same human PBMC control to verify experimental consistency. Aliquots of rhesus PBMC were incubated with the appropriate mAb (7B2 IgG, A32 IgG, CH22 IgG, CH65 IgG [negative control], or mock) at 4°C for 30 min. For experiments using gp120 Env proteins for detection, cells were then blocked with an anti-CD4 mAb (clone SK3; Biolegend, San Diego, CA) at 4°C for 15 min. Cells were washed with 1x PBS + 1% BSA and stained at 4°C for 30 min with an NK cell panel consisting of CD3 PerCP-Cy5.5, CD4 PE-Cy7, CD14 PE, CD20 FITC, CD32 APC (BD Biosciences, San Jose, CA); CD16 BV570, CD64 APC-Cy7 (Biolegend, San Diego, CA); and CD8 PE-TexasRed (Invitrogen, Carlsbad, CA). The NK cell cocktail also contained the detection reagent matched to the FcR-bound mAb: 7b2 detected with gp41 immunodominant region peptide tetramer (sequence biotin-GGGKQLQARVLAVERYLKDQQLLGIWGCSGKLICTTAV); CH22 detected with gp120 clade B consensus V3 peptide tetramer (sequence biotin-GGGTRPNNNTRKSIHIGPGRAFYTTGEIIGDIRQAH); A32 detected with gp120 A244 Env tetramer. Cells were washed again and resuspended in 2% formaldehyde in PBS and stored at 4°C prior to acquisition on a BD LSRII flow cytometer (BD Biosciences, San Jose, CA). Data were analyzed in FlowJo (TreeStar, Ashland, OR). Rhesus NK cells were gated as negative for CD3 and CD14, CD20 dim/negative, CD8 bright [108] Mean fluorescence intensity of detection reagent binding was assessed for CD16+ NK cells only.
Briefly, a 1ml packed suspension of pooled sheep red blood cells (SRBCs) were washed 3 times in 0.85% NaCl and spun at 3,000 rpm for 5 min. The SRBCs were then added to 1ml of a 0.5mg/ml gp41 immunodominant peptide (SP400: sequence biotin-GGGKQLQARVLAVERYLKDQQLLGIWGCSGKLICTTAV) antigen solution and 10 ml of chromium chloride solution (0.1 mg/ml). The coupling suspension was incubated at 30°C for 40 minutes in a shaking incubator (100 rpm) and then centrifuged for 10 minutes at 1000xg, 4°C. Coupled cells were then washed with 0.85% NaCl and centrifuged again for 10 minutes at 1000xg, 4°C. Coupled cells were resuspended as a 25% solution (v/v) in HBSS. The antigen coupling of sp400 to SRBCs was analyzed via flow cytometry. Antibodies 17b IgG mAb and 7B2 IgG_AAA (10 μg/ml) were incubated with 100 μl of a 2% solution (v/v in 0.85% NaCl) of both uncoupled and sp400 coupled SRBCs at room temperature for 30 minutes. Cells were washed twice with 3ml FACS Buffer (1x PBS + 1% BSA, pH 7.2) and centrifuged at 3000 rpm, 5 min to pellet cells. Supernatant was aspirated, and cells were gently re-suspended via agitation. Goat anti-human FITC secondary antibody (1:50 dilution in 1xPBS) was then added to the SRBC suspensions and incubated for 30 min at room temperature. Cells were again washed twice with 3ml FACS Buffer and then resuspended in 1ml FACS buffer.
Antibodies 7B2 IgG_AAA and palivizumab (both at 10 μg/ml) were incubated with gp41 immunodominant peptide coupled SRBCs as well as uncoupled (control) SRBCs for 45 minutes on ice. SRBCs were then washed 3 times with PBS and centrifuged at 3000rpm, 5 min. In 4ml tubes, 5x106 coupled or sp400 coupled SRBCs were mixed with 1x106 human or rhesus PBMCs in 1ml RPMI 1640. Cells were pelleted by centrifugation at 3000 rpm for 5 minutes and incubated as a pellet at 37°C for 30 min. After incubation, the pellet was resuspended in the 1ml media supernatant. Cytopreps were prepared and stained with Wrights stain.
MAb binding to recombinant purified Fc receptors (FcRs) was performed using Surface Plasmon Resonance. Full sequences of the Fc receptors (recombinant FcRγI (CD64), FcRγIIa (CD32) and FCRγIIIb (CD16) (R&D Systems)).and the purification protocol of the rhesus FcRs is described elsewhere (Cocklin et al, 2014 in preparation). Briefly, rhesus FcR allelic variants were synthesized by Genscript with a hexa-histidine tag and cloned into the pcDNA3.3 mammalian expression vector (Invitrogen, Carlsabad, CA). Stable 293 HEK cell lines expressing rhesus FcR variants were prepared by nucleofection (AMAXA, Lonza) and antibiotic selection. Supernatants containing the soluble FcRs were harvested, centrifuged at 3500rpm for 1h and the supernatant was filtered through 0.45μm membranes prior to purification by Immobilized metal affinity chromatography (IMAC) using the Profinia protein purification system (Biorad, Hercules, CA). IMAC eluates were incubated with human pooled IgG sepharose for to remove serum protein contamination and for functional confirmation. IgG eluates were concentrated and dialysed into PBS, pH7.4. Purity was confirmed at >95% by SDS PAGE and coomassie staining and bicinchoninic acid (BCA) assay was used to determine the protein concentration.
PBMC and HIV-1 envelope pseudotyped virus neutralization assays were performed as described [109,110]. Ability of the mAb to inhibit HIV-1 infection of monocyte/macrophages was performed as described [9]. Briefly, human peripheral blood (PB) monocytes were differentiated into macrophages and seeded on 48 well plates. Cells were infected with primary R5 HIV-1 BaL at a concentration of 200 ng/mL viral p24 antigen in the presence or absence of different concentrations of 7B2 IgG_AAA Ab and cultured in AIM plus Glutamax 1 and GM-CSF (10 ng/mL, R&D System). Productive infection was quantified by flow cytometry by detection of intracellular viral p24 antigen in MDM after 48 hours of culture. The percentage of infected cells in presence of Abs compared to control infected cells was determined. The TZM-bl neutralization assay was performed with the SHIV BaL challenge stock and the antibody concentration (μg/ml) at which relative luminescence units (RLUs) were reduced 50% compared to virus control wells were reported.
Surgically resected specimens of intestinal tissue were collected at St Mary’s Hospital, London, United Kingdom, after receipt of signed informed consent. All patients were HIV negative. All tissues were collected under protocols approved by the Local Research Ethics Committee. Colorectal tissue was obtained from patients undergoing rectocele repair and colectomy for colorectal cancer. Only healthy tissue obtained 10 to 15 cm away from any tumor was employed. Following transport to the laboratory, muscle was stripped from the resected tissue, which was then cut into 2- to 3-mm3 explants comprising both epithelial and muscularis mucosae. Colorectal explants were maintained with Dulbecco’s minimal essential medium containing 10% fetal calf serum, 2 mM L-glutamine, 2.5 μg/ml Fungizone (Life Technologies) and antibiotics (100 U of penicillin/ml, 100 μg of streptomycin/ml, and 80 μg of gentamicin/ml) at 37°C under an atmosphere containing 5% CO2 as previously described [66]. Antibodies at 50 μg/ml were pre-incubated with cell free HIV-1BaL (5 x 104 TCID50) for 1 hour at 37°C. The 2-3mm3 dissected colorectal tissue explants were then exposed to virus and antibody for 2 hours at 37°C. Following viral incubation explants were washed three times with PBS and placed into 96-well tissue culture plates and cultured with fresh media containing antibody at the same concentration. The next day the tissue explants were transferred into a new 96-well tissue culture plate and washed twice with PBS. Tissue explants were subsequently cultured for 14 days in 200 μl of medium supplemented. On days 4, 7, 11 and 14 post-infection, 100 μl of supernatant was harvested and replaced with 100 μl of fresh media without further antibody exposure. To assess migratory cells present in the overnight culture of explanted tissue, cells were washed twice with PBS and co-cultured with 4 x 104 PM-1 indicator T cells in 200 μl of medium containing antibodies at 50 μg. The supernatant was collected on days 4, 7, 11 and 14 post-infection and replaced with fresh media without further antibody exposure. All assays were performed in triplicate unless otherwise stated and controls included: virus only; medium only; and antibody isotype controls. Levels of p24 in tissue explant supernatants at day 11 post-infection were quantified using HIV-1 p24 enzyme-linked immunosorbent assays (ELISA; SAIC-Frederick, Inc., Frederick, MD) according to the manufacturer’s instructions. Each experiment was performed in triplicate, using tissues from different donors.
Replication competent HIV -1 BAL was utilized to infect tissue macrophages and blood-derived macrophages. Peritoneal tissue macrophages were obtained under informed consent by IRB approval at Duke University. Virus input was normalized to RNA copies/mL. HIV replication was quantified by measuring the amount of luciferase in macrophage lysates at either day 4 or day 14 post-infection. HIV production was quantified by measuring luciferase in TZM-bl reporter cells infected with macrophage supernatants collected at regular intervals (4, 7, 11, and 14 days post infection).
HIV-1 reporter viruses (provided by Dr. John Kappes and Christina Ochsenbauer, University of Alabama) used were replication-competent infectious molecular clones (IMC) designed to encode the BaL (subtype B) env genes in cis within an isogenic backbone that also expresses the Renilla luciferase reporter gene and preserves all viral ORFs. The Env-IMC-LucR virus used was NL-LucR.T2A-BaL.ecto (IMCBaL) [111,112].
IMCs were titrated in order to achieve maximum expression within 36 hours post-infection by detection of Luciferase activity and intra-cellular p24 expression. We infected 2x106 cells with IMCBAL by incubation with the appropriate TCID50/cell dose of IMC for 0.5 hour at 37°C and 5% CO2 in presence of DEAE-Dextran (7.5 μg/ml). The cells were subsequently resuspended at 0.5 x 106/ml and cultured for 36 hours in complete medium containing 7.5 μg/ml DEAE-Dextran. Infection was monitored by measuring the frequency of cells expressing intracellular p24. The assays performed using the IMC-infected target cells were considered reliable if the percentage of viable p24+ target cells on assay day was ≥ 20%.
Binding of mAbs to the surface of infected cells was performed as previously described [18]. Polyclonal activated CD4+-enriched T cells were obtained and infected by spinoculation (1200 x g for 2 hours; [113]) with NL-LucR.T2A-BaL.ecto (IMCBaL). Cells spinoculated in the absence of virus (mock-infected) were used as a negative infection control. Following 72 hours of infection in RPMI 1640 medium (Invitrogen, Carlsbad, CA), supplemented with 20% Fetal Bovine Serum (FBS) (Gemini Bio-Products, West Sacramento, CA) (R20) in presence of 20 U/ml rhIL-2 (Peprotech, Rocky Hill, NJ) (R20-IL2), CD4+-enriched T cells were washed in PBS, dispensed in 96-well V-bottom plates at 1x105 viable cells/well, and stained with a vital dye (LIVE/DEAD Fixable Aqua Dead Cell Stain, Invitrogen) to exclude non-viable cells from subsequent analyses. The cells were then incubated at 4°C for 25 minutes with the primary Ab. After two washes, cells were stained with Phycoerythrin (PE)-conjugated goat anti-human IgG secondary (2ary) Ab (Southern Biotech, Birmingham, AL) for 2 hours at 37°C. Cells were subsequently washed 3 times and fixed with 1% formaldehyde PBS. The samples were acquired on a LSRII (BD Biosciences) within 24 hours. A minimum of 10,000 total singlet events was acquired for each test to identify live events. Data analysis was performed using FlowJo 9.6.4 software (TreeStar Inc., Ashland, OR).
Luciferase-based ADCC assays were performed as previously described luciferase-based assay [65,114]. The HIV-1 IMCBAL infected CEM.NKRCCR5 cell line (NIH AIDS Research and Reference Reagent Repository) was used as target cells. The target cells were incubated in the presence of 4-fold serial concentrations of mAbs starting at 40 μg/ml. Purified CD3-CD16+ NK cells obtained from a HIV seronegative donor with the F/F Fc-gamma Receptor (FcRγ) IIIa phenotype were used as effector cells. The cells were isolated from the cryopreserved PBMCs by negative selection with magnetic beads (Miltenyi Biotec GmbH, Germany) after overnight resting. The NK cells were used as effector cells at an effector to target ratio of 5:1. The cells and mAbs were incubated in duplicate wells for 6 hours at 37°C in 5% CO2. The final read-out was the luminescence intensity generated by the presence of residual intact target cells that have not been lysed by the effector population in presence of ADCC-mediating mAb. The % of killing was calculated using the formula: % killing = (RLU of Target + Effector well)—(RLU of test well)/ (RLU of Target + Effector well) * 100. In this analysis, the RLU of the target plus effector wells represents spontaneous lysis in absence of any source of Ab. The humanized monoclonal antibody (IgG1k) directed to an epitope in the A antigenic site of the F protein of respiratory syncytial virus, (Palivizumab (MedImmune, LLC; Gaithersburg, MD) was purchased from the manufacturer and used as a control.
To measure the captured infectious IC, we adopted the Ig-virus capture assay described previously [26,64]. Briefly, Microplates (NUNC) were coated overnight at 4°C with mouse monoclonal anti-human IgG (Southern Biotech) at a concentration of 1 μg/ml diluted in PBS. After coating and washing, coated plates were blocked for 2 h with PBS supplemented with 5% Goat serum, 5% milk, 0.05% Tween. The indicated concentration of antibodies was mixed with the viral stock containing 5X 106 viral RNA and then centrifuged 90 min at 2,000 rpm. Then the mixture was centrifuged at 21,000 x g for 45 min at 4°C to remove the virus free antibody [116] the pellet was resuspended in the same volume of PBS. 50 μl of the IC mixture was added to each coated well in triple wells for a 90 min incubation. Then, the wells were washed 4 times and the indicator cell line (M7-luc or TZM-bl) was added. HIV-1 replication was assessed on day 5 after infection for M7-luc and on day 3 for TZM-bl. The infection was measured by the firefly luciferase assay and was expressed as RLU.
T/F viral sequences were obtained by single genome amplification (SGA) followed by direct amplicon sequencing by methods modified from Keele et al. [73] and published in Klein et al. [117]. Viral RNA was purified from the first or second virus positive plasma sample from each animal by the Qiagen QiaAmp viral RNA mini kit and subjected to cDNA synthesis using 1X reaction buffer, 0.5 mM of each deoxynucleoside triphosphate (dNTP), 5 mM DTT, 2 U/mL RNaseOUT, 10 U/mL of SuperScript III reverse transcription mix (Invitrogen), and 0.25 mM antisense primer SHIVBalEnvR1 5’- CTG TAA TAA ATC CCT TCC AGT CC -3’ located in the nef open reading frame (nt 9458–9480 in SIVsmm239). The resulting cDNA was end-point diluted in 96 well plates (Applied Biosystems, Inc.) and PCR amplified using Platinum Taq DNA polymerase High Fidelity (Invitrogen) so that ≤30% of reactions were positive in order to maximize the likelihood of amplification from a single genome. A second round of PCR amplification was conducted using 1 μl of the first round products as template. SHIVBalEnvR1 and SIVsm/macEnvF1 5’-CCT CCC CCT CCA GGA CTA GC-3’ (nt 6130–6146 in SIVsmm239 vpx) were used in the first round PCR amplification step, followed by a second round with primers envB5-in 5’- TTA GGC ATC TCC TAT GGC AGG AAG AAG -3’ (nt 5960–5983 in the HXB2 tat coding region) and BKSIVsm/macEnvR261 5’- ATG AGA CAT RTC TAT TGC CAA TTT GTA -3’ (nt 9413–9436 in SIVsmm239 nef). PCR was carried out using 1X buffer, 2 mM MgSO4, 0.2 mM of each dNTP, 0.2 μM of each primer, and 0.025 U/μL Platinum Taq High Fidelity polymerase (Invitrogen) in a 20-μL reaction. Round 1 amplification conditions were 1 cycle of 94°C for 2 minutes, 35 cycles of 94°C for 15 seconds, 58°C for 30 seconds, and 68°C for 4 minutes, followed by 1 cycle of 68°C for 10 minutes. Round 2 conditions were one cycle of 94°C for 2 minutes, 45 cycles of 94°C for 15 seconds, 58°C for 30 seconds, and 68°C for 4 minutes, followed by 1 cycle of 68°C for 10 minutes. Round 2 PCR amplicons were visualized by agarose gel electrophoresis and directly sequenced using an ABI3730xl genetic analyzer (Applied Biosystems). The final amplification product was ~3160 nucleotides in length exclusive of primer sequences and included all of rev and env gp160, and 336 nucleotides of nef. Partially overlapping sequences from each amplicon were assembled and edited using Sequencher (Gene Codes, Inc). Sequences with ≥2 double peaks indicating amplification from multiple templates were discarded. Sequences with one double peak were retained as this most likely represents a Taq polymerase error in an early round of PCR rather than multiple template amplification; such sequence ambiguities were read as the consensus nucleotide. Sequence alignments and phylogenetic trees were constructed using ClustalW and Highlighter plots were created using the tool at http://www.lanl.gov. Detailed descriptions of the mathematical models of early sequence evolution, star phylogeny determination, estimates of viral fitness, and power calculations for estimating the likelihood of detecting transmitted envelope variants with minor representation are provided elsewhere [118].
In order to compare T/F numbers accurately across different control and antibody-infused groups, we ensured that the number of sequences per group was consistent. In the 6 animals infused with 7b2 and the 6 matched controls, T/F numbers were determined by using 60 sequences per animal, for a total of 720 env sequences. We can be 95% confident of detecting all variants that are at greater than 5% prevalence in each animal. For the A32-treated and matched controls, a mean of 37 sequences were used from each group (range 33–42 over all 12 animals). For the CH22 study, 6 control animals had a mean of 40 sequences each, with a range of 35–45, while the 2 infected animals with CH22 infusion had 40 and 42 sequences, respectively. The A32 and CH22 studies were powered to detect any variants greater than 8% prevalence with 95% confidence. Therefore, all T/F numbers shown are a minimum estimate.
The following rules were followed in order to identify and enumerate T/F variants: In clusters of related sequences determined through visual analysis of phylogenetic trees [FigTree version 1.4 (http://tree.bio.ed.ac.uk/software/figtree/)], we allowed sequences with up to 3 mutations to be part of the same cluster from days 7–14 post-infection. At day 21 or 28 we allowed up to 4 changes in any one sequence. Variants exceeding these limits were identified as the progeny of distinct T/F genomes. Changes at positions predicted by the Hypermut algorithm 2.0 to be potential G-A hypermutation caused by APOBEC 3G/3F were reverted for analysis if there were ≤ 2. Sequences that had ≥ 3 potential APOBEC 3G/3F mutated positions were not considered for the T/F analyses (Hypermut, http://www.hiv.lanl.gov [119]). Sequence clusters of ≥ 2 sequences with ≥ 2 shared mutations were classified as distinct T/F variants. Sequences representing recombinants between two distinct T/F lineages were identified as previously described [33] and were excluded from the analyses. Sequences were deposited in Genbank under accession numbers KR608795—KR610312 (http://www.ncbi.nlm.nih.gov/Genbank/).
Mamu-A01-negative Indian-origin rhesus monkeys were housed and maintained in an Association for Assessment and Accreditation of Laboratory Animal Care accredited institution in accordance with the principles of the National Institute of Health. All antibodies were produced in CHO cells (Catalent, Inc., Somerset, NJ) and were administered by the intravenous route in rhesus monkeys. Monkeys were challenged with 1 ml of SHIV-BaL (2 × 105 TCID50) by the intra-rectal route at times specified in the “Results” section for each of the antibodies tested. The SHIV-BaL challenge stock used in the present study was uncloned. The original SHIV-BaL molecular clone was developed by Pal et al. [120], which was transfected into cells to derive a virus isolate, which in turn was serially passaged in 4 pig-tail macaques and then expanded in the PM-1 cell line. Our laboratory expanded this SHIV-BaL isolate further in human PBMCs and generated a large volume stock for rhesus challenge experiments. Mean genetic diversity of Env determined by single genome sequencing in this stock is 0.3%, with maximum diversity of 0.7%.
SIV plasma viral RNA measurements were performed at the Immunology Virology Quality Assessment (IVQA) Center Laboratory Shared Resource, Duke Human Vaccine Institute, Durham, NC. Plasma viral loads were assessed using a Qiagen QIAsymphony DSP Virus/Pathogen Midi Kit using the QIAsymphony SP platform and real-time PCR reaction carried out on the StepOnePlus (Applied Biosystems) instrument. Data from the real-time PCR reaction was analyzed using the StepOnePlus software. The sensitivity of this SIV viral load assay is 250 copies per ml.
CD4+ T lymphocyte subsets were determined by multi-channel flow cytometry for CD3, CD4, CD8, CD28, CD95, CCR5 and CCR7. CD4+ T lymphocyte counts were calculated by multiplying the total lymphocyte count by the percentage of CD3+ CD4+ T cells. Briefly, 100 μl of EDTA-anticoagulated whole blood was stained with anti-CD3-A700 (clone SP34.2), anti-CD4-PerCP Cy5.5 (clone L200), anti-CD8-APC H7 (clone SK1), anti-CCR5-PE (clone 3A9), anti-CD95 APC (clone DX2) all from BD Biosciences, anti-CD28-PE CY7 (clone CD28.2; eBiosciences), and anti-CCR7-FITC (clone 150503; R&D Systems). Fixed cells were collected (30,000 events) on a LSRII instrument using FACSDiva software version 6.1.1 (BD Biosciences) and data were analyzed using FlowJo Software (TreeStar, Ashland, OR).
|
10.1371/journal.pcbi.1000608 | Predicting Functional Alternative Splicing by Measuring RNA Selection Pressure from Multigenome Alignments | High-throughput methods such as EST sequencing, microarrays and deep sequencing have identified large numbers of alternative splicing (AS) events, but studies have shown that only a subset of these may be functional. Here we report a sensitive bioinformatics approach that identifies exons with evidence of a strong RNA selection pressure ratio (RSPR) —i.e., evolutionary selection against mutations that change only the mRNA sequence while leaving the protein sequence unchanged—measured across an entire evolutionary family, which greatly amplifies its predictive power. Using the UCSC 28 vertebrate genome alignment, this approach correctly predicted half to three-quarters of AS exons that are known binding targets of the NOVA splicing regulatory factor, and predicted 345 strongly selected alternative splicing events in human, and 262 in mouse. These predictions were strongly validated by several experimental criteria of functional AS such as independent detection of the same AS event in other species, reading frame-preservation, and experimental evidence of tissue-specific regulation: 75% (15/20) of a sample of high-RSPR exons displayed tissue specific regulation in a panel of ten tissues, vs. only 20% (4/20) among a sample of low-RSPR exons. These data suggest that RSPR can identify exons with functionally important splicing regulation, and provides biologists with a dataset of over 600 such exons. We present several case studies, including both well-studied examples (GRIN1) and novel examples (EXOC7). These data also show that RSPR strongly outperforms other approaches such as standard sequence conservation (which fails to distinguish amino acid selection pressure from RNA selection pressure), or pairwise genome comparison (which lacks adequate statistical power for predicting individual exons).
| Alternative splicing is an important mechanism for regulating gene function in complex organisms, and has been shown to play a key role in human diseases such as cancer. Recently, high-throughput technologies have been used in an effort to detect alternative splicing events throughout the human genome. However, validating the results of these automated detection methods, and showing that the minor splice forms they detected play an important role in regulating biological functions, have traditionally required time-consuming experiments. In this study we show that such regulatory functions can very often be detected by a distinctive pattern of strong selection on RNA sequence motifs within the alternatively spliced region. We have measured this “RNA selection pressure ratio” (RSPR) across 28 animal species representing 400 million years of evolution, and show that this metric successfully predicts known patterns of alternative splicing, and also have validated its predictions experimentally. For example, whereas high-RSPR alternative splices were found experimentally to undergo tissue-specific regulation in 75% of cases, only 20% of low-RSPR cases were found to be tissue-specific. Using RSPR, we have predicted over 600 human and mouse alternative splicing events that appear to be under strong selection. These data should be valuable for biologists seeking to understand the functional effects and underlying mechanisms of splicing regulation.
| Global analyses of alternative splicing (AS) have established its importance in protein diversity and gene regulation in higher eukaryotes [1],[2]. Alternative splicing can regulate biological function by altering the sequence of protein products and modulating transcript expression levels [3]. Alternative splicing can modify binding properties, intracellular localization, enzymatic activity, protein stability or post-translational modifications[4]. Alternative splicing is often regulated in a tissue-specific manner [5] and can undergo important changes in disease states such as cancer [6],[7]. All of these illustrate that it is necessary to study the functional effects of alternative splicing to understand the complexity of biological system and human disease.
One major challenge is the identification of functional alternative splicing events. In general, experimental methods that can directly demonstrate a biological function for an AS event are time-consuming, ad hoc, and impractical on a genome-wide scale. By contrast, high-throughput methods for surveying the transcriptome, such as EST sequencing [8], microarrays [9] or deep sequencing [10], mainly enable detection of whether a given splicing event is “present” or “absent” in a sample. Genome-wide analysis of such datasets has produced large databases of detected alternative splicing events, but with limited guidance for biologists about which ones are likely to be functional. More importantly, a number of studies have shown that a significant fraction of these detected events are probably not functional [2],[11],[12]. In this context, biologists need improved ways of distinguishing AS events that are likely to have important biological functions, before initiating costly experiments, such as high-throughput studies of regulation [13],[14].
There are multiple aspects of function that can be assessed using bioinformatics. Many studies have used the independent observation of the same alternative splicing event in multiple species as evidence that it is functional [15]–[18]. By contrast, introduction of a STOP codon more than 50 nt from the last exon-exon junction is predicted to cause nonsense-mediated decay; such a splice form will not produce a functional protein product (although the AS event itself might still play an important role in regulating function by down-regulating the transcript level) [19]–. Mapping of the AS exon to known protein domains or structures has also been suggested as an indicator of useful biological function [23],[24]. Many studies have indicated that an AS sequence segment consisting of an exact multiple of 3 nt length is more likely to be functional, since it can be alternatively spliced without affecting the protein reading frame [25],[26]. Evidence of tissue-specific regulation (from EST or microarray data) has also been taken as evidence of function [27]–[30]. While all of these criteria have been shown to be useful indicators of “function”, it should be emphasized that no one method can adequately capture this ill-defined concept, precisely because it has many different aspects.
Functions that are important for reproductive success are subject to selection pressure, which can be defined as a reduction in the frequency of observed mutations relative to that expected under a neutral model. For example, sequence conservation both within alternatively spliced exons and in flanking introns has been cited as evidence of important regulatory motifs. One useful extension of this principle is to separate total conservation into non-synonymous sites (i.e. where mutations will change the amino acid sequence) vs. synonymous sites (where mutations leave the amino acid sequence unchanged). Whereas alternative exons show poorer conservation than constitutive exons by total conservation metrics like phastCons [31], several studies have reported that measures of synonymous mutations (Ks, or ds) drop dramatically in certain types of alternative exons [26],[32],[33], particularly those that show tissue-specific regulation. Unlike standard conservation, such Ks effects cannot be attributed to protein function, and have thus been used as a measure of “RNA selection pressure” for features such as splicing factor binding sites, RNA secondary structure etc. [26], [34]–[40].
With the rapid growth in complete genome sequences for animals and plants [41], the strategy of seeking to detect RNA selection pressure gains increasing power as a way of predicting strongly selected AS regions [42],[43]. Past applications of Ks to this problem typically relied upon comparing a single pair of related genomes (e.g. human vs. mouse). Given the high level of identity seen in such exon comparisons (around 87% for human vs. mouse), the number of synonymous mutations expected in an alternative exon (just based on size, with no RNA selection pressure) is low, perhaps ten or fewer. Even if the observed number of synonymous mutations were three-fold lower, i.e. three or fewer, implying strong RNA selection pressure, the result would not be statistically significant, due to the small number of counts being compared. For this reason, such studies have typically not tried to predict which individual AS exons are strongly selected, but instead to compare entire groups of exons, e.g. all “minor-form exons” vs. all constitutive exons.
However, large multigenome alignments such as the UCSC 28 vertebrate genome alignment could improve predictive power, by measuring Ks simultaneously in many separate species. This has two benefits. First, the dataset of mutation counts for any given exon is greatly increased (naively, by a factor of up to 20-fold compared with a single pair of species), increasing the statistical power for detecting real selection pressure cases. Second, the ability to discriminate spurious cases is enhanced by utilizing a much more diverse set of genomes: various types of artifacts that might occur in one genome (e.g. a mutation “cold-spot” in mouse evolution) are unlikely to be conserved over 28 genomes spanning 300 million years of vertebrate evolution (such conservation would in fact indicate consistently strong selection).
In this paper we present a robust method for applying this approach, combining multigenome alignment data and RNA selection pressure calculations. This approach enables the detection of statistically significant RNA selection pressure for each individual exon, providing a direct prediction of whether that AS event has been strongly selected during vertebrate evolution. We have tested these prediction using a wide variety of data including known sets of regulatory targets, large-scale EST and microarray data, and RT-PCR analysis of tissue-specific splicing regulation.
As an initial test of the RNA selection pressure ratio (RSPR) metric, we applied it to GRIN1, a gene whose alternative splicing is well understood [44]. Of the 21 coding region exons, two exons (exons 5 and 21) show evidence of dramatically increased RSPR (RSPR values of 7 and 10 respectively) compared with the remaining exons (Figure 1). These two exons correspond to the well-studied N1 and C1 alternative exons, which have been shown to be regulated by PTB, NOVA2, hnRNP H and hnRNP A1 [45], and in turn control receptor desensitization [46] and NMDA receptor interactions [47]. The RSPR data indicate that these two exons have synonymous mutation rate approximately ten-fold lower than the surrounding exons (p-value = 7.7×10−25). This very strong signal suggests that it should be possible to detect regulated alternative exons using the RSPR metric. These data suggest that more than half of the synonymous sites in the N1 and C1 exons are under some kind of negative selection pressure. This seems compatible with existing splicing factor motif databases. RESCUE-ESE [48] and FAS-ESS [49] predicted 43% of the N1 exon sequence and 50% of the C1 exon sequence as splicing factor binding sites (Figure 1D).
The GRIN1 data provide evidence that using multigenome alignments (in this case GRIN1 sequences from 16 species) can make RSPR more sensitive than simply comparing a pair of genomes. Figure 1C shows the synonymous substitution data (Ks) used to compute RSPR, annotated on the lineages leading to each species. The Ks values for the GRIN1 N1 and C1 alternative exons are about ten-fold lower than the Ks values for the other exons, on each lineage. For example, within the primate lineage, these alternative exons showed a Ks level of 0.027 vs. 0.26 in the neighboring exons. Independent data for other mammal lineages (mouse, rat, dog) show a similar pattern, with total Ks values of 0.05 vs. 0.70 respectively. Moreover, frog, and zebrafish also show big decreases in Ks for the two alternative exons. These data indicate that the GRIN1 N1 and C1 exons have been under a consistently strong negative RNA selection pressure for over 450 My. They also show why this computational approach can be more sensitive than simply comparing a pair of species, since each additional genome contributes further statistical evidence for the significance of this pattern.
Next, to estimate the sensitivity of RSPR and compare it with existing methods, we applied it to a test set of alternative exons that are known targets of NOVA, generated by Jelen et al [50]. Using a conservative criterion (RSPR>3, P_RSPR<0.001), RSPR detected 55% (25/45) to 73% (11/15) of these alternative exons as having strong RNA selection pressure (Figure 2; Table S1), depending on whether RSPR was computed for the target exon vs. all other exons in the gene, or vs. constitutive exons in the gene. An additional 24% (11/45) showed evidence of substantial selection pressure (RSPR>1.4) and had significant p-values, but fell below our cutoff of RSPR>3. To assess the specificity of RSPR, we also tested it on a dataset of alternative exons not known to be regulated. In the absence of a gold standard dataset of “true negative” exons that are known not to be targets of splicing regulation, we used a test set of 295 mouse alternative exons whose inclusion level remained relatively constant across a set of 10 different mouse tissues [51]. We detected RSPR>3 in only 9.5% (28/295) of these exons. This suggests that RSPR is relatively specific to exons that are targets of regulation.
We also tested two standard existing methods: pairwise genome comparison [52]; and standard measures of sequence conservation computed over the same multigenome alignments as the RSPR. RSPR differs from previous methods in combining two distinct approaches: first, instead of calculating simple sequence conservation, it separates evidence of nucleotide selection pressure from amino acid selection pressure (and ignores the latter); second, it measures this pressure across an entire family of aligned genomes. Ordinarily, studies of RNA selection pressure have compared a pair of genomes (e.g. human vs. mouse), but the pairwise method detected only 4% of the NOVA targets (Figure 2). Furthermore, even the successful pairwise predictions had much weaker p-values than from the multigenome RSPR calculation, by about 4 to 32 orders of magnitude. Secondly, we tested two standard measures of sequence conservation that are computed from all aligned genomes: baseml [53], and phastCons [54]. Baseml is calculated almost identically as RSPR, using the same software package (PAML), but unlike RSPR does not specifically estimate RNA selection pressure. With optimized signal-to-noise cutoffs, standard sequence conservation (baseml) detected 38% (17/45) of the NOVA targets, and phastCons detected 40% (18/45) of the NOVA targets.
We next sought to compare RSPR's performance on a much larger dataset (Table 1). Numerous studies have reported that an alternative splicing event can be characterized as functional if it is independently observed in expression data from divergent species such as human and mouse [15],[16],[18]. We therefore tested each human exon in our dataset to see whether the orthologous mouse exon is observed to be alternative spliced in mouse EST data, and graphed the results as a function of the RSPR value (Figure 3A). For values of RSPR<1, only a small fraction of exons (approximately 10%) were also observed to be alternatively spliced in mouse, but the rate increased rapidly to 60–70% for values of RSPR of 3 or higher. These data suggest again that RSPR can help distinguish AS events that are likely to be functional, and that a large fraction of the exons with RSPR>3 are actually functional in both human and mouse. Furthermore, this pattern of validation can be seen not only in mouse EST data, but holds true in independent EST data from all of the vertebrate species in our study (Figure 3A). Although the total percentage of validation depends on the level of EST coverage for each organism (highest in mouse; lowest in zebrafish), the pattern is consistent in every EST dataset from mammals, chicken, frog, or zebrafish: high RSPR values predict which exons' alternative splicing will be conserved in other species. These results are consistent with the analysis by Plass and Eyras [39]. Throughout all species, an RSPR value of 3 appears to be a threshold for predicting strongly selected alternative splicing events.
We have also used the AS validation rate in mouse to estimate the true positive vs. false positive rates for RSPR (Figure 3B). We classified exons that were independently observed to be alternatively spliced in mouse EST data as “true positives” (validated), and exons that were not observed to be alternatively spliced in the mouse EST data as “false positives” (not validated). We graphed the rate of true positives vs. false positives both as a function of increasing RSPR (Figure 3B). These data show that RSPR outperforms standard measures of conservation (e.g. 60% true positive rate for RSPR vs. 50% true positive rate for baseml vs. 40% true positive rate for phastCons, at the 20% false positive point) over the whole range of error rates. It should be emphasized that the actual false positive rate is likely to be significantly less, because mouse EST coverage is poor, resulting in many AS events that fail to be observed simply due to insufficient EST sampling.
To further assess the robustness of RSPR predictions, we subdivided the exons based on whether they showed standard conservation or not (measured using baseml), and evaluated the predictive power of RSPR in both cases (Figure 4A). RSPR showed a high level of predictive power both for exons with strong standard conservation (baseml mutation ratio>3) and weak conservation (mutation ratio<3). Thus RSPR is robust in predicting functional alternative splicing (regardless of the exact baseml value; Figure 4B ).
We have also found that it is possible to combine RSPR (measured from the exon sequence) with conservation measured in the flanking introns (Figure 4C). We computed the intron mutation ratio in the 50 nt on each side, using baseml. Exons with low RSPR showed flanking intron mutation ratios around one (i.e. neutral selection). By contrast, exons with high RSPR (RSPR>3) nearly all display higher levels of conservation in their flanking introns. This difference was statistically significant (p-value <10−300). On average, the amount of RNA selection pressure measured within the exon (RSPR) is matched by a proportional amount of selection pressure in the flanking 50 nt (intron mutation ratio). However, the intron mutation ratio displays a large variance (as shown by the error bars). These data suggest that exon-based RSPR can be complemented by the independent signal supplied by intron conservation, to improve prediction of strongly selected AS.
We also analyzed the effect of unusually small trees (only 5–9 species; Figure S1B) and of trees containing only placental mammal species (Figure S1C) compared with trees spanning both placental mammals and more distantly related vertebrates (Figure S1D). These data showed that the RSPR calculated using only 5–9 species successfully predicts whether the exon will be observed to be alternatively spliced in another species (in this case, mouse), just as we previously showed for our complete dataset (Figure S1B). Similarly, the RSPR calculations that only used placental mammal species showed the same direct correlation between RSPR and the AS validation rate. It should be noted that these are both unusual cases in our dataset: 86% of the RSPR calculations used 10 or more species, and 78% of the RSPR calculations spanned both placental mammals and more distant vertebrate species (Figure S1A). Finally, it should be emphasized that the highest RSPR values (those significantly above RSPR = 3) come almost exclusively from the cases with broad species representation, as one can see by comparing the RSPR range for the Placental Mammals set (Figure S1C) vs. the set that spans both Placental Mammals + Other Vertebrates (Figure S1D).
We calculated RSPR for a set of 4626 alternative exons in human and 1935 exons in mouse, and applied a threshold of RSPR>3 and P_RSPR<0.001, yielding 345 exons in human (vs. 4.5 expected by random chance) and 262 exons in mouse (Table 1 and Table S2). Some of the alternative splicing events detected by RSPR have already been demonstrated to be functional by experimental studies, such as LRP8 [55], SYK [56], DGKH [57] and WT1 [58] (Table S3). For example, the alternative exon in LRP8 (ApoER2) encodes a 13 amino acid insertion containing a furin cleavage site important for LRP8 function. SYK alternative splicing appears to play an important role in cancer; the exon-skip form (Syk(S), deleting 23 amino acid residues) occurs frequently in primary tumors but never in matched normal mammary tissues. DGKH has been shown to produce two splice forms, DGKH1 and DGKH2 which differ in the inclusion of a serile alpha motif (SAM) domain [57]. DGKH2 was detected only in testis, kidney and colon while DGKH1 was ubiquitously distributed in various tissues. WT1 is a transcriptional regulator with an alternative exon encoding a 17 amino acids insertion, which appears to play a role in regulating cell survival and proliferation.
First, we performed RT-PCR experiments to test a subset of the RSPR predictions (Figure 5; Table S4). Tissue-specific splicing is one type of functional regulation that can be assayed easily in a panel of different tissue samples. We therefore extracted a random sample of 20 high-RSPR exons and 20 low-RSPR exons, and assayed their splicing by RT-PCR in cerebellum, heart, kidney and seven other human tissues. We classified an alternative exon as tissue specific if it was observed to be expressed as the major-form in at least one tissue, and as a minor-form in at least one other tissue. 75% (15/20) of the high-RSPR exons were found to be regulated in a tissue specific manner in this panel, vs. only 20% (4/20) among the low-RSPR exons (p-value 6.1×10−4). Thus a large fraction of high-RSPR exons were validated by experimental evidence of regulation. Low RSPR may indicate a lower probability of functional AS regulation; nearly half of the low-RSPR exons (8/20) did not even show evidence of alternative splicing in this panel (whereas the expected AS pattern was detected in this panel in 100% of high-RSPR exons (20/20).
We tested all the RSPR predictions using large-scale EST and microarray data. If RNA selection pressure is associated with exons whose splicing is tightly regulated (e.g. tissue-specific), then high-RSPR exons should not be expressed ubiquitously, but only in a subset of tissues or cells, and thus would be classified in EST datasets as minor-form (included in less than a third of a gene's transcripts) rather than major-form (included in more than 2/3 of transcripts). To test this hypothesis, we compared the ratio of major-form vs. minor-form exons in low-RSPR vs. high-RSPR groups (Table 1). Whereas the minor/major ratio was 0.0795 in the low-RSPR set, it rose to 1.79 in the high-RSPR set, more than a 20-fold increase. The ratio of intermediate-form over major-form also increased, from 0.158 in the low-RSPR set, to 1.27 in the high-RSPR set. Overall, only 24.6% of the high-RSPR exons were classified as major-form, vs. 74.7% in the low-RSPR set. A large fraction of exons in the high-RSPR dataset (44%) are annotated to be minor form. Thus, high levels of RSPR appear to be associated with exons that are included in a non-ubiquitous fashion, i.e. in a minority of a gene's transcripts.
To assess whether some of this pattern can be traced to tissue-specific splicing, we compared our RSPR results with a mouse microarray study that measured the inclusion level of mouse exons in 10 tissues [51]. Exons with a tissue-switched splicing pattern (i.e. minor-form in some tissue(s), but major-form in other tissue(s) [59] were three times as abundant in the high-RSPR set (28.2%, 11/39) than in the low-RSPR set (8.6%, 25/292). This difference was statistically significant, p-value = 0.0011 (Fisher exact test). We also saw a significant enrichment of tissue-specific exons detected by a previous EST analysis [5]. In the high-RSPR set, 9.3% (30/324) had strong evidence of tissue-specificity in the EST data (LOD>3), approximately double that observed in the low-RSPR set (5.2%, 204/3958). The p-value for this difference was 0.0026. To verify the EST analyses of tissue-specificity, we tested a sample of 10 high-RSPR exons with putative brain-specific splicing in the EST data, using RT-PCR in ten human tissues. 80% (8/10) showed tissue-specific splicing in this experiment (Table S5).
Numerous studies have examined the evolutionary patterns associated with different AS inclusion levels, such as minor-form, major-form, and intermediate form [60]–[63]. We have analyzed the RSPR distributions of these different forms of alternative splicing.
Exons with different inclusion levels show strikingly different distributions of RNA selection pressure as measured by RSPR (Figure 6A). Major-form exons form a tight, symmetric distribution centered on RSPR = 1 (neutral selection). By contrast, minor-form exons display a broader distribution with a peak at RSPR = 3. At least within the constraints of this study (which was limited to exons that have been retained in mammalian genomes long enough to be found in multiple species, so that we can measure an RSPR value), 36% of conserved minor-form exons display strong RNA selection pressure, whereas only 2.3% of major-form exons had strong RSPR, which is in agreement with Xing et.al [26]. Intriguingly, intermediate-form exons revealed a bimodal distribution, with a main peak that closely follows the profile of the major-form distribution, and a small peak extending above RSPR = 3.
Validation by mouse EST data, and by frame-preservation, also highlight interesting differences between major- vs. minor-form exons. First, it is striking that high values of RSPR are predictive of functional AS (as measured by observation of alternative splicing of the orthologous mouse exon, in mouse EST data), at all three inclusion levels, even in major-form exons (Figure 6B). Thus, it appears that a small fraction of major-form exons are under RNA selection pressure for maintaining an important alternative splicing function. However, the total level of AS validation (by mouse EST data) was approximately two-fold higher for minor-form exons (over 80% for RSPR>3, vs. about 40% for major-form exons). Second, RSPR was predictive of functional AS at much lower values of RSPR for minor-form exons than for major-form exons. Third, frame-preservation revealed another interesting difference (Figure 6C). Whereas both minor- and intermediate-form exons displayed strong increases in frame-preservation with increasing RSPR, major-form exons remained at background frame-preservation levels (around 40%) across the whole range of RSPR values. This implies that RNA selection pressure is associated with “modular” protein sequence insertions for minor- and intermediate-form exons (i.e. insertions that do not alter the reading frame of the downstream protein sequence), but not for major-form exons.
Our RSPR results predict hundreds of alternative exons as strongly selected alternative splicing events. As one example, EXOC7 is a component of the exocyst, an evolutionarily conserved octameric protein complex essential for exocytosis. The protein structure of mouse Exoc7 [64], published recently, includes 19 α-helices linked with loops. EST data reveal several alternative exons, including two groups that map to disordered regions in the protein structure, one between helix 6 and helix 7, and the other between helix 12 and helix 13 (see Table S6). The latter (exon_id 25261 in ASAP II) contains 39 nucleotides, and is alternatively spliced within independent EST data not only for human, but also for mouse, dog, cow and frog (Figure 7). Our RSPR analysis detected this exon as having strong RNA selection pressure: RSPR = 6.24, measured over Human, Chimpanzee, Rhesus, Rat, Mouse, Hedgehog, Dog, Cat, Horse, Cow, Opossum, Platypus, Chicken, Lizard, Frog, Tetraodon, Fugu, Medaka and Zebrafish, with a P_RSPR of 1.67×10−10. ASAP2 reports this splicing event as tissue-specific to brain_nerve (with LOD 2.5) and retina (LOD 2.2) [5], and RT-PCR experiments confirmed its brain-specific splicing pattern (Figure 7D). NetPhos analysis [65] indicates a set of phosphorylation sites nearby (See Figure 7B). FAS-ESS (http://genes.mit.edu/fas-ess/) [49] identifies four ESS motifs in this exon (Figure 7E). Although the function of this exon is unknown, all these clues suggest that its alternative splicing has played an important functional role over a very long period of vertebrate evolution.
We have presented an effective method for estimating RNA selection pressure within an individual exon, and have tested its predictions against a variety of empirical measures of functional alternative splicing, such as known NOVA-regulated exons, conserved alternative splicing, frame preservation, and tissue-specific splicing patterns. We have also predicted a large dataset of strongly selected AS exons that can be useful targets for biologists to study the regulation of alternative splicing. Not only can the high-RSPR dataset furnish biologists with new insights into well-studied genes, but also identifies many new targets worthy of experimental study, in the form of strongly selected alternative splicing events.
These data suggest several possible benefits of RSPR. It provides a general way for distinguishing selection pressures that operate at the nucleotide level rather than protein level. RNA selection pressure may reflect many possible functional mechanisms, such as binding sites of splicing regulators including exon splicing enhancers (ESE) and exon splicing silencers (ESS) [48],[49]. As an example, ESE/ESS analyses for GRIN1 and EXOC7 annotated slightly under one-half of sites in these high-RSPR exons as ESEs or ESSs. Finally, RSPR integrates several powerful tools in comparative genomics, such as MULTIZ multiple genome alignments and PAML evolutionary model inference, and can in principle be applied to any genome.
Previous studies have reported that minor-form exons were associated with increased values of Ka/Ks and Ka compared with neighboring constitutive exons [26], [35]–[37] (for a review see [42]). Our results are consistent with this pattern; the distributions of Ka/Ks and Ka for minor-form exons showed significant increases relative to their control regions (neighboring consitutive exons; Figure S2 AB). Moreover, this pattern was also observed for minor-form exons with high RSPR values (Figure S2 C ).
Our approach has important limitations. First, it is important to emphasize that it seeks to detect RNA selection pressure, but gives no suggestion of what specific functional mechanism might cause it. Other possible patterns of selection that might be detected by RSPR include RNA secondary structure present within the pre-mRNA [42], miRNA binding sites, or binding sites in the parent DNA sequence. While some RSPR may be associated with ESE and ESS motifs, we cannot assume that they fully explain the RSPR in alternative exons. Consistent with several previous studies [35],[66] (see [43] for a review), we did not observe an overall correlation between the RESCUE-ESE density and RSPR, or between FAS-ESS density and RSPR in alternative exons (data not shown). Second, in this paper we have focused on demonstrating the predictive value of calculating RSPR within exonic sequence, without taking into account other useful information such as the flanking intron sequence, frame preservation, expression data from multiple species etc. An integrated prediction method would presumably make use of all available information [18]. A naive initial approach, based on simply multiplying the p-values from baseml (for the flanking intron conservation) and from codeml (for the exon RSPR conservation) did not appear to give major improvements of prediction accuracy in our preliminary tests, compared with simply using the codeml p-value. Since the two calculations use different programs (baseml vs. codeml) and different mutation models (single-nucleotide based vs. codon based), combining them in a single integrated calculation is not trivial. Third, RSPR is calculated based on the multiple genome alignment, and thus requires that an exon be sufficiently conserved among several genomes to be aligned.
We obtained data for alternative exons (exon skipping) and constitutive exons in the same gene, from the ASAP II database [67]. Based on EST data, ASAPII classified each alternative exon as Major form (exon inclusion level greater than 2/3), Minor-form (exon inclusion level less than 1/3) and Intermediate-form (exon inclusion level between 1/3 and 2/3). We defined a constitutive exon as an exon that is included in all transcript isoforms of the gene (inclusion level 100%). We used several additional kinds of data to validate the prediction of functional alternative splicing, such as conserved alternative splicing based on independent EST data from multiple species assembled in the ASAPII database [67], tissue-switched alternative exons identified in the mouse microarray data of Pan et al. [51],[59].
We obtained all genome alignment information used in this study from the UCSC 28 vertebrate genome alignment hg18_multiz28way [68], available from ftp://hgdownload.cse.ucsc.edu/. This alignment includes the following complete genomes: Human, Armadillo, Bushbaby, Cat, Chicken, Chimpanzee, Cow, Dog, Elephant, Frog, Fugu, Guinea Pig, Hedgehog, Horse, Lizard, Medaka, Mouse, Opossum, Platypus, Rabbit, Rat, Rhesus, Shrew, Stickleback, Tenrec, Tetraodon, Tree shrew and Zebrafish.
We used the list of NOVA target exons of Jelen et al. [50] as a validation testset for our RSPR predictions. We were able to obtain genomic coordinates and genome alignments for 45 of NOVA targets published in Jelen et al., which were used for the validation tests presented in Figure 1D and Table S1.
We calculated the RNA Selection Pressure Ratio (RSPR) for each alternative exon compared with the constitutive exons within the same transcript isoform. Here we briefly summarize each step (See Figure S3): 1) each exon was mapped to orthologous exons in the 28 aligned genomes using the NLMSA alignment query tool [69] in the Pygr software package (http://code.google.com/p/pygr/). 2) each orthologous exon was required to retain the aligned splice sites and maintain a minimum of 70% amino acid identity (calculated by needle in EMBOSS [70]. For each exon, a minimum of 5 species was required. We ranked constitutive exons in a given gene by the number of species with orthologous exons, and identified the top third (i.e. most widely conserved exons), or a minimum of four constitutive exons, to represent that gene. 3) We then found the subset of species that were each aligned to all of these exons as well as to the alternative exon. That is, we computed the intersection of the sets of species that are aligned to each of these exons. This yielded the subset of species that we used for the RSPR calculation. 4) Next, we generated the list of constitutive exons that were aligned to all of these species, and used these as the control region for the RSPR calculation. 5) We extracted the alignment consisting of just this subset of species, for the control region + the alternative exon. 6) Prior to calculating RSPR, gaps were removed from the alignment. Specifically, only codons that were present in each of the species in the subset were retained in the alignment: amino acid sequences for the orthologous exons were aligned using clustalw [71], columns containing gaps were removed. This procedure ensures that RSPR is calculated using the exact same tree of species for the control region as for the alternative exon.
We used the multi-partition model D [72] of the PAML program codeml (http://abacus.gene.ucl.ac.uk/software/paml.html) to calculate maximum likelihood estimates of the RSPR, and a p-value for the null hypothesis of neutral RNA selection (i.e. RSPR = 1). Codeml uses a codon substitution model that is similar to the HKY85 nucleotide substitution model. Although codeml does not directly compute RSPR, its output parameters can be used to calculate RSPR. The gap-trimmed nucleotide sequences and a tree file including the phylogenetic tree for the subset of aligned species[68] were submitted to the codeml program, which estimates a set of evolutionary parameters from the whole tree, by maximum likelihood. We defined the RSPR as the ratio of synonymous mutation rate Ks for the alternative exon vs. the constitutive exons among all branches in the phylogenetic tree:(1)where subscript 0 indicates the constitutive exons, and subscript 1 indicates the alternative exon. Based on this definition, a high RSPR value implies the alternative exon is under stronger negative RNA selection pressure than the corresponding constitutive exons. Ordinarily, for each branch in the tree, codeml estimates the total branch length t [73], which is related to the synonymous and non-synonymous substitution densities Ks and Ka via the ratio(2)where S and N are the number of synonymous and non-synonymous sites respectively. In multiple partition mode, tb for each branch b is replaced by t0b (for the constitutive exons) and t1b (for the alternative exon), related by the partition ratio(3)which has a single value over the entire tree (i.e. for every branch b, t0b is constrained to be equal to t0b = r t1b). Based on the input phylogenetic tree and sequence alignment, codeml simultaneously estimates t1b (for each branch), r (for the whole tree), as well as the πj, κ0, κ1, and ω0, ω1 values (for the whole tree). Finally, combining equations 1, 2, and 3, we computed the RNA selection pressure ratio ρ:(4)where S0, S1 and N0, N1 are the number of synonymous and non-synonymous sites in the constitutive exons vs. the alternative exon, respectively.
We also modified codeml to be able to compute the likelihood under the constraint ρ = 1. We computed the p-value P_RSPR for the null hypothesis RSPR = 1 based on the log-odds ratio 2log(L(ρML)/L(ρ = 1)), which follows a χ2 distribution with one degree of freedom, where ρML is the original maximum likelihood estimate of ρ obtained above.
We used two different standard methods for computing sequence conservation, baseml [53] and phastCons [54]. The baseml calculation used the HKY85 nucleotide substitution model (model = 4) and the Mgene = 3 multiple partition mode, similar to our codeml calculation. RSPR and baseml are calculated almost identically using the PAML package; the only difference is that whereas RSPR is calculated from the Ks ratio (eq. 1, above), for baseml we simply used the total nucleotide substitution ratio r (eq. 3, above). We calculated this ratio both for the alternative exon, and its flanking intronic regions (50 nt flanking each exon on each side).
PhastCons [54] is a widely used method for measuring sequence conservation in multiple genome alignments, used for example in the UCSC genome browser. We used phastCons to compute the ratio η for a region of interest compared with a control region (analogous to the RSPR), as follows:(5)where is the average probability of the phastCons non-conserved state in the control region, vs. in the region of interest, and is the average phastCons score in the control region, vs. in the region of interest. We applied this both to exons (constitutive exons vs. alternative exon), and their flanking introns (50 nt flanking each exon on each side). We also calculated a P value based on the phastCons score for the null hypothesis that the mutation density is equal in the control region vs. region of interest. Specifically, we performed the Wilcoxon rank sum test on the phastCons scores for each nucleotide from the control region, vs. for each nucleotide from the region of interest.
For performing the NOVA analysis (Figure 2), we first determined cutoffs for the baseml ratio and phastCons ratio that yielded the same false positive rate in our ROC analysis (Figure 3B) as our RSPR cutoff (RSPR = 3): baseml ratio = 2.9; phastCons ratio = 45. Thus the NOVA analysis compares the sensitivity of these different methods, when calibrated to the same level of specificity.
We generated a random sample of 20 high-RSPR exons (RSPR>3 and p<0.001), and a random sample of 20 low-RSPR exons (RSPR<1.0 and p<0.001). We then designed primers and performed RT-PCR as described below. As a separate test to confirm putative brain-specific splicing identified from EST data, we performed a join of the high-RSPR exon set and a previous database of EST evidence of brain-specific splicing [5]. We selected ten exons from this group, and performed RT-PCR as described below.
Total RNA samples from 10 human tissues were purchased from Clontech (Mountain View, CA). Single-pass cDNA was synthesized using High-Capacity cDNA Reverse Transcription Kit (Applied Biosystems, Foster City, CA) according to manufacturer's instructions. For each tested exon, we designed a pair of forward and reverse PCR primers at flanking constitutive exons using PRIMER3. Two µg of total RNA were used for each 20 ul cDNA synthesis reaction. For each exon, 15 ng total RNA equivalent of cDNA were used for the amplification in a 10 µl PCR reaction. For each exon tested, three DNA polymerase systems were used to optimize RT-PCR reaction: Herculase® II Fusion DNA Polymerase (Stratagene, La Jolla, CA), HotStarTaq DNA Polymerase (Qiagen, Valencia CA) and Phire® Hot Start DNA Polymerase (NEB, Ipswich, MA). PCR reactions were run between 25 to 35 cycles (optimized for each exon) in a Bio-Rad thermocycler with an annealing temperature of 62 to 66°C (optimized for each exon). The reaction products were resolved on 2% TAE/agarose gels or 5% TBE polyacrylamide gels. Each result was a representation of 3–6 RT-PCR replications. DNA fragments with ambiguous sizes were cloned for sequencing using Zero Blunt® TOPO® PCR Cloning Kit (Invitrogen, Carlsbad, CA). Gel images (Table S4 and Table S5) were visually assessed as tissue specific if the alternative exon's splicing fraction (percentage of the exon-inclusion isoform vs. the exon-skip isoform) changed by a factor of two or greater in different tissues.
|
10.1371/journal.ppat.1003213 | An Overexpression Screen of Toxoplasma gondii Rab-GTPases Reveals Distinct Transport Routes to the Micronemes | The basic organisation of the endomembrane system is conserved in all eukaryotes and comparative genome analyses provides compelling evidence that the endomembrane system of the last common eukaryotic ancestor (LCEA) is complex with many genes required for regulated traffic being present. Although apicomplexan parasites, causative agents of severe human and animal diseases, appear to have only a basic set of trafficking factors such as Rab-GTPases, they evolved unique secretory organelles (micronemes, rhoptries and dense granules) that are sequentially secreted during invasion of the host cell. In order to define the secretory pathway of apicomplexans, we performed an overexpression screen of Rabs in Toxoplasma gondii and identified Rab5A and Rab5C as important regulators of traffic to micronemes and rhoptries. Intriguingly, we found that not all microneme proteins traffic depends on functional Rab5A and Rab5C, indicating the existence of redundant microneme targeting pathways. Using two-colour super-resolution stimulated emission depletion (STED) we verified distinct localisations of independent microneme proteins and demonstrate that micronemal organelles are organised in distinct subsets or subcompartments. Our results suggest that apicomplexan parasites modify classical regulators of the endocytic system to carryout essential parasite-specific roles in the biogenesis of their unique secretory organelles.
| Eukaryotic cells evolved a highly complex endomembrane system, consisting of secretory and endocytic organelles. In the case of apicomplexan parasites unique secretory organelles have evolved that are essential for the invasion of the host cell. Surprisingly these protozoans show a paucity of trafficking factors, such as Rabs and it appears that they lost several factors involved in endocytosis. Here, we demonstrate that Rab5A and Rab5C, normally involved in endocytic uptake, actually regulate secretion in Toxoplasma gondii, since functional ablation of Rab5A or Rab5C results in aberrant transport of proteins to specialised secretory organelles called micronemes and rhoptries. Furthermore, we demonstrate that independent transport routes to micronemes exist indicating that apicomplexans have remodelled Rab5-mediated vesicular traffic into a secretory system that is essential for host cell invasion.
| Eukaryotic cells evolved a complex internal membrane system, giving rise to specialised organelles that are linked to the endocytic or exocytic pathway. The basic organisation of the endomembrane system is conserved in all eukaryotes and includes the ER, Golgi and major exocytic pathways [1]. Indeed, recent efforts to compare trafficking factors in the genome of diverse taxa provided compelling evidence that the endomembrane system of the last common eukaryotic ancestor (LCEA) was very complex with most trafficking factors, like Rab-GTPases, present [2], [3].
Rabs constitute the largest family of small G-proteins that function as molecular switches in vesicular traffic [4]. They are required for the specific transport of vesicles from a donor to an acceptor compartment. Importantly, several studies suggest that the function of orthologous Rabs is highly conserved across different taxa. For example, Rab1 and Rab2 appear to have a highly conserved role in ER/Golgi transport [5], [6] and are present in most eukaryotes. Similarly, at least one copy of Rab5, Rab6, Rab7 and Rab11 can be found across most taxa [2]. Other Rabs are present in different taxa, but have been secondarily lost in some species. A good example is Rab4 that although present in many diverse eukaryotes has been lost in several instances [2]. Others like Rab5 or Rab11 display lineage specific expansion most likely due to gene duplications [7], [8]. This expansion of trafficking factors allowed an increase in organellar complexity in the respective species [9] consistently, in more complex organisms the number of Rabs increased [10]. However, this view can be challenged by the fact that many unicellular organisms contain a huge repertoire of Rabs. A recent comprehensive analysis of 56 Rabs in the ciliate Tetrahymena thermophila suggests that this protozoan evolved a highly dynamic flexibility in vesicular trafficking pathways [3], [11]. Similarly, the ciliate Paramecium tetrauelia contains more than 200 Rabs [12].
Ciliates and apicomplexans belong to the recently recognised infrakingdom alveolata [13]. Despite the huge expansion of Rab-GTPases seen in ciliates, apicomplexan parasites show only a relatively basic set of Rabs [14]. Only few members, like Rab1A or Rab11B have been described as unique to apicomplexans [8], [15]. Furthermore, it appears that apicomplexans lost other components of their vesicular trafficking system, such as adaptin complex genes [16], or components of the Endomembrane Sorting Complex (ESCRT) [17], [18]. The ESCRT-system consists of four distinct complexes (ESCRT0, -I, -II and -III) and is required for ubiquitin-mediated endocytosis of surface proteins [19]. Although this paucity might suggest the absence of a functional endocytic system, endosomal like compartments [20] and lysosome-like vacuoles have been described [21], [22]. We recently demonstrated the role of the dynamin related protein B (DrpB) in vesicular traffic to the rhoptries and micronemes and showed that in its absence, organelle-specific proteins enter the constitutive secretion pathway [23]. Similarly, a recent study identified the T.gondii homologue of yeast VPS10 (TgSORTLR) as an essential cargo receptor to transport microneme and rhoptry proteins from the Golgi to endosomal-related compartments [24]. As is with DrpB, ablation of VPS10 resulted in the constitutive secretion of microneme and rhoptry proteins and an organellar biogenesis defect. While these studies confirm the crucial role of the Golgi/TGN (trans-Golgi network) and the highly conserved retromer complex in the specific transport of microneme and rhoptry proteins en route to the endosomal-like compartments (ELCs), it is unknown how the parasite coordinates the transport to and from the ELCs.
Given their crucial role as regulators of vesicular transport we performed a systematic analysis of T.gondii Rabs to examine their role in the delivery of proteins to the specialised secretory organelles. We found that overexpression of either wild type and/or dominant negative expression of Rab5A and RabC mutants resulted in a specific defect in transport to micronemes and rhoptries. Surprisingly, only a subset of microneme specific proteins requires Rab5A and Rab5C for their transport to the apical complex. Together our data suggest that apicomplexans have altered part of their endocytic system into unique secretory organelles.
Previously, 15 genes encoding Rab-like proteins have been identified in the genome of T. gondii, whereas analyses of other apicomplexan genomes indicated the presence of 9 Rab-GTPases in Theileria, Cryptosporidium and Babesia and 11 in Plasmodium [14] (Table S1). We excluded 3 of the previous identified putative Rab-GTPases in T.gondii from our analysis based on the absence of Rab-family specific motifs or expression evidence in the available database (www.toxoDB.org) and our inability to amplify cDNA for the respective genes (Table S2). We amplified full-length cDNA for each Tgrab and verified their predicted amino acid sequence (Figure S1, Table S1&S2).
Given the presence of unique secretory organelles in apicomplexan parasites, we wondered if some of the identified Rabs show unique features that would classify them as apicomplexan or alveolate specific, as previously demonstrated for Rab11B [8], or Rab1A [15]. Surprisingly, it appeared that apicomplexans have a highly conserved, minimal set of Rabs that has been maintained across most eukaryotic lineages (Figure 1A, Figure S2). Importantly, the lack of lineage specific expansions of Rabs in apicomplexans raised the question as to how these parasites have evolved an elaborate endomembrane system with several unique secretory organelles [7], [9], [10]. We speculated that some of the highly conserved Rabs (i.e. Rab4, Rab5, Rab7) might have acquired a novel role in regulating vesicular transport in apicomplexan parasites.
For the generation of stable transfected parasites expressing Rabs we employed the ddFKBP-system that allows tuneable regulation of protein levels in dependence of the ligand Shield-1 (Shld-1) [25] and hence minimises the risk of mis-localisations due to overexpression of the respective Rab. As Toxoplasma Rab6 [26], Rab11A and Rab11B [8], [27] have been described previously they were excluded from this analysis.
We generated parasite lines expressing ddFKBPmyc-tagged versions of Rab1A, 1B, 2, 4, 5A, 5C, 7, 18 and Rab5B-ddFKBPHA and performed co-localisation studies. Hereafter, these transgenic strains are referred to as simply Rab1A, Rab1B, Rab2, Rab4, Rab5A, Rab5B, Rab5C, Rab7 and Rab18. For all experiments we confirmed Shield-1 dependent regulation of protein levels (Figure S3) and adjusted the Shield-1 concentration to a low level to minimise the risk of artefacts due to overexpression of the respective protein. A summary of the different localisations is shown in Figure 1B. Intriguingly, we found that none of the Rabs co-localised with the apical secretory organelles. Instead, we found that all Rabs localise to the early secretory pathway (Rab1B, 2 and 18), the Golgi (Rab4), or the late secretory pathway (Rab5A, Rab5B, Rab5C and Rab7). For Rab1A we were unable to determine its exact localisation, since co-staining with diverse markers of both the early and late secretory system was observed, possibly indicating multiple roles of this Rab (Figure S4). For a more detailed discussion of the individual localisations, see supplements (Text S1 and Figure S4&S5).
Overexpression of Rab-GTPases is an efficient method to analyse their function and has been successfully employed in different eukaryotes [28], [29]. To screen for phenotypes caused by overexpression of the different ddFKBPmyc/HA-tagged Rabs, we inoculated each of the above transgenic parasite strains on HFF monolayers for 5 days in presence and absence of 1 µM Shield-1, which ensured maximal stabilisation of the respective ddFKBP-tagged protein [25] (Figure S3) and found that parasite growth was ablated for Rab2,4 5A,5B and 5C (Figure 2A). Rab1A and Rab1B displayed a growth defect, while in contrast for Rab7 and Rab18 only minor differences in parasite growth were detected (Figure 2A)
In order to identify Rabs that play a crucial role in vesicular transport to the apicomplexan-specific secretory organelles we grew parasites overexpressing Rabs 2,4,5A,5B and 5C for 24 hours in presence of Shield-1 and analysed the location of the microneme proteins MIC2 and MIC3 and of the rhoptry proteins ROP2,3 and 4 (Figure 2B). While we were unable to detect aberrant targeting of these proteins in Rab2, Rab4 and Rab5B parasites, overexpression of Rab5A and 5C resulted in an aberrant localisation of rhoptry proteins and MIC3 in the lumen of the parasitophorous vacuole. This indicates that these proteins have entered the constitutive secretory pathway in these parasites. Curiously, we were unable to detect a similar defect in trafficking of MIC2, which displayed a normal microneme location in all lines (Figure 2B). We conclude therefore that overexpression of Rab5A and Rab5C results in a specific trafficking defect to rhoptries and micronemes and the different behaviour of MIC2 and MIC3 suggest that specific transport pathways exist for a subset of microneme proteins.
Similar to Rab5A and 5C we found Rab1A and Rab7 on post-Golgi membranes (Figure 1B, Figure S4&S5), indicating a role in the late secretory pathway. We established parasites conditionally expressing ddFKBPmyc-tagged, trans-dominant versions of Rab1A and Rab7 (Figure S6&7). Expression of dominant negative Rab1A(N126I) resulted only in a slight, but not significant growth defect (Figure S6) and no adverse effects were evident on secretory or other organelles (data not shown).
The high conservation in eukaryotes and its localisation to ELCs in apicomplexans suggests an important role for Rab7, possibly during the vesicular transport to a plant like vacuole and/or secretory organelles [21], [22]. We generated parasite strains conditionally expressing dominant negative (N124I) and dominant active (G18E) [30] Rab7 (Figure S7). While expression of Rab7(N124I) did not affect parasite proliferation, expression of Rab7(G18E) blocked growth (Figure S7). However, this growth deficiency was not apparently linked to vesicle targeting to secretory organelles, since all markers tested (proM2AP, TgVP1, TgCPL, M2AP, MIC3, Rop2-4) showed a normal localisation in parasites expressing Rab7(G18E) (Figure S7). Therefore, in the absence of additional markers for the ELCs we were unable to define the precise trafficking step regulated by Rab7 during the asexual life cycle of the parasite.
Although Rab5B showed a similar localisation at the ELCs as Rab5A and Rab5C, we failed to detect an immediate effect on trafficking to the micronemes or rhoptries (Figure 2B). However, since prolonged overexpression of Rab5B-ddFKBPHA resulted in the gradual mislocalisation of microneme and rhoptry proteins (Figure S8), we wished to exclude a direct role of Rab5B in the transport to the secretory organelles. Therefore, we generated transgenic parasites conditionally expressing a dominant negative version of Rab5B (ddFKBPmyc-Rab5B(N152I)). As expected, we found that parasite growth was almost completely blocked (Figure S8). However, no effect on the location of microneme and rhoptry proteins could be detected, demonstrating that Rab5B plays no direct role in the vesicular transport to the secretory organelles of the parasite (Figure S8).
Since we found in the initial screen that overexpression of both Rab5A and Rab5C resulted in the constitutive secretion of MIC3, but not MIC2 (Figure 2B), we analysed further the localisation of different microneme proteins to determine their dependence on functional Rab5A and Rab5C. While MIC2, M2AP and AMA1 showed normal localisation in the micronemes, MIC3, 8 and 11 entered the constitutive secretory pathway upon overexpression of Rab5A and Rab5C (Figure S9). In good agreement with other mutants that show a defect in microneme and rhoptry biogenesis [23], [24], we found that parasites overexpressing Rab5A or Rab5C displayed no proliferation defect, while host cell egress and invasion was significantly blocked (Figure S9).
To ensure specificity of the observed overexpression phenotype, we established parasites expressing dominant negative versions of Rab5A(N158I) and Rab5C(N153I). We verified efficient Shield-1 dependent regulation of both dominant negative Rab mutants and found a severe block in parasite growth upon their induction (Figure S10). Expression of dominant negative Rab5A(N158I) and Rab5C(N153I) resulted in a phenotype identical to that observed for overexpression of their wild type versions. We found that all rhoptry proteins analysed (Rop2-4,5) enter the constitutive secretory pathway. In sharp contrast, only a subset of micronemal proteins (MIC3,8,11) were mislocalised upon expression of Rab5A(N158I) and Rab5C(N153I) (compare Figure 3 and Figure S9). While intracellular replication was not affected (Figure 3B), we observed that the mislocalisation of essential microneme and rhoptry proteins results in parasites that are blocked in host cell egress (Figure 3C) and invasion (Figure 3D). This reconfirms that secretory organelles function primarily during host cell egress and invasion and are at least in vitro dispensable for parasite replication [23], [24].
Next, we performed a time course analysis following Shield-1 induction and found MIC3 and MIC8 to be aberrantly targeted within 12 hours (∼80%), whereas ∼20% of parasites showed mislocalisation of MIC2 and M2AP after this time (Figure 4A and B). Importantly, in these cases we found that MIC2 and M2AP show a non-specific, intracellular localisation (Figure 4A, arrowhead), whereas MIC3 and MIC8 are trafficked via the constitutive secretory pathway (Figure 4A and B).
We also verified that other organelles, such as Golgi, apicoplast, mitochondria, dense granules and the inner membrane complex (IMC) are not affected by expression of dominant negative Rab5A (Figure S11) or Rab5C (data not shown), demonstrating that Rab5A and Rab5C are specifically required for the transport of rhoptry proteins and a subset of microneme proteins.
To gain more insight into the role of Rab5A and RabC in vesicular transport, we analysed the localisation of ELC markers, such as proM2AP, TgCPL and TgVP1 [21], [22]. None of these markers showed a significant relocation when either wild type or dominant-negative Rab5A and Rab5C were expressed (Figure 5A), indicating that in T.gondii functional ELCs are maintained in the absence of functional Rab5A or Rab5C.
Many microneme proteins, such as MIC3 and M2AP undergo proteolytic maturation during their transit through the ELCs [20], [22], [31]. Since MIC3, but not M2AP is constitutively secreted in parasites expressing dominant negative Rab5A, we investigated at which step this rerouting occurs. If rerouting occured directly at the Golgi MIC3 would be secreted as an immature proMIC3. In contrast, if rerouting occurred at the ELCs, processing of the propeptide would take place, resulting in secretion of mature MIC3. Therefore, we performed a pulse-chase experiment and compared pro-peptide processing of MIC3 in wild type (RH) parasites and parasites expressing ddFKBPmyc-Rab5A(N158I) in presence and absence of Shield-1 (Figure 5B). We were unable to detect any differences in propeptide processing, strongly suggesting that the rerouting of MIC3 occurs post-Golgi, after processing in the ELCs (Figure 5C).
The results suggest that microneme proteins reach their destination using at least two distinct transport routes, with one depending on functional Rab5A and/or Rab5C. Consequently we speculated that micronemes might be organised into different subsets with different protein content. Dense clustering of secretory organelles within the apical complex of the parasite and limitations in optical resolution make it difficult to differentiate individual compartments using standard microscopy techniques. For these reasons we used super-resolution two-colour STED (Stimulated Emission Depletion) microscopy [32] to finely pinpoint the subcellular localisation of microneme proteins (Figure 6). As expected MIC2 and its interaction partner M2AP [33] exhibit an extensive co-localisation pattern (Figure 6A). In contrast, M2AP and MIC3 localised to independent subsets with minimal overlap in the sub-apical region of the parasites (Figure 6A). This is in good agreement with the genetic data presented above and demonstrates that MIC3- and M2AP-positive vesicles are independently transported to the apical tip of the parasite. Due to limitations in z-resolution, we were unable to resolve if MIC3 and M2AP show also different locations at the densely packed apical tip of the parasite. Therefore we performed two-colour STED on 100 nm ultrathin-melamine sections of parasites labelled with antibodies against different microneme proteins (Figure 6B). Examination of thin-sections of the apical tip revealed a near perfect co-localisation of MIC2 and M2AP (as expected), whereas all other microneme protein combinations exhibited lower co-localisation correlations (Figure 6B and C). Together, the results demonstrate that microneme proteins are transported via independent routes to the micronemes, where they are stored either in different subsets or sub-compartments, rather than a unique, homogenous organelle.
Next, we analysed the ultrastructure of parasites overexpressing Rab5A(N158I). In good agreement with our previous analysis, these parasites are devoid of rhoptries and only very few micronemes are identified (Figure 7A,B). In particular, interference with Rab5A function resulted in a significant loss of micronemes (∼70%) (Figure 7E). STED analysis showed that M2AP has a normal localisation at the apical pole of the parasite, whereas MIC3 is now detected in the parasitophorous vacuole (Figure 7D). Together, these results indicate that M2AP is transported to a subset of micronemes in a Rab5A-independent manner, while MIC3 transport cannot occur in absence of functional Rab5A.
As Apicomplexa parasites invade the host cells, microneme content is the first to be released, as they contain virulence factors that act in a sequential manner during host cell egress and reinvasion. The microneme transmembrane protein MIC2 is involved in host cell attachment and gliding motility [34] and MIC8 is required for rhoptry secretion [35] leading to the establishment of the moving junction (MJ). Finally, AMA-1 has been suggested to interact with the MJ [36], [37]. How Toxoplasma achieves sequential secretion of proteins present in the same organelle has remained a mystery. This is partly due to the dense packing of secretory organelles within the apical complex of the parasite that limits their optical resolution, rendering it difficult to differentiate between individual sub-compartments using standard microscopy. To test whether micronemes are in fact made up of multiple subsets we employed two-colour STED measurements on ultra-thin sections to finely pinpoint the location of different microneme proteins. We found that only few microneme proteins, such as MIC2 and M2AP that are known to form a complex [38], co-localise. In contrast substantially less co-localisation was observed for several other microneme proteins, suggesting the presence of different subsets of micronemes with independent protein content. Alternatively, the differential localisation could reflect organisation of micronemal proteins into sub-compartments, similar to the rhoptries, where proteins are either localised in the bulb, or the neck of the organelle [39].
To identify apicomplexan Rabs involved in regulating vesicular transport to unique parasite secretory organelles, we employed an overexpression screen and identified T. gondii Rab5A and Rab5C as key regulators. Intriguingly, we found that Rab5A and Rab5C regulate the localisation of MIC3, 8, 11 and not MIC2, PLP1, AMA1 and M2AP. Combined with our STED analysis it strongly suggests that micronemes are organised into at least 2 different subsets. Similarly, ultrastructural analysis of Rab5A and Rab5C mutants showed only a subset of micronemes to be ablated. In contrast to micronemes, interference with Rab5A and Rab5C function ablated rhoptry organelles indicating that Rab5A and Rab5C are essential for rhoptry biogenesis. Moreover, pulse-chase experiments indicate that the defect in microneme transport occurs post-Golgi, since MIC3 maturation that occurs in the endosomal system [20], [22], [31] was not affected upon functional ablation of Rab5A and Rab5C.
We speculate that rhoptries and one subset of micronemes share the same Rab5A/C trafficking pathway, see model presented in Figure 7F. It appears that apicomplexans altered their endocytic system to evolve unique secretory organelles. Furthermore we demonstrate that microneme proteins consist of at least two independent populations, with distinct transport pathways and subcellular localisations. One involves transport of MIC3, 8 and 11, whereas the other involves MIC2, M2AP, PLP1 and AMA1. It is possible that redundant pathways are in place that can complement functional abrogation of Rab5A and C (Figure 7F).
Unfortunately we were unable to identify the trafficking pathways involved in the transport of the second subset of microneme proteins (MIC2, M2AP, PLP1 and AMA1).
In other eukaryotes, Rab5-GTPases are known to function in the transport of vesicles in the early endocytic pathway [28], where they act as master regulators for endosome biogenesis [40]. In contrast, Rab5A and Rab5C appear to be essential for the transport of proteins to the unique secretory organelles in T.gondii, demonstrating a previously unknown functional plasticity that has allowed apicomplexans to evolve micronemes and rhoptries. In good agreement with our data, all microneme and rhoptry trafficking mutants described so far correspond to homologues of the yeast VPS (vacuolar protein sorting) system. The dynamin related protein B (DrpB) is a homologue of VPS1 [23], Sortilin (TgSORTLR) of VPS10 and Rab5A and RabC are homologues of VPS21. In yeast these mutants were identified by screening for transport defects of carboxypeptidase Y (CPY) to the yeast vacuole, which is analogous to the lysosome [41]–[44]. Interestingly, their abrogation in yeast leads to the constitutive secretion of CPY, a phenotype we observed here for rhoptry and microneme proteins. Similarities between rhoptries and secretory lysosomes have been pointed out [45] and it is tempting to speculate that micronemes and rhoptries are derived from lysosomal organelles. Therefore the data presented in this study are consistent with the parasite modifying parts of its endocytic system giving rise to the formation of unique organelles, required for intracellular parasitism.
T. gondii tachyzoites (RH hxgprt−) were grown in human foreskin fibroblasts (HFF). To generate stable transformants parasites were transfected and selected as previously described [46]. The selections based on pyrimethamine and chloramphenicole resistance were achieved as described previously [47], [48].
For ddFKBPmyc-tagging of Rab-GTPases full-length cDNAs (see Table S3) were introduced into p5RT70DDmycGFP-HXGPRT [25] using indicated restriction enzymes (Table S3). For the generation of trans dominant Rab-GTPases point mutations were introduced using megaprimer method [49] with indicated oligonucleotides (Table S3). For expression of N-terminal tagged TyRab5A the ddFKBPmyc tag in construct ddFKBPmyc-Rab5A was exchanged for the Ty-tag using EcoRI/NsiI.
HFF monolayers grown on 24 well coverslips were infected with tachyzoites of the strains to be analysed and grown in presence or absence of 1 µM Shield-1 for 12–24 hrs. Immunofluorescence analysis was performed as described [23]. Z-stack images of 0.15 µm increment were collected on a PerkinElmer Ultra-View spinning disc confocal Nikon Ti inverted microscope, using a 100× NA 1.6 oil immersion lens or DeltaVision Core microscope. For deconvolution Huygens or softWoRx software was used. For image processing ImageJ 1.44 and Adobe Photoshop CS4 were utilised. The Pearson's correlation coefficient was calculated using ImageJ or softWoRx software.
Freshly lysed extracellular parasites were incubated in culture media in presence and absence of 1 µM Shld-1 and incubated for 4–8 hrs. Parasite pellets (corresponding to 2×106 to 5×106 parasites) were analysed with indicated antibodies as described [25].
The following previously described primary antibodies have been used in these analysis: α-IMC1 [50], α- AMA1 [36], Gra9 [51], α-Rop5 [52], α-Rop2,3,4 [53], α-MIC3 [54], α-proMIC3 [31], α-MIC8 [55], α-MIC2 [56], α-M2AP [20], α-MIC11 [57], α-PLP1 [58], α-CPL [22], α-proM2AP [20], α-VP1[21], α-Catalase [59], GRASP-RFP, TgERD2-GFP, GalNac-YFP [60], FNR-RFP [61], HSP60-RFP [61]
Plaque, Replication, Invasion and Egress assays were performed as previously described [23], [62]
Parasites were grown for 24 hours in presence or absence of inducer. Pulse-chase analysis of MIC3 processing has been performed as previously described [22].
Monolayers were infected with parasites and cultured for 12 and 24 hours in the presence or absence of 1 µM Shld-1 prior to fixation for routine electron microscopy as described previously [23].
Thin sections of 100 nm were processed as described previously [63]. Briefly, immunolabelled (see below) parasites in HFF monolayers on 12 mm coverslips were embedded in melamine. After polymerization the block with the parasites-containing cells was detached from the glass coverslip and processed with an ultramicrotome (EM UC6, Leica Microsys- tems GmbH, Wetzlar, Germany) into sections of 100 nm thickness.
STED immunofluorescence analysis with intracellular parasites was performed using α-mouse ATTO 565 and α-rabbit Dyomics 485 (DY 485 XL) secondary antibodies. STED immunofluorescence analysis of 100 nm thin-sections was performed using the dye pairs ATTO 594 and ATTO 647N (ATTO-TEC, Siegen, Germany) coupled to α-rabbit and/or α-mouse secondary antibodies (Dianova, Hamburg, Germany). The whole-mount samples and thin-sections were embedded in Mowiol 4-88/DABCO mounting media.
The two-colour STED measurements of whole-mount samples were performed on a home-built setup. The two colour channels were realised using the dyes DY 485 XL (excited at 470 nm) and ATTO 565 (excited at 532 nm). Both fluorophores can be efficiently silenced by the same STED wavelength at 647 nm due to the large Stoke's shift of DY 485 XL. Two pulsed laser diodes served as excitation sources (Picoquant, Berlin, Germany) which were triggered by the STED pulses – generated by an actively mode locked (APE, Berlin, Germany) Ar-Kr laser (Spectra Physics-Division of Newport Corporation, Irvine, USA). The synchronized pulses were combined using acousto-optical tunable filters (AOTF) (Crystal Technologies, Palo Alto, USA) and coupled into a microscope (DMI 4000B with an objective lens ACS APO 63x/1.3NA, Leica Microsystems GmbH, Mannheim, Germany) equipped with a three axis piezo stage-scanner (PI, Karlsruhe, Germany) which also imaged the fluorescence signal onto a confocally arranged aperture of a photon counting module (SPCM-AQR-13-FC, PerkinElmer, Canada). The AOTFs also served as fast shutters and independent power controllers for each laser beam as well as a filter system selecting the fluorescence signal. For additional filtering, a band-pass filter (580/40, AHF Analysentechnik, Tübingen, Germany) was used. The doughnut-shaped intensity profile of the STED focus was generated by inserting a glass phase plate (RPC Photonics, NY, USA) which induced a helical phase ramp from 0 to 2 on the initially flat wave front.
Two-colour STED imaging of melamine sections was performed on a home-built setup as previously described [64]. Two excitation lasers at 570 nm +/− 5 nm (for ATTO 594) and 647 nm +/− 5 nm (for ATTO 647N) originated from a single supercontinuum laser source while two additional high-power STED lasers at 711 nm +/− 3 nm and 745 nm +/− 3 nm, respectively, emerged from the same laser source (SC-450-PP-HE, Fianium, Ltd., Southampton, UK). A pulse-interleaved acquisition scheme was used to image both colour channels in a quasi-simultaneous recording mode (25 ns time-shift). The fluorescent signals were detected by two high-sensitive avalanche photodiodes at 620 nm +/− 20 nm (ATTO 594) and 670 nm +/− 15 nm (ATTO 647N). STED images were deconvolved by a linear deconvolution algorithm using the Software ImSpector (www.imspector.de). The Pearson's correlation coefficient (Figure 6C) was calculated using Image J.
Rab-GTPases from apicomplexan parasites have been previously identified [14]. Orthologue groups were obtained from OrthoMCL DB (http://www.orthomcl.org/cgi-bin/OrthoMclWeb.cgi). From each orthologue group sequences were collected if they were derived from the following species groups: Canonical species – representative species from the major groups of the eukaryotes (Baldauf) [except for Rhizaria – no genomic data set] Homo sapiens (Opisthokonts), Dictyostelium discoideum (amoebazoa), Arabidopsis thaliana (archaeplastidae,viridiplantae), Plasmodium falciparum (alveolates), Thalassiosira pseudonana (stamenopiles), Trypanosoma brucei (discicristates), Giardia lamblia (excavates), Species drawn from the Alveolate group: Plasmodium bergei, Cryptosproridium parvum, Neospora caninum, Theileria anuulata, Toxoplasma gondii, Red algae (C. merolae). Candidate sequences were assessed for the presence of ras domains using the hmmsearch option of the HMMER package using the profile PF00071 from Pfam. Sequences were checked for the presence of potential prenylation or myristilation sites in the C-terminal region. The ras domains from the conformant sequences were identified by hmmalign option of the HMMER package and extracted as described previously [65]. Blastp searches of the conformant sequences were performed agains the appropriate Uniprot KB species specific proteome sets. Multiple sequence alignment of the set of ras domains was performed by three independent methods; ClustalW [66], t-coffee [67] and hmmalign [68] guided by the appropriate profile. The alignments used the default settings for each method. Alignments were combined under t-coffee, and quality of alignment assessed - columns displaying low consistency (score < 5) or significant numbers of gaps (> 15%) were removed. The phylogeny was visualised as an unrooted neighbour joining tree by the Splits-Tree program. Major clades containing T.gondii sequences were identified, and the alignments of sequences within these clades extracted. The phylogenies were visualised as rooted Neighbour Joining phylograms, using an appropriate C. merolae sequences as outgroup. The robustness of the phylogeny was established by bootstrap analysis (1000 iterations).
|
10.1371/journal.ppat.1005126 | Extracellular Adenosine Protects against Streptococcus pneumoniae Lung Infection by Regulating Pulmonary Neutrophil Recruitment | An important determinant of disease following Streptococcus pneumoniae (pneumococcus) lung infection is pulmonary inflammation mediated by polymorphonuclear leukocytes (PMNs). We found that upon intratracheal challenge of mice, recruitment of PMNs into the lungs within the first 3 hours coincided with decreased pulmonary pneumococci, whereas large numbers of pulmonary PMNs beyond 12 hours correlated with a greater bacterial burden. Indeed, mice that survived infection largely resolved inflammation by 72 hours, and PMN depletion at peak infiltration, i.e. 18 hours post-infection, lowered bacterial numbers and enhanced survival. We investigated host signaling pathways that influence both pneumococcus clearance and pulmonary inflammation. Pharmacologic inhibition and/or genetic ablation of enzymes that generate extracellular adenosine (EAD) (e.g. the ectoenzyme CD73) or degrade EAD (e.g. adenosine deaminase) revealed that EAD dramatically increases murine resistance to S. pneumoniae lung infection. Moreover, adenosine diminished PMN movement across endothelial monolayers in vitro, and although inhibition or deficiency of CD73 had no discernible impact on PMN recruitment within the first 6 hours after intratracheal inoculation of mice, these measures enhanced PMN numbers in the pulmonary interstitium after 18 hours of infection, culminating in dramatically elevated numbers of pulmonary PMNs at three days post-infection. When assessed at this time point, CD73-/- mice displayed increased levels of cellular factors that promote leukocyte migration, such as CXCL2 chemokine in the murine lung, as well as CXCR2 and β-2 integrin on the surface of pulmonary PMNs. The enhanced pneumococcal susceptibility of CD73-/- mice was significantly reversed by PMN depletion following infection, suggesting that EAD-mediated resistance is largely mediated by its effects on PMNs. Finally, CD73-inhibition diminished the ability of PMNs to kill pneumococci in vitro, suggesting that EAD alters both the recruitment and bacteriocidal function of PMNs. The EAD-pathway may provide a therapeutic target for regulating potentially harmful inflammatory host responses during Gram-positive bacterial pneumonia.
| Despite the presence of vaccines and antibiotic therapies, invasive Streptococcus pneumoniae (pneumococcus) infections, such as pneumonia, bacteremia and meningitis, remain a leading cause of mortality and morbidity worldwide. Understanding the host factors that influence the outcome of S. pneumoniae infection will allow us to design better therapies. Here, we elucidate the role of rapidly responding innate immune cells termed neutrophils, or PMNs (polymorphonuclear leukocytes), whose role in S. pneumoniae infection has long been controversial. We found that PMNs are initially required for controlling bacterial numbers, but their extended presence in the lungs leads to significant damage and poor control of infection. The signals that control the movement of PMNs into the infected lungs are not well understood. Here, we identified extracellular adenosine (EAD), a molecule produced by the host in response to cellular damage, as important in limiting PMN movement into the lungs upon pneumococcal challenge. Importantly, EAD-mediated control of PMNs was crucial for fighting lung infection by S. pneumoniae. This study may lead to the potential use of clinically available adenosine-based therapies to combat pneumococcal pneumonia in the future.
| Despite vaccines and antibiotic therapies, invasive Streptococcus pneumoniae (pneumococcus) infections such as pneumonia, meningitis and bacteremia remain a considerable health and economic burden [1,2]. A major determinant of disease following S. pneumoniae infection is pulmonary inflammation, which, if excessive, can result in tissue destruction, compromised gas exchange, and/or acute respiratory distress syndrome [3]. Many conditions associated with enhanced inflammation, including influenza infection [4–6] and aging [7,8], lead to increased susceptibility to pneumococcal pneumonia.
Effective inflammatory responses to infection balance host defense with the potentially competing demand of a rapid return to homeostasis. Indeed, pneumococcal pneumonia triggers a massive neutrophil, or polymorphonuclear leukocyte (PMN), influx into the alveolar spaces [9,10], but the role of these innate immune cells during infection is complex. Several findings suggest that PMNs are needed to control the infection: neutropenic patients are at increased risk for pneumonia [11], and in several mouse studies, depletion of PMNs prior to S. pneumoniae infection [12,13] or delay in PMN recruitment into the lungs [14,15] resulted in higher pulmonary bacterial loads and lethal septicemia. Paradoxically however, conditions associated with increased numbers of PMNs in the lungs several days after S. pneumoniae infection of mice, such as advanced age [8,16], deficiency in regulatory T cells [17], or influenza infection [18], result in more severe systemic infection and reduced survival. Conversely, reducing PMN influx into mouse airways dramatically decreases bacteremia, resulting in uniform survival to a normally lethal pneumococcal pulmonary challenge [9]. These findings suggest that host survival may require an initial acute PMN response that is rapidly resolved later in the course of S. pneumoniae infection.
To reach S. pneumoniae in alveolar spaces, circulating PMNs cross the endothelium, enter into the interstitial space, then breach the lung epithelium to access the airway spaces [19]. This complex process involves multiple pathways of chemotaxis, including those mediated by eicosanoids [9] or chemokines [19] [20], as well as a network of ligand-receptor interactions, including those mediated by lectins or integrins [15]. Although many studies have focused on positive regulators of PMN recruitment into the lungs following pneumococcal challenge [9,14,15], signals that negatively regulate this process and ultimately promote resolution of this response are poorly understood.
Extracellular adenosine (EAD) is a potentially crucial regulator of PMN-mediated pulmonary inflammation. Basal EAD levels in tissues are typically low (<1μM) [21], but can increase more than ten-fold during pathological conditions [22]. Upon cellular insult, such as infection [23], ATP is released from cells and metabolized to adenosine by the sequential action of two extracellular enzymes, CD39, which converts ATP to AMP, and CD73, an ecto-5’-nucleotidase that de-phosphorylates AMP to EAD [22]. EAD is recognized by four G-protein coupled receptors, A1, A2A, A2B and A3 [23] leading to enhanced or diminished acute inflammation, depending on the target receptor, cell type, and/or EAD concentration [23]. Thus, the EAD pathway may provide a means for complex regulation of PMN movement [22].
Several non-infectious acute pulmonary injury models indicate that EAD generated by endothelial cell CD73 binds to cognate adenosine receptors on PMNs, leading to reduced PMN-endothelial cell adhesion, inflammation, and tissue damage [24–26]. Lung epithelial cells are both an important EAD source [25] and, given that they produce all four adenosine receptors [21], a potential EAD target. CD73-/- mice show impaired clearance of bacteremia and enhanced pulmonary inflammation in a cecal puncture model [27], whereas deficiency of adenosine A2B or A1 receptors was protective against Klebsiella pneumoniae [28] or influenza lung infection [29], respectively. Thus, the role of EAD in pathogen lung burden, inflammation, and injury during bacterial infection is not fully characterized.
In this study, we characterized the kinetics of PMN entry into the lung during murine pneumococcal challenge with an invasive S. pneumoniae strain, and addressed potential beneficial and detrimental roles of PMNs in disease. We found that PMNs promoted microbial control early, but inhibited bacterial clearance later in infection. We identified the EAD pathway as a regulator of endothelial transmigration and PMN recruitment into the lung at later time points after pneumococcal infection, as well as PMNs anti- pneumococci function. This study is a first step in elucidating the potentially complex role of the EAD-pathway in regulating pulmonary inflammation and host defense against Gram-positive bacterial pneumonia.
To better understand the role of PMNs following pneumococcal infection, C57BL/6 mice were infected intra-tracheally (I.T.) with 5x105 colony-forming units (CFU) of S. pneumoniae TIGR4 strain and pulmonary PMN influx and bacterial burdens in the lungs and blood were monitored for 72 hours. The total number of pulmonary PMNs, determined by flow cytometric analysis of a single-cell suspension of whole lung, increased four-fold in the first three hours post-infection, then underwent a dramatic increase, peaking at 30 million, i.e. ~100-fold greater than uninfected controls, at 18 hours post-infection (Fig 1A). Between 24 and 72 hours post-infection, as mice started to succumb to the disease, surviving mice experienced an ~10-fold decrease in pulmonary PMNs (Fig 1A).
Quantitation of bacterial numbers in the lung revealed two phases of infection control. In spite of the fact that S. pneumoniae TIGR4 is a virulent strain capable of replication in the murine lung [30], bacterial numbers in the lung decreased ~30-fold in the first 12 hours of infection, a period in which PMN numbers increased dramatically (Fig 1A). However, between 12 and 18 hours, during which mice continued to experience a striking increase in pulmonary PMNs, bacterial lung burden increased 5-fold to approximately 2 x 105, and this level of infection or higher was maintained for the remainder of the 72-hour experiment. Moreover, the three-fold increase in pulmonary PMNs, peaking at 18 hours post-challenge correlated with a large increase in bacterial numbers in the circulation, with titers of more than 104/ml, consistent with our previous findings that PMN entry into the lung facilitates bacterial spread [9]. Over the next 30 hours, a majority of infected mice succumbed to infection (Fig 1A) and even among survivors, bacterial titers in the blood increased 100-fold to over a million CFU/ml. Thus, although the initial increase in PMN influx into the lungs corresponded to a transient control of infection during the first 12 hours, the further accumulation of PMNs after this time point, peaking at 18 hours post-infection, coincided with the development of serious systemic infection.
To experimentally address the role of PMNs during lung infection by S. pneumoniae, we depleted PMNs with intra-peritoneal (i.p.) injections of the anti-Ly6G antibody (IA8) either one day before I.T. infection with ~5x105 colony forming units (CFU) of S. pneumoniae TIGR4 strain, or 18 hours post-infection (see Methods), a time point that corresponded to peak pulmonary infiltration by PMNs (Fig 1A). At both time points, treatment with the anti-Ly6G antibody resulted in >90% depletion of lung and circulating PMNs compared to isotype-treated controls (see Methods). Survival and bacteremia, as well bacterial burdens in the lungs and blood at day three following infection were compared between the groups (Fig 1). Consistent with previous reports [12,13], mice depleted of PMNs pre-infection were extremely susceptible to S. pneumoniae (Fig 1B). In comparison to the matched isotype-treated control group, the pre-depleted mice suffered more than a thousand-fold greater bacterial load in the bloodstream (Fig 1C and 1D), and failed to survive the infection (Fig 1B).
In contrast, depletion of PMNs at 18 hours post-infection significantly increased the survival rate (Fig 1B) and lowered bacterial burdens a hundred-fold in both the lungs and blood (Fig 1C and 1D). Our findings strongly support the hypothesis that while PMNs are required for bacterial control at the beginning of pneumococcal infection, their persistence following infection is detrimental to the host.
A potentially crucial regulator of PMN-mediated pulmonary inflammation is extracellular adenosine (EAD; [23]). To test its role in resistance against serious infection by S. pneumoniae, we first inhibited adenosine deaminase (ADA), an enzyme responsible for the breakdown of adenosine [31]. Mice were subjected to i.p. injections of EHNA hydrochloride, a pharmacological inhibitor of ADA [32] that was previously shown to increase EAD levels in mice [33]. The mice were then challenged I.T. with 5x105 CFU, a normally lethal dose of S. pneumoniae, and bacterial burdens in the lung and blood were determined three days post-infection. Whereas mock-treated mice suffered high levels of bacteria in the lungs and blood, ADA-inhibited mice had on average ten thousand-fold fewer pneumococci in the lungs (Fig 2A) and were free of detectable bacteremia (Fig 2B). By day 3 post-infection, 40% of the mock-treated mice had succumbed to the infection, compared to 10% of the ADA-inhibited mice (Fig 2C). Although in some infection models, adenosine-mediated protection is due to the direct effects of adenosine on the infectious agent [34], we found that adenosine concentrations typically present in inflamed tissues [22] had no effect on bacterial growth in vitro (S1 Fig).
Since the ADA inhibitor EHNA hydrochloride may also target other enzymes [35], we tested whether the protective phenotype was dependent on adenosine signaling. Adenosine receptor blockade partially reversed the protective effect of inhibition of adenosine breakdown by ADA (S2 Fig), consistent with the hypothesis that the protection we observed upon treatment with EHNA hydrochloride is at least partly mediated via the interaction between adenosine and its receptors.
A prediction of the hypothesis that EAD is responsible for the protective effect of ADA inhibition is that inhibition of adenosine production should enhance susceptibility to infection. To test that, wild-type C57BL/6 mice were either mock-treated or injected intra-peritoneally with the CD73 inhibitor α,β methylene ADP which was shown to drastically lower adenosine levels in mice [24]. The mice were then challenged I.T. with 5x103 S. pneumoniae TIGR4, a dose ~2-fold below the LD50. Neither the addition of adenosine or α,β methylene ADP had any effect on S. pneumoniae growth in vitro (S1 Fig). However, at day 3 post-infection, mice treated with the inhibitor suffered approximately a million-fold higher bacterial burden in their lungs compared to mock-treated controls (Fig 3A). In addition, whereas mock-treated mice suffered low-level bacteremia that eventually resolved, CD73-inhibited mice suffered bacteremia that reached a million CFU per milliliter of blood at day 3 post-infection (Fig 3B). To determine if CD73 inhibition resulted in enhanced bacteremia in a mouse strain more resistant to S. pneumoniae infection [36], we I.T. challenged either mock-treated or CD73-inhibited BALB/c mice and found that CD73 inhibition was associated with a 100- to 1000-fold increase in bacteremia after day 3 post-infection (S3 Fig). In C57BL/6 mice, CD73 inhibition was also associated with apparent neurological dysfunction, such as hind limb twitching, weakness, paralysis and an inability to walk normally. By day 4 post-infection, 85% of CD73-inhibited C57BL/6 mice succumbed to the infection (Fig 3C).
To test whether the increased systemic spread of pneumococci was simply a reflection of increased bacterial loads in the lungs, we challenged mice with a high dose of S. pneumoniae, i.e. 2x107 CFU. CD73-inhibited mice suffered only a 1.1-fold higher (and statistically indistinguishable) bacterial lung burden than mock-treated mice (Fig 3D). Despite similar numbers of bacteria in the lung, CD73-inhibition resulted in 1000-fold higher levels of bacteremia. Our findings suggest that in addition to impacting the ability of the host to control lung infection, CD73 inhibition also promotes systemic spread of S. pneumoniae from the lungs.
To test the role of CD73 during pneumococcal infection using genetic rather than pharmacological means, and to determine whether CD73 activity alters bacterial load early in infection, we inoculated CD73-/- mice I.T with 5x 103 S. pneumoniae and followed lung and blood CFU over time. CD73-deficiency had no significant effect on bacterial burden at either site at 6 or 18 hours post-infection (Fig 3E), suggesting that EAD does not play a major role in controlling bacterial numbers at the early stages of infection. Beyond 18 hours post-infection, CD73-/- mice were incapable of controlling pneumococcal burdens, reflected in a 100- to 1000-fold increase in both infection sites (Fig 3E). In contrast, bacterial numbers in the lung and blood of wild-type mice increased only slightly in the first 48 hours of infection and were largely cleared by 72 hours. Thus, pharmacological inhibition or genetic ablation of CD73, an enzyme required for EAD production [24], drastically increased the S. pneumoniae lung burden and susceptibility to systemic disease.
To test whether EAD-mediated protection upon pneumococcal infection was dependent on signaling via adenosine receptors in the host, mice were treated with the pan-adenosine receptor antagonist CGS-15943 [37] prior to challenge with S. pneumoniae. This inhibitor targets all four adenosine receptors, with Ki values of 3.5, 4.2, 16 and 51 nM for human A1, A2A, A2B and A3 receptors respectively [37,38]. Although CGS-15943 had no effect on the viability of S. pneumoniae in vitro (S1 Fig), treatment of mice with this inhibitor resulted in increased susceptibility to S. pneumoniae lung challenge that was virtually identical to that observed upon inhibition of CD73-mediated EAD production (Fig 3). Compared to mock-treated controls, mice treated with the adenosine receptors antagonist suffered ten thousand-fold higher bacterial loads in their lungs (Fig 3A) as well as bacteremia exceeding 103 CFU/ml (Fig 3B). The mice treated with the adenosine receptors antagonist also displayed a significant survival defect compared to mock-treated mice following pneumococcal lung challenge (Fig 3C). These findings clearly show that inhibition of adenosine receptors signaling render mice highly susceptible to pneumococcal challenge.
PMNs are recruited into the pulmonary airways via a multistep process involving first movement from the vasculature into the thin interstitial space and then across the lung epithelium into the airways. In several models, EAD limits the movement of PMNs across endothelial barriers [24,26]. To assess whether EAD targets the first step of pulmonary PMN recruitment, we measured the apical to basolateral movement of PMNs across monolayers of human umbilical vascular endothelial cells (HUVECS) grown on filter membranes in response to S. pneumoniae infection (see Methods; [39]). PMN migration was dependent on the infection of the endothelial cell monolayer by S. pneumoniae (Fig 4A), suggesting that, in this model, pneumococcal infection activates the endothelium to trigger PMN transmigration.
To test the role of EAD production on PMN migration in this system, we added the pharmacological inhibitor of CD73, α,β methylene ADP, to the media during the migration process, and found that it resulted in a significant dose-dependent increase in PMN migration in response to pneumococcal infection (Fig 4A, left panel). Importantly, a similar increase in PMN migration was observed when adenosine receptor signaling was blocked using the pan-adenosine receptor inhibitor CGS-15943 (Fig 4A, right panel).
We previously showed that blocking the movement of PMNs across the lung epithelium and into the airways protected mice against an otherwise lethal S. pneumoniae infection [9]. To test whether EAD also regulates PMN movement across this barrier, we utilized a well-established in vitro human PMN trans-epithelial migration assay [9]. As previously observed, apical pneumococcal infection of confluent polarized lung epithelial cells grown on filter membranes elicited robust basolateral to apical migration of PMNs (Fig 4B). Addition of the CD73 inhibitor, or exogenous adenosine to this assay had no significant effect on migration (Fig 4B). Together with our studies on endothelium, these results indicate that EAD negatively regulates PMN transmigration across endothelial but not epithelial monolayers in response to S. pneumoniae infection.
To test whether EAD regulates transmigration specifically across pulmonary endothelium during infection, we administered 5x103 CFU of S. pneumoniae I.T to mock-treated or CD73-inhibited mice, as well as to CD73-/- or wild type C57BL/6 mice. We assessed cellular recruitment into the lungs at day three post-infection, a time point that coincides with the resolution of pulmonary inflammation following a sub-lethal pneumococcal infection [40]. Consistent with the hypothesis that blocking EAD synthesis results in enhanced egress of PMNs from the vasculature, histological analysis of H&E stained lung sections at 3 days post-infection, revealed an increase in cellular infiltrates into the lungs of both CD73-/- mice and wild type C57BL/6 mice treated with the CD73 inhibitor (Fig 5A).
To quantify the apparent increase in pulmonary PMNs upon inhibition of EAD production or signaling, we measured pulmonary PMNs of mice that had been treated with the CD73 or the pan-adenosine receptor inhibitors and previously analyzed for lung and blood CFU in Fig 3A and 3B. Single-cell suspensions of lung tissue at day 3 post-infection were analyzed by flow cytometry after staining with antibody directed against the PMN marker Ly6G. Genetic ablation of CD73 (Fig 5D), as well as inhibition of CD73 or adenosine receptors (Fig 5B), resulted in a 6- to 8-fold increase in pulmonary PMNs, respectively, compared to mock-treated controls.
The failure of EAD to regulate PMN transmigration across human epithelial monolayers in vitro predicts that migration of PMNs into the airway spaces should be unaltered by manipulation of EAD signaling. To estimate the number of airway PMN, the number of PMNs in bronchoalveolar lavage fluid (BALF) of mock-treated and CD73-inhibited mice, or CD73-/- and wildtype control mice, three days after I.T. infection was determined by flow cytometry. In spite of the large increase in total pulmonary PMNs, no significant increase in the number of PMNs in BALF was observed upon pharmacological inhibition or genetic ablation of CD73 compared to control mice (Fig 5C). Importantly, the increase in pulmonary PMNs in the absence of CD73 was not simply a reflection of an increase in circulating PMNs, because both control and CD73-/- mice had comparable numbers of PMNs in the blood at 72 hours post-infection (S4 Fig). Our findings suggest that during pneumococcoal infection, EAD production and signaling are crucial for regulating the movement of PMNs specifically from the bloodstream across the endothelium, highlighting the differences in the regulation of PMN trafficking across the distinct endothelial and epithelial barriers.
To determine whether EAD production was important for regulating initial PMN recruitment, we assessed PMN influx into the lungs of CD73-/- and control mice infected I.T. with 5x103 CFU of S. pneumoniae in both the early and later phases of infection in the set of mice previously analyzed for lung and blood CFU in Fig 3D. The two mouse strains displayed indistinguishable numbers of pulmonary PMNs at 6 hours post-infection, indicating the CD73-deficiency had no discernable effect on PMN influx into the lung in the first few hours of infection (Fig 5D). At 18 hours post-infection, the number of pulmonary PMNs in CD73-/- mice was 1.7-fold higher (p = 0.068) than in control in wild type mice. By 72 hours post-infection, PMN numbers returned to near-baseline levels in wild type mice but had increased two-fold in CD73-/- mice, reaching numbers ~four-fold higher compared to wild type (Fig 5D). These results suggest that EAD signaling has little or no effect on PMN recruitment early (i.e. 6 hours) after inoculation, but has a dramatic effect on PMN numbers 12 hours later (i.e. at 18 hours) and beyond, thereby interfering with the resolution of pulmonary inflammation during S. pneumoniae infection.
PMN recruitment from the vasculature into the lung interstitial space involves a complex combination of chemotaxis signaling and cell adhesion molecule interactions [15,20]. In assessing whether EAD regulated some of the key molecules implicated in PMN recruitment into the lungs during pneumococcal infection, we found by ELISA that upon infection, CD73-/- mice had 7-fold higher levels of the chemokine CXCL2 in their lungs than did wild type mice (Fig 6A). Flow cytometric analysis revealed that the expression of the cognate receptor, CXCR2, was increased by 1.5-fold on the surface of PMNs from CD73-/- compared to wild type mice (Fig 6B). Similarly, levels of surface expressed β2 integrin (CD18), an adhesion molecule critical for PMN transmigration across endothelial barriers, was more than 2-fold higher on PMNs isolated from CD73-/- mice compared to wild type mice (Fig 6C). These findings suggest that EAD may be involved in regulating both chemotactic and cell adhesion steps during endothelial transmigration by PMNs.
Although CD73 inhibition resulted in enhanced recruitment of PMNs into the lungs, these PMNs failed to control infection. Indeed, by day 3 post-infection, CD73-inhibition was associated with ~100,000-fold more pulmonary pneumococci compared to mock-treatment (Fig 3). Thus, the absence of CD73 activity appeared to diminish the ability of PMNs recruited to the site of infection to clear the infection. We compared opsonophagocytic killing of pneumococci by PMNs isolated from the blood and bone marrow of CD73-inhibited or mock-treated mice. PMNs isolated from both the blood and bone marrow of CD73-inhibited mice displayed a ~5-fold defect in bacterial killing as compared to PMNs from mock-treated mice (Fig 7). These data are consistent with the suggestion that, in addition to a role for EAD in modulating transendothelial migration by PMNS, EAD may enhance bacteriocidal functions of PMNs.
To determine whether the heightened S. pneumoniae susceptibility of mice inhibited for EAD signaling was due to a dysregulated recruitment and function of PMN, CD73-/- or wildtype control mice were treated with the Ly6G antibody 18 hours after I.T. infection with 5x103 CFU of S. pneumoniae and bacterial burdens in the lungs, as well as spread to the blood were assessed. Treatment with the Ly6G antibodies resulted in ~80% PMN reduction in CD73-/- mice at day 3 post-infection compared to isotype-treated controls. In comparison to the untreated CD73-/- mice, PMN depleted CD73-/- mice had significantly lower bacterial burdens in both the lung and the blood, statistically indistinguishable (albeit slightly higher) from those of wildtype control mice (Fig 8). Our data suggest that the increased susceptibility of mice diminished for EAD signaling during pneumococcal infection is at least in part mediated by PMNs.
Acute inflammation following microbial infection may have either beneficial or detrimental effects. We investigated the role of PMNs in shaping the course of disease caused by the global pathogen S. pneumoniae. We first showed that within the first 12 hours after I.T. inoculation of mice, PMN entry into the lungs correlates with initial control of pulmonary bacterial burdens and that depletion of PMNs prior to pulmonary challenge with S. pneumoniae results in increased susceptibility and lethal septicemia. Although S. pneumoniae strains are quite heterogeneous, and PMN depletion enhances survival during murine infection by a serotype 8 pneumococcal strain [41], our current findings with strain TIGR4, a serotype 4 strain, are consistent with previous work indicating that PMNs, which phagocytose and kill pneumococci [42], are crucial for host defense against many serotypes of S. pneumoniae [13–15,43].
We also found that in the next phase of infection, beginning at approximately twelve hours after inoculation, PMN influx into the lungs corresponded with increased bacterial lung burdens and pneumoccocal spread into the systemic circulation. Depletion of PMNs 18 h after pulmonary challenge resulted in lower bacterial loads and enhanced survival, suggesting that timely resolution of inflammation may diminish deleterious effects of an over-exuberant host response. Indeed, mice that survived infection had drastically fewer pulmonary PMNs at day 3 post-infection, and many studies have shown that conditions that result in increased numbers of PMNs in the lungs several days after S. pneumoniae lung infection, such as influenza virus infection [18], aging [8,16,44] or deficiency in regulatory T cells [17], suffer more severe systemic spread and reduced survival. Conversely, reducing chemotaxis of PMNs into airways after I.T. pneumococcal challenge of mice resulted in uniform survival after an otherwise lethal pneumococcal pulmonary challenge [9]. Mice protected from S. pneumoniae challenge by treatment with anti-capsular antibody experience only transient influx in PMNs into the lung followed by resolution by 24 hours post-infection [45]. Thus, although PMNs are initially needed to clear S. pneumoniae infection, later in infection they function in ways that are detrimental to the host, suggesting that regulation of PMN influx is crucial to protect against disease.
Although EAD is a crucial regulator of acute pulmonary inflammation in several sterile lung injury models [24,25,46,47], its role in infection-induced inflammation remains relatively unexplored. EAD is recognized by four distinct adenosine receptors, termed A1, A2A, A2B and A3, and stimulation of a particular adenosine receptor may have a positive or negative effect on pulmonary inflammation depending on the type of lung injury [23]. A1 receptor stimulation diminished PMN infiltration and tissue damage in murine lung injury models [48–50] but promoted damaging lung inflammation during influenza infection [29]. Stimulation of the A2B adenosine receptor blocked LPS-mediated PMN recruitment into the lungs in mice [46,47,51], but had no effect on leukocyte recruitment following pulmonary infection by the Gram-negative bacterium K. pneumoniae [28]. In the context of the Gram-positive pathogen, S. pneumoniae, we found here that EAD negatively regulates trans-endothelial migration in vitro, and inhibition of EAD signaling by pan-adenosine receptor blockade, or by genetic ablation or chemical inhibition of CD73, resulted in a four- to 20-fold fold increase in pulmonary PMNs three days following I.T. pneumococcal challenge. Adenosine enhanced basolateral-to-apical transmigration of PMNs across endothelial monolayers in vitro, but did not regulate PMN migration across epithelial monolayers. Correspondingly, the increase in pulmonary PMNs during murine infection was not reflected in an increase in airway PMNs, as sampled by bronchoalveolar lavage. Thus, similar to previous findings after A2B receptor inhibition in an LPS-induced lung injury model [51], upon disruption of EAD signaling, PMNs accumulated predominantly in the interstitium.
The mechanism by which EAD modulates PMN transendothelial migration during pneumococcal infection could involve chemotactic signals or molecules that directly mediate PMN-endothelial cell interactions, or both. In vitro, the production of the chemokine CXCL-8 (IL-8) by endothelial monolayers is diminished by adenosine [52], and in a murine LPS-induced lung injury model, the level of CXCL2/3 (i.e. the murine paralog of IL-8), a chemokine that promotes PMN and macrophage recruitment during murine pneumococcal infection [20], is diminished by A1 receptor stimulation [48]. On activated PMNs, adenosine inhibits up-regulation of the β2 integrin CD11b/CD18 [53], which has been implicated in pulmonary PMN recruitment during pneumococcal murine infection[15]. We found that, following pneumococcal infection, the level of pulmonary CXCL-2 was significantly elevated in CD73-/- mice compared to wild-type mice. In addition, levels of CXCR-2 (i.e. the CXCL-2 receptor) and the integrin CD18 were elevated on CD73-/- PMNs. Thus, EAD likely regulates multiple signals involved in pulmonary recruitment of PMNs in response to S. pneumoniae infection.
A striking finding was that disruption of EAD production or signaling resulted in an increase of many orders of magnitude in bacterial numbers in the lung and blood, as well as significantly higher mortality rates. Conversely, inhibition of EAD breakdown decreased bacterial loads and diminished lethality. Although we cannot rule out that altering extracellular ATP or adenosine levels in the host may have direct effects on S. pneumoniae, especially given their far ranging metabolic and/or regulatory effects on pneumococcus [5], neither adenosine nor the ADA or CD73 inhibitors altered S. pneumoniae viability in vitro. EAD can regulate PMN phagocytosis and degranulation in vitro [22], features that are crucial for anti-pneumococcal activity of PMNs [42]. Interestingly, several Gram-positive pathogens (although likely not S. pneumoniae) express ectonucleotidases that produce EAD that inhibits PMN-mediated phagocytosis [54] and oxidative killing [55]. A2B-deficient PMNs form neutrophil extracellular traps (NETs) and clear K. pneumoniae more efficiently than wild type PMNs [28]. In contrast, here we found that pharmacologic blockade of CD73 impaired opsonophagocytic killing of S. pneumoniae by PMNs ex vivo. Phagocytic killing of pneumococci by PMNs requires serine proteases but is independent of oxidative burst [13,42], raising the possibility that the effect of EAD on a given infection may depend on the specific mechanism(s) by which PMNs kill the particular infecting microbe.
Importantly, however, the effects of CD73 ablation or inhibition and adenosine signaling blockade on lung infection cannot be fully explained by the loss of a putative PMN defense function, because depletion of these cells 18 hours after inoculation significantly mitigated the susceptibility of CD73-/- mice. Thus, EAD appears to limit disease by blunting the detrimental effect of PMNs later in infection. The nature of this PMN-mediated harmful effect on immune control is unknown, but it is possible that once bacterial burden reaches a threshold beyond which PMNs can no longer control the infection, they instead contribute to an environment permissive for bacterial persistence and growth. Some pathogens, such as Salmonella enterica, harbor metabolic capacities well adapted to the inflamed environment [56], and given that sugar utilization and other metabolic pathways have been shown to be critical determinants of pneumococcal virulence in vivo [57], PMN-derived products in inflamed tissue might make available growth-limiting nutrients utilized by this organism [58]. PMNs are also known to modulate other arms of the host immune response, such as the recruitment and function of T cells [59] and monocytes [60], and may influence pneumococcal persistence indirectly. Finally, although PMN depletion following infection significantly mitigated the susceptibility of CD73-/- mice, these mice still suffered somewhat (albeit not statistical significant) higher bacterial burdens than wild-type mice. Thus, EAD, which regulates the function of immune cells such as macrophages [61] and regulatory T- cells [62,63] that promote pneumococcal defense [17,64], may also enhance resistance by PMN-independent mechanisms.
In addition to reducing bacterial burden in the lungs, we found a strong correlation between pulmonary inflammation and systemic spread. Inhibition of EAD production or receptor signaling resulted in high levels of both pulmonary inflammation and bacteremia, whereas PMN depletion 18 hours post-infection or chemical inhibition of adenosine breakdown reduced bacterial spread. Although the reduced spread may partially reflect lower bacterial burden in the lung, CD73-inhibited mice challenged with a high (107) dose of pneumococci harbored numbers of bacteria in the lung equivalent to untreated controls, yet suffered greater bloodstream spread. In other infection models, PMN influx into infected tissues was associated with tissue damage and poor infection outcome, without altering pathogen numbers [65]. In addition, we previously showed that transmigration of PMN across a respiratory epithelial monolayer disrupted its barrier function in vitro and inhibition of PMN influx into the airways prevented lethal septicemia in mice [9]. Given that we found here that EAD signaling controls transmigration across endothelium but not epithelium, inflammation may promote disseminated pneumococcal disease by multiple mechanisms.
All four EAD receptors are produced in the lung [21] and on PMNs [22], and in future studies it will be essential to characterize the adenosine receptor(s) that influence the course of pneumococcal infection. Adenosine receptors vary in both their effect on pulmonary inflammation and their affinity for adenosine, with EC50‘s varying from <0.5 to 64μM, raising the possibility that EAD could be pro- or anti-inflammatory depending on EAD tissue concentration. Previous studies indicate that administration of the ADA inhibitor EHNA-hydrochloride and the CD73 inhibitor α,β methylene ADP to mice results in the predicted effects on adenosine [24,33] concentration, but we did not directly measure changes in EAD levels during pneumococcal infection. Changes in the expression of adenosine receptors [61] could also raise another dynamic variable that may influence EAD signaling. Adenosine receptor signaling resulted in either a pro-, or anti-inflammatory T-cell response during autoimmune uveitis depending on the phase of the disease [66], and one might imagine that the effect of EAD signaling may differ with phase of pneumococcal infection, providing a rationale for the lack of discernable effect of CD73 inhibition or ablation soon after I.T. inoculation, but a dramatic effect on the resolution of pulmonary inflammation later in infection. The use of receptor-specific agonists and antagonist or mice that are genetically ablated for a specific adenosine receptor provide future avenues to better define specific pathways that control inflammation and disease during pneumococcal infection, potentially revealing new therapeutic strategies to combat this important disease.
This work was performed in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals published by the National Institutes of Health. All procedures were reviewed and approved by Tufts University Institutional Animal Care and Use Committee (IACUC) and are under protocol # B2014-86.
Wild type BALB/c/By/J (BALB/c), C57BL/6J (B6) and CD73 knockout (CD73-/-) mice on a B6 background were purchased from The Jackson Laboratory (Bar Harbor, ME) and bred at Tufts University. Mice were matched for age and sex and maintained in a specific-pathogen free facility at Tufts University.
Mice were challenged I.T. with S. pneumoniae TIGR4 grown at 37°C in 5% CO2 in Todd-Hewitt broth (BD Biosciences) supplemented with 0.5% yeast extract (THY) and oxyrase as previously described [9]. For every experiment, the inoculum was plated on blood agar plates for CFU enumeration. If the bacterial inoculae differed by less than 20%, the data from separate experiments were pooled. If the titers varied by more 20% between individual experiments, then representative data are shown.
The effect of EAD on infection was assessed using the following: The selective and competitive inhibitor of CD73, α,β methylene ADP; the pan-adenosine receptor antagonist CGS-15943; and the adenosine deaminase inhibitor EHNA hydrochloride. All chemicals were purchased from Sigma Aldrich, dissolved in DMSO and filter sterilized by passing through a 0.22μm filter. The mice were then given intraperitoneal (i.p.) injections of 10mg/kg daily at days 0 (immediately before I.T. infection), 1 and 2 post-infection. Control mice were mock-treated with the vehicle control.
For enumeration of bacterial numbers, lung and blood samples were harvested from the live mice and plated on TSA plates supplemented with 5% sheep blood agar (Northeast Laboratory Services). The limit of detection was 20 CFU (1.3 Log10) per lung and 200 CFU (2.3 Log10) per ml blood. When no colonies were detected on the plates, the numbers of bacteria were assumed to be slightly under our limit of detection (2.0 Log10 bacteria per 1ml of blood and 1.0 Log10 bacteria in the lungs). For survival analysis, the mice were monitored for 7 days following infection. At a dose of 5x105 CFU S. pneumoniae TIGR4, we typically observed ~ 60% total mortality rate over the course of a week, with all deaths occurring on days 2, 3 or 4. However, the specific kinetics of death over this three-day period varied between experiments. We consistently observed an ~70% survival rate in mice inoculated with 5x103 CFU of S. pneumoniae TIGR4. Despite slight variation in kinetics between experiments, the differences between experimental groups were consistent from experiment to experiment.
Frozen aliquots of log-phase S. pneumoniae TIGR4 strain (serotype 4) were thawed, washed and diluted to an OD600 ~ 0.1 in THY liquid media supplemented with oxyrase. To measure the effect of the CD73 inhibitor on bacterial growth, 40μg/ml of the drug was added to the media. To measure the effect of adenosine on growth, the chemical (Sigma Aldrich) was added exogenously to a final concentration of 10μM or 100μM. The compounds were added at 0 h and bacterial growth in 37°C / 5% CO2 was monitored overtime by measuring OD600 and compared to growth in media alone. To measure the effect of the adenosine deaminase inhibitor and the pan-adenosine receptor inhibitor on bacterial viability, 40μg/ml of the drugs was added to the media at 37°C / 5% CO2 and two hours later bacterial viability was measured by plating on blood agar plates for CFU enumeration.
Mice were injected i.p. with 100 μg of the Ly6G-depleting antibody IA8 or isotype IgG control (BD Bioscience). For preinfection depletion, mice received one injection per day at 24 hours pre-infection, the time of infection plus 18 and 48 hours post-infection. For depletion 18 hours post-infection,mice were given one injection per day at 18 and 48 hours post-infection. Treatments resulted in depletion of >90% in wildtype mice and ~80% in CD73-/- mice of circulating and lung PMNs at day 3 post-infection as compared to isotype-treated controls.
Mice were euthanized at the indicated times post-infection and the bronchio-alveolar lavage fluid (BALF) obtained by washing the lungs with PBS. The lungs were then digested with Type II collagenase (Worthington) and DNase (Worthington) and single-cell suspensions obtained as previously described [8].
Cells were stained with anti-mouse Ly6G (clone 1A8, BD Biosciences), CD18 (Clone M18/2, Biolegend) and CXCR2 (Clone SA045E1, Biolegend) antibodies. Fluorescence intensities were measured on a FACSCalibur and at least 25,000 events for lung tissue and 10,000 events for BALF were analyzed using FlowJo.
Three days post-infection, the lungs were harvested, homogenized in sterile PBS and the resulting supernatants were used CXCL-2 concentrations using the mouse MIP2/CXCL-2 ELISA kit (Sigma-Aldrich) following the manufacturer’s protocol.
For histological analysis mice were euthanized 3 days post-infection and whole lungs were fixed in 10% buffered formalin (Sigma-Aldrich). Lungs were then embedded in paraffin, sectioned at 5μm, stained with hematoxylin and eosin (H&E) and analyzed using a Nikon eclipse TE2000-U microscope.
Healthy human volunteers were recruited in accordance to IRB protocols and signed informed consent forms. Whole blood was obtained and anticoagulated with acid citrate/dextrose. PMNs were isolated using a 2% gelatin sedimentation technique as previously described [9].
Human pulmonary mucoepidermoid carcinoma-derived NCI-H292 (H292) cells were grown on the underside of collagen-coated Transwell filters (0.33-cm2, Corning Life Sciences) in RPMI 1640 medium (ATCC) with 2 mM L-glutamine, 10% FBS, and 100 U penicillin/streptomycin following a previously described protocol [9].
Transmigration assay was performed as previously described [9] with pneumococcal infection and PMN migration time of 2.5 h. When indicated, the migration was allowed to occur in HBSS +/- EAD, the CD73 or adenosine receptors inhibitor at the indicated concentrations. PMNs that transmigrated into the apical chamber were measured by the myeloperoxidase ELISA following a well established assay [67] after their collection and lysis in 10% Triton-X 100 to release the myeloperoxidase. Myeloperoxidase ELISA of serial dilutions of known numbers of neutrophils were used to establish a standard curve, which was then used to quantitated migrated neutrophils.
Human umbilical vascular endothelial cells (HUVECs) were seeded on the inner chamber of collagen-coated Transwell filters (0.33-cm2, Corning Life Sciences) in M199 medium (Biowhittaker) supplemented with 2 mM L-glutamine, 10% FBS, 10μg/ml endothelium mitogen (Fisher), 20μg/ml heparin sodium salt (Sigma) and 100 U penicillin/streptomycin following a previously described protocol [68]. The cell monolayer was allowed to form over 4–5 days.
The PMN migration assay across the endothelium was performed as previously described [39] with the following modifications. The endothelial cells seeded on Transwells were infected for 3 h by S. pneumoniae added to the lower chamber. 5x105 PMNs were added to the upper chamber and migration was allowed to occur +/- CD73 or adenosine receptors inhibitors for 3 h at 37°C/ 5% CO2. Since this assay utilizes standard RPMI with phenol red, which precludes colorimetric assays such as MPO, the number of PMNs that migrated was determined by counting in a hemacytometer in triplicate, as previously described [39]. The inhibitors had no significant affect on cell viability within the timeframe of the assay as measured by trypan-blue exclusion.
Bone marrow PMNs were isolated from the femurs of mice as previously described [43] and enriched using Percoll (Sigma) density gradient centrifugation. For isolation of PMNs from the circulation, blood was collected by cardiac puncture using acid citrate/dextrose as an anticoagulant. PMNs were then enriched by Ficoll density gradient centrifugation in Mono-poly (MP-Biomedicals) resolving medium based on the manufacturer’s instructions. The enriched cells were ~ 85–90% Ly6G+ by flowcytometry.
The ability of PMNs to kill pneumococci was assessed ex vivo as previously described. Briefly, 200μl reactions in Hank’s buffer/0.1% gelatin consisted of 1x105 PMNs incubated with 1x102 bacteria grown to mid log phase and pre-opsonized with 20μl mouse sera. The reactions were incubated rotating for 45 minutes at 37°C. Percent killing relative to parallel incubations without PMNs was determined by plating serial dilutions on blood agar plates.
All statistical analysis was performed using Prism4 for Macintosh (Graph Pad). For analysis of survival curves, Log-rank (Mantel-Cox) test was performed. CFU data were log-transformed to normalize distribution. Student t-test was used for comparison between groups. p values less than 0.05 were considered significant. For all graphs, the mean values +/- SEM are shown.
|
10.1371/journal.pgen.1002506 | A Regulatory Network for Coordinated Flower Maturation | For self-pollinating plants to reproduce, male and female organ development must be coordinated as flowers mature. The Arabidopsis transcription factors AUXIN RESPONSE FACTOR 6 (ARF6) and ARF8 regulate this complex process by promoting petal expansion, stamen filament elongation, anther dehiscence, and gynoecium maturation, thereby ensuring that pollen released from the anthers is deposited on the stigma of a receptive gynoecium. ARF6 and ARF8 induce jasmonate production, which in turn triggers expression of MYB21 and MYB24, encoding R2R3 MYB transcription factors that promote petal and stamen growth. To understand the dynamics of this flower maturation regulatory network, we have characterized morphological, chemical, and global gene expression phenotypes of arf, myb, and jasmonate pathway mutant flowers. We found that MYB21 and MYB24 promoted not only petal and stamen development but also gynoecium growth. As well as regulating reproductive competence, both the ARF and MYB factors promoted nectary development or function and volatile sesquiterpene production, which may attract insect pollinators and/or repel pathogens. Mutants lacking jasmonate synthesis or response had decreased MYB21 expression and stamen and petal growth at the stage when flowers normally open, but had increased MYB21 expression in petals of older flowers, resulting in renewed and persistent petal expansion at later stages. Both auxin response and jasmonate synthesis promoted positive feedbacks that may ensure rapid petal and stamen growth as flowers open. MYB21 also fed back negatively on expression of jasmonate biosynthesis pathway genes to decrease flower jasmonate level, which correlated with termination of growth after flowers have opened. These dynamic feedbacks may promote timely, coordinated, and transient growth of flower organs.
| Perfect flowers have both male organs that produce and release pollen and female organs that make and harbor seeds. Flowers also often attract pollinators using visual or chemical signals. So that male, female, and pollinator attraction functions occur at the right time, flower organs must grow and mature in a coordinated fashion. In the model self-pollinating plant Arabidopsis, a transcriptional network regulates genes that ensure coordinated growth of different flower organs, as well as pollen release and gynoecium (female) competence to support pollination. This network also regulates nectary development and production of volatile chemicals that may attract or repel insects. We have studied growth, chemical signal levels, and gene expression in mutants affected in components of this network, in order to determine how flower growth is controlled. Several plant hormones act in a cascade that promotes flower maturation. Moreover, regulatory feedback loops affect the timing and extent of developmental steps. Positive feedbacks may ensure that the development of different flower organs is coordinated and rapid, whereas negative feedbacks may allow growth to cease once flowers have opened. Our results provide a framework to understand how flower opening and reproduction are coordinated in Arabidopsis and other flowering plants.
| In typical angiosperms, late in flower development, sepals open to expose the inner organs; the petals, stamen filaments, and style elongate; the anthers dehisce to release pollen; and the stigma and transmitting tract mature so as to permit pollen germination and pollen tube growth. These events often occur quite quickly, and are transient, so that flowers open and pollinate, but then stop growing. Effective reproduction therefore requires accurate coordination of these events. Variation in spatial arrangement and timing of maturation of different organs may affect the pollination mode and the mating system. In plants with self-pollinating flowers such as Arabidopsis thaliana, stamens and gynoecium grow to about the same length and mature synchronously, allowing efficient self-fertilization [1]. In outcrossing plants, differential growth of stamens and style or staggered timing of anther and gynoecium maturation can instead promote cross-pollination.
The Arabidopsis transcription factors AUXIN RESPONSE FACTOR 6 (ARF6/At1g30330) and ARF8/At5g37020 act partially redundantly to promote late stages in petal, stamen and gynoecium development. arf6 arf8 double null mutant flowers arrest at flower stage 12 as closed buds with short petals, short stamen filaments, undehisced anthers, and immature gynoecia with short stigmatic papillae and poor support of pollen tube growth, and are largely male- and female-sterile [2]–[4]. arf6 and arf8 single mutants and sesquimutants (homozygous for one mutation and heterozygous for the other) have delayed stamen filament elongation and decreased fecundity. ARF6 and ARF8 are each expressed in multiple flower tissues including sepals, petals, stamen filaments, style, transmitting tract, ovule funiculi, and nectaries [2], [4]. ARF6 and ARF8 thus act in several organs to promote the transition from closed buds to mature fertile flowers, and to ensure coordinated development of male and female organs, leading to efficient self-fertilization.
arf6-2 arf8-3 flowers have very low jasmonic acid (JA) levels and decreased expression of several jasmonate biosynthesis genes, and exogenous methyl jasmonate (MeJA) rescued the petal elongation and anther dehiscence defects, but not the stamen elongation defect or gynoecium arrest, of arf6 arf8 flowers [2], [3]. Mutants affected in jasmonate synthesis or signaling similarly have delayed stamen growth and indehiscent anthers [5]–[8]. Similarly to stamens, petals of jasmonate pathway mutants have been reported to have delayed growth [6]. However, in contrast, other groups have reported that petals of jasmonate pathway mutants are larger than those of wild-type flowers [5], [8], [9]. Jasmonates can inhibit petal expansion by activating alternative splicing of a bHLH31/BPE/At1g59640 transcript [9], [10]. arf8 mutants also had enlarged petals, suggesting that ARF8 and BPE act in a common pathway [11]. These results indicate that ARF6 and ARF8 trigger anther dehiscence by promoting jasmonate production, can promote or inhibit petal growth through jasmonate-dependent pathways, and regulate other aspects of flower maturation independently of jasmonate.
The role of jasmonate in stamen development was investigated in more detail by examining MeJA-induced global gene expression changes in the stamens of jasmonate-deficient opr3 mutant plants [8], [12]. Two closely related R2R3 MYB transcription factor genes, MYB21/At3g27810 and MYB24/At5g40350 [13], were rapidly induced by jasmonate. myb21 mutants had short stamen filaments and petals, and myb21 myb24 double mutants had indehiscent anthers. These phenotypes were not rescued by exogenous JA or MeJA application, indicating that MYB21 and MYB24 act downstream of jasmonate signaling to promote stamen and petal development [12], [14]. Gibberellin-deficient mutants also have delayed stamen development, decreased JA level, and decreased expression of MYB21, MYB24, and a third closely related gene, MYB57/At3g01530 [14]. A fourth closely related gene, MYB108/At3g06490, also contributes to stamen development partially redundantly with MYB24 [15]. MYB57 and MYB108 are also induced by jasmonate. MYB108 has also been isolated as BOTRYTIS OVERSENSITIVE 1 (BOS1), and is required for JA-mediated biotic and abiotic stress responses [16]. MYB21 and MYB24 can activate transcription, and overexpression of MYB21 or MYB24 caused defects in flower development [17]–[20]. Other genes encoding members of this clade, MYB2, MYB62, MYB78, MYB112 and MYB116, were not appreciably expressed in flowers [21].
To understand how these components interact to regulate flower maturation, we have analyzed the relative timing of flower organ growth in arf, myb, and jasmonate pathway mutants, and compared expression of MYB and jasmonate pathway genes in wild-type and mutant flowers. These analyses suggest a hierarchical regulatory pathway that triggers flower maturation, and also reveal contrasting effects of jasmonate on petal growth at different developmental stages. Analyses of global gene expression patterns in wild-type, myb21 myb24, and arf6 arf8 flowers reveal that the flower maturation network controls putative chemical attractant functions of flowers, and that both positive and negative feedback loops control auxin and jasmonate responses during flower maturation.
Before characterizing mutant phenotypes, we examined expression of MYB genes in wild-type, arf6-2 arf8-3, and jasmonate pathway mutant flowers. In wild-type flowers, MYB21 and MYB24 were first expressed at stages 11–12 shortly before flower opening, whereas in arf6-2 arf8-3 flowers MYB21 and MYB24 mRNAs were almost undetectable (Figure 1A; Figure S1A, S1C) [2]. Conversely, ARF6 and ARF8 mRNA levels were normal in myb21-5 myb24-5 flowers (Figure 1C, Table S3). MYB21 and MYB24 were also underexpressed in jasmonate-deficient aos-2 mutant inflorescence apices and in jasmonate-resistant coi1-1 apices (Figure 1B). Both MYB21 and MYB24 baseline expression levels were lower in arf6-2 arf8-3 inflorescences than in aos-2 or coi1-1 inflorescences (Figure 1B). Exogenous methyl jasmonate induced MYB21 and MYB24 genes in arf6-2 arf8-3 and aos-2 mutant inflorescence apices, but not in coi1-1 apices (Figure 1B) [12], [14], [15]. P35S:ARF6 plants that overexpress ARF6 did not have increased MYB21 mRNA level (Figure S1A); and PARF6:mARF6 plants expressing an ARF6 transgene that is immune to regulation by miR167, and which have an expanded ARF6 expression domain in the ovules [4], did not have a similarly expanded MYB21 expression domain (Figure 2F, 2G). These results suggest that ARF6 and ARF8 induce these MYB genes indirectly, at least partly by increasing jasmonate levels. Exogenous MeJA only partially restored MYB21 and MYB24 expression and stamen and petal growth in arf6-2 arf8-3 flowers (Figure 1B) [2], raising the possibility that ARF6 and ARF8 may also regulate MYB21 and MYB24 by additional jasmonate-independent mechanisms.
By in situ hybridization and using transgenic plants carrying a PMYB21:MYB21:GUS protein fusion reporter, we detected MYB21 expression in sepals, petals, the apical part of stamen filaments, the style, ovule funiculi, and nectaries of stage 13 and 14 flowers (Figure 2A–2C, 2F, 2J–2L; Figure S1D). In the aos-2 jasmonate-deficient background, PMYB21:MYB21:GUS expression was decreased in these organs, but was restored by exogenous methyl jasmonate (Figure 2M, 2N). MYB24 and PMYB24:MYB24:GUS were likewise expressed in stamen filaments, style, and nectaries of stage 13 and 14 flowers, but not in ovule funiculi (Figure 2D, 2E; Figure S1E). Available microarray expression data are consistent with expression of both MYB21 and MYB24 in sepals, petals, stamens and carpels [21]. Expression of PMYB21:MYB21:GUS and PMYB24:MYB24:GUS in anthers or pollen (Figure 2, Figure S1) is likely an artifact of our fusion constructs, because in situ hybridizations revealed stamen filament but not anther expression (Figure 2C, 2E) [14]; microarray data from dry or germinated pollen revealed no expression of MYB21 or MYB24 [22]; and X-Gluc staining was present in anthers of arf6-2 arf8-3 PMYB21:MYB21:GUS plants although arf6-2 arf8-3 flowers lacked detectable MYB21 transcript (Figure 1A, 1B; Figure S1A–S1C).
To determine timing of stamen and petal growth in flowers of different genotypes, we measured organ lengths of flowers along a developmental series from closed buds to open flowers (Figure 3) [23]. Wild-type gynoecia elongated at a fairly constant rate through these stages, so that gynoecium length provided an internal reference for developmental stage. In addition, in independent experiments we measured organ lengths of flowers at defined positions on the inflorescence relative to the position of the first open flower in wild-type plants (Table S1). In wild-type Arabidopsis flowers, sepals stopped growing at stage 12, shortly before flowers opened [1], [23]. Petals and stamens grew slowly through early stages, but grew much more rapidly at stage 12 and stage 13, when the flowers opened (Figure 3A, 3I, 3J). Wild-type flowers generally self-pollinated as they opened. Just after this stage, petals and stamens stopped elongating, and about two days thereafter they began to senesce [1], [24].
arf6-2 and arf8-3 single mutants had delayed petal and stamen growth compared to wild type, but at a slightly later stage arf6-2 and arf8-3 mutant petals and stamens did reach wild-type lengths relative to gynoecium length (Figure 3C, Figure S2A). Although arf8-3 mutants have been reported to have longer and wider petals than wild type [11], under our growth conditions petals of arf6-2 and arf8-3 flowers appeared wider but were not longer than wild-type petals. arf6-2 arf8-3 double mutant flowers arrested with short stamens, petals, and gynoecia (Figure S2A) [2], [3].
We recovered the presumed null mutations myb21-4 and myb21-5, each of which has a stop codon in the MYB21 coding sequence, in a screen for arf6-2 enhancers (Figure S3); and we used available T-DNA insertion alleles in MYB24 (Materials and Methods, Figure S4, Table S2). arf6-2 myb21-4 and arf6-2 myb21-5 plants had flower buds with small unreflexed petals and short stamens, and set seed only when manually pollinated (Figure 3D, Table S1). The myb21-5 mutation also enhanced arf8-3 phenotypes, but did not affect organ lengths of the more severely affected arf6-2 arf8-3 flowers (Table S1), indicating that MYB21 can be placed in the same genetic pathway as ARF6 and ARF8.
Similarly to other myb21 mutants [12], [14], [20], myb21-4 and myb21-5 single mutant flowers had short petals, short stamens with reduced epidermal cell length, and delayed flower opening and anther dehiscence (Figure 3B, 3I, Figure S2C, Figure S5A–S5C, Table S1). The myb21-4 and myb21-5 mutants had stronger phenotypes than the myb21-2 T-DNA insertion allele, which is in an intron and makes some full-length transcript (Table S1, Figure S4) [12]. Flowers of myb24-2 and myb24-5 single mutant plants appeared normal (Table S1, Figure 3E). Flowers of myb21-5 myb24-5 double mutants grew similarly to myb21-5 flowers up to stage 13 (Figure 3B, 3F, 3I; Table S1). However, whereas myb21-5 flowers sometimes opened, myb21-5 myb24-5 flower buds remained closed (Figure 3B, 3F). Moreover, myb21-5 myb24-5 anthers failed to release pollen until after the flowers started to senesce, and treatment with exogenous MeJA failed to accelerate pollen release (Table S1).
As well as acting in petals and stamens, MYB21 and MYB24 are expressed in the gynoecium, suggesting that they may regulate aspects of gynoecium development or function. Gynoecia of wild-type, arf6-2, and arf8-3 flowers grew to at least 4 mm long even if unpollinated (Figure S2A). Gynoecia of arf6-2 arf8-3, myb21 and myb21 myb24 flowers were shorter than wild-type gynoecia, and arrested at about 3 mm long (Table S1; Figure 3I, 3J; Figure S2A). This phenotype was largely attributable to decreased valve lengths in the mutants (Figure S6F). myb21 mutations also decreased stigma lengths, although this effect was only statistically significant for both tested myb21 alleles in myb24-5 or arf6-2/+ arf8-3 genetic backgrounds (Figure S6A–S6D, S6G). In the arf6-2/+ arf8-3 genetic background, myb21 mutations also decreased the proportion of ovules that were fertilized by wild-type pollen, from about 78% in arf6-2/ARF6 arf8-3 ovules, to just 35–40% in arf6-2/ARF6 arf8-3 myb21-4 or arf6-2/ARF6 arf8-3 myb21-5 ovules. In many poorly fertilized gynoecia, pollen tubes only entered the apical part of the transmitting tract. These stigma and fertilization phenotypes were similar to, although less severe than, those observed for arf6-2 arf8-3 plants (Figure S6E) [4].
Flowers of jasmonate-deficient (aos-2) or -insensitive (coi1-1) mutants had short stamens and indehiscent anthers similar to those of myb21 myb24 mutants (Figure 3G, 3J; Figure S2C; Table S1). Similarly, at the time of wild-type flower opening (staged according to gynoecium length), aos-2 and coi1-1 flowers had delayed petal growth just as myb21 and myb21 myb24 flowers did, indicating that jasmonates promote petal growth at stage 12 (Figure 3G, 3J; Figure S2C). However, at stages 14–15 after pollination has normally occurred in wild-type flowers, petals of aos-2 and coi1-1 flowers continued to grow, so as to become larger than wild-type petals (Figure 3G, 3J; Figure S2C). Mutant flowers also senesced later than wild-type flowers, possibly accounting in part for the prolonged growth phase of these petals.
Gynoecia and valves of aos-2 and coi1-1 mutant flowers grew slightly less than those of unfertilized wild-type flowers, but more than those of myb21 or myb21 myb24 flowers (Figure 3J, Figure S2C, Figure S6F). Stigmas of aos-2 and coi1-1 flowers were as long as those of wild-type flowers, and aos-2 and coi1-1 gynoecia supported full fertilization after being pollinated manually (Figure S6G). Thus, myb21 mutations had stronger effects on both petal and gynoecium growth than did aos-2 or coi1-1 mutations. The weaker phenotypes of aos-2 and coi1-1 than myb21 and myb21 myb24 mutants appears inconsistent with the hierarchical model in which jasmonates induce MYB genes which in turn cause petal expansion. These results might have arisen if the aos-2 and coi1-1 mutants each retain some jasmonate response. However, we detected no cis-JA in aos-2 flowers (Figure 1F), and the coi1-1 mutation is a null mutation in the only known JA-Ile receptor. Moreover, flowers of aos-2 coi1-1 double mutant plants had enlarged petals and delayed senescence as did flowers of either single mutant (data not shown), suggesting that aos-2 and coi1-1 mutations each eliminated jasmonate response in flowers.
We therefore explored in more detail how the jasmonate pathway affects MYB21 expression. In wild-type flowers, MYB21 expression was high at stage 12, and then decreased at stages 13 and 14 (Figure 4, Table S3). Whereas at stage 12 aos-2 and coi1-1 flowers had lower expression of MYB21 than did wild-type flowers, at stage 14 they had higher expression (Figure 4). Similarly, aos-2 PMYB21:MYB21:GUS plants had reduced X-Gluc staining at stage 13, but had X-Gluc staining in petals at stage 15 (Figure 2M, 2O). Thus, in both wild-type and jasmonate pathway mutant plants, petal growth correlated with MYB21 expression. Moreover, petals of myb21-4 aos-2, myb21-5 aos-2, and coi1-1 myb21-4 double mutant flowers failed to enlarge at late stages, and flower buds of these double mutants never opened (Table S1; Figure 3H, 3J; Figure S2C). Thus, MYB21 is active and promotes petal elongation in stage 14 aos-2 and coi1-1 flowers.
These analyses revealed that starting at flower stage 12, ARF6 and ARF8 promote MYB21 and MYB24 expression in multiple flower organs largely by increasing jasmonate levels. MYB21 and MYB24 in turn promote petal and stamen filament growth, anther dehiscence, and gynoecium growth and maturation, with MYB21 having a predominant role. To explore gene expression patterns underlying this regulatory hierarchy, we used Affymetrix ATH1 gene chip arrays to monitor global gene expression in wild-type, arf6-2 arf8-3 and myb21-5 myb24-5 closed buds (stage 12 flowers) and newly open flowers (stage 13). Expression data for each array probe set were compared statistically between genotypes, and in addition a two-fold expression ratio cutoff was applied to remove genes with statistically significant but small relative differences in expression level (Figure 5, Table S3). We focussed our analyses on gene expression changes at stage 12, when flowers of both double mutants have similar morphology to wild-type flowers. Stage 13 data are presented for reference (Table S3), but presumably include many indirect effects caused by developmental arrest of mutant flowers at stage 12. As most array probe sets correspond uniquely to a single gene, in the following analyses we refer to probe sets as “genes.”
At flower stage 12, 624 genes were expressed at a lower level in arf6-2 arf8-3 flowers than in wild-type flowers, and 312 genes were expressed at a higher level (Figure 5). In myb21-5 myb24-5 stage 12 flowers, 356 genes were underexpressed and 97 were overexpressed relative to wild-type flowers. Of the genes underexpressed in arf6-2 arf8-3 flowers, 33% (209/624) were also underexpressed in myb21-5 myb24-5 flowers, and 2% (14/624) were overexpressed. Of the genes overexpressed in arf6-2 arf8-3 flowers, 6% (18/312) were also overexpressed in myb21-5 myb24-5 flowers, and none was underexpressed. Thus, the myb mutations affected a greater proportion of genes that were underexpressed in arf6-2 arf8-3 flowers than of genes that were overexpressed in arf6-2 arf8-3 flowers. As MYB21 and MYB24 can activate genes [17], [18], the 209 genes underexpressed in both myb21-5 myb24-5 and arf6-2 arf8-3 flowers may include genes that the MYB proteins activate. Independent RNA blot hybridization and qRT-PCR experiments confirmed expression characteristics deduced from the array data for about 15 genes of interest (Figure 1, Figure 4, Figure S1C).
To discern patterns in the gene expression data, we compared our data to global gene expression datasets generated by other workers (Table S3). Gibberellins, acting in part through derepression of DELLA protein activity, also promote late stages of petal, stamen, and gynoecium development [14], [25]–[28]. We compared our gene expression results to a list of genes that were over- or under-expressed in ga1-3 gibberellin-deficient mutant flowers [29]. 28% (172/624) of genes that were underexpressed in arf6-2 arf8-3 flowers were also underexpressed in ga1-3 flowers, and just 1.3% (8/624) were overexpressed in ga1-3 flowers (Table S3). Similarly, 25% (77/312) of genes that were overexpressed in arf6-2 arf8-3 flowers were also overexpressed in ga1-3 flowers, and just 4.5% (14/312) were underexpressed in ga1-3 flowers. Thus, ARF6 and ARF8 and gibberellin induce and repress an overlapping set of downstream responses in flowers, in most cases in the same direction.
We used data on gene expression in wild-type stage 12 sepals, petals, stamens and carpels [21] to determine in which organs each gene affected in arf6-2 arf8-3 or myb21-5 myb24-5 flowers was expressed (Figure 5, Table S3). Although most of the affected genes were expressed in multiple flower organs, to identify trends in the data it proved convenient to bin genes according to the organ in which they had highest expression in wild-type stage 12 flowers. Of the genes that were underexpressed in arf6-2 arf8-3 flowers, substantial numbers were most highly expressed in sepals (79/624, 13%), petals (185/624, 30%), stamens (246/624, 39%) or carpels (114/624, 18%) of wild-type flowers. In contrast, of the genes that were overexpressed in arf6-2 arf8-3 flowers, over half (161/312, 52%) were most highly expressed in sepals of wild-type flowers, whereas just 22% (69/312) were most highly expressed in wild-type stamens. In myb21-5 myb24-5 flowers, 66% (234/356) of underexpressed genes had highest expression in stamens of wild-type flowers. Of the genes that were overexpressed in the myb21-5 myb24-5 flowers, an equal number had highest expression in wild-type sepals as in wild-type stamens (32/97 in each case).
Among the genes with decreased expression in arf6-2 arf8-3 flowers were several known auxin-inducible genes including IAA1, SHY2/IAA3, IAA6, AXR2/IAA7, IAA17, IAA19, SAUR9 (SMALL AUXIN UP RNA9), SAUR23, SAUR25, SAUR27, SAUR35, SAUR42, SAUR62-SAUR68, and SAUR70 (Table S3). Many of these have auxin response elements in their presumed promoters and are good candidates to be direct targets of ARF6 and ARF8 [30]. Although the hierarchical regulatory model does not predict that MYB21 or MYB24 should affect expression of direct ARF targets, several of these IAA and SAUR genes (IAA6, IAA19, SAUR9, 25, 35, 64, 66, 67, and 68) were also underexpressed in stage 12 myb21-5 myb24-5 flowers (Table S3). RNA gel blot hybridization experiments confirmed that IAA19 and SAUR63 were underexpressed in myb21-5 myb24-5 flowers, and that in addition IAA2, SHY2/IAA3 and AXR2/IAA7 were more modestly underexpressed in myb21-5 myb24-5 flowers (Figure 1D). These results suggest that MYB21 and MYB24 participate in positive feedback loops that promote ARF activity.
Additional ARF and MYB genes were underexpressed in mutant flowers, and might also constitute positive feedbacks if they share targets with ARF6 and ARF8 or MYB21. The ARF16 (At4g30080) gene encodes an Auxin Response Factor that is phylogenetically distant from ARF6 and ARF8, and regulates root cap differentiation together with its closest paralog ARF10 [31], [32]. ARF16 was underexpressed in arf6-2 arf8-3 flowers but had normal expression level in myb21-5 myb24-5 flowers. arf10-3 arf16-2 flowers, as well as P35S:MIR160c flowers overexpressing a microRNA that targets ARF10 and ARF16 [31], had delayed stamen and petal growth, similarly to arf6-2 or arf8-3 single mutant flowers (Figure S2A, S2B). These results suggest that ARF10 and ARF16 act downstream of ARF6 and ARF8 to amplify stamen and petal growth at stage 12.
Analogously, MYB57 and MYB108, closely related genes to MYB21 and MYB24, were underexpressed in arf6-2 arf8-3 and myb21-5 myb24-5 flowers (Figure 1A, 1B; Table S3). A PMYB57:MYB57:GUS reporter was expressed in stamen filaments of opened wild-type flowers (Figure S1F). A PMYB108:GUS reporter was expressed in sepals and stamen filaments, particularly in the vasculature of these organs, and in the style (Figure S1G). The myb57-1 mutant (Figure S3) had no obvious floral phenotypes (data not shown), although myb57-1 can enhance a myb21 mutation [14]. Flowers of myb108 mutants (Figure S3) had normal organ lengths at stages 12 and 13, but had slightly delayed anther dehiscence (Table S1; Figure 3I; Figure S5E, S5F) [15]. In addition, myb108-4 petals continued to grow after wild-type petals had stopped expanding, resulting in slightly longer petals at stage 14 (Figure 3I, Figure S5D–S5F). Similarly to the jasmonate pathway mutants, stage 14 myb108-4 flowers had elevated MYB21 expression, and myb21-5 myb108-4 flowers had small petals (Figure 4, Table S1). Thus, increased MYB21 expression may also cause persistent petal growth in myb108 mutants.
ARF6, ARF8, MYB21 and MYB24 are each expressed in nectaries. A previous study identified 270 genes that were preferentially expressed in nectaries [33]. Of these, 18 were underexpressed in arf6-2 arf8-3 only, 6 were underexpressed in myb21-5 myb24-5 only, and 14 were underexpressed in both mutants (Table S3). In contrast, just 5 of the nectary-enriched genes were overexpressed in either mutant. Among the underexpressed genes were CRABSCLAW (CRC/At1g69180), which is required for nectary formation [34]; YABBY5 (At2g26580) encoding a protein closely related to CRC; CWIV4 (At2g36190) encoding a cell wall invertase required for nectary sink strength and nectar production [35]; SWEET9 (At2g39060) encoding a nectary-specific glucose transporter [36]; and JMT (At1g19640) encoding S-adenosyl-L-methionine:jasmonic acid carboxyl methyltransferase, which makes the volatile compound methyl jasmonate [37]. Each of these genes was underexpressed in both arf6-2 arf8-3 and myb21-5 myb24-5 flowers, except for CRC which was underexpressed in arf6-2 arf8-3 flowers only. Consistent with these gene expression changes, nectaries in arf6-2 arf8-3 flowers were very small and only apparent by light microscopy in a fraction of flowers (Figure 6A, 6B). Nectaries in coi1-1, arf6-2 and arf8-3 single mutants and in myb21-5 myb24-5 double mutant flowers appeared normal (Figure 6C, 6D; data not shown). These morphological and gene expression results indicate that ARF6 and ARF8 affect nectary growth and function, and that MYB21 and MYB24 affect nectary gene expression but not nectary formation.
As flowers open, they emit volatile compounds, which may attract insect pollinators or predators, or may have a role in pathogen defense [38]–[41]. The Arabidopsis terpene synthase genes TPS11 (At5g44630) and TPS21 (At5g23960) synthesize a mixture of volatile sesquiterpenes emitted from flowers [42], [43]. Both genes were highly expressed in wild-type carpels, and TPS11 was also expressed in nectaries [33], [43]. Both TPS11 and TPS21 were underexpressed in arf6-2 arf8-3 and myb21-5 myb24-5 flowers (Figure 1C, Table S3). Consistent with these patterns, arf6-2 arf8-3 flowers emitted dramatically less sesquiterpenes produced by both TPS11 and TPS21 (Figure 6F). Similarly, myb21-5 flowers had strongly reduced emission of sesquiterpenes produced by TPS21 (e.g. (E)-β-caryophyllene, α-humulene), and partially reduced levels of volatile sesquiterpenes produced by TPS11 (e.g. thujopsene, β-chamigrene) (Figure 6E, 6F). These effects are consistent with the gene expression patterns, as TPS11 expression was reduced in myb21-5 myb24-5 flowers by less than was TPS21 expression (Table S3, Figure 1C). The myb24-5 mutation did not affect emission of volatile sesquiterpenes, either by itself or in combination with myb21-5 (Figure 6E). (E)-β-caryophyllene and thujopsene emissions were also reduced in flowers of the opr3 jasmonate-deficient mutant (Figure S7).
In a gene chip array dataset of gene expression in stamens of jasmonate-deficient opr3 mutant stage 12 flowers treated with exogenous methyl jasmonate (MeJA), 31 genes were induced by at least 2-fold after 30 minutes of MeJA treatment, 179 additional genes were first induced after 2 hours, and 393 more genes were first induced after 8 hours [12]. MYB21 and MYB24 were themselves induced at the two hour time point in this dataset. Consistent with their reduced jasmonate production, arf6-2 arf8-3 flowers underexpressed many of these MeJA-responsive genes, with the greatest proportional effect on the earliest MeJA-responsive genes. Thus, about 45% (14/31) of the genes induced by MeJA in stamens within 30 minutes were underexpressed in arf6-2 arf8-3 flowers (Table S3).
In the myb21-5 myb24-5 flowers, none of the early MeJA-inducible genes was underexpressed, and the proportion of MeJA-responsive genes affected was highest among those induced by MeJA at 8 hours. Of 86 late (8 hour) MeJA-inducible genes underexpressed in arf6-2 arf8-3 flowers in our experiment, 50 (58%) were also underexpressed in myb21-5 myb24-5 flowers, indicating that MYB21 and MYB24 mediate a large portion of late responses to jasmonate in flowers (Table S3).
Strikingly, 13 of the 14 genes that were underexpressed in arf6-2 arf8-3 flowers but overexpressed in myb21-5 myb24-5 flowers were MeJA-induced in stamens (Table S3). Using a less stringent 1.3-fold expression ratio cutoff, 71 genes were underexpressed in arf6-2 arf8-3 flowers and overexpressed in myb21-5 myb24-5 flowers, and 44 of these were MeJA-induced in stamens (13 of these at the earliest 0.5 h time point) (Table S4). Among these genes were MYC2 (At1g32640), which binds to jasmonate-inducible promoters to mediate induction [44]; seven JAZ genes encoding negative regulators of jasmonate response [45]; and several genes encoding known or putative enzymes in the jasmonate biosynthesis pathway. These included LOX2 (At3g45140) and LOX4 (At1g72520) encoding lipoxygenases involved in generating the fatty acid precursor [46], [47]; AOS (At5g42650) encoding allene oxide synthase [48]; OPR3 (At2g06050) encoding 12-oxophytodienoate reductase [8]; and 4CL11 (At5g38120) and 4CL9/OPCL1 (At1g20510), encoding 4-coumarate CoA ligases [49] (Table S4). RNA blot hybridization with polyA+ mRNA and qRT-PCR experiments confirmed increased expression of LOX2 and AOS in myb21-4 myb24-5 and myb21-5 myb24-5 flowers (Figure 1E, Figure 4). The phospholipase DAD1 (At2g44810) was expressed at a low level in all samples in the gene chip array experiment, but was also seen to be overexpressed in myb21-5 myb24-5 flowers by RNA blot hybridization (Figure 1E). Consistent with their increased expression of jasmonate biosynthetic genes, stage 12–13 myb21-5 myb24-5 flowers had about 12-fold higher level of cis-JA than did wild-type flowers (Figure 1F). The AOS, LOX2, JAZ5, and JAZ7 genes were also overexpressed in myb21-4 single mutant flowers, to the same degree as in myb21-4 myb24-5 double mutant flowers (Figure 4). These data indicate that MYB21 acts within a negative feedback loop that regulates expression of multiple JA biosynthetic genes.
As mentioned above, the nectary-expressed JMT (At1g19640) gene whose product makes methyl jasmonate was underexpressed in both arf6-2 arf8-3 and myb21-5 myb24-5 flowers. However, the At3g11480 gene encoding a JMT-related protein was overexpressed in myb21-5 myb24-5 flowers, suggesting that At3g11480 rather than JMT/At1g19640 might produce MeJA as part of the MYB-regulated negative feedback loop. JAR1 (At2g46370), encoding an enzyme that synthesizes the active JA-Ile, did not show statistically different expression between wild-type and mutant flowers. jar1 plants are male-fertile, suggesting that another enzyme produces JA-Ile in flowers [50]. The most closely related Arabidopsis gene to JAR1 is GH3-10/DFL2 (At4g03400), which had normal expression in both mutants at stage 12, but was underexpressed in both mutants at stage 13 (data not shown).
In leaves, jasmonate induces genes encoding enzymes in the jasmonate biosynthesis pathway, indicating that a positive feedback loop amplifies jasmonate synthesis [7], [8], [47], [51]–[54]. In qRT-PCR assays, stage 12, 13, and 14 aos-2 and coi1-1 flowers had lower levels of AOS and LOX2 than did wild-type flowers (Figure 4), confirming that such a positive feedback loop operates in flowers. To explore how the MYB21-mediated negative feedback and the COI1-mediated positive feedback interact, we assessed expression of these genes in aos-2 myb21-4 and coi1-1 myb21-4 double mutant flowers. In flowers of both double mutants, AOS and LOX2 levels were as low as in aos-2 or coi1-1 mutant flowers. These results indicate that COI1 is required to activate jasmonate biosynthesis in myb21-4 flowers, and suggest that MYB21 acts by inhibiting the COI1-mediated positive feedback loop in jasmonate biosynthesis. AOS, LOX2, JAZ5, and JAZ7 were also underexpressed in arf6-2 arf8-3 myb21-4 triple mutant flowers (Figure S1C), indicating that jasmonate overproduction in myb21 mutant flowers also depends on ARF6 and ARF8.
The phenotypic and gene expression analyses presented here show that, in addition to previously described petal, stamen, and gynoecium growth and maturation [2], [3], the ARF6 and ARF8 regulatory network promotes nectary development and floral scent production. This regulatory network should promote reproduction by both self-pollination and outcrossing. Thus, coordination of timing of stamen filament elongation, pollen release, stigma growth, and style and transmitting tract support of pollen tube growth ensures efficient self-fertilization; whereas coordination of petal growth, nectary development, and sesquiterpene production with stamen and gynoecium development would attract pollinators to flowers when they are reproductively competent. Although Arabidopsis self-pollinates efficiently, outcrossing by insect pollination has been observed in field populations [55], [56]. Terpene formation coordinated with gynoecium development also helps to protect reproductive organs against invasion by microbial pathogens (M. Huang, A. M. Sanchez-Moreiras, C. Abel, J. Gershenzon, and D. Tholl, unpublished results).
ARF6 and ARF8 activate jasmonate biosynthesis, which in turn activates MYB21 and MYB24. Genes underexpressed in arf6-2 arf8-3 and myb21-5 myb24-5 flowers may promote aspects of flower maturation deficient in both mutants. Such genes include MYB108, which promotes anther dehiscence; several SAUR genes that promote organ elongation (K. Chae, C. G. Isaacs, P. H. Reeves, G. S. Maloney, G. K. Muday, and J. W. Reed, unpublished results); and the TPS11 and TPS21 genes required for sesquiterpene production. Genes affected in arf6-2 arf8-3 but not myb21-5 myb24-5 flowers must act upstream of MYB21 or mediate MYB21-independent functions. These include ARF16, which contributes to petal and stamen elongation; several genes involved in nectary formation or function; and three closely related bHLH transcription factors, HALF-FILLED(HAF)/bHLH075, BRASSINOSTEROID ENHANCED EXPRESSION1 (BEE1)/bHLH044 and BEE3/bHLH050, which act redundantly to promote transmitting tract differentiation (Table S3) [57], [58]. Other genes identified in this dataset may allow further dissection of general and organ-specific aspects of flower maturation, such as stylar factors that promote stigma growth non-cell-autonomously and/or potentiate pollen tube growth [22].
Sepal growth normally ceases at stage 12 when petal and stamen filament growth accelerates, and sepals of mutant flowers appeared outwardly normal. Nevertheless, 240 genes having preferential expression in wild-type sepals had altered expression in arf6-2 arf8-3 flowers, and about two thirds (161/240) of these were overexpressed. In contrast, most affected genes with preferential expression in wild-type petals, stamens, or gynoecia were underexpressed in arf6-2 arf8-3 flowers (91%, 78%, and 64%, respectively). Internal organs in the mutant flowers might have decreased sink strength, which might induce gene expression changes in sepals indirectly, or might cause internal organs to resemble sepals physiologically and express higher levels of “sepal” genes.
Three mobile hormone signals - auxin, gibberellin, and jasmonate - regulate flower maturation, and this network incorporates both signal amplification and feedback mechanisms (Figure 7). Auxin can activate ARF6 and ARF8 activity by destabilizing Aux/IAA transcriptional repressor proteins, and both msg2/iaa19 gain-of-function mutants and yucca2 yucca6 mutants deficient in auxin biosynthesis have delayed stamen development [23], [59]–[64]. These results indicate that auxin indeed promotes wild-type flower maturation. Temperature stress, shade light and the circadian rhythm can each regulate auxin levels and/or response [65]–[71], and these environmental factors might thereby regulate flower growth according to light or temperature conditions, or ensure appropriate diurnal timing of flower opening and pollination.
Similarly to arf6 arf8 mutants, gibberellin-deficient mutants have arrested petal, stamen, and gynoecium development, are deficient in jasmonate production, and are both male- and female-sterile [14], [26], [28]. Although the two pathways had overlapping effects on gene expression, based on our gene chip expression data, arf6-2 arf8-3 flowers had normal gibberellin biosynthetic gene expression levels, and known auxin biosynthetic genes did not appear in the gibberellin-responsive gene lists. Thus, the two pathways may be integrated through shared downstream targets rather than acting hierarchically. Auxins and gibberellins also each regulate hypocotyl elongation and fruit growth, by both hierarchical and parallel mechanisms [72]–[75].
ARF6 and ARF8 and gibberellins each activate jasmonate biosynthesis. ARF6 and ARF8 may do this in part through TCP4 (At3g15030), which was underexpressed in arf6-2 arf8-3 flowers and activates developmental expression of LOX2 [76]. JA-Ile in turn activates a positive feedback loop of jasmonate synthesis by causing COI1-dependent turnover of JAZ transcriptional repressor proteins, which then (at least in leaves) releases the bHLH proteins MYC2, MYC3, and MYC4 to activate transcription of jasmonate biosynthesis genes as well as MYC2 itself [77]–[79]. Jasmonate synthesis has been postulated to occur in stamen filaments, based on the expression pattern of DAD1 [6], [80]. However, other genes can act redundantly with DAD1 during wound-induced jasmonate production [81], and other jasmonate biosynthetic genes were expressed in multiple flower organs (Table S3) [21], [47], suggesting that jasmonates are synthesized broadly throughout the flower. If synthesis were first triggered in stamen filaments, the positive feedback of jasmonate synthesis and movement of MeJA or another jasmonate pathway compound might amplify jasmonate production throughout the flower, thereby causing a coordinated burst of stamen and petal growth and emission of floral scents.
Jasmonates induce MYB21 and MYB24, and MYB21 and MYB24 then activate secondary gene expression responses to jasmonate leading to petal and stamen filament elongation and anther dehiscence. MYB21 and MYB24 are also required for expression of several known primary auxin-responsive genes. This finding suggests that MYB21 and MYB24 also affect ARF6 and ARF8 activity, and that a portion of the myb21 myb24 flower phenotypes may be caused by decreased ARF activity.
MYB21 also induces a negative feedback on jasmonate biosynthesis. Jasmonate overproduction in myb21 flowers requires the COI1-dependent positive feedback pathway that activates jasmonate biosynthesis genes, suggesting that MYB21 acts on a component of this pathway. JAZ genes encoding repressors of jasmonate response are themselves jasmonate-inducible, and the MYB proteins might amplify this negative feedback loop if they activate JAZ gene expression. However, the increased rather than decreased expression of JAZ and other primary jasmonate responsive genes in myb21-5 myb24-5 flowers suggests that other proteins such as MYC2 are sufficient to activate primary jasmonate response. Alternatively, as suggested by the recent discovery that MYB21 and MYB24 proteins can interact with JAZ1, JAZ8, JAZ10, and JAZ11 proteins [20], MYB21 might stabilize JAZ proteins by interfering with their COI1-mediated turnover. This negative feedback pathway may also act in flowers of the jar1-1 mutant deficient in the enzyme that makes active JA-Ile, which similarly had elevated jasmonic acid levels [82].
In wild-type flowers, jasmonic acid levels increase at stages 11–12 just before flowers open, and then decrease at stages 13–14, when flower organs stop growing [2]. Mathematical modeling suggests that after wounding of leaves, positive feedback increases the amplitude of jasmonate synthesis, whereas negative feedback mediated by the JAZ proteins determines the duration of the jasmonate pulse [83]. In flowers, the linked positive and negative feedback loops regulating jasmonate production and auxin response provide a plausible mechanism for inducing coordinated rapid increase in petal and stamen growth at stage 12, followed by a quick cessation of growth after stage 13 once the flower has opened and pollen has been released. MYB21 and MYB24 are not expressed in wounded leaves, and recruitment of the MYB factors into the feedback mechanisms may be an evolutionary innovation that has contributed to the adaptation of this network to regulate flower opening.
The prolonged growth seen in petals of stage 13–14 jasmonate pathway mutant flowers arises from jasmonate-independent MYB21 expression. As arf6 arf8 flowers do not express MYB21, an ARF-dependent but jasmonate-independent mechanism can apparently activate MYB21. This or a similar pathway apparently also acts in stage 14 arf8 and myb108 mutant flowers [11]. BIGPETAL (BPE)/bHLH31 (At1g59640) is activated in petals by jasmonate-induced alternative splicing and represses petal growth [9], [10], [12], [58], and it will be interesting to test whether it represses MYB21.
In tobacco and petunia, putative orthologs of MYB21 and MYB24 regulate both floral scent production and flower opening [84]–[86]. The network described here may thus provide a useful context to understand flower maturation in other angiosperms, and the roles of genes responsible for natural variation in flower morphology [87]. For example, variation in the expression level of the tomato Style2.1 gene determines the extent of style growth, which in turn affects whether the plant self-pollinates or outcrosses [88]. An Arabidopsis homolog of Style2.1, PRE1/bHLH136/BNQ1 (At5g39860), is underexpressed in arf6-2 arf8-3 flowers and may contribute to Arabidopsis flower organ elongation [89], [90].
All genotypes were in the Columbia ecotype of Arabidopsis thaliana. arf6-2 and arf8-3 mutants were previously described [2]. The myb21-4 and myb21-5 mutants were isolated from an EMS mutagenesis screen for enhancers of the arf6-2 mutant. arf6-2 seeds were treated with 0.2% EMS for 16 hours, and 10,000 M2 plants derived from approximately 5000 M1 parents were screened for reduced fecundity. In addition to the myb21 mutations described here, we isolated three new arf8 alleles in this screen (Table S2). arf6-2 myb21 plants from the screen were back-crossed once to arf6-2, and then crossed twice to wild type prior to further analysis. Backcrosses indicated that the myb21 phenotype was caused by a recessive mutation at a single genetic locus. To map the mutations, arf6-2 enhancer mutants were crossed to an arf6-2 line that had been introgressed into the Landsberg erecta ecotype. A bulked-segregant analysis approach using 29 markers evenly distributed over the genome was used to establish a preliminary map position [91], and the map position was then refined using closely linked SSLP, CAPS and dCAPS markers (Figure S3).
T-DNA insertion mutations in MYB21, MYB24, MYB57, MYB108, and AOS from the SALK Genomic Analysis Laboratory were provided by the Arabidopsis Biological Resource Centre [92]. Homozygous mutants were identified within segregating T3 and T4 populations. Details on these mutants, and PCR primers used to identify mutant alleles, are provided in Figure S4, Table S2, and Table S5. Double and triple mutants were identified in the F2 progeny of crosses from the respective single or double mutant parents. Most genotypes were fertile when manually self-pollinated. However, myb21 myb24, myb21 aos-2 and myb21-5 arf6-2 arf8-3 plants were maintained as myb21/+ myb24, myb21 aos-2/+ and myb21 arf8-3 arf6-2/+ stocks. coi1-1 seeds were provided by John Turner (University of East Anglia, Norwich, UK), and PLAT52:GUS seed were provided by Mark Johnson (Brown University, Providence, RI). ams seeds (SALK_152147) [93] were provided by Hong Ma (Pennsylvania State University, College Station, PA).
To measure flower organs across a developmental series, flower buds were dissected, and flower organs were placed on an agar plate and measured using a camera lucida attachment on a dissecting microscope. For measurements of floral organ lengths and timing of anther dehiscence in Table S1, the first open flower of wild-type plants was designated as flower 1 (stage 13) [1]. For genotypes in which flower opening was impaired, equivalent stage flowers were identified based upon bud size and position on the inflorescence stem. Scanning electron microscopy was carried out as previously described [2]. Fertilization frequencies were assessed by X-Gluc staining 24 hours after pollination with the pollen-specific reporter line PLAT52:GUS [94]. In these assays, 87% or more of wild-type, ams male-sterile, myb21-5, myb21 myb24, and myb21 myb24 myb108 ovules were fertilized, as judged by strong X-Gluc staining in ovules in which a pollen tube had ruptured.
To make GUS reporter constructs, promoter and genomic sequences lacking the endogenous stop codon were cloned into the Gateway pENTR/D-TOPO vector (Invitrogen Life Technologies, Carlsbad, CA). The upstream region (2266 bp) and first exon of MYB21 were amplified by PCR using the primers MYB21 PF and MYB21 R4 (Table S5). The introns and second and third exons were amplified using the primers MYB21 F4 and MYB21 R7. These two PCR products were cloned separately into pENTR/D-TOPO, and then the promoter and first exon of MYB21 were excised and ligated into the construct containing the 3′end of the MYB21 gene using NotI and PstI. The upstream region (2207 bp) and first exon of MYB24 was amplified by PCR using the primers MYB24 PF and MYB24 R3. The entire predicted coding region of MYB24 was amplified using the primers MYB24 F2 and MYB24 R2. These two PCR products were cloned separately into pENTR/D-TOPO, and then the promoter and first exon were excised and ligated into the construct containing the 3′end of MYB24 gene using NotI and PstI. The MYB57 upstream region (2414 bp) and predicted coding region were amplified using the primers MYB57 PF and MYB57 R2 and cloned into pENTR/D-TOPO. For MYB108, only the promoter was used in the GUS reporter construct. The upstream region of MYB108 (2090 bp) was amplified using the primers MYB108 PF and MYB108 R2 and cloned into pENTR/D-TOPO. The promoter and genomic sequences were fused to the GUS reporter by recombining entry clones into the destination vector pGWB3 [95] using LR clonase (Invitrogen). Transformation of Arabidopsis plants and histochemical staining were performed as described previously [96], [97]. The PMYB21:MYB21:GUS and PMYB24:MYB24:GUS constructs partially rescued the phenotype of myb21-2 myb24-2 flowers, indicating that they retained some MYB21 and MYB24 activity.
For P35S:MYB21 and P35S:Green fluorescent protein(GFP):MYB21 constructs, a genomic MYB21 fragment was amplified using the primers MYB21 GWF and MYB21 GWR* (Table S5), and cloned into the Gateway pENTR/D-TOPO vector. For P35S:MYB21, the entry clone was recombined into pB2GW7 [98]. For P35S:GFP:MYB21 the entry clone was recombined into pGWB6 [95]. Transgenic T1 P35S:MYB21 and P35S:GFP:MYB21 plants showed a range of phenotypes, including narrow leaves, dwarfism, and floral defects, similar to those previously described [19], [20]. For our analyses, we used weaker lines that had less severe phenotypes.
For gene expression analyses, plants were sprayed with 1 mM MeJA (Bedoukian Research, Inc, Danbury, CT) or 10 µM IAA (Sigma) in 1% methanol 0.05% Tween-20, or with solvent alone, and were harvested after two hours (MeJA) or the specified time periods (IAA). To restore fertility to JA-deficient plants and to assess the effect of jasmonate on aos-2 PMYB21:MYB21:GUS plants, flowers were sprayed with 1 mM MeJA daily for 4 days.
Flowers or whole inflorescences were frozen in liquid N2 and total RNA was isolated using Trizol reagent (Invitrogen Life Technologies, Carlsbad, CA) or with RNeasy plant mini kits (Qiagen). Poly (A+) RNA was extracted from 50 µg of total RNA using oligo(dT)25 Dynabeads according to manufacturers' instructions (Dynal A.S., Oslo, Norway). RNA gel blot hybridizations were performed as described [99]. Probes were created by PCR using genomic DNA or Peking-Yale cDNA clones (Arabidopsis Biological Resource Center) as template using primer pairs listed in Table S5.
For real-time quantitative RT-PCR analyses, total RNA was extracted from stage 12, 13 and 14 flowers in the morning between 2 and 4 hours after subjective dawn. cDNA was synthesized using the Reverse Transcription System (Promega A3500) with random primers according to the manufacturer's instructions. 0.1 µg of total RNA was used for the 20 µl volume reaction and incubated for 1 hr at 42°C. The RT reaction mixture was diluted 10-fold and 1 µl was used as a template in 10 µl PCR reactions using the Applied Biosystems real-time PCR systems in standard mode with SYBR Green Master Mix (Applied Biosystems) following the manufacturer's protocol. The primers used for qRT-PCR analysis are listed in Table S5. Reactions were performed in triplicate and the products were checked by melting curve analysis. Transcript levels were normalized to the level of reference transcript UBQ10.
For Affymetrix gene chip gene expression analyses, RNA was isolated from stage 12 (largest closed buds) and stage 13 (first open flowers) harvested in the morning between 2 and 4 hours after subjective dawn. Three biological replicates were performed. Probe synthesis and gene chip hybridizations were performed by the UNC-CH Functional Genomics Core Facility. Total RNA (1000 ng) was used to synthesize cDNA followed by aRNA. The MessageAmp II-Biotin Enhanced Kit (Ambion) was used to generate biotinylated aRNA from the cDNA reaction. The aRNA was then fragmented in fragmentation buffer from the Ambion kit at 94°C for 35 minutes before the chip hybridization. Fragmented aRNA (15 µg) was then added to a hybridization cocktail (0.05 µg/µl fragmented aRNA, 50 pM control oligonucleotide B2, BioB, BioC, BioD and cre hybridization controls, 0.1 mg/ml herring sperm DNA, 0.5 mg/ml acetylated BSA, 100 mM MES, 1 M [Na+], 20 mM EDTA, 0.01% Tween 20). aRNA (10 µg) was used for hybridization in a volume of 200 µl per slide. ATH1 arrays [100] (Affymetrix, Santa Clara, CA) were hybridized for 16 hours at 45°C in the GeneChip Hybridization Oven 640 (Affymetrix). The arrays were washed and stained with R-phycoerythrin streptavidin in the GeneChip Fluidics Station 450 (Affymetrix) using wash protocol EukGE-WS2v4, and arrays were scanned with the GeneChip Scanner 3000 7G Plus with autoloader. Affymetrix MAS 5.0 GeneChip Operating Software was used for washing, scanning and basic analysis. Sample quality was assessed by examination of 3′ to 5′ intensity ratios of selected genes. Data were analyzed using Genespring GX 10.0.1 software. Raw data were background corrected and normalized using the RMA algorithm with no baseline correction. Means for each gene over the three biological replicates were calculated, and statistical differences between wild-type and mutant expression levels assessed by t-test without multiple testing correction. Genes reported in Table S3 are those with P<0.05 and having 2-fold or greater expression level difference from the corresponding wild-type sample. Gene chip hybridization data have been deposited in the NCBI GEO database (http://www.ncbi.nlm.nih.gov/geo/) with accession number GSE32193.
In situ hybridizations were carried out as previously described [4]. MYB21 and MYB24 probes were PCR amplified from genomic DNA using primers that spanned the last exon (MYB21 ins-HindIII F, MYB21 ins-BamHI R; MYB24 ins-HindIII F, MYB24 ins-BamHI R) (Table S5). PCR products were then cloned into the pGEM-T vector (Promega). MYB21 and MYB24 sense probes produced no signal in wild-type flowers.
Jasmonic acid was measured as described [101] from stage 12–13 flowers collected in the morning and frozen in liquid nitrogen. To measure sesquiterpenes, volatile compounds were collected in 1 L bell jars with 40 detached inflorescences placed in a small glass beaker filled with tap water, under controlled temperature and light conditions (22°C, 150 µmol m−2 s−1 PAR). Emitted volatile compounds were collected for 7 h on 5 mg Charcoal filter traps (Gränicher and Quartero, Daumazan, France) in a closed-loop stripping procedure [102] and then eluted from the traps with 40 µl CH2Cl2 containing 1-bromodecane (20 ng/µl) as a standard. Sample analysis and quantification of terpenes was performed by gas chromatography–mass spectrometry (GC-MS) on a Shimadzu QP 2010s GC-MS instrument as described previously [103]. Separation was performed on a (5%-phenyl)-methylpolysiloxane (DB5) column (Restek, 30 m×0.25 mm i.d.×0.25 m –thickness). Helium was the carrier gas (flow rate 1.4 ml min−1), a splitless injection (injection volume 1 µl) was used, and a temperature gradient of 5°C/min from 40°C (2 min hold) to 220°C was applied. Compounds were identified by comparison of retention times and mass spectra with those of authentic standards. Trapping and GC-MS analysis of volatiles from flowers of opr3 and Wassilewskija wild type were performed as described in [43]. Statistical significance of differences in volatile emission was determined with SAS9.1 (SAS Institute Inc., Cary, NC, USA) using student's t-test or ANOVA with Tukey post-hoc test.
For an alternative fast sampling and analysis of volatile compounds, 20 inflorescences were placed in a 20 ml screw cap glass vial containing 4 ml of water. Inflorescences were incubated in the sealed vial for 2 h under the conditions described above. Volatile compounds were then trapped by solid phase microextraction (SPME) for 30 min at 40°C and injected into the GC by thermal desorption using an automated SPME sampling device (Combi-PAL, CTC Analytics, Zwingen, Switzerland).
Arabidopsis Genome Initiative locus identifiers for the genes studied in this article are as follows: AOS (At5g42650); ARF6 (At1g30330); ARF8 (At5g37020); COI1 (At2g39440); IAA2 (At 3g23030); IAA3 (At1g04240); IAA4 (At5g43700); IAA7 (At3g23050); IAA13 (At2g33310); IAA16 (At3g04730); IAA19 (At3g15540); MYB21 (At3g27810); MYB24 (At5g40350); MYB57 (At3g01530); MYB108 (At3g06490); OPR3 (At2g06050); SAUR63 (At1g29440); AtTPS11 (At5g44630); AtTPS21 (At5g23960).
|
10.1371/journal.pbio.2005750 | A systems genetics resource and analysis of sleep regulation in the mouse | Sleep is essential for optimal brain functioning and health, but the biological substrates through which sleep delivers these beneficial effects remain largely unknown. We used a systems genetics approach in the BXD genetic reference population (GRP) of mice and assembled a comprehensive experimental knowledge base comprising a deep “sleep-wake” phenome, central and peripheral transcriptomes, and plasma metabolome data, collected under undisturbed baseline conditions and after sleep deprivation (SD). We present analytical tools to interactively interrogate the database, visualize the molecular networks altered by sleep loss, and prioritize candidate genes. We found that a one-time, short disruption of sleep already extensively reshaped the systems genetics landscape by altering 60%–78% of the transcriptomes and the metabolome, with numerous genetic loci affecting the magnitude and direction of change. Systems genetics integrative analyses drawing on all levels of organization imply α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) receptor trafficking and fatty acid turnover as substrates of the negative effects of insufficient sleep. Our analyses demonstrate that genetic heterogeneity and the effects of insufficient sleep itself on the transcriptome and metabolome are far more widespread than previously reported.
| Sleep is essential for optimal brain functioning and health, but the biological substrates through which sleep delivers these beneficial effects remain largely unknown. We used a systems genetics approach in a large, diverse reference population of mice and assembled a comprehensive experimental knowledge base comprising “sleep-wake” data, central and peripheral gene expression, and plasma metabolic indicators, collected under undisturbed baseline conditions and after sleep deprivation (SD). We present analytical tools to interactively interrogate the database, visualize the molecular networks altered by sleep loss, and prioritize candidate genes. We found that a brief, one-time disruption of sleep extensively reshaped the transcriptome in cerebral cortex and liver, and the plasma metabolome, with numerous genetic loci affecting the magnitude and direction of change. Integrative analyses drawing on multiple sources of data imply α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) receptor trafficking and fatty acid turnover as substrates of the negative effects of insufficient sleep. Our analyses demonstrate that genetic heterogeneity and the effects of insufficient sleep on gene expression and metabolism are far more widespread than previously reported.
| Insufficient or disrupted sleep characterizes the 24 h lifestyle of modern society and represents a serious public health concern, as it is associated with increased risk for, e.g., obesity, diabetes, and high blood pressure, and impairs cognitive performance, which in turn increases the likelihood of accidents, medical errors, and loss of productivity [1,2]. Several hypotheses concerning sleep’s still elusive function converge on the notion that staying awake imposes a burden that can only be efficiently alleviated during sleep [3–7]. This concept of a need for sleep accumulating during wakefulness and recovering while asleep is central in sleep research and is referred to as sleep homeostasis. Insight into the molecular substrates of the sleep homeostatic process is instrumental in advancing our basic understanding of sleep need under both physiological and pathological conditions.
The impact of acute sleep deprivation (SD) on recovery sleep and cognitive performance is under strong genetic control [8–13], and genetic approaches therefore seem promising in uncovering the molecular pathways important in sleep homeostasis. Reductionist studies in mice and flies deleting genes through gene targeting (for review, see [8]) or in mutagenesis screens [14–16] have demonstrated that single genes can have large effects on various aspects of sleep, including its homeostatic regulation. Such large single-gene (mendelian) effects—often assessed on 1 genetic background only—are, however, likely to be the exception. Indeed, susceptibility to sleep loss in the general population is assumed to be determined by the interactions of many genes, their natural allelic variants, and their interaction with the environment (lifestyle), a complexity that only recently has begun to be appreciated. Such complexity can best be assessed in so-called genetic reference populations (GRPs), which are designed for the study of complex traits inherited in a nonmendelian fashion. The BXD panel of advanced recombinant inbred lines (ARILs) is the largest and best-characterized GRP to date, consisting of well over 150 lines in which 2 parental (C57BL/6J [B6] and DBA/2J [D2]), now fully sequenced genomes are segregating (www.genenetwork.org; [17]). As each line represents a reproducible clone of animals, many mutually reinforcing datasets can be collected and compared at multiple levels across many biological systems. This approach has been termed “systems genetics,” which in essence allows for making inferences about biological phenomena by assessing the flow of information from DNA to phenotype at the level of a population and how this flow is perturbed by environmental challenges. Because systems genetics generalizes results to a population level, it is considered critical for predicting disease susceptibility [18]. Systems genetics has been applied with great success in the BXD set for, e.g., mitochondrial function and metabolic- and aging-related phenotypes [19–21].
Systems genetics approaches for sleep have been pioneered in the fly and mouse [22,23], but neither study reported on the effects of sleep loss on intermediate phenotypes, such as the metabolome and transcriptome. Here, we present an extensive and comprehensive dataset interrogating the BXD set at the levels of the genome, the brain and liver transcriptomes, the plasma metabolome, and finally, the phenome including sleep-wake state, electroencephalography (EEG)-, and locomotor activity (LMA)-related phenotypes, both under undisturbed baseline conditions and after an acute SD challenging the sleep homeostatic process. We observed that SD profoundly impacted all 3 phenotypic levels and that genetic background not only determined the magnitude but also the direction of the SD-evoked changes. The molecular pathways associated with these effects will be illustrated here to introduce our integrated data resource. The molecular signaling circuitry underlying the equally profound phenotypic differences observed under baseline conditions will be reported in subsequent molecular-driven validations.
Systems genetics is an emerging field, and innovative ways to improve data access, portability, and reproducibility; tools to display and mine these data; and statistical models to extract the multidimensional relationships across datasets are areas of intense research [24]. The size and complexity of our current dataset necessitated the development of new analytical tools and data sharing strategies such as (i) a supervised machine learning–based algorithm to annotate sleep-wake states on EEG/electromyogram (EMG) tracks, (ii) a gene-prioritization strategy that draws on all levels of the experimental dataset to assist the search for candidate genes within quantitative trait locus (QTL) intervals, and (iii) the implementation and integration of a recently developed systems genetics visualization tool [25] in a dynamic web-based interface that, in addition, provides access to the data presented and enables interactive data mining (https://bxd.vital-it.ch).
This section is organized as follows: study design and the types of data contributing to our resource are shortly described first. We then ascertain the contribution of genetic factors to all the intermediate and end phenotypes we quantified. Next, the tools to interactively visualize the systems genetics relationships and to prioritize candidate genes will be described in detail. Because our current focus is on the effects of enforced wakefulness, we describe the SD-evoked changes at the level of the GRP, as well as the genetic effects thereon, before closing with 4 examples that, aided with the prioritization tool, point to novel molecular pathways shaping the marked genetic variability in the response to sleep loss at all levels of organization.
We subjected mice from 33 BXD/RwwJ lines (see https://bxd.vital-it.ch; Downloads, General_Information.xlsx for a listing), the 2 parental strains (B6 and D2), and F1 individuals from reciprocal crosses between the parental lines to a deep behavioral and molecular phenotyping across 4 levels of organization. In 1 set of mice, we recorded sleep-wake behavior, brain activity (by EEG), and LMA for 4 d (Fig 1, Experiment 1). On day 3, mice underwent an SD challenge during the first 6 h of the light period, when mice normally sleep most of the time. During SD, an average of 8.6 ± 0.7 successful attempts at sleep were observed lasting 14.2 ± 0.6 s on average, resulting in a total of 1.8 ± 0.1 min (range: 0.0–9.8 min, n = 198 over the 33 BXD lines) of sleep or 0.5% of the 6 h intervention. Both the number of sleep episodes and total time spent asleep varied according to BXD line (1-way ANOVA, p < 0.0001 for both variables), while response time of the experimenter (i.e., episode duration) did not (p = 0.66). Aided by a specifically developed, supervised machine learning–based algorithm (see Materials and methods and S1 Fig), we could extract a comprehensive set of EEG/behavioral phenotypes (see https://bxd.vital-it.ch; Downloads, General_Information.xlsx), which were separated into 3 main biological categories related to (i) LMA, (ii) EEG signal features, and (iii) the prevalence and time structure of sleep-wake state, collectively referred to as “LMA,” “EEG,” and “State,” respectively. The 3 phenotypic categories were divided further into subcategories (see Materials and methods) and by experimental condition (baseline, SD, and recovery). Because some of the 341 phenotypes we quantified were tightly linked (e.g., the time spent in non-REM [NREM] sleep and wakefulness), we estimated the total number of distinct phenotypic clusters or modules to be 120 or 148 when considering phenotypes of different subclasses (e.g., “EEG,” “State,” or “LMA”) within a given module as separate (S2 Fig, Materials and methods, and https://bxd.vital-it.ch; Downloads, General_Information.xlsx). Most phenotypes were unique or were grouped in modules of 2 phenotypes only (67%; median: 2 phenotypes/module, range: 1–13). Several of these modules (49/120) were associated into 3 larger “superclusters” (Supercluster I–III; S2 Fig), containing 18, 20, and 11 modules, respectively. Supercluster I grouped almost exclusively “State”-related phenotypes (80/83), while Supercluster II was composed mostly of “EEG”-related phenotypes (65/73). Supercluster III was composed of 10 “LMA”-related and 30 “State”-related phenotypes. However, in our analyses, we still used all available phenotypes to detect potential regulatory differences among even closely related phenotypes and to avoid analysis bias arising from selecting a “representative” phenotype.
A second set of mice, representing the same lines, was processed in parallel for collection of brain, liver, and plasma (Fig 1, Experiment 2) to measure gene expression in cortex and liver and metabolites in plasma. These transcriptomic and metabolomic data are collectively referred to as (intermediate) molecular phenotypes. We quantified 124 metabolites (see https://bxd.vital-it.ch; Downloads, General_Information.xlsx) using targeted metabolomics covering 5 important metabolite classes (i.e., amino acids, biogenic amines, acylcarnitines, sphingolipids, and glycerophospholipids). Cortex and liver transcript levels were measured using RNA sequencing (RNA-seq), and we detected about 14,900 expressed genes in the cortex and about 14,100 genes in the liver after filtering and normalization.
We used the RNA-seq alignments also to genotype the lines to verify that no mix-up occurred during the breeding and data collection phase, and to increase mapping resolution. We compared the around 500,000 detected genotypes with the publicly available 3,500-genotype set for the same BXD lines from GeneNetwork (2005 release; see Materials and methods). We observed only an approximately 1% discrepancy and merged both genotype sets, resulting in a set of about 11,000 tag variations, which increased the number of haplotype blocks from 551 (GeneNetwork) to 1,071 (RNA-seq + GeneNetwork). All analyses we report here were based on our merged map (see https://bxd.vital-it.ch; Downloads, Genotypes.GeneNetwork2005AndRNAseq.geno). Of note, by the completion of this publication, an updated set of BXD genotypes was released with an estimated haplotype block number of 816 for the specific lines we used (GeneNetwork, 2017 release http://genenetwork.org). Of the 61 significant phenotypic quantitative trait loci (phQTLs) we detected (see below), 54 were also detected using either GeneNetwork genotypes (the 2005 or 2017 release), while the remaining 7 significant phQTLs were unique to our merged genotype map.
To obtain a first sense of the contribution of genetic factors to the phenotypic variability contained within our BXD set, we examined the heritability of the EEG/behavioral and metabolic phenotypes. The estimated narrow sense heritability [26] among the EEG/behavioral phenotypes was high overall (median h2 = 0.68, Fig 2A), consistent with what has been reported in previous human and mouse studies [27]. We also confirm that various aspects of the EEG signal are among the most heritable traits with, in our dataset, theta-peak frequency (TPF) in REM sleep ranking highest (h2 = 0.89). The heritability for differential EEG/behavioral phenotypes (i.e., recovery versus baseline; green symbols in Fig 2A) were consistently lower by around 0.2 points compared with the heritabilities obtained for recovery or the baseline values per se. By contrasting individual recovery values to the baseline strain averages, instead of to each animal’s individual baseline value (thereby keeping within strain variance similar to that of the absolute recovery values), we found that this effect did not simply reflect increased variability due to combining recovery and baseline values and thus suggests a smaller genetic contribution to the response to sleep loss. The overall heritability of plasma metabolite levels was somewhat lower than for EEG/behavioral phenotypes (median h2 = 0.50), with alpha-aminoadipic acid (α-AAA) displaying the highest heritability (h2 = 0.88; Fig 2A).
Average-to-high heritabilities are a requirement to attribute phenotypic variation to gene loci, but even then, there is no guarantee to find genome-wide significant QTL(s); e.g., for the TPF in REM sleep phenotype mentioned above, only 4 suggestive phQTLs of small effect size were identified (see https://bxd.vital-it.ch; Downloads, QTL_Mapping.xlsx) that together could nevertheless account for 58% of the variance (estimated using an additive model, see Materials and methods), suggesting that perhaps higher-order loci interactions (e.g., epistasis), which cannot be captured using the single-marker linkage analysis we used here, underlie differences in this EEG trait. Genome scans revealed a total of 61 “significant” (false discovery rate [FDR] ≤ 0.05), 65 “highly suggestive” (0.05 < FDR ≤ 0.10), and 923 “suggestive” (0.10 < FDR ≤ 0.63) [28,29] phQTLs and 21 significant, 40 highly suggestive, and 528 suggestive metabolic quantitative trait loci (mQTLs; Fig 2B).
Several phenotypes from distinct phenotypic categories were associated with overlapping genomic regions. For example, differences in baseline wake consolidation, gain in REM sleep time after SD, EEG delta power (1.0–4.0 Hz) in REM sleep, baseline levels of serotonin and phosphatidylcholine acyl-alkyl (PC-aa)-C34:4, and levels of PC-aa-C34:4 and PC-aa-C36:6 after SD all mapped to one 30 Mb region on chromosome 10 (50–80 Mb), each with a significant or highly suggestive QTL (Fig 2B). These overlapping QTLs may point to pleiotropic effects of 1 underlying gene or close but distinct underlying QTLs.
We also performed QTL analysis for gene expression, but because many more linkage tests were required for transcriptome mapping, we used a more suitable method than for ph- and mQTL mapping. The format for reporting expression quantitative trait loci (eQTLs) will therefore differ from that used for ph- and mQTLs (see Materials and methods). The expression of individual genes was mapped separately for cis-eQTLs with genetic markers within a 2 Mb window and trans-eQTLs with markers positioned throughout the genome (see Materials and methods). The transcriptome of BXD mice showed strong linkage with genotypic variation. For example, in the cortex, the expression of 5,704 genes (i.e., 38% of all expressed genes) was significantly driven by a cis-variation (Fig 2C and https://bxd.vital-it.ch; Downloads, cis_eQTL.xlsx). Moreover, 2,465 (34%) of all genes under cis-eQTL effect in both tissues passed the 0.05 FDR cutoff in a single condition and tissue. Factors contributing to this tissue/condition specificity are the absence of gene expression in one of the 2 tissues or a different gene regulatory environment on which SD had pervasive effects (see Pervasive effects of SD at all levels). This important tissue/condition specificity also applied to trans-eQTLs with 5,537 (53% of 10,450) being under trans-eQTL effect only in one specific tissue or condition. Although the observation that a large portion of eQTLs reached significance in 1 tissue and condition only does suggest widespread gene × environment interactions regulating gene expression, reaching the 0.05 FDR threshold or not does not prove this. We therefore compared linkage strength of significant cis-eQTLs that were specific for 1 tissue and condition with that in the 3 other RNA-seq sets. Among the 870 genes with a significant cis-eQTL effect in sleep-deprived cortex only (Fig 2C), 175 (20%) showed a significant difference in linkage signal (FDR < 0.05). This proportion was similar in the control cortex and liver (19% and 21%, respectively) and somewhat higher in sleep-deprived livers (32%).
The complexity of multilevel networks can only be appreciated through visual aids. Because the widely used “hairball” representation, in which biological factors are represented as “nodes” and their interconnections as “edges,” is hardly interpretable due to its nondeterministic structure (Fig 3A), we opted for a structured representation more suitable for the visualization of complex systems, namely, “hiveplots” [25]. The hiveplots were laid out as follows: each plot represents 1 EEG/behavioral phenotype and its associated molecular network—i.e., only the genes and metabolites strongly correlated with a given phenotype are displayed (Fig 3B; see Materials and methods for details). Each hiveplot is composed of 3 radial axes containing the molecular data with nodes assigned to the 2 bottom axes for genes expressed in the cortex (Fig 3B left, in blue) and liver (Fig 3B right, in red), while nodes on the vertical axis (Fig 3B top, in yellow) represent metabolites. On top, we added a separate “genetic” axis (Fig 3B top, white) containing the genotypes. The node position on the 3 (molecular) radial axes was determined by the response to SD—i.e., molecules positioned closer to the center were down-regulated more strongly, while more up-regulated genes/metabolites can be found closer to the axes’ perimeter. Edges connecting nodes represent positive/negative correlations (red/blue, respectively) between measurements of expression/metabolite levels. Genetic markers linked to genes by eQTLs connect the genetic and molecular space.
The hiveplot representation allows investigation of the molecular network associated with an EEG/behavioral phenotype in a structured manner and comparison of phenotypes using all intermediate phenotypic layers available in the dataset. The difference in presence or absence of nodes/edges between 2 phenotypes indicates which association was gained or lost. Furthermore, the importance of the SD effect on these nodes can be visually estimated by their position along the axis (Fig 3C). Although the interphenotype connectivity present in the hairball representation is lost in the printed format of these hiveplots, this aspect can be easily accessed through our web interface (https://bxd.vital-it.ch) by highlighting common edges. The web interface also allows for an in-depth exploration of the data by displaying node details, such as gene and metabolite name, and variation identifiers. It also lets the user modify the parameter settings, such as the correlation strength used to include correlated genes and metabolites, with which the hiveplots are generated (see S3 Fig and the tutorial on https://bxd.vital-it.ch; Help).
We developed an unbiased, data-driven approach to select candidate genes associated with our EEG/behavioral and metabolic phenotypes. We focused on genes located in the associated genomic regions found by QTL analyses (see Fig 2B). To investigate these often quite large regions (mean = 9.8 Mb, range = 0.7–34.7 Mb for significant and highly suggestive phQTLs), we implemented a scoring strategy inspired by the “similarity profiling prioritization strategy” [30], which combines multiple sources to prioritize a gene. For each gene, we computed an integrated score composed of (i) the genomic position of the gene with respect to the ph-/mQTL peak, (ii) a detected cis-eQTL driving the expression of the gene, (iii) a protein-damaging annotation of a variant, (iv) differential expression (DE) after SD, and (v) correlation between expression and phenotype of interest (Fig 3D, S4 Fig, see Materials and methods for details). Our prioritization strategy thus aimed at identifying genes that are sensitive to sleep loss, correlated with the phenotype being evaluated, associated to a cis-eQTL, and/or carrying a protein-damaging variant that could contribute to trait variance. A Henikoff weighting algorithm was applied to correct for intrinsic correlations among the 5 analysis scores. One informative example of such intrinsic correlation is a cis-eQTL located within a phQTL region, in which case the phenotype–gene expression correlation will be influenced by linkage. The algorithm decreases the cis-eQTL score accordingly, and cis-eQTLs therefore usually contributed with a low score to the prioritization (see S1 Table for examples). The integrated score for each gene was computed with the given formula (Fig 3D), and an FDR was computed by performing 10,000 permutations (S4 Fig and Materials and methods). For each QTL, we kept the gene with the highest significant integrated score. This scoring strategy was applied to cortex and liver data separately.
To illustrate our prioritization algorithm, we applied it to the metabolite with the highest heritability, α-AAA (see above), and for which we obtained a highly significant mQTL on chromosome 2 (logarithm of odds ratio [LOD] = 9.25, 1–11 Mb). We readily identified Dhtkd1 as the top-ranked significant candidate gene in liver within the chromosome 2 mQTL (Fig 3E) because of (i) the strong correlation of Dhtkd1 expression with α-AAA levels, (ii) Dhtkd1 is under a cis-eQTL effect (rs222492362, chr2: 5.8 Mb, q = 1.5e−17), (iii) the marker of the cis-eQTL is located within the peak of the mQTL, and (iv) both α-AAA and Dhtkd1 levels are affected similarly by SD. The 5 scores and weights of this example and those obtained in Examples 1–4 (see below) are detailed in S1 Table.
This result can be taken as a first validation of our scoring strategy because Dhtkd1 encodes an enzyme subunit involved in lysine degradation known to control α-AAA levels in BXD lines [31]. Although with this particular example, the prioritization tool did successfully select the causative gene underlying the α-AAA mQTL, it is important to note that, as opposed to other tools that have been developed (e.g., [32,33]), our algorithm cannot infer causality and is designed to help select likely candidate genes within m- and phQTLs.
The EEG/behavioral and molecular phenotypes were assessed both under undisturbed baseline conditions and after 6 h SD. SD profoundly and significantly impacted a majority of measurements at all levels. We observed the well-known increase in EEG delta power (1.0–4.0 Hz) during NREM sleep as well as the increase in the time spent asleep (Fig 4A), both reflective of an accumulated homeostatic sleep pressure during SD. The gain in time spent in NREM sleep was strongest during the initial 12 h following the SD, with an average gain of +23 min (compared with values reached during corresponding baseline hours) during the first 6 h after the SD (zeitgeber time [ZT]6–12) and +32 min during the first 6 h of the following dark period (ZT12–18). The most strongly affected sleep phenotype concerned time spent in REM sleep, which displayed a 3.3-fold gain during the first 6 h of darkness (ZT12–18) after SD (Fig 4A). SD thereby doubled the proportion of REM sleep to NREM sleep in this interval. Locomotor activity and waking phenotypes were generally decreased during the light period immediately following the SD (ZT6–12).
In addition, the plasma metabolome was profoundly altered by SD. Of the 124 measured metabolites, 75 (60%) were significantly up- or down-regulated. The levels of all amino acids were significantly altered after SD, the majority being down-regulated, with the exception of glutamine, glutamate, and tryptophan, which were up-regulated (Fig 4B). A recent publication reported similar effects on amino acid levels in brain dialysates of sleep-deprived rats [35], suggesting that plasma can report on central changes in amino acid levels. By contrast, tryptophan was the only amino acid that was found significantly changed during SD in humans using the same methodology [36]. The 2 acylcarnitines present in our dataset (C18:1 and C18:2) were both strongly up-regulated with a greater than 2-fold change. Similar results were found in humans, with acylcarnitines levels increased in blood and carnitines increased in urine after sleep loss [36,37].
The transcriptome was especially sensitive to SD, with 78% of all expressed genes being differentially expressed in cortex and 60% in liver. In cortex, the most strongly differentially expressed genes were activity-regulated cytoskeletal-associated protein (Arc), early growth response 2 (Egr2), and perilipin 4 (Plin4), with an almost 8-fold increase in expression after SD (see S2 Table). Arc is an immediate early gene crucial for long-term synaptic plasticity and memory formation [38]. Arc is among the most consistently up-regulated transcripts after SD [39] and features in a short list of 78 genes, the expression of which we found reliably and significantly changed by extended wakefulness under a number of experimental conditions [34]. Forty-nine other genes in this short list featured among the top 5% most affected transcripts of the current experiment (S2 Table and blue symbols in Fig 4C left; enrichment p = 5.6e−43, Fisher test). The remaining 29 of this short list were all significantly affected by SD also in the current study, 15 of which were found in the 5%–10% tile, and all ranked in the top 26% of most differentially expressed genes. Similarly, Egr2 is 1 of 3 Egr genes that are rapidly induced by SD in several species [39]. Egr1 and Egr3 appear on our short list of 78, and all 3 Egrs are among the top-100 differentially expressed cortical genes in the current study (S2 Table). The Egr family are immediate early genes encoding transcription factors important in neuronal plasticity [40]. Plin4, which encodes a lipid droplet–associated protein involved in lipid storage [41], has not been reported previously as part of the SD response. Tubulin tyrosine ligase-like family 8 (Ttll8), encoding a ligase that glycylates microtubules [42], and family with sequence similarity 107, A (Fam107a), a stress- and glucocorticoid-regulated gene [43,44], were the top differentially expressed genes in liver (S3 Table). Although the short list of 78 was based on forebrain samples, 17 genes were also present in the top 5% differentially expressed genes in the liver (blue symbols in Fig 4C right). Moreover, 13 genes were common to the top 5% list in cortex, liver, and the 78 genes of the short list (Hspa1a/b, Cirbp, Fos, P4ha1, Chordc1, Dusp1, Slc5a3, Hsph1, Creld2, Tra2a, Zbtb40, and Pfkfb3). These genes might be interesting candidates for tissue-independent biomarkers of sleep pressure.
In the context of our project, a key question is whether genetic background modifies these pervasive effects of SD. We found evidence for this at all 3 levels of organization and detected genomic loci predicting differences not only in the magnitude of the response to SD but also in the direction of the response (illustrated in Fig 4D–4F). In the analyses, we included both the levels reached after the SD and these levels contrasted with their baseline levels. These contrasts will be referred to as “change,” “increase,” “gain,” “decrease,” or “DE”.
For 7 EEG/behavioral “gain” phenotypes we discovered a significant QTL (https://bxd.vital-it.ch; Downloads, QTL_Mapping.xlsx). Illustrated in Fig 4D is the gain in time spent in REM sleep, which mapped significantly to chromosome 18 (LOD = 3.9; 57–62 Mb) with B6-allele carriers gaining more REM sleep than D2-carriers (genotype × SD interaction: p = 2.0e−5). Three more “gain” phenotypes will be discussed in detail below (see Example 1, 3, and 4 in the Systems genetics of the effects of SD section). Also illustrated in Fig 4D is an EEG/behavioral gain phenotype with a pronounced genotype effect on the direction of change. The SD-induced changes in EEG activity in the fast gamma band (55–80 Hz) in NREM sleep mapped suggestively to chromosome 6 (LOD = 2.83; 77–89 Mb), with a majority of B6-allele carriers at the QTL peak position having a significant decrease in fast gamma, while several D2-allele carriers showed a significant increase (genotype × SD interaction: p = 1.0e−5).
Examples of 2 genetically driven metabolic responses to SD are illustrated in Fig 4E. The change in PC-ae-C32:2 after SD mapped significantly to chromosome 5 (LOD = 3.6; 58–69 Mb; genotype × SD interaction: p = 2.0e−3). The change in acylcarnitine C18:1, the strongest among all metabolites assayed (Fig 4B), mapped suggestively to chromosome 18 (LOD = 3.6; 73–75 Mb; genotype × SD interaction: p = 2.0e−3). For an additional 79 metabolites, a significant genotype × SD interaction was obtained that mapped at the suggestive level (see https://bxd.vital-it.ch; Downloads, Genotype_SD_Interaction.xlsx). Finally, significant cis-eQTLs were detected for the DE (i.e., recovery versus control) of 195 genes after SD in cortex and 62 in liver (see https://bxd.vital-it.ch; Downloads, Genotype_SD_Interaction.xlsx and cis_eQTL.xlsx). The strongest cis-allele in cortex was found for the DE of phospholipase A2, group IVE (Pla2g4e; rs47077493, chr2: 118.3 Mb, q = 1.2e−9) with a down-regulation that was 2-fold larger in B6- than in D2-allele carriers (genotype × SD interaction: p = 1.0e−9; Fig 4F). Also illustrated are the effects of SD on malonyl-CoA decarboxylase (Mlycd) expression for which a cis-eQTL was identified (rs33610973, chr8: 120.8 Mb, q = 1.9e−5). In BXD lines carrying a B6-allele at the cis-eQTL position, a down-regulation of Mlycd was observed, while the opposite was true for D2-allele carriers (genotype × SD interaction: p = 2.0e−4; Fig 4F). Pla2g4e encodes a phospholipase promoting the formation of free fatty acids (FFAs), while Mlycd encodes an enzyme promoting mitochondrial fatty acid oxidation. One last example of a significant differential cis-eQTL, for Werner syndrome RecQ like helicase (Wrn), will be discussed in detail below (see Example 1 in the Systems genetics of the effects of SD section). It should be noted that for most of the significant differential cis-eQTLs, including Wrn, DE and the absolute expression after SD were highly correlated (>0.5; 140/195 in cortex), and both were regulated by shared cis-eQTLs (161/195).
In the following 4 sections, we highlight 4 phenotypes quantified during recovery from SD that emerged from our systems genetics analyses because of the presence of strong genetic evidence at all levels of organization. Two concern the levels of EEG delta power reached after SD, 1 concerns the gain in time spent in NREM sleep during recovery, and, as a last example, the changes in TPF during REM sleep in recovery. While for the first 3 phenotypes abundant evidence exists documenting their change with SD and their relevance in optimal daytime functioning and health, the latter phenotype (which has not been reported on previously) illustrates that, depending on genotype, a phenotype can either increase or decrease after sleep loss. Moreover, this example shows that phenotypes considered strictly “central” (i.e., the frequency of hippocampal theta oscillations) are strongly associated with genomic loci affecting gene expression in the periphery and not in the brain. It is important to point out that the genomic loci identified for these 4 recovery phenotypes appear after SD only and not (even at the suggestive level) under baseline conditions. Of equal importance is pointing out that our analyses cannot provide causal proof; instead, the systems genetics approach’s power lies in generating new hypotheses that need experimental confirmation. A first step in that direction was made in Example 4 below.
We have generated a rich, multidimensional, experimentally determined knowledge base, drawing on 4 levels of organization from the DNA level to steady-state RNA levels in brain and liver, circulating metabolites, and a deep phenome of sleep-wake-related phenotypes, all under 2 experimental conditions. At the core of this knowledge base is the BXD ARIL resource. This mouse GRP provides a “population model” with a controlled and stable degree of genetic variation, each line carrying a fixed and unique pattern of recombination of the 2 parental chromosomes [17]. The panel segregates for approximately 5.2 million sequence variants corresponding to about half of all common genetic variation among classic laboratory mouse strains [97]. This level of genetic complexity exceeds that in many human populations, such as the Icelandic and Finnish populations that have been so useful in genetics of disease [98–100]. Our results underscore the power of the BXD panel in discovering the genetic and molecular underpinnings of clinically relevant traits already demonstrated in other research fields [19–21].
We extracted 341 sleep-wake-related phenotypes belonging to 120 distinct phenotypic modules from each individual mouse. Half of these phenotypes had higher than 0.68 heritability, indicating that they are amenable to genetic dissection even when using only 33 ARILs. Although numerous knockout studies have shown that (lack of) single genes impact many of the phenotypes we quantified (for review, see [8,101]), we demonstrate here that even highly heritable traits are determined by the interaction of several small-effect loci. Two striking examples of such traits are TPF during REM sleep and the gain in δ2 power after SD, for which we identified 4 and 5 suggestive QTLs, respectively, that together explained 58% and 75% of the genetic variance in these 2 traits. Thus, while reductionist approaches have been successful at identifying genes affecting sleep in a mendelian fashion, when studied at a more natural population level, most of these phenotypes represent complex traits, and mendelian (or null) alleles are likely to play a lesser role. To systematically explore these nonadditive, multiloci interactions at the level of the whole genome, innovative algorithms in the area of machine learning are needed. Currently, more than 2-way epistatic interactions are computationally challenging. We are therefore now exploring novel multiloci epistatic approaches to extract this type of information (see, e.g., [102,103]).
With the 4 examples described, we could only illustrate a fraction of all the novel information contained in our experimentally derived knowledge base. Here, we focused on the effects of sleep loss exclusively because systems genetics resources in this research domain are lacking and because of the immediate clinical relevance of these effects. Importantly, the pathways we identified were unique to the sleep-deprivation condition and did not explain phenotypic variance of the respective traits under undisturbed baseline conditions. This illustrates that already a relatively mild sleep disruption (preventing sleep during half of the rest phase) extensively reshapes the systems genetics landscape.
The power of systems genetics lies in generating hypotheses. In the current dataset, several observations imply SD to challenge fatty acid turnover. Besides Acot11—which regulates the levels of FFAs and, as we show here, the recovery of NREM sleep—also Cyp4a32, which contributes to the SD-induced shift in the frequency of theta oscillation in REM sleep, encodes an enzyme regulating fatty acid levels. This frequency shift was strongly correlated with levels of the branched amino acids leucine, isoleucine, and valine, which, in turn, are part of a fatty acid biosynthesis pathway. The link between Cyp4a32 and the dominant frequency of theta oscillatory activity also illustrates the importance of a peripheral molecular pathway in regulating brain activity, as Cyp4a32 was not expressed in brain. This finding is of relevance because although many studies have emphasized the deleterious effect of sleep loss on peripheral systems, research on the substrate of sleep need largely remains brain centric. In addition, Pla2g4e and Mlycd, the 2 genes with the strongest cis-eQTL effect for their DE after SD, both encode enzymes affecting fatty acid metabolism. Acot11, the Cyp4a gene family, FFA levels, and sleep restriction have all been linked to obesity and insulin resistance [82,91,93–96]. Another pathway of importance in mediating the effects of sleep loss concerns AMPA-R trafficking supported by the 8-fold increase in cortical Arc expression and Kif16b’s role in shaping δ2 power after SD. Both genes encode proteins involved in the endosomal trafficking of AMPA-Rs (see Results) that have already been explored as therapeutic targets to counter the deleterious effects of SD on cognition [73,76]. Finally, Wrn‘s association with EEG slow waves during NREM sleep offers a model system to mechanistically study the molecular pathways underlying the characteristic age-related decrease in the prevalence of EEG slow waves and sleep quality.
Hypotheses concerning the involvement of the pathways in the sleep homeostatic process we discovered need to be further tested experimentally. With a reverse genetics approach, we could already confirm Acot11’s role in the recovery of sleep time lost. This approach is, however, not always informative or possible, because a lack of protein on a given genetic background is unlikely to mimic the impact of an allelic variant in a genetically diverse population, or the knockout might be lethal, as is the case for Kif16b [69]. Efforts to comprehensively phenotype (including sleep) knockouts for all known and predicted mouse genes by the International Mouse Phenotyping Consortium (IMPC; www.mousephenotype.org) are ongoing, but unfortunately, no knockouts for the 4 genes we highlight here have been submitted for phenotyping. Another important community resource is the mostly mouse-oriented database GeneNetwork (www.genenetwork.org), which hosts a massive amount of phenotypic and molecular information collected by the many researchers using the same BXD resource. We are in the process of structuring our database to enable sharing of the integrated data in GeneNetwork according to the FAIR data management concepts [104]. Furthermore, cross-species validation in, e.g., humans, flies, and Caenorhabditis elegans and Genome-Wide Association Study (GWAS) and biobank database searches are important additional ways of validating and extending our mouse observations. According to the human GWAS databases grasp.nhlbi.nih.gov and www.ebi.ac.uk/gwas/, SNP variants in Acot11 are significantly associated with (among others) the rate of cognitive decline in Alzheimer disease, behavioral disinhibition, cardiovascular disease, and triglyceride levels. Variants in Wrn are associated with aging and time to death, cardiovascular disease, cholesterol, and daytime rest. Finally, variants in the human ortholog of Cyp4a32, CYP4A11, are associated with blood metabolite levels, including amino acids and acyl carnitines, and Kif16b variants with intelligence.
A first evaluation of the systems genetics field has highlighted a clear need for better communication, “open science,” and collaboration among groups [24]. Toward this aim, we have shared our results and analyses through an easily accessible and reproducibility-oriented web interface that accompanies this publication. We hope that the interactivity of the web interface will encourage the reader to further mine our data, thereby reproducing our conclusions and, hopefully, discovering other key regulators and pathways. In our analyses, we have also strived to follow the concepts of the FAIR data management approach [104], resulting in a data life cycle management plan, open access provided by the web interface for data mining, and, importantly, interoperability. The implementation of the FAIR approach will be illustrated in an accompanying publication.
In summary, we have applied a systems genetics approach to uncover new genes and pathways associated with the effects of sleep loss, an approach thought critical for predicting disease susceptibility [18]. This integrative, multilevel approach allowed us to follow the flow of information from DNA variants to molecular intermediate phenotypes to behavioral and electrophysiological end phenotypes, and to assess how this network of multiscale effects is perturbed by an environmental challenge. The information gained could not have been achieved through other genetic approaches that are based on the “1-gene-to-1-phenotype” approach. Moreover, with the tools and web interface we developed, our open-access knowledge base provides a unique resource that goes well beyond merely cataloguing and ranking ph-, m-, and eQTLs. Furthermore, owing to the use of a GRP, the database and its content are easily scalable. A first challenge will be to complement the dataset with females of the same lines. In addition, we are expanding the database with an additional intermediate phenotype—namely, the SD-induced changes in chromatin accessibility—aiming to identify the variants in noncoding regulatory elements that could predict the varying molecular and phenotypic response to sleep loss. Proteome, microbiome, and inflammasome data are obvious other intermediate phenotypes that will further strengthen this knowledge base and increase its value to, e.g., assist with identifying biomarkers gauging sleep pressure and potential therapeutic targets for sleep-wake-related disorders.
All experiments followed international guidelines and were approved by the veterinary authorities of the state of Vaud, Switzerland (SCAV authorization #2534). Animals assigned to Experiment 1 (see Experimental design below and Fig 1) were equipped with chronic EEG and EMG electrodes under deep anesthesia according to methods described in detail in [105]. In short, IP injection of Xylazine (10 mg/kg)/Ketamine (100 mg/kg) ensures a deep plane of anesthesia for the duration of the surgery (i.e., around 30 min). Analgesia was provided the evening prior and the 3 d after surgery with Dafalgan in the drinking water (200–300 mg/kg). Mice were allowed to recover for at least 10 d prior to baseline recordings. Animals assigned to Experiment 2 (see Experimental design below and Fig 1) were killed by decapitation after being anesthetized with isoflurane, upon which blood, cerebral cortex, and liver samples were collected immediately.
We phenotyped 33 BXD RI strains originating from the University of Tennessee Health Science Center (Memphis, TN, United States of America). The 33 lines were randomly chosen from the then available, newly generated ARIL panel [17], although lines with documented poor breeding performance were not considered. Two breeding trios per BXD strain were purchased from a local facility (EPFL-SV, Lausanne, Switzerland) and bred in-house until sufficient offspring was obtained. The parental strains D2 and B6 and their reciprocal F1 offspring (B6D2F1 [BD-F1] and D2B6F1 [DB-F1]) were bred and phenotyped alongside. Suitable (age and sex) offspring was transferred to our sleep-recording facility, where they were singly housed, with food and water available ad libitum, at a constant temperature of 25°C and under a 12 h light/12 h dark cycle (LD12:12, fluorescent lights, intensity 6.6 cds/m2, with ZT0 and ZT12 designating light and dark onset, respectively). Male mice aged 11–14 wk at the time of experiment were used for phenotyping, with a mean of 12 animals per BXD line among all experiments. Note that 3 BXD lines had a lower replicate number (n), with respectively BXD79 (n = 6), BXD85 (n = 5), and BXD101 (n = 4) because of poor breeding success. For the remaining 30 BXD lines, replicates were distributed as follows: for EEG/behavioral phenotyping (Experiment 1 in Fig 1; mean = 6.2/line; 5 ≤ n ≤ 7) and for molecular phenotyping (Experiment 2 in Fig 1; mean = 6.8/line; 6 ≤ n ≤ 9). Additionally, to assess the stability of outcome variables over time, parental lines were phenotyped twice—i.e., at the start (labeled B6-1 and D2-1) and end (labeled B6-2 and D2-2) of the breeding and data-collecting phase, which spanned 2 y (March 2012–December 2013). To summarize, distributed over 32 experimental cohorts, 227 individual mice were used for behavioral/EEG phenotyping (Experiment 1) and 256 mice for tissue collection for transcriptome and metabolome analyses (Experiment 2), the latter being divided into sleep deprived (SD) and controls (“Ctr”; see Experimental design section below). We strived to randomize the lines across the experimental cohorts so that biological replicates of 1 line were collected/recorded on more than 1 occasion while also ensuring that an even number of mice per line was included for tissue collection so as to pair SD and “Ctr” individuals within each cohort (for behavioral/EEG phenotyping, each mouse serves as its own control).
The study consisted of 2 experiments, i.e., Experiments 1 and 2 (Fig 1). Animals of both experiments were maintained under the same housing conditions. Animals in Experiment 1 underwent surgery and, after a >10 d recovery period, EEG and LMA were recorded continuously for a 4 d period starting at ZT0. The first 2 d were considered baseline (B1 and B2). The first 6 h of Day 3 (ZT0–6), animals were sleep deprived in their home cage by “gentle handling” [105]. The remaining 18 h of Day 3 and Day 4 were considered recovery (R1 and R2). Half of the animals included in Experiment 2 were sleep deprived (SD) alongside the animals of Experiment 1. The other half was left undisturbed in another room (i.e., control or Ctr). Both SD and “Ctr” mice of Experiment 2 were killed at ZT6 (i.e., immediately after the end of the SD) for sampling of liver and cerebral cortex tissue as well as trunk blood. All mice were left undisturbed for at least 2 d prior to SD.
RNA-seq data were processed using the Illumina Pipeline Software version 1.82. All RNA-seq samples passed FastQC quality thresholds (version 0.10.1) and could thus be used in subsequent analysis. For gene expression quantification, we used a standard pipeline that was already applied in a previous study [111]. Reads were mapped to MGSCv37/mm9 using the STAR splice aligner with the 2pass pipeline [112]. Count data was generated using htseq-count from the HTseq package using parameters “stranded = reverse” and “mode = union” [113]. Gene boundaries were extracted from the mm9/refseq/reflat dataset of the UCSC table browser. EdgeR was then used to normalize read counts by library size. Genes with a mean raw read count below 10 were excluded from the analysis, and the raw read counts were normalized using the TMM normalization [114] and converted to log counts per million (CPM). Although for both tissues, the RNA-seq samples passed all quality thresholds, and among-strain variability was small, more reads were mapped in cortex than in liver (S6 Fig), and we observed a somewhat higher coefficient of variation in the raw gene read count in liver than in cortex (S6 Fig). To assess the DE between the sleep-deprived and control conditions, we used the R package limma [115] with the voom weighting function followed by the limma empirical Bayes method [116]. RNA-seq data are deposited in NCBI GEO (accession code GSE114845).
The RNA-seq dataset was also used to complement the publicly available GeneNetwork genetic map (www.genenetwork.org), thus increasing its resolution. RNA-seq variant calling was performed using the Genome Analysis ToolKit (GATK) from the Broad Institute, using the recommended workflow for RNA-seq data [117]. To improve coverage depth, 2 additional RNA-seq datasets from other projects using the same BXD lines were added [111]. In total, 6 BXD datasets from 4 different tissues (cortex, hypothalamus, brainstem, and liver) were used. A hard filtering procedure was applied as suggested by the GATK pipeline [117–119]. Furthermore, genotypes with more than 10% missing information, low quality (<5,000), and redundant information were removed. GeneNetwork genotypes, which were discrepant with our RNA-seq experiment, were tagged as “unknown” (mean of 1% of the GeneNetwork genotypes/strain [0.05% ≤ n ≤ 8%]). Finally, GeneNetwork and our RNA-seq genotypes were merged into a unique set of around 11,000 genotypes, which was used for all subsequent analyses. This set of genotypes was already used successfully in a previous study of BXD lines [111] and is available through our “Swiss-BXD” web interface (https://bxd.vital-it.ch; Downloads, Genotypes.GeneNetwork2005AndRNAseq.geno).
Although overall, a close to 50/50 balance between B6 and D2 genotypes was observed across the genome, a minority of sites displayed a strong imbalance toward either genotype (S7 Fig). We also confirmed a minor but general trend toward more D2 than B6 genotypes per strain (S7 Fig), which was also found in the GeneNetwork genotypes for the BXD strains used in our study.
The R package qtl/r [120] was used for interval mapping of behavioral/EEG phenotypes (phQTLs) and metabolites (mQTLs). Pseudomarkers were imputed every cM, and genome-wide associations were calculated using the Expected-Maximization (EM) algorithm. p-values were corrected for FDR using permutation tests with 1,000 random shuffles. The significance threshold was set to 0.05 FDR, a suggestive threshold to 0.63 FDR, and a highly suggestive threshold to 0.10 FDR according to [28,29]. QTL boundaries were determined using a 1.5 LOD support interval. To preserve sensitivity in QTL detection, we did not apply further p-value correction for the many phenotypes tested. Effect size of single QTLs was estimated using 2 methods. Method 1 does not consider eventual other QTLs present and computes effect size according to 1 − 10^(−(2/n)*LOD). Method 2 does consider multi-QTL effects and computes effect size by each contributing QTL by calculating first the full, additive model for all QTLs identified and, subsequently, estimating the effects of each contributing QTL by computing the variance lost when removing that QTL from the full model (“drop-one-term” analysis). For Method 2, the additive effect of multiple suggestive, highly suggestive, and significant QTLs was calculated using the fitqtl function of the qtl/r package [121]. With this method, the sum of single QTL effect estimation can be lower than the full model because of association between genotypes. In the Results section, Method 1 was used to estimate effect size, unless specified otherwise. It is important to note that the effect size estimated for a QTL represents the variance explained of the genetic portion of the variance (between-strain variability) quantified as heritability and not of the total variance observed for a given phenotype (i.e., within- plus between-strain variability).
For detection of eQTLs, cis-eQTLs were mapped using FastQTL [122] within a 2 Mb window for which adjusted p-values were computed with 1,000 permutations and beta distribution fitting. The R package qvalue [123] was then used for multiple-testing correction as proposed by [122]. Only the q-values are reported for each cis-eQTL in the text. Trans-eQTL detection was performed using a modified version of FastEpistasis [124], on several million associations (approximately 15,000 genes × 11,000 markers), applying a global, hard p-value threshold of 1E−4.
Variants detected by our RNA-seq variant calling were annotated using Annovar [125] with the RefSeq annotation dataset. Nonsynonymous variations were further investigated for protein disruption using Polyphen-2 version 2.2.2 [126], which was adapted for use in the mouse according to recommended configuration.
Hiveplots were constructed with the R package HiveR [25] for each phenotype. Gene expression and metabolite levels represented in the hiveplots come from either the “Ctr” (control) or SD molecular datasets according to the phenotype represented in the hiveplot; i.e., the “Ctr” dataset is represented for phenotypes related to the baseline (“bsl”) condition, while the SD dataset is shown for phenotypes related to recovery (“rec” and “rec/bsl”). For a given hiveplot, only those genes and metabolites were included (depicted as nodes on the axes) for which the Pearson correlation coefficient between the phenotype concerned and the molecule passed a data-driven threshold set to the top 0.5% of all absolute correlations between all phenotypes on the one hand and all molecular (gene expression and metabolites) on the other. This threshold was calculated separately for “Bsl” phenotypes and for “Rec” and “Rec/Bsl” phenotypes and amounted to absolute correlation thresholds of 0.510 and 0.485, respectively. The latter was used for the recovery phenotypes in Results Examples 1–4 and for the printed hiveplots (other thresholds can be chosen in the interactive website https://bxd.vital-it.ch). Cross-associations between genes and metabolites represented by the edges in the hiveplot were filtered using quantile thresholds (top 0.05% gene–gene associations, top 0.5% gene–metabolite associations). We corrected for cis-eQTL confounding effects by computing partial correlations between all possible pairs of genes (see Results and Fig 4B and 4C for details).
In order to prioritize genes in identified QTL regions, we chose to combine the results of the following analyses: (i) QTL mapping (phQTL or mQTL, Fig 2C), (ii) correlation analysis, (iii) expression QTL (eQTL, Fig 2B), (iv) protein damaging–variation prediction, and (v) DE (Fig 3A). Each result was transformed into an “analysis score” using a min/max normalization, in which the contribution of extreme values was reduced by a winsorization of the results (S4 Fig). These analysis scores were first associated with each gene (see below) and then integrated into a single "integrated score" computed separately for each tissue, yielding 1 integrated score in cortex and 1 in liver. The correlation analysis score, eQTL score, DE score, and protein damaging–variation score are already associated to genes, and these values were therefore simply attributed to the corresponding gene. To associate a gene with the ph-/mQTL analysis score (which is associated to markers), we used the central position of the gene to infer the associated ph-/mQTL analysis score at that position. In case of a cis-eQTL linked to a gene or a damaging variation within the gene, we used the position of the associated marker instead (S4 Fig). To emphasize diversity and reduce analysis score information redundancy, we weighted each analysis score using the Henikoff algorithm. The individual scores were discretized before using the Henikoff algorithm, which was applied on all the genes within the ph-/mQTL region associated with each phenotype (S4 Fig). The integrated score (formula in Fig 4D) was calculated separately for cortex and liver. We performed a 10,000-permutation procedure to compute an FDR for the integrated scores. For each permutation procedure, all 5 analysis scores were permutated, and a novel integrated score was computed again. The maximal integrated score for each permutation procedure was kept, and a significance threshold was set at quantile 95. Applying the Henikoff weighting improved the sensitivity of the gene prioritization. E.g., among the 91 behavioral/EEG phenotypes quantified with 1 or more suggestive/significant QTL after SD, 40 had at least 1 gene significantly prioritized with Henikoff weighting, against 32 without. Examples of analysis scores and weight can be found in S1 Table.
|
10.1371/journal.pntd.0003922 | Elimination of Onchocerciasis from Mexico | Mexico is one of the six countries formerly endemic for onchocerciasis in Latin America. Transmission has been interrupted in the three endemic foci of that country and mass drug distribution has ceased. Three years after mass drug distribution ended, post-treatment surveillance (PTS) surveys were undertaken which employed entomological indicators to check for transmission recrudescence.
In-depth entomologic assessments were performed in 18 communities in the three endemic foci of Mexico. None of the 108,212 Simulium ochraceum s.l. collected from the three foci were found to contain parasite DNA when tested by polymerase chain reaction-enzyme-linked immunosorbent assay (PCR-ELISA), resulting in a maximum upper bound of the 95% confidence interval (95%-ULCI) of the infective rate in the vectors of 0.035/2,000 flies examined. This is an order of magnitude below the threshold of a 95%-ULCI of less than one infective fly per 2,000 flies tested, the current entomological criterion for interruption of transmission developed by the international community. The point estimate of seasonal transmission potential (STP) was zero, and the upper bound of the 95% confidence interval for the STP ranged from 1.2 to 1.7 L3/person/season in the different foci. This value is below all previous estimates for the minimum transmission potential required to maintain the parasite population.
The results from the in-depth entomological post treatment surveillance surveys strongly suggest that transmission has not resumed in the three foci of Mexico during the three years since the last distribution of ivermectin occurred; it was concluded that transmission remains undetectable without intervention, and Onchocerca volvulus has been eliminated from Mexico.
| Onchocerciasis, or river blindness, is one of the neglected tropical diseases targeted by the international community for elimination. In Mexico, onchocerciasis was historically endemic in three foci, which included Northern Chiapias, Southern Chiapas and Oaxaca. Both the criteria for verification of elimination and for post-treatment surveillance developed by the international community rely heavily on the use of entomological metrics. The absence of evidence of ongoing transmission of the parasite three years after mass drug distribution has been halted is considered to be evidence that elimination efforts have been successful. In the present study, we report entomological assessments carried out in the three endemic foci in Mexico that were performed three years following the end of mass drug distribution in each focus. None of the over 100,000 Simulium ochraceum s.l. vector black flies collected from sentinel and extra-sentinel communities in these foci were found to contain parasite DNA, suggesting vector parasite contact was non-existent. This data suggest that elimination of onchocerciasis from Mexico has been achieved.
| Onchocerciasis (river blindness) is caused by chronic infection with Onchocerca volvulus, a filarial nematode that is transmitted by Simulium spp. (Diptera: Simuliidae). The disease historically has constituted a serious public health concern and an enormous source of socio-economic disruption in many developing countries, most severely in sub-Saharan Africa and to a lesser extent in Latin America, where the parasite was introduced from Africa several centuries ago [1–6]. The current strategy for the elimination of onchocerciasis relies on mass treatment of endemic communities with ivermectin (Mectizan, donated by Merck & Co.). A variety of treatment regimens, i.e., quarterly and semi-annual treatment, have proven effective in interrupting transmission and eliminating the parasite throughout much of Latin America; semi-annual and annual treatments have also succeeded in isolated foci in Africa [7–12]. High coverage (≥ 85% of eligible persons), community-wide treatment of residents is believed to be sufficient to reduce the load of microfilariae in human hosts below the threshold that can sustain transmission by black fly vectors, thus locally eliminating the infection [7].
The elimination guidelines set forth by the Onchocerciasis Elimination Program for the Americas (OEPA) and the World Health Organization (WHO) use the prevalence of O. volvulus infective stage larvae (L3) in the black fly vectors as a major metric for determining whether or not transmission has been successfully interrupted in an endemic community [13, 14]. In Latin America, the threshold used for declaring interruption of transmission is an upper bound of the 95% confidence interval (95%-ULCI) for the point estimate of the prevalence of vectors carrying L3 of less than 1/2,000 per endemic community [14]. At least 6,000 flies must be tested and all must be found to be L3-free to satisfy this standard [15,16]. In addition to the 1/2,000 infective fly threshold, OEPA recommends the use of the Annual Transmission Potential (which in the present situation is equivalent to seasonal transmission potential [STP]) to assess the status of O. volvulus transmission, because transmission potentials take into account both the biting rate and the prevalence of infective flies. OEPA/WHO verification guidelines for onchocerciasis elimination stipulate that in areas where transmission has been interrupted and mass drug distribution has been stopped, a post-treatment surveillance (PTS) period of at least 3 years is needed [14, 17]. If surveys conducted after the PTS period show no evidence for recrudescence of transmission, then O. volvulus is considered to have been eliminated, and the resident population is no longer at risk.
In Mexico, onchocerciasis was historically endemic in three distinct foci; Southern Chiapas, Northern Chiapas, and Oaxaca (Table 1). In 1960, a total of 20,090 individuals harbored nodules (a prevalence of 15%); 135 individuals blinded by onchocerciasis were reported (representing a prevalence of 0.1%). In the Oaxaca focus 5,800 cases (i.e. individuals diagnosed positive for onchocerciasis by any of the available methods, Mazzotti reaction, nodule palpation, or skin biopsies (snips) during active case finding campaigns conducted by the Mexican onchocerciasis brigades) were reported in 1960. This represented a prevalence of 13% in the at risk population of 45,000 individuals residing in 154 communities. In contrast, in 1960, the Northern Chiapas focus had just 4,000 imported cases (residents that had regularly visited other foci where they likely acquired the infection) in an at risk population of 22,500 individuals residing in 133 communities, representing a prevalence of 18%; by 1993 only 180 cases were reported in a population of 15,539 at risk individuals in this focus, representing a prevalence of just 1%. The Oaxaca and Northern Chiapas foci were therefore considered as hypo-endemic for onchocerciasis, as the prevalence in these foci was less than 20%.
The Southern Chiapas focus was the major focus in Mexico, given its large size (12,000 km2) and a well-documented history of intense transmission (Table 1). In 1960, 26,003 cases were reported in an at risk population of 61,619 individuals residing in 837 communities were reported (a prevalence of 42%). In 1999, 22,361 cases were reported, of which 274 were classified as “new” clinical cases (i.e. individuals diagnosed positive by Mazzotti reaction, nodules, or skin biopsies for the first time, during active case finding campaigns conducted by the onchocerciasis brigades), while 782 individuals harbored nodules in an at risk population of 219,923; 31 individuals blinded by onchocerciasis were reported [18–20]. The prevalence of cases, “new” clinical cases, nodules, and onchocercal blindness were 10%, 0.1%, 0.3%, and 0.01%, respectively. In 2012, Mexico was surpassed only by Guatemala in the Americas in terms of the at risk population for onchocerciasis. The total population at risk in Guatemala and Mexico together represented 71% of Latin America´s total at risk population of 565,232 individuals.
Onchocerciasis was discovered in the Americas in 1915 by Dr. Rodolfo Robles, who described the first clinical cases in Guatemala. In Mexico, the first cases of onchocerciasis were documented in Southern Chiapas in 1923. The disease was probably introduced to this area due to the seasonal migration of coffee workers from the endemic foci of Guatemala. Similarly, the Oaxaca and Northern Chiapas foci may have resulted from the expansion of coffee cultivation into these areas and the corresponding migration of workers from the established foci of Southern Chiapas and Guatemala [21]. One of the first programs to combat onchocerciasis in the world was established in Mexico in 1930; this program has been operating continuously since then. From 1930 through 1946, the Mexican onchocerciasis control program carried out sporadic vector control campaigns, treating breeding sites with creosote as a larvicide to reduce vector populations, and nodulectomy campaigns (removal of nodules harboring adult worms) to reduce the most severe cases of the disease. Administration of diethylcarbamazine (DEC) began in 1947, when it was tested in six infected individuals. In 1949, DEC began to be provided to all clinical cases of onchocerciasis. This was augmented in 1952 with sporadic applications of dichlorodiphenyltrichloroethane (DDT) to control the vector population [21]. In 1990, ivermectin (Mectizan, Merck & Co., Inc., Whitehouse Station, NJ) replaced DEC. The Mexican program initially used ivermectin only in symptomatic individuals. However, beginning in 1997, ivermectin was provided to all individuals living in endemic communities, using a strategy of administering two rounds of treatments per year (semi-annual regimen); this was followed by a distribution of four rounds of treatment per year (quarterly regimen) from 2003 through 2011 in the Southern Chiapas focus [18]. The increase in the frequency of ivermectin treatments proved to be a good strategy, accelerating the interruption of parasite transmission in this focus [8]. Recent in-depth epidemiological studies based on entomological, parasitological, serological, and ophthalmological surveys conducted in individuals of endemic communities have demonstrated the interruption of parasite transmission in all three endemic foci in Mexico [18–20]. These results led to the cessation of the treatment with ivermectin by the Ministry of Health of Mexico. The endemic communities then entered the post-treatment surveillance (PTS) period.
Here, we present the results of PTS entomological surveys carried out in the three endemic foci of Mexico. Taken together, the results demonstrate that transmission has not resumed in the three years since the last distribution of ivermectin occurred. Mexico has entered the post-endemic era and now appears to be free of the scourge of onchocerciasis.
Flies were collected using human attractants from 18 sentinel and extra-sentinel communities as previously described [7, 8]. Four out of 13 endemic communities in the Northern Chiapas focus, 6 out of 98 endemic communities in the Oaxaca focus, and 8 out of 559 in the Southern Chiapas focus were included in the surveys (Fig 1). All communities were previously either meso- or hyper-endemic for onchocerciasis, and they were generally the communities with the most intense transmission in each focus before interventions began (Fig 2). In the two endemic states of Mexico, 39 communities were hyper-endemic and 220 and 411 were meso and hypo-endemic respectively. Meso-endemic communities were defined as having a historical onchocerciasis prevalence of more than 20% but less than 60% while hyper-endemic communities had a historical prevalence of 60% or higher. Hypo-endemic communities were those with a historical prevalence less than or equal to 20%.
In the three endemic foci, the ivermectin treatment regimen was generally provided on a semi-annual basis. A quarterly treatment regimen was employed from 2003–2011 in communities of the Southern Chiapas focus, as described above. MDA coverage rates (percent) of the eligible population in the three foci are summarized in Fig 3.
All procedures involving use of humans for fly collections were reviewed and approved by the Bioethics Committee of the Center for Research and Development in Health Sciences of the Autonomous University of Nuevo León (Monterrey, Nuevo León, Mexico). Written informed consent was obtained from all fly collectors.
Black fly collection was carried out by two teams in each community, with each team consisting of a fly collector and a human attractant. One team was located at a randomly selected location within the community while the second team conducted collections at nearby coffee plantation. Black fly collections were performed during the dry seasons of November 2009 through February 2010, January through March 2011, and December 2013 through May 2014 in the foci of Northern Chiapas, Oaxaca, and Southern Chiapas, respectively. These collection periods coincided with the peak Simulium ochraceum sensu lato population densities and the peak O. volvulus transmission season [22]. S. ochraceum s.l. is the major vector of O. volvulus in Mexico. No other species of black fly has been documented to be an important vector in the Mexican foci [22].
Collections were performed during the first 50 min of each hour, beginning at 11:00 h and ending at 16:50 hrs. Collectors received ivermectin 1 week before beginning the collection process. The black fly collections were supervised by federal health officials to ensure that collections were conducted throughout the entire 50 min collection period of each hour; the remaining 10 min of each hour was utilized as a break period. The officials also ensured that the black flies were stored appropriately. The black flies were transported every hour to a field station for observation under a dissecting microscope. The flies were identified to the species level; flies other than S. ochraceum s.l. and any fly containing evidence of having taken a recent blood meal were discarded. The number of S. ochraceum s.l. collected by each team in each 50-minute period was recorded. The S. ochraceum s.l. females were then combined into pools using a pool size of 50 flies/pool for the samples collected in the Oaxaca and Northern Chiapas foci and a pool size of 200 flies/pool for those samples collected in the Southern Chiapas focus.
Black flies were collected before they began blood-feeding. The landing rate as measured from the collections was taken as an estimate of the biting rate, although this probably overestimated the biting rate, as a proportion of the landing flies in a natural setting do not successfully obtain a blood meal. Flies were combined into pools as described above and the heads and bodies separated by freezing and agitation, as previously described [7, 8, 23, 24]. The separated bodies were tested for O. volvulus parasites by using a PCR assay specific for O. volvulus, as previously described [7, 8, 23, 24]. Screenings initially focused on pools of bodies, as previous studies have shown that infection rates in bodies, which contain multiple life cycle stages of the parasite, provide a more sensitive indicator of parasite-vector contact than testing heads, which only contain L3 larvae [23, 25]. PoolScreen v2.0 was used to estimate the upper bound of the 95% confidence interval for the prevalence of flies carrying O. volvulus [26]. The seasonal transmission potential (STP) was calculated as the product of the seasonal biting rate, the proportion of flies carrying L3 and the average number of L3 larvae in each infective fly. After multiple rounds of ivermectin treatment, the number of L3 present in each infective fly was assumed to be one, as previously described [7, 8, 18]. The seasonal biting rate was calculated as the product of the arithmetic mean of the number of flies collected per person per day and the total number of days in the transmission season. Because S. ochraceum s.l. females were not collected throughout the year, it was not possible to precisely calculate the annual transmission potential (ATP). However, given the low abundance of vector black flies present outside the normal transmission season, the transmission potential outside of the peak transmission period is probably zero or near zero [22]. The STP (transmission occurring during the peak transmission dry season of December through May) thus likely represented a fairly accurate estimate of the ATP.
In the Northern Chiapas focus, totals of 5,731 and 5,476 host-seeking S. ochraceum s.l. females were collected from four extra-sentinel communities in the community and coffee plantation collection sites respectively. These were divided into a total of 230 pools, each containing a maximum of 50 individuals for PCR analysis. In the Oaxaca focus, a total of 11,148 and 17,494 host-seeking S. ochraceum s.l. females were collected in the community and coffee plantation sites, respectively. These were divided into a total of 582 pools, each containing a maximum of 50 individuals for PCR. Finally, a total of 40,001 and 28,362 host-seeking S. ochraceum s.l. females were collected in the community and coffee plantation sites of Southern Chiapas respectively. These were divided into a total of 362 pools, each containing a maximum of 200 individuals. The number of vectors collected was sufficient to comply with the WHO/OEPA guideline of having at least 6,000 flies tested from each focus.
None of the pools of S. ochraceum s.l. collected in 2010 and 2011 in the Northern Chiapas and Oaxaca foci (11,207 and 28,642 flies respectively) were found to be positive in the PCR assay. Thus, the associated upper limit of the 95% confidence interval (95% ULCI) for the prevalence of flies carrying O. volvulus were 0.3 and 0.13/2,000 flies for Northern Chiapas and Oaxaca respectively, both of which were below the threshold of a 95%-ULCI of 1/2,000 mandated by the international community as sufficient to declare that transmission had been eliminated (Table 2). Similarly, the 68,383 flies collected from the Southern Chiapas focus were also all found to be negative for parasite DNA. In this case, the 95%-ULCI for the prevalence of infection in the vector was just 0.1/2000 (Table 2).
The point estimate for the STP in all foci was zero. The 95%-ULCI for the STP was 1.3 and 1.2 L3/person/season in the Northern Chiapas and Oaxaca foci respectively, while in the Southern Chiapas focus the 95%-ULCI for the STP was 1.7 L3/person/season (Table 2). These values were well below the ATP breakpoint for transmission, which has been estimated by various sources to be between 5–20 L3/person/year [27, 28]. Taken together, these data suggest that no parasite-vector contact was occurring in any of the foci in Mexico three years following the end of mass Mectizan distribution.
Onchocerciasis was historically endemic in three foci in Southeastern Mexico; Northern Chiapas, Southern Chiapas and Oaxaca. The smallest of these, Northern Chiapas, was also the first in which transmission of O. volvulus was reported to have been interrupted following exhaustive clinical, epidemiological and entomological surveys [19]. This occurred in 2007, following 10 years of semi-annual Mectizan mass treatment of the at-risk communities [19, 23, 29, 30]. Northern Chiapas was followed in 2008 by the second largest focus in Mexico, Oaxaca (after 13 years of semi-annual treatment) [7, 20], and finally by the largest focus, Southern Chiapas, in 2011 (following 17 years of semi-annual and quarterly treatments aimed at hastening onchocerciasis elimination) [18]. Following the declaration of the interruption of transmission in each of these foci, the Mexican Ministry of Health accepted the recommendations of the Program Coordinating Committee (PCC) of OEPA, and discontinued ivermectin treatments in these foci. Following the recommendations by OEPA and WHO [13, 14, 17], post treatment surveys were conducted three years after the end of mass drug distribution activities in each focus. These surveys focused on entomological monitoring for evidence of transmission, as this represents the earliest indicator of ongoing transmission available [14]. The data collected from these surveys, which are reported here, uncovered no evidence of transmission three years after treatment was halted in any of the three foci in Mexico. These results suggest that parasite transmission has not resumed in the three years since drug pressure was removed, and that O. volvulus has therefore been eliminated from each of the foci. Mexico as a country can now declare national elimination of onchocerciasis, and request WHO for verification of elimination.
The 95%-ULCI for the prevalence of flies carrying O. volvulus parasites in all three foci was found to be much less than the 1/2000 threshold developed by OEPA and WHO; furthermore more than 10,000 flies were tested from each focus, satisfying the criterion developed by WHO for verifying onchocerciasis elimination [17]. However, it has been pointed out that the risk of recrudescence is in part dependent upon the number of vectors biting residents of affected communities, and that measurements of ATP, which take biting rates into account, may be a better indicator of the risk of recrudescence than the prevalence of infection in the vector population alone [31]. Estimates of the ATP necessary to maintain the parasite population (the transmission breakpoint) range from 5 to 54 L3/person/year based on mathematical modeling [27] and from 7.6 to 18 L3/person/year based on field observations [28]. In the PTS surveys reported here, the point estimates for the STP for all foci were all zero, with the 95%-ULCI for all foci falling below 2 (Table 2). Thus, these data confirm that three years following the cessation of mass drug treatments in the population, transmission values remained well below the transmission breakpoint.
The Northern Chiapas focus historically had little autochthonous transmission; onchocerciasis cases in this focus were believed to have resulted from importation from Southern Chiapas and/or Guatemala. Oaxaca, while clearly having autochthonous transmission, was a much smaller focus than Southern Chiapas, both in terms of area and at risk population (Table 1). Interruption of transmission was successfully achieved in the two smaller foci using a semi-annual ivermectin treatment regimen. In contrast, progress in Southern Chiapas was delayed relative to Northern Chiapas and Oaxaca, and it was necessary to move to a quarterly treatment regimen to accelerate the process towards the interruption of transmission (Fig 3 Panel D). In all foci, the major challenge faced by the program was in obtaining and maintaining the coverage rates needed to ensure interruption of transmission. Mainly, this challenge resulted from two classes of individuals who were not receiving ivermectin; those who were chronically absent during the days their community was treated and those who were consistently non-compliant with respect to the program. A third untreated group were those ineligible for treatment, i.e. individuals under the age of five, or pregnant or lactating women. Individuals in the latter group generally received treatment once reaching eligible age, or when no longer pregnant or lactating. With respect to those who were chronically absent during the treatment period, coverage was found to improve significantly when the program moved from semi-annual to quarterly treatments. This change of strategy allowed the brigades to locate and treat people who were absent during the semi-annual visits. The brigades also performed many educational campaigns promoting health and preventing disease; these had the effect of informing the population reaching the people who were not compliant at the beginning of the program and convincing them of the importance of being treated, improving the coverage rate as a whole.
It is unlikely that onchocerciasis will be re-introduced to the Mexican foci from elsewhere, as these foci are well isolated from others in Latin America, and, unlike some species in Africa, the vectors responsible for transmission in Latin America do not migrate seasonally [32]. Similarly, migrant workers who cross the Mexico–Guatemala border are unlikely to pose a threat for re-introduction of the infection, as both nations have now interrupted transmission [33]. Furthermore, transmission has already been interrupted in most of the other foci in the Americas, which themselves are entering the post-treatment surveillance phase [33]. Nonetheless, it will be important to continue some surveillance activities for the next few years to ensure that transmission does not re-occur.
The studies reported above report collections that were carried out at a number of sentinel and extra-sentinel communities in each focus and do not represent comprehensive surveys of all communities at risk in each focus. For example, the total population of the 18 communities studied here was 6,738 individuals, which represents only 4% of a total at risk population in the 670 communities under PTS. Thus, although the communities chosen were generally those with the most intense transmission before the program began, and thus are expected to represent the worst case scenario, it is still possible that a low level of transmission still might occur in communities that were not included in the surveys. Indeed, mathematical models predict low levels of transmission are likely to occur, but that the reproductive rate will remain well below 1.0 and the overall parasite population would inevitably decline to extinction [14]. In light of the PTS entomologic findings in the three endemic foci, it appears that the parasite reproductive rate is now at a negligible level. This has resulted in a level of transmission that is no longer detectable, even in the absence of intervention, making it probable that the parasite population will not be able to recover even in the absence of any control measures. Despite this, the Mexican Ministry of Health has recognized it will be prudent to continue to conduct periodic clinical and entomological surveys in the formerly endemic states to ensure that transmission does not recrudesce [34]. The recent development of traps that can replace human landing collections for S. ochraceum s.l., the major vector of onchocerciasis in Mexico [35] should facilitate this process.
Additional data suggest that clinical onchocerciasis has also been eliminated in three endemic foci (Fig 4). There was a near absence of new clinically defined cases of onchocerciasis during the last years before the PTS phase began. Only five new clinical cases (i.e., individuals diagnosed positive for nodules or skin microfilariae for the first time) were reported in 1996 in the Northern Chiapas focus, while just nine new clinical cases were reported in Southern Chiapas in 2010 (Fig 4; Panel A). Similarly, no new clinical cases were reported in the Oaxaca focus from 2000 through 2007 (Fig 4; Panel B). These findings suggest that onchocerciasis no longer represents a health problem in the formerly endemic communities in Mexico.
In addition to Mexico, the programs of Ecuador [36] and Guatemala [37–39] have reported success in eliminating transmission of Onchocerca volvulus by the use of community-wide Mectizan distribution. The data presented here suggest that transmission interruption has been achieved in Mexico, resulting in elimination of this disease from the entire country. Coupled with recent studies in Mali, Nigeria, Senegal, Uganda and Northern Sudan that have indicated that ivermectin distribution may lead to focal elimination of onchocerciasis in certain African settings [40, 9–11, 41, 42], these findings give hope to the concept that worldwide elimination of this parasite is indeed possible.
|
10.1371/journal.pgen.1007194 | Paternal lineage early onset hereditary ovarian cancers: A Familial Ovarian Cancer Registry study | Given prior evidence that an affected woman conveys a higher risk of ovarian cancer to her sister than to her mother, we hypothesized that there exists an X-linked variant evidenced by transmission to a woman from her paternal grandmother via her father. We ascertained 3,499 grandmother/granddaughter pairs from the Familial Ovarian Cancer Registry at the Roswell Park Cancer Institute observing 892 informative pairs with 157 affected granddaughters. We performed germline X-chromosome exome sequencing on 186 women with ovarian cancer from the registry. The rate of cancers was 28.4% in paternal grandmother/granddaughter pairs and 13.9% in maternal pairs consistent with an X-linked dominant model (Chi-square test X2 = 0.02, p = 0.89) and inconsistent with an autosomal dominant model (X2 = 20.4, p<0.001). Paternal grandmother cases had an earlier age-of-onset versus maternal cases (hazard ratio HR = 1.59, 95%CI: 1.12–2.25) independent of BRCA1/2 status. Reinforcing the X-linked hypothesis, we observed an association between prostate cancer in men and ovarian cancer in his mother and daughters (odds ratio, OR = 2.34, p = 0.034). Unaffected mothers with affected daughters produced significantly more daughters than sons (ratio = 1.96, p<0.005). We performed exome sequencing in reported BRCA negative cases from the registry. Considering age-of-onset, one missense variant (rs176026 in MAGEC3) reached chromosome-wide significance (Hazard ratio HR = 2.85, 95%CI: 1.75–4.65) advancing the age of onset by 6.7 years. In addition to the well-known contribution of BRCA, we demonstrate that a genetic locus on the X-chromosome contributes to ovarian cancer risk. An X-linked pattern of inheritance has implications for genetic risk stratification. Women with an affected paternal grandmother and sisters of affected women are at increased risk for ovarian cancer. Further work is required to validate this variant and to characterize carrier families.
| Our article uses the largest familial study of ovarian cancer to argue that there exists an ovarian cancer susceptibility gene on the X-chromosome acting independently of BRCA1 and BRCA2. This observation implies that there may be many cases of seemingly sporadic ovarian cancer that are actually inherited; for example, only daughters who inherit risk from their fathers. This X-linked pattern implies novel ways to prioritize families for screening even without additional testing—sisters must both be carriers or neither; fathers of women with potentially inherited ovarian cancer may receive new attention. In addition, we found evidence that other cancers affect fathers and sons in these families. Using sequencing technology, we isolated a candidate gene, MAGEC3, that may be associated with earlier onset of ovarian cancer. The further study of this gene and the X-linked pattern will require additional study.
| A history of ovarian cancer among first-order relatives remains the strongest and best-characterized predictor of ovarian cancer risk [1–3] and a main determinant of genetic testing referral [4, 5]. The evidence for a monogenic, autosomal dominant mode of inherited risk dates to the pre-BRCA era where studies focused on assessing heritability [6,7] using affected first-order and second-order [8] female relatives. In a systematic review, Stratton and colleagues noted that, “not explicable in terms of any genetic model,” an affected woman’s sisters are at higher risk of disease than their mother [1]. We propose an explanation to this paradox is the existence of an X-linked gene that must pass preferentially from a carrier father to each of his daughters.
Genetic evidence of X-linkage has appeared in cytogenetic studies where loss of X-chromosome inactivation (XCI) can be visualized by loss of heterochromatin based Barr bodies [9]. Studies of ovarian tumors’ genomic profiles show loss of heterozygosity around Xq25 and Xp [10,11] as well as patterns of XCI [12,13] possibly associated with tumors of low malignant potential [14]. Studies investigating a mechanistic connection between BRCA1 and XCI [15,16], especially in tissue after transformation [17], are mixed but tend to conclude that XCI dysregulation is BRCA independent [9,18,19].
The Familial Ovarian Cancer Registry housed at Roswell Park Cancer Institute (Buffalo, NY), for over 35 years comprises over 50,000 participants and 5,600 cancers in 2,600 families. To leverage the deep pedigree data in this study, we reasoned that, if the disease allele passes through the father’s side of the family, it could be inferred by disease in a woman’s father’s mother. That is, by considering the frequency of disease transmission in grandmother/granddaughter pairs with an intermediate son/father. Under an autosomal dominant model, an affected grandmother (either maternal or paternal) passes the disease allele to her granddaughter with probability 1/4. This means that data previously presented by affected mothers and sisters are not able to discriminate between autosomal and X-linked models and these effects have been previously indistinguishable by segregation analysis due to disease censoring in fathers. In the X-linked dominant model, while a maternal grandmother again passes the disease allele to her granddaughter with probability 1/4, a paternal grandmother passes the allele to her granddaughter with probability 1/2 due to deterministic transmission by the obligate carrier son/father. Therefore, we might discriminate between autosomal versus X-linked models by considering the rate of cancers in granddaughters with exactly one affected grandmother. An autosomal dominant model predicts an equal rate of cancers in maternal-lineage and paternal-lineage pairs while an X-linked dominant model predicts that paternal-lineage families will have twice the rate of cancers (Fig 1).
We collected almost 3,500 grandmother/granddaughter pairs within the registry to test this paternal lineage hypothesis and were able to sequence 159 germlines to search for candidate variants.
We recapitulated Stratton’s paradox as outlined in Table 1: while mothers of affected women had a high risk of cancer (35%, 95% confidence interval (CI): 33–37%), sisters had a further elevated risk (66%, 65–67%). Unaffected women yielded an increased maternal risk (24%, 23–26%) while their sisters had the lowest risk (15%, 15–16%). The relative risks (RR) in the registry study (mothers RR = 1.43, sisters RR = 4.42) are close to Stratton’s original estimates (mothers RR = 1.1, sisters RR = 3.6) [1]. The effect remained for the eldest daughter in each family versus her sisters, obviating correlation due to multiple sister-pairs. We also noted that these effects were robust to stratification by familial BRCA status and families manifesting a hereditary breast and ovary pattern of disease versus families with site-specific ovary disease.
Among granddaughters with one affected grandmother (Table 2), we observed a paternal-lineage cancer rate of 28.4% (95% CI: 22.8–34.8%) and maternal-lineage cancer rate of 13.9% (11.4–16.8%). The paternal-lineage women had 2.04 times the risk (1.55–2.71) of maternal-lineage women, consistent with the X-linked dominant model that assumes the rate of paternal-lineage cancers is twice the maternal-lineage rate (goodness-of-fit, chi-square X2 = 0.2, p = 0.89). The autosomal dominant model predicted too many maternal cancers and too few paternal cancers (X2 = 20.4, p<0.001). The X-linked effect was robust versus ascertainment bias; we repeated the analysis excluding the granddaughters who were probands and observed a nearly identical relative risk (RR = 2.03, 1.28–3.23).
We observed a significant acceleration in the development of disease in granddaughters with an affected paternal grandmother versus maternal grandmother (log-rank test p = 0.009; hazard ratio HR = 1.59, 95% CI: 1.12–2.25). Granddaughters with an affected maternal grandmother were not more likely to manifest early onset disease versus women with two unaffected grandmothers (N = 2293, Log-rank p = 0.87, HR = 0.97, 95%CI: 0.75–1.24). While it does not affect the paternal/maternal effect conclusion, the subset of grandmother-granddaughter pairs possesses a higher risk than the average woman in the registry reflecting the selection bias towards families with more genetic follow up.
We considered whether men in the path of transmission were more likely to develop other cancers. In the grandmothers-granddaughter trios with affected granddaughters, the intermediate father was more likely to report a prostate cancer diagnosis if his mother had had ovarian cancer (OR = 2.34, 95%CI: 1.07–5.06, Fisher's exact test p = 0.0336) implying that the three generation pattern—ovary, prostate, ovary—was unusually common.
Because cancer-causing germline variants on X might affect the fitness of offspring and lead to an imbalance in sex birth ratio, we examined whether families were more likely to report female offspring. Removing the probands from the analysis, among BRCA negative families, there was a strong bias towards producing daughters when their mother was unaffected (female to male ratio = 1.96, difference from 1.0, p < 0.0001), putatively due to an allele transmitted by the father. The effect was attenuated for affected mothers but still significantly favored daughters (ratio = 1.21, p = 0.0131). Conservatively, this number establishes a baseline for reporting bias in these families. Among BRCA1 mutation carriers, when the mother was unaffected thereby favoring paternal transmission, there was a strong bias towards daughters (ratio = 1.19, p = 0.005). Among BRCA1 mutation families there was no difference in sex ratio when the mother was affected (ratio = 1.07, p = 0.562). This number is close to the expected population ratio suggesting minimal ascertainment bias. There were insufficient BRCA2 carrying families to make an assessment.
Among children of men who reported cancers other than prostate cancer, the sex ratio was consistent with no sex bias (ratio = 1.05, p = 0.554) and, while men with prostate cancer reported an excess of daughters (ratio = 1.12, p = 0.09), the effect was shy of statistical significance.
To estimate the potential impact of X-linked disease, we evaluated the kinship-based likelihood for autosomal dominant and X-linked models [20] for 1,386 registry pedigrees with at least two confirmed cases of ovarian cancer. Of these pedigrees, 14 (1.0%) clearly ruled out X-linked disease due to father-son transmission while 566 (40.8%) were equally likely under X-linked or autosomal likelihood models. A quarter of families slightly favored X-linkage (338, 24.4%) and 468 (33.7%) had a likelihood ratio greater than 1.5.
We illustrated this imbalanced likelihood in the pedigrees represented in Fig 2, where in the first family, 4 of 5 daughters at risk developed ovarian cancers between the ages of 43 and 51 (a sixth daughter underwent prophylactic oophorectomy). This situation favors X-linkage: assuming a carrier penetrance of 0.65 [21] and a non-carrier penetrance of 0.15 (the rate of familial ovarian cancer), the likelihood of this observation is 0.077 under the autosomal model and 0.315 under the X-linked model; a likelihood ratio of 4.10. This pattern is therefore apparent in cases of strong familial aggregation within a generation and could be inferred probabilistically. The remaining pedigrees show the paternal grandmother/granddaughter genetic logic, selecting for pedigrees where the intervening male developed prostate cancer.
To identify candidate X-linked loci, we sequenced the germline X-chromosome exome and BRCA1 coding region for 159 affected registry women who reported a negative BRCA test. The set comprised 49 cases with an affected mother only, 46 with an affected sister only, and 7 with both. Among the 2,161 common variants, one exceeded the chromosome-wide significance level (S1A Fig) at position ChrX:140,983,127 (GRCh37/hg19) with LOD = 4.91. This position mapped to rs176026, a missense SNP (Q8TD91, p.A328T) in the MAGEC3 gene on Xq27.2.
We observed 138 ovarian cancer cases with the A/A genotype, 20 women with the A/G genotype and 1 woman with the G/G genotype (Hardy-Weinberg Equilibrium (HWE), X2 = 0.086, p = 0.96) (Table 3). While the observed minor allele frequency of 6.9% (22 of 318 alleles) was not higher than the HapMap CEU frequency (5.3%, t-test p = 0.24), the frequency of A/G genotypes was significantly higher in registry women (12.6% versus 5.3%, p<0.001). Among carriers, serous histologies were common (15/21, 71%) and we observed other common epithelial types (2 endometrioid, 2 mucinous, 1 clear cell) as well as a granulosa cell tumor; these frequencies are similar to the overall distribution of ovarian cancer histologies. Women with A/A genotypes had a median age-of-onset of 50.3 years (95%CI: 47.8–52.9) reflecting the selection for family history. Women with A/G genotypes were significantly younger at diagnosis (43.6 years, 95%CI: 40.0–48.1; log-rank p<0.001, S1B Fig). Notably, all women with A/G genotypes developed cancer before age 53. The A/G effect was pronounced for women with affected sisters only (HR = 6.86, 2.18–21.56) with a median 11.0 years earlier onset (p<0.001). While women who either carry the rs176026 risk allele or were from BRCA1 variant families had similar age-of-onset (43 years), rs176026 carriers without BRCA1 variants had a stronger hazard (HR = 3.17, 1.80–5.58, p<0.001) suggesting that the X-linked variant has the stronger effect in these families.
The candidate SNP was in strong linkage with rs176024 (R2 = 0.7395, D’ = 1.0) another misssense SNP 63 base pairs away. Within the MAGEC3 coding sequence, two intronic SNPs (rs73577987, rs73577990) were also weaky linked to rs176026 (R2 = 0.1153, D’ = 0.3395 and R2 = 0.3103, D’ = 0.6478) and retained an age of onset effect. Altogether 28 women had a variant in any of rs176026 (N = 21), rs176024 (N = 14), rs73577987 (N = 18), or rs73577990 (N = 11) with a median age of onset of 53.4 (95%CI: 50.4–57.9, log-rank p<0.0001), more than 9.4 years earlier than the rest of the familial cases (HR = 2.8, 95%CI = 1.8–4.3). The frequency of the possible haploblock (rs73577987-A, rs73577990-C, rs176024-G, rs176026-G) was 1.96% (15/766) in all European populations, 0.7% (7/1003) in all African populations, and 26.8% (205/764) in all East Asian populations.
HaploReg (v4) analysis noted that the rs176026 alternate allele increase the binding affinity for a HOXA13 motif and a PU.1 (SPI1) motif. rs176024 directly affects a predicted ERα binding motif reducing the PWM score 80%. While intronic, rs73577987 affected a STAT motif and an EWSR1-FLI1 motif; rs73577990 affected BDP1, ERα, NRSF and PU.1 motifs. Functional prediction suggested that rs176026 was more likely to be deleterious to secondary protein structure: PolyPhen-2 predicted rs176026 to have a probably damaging effect (Polyphen score 0.980, sensitivity 0.75, specificity 0.96) [22] and to alter the tertiary structure of MAGEC3 [23] forcing a conformational change that impedes access to the MAGE binding domains (S1C Fig). In contrast, rs176024 was scored benign (score 0.001, sensitivity 0.99, specificity 0.15).
The canonical isoform of MAGEC3 (NM_138702.1, Uniprot: Q8TD91-1) possesses two copies of the MAGE homology domain (MHD) that defines members of the MAGE gene duplication family. Unlike other cancer-testis antigens, which are expected to have no expression in normal tissue, MAGEC3 appears to have low to moderate expression in a range of normal tissues tested by the GTEx project [24] (S2 Fig). Using the classic CT antigen MAGEA1 to provide a reference for lower limit of detection, we saw that MAGEC3 had a median expression 83x (brain) higher than MAGEA1, 47.5x (blood vessel endothelium), 9.2x (ovary) (S3 Fig). Results were similar for classic CT antigens NY-ESO-1 (CTAG1B), MAGEA3, MAGEC1 while in contrast CT-like antigens (NY-BR-1, OY-TES-1) showed higher expression in other tissues.
In contrast, the TCGA mRNA data the level was 1.2x higher (95%CI: 1.19–1.30, Affymetrix Human Exon 1.0 array) and 0.80x (95%CI: 0.65–0.98, RNAseq, Z-scores). Notably in the cancers, there were three groups present: the set around the unity line where MAGEC3 expression was at the limit of detection, a set left of the line (7.1% 23/306) where MAGEA1 is likely expressed, a set right of the line (12.4%, 38/306) where MAGEC3 may be expressed. Even for the fraction of tumors expressing MAGEC3, the level was 2.65 fold lower than normal tissue (t-test p = 0.0009). These patterns are consistent with the idea that MAGEC3 may perform a tumor suppressive function like many inherited cancer genes.
As we noted, the G allele frequency for rs176026 in the CEU HapMap study was 5%; in the African population (YRI), G is the major allele with 65% frequency (S4 Fig). These frequencies are correlated with the IARC reported incidence of ovarian cancer (Pearson’s r = 0.557 all HapMap populations) especially when excluding the Chinese American (CHD), African American (ASW) and Mexican ancestry (MEX) populations living in the United States (r = 0.858). The 1000 genomes populations were similar (r = 0.903); notably the Japanese in Tokyo, Japan (JPT), Masaii in Kinyawa, Kenya (MKK) and Luhya in Webuye, Kenya (LWK) had unusually high incidence of ovarian cancer given their minor allele frequency.
We considered whether the increased frequency of granddaughter cancers might be a result of ascertainment. Because two qualifying cancer cases are required to register a family, if a mother cannot contribute the case, then one may suspect the granddaughter is a qualifying case. In paternal grandmother/granddaughter pairs (N = 229), we noted that 164 granddaughters did not have cancer and could not be qualifying cases. The second qualifying case was a sister or paternal aunt in 79.1% of cases. In paternal pairs, 42.3% had an affected paternal aunt and 6.7% an affected maternal aunt, confirming evidence of specific lineage. The paternal/maternal granddaughter cancer rate was similar when stratifying pairs by initial ascertainment by mother/daughter (RR = 2.18, 95%CI:1.76–2.72, N = 220), sister/sister (RR = 2.51, 2.11–2.97, N = 376), or 2nd degree (RR = 2.15, 1.56–2.93, N = 203) pair. We conclude that the ascertainment bias due to an unaffected mother is minimal and the notion that the relevant pairs represent paternal lineage is well supported.
Reflecting a maternal-lineage ascertainment bias, there were more maternal than paternal pairs (663 versus 229). Mean accession identifiers for maternal and paternal grandmothers are 10.3 and 14.9, respectively (two-sample t-test, p<0.001; medians 9 and 14), suggesting that we have preferentially contacted maternal grandmothers. The direction of this potential bias is consistent with our suspicion that paternal pairs are underreported and reduces the precision of the paternal-lineage cancer rate estimate. Granddaughters with ovarian cancer in maternal families were not older than paternal-lineage women (mean difference 1.8 years, two-sample t-test p = 0.39), but unaffected women were 4.2 years older (p<0.001). While statistically significant, this difference was not likely to be the source of a 14% increase in cancer risk.
The rate of granddaughter ovarian cancers was not different between families with and without BRCA1/2 mutations (37.3% versus 38.2%, two-sample t-test p = 0.89), therefore BRCA status cannot be a confounding variable for women ascertained for family history. Alternatively, we observed the risk- doubling effect when stratifying 864 pairs into site-specific ovary (RR = 2.06, 1.45–2.92) or breast/ovary syndrome families (RR = 1.84, 1.22–2.78); there was no confounding by a family history of breast cancer. Therefore, an X-linked gene might confer ovarian cancer-specific risk independent of BRCA-type disease.
We have presented evidence that there may exist an X-linked model of transmission of an ovarian cancer susceptibility gene. Our observations are supported by a large familial study and the novel use of grandmother/granddaughter pairs to observe an increased rate of cancer among paternal granddaughters, an earlier age of onset, and a bias towards families with more daughters. We sequenced the X chromosomes of a small number of registry members in order to isolate a candidate gene, but we cannot rule out the possibility that our reported variant is in linkage with the true variant. However, the segregation analysis and age of onset analyses do suggest that it is likely to lie on the X chromosome.
Future studies are warranted to confirm the identity and function of the X-linked gene that contributes to familial transmission of ovarian cancer. Limitations of our study include the case-only design, which has required us to forgo investigating common variants. While the number of pedigrees in the registry is large, unrelated case-control studies are much larger and would likely yield other potential variants. Our study population is nearly exclusively Caucasian and our results may not extend to other populations. Our exome sequencing approach focuses on the coding regions of the X-chromosome only. This design is unable to identify intragenic variants and complex rearrangements not involving exons.
Evidence of X-linkage is not inconsistent with the prevailing autosomal dominant BRCA1/2 with polygenic weak variant effects model for ovarian cancer [8]. Ramus and colleagues [21] previously noted a lack of BRCA mutations in more than 33% of families with 3 or more ovarian cancers and 35% of families with breast/ovary cancers and concluded that there is a missing susceptibility gene. In families with two cases of ovarian cancer, the rate of BRCA mutations increased from 27% with no breast cancers to 83% with two breast cancer cases suggesting that BRCA mutations may be more specific to breast cancers. Schildkraut and colleagues [25] inferred that there must exist both shared and disease-specific genes after estimating the heritable correlation between breast and ovary cancers at h2 = 0.48. The missing gene might be ovarian cancer specific.
We suspect that the difficulty identifying this missing heritability may be due, in part, to historically inconsistent disease definition. In our literature review, we noted that studies that ascertained patients for breast cancer first and then acquired family members with ovarian cancer only saw increased risk to mothers [26]. Indeed, aggregated breast/ovary cancer studies [27] tend to show the autosomal dominant model while studies that carefully isolate ovary cancers uncover the X-linked, sister/mother effect [1]: given a family history of breast cancer, a mother with breast cancer increases the ovarian cancer risk to her daughter (OR = 2.3) while an affected sister yields a negligible odds ratio (OR = 1.1). Conversely in families with a history of ovarian cancer only, a mother’s ovarian cancer raised her daughters’ ovarian cancer risk (OR = 2.3, but p>0.05) while a sister’s ovarian cancer nearly quadruples her sister’s risk (OR = 3.92) [28]. Therefore, the autosomal dominant genes may be common to both breast and ovary cancers while the X-linked gene may be ovary-specific. We emphasize that future studies should be carefully designed to isolate X-linked versus autosomal and ovary-specific versus breast-ovary associations and to distinguish sporadic and hereditary ovarian cancers. Identifying a significant X-linked contribution to familial ovarian cancer risk has implications for clinical genetics: with suspicion of paternal lineage, an affected woman’s sisters are at significantly increased risk for ovarian cancer and ought to be counseled. If the affected woman carries the X-linked gene through her father, her sisters must also be carriers. It is reasonable to conjecture that maternal-lineage bias may have affected how patients, physicians, and researchers view family history and so the X-linked pattern may imply a familial origin for ovarian cancers previously thought to be sporadic cases. In particular, if the disease transmits through the father’s side, cases manifesting in only children or a woman with only brothers may not appear overtly hereditary. Using the rate of second generation grandmother pairs, we observed twice as many affected maternal grandmothers versus paternal. Without ascertainment bias, we would have expected a balanced rate, so we might predict that, other things equal, we have missed almost two paternal cases for every observed one.
We have provided some evidence that MAGEC3 is a potential candidate for the X-linked gene near previous linkage loci and it possesses a missense variant with large effect, rare prevalence and is associated with earlier onset. While MAGEC3 is thought to be a cancer testis antigen, it shows some expression in normal heart, brain, fallopian tube and pituitary gland tissues suggesting that the loss of expression of MAGEC3 plays some role in cancer formation. We have previously shown that co-expression patterns of the MAGE genes are non-random in ovarian tumors [29] and that other X-linked CT antigens (NY-ESO-1 encoded by CTAG1B) signaling highly aggressive tumors [30].
The only MHD-carrying yeast homologous gene (NSE3) binds with NSE1 and complexes with SMC5/SMC6 to repair double strand breaks via homologous recombination [31–33]. On the other hand, family members MAGEA [34, 35], MAGEC2 [36], carrying a paralogous MHD, have been shown to bind to RING domain proteins to form a p53 interacting E3 ubiquitin ligase that promotes tumorigenesis. Recently, computational modeling of sex-bias in cancers affecting both sexes has identified MAGEC3 directly as a putative X-linked tumor suppressor [37]. Reinforcing our observation that men in X-linked families may be at increased risk of prostate cancer, the cytoband housing MAGEC3 (Xq27.2) has been previously linked to these hereditary cancers [38], raising the possibility that there is a common hereditary X-linked locus responsible for reproductive tract-specific cancers.
While not unexpected for the X-chromosome [39], the candidate SNP appeared to rest in HWE while the HapMap populations reject HWE for this SNP. Given that the population frequencies are correlated with ovarian cancer incidence, we conjecture that the general population is under selection against the G allele (and against ovarian cancer) and it is the registry population that inherits neutrally and therefore more often manifests disease. This conjecture is consistent with the observations that CT antigens, especially those on the X chromosome, are under strong positive selection [40] and that the region containing MAGEC3 shows strong inter-population difference [41]. While the latter study localized the effect to MAGEC2 and not MAGEC3, their criterion for the frequency difference was aggressively high (delta > 0.90) which would preclude the SNPs that may be still be in the middle of a soft selective sweep. That the beneficial A allele in rs176026 has not yet fixed may explain why we have found a common variant, but it may also imply that the true cancer phenotype variant is hitchhiking along with the selective pressure behind the CT antigens.
Families in the Familial Ovarian Cancer Registry (formerly Gilda Radner Familial Ovarian Cancer Registry, Buffalo, NY) have been accessioned continuously from 1981 to present as described previously [8,42]. Briefly, qualifying families must have (a) two or more cases of ovarian cancer, (b) one ovarian cancer with two or more other cancers or (c) an early onset (age 45) ovarian cancer and at least one other cancer. Families provide written informed consent under Roswell Park Cancer Institute protocol CIC95-27. Cases are verified by medical record and/or death certificate when required and stage and histology are verified by a registry pathologist. The registry comprises 50,401 individuals including 5,614 ovarian cancers from 2,636 unique families. Families are also classified by disease pattern: families manifesting only ovarian cancer are termed “site-specific ovary” families and families with a number of breast cancers as well as ovarian cancers are “breast and ovary” families.
Considering women who (a) were at least 45 without disease at last contact or (b) had died without disease and those with confirmed ovarian cancer, we observed over 8,900 mother-daughter pairs and 27,000 sister-sister pairs. From large registry pedigrees, we ascertained 3,499 women with two grandmothers who possessed a recorded disease status. Of these granddaughters, 2,569 reported no affected grandmothers (73.4%), 892 had exactly one affected grandmother (25.5%) and 38 had two affected grandmothers (1.1%). These women came from large pedigrees where the average family under study has 27.3 individuals (range: 8–330). Of the 3,499 pairs, 619 belonged to high-risk families tested for BRCA mutations as previously described [8]. Families are classified as BRCA1 and/or BRCA2 positive if any one family member tests positive for a deleterious mutation. If every family member tests negative, the family is classified as BRCA negative.
Due to the age of the study cohort and availability of blood samples, we could not sequence all of the grandmother/granddaughter pairs. We focused on 159 women who reported a negative BRCA test and had an available DNA sample. DNA samples were whole-exome sequenced using Agilent SureSelect Human All Exome 50Mb kits v3 and v5. Raw sequence reads were aligned to the Human Reference Genome (NCBI Build 37) using the Burrows-Wheeler Aligner (BWA) [43], Picard [44] and GATK [45]. We retained variants within the X-chromosome exome with at least a 10% rate of non-reference genotypes, using the total, 2,161, to set the chromosome-wide significance threshold for the log-rank test of age-of-onset association at–log10(0.05/2161) = 4.636 on the log odds (LOD) scale. BRCA status was re-evaluated based on the sequencing results.
Ovarian cancer incidence data was downloaded from the IARC website [46]. HapMap allele frequencies were accessed via dbSNP and 1000 genomes populations via the Phase 3 1000 genomes browser. Correlations between the incidence and allele frequency were assessed by simple linear regression. The sequenced women all self-report Caucasian ancestry which we confirmed through principal components analysis.
Assume that we observe W = Wn + Wc women with ovarian cancer among a sistership of N = Nn + Nc women, where the subscripts n and c refer to non-carriers and variant carriers. The likelihood of W given the probability of disease in carriers (pc) and non-carriers (pn) can be constructed by assuming that, conditional on Nn and Nc, Wn and Wc are simply binomial random variables. Under an autosomal model, Nc is Binomial (N, 0.5). Under the X-linked model, P(Nc = N) = P(Nc = 0). Evaluating these likelihoods by enumerating admissible combinations of (Wn, Wc, Nn, Nc) is straightforward.
We downloaded GTEx v7 data (https://www.gtexportal.org/) aligning reported female cases only with known tissue sample types. We normalized the MAGEC3 RNAseq levels to MAGEA1 levels on a per sample basis. TCGA ovary data on the quantile normalized HuEx array were downloaded from the GDAC Firehose (https://gdac.broadinstitute.org/) and cBioPortal’s normalized RNAseq Z-scores.
We examined SNP-based linkage via LDLink, LDproxy [47] using the CEU population for reference. The functional predictions for missense SNPs were scored by PolyPhen [22] and the regulation was scored by HaploReg (v4) [48] using the EUR reference and its default position weight-matrix scoring algorithms [49].
Expected frequencies under the autosomal model were based on the pooled case frequency (157/892 = 17.6%). It can be shown that the X-linked likelihood was maximized by a granddaughter cancer rate of 14.0%. Goodness-of-fit was tested versus a chi-square distribution. Pedigree likelihoods were evaluated using the kinship2 [20] algorithms. Relative risk and odds ratio confidence intervals were computed under the log transform. Age-of-onset was defined as the shortest time to death, ovarian cancer diagnosis or prophylactic oophorectomy censored by age at last contact. The majority of granddaughter ages were observed (89.0%, 3080/3461). Risk of disease was estimated using the product-limit (Kaplan-Meier) estimate, tested with the log-rank test and hazard ratios estimated through Cox’s partial likelihood with graphical diagnostics for proportional hazards. All tests are two-sided and analyses were performed in R3.3.1 including the survival package.
|
10.1371/journal.pcbi.1003962 | Intrinsic Neuronal Properties Switch the Mode of Information Transmission in Networks | Diverse ion channels and their dynamics endow single neurons with complex biophysical properties. These properties determine the heterogeneity of cell types that make up the brain, as constituents of neural circuits tuned to perform highly specific computations. How do biophysical properties of single neurons impact network function? We study a set of biophysical properties that emerge in cortical neurons during the first week of development, eventually allowing these neurons to adaptively scale the gain of their response to the amplitude of the fluctuations they encounter. During the same time period, these same neurons participate in large-scale waves of spontaneously generated electrical activity. We investigate the potential role of experimentally observed changes in intrinsic neuronal properties in determining the ability of cortical networks to propagate waves of activity. We show that such changes can strongly affect the ability of multi-layered feedforward networks to represent and transmit information on multiple timescales. With properties modeled on those observed at early stages of development, neurons are relatively insensitive to rapid fluctuations and tend to fire synchronously in response to wave-like events of large amplitude. Following developmental changes in voltage-dependent conductances, these same neurons become efficient encoders of fast input fluctuations over few layers, but lose the ability to transmit slower, population-wide input variations across many layers. Depending on the neurons' intrinsic properties, noise plays different roles in modulating neuronal input-output curves, which can dramatically impact network transmission. The developmental change in intrinsic properties supports a transformation of a networks function from the propagation of network-wide information to one in which computations are scaled to local activity. This work underscores the significance of simple changes in conductance parameters in governing how neurons represent and propagate information, and suggests a role for background synaptic noise in switching the mode of information transmission.
| Differences in ion channel composition endow different neuronal types with distinct computational properties. Understanding how these biophysical differences affect network-level computation is an important frontier. We focus on a set of biophysical properties, experimentally observed in developing cortical neurons, that allow these neurons to efficiently encode their inputs despite time-varying changes in the statistical context. Large-scale propagating waves are autonomously generated by the developing brain even before the onset of sensory experience. Using multi-layered feedforward networks, we examine how changes in intrinsic properties can lead to changes in the network's ability to represent and transmit information on multiple timescales. We demonstrate that measured changes in the computational properties of immature single neurons enable the propagation of slow-varying wave-like inputs. In contrast, neurons with more mature properties are more sensitive to fast fluctuations, which modulate the slow-varying information. While slow events are transmitted with high fidelity in initial network layers, noise degrades transmission in downstream network layers. Our results show how short-term adaptation and modulation of the neurons' input-output firing curves by background synaptic noise determine the ability of neural networks to transmit information on multiple timescales.
| Gain scaling refers to the ability of neurons to scale the gain of their responses when stimulated with currents of different amplitudes. A common property of neural systems, gain scaling adjusts the system's response to the size of the input relative to the input's standard deviation [1]. This form of adaptation maximizes information transmission for different input distributions [1]–[3]. Though this property is typically observed with respect to the coding of external stimuli by neural circuits [1], [3]–[7], Mease et al. [8] have recently shown that single neurons during early development of mouse cortex automatically adjust the dynamic range of coding to the scale of input stimuli through a modulation of the slope of their effective input-output relationship. In contrast to previous work, perfect gain scaling in the input-output relation occurs for certain values of ionic conductances and does not require any explicit adaptive processes that adjust the gain through spike-driven negative feedback, such as slow sodium inactivation [4], [9], [10] and slow afterhyperpolarization (AHP) currents [10], [11]. However, these experiments found that gain scaling is not a static property during development. At birth, or P0 (postnatal day 0), cortical neurons show limited gain scaling; in contrast, at P8, neurons showed pronounced gain-scaling abilities [8]. Here, we examined how the emergence of the gain-scaling property in single cortical neurons during the first week of development might affect signal transmission over multiple timescales across the cortical network.
Along with the emergence of gain scaling during the first week of neural development, single neurons in the developing cortex participate in large-scale spontaneously generated activity which travels across different regions in the form of waves [12]–[14]. Pacemaker neurons located in the ventrolateral (piriform) cortex initiate spontaneous waves that continue to propagate dorsally across the neocortex [13]. Experimentally, much attention has been focused on synaptic interactions in initiating and propagating activity, with a particular emphasis on the role of GABAergic circuits, which are depolarizing in early development [15], [16]. While multiple network properties play an important role in the generation of spontaneous waves, here we ask how the intrinsic computational properties of cortical neurons, in particular gain scaling, can affect the generation and propagation of spontaneous activity. Changes in intrinsic properties may play a role in wave propagation during development, and the eventual disappearance of this activity as sensory circuits become mature.
A simple model for propagating activity, like that observed during spontaneous waves, is a feedforward network in which activity is carried from one population, or layer, of neurons to the next without affecting previous layers [17]. We compare the behavior of networks composed of conductance-based neurons with either immature (nongain-scaling) or mature (gain-scaling) computational properties [8]. These networks exhibit different information processing properties with respect to both fast and slow timescales of the input. We determine how rapid input fluctuations are encoded in the precise spike timing of the output by the use of linear-nonlinear models [18], [19], and use noise-modulated frequency-current relationships to predict the transmission of slow variations in the input [20], [21].
We find that networks built from neuron types with different gain-scaling ability propagate information in strikingly different ways. Networks of gain-scaling (GS) neurons convey a large amount of fast-varying information from neuron to neuron, and transmit slow-varying information at the population level, but only across a few layers in the network; over multiple layers the slow-varying information disappears. In contrast, nongain-scaling (NGS) neurons are worse at processing fast-varying information at the single neuron level; however, subsequent network layers transmit slow-varying signals faithfully, reproducing wave-like behavior. We qualitatively explain these results in terms of the differences in the noise-modulated frequency-current curves of the neuron types through a mean field approach: this approach allows us to characterize how the mean firing rate of a neuronal population in a given layer depends on the firing rate of the neuronal population in the previous layer through the mean synaptic currents exchanged between the two layers. Our results suggest that the experimentally observed changes in intrinsic properties may contribute to the transition from spontaneous wave propagation in developing cortex to sensitivity to local input fluctuations in more mature networks, priming cortical networks to become capable of processing functionally relevant stimuli.
Single cortical neurons acquire the ability to scale the gain of their responses in the first week of development, as shown in cortical slice experiments [8]. Here, we described gain scaling by characterizing a single neuron's response to white noise using linear/nonlinear (LN) models (see below). Before becoming efficient encoders of fast stimulus fluctuations, the neurons participate in network-wide activity events that propagate along stereotypical directions, known as spontaneous cortical waves [13], [22]. Although many parameters regulate these waves in the developing cortex, we sought to understand the effect of gain scaling in single neurons on the ability of cortical networks to propagate information about inputs over long timescales, as occur during waves, and over short timescales, as occur when waves disappear and single neurons become efficient gain scalers. More broadly, we use waves in developing cortex as an example of a broader issue: how do changes in intrinsic properties of biophysically realistic model neurons affect how a network of such neurons processes and transmits information?
We have shown that in cortical neurons in brain slices, developmental increases in the maximal sodium () to potassium () conductance ratio can explain the parallel transition from nongain-scaling to gain scaling behavior [8]. Furthermore, the gain scaling ability can be controlled by pharmacological manipulation of the maximal to ratio [8]. The gain scaling property can also be captured by changing this ratio in single conductance-based model neurons [8]. Therefore, we first examined networks consisting of two types of neurons: where the ratio of to was set to either 0.6 (representing immature, nongain-scaling neurons) or 1.5 (representing mature, gain-scaling neurons).
We first characterized neuronal responses of conductance-based model neurons using methods previously applied to experimentally recorded neurons driven with white noise. The neuron's gain scaling ability is defined by a rescaling of the input/output function of a linear/nonlinear (LN) model by the stimulus standard deviation [8]. Using a white noise input current, we extracted LN models describing the response properties of the two neuron types to rapid fluctuations, while fixing the mean (DC) of the input current. The LN model [18], [19], [23] predicts the instantaneous time-varying firing rate of a single neuron by first identifying a relevant feature of the input, and after linearly filtering the input stimulus with this feature, a nonlinear input-output curve that relates the magnitude of that feature in the input (the filtered stimulus) to the probability of firing. We computed the spike-triggered average (STA) as the relevant feature of the input [18], [24], and then constructed the nonlinear response function as the probability of firing given the stimulus linearly filtered by the STA.
Repeating this procedure for noise stimuli with a range of standard deviations () produces a family of curves for both neuron types (Figure 1A). While the linear feature is relatively constant as a function of the magnitude of the rapid fluctuations, , the nonlinear input-output curves change, similar to experimental observations in single neurons in cortical slices [8]. When the input is normalized by , the mature neurons have a common input-output curve with respect to the normalized stimulus (Figure 1B, red) [8] over a wide range of input DC. In contrast, the input-output curves of immature neurons have a different slope when compared in units of the normalized stimulus (Figure 1B, blue). Gain scaling has previously been shown to support a high rate of information transmission about stimulus fluctuations in the face of changing stimulus amplitude [1]. Indeed, these GS neurons have higher output entropy, and therefore transmit more information, than NGS neurons (Figure 1E). The output entropy is approximately constant regardless of for a range of mean (DC) inputs – this is a hallmark of their gain-scaling ability. The changing shape of the input-output curve for the NGS neurons results in an increasing output entropy as a function of (Figure 1E). With the addition of DC, the output entropy of the NGS neurons' firing eventually approaches that of the GS neurons; this is accompanied with a simultaneous decrease in the distance between rest and threshold membrane potential of the NGS neurons as shown previously [8]. Thus, GS neurons are better at encoding fast fluctuations, a property which might enable efficient local computation independent of the background signal amplitude in more mature circuits after waves disappear.
The response of a neuron to slow input variations may be described in terms of its firing rate as a function of the mean input through a frequency-current (–) curve. This description averages over the details of the rapid fluctuations. The shape of this – curve can be modulated by the standard deviation () of the background noise [20], [21]. Here, the "background noise'' is a rapidly-varying input that is not considered to convey specific stimulus information but rather, provides a statistical context that modulates the signaled information assumed to be contained in the slow-varying mean input. Thus, a neuron's slow-varying responses can be characterized in terms of a family of – curves parameterized by .
Comparing the – curves for the two neuron types using the same conductance-based models reveals substantial differences in their firing thresholds and also in their modulability by (Figure 1C,D). NGS neurons have a relatively high threshold at low , and the – curves are significantly modulated by the addition of noise, i.e. with increasing (Figure 1C). In contrast, the – curves of GS neurons have lower thresholds, and show minimal modulation with the level of noise (Figure 1D). This behavior is reflected in the information that each neuron type transmits about firing rate for a range of (Figure 1F). This information quantification determines how well a distribution of input DC can be distinguished at the level of the neuron's output firing rate while averaging out the fast fluctuations. The information would be low for neurons whose output firing rates are indistinguishable for a range of DC inputs, and high for neurons whose output firing rates unambiguously differ for different DC inputs. The two neuron types convey similar information for large where the – curves are almost invariant to noise magnitude. For GS neurons, most information is conveyed about the input rate at low where the – curve encodes the largest range of firing rates (0 to 30 Hz). The information encoded by NGS neurons is non-monotonic: at low these neurons transmit less information because of their high thresholds, compressing the range of inputs being encoded. Information transmission is maximized at for which the – curve approaches linearity, simultaneously maximizing the range of inputs and outputs encoded by the neuron. For both neuron types, the general trend of decreasing information as increases is the result of compressing the range of outputs (10 to 30 Hz).
These two descriptions characterize the different processing abilities of the two neuron types. GS neurons with their -invariant input-output relations of the LN model are better suited to efficiently encode fast current fluctuations because information transmission is independent of . However, NGS neurons with their -modulatable – curves are better at representing a range of mean inputs, as illustrated by their ability to preserve the range of input currents in the range of output firing rates.
To characterize the spectrum of intrinsic properties that might arise as a result of different maximal conductances, and , we determined the – curves for a range of maximal conductances in the conductance-based model neurons (Figure 2). Mease et al. [8] previously classified neurons as spontaneously active, excitable or silent, and based on the neurons' LN models determined gain-scaling ability as a function of the individual and for excitable neurons. Models with low had nonlinear input-output relations that did not scale completely with , while models with high had almost identical nonlinear input-output relations for all [8]. Therefore, gain scaling ability increased with increasing ratio, independent of each individual conductance.
We examined the modulability of – curves by in excitable model neurons while independently varying and (Figure 2). Like gain scaling, the modulability by also depended only on the ratio , rather than either conductance alone, with larger modulability observed for smaller ratios. To further explore the implications of such modulability by , we computed the mutual information that each model neuron transmits about mean inputs for a range of (Figure 2). Neurons with behaved like GS neurons in Figure 1F, while neurons with behaved like NGS neurons.
These results suggest that the ability of single neurons to represent a distribution of mean input currents by their distribution of output firing rates can be captured only by changing the ratio of and . Therefore, we focused on studying two neuron types with in the two extremes of the conductance range of excitable neurons: GS neurons with and NGS neurons with .
Upon characterizing single neuron responses of the two neuron types to fast-varying information via the LN models and to slow-varying information via the – curves, we compared their population responses to stimuli with fast and slow timescales. A population of uncoupled neurons of each type was stimulated with a common slow ramp of input current, and superimposed fast-varying noise inputs, generated independently for each neuron (Figure 3A). The population of NGS neurons fired synchronously with respect to the ramp input and only during the peak of the ramp (Figure 3B), while the GS neurons were more sensitive to the background noise and fired asynchronously during the ramp (Figure 3C) with a firing rate that was continuously modulated by the ramp input. This suggests that the sensitivity to noise fluctuations of the GS neurons at the single neuron level allows them to better encode slower variations in the common signal at the population level [25]–[27], in contrast to the NGS population which only responds to events of large amplitude independent of the background noise.
During cortical development, wave-like activity on longer timescales occurs in the midst of fast-varying random synaptic fluctuations [13], [14], [28], [29]. Therefore, we compared the population responses of GS and NGS neurons to a slow-varying input (500 ms correlation time constant) common to all neurons with fast-varying background noise input (1 ms correlation time constant) independent for all neurons (Figure 3D). The distinction between the two neuron types is evident in the mean population responses (peristimulus time histogram, i.e. PSTH). The NGS population only captured the stimulus peaks (Figure 3E) while the GS population faithfully captured the temporal fluctuations of the common signal, aided by each neuron's temporal jitter caused by the independent noise fluctuations (Figure 3F). Although not an exact model of cortical wave development, this comparison supports the hypothesis that the intrinsic properties of single neurons can lead to different information transmission capabilities of cortical networks at different developmental time points, and the transition from wave propagation to wave cessation.
The observed difference between the population responses of the GS and NGS neurons to the slow-varying stimulus in the presence of fast background fluctuations (Figure 3D–F) suggested that the two neuron types differ in their ability to transmit information at slow timescales. Therefore, we next examined how the identified single neuron properties affect information transmission across multiple layers in feedforward networks. Networks consisted of 10 layers of 2000 identical neurons of the two different types (Figure 4A). The neurons in the first layer receive a common temporally fluctuating stimulus with a long correlation time constant (1 s, see Methods); neurons in deeper layers receive synaptic input from neurons in the previous layer via conductance-based synapses. Each neuron in the network also receives a rapidly varying independent noise input (with a correlation time constant of 1 ms) to simulate fast-varying synaptic fluctuations. The noise input here is a rapidly-varying input that sets the statistical context for the slow-varying information; it does not transmit specific stimulus information itself. The GS and NGS networks have strikingly different spiking dynamics (Figure 4B). The GS network responds with higher mean firing rates in each layer, as would be expected from the – curves characterizing intrinsic neuronal properties (Figure 1C,D). While the GS neurons have a baseline firing rate even at zero input current, the NGS neurons only fire for large input currents, with a threshold dependent on the level of intrinsic noise; thus, the two neuron types have different firing rates. To evaluate how the networks transmit fluctuations of the slow-varying common input signal, independent of the overall firing rates, we evaluated the averaged population (PSTH) response of each layer, normalized to have a mean equal to 0 and a variance equal to 1 (Figure 4C).
The first few layers of the GS network robustly propagate the slow-varying signal as a result of the temporally jittered response produced by the sensitivity to fast fluctuations at the single neuron level, consistent with the population response in Figure 3F. However, due to the effects of these same noise fluctuations, this population response degrades in deeper layers (Figure 4C, left, see also Figure S1 for ). In contrast, the NGS network is insensitive to the fast fluctuations and thresholds the slow-varying input at the first layer, as in Figure 3E. Despite the presence of fast-varying background noise, the NGS network robustly transmits the large peaks of this stimulus to deeper layers without distortion (Figure 4C, right).
This difference in the transmission of information through the two network types is captured in the information between the population response and the slow-varying stimulus in Figure 4D. The GS network initially carries more information about the slow-varying stimulus than the NGS network; however, this information degrades in deeper layers when virtually all the input structure is lost, and drops below the NGS network beyond layer four (Figure 4D, bottom). While the information carried by the NGS network is initially lower than the GS network (due to signal thresholding), this information is preserved across layers and eventually exceeds the GS information.
The observed differences in the propagation of slow-varying inputs between the two network types resemble changes in wave propagation during development. While spontaneous waves cross cortex in stereotyped activity events that simultaneously activate large populations of neurons at birth, these waves disappear after the first postnatal week [13], [16]. We have demonstrated that immature neurons lacking the gain-scaling ability can indeed propagate slow-varying wave-like input of large amplitude as population activity across many layers. As these same neurons acquire the ability to locally scale the gain of their inputs and efficiently encode fast fluctuations, they lose the ability to propagate large amplitude events at the population level, consistent with the disappearance of waves in the second postnatal week [13]. While many parameters regulate the propagation of waves [14], [29], our network models demonstrate that varying the intrinsic properties of single neurons can capture substantial differences in the ability of networks to propagate slow-varying information. Thus, changes in single neuron properties can contribute to both spontaneous wave generation and propagation early in development and the waves' disappearance later in development.
The layer-by-layer propagation of a slow-varying signal through the population responses of the two networks can be qualitatively predicted using a mean field approach that bridges descriptions of single neuron and network properties. Since network dynamics varies on faster timescales than the correlation timescale of the slow-varying signal, the propagation of a slow-varying signal can be studied by considering how a range of mean inputs propagate through each network. The intrinsic response of the neuron to a mean (DC) current input is quantified by the – curve which averages over the details of the fast background fluctuations; yet, the magnitude of background noise, , can change the shape and gain of this curve [20], [21]. Thus, for a given neuron type, there is a different – curve depending on the level of noise , (Figure 1C,D). One can approximate the mean current input to a neuron in a given layer , , from the firing rate in the previous layer through a linear input-output relationship, with a slope dependent on network properties (connection probability and synaptic strength, see Eq. 15). Given the estimated mean input current for a given neuron in layer , , the resulting firing rate of layer , , can then be computed by evaluating the appropriate – curve, , which characterizes the neuron's intrinsic computation(1)
Thus, these two curves serve as an iterated map whereby an estimate of the firing rate in the Lth layer, , is converted into a mean input current to the next layer, , which can be further converted into , propagating mean activity across multiple layers in the network (Figures 5, 6). While for neurons in the first layer, the selected – curve is the one corresponding to the level of intrinsic noise injected into the first layer, , for neurons in deeper layers, the choice of – curve depends not only on the magnitude of the independent noise fluctuations injected into each neuron, but also on the fluctuations arising from the input from the previous layer (see Eq. 16 in Methods). The behavior of this iterated map is shaped by its fixed points, the points of intersection of the – curve with the input-output line , which organize the way in which signals are propagated from layer to layer. The number, location and stability of these fixed points depend on the curvature of and on (Figure 5). When the slope of at the fixed point is less than , the fixed point is stable. This implies that the entire range of initial DC inputs (into layer 1) will tend to iterate toward the value at the fixed point as the mean current is propagated through downstream layers in the network (Figure 5, left). Therefore, all downstream layers will converge to the same population firing rate that corresponds to the fixed point. In the interesting case that becomes tangent to the linear input-output relation, i.e. the – curve has a slope equal to , the map exhibits a line attractor: there appears an entire line of stable fixed points (Figure 5, middle). This ensures the robust propagation of many input currents and population rates across the network. Interestingly, the – curves of the GS and NGS neurons for different values of fall into one of the regimes illustrated in Figure 5: GS neurons with their -invariant – curves have a single stable fixed point (Figure 5, left), while the NGS neurons have line attractors with exact details depending on (Figure 5, middle and right). The mechanics of generating a line attractor have been most extensively explored in the context of oculomotor control (where persistent activity has been interpreted as a short-term memory of eye position that keeps the eyes still between saccades) and decision making in primates (where persistent neural activity has been interpreted as the basis of working memory) [30].
Indeed, Figure 6A,B shows that the – curves for GS neurons at two values of , one low and one high, are very similar. The mean field analysis predicts that all initial DC inputs applied to layer 1 will converge to the same stable fixed point during propagation to downstream layers. Numerical simulations corroborate these predictions (Figure 6A,B, bottom). A combination of single neuron and network properties determine the steady state firing rate through (Eq. 15). Activity in the GS networks can propagate from one layer onto the next with relatively weak synaptic strength even when the networks are sparsely connected (5% connection probability), as a result of the low thresholds of these neurons (Figure 1D). The specific synaptic strength in Figure 6A,B was chosen arbitrarily so that the – curve intersects the input-output line with slope , but choosing different synaptic strength produces qualitatively similar network behavior (Figure S2). The parameter can be modulated by changing either the connectivity probability or the synaptic strength in the network; as long as their product is preserved, remains constant and the resulting network dynamics does not change (Figure S2). Furthermore, as a result of the lack of modulability of GS – curves by (Figure 1D), the network dynamics remains largely invariant to the amplitude of background noise.
In contrast, the amplitude of background noise fluctuations, , has a much larger impact on the shape of NGS – curves (Figure 1C) and on the resulting network dynamics (Figure 5). When the combination of sparse connection probability and weak synaptic strength leads to the slope being too steep (weak connectivity in GS networks, Figure 6A,B), there may be no point of intersection with the NGS – curves: all DC inputs are mapped below threshold and activity does not propagate to downstream layers. Keeping the same sparse connection probability of and increasing synaptic strength enables the propagation of neuronal activity initiated in the first layer to subsequent layers in NGS networks. For a particular value of , there is an entire line of stable fixed points in the network dynamics (Figure 5, middle), so that a large range of input currents are robustly transmitted through the network. More commonly, however, the map has three fixed points: stable fixed points at a high value and at zero, and an intermediate unstable fixed point (Figure 6C,D). In this case, mean field theory predicts that DC inputs above the unstable fixed point should flow toward the high value, while inputs below it should iterate toward zero, causing the network to stop firing. However, the map still behaves as though the – curve and the input-output transformation are effectively tangent to one another over a wide range of input rates (green box in Figure 6C,D), creating an effective line of fixed points for which a large range of DC inputs is stably propagated through the network; this is generically true for a wide range of noise values, although the exact region of stable propagation depends on the value of (Figure 5, middle and right, Figure S3). The best input signal transmission is observed when the network noise selects the most linear – curve that simultaneously maximizes the range of DC inputs and population firing rates of the neurons (Figure 5, middle). This is approximately the noise value selected in Figure 6C,D. We call this a stable region of propagation for the network since a large range of mean DC inputs can be propagated across the network layers so that the population firing rates at each layer remain distinct. Our results resemble those of van Rossum et al. [31] where regimes of stable signal propagation were observed in networks of integrate-and-fire neurons by varying the DC input and an additional background noise. The best regime for stable signal propagation occurred for additive noise that was large enough to ensure that the population of neurons independently estimated the stimulus, as in our NGS networks (Figure 5, middle and right, Figure S3).
The emergence of extended regions of stable rate propagation implies that the NGS mean field predictions (Figure 6C,D, bottom) are less accurate than for the GS networks where the convergence to the stable fixed points is exact (Figure 6A,B). However, the NGS mean field predictions show qualitative agreement with the simulation results, in particular in the initial network layers where the approach to the nonzero stable fixed point is much slower than in the GS networks, i.e. occurs over a larger number of layers. Along with the slow convergence of firing rates toward a single population firing rate, the ability of network noise to modulate the NGS – curves suggests that multiple – curves can be used to predict network dynamics by combining added and intrinsically generated noise (see Eq. 16). As a result, for some input currents (e.g. arrow in Figure 6C) the firing rate goes down in the first three layers where network dynamics predicts convergence to the zero stable fixed point. The initial decrease of firing rate is due to the disappearance of weak synaptic inputs that cannot trigger the cells to spike. Network noise then selects a different – curve that shifts the dynamics into the rate stabilization region (Figure 6C, green box) where firing rates are stably propagated. The onset of synchronous firing of the neuronal population in each layer also contributes to rate stabilization. Population firing rates in deeper layers increase to a saturating value lower than the mean field predicted value. Similar results have been observed experimentally [32] and in networks of Hodgkin-Huxley neurons [33]. We find similar network dynamics for a more weakly connected NGS network using the smallest possible synaptic strength that allows activity to propagate through the network (Figure S2). As for the GS networks, as long as the product of connection probability and synaptic strength is constant, the slope of the input-output linear relationship , and the network dynamics remain unchanged, even if these network parameters change individually (Figure S2).
An exception to this result is observed at very sparse connectivity (2%), where network behavior is more similar to the GS networks (Figure S2, bottom right). At this sparse connectivity, independent noise reduces the common input across different neurons and synchrony is less pronounced. This argues that the emergence of synchrony plays a fundamental role in achieving reliable propagation of a range of DC inputs (and correspondingly population firing rates) in the NGS networks. Although experimental measurements of the connectivity probability in developing cortical networks are lacking, calcium imaging of single neurons demonstrates that activity across many neurons during wave propagation is synchronous [34]. Intracellular recordings of adult cultured cortical networks also demonstrate that synchronous neuronal firing activity is transmitted in multiple layers [32].
To examine network behavior for comparable connectivity strength, we repeated the network simulations and mean field predictions of mean DC input propagation in GS networks with the same increased synaptic strength needed for propagation of activity in the NGS networks. We found that the behavior was similar to the weakly connected GS network: Regardless of the initial input current, the network output converged to a single output firing rate by layer 5 (Figure 6E,F), making these networks incapable of robustly propagating slow-varying signals without distortion. As for the strongly connected NGS networks, neurons across the different layers in these strongly connected GS networks developed synchronous firing. This synchrony led to a small difference (several Hz) between the final firing rate approached by each network compared with the firing rate predicted from the mean field analysis. Although both the strongly connected GS and NGS networks developed synchronous firing, the behavior of the two types of networks remained different (Figure 6).
The results in this section indicate that firing rate transmission depends on the details of single neuron properties, including their sensitivity to fast fluctuations as characterized by the LN models (Figure 1A,B). Firing rate transmission also depends on the modulability of the – curves by the noise amplitude (Figure 1C,D). Because of these differences in intrinsic computation, the GS and NGS networks show distinct patterns of information transmission (Figure 5): firing rate convergence to a unique fixed point, or a line of fixed points ensuring stable propagation of firing rates which can be reliably distinguished at the output, respectively. In the latter case, even when a line of fixed point is not precisely realized as in Figure 5 (middle), competition between the slow convergence of firing rates to the mean field fixed point and the emergence of synchrony enable the propagation of firing rates through the different network layers, aided by the range of – curves sampled by network noise with amplitude .
Given the predicted signal propagation dynamics, we now directly compute the mutual information between the mean DC input injected into layer 1 and the population firing rates at a given layer for each magnitude of the independent noise (Figure 7). This measures how distinguishable network firing rate outputs at each layer are for different initial mean inputs. The convergence of population firing rates across layers to a single value in the GS networks leads to a drop in information towards zero for both the weakly (Figure 6A,B) and strongly connected GS networks (Figure 6E,F) as a function of layer number and for a wide range of network noise (Figure 7A,C). NGS networks can transmit a range of mean DC inputs without distortion (Figure 6C,D); thus, the information between input DC and population firing rate remains relatively constant in subsequent layers (Figure 7B). The information slightly increases in deeper layers due to the emergence of synchronization, which locks the network output into a specific distribution of population firing rates. As noise amplitude increases, the selected – curve becomes tangent to the linear input-output relationship over a larger range of input firing rates (Figure 6C,D); hence, a larger range of inputs is stably transmitted across network layers. Counterintuitively, this suggests that increasing noise in the NGS networks can serve to increase the information such networks carry about a distribution of mean inputs.
The differential ability of GS and NGS networks to reliably propagate mean input signals is predicted by the modulability of the – curves by the network noise . To understand the dynamical origins of this difference, we analytically reduced the neuron model (Eq. 2) to a system of two first order differential equations describing the dynamics of the membrane potential and an auxiliary slower-varying potential variable (Methods) [35]. We analyzed the dynamics in the phase plane by plotting vs. . The nullclines, curves along which the change in either or is 0, organize the flows of and (Figure 8); these lines intersect at the fixed points of the neuron's dynamics. We studied the fixed points at different ratios of and , with a particular focus on the values discussed above ( and ). These exhibit substantial differences in the type and stability of the fixed points, as well as the emergent bifurcations where the fixed points change stability as one varies the mean DC input current into the neuron (Figure 8).
For a large range of DC inputs, the NGS neuron () has a single stable fixed point (either a node or a focus) (Figure 8A). In this case, the only perturbation that can trigger the system to fire an action potential is a large-amplitude noise current fluctuation. The of the current then determines the number of action potentials that will be fired in a given trial and strongly modulates the firing rate of the neuron. We show two trajectories at pA and 50 pA and at two different DC values of 0 and 30 pA (Figure 8A), at which the – curves are strongly noise-modulated (Figure 1C). As the DC increases beyond 62 pA, the fixed point becomes unstable and a stable limit cycle emerges (not shown). In this case, any will move the trajectories into the stable limit cycle and the neuron will continuously generate action potentials, with a firing rate independent of . Indeed, Figure 1C shows that the – curves become less effectively modulated by for DC values greater than 62 pA.
As the conductance ratio increases, the range of DC values for which the system has a single fixed point decreases (Figure 8B). Indeed, the GS neuron () has a stable limit cycle for the majority of DC values (Figure 8C). This implies that GS neurons are reliably driven to fire action potentials for any and their firing rate is not very sensitive to . For low DC values, the stable limit cycle coexists with a stable fixed point, so in this case of the noise can modulate the firing rate more effectively, as is seen in Figure 1D.
This analysis highlights the origins for the differential modulability of firing rate in NGS and GS neurons. Although the model reduction sacrifices some of the accuracy of the original model, it retains the essential features of action potential generation: the sudden rise of the action potential which turns on a positive inward sodium current, and its termination by a slower decrease in membrane potential which shuts off the sodium current and initiates a positive outward potassium current hyperpolarizing the cell. Although simpler neuron models (e.g. binary and integrate-and-fire [36]–[38]) allow simple changes in firing thresholds, the dynamical features inherent in the conductance-based neurons studied here are needed to capture noise-dependent modulation.
The adult brain exhibits a diversity of cell types with a range of biophysical properties. Organized into intricate circuits, these cell types contribute to network computation, but the role of intrinsic properties is unclear. Recently, we have shown that during early development, single cortical neurons acquire the ability to represent fast-fluctuating inputs despite variability in input amplitudes by scaling the gain of their responses relative to the scale of the inputs they encounter [8]. Before these intrinsic properties shift, the developing cortex generates and propagates spontaneous waves of large-scale activity [13], [22], [39], [40], which regulate developmental changes in ion channel expression, synaptic growth and synaptic refinement processes [29], [41], [42]. How do experimentally observed biophysical properties affect ongoing network dynamics at this time? Using model neurons with conductance properties chosen to reproduce this developmental change in gain scaling, we investigated the implications of this change on the ability of feedforward networks to robustly transmit slow-varying wave-like signals. The conductance-based models that we considered are not intended as an exact biophysical model for developing cortical neurons; rather they allow us to study the more fundamental question of the role of single neuron computation on network behavior in a case with a well-defined and physiologically relevant network level property.
We add to previous studies by considering first, the fidelity of propagation of temporally varying patterns by biophysically realistic neurons, basing our work in a biological context where the brain naturally enters a state of wave propagation. Second, our work highlights a role of cellular processes in large-scale network behavior that has rarely been studied. Our results implicate intrinsic conductance change as a way to switch between global synchronization and local responsiveness, rather than synaptic plasticity, which is typically used to evoke such a global network change [17]. Related changes in excitability that accompany the cessation of spontaneous activity have been observed in the mouse embryonic hindbrain, where they have been ascribed to hyperpolarization of resting membrane potential and increased resting conductance of channels [43]. Finally, we analyze network information transmission on two different timescales (local fluctuations and network-wide wave-like events) and thereby generalize previous classification of feedforward network propagation into either synchrony-based coding [32], [44], and rate-based coding [31], [45].
We use two different descriptions of neuronal properties to characterize the neuron's ability to propagate information at these different time- and lengthscales. The processing of fast input fluctuations can be characterized using LN models [8], [46]–[48]. While single neuron properties affect the linear feature [46], [48], [49], here we focus on the scaling of the nonlinearity in the LN model to stimuli of different amplitudes. Information about slowly modulated input is described using noise-modulated – curves [20], [21], [50]. This ability of developing neurons to transmit distinct information at two different timescales is an example of a temporally multiplexed code [3], [51]–[53]. Here, GS neurons perform temporal multiplexing as they simultaneously convey distinct information about fast and slow fluctuations, reliably encoding slowly varying stimuli, albeit only for a few network layers. The NGS neurons also implement a multiplexed code because of their dual role to transmit firing rates while maintaining synchrony.
The above characterizations predict the success of global information propagation across multiple network layers [49], [50]. In integrate-and-fire network models with a fixed – curve, different network dynamics has been achieved by varying connectivity probability and synaptic strength [31], [45], [54], [55]. Here, in addition we considered the modulation of the – curves by the combined effects of injected independent noise and measured correlated noise from network interactions, permitting a description of network responses dependent on the input statistics, intrinsic single neuron properties and network connectivity (Figure 6). The role of -modulated – curves has also been fundamental in understanding how intrinsic neuron properties affect correlation transfer and encoding of rate- and synchrony-based signals in reduced networks of two neurons stimulated with a common input signal and independent noise [48], [49], [52], [53], [56]. We expect that generalizations of these methods will enable improved theoretical predictions for firing rate and correlation transfer beyond mean field, by computing the effects of temporal correlations such as we observe.
Firing rate transmission in our NGS networks co-occurs with the development of precise spike-time synchronization over a wide range of stimulus statistics and network connectivity (Figure 6). This synchronization might be a feature of biologically inspired networks because similar patterns were reported in experimentally simulated feedforward networks in vitro [32] and Hodgkin-Huxley-based simulations [33], but not in networks of threshold binary neurons [36], [57], nor integrate-and-fire neurons [55]. Several manipulations to single neuron or network properties might reduce this synchrony. These include: introducing sparse connectivity with strong synapses [17], [37], increasing independent noise input [31], [36], or embedding the feedforward into recurrent networks with inhibition to generate asynchronous background activity [37], [38], [55], [58]; but these typically result in signal degradation or implausible assumptions in our models. We did not find a regime supporting reliable asynchronous rate propagation, consistent with other studies [32], [33], [36], [44].
We identified the biophysical basis of the single-unit properties that underlies our results. The change in gain scaling is accompanied by a difference in the distance from rest to threshold membrane potential [8]: GS neurons have a smaller distance to threshold and are more likely to fire driven by noise fluctuations, while NGS neurons have a larger distance to threshold and must integrate many coincident inputs to fire. Indeed, a change in spiking threshold in simpler model neurons has been shown to modulate the mode of signal transmission in a feedforward network [36], [59], [60]. However, our mean-field and phase-plane dynamical analyses together show that threshold is not the only factor at work: the nature of rate propagation is intimately connected with the bifurcation properties of the neuron model. While we focused on two representative contrasting cases, these properties vary systematically with the conductance ratio of the neuron and we have mapped out the spectrum of possible behaviors of this model.
The robustness of information propagation across network layers is likely to have important implications for how developmental information contained in wave propagation patterns is transmitted across the cortex. We have previously shown that cortical waves are initiated in a pacemaker circuit contained within the piriform cortex [12]–[14], which is likely to provide the strong input necessary to drive NGS neurons. The waves propagate dorsally across the neocortex so that throughout the developmental period of wave generation, the neocortex acts as a follower region in the sequence of wave propagation. The reliability with which firing patterns of piriform neurons are retained as waves propagate into the neocortex will determine the nature of developmental information that the neocortex receives from those waves during its development. As gain scaling develops, more mature neurons can support efficient coding of local fluctuations and discard information about network-wide events. Therefore, the alteration of a single developmentally regulated conductance parameter can shift cortical neurons from synchrony-based encoders of slow inputs to noise-sensitive units that respond with high fidelity to local fluctuations independent of the overall scale. The growing sensitivity to noise of cortical neurons in the first postnatal week might help to prevent large-scale wave activity from dominating adult neural circuits, thus discouraging epileptiform patterns of network activity. At the same time, the emergence of gain scaling supports a transition to a state in which cortical circuits, rather than participating in network-wide events, can respond optimally to appropriately scaled local information, breaking up the cortical sheet into smaller information-processing units.
The mature cortex is also capable of generating spontaneous activity that propagates over large distances in the absence of sensory stimulation [61]–[63]. Such wave activity is postulated to be involved in short-term memory and the consolidation of recent transient sensory experience into long-lasting cortical modifications. For example, recent in vivo experiments proposed that synaptic plasticity is enforced by slow waves that occur during sleep [64, 65]. Spontaneous propagation activity patterns emerge from the interplay of intrinsic cellular conductances and local circuit properties [63]; our results raise the possibility that modulation of intrinsic properties through slow Na+ inactivation or neuromodulation could have multiple short-term effects on cortical information processing.
While we have examined the effect of gain scaling as a specific form of adaptation emerging during development, other adaptation mechanisms also likely play an important role in information transmission in feedforward networks. For instance, spike frequency adaptation has been shown to have effects that accumulate across multiple layers of feed-forward networks [31]. This widely observed form of adaptation can arise from calcium-dependent potassium conductances which generate AHPs [21], [66], [67]. Indeed, we and others have found that AHP-generating conductances can also support gain scaling behavior by single neurons [9], [68]. Independent of AHP conductances, slow sodium channel inactivation can also contribute to spike frequency adaptation [69], [70]. Incorporating such slow-timescale channel dynamics will require taking into account temporal aspects of the coding of mean (or variance) [71] that are presently ignored in our mean-field analysis based on modulated – curves. These slow dynamics may contribute to successive layers of filtering that affect information transmission [10]. An analytical characterization of the impact of slow neuronal dynamics on networks is likely to require novel theoretical approaches beyond those used here.
Similarly, other factors beyond the specific changing intrinsic neuronal properties addressed here contribute to the generation of spontaneous cortical waves with complex spatio-temporal properties. During the same developmental time period, the cortex undergoes substantial changes in information processing capacity that are beyond the scope of the present study [72]–[74]. Activity-dependent modification of synaptic connections driven by developmental cues contained in spontaneous wave patterns are likely to refine cortical networks into their mature state [14], [16], [39], [42], [73]. Furthermore, the emergence of synaptic inhibition as GABA becomes more hyperpolarizing contributes to diminishing the wave-like activity generated by the immature excitatory network [14], [73]. Thus, synaptic plasticity and intrinsic neuronal properties interact to modulate the emergence, propagation and the eventual disappearance of spontaneous waves in the developing cortex, and also to endow spatially-distinct regions at different time points with different information processing capabilities.
We studied a modified version of a Hodgkin-Huxley style model adapted by Mainen et al. [75] for spike initiation in neocortical pyramidal neurons. The model consists of a leak current, mammalian voltage-gated transient sodium and delayed-rectified potassium currents with maximal conductances , and , and reversal potentials mV, mV and mV:(2)
where µF/cm is the specific membrane capacitance and is the input current with denoting the area of the membrane patch with radius of 30 µm. The leak conductance was set to pS/µm2 such that the membrane time constant at the resting potential was 40 ms (any values between 25 and 50 ms were consistent with experimental data) [8]. The active conductances can be expressed via the gating variables , and such that and . We used pS/µm2 and pS/µm2 for the maximal conductances of the GS neurons, so that their ratio was ; and pS/µm2 and pS/µm2 for the maximal conductances of the NGS neurons, so that their ratio was . We also studied a larger range of these maximal conductances in Figure 2. The gating variables have the following kinetics: with where can be , or , and:(3)(4)(5)
The rate coefficients, and are of the form and and the kinematic parameters are provided in Table 1.
The equations were numerically solved using a first-order Euler method with an integration time step of ms. We used a threshold of −20 mV to detect spikes, although our results did not depend on the exact value of this parameter.
For spike-triggered characterization we injected Gaussian noise current, , with mean, , and standard deviation, , to elicit spike trains in ten 1000-second long trials. All input current traces were realizations of the Ornstein-Uhlenbeck process [76] expressed as:(6)
where has unit variance and correlation time of 1 ms to match experimental conditions [8].
Intrinsic computation in these neuron types was previously characterized in experiments and model neurons [8] using a one-dimensional Linear-Nonlinear (LN) cascade model of output spike times to the input Gaussian current stimulus with standard deviation [23]. The first component of the LN model is a feature which linearly filters the stimulus producing the amplitude of the feature present in the input; the second component is a nonlinear function which gives the instantaneous firing rate for each value of the filtered stimulus. We take the feature to be the spike-triggered average (STA) [18], [24], and obtain the expression for the nonlinear response function from Bayes' law:(7)
where is the mean firing rates for fixed input mean and standard deviation , is the prior distribution which is a Gaussian with mean zero and variance , is the spike-triggered stimulus distribution obtained from the histogram of filtered stimulus values when the spikes occur.
We refer to the neurons with ratio equal to 1.5 as gain-scaling, because scaling the stimulus by produces a nonlinearity in the LN model that is independent of , i.e. for inputs with two different standard deviations and (mean fixed to zero in Figure 1A,B, red) [8]. The neurons with ratio equal to 0.6 are termed nongain-scaling, because nonlinearities in the LN model vary with different values of the standard deviation when the stimulus is scaled by (Figure 1A,B, blue). The gain-scaling properties of single neurons hold for all [8].
We considered a feedforward network architecture with layers, each layer consisting of neurons (Figure 4A). We considered networks of neurons (the results remain the same as long as ). A common temporally fluctuating input current was injected to all neurons in the first layer. The common input was generated using(8)
where is a random phase in , and is the total length of the stimulus. The exact properties of this stimulus (size of the window , the cutoff frequency of 1 Hz) were not important, as long as the correlation timescale of this stimulus was much longer than the correlation timescale of the fast fluctuations ( ms) independently injected into each neuron.
Instead of , neurons in deeper layers (beyond the first) received synaptic input from neurons in the previous layer via conductance-based synapses. In contrast to current-based synapses, conductance-based synapses have been shown to support the stable propagation of synfire chains [38] and a larger range of firing rates [37]. The synaptic input current into a neuron in layer in the network (which receives inputs from a subset of neurons in the previous layers) is given by(9)
where mV is the excitatory reversal potential and is the membrane potential of the neuron. The synaptic conductance is a continuous variable which increases with the spike times of each input by the excitatory postsynaptic potential (EPSP) scaled by the corresponding synaptic strength . We used exponentially decaying EPSPs with a time constant ms. Then we can write the synaptic conductance as(10)
where is the delta spike train of the -th neuron in the previous layer with spikes at times and when is the EPSP. denotes a random subset of the 2000 neurons in the previous layer providing synaptic input into the given neuron.
There were no recurrent connections among the neurons. Each neuron in the network also received an independent noise input with mean 0 and standard deviation that fluctuates on a timescale significantly shorter than the timescale of the common input to represent random synaptic input that cortical networks experience during early development [28]. In all models, the noise stimulus added to each neuron was independent from the mean stimulus and correlated with a correlation time of 1 ms. Note that for the mean field analysis (see below), simulations were performed with a constant mean (Figure 6), rather than the time-dependent (Equation 8). The range of stimulus standard deviations was chosen to produce firing rates larger than 3 Hz and such that voltages were not hyperpolarized below mV to match the corresponding experiments [8].
Given an input current , the output firing rate can be expressed by the -dependent – curve: . We computed the – curves for the GS and NGS neurons for a range of mean inputs and fluctuation amplitudes (Figure 1C,D) from 100 second long simulations. The mean current ranged from 0 to 120 pA in steps of 2.5 pA and the standard deviation from 5 to 150 pA in steps of 2.5 pA.
The mean field analysis was used to predict firing rate transmission across the network (Figure 6). Given the synaptic current into a neuron in layer in the network (which receives inputs from a subset of neurons in the previous layers connected with weights of strengths ), the average synaptic current received by a neuron in one layer from a subset (or all) of neurons in the previous layer can be written as:(11)
where the angle brackets denote average over time. In the limit that and are uncorrelated, then(12)
The average synaptic conductance can be written as(13)
where is the average firing rate of neuron . We let denote the connection probability between neurons in two consecutive layers; therefore, the subset has approximately neurons. We examined connectivity probability ranging between 0.5%, 5% and 10% while keeping the product of the connectivity probability and synaptic strength fixed, and observed no differences in how effectively firing rates were propagated across different layers in the network (Figure S2). The main results use . For the two network types, we chose synaptic strength sufficiently strong to allow for activity to be maintained in each network. For the NGS network we used , while for the GS network we explored in addition weaker synaptic strength of ; although the exact values used were not too important as long as the iterated map dynamics predicting the mean firing rates across the network had the same structure (for example, number of fixed points) (Figure S2). Since all synapses in our network are identical to , we can approximate ; similarly, all the neurons in a given layer are identical so . Then the average synaptic current into a neuron in a given layer can be approximated as(14)
From the – relationship, the firing rate in layer can be expressed as a function of the firing rate of the neurons in the previous layer (see Eq. 1) where the scaling coefficient is given by(15)
When computing we used only subthreshold voltage fluctuations.
The input-output relationship plotted in Figure 6 (black line) corresponds to the line of slope . We also computed the standard deviation of the subthreshold voltage fluctuations and thus estimated where was obtained using Equation 15 with instead of . Figure 6 text shows this as a gray boundary around the line with slope , which was used further to interpret the variability of propagation of firing rates.
Furthermore, we note that when predicting the propagation of firing rates across subsequent layers in this mean field analysis, the – curve in Equation 1 was chosen such that was obtained by combining the standard deviation of the independent noise fluctuations added in each layer , and the standard deviation of the synaptic current recorded in each layer , where(16)
We first measured information transmission in the network about slow variations in the input (Figure 4). The mutual information of stimulus and response was computed by testing a particular encoding model (Figure 4D). Typically, this method assumes a model for estimating the stimulus and provides a lower bound on the information transfer because the model does not capture all aspects of the information [77]. We chose the stimulus reconstruction to be a simple population average of the neuronal response (the PSTH), so that the stimulus estimate in layer , , is given by the mean neuronal response obtained from many repetitions of the identical slow stimulus, but different realizations of the fast fluctuations. We computed the information in the -th layer using the equation for a dynamic Gaussian channel [24](17)
where the signal-to-noise ratio can be written as(18)
Assuming Gaussian probability distributions, the noise is(19)
This quantity computes the information between stimulus and response by taking into account how similar the response (reconstructed stimulus) is to the original stimulus. Due to the different firing rates evoked in the different networks, when computing the information we normalized the reconstructed stimulus (the PSTH) to have zero mean and unit variance.
To quantify the information about fast fluctuations as a function of the mean and of the input current injected into single neurons (Figure 1E), we used the output entropy of the predicted firing rate probability in the LN model, , using the nonlinear response function expression from Equation 7.
When examining the fidelity of firing rate transfer in networks composed of the two neuron types, we wanted a measure of how distinguishable is a discrete set of output firing rates in each layer given a set of input currents in the first layer (see Figure 7, note that Figure 1F is like the data in Figure 7 layer 1). This was the information conveyed by the network response of each layer about a stationary mean input , in the presence of background noise (Figure 7). We obtained the firing rate response of strongly connected NGS and GS networks (synaptic strength ) and weakly connected GS networks () for different layers, noise conditions and ranges of input. For the strongly connected NGS and GS networks, we used a range of 28 input currents uniformly distributed between 0 and 70 pA, and for the weakly connected GS networks, the same number of input currents uniformly distributed in the range of 0 to 22 pA. The noise values that we examined spanned the range of from 15 to 75 pA–which produced biologically relevant output firing rates and subthreshold voltage fluctuations in a valid regime mV. The output firing rates were obtained using 2 second long bins (total length of the trial was 20,000 seconds). Qualitative trends in the information curves were maintained for 1, 5 and 10 second long bins. Then, given the set of firing rate responses of the neurons of the -th layer for the input currents, we constructed by computing histograms of the output firing rates binned into the same 28 bins. We computed the mutual information for each layer(20)
where is the probability distribution of the output firing rates [77]. denotes the prior probability of input stimuli which we took to be a uniform distribution so that each stimulus had the same probability 1/28 of occurrence. Although the exact value of the information will depend on the binning choice (here into 28 bins), the contrast in performance of the GS and NGS neurons (which was our goal) was preserved for other binning choices.
To reduce the full conductance-based model (Eq. 2) that depends on four variables, , , and , to a system of two first-order differential equations, we followed the procedure described by Abbott and Kepler [35] for the Hodgkin-Huxley model. Although the neuron's membrane potential is affected by the three dynamic variables, , and , these three do not directly couple to each other but only interact through . This property allows us to approximate their dynamics by introducing an auxiliary potential variable. Since the time constant that governs the behavior for is much smaller than the time constants for and , then will reach its asymptotic value more rapidly than other changes in the model. Therefore, we lose some accuracy in the generation of spikes, but can write . Because of their longer time constants, and lag behind and reach their asymptotic values more slowly. This can be implemented by introducing an auxiliary voltage variable and then replacing and by and , since the functions and are well separated as a function of the dependent variable, in this case . To choose , we ask for the time dependence of in and the time dependence that the slowly changing and induce into in the full model to match – this is achieved by equating the time derivatives of at constant in the full and reduced models. Hence, we convert the full model (Eq. 2) into the following system of first-order differential equations:(21)(22)
where(23)
and where(24)
where and are evaluated at and .
To study the dynamics of this system in Figure 8, we plotted the nullclines, i.e. the curves where and . The points where these two curves intersect are the fixed points of the two-dimensional dynamics. In Figure 8 we use arrows in the phase planes to denote the flows around the nullclines.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.